Commit | Line | Data |
---|---|---|
a497ee34 | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
aee69d78 PV |
2 | /* |
3 | * Budget Fair Queueing (BFQ) I/O scheduler. | |
4 | * | |
5 | * Based on ideas and code from CFQ: | |
6 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> | |
7 | * | |
8 | * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it> | |
9 | * Paolo Valente <paolo.valente@unimore.it> | |
10 | * | |
11 | * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it> | |
12 | * Arianna Avanzini <avanzini@google.com> | |
13 | * | |
14 | * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org> | |
15 | * | |
aee69d78 PV |
16 | * BFQ is a proportional-share I/O scheduler, with some extra |
17 | * low-latency capabilities. BFQ also supports full hierarchical | |
18 | * scheduling through cgroups. Next paragraphs provide an introduction | |
19 | * on BFQ inner workings. Details on BFQ benefits, usage and | |
898bd37a | 20 | * limitations can be found in Documentation/block/bfq-iosched.rst. |
aee69d78 PV |
21 | * |
22 | * BFQ is a proportional-share storage-I/O scheduling algorithm based | |
23 | * on the slice-by-slice service scheme of CFQ. But BFQ assigns | |
24 | * budgets, measured in number of sectors, to processes instead of | |
25 | * time slices. The device is not granted to the in-service process | |
26 | * for a given time slice, but until it has exhausted its assigned | |
27 | * budget. This change from the time to the service domain enables BFQ | |
28 | * to distribute the device throughput among processes as desired, | |
29 | * without any distortion due to throughput fluctuations, or to device | |
30 | * internal queueing. BFQ uses an ad hoc internal scheduler, called | |
31 | * B-WF2Q+, to schedule processes according to their budgets. More | |
32 | * precisely, BFQ schedules queues associated with processes. Each | |
33 | * process/queue is assigned a user-configurable weight, and B-WF2Q+ | |
34 | * guarantees that each queue receives a fraction of the throughput | |
35 | * proportional to its weight. Thanks to the accurate policy of | |
36 | * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound | |
37 | * processes issuing sequential requests (to boost the throughput), | |
38 | * and yet guarantee a low latency to interactive and soft real-time | |
39 | * applications. | |
40 | * | |
41 | * In particular, to provide these low-latency guarantees, BFQ | |
42 | * explicitly privileges the I/O of two classes of time-sensitive | |
4029eef1 PV |
43 | * applications: interactive and soft real-time. In more detail, BFQ |
44 | * behaves this way if the low_latency parameter is set (default | |
45 | * configuration). This feature enables BFQ to provide applications in | |
46 | * these classes with a very low latency. | |
47 | * | |
48 | * To implement this feature, BFQ constantly tries to detect whether | |
49 | * the I/O requests in a bfq_queue come from an interactive or a soft | |
50 | * real-time application. For brevity, in these cases, the queue is | |
51 | * said to be interactive or soft real-time. In both cases, BFQ | |
52 | * privileges the service of the queue, over that of non-interactive | |
53 | * and non-soft-real-time queues. This privileging is performed, | |
54 | * mainly, by raising the weight of the queue. So, for brevity, we | |
55 | * call just weight-raising periods the time periods during which a | |
56 | * queue is privileged, because deemed interactive or soft real-time. | |
57 | * | |
58 | * The detection of soft real-time queues/applications is described in | |
59 | * detail in the comments on the function | |
60 | * bfq_bfqq_softrt_next_start. On the other hand, the detection of an | |
61 | * interactive queue works as follows: a queue is deemed interactive | |
62 | * if it is constantly non empty only for a limited time interval, | |
63 | * after which it does become empty. The queue may be deemed | |
64 | * interactive again (for a limited time), if it restarts being | |
65 | * constantly non empty, provided that this happens only after the | |
66 | * queue has remained empty for a given minimum idle time. | |
67 | * | |
68 | * By default, BFQ computes automatically the above maximum time | |
69 | * interval, i.e., the time interval after which a constantly | |
70 | * non-empty queue stops being deemed interactive. Since a queue is | |
71 | * weight-raised while it is deemed interactive, this maximum time | |
72 | * interval happens to coincide with the (maximum) duration of the | |
73 | * weight-raising for interactive queues. | |
74 | * | |
75 | * Finally, BFQ also features additional heuristics for | |
aee69d78 PV |
76 | * preserving both a low latency and a high throughput on NCQ-capable, |
77 | * rotational or flash-based devices, and to get the job done quickly | |
78 | * for applications consisting in many I/O-bound processes. | |
79 | * | |
43c1b3d6 PV |
80 | * NOTE: if the main or only goal, with a given device, is to achieve |
81 | * the maximum-possible throughput at all times, then do switch off | |
82 | * all low-latency heuristics for that device, by setting low_latency | |
83 | * to 0. | |
84 | * | |
4029eef1 PV |
85 | * BFQ is described in [1], where also a reference to the initial, |
86 | * more theoretical paper on BFQ can be found. The interested reader | |
87 | * can find in the latter paper full details on the main algorithm, as | |
88 | * well as formulas of the guarantees and formal proofs of all the | |
89 | * properties. With respect to the version of BFQ presented in these | |
90 | * papers, this implementation adds a few more heuristics, such as the | |
91 | * ones that guarantee a low latency to interactive and soft real-time | |
92 | * applications, and a hierarchical extension based on H-WF2Q+. | |
aee69d78 PV |
93 | * |
94 | * B-WF2Q+ is based on WF2Q+, which is described in [2], together with | |
95 | * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+ | |
96 | * with O(log N) complexity derives from the one introduced with EEVDF | |
97 | * in [3]. | |
98 | * | |
99 | * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O | |
100 | * Scheduler", Proceedings of the First Workshop on Mobile System | |
101 | * Technologies (MST-2015), May 2015. | |
102 | * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf | |
103 | * | |
104 | * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing | |
105 | * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689, | |
106 | * Oct 1997. | |
107 | * | |
108 | * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz | |
109 | * | |
110 | * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline | |
111 | * First: A Flexible and Accurate Mechanism for Proportional Share | |
112 | * Resource Allocation", technical report. | |
113 | * | |
114 | * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf | |
115 | */ | |
116 | #include <linux/module.h> | |
117 | #include <linux/slab.h> | |
118 | #include <linux/blkdev.h> | |
e21b7a0b | 119 | #include <linux/cgroup.h> |
aee69d78 PV |
120 | #include <linux/ktime.h> |
121 | #include <linux/rbtree.h> | |
122 | #include <linux/ioprio.h> | |
123 | #include <linux/sbitmap.h> | |
124 | #include <linux/delay.h> | |
d51cfc53 | 125 | #include <linux/backing-dev.h> |
aee69d78 | 126 | |
b357e4a6 CK |
127 | #include <trace/events/block.h> |
128 | ||
2e9bc346 | 129 | #include "elevator.h" |
aee69d78 PV |
130 | #include "blk.h" |
131 | #include "blk-mq.h" | |
aee69d78 | 132 | #include "blk-mq-sched.h" |
ea25da48 | 133 | #include "bfq-iosched.h" |
b5dc5d4d | 134 | #include "blk-wbt.h" |
aee69d78 | 135 | |
ea25da48 PV |
136 | #define BFQ_BFQQ_FNS(name) \ |
137 | void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ | |
138 | { \ | |
139 | __set_bit(BFQQF_##name, &(bfqq)->flags); \ | |
140 | } \ | |
141 | void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ | |
142 | { \ | |
143 | __clear_bit(BFQQF_##name, &(bfqq)->flags); \ | |
144 | } \ | |
145 | int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ | |
146 | { \ | |
147 | return test_bit(BFQQF_##name, &(bfqq)->flags); \ | |
44e44a1b PV |
148 | } |
149 | ||
ea25da48 PV |
150 | BFQ_BFQQ_FNS(just_created); |
151 | BFQ_BFQQ_FNS(busy); | |
152 | BFQ_BFQQ_FNS(wait_request); | |
153 | BFQ_BFQQ_FNS(non_blocking_wait_rq); | |
154 | BFQ_BFQQ_FNS(fifo_expire); | |
d5be3fef | 155 | BFQ_BFQQ_FNS(has_short_ttime); |
ea25da48 PV |
156 | BFQ_BFQQ_FNS(sync); |
157 | BFQ_BFQQ_FNS(IO_bound); | |
158 | BFQ_BFQQ_FNS(in_large_burst); | |
159 | BFQ_BFQQ_FNS(coop); | |
160 | BFQ_BFQQ_FNS(split_coop); | |
161 | BFQ_BFQQ_FNS(softrt_update); | |
162 | #undef BFQ_BFQQ_FNS \ | |
aee69d78 | 163 | |
4168a8d2 | 164 | /* Expiration time of async (0) and sync (1) requests, in ns. */ |
ea25da48 | 165 | static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 }; |
aee69d78 | 166 | |
ea25da48 PV |
167 | /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */ |
168 | static const int bfq_back_max = 16 * 1024; | |
aee69d78 | 169 | |
ea25da48 PV |
170 | /* Penalty of a backwards seek, in number of sectors. */ |
171 | static const int bfq_back_penalty = 2; | |
e21b7a0b | 172 | |
ea25da48 PV |
173 | /* Idling period duration, in ns. */ |
174 | static u64 bfq_slice_idle = NSEC_PER_SEC / 125; | |
aee69d78 | 175 | |
ea25da48 PV |
176 | /* Minimum number of assigned budgets for which stats are safe to compute. */ |
177 | static const int bfq_stats_min_budgets = 194; | |
aee69d78 | 178 | |
ea25da48 PV |
179 | /* Default maximum budget values, in sectors and number of requests. */ |
180 | static const int bfq_default_max_budget = 16 * 1024; | |
e21b7a0b | 181 | |
ea25da48 | 182 | /* |
d5801088 PV |
183 | * When a sync request is dispatched, the queue that contains that |
184 | * request, and all the ancestor entities of that queue, are charged | |
636b8fe8 | 185 | * with the number of sectors of the request. In contrast, if the |
d5801088 PV |
186 | * request is async, then the queue and its ancestor entities are |
187 | * charged with the number of sectors of the request, multiplied by | |
188 | * the factor below. This throttles the bandwidth for async I/O, | |
189 | * w.r.t. to sync I/O, and it is done to counter the tendency of async | |
190 | * writes to steal I/O throughput to reads. | |
191 | * | |
192 | * The current value of this parameter is the result of a tuning with | |
193 | * several hardware and software configurations. We tried to find the | |
194 | * lowest value for which writes do not cause noticeable problems to | |
195 | * reads. In fact, the lower this parameter, the stabler I/O control, | |
196 | * in the following respect. The lower this parameter is, the less | |
197 | * the bandwidth enjoyed by a group decreases | |
198 | * - when the group does writes, w.r.t. to when it does reads; | |
199 | * - when other groups do reads, w.r.t. to when they do writes. | |
ea25da48 | 200 | */ |
d5801088 | 201 | static const int bfq_async_charge_factor = 3; |
aee69d78 | 202 | |
ea25da48 PV |
203 | /* Default timeout values, in jiffies, approximating CFQ defaults. */ |
204 | const int bfq_timeout = HZ / 8; | |
aee69d78 | 205 | |
7b8fa3b9 PV |
206 | /* |
207 | * Time limit for merging (see comments in bfq_setup_cooperator). Set | |
208 | * to the slowest value that, in our tests, proved to be effective in | |
209 | * removing false positives, while not causing true positives to miss | |
210 | * queue merging. | |
211 | * | |
212 | * As can be deduced from the low time limit below, queue merging, if | |
636b8fe8 | 213 | * successful, happens at the very beginning of the I/O of the involved |
7b8fa3b9 PV |
214 | * cooperating processes, as a consequence of the arrival of the very |
215 | * first requests from each cooperator. After that, there is very | |
216 | * little chance to find cooperators. | |
217 | */ | |
218 | static const unsigned long bfq_merge_time_limit = HZ/10; | |
219 | ||
ea25da48 | 220 | static struct kmem_cache *bfq_pool; |
e21b7a0b | 221 | |
ea25da48 PV |
222 | /* Below this threshold (in ns), we consider thinktime immediate. */ |
223 | #define BFQ_MIN_TT (2 * NSEC_PER_MSEC) | |
e21b7a0b | 224 | |
ea25da48 | 225 | /* hw_tag detection: parallel requests threshold and min samples needed. */ |
a3c92560 | 226 | #define BFQ_HW_QUEUE_THRESHOLD 3 |
ea25da48 | 227 | #define BFQ_HW_QUEUE_SAMPLES 32 |
aee69d78 | 228 | |
ea25da48 PV |
229 | #define BFQQ_SEEK_THR (sector_t)(8 * 100) |
230 | #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32) | |
d87447d8 PV |
231 | #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \ |
232 | (get_sdist(last_pos, rq) > \ | |
233 | BFQQ_SEEK_THR && \ | |
234 | (!blk_queue_nonrot(bfqd->queue) || \ | |
235 | blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT)) | |
ea25da48 | 236 | #define BFQQ_CLOSE_THR (sector_t)(8 * 1024) |
f0ba5ea2 | 237 | #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19) |
7074f076 PV |
238 | /* |
239 | * Sync random I/O is likely to be confused with soft real-time I/O, | |
240 | * because it is characterized by limited throughput and apparently | |
241 | * isochronous arrival pattern. To avoid false positives, queues | |
242 | * containing only random (seeky) I/O are prevented from being tagged | |
243 | * as soft real-time. | |
244 | */ | |
e6feaf21 | 245 | #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1) |
aee69d78 | 246 | |
ea25da48 PV |
247 | /* Min number of samples required to perform peak-rate update */ |
248 | #define BFQ_RATE_MIN_SAMPLES 32 | |
249 | /* Min observation time interval required to perform a peak-rate update (ns) */ | |
250 | #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC) | |
251 | /* Target observation time interval for a peak-rate update (ns) */ | |
252 | #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC | |
aee69d78 | 253 | |
bc56e2ca PV |
254 | /* |
255 | * Shift used for peak-rate fixed precision calculations. | |
256 | * With | |
257 | * - the current shift: 16 positions | |
258 | * - the current type used to store rate: u32 | |
259 | * - the current unit of measure for rate: [sectors/usec], or, more precisely, | |
260 | * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift, | |
261 | * the range of rates that can be stored is | |
262 | * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec = | |
263 | * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec = | |
264 | * [15, 65G] sectors/sec | |
265 | * Which, assuming a sector size of 512B, corresponds to a range of | |
266 | * [7.5K, 33T] B/sec | |
267 | */ | |
ea25da48 | 268 | #define BFQ_RATE_SHIFT 16 |
aee69d78 | 269 | |
ea25da48 | 270 | /* |
4029eef1 PV |
271 | * When configured for computing the duration of the weight-raising |
272 | * for interactive queues automatically (see the comments at the | |
273 | * beginning of this file), BFQ does it using the following formula: | |
e24f1c24 PV |
274 | * duration = (ref_rate / r) * ref_wr_duration, |
275 | * where r is the peak rate of the device, and ref_rate and | |
276 | * ref_wr_duration are two reference parameters. In particular, | |
277 | * ref_rate is the peak rate of the reference storage device (see | |
278 | * below), and ref_wr_duration is about the maximum time needed, with | |
279 | * BFQ and while reading two files in parallel, to load typical large | |
280 | * applications on the reference device (see the comments on | |
281 | * max_service_from_wr below, for more details on how ref_wr_duration | |
282 | * is obtained). In practice, the slower/faster the device at hand | |
283 | * is, the more/less it takes to load applications with respect to the | |
4029eef1 PV |
284 | * reference device. Accordingly, the longer/shorter BFQ grants |
285 | * weight raising to interactive applications. | |
ea25da48 | 286 | * |
e24f1c24 PV |
287 | * BFQ uses two different reference pairs (ref_rate, ref_wr_duration), |
288 | * depending on whether the device is rotational or non-rotational. | |
ea25da48 | 289 | * |
e24f1c24 PV |
290 | * In the following definitions, ref_rate[0] and ref_wr_duration[0] |
291 | * are the reference values for a rotational device, whereas | |
292 | * ref_rate[1] and ref_wr_duration[1] are the reference values for a | |
293 | * non-rotational device. The reference rates are not the actual peak | |
294 | * rates of the devices used as a reference, but slightly lower | |
295 | * values. The reason for using slightly lower values is that the | |
296 | * peak-rate estimator tends to yield slightly lower values than the | |
297 | * actual peak rate (it can yield the actual peak rate only if there | |
298 | * is only one process doing I/O, and the process does sequential | |
299 | * I/O). | |
ea25da48 | 300 | * |
e24f1c24 PV |
301 | * The reference peak rates are measured in sectors/usec, left-shifted |
302 | * by BFQ_RATE_SHIFT. | |
ea25da48 | 303 | */ |
e24f1c24 | 304 | static int ref_rate[2] = {14000, 33000}; |
ea25da48 | 305 | /* |
e24f1c24 PV |
306 | * To improve readability, a conversion function is used to initialize |
307 | * the following array, which entails that the array can be | |
308 | * initialized only in a function. | |
ea25da48 | 309 | */ |
e24f1c24 | 310 | static int ref_wr_duration[2]; |
aee69d78 | 311 | |
8a8747dc PV |
312 | /* |
313 | * BFQ uses the above-detailed, time-based weight-raising mechanism to | |
314 | * privilege interactive tasks. This mechanism is vulnerable to the | |
315 | * following false positives: I/O-bound applications that will go on | |
316 | * doing I/O for much longer than the duration of weight | |
317 | * raising. These applications have basically no benefit from being | |
318 | * weight-raised at the beginning of their I/O. On the opposite end, | |
319 | * while being weight-raised, these applications | |
320 | * a) unjustly steal throughput to applications that may actually need | |
321 | * low latency; | |
322 | * b) make BFQ uselessly perform device idling; device idling results | |
323 | * in loss of device throughput with most flash-based storage, and may | |
324 | * increase latencies when used purposelessly. | |
325 | * | |
326 | * BFQ tries to reduce these problems, by adopting the following | |
327 | * countermeasure. To introduce this countermeasure, we need first to | |
328 | * finish explaining how the duration of weight-raising for | |
329 | * interactive tasks is computed. | |
330 | * | |
331 | * For a bfq_queue deemed as interactive, the duration of weight | |
332 | * raising is dynamically adjusted, as a function of the estimated | |
333 | * peak rate of the device, so as to be equal to the time needed to | |
334 | * execute the 'largest' interactive task we benchmarked so far. By | |
335 | * largest task, we mean the task for which each involved process has | |
336 | * to do more I/O than for any of the other tasks we benchmarked. This | |
337 | * reference interactive task is the start-up of LibreOffice Writer, | |
338 | * and in this task each process/bfq_queue needs to have at most ~110K | |
339 | * sectors transferred. | |
340 | * | |
341 | * This last piece of information enables BFQ to reduce the actual | |
342 | * duration of weight-raising for at least one class of I/O-bound | |
343 | * applications: those doing sequential or quasi-sequential I/O. An | |
344 | * example is file copy. In fact, once started, the main I/O-bound | |
345 | * processes of these applications usually consume the above 110K | |
346 | * sectors in much less time than the processes of an application that | |
347 | * is starting, because these I/O-bound processes will greedily devote | |
348 | * almost all their CPU cycles only to their target, | |
349 | * throughput-friendly I/O operations. This is even more true if BFQ | |
350 | * happens to be underestimating the device peak rate, and thus | |
351 | * overestimating the duration of weight raising. But, according to | |
352 | * our measurements, once transferred 110K sectors, these processes | |
353 | * have no right to be weight-raised any longer. | |
354 | * | |
355 | * Basing on the last consideration, BFQ ends weight-raising for a | |
356 | * bfq_queue if the latter happens to have received an amount of | |
357 | * service at least equal to the following constant. The constant is | |
358 | * set to slightly more than 110K, to have a minimum safety margin. | |
359 | * | |
360 | * This early ending of weight-raising reduces the amount of time | |
361 | * during which interactive false positives cause the two problems | |
362 | * described at the beginning of these comments. | |
363 | */ | |
364 | static const unsigned long max_service_from_wr = 120000; | |
365 | ||
7812472f PP |
366 | /* |
367 | * Maximum time between the creation of two queues, for stable merge | |
368 | * to be activated (in ms) | |
369 | */ | |
370 | static const unsigned long bfq_activation_stable_merging = 600; | |
371 | /* | |
372 | * Minimum time to be waited before evaluating delayed stable merge (in ms) | |
373 | */ | |
374 | static const unsigned long bfq_late_stable_merging = 600; | |
375 | ||
e79cf889 | 376 | #define RQ_BIC(rq) ((struct bfq_io_cq *)((rq)->elv.priv[0])) |
ea25da48 | 377 | #define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
aee69d78 | 378 | |
9778369a PV |
379 | struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync, |
380 | unsigned int actuator_idx) | |
e21b7a0b | 381 | { |
9778369a PV |
382 | if (is_sync) |
383 | return bic->bfqq[1][actuator_idx]; | |
384 | ||
385 | return bic->bfqq[0][actuator_idx]; | |
aee69d78 PV |
386 | } |
387 | ||
7ea96eef PV |
388 | static void bfq_put_stable_ref(struct bfq_queue *bfqq); |
389 | ||
9778369a PV |
390 | void bic_set_bfqq(struct bfq_io_cq *bic, |
391 | struct bfq_queue *bfqq, | |
392 | bool is_sync, | |
393 | unsigned int actuator_idx) | |
aee69d78 | 394 | { |
9778369a | 395 | struct bfq_queue *old_bfqq = bic->bfqq[is_sync][actuator_idx]; |
64dc8c73 | 396 | |
7ea96eef PV |
397 | /* |
398 | * If bfqq != NULL, then a non-stable queue merge between | |
399 | * bic->bfqq and bfqq is happening here. This causes troubles | |
400 | * in the following case: bic->bfqq has also been scheduled | |
401 | * for a possible stable merge with bic->stable_merge_bfqq, | |
402 | * and bic->stable_merge_bfqq == bfqq happens to | |
403 | * hold. Troubles occur because bfqq may then undergo a split, | |
404 | * thereby becoming eligible for a stable merge. Yet, if | |
405 | * bic->stable_merge_bfqq points exactly to bfqq, then bfqq | |
406 | * would be stably merged with itself. To avoid this anomaly, | |
407 | * we cancel the stable merge if | |
408 | * bic->stable_merge_bfqq == bfqq. | |
409 | */ | |
fd571df0 | 410 | struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[actuator_idx]; |
a6123047 PV |
411 | |
412 | /* Clear bic pointer if bfqq is detached from this bic */ | |
413 | if (old_bfqq && old_bfqq->bic == bic) | |
414 | old_bfqq->bic = NULL; | |
7ea96eef | 415 | |
9778369a PV |
416 | if (is_sync) |
417 | bic->bfqq[1][actuator_idx] = bfqq; | |
418 | else | |
419 | bic->bfqq[0][actuator_idx] = bfqq; | |
7ea96eef | 420 | |
a6123047 | 421 | if (bfqq && bfqq_data->stable_merge_bfqq == bfqq) { |
7ea96eef PV |
422 | /* |
423 | * Actually, these same instructions are executed also | |
424 | * in bfq_setup_cooperator, in case of abort or actual | |
425 | * execution of a stable merge. We could avoid | |
426 | * repeating these instructions there too, but if we | |
427 | * did so, we would nest even more complexity in this | |
428 | * function. | |
429 | */ | |
a6123047 | 430 | bfq_put_stable_ref(bfqq_data->stable_merge_bfqq); |
7ea96eef | 431 | |
a6123047 | 432 | bfqq_data->stable_merge_bfqq = NULL; |
7ea96eef | 433 | } |
aee69d78 PV |
434 | } |
435 | ||
ea25da48 | 436 | struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
aee69d78 | 437 | { |
ea25da48 | 438 | return bic->icq.q->elevator->elevator_data; |
e21b7a0b | 439 | } |
aee69d78 | 440 | |
ea25da48 PV |
441 | /** |
442 | * icq_to_bic - convert iocontext queue structure to bfq_io_cq. | |
443 | * @icq: the iocontext queue. | |
444 | */ | |
445 | static struct bfq_io_cq *icq_to_bic(struct io_cq *icq) | |
e21b7a0b | 446 | { |
ea25da48 PV |
447 | /* bic->icq is the first member, %NULL will convert to %NULL */ |
448 | return container_of(icq, struct bfq_io_cq, icq); | |
e21b7a0b | 449 | } |
aee69d78 | 450 | |
ea25da48 PV |
451 | /** |
452 | * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. | |
ea25da48 PV |
453 | * @q: the request queue. |
454 | */ | |
836b394b | 455 | static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q) |
e21b7a0b | 456 | { |
836b394b CH |
457 | struct bfq_io_cq *icq; |
458 | unsigned long flags; | |
aee69d78 | 459 | |
836b394b CH |
460 | if (!current->io_context) |
461 | return NULL; | |
aee69d78 | 462 | |
836b394b | 463 | spin_lock_irqsave(&q->queue_lock, flags); |
eca5892a | 464 | icq = icq_to_bic(ioc_lookup_icq(q)); |
836b394b | 465 | spin_unlock_irqrestore(&q->queue_lock, flags); |
e21b7a0b | 466 | |
836b394b | 467 | return icq; |
aee69d78 PV |
468 | } |
469 | ||
ea25da48 PV |
470 | /* |
471 | * Scheduler run of queue, if there are requests pending and no one in the | |
472 | * driver that will restart queueing. | |
473 | */ | |
474 | void bfq_schedule_dispatch(struct bfq_data *bfqd) | |
aee69d78 | 475 | { |
181490d5 YK |
476 | lockdep_assert_held(&bfqd->lock); |
477 | ||
ea25da48 PV |
478 | if (bfqd->queued != 0) { |
479 | bfq_log(bfqd, "schedule dispatch"); | |
480 | blk_mq_run_hw_queues(bfqd->queue, true); | |
e21b7a0b | 481 | } |
aee69d78 PV |
482 | } |
483 | ||
484 | #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
aee69d78 PV |
485 | |
486 | #define bfq_sample_valid(samples) ((samples) > 80) | |
487 | ||
aee69d78 PV |
488 | /* |
489 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
636b8fe8 | 490 | * We choose the request that is closer to the head right now. Distance |
aee69d78 PV |
491 | * behind the head is penalized and only allowed to a certain extent. |
492 | */ | |
493 | static struct request *bfq_choose_req(struct bfq_data *bfqd, | |
494 | struct request *rq1, | |
495 | struct request *rq2, | |
496 | sector_t last) | |
497 | { | |
498 | sector_t s1, s2, d1 = 0, d2 = 0; | |
499 | unsigned long back_max; | |
500 | #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
501 | #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
502 | unsigned int wrap = 0; /* bit mask: requests behind the disk head? */ | |
503 | ||
504 | if (!rq1 || rq1 == rq2) | |
505 | return rq2; | |
506 | if (!rq2) | |
507 | return rq1; | |
508 | ||
509 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) | |
510 | return rq1; | |
511 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
512 | return rq2; | |
513 | if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) | |
514 | return rq1; | |
515 | else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) | |
516 | return rq2; | |
517 | ||
518 | s1 = blk_rq_pos(rq1); | |
519 | s2 = blk_rq_pos(rq2); | |
520 | ||
521 | /* | |
522 | * By definition, 1KiB is 2 sectors. | |
523 | */ | |
524 | back_max = bfqd->bfq_back_max * 2; | |
525 | ||
526 | /* | |
527 | * Strict one way elevator _except_ in the case where we allow | |
528 | * short backward seeks which are biased as twice the cost of a | |
529 | * similar forward seek. | |
530 | */ | |
531 | if (s1 >= last) | |
532 | d1 = s1 - last; | |
533 | else if (s1 + back_max >= last) | |
534 | d1 = (last - s1) * bfqd->bfq_back_penalty; | |
535 | else | |
536 | wrap |= BFQ_RQ1_WRAP; | |
537 | ||
538 | if (s2 >= last) | |
539 | d2 = s2 - last; | |
540 | else if (s2 + back_max >= last) | |
541 | d2 = (last - s2) * bfqd->bfq_back_penalty; | |
542 | else | |
543 | wrap |= BFQ_RQ2_WRAP; | |
544 | ||
545 | /* Found required data */ | |
546 | ||
547 | /* | |
548 | * By doing switch() on the bit mask "wrap" we avoid having to | |
549 | * check two variables for all permutations: --> faster! | |
550 | */ | |
551 | switch (wrap) { | |
552 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
553 | if (d1 < d2) | |
554 | return rq1; | |
555 | else if (d2 < d1) | |
556 | return rq2; | |
557 | ||
558 | if (s1 >= s2) | |
559 | return rq1; | |
560 | else | |
561 | return rq2; | |
562 | ||
563 | case BFQ_RQ2_WRAP: | |
564 | return rq1; | |
565 | case BFQ_RQ1_WRAP: | |
566 | return rq2; | |
567 | case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */ | |
568 | default: | |
569 | /* | |
570 | * Since both rqs are wrapped, | |
571 | * start with the one that's further behind head | |
572 | * (--> only *one* back seek required), | |
573 | * since back seek takes more time than forward. | |
574 | */ | |
575 | if (s1 <= s2) | |
576 | return rq1; | |
577 | else | |
578 | return rq2; | |
579 | } | |
580 | } | |
581 | ||
76f1df88 JK |
582 | #define BFQ_LIMIT_INLINE_DEPTH 16 |
583 | ||
584 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
585 | static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit) | |
586 | { | |
587 | struct bfq_data *bfqd = bfqq->bfqd; | |
588 | struct bfq_entity *entity = &bfqq->entity; | |
589 | struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH]; | |
590 | struct bfq_entity **entities = inline_entities; | |
09df6a75 | 591 | int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH; |
76f1df88 JK |
592 | int class_idx = bfqq->ioprio_class - 1; |
593 | struct bfq_sched_data *sched_data; | |
594 | unsigned long wsum; | |
595 | bool ret = false; | |
596 | ||
597 | if (!entity->on_st_or_in_serv) | |
598 | return false; | |
599 | ||
09df6a75 JK |
600 | retry: |
601 | spin_lock_irq(&bfqd->lock); | |
76f1df88 JK |
602 | /* +1 for bfqq entity, root cgroup not included */ |
603 | depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1; | |
09df6a75 JK |
604 | if (depth > alloc_depth) { |
605 | spin_unlock_irq(&bfqd->lock); | |
606 | if (entities != inline_entities) | |
607 | kfree(entities); | |
76f1df88 JK |
608 | entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO); |
609 | if (!entities) | |
610 | return false; | |
09df6a75 JK |
611 | alloc_depth = depth; |
612 | goto retry; | |
76f1df88 JK |
613 | } |
614 | ||
76f1df88 JK |
615 | sched_data = entity->sched_data; |
616 | /* Gather our ancestors as we need to traverse them in reverse order */ | |
617 | level = 0; | |
618 | for_each_entity(entity) { | |
619 | /* | |
620 | * If at some level entity is not even active, allow request | |
621 | * queueing so that BFQ knows there's work to do and activate | |
622 | * entities. | |
623 | */ | |
624 | if (!entity->on_st_or_in_serv) | |
625 | goto out; | |
626 | /* Uh, more parents than cgroup subsystem thinks? */ | |
627 | if (WARN_ON_ONCE(level >= depth)) | |
628 | break; | |
629 | entities[level++] = entity; | |
630 | } | |
631 | WARN_ON_ONCE(level != depth); | |
632 | for (level--; level >= 0; level--) { | |
633 | entity = entities[level]; | |
634 | if (level > 0) { | |
635 | wsum = bfq_entity_service_tree(entity)->wsum; | |
636 | } else { | |
637 | int i; | |
638 | /* | |
639 | * For bfqq itself we take into account service trees | |
640 | * of all higher priority classes and multiply their | |
641 | * weights so that low prio queue from higher class | |
642 | * gets more requests than high prio queue from lower | |
643 | * class. | |
644 | */ | |
645 | wsum = 0; | |
646 | for (i = 0; i <= class_idx; i++) { | |
647 | wsum = wsum * IOPRIO_BE_NR + | |
648 | sched_data->service_tree[i].wsum; | |
649 | } | |
650 | } | |
e53413f8 CIK |
651 | if (!wsum) |
652 | continue; | |
76f1df88 JK |
653 | limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum); |
654 | if (entity->allocated >= limit) { | |
655 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
656 | "too many requests: allocated %d limit %d level %d", | |
657 | entity->allocated, limit, level); | |
658 | ret = true; | |
659 | break; | |
660 | } | |
661 | } | |
662 | out: | |
663 | spin_unlock_irq(&bfqd->lock); | |
664 | if (entities != inline_entities) | |
665 | kfree(entities); | |
666 | return ret; | |
667 | } | |
668 | #else | |
669 | static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit) | |
670 | { | |
671 | return false; | |
672 | } | |
673 | #endif | |
674 | ||
a52a69ea PV |
675 | /* |
676 | * Async I/O can easily starve sync I/O (both sync reads and sync | |
677 | * writes), by consuming all tags. Similarly, storms of sync writes, | |
678 | * such as those that sync(2) may trigger, can starve sync reads. | |
679 | * Limit depths of async I/O and sync writes so as to counter both | |
680 | * problems. | |
76f1df88 JK |
681 | * |
682 | * Also if a bfq queue or its parent cgroup consume more tags than would be | |
683 | * appropriate for their weight, we trim the available tag depth to 1. This | |
684 | * avoids a situation where one cgroup can starve another cgroup from tags and | |
685 | * thus block service differentiation among cgroups. Note that because the | |
686 | * queue / cgroup already has many requests allocated and queued, this does not | |
687 | * significantly affect service guarantees coming from the BFQ scheduling | |
688 | * algorithm. | |
a52a69ea | 689 | */ |
dc469ba2 | 690 | static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data) |
a52a69ea | 691 | { |
a52a69ea | 692 | struct bfq_data *bfqd = data->q->elevator->elevator_data; |
a0725c22 | 693 | struct bfq_io_cq *bic = bfq_bic_lookup(data->q); |
76f1df88 JK |
694 | int depth; |
695 | unsigned limit = data->q->nr_requests; | |
9778369a | 696 | unsigned int act_idx; |
76f1df88 JK |
697 | |
698 | /* Sync reads have full depth available */ | |
dc469ba2 | 699 | if (op_is_sync(opf) && !op_is_write(opf)) { |
76f1df88 JK |
700 | depth = 0; |
701 | } else { | |
dc469ba2 | 702 | depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)]; |
76f1df88 JK |
703 | limit = (limit * depth) >> bfqd->full_depth_shift; |
704 | } | |
a52a69ea | 705 | |
9778369a PV |
706 | for (act_idx = 0; bic && act_idx < bfqd->num_actuators; act_idx++) { |
707 | struct bfq_queue *bfqq = | |
708 | bic_to_bfqq(bic, op_is_sync(opf), act_idx); | |
a52a69ea | 709 | |
9778369a PV |
710 | /* |
711 | * Does queue (or any parent entity) exceed number of | |
712 | * requests that should be available to it? Heavily | |
713 | * limit depth so that it cannot consume more | |
714 | * available requests and thus starve other entities. | |
715 | */ | |
716 | if (bfqq && bfqq_request_over_limit(bfqq, limit)) { | |
717 | depth = 1; | |
718 | break; | |
719 | } | |
720 | } | |
a52a69ea | 721 | bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u", |
dc469ba2 | 722 | __func__, bfqd->wr_busy_queues, op_is_sync(opf), depth); |
76f1df88 JK |
723 | if (depth) |
724 | data->shallow_depth = depth; | |
a52a69ea PV |
725 | } |
726 | ||
36eca894 AA |
727 | static struct bfq_queue * |
728 | bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, | |
729 | sector_t sector, struct rb_node **ret_parent, | |
730 | struct rb_node ***rb_link) | |
731 | { | |
732 | struct rb_node **p, *parent; | |
733 | struct bfq_queue *bfqq = NULL; | |
734 | ||
735 | parent = NULL; | |
736 | p = &root->rb_node; | |
737 | while (*p) { | |
738 | struct rb_node **n; | |
739 | ||
740 | parent = *p; | |
741 | bfqq = rb_entry(parent, struct bfq_queue, pos_node); | |
742 | ||
743 | /* | |
744 | * Sort strictly based on sector. Smallest to the left, | |
745 | * largest to the right. | |
746 | */ | |
747 | if (sector > blk_rq_pos(bfqq->next_rq)) | |
748 | n = &(*p)->rb_right; | |
749 | else if (sector < blk_rq_pos(bfqq->next_rq)) | |
750 | n = &(*p)->rb_left; | |
751 | else | |
752 | break; | |
753 | p = n; | |
754 | bfqq = NULL; | |
755 | } | |
756 | ||
757 | *ret_parent = parent; | |
758 | if (rb_link) | |
759 | *rb_link = p; | |
760 | ||
761 | bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", | |
762 | (unsigned long long)sector, | |
763 | bfqq ? bfqq->pid : 0); | |
764 | ||
765 | return bfqq; | |
766 | } | |
767 | ||
7b8fa3b9 PV |
768 | static bool bfq_too_late_for_merging(struct bfq_queue *bfqq) |
769 | { | |
770 | return bfqq->service_from_backlogged > 0 && | |
771 | time_is_before_jiffies(bfqq->first_IO_time + | |
772 | bfq_merge_time_limit); | |
773 | } | |
774 | ||
8cacc5ab PV |
775 | /* |
776 | * The following function is not marked as __cold because it is | |
777 | * actually cold, but for the same performance goal described in the | |
778 | * comments on the likely() at the beginning of | |
779 | * bfq_setup_cooperator(). Unexpectedly, to reach an even lower | |
780 | * execution time for the case where this function is not invoked, we | |
781 | * had to add an unlikely() in each involved if(). | |
782 | */ | |
783 | void __cold | |
784 | bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
36eca894 AA |
785 | { |
786 | struct rb_node **p, *parent; | |
787 | struct bfq_queue *__bfqq; | |
788 | ||
789 | if (bfqq->pos_root) { | |
790 | rb_erase(&bfqq->pos_node, bfqq->pos_root); | |
791 | bfqq->pos_root = NULL; | |
792 | } | |
793 | ||
32c59e3a PV |
794 | /* oom_bfqq does not participate in queue merging */ |
795 | if (bfqq == &bfqd->oom_bfqq) | |
796 | return; | |
797 | ||
7b8fa3b9 PV |
798 | /* |
799 | * bfqq cannot be merged any longer (see comments in | |
800 | * bfq_setup_cooperator): no point in adding bfqq into the | |
801 | * position tree. | |
802 | */ | |
803 | if (bfq_too_late_for_merging(bfqq)) | |
804 | return; | |
805 | ||
36eca894 AA |
806 | if (bfq_class_idle(bfqq)) |
807 | return; | |
808 | if (!bfqq->next_rq) | |
809 | return; | |
810 | ||
43a4b1fe | 811 | bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree; |
36eca894 AA |
812 | __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, |
813 | blk_rq_pos(bfqq->next_rq), &parent, &p); | |
814 | if (!__bfqq) { | |
815 | rb_link_node(&bfqq->pos_node, parent, p); | |
816 | rb_insert_color(&bfqq->pos_node, bfqq->pos_root); | |
817 | } else | |
818 | bfqq->pos_root = NULL; | |
819 | } | |
820 | ||
1de0c4cd | 821 | /* |
fb53ac6c PV |
822 | * The following function returns false either if every active queue |
823 | * must receive the same share of the throughput (symmetric scenario), | |
824 | * or, as a special case, if bfqq must receive a share of the | |
825 | * throughput lower than or equal to the share that every other active | |
826 | * queue must receive. If bfqq does sync I/O, then these are the only | |
827 | * two cases where bfqq happens to be guaranteed its share of the | |
828 | * throughput even if I/O dispatching is not plugged when bfqq remains | |
829 | * temporarily empty (for more details, see the comments in the | |
830 | * function bfq_better_to_idle()). For this reason, the return value | |
831 | * of this function is used to check whether I/O-dispatch plugging can | |
832 | * be avoided. | |
1de0c4cd | 833 | * |
fb53ac6c | 834 | * The above first case (symmetric scenario) occurs when: |
1de0c4cd | 835 | * 1) all active queues have the same weight, |
73d58118 | 836 | * 2) all active queues belong to the same I/O-priority class, |
1de0c4cd | 837 | * 3) all active groups at the same level in the groups tree have the same |
73d58118 PV |
838 | * weight, |
839 | * 4) all active groups at the same level in the groups tree have the same | |
1de0c4cd AA |
840 | * number of children. |
841 | * | |
2d29c9f8 FM |
842 | * Unfortunately, keeping the necessary state for evaluating exactly |
843 | * the last two symmetry sub-conditions above would be quite complex | |
73d58118 PV |
844 | * and time consuming. Therefore this function evaluates, instead, |
845 | * only the following stronger three sub-conditions, for which it is | |
2d29c9f8 | 846 | * much easier to maintain the needed state: |
1de0c4cd | 847 | * 1) all active queues have the same weight, |
73d58118 | 848 | * 2) all active queues belong to the same I/O-priority class, |
eed3ecc9 | 849 | * 3) there is at most one active group. |
2d29c9f8 FM |
850 | * In particular, the last condition is always true if hierarchical |
851 | * support or the cgroups interface are not enabled, thus no state | |
852 | * needs to be maintained in this case. | |
1de0c4cd | 853 | */ |
fb53ac6c PV |
854 | static bool bfq_asymmetric_scenario(struct bfq_data *bfqd, |
855 | struct bfq_queue *bfqq) | |
1de0c4cd | 856 | { |
fb53ac6c PV |
857 | bool smallest_weight = bfqq && |
858 | bfqq->weight_counter && | |
859 | bfqq->weight_counter == | |
860 | container_of( | |
861 | rb_first_cached(&bfqd->queue_weights_tree), | |
862 | struct bfq_weight_counter, | |
863 | weights_node); | |
864 | ||
73d58118 PV |
865 | /* |
866 | * For queue weights to differ, queue_weights_tree must contain | |
867 | * at least two nodes. | |
868 | */ | |
fb53ac6c PV |
869 | bool varied_queue_weights = !smallest_weight && |
870 | !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) && | |
871 | (bfqd->queue_weights_tree.rb_root.rb_node->rb_left || | |
872 | bfqd->queue_weights_tree.rb_root.rb_node->rb_right); | |
73d58118 PV |
873 | |
874 | bool multiple_classes_busy = | |
875 | (bfqd->busy_queues[0] && bfqd->busy_queues[1]) || | |
876 | (bfqd->busy_queues[0] && bfqd->busy_queues[2]) || | |
877 | (bfqd->busy_queues[1] && bfqd->busy_queues[2]); | |
878 | ||
fb53ac6c | 879 | return varied_queue_weights || multiple_classes_busy |
42b1bd33 | 880 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
eed3ecc9 | 881 | || bfqd->num_groups_with_pending_reqs > 1 |
73d58118 | 882 | #endif |
fb53ac6c | 883 | ; |
1de0c4cd AA |
884 | } |
885 | ||
886 | /* | |
887 | * If the weight-counter tree passed as input contains no counter for | |
2d29c9f8 | 888 | * the weight of the input queue, then add that counter; otherwise just |
1de0c4cd AA |
889 | * increment the existing counter. |
890 | * | |
891 | * Note that weight-counter trees contain few nodes in mostly symmetric | |
892 | * scenarios. For example, if all queues have the same weight, then the | |
893 | * weight-counter tree for the queues may contain at most one node. | |
894 | * This holds even if low_latency is on, because weight-raised queues | |
895 | * are not inserted in the tree. | |
896 | * In most scenarios, the rate at which nodes are created/destroyed | |
897 | * should be low too. | |
898 | */ | |
afdba146 | 899 | void bfq_weights_tree_add(struct bfq_queue *bfqq) |
1de0c4cd | 900 | { |
afdba146 | 901 | struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree; |
2d29c9f8 | 902 | struct bfq_entity *entity = &bfqq->entity; |
fb53ac6c PV |
903 | struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; |
904 | bool leftmost = true; | |
1de0c4cd AA |
905 | |
906 | /* | |
2d29c9f8 | 907 | * Do not insert if the queue is already associated with a |
1de0c4cd | 908 | * counter, which happens if: |
2d29c9f8 | 909 | * 1) a request arrival has caused the queue to become both |
1de0c4cd AA |
910 | * non-weight-raised, and hence change its weight, and |
911 | * backlogged; in this respect, each of the two events | |
912 | * causes an invocation of this function, | |
2d29c9f8 | 913 | * 2) this is the invocation of this function caused by the |
1de0c4cd AA |
914 | * second event. This second invocation is actually useless, |
915 | * and we handle this fact by exiting immediately. More | |
916 | * efficient or clearer solutions might possibly be adopted. | |
917 | */ | |
2d29c9f8 | 918 | if (bfqq->weight_counter) |
1de0c4cd AA |
919 | return; |
920 | ||
921 | while (*new) { | |
922 | struct bfq_weight_counter *__counter = container_of(*new, | |
923 | struct bfq_weight_counter, | |
924 | weights_node); | |
925 | parent = *new; | |
926 | ||
927 | if (entity->weight == __counter->weight) { | |
2d29c9f8 | 928 | bfqq->weight_counter = __counter; |
1de0c4cd AA |
929 | goto inc_counter; |
930 | } | |
931 | if (entity->weight < __counter->weight) | |
932 | new = &((*new)->rb_left); | |
fb53ac6c | 933 | else { |
1de0c4cd | 934 | new = &((*new)->rb_right); |
fb53ac6c PV |
935 | leftmost = false; |
936 | } | |
1de0c4cd AA |
937 | } |
938 | ||
2d29c9f8 FM |
939 | bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter), |
940 | GFP_ATOMIC); | |
1de0c4cd AA |
941 | |
942 | /* | |
943 | * In the unlucky event of an allocation failure, we just | |
2d29c9f8 | 944 | * exit. This will cause the weight of queue to not be |
fb53ac6c | 945 | * considered in bfq_asymmetric_scenario, which, in its turn, |
73d58118 PV |
946 | * causes the scenario to be deemed wrongly symmetric in case |
947 | * bfqq's weight would have been the only weight making the | |
948 | * scenario asymmetric. On the bright side, no unbalance will | |
949 | * however occur when bfqq becomes inactive again (the | |
950 | * invocation of this function is triggered by an activation | |
951 | * of queue). In fact, bfq_weights_tree_remove does nothing | |
952 | * if !bfqq->weight_counter. | |
1de0c4cd | 953 | */ |
2d29c9f8 | 954 | if (unlikely(!bfqq->weight_counter)) |
1de0c4cd AA |
955 | return; |
956 | ||
2d29c9f8 FM |
957 | bfqq->weight_counter->weight = entity->weight; |
958 | rb_link_node(&bfqq->weight_counter->weights_node, parent, new); | |
fb53ac6c PV |
959 | rb_insert_color_cached(&bfqq->weight_counter->weights_node, root, |
960 | leftmost); | |
1de0c4cd AA |
961 | |
962 | inc_counter: | |
2d29c9f8 | 963 | bfqq->weight_counter->num_active++; |
9dee8b3b | 964 | bfqq->ref++; |
1de0c4cd AA |
965 | } |
966 | ||
967 | /* | |
2d29c9f8 | 968 | * Decrement the weight counter associated with the queue, and, if the |
1de0c4cd AA |
969 | * counter reaches 0, remove the counter from the tree. |
970 | * See the comments to the function bfq_weights_tree_add() for considerations | |
971 | * about overhead. | |
972 | */ | |
eb5bca73 | 973 | void bfq_weights_tree_remove(struct bfq_queue *bfqq) |
1de0c4cd | 974 | { |
afdba146 YK |
975 | struct rb_root_cached *root; |
976 | ||
2d29c9f8 | 977 | if (!bfqq->weight_counter) |
1de0c4cd AA |
978 | return; |
979 | ||
afdba146 | 980 | root = &bfqq->bfqd->queue_weights_tree; |
2d29c9f8 FM |
981 | bfqq->weight_counter->num_active--; |
982 | if (bfqq->weight_counter->num_active > 0) | |
1de0c4cd AA |
983 | goto reset_entity_pointer; |
984 | ||
fb53ac6c | 985 | rb_erase_cached(&bfqq->weight_counter->weights_node, root); |
2d29c9f8 | 986 | kfree(bfqq->weight_counter); |
1de0c4cd AA |
987 | |
988 | reset_entity_pointer: | |
2d29c9f8 | 989 | bfqq->weight_counter = NULL; |
9dee8b3b | 990 | bfq_put_queue(bfqq); |
1de0c4cd AA |
991 | } |
992 | ||
aee69d78 PV |
993 | /* |
994 | * Return expired entry, or NULL to just start from scratch in rbtree. | |
995 | */ | |
996 | static struct request *bfq_check_fifo(struct bfq_queue *bfqq, | |
997 | struct request *last) | |
998 | { | |
999 | struct request *rq; | |
1000 | ||
1001 | if (bfq_bfqq_fifo_expire(bfqq)) | |
1002 | return NULL; | |
1003 | ||
1004 | bfq_mark_bfqq_fifo_expire(bfqq); | |
1005 | ||
1006 | rq = rq_entry_fifo(bfqq->fifo.next); | |
1007 | ||
1008 | if (rq == last || ktime_get_ns() < rq->fifo_time) | |
1009 | return NULL; | |
1010 | ||
1011 | bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq); | |
1012 | return rq; | |
1013 | } | |
1014 | ||
1015 | static struct request *bfq_find_next_rq(struct bfq_data *bfqd, | |
1016 | struct bfq_queue *bfqq, | |
1017 | struct request *last) | |
1018 | { | |
1019 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
1020 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
1021 | struct request *next, *prev = NULL; | |
1022 | ||
1023 | /* Follow expired path, else get first next available. */ | |
1024 | next = bfq_check_fifo(bfqq, last); | |
1025 | if (next) | |
1026 | return next; | |
1027 | ||
1028 | if (rbprev) | |
1029 | prev = rb_entry_rq(rbprev); | |
1030 | ||
1031 | if (rbnext) | |
1032 | next = rb_entry_rq(rbnext); | |
1033 | else { | |
1034 | rbnext = rb_first(&bfqq->sort_list); | |
1035 | if (rbnext && rbnext != &last->rb_node) | |
1036 | next = rb_entry_rq(rbnext); | |
1037 | } | |
1038 | ||
1039 | return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); | |
1040 | } | |
1041 | ||
c074170e | 1042 | /* see the definition of bfq_async_charge_factor for details */ |
aee69d78 PV |
1043 | static unsigned long bfq_serv_to_charge(struct request *rq, |
1044 | struct bfq_queue *bfqq) | |
1045 | { | |
02a6d787 | 1046 | if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 || |
fb53ac6c | 1047 | bfq_asymmetric_scenario(bfqq->bfqd, bfqq)) |
c074170e PV |
1048 | return blk_rq_sectors(rq); |
1049 | ||
d5801088 | 1050 | return blk_rq_sectors(rq) * bfq_async_charge_factor; |
aee69d78 PV |
1051 | } |
1052 | ||
1053 | /** | |
1054 | * bfq_updated_next_req - update the queue after a new next_rq selection. | |
1055 | * @bfqd: the device data the queue belongs to. | |
1056 | * @bfqq: the queue to update. | |
1057 | * | |
1058 | * If the first request of a queue changes we make sure that the queue | |
1059 | * has enough budget to serve at least its first request (if the | |
1060 | * request has grown). We do this because if the queue has not enough | |
1061 | * budget for its first request, it has to go through two dispatch | |
1062 | * rounds to actually get it dispatched. | |
1063 | */ | |
1064 | static void bfq_updated_next_req(struct bfq_data *bfqd, | |
1065 | struct bfq_queue *bfqq) | |
1066 | { | |
1067 | struct bfq_entity *entity = &bfqq->entity; | |
1068 | struct request *next_rq = bfqq->next_rq; | |
1069 | unsigned long new_budget; | |
1070 | ||
1071 | if (!next_rq) | |
1072 | return; | |
1073 | ||
1074 | if (bfqq == bfqd->in_service_queue) | |
1075 | /* | |
1076 | * In order not to break guarantees, budgets cannot be | |
1077 | * changed after an entity has been selected. | |
1078 | */ | |
1079 | return; | |
1080 | ||
f3218ad8 PV |
1081 | new_budget = max_t(unsigned long, |
1082 | max_t(unsigned long, bfqq->max_budget, | |
1083 | bfq_serv_to_charge(next_rq, bfqq)), | |
1084 | entity->service); | |
aee69d78 PV |
1085 | if (entity->budget != new_budget) { |
1086 | entity->budget = new_budget; | |
1087 | bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", | |
1088 | new_budget); | |
80294c3b | 1089 | bfq_requeue_bfqq(bfqd, bfqq, false); |
aee69d78 PV |
1090 | } |
1091 | } | |
1092 | ||
3e2bdd6d PV |
1093 | static unsigned int bfq_wr_duration(struct bfq_data *bfqd) |
1094 | { | |
1095 | u64 dur; | |
1096 | ||
e24f1c24 | 1097 | dur = bfqd->rate_dur_prod; |
3e2bdd6d PV |
1098 | do_div(dur, bfqd->peak_rate); |
1099 | ||
1100 | /* | |
d450542e DS |
1101 | * Limit duration between 3 and 25 seconds. The upper limit |
1102 | * has been conservatively set after the following worst case: | |
1103 | * on a QEMU/KVM virtual machine | |
1104 | * - running in a slow PC | |
1105 | * - with a virtual disk stacked on a slow low-end 5400rpm HDD | |
1106 | * - serving a heavy I/O workload, such as the sequential reading | |
1107 | * of several files | |
1108 | * mplayer took 23 seconds to start, if constantly weight-raised. | |
1109 | * | |
636b8fe8 | 1110 | * As for higher values than that accommodating the above bad |
d450542e DS |
1111 | * scenario, tests show that higher values would often yield |
1112 | * the opposite of the desired result, i.e., would worsen | |
1113 | * responsiveness by allowing non-interactive applications to | |
1114 | * preserve weight raising for too long. | |
3e2bdd6d PV |
1115 | * |
1116 | * On the other end, lower values than 3 seconds make it | |
1117 | * difficult for most interactive tasks to complete their jobs | |
1118 | * before weight-raising finishes. | |
1119 | */ | |
d450542e | 1120 | return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000)); |
3e2bdd6d PV |
1121 | } |
1122 | ||
1123 | /* switch back from soft real-time to interactive weight raising */ | |
1124 | static void switch_back_to_interactive_wr(struct bfq_queue *bfqq, | |
1125 | struct bfq_data *bfqd) | |
1126 | { | |
1127 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1128 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1129 | bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt; | |
1130 | } | |
1131 | ||
36eca894 | 1132 | static void |
13c931bd PV |
1133 | bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd, |
1134 | struct bfq_io_cq *bic, bool bfq_already_existing) | |
36eca894 | 1135 | { |
8c544770 | 1136 | unsigned int old_wr_coeff = 1; |
13c931bd | 1137 | bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq); |
fd571df0 PV |
1138 | unsigned int a_idx = bfqq->actuator_idx; |
1139 | struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx]; | |
13c931bd | 1140 | |
a6123047 | 1141 | if (bfqq_data->saved_has_short_ttime) |
d5be3fef | 1142 | bfq_mark_bfqq_has_short_ttime(bfqq); |
36eca894 | 1143 | else |
d5be3fef | 1144 | bfq_clear_bfqq_has_short_ttime(bfqq); |
36eca894 | 1145 | |
a6123047 | 1146 | if (bfqq_data->saved_IO_bound) |
36eca894 AA |
1147 | bfq_mark_bfqq_IO_bound(bfqq); |
1148 | else | |
1149 | bfq_clear_bfqq_IO_bound(bfqq); | |
1150 | ||
a6123047 PV |
1151 | bfqq->last_serv_time_ns = bfqq_data->saved_last_serv_time_ns; |
1152 | bfqq->inject_limit = bfqq_data->saved_inject_limit; | |
1153 | bfqq->decrease_time_jif = bfqq_data->saved_decrease_time_jif; | |
5a5436b9 | 1154 | |
a6123047 PV |
1155 | bfqq->entity.new_weight = bfqq_data->saved_weight; |
1156 | bfqq->ttime = bfqq_data->saved_ttime; | |
1157 | bfqq->io_start_time = bfqq_data->saved_io_start_time; | |
1158 | bfqq->tot_idle_time = bfqq_data->saved_tot_idle_time; | |
8c544770 PV |
1159 | /* |
1160 | * Restore weight coefficient only if low_latency is on | |
1161 | */ | |
1162 | if (bfqd->low_latency) { | |
1163 | old_wr_coeff = bfqq->wr_coeff; | |
a6123047 | 1164 | bfqq->wr_coeff = bfqq_data->saved_wr_coeff; |
8c544770 | 1165 | } |
a6123047 PV |
1166 | bfqq->service_from_wr = bfqq_data->saved_service_from_wr; |
1167 | bfqq->wr_start_at_switch_to_srt = | |
1168 | bfqq_data->saved_wr_start_at_switch_to_srt; | |
1169 | bfqq->last_wr_start_finish = bfqq_data->saved_last_wr_start_finish; | |
1170 | bfqq->wr_cur_max_time = bfqq_data->saved_wr_cur_max_time; | |
36eca894 | 1171 | |
e1b2324d | 1172 | if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) || |
36eca894 | 1173 | time_is_before_jiffies(bfqq->last_wr_start_finish + |
e1b2324d | 1174 | bfqq->wr_cur_max_time))) { |
3e2bdd6d PV |
1175 | if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && |
1176 | !bfq_bfqq_in_large_burst(bfqq) && | |
1177 | time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt + | |
1178 | bfq_wr_duration(bfqd))) { | |
1179 | switch_back_to_interactive_wr(bfqq, bfqd); | |
1180 | } else { | |
1181 | bfqq->wr_coeff = 1; | |
1182 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
1183 | "resume state: switching off wr"); | |
1184 | } | |
36eca894 AA |
1185 | } |
1186 | ||
1187 | /* make sure weight will be updated, however we got here */ | |
1188 | bfqq->entity.prio_changed = 1; | |
13c931bd PV |
1189 | |
1190 | if (likely(!busy)) | |
1191 | return; | |
1192 | ||
1193 | if (old_wr_coeff == 1 && bfqq->wr_coeff > 1) | |
1194 | bfqd->wr_busy_queues++; | |
1195 | else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1) | |
1196 | bfqd->wr_busy_queues--; | |
36eca894 AA |
1197 | } |
1198 | ||
1199 | static int bfqq_process_refs(struct bfq_queue *bfqq) | |
1200 | { | |
98f04499 JK |
1201 | return bfqq->ref - bfqq->entity.allocated - |
1202 | bfqq->entity.on_st_or_in_serv - | |
430a67f9 | 1203 | (bfqq->weight_counter != NULL) - bfqq->stable_ref; |
36eca894 AA |
1204 | } |
1205 | ||
e1b2324d AA |
1206 | /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */ |
1207 | static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1208 | { | |
1209 | struct bfq_queue *item; | |
1210 | struct hlist_node *n; | |
1211 | ||
1212 | hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node) | |
1213 | hlist_del_init(&item->burst_list_node); | |
84a74689 PV |
1214 | |
1215 | /* | |
1216 | * Start the creation of a new burst list only if there is no | |
1217 | * active queue. See comments on the conditional invocation of | |
1218 | * bfq_handle_burst(). | |
1219 | */ | |
1220 | if (bfq_tot_busy_queues(bfqd) == 0) { | |
1221 | hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list); | |
1222 | bfqd->burst_size = 1; | |
1223 | } else | |
1224 | bfqd->burst_size = 0; | |
1225 | ||
e1b2324d AA |
1226 | bfqd->burst_parent_entity = bfqq->entity.parent; |
1227 | } | |
1228 | ||
1229 | /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */ | |
1230 | static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1231 | { | |
1232 | /* Increment burst size to take into account also bfqq */ | |
1233 | bfqd->burst_size++; | |
1234 | ||
1235 | if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) { | |
1236 | struct bfq_queue *pos, *bfqq_item; | |
1237 | struct hlist_node *n; | |
1238 | ||
1239 | /* | |
1240 | * Enough queues have been activated shortly after each | |
1241 | * other to consider this burst as large. | |
1242 | */ | |
1243 | bfqd->large_burst = true; | |
1244 | ||
1245 | /* | |
1246 | * We can now mark all queues in the burst list as | |
1247 | * belonging to a large burst. | |
1248 | */ | |
1249 | hlist_for_each_entry(bfqq_item, &bfqd->burst_list, | |
1250 | burst_list_node) | |
1251 | bfq_mark_bfqq_in_large_burst(bfqq_item); | |
1252 | bfq_mark_bfqq_in_large_burst(bfqq); | |
1253 | ||
1254 | /* | |
1255 | * From now on, and until the current burst finishes, any | |
1256 | * new queue being activated shortly after the last queue | |
1257 | * was inserted in the burst can be immediately marked as | |
1258 | * belonging to a large burst. So the burst list is not | |
1259 | * needed any more. Remove it. | |
1260 | */ | |
1261 | hlist_for_each_entry_safe(pos, n, &bfqd->burst_list, | |
1262 | burst_list_node) | |
1263 | hlist_del_init(&pos->burst_list_node); | |
1264 | } else /* | |
1265 | * Burst not yet large: add bfqq to the burst list. Do | |
1266 | * not increment the ref counter for bfqq, because bfqq | |
1267 | * is removed from the burst list before freeing bfqq | |
1268 | * in put_queue. | |
1269 | */ | |
1270 | hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list); | |
1271 | } | |
1272 | ||
1273 | /* | |
1274 | * If many queues belonging to the same group happen to be created | |
1275 | * shortly after each other, then the processes associated with these | |
1276 | * queues have typically a common goal. In particular, bursts of queue | |
1277 | * creations are usually caused by services or applications that spawn | |
1278 | * many parallel threads/processes. Examples are systemd during boot, | |
1279 | * or git grep. To help these processes get their job done as soon as | |
1280 | * possible, it is usually better to not grant either weight-raising | |
84a74689 PV |
1281 | * or device idling to their queues, unless these queues must be |
1282 | * protected from the I/O flowing through other active queues. | |
e1b2324d AA |
1283 | * |
1284 | * In this comment we describe, firstly, the reasons why this fact | |
1285 | * holds, and, secondly, the next function, which implements the main | |
1286 | * steps needed to properly mark these queues so that they can then be | |
1287 | * treated in a different way. | |
1288 | * | |
1289 | * The above services or applications benefit mostly from a high | |
1290 | * throughput: the quicker the requests of the activated queues are | |
1291 | * cumulatively served, the sooner the target job of these queues gets | |
1292 | * completed. As a consequence, weight-raising any of these queues, | |
1293 | * which also implies idling the device for it, is almost always | |
84a74689 PV |
1294 | * counterproductive, unless there are other active queues to isolate |
1295 | * these new queues from. If there no other active queues, then | |
1296 | * weight-raising these new queues just lowers throughput in most | |
1297 | * cases. | |
e1b2324d AA |
1298 | * |
1299 | * On the other hand, a burst of queue creations may be caused also by | |
1300 | * the start of an application that does not consist of a lot of | |
1301 | * parallel I/O-bound threads. In fact, with a complex application, | |
1302 | * several short processes may need to be executed to start-up the | |
1303 | * application. In this respect, to start an application as quickly as | |
1304 | * possible, the best thing to do is in any case to privilege the I/O | |
1305 | * related to the application with respect to all other | |
1306 | * I/O. Therefore, the best strategy to start as quickly as possible | |
1307 | * an application that causes a burst of queue creations is to | |
1308 | * weight-raise all the queues created during the burst. This is the | |
1309 | * exact opposite of the best strategy for the other type of bursts. | |
1310 | * | |
1311 | * In the end, to take the best action for each of the two cases, the | |
1312 | * two types of bursts need to be distinguished. Fortunately, this | |
1313 | * seems relatively easy, by looking at the sizes of the bursts. In | |
1314 | * particular, we found a threshold such that only bursts with a | |
1315 | * larger size than that threshold are apparently caused by | |
1316 | * services or commands such as systemd or git grep. For brevity, | |
1317 | * hereafter we call just 'large' these bursts. BFQ *does not* | |
1318 | * weight-raise queues whose creation occurs in a large burst. In | |
1319 | * addition, for each of these queues BFQ performs or does not perform | |
1320 | * idling depending on which choice boosts the throughput more. The | |
1321 | * exact choice depends on the device and request pattern at | |
1322 | * hand. | |
1323 | * | |
1324 | * Unfortunately, false positives may occur while an interactive task | |
1325 | * is starting (e.g., an application is being started). The | |
1326 | * consequence is that the queues associated with the task do not | |
1327 | * enjoy weight raising as expected. Fortunately these false positives | |
1328 | * are very rare. They typically occur if some service happens to | |
1329 | * start doing I/O exactly when the interactive task starts. | |
1330 | * | |
84a74689 PV |
1331 | * Turning back to the next function, it is invoked only if there are |
1332 | * no active queues (apart from active queues that would belong to the | |
1333 | * same, possible burst bfqq would belong to), and it implements all | |
1334 | * the steps needed to detect the occurrence of a large burst and to | |
1335 | * properly mark all the queues belonging to it (so that they can then | |
1336 | * be treated in a different way). This goal is achieved by | |
1337 | * maintaining a "burst list" that holds, temporarily, the queues that | |
1338 | * belong to the burst in progress. The list is then used to mark | |
1339 | * these queues as belonging to a large burst if the burst does become | |
1340 | * large. The main steps are the following. | |
e1b2324d AA |
1341 | * |
1342 | * . when the very first queue is created, the queue is inserted into the | |
1343 | * list (as it could be the first queue in a possible burst) | |
1344 | * | |
1345 | * . if the current burst has not yet become large, and a queue Q that does | |
1346 | * not yet belong to the burst is activated shortly after the last time | |
1347 | * at which a new queue entered the burst list, then the function appends | |
1348 | * Q to the burst list | |
1349 | * | |
1350 | * . if, as a consequence of the previous step, the burst size reaches | |
1351 | * the large-burst threshold, then | |
1352 | * | |
1353 | * . all the queues in the burst list are marked as belonging to a | |
1354 | * large burst | |
1355 | * | |
1356 | * . the burst list is deleted; in fact, the burst list already served | |
1357 | * its purpose (keeping temporarily track of the queues in a burst, | |
1358 | * so as to be able to mark them as belonging to a large burst in the | |
1359 | * previous sub-step), and now is not needed any more | |
1360 | * | |
1361 | * . the device enters a large-burst mode | |
1362 | * | |
1363 | * . if a queue Q that does not belong to the burst is created while | |
1364 | * the device is in large-burst mode and shortly after the last time | |
1365 | * at which a queue either entered the burst list or was marked as | |
1366 | * belonging to the current large burst, then Q is immediately marked | |
1367 | * as belonging to a large burst. | |
1368 | * | |
1369 | * . if a queue Q that does not belong to the burst is created a while | |
1370 | * later, i.e., not shortly after, than the last time at which a queue | |
1371 | * either entered the burst list or was marked as belonging to the | |
1372 | * current large burst, then the current burst is deemed as finished and: | |
1373 | * | |
1374 | * . the large-burst mode is reset if set | |
1375 | * | |
1376 | * . the burst list is emptied | |
1377 | * | |
1378 | * . Q is inserted in the burst list, as Q may be the first queue | |
1379 | * in a possible new burst (then the burst list contains just Q | |
1380 | * after this step). | |
1381 | */ | |
1382 | static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1383 | { | |
1384 | /* | |
1385 | * If bfqq is already in the burst list or is part of a large | |
1386 | * burst, or finally has just been split, then there is | |
1387 | * nothing else to do. | |
1388 | */ | |
1389 | if (!hlist_unhashed(&bfqq->burst_list_node) || | |
1390 | bfq_bfqq_in_large_burst(bfqq) || | |
1391 | time_is_after_eq_jiffies(bfqq->split_time + | |
1392 | msecs_to_jiffies(10))) | |
1393 | return; | |
1394 | ||
1395 | /* | |
1396 | * If bfqq's creation happens late enough, or bfqq belongs to | |
1397 | * a different group than the burst group, then the current | |
1398 | * burst is finished, and related data structures must be | |
1399 | * reset. | |
1400 | * | |
1401 | * In this respect, consider the special case where bfqq is | |
1402 | * the very first queue created after BFQ is selected for this | |
1403 | * device. In this case, last_ins_in_burst and | |
1404 | * burst_parent_entity are not yet significant when we get | |
1405 | * here. But it is easy to verify that, whether or not the | |
1406 | * following condition is true, bfqq will end up being | |
1407 | * inserted into the burst list. In particular the list will | |
1408 | * happen to contain only bfqq. And this is exactly what has | |
1409 | * to happen, as bfqq may be the first queue of the first | |
1410 | * burst. | |
1411 | */ | |
1412 | if (time_is_before_jiffies(bfqd->last_ins_in_burst + | |
1413 | bfqd->bfq_burst_interval) || | |
1414 | bfqq->entity.parent != bfqd->burst_parent_entity) { | |
1415 | bfqd->large_burst = false; | |
1416 | bfq_reset_burst_list(bfqd, bfqq); | |
1417 | goto end; | |
1418 | } | |
1419 | ||
1420 | /* | |
1421 | * If we get here, then bfqq is being activated shortly after the | |
1422 | * last queue. So, if the current burst is also large, we can mark | |
1423 | * bfqq as belonging to this large burst immediately. | |
1424 | */ | |
1425 | if (bfqd->large_burst) { | |
1426 | bfq_mark_bfqq_in_large_burst(bfqq); | |
1427 | goto end; | |
1428 | } | |
1429 | ||
1430 | /* | |
1431 | * If we get here, then a large-burst state has not yet been | |
1432 | * reached, but bfqq is being activated shortly after the last | |
1433 | * queue. Then we add bfqq to the burst. | |
1434 | */ | |
1435 | bfq_add_to_burst(bfqd, bfqq); | |
1436 | end: | |
1437 | /* | |
1438 | * At this point, bfqq either has been added to the current | |
1439 | * burst or has caused the current burst to terminate and a | |
1440 | * possible new burst to start. In particular, in the second | |
1441 | * case, bfqq has become the first queue in the possible new | |
1442 | * burst. In both cases last_ins_in_burst needs to be moved | |
1443 | * forward. | |
1444 | */ | |
1445 | bfqd->last_ins_in_burst = jiffies; | |
1446 | } | |
1447 | ||
aee69d78 PV |
1448 | static int bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
1449 | { | |
1450 | struct bfq_entity *entity = &bfqq->entity; | |
1451 | ||
1452 | return entity->budget - entity->service; | |
1453 | } | |
1454 | ||
1455 | /* | |
1456 | * If enough samples have been computed, return the current max budget | |
1457 | * stored in bfqd, which is dynamically updated according to the | |
1458 | * estimated disk peak rate; otherwise return the default max budget | |
1459 | */ | |
1460 | static int bfq_max_budget(struct bfq_data *bfqd) | |
1461 | { | |
1462 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
1463 | return bfq_default_max_budget; | |
1464 | else | |
1465 | return bfqd->bfq_max_budget; | |
1466 | } | |
1467 | ||
1468 | /* | |
1469 | * Return min budget, which is a fraction of the current or default | |
1470 | * max budget (trying with 1/32) | |
1471 | */ | |
1472 | static int bfq_min_budget(struct bfq_data *bfqd) | |
1473 | { | |
1474 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
1475 | return bfq_default_max_budget / 32; | |
1476 | else | |
1477 | return bfqd->bfq_max_budget / 32; | |
1478 | } | |
1479 | ||
aee69d78 PV |
1480 | /* |
1481 | * The next function, invoked after the input queue bfqq switches from | |
1482 | * idle to busy, updates the budget of bfqq. The function also tells | |
1483 | * whether the in-service queue should be expired, by returning | |
1484 | * true. The purpose of expiring the in-service queue is to give bfqq | |
1485 | * the chance to possibly preempt the in-service queue, and the reason | |
44e44a1b PV |
1486 | * for preempting the in-service queue is to achieve one of the two |
1487 | * goals below. | |
aee69d78 | 1488 | * |
44e44a1b PV |
1489 | * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has |
1490 | * expired because it has remained idle. In particular, bfqq may have | |
1491 | * expired for one of the following two reasons: | |
aee69d78 PV |
1492 | * |
1493 | * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling | |
1494 | * and did not make it to issue a new request before its last | |
1495 | * request was served; | |
1496 | * | |
1497 | * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue | |
1498 | * a new request before the expiration of the idling-time. | |
1499 | * | |
1500 | * Even if bfqq has expired for one of the above reasons, the process | |
1501 | * associated with the queue may be however issuing requests greedily, | |
1502 | * and thus be sensitive to the bandwidth it receives (bfqq may have | |
1503 | * remained idle for other reasons: CPU high load, bfqq not enjoying | |
1504 | * idling, I/O throttling somewhere in the path from the process to | |
1505 | * the I/O scheduler, ...). But if, after every expiration for one of | |
1506 | * the above two reasons, bfqq has to wait for the service of at least | |
1507 | * one full budget of another queue before being served again, then | |
1508 | * bfqq is likely to get a much lower bandwidth or resource time than | |
1509 | * its reserved ones. To address this issue, two countermeasures need | |
1510 | * to be taken. | |
1511 | * | |
1512 | * First, the budget and the timestamps of bfqq need to be updated in | |
1513 | * a special way on bfqq reactivation: they need to be updated as if | |
1514 | * bfqq did not remain idle and did not expire. In fact, if they are | |
1515 | * computed as if bfqq expired and remained idle until reactivation, | |
1516 | * then the process associated with bfqq is treated as if, instead of | |
1517 | * being greedy, it stopped issuing requests when bfqq remained idle, | |
1518 | * and restarts issuing requests only on this reactivation. In other | |
1519 | * words, the scheduler does not help the process recover the "service | |
1520 | * hole" between bfqq expiration and reactivation. As a consequence, | |
1521 | * the process receives a lower bandwidth than its reserved one. In | |
1522 | * contrast, to recover this hole, the budget must be updated as if | |
1523 | * bfqq was not expired at all before this reactivation, i.e., it must | |
1524 | * be set to the value of the remaining budget when bfqq was | |
1525 | * expired. Along the same line, timestamps need to be assigned the | |
1526 | * value they had the last time bfqq was selected for service, i.e., | |
1527 | * before last expiration. Thus timestamps need to be back-shifted | |
1528 | * with respect to their normal computation (see [1] for more details | |
1529 | * on this tricky aspect). | |
1530 | * | |
1531 | * Secondly, to allow the process to recover the hole, the in-service | |
1532 | * queue must be expired too, to give bfqq the chance to preempt it | |
1533 | * immediately. In fact, if bfqq has to wait for a full budget of the | |
1534 | * in-service queue to be completed, then it may become impossible to | |
1535 | * let the process recover the hole, even if the back-shifted | |
1536 | * timestamps of bfqq are lower than those of the in-service queue. If | |
1537 | * this happens for most or all of the holes, then the process may not | |
1538 | * receive its reserved bandwidth. In this respect, it is worth noting | |
1539 | * that, being the service of outstanding requests unpreemptible, a | |
1540 | * little fraction of the holes may however be unrecoverable, thereby | |
1541 | * causing a little loss of bandwidth. | |
1542 | * | |
1543 | * The last important point is detecting whether bfqq does need this | |
1544 | * bandwidth recovery. In this respect, the next function deems the | |
1545 | * process associated with bfqq greedy, and thus allows it to recover | |
1546 | * the hole, if: 1) the process is waiting for the arrival of a new | |
1547 | * request (which implies that bfqq expired for one of the above two | |
1548 | * reasons), and 2) such a request has arrived soon. The first | |
1549 | * condition is controlled through the flag non_blocking_wait_rq, | |
1550 | * while the second through the flag arrived_in_time. If both | |
1551 | * conditions hold, then the function computes the budget in the | |
1552 | * above-described special way, and signals that the in-service queue | |
1553 | * should be expired. Timestamp back-shifting is done later in | |
1554 | * __bfq_activate_entity. | |
44e44a1b PV |
1555 | * |
1556 | * 2. Reduce latency. Even if timestamps are not backshifted to let | |
1557 | * the process associated with bfqq recover a service hole, bfqq may | |
1558 | * however happen to have, after being (re)activated, a lower finish | |
1559 | * timestamp than the in-service queue. That is, the next budget of | |
1560 | * bfqq may have to be completed before the one of the in-service | |
1561 | * queue. If this is the case, then preempting the in-service queue | |
1562 | * allows this goal to be achieved, apart from the unpreemptible, | |
1563 | * outstanding requests mentioned above. | |
1564 | * | |
1565 | * Unfortunately, regardless of which of the above two goals one wants | |
1566 | * to achieve, service trees need first to be updated to know whether | |
1567 | * the in-service queue must be preempted. To have service trees | |
1568 | * correctly updated, the in-service queue must be expired and | |
1569 | * rescheduled, and bfqq must be scheduled too. This is one of the | |
1570 | * most costly operations (in future versions, the scheduling | |
1571 | * mechanism may be re-designed in such a way to make it possible to | |
1572 | * know whether preemption is needed without needing to update service | |
1573 | * trees). In addition, queue preemptions almost always cause random | |
96a291c3 PV |
1574 | * I/O, which may in turn cause loss of throughput. Finally, there may |
1575 | * even be no in-service queue when the next function is invoked (so, | |
1576 | * no queue to compare timestamps with). Because of these facts, the | |
1577 | * next function adopts the following simple scheme to avoid costly | |
1578 | * operations, too frequent preemptions and too many dependencies on | |
1579 | * the state of the scheduler: it requests the expiration of the | |
1580 | * in-service queue (unconditionally) only for queues that need to | |
1581 | * recover a hole. Then it delegates to other parts of the code the | |
1582 | * responsibility of handling the above case 2. | |
aee69d78 PV |
1583 | */ |
1584 | static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd, | |
1585 | struct bfq_queue *bfqq, | |
96a291c3 | 1586 | bool arrived_in_time) |
aee69d78 PV |
1587 | { |
1588 | struct bfq_entity *entity = &bfqq->entity; | |
1589 | ||
218cb897 PV |
1590 | /* |
1591 | * In the next compound condition, we check also whether there | |
1592 | * is some budget left, because otherwise there is no point in | |
1593 | * trying to go on serving bfqq with this same budget: bfqq | |
1594 | * would be expired immediately after being selected for | |
1595 | * service. This would only cause useless overhead. | |
1596 | */ | |
1597 | if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time && | |
1598 | bfq_bfqq_budget_left(bfqq) > 0) { | |
aee69d78 PV |
1599 | /* |
1600 | * We do not clear the flag non_blocking_wait_rq here, as | |
1601 | * the latter is used in bfq_activate_bfqq to signal | |
1602 | * that timestamps need to be back-shifted (and is | |
1603 | * cleared right after). | |
1604 | */ | |
1605 | ||
1606 | /* | |
1607 | * In next assignment we rely on that either | |
1608 | * entity->service or entity->budget are not updated | |
1609 | * on expiration if bfqq is empty (see | |
1610 | * __bfq_bfqq_recalc_budget). Thus both quantities | |
1611 | * remain unchanged after such an expiration, and the | |
1612 | * following statement therefore assigns to | |
1613 | * entity->budget the remaining budget on such an | |
9fae8dd5 | 1614 | * expiration. |
aee69d78 PV |
1615 | */ |
1616 | entity->budget = min_t(unsigned long, | |
1617 | bfq_bfqq_budget_left(bfqq), | |
1618 | bfqq->max_budget); | |
1619 | ||
9fae8dd5 PV |
1620 | /* |
1621 | * At this point, we have used entity->service to get | |
1622 | * the budget left (needed for updating | |
1623 | * entity->budget). Thus we finally can, and have to, | |
1624 | * reset entity->service. The latter must be reset | |
1625 | * because bfqq would otherwise be charged again for | |
1626 | * the service it has received during its previous | |
1627 | * service slot(s). | |
1628 | */ | |
1629 | entity->service = 0; | |
1630 | ||
aee69d78 PV |
1631 | return true; |
1632 | } | |
1633 | ||
9fae8dd5 PV |
1634 | /* |
1635 | * We can finally complete expiration, by setting service to 0. | |
1636 | */ | |
1637 | entity->service = 0; | |
aee69d78 PV |
1638 | entity->budget = max_t(unsigned long, bfqq->max_budget, |
1639 | bfq_serv_to_charge(bfqq->next_rq, bfqq)); | |
1640 | bfq_clear_bfqq_non_blocking_wait_rq(bfqq); | |
96a291c3 | 1641 | return false; |
44e44a1b PV |
1642 | } |
1643 | ||
4baa8bb1 PV |
1644 | /* |
1645 | * Return the farthest past time instant according to jiffies | |
1646 | * macros. | |
1647 | */ | |
1648 | static unsigned long bfq_smallest_from_now(void) | |
1649 | { | |
1650 | return jiffies - MAX_JIFFY_OFFSET; | |
1651 | } | |
1652 | ||
44e44a1b PV |
1653 | static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd, |
1654 | struct bfq_queue *bfqq, | |
1655 | unsigned int old_wr_coeff, | |
1656 | bool wr_or_deserves_wr, | |
77b7dcea | 1657 | bool interactive, |
e1b2324d | 1658 | bool in_burst, |
77b7dcea | 1659 | bool soft_rt) |
44e44a1b PV |
1660 | { |
1661 | if (old_wr_coeff == 1 && wr_or_deserves_wr) { | |
1662 | /* start a weight-raising period */ | |
77b7dcea | 1663 | if (interactive) { |
8a8747dc | 1664 | bfqq->service_from_wr = 0; |
77b7dcea PV |
1665 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; |
1666 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1667 | } else { | |
4baa8bb1 PV |
1668 | /* |
1669 | * No interactive weight raising in progress | |
1670 | * here: assign minus infinity to | |
1671 | * wr_start_at_switch_to_srt, to make sure | |
1672 | * that, at the end of the soft-real-time | |
1673 | * weight raising periods that is starting | |
1674 | * now, no interactive weight-raising period | |
1675 | * may be wrongly considered as still in | |
1676 | * progress (and thus actually started by | |
1677 | * mistake). | |
1678 | */ | |
1679 | bfqq->wr_start_at_switch_to_srt = | |
1680 | bfq_smallest_from_now(); | |
77b7dcea PV |
1681 | bfqq->wr_coeff = bfqd->bfq_wr_coeff * |
1682 | BFQ_SOFTRT_WEIGHT_FACTOR; | |
1683 | bfqq->wr_cur_max_time = | |
1684 | bfqd->bfq_wr_rt_max_time; | |
1685 | } | |
44e44a1b PV |
1686 | |
1687 | /* | |
1688 | * If needed, further reduce budget to make sure it is | |
1689 | * close to bfqq's backlog, so as to reduce the | |
1690 | * scheduling-error component due to a too large | |
1691 | * budget. Do not care about throughput consequences, | |
1692 | * but only about latency. Finally, do not assign a | |
1693 | * too small budget either, to avoid increasing | |
1694 | * latency by causing too frequent expirations. | |
1695 | */ | |
1696 | bfqq->entity.budget = min_t(unsigned long, | |
1697 | bfqq->entity.budget, | |
1698 | 2 * bfq_min_budget(bfqd)); | |
1699 | } else if (old_wr_coeff > 1) { | |
77b7dcea PV |
1700 | if (interactive) { /* update wr coeff and duration */ |
1701 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1702 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
e1b2324d AA |
1703 | } else if (in_burst) |
1704 | bfqq->wr_coeff = 1; | |
1705 | else if (soft_rt) { | |
77b7dcea PV |
1706 | /* |
1707 | * The application is now or still meeting the | |
1708 | * requirements for being deemed soft rt. We | |
1709 | * can then correctly and safely (re)charge | |
1710 | * the weight-raising duration for the | |
1711 | * application with the weight-raising | |
1712 | * duration for soft rt applications. | |
1713 | * | |
1714 | * In particular, doing this recharge now, i.e., | |
1715 | * before the weight-raising period for the | |
1716 | * application finishes, reduces the probability | |
1717 | * of the following negative scenario: | |
1718 | * 1) the weight of a soft rt application is | |
1719 | * raised at startup (as for any newly | |
1720 | * created application), | |
1721 | * 2) since the application is not interactive, | |
1722 | * at a certain time weight-raising is | |
1723 | * stopped for the application, | |
1724 | * 3) at that time the application happens to | |
1725 | * still have pending requests, and hence | |
1726 | * is destined to not have a chance to be | |
1727 | * deemed soft rt before these requests are | |
1728 | * completed (see the comments to the | |
1729 | * function bfq_bfqq_softrt_next_start() | |
1730 | * for details on soft rt detection), | |
1731 | * 4) these pending requests experience a high | |
1732 | * latency because the application is not | |
1733 | * weight-raised while they are pending. | |
1734 | */ | |
1735 | if (bfqq->wr_cur_max_time != | |
1736 | bfqd->bfq_wr_rt_max_time) { | |
1737 | bfqq->wr_start_at_switch_to_srt = | |
1738 | bfqq->last_wr_start_finish; | |
1739 | ||
1740 | bfqq->wr_cur_max_time = | |
1741 | bfqd->bfq_wr_rt_max_time; | |
1742 | bfqq->wr_coeff = bfqd->bfq_wr_coeff * | |
1743 | BFQ_SOFTRT_WEIGHT_FACTOR; | |
1744 | } | |
1745 | bfqq->last_wr_start_finish = jiffies; | |
1746 | } | |
44e44a1b PV |
1747 | } |
1748 | } | |
1749 | ||
1750 | static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd, | |
1751 | struct bfq_queue *bfqq) | |
1752 | { | |
1753 | return bfqq->dispatched == 0 && | |
1754 | time_is_before_jiffies( | |
1755 | bfqq->budget_timeout + | |
1756 | bfqd->bfq_wr_min_idle_time); | |
aee69d78 PV |
1757 | } |
1758 | ||
96a291c3 PV |
1759 | |
1760 | /* | |
1761 | * Return true if bfqq is in a higher priority class, or has a higher | |
1762 | * weight than the in-service queue. | |
1763 | */ | |
1764 | static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq, | |
1765 | struct bfq_queue *in_serv_bfqq) | |
1766 | { | |
1767 | int bfqq_weight, in_serv_weight; | |
1768 | ||
1769 | if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class) | |
1770 | return true; | |
1771 | ||
1772 | if (in_serv_bfqq->entity.parent == bfqq->entity.parent) { | |
1773 | bfqq_weight = bfqq->entity.weight; | |
1774 | in_serv_weight = in_serv_bfqq->entity.weight; | |
1775 | } else { | |
1776 | if (bfqq->entity.parent) | |
1777 | bfqq_weight = bfqq->entity.parent->weight; | |
1778 | else | |
1779 | bfqq_weight = bfqq->entity.weight; | |
1780 | if (in_serv_bfqq->entity.parent) | |
1781 | in_serv_weight = in_serv_bfqq->entity.parent->weight; | |
1782 | else | |
1783 | in_serv_weight = in_serv_bfqq->entity.weight; | |
1784 | } | |
1785 | ||
1786 | return bfqq_weight > in_serv_weight; | |
1787 | } | |
1788 | ||
9778369a PV |
1789 | /* |
1790 | * Get the index of the actuator that will serve bio. | |
1791 | */ | |
1792 | static unsigned int bfq_actuator_index(struct bfq_data *bfqd, struct bio *bio) | |
1793 | { | |
4fdb3b9f FG |
1794 | unsigned int i; |
1795 | sector_t end; | |
1796 | ||
1797 | /* no search needed if one or zero ranges present */ | |
1798 | if (bfqd->num_actuators == 1) | |
1799 | return 0; | |
1800 | ||
1801 | /* bio_end_sector(bio) gives the sector after the last one */ | |
1802 | end = bio_end_sector(bio) - 1; | |
1803 | ||
1804 | for (i = 0; i < bfqd->num_actuators; i++) { | |
1805 | if (end >= bfqd->sector[i] && | |
1806 | end < bfqd->sector[i] + bfqd->nr_sectors[i]) | |
1807 | return i; | |
1808 | } | |
1809 | ||
1810 | WARN_ONCE(true, | |
1811 | "bfq_actuator_index: bio sector out of ranges: end=%llu\n", | |
1812 | end); | |
9778369a PV |
1813 | return 0; |
1814 | } | |
1815 | ||
7f1995c2 PV |
1816 | static bool bfq_better_to_idle(struct bfq_queue *bfqq); |
1817 | ||
aee69d78 PV |
1818 | static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd, |
1819 | struct bfq_queue *bfqq, | |
44e44a1b PV |
1820 | int old_wr_coeff, |
1821 | struct request *rq, | |
1822 | bool *interactive) | |
aee69d78 | 1823 | { |
e1b2324d AA |
1824 | bool soft_rt, in_burst, wr_or_deserves_wr, |
1825 | bfqq_wants_to_preempt, | |
44e44a1b | 1826 | idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq), |
aee69d78 PV |
1827 | /* |
1828 | * See the comments on | |
1829 | * bfq_bfqq_update_budg_for_activation for | |
1830 | * details on the usage of the next variable. | |
1831 | */ | |
1832 | arrived_in_time = ktime_get_ns() <= | |
1833 | bfqq->ttime.last_end_request + | |
1834 | bfqd->bfq_slice_idle * 3; | |
fd571df0 PV |
1835 | unsigned int act_idx = bfq_actuator_index(bfqd, rq->bio); |
1836 | bool bfqq_non_merged_or_stably_merged = | |
1837 | bfqq->bic || RQ_BIC(rq)->bfqq_data[act_idx].stably_merged; | |
e21b7a0b | 1838 | |
aee69d78 | 1839 | /* |
44e44a1b PV |
1840 | * bfqq deserves to be weight-raised if: |
1841 | * - it is sync, | |
e1b2324d | 1842 | * - it does not belong to a large burst, |
36eca894 | 1843 | * - it has been idle for enough time or is soft real-time, |
91b896f6 PV |
1844 | * - is linked to a bfq_io_cq (it is not shared in any sense), |
1845 | * - has a default weight (otherwise we assume the user wanted | |
1846 | * to control its weight explicitly) | |
44e44a1b | 1847 | */ |
e1b2324d | 1848 | in_burst = bfq_bfqq_in_large_burst(bfqq); |
77b7dcea | 1849 | soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 && |
7074f076 | 1850 | !BFQQ_TOTALLY_SEEKY(bfqq) && |
e1b2324d | 1851 | !in_burst && |
f6c3ca0e | 1852 | time_is_before_jiffies(bfqq->soft_rt_next_start) && |
91b896f6 PV |
1853 | bfqq->dispatched == 0 && |
1854 | bfqq->entity.new_weight == 40; | |
1855 | *interactive = !in_burst && idle_for_long_time && | |
1856 | bfqq->entity.new_weight == 40; | |
511a2699 PV |
1857 | /* |
1858 | * Merged bfq_queues are kept out of weight-raising | |
1859 | * (low-latency) mechanisms. The reason is that these queues | |
1860 | * are usually created for non-interactive and | |
1861 | * non-soft-real-time tasks. Yet this is not the case for | |
1862 | * stably-merged queues. These queues are merged just because | |
1863 | * they are created shortly after each other. So they may | |
1864 | * easily serve the I/O of an interactive or soft-real time | |
1865 | * application, if the application happens to spawn multiple | |
1866 | * processes. So let also stably-merged queued enjoy weight | |
1867 | * raising. | |
1868 | */ | |
44e44a1b PV |
1869 | wr_or_deserves_wr = bfqd->low_latency && |
1870 | (bfqq->wr_coeff > 1 || | |
fd571df0 PV |
1871 | (bfq_bfqq_sync(bfqq) && bfqq_non_merged_or_stably_merged && |
1872 | (*interactive || soft_rt))); | |
44e44a1b PV |
1873 | |
1874 | /* | |
1875 | * Using the last flag, update budget and check whether bfqq | |
1876 | * may want to preempt the in-service queue. | |
aee69d78 PV |
1877 | */ |
1878 | bfqq_wants_to_preempt = | |
1879 | bfq_bfqq_update_budg_for_activation(bfqd, bfqq, | |
96a291c3 | 1880 | arrived_in_time); |
aee69d78 | 1881 | |
e1b2324d AA |
1882 | /* |
1883 | * If bfqq happened to be activated in a burst, but has been | |
1884 | * idle for much more than an interactive queue, then we | |
1885 | * assume that, in the overall I/O initiated in the burst, the | |
1886 | * I/O associated with bfqq is finished. So bfqq does not need | |
1887 | * to be treated as a queue belonging to a burst | |
1888 | * anymore. Accordingly, we reset bfqq's in_large_burst flag | |
1889 | * if set, and remove bfqq from the burst list if it's | |
1890 | * there. We do not decrement burst_size, because the fact | |
1891 | * that bfqq does not need to belong to the burst list any | |
1892 | * more does not invalidate the fact that bfqq was created in | |
1893 | * a burst. | |
1894 | */ | |
1895 | if (likely(!bfq_bfqq_just_created(bfqq)) && | |
1896 | idle_for_long_time && | |
1897 | time_is_before_jiffies( | |
1898 | bfqq->budget_timeout + | |
1899 | msecs_to_jiffies(10000))) { | |
1900 | hlist_del_init(&bfqq->burst_list_node); | |
1901 | bfq_clear_bfqq_in_large_burst(bfqq); | |
1902 | } | |
1903 | ||
1904 | bfq_clear_bfqq_just_created(bfqq); | |
1905 | ||
44e44a1b | 1906 | if (bfqd->low_latency) { |
36eca894 AA |
1907 | if (unlikely(time_is_after_jiffies(bfqq->split_time))) |
1908 | /* wraparound */ | |
1909 | bfqq->split_time = | |
1910 | jiffies - bfqd->bfq_wr_min_idle_time - 1; | |
1911 | ||
1912 | if (time_is_before_jiffies(bfqq->split_time + | |
1913 | bfqd->bfq_wr_min_idle_time)) { | |
1914 | bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq, | |
1915 | old_wr_coeff, | |
1916 | wr_or_deserves_wr, | |
1917 | *interactive, | |
e1b2324d | 1918 | in_burst, |
36eca894 AA |
1919 | soft_rt); |
1920 | ||
1921 | if (old_wr_coeff != bfqq->wr_coeff) | |
1922 | bfqq->entity.prio_changed = 1; | |
1923 | } | |
44e44a1b PV |
1924 | } |
1925 | ||
77b7dcea PV |
1926 | bfqq->last_idle_bklogged = jiffies; |
1927 | bfqq->service_from_backlogged = 0; | |
1928 | bfq_clear_bfqq_softrt_update(bfqq); | |
1929 | ||
d322f355 | 1930 | bfq_add_bfqq_busy(bfqq); |
aee69d78 PV |
1931 | |
1932 | /* | |
7f1995c2 PV |
1933 | * Expire in-service queue if preemption may be needed for |
1934 | * guarantees or throughput. As for guarantees, we care | |
1935 | * explicitly about two cases. The first is that bfqq has to | |
1936 | * recover a service hole, as explained in the comments on | |
96a291c3 PV |
1937 | * bfq_bfqq_update_budg_for_activation(), i.e., that |
1938 | * bfqq_wants_to_preempt is true. However, if bfqq does not | |
1939 | * carry time-critical I/O, then bfqq's bandwidth is less | |
1940 | * important than that of queues that carry time-critical I/O. | |
1941 | * So, as a further constraint, we consider this case only if | |
1942 | * bfqq is at least as weight-raised, i.e., at least as time | |
1943 | * critical, as the in-service queue. | |
1944 | * | |
1945 | * The second case is that bfqq is in a higher priority class, | |
1946 | * or has a higher weight than the in-service queue. If this | |
1947 | * condition does not hold, we don't care because, even if | |
1948 | * bfqq does not start to be served immediately, the resulting | |
1949 | * delay for bfqq's I/O is however lower or much lower than | |
1950 | * the ideal completion time to be guaranteed to bfqq's I/O. | |
1951 | * | |
1952 | * In both cases, preemption is needed only if, according to | |
1953 | * the timestamps of both bfqq and of the in-service queue, | |
1954 | * bfqq actually is the next queue to serve. So, to reduce | |
1955 | * useless preemptions, the return value of | |
1956 | * next_queue_may_preempt() is considered in the next compound | |
1957 | * condition too. Yet next_queue_may_preempt() just checks a | |
1958 | * simple, necessary condition for bfqq to be the next queue | |
1959 | * to serve. In fact, to evaluate a sufficient condition, the | |
1960 | * timestamps of the in-service queue would need to be | |
1961 | * updated, and this operation is quite costly (see the | |
1962 | * comments on bfq_bfqq_update_budg_for_activation()). | |
7f1995c2 PV |
1963 | * |
1964 | * As for throughput, we ask bfq_better_to_idle() whether we | |
1965 | * still need to plug I/O dispatching. If bfq_better_to_idle() | |
1966 | * says no, then plugging is not needed any longer, either to | |
1967 | * boost throughput or to perserve service guarantees. Then | |
1968 | * the best option is to stop plugging I/O, as not doing so | |
1969 | * would certainly lower throughput. We may end up in this | |
1970 | * case if: (1) upon a dispatch attempt, we detected that it | |
1971 | * was better to plug I/O dispatch, and to wait for a new | |
1972 | * request to arrive for the currently in-service queue, but | |
1973 | * (2) this switch of bfqq to busy changes the scenario. | |
aee69d78 | 1974 | */ |
96a291c3 PV |
1975 | if (bfqd->in_service_queue && |
1976 | ((bfqq_wants_to_preempt && | |
1977 | bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) || | |
7f1995c2 PV |
1978 | bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) || |
1979 | !bfq_better_to_idle(bfqd->in_service_queue)) && | |
aee69d78 PV |
1980 | next_queue_may_preempt(bfqd)) |
1981 | bfq_bfqq_expire(bfqd, bfqd->in_service_queue, | |
1982 | false, BFQQE_PREEMPTED); | |
1983 | } | |
1984 | ||
766d6141 PV |
1985 | static void bfq_reset_inject_limit(struct bfq_data *bfqd, |
1986 | struct bfq_queue *bfqq) | |
1987 | { | |
1988 | /* invalidate baseline total service time */ | |
1989 | bfqq->last_serv_time_ns = 0; | |
1990 | ||
1991 | /* | |
1992 | * Reset pointer in case we are waiting for | |
1993 | * some request completion. | |
1994 | */ | |
1995 | bfqd->waited_rq = NULL; | |
1996 | ||
1997 | /* | |
1998 | * If bfqq has a short think time, then start by setting the | |
1999 | * inject limit to 0 prudentially, because the service time of | |
2000 | * an injected I/O request may be higher than the think time | |
2001 | * of bfqq, and therefore, if one request was injected when | |
2002 | * bfqq remains empty, this injected request might delay the | |
2003 | * service of the next I/O request for bfqq significantly. In | |
2004 | * case bfqq can actually tolerate some injection, then the | |
2005 | * adaptive update will however raise the limit soon. This | |
2006 | * lucky circumstance holds exactly because bfqq has a short | |
2007 | * think time, and thus, after remaining empty, is likely to | |
2008 | * get new I/O enqueued---and then completed---before being | |
2009 | * expired. This is the very pattern that gives the | |
2010 | * limit-update algorithm the chance to measure the effect of | |
2011 | * injection on request service times, and then to update the | |
2012 | * limit accordingly. | |
2013 | * | |
2014 | * However, in the following special case, the inject limit is | |
2015 | * left to 1 even if the think time is short: bfqq's I/O is | |
2016 | * synchronized with that of some other queue, i.e., bfqq may | |
2017 | * receive new I/O only after the I/O of the other queue is | |
2018 | * completed. Keeping the inject limit to 1 allows the | |
2019 | * blocking I/O to be served while bfqq is in service. And | |
2020 | * this is very convenient both for bfqq and for overall | |
2021 | * throughput, as explained in detail in the comments in | |
2022 | * bfq_update_has_short_ttime(). | |
2023 | * | |
2024 | * On the opposite end, if bfqq has a long think time, then | |
2025 | * start directly by 1, because: | |
2026 | * a) on the bright side, keeping at most one request in | |
2027 | * service in the drive is unlikely to cause any harm to the | |
2028 | * latency of bfqq's requests, as the service time of a single | |
2029 | * request is likely to be lower than the think time of bfqq; | |
2030 | * b) on the downside, after becoming empty, bfqq is likely to | |
2031 | * expire before getting its next request. With this request | |
2032 | * arrival pattern, it is very hard to sample total service | |
2033 | * times and update the inject limit accordingly (see comments | |
2034 | * on bfq_update_inject_limit()). So the limit is likely to be | |
2035 | * never, or at least seldom, updated. As a consequence, by | |
2036 | * setting the limit to 1, we avoid that no injection ever | |
2037 | * occurs with bfqq. On the downside, this proactive step | |
2038 | * further reduces chances to actually compute the baseline | |
2039 | * total service time. Thus it reduces chances to execute the | |
2040 | * limit-update algorithm and possibly raise the limit to more | |
2041 | * than 1. | |
2042 | */ | |
2043 | if (bfq_bfqq_has_short_ttime(bfqq)) | |
2044 | bfqq->inject_limit = 0; | |
2045 | else | |
2046 | bfqq->inject_limit = 1; | |
2047 | ||
2048 | bfqq->decrease_time_jif = jiffies; | |
2049 | } | |
2050 | ||
eb2fd80f PV |
2051 | static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns) |
2052 | { | |
2053 | u64 tot_io_time = now_ns - bfqq->io_start_time; | |
2054 | ||
2055 | if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0) | |
2056 | bfqq->tot_idle_time += | |
2057 | now_ns - bfqq->ttime.last_end_request; | |
2058 | ||
2059 | if (unlikely(bfq_bfqq_just_created(bfqq))) | |
2060 | return; | |
2061 | ||
2062 | /* | |
2063 | * Must be busy for at least about 80% of the time to be | |
2064 | * considered I/O bound. | |
2065 | */ | |
2066 | if (bfqq->tot_idle_time * 5 > tot_io_time) | |
2067 | bfq_clear_bfqq_IO_bound(bfqq); | |
2068 | else | |
2069 | bfq_mark_bfqq_IO_bound(bfqq); | |
2070 | ||
2071 | /* | |
2072 | * Keep an observation window of at most 200 ms in the past | |
2073 | * from now. | |
2074 | */ | |
2075 | if (tot_io_time > 200 * NSEC_PER_MSEC) { | |
2076 | bfqq->io_start_time = now_ns - (tot_io_time>>1); | |
2077 | bfqq->tot_idle_time >>= 1; | |
2078 | } | |
2079 | } | |
2080 | ||
71217df3 PV |
2081 | /* |
2082 | * Detect whether bfqq's I/O seems synchronized with that of some | |
2083 | * other queue, i.e., whether bfqq, after remaining empty, happens to | |
2084 | * receive new I/O only right after some I/O request of the other | |
2085 | * queue has been completed. We call waker queue the other queue, and | |
2086 | * we assume, for simplicity, that bfqq may have at most one waker | |
2087 | * queue. | |
2088 | * | |
2089 | * A remarkable throughput boost can be reached by unconditionally | |
2090 | * injecting the I/O of the waker queue, every time a new | |
2091 | * bfq_dispatch_request happens to be invoked while I/O is being | |
2092 | * plugged for bfqq. In addition to boosting throughput, this | |
2093 | * unblocks bfqq's I/O, thereby improving bandwidth and latency for | |
2094 | * bfqq. Note that these same results may be achieved with the general | |
2095 | * injection mechanism, but less effectively. For details on this | |
2096 | * aspect, see the comments on the choice of the queue for injection | |
2097 | * in bfq_select_queue(). | |
2098 | * | |
1f18b700 JK |
2099 | * Turning back to the detection of a waker queue, a queue Q is deemed as a |
2100 | * waker queue for bfqq if, for three consecutive times, bfqq happens to become | |
2101 | * non empty right after a request of Q has been completed within given | |
2102 | * timeout. In this respect, even if bfqq is empty, we do not check for a waker | |
2103 | * if it still has some in-flight I/O. In fact, in this case bfqq is actually | |
2104 | * still being served by the drive, and may receive new I/O on the completion | |
2105 | * of some of the in-flight requests. In particular, on the first time, Q is | |
2106 | * tentatively set as a candidate waker queue, while on the third consecutive | |
2107 | * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q | |
2108 | * is a waker queue for bfqq. These detection steps are performed only if bfqq | |
2109 | * has a long think time, so as to make it more likely that bfqq's I/O is | |
2110 | * actually being blocked by a synchronization. This last filter, plus the | |
2111 | * above three-times requirement and time limit for detection, make false | |
efc72524 | 2112 | * positives less likely. |
71217df3 PV |
2113 | * |
2114 | * NOTE | |
2115 | * | |
2116 | * The sooner a waker queue is detected, the sooner throughput can be | |
2117 | * boosted by injecting I/O from the waker queue. Fortunately, | |
2118 | * detection is likely to be actually fast, for the following | |
2119 | * reasons. While blocked by synchronization, bfqq has a long think | |
2120 | * time. This implies that bfqq's inject limit is at least equal to 1 | |
2121 | * (see the comments in bfq_update_inject_limit()). So, thanks to | |
2122 | * injection, the waker queue is likely to be served during the very | |
2123 | * first I/O-plugging time interval for bfqq. This triggers the first | |
2124 | * step of the detection mechanism. Thanks again to injection, the | |
2125 | * candidate waker queue is then likely to be confirmed no later than | |
2126 | * during the next I/O-plugging interval for bfqq. | |
2127 | * | |
2128 | * ISSUE | |
2129 | * | |
2130 | * On queue merging all waker information is lost. | |
2131 | */ | |
a5bf0a92 JA |
2132 | static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
2133 | u64 now_ns) | |
71217df3 | 2134 | { |
1eb17f5e JK |
2135 | char waker_name[MAX_BFQQ_NAME_LENGTH]; |
2136 | ||
71217df3 PV |
2137 | if (!bfqd->last_completed_rq_bfqq || |
2138 | bfqd->last_completed_rq_bfqq == bfqq || | |
2139 | bfq_bfqq_has_short_ttime(bfqq) || | |
99771d73 KK |
2140 | now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC || |
2141 | bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq || | |
2142 | bfqq == &bfqd->oom_bfqq) | |
71217df3 PV |
2143 | return; |
2144 | ||
1f18b700 JK |
2145 | /* |
2146 | * We reset waker detection logic also if too much time has passed | |
2147 | * since the first detection. If wakeups are rare, pointless idling | |
2148 | * doesn't hurt throughput that much. The condition below makes sure | |
9778369a | 2149 | * we do not uselessly idle blocking waker in more than 1/64 cases. |
1f18b700 | 2150 | */ |
71217df3 | 2151 | if (bfqd->last_completed_rq_bfqq != |
1f18b700 JK |
2152 | bfqq->tentative_waker_bfqq || |
2153 | now_ns > bfqq->waker_detection_started + | |
2154 | 128 * (u64)bfqd->bfq_slice_idle) { | |
71217df3 PV |
2155 | /* |
2156 | * First synchronization detected with a | |
2157 | * candidate waker queue, or with a different | |
2158 | * candidate waker queue from the current one. | |
2159 | */ | |
2160 | bfqq->tentative_waker_bfqq = | |
2161 | bfqd->last_completed_rq_bfqq; | |
2162 | bfqq->num_waker_detections = 1; | |
1f18b700 | 2163 | bfqq->waker_detection_started = now_ns; |
1eb17f5e JK |
2164 | bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name, |
2165 | MAX_BFQQ_NAME_LENGTH); | |
8ef22dc4 | 2166 | bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name); |
71217df3 PV |
2167 | } else /* Same tentative waker queue detected again */ |
2168 | bfqq->num_waker_detections++; | |
2169 | ||
2170 | if (bfqq->num_waker_detections == 3) { | |
2171 | bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq; | |
2172 | bfqq->tentative_waker_bfqq = NULL; | |
1eb17f5e JK |
2173 | bfq_bfqq_name(bfqq->waker_bfqq, waker_name, |
2174 | MAX_BFQQ_NAME_LENGTH); | |
2175 | bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name); | |
71217df3 PV |
2176 | |
2177 | /* | |
2178 | * If the waker queue disappears, then | |
2179 | * bfqq->waker_bfqq must be reset. To | |
2180 | * this goal, we maintain in each | |
2181 | * waker queue a list, woken_list, of | |
2182 | * all the queues that reference the | |
2183 | * waker queue through their | |
2184 | * waker_bfqq pointer. When the waker | |
2185 | * queue exits, the waker_bfqq pointer | |
2186 | * of all the queues in the woken_list | |
2187 | * is reset. | |
2188 | * | |
2189 | * In addition, if bfqq is already in | |
2190 | * the woken_list of a waker queue, | |
2191 | * then, before being inserted into | |
2192 | * the woken_list of a new waker | |
2193 | * queue, bfqq must be removed from | |
2194 | * the woken_list of the old waker | |
2195 | * queue. | |
2196 | */ | |
2197 | if (!hlist_unhashed(&bfqq->woken_list_node)) | |
2198 | hlist_del_init(&bfqq->woken_list_node); | |
2199 | hlist_add_head(&bfqq->woken_list_node, | |
2200 | &bfqd->last_completed_rq_bfqq->woken_list); | |
2201 | } | |
2202 | } | |
2203 | ||
aee69d78 PV |
2204 | static void bfq_add_request(struct request *rq) |
2205 | { | |
2206 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
2207 | struct bfq_data *bfqd = bfqq->bfqd; | |
2208 | struct request *next_rq, *prev; | |
44e44a1b PV |
2209 | unsigned int old_wr_coeff = bfqq->wr_coeff; |
2210 | bool interactive = false; | |
eb2fd80f | 2211 | u64 now_ns = ktime_get_ns(); |
aee69d78 PV |
2212 | |
2213 | bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq)); | |
2214 | bfqq->queued[rq_is_sync(rq)]++; | |
ddc25c86 YK |
2215 | /* |
2216 | * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it | |
2217 | * may be read without holding the lock in bfq_has_work(). | |
2218 | */ | |
2219 | WRITE_ONCE(bfqd->queued, bfqd->queued + 1); | |
aee69d78 | 2220 | |
f9506673 | 2221 | if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) { |
71217df3 | 2222 | bfq_check_waker(bfqd, bfqq, now_ns); |
13a857a4 | 2223 | |
2341d662 PV |
2224 | /* |
2225 | * Periodically reset inject limit, to make sure that | |
2226 | * the latter eventually drops in case workload | |
2227 | * changes, see step (3) in the comments on | |
2228 | * bfq_update_inject_limit(). | |
2229 | */ | |
2230 | if (time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
766d6141 PV |
2231 | msecs_to_jiffies(1000))) |
2232 | bfq_reset_inject_limit(bfqd, bfqq); | |
2341d662 PV |
2233 | |
2234 | /* | |
2235 | * The following conditions must hold to setup a new | |
2236 | * sampling of total service time, and then a new | |
2237 | * update of the inject limit: | |
2238 | * - bfqq is in service, because the total service | |
2239 | * time is evaluated only for the I/O requests of | |
2240 | * the queues in service; | |
2241 | * - this is the right occasion to compute or to | |
2242 | * lower the baseline total service time, because | |
2243 | * there are actually no requests in the drive, | |
2244 | * or | |
2245 | * the baseline total service time is available, and | |
2246 | * this is the right occasion to compute the other | |
2247 | * quantity needed to update the inject limit, i.e., | |
2248 | * the total service time caused by the amount of | |
2249 | * injection allowed by the current value of the | |
2250 | * limit. It is the right occasion because injection | |
2251 | * has actually been performed during the service | |
2252 | * hole, and there are still in-flight requests, | |
2253 | * which are very likely to be exactly the injected | |
2254 | * requests, or part of them; | |
2255 | * - the minimum interval for sampling the total | |
2256 | * service time and updating the inject limit has | |
2257 | * elapsed. | |
2258 | */ | |
2259 | if (bfqq == bfqd->in_service_queue && | |
2d31c684 | 2260 | (bfqd->tot_rq_in_driver == 0 || |
2341d662 | 2261 | (bfqq->last_serv_time_ns > 0 && |
2d31c684 | 2262 | bfqd->rqs_injected && bfqd->tot_rq_in_driver > 0)) && |
2341d662 | 2263 | time_is_before_eq_jiffies(bfqq->decrease_time_jif + |
17c3d266 | 2264 | msecs_to_jiffies(10))) { |
2341d662 PV |
2265 | bfqd->last_empty_occupied_ns = ktime_get_ns(); |
2266 | /* | |
2267 | * Start the state machine for measuring the | |
2268 | * total service time of rq: setting | |
2269 | * wait_dispatch will cause bfqd->waited_rq to | |
2270 | * be set when rq will be dispatched. | |
2271 | */ | |
2272 | bfqd->wait_dispatch = true; | |
23ed570a PV |
2273 | /* |
2274 | * If there is no I/O in service in the drive, | |
2275 | * then possible injection occurred before the | |
2276 | * arrival of rq will not affect the total | |
2277 | * service time of rq. So the injection limit | |
2278 | * must not be updated as a function of such | |
2279 | * total service time, unless new injection | |
2280 | * occurs before rq is completed. To have the | |
2281 | * injection limit updated only in the latter | |
2282 | * case, reset rqs_injected here (rqs_injected | |
2283 | * will be set in case injection is performed | |
2284 | * on bfqq before rq is completed). | |
2285 | */ | |
2d31c684 | 2286 | if (bfqd->tot_rq_in_driver == 0) |
23ed570a | 2287 | bfqd->rqs_injected = false; |
2341d662 PV |
2288 | } |
2289 | } | |
2290 | ||
eb2fd80f PV |
2291 | if (bfq_bfqq_sync(bfqq)) |
2292 | bfq_update_io_intensity(bfqq, now_ns); | |
2293 | ||
aee69d78 PV |
2294 | elv_rb_add(&bfqq->sort_list, rq); |
2295 | ||
2296 | /* | |
2297 | * Check if this request is a better next-serve candidate. | |
2298 | */ | |
2299 | prev = bfqq->next_rq; | |
2300 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); | |
2301 | bfqq->next_rq = next_rq; | |
2302 | ||
36eca894 AA |
2303 | /* |
2304 | * Adjust priority tree position, if next_rq changes. | |
8cacc5ab | 2305 | * See comments on bfq_pos_tree_add_move() for the unlikely(). |
36eca894 | 2306 | */ |
8cacc5ab | 2307 | if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq)) |
36eca894 AA |
2308 | bfq_pos_tree_add_move(bfqd, bfqq); |
2309 | ||
aee69d78 | 2310 | if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */ |
44e44a1b PV |
2311 | bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff, |
2312 | rq, &interactive); | |
2313 | else { | |
2314 | if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) && | |
2315 | time_is_before_jiffies( | |
2316 | bfqq->last_wr_start_finish + | |
2317 | bfqd->bfq_wr_min_inter_arr_async)) { | |
2318 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
2319 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
2320 | ||
cfd69712 | 2321 | bfqd->wr_busy_queues++; |
44e44a1b PV |
2322 | bfqq->entity.prio_changed = 1; |
2323 | } | |
2324 | if (prev != bfqq->next_rq) | |
2325 | bfq_updated_next_req(bfqd, bfqq); | |
2326 | } | |
2327 | ||
2328 | /* | |
2329 | * Assign jiffies to last_wr_start_finish in the following | |
2330 | * cases: | |
2331 | * | |
2332 | * . if bfqq is not going to be weight-raised, because, for | |
2333 | * non weight-raised queues, last_wr_start_finish stores the | |
2334 | * arrival time of the last request; as of now, this piece | |
2335 | * of information is used only for deciding whether to | |
2336 | * weight-raise async queues | |
2337 | * | |
2338 | * . if bfqq is not weight-raised, because, if bfqq is now | |
2339 | * switching to weight-raised, then last_wr_start_finish | |
2340 | * stores the time when weight-raising starts | |
2341 | * | |
2342 | * . if bfqq is interactive, because, regardless of whether | |
2343 | * bfqq is currently weight-raised, the weight-raising | |
2344 | * period must start or restart (this case is considered | |
2345 | * separately because it is not detected by the above | |
2346 | * conditions, if bfqq is already weight-raised) | |
77b7dcea PV |
2347 | * |
2348 | * last_wr_start_finish has to be updated also if bfqq is soft | |
2349 | * real-time, because the weight-raising period is constantly | |
2350 | * restarted on idle-to-busy transitions for these queues, but | |
2351 | * this is already done in bfq_bfqq_handle_idle_busy_switch if | |
2352 | * needed. | |
44e44a1b PV |
2353 | */ |
2354 | if (bfqd->low_latency && | |
2355 | (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive)) | |
2356 | bfqq->last_wr_start_finish = jiffies; | |
aee69d78 PV |
2357 | } |
2358 | ||
2359 | static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, | |
2360 | struct bio *bio, | |
2361 | struct request_queue *q) | |
2362 | { | |
2363 | struct bfq_queue *bfqq = bfqd->bio_bfqq; | |
2364 | ||
2365 | ||
2366 | if (bfqq) | |
2367 | return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); | |
2368 | ||
2369 | return NULL; | |
2370 | } | |
2371 | ||
ab0e43e9 PV |
2372 | static sector_t get_sdist(sector_t last_pos, struct request *rq) |
2373 | { | |
2374 | if (last_pos) | |
2375 | return abs(blk_rq_pos(rq) - last_pos); | |
2376 | ||
2377 | return 0; | |
2378 | } | |
2379 | ||
aee69d78 PV |
2380 | static void bfq_remove_request(struct request_queue *q, |
2381 | struct request *rq) | |
2382 | { | |
2383 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
2384 | struct bfq_data *bfqd = bfqq->bfqd; | |
2385 | const int sync = rq_is_sync(rq); | |
2386 | ||
2387 | if (bfqq->next_rq == rq) { | |
2388 | bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); | |
2389 | bfq_updated_next_req(bfqd, bfqq); | |
2390 | } | |
2391 | ||
2392 | if (rq->queuelist.prev != &rq->queuelist) | |
2393 | list_del_init(&rq->queuelist); | |
2394 | bfqq->queued[sync]--; | |
ddc25c86 YK |
2395 | /* |
2396 | * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it | |
2397 | * may be read without holding the lock in bfq_has_work(). | |
2398 | */ | |
2399 | WRITE_ONCE(bfqd->queued, bfqd->queued - 1); | |
aee69d78 PV |
2400 | elv_rb_del(&bfqq->sort_list, rq); |
2401 | ||
2402 | elv_rqhash_del(q, rq); | |
2403 | if (q->last_merge == rq) | |
2404 | q->last_merge = NULL; | |
2405 | ||
2406 | if (RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
2407 | bfqq->next_rq = NULL; | |
2408 | ||
2409 | if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) { | |
d322f355 | 2410 | bfq_del_bfqq_busy(bfqq, false); |
aee69d78 PV |
2411 | /* |
2412 | * bfqq emptied. In normal operation, when | |
2413 | * bfqq is empty, bfqq->entity.service and | |
2414 | * bfqq->entity.budget must contain, | |
2415 | * respectively, the service received and the | |
2416 | * budget used last time bfqq emptied. These | |
2417 | * facts do not hold in this case, as at least | |
2418 | * this last removal occurred while bfqq is | |
2419 | * not in service. To avoid inconsistencies, | |
2420 | * reset both bfqq->entity.service and | |
2421 | * bfqq->entity.budget, if bfqq has still a | |
2422 | * process that may issue I/O requests to it. | |
2423 | */ | |
2424 | bfqq->entity.budget = bfqq->entity.service = 0; | |
2425 | } | |
36eca894 AA |
2426 | |
2427 | /* | |
2428 | * Remove queue from request-position tree as it is empty. | |
2429 | */ | |
2430 | if (bfqq->pos_root) { | |
2431 | rb_erase(&bfqq->pos_node, bfqq->pos_root); | |
2432 | bfqq->pos_root = NULL; | |
2433 | } | |
05e90283 | 2434 | } else { |
8cacc5ab PV |
2435 | /* see comments on bfq_pos_tree_add_move() for the unlikely() */ |
2436 | if (unlikely(!bfqd->nonrot_with_queueing)) | |
2437 | bfq_pos_tree_add_move(bfqd, bfqq); | |
aee69d78 PV |
2438 | } |
2439 | ||
2440 | if (rq->cmd_flags & REQ_META) | |
2441 | bfqq->meta_pending--; | |
e21b7a0b | 2442 | |
aee69d78 PV |
2443 | } |
2444 | ||
efed9a33 | 2445 | static bool bfq_bio_merge(struct request_queue *q, struct bio *bio, |
14ccb66b | 2446 | unsigned int nr_segs) |
aee69d78 | 2447 | { |
aee69d78 PV |
2448 | struct bfq_data *bfqd = q->elevator->elevator_data; |
2449 | struct request *free = NULL; | |
2450 | /* | |
2451 | * bfq_bic_lookup grabs the queue_lock: invoke it now and | |
2452 | * store its return value for later use, to avoid nesting | |
2453 | * queue_lock inside the bfqd->lock. We assume that the bic | |
2454 | * returned by bfq_bic_lookup does not go away before | |
2455 | * bfqd->lock is taken. | |
2456 | */ | |
836b394b | 2457 | struct bfq_io_cq *bic = bfq_bic_lookup(q); |
aee69d78 PV |
2458 | bool ret; |
2459 | ||
2460 | spin_lock_irq(&bfqd->lock); | |
2461 | ||
ea591cd4 JK |
2462 | if (bic) { |
2463 | /* | |
2464 | * Make sure cgroup info is uptodate for current process before | |
2465 | * considering the merge. | |
2466 | */ | |
2467 | bfq_bic_update_cgroup(bic, bio); | |
2468 | ||
9778369a PV |
2469 | bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf), |
2470 | bfq_actuator_index(bfqd, bio)); | |
ea591cd4 | 2471 | } else { |
aee69d78 | 2472 | bfqd->bio_bfqq = NULL; |
ea591cd4 | 2473 | } |
aee69d78 PV |
2474 | bfqd->bio_bic = bic; |
2475 | ||
14ccb66b | 2476 | ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free); |
aee69d78 | 2477 | |
fd2ef39c | 2478 | spin_unlock_irq(&bfqd->lock); |
aee69d78 PV |
2479 | if (free) |
2480 | blk_mq_free_request(free); | |
aee69d78 PV |
2481 | |
2482 | return ret; | |
2483 | } | |
2484 | ||
2485 | static int bfq_request_merge(struct request_queue *q, struct request **req, | |
2486 | struct bio *bio) | |
2487 | { | |
2488 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2489 | struct request *__rq; | |
2490 | ||
2491 | __rq = bfq_find_rq_fmerge(bfqd, bio, q); | |
2492 | if (__rq && elv_bio_merge_ok(__rq, bio)) { | |
2493 | *req = __rq; | |
866663b7 ML |
2494 | |
2495 | if (blk_discard_mergable(__rq)) | |
2496 | return ELEVATOR_DISCARD_MERGE; | |
aee69d78 PV |
2497 | return ELEVATOR_FRONT_MERGE; |
2498 | } | |
2499 | ||
2500 | return ELEVATOR_NO_MERGE; | |
2501 | } | |
2502 | ||
2503 | static void bfq_request_merged(struct request_queue *q, struct request *req, | |
2504 | enum elv_merge type) | |
2505 | { | |
2506 | if (type == ELEVATOR_FRONT_MERGE && | |
2507 | rb_prev(&req->rb_node) && | |
2508 | blk_rq_pos(req) < | |
2509 | blk_rq_pos(container_of(rb_prev(&req->rb_node), | |
2510 | struct request, rb_node))) { | |
5f550ede | 2511 | struct bfq_queue *bfqq = RQ_BFQQ(req); |
fd03177c | 2512 | struct bfq_data *bfqd; |
aee69d78 PV |
2513 | struct request *prev, *next_rq; |
2514 | ||
fd03177c PV |
2515 | if (!bfqq) |
2516 | return; | |
2517 | ||
2518 | bfqd = bfqq->bfqd; | |
2519 | ||
aee69d78 PV |
2520 | /* Reposition request in its sort_list */ |
2521 | elv_rb_del(&bfqq->sort_list, req); | |
2522 | elv_rb_add(&bfqq->sort_list, req); | |
2523 | ||
2524 | /* Choose next request to be served for bfqq */ | |
2525 | prev = bfqq->next_rq; | |
2526 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req, | |
2527 | bfqd->last_position); | |
2528 | bfqq->next_rq = next_rq; | |
2529 | /* | |
36eca894 AA |
2530 | * If next_rq changes, update both the queue's budget to |
2531 | * fit the new request and the queue's position in its | |
2532 | * rq_pos_tree. | |
aee69d78 | 2533 | */ |
36eca894 | 2534 | if (prev != bfqq->next_rq) { |
aee69d78 | 2535 | bfq_updated_next_req(bfqd, bfqq); |
8cacc5ab PV |
2536 | /* |
2537 | * See comments on bfq_pos_tree_add_move() for | |
2538 | * the unlikely(). | |
2539 | */ | |
2540 | if (unlikely(!bfqd->nonrot_with_queueing)) | |
2541 | bfq_pos_tree_add_move(bfqd, bfqq); | |
36eca894 | 2542 | } |
aee69d78 PV |
2543 | } |
2544 | } | |
2545 | ||
8abfa4d6 PV |
2546 | /* |
2547 | * This function is called to notify the scheduler that the requests | |
2548 | * rq and 'next' have been merged, with 'next' going away. BFQ | |
2549 | * exploits this hook to address the following issue: if 'next' has a | |
2550 | * fifo_time lower that rq, then the fifo_time of rq must be set to | |
2551 | * the value of 'next', to not forget the greater age of 'next'. | |
8abfa4d6 PV |
2552 | * |
2553 | * NOTE: in this function we assume that rq is in a bfq_queue, basing | |
2554 | * on that rq is picked from the hash table q->elevator->hash, which, | |
2555 | * in its turn, is filled only with I/O requests present in | |
2556 | * bfq_queues, while BFQ is in use for the request queue q. In fact, | |
2557 | * the function that fills this hash table (elv_rqhash_add) is called | |
2558 | * only by bfq_insert_request. | |
2559 | */ | |
aee69d78 PV |
2560 | static void bfq_requests_merged(struct request_queue *q, struct request *rq, |
2561 | struct request *next) | |
2562 | { | |
5f550ede JK |
2563 | struct bfq_queue *bfqq = RQ_BFQQ(rq), |
2564 | *next_bfqq = RQ_BFQQ(next); | |
aee69d78 | 2565 | |
fd03177c | 2566 | if (!bfqq) |
a921c655 | 2567 | goto remove; |
fd03177c | 2568 | |
aee69d78 PV |
2569 | /* |
2570 | * If next and rq belong to the same bfq_queue and next is older | |
2571 | * than rq, then reposition rq in the fifo (by substituting next | |
2572 | * with rq). Otherwise, if next and rq belong to different | |
2573 | * bfq_queues, never reposition rq: in fact, we would have to | |
2574 | * reposition it with respect to next's position in its own fifo, | |
2575 | * which would most certainly be too expensive with respect to | |
2576 | * the benefits. | |
2577 | */ | |
2578 | if (bfqq == next_bfqq && | |
2579 | !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
2580 | next->fifo_time < rq->fifo_time) { | |
2581 | list_del_init(&rq->queuelist); | |
2582 | list_replace_init(&next->queuelist, &rq->queuelist); | |
2583 | rq->fifo_time = next->fifo_time; | |
2584 | } | |
2585 | ||
2586 | if (bfqq->next_rq == next) | |
2587 | bfqq->next_rq = rq; | |
2588 | ||
e21b7a0b | 2589 | bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags); |
a921c655 JK |
2590 | remove: |
2591 | /* Merged request may be in the IO scheduler. Remove it. */ | |
2592 | if (!RB_EMPTY_NODE(&next->rb_node)) { | |
2593 | bfq_remove_request(next->q, next); | |
2594 | if (next_bfqq) | |
2595 | bfqg_stats_update_io_remove(bfqq_group(next_bfqq), | |
2596 | next->cmd_flags); | |
2597 | } | |
aee69d78 PV |
2598 | } |
2599 | ||
44e44a1b PV |
2600 | /* Must be called with bfqq != NULL */ |
2601 | static void bfq_bfqq_end_wr(struct bfq_queue *bfqq) | |
2602 | { | |
3c337690 PV |
2603 | /* |
2604 | * If bfqq has been enjoying interactive weight-raising, then | |
2605 | * reset soft_rt_next_start. We do it for the following | |
2606 | * reason. bfqq may have been conveying the I/O needed to load | |
2607 | * a soft real-time application. Such an application actually | |
2608 | * exhibits a soft real-time I/O pattern after it finishes | |
2609 | * loading, and finally starts doing its job. But, if bfqq has | |
2610 | * been receiving a lot of bandwidth so far (likely to happen | |
2611 | * on a fast device), then soft_rt_next_start now contains a | |
2612 | * high value that. So, without this reset, bfqq would be | |
2613 | * prevented from being possibly considered as soft_rt for a | |
2614 | * very long time. | |
2615 | */ | |
2616 | ||
2617 | if (bfqq->wr_cur_max_time != | |
2618 | bfqq->bfqd->bfq_wr_rt_max_time) | |
2619 | bfqq->soft_rt_next_start = jiffies; | |
2620 | ||
cfd69712 PV |
2621 | if (bfq_bfqq_busy(bfqq)) |
2622 | bfqq->bfqd->wr_busy_queues--; | |
44e44a1b PV |
2623 | bfqq->wr_coeff = 1; |
2624 | bfqq->wr_cur_max_time = 0; | |
77b7dcea | 2625 | bfqq->last_wr_start_finish = jiffies; |
44e44a1b PV |
2626 | /* |
2627 | * Trigger a weight change on the next invocation of | |
2628 | * __bfq_entity_update_weight_prio. | |
2629 | */ | |
2630 | bfqq->entity.prio_changed = 1; | |
2631 | } | |
2632 | ||
ea25da48 PV |
2633 | void bfq_end_wr_async_queues(struct bfq_data *bfqd, |
2634 | struct bfq_group *bfqg) | |
44e44a1b | 2635 | { |
8b7fd741 | 2636 | int i, j, k; |
44e44a1b | 2637 | |
8b7fd741 DZ |
2638 | for (k = 0; k < bfqd->num_actuators; k++) { |
2639 | for (i = 0; i < 2; i++) | |
2640 | for (j = 0; j < IOPRIO_NR_LEVELS; j++) | |
2641 | if (bfqg->async_bfqq[i][j][k]) | |
2642 | bfq_bfqq_end_wr(bfqg->async_bfqq[i][j][k]); | |
2643 | if (bfqg->async_idle_bfqq[k]) | |
2644 | bfq_bfqq_end_wr(bfqg->async_idle_bfqq[k]); | |
2645 | } | |
44e44a1b PV |
2646 | } |
2647 | ||
2648 | static void bfq_end_wr(struct bfq_data *bfqd) | |
2649 | { | |
2650 | struct bfq_queue *bfqq; | |
2d31c684 | 2651 | int i; |
44e44a1b PV |
2652 | |
2653 | spin_lock_irq(&bfqd->lock); | |
2654 | ||
2d31c684 DZ |
2655 | for (i = 0; i < bfqd->num_actuators; i++) { |
2656 | list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list) | |
2657 | bfq_bfqq_end_wr(bfqq); | |
2658 | } | |
44e44a1b PV |
2659 | list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) |
2660 | bfq_bfqq_end_wr(bfqq); | |
2661 | bfq_end_wr_async(bfqd); | |
2662 | ||
2663 | spin_unlock_irq(&bfqd->lock); | |
2664 | } | |
2665 | ||
36eca894 AA |
2666 | static sector_t bfq_io_struct_pos(void *io_struct, bool request) |
2667 | { | |
2668 | if (request) | |
2669 | return blk_rq_pos(io_struct); | |
2670 | else | |
2671 | return ((struct bio *)io_struct)->bi_iter.bi_sector; | |
2672 | } | |
2673 | ||
2674 | static int bfq_rq_close_to_sector(void *io_struct, bool request, | |
2675 | sector_t sector) | |
2676 | { | |
2677 | return abs(bfq_io_struct_pos(io_struct, request) - sector) <= | |
2678 | BFQQ_CLOSE_THR; | |
2679 | } | |
2680 | ||
2681 | static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd, | |
2682 | struct bfq_queue *bfqq, | |
2683 | sector_t sector) | |
2684 | { | |
43a4b1fe | 2685 | struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree; |
36eca894 AA |
2686 | struct rb_node *parent, *node; |
2687 | struct bfq_queue *__bfqq; | |
2688 | ||
2689 | if (RB_EMPTY_ROOT(root)) | |
2690 | return NULL; | |
2691 | ||
2692 | /* | |
2693 | * First, if we find a request starting at the end of the last | |
2694 | * request, choose it. | |
2695 | */ | |
2696 | __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); | |
2697 | if (__bfqq) | |
2698 | return __bfqq; | |
2699 | ||
2700 | /* | |
2701 | * If the exact sector wasn't found, the parent of the NULL leaf | |
2702 | * will contain the closest sector (rq_pos_tree sorted by | |
2703 | * next_request position). | |
2704 | */ | |
2705 | __bfqq = rb_entry(parent, struct bfq_queue, pos_node); | |
2706 | if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) | |
2707 | return __bfqq; | |
2708 | ||
2709 | if (blk_rq_pos(__bfqq->next_rq) < sector) | |
2710 | node = rb_next(&__bfqq->pos_node); | |
2711 | else | |
2712 | node = rb_prev(&__bfqq->pos_node); | |
2713 | if (!node) | |
2714 | return NULL; | |
2715 | ||
2716 | __bfqq = rb_entry(node, struct bfq_queue, pos_node); | |
2717 | if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) | |
2718 | return __bfqq; | |
2719 | ||
2720 | return NULL; | |
2721 | } | |
2722 | ||
2723 | static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd, | |
2724 | struct bfq_queue *cur_bfqq, | |
2725 | sector_t sector) | |
2726 | { | |
2727 | struct bfq_queue *bfqq; | |
2728 | ||
2729 | /* | |
2730 | * We shall notice if some of the queues are cooperating, | |
2731 | * e.g., working closely on the same area of the device. In | |
2732 | * that case, we can group them together and: 1) don't waste | |
2733 | * time idling, and 2) serve the union of their requests in | |
2734 | * the best possible order for throughput. | |
2735 | */ | |
2736 | bfqq = bfqq_find_close(bfqd, cur_bfqq, sector); | |
2737 | if (!bfqq || bfqq == cur_bfqq) | |
2738 | return NULL; | |
2739 | ||
2740 | return bfqq; | |
2741 | } | |
2742 | ||
2743 | static struct bfq_queue * | |
2744 | bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) | |
2745 | { | |
2746 | int process_refs, new_process_refs; | |
2747 | struct bfq_queue *__bfqq; | |
2748 | ||
2749 | /* | |
2750 | * If there are no process references on the new_bfqq, then it is | |
2751 | * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain | |
2752 | * may have dropped their last reference (not just their last process | |
2753 | * reference). | |
2754 | */ | |
2755 | if (!bfqq_process_refs(new_bfqq)) | |
2756 | return NULL; | |
2757 | ||
2758 | /* Avoid a circular list and skip interim queue merges. */ | |
2759 | while ((__bfqq = new_bfqq->new_bfqq)) { | |
2760 | if (__bfqq == bfqq) | |
2761 | return NULL; | |
2762 | new_bfqq = __bfqq; | |
2763 | } | |
2764 | ||
2765 | process_refs = bfqq_process_refs(bfqq); | |
2766 | new_process_refs = bfqq_process_refs(new_bfqq); | |
2767 | /* | |
2768 | * If the process for the bfqq has gone away, there is no | |
2769 | * sense in merging the queues. | |
2770 | */ | |
2771 | if (process_refs == 0 || new_process_refs == 0) | |
2772 | return NULL; | |
2773 | ||
c1cee4ab JK |
2774 | /* |
2775 | * Make sure merged queues belong to the same parent. Parents could | |
2776 | * have changed since the time we decided the two queues are suitable | |
2777 | * for merging. | |
2778 | */ | |
2779 | if (new_bfqq->entity.parent != bfqq->entity.parent) | |
2780 | return NULL; | |
2781 | ||
36eca894 AA |
2782 | bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
2783 | new_bfqq->pid); | |
2784 | ||
2785 | /* | |
2786 | * Merging is just a redirection: the requests of the process | |
2787 | * owning one of the two queues are redirected to the other queue. | |
2788 | * The latter queue, in its turn, is set as shared if this is the | |
2789 | * first time that the requests of some process are redirected to | |
2790 | * it. | |
2791 | * | |
6fa3e8d3 PV |
2792 | * We redirect bfqq to new_bfqq and not the opposite, because |
2793 | * we are in the context of the process owning bfqq, thus we | |
2794 | * have the io_cq of this process. So we can immediately | |
2795 | * configure this io_cq to redirect the requests of the | |
2796 | * process to new_bfqq. In contrast, the io_cq of new_bfqq is | |
2797 | * not available any more (new_bfqq->bic == NULL). | |
36eca894 | 2798 | * |
6fa3e8d3 PV |
2799 | * Anyway, even in case new_bfqq coincides with the in-service |
2800 | * queue, redirecting requests the in-service queue is the | |
2801 | * best option, as we feed the in-service queue with new | |
2802 | * requests close to the last request served and, by doing so, | |
2803 | * are likely to increase the throughput. | |
36eca894 AA |
2804 | */ |
2805 | bfqq->new_bfqq = new_bfqq; | |
15729ff8 PV |
2806 | /* |
2807 | * The above assignment schedules the following redirections: | |
2808 | * each time some I/O for bfqq arrives, the process that | |
2809 | * generated that I/O is disassociated from bfqq and | |
2810 | * associated with new_bfqq. Here we increases new_bfqq->ref | |
2811 | * in advance, adding the number of processes that are | |
2812 | * expected to be associated with new_bfqq as they happen to | |
2813 | * issue I/O. | |
2814 | */ | |
36eca894 AA |
2815 | new_bfqq->ref += process_refs; |
2816 | return new_bfqq; | |
2817 | } | |
2818 | ||
2819 | static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq, | |
2820 | struct bfq_queue *new_bfqq) | |
2821 | { | |
7b8fa3b9 PV |
2822 | if (bfq_too_late_for_merging(new_bfqq)) |
2823 | return false; | |
2824 | ||
36eca894 AA |
2825 | if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) || |
2826 | (bfqq->ioprio_class != new_bfqq->ioprio_class)) | |
2827 | return false; | |
2828 | ||
2829 | /* | |
2830 | * If either of the queues has already been detected as seeky, | |
2831 | * then merging it with the other queue is unlikely to lead to | |
2832 | * sequential I/O. | |
2833 | */ | |
2834 | if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq)) | |
2835 | return false; | |
2836 | ||
2837 | /* | |
2838 | * Interleaved I/O is known to be done by (some) applications | |
2839 | * only for reads, so it does not make sense to merge async | |
2840 | * queues. | |
2841 | */ | |
2842 | if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq)) | |
2843 | return false; | |
2844 | ||
2845 | return true; | |
2846 | } | |
2847 | ||
430a67f9 PV |
2848 | static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd, |
2849 | struct bfq_queue *bfqq); | |
2850 | ||
fd571df0 PV |
2851 | static struct bfq_queue * |
2852 | bfq_setup_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
2853 | struct bfq_queue *stable_merge_bfqq, | |
2854 | struct bfq_iocq_bfqq_data *bfqq_data) | |
2855 | { | |
2856 | int proc_ref = min(bfqq_process_refs(bfqq), | |
2857 | bfqq_process_refs(stable_merge_bfqq)); | |
e2f2a394 | 2858 | struct bfq_queue *new_bfqq = NULL; |
fd571df0 | 2859 | |
e2f2a394 YK |
2860 | bfqq_data->stable_merge_bfqq = NULL; |
2861 | if (idling_boosts_thr_without_issues(bfqd, bfqq) || proc_ref == 0) | |
2862 | goto out; | |
fd571df0 PV |
2863 | |
2864 | /* next function will take at least one ref */ | |
2865 | new_bfqq = bfq_setup_merge(bfqq, stable_merge_bfqq); | |
2866 | ||
2867 | if (new_bfqq) { | |
2868 | bfqq_data->stably_merged = true; | |
2869 | if (new_bfqq->bic) { | |
2870 | unsigned int new_a_idx = new_bfqq->actuator_idx; | |
2871 | struct bfq_iocq_bfqq_data *new_bfqq_data = | |
2872 | &new_bfqq->bic->bfqq_data[new_a_idx]; | |
2873 | ||
2874 | new_bfqq_data->stably_merged = true; | |
2875 | } | |
2876 | } | |
e2f2a394 YK |
2877 | |
2878 | out: | |
2879 | /* deschedule stable merge, because done or aborted here */ | |
2880 | bfq_put_stable_ref(stable_merge_bfqq); | |
2881 | ||
fd571df0 PV |
2882 | return new_bfqq; |
2883 | } | |
2884 | ||
36eca894 AA |
2885 | /* |
2886 | * Attempt to schedule a merge of bfqq with the currently in-service | |
2887 | * queue or with a close queue among the scheduled queues. Return | |
2888 | * NULL if no merge was scheduled, a pointer to the shared bfq_queue | |
2889 | * structure otherwise. | |
2890 | * | |
2891 | * The OOM queue is not allowed to participate to cooperation: in fact, since | |
2892 | * the requests temporarily redirected to the OOM queue could be redirected | |
2893 | * again to dedicated queues at any time, the state needed to correctly | |
2894 | * handle merging with the OOM queue would be quite complex and expensive | |
2895 | * to maintain. Besides, in such a critical condition as an out of memory, | |
2896 | * the benefits of queue merging may be little relevant, or even negligible. | |
2897 | * | |
36eca894 AA |
2898 | * WARNING: queue merging may impair fairness among non-weight raised |
2899 | * queues, for at least two reasons: 1) the original weight of a | |
2900 | * merged queue may change during the merged state, 2) even being the | |
2901 | * weight the same, a merged queue may be bloated with many more | |
2902 | * requests than the ones produced by its originally-associated | |
2903 | * process. | |
2904 | */ | |
2905 | static struct bfq_queue * | |
2906 | bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
430a67f9 | 2907 | void *io_struct, bool request, struct bfq_io_cq *bic) |
36eca894 AA |
2908 | { |
2909 | struct bfq_queue *in_service_bfqq, *new_bfqq; | |
fd571df0 PV |
2910 | unsigned int a_idx = bfqq->actuator_idx; |
2911 | struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx]; | |
36eca894 | 2912 | |
15729ff8 PV |
2913 | /* if a merge has already been setup, then proceed with that first */ |
2914 | if (bfqq->new_bfqq) | |
2915 | return bfqq->new_bfqq; | |
2916 | ||
430a67f9 PV |
2917 | /* |
2918 | * Check delayed stable merge for rotational or non-queueing | |
2919 | * devs. For this branch to be executed, bfqq must not be | |
2920 | * currently merged with some other queue (i.e., bfqq->bic | |
2921 | * must be non null). If we considered also merged queues, | |
2922 | * then we should also check whether bfqq has already been | |
2923 | * merged with bic->stable_merge_bfqq. But this would be | |
2924 | * costly and complicated. | |
2925 | */ | |
2926 | if (unlikely(!bfqd->nonrot_with_queueing)) { | |
bd3664b3 PV |
2927 | /* |
2928 | * Make sure also that bfqq is sync, because | |
2929 | * bic->stable_merge_bfqq may point to some queue (for | |
2930 | * stable merging) also if bic is associated with a | |
2931 | * sync queue, but this bfqq is async | |
2932 | */ | |
a6123047 | 2933 | if (bfq_bfqq_sync(bfqq) && bfqq_data->stable_merge_bfqq && |
430a67f9 | 2934 | !bfq_bfqq_just_created(bfqq) && |
e03f2ab7 | 2935 | time_is_before_jiffies(bfqq->split_time + |
7812472f | 2936 | msecs_to_jiffies(bfq_late_stable_merging)) && |
d4f49983 | 2937 | time_is_before_jiffies(bfqq->creation_time + |
7812472f | 2938 | msecs_to_jiffies(bfq_late_stable_merging))) { |
430a67f9 | 2939 | struct bfq_queue *stable_merge_bfqq = |
a6123047 | 2940 | bfqq_data->stable_merge_bfqq; |
430a67f9 | 2941 | |
fd571df0 PV |
2942 | return bfq_setup_stable_merge(bfqd, bfqq, |
2943 | stable_merge_bfqq, | |
2944 | bfqq_data); | |
430a67f9 PV |
2945 | } |
2946 | } | |
2947 | ||
8cacc5ab PV |
2948 | /* |
2949 | * Do not perform queue merging if the device is non | |
2950 | * rotational and performs internal queueing. In fact, such a | |
2951 | * device reaches a high speed through internal parallelism | |
2952 | * and pipelining. This means that, to reach a high | |
2953 | * throughput, it must have many requests enqueued at the same | |
2954 | * time. But, in this configuration, the internal scheduling | |
2955 | * algorithm of the device does exactly the job of queue | |
2956 | * merging: it reorders requests so as to obtain as much as | |
2957 | * possible a sequential I/O pattern. As a consequence, with | |
2958 | * the workload generated by processes doing interleaved I/O, | |
2959 | * the throughput reached by the device is likely to be the | |
2960 | * same, with and without queue merging. | |
2961 | * | |
2962 | * Disabling merging also provides a remarkable benefit in | |
2963 | * terms of throughput. Merging tends to make many workloads | |
2964 | * artificially more uneven, because of shared queues | |
2965 | * remaining non empty for incomparably more time than | |
2966 | * non-merged queues. This may accentuate workload | |
2967 | * asymmetries. For example, if one of the queues in a set of | |
2968 | * merged queues has a higher weight than a normal queue, then | |
2969 | * the shared queue may inherit such a high weight and, by | |
2970 | * staying almost always active, may force BFQ to perform I/O | |
2971 | * plugging most of the time. This evidently makes it harder | |
2972 | * for BFQ to let the device reach a high throughput. | |
2973 | * | |
2974 | * Finally, the likely() macro below is not used because one | |
2975 | * of the two branches is more likely than the other, but to | |
2976 | * have the code path after the following if() executed as | |
2977 | * fast as possible for the case of a non rotational device | |
2978 | * with queueing. We want it because this is the fastest kind | |
2979 | * of device. On the opposite end, the likely() may lengthen | |
2980 | * the execution time of BFQ for the case of slower devices | |
2981 | * (rotational or at least without queueing). But in this case | |
2982 | * the execution time of BFQ matters very little, if not at | |
2983 | * all. | |
2984 | */ | |
2985 | if (likely(bfqd->nonrot_with_queueing)) | |
2986 | return NULL; | |
2987 | ||
7b8fa3b9 PV |
2988 | /* |
2989 | * Prevent bfqq from being merged if it has been created too | |
2990 | * long ago. The idea is that true cooperating processes, and | |
2991 | * thus their associated bfq_queues, are supposed to be | |
2992 | * created shortly after each other. This is the case, e.g., | |
2993 | * for KVM/QEMU and dump I/O threads. Basing on this | |
2994 | * assumption, the following filtering greatly reduces the | |
2995 | * probability that two non-cooperating processes, which just | |
2996 | * happen to do close I/O for some short time interval, have | |
2997 | * their queues merged by mistake. | |
2998 | */ | |
2999 | if (bfq_too_late_for_merging(bfqq)) | |
3000 | return NULL; | |
3001 | ||
4403e4e4 | 3002 | if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq)) |
36eca894 AA |
3003 | return NULL; |
3004 | ||
3005 | /* If there is only one backlogged queue, don't search. */ | |
73d58118 | 3006 | if (bfq_tot_busy_queues(bfqd) == 1) |
36eca894 AA |
3007 | return NULL; |
3008 | ||
3009 | in_service_bfqq = bfqd->in_service_queue; | |
3010 | ||
4403e4e4 AR |
3011 | if (in_service_bfqq && in_service_bfqq != bfqq && |
3012 | likely(in_service_bfqq != &bfqd->oom_bfqq) && | |
058fdecc PV |
3013 | bfq_rq_close_to_sector(io_struct, request, |
3014 | bfqd->in_serv_last_pos) && | |
36eca894 AA |
3015 | bfqq->entity.parent == in_service_bfqq->entity.parent && |
3016 | bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) { | |
3017 | new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq); | |
3018 | if (new_bfqq) | |
3019 | return new_bfqq; | |
3020 | } | |
3021 | /* | |
3022 | * Check whether there is a cooperator among currently scheduled | |
3023 | * queues. The only thing we need is that the bio/request is not | |
3024 | * NULL, as we need it to establish whether a cooperator exists. | |
3025 | */ | |
36eca894 AA |
3026 | new_bfqq = bfq_find_close_cooperator(bfqd, bfqq, |
3027 | bfq_io_struct_pos(io_struct, request)); | |
3028 | ||
4403e4e4 | 3029 | if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) && |
36eca894 AA |
3030 | bfq_may_be_close_cooperator(bfqq, new_bfqq)) |
3031 | return bfq_setup_merge(bfqq, new_bfqq); | |
3032 | ||
3033 | return NULL; | |
3034 | } | |
3035 | ||
3036 | static void bfq_bfqq_save_state(struct bfq_queue *bfqq) | |
3037 | { | |
3038 | struct bfq_io_cq *bic = bfqq->bic; | |
fd571df0 PV |
3039 | unsigned int a_idx = bfqq->actuator_idx; |
3040 | struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx]; | |
36eca894 AA |
3041 | |
3042 | /* | |
3043 | * If !bfqq->bic, the queue is already shared or its requests | |
3044 | * have already been redirected to a shared queue; both idle window | |
3045 | * and weight raising state have already been saved. Do nothing. | |
3046 | */ | |
3047 | if (!bic) | |
3048 | return; | |
3049 | ||
a6123047 | 3050 | bfqq_data->saved_last_serv_time_ns = bfqq->last_serv_time_ns; |
fd571df0 | 3051 | bfqq_data->saved_inject_limit = bfqq->inject_limit; |
a6123047 PV |
3052 | bfqq_data->saved_decrease_time_jif = bfqq->decrease_time_jif; |
3053 | ||
3054 | bfqq_data->saved_weight = bfqq->entity.orig_weight; | |
3055 | bfqq_data->saved_ttime = bfqq->ttime; | |
3056 | bfqq_data->saved_has_short_ttime = | |
3057 | bfq_bfqq_has_short_ttime(bfqq); | |
3058 | bfqq_data->saved_IO_bound = bfq_bfqq_IO_bound(bfqq); | |
3059 | bfqq_data->saved_io_start_time = bfqq->io_start_time; | |
3060 | bfqq_data->saved_tot_idle_time = bfqq->tot_idle_time; | |
3061 | bfqq_data->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq); | |
3062 | bfqq_data->was_in_burst_list = | |
3063 | !hlist_unhashed(&bfqq->burst_list_node); | |
3064 | ||
894df937 | 3065 | if (unlikely(bfq_bfqq_just_created(bfqq) && |
1be6e8a9 AR |
3066 | !bfq_bfqq_in_large_burst(bfqq) && |
3067 | bfqq->bfqd->low_latency)) { | |
894df937 PV |
3068 | /* |
3069 | * bfqq being merged right after being created: bfqq | |
3070 | * would have deserved interactive weight raising, but | |
3071 | * did not make it to be set in a weight-raised state, | |
3072 | * because of this early merge. Store directly the | |
3073 | * weight-raising state that would have been assigned | |
3074 | * to bfqq, so that to avoid that bfqq unjustly fails | |
3075 | * to enjoy weight raising if split soon. | |
3076 | */ | |
a6123047 PV |
3077 | bfqq_data->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff; |
3078 | bfqq_data->saved_wr_start_at_switch_to_srt = | |
3079 | bfq_smallest_from_now(); | |
3080 | bfqq_data->saved_wr_cur_max_time = | |
3081 | bfq_wr_duration(bfqq->bfqd); | |
3082 | bfqq_data->saved_last_wr_start_finish = jiffies; | |
894df937 | 3083 | } else { |
a6123047 PV |
3084 | bfqq_data->saved_wr_coeff = bfqq->wr_coeff; |
3085 | bfqq_data->saved_wr_start_at_switch_to_srt = | |
894df937 | 3086 | bfqq->wr_start_at_switch_to_srt; |
a6123047 PV |
3087 | bfqq_data->saved_service_from_wr = |
3088 | bfqq->service_from_wr; | |
3089 | bfqq_data->saved_last_wr_start_finish = | |
3090 | bfqq->last_wr_start_finish; | |
3091 | bfqq_data->saved_wr_cur_max_time = bfqq->wr_cur_max_time; | |
894df937 | 3092 | } |
36eca894 AA |
3093 | } |
3094 | ||
430a67f9 PV |
3095 | |
3096 | static void | |
3097 | bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq) | |
3098 | { | |
3099 | if (cur_bfqq->entity.parent && | |
3100 | cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq) | |
3101 | cur_bfqq->entity.parent->last_bfqq_created = new_bfqq; | |
3102 | else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq) | |
3103 | cur_bfqq->bfqd->last_bfqq_created = new_bfqq; | |
3104 | } | |
3105 | ||
478de338 PV |
3106 | void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
3107 | { | |
3108 | /* | |
3109 | * To prevent bfqq's service guarantees from being violated, | |
3110 | * bfqq may be left busy, i.e., queued for service, even if | |
3111 | * empty (see comments in __bfq_bfqq_expire() for | |
3112 | * details). But, if no process will send requests to bfqq any | |
3113 | * longer, then there is no point in keeping bfqq queued for | |
3114 | * service. In addition, keeping bfqq queued for service, but | |
3115 | * with no process ref any longer, may have caused bfqq to be | |
3116 | * freed when dequeued from service. But this is assumed to | |
3117 | * never happen. | |
3118 | */ | |
3119 | if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) && | |
3120 | bfqq != bfqd->in_service_queue) | |
d322f355 | 3121 | bfq_del_bfqq_busy(bfqq, false); |
478de338 | 3122 | |
430a67f9 PV |
3123 | bfq_reassign_last_bfqq(bfqq, NULL); |
3124 | ||
478de338 PV |
3125 | bfq_put_queue(bfqq); |
3126 | } | |
3127 | ||
36eca894 AA |
3128 | static void |
3129 | bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, | |
3130 | struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) | |
3131 | { | |
3132 | bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", | |
3133 | (unsigned long)new_bfqq->pid); | |
3134 | /* Save weight raising and idle window of the merged queues */ | |
3135 | bfq_bfqq_save_state(bfqq); | |
3136 | bfq_bfqq_save_state(new_bfqq); | |
3137 | if (bfq_bfqq_IO_bound(bfqq)) | |
3138 | bfq_mark_bfqq_IO_bound(new_bfqq); | |
3139 | bfq_clear_bfqq_IO_bound(bfqq); | |
3140 | ||
8ef3fc3a PV |
3141 | /* |
3142 | * The processes associated with bfqq are cooperators of the | |
3143 | * processes associated with new_bfqq. So, if bfqq has a | |
3144 | * waker, then assume that all these processes will be happy | |
3145 | * to let bfqq's waker freely inject I/O when they have no | |
3146 | * I/O. | |
3147 | */ | |
3148 | if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq && | |
3149 | bfqq->waker_bfqq != new_bfqq) { | |
3150 | new_bfqq->waker_bfqq = bfqq->waker_bfqq; | |
3151 | new_bfqq->tentative_waker_bfqq = NULL; | |
3152 | ||
3153 | /* | |
3154 | * If the waker queue disappears, then | |
3155 | * new_bfqq->waker_bfqq must be reset. So insert | |
3156 | * new_bfqq into the woken_list of the waker. See | |
3157 | * bfq_check_waker for details. | |
3158 | */ | |
3159 | hlist_add_head(&new_bfqq->woken_list_node, | |
3160 | &new_bfqq->waker_bfqq->woken_list); | |
3161 | ||
3162 | } | |
3163 | ||
36eca894 AA |
3164 | /* |
3165 | * If bfqq is weight-raised, then let new_bfqq inherit | |
3166 | * weight-raising. To reduce false positives, neglect the case | |
3167 | * where bfqq has just been created, but has not yet made it | |
3168 | * to be weight-raised (which may happen because EQM may merge | |
3169 | * bfqq even before bfq_add_request is executed for the first | |
e1b2324d AA |
3170 | * time for bfqq). Handling this case would however be very |
3171 | * easy, thanks to the flag just_created. | |
36eca894 AA |
3172 | */ |
3173 | if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) { | |
3174 | new_bfqq->wr_coeff = bfqq->wr_coeff; | |
3175 | new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time; | |
3176 | new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish; | |
3177 | new_bfqq->wr_start_at_switch_to_srt = | |
3178 | bfqq->wr_start_at_switch_to_srt; | |
3179 | if (bfq_bfqq_busy(new_bfqq)) | |
3180 | bfqd->wr_busy_queues++; | |
3181 | new_bfqq->entity.prio_changed = 1; | |
3182 | } | |
3183 | ||
3184 | if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */ | |
3185 | bfqq->wr_coeff = 1; | |
3186 | bfqq->entity.prio_changed = 1; | |
3187 | if (bfq_bfqq_busy(bfqq)) | |
3188 | bfqd->wr_busy_queues--; | |
3189 | } | |
3190 | ||
3191 | bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d", | |
3192 | bfqd->wr_busy_queues); | |
3193 | ||
36eca894 AA |
3194 | /* |
3195 | * Merge queues (that is, let bic redirect its requests to new_bfqq) | |
3196 | */ | |
9778369a | 3197 | bic_set_bfqq(bic, new_bfqq, true, bfqq->actuator_idx); |
36eca894 AA |
3198 | bfq_mark_bfqq_coop(new_bfqq); |
3199 | /* | |
3200 | * new_bfqq now belongs to at least two bics (it is a shared queue): | |
3201 | * set new_bfqq->bic to NULL. bfqq either: | |
3202 | * - does not belong to any bic any more, and hence bfqq->bic must | |
3203 | * be set to NULL, or | |
3204 | * - is a queue whose owning bics have already been redirected to a | |
3205 | * different queue, hence the queue is destined to not belong to | |
3206 | * any bic soon and bfqq->bic is already NULL (therefore the next | |
3207 | * assignment causes no harm). | |
3208 | */ | |
3209 | new_bfqq->bic = NULL; | |
1e66413c FP |
3210 | /* |
3211 | * If the queue is shared, the pid is the pid of one of the associated | |
3212 | * processes. Which pid depends on the exact sequence of merge events | |
3213 | * the queue underwent. So printing such a pid is useless and confusing | |
3214 | * because it reports a random pid between those of the associated | |
3215 | * processes. | |
3216 | * We mark such a queue with a pid -1, and then print SHARED instead of | |
3217 | * a pid in logging messages. | |
3218 | */ | |
3219 | new_bfqq->pid = -1; | |
36eca894 | 3220 | bfqq->bic = NULL; |
430a67f9 PV |
3221 | |
3222 | bfq_reassign_last_bfqq(bfqq, new_bfqq); | |
3223 | ||
478de338 | 3224 | bfq_release_process_ref(bfqd, bfqq); |
36eca894 AA |
3225 | } |
3226 | ||
aee69d78 PV |
3227 | static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq, |
3228 | struct bio *bio) | |
3229 | { | |
3230 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
3231 | bool is_sync = op_is_sync(bio->bi_opf); | |
36eca894 | 3232 | struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq; |
aee69d78 PV |
3233 | |
3234 | /* | |
3235 | * Disallow merge of a sync bio into an async request. | |
3236 | */ | |
3237 | if (is_sync && !rq_is_sync(rq)) | |
3238 | return false; | |
3239 | ||
3240 | /* | |
3241 | * Lookup the bfqq that this bio will be queued with. Allow | |
3242 | * merge only if rq is queued there. | |
3243 | */ | |
3244 | if (!bfqq) | |
3245 | return false; | |
3246 | ||
36eca894 AA |
3247 | /* |
3248 | * We take advantage of this function to perform an early merge | |
3249 | * of the queues of possible cooperating processes. | |
3250 | */ | |
430a67f9 | 3251 | new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic); |
36eca894 AA |
3252 | if (new_bfqq) { |
3253 | /* | |
3254 | * bic still points to bfqq, then it has not yet been | |
3255 | * redirected to some other bfq_queue, and a queue | |
636b8fe8 AR |
3256 | * merge between bfqq and new_bfqq can be safely |
3257 | * fulfilled, i.e., bic can be redirected to new_bfqq | |
36eca894 AA |
3258 | * and bfqq can be put. |
3259 | */ | |
3260 | bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq, | |
3261 | new_bfqq); | |
3262 | /* | |
3263 | * If we get here, bio will be queued into new_queue, | |
3264 | * so use new_bfqq to decide whether bio and rq can be | |
3265 | * merged. | |
3266 | */ | |
3267 | bfqq = new_bfqq; | |
3268 | ||
3269 | /* | |
3270 | * Change also bqfd->bio_bfqq, as | |
3271 | * bfqd->bio_bic now points to new_bfqq, and | |
3272 | * this function may be invoked again (and then may | |
3273 | * use again bqfd->bio_bfqq). | |
3274 | */ | |
3275 | bfqd->bio_bfqq = bfqq; | |
3276 | } | |
3277 | ||
aee69d78 PV |
3278 | return bfqq == RQ_BFQQ(rq); |
3279 | } | |
3280 | ||
44e44a1b PV |
3281 | /* |
3282 | * Set the maximum time for the in-service queue to consume its | |
3283 | * budget. This prevents seeky processes from lowering the throughput. | |
3284 | * In practice, a time-slice service scheme is used with seeky | |
3285 | * processes. | |
3286 | */ | |
3287 | static void bfq_set_budget_timeout(struct bfq_data *bfqd, | |
3288 | struct bfq_queue *bfqq) | |
3289 | { | |
77b7dcea PV |
3290 | unsigned int timeout_coeff; |
3291 | ||
3292 | if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time) | |
3293 | timeout_coeff = 1; | |
3294 | else | |
3295 | timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; | |
3296 | ||
44e44a1b PV |
3297 | bfqd->last_budget_start = ktime_get(); |
3298 | ||
3299 | bfqq->budget_timeout = jiffies + | |
77b7dcea | 3300 | bfqd->bfq_timeout * timeout_coeff; |
44e44a1b PV |
3301 | } |
3302 | ||
aee69d78 PV |
3303 | static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
3304 | struct bfq_queue *bfqq) | |
3305 | { | |
3306 | if (bfqq) { | |
aee69d78 PV |
3307 | bfq_clear_bfqq_fifo_expire(bfqq); |
3308 | ||
3309 | bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8; | |
3310 | ||
77b7dcea PV |
3311 | if (time_is_before_jiffies(bfqq->last_wr_start_finish) && |
3312 | bfqq->wr_coeff > 1 && | |
3313 | bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && | |
3314 | time_is_before_jiffies(bfqq->budget_timeout)) { | |
3315 | /* | |
3316 | * For soft real-time queues, move the start | |
3317 | * of the weight-raising period forward by the | |
3318 | * time the queue has not received any | |
3319 | * service. Otherwise, a relatively long | |
3320 | * service delay is likely to cause the | |
3321 | * weight-raising period of the queue to end, | |
3322 | * because of the short duration of the | |
3323 | * weight-raising period of a soft real-time | |
3324 | * queue. It is worth noting that this move | |
3325 | * is not so dangerous for the other queues, | |
3326 | * because soft real-time queues are not | |
3327 | * greedy. | |
3328 | * | |
3329 | * To not add a further variable, we use the | |
3330 | * overloaded field budget_timeout to | |
3331 | * determine for how long the queue has not | |
3332 | * received service, i.e., how much time has | |
3333 | * elapsed since the queue expired. However, | |
3334 | * this is a little imprecise, because | |
3335 | * budget_timeout is set to jiffies if bfqq | |
3336 | * not only expires, but also remains with no | |
3337 | * request. | |
3338 | */ | |
3339 | if (time_after(bfqq->budget_timeout, | |
3340 | bfqq->last_wr_start_finish)) | |
3341 | bfqq->last_wr_start_finish += | |
3342 | jiffies - bfqq->budget_timeout; | |
3343 | else | |
3344 | bfqq->last_wr_start_finish = jiffies; | |
3345 | } | |
3346 | ||
44e44a1b | 3347 | bfq_set_budget_timeout(bfqd, bfqq); |
aee69d78 PV |
3348 | bfq_log_bfqq(bfqd, bfqq, |
3349 | "set_in_service_queue, cur-budget = %d", | |
3350 | bfqq->entity.budget); | |
3351 | } | |
3352 | ||
3353 | bfqd->in_service_queue = bfqq; | |
41e76c85 | 3354 | bfqd->in_serv_last_pos = 0; |
aee69d78 PV |
3355 | } |
3356 | ||
3357 | /* | |
3358 | * Get and set a new queue for service. | |
3359 | */ | |
3360 | static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd) | |
3361 | { | |
3362 | struct bfq_queue *bfqq = bfq_get_next_queue(bfqd); | |
3363 | ||
3364 | __bfq_set_in_service_queue(bfqd, bfqq); | |
3365 | return bfqq; | |
3366 | } | |
3367 | ||
aee69d78 PV |
3368 | static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
3369 | { | |
3370 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
aee69d78 PV |
3371 | u32 sl; |
3372 | ||
aee69d78 PV |
3373 | bfq_mark_bfqq_wait_request(bfqq); |
3374 | ||
3375 | /* | |
3376 | * We don't want to idle for seeks, but we do want to allow | |
3377 | * fair distribution of slice time for a process doing back-to-back | |
3378 | * seeks. So allow a little bit of time for him to submit a new rq. | |
3379 | */ | |
3380 | sl = bfqd->bfq_slice_idle; | |
3381 | /* | |
1de0c4cd AA |
3382 | * Unless the queue is being weight-raised or the scenario is |
3383 | * asymmetric, grant only minimum idle time if the queue | |
3384 | * is seeky. A long idling is preserved for a weight-raised | |
3385 | * queue, or, more in general, in an asymmetric scenario, | |
3386 | * because a long idling is needed for guaranteeing to a queue | |
3387 | * its reserved share of the throughput (in particular, it is | |
3388 | * needed if the queue has a higher weight than some other | |
3389 | * queue). | |
aee69d78 | 3390 | */ |
1de0c4cd | 3391 | if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 && |
fb53ac6c | 3392 | !bfq_asymmetric_scenario(bfqd, bfqq)) |
aee69d78 | 3393 | sl = min_t(u64, sl, BFQ_MIN_TT); |
778c02a2 PV |
3394 | else if (bfqq->wr_coeff > 1) |
3395 | sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC); | |
aee69d78 PV |
3396 | |
3397 | bfqd->last_idling_start = ktime_get(); | |
2341d662 PV |
3398 | bfqd->last_idling_start_jiffies = jiffies; |
3399 | ||
aee69d78 PV |
3400 | hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl), |
3401 | HRTIMER_MODE_REL); | |
e21b7a0b | 3402 | bfqg_stats_set_start_idle_time(bfqq_group(bfqq)); |
aee69d78 PV |
3403 | } |
3404 | ||
ab0e43e9 PV |
3405 | /* |
3406 | * In autotuning mode, max_budget is dynamically recomputed as the | |
3407 | * amount of sectors transferred in timeout at the estimated peak | |
3408 | * rate. This enables BFQ to utilize a full timeslice with a full | |
3409 | * budget, even if the in-service queue is served at peak rate. And | |
3410 | * this maximises throughput with sequential workloads. | |
3411 | */ | |
3412 | static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd) | |
3413 | { | |
3414 | return (u64)bfqd->peak_rate * USEC_PER_MSEC * | |
3415 | jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT; | |
3416 | } | |
3417 | ||
44e44a1b PV |
3418 | /* |
3419 | * Update parameters related to throughput and responsiveness, as a | |
3420 | * function of the estimated peak rate. See comments on | |
e24f1c24 | 3421 | * bfq_calc_max_budget(), and on the ref_wr_duration array. |
44e44a1b PV |
3422 | */ |
3423 | static void update_thr_responsiveness_params(struct bfq_data *bfqd) | |
3424 | { | |
e24f1c24 | 3425 | if (bfqd->bfq_user_max_budget == 0) { |
44e44a1b PV |
3426 | bfqd->bfq_max_budget = |
3427 | bfq_calc_max_budget(bfqd); | |
e24f1c24 | 3428 | bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget); |
44e44a1b | 3429 | } |
44e44a1b PV |
3430 | } |
3431 | ||
ab0e43e9 PV |
3432 | static void bfq_reset_rate_computation(struct bfq_data *bfqd, |
3433 | struct request *rq) | |
3434 | { | |
3435 | if (rq != NULL) { /* new rq dispatch now, reset accordingly */ | |
3436 | bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns(); | |
3437 | bfqd->peak_rate_samples = 1; | |
3438 | bfqd->sequential_samples = 0; | |
3439 | bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size = | |
3440 | blk_rq_sectors(rq); | |
3441 | } else /* no new rq dispatched, just reset the number of samples */ | |
3442 | bfqd->peak_rate_samples = 0; /* full re-init on next disp. */ | |
3443 | ||
3444 | bfq_log(bfqd, | |
3445 | "reset_rate_computation at end, sample %u/%u tot_sects %llu", | |
3446 | bfqd->peak_rate_samples, bfqd->sequential_samples, | |
3447 | bfqd->tot_sectors_dispatched); | |
3448 | } | |
3449 | ||
3450 | static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq) | |
3451 | { | |
3452 | u32 rate, weight, divisor; | |
3453 | ||
3454 | /* | |
3455 | * For the convergence property to hold (see comments on | |
3456 | * bfq_update_peak_rate()) and for the assessment to be | |
3457 | * reliable, a minimum number of samples must be present, and | |
3458 | * a minimum amount of time must have elapsed. If not so, do | |
3459 | * not compute new rate. Just reset parameters, to get ready | |
3460 | * for a new evaluation attempt. | |
3461 | */ | |
3462 | if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES || | |
3463 | bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL) | |
3464 | goto reset_computation; | |
3465 | ||
3466 | /* | |
3467 | * If a new request completion has occurred after last | |
3468 | * dispatch, then, to approximate the rate at which requests | |
3469 | * have been served by the device, it is more precise to | |
3470 | * extend the observation interval to the last completion. | |
3471 | */ | |
3472 | bfqd->delta_from_first = | |
3473 | max_t(u64, bfqd->delta_from_first, | |
3474 | bfqd->last_completion - bfqd->first_dispatch); | |
3475 | ||
3476 | /* | |
3477 | * Rate computed in sects/usec, and not sects/nsec, for | |
3478 | * precision issues. | |
3479 | */ | |
3480 | rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT, | |
3481 | div_u64(bfqd->delta_from_first, NSEC_PER_USEC)); | |
3482 | ||
3483 | /* | |
3484 | * Peak rate not updated if: | |
3485 | * - the percentage of sequential dispatches is below 3/4 of the | |
3486 | * total, and rate is below the current estimated peak rate | |
3487 | * - rate is unreasonably high (> 20M sectors/sec) | |
3488 | */ | |
3489 | if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 && | |
3490 | rate <= bfqd->peak_rate) || | |
3491 | rate > 20<<BFQ_RATE_SHIFT) | |
3492 | goto reset_computation; | |
3493 | ||
3494 | /* | |
3495 | * We have to update the peak rate, at last! To this purpose, | |
3496 | * we use a low-pass filter. We compute the smoothing constant | |
3497 | * of the filter as a function of the 'weight' of the new | |
3498 | * measured rate. | |
3499 | * | |
3500 | * As can be seen in next formulas, we define this weight as a | |
3501 | * quantity proportional to how sequential the workload is, | |
3502 | * and to how long the observation time interval is. | |
3503 | * | |
3504 | * The weight runs from 0 to 8. The maximum value of the | |
3505 | * weight, 8, yields the minimum value for the smoothing | |
3506 | * constant. At this minimum value for the smoothing constant, | |
3507 | * the measured rate contributes for half of the next value of | |
3508 | * the estimated peak rate. | |
3509 | * | |
3510 | * So, the first step is to compute the weight as a function | |
3511 | * of how sequential the workload is. Note that the weight | |
3512 | * cannot reach 9, because bfqd->sequential_samples cannot | |
3513 | * become equal to bfqd->peak_rate_samples, which, in its | |
3514 | * turn, holds true because bfqd->sequential_samples is not | |
3515 | * incremented for the first sample. | |
3516 | */ | |
3517 | weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples; | |
3518 | ||
3519 | /* | |
3520 | * Second step: further refine the weight as a function of the | |
3521 | * duration of the observation interval. | |
3522 | */ | |
3523 | weight = min_t(u32, 8, | |
3524 | div_u64(weight * bfqd->delta_from_first, | |
3525 | BFQ_RATE_REF_INTERVAL)); | |
3526 | ||
3527 | /* | |
3528 | * Divisor ranging from 10, for minimum weight, to 2, for | |
3529 | * maximum weight. | |
3530 | */ | |
3531 | divisor = 10 - weight; | |
3532 | ||
3533 | /* | |
3534 | * Finally, update peak rate: | |
3535 | * | |
3536 | * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor | |
3537 | */ | |
3538 | bfqd->peak_rate *= divisor-1; | |
3539 | bfqd->peak_rate /= divisor; | |
3540 | rate /= divisor; /* smoothing constant alpha = 1/divisor */ | |
3541 | ||
3542 | bfqd->peak_rate += rate; | |
bc56e2ca PV |
3543 | |
3544 | /* | |
3545 | * For a very slow device, bfqd->peak_rate can reach 0 (see | |
3546 | * the minimum representable values reported in the comments | |
3547 | * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid | |
3548 | * divisions by zero where bfqd->peak_rate is used as a | |
3549 | * divisor. | |
3550 | */ | |
3551 | bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate); | |
3552 | ||
44e44a1b | 3553 | update_thr_responsiveness_params(bfqd); |
ab0e43e9 PV |
3554 | |
3555 | reset_computation: | |
3556 | bfq_reset_rate_computation(bfqd, rq); | |
3557 | } | |
3558 | ||
3559 | /* | |
3560 | * Update the read/write peak rate (the main quantity used for | |
3561 | * auto-tuning, see update_thr_responsiveness_params()). | |
3562 | * | |
3563 | * It is not trivial to estimate the peak rate (correctly): because of | |
3564 | * the presence of sw and hw queues between the scheduler and the | |
3565 | * device components that finally serve I/O requests, it is hard to | |
3566 | * say exactly when a given dispatched request is served inside the | |
3567 | * device, and for how long. As a consequence, it is hard to know | |
3568 | * precisely at what rate a given set of requests is actually served | |
3569 | * by the device. | |
3570 | * | |
3571 | * On the opposite end, the dispatch time of any request is trivially | |
3572 | * available, and, from this piece of information, the "dispatch rate" | |
3573 | * of requests can be immediately computed. So, the idea in the next | |
3574 | * function is to use what is known, namely request dispatch times | |
3575 | * (plus, when useful, request completion times), to estimate what is | |
3576 | * unknown, namely in-device request service rate. | |
3577 | * | |
3578 | * The main issue is that, because of the above facts, the rate at | |
3579 | * which a certain set of requests is dispatched over a certain time | |
3580 | * interval can vary greatly with respect to the rate at which the | |
3581 | * same requests are then served. But, since the size of any | |
3582 | * intermediate queue is limited, and the service scheme is lossless | |
3583 | * (no request is silently dropped), the following obvious convergence | |
3584 | * property holds: the number of requests dispatched MUST become | |
3585 | * closer and closer to the number of requests completed as the | |
3586 | * observation interval grows. This is the key property used in | |
3587 | * the next function to estimate the peak service rate as a function | |
3588 | * of the observed dispatch rate. The function assumes to be invoked | |
3589 | * on every request dispatch. | |
3590 | */ | |
3591 | static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq) | |
3592 | { | |
3593 | u64 now_ns = ktime_get_ns(); | |
3594 | ||
3595 | if (bfqd->peak_rate_samples == 0) { /* first dispatch */ | |
3596 | bfq_log(bfqd, "update_peak_rate: goto reset, samples %d", | |
3597 | bfqd->peak_rate_samples); | |
3598 | bfq_reset_rate_computation(bfqd, rq); | |
3599 | goto update_last_values; /* will add one sample */ | |
3600 | } | |
3601 | ||
3602 | /* | |
3603 | * Device idle for very long: the observation interval lasting | |
3604 | * up to this dispatch cannot be a valid observation interval | |
3605 | * for computing a new peak rate (similarly to the late- | |
3606 | * completion event in bfq_completed_request()). Go to | |
3607 | * update_rate_and_reset to have the following three steps | |
3608 | * taken: | |
3609 | * - close the observation interval at the last (previous) | |
3610 | * request dispatch or completion | |
3611 | * - compute rate, if possible, for that observation interval | |
3612 | * - start a new observation interval with this dispatch | |
3613 | */ | |
3614 | if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC && | |
2d31c684 | 3615 | bfqd->tot_rq_in_driver == 0) |
ab0e43e9 PV |
3616 | goto update_rate_and_reset; |
3617 | ||
3618 | /* Update sampling information */ | |
3619 | bfqd->peak_rate_samples++; | |
3620 | ||
2d31c684 | 3621 | if ((bfqd->tot_rq_in_driver > 0 || |
ab0e43e9 | 3622 | now_ns - bfqd->last_completion < BFQ_MIN_TT) |
d87447d8 | 3623 | && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq)) |
ab0e43e9 PV |
3624 | bfqd->sequential_samples++; |
3625 | ||
3626 | bfqd->tot_sectors_dispatched += blk_rq_sectors(rq); | |
3627 | ||
3628 | /* Reset max observed rq size every 32 dispatches */ | |
3629 | if (likely(bfqd->peak_rate_samples % 32)) | |
3630 | bfqd->last_rq_max_size = | |
3631 | max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size); | |
3632 | else | |
3633 | bfqd->last_rq_max_size = blk_rq_sectors(rq); | |
3634 | ||
3635 | bfqd->delta_from_first = now_ns - bfqd->first_dispatch; | |
3636 | ||
3637 | /* Target observation interval not yet reached, go on sampling */ | |
3638 | if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL) | |
3639 | goto update_last_values; | |
3640 | ||
3641 | update_rate_and_reset: | |
3642 | bfq_update_rate_reset(bfqd, rq); | |
3643 | update_last_values: | |
3644 | bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
058fdecc PV |
3645 | if (RQ_BFQQ(rq) == bfqd->in_service_queue) |
3646 | bfqd->in_serv_last_pos = bfqd->last_position; | |
ab0e43e9 PV |
3647 | bfqd->last_dispatch = now_ns; |
3648 | } | |
3649 | ||
aee69d78 PV |
3650 | /* |
3651 | * Remove request from internal lists. | |
3652 | */ | |
3653 | static void bfq_dispatch_remove(struct request_queue *q, struct request *rq) | |
3654 | { | |
3655 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
3656 | ||
3657 | /* | |
3658 | * For consistency, the next instruction should have been | |
3659 | * executed after removing the request from the queue and | |
3660 | * dispatching it. We execute instead this instruction before | |
3661 | * bfq_remove_request() (and hence introduce a temporary | |
3662 | * inconsistency), for efficiency. In fact, should this | |
3663 | * dispatch occur for a non in-service bfqq, this anticipated | |
3664 | * increment prevents two counters related to bfqq->dispatched | |
3665 | * from risking to be, first, uselessly decremented, and then | |
3666 | * incremented again when the (new) value of bfqq->dispatched | |
3667 | * happens to be taken into account. | |
3668 | */ | |
3669 | bfqq->dispatched++; | |
ab0e43e9 | 3670 | bfq_update_peak_rate(q->elevator->elevator_data, rq); |
aee69d78 PV |
3671 | |
3672 | bfq_remove_request(q, rq); | |
3673 | } | |
3674 | ||
3726112e PV |
3675 | /* |
3676 | * There is a case where idling does not have to be performed for | |
3677 | * throughput concerns, but to preserve the throughput share of | |
3678 | * the process associated with bfqq. | |
3679 | * | |
3680 | * To introduce this case, we can note that allowing the drive | |
3681 | * to enqueue more than one request at a time, and hence | |
3682 | * delegating de facto final scheduling decisions to the | |
3683 | * drive's internal scheduler, entails loss of control on the | |
3684 | * actual request service order. In particular, the critical | |
3685 | * situation is when requests from different processes happen | |
3686 | * to be present, at the same time, in the internal queue(s) | |
3687 | * of the drive. In such a situation, the drive, by deciding | |
3688 | * the service order of the internally-queued requests, does | |
3689 | * determine also the actual throughput distribution among | |
3690 | * these processes. But the drive typically has no notion or | |
3691 | * concern about per-process throughput distribution, and | |
3692 | * makes its decisions only on a per-request basis. Therefore, | |
3693 | * the service distribution enforced by the drive's internal | |
3694 | * scheduler is likely to coincide with the desired throughput | |
3695 | * distribution only in a completely symmetric, or favorably | |
3696 | * skewed scenario where: | |
3697 | * (i-a) each of these processes must get the same throughput as | |
3698 | * the others, | |
3699 | * (i-b) in case (i-a) does not hold, it holds that the process | |
3700 | * associated with bfqq must receive a lower or equal | |
3701 | * throughput than any of the other processes; | |
3702 | * (ii) the I/O of each process has the same properties, in | |
3703 | * terms of locality (sequential or random), direction | |
3704 | * (reads or writes), request sizes, greediness | |
3705 | * (from I/O-bound to sporadic), and so on; | |
3706 | ||
3707 | * In fact, in such a scenario, the drive tends to treat the requests | |
3708 | * of each process in about the same way as the requests of the | |
3709 | * others, and thus to provide each of these processes with about the | |
3710 | * same throughput. This is exactly the desired throughput | |
3711 | * distribution if (i-a) holds, or, if (i-b) holds instead, this is an | |
3712 | * even more convenient distribution for (the process associated with) | |
3713 | * bfqq. | |
3714 | * | |
3715 | * In contrast, in any asymmetric or unfavorable scenario, device | |
3716 | * idling (I/O-dispatch plugging) is certainly needed to guarantee | |
3717 | * that bfqq receives its assigned fraction of the device throughput | |
3718 | * (see [1] for details). | |
3719 | * | |
3720 | * The problem is that idling may significantly reduce throughput with | |
3721 | * certain combinations of types of I/O and devices. An important | |
3722 | * example is sync random I/O on flash storage with command | |
3723 | * queueing. So, unless bfqq falls in cases where idling also boosts | |
3724 | * throughput, it is important to check conditions (i-a), i(-b) and | |
3725 | * (ii) accurately, so as to avoid idling when not strictly needed for | |
3726 | * service guarantees. | |
3727 | * | |
3728 | * Unfortunately, it is extremely difficult to thoroughly check | |
3729 | * condition (ii). And, in case there are active groups, it becomes | |
3730 | * very difficult to check conditions (i-a) and (i-b) too. In fact, | |
3731 | * if there are active groups, then, for conditions (i-a) or (i-b) to | |
3732 | * become false 'indirectly', it is enough that an active group | |
3733 | * contains more active processes or sub-groups than some other active | |
3734 | * group. More precisely, for conditions (i-a) or (i-b) to become | |
3735 | * false because of such a group, it is not even necessary that the | |
3736 | * group is (still) active: it is sufficient that, even if the group | |
3737 | * has become inactive, some of its descendant processes still have | |
3738 | * some request already dispatched but still waiting for | |
3739 | * completion. In fact, requests have still to be guaranteed their | |
3740 | * share of the throughput even after being dispatched. In this | |
3741 | * respect, it is easy to show that, if a group frequently becomes | |
3742 | * inactive while still having in-flight requests, and if, when this | |
3743 | * happens, the group is not considered in the calculation of whether | |
3744 | * the scenario is asymmetric, then the group may fail to be | |
3745 | * guaranteed its fair share of the throughput (basically because | |
3746 | * idling may not be performed for the descendant processes of the | |
3747 | * group, but it had to be). We address this issue with the following | |
3748 | * bi-modal behavior, implemented in the function | |
3749 | * bfq_asymmetric_scenario(). | |
3750 | * | |
3751 | * If there are groups with requests waiting for completion | |
3752 | * (as commented above, some of these groups may even be | |
3753 | * already inactive), then the scenario is tagged as | |
3754 | * asymmetric, conservatively, without checking any of the | |
3755 | * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq. | |
3756 | * This behavior matches also the fact that groups are created | |
3757 | * exactly if controlling I/O is a primary concern (to | |
3758 | * preserve bandwidth and latency guarantees). | |
3759 | * | |
3760 | * On the opposite end, if there are no groups with requests waiting | |
3761 | * for completion, then only conditions (i-a) and (i-b) are actually | |
3762 | * controlled, i.e., provided that conditions (i-a) or (i-b) holds, | |
3763 | * idling is not performed, regardless of whether condition (ii) | |
3764 | * holds. In other words, only if conditions (i-a) and (i-b) do not | |
3765 | * hold, then idling is allowed, and the device tends to be prevented | |
3766 | * from queueing many requests, possibly of several processes. Since | |
3767 | * there are no groups with requests waiting for completion, then, to | |
3768 | * control conditions (i-a) and (i-b) it is enough to check just | |
3769 | * whether all the queues with requests waiting for completion also | |
3770 | * have the same weight. | |
3771 | * | |
3772 | * Not checking condition (ii) evidently exposes bfqq to the | |
3773 | * risk of getting less throughput than its fair share. | |
3774 | * However, for queues with the same weight, a further | |
3775 | * mechanism, preemption, mitigates or even eliminates this | |
3776 | * problem. And it does so without consequences on overall | |
3777 | * throughput. This mechanism and its benefits are explained | |
3778 | * in the next three paragraphs. | |
3779 | * | |
3780 | * Even if a queue, say Q, is expired when it remains idle, Q | |
3781 | * can still preempt the new in-service queue if the next | |
3782 | * request of Q arrives soon (see the comments on | |
3783 | * bfq_bfqq_update_budg_for_activation). If all queues and | |
3784 | * groups have the same weight, this form of preemption, | |
3785 | * combined with the hole-recovery heuristic described in the | |
3786 | * comments on function bfq_bfqq_update_budg_for_activation, | |
3787 | * are enough to preserve a correct bandwidth distribution in | |
3788 | * the mid term, even without idling. In fact, even if not | |
3789 | * idling allows the internal queues of the device to contain | |
3790 | * many requests, and thus to reorder requests, we can rather | |
3791 | * safely assume that the internal scheduler still preserves a | |
3792 | * minimum of mid-term fairness. | |
3793 | * | |
3794 | * More precisely, this preemption-based, idleless approach | |
3795 | * provides fairness in terms of IOPS, and not sectors per | |
3796 | * second. This can be seen with a simple example. Suppose | |
3797 | * that there are two queues with the same weight, but that | |
3798 | * the first queue receives requests of 8 sectors, while the | |
3799 | * second queue receives requests of 1024 sectors. In | |
3800 | * addition, suppose that each of the two queues contains at | |
3801 | * most one request at a time, which implies that each queue | |
3802 | * always remains idle after it is served. Finally, after | |
3803 | * remaining idle, each queue receives very quickly a new | |
3804 | * request. It follows that the two queues are served | |
3805 | * alternatively, preempting each other if needed. This | |
3806 | * implies that, although both queues have the same weight, | |
3807 | * the queue with large requests receives a service that is | |
3808 | * 1024/8 times as high as the service received by the other | |
3809 | * queue. | |
3810 | * | |
3811 | * The motivation for using preemption instead of idling (for | |
3812 | * queues with the same weight) is that, by not idling, | |
3813 | * service guarantees are preserved (completely or at least in | |
3814 | * part) without minimally sacrificing throughput. And, if | |
3815 | * there is no active group, then the primary expectation for | |
3816 | * this device is probably a high throughput. | |
3817 | * | |
b5e02b48 PV |
3818 | * We are now left only with explaining the two sub-conditions in the |
3819 | * additional compound condition that is checked below for deciding | |
3820 | * whether the scenario is asymmetric. To explain the first | |
3821 | * sub-condition, we need to add that the function | |
3726112e | 3822 | * bfq_asymmetric_scenario checks the weights of only |
b5e02b48 PV |
3823 | * non-weight-raised queues, for efficiency reasons (see comments on |
3824 | * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised | |
3825 | * is checked explicitly here. More precisely, the compound condition | |
3826 | * below takes into account also the fact that, even if bfqq is being | |
3827 | * weight-raised, the scenario is still symmetric if all queues with | |
3828 | * requests waiting for completion happen to be | |
3829 | * weight-raised. Actually, we should be even more precise here, and | |
3830 | * differentiate between interactive weight raising and soft real-time | |
3831 | * weight raising. | |
3832 | * | |
3833 | * The second sub-condition checked in the compound condition is | |
3834 | * whether there is a fair amount of already in-flight I/O not | |
3835 | * belonging to bfqq. If so, I/O dispatching is to be plugged, for the | |
3836 | * following reason. The drive may decide to serve in-flight | |
3837 | * non-bfqq's I/O requests before bfqq's ones, thereby delaying the | |
3838 | * arrival of new I/O requests for bfqq (recall that bfqq is sync). If | |
3839 | * I/O-dispatching is not plugged, then, while bfqq remains empty, a | |
3840 | * basically uncontrolled amount of I/O from other queues may be | |
3841 | * dispatched too, possibly causing the service of bfqq's I/O to be | |
3842 | * delayed even longer in the drive. This problem gets more and more | |
3843 | * serious as the speed and the queue depth of the drive grow, | |
3844 | * because, as these two quantities grow, the probability to find no | |
3845 | * queue busy but many requests in flight grows too. By contrast, | |
3846 | * plugging I/O dispatching minimizes the delay induced by already | |
3847 | * in-flight I/O, and enables bfqq to recover the bandwidth it may | |
3848 | * lose because of this delay. | |
3726112e PV |
3849 | * |
3850 | * As a side note, it is worth considering that the above | |
b5e02b48 PV |
3851 | * device-idling countermeasures may however fail in the following |
3852 | * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled | |
3853 | * in a time period during which all symmetry sub-conditions hold, and | |
3854 | * therefore the device is allowed to enqueue many requests, but at | |
3855 | * some later point in time some sub-condition stops to hold, then it | |
3856 | * may become impossible to make requests be served in the desired | |
3857 | * order until all the requests already queued in the device have been | |
3858 | * served. The last sub-condition commented above somewhat mitigates | |
3859 | * this problem for weight-raised queues. | |
2391d13e PV |
3860 | * |
3861 | * However, as an additional mitigation for this problem, we preserve | |
3862 | * plugging for a special symmetric case that may suddenly turn into | |
3863 | * asymmetric: the case where only bfqq is busy. In this case, not | |
3864 | * expiring bfqq does not cause any harm to any other queues in terms | |
3865 | * of service guarantees. In contrast, it avoids the following unlucky | |
3866 | * sequence of events: (1) bfqq is expired, (2) a new queue with a | |
3867 | * lower weight than bfqq becomes busy (or more queues), (3) the new | |
3868 | * queue is served until a new request arrives for bfqq, (4) when bfqq | |
3869 | * is finally served, there are so many requests of the new queue in | |
3870 | * the drive that the pending requests for bfqq take a lot of time to | |
3871 | * be served. In particular, event (2) may case even already | |
3872 | * dispatched requests of bfqq to be delayed, inside the drive. So, to | |
3873 | * avoid this series of events, the scenario is preventively declared | |
3874 | * as asymmetric also if bfqq is the only busy queues | |
3726112e PV |
3875 | */ |
3876 | static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd, | |
3877 | struct bfq_queue *bfqq) | |
3878 | { | |
2391d13e PV |
3879 | int tot_busy_queues = bfq_tot_busy_queues(bfqd); |
3880 | ||
f718b093 PV |
3881 | /* No point in idling for bfqq if it won't get requests any longer */ |
3882 | if (unlikely(!bfqq_process_refs(bfqq))) | |
3883 | return false; | |
3884 | ||
3726112e | 3885 | return (bfqq->wr_coeff > 1 && |
2d31c684 DZ |
3886 | (bfqd->wr_busy_queues < tot_busy_queues || |
3887 | bfqd->tot_rq_in_driver >= bfqq->dispatched + 4)) || | |
2391d13e PV |
3888 | bfq_asymmetric_scenario(bfqd, bfqq) || |
3889 | tot_busy_queues == 1; | |
3726112e PV |
3890 | } |
3891 | ||
3892 | static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
3893 | enum bfqq_expiration reason) | |
aee69d78 | 3894 | { |
36eca894 AA |
3895 | /* |
3896 | * If this bfqq is shared between multiple processes, check | |
3897 | * to make sure that those processes are still issuing I/Os | |
3898 | * within the mean seek distance. If not, it may be time to | |
3899 | * break the queues apart again. | |
3900 | */ | |
3901 | if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) | |
3902 | bfq_mark_bfqq_split_coop(bfqq); | |
3903 | ||
3726112e PV |
3904 | /* |
3905 | * Consider queues with a higher finish virtual time than | |
3906 | * bfqq. If idling_needed_for_service_guarantees(bfqq) returns | |
3907 | * true, then bfqq's bandwidth would be violated if an | |
3908 | * uncontrolled amount of I/O from these queues were | |
3909 | * dispatched while bfqq is waiting for its new I/O to | |
3910 | * arrive. This is exactly what may happen if this is a forced | |
3911 | * expiration caused by a preemption attempt, and if bfqq is | |
3912 | * not re-scheduled. To prevent this from happening, re-queue | |
3913 | * bfqq if it needs I/O-dispatch plugging, even if it is | |
3914 | * empty. By doing so, bfqq is granted to be served before the | |
3915 | * above queues (provided that bfqq is of course eligible). | |
3916 | */ | |
3917 | if (RB_EMPTY_ROOT(&bfqq->sort_list) && | |
3918 | !(reason == BFQQE_PREEMPTED && | |
3919 | idling_needed_for_service_guarantees(bfqd, bfqq))) { | |
44e44a1b PV |
3920 | if (bfqq->dispatched == 0) |
3921 | /* | |
3922 | * Overloading budget_timeout field to store | |
3923 | * the time at which the queue remains with no | |
3924 | * backlog and no outstanding request; used by | |
3925 | * the weight-raising mechanism. | |
3926 | */ | |
3927 | bfqq->budget_timeout = jiffies; | |
3928 | ||
d322f355 | 3929 | bfq_del_bfqq_busy(bfqq, true); |
36eca894 | 3930 | } else { |
80294c3b | 3931 | bfq_requeue_bfqq(bfqd, bfqq, true); |
36eca894 AA |
3932 | /* |
3933 | * Resort priority tree of potential close cooperators. | |
8cacc5ab | 3934 | * See comments on bfq_pos_tree_add_move() for the unlikely(). |
36eca894 | 3935 | */ |
3726112e PV |
3936 | if (unlikely(!bfqd->nonrot_with_queueing && |
3937 | !RB_EMPTY_ROOT(&bfqq->sort_list))) | |
8cacc5ab | 3938 | bfq_pos_tree_add_move(bfqd, bfqq); |
36eca894 | 3939 | } |
e21b7a0b AA |
3940 | |
3941 | /* | |
3942 | * All in-service entities must have been properly deactivated | |
3943 | * or requeued before executing the next function, which | |
eed47d19 PV |
3944 | * resets all in-service entities as no more in service. This |
3945 | * may cause bfqq to be freed. If this happens, the next | |
3946 | * function returns true. | |
e21b7a0b | 3947 | */ |
eed47d19 | 3948 | return __bfq_bfqd_reset_in_service(bfqd); |
aee69d78 PV |
3949 | } |
3950 | ||
3951 | /** | |
3952 | * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. | |
3953 | * @bfqd: device data. | |
3954 | * @bfqq: queue to update. | |
3955 | * @reason: reason for expiration. | |
3956 | * | |
3957 | * Handle the feedback on @bfqq budget at queue expiration. | |
3958 | * See the body for detailed comments. | |
3959 | */ | |
3960 | static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, | |
3961 | struct bfq_queue *bfqq, | |
3962 | enum bfqq_expiration reason) | |
3963 | { | |
3964 | struct request *next_rq; | |
3965 | int budget, min_budget; | |
3966 | ||
aee69d78 PV |
3967 | min_budget = bfq_min_budget(bfqd); |
3968 | ||
44e44a1b PV |
3969 | if (bfqq->wr_coeff == 1) |
3970 | budget = bfqq->max_budget; | |
3971 | else /* | |
3972 | * Use a constant, low budget for weight-raised queues, | |
3973 | * to help achieve a low latency. Keep it slightly higher | |
3974 | * than the minimum possible budget, to cause a little | |
3975 | * bit fewer expirations. | |
3976 | */ | |
3977 | budget = 2 * min_budget; | |
3978 | ||
aee69d78 PV |
3979 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d", |
3980 | bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); | |
3981 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d", | |
3982 | budget, bfq_min_budget(bfqd)); | |
3983 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", | |
3984 | bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); | |
3985 | ||
44e44a1b | 3986 | if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) { |
aee69d78 PV |
3987 | switch (reason) { |
3988 | /* | |
3989 | * Caveat: in all the following cases we trade latency | |
3990 | * for throughput. | |
3991 | */ | |
3992 | case BFQQE_TOO_IDLE: | |
54b60456 PV |
3993 | /* |
3994 | * This is the only case where we may reduce | |
3995 | * the budget: if there is no request of the | |
3996 | * process still waiting for completion, then | |
3997 | * we assume (tentatively) that the timer has | |
3998 | * expired because the batch of requests of | |
3999 | * the process could have been served with a | |
4000 | * smaller budget. Hence, betting that | |
4001 | * process will behave in the same way when it | |
4002 | * becomes backlogged again, we reduce its | |
4003 | * next budget. As long as we guess right, | |
4004 | * this budget cut reduces the latency | |
4005 | * experienced by the process. | |
4006 | * | |
4007 | * However, if there are still outstanding | |
4008 | * requests, then the process may have not yet | |
4009 | * issued its next request just because it is | |
4010 | * still waiting for the completion of some of | |
4011 | * the still outstanding ones. So in this | |
4012 | * subcase we do not reduce its budget, on the | |
4013 | * contrary we increase it to possibly boost | |
4014 | * the throughput, as discussed in the | |
4015 | * comments to the BUDGET_TIMEOUT case. | |
4016 | */ | |
4017 | if (bfqq->dispatched > 0) /* still outstanding reqs */ | |
4018 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
4019 | else { | |
4020 | if (budget > 5 * min_budget) | |
4021 | budget -= 4 * min_budget; | |
4022 | else | |
4023 | budget = min_budget; | |
4024 | } | |
aee69d78 PV |
4025 | break; |
4026 | case BFQQE_BUDGET_TIMEOUT: | |
54b60456 PV |
4027 | /* |
4028 | * We double the budget here because it gives | |
4029 | * the chance to boost the throughput if this | |
4030 | * is not a seeky process (and has bumped into | |
4031 | * this timeout because of, e.g., ZBR). | |
4032 | */ | |
4033 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
aee69d78 PV |
4034 | break; |
4035 | case BFQQE_BUDGET_EXHAUSTED: | |
4036 | /* | |
4037 | * The process still has backlog, and did not | |
4038 | * let either the budget timeout or the disk | |
4039 | * idling timeout expire. Hence it is not | |
4040 | * seeky, has a short thinktime and may be | |
4041 | * happy with a higher budget too. So | |
4042 | * definitely increase the budget of this good | |
4043 | * candidate to boost the disk throughput. | |
4044 | */ | |
54b60456 | 4045 | budget = min(budget * 4, bfqd->bfq_max_budget); |
aee69d78 PV |
4046 | break; |
4047 | case BFQQE_NO_MORE_REQUESTS: | |
4048 | /* | |
4049 | * For queues that expire for this reason, it | |
4050 | * is particularly important to keep the | |
4051 | * budget close to the actual service they | |
4052 | * need. Doing so reduces the timestamp | |
4053 | * misalignment problem described in the | |
4054 | * comments in the body of | |
4055 | * __bfq_activate_entity. In fact, suppose | |
4056 | * that a queue systematically expires for | |
4057 | * BFQQE_NO_MORE_REQUESTS and presents a | |
4058 | * new request in time to enjoy timestamp | |
4059 | * back-shifting. The larger the budget of the | |
4060 | * queue is with respect to the service the | |
4061 | * queue actually requests in each service | |
4062 | * slot, the more times the queue can be | |
4063 | * reactivated with the same virtual finish | |
4064 | * time. It follows that, even if this finish | |
4065 | * time is pushed to the system virtual time | |
4066 | * to reduce the consequent timestamp | |
4067 | * misalignment, the queue unjustly enjoys for | |
4068 | * many re-activations a lower finish time | |
4069 | * than all newly activated queues. | |
4070 | * | |
4071 | * The service needed by bfqq is measured | |
4072 | * quite precisely by bfqq->entity.service. | |
4073 | * Since bfqq does not enjoy device idling, | |
4074 | * bfqq->entity.service is equal to the number | |
4075 | * of sectors that the process associated with | |
4076 | * bfqq requested to read/write before waiting | |
4077 | * for request completions, or blocking for | |
4078 | * other reasons. | |
4079 | */ | |
4080 | budget = max_t(int, bfqq->entity.service, min_budget); | |
4081 | break; | |
4082 | default: | |
4083 | return; | |
4084 | } | |
44e44a1b | 4085 | } else if (!bfq_bfqq_sync(bfqq)) { |
aee69d78 PV |
4086 | /* |
4087 | * Async queues get always the maximum possible | |
4088 | * budget, as for them we do not care about latency | |
4089 | * (in addition, their ability to dispatch is limited | |
4090 | * by the charging factor). | |
4091 | */ | |
4092 | budget = bfqd->bfq_max_budget; | |
4093 | } | |
4094 | ||
4095 | bfqq->max_budget = budget; | |
4096 | ||
4097 | if (bfqd->budgets_assigned >= bfq_stats_min_budgets && | |
4098 | !bfqd->bfq_user_max_budget) | |
4099 | bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget); | |
4100 | ||
4101 | /* | |
4102 | * If there is still backlog, then assign a new budget, making | |
4103 | * sure that it is large enough for the next request. Since | |
4104 | * the finish time of bfqq must be kept in sync with the | |
4105 | * budget, be sure to call __bfq_bfqq_expire() *after* this | |
4106 | * update. | |
4107 | * | |
4108 | * If there is no backlog, then no need to update the budget; | |
4109 | * it will be updated on the arrival of a new request. | |
4110 | */ | |
4111 | next_rq = bfqq->next_rq; | |
4112 | if (next_rq) | |
4113 | bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, | |
4114 | bfq_serv_to_charge(next_rq, bfqq)); | |
4115 | ||
4116 | bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d", | |
4117 | next_rq ? blk_rq_sectors(next_rq) : 0, | |
4118 | bfqq->entity.budget); | |
4119 | } | |
4120 | ||
aee69d78 | 4121 | /* |
ab0e43e9 PV |
4122 | * Return true if the process associated with bfqq is "slow". The slow |
4123 | * flag is used, in addition to the budget timeout, to reduce the | |
4124 | * amount of service provided to seeky processes, and thus reduce | |
4125 | * their chances to lower the throughput. More details in the comments | |
4126 | * on the function bfq_bfqq_expire(). | |
4127 | * | |
4128 | * An important observation is in order: as discussed in the comments | |
4129 | * on the function bfq_update_peak_rate(), with devices with internal | |
4130 | * queues, it is hard if ever possible to know when and for how long | |
4131 | * an I/O request is processed by the device (apart from the trivial | |
4132 | * I/O pattern where a new request is dispatched only after the | |
4133 | * previous one has been completed). This makes it hard to evaluate | |
4134 | * the real rate at which the I/O requests of each bfq_queue are | |
4135 | * served. In fact, for an I/O scheduler like BFQ, serving a | |
4136 | * bfq_queue means just dispatching its requests during its service | |
4137 | * slot (i.e., until the budget of the queue is exhausted, or the | |
4138 | * queue remains idle, or, finally, a timeout fires). But, during the | |
4139 | * service slot of a bfq_queue, around 100 ms at most, the device may | |
4140 | * be even still processing requests of bfq_queues served in previous | |
4141 | * service slots. On the opposite end, the requests of the in-service | |
4142 | * bfq_queue may be completed after the service slot of the queue | |
4143 | * finishes. | |
4144 | * | |
4145 | * Anyway, unless more sophisticated solutions are used | |
4146 | * (where possible), the sum of the sizes of the requests dispatched | |
4147 | * during the service slot of a bfq_queue is probably the only | |
4148 | * approximation available for the service received by the bfq_queue | |
4149 | * during its service slot. And this sum is the quantity used in this | |
4150 | * function to evaluate the I/O speed of a process. | |
aee69d78 | 4151 | */ |
ab0e43e9 | 4152 | static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
bebeb9e5 | 4153 | bool compensate, unsigned long *delta_ms) |
aee69d78 | 4154 | { |
ab0e43e9 PV |
4155 | ktime_t delta_ktime; |
4156 | u32 delta_usecs; | |
4157 | bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */ | |
aee69d78 | 4158 | |
ab0e43e9 | 4159 | if (!bfq_bfqq_sync(bfqq)) |
aee69d78 PV |
4160 | return false; |
4161 | ||
4162 | if (compensate) | |
ab0e43e9 | 4163 | delta_ktime = bfqd->last_idling_start; |
aee69d78 | 4164 | else |
ab0e43e9 PV |
4165 | delta_ktime = ktime_get(); |
4166 | delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start); | |
4167 | delta_usecs = ktime_to_us(delta_ktime); | |
aee69d78 PV |
4168 | |
4169 | /* don't use too short time intervals */ | |
ab0e43e9 PV |
4170 | if (delta_usecs < 1000) { |
4171 | if (blk_queue_nonrot(bfqd->queue)) | |
4172 | /* | |
4173 | * give same worst-case guarantees as idling | |
4174 | * for seeky | |
4175 | */ | |
4176 | *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC; | |
4177 | else /* charge at least one seek */ | |
4178 | *delta_ms = bfq_slice_idle / NSEC_PER_MSEC; | |
4179 | ||
4180 | return slow; | |
4181 | } | |
aee69d78 | 4182 | |
ab0e43e9 | 4183 | *delta_ms = delta_usecs / USEC_PER_MSEC; |
aee69d78 PV |
4184 | |
4185 | /* | |
ab0e43e9 PV |
4186 | * Use only long (> 20ms) intervals to filter out excessive |
4187 | * spikes in service rate estimation. | |
aee69d78 | 4188 | */ |
ab0e43e9 PV |
4189 | if (delta_usecs > 20000) { |
4190 | /* | |
4191 | * Caveat for rotational devices: processes doing I/O | |
4192 | * in the slower disk zones tend to be slow(er) even | |
4193 | * if not seeky. In this respect, the estimated peak | |
4194 | * rate is likely to be an average over the disk | |
4195 | * surface. Accordingly, to not be too harsh with | |
4196 | * unlucky processes, a process is deemed slow only if | |
4197 | * its rate has been lower than half of the estimated | |
4198 | * peak rate. | |
4199 | */ | |
4200 | slow = bfqq->entity.service < bfqd->bfq_max_budget / 2; | |
aee69d78 PV |
4201 | } |
4202 | ||
ab0e43e9 | 4203 | bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow); |
aee69d78 | 4204 | |
ab0e43e9 | 4205 | return slow; |
aee69d78 PV |
4206 | } |
4207 | ||
77b7dcea PV |
4208 | /* |
4209 | * To be deemed as soft real-time, an application must meet two | |
4210 | * requirements. First, the application must not require an average | |
4211 | * bandwidth higher than the approximate bandwidth required to playback or | |
4212 | * record a compressed high-definition video. | |
4213 | * The next function is invoked on the completion of the last request of a | |
4214 | * batch, to compute the next-start time instant, soft_rt_next_start, such | |
4215 | * that, if the next request of the application does not arrive before | |
4216 | * soft_rt_next_start, then the above requirement on the bandwidth is met. | |
4217 | * | |
4218 | * The second requirement is that the request pattern of the application is | |
4219 | * isochronous, i.e., that, after issuing a request or a batch of requests, | |
4220 | * the application stops issuing new requests until all its pending requests | |
4221 | * have been completed. After that, the application may issue a new batch, | |
4222 | * and so on. | |
4223 | * For this reason the next function is invoked to compute | |
4224 | * soft_rt_next_start only for applications that meet this requirement, | |
4225 | * whereas soft_rt_next_start is set to infinity for applications that do | |
4226 | * not. | |
4227 | * | |
a34b0244 PV |
4228 | * Unfortunately, even a greedy (i.e., I/O-bound) application may |
4229 | * happen to meet, occasionally or systematically, both the above | |
4230 | * bandwidth and isochrony requirements. This may happen at least in | |
4231 | * the following circumstances. First, if the CPU load is high. The | |
4232 | * application may stop issuing requests while the CPUs are busy | |
4233 | * serving other processes, then restart, then stop again for a while, | |
4234 | * and so on. The other circumstances are related to the storage | |
4235 | * device: the storage device is highly loaded or reaches a low-enough | |
4236 | * throughput with the I/O of the application (e.g., because the I/O | |
4237 | * is random and/or the device is slow). In all these cases, the | |
4238 | * I/O of the application may be simply slowed down enough to meet | |
4239 | * the bandwidth and isochrony requirements. To reduce the probability | |
4240 | * that greedy applications are deemed as soft real-time in these | |
4241 | * corner cases, a further rule is used in the computation of | |
4242 | * soft_rt_next_start: the return value of this function is forced to | |
4243 | * be higher than the maximum between the following two quantities. | |
4244 | * | |
4245 | * (a) Current time plus: (1) the maximum time for which the arrival | |
4246 | * of a request is waited for when a sync queue becomes idle, | |
4247 | * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We | |
4248 | * postpone for a moment the reason for adding a few extra | |
4249 | * jiffies; we get back to it after next item (b). Lower-bounding | |
4250 | * the return value of this function with the current time plus | |
4251 | * bfqd->bfq_slice_idle tends to filter out greedy applications, | |
4252 | * because the latter issue their next request as soon as possible | |
4253 | * after the last one has been completed. In contrast, a soft | |
4254 | * real-time application spends some time processing data, after a | |
4255 | * batch of its requests has been completed. | |
4256 | * | |
4257 | * (b) Current value of bfqq->soft_rt_next_start. As pointed out | |
4258 | * above, greedy applications may happen to meet both the | |
4259 | * bandwidth and isochrony requirements under heavy CPU or | |
4260 | * storage-device load. In more detail, in these scenarios, these | |
4261 | * applications happen, only for limited time periods, to do I/O | |
4262 | * slowly enough to meet all the requirements described so far, | |
4263 | * including the filtering in above item (a). These slow-speed | |
4264 | * time intervals are usually interspersed between other time | |
4265 | * intervals during which these applications do I/O at a very high | |
4266 | * speed. Fortunately, exactly because of the high speed of the | |
4267 | * I/O in the high-speed intervals, the values returned by this | |
4268 | * function happen to be so high, near the end of any such | |
4269 | * high-speed interval, to be likely to fall *after* the end of | |
4270 | * the low-speed time interval that follows. These high values are | |
4271 | * stored in bfqq->soft_rt_next_start after each invocation of | |
4272 | * this function. As a consequence, if the last value of | |
4273 | * bfqq->soft_rt_next_start is constantly used to lower-bound the | |
4274 | * next value that this function may return, then, from the very | |
4275 | * beginning of a low-speed interval, bfqq->soft_rt_next_start is | |
4276 | * likely to be constantly kept so high that any I/O request | |
4277 | * issued during the low-speed interval is considered as arriving | |
4278 | * to soon for the application to be deemed as soft | |
4279 | * real-time. Then, in the high-speed interval that follows, the | |
4280 | * application will not be deemed as soft real-time, just because | |
4281 | * it will do I/O at a high speed. And so on. | |
4282 | * | |
4283 | * Getting back to the filtering in item (a), in the following two | |
4284 | * cases this filtering might be easily passed by a greedy | |
4285 | * application, if the reference quantity was just | |
4286 | * bfqd->bfq_slice_idle: | |
4287 | * 1) HZ is so low that the duration of a jiffy is comparable to or | |
4288 | * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow | |
4289 | * devices with HZ=100. The time granularity may be so coarse | |
4290 | * that the approximation, in jiffies, of bfqd->bfq_slice_idle | |
4291 | * is rather lower than the exact value. | |
77b7dcea PV |
4292 | * 2) jiffies, instead of increasing at a constant rate, may stop increasing |
4293 | * for a while, then suddenly 'jump' by several units to recover the lost | |
4294 | * increments. This seems to happen, e.g., inside virtual machines. | |
a34b0244 PV |
4295 | * To address this issue, in the filtering in (a) we do not use as a |
4296 | * reference time interval just bfqd->bfq_slice_idle, but | |
4297 | * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the | |
4298 | * minimum number of jiffies for which the filter seems to be quite | |
4299 | * precise also in embedded systems and KVM/QEMU virtual machines. | |
77b7dcea PV |
4300 | */ |
4301 | static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd, | |
4302 | struct bfq_queue *bfqq) | |
4303 | { | |
a34b0244 PV |
4304 | return max3(bfqq->soft_rt_next_start, |
4305 | bfqq->last_idle_bklogged + | |
4306 | HZ * bfqq->service_from_backlogged / | |
4307 | bfqd->bfq_wr_max_softrt_rate, | |
4308 | jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4); | |
77b7dcea PV |
4309 | } |
4310 | ||
aee69d78 PV |
4311 | /** |
4312 | * bfq_bfqq_expire - expire a queue. | |
4313 | * @bfqd: device owning the queue. | |
4314 | * @bfqq: the queue to expire. | |
4315 | * @compensate: if true, compensate for the time spent idling. | |
4316 | * @reason: the reason causing the expiration. | |
4317 | * | |
c074170e PV |
4318 | * If the process associated with bfqq does slow I/O (e.g., because it |
4319 | * issues random requests), we charge bfqq with the time it has been | |
4320 | * in service instead of the service it has received (see | |
4321 | * bfq_bfqq_charge_time for details on how this goal is achieved). As | |
4322 | * a consequence, bfqq will typically get higher timestamps upon | |
4323 | * reactivation, and hence it will be rescheduled as if it had | |
4324 | * received more service than what it has actually received. In the | |
4325 | * end, bfqq receives less service in proportion to how slowly its | |
4326 | * associated process consumes its budgets (and hence how seriously it | |
4327 | * tends to lower the throughput). In addition, this time-charging | |
4328 | * strategy guarantees time fairness among slow processes. In | |
4329 | * contrast, if the process associated with bfqq is not slow, we | |
4330 | * charge bfqq exactly with the service it has received. | |
aee69d78 | 4331 | * |
c074170e PV |
4332 | * Charging time to the first type of queues and the exact service to |
4333 | * the other has the effect of using the WF2Q+ policy to schedule the | |
4334 | * former on a timeslice basis, without violating service domain | |
4335 | * guarantees among the latter. | |
aee69d78 | 4336 | */ |
ea25da48 PV |
4337 | void bfq_bfqq_expire(struct bfq_data *bfqd, |
4338 | struct bfq_queue *bfqq, | |
4339 | bool compensate, | |
4340 | enum bfqq_expiration reason) | |
aee69d78 PV |
4341 | { |
4342 | bool slow; | |
ab0e43e9 PV |
4343 | unsigned long delta = 0; |
4344 | struct bfq_entity *entity = &bfqq->entity; | |
aee69d78 PV |
4345 | |
4346 | /* | |
ab0e43e9 | 4347 | * Check whether the process is slow (see bfq_bfqq_is_slow). |
aee69d78 | 4348 | */ |
bebeb9e5 | 4349 | slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, &delta); |
aee69d78 PV |
4350 | |
4351 | /* | |
c074170e PV |
4352 | * As above explained, charge slow (typically seeky) and |
4353 | * timed-out queues with the time and not the service | |
4354 | * received, to favor sequential workloads. | |
4355 | * | |
4356 | * Processes doing I/O in the slower disk zones will tend to | |
4357 | * be slow(er) even if not seeky. Therefore, since the | |
4358 | * estimated peak rate is actually an average over the disk | |
4359 | * surface, these processes may timeout just for bad luck. To | |
4360 | * avoid punishing them, do not charge time to processes that | |
4361 | * succeeded in consuming at least 2/3 of their budget. This | |
4362 | * allows BFQ to preserve enough elasticity to still perform | |
4363 | * bandwidth, and not time, distribution with little unlucky | |
4364 | * or quasi-sequential processes. | |
aee69d78 | 4365 | */ |
44e44a1b PV |
4366 | if (bfqq->wr_coeff == 1 && |
4367 | (slow || | |
4368 | (reason == BFQQE_BUDGET_TIMEOUT && | |
4369 | bfq_bfqq_budget_left(bfqq) >= entity->budget / 3))) | |
c074170e | 4370 | bfq_bfqq_charge_time(bfqd, bfqq, delta); |
aee69d78 | 4371 | |
44e44a1b PV |
4372 | if (bfqd->low_latency && bfqq->wr_coeff == 1) |
4373 | bfqq->last_wr_start_finish = jiffies; | |
4374 | ||
77b7dcea PV |
4375 | if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 && |
4376 | RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
4377 | /* | |
4378 | * If we get here, and there are no outstanding | |
4379 | * requests, then the request pattern is isochronous | |
4380 | * (see the comments on the function | |
3c337690 PV |
4381 | * bfq_bfqq_softrt_next_start()). Therefore we can |
4382 | * compute soft_rt_next_start. | |
20cd3245 PV |
4383 | * |
4384 | * If, instead, the queue still has outstanding | |
4385 | * requests, then we have to wait for the completion | |
4386 | * of all the outstanding requests to discover whether | |
4387 | * the request pattern is actually isochronous. | |
77b7dcea | 4388 | */ |
3c337690 | 4389 | if (bfqq->dispatched == 0) |
77b7dcea PV |
4390 | bfqq->soft_rt_next_start = |
4391 | bfq_bfqq_softrt_next_start(bfqd, bfqq); | |
20cd3245 | 4392 | else if (bfqq->dispatched > 0) { |
77b7dcea PV |
4393 | /* |
4394 | * Schedule an update of soft_rt_next_start to when | |
4395 | * the task may be discovered to be isochronous. | |
4396 | */ | |
4397 | bfq_mark_bfqq_softrt_update(bfqq); | |
4398 | } | |
4399 | } | |
4400 | ||
aee69d78 | 4401 | bfq_log_bfqq(bfqd, bfqq, |
d5be3fef PV |
4402 | "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason, |
4403 | slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq)); | |
aee69d78 | 4404 | |
2341d662 PV |
4405 | /* |
4406 | * bfqq expired, so no total service time needs to be computed | |
4407 | * any longer: reset state machine for measuring total service | |
4408 | * times. | |
4409 | */ | |
4410 | bfqd->rqs_injected = bfqd->wait_dispatch = false; | |
4411 | bfqd->waited_rq = NULL; | |
4412 | ||
aee69d78 PV |
4413 | /* |
4414 | * Increase, decrease or leave budget unchanged according to | |
4415 | * reason. | |
4416 | */ | |
4417 | __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); | |
3726112e | 4418 | if (__bfq_bfqq_expire(bfqd, bfqq, reason)) |
eed47d19 | 4419 | /* bfqq is gone, no more actions on it */ |
9fae8dd5 PV |
4420 | return; |
4421 | ||
aee69d78 | 4422 | /* mark bfqq as waiting a request only if a bic still points to it */ |
9fae8dd5 | 4423 | if (!bfq_bfqq_busy(bfqq) && |
aee69d78 | 4424 | reason != BFQQE_BUDGET_TIMEOUT && |
9fae8dd5 | 4425 | reason != BFQQE_BUDGET_EXHAUSTED) { |
aee69d78 | 4426 | bfq_mark_bfqq_non_blocking_wait_rq(bfqq); |
9fae8dd5 PV |
4427 | /* |
4428 | * Not setting service to 0, because, if the next rq | |
4429 | * arrives in time, the queue will go on receiving | |
4430 | * service with this same budget (as if it never expired) | |
4431 | */ | |
4432 | } else | |
4433 | entity->service = 0; | |
8a511ba5 PV |
4434 | |
4435 | /* | |
4436 | * Reset the received-service counter for every parent entity. | |
4437 | * Differently from what happens with bfqq->entity.service, | |
4438 | * the resetting of this counter never needs to be postponed | |
4439 | * for parent entities. In fact, in case bfqq may have a | |
4440 | * chance to go on being served using the last, partially | |
4441 | * consumed budget, bfqq->entity.service needs to be kept, | |
4442 | * because if bfqq then actually goes on being served using | |
4443 | * the same budget, the last value of bfqq->entity.service is | |
4444 | * needed to properly decrement bfqq->entity.budget by the | |
4445 | * portion already consumed. In contrast, it is not necessary | |
4446 | * to keep entity->service for parent entities too, because | |
4447 | * the bubble up of the new value of bfqq->entity.budget will | |
4448 | * make sure that the budgets of parent entities are correct, | |
4449 | * even in case bfqq and thus parent entities go on receiving | |
4450 | * service with the same budget. | |
4451 | */ | |
4452 | entity = entity->parent; | |
4453 | for_each_entity(entity) | |
4454 | entity->service = 0; | |
aee69d78 PV |
4455 | } |
4456 | ||
4457 | /* | |
4458 | * Budget timeout is not implemented through a dedicated timer, but | |
4459 | * just checked on request arrivals and completions, as well as on | |
4460 | * idle timer expirations. | |
4461 | */ | |
4462 | static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) | |
4463 | { | |
44e44a1b | 4464 | return time_is_before_eq_jiffies(bfqq->budget_timeout); |
aee69d78 PV |
4465 | } |
4466 | ||
4467 | /* | |
4468 | * If we expire a queue that is actively waiting (i.e., with the | |
4469 | * device idled) for the arrival of a new request, then we may incur | |
4470 | * the timestamp misalignment problem described in the body of the | |
4471 | * function __bfq_activate_entity. Hence we return true only if this | |
4472 | * condition does not hold, or if the queue is slow enough to deserve | |
4473 | * only to be kicked off for preserving a high throughput. | |
4474 | */ | |
4475 | static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) | |
4476 | { | |
4477 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
4478 | "may_budget_timeout: wait_request %d left %d timeout %d", | |
4479 | bfq_bfqq_wait_request(bfqq), | |
4480 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, | |
4481 | bfq_bfqq_budget_timeout(bfqq)); | |
4482 | ||
4483 | return (!bfq_bfqq_wait_request(bfqq) || | |
4484 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) | |
4485 | && | |
4486 | bfq_bfqq_budget_timeout(bfqq); | |
4487 | } | |
4488 | ||
05c2f5c3 PV |
4489 | static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd, |
4490 | struct bfq_queue *bfqq) | |
aee69d78 | 4491 | { |
edaf9428 PV |
4492 | bool rot_without_queueing = |
4493 | !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag, | |
4494 | bfqq_sequential_and_IO_bound, | |
05c2f5c3 | 4495 | idling_boosts_thr; |
d5be3fef | 4496 | |
f718b093 PV |
4497 | /* No point in idling for bfqq if it won't get requests any longer */ |
4498 | if (unlikely(!bfqq_process_refs(bfqq))) | |
4499 | return false; | |
4500 | ||
edaf9428 PV |
4501 | bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) && |
4502 | bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq); | |
4503 | ||
aee69d78 | 4504 | /* |
44e44a1b PV |
4505 | * The next variable takes into account the cases where idling |
4506 | * boosts the throughput. | |
4507 | * | |
e01eff01 PV |
4508 | * The value of the variable is computed considering, first, that |
4509 | * idling is virtually always beneficial for the throughput if: | |
edaf9428 PV |
4510 | * (a) the device is not NCQ-capable and rotational, or |
4511 | * (b) regardless of the presence of NCQ, the device is rotational and | |
4512 | * the request pattern for bfqq is I/O-bound and sequential, or | |
4513 | * (c) regardless of whether it is rotational, the device is | |
4514 | * not NCQ-capable and the request pattern for bfqq is | |
4515 | * I/O-bound and sequential. | |
bf2b79e7 PV |
4516 | * |
4517 | * Secondly, and in contrast to the above item (b), idling an | |
4518 | * NCQ-capable flash-based device would not boost the | |
e01eff01 | 4519 | * throughput even with sequential I/O; rather it would lower |
bf2b79e7 PV |
4520 | * the throughput in proportion to how fast the device |
4521 | * is. Accordingly, the next variable is true if any of the | |
edaf9428 PV |
4522 | * above conditions (a), (b) or (c) is true, and, in |
4523 | * particular, happens to be false if bfqd is an NCQ-capable | |
4524 | * flash-based device. | |
aee69d78 | 4525 | */ |
edaf9428 PV |
4526 | idling_boosts_thr = rot_without_queueing || |
4527 | ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) && | |
4528 | bfqq_sequential_and_IO_bound); | |
aee69d78 | 4529 | |
cfd69712 | 4530 | /* |
05c2f5c3 | 4531 | * The return value of this function is equal to that of |
cfd69712 PV |
4532 | * idling_boosts_thr, unless a special case holds. In this |
4533 | * special case, described below, idling may cause problems to | |
4534 | * weight-raised queues. | |
4535 | * | |
4536 | * When the request pool is saturated (e.g., in the presence | |
4537 | * of write hogs), if the processes associated with | |
4538 | * non-weight-raised queues ask for requests at a lower rate, | |
4539 | * then processes associated with weight-raised queues have a | |
4540 | * higher probability to get a request from the pool | |
4541 | * immediately (or at least soon) when they need one. Thus | |
4542 | * they have a higher probability to actually get a fraction | |
4543 | * of the device throughput proportional to their high | |
4544 | * weight. This is especially true with NCQ-capable drives, | |
4545 | * which enqueue several requests in advance, and further | |
4546 | * reorder internally-queued requests. | |
4547 | * | |
05c2f5c3 PV |
4548 | * For this reason, we force to false the return value if |
4549 | * there are weight-raised busy queues. In this case, and if | |
4550 | * bfqq is not weight-raised, this guarantees that the device | |
4551 | * is not idled for bfqq (if, instead, bfqq is weight-raised, | |
4552 | * then idling will be guaranteed by another variable, see | |
4553 | * below). Combined with the timestamping rules of BFQ (see | |
4554 | * [1] for details), this behavior causes bfqq, and hence any | |
4555 | * sync non-weight-raised queue, to get a lower number of | |
4556 | * requests served, and thus to ask for a lower number of | |
4557 | * requests from the request pool, before the busy | |
4558 | * weight-raised queues get served again. This often mitigates | |
4559 | * starvation problems in the presence of heavy write | |
4560 | * workloads and NCQ, thereby guaranteeing a higher | |
4561 | * application and system responsiveness in these hostile | |
4562 | * scenarios. | |
4563 | */ | |
4564 | return idling_boosts_thr && | |
cfd69712 | 4565 | bfqd->wr_busy_queues == 0; |
05c2f5c3 | 4566 | } |
cfd69712 | 4567 | |
05c2f5c3 PV |
4568 | /* |
4569 | * For a queue that becomes empty, device idling is allowed only if | |
4570 | * this function returns true for that queue. As a consequence, since | |
4571 | * device idling plays a critical role for both throughput boosting | |
4572 | * and service guarantees, the return value of this function plays a | |
4573 | * critical role as well. | |
4574 | * | |
4575 | * In a nutshell, this function returns true only if idling is | |
4576 | * beneficial for throughput or, even if detrimental for throughput, | |
4577 | * idling is however necessary to preserve service guarantees (low | |
4578 | * latency, desired throughput distribution, ...). In particular, on | |
4579 | * NCQ-capable devices, this function tries to return false, so as to | |
4580 | * help keep the drives' internal queues full, whenever this helps the | |
4581 | * device boost the throughput without causing any service-guarantee | |
4582 | * issue. | |
4583 | * | |
4584 | * Most of the issues taken into account to get the return value of | |
4585 | * this function are not trivial. We discuss these issues in the two | |
4586 | * functions providing the main pieces of information needed by this | |
4587 | * function. | |
4588 | */ | |
4589 | static bool bfq_better_to_idle(struct bfq_queue *bfqq) | |
4590 | { | |
4591 | struct bfq_data *bfqd = bfqq->bfqd; | |
4592 | bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar; | |
4593 | ||
f718b093 PV |
4594 | /* No point in idling for bfqq if it won't get requests any longer */ |
4595 | if (unlikely(!bfqq_process_refs(bfqq))) | |
4596 | return false; | |
4597 | ||
05c2f5c3 PV |
4598 | if (unlikely(bfqd->strict_guarantees)) |
4599 | return true; | |
4600 | ||
4601 | /* | |
4602 | * Idling is performed only if slice_idle > 0. In addition, we | |
4603 | * do not idle if | |
4604 | * (a) bfqq is async | |
4605 | * (b) bfqq is in the idle io prio class: in this case we do | |
4606 | * not idle because we want to minimize the bandwidth that | |
4607 | * queues in this class can steal to higher-priority queues | |
4608 | */ | |
4609 | if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) || | |
4610 | bfq_class_idle(bfqq)) | |
4611 | return false; | |
4612 | ||
4613 | idling_boosts_thr_with_no_issue = | |
4614 | idling_boosts_thr_without_issues(bfqd, bfqq); | |
4615 | ||
4616 | idling_needed_for_service_guar = | |
4617 | idling_needed_for_service_guarantees(bfqd, bfqq); | |
e1b2324d | 4618 | |
44e44a1b | 4619 | /* |
05c2f5c3 | 4620 | * We have now the two components we need to compute the |
d5be3fef PV |
4621 | * return value of the function, which is true only if idling |
4622 | * either boosts the throughput (without issues), or is | |
4623 | * necessary to preserve service guarantees. | |
aee69d78 | 4624 | */ |
05c2f5c3 PV |
4625 | return idling_boosts_thr_with_no_issue || |
4626 | idling_needed_for_service_guar; | |
aee69d78 PV |
4627 | } |
4628 | ||
4629 | /* | |
277a4a9b | 4630 | * If the in-service queue is empty but the function bfq_better_to_idle |
aee69d78 PV |
4631 | * returns true, then: |
4632 | * 1) the queue must remain in service and cannot be expired, and | |
4633 | * 2) the device must be idled to wait for the possible arrival of a new | |
4634 | * request for the queue. | |
277a4a9b | 4635 | * See the comments on the function bfq_better_to_idle for the reasons |
aee69d78 | 4636 | * why performing device idling is the best choice to boost the throughput |
277a4a9b | 4637 | * and preserve service guarantees when bfq_better_to_idle itself |
aee69d78 PV |
4638 | * returns true. |
4639 | */ | |
4640 | static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) | |
4641 | { | |
277a4a9b | 4642 | return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq); |
aee69d78 PV |
4643 | } |
4644 | ||
2341d662 PV |
4645 | /* |
4646 | * This function chooses the queue from which to pick the next extra | |
4647 | * I/O request to inject, if it finds a compatible queue. See the | |
4648 | * comments on bfq_update_inject_limit() for details on the injection | |
4649 | * mechanism, and for the definitions of the quantities mentioned | |
4650 | * below. | |
4651 | */ | |
4652 | static struct bfq_queue * | |
4653 | bfq_choose_bfqq_for_injection(struct bfq_data *bfqd) | |
d0edc247 | 4654 | { |
2341d662 PV |
4655 | struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue; |
4656 | unsigned int limit = in_serv_bfqq->inject_limit; | |
2d31c684 DZ |
4657 | int i; |
4658 | ||
2341d662 PV |
4659 | /* |
4660 | * If | |
4661 | * - bfqq is not weight-raised and therefore does not carry | |
4662 | * time-critical I/O, | |
4663 | * or | |
4664 | * - regardless of whether bfqq is weight-raised, bfqq has | |
4665 | * however a long think time, during which it can absorb the | |
4666 | * effect of an appropriate number of extra I/O requests | |
4667 | * from other queues (see bfq_update_inject_limit for | |
4668 | * details on the computation of this number); | |
4669 | * then injection can be performed without restrictions. | |
4670 | */ | |
4671 | bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 || | |
4672 | !bfq_bfqq_has_short_ttime(in_serv_bfqq); | |
d0edc247 PV |
4673 | |
4674 | /* | |
2341d662 PV |
4675 | * If |
4676 | * - the baseline total service time could not be sampled yet, | |
4677 | * so the inject limit happens to be still 0, and | |
4678 | * - a lot of time has elapsed since the plugging of I/O | |
4679 | * dispatching started, so drive speed is being wasted | |
4680 | * significantly; | |
4681 | * then temporarily raise inject limit to one request. | |
4682 | */ | |
4683 | if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 && | |
4684 | bfq_bfqq_wait_request(in_serv_bfqq) && | |
4685 | time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies + | |
4686 | bfqd->bfq_slice_idle) | |
4687 | ) | |
4688 | limit = 1; | |
4689 | ||
2d31c684 | 4690 | if (bfqd->tot_rq_in_driver >= limit) |
2341d662 PV |
4691 | return NULL; |
4692 | ||
4693 | /* | |
4694 | * Linear search of the source queue for injection; but, with | |
4695 | * a high probability, very few steps are needed to find a | |
4696 | * candidate queue, i.e., a queue with enough budget left for | |
4697 | * its next request. In fact: | |
d0edc247 PV |
4698 | * - BFQ dynamically updates the budget of every queue so as |
4699 | * to accommodate the expected backlog of the queue; | |
4700 | * - if a queue gets all its requests dispatched as injected | |
4701 | * service, then the queue is removed from the active list | |
2341d662 PV |
4702 | * (and re-added only if it gets new requests, but then it |
4703 | * is assigned again enough budget for its new backlog). | |
d0edc247 | 4704 | */ |
2d31c684 DZ |
4705 | for (i = 0; i < bfqd->num_actuators; i++) { |
4706 | list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list) | |
4707 | if (!RB_EMPTY_ROOT(&bfqq->sort_list) && | |
4708 | (in_serv_always_inject || bfqq->wr_coeff > 1) && | |
4709 | bfq_serv_to_charge(bfqq->next_rq, bfqq) <= | |
4710 | bfq_bfqq_budget_left(bfqq)) { | |
2341d662 PV |
4711 | /* |
4712 | * Allow for only one large in-flight request | |
4713 | * on non-rotational devices, for the | |
4714 | * following reason. On non-rotationl drives, | |
4715 | * large requests take much longer than | |
4716 | * smaller requests to be served. In addition, | |
4717 | * the drive prefers to serve large requests | |
4718 | * w.r.t. to small ones, if it can choose. So, | |
4719 | * having more than one large requests queued | |
4720 | * in the drive may easily make the next first | |
4721 | * request of the in-service queue wait for so | |
4722 | * long to break bfqq's service guarantees. On | |
4723 | * the bright side, large requests let the | |
4724 | * drive reach a very high throughput, even if | |
4725 | * there is only one in-flight large request | |
4726 | * at a time. | |
4727 | */ | |
4728 | if (blk_queue_nonrot(bfqd->queue) && | |
4729 | blk_rq_sectors(bfqq->next_rq) >= | |
0c3e09e8 KS |
4730 | BFQQ_SECT_THR_NONROT && |
4731 | bfqd->tot_rq_in_driver >= 1) | |
4732 | continue; | |
4733 | else { | |
2341d662 PV |
4734 | bfqd->rqs_injected = true; |
4735 | return bfqq; | |
4736 | } | |
4737 | } | |
2d31c684 | 4738 | } |
d0edc247 PV |
4739 | |
4740 | return NULL; | |
4741 | } | |
4742 | ||
2d31c684 DZ |
4743 | static struct bfq_queue * |
4744 | bfq_find_active_bfqq_for_actuator(struct bfq_data *bfqd, int idx) | |
4745 | { | |
4746 | struct bfq_queue *bfqq; | |
4747 | ||
4748 | if (bfqd->in_service_queue && | |
4749 | bfqd->in_service_queue->actuator_idx == idx) | |
4750 | return bfqd->in_service_queue; | |
4751 | ||
4752 | list_for_each_entry(bfqq, &bfqd->active_list[idx], bfqq_list) { | |
4753 | if (!RB_EMPTY_ROOT(&bfqq->sort_list) && | |
4754 | bfq_serv_to_charge(bfqq->next_rq, bfqq) <= | |
4755 | bfq_bfqq_budget_left(bfqq)) { | |
4756 | return bfqq; | |
4757 | } | |
4758 | } | |
d0edc247 PV |
4759 | |
4760 | return NULL; | |
4761 | } | |
4762 | ||
2d31c684 DZ |
4763 | /* |
4764 | * Perform a linear scan of each actuator, until an actuator is found | |
1bd43e19 DZ |
4765 | * for which the following three conditions hold: the load of the |
4766 | * actuator is below the threshold (see comments on | |
4767 | * actuator_load_threshold for details) and lower than that of the | |
4768 | * next actuator (comments on this extra condition below), and there | |
4769 | * is a queue that contains I/O for that actuator. On success, return | |
4770 | * that queue. | |
4771 | * | |
4772 | * Performing a plain linear scan entails a prioritization among | |
4773 | * actuators. The extra condition above breaks this prioritization and | |
4774 | * tends to distribute injection uniformly across actuators. | |
2d31c684 DZ |
4775 | */ |
4776 | static struct bfq_queue * | |
4777 | bfq_find_bfqq_for_underused_actuator(struct bfq_data *bfqd) | |
4778 | { | |
4779 | int i; | |
4780 | ||
4781 | for (i = 0 ; i < bfqd->num_actuators; i++) { | |
1bd43e19 DZ |
4782 | if (bfqd->rq_in_driver[i] < bfqd->actuator_load_threshold && |
4783 | (i == bfqd->num_actuators - 1 || | |
4784 | bfqd->rq_in_driver[i] < bfqd->rq_in_driver[i+1])) { | |
2d31c684 DZ |
4785 | struct bfq_queue *bfqq = |
4786 | bfq_find_active_bfqq_for_actuator(bfqd, i); | |
4787 | ||
4788 | if (bfqq) | |
4789 | return bfqq; | |
4790 | } | |
4791 | } | |
4792 | ||
4793 | return NULL; | |
4794 | } | |
4795 | ||
4796 | ||
aee69d78 PV |
4797 | /* |
4798 | * Select a queue for service. If we have a current queue in service, | |
4799 | * check whether to continue servicing it, or retrieve and set a new one. | |
4800 | */ | |
4801 | static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) | |
4802 | { | |
2d31c684 | 4803 | struct bfq_queue *bfqq, *inject_bfqq; |
aee69d78 PV |
4804 | struct request *next_rq; |
4805 | enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT; | |
4806 | ||
4807 | bfqq = bfqd->in_service_queue; | |
4808 | if (!bfqq) | |
4809 | goto new_queue; | |
4810 | ||
4811 | bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); | |
4812 | ||
4420b095 PV |
4813 | /* |
4814 | * Do not expire bfqq for budget timeout if bfqq may be about | |
4815 | * to enjoy device idling. The reason why, in this case, we | |
4816 | * prevent bfqq from expiring is the same as in the comments | |
4817 | * on the case where bfq_bfqq_must_idle() returns true, in | |
4818 | * bfq_completed_request(). | |
4819 | */ | |
aee69d78 | 4820 | if (bfq_may_expire_for_budg_timeout(bfqq) && |
aee69d78 PV |
4821 | !bfq_bfqq_must_idle(bfqq)) |
4822 | goto expire; | |
4823 | ||
4824 | check_queue: | |
2d31c684 DZ |
4825 | /* |
4826 | * If some actuator is underutilized, but the in-service | |
4827 | * queue does not contain I/O for that actuator, then try to | |
4828 | * inject I/O for that actuator. | |
4829 | */ | |
4830 | inject_bfqq = bfq_find_bfqq_for_underused_actuator(bfqd); | |
4831 | if (inject_bfqq && inject_bfqq != bfqq) | |
4832 | return inject_bfqq; | |
4833 | ||
aee69d78 PV |
4834 | /* |
4835 | * This loop is rarely executed more than once. Even when it | |
4836 | * happens, it is much more convenient to re-execute this loop | |
4837 | * than to return NULL and trigger a new dispatch to get a | |
4838 | * request served. | |
4839 | */ | |
4840 | next_rq = bfqq->next_rq; | |
4841 | /* | |
4842 | * If bfqq has requests queued and it has enough budget left to | |
4843 | * serve them, keep the queue, otherwise expire it. | |
4844 | */ | |
4845 | if (next_rq) { | |
4846 | if (bfq_serv_to_charge(next_rq, bfqq) > | |
4847 | bfq_bfqq_budget_left(bfqq)) { | |
4848 | /* | |
4849 | * Expire the queue for budget exhaustion, | |
4850 | * which makes sure that the next budget is | |
4851 | * enough to serve the next request, even if | |
4852 | * it comes from the fifo expired path. | |
4853 | */ | |
4854 | reason = BFQQE_BUDGET_EXHAUSTED; | |
4855 | goto expire; | |
4856 | } else { | |
4857 | /* | |
4858 | * The idle timer may be pending because we may | |
4859 | * not disable disk idling even when a new request | |
4860 | * arrives. | |
4861 | */ | |
4862 | if (bfq_bfqq_wait_request(bfqq)) { | |
4863 | /* | |
4864 | * If we get here: 1) at least a new request | |
4865 | * has arrived but we have not disabled the | |
4866 | * timer because the request was too small, | |
4867 | * 2) then the block layer has unplugged | |
4868 | * the device, causing the dispatch to be | |
4869 | * invoked. | |
4870 | * | |
4871 | * Since the device is unplugged, now the | |
4872 | * requests are probably large enough to | |
4873 | * provide a reasonable throughput. | |
4874 | * So we disable idling. | |
4875 | */ | |
4876 | bfq_clear_bfqq_wait_request(bfqq); | |
4877 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
4878 | } | |
4879 | goto keep_queue; | |
4880 | } | |
4881 | } | |
4882 | ||
4883 | /* | |
4884 | * No requests pending. However, if the in-service queue is idling | |
4885 | * for a new request, or has requests waiting for a completion and | |
4886 | * may idle after their completion, then keep it anyway. | |
d0edc247 | 4887 | * |
2341d662 PV |
4888 | * Yet, inject service from other queues if it boosts |
4889 | * throughput and is possible. | |
aee69d78 PV |
4890 | */ |
4891 | if (bfq_bfqq_wait_request(bfqq) || | |
277a4a9b | 4892 | (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) { |
9778369a PV |
4893 | unsigned int act_idx = bfqq->actuator_idx; |
4894 | struct bfq_queue *async_bfqq = NULL; | |
2ec5a5c4 PV |
4895 | struct bfq_queue *blocked_bfqq = |
4896 | !hlist_empty(&bfqq->woken_list) ? | |
4897 | container_of(bfqq->woken_list.first, | |
4898 | struct bfq_queue, | |
4899 | woken_list_node) | |
4900 | : NULL; | |
2341d662 | 4901 | |
9778369a PV |
4902 | if (bfqq->bic && bfqq->bic->bfqq[0][act_idx] && |
4903 | bfq_bfqq_busy(bfqq->bic->bfqq[0][act_idx]) && | |
4904 | bfqq->bic->bfqq[0][act_idx]->next_rq) | |
4905 | async_bfqq = bfqq->bic->bfqq[0][act_idx]; | |
2341d662 | 4906 | /* |
2ec5a5c4 | 4907 | * The next four mutually-exclusive ifs decide |
13a857a4 PV |
4908 | * whether to try injection, and choose the queue to |
4909 | * pick an I/O request from. | |
4910 | * | |
4911 | * The first if checks whether the process associated | |
4912 | * with bfqq has also async I/O pending. If so, it | |
4913 | * injects such I/O unconditionally. Injecting async | |
4914 | * I/O from the same process can cause no harm to the | |
4915 | * process. On the contrary, it can only increase | |
4916 | * bandwidth and reduce latency for the process. | |
4917 | * | |
4918 | * The second if checks whether there happens to be a | |
4919 | * non-empty waker queue for bfqq, i.e., a queue whose | |
4920 | * I/O needs to be completed for bfqq to receive new | |
4921 | * I/O. This happens, e.g., if bfqq is associated with | |
4922 | * a process that does some sync. A sync generates | |
4923 | * extra blocking I/O, which must be completed before | |
4924 | * the process associated with bfqq can go on with its | |
4925 | * I/O. If the I/O of the waker queue is not served, | |
4926 | * then bfqq remains empty, and no I/O is dispatched, | |
4927 | * until the idle timeout fires for bfqq. This is | |
4928 | * likely to result in lower bandwidth and higher | |
4929 | * latencies for bfqq, and in a severe loss of total | |
4930 | * throughput. The best action to take is therefore to | |
4931 | * serve the waker queue as soon as possible. So do it | |
4932 | * (without relying on the third alternative below for | |
4933 | * eventually serving waker_bfqq's I/O; see the last | |
4934 | * paragraph for further details). This systematic | |
4935 | * injection of I/O from the waker queue does not | |
4936 | * cause any delay to bfqq's I/O. On the contrary, | |
4937 | * next bfqq's I/O is brought forward dramatically, | |
4938 | * for it is not blocked for milliseconds. | |
4939 | * | |
2ec5a5c4 PV |
4940 | * The third if checks whether there is a queue woken |
4941 | * by bfqq, and currently with pending I/O. Such a | |
4942 | * woken queue does not steal bandwidth from bfqq, | |
4943 | * because it remains soon without I/O if bfqq is not | |
4944 | * served. So there is virtually no risk of loss of | |
4945 | * bandwidth for bfqq if this woken queue has I/O | |
4946 | * dispatched while bfqq is waiting for new I/O. | |
4947 | * | |
4948 | * The fourth if checks whether bfqq is a queue for | |
13a857a4 PV |
4949 | * which it is better to avoid injection. It is so if |
4950 | * bfqq delivers more throughput when served without | |
4951 | * any further I/O from other queues in the middle, or | |
4952 | * if the service times of bfqq's I/O requests both | |
4953 | * count more than overall throughput, and may be | |
4954 | * easily increased by injection (this happens if bfqq | |
4955 | * has a short think time). If none of these | |
4956 | * conditions holds, then a candidate queue for | |
4957 | * injection is looked for through | |
4958 | * bfq_choose_bfqq_for_injection(). Note that the | |
4959 | * latter may return NULL (for example if the inject | |
4960 | * limit for bfqq is currently 0). | |
4961 | * | |
4962 | * NOTE: motivation for the second alternative | |
4963 | * | |
4964 | * Thanks to the way the inject limit is updated in | |
4965 | * bfq_update_has_short_ttime(), it is rather likely | |
4966 | * that, if I/O is being plugged for bfqq and the | |
4967 | * waker queue has pending I/O requests that are | |
2ec5a5c4 | 4968 | * blocking bfqq's I/O, then the fourth alternative |
13a857a4 PV |
4969 | * above lets the waker queue get served before the |
4970 | * I/O-plugging timeout fires. So one may deem the | |
4971 | * second alternative superfluous. It is not, because | |
2ec5a5c4 | 4972 | * the fourth alternative may be way less effective in |
13a857a4 PV |
4973 | * case of a synchronization. For two main |
4974 | * reasons. First, throughput may be low because the | |
4975 | * inject limit may be too low to guarantee the same | |
4976 | * amount of injected I/O, from the waker queue or | |
4977 | * other queues, that the second alternative | |
4978 | * guarantees (the second alternative unconditionally | |
4979 | * injects a pending I/O request of the waker queue | |
4980 | * for each bfq_dispatch_request()). Second, with the | |
2ec5a5c4 | 4981 | * fourth alternative, the duration of the plugging, |
13a857a4 PV |
4982 | * i.e., the time before bfqq finally receives new I/O, |
4983 | * may not be minimized, because the waker queue may | |
4984 | * happen to be served only after other queues. | |
2341d662 PV |
4985 | */ |
4986 | if (async_bfqq && | |
4987 | icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic && | |
4988 | bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <= | |
4989 | bfq_bfqq_budget_left(async_bfqq)) | |
86f8382e | 4990 | bfqq = async_bfqq; |
71217df3 | 4991 | else if (bfqq->waker_bfqq && |
13a857a4 | 4992 | bfq_bfqq_busy(bfqq->waker_bfqq) && |
d4fc3640 | 4993 | bfqq->waker_bfqq->next_rq && |
13a857a4 PV |
4994 | bfq_serv_to_charge(bfqq->waker_bfqq->next_rq, |
4995 | bfqq->waker_bfqq) <= | |
4996 | bfq_bfqq_budget_left(bfqq->waker_bfqq) | |
4997 | ) | |
4998 | bfqq = bfqq->waker_bfqq; | |
2ec5a5c4 PV |
4999 | else if (blocked_bfqq && |
5000 | bfq_bfqq_busy(blocked_bfqq) && | |
5001 | blocked_bfqq->next_rq && | |
5002 | bfq_serv_to_charge(blocked_bfqq->next_rq, | |
5003 | blocked_bfqq) <= | |
5004 | bfq_bfqq_budget_left(blocked_bfqq) | |
5005 | ) | |
5006 | bfqq = blocked_bfqq; | |
2341d662 PV |
5007 | else if (!idling_boosts_thr_without_issues(bfqd, bfqq) && |
5008 | (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 || | |
5009 | !bfq_bfqq_has_short_ttime(bfqq))) | |
d0edc247 PV |
5010 | bfqq = bfq_choose_bfqq_for_injection(bfqd); |
5011 | else | |
5012 | bfqq = NULL; | |
5013 | ||
aee69d78 PV |
5014 | goto keep_queue; |
5015 | } | |
5016 | ||
5017 | reason = BFQQE_NO_MORE_REQUESTS; | |
5018 | expire: | |
5019 | bfq_bfqq_expire(bfqd, bfqq, false, reason); | |
5020 | new_queue: | |
5021 | bfqq = bfq_set_in_service_queue(bfqd); | |
5022 | if (bfqq) { | |
5023 | bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue"); | |
5024 | goto check_queue; | |
5025 | } | |
5026 | keep_queue: | |
5027 | if (bfqq) | |
5028 | bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue"); | |
5029 | else | |
5030 | bfq_log(bfqd, "select_queue: no queue returned"); | |
5031 | ||
5032 | return bfqq; | |
5033 | } | |
5034 | ||
44e44a1b PV |
5035 | static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5036 | { | |
5037 | struct bfq_entity *entity = &bfqq->entity; | |
5038 | ||
5039 | if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */ | |
5040 | bfq_log_bfqq(bfqd, bfqq, | |
5041 | "raising period dur %u/%u msec, old coeff %u, w %d(%d)", | |
5042 | jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish), | |
5043 | jiffies_to_msecs(bfqq->wr_cur_max_time), | |
5044 | bfqq->wr_coeff, | |
5045 | bfqq->entity.weight, bfqq->entity.orig_weight); | |
5046 | ||
5047 | if (entity->prio_changed) | |
5048 | bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change"); | |
5049 | ||
5050 | /* | |
e1b2324d AA |
5051 | * If the queue was activated in a burst, or too much |
5052 | * time has elapsed from the beginning of this | |
5053 | * weight-raising period, then end weight raising. | |
44e44a1b | 5054 | */ |
e1b2324d AA |
5055 | if (bfq_bfqq_in_large_burst(bfqq)) |
5056 | bfq_bfqq_end_wr(bfqq); | |
5057 | else if (time_is_before_jiffies(bfqq->last_wr_start_finish + | |
5058 | bfqq->wr_cur_max_time)) { | |
77b7dcea PV |
5059 | if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time || |
5060 | time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt + | |
3c337690 PV |
5061 | bfq_wr_duration(bfqd))) { |
5062 | /* | |
5063 | * Either in interactive weight | |
5064 | * raising, or in soft_rt weight | |
5065 | * raising with the | |
5066 | * interactive-weight-raising period | |
5067 | * elapsed (so no switch back to | |
5068 | * interactive weight raising). | |
5069 | */ | |
77b7dcea | 5070 | bfq_bfqq_end_wr(bfqq); |
3c337690 PV |
5071 | } else { /* |
5072 | * soft_rt finishing while still in | |
5073 | * interactive period, switch back to | |
5074 | * interactive weight raising | |
5075 | */ | |
3e2bdd6d | 5076 | switch_back_to_interactive_wr(bfqq, bfqd); |
77b7dcea PV |
5077 | bfqq->entity.prio_changed = 1; |
5078 | } | |
44e44a1b | 5079 | } |
8a8747dc PV |
5080 | if (bfqq->wr_coeff > 1 && |
5081 | bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time && | |
5082 | bfqq->service_from_wr > max_service_from_wr) { | |
5083 | /* see comments on max_service_from_wr */ | |
5084 | bfq_bfqq_end_wr(bfqq); | |
5085 | } | |
44e44a1b | 5086 | } |
431b17f9 PV |
5087 | /* |
5088 | * To improve latency (for this or other queues), immediately | |
5089 | * update weight both if it must be raised and if it must be | |
5090 | * lowered. Since, entity may be on some active tree here, and | |
5091 | * might have a pending change of its ioprio class, invoke | |
5092 | * next function with the last parameter unset (see the | |
5093 | * comments on the function). | |
5094 | */ | |
44e44a1b | 5095 | if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1)) |
431b17f9 PV |
5096 | __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity), |
5097 | entity, false); | |
44e44a1b PV |
5098 | } |
5099 | ||
aee69d78 PV |
5100 | /* |
5101 | * Dispatch next request from bfqq. | |
5102 | */ | |
5103 | static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd, | |
5104 | struct bfq_queue *bfqq) | |
5105 | { | |
5106 | struct request *rq = bfqq->next_rq; | |
5107 | unsigned long service_to_charge; | |
5108 | ||
5109 | service_to_charge = bfq_serv_to_charge(rq, bfqq); | |
5110 | ||
5111 | bfq_bfqq_served(bfqq, service_to_charge); | |
5112 | ||
2341d662 PV |
5113 | if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) { |
5114 | bfqd->wait_dispatch = false; | |
5115 | bfqd->waited_rq = rq; | |
5116 | } | |
aee69d78 | 5117 | |
2341d662 | 5118 | bfq_dispatch_remove(bfqd->queue, rq); |
d0edc247 | 5119 | |
2341d662 | 5120 | if (bfqq != bfqd->in_service_queue) |
87c971de | 5121 | return rq; |
d0edc247 | 5122 | |
44e44a1b PV |
5123 | /* |
5124 | * If weight raising has to terminate for bfqq, then next | |
5125 | * function causes an immediate update of bfqq's weight, | |
5126 | * without waiting for next activation. As a consequence, on | |
5127 | * expiration, bfqq will be timestamped as if has never been | |
5128 | * weight-raised during this service slot, even if it has | |
5129 | * received part or even most of the service as a | |
5130 | * weight-raised queue. This inflates bfqq's timestamps, which | |
5131 | * is beneficial, as bfqq is then more willing to leave the | |
5132 | * device immediately to possible other weight-raised queues. | |
5133 | */ | |
5134 | bfq_update_wr_data(bfqd, bfqq); | |
5135 | ||
aee69d78 PV |
5136 | /* |
5137 | * Expire bfqq, pretending that its budget expired, if bfqq | |
5138 | * belongs to CLASS_IDLE and other queues are waiting for | |
5139 | * service. | |
5140 | */ | |
87c971de KS |
5141 | if (bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)) |
5142 | bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED); | |
d0edc247 | 5143 | |
aee69d78 PV |
5144 | return rq; |
5145 | } | |
5146 | ||
5147 | static bool bfq_has_work(struct blk_mq_hw_ctx *hctx) | |
5148 | { | |
5149 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5150 | ||
5151 | /* | |
ddc25c86 | 5152 | * Avoiding lock: a race on bfqd->queued should cause at |
aee69d78 PV |
5153 | * most a call to dispatch for nothing |
5154 | */ | |
5155 | return !list_empty_careful(&bfqd->dispatch) || | |
ddc25c86 | 5156 | READ_ONCE(bfqd->queued); |
aee69d78 PV |
5157 | } |
5158 | ||
5159 | static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) | |
5160 | { | |
5161 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5162 | struct request *rq = NULL; | |
5163 | struct bfq_queue *bfqq = NULL; | |
5164 | ||
5165 | if (!list_empty(&bfqd->dispatch)) { | |
5166 | rq = list_first_entry(&bfqd->dispatch, struct request, | |
5167 | queuelist); | |
5168 | list_del_init(&rq->queuelist); | |
5169 | ||
5170 | bfqq = RQ_BFQQ(rq); | |
5171 | ||
5172 | if (bfqq) { | |
5173 | /* | |
5174 | * Increment counters here, because this | |
5175 | * dispatch does not follow the standard | |
5176 | * dispatch flow (where counters are | |
5177 | * incremented) | |
5178 | */ | |
5179 | bfqq->dispatched++; | |
5180 | ||
5181 | goto inc_in_driver_start_rq; | |
5182 | } | |
5183 | ||
5184 | /* | |
a7877390 | 5185 | * We exploit the bfq_finish_requeue_request hook to |
2d31c684 | 5186 | * decrement tot_rq_in_driver, but |
a7877390 PV |
5187 | * bfq_finish_requeue_request will not be invoked on |
5188 | * this request. So, to avoid unbalance, just start | |
2d31c684 DZ |
5189 | * this request, without incrementing tot_rq_in_driver. As |
5190 | * a negative consequence, tot_rq_in_driver is deceptively | |
a7877390 PV |
5191 | * lower than it should be while this request is in |
5192 | * service. This may cause bfq_schedule_dispatch to be | |
5193 | * invoked uselessly. | |
aee69d78 PV |
5194 | * |
5195 | * As for implementing an exact solution, the | |
a7877390 PV |
5196 | * bfq_finish_requeue_request hook, if defined, is |
5197 | * probably invoked also on this request. So, by | |
5198 | * exploiting this hook, we could 1) increment | |
2d31c684 | 5199 | * tot_rq_in_driver here, and 2) decrement it in |
a7877390 PV |
5200 | * bfq_finish_requeue_request. Such a solution would |
5201 | * let the value of the counter be always accurate, | |
5202 | * but it would entail using an extra interface | |
5203 | * function. This cost seems higher than the benefit, | |
5204 | * being the frequency of non-elevator-private | |
aee69d78 PV |
5205 | * requests very low. |
5206 | */ | |
5207 | goto start_rq; | |
5208 | } | |
5209 | ||
73d58118 PV |
5210 | bfq_log(bfqd, "dispatch requests: %d busy queues", |
5211 | bfq_tot_busy_queues(bfqd)); | |
aee69d78 | 5212 | |
73d58118 | 5213 | if (bfq_tot_busy_queues(bfqd) == 0) |
aee69d78 PV |
5214 | goto exit; |
5215 | ||
5216 | /* | |
5217 | * Force device to serve one request at a time if | |
5218 | * strict_guarantees is true. Forcing this service scheme is | |
5219 | * currently the ONLY way to guarantee that the request | |
5220 | * service order enforced by the scheduler is respected by a | |
5221 | * queueing device. Otherwise the device is free even to make | |
5222 | * some unlucky request wait for as long as the device | |
5223 | * wishes. | |
5224 | * | |
f06678af | 5225 | * Of course, serving one request at a time may cause loss of |
aee69d78 PV |
5226 | * throughput. |
5227 | */ | |
2d31c684 | 5228 | if (bfqd->strict_guarantees && bfqd->tot_rq_in_driver > 0) |
aee69d78 PV |
5229 | goto exit; |
5230 | ||
5231 | bfqq = bfq_select_queue(bfqd); | |
5232 | if (!bfqq) | |
5233 | goto exit; | |
5234 | ||
5235 | rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq); | |
5236 | ||
5237 | if (rq) { | |
5238 | inc_in_driver_start_rq: | |
2d31c684 DZ |
5239 | bfqd->rq_in_driver[bfqq->actuator_idx]++; |
5240 | bfqd->tot_rq_in_driver++; | |
aee69d78 PV |
5241 | start_rq: |
5242 | rq->rq_flags |= RQF_STARTED; | |
5243 | } | |
5244 | exit: | |
5245 | return rq; | |
5246 | } | |
5247 | ||
8060c47b | 5248 | #ifdef CONFIG_BFQ_CGROUP_DEBUG |
9b25bd03 PV |
5249 | static void bfq_update_dispatch_stats(struct request_queue *q, |
5250 | struct request *rq, | |
5251 | struct bfq_queue *in_serv_queue, | |
5252 | bool idle_timer_disabled) | |
5253 | { | |
5254 | struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL; | |
aee69d78 | 5255 | |
24bfd19b | 5256 | if (!idle_timer_disabled && !bfqq) |
9b25bd03 | 5257 | return; |
24bfd19b PV |
5258 | |
5259 | /* | |
5260 | * rq and bfqq are guaranteed to exist until this function | |
5261 | * ends, for the following reasons. First, rq can be | |
5262 | * dispatched to the device, and then can be completed and | |
5263 | * freed, only after this function ends. Second, rq cannot be | |
5264 | * merged (and thus freed because of a merge) any longer, | |
5265 | * because it has already started. Thus rq cannot be freed | |
5266 | * before this function ends, and, since rq has a reference to | |
5267 | * bfqq, the same guarantee holds for bfqq too. | |
5268 | * | |
5269 | * In addition, the following queue lock guarantees that | |
5270 | * bfqq_group(bfqq) exists as well. | |
5271 | */ | |
0d945c1f | 5272 | spin_lock_irq(&q->queue_lock); |
24bfd19b PV |
5273 | if (idle_timer_disabled) |
5274 | /* | |
5275 | * Since the idle timer has been disabled, | |
5276 | * in_serv_queue contained some request when | |
5277 | * __bfq_dispatch_request was invoked above, which | |
5278 | * implies that rq was picked exactly from | |
5279 | * in_serv_queue. Thus in_serv_queue == bfqq, and is | |
5280 | * therefore guaranteed to exist because of the above | |
5281 | * arguments. | |
5282 | */ | |
5283 | bfqg_stats_update_idle_time(bfqq_group(in_serv_queue)); | |
5284 | if (bfqq) { | |
5285 | struct bfq_group *bfqg = bfqq_group(bfqq); | |
5286 | ||
5287 | bfqg_stats_update_avg_queue_size(bfqg); | |
5288 | bfqg_stats_set_start_empty_time(bfqg); | |
5289 | bfqg_stats_update_io_remove(bfqg, rq->cmd_flags); | |
5290 | } | |
0d945c1f | 5291 | spin_unlock_irq(&q->queue_lock); |
9b25bd03 PV |
5292 | } |
5293 | #else | |
5294 | static inline void bfq_update_dispatch_stats(struct request_queue *q, | |
5295 | struct request *rq, | |
5296 | struct bfq_queue *in_serv_queue, | |
5297 | bool idle_timer_disabled) {} | |
8060c47b | 5298 | #endif /* CONFIG_BFQ_CGROUP_DEBUG */ |
24bfd19b | 5299 | |
9b25bd03 PV |
5300 | static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) |
5301 | { | |
5302 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
5303 | struct request *rq; | |
5304 | struct bfq_queue *in_serv_queue; | |
ab552fcb | 5305 | bool waiting_rq, idle_timer_disabled = false; |
9b25bd03 PV |
5306 | |
5307 | spin_lock_irq(&bfqd->lock); | |
5308 | ||
5309 | in_serv_queue = bfqd->in_service_queue; | |
5310 | waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue); | |
5311 | ||
5312 | rq = __bfq_dispatch_request(hctx); | |
ab552fcb ZW |
5313 | if (in_serv_queue == bfqd->in_service_queue) { |
5314 | idle_timer_disabled = | |
5315 | waiting_rq && !bfq_bfqq_wait_request(in_serv_queue); | |
5316 | } | |
9b25bd03 PV |
5317 | |
5318 | spin_unlock_irq(&bfqd->lock); | |
ab552fcb ZW |
5319 | bfq_update_dispatch_stats(hctx->queue, rq, |
5320 | idle_timer_disabled ? in_serv_queue : NULL, | |
5321 | idle_timer_disabled); | |
9b25bd03 | 5322 | |
aee69d78 PV |
5323 | return rq; |
5324 | } | |
5325 | ||
5326 | /* | |
5327 | * Task holds one reference to the queue, dropped when task exits. Each rq | |
5328 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
5329 | * | |
5330 | * Scheduler lock must be held here. Recall not to use bfqq after calling | |
5331 | * this function on it. | |
5332 | */ | |
ea25da48 | 5333 | void bfq_put_queue(struct bfq_queue *bfqq) |
aee69d78 | 5334 | { |
3f758e84 PV |
5335 | struct bfq_queue *item; |
5336 | struct hlist_node *n; | |
e21b7a0b | 5337 | struct bfq_group *bfqg = bfqq_group(bfqq); |
e21b7a0b | 5338 | |
1e3cc212 | 5339 | bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref); |
aee69d78 PV |
5340 | |
5341 | bfqq->ref--; | |
5342 | if (bfqq->ref) | |
5343 | return; | |
5344 | ||
99fead8d | 5345 | if (!hlist_unhashed(&bfqq->burst_list_node)) { |
e1b2324d | 5346 | hlist_del_init(&bfqq->burst_list_node); |
99fead8d PV |
5347 | /* |
5348 | * Decrement also burst size after the removal, if the | |
5349 | * process associated with bfqq is exiting, and thus | |
5350 | * does not contribute to the burst any longer. This | |
5351 | * decrement helps filter out false positives of large | |
5352 | * bursts, when some short-lived process (often due to | |
5353 | * the execution of commands by some service) happens | |
5354 | * to start and exit while a complex application is | |
5355 | * starting, and thus spawning several processes that | |
5356 | * do I/O (and that *must not* be treated as a large | |
5357 | * burst, see comments on bfq_handle_burst). | |
5358 | * | |
5359 | * In particular, the decrement is performed only if: | |
5360 | * 1) bfqq is not a merged queue, because, if it is, | |
5361 | * then this free of bfqq is not triggered by the exit | |
5362 | * of the process bfqq is associated with, but exactly | |
5363 | * by the fact that bfqq has just been merged. | |
5364 | * 2) burst_size is greater than 0, to handle | |
5365 | * unbalanced decrements. Unbalanced decrements may | |
5366 | * happen in te following case: bfqq is inserted into | |
5367 | * the current burst list--without incrementing | |
5368 | * bust_size--because of a split, but the current | |
5369 | * burst list is not the burst list bfqq belonged to | |
5370 | * (see comments on the case of a split in | |
5371 | * bfq_set_request). | |
5372 | */ | |
5373 | if (bfqq->bic && bfqq->bfqd->burst_size > 0) | |
5374 | bfqq->bfqd->burst_size--; | |
7cb04004 | 5375 | } |
e21b7a0b | 5376 | |
3f758e84 PV |
5377 | /* |
5378 | * bfqq does not exist any longer, so it cannot be woken by | |
5379 | * any other queue, and cannot wake any other queue. Then bfqq | |
5380 | * must be removed from the woken list of its possible waker | |
5381 | * queue, and all queues in the woken list of bfqq must stop | |
5382 | * having a waker queue. Strictly speaking, these updates | |
5383 | * should be performed when bfqq remains with no I/O source | |
5384 | * attached to it, which happens before bfqq gets freed. In | |
5385 | * particular, this happens when the last process associated | |
5386 | * with bfqq exits or gets associated with a different | |
5387 | * queue. However, both events lead to bfqq being freed soon, | |
5388 | * and dangling references would come out only after bfqq gets | |
5389 | * freed. So these updates are done here, as a simple and safe | |
5390 | * way to handle all cases. | |
5391 | */ | |
5392 | /* remove bfqq from woken list */ | |
5393 | if (!hlist_unhashed(&bfqq->woken_list_node)) | |
5394 | hlist_del_init(&bfqq->woken_list_node); | |
5395 | ||
5396 | /* reset waker for all queues in woken list */ | |
5397 | hlist_for_each_entry_safe(item, n, &bfqq->woken_list, | |
5398 | woken_list_node) { | |
5399 | item->waker_bfqq = NULL; | |
3f758e84 PV |
5400 | hlist_del_init(&item->woken_list_node); |
5401 | } | |
5402 | ||
1e3cc212 | 5403 | if (bfqq->bfqd->last_completed_rq_bfqq == bfqq) |
08d383a7 PV |
5404 | bfqq->bfqd->last_completed_rq_bfqq = NULL; |
5405 | ||
3e49c1e4 BVA |
5406 | WARN_ON_ONCE(!list_empty(&bfqq->fifo)); |
5407 | WARN_ON_ONCE(!RB_EMPTY_ROOT(&bfqq->sort_list)); | |
f80dd11d | 5408 | WARN_ON_ONCE(bfqq->dispatched); |
3e49c1e4 | 5409 | |
aee69d78 | 5410 | kmem_cache_free(bfq_pool, bfqq); |
8f9bebc3 | 5411 | bfqg_and_blkg_put(bfqg); |
aee69d78 PV |
5412 | } |
5413 | ||
430a67f9 PV |
5414 | static void bfq_put_stable_ref(struct bfq_queue *bfqq) |
5415 | { | |
5416 | bfqq->stable_ref--; | |
5417 | bfq_put_queue(bfqq); | |
5418 | } | |
5419 | ||
3bc5e683 | 5420 | void bfq_put_cooperator(struct bfq_queue *bfqq) |
36eca894 AA |
5421 | { |
5422 | struct bfq_queue *__bfqq, *next; | |
5423 | ||
5424 | /* | |
5425 | * If this queue was scheduled to merge with another queue, be | |
5426 | * sure to drop the reference taken on that queue (and others in | |
5427 | * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. | |
5428 | */ | |
5429 | __bfqq = bfqq->new_bfqq; | |
5430 | while (__bfqq) { | |
36eca894 AA |
5431 | next = __bfqq->new_bfqq; |
5432 | bfq_put_queue(__bfqq); | |
5433 | __bfqq = next; | |
5434 | } | |
5435 | } | |
5436 | ||
aee69d78 PV |
5437 | static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5438 | { | |
5439 | if (bfqq == bfqd->in_service_queue) { | |
3726112e | 5440 | __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT); |
aee69d78 PV |
5441 | bfq_schedule_dispatch(bfqd); |
5442 | } | |
5443 | ||
5444 | bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref); | |
5445 | ||
36eca894 AA |
5446 | bfq_put_cooperator(bfqq); |
5447 | ||
478de338 | 5448 | bfq_release_process_ref(bfqd, bfqq); |
aee69d78 PV |
5449 | } |
5450 | ||
9778369a PV |
5451 | static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync, |
5452 | unsigned int actuator_idx) | |
aee69d78 | 5453 | { |
9778369a | 5454 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, actuator_idx); |
aee69d78 PV |
5455 | struct bfq_data *bfqd; |
5456 | ||
5457 | if (bfqq) | |
5458 | bfqd = bfqq->bfqd; /* NULL if scheduler already exited */ | |
5459 | ||
5460 | if (bfqq && bfqd) { | |
9778369a | 5461 | bic_set_bfqq(bic, NULL, is_sync, actuator_idx); |
246cf66e | 5462 | bfq_exit_bfqq(bfqd, bfqq); |
aee69d78 PV |
5463 | } |
5464 | } | |
5465 | ||
5466 | static void bfq_exit_icq(struct io_cq *icq) | |
5467 | { | |
5468 | struct bfq_io_cq *bic = icq_to_bic(icq); | |
9778369a PV |
5469 | struct bfq_data *bfqd = bic_to_bfqd(bic); |
5470 | unsigned long flags; | |
5471 | unsigned int act_idx; | |
5472 | /* | |
5473 | * If bfqd and thus bfqd->num_actuators is not available any | |
5474 | * longer, then cycle over all possible per-actuator bfqqs in | |
5475 | * next loop. We rely on bic being zeroed on creation, and | |
5476 | * therefore on its unused per-actuator fields being NULL. | |
5477 | */ | |
5478 | unsigned int num_actuators = BFQ_MAX_ACTUATORS; | |
fd571df0 | 5479 | struct bfq_iocq_bfqq_data *bfqq_data = bic->bfqq_data; |
aee69d78 | 5480 | |
9778369a PV |
5481 | /* |
5482 | * bfqd is NULL if scheduler already exited, and in that case | |
5483 | * this is the last time these queues are accessed. | |
5484 | */ | |
5485 | if (bfqd) { | |
5486 | spin_lock_irqsave(&bfqd->lock, flags); | |
5487 | num_actuators = bfqd->num_actuators; | |
5488 | } | |
430a67f9 | 5489 | |
9778369a | 5490 | for (act_idx = 0; act_idx < num_actuators; act_idx++) { |
fd571df0 PV |
5491 | if (bfqq_data[act_idx].stable_merge_bfqq) |
5492 | bfq_put_stable_ref(bfqq_data[act_idx].stable_merge_bfqq); | |
430a67f9 | 5493 | |
9778369a PV |
5494 | bfq_exit_icq_bfqq(bic, true, act_idx); |
5495 | bfq_exit_icq_bfqq(bic, false, act_idx); | |
430a67f9 PV |
5496 | } |
5497 | ||
9778369a PV |
5498 | if (bfqd) |
5499 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
aee69d78 PV |
5500 | } |
5501 | ||
5502 | /* | |
5503 | * Update the entity prio values; note that the new values will not | |
5504 | * be used until the next (re)activation. | |
5505 | */ | |
5506 | static void | |
5507 | bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) | |
5508 | { | |
5509 | struct task_struct *tsk = current; | |
5510 | int ioprio_class; | |
5511 | struct bfq_data *bfqd = bfqq->bfqd; | |
5512 | ||
5513 | if (!bfqd) | |
5514 | return; | |
5515 | ||
5516 | ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
5517 | switch (ioprio_class) { | |
5518 | default: | |
d51cfc53 | 5519 | pr_err("bdi %s: bfq: bad prio class %d\n", |
d152c682 | 5520 | bdi_dev_name(bfqq->bfqd->queue->disk->bdi), |
edb0872f | 5521 | ioprio_class); |
df561f66 | 5522 | fallthrough; |
aee69d78 PV |
5523 | case IOPRIO_CLASS_NONE: |
5524 | /* | |
5525 | * No prio set, inherit CPU scheduling settings. | |
5526 | */ | |
5527 | bfqq->new_ioprio = task_nice_ioprio(tsk); | |
5528 | bfqq->new_ioprio_class = task_nice_ioclass(tsk); | |
5529 | break; | |
5530 | case IOPRIO_CLASS_RT: | |
5531 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5532 | bfqq->new_ioprio_class = IOPRIO_CLASS_RT; | |
5533 | break; | |
5534 | case IOPRIO_CLASS_BE: | |
5535 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5536 | bfqq->new_ioprio_class = IOPRIO_CLASS_BE; | |
5537 | break; | |
5538 | case IOPRIO_CLASS_IDLE: | |
5539 | bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE; | |
5540 | bfqq->new_ioprio = 7; | |
aee69d78 PV |
5541 | break; |
5542 | } | |
5543 | ||
202bc942 | 5544 | if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) { |
aee69d78 PV |
5545 | pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n", |
5546 | bfqq->new_ioprio); | |
202bc942 | 5547 | bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1; |
aee69d78 PV |
5548 | } |
5549 | ||
5550 | bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio); | |
3c337690 PV |
5551 | bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d", |
5552 | bfqq->new_ioprio, bfqq->entity.new_weight); | |
aee69d78 PV |
5553 | bfqq->entity.prio_changed = 1; |
5554 | } | |
5555 | ||
ea25da48 PV |
5556 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
5557 | struct bio *bio, bool is_sync, | |
430a67f9 PV |
5558 | struct bfq_io_cq *bic, |
5559 | bool respawn); | |
ea25da48 | 5560 | |
aee69d78 PV |
5561 | static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio) |
5562 | { | |
5563 | struct bfq_data *bfqd = bic_to_bfqd(bic); | |
5564 | struct bfq_queue *bfqq; | |
5565 | int ioprio = bic->icq.ioc->ioprio; | |
5566 | ||
5567 | /* | |
5568 | * This condition may trigger on a newly created bic, be sure to | |
5569 | * drop the lock before returning. | |
5570 | */ | |
5571 | if (unlikely(!bfqd) || likely(bic->ioprio == ioprio)) | |
5572 | return; | |
5573 | ||
5574 | bic->ioprio = ioprio; | |
5575 | ||
9778369a | 5576 | bfqq = bic_to_bfqq(bic, false, bfq_actuator_index(bfqd, bio)); |
aee69d78 | 5577 | if (bfqq) { |
b600de2d YK |
5578 | struct bfq_queue *old_bfqq = bfqq; |
5579 | ||
f6bad159 | 5580 | bfqq = bfq_get_queue(bfqd, bio, false, bic, true); |
9778369a | 5581 | bic_set_bfqq(bic, bfqq, false, bfq_actuator_index(bfqd, bio)); |
b600de2d | 5582 | bfq_release_process_ref(bfqd, old_bfqq); |
aee69d78 PV |
5583 | } |
5584 | ||
9778369a | 5585 | bfqq = bic_to_bfqq(bic, true, bfq_actuator_index(bfqd, bio)); |
aee69d78 PV |
5586 | if (bfqq) |
5587 | bfq_set_next_ioprio_data(bfqq, bic); | |
5588 | } | |
5589 | ||
5590 | static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
9778369a PV |
5591 | struct bfq_io_cq *bic, pid_t pid, int is_sync, |
5592 | unsigned int act_idx) | |
aee69d78 | 5593 | { |
eb2fd80f PV |
5594 | u64 now_ns = ktime_get_ns(); |
5595 | ||
9778369a | 5596 | bfqq->actuator_idx = act_idx; |
aee69d78 PV |
5597 | RB_CLEAR_NODE(&bfqq->entity.rb_node); |
5598 | INIT_LIST_HEAD(&bfqq->fifo); | |
e1b2324d | 5599 | INIT_HLIST_NODE(&bfqq->burst_list_node); |
13a857a4 PV |
5600 | INIT_HLIST_NODE(&bfqq->woken_list_node); |
5601 | INIT_HLIST_HEAD(&bfqq->woken_list); | |
aee69d78 PV |
5602 | |
5603 | bfqq->ref = 0; | |
5604 | bfqq->bfqd = bfqd; | |
5605 | ||
5606 | if (bic) | |
5607 | bfq_set_next_ioprio_data(bfqq, bic); | |
5608 | ||
5609 | if (is_sync) { | |
d5be3fef PV |
5610 | /* |
5611 | * No need to mark as has_short_ttime if in | |
5612 | * idle_class, because no device idling is performed | |
5613 | * for queues in idle class | |
5614 | */ | |
aee69d78 | 5615 | if (!bfq_class_idle(bfqq)) |
d5be3fef PV |
5616 | /* tentatively mark as has_short_ttime */ |
5617 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
aee69d78 | 5618 | bfq_mark_bfqq_sync(bfqq); |
e1b2324d | 5619 | bfq_mark_bfqq_just_created(bfqq); |
aee69d78 PV |
5620 | } else |
5621 | bfq_clear_bfqq_sync(bfqq); | |
5622 | ||
5623 | /* set end request to minus infinity from now */ | |
eb2fd80f PV |
5624 | bfqq->ttime.last_end_request = now_ns + 1; |
5625 | ||
430a67f9 PV |
5626 | bfqq->creation_time = jiffies; |
5627 | ||
eb2fd80f | 5628 | bfqq->io_start_time = now_ns; |
aee69d78 PV |
5629 | |
5630 | bfq_mark_bfqq_IO_bound(bfqq); | |
5631 | ||
5632 | bfqq->pid = pid; | |
5633 | ||
5634 | /* Tentative initial value to trade off between thr and lat */ | |
54b60456 | 5635 | bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
aee69d78 | 5636 | bfqq->budget_timeout = bfq_smallest_from_now(); |
aee69d78 | 5637 | |
44e44a1b | 5638 | bfqq->wr_coeff = 1; |
36eca894 | 5639 | bfqq->last_wr_start_finish = jiffies; |
77b7dcea | 5640 | bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now(); |
36eca894 | 5641 | bfqq->split_time = bfq_smallest_from_now(); |
77b7dcea PV |
5642 | |
5643 | /* | |
a34b0244 PV |
5644 | * To not forget the possibly high bandwidth consumed by a |
5645 | * process/queue in the recent past, | |
5646 | * bfq_bfqq_softrt_next_start() returns a value at least equal | |
5647 | * to the current value of bfqq->soft_rt_next_start (see | |
5648 | * comments on bfq_bfqq_softrt_next_start). Set | |
5649 | * soft_rt_next_start to now, to mean that bfqq has consumed | |
5650 | * no bandwidth so far. | |
77b7dcea | 5651 | */ |
a34b0244 | 5652 | bfqq->soft_rt_next_start = jiffies; |
44e44a1b | 5653 | |
aee69d78 PV |
5654 | /* first request is almost certainly seeky */ |
5655 | bfqq->seek_history = 1; | |
1c970450 KS |
5656 | |
5657 | bfqq->decrease_time_jif = jiffies; | |
aee69d78 PV |
5658 | } |
5659 | ||
5660 | static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, | |
e21b7a0b | 5661 | struct bfq_group *bfqg, |
8b7fd741 | 5662 | int ioprio_class, int ioprio, int act_idx) |
aee69d78 PV |
5663 | { |
5664 | switch (ioprio_class) { | |
5665 | case IOPRIO_CLASS_RT: | |
8b7fd741 | 5666 | return &bfqg->async_bfqq[0][ioprio][act_idx]; |
aee69d78 | 5667 | case IOPRIO_CLASS_NONE: |
e70344c0 | 5668 | ioprio = IOPRIO_BE_NORM; |
df561f66 | 5669 | fallthrough; |
aee69d78 | 5670 | case IOPRIO_CLASS_BE: |
8b7fd741 | 5671 | return &bfqg->async_bfqq[1][ioprio][act_idx]; |
aee69d78 | 5672 | case IOPRIO_CLASS_IDLE: |
8b7fd741 | 5673 | return &bfqg->async_idle_bfqq[act_idx]; |
aee69d78 PV |
5674 | default: |
5675 | return NULL; | |
5676 | } | |
5677 | } | |
5678 | ||
430a67f9 PV |
5679 | static struct bfq_queue * |
5680 | bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5681 | struct bfq_io_cq *bic, | |
5682 | struct bfq_queue *last_bfqq_created) | |
5683 | { | |
fd571df0 | 5684 | unsigned int a_idx = last_bfqq_created->actuator_idx; |
430a67f9 PV |
5685 | struct bfq_queue *new_bfqq = |
5686 | bfq_setup_merge(bfqq, last_bfqq_created); | |
5687 | ||
5688 | if (!new_bfqq) | |
5689 | return bfqq; | |
5690 | ||
5691 | if (new_bfqq->bic) | |
fd571df0 PV |
5692 | new_bfqq->bic->bfqq_data[a_idx].stably_merged = true; |
5693 | bic->bfqq_data[a_idx].stably_merged = true; | |
430a67f9 PV |
5694 | |
5695 | /* | |
5696 | * Reusing merge functions. This implies that | |
5697 | * bfqq->bic must be set too, for | |
5698 | * bfq_merge_bfqqs to correctly save bfqq's | |
5699 | * state before killing it. | |
5700 | */ | |
5701 | bfqq->bic = bic; | |
5702 | bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq); | |
5703 | ||
5704 | return new_bfqq; | |
5705 | } | |
5706 | ||
5707 | /* | |
5708 | * Many throughput-sensitive workloads are made of several parallel | |
5709 | * I/O flows, with all flows generated by the same application, or | |
5710 | * more generically by the same task (e.g., system boot). The most | |
5711 | * counterproductive action with these workloads is plugging I/O | |
5712 | * dispatch when one of the bfq_queues associated with these flows | |
5713 | * remains temporarily empty. | |
5714 | * | |
5715 | * To avoid this plugging, BFQ has been using a burst-handling | |
5716 | * mechanism for years now. This mechanism has proven effective for | |
5717 | * throughput, and not detrimental for service guarantees. The | |
5718 | * following function pushes this mechanism a little bit further, | |
5719 | * basing on the following two facts. | |
5720 | * | |
5721 | * First, all the I/O flows of a the same application or task | |
5722 | * contribute to the execution/completion of that common application | |
5723 | * or task. So the performance figures that matter are total | |
5724 | * throughput of the flows and task-wide I/O latency. In particular, | |
5725 | * these flows do not need to be protected from each other, in terms | |
5726 | * of individual bandwidth or latency. | |
5727 | * | |
5728 | * Second, the above fact holds regardless of the number of flows. | |
5729 | * | |
5730 | * Putting these two facts together, this commits merges stably the | |
5731 | * bfq_queues associated with these I/O flows, i.e., with the | |
5732 | * processes that generate these IO/ flows, regardless of how many the | |
5733 | * involved processes are. | |
5734 | * | |
5735 | * To decide whether a set of bfq_queues is actually associated with | |
5736 | * the I/O flows of a common application or task, and to merge these | |
5737 | * queues stably, this function operates as follows: given a bfq_queue, | |
5738 | * say Q2, currently being created, and the last bfq_queue, say Q1, | |
5739 | * created before Q2, Q2 is merged stably with Q1 if | |
5740 | * - very little time has elapsed since when Q1 was created | |
5741 | * - Q2 has the same ioprio as Q1 | |
5742 | * - Q2 belongs to the same group as Q1 | |
5743 | * | |
5744 | * Merging bfq_queues also reduces scheduling overhead. A fio test | |
5745 | * with ten random readers on /dev/nullb shows a throughput boost of | |
5746 | * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of | |
5747 | * the total per-request processing time, the above throughput boost | |
5748 | * implies that BFQ's overhead is reduced by more than 50%. | |
5749 | * | |
5750 | * This new mechanism most certainly obsoletes the current | |
5751 | * burst-handling heuristics. We keep those heuristics for the moment. | |
5752 | */ | |
5753 | static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd, | |
5754 | struct bfq_queue *bfqq, | |
5755 | struct bfq_io_cq *bic) | |
5756 | { | |
5757 | struct bfq_queue **source_bfqq = bfqq->entity.parent ? | |
5758 | &bfqq->entity.parent->last_bfqq_created : | |
5759 | &bfqd->last_bfqq_created; | |
5760 | ||
5761 | struct bfq_queue *last_bfqq_created = *source_bfqq; | |
5762 | ||
5763 | /* | |
5764 | * If last_bfqq_created has not been set yet, then init it. If | |
5765 | * it has been set already, but too long ago, then move it | |
5766 | * forward to bfqq. Finally, move also if bfqq belongs to a | |
5767 | * different group than last_bfqq_created, or if bfqq has a | |
b7529898 PV |
5768 | * different ioprio, ioprio_class or actuator_idx. If none of |
5769 | * these conditions holds true, then try an early stable merge | |
5770 | * or schedule a delayed stable merge. As for the condition on | |
5771 | * actuator_idx, the reason is that, if queues associated with | |
5772 | * different actuators are merged, then control is lost on | |
5773 | * each actuator. Therefore some actuator may be | |
5774 | * underutilized, and throughput may decrease. | |
430a67f9 PV |
5775 | * |
5776 | * A delayed merge is scheduled (instead of performing an | |
5777 | * early merge), in case bfqq might soon prove to be more | |
5778 | * throughput-beneficial if not merged. Currently this is | |
5779 | * possible only if bfqd is rotational with no queueing. For | |
5780 | * such a drive, not merging bfqq is better for throughput if | |
5781 | * bfqq happens to contain sequential I/O. So, we wait a | |
5782 | * little bit for enough I/O to flow through bfqq. After that, | |
5783 | * if such an I/O is sequential, then the merge is | |
5784 | * canceled. Otherwise the merge is finally performed. | |
5785 | */ | |
5786 | if (!last_bfqq_created || | |
5787 | time_before(last_bfqq_created->creation_time + | |
7812472f | 5788 | msecs_to_jiffies(bfq_activation_stable_merging), |
430a67f9 PV |
5789 | bfqq->creation_time) || |
5790 | bfqq->entity.parent != last_bfqq_created->entity.parent || | |
5791 | bfqq->ioprio != last_bfqq_created->ioprio || | |
b7529898 PV |
5792 | bfqq->ioprio_class != last_bfqq_created->ioprio_class || |
5793 | bfqq->actuator_idx != last_bfqq_created->actuator_idx) | |
430a67f9 PV |
5794 | *source_bfqq = bfqq; |
5795 | else if (time_after_eq(last_bfqq_created->creation_time + | |
5796 | bfqd->bfq_burst_interval, | |
5797 | bfqq->creation_time)) { | |
5798 | if (likely(bfqd->nonrot_with_queueing)) | |
5799 | /* | |
5800 | * With this type of drive, leaving | |
5801 | * bfqq alone may provide no | |
5802 | * throughput benefits compared with | |
5803 | * merging bfqq. So merge bfqq now. | |
5804 | */ | |
5805 | bfqq = bfq_do_early_stable_merge(bfqd, bfqq, | |
5806 | bic, | |
5807 | last_bfqq_created); | |
5808 | else { /* schedule tentative stable merge */ | |
5809 | /* | |
5810 | * get reference on last_bfqq_created, | |
5811 | * to prevent it from being freed, | |
5812 | * until we decide whether to merge | |
5813 | */ | |
5814 | last_bfqq_created->ref++; | |
5815 | /* | |
5816 | * need to keep track of stable refs, to | |
5817 | * compute process refs correctly | |
5818 | */ | |
5819 | last_bfqq_created->stable_ref++; | |
5820 | /* | |
5821 | * Record the bfqq to merge to. | |
5822 | */ | |
fd571df0 PV |
5823 | bic->bfqq_data[last_bfqq_created->actuator_idx].stable_merge_bfqq = |
5824 | last_bfqq_created; | |
430a67f9 PV |
5825 | } |
5826 | } | |
5827 | ||
5828 | return bfqq; | |
5829 | } | |
5830 | ||
5831 | ||
aee69d78 PV |
5832 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
5833 | struct bio *bio, bool is_sync, | |
430a67f9 PV |
5834 | struct bfq_io_cq *bic, |
5835 | bool respawn) | |
aee69d78 PV |
5836 | { |
5837 | const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5838 | const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
5839 | struct bfq_queue **async_bfqq = NULL; | |
5840 | struct bfq_queue *bfqq; | |
e21b7a0b | 5841 | struct bfq_group *bfqg; |
aee69d78 | 5842 | |
4e54a249 | 5843 | bfqg = bfq_bio_bfqg(bfqd, bio); |
aee69d78 | 5844 | if (!is_sync) { |
e21b7a0b | 5845 | async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
8b7fd741 DZ |
5846 | ioprio, |
5847 | bfq_actuator_index(bfqd, bio)); | |
aee69d78 PV |
5848 | bfqq = *async_bfqq; |
5849 | if (bfqq) | |
5850 | goto out; | |
5851 | } | |
5852 | ||
5853 | bfqq = kmem_cache_alloc_node(bfq_pool, | |
5854 | GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN, | |
5855 | bfqd->queue->node); | |
5856 | ||
5857 | if (bfqq) { | |
5858 | bfq_init_bfqq(bfqd, bfqq, bic, current->pid, | |
9778369a | 5859 | is_sync, bfq_actuator_index(bfqd, bio)); |
e21b7a0b | 5860 | bfq_init_entity(&bfqq->entity, bfqg); |
aee69d78 PV |
5861 | bfq_log_bfqq(bfqd, bfqq, "allocated"); |
5862 | } else { | |
5863 | bfqq = &bfqd->oom_bfqq; | |
5864 | bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); | |
5865 | goto out; | |
5866 | } | |
5867 | ||
5868 | /* | |
5869 | * Pin the queue now that it's allocated, scheduler exit will | |
5870 | * prune it. | |
5871 | */ | |
5872 | if (async_bfqq) { | |
e21b7a0b AA |
5873 | bfqq->ref++; /* |
5874 | * Extra group reference, w.r.t. sync | |
5875 | * queue. This extra reference is removed | |
5876 | * only if bfqq->bfqg disappears, to | |
5877 | * guarantee that this queue is not freed | |
5878 | * until its group goes away. | |
5879 | */ | |
5880 | bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", | |
aee69d78 PV |
5881 | bfqq, bfqq->ref); |
5882 | *async_bfqq = bfqq; | |
5883 | } | |
5884 | ||
5885 | out: | |
5886 | bfqq->ref++; /* get a process reference to this queue */ | |
430a67f9 PV |
5887 | |
5888 | if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn) | |
5889 | bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic); | |
aee69d78 PV |
5890 | return bfqq; |
5891 | } | |
5892 | ||
5893 | static void bfq_update_io_thinktime(struct bfq_data *bfqd, | |
5894 | struct bfq_queue *bfqq) | |
5895 | { | |
5896 | struct bfq_ttime *ttime = &bfqq->ttime; | |
7684fbde | 5897 | u64 elapsed; |
aee69d78 | 5898 | |
7684fbde JK |
5899 | /* |
5900 | * We are really interested in how long it takes for the queue to | |
5901 | * become busy when there is no outstanding IO for this queue. So | |
5902 | * ignore cases when the bfq queue has already IO queued. | |
5903 | */ | |
5904 | if (bfqq->dispatched || bfq_bfqq_busy(bfqq)) | |
5905 | return; | |
5906 | elapsed = ktime_get_ns() - bfqq->ttime.last_end_request; | |
aee69d78 PV |
5907 | elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle); |
5908 | ||
28c6def0 | 5909 | ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; |
aee69d78 PV |
5910 | ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8); |
5911 | ttime->ttime_mean = div64_ul(ttime->ttime_total + 128, | |
5912 | ttime->ttime_samples); | |
5913 | } | |
5914 | ||
5915 | static void | |
5916 | bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5917 | struct request *rq) | |
5918 | { | |
aee69d78 | 5919 | bfqq->seek_history <<= 1; |
d87447d8 | 5920 | bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq); |
7074f076 PV |
5921 | |
5922 | if (bfqq->wr_coeff > 1 && | |
5923 | bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && | |
d1f600fa PV |
5924 | BFQQ_TOTALLY_SEEKY(bfqq)) { |
5925 | if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt + | |
5926 | bfq_wr_duration(bfqd))) { | |
5927 | /* | |
5928 | * In soft_rt weight raising with the | |
5929 | * interactive-weight-raising period | |
5930 | * elapsed (so no switch back to | |
5931 | * interactive weight raising). | |
5932 | */ | |
5933 | bfq_bfqq_end_wr(bfqq); | |
5934 | } else { /* | |
5935 | * stopping soft_rt weight raising | |
5936 | * while still in interactive period, | |
5937 | * switch back to interactive weight | |
5938 | * raising | |
5939 | */ | |
5940 | switch_back_to_interactive_wr(bfqq, bfqd); | |
5941 | bfqq->entity.prio_changed = 1; | |
5942 | } | |
5943 | } | |
aee69d78 PV |
5944 | } |
5945 | ||
d5be3fef PV |
5946 | static void bfq_update_has_short_ttime(struct bfq_data *bfqd, |
5947 | struct bfq_queue *bfqq, | |
5948 | struct bfq_io_cq *bic) | |
aee69d78 | 5949 | { |
766d6141 | 5950 | bool has_short_ttime = true, state_changed; |
aee69d78 | 5951 | |
d5be3fef PV |
5952 | /* |
5953 | * No need to update has_short_ttime if bfqq is async or in | |
5954 | * idle io prio class, or if bfq_slice_idle is zero, because | |
5955 | * no device idling is performed for bfqq in this case. | |
5956 | */ | |
5957 | if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) || | |
5958 | bfqd->bfq_slice_idle == 0) | |
aee69d78 PV |
5959 | return; |
5960 | ||
36eca894 AA |
5961 | /* Idle window just restored, statistics are meaningless. */ |
5962 | if (time_is_after_eq_jiffies(bfqq->split_time + | |
5963 | bfqd->bfq_wr_min_idle_time)) | |
5964 | return; | |
5965 | ||
d5be3fef | 5966 | /* Think time is infinite if no process is linked to |
b5f74eca PV |
5967 | * bfqq. Otherwise check average think time to decide whether |
5968 | * to mark as has_short_ttime. To this goal, compare average | |
5969 | * think time with half the I/O-plugging timeout. | |
d5be3fef | 5970 | */ |
aee69d78 | 5971 | if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
d5be3fef | 5972 | (bfq_sample_valid(bfqq->ttime.ttime_samples) && |
b5f74eca | 5973 | bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1)) |
d5be3fef PV |
5974 | has_short_ttime = false; |
5975 | ||
766d6141 | 5976 | state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq); |
aee69d78 | 5977 | |
d5be3fef PV |
5978 | if (has_short_ttime) |
5979 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
aee69d78 | 5980 | else |
d5be3fef | 5981 | bfq_clear_bfqq_has_short_ttime(bfqq); |
766d6141 PV |
5982 | |
5983 | /* | |
5984 | * Until the base value for the total service time gets | |
5985 | * finally computed for bfqq, the inject limit does depend on | |
5986 | * the think-time state (short|long). In particular, the limit | |
5987 | * is 0 or 1 if the think time is deemed, respectively, as | |
5988 | * short or long (details in the comments in | |
5989 | * bfq_update_inject_limit()). Accordingly, the next | |
5990 | * instructions reset the inject limit if the think-time state | |
5991 | * has changed and the above base value is still to be | |
5992 | * computed. | |
5993 | * | |
5994 | * However, the reset is performed only if more than 100 ms | |
5995 | * have elapsed since the last update of the inject limit, or | |
5996 | * (inclusive) if the change is from short to long think | |
5997 | * time. The reason for this waiting is as follows. | |
5998 | * | |
5999 | * bfqq may have a long think time because of a | |
6000 | * synchronization with some other queue, i.e., because the | |
6001 | * I/O of some other queue may need to be completed for bfqq | |
13a857a4 PV |
6002 | * to receive new I/O. Details in the comments on the choice |
6003 | * of the queue for injection in bfq_select_queue(). | |
766d6141 | 6004 | * |
13a857a4 PV |
6005 | * As stressed in those comments, if such a synchronization is |
6006 | * actually in place, then, without injection on bfqq, the | |
6007 | * blocking I/O cannot happen to served while bfqq is in | |
6008 | * service. As a consequence, if bfqq is granted | |
6009 | * I/O-dispatch-plugging, then bfqq remains empty, and no I/O | |
6010 | * is dispatched, until the idle timeout fires. This is likely | |
6011 | * to result in lower bandwidth and higher latencies for bfqq, | |
6012 | * and in a severe loss of total throughput. | |
766d6141 PV |
6013 | * |
6014 | * On the opposite end, a non-zero inject limit may allow the | |
6015 | * I/O that blocks bfqq to be executed soon, and therefore | |
13a857a4 PV |
6016 | * bfqq to receive new I/O soon. |
6017 | * | |
6018 | * But, if the blocking gets actually eliminated, then the | |
6019 | * next think-time sample for bfqq may be very low. This in | |
6020 | * turn may cause bfqq's think time to be deemed | |
6021 | * short. Without the 100 ms barrier, this new state change | |
6022 | * would cause the body of the next if to be executed | |
766d6141 PV |
6023 | * immediately. But this would set to 0 the inject |
6024 | * limit. Without injection, the blocking I/O would cause the | |
6025 | * think time of bfqq to become long again, and therefore the | |
6026 | * inject limit to be raised again, and so on. The only effect | |
6027 | * of such a steady oscillation between the two think-time | |
6028 | * states would be to prevent effective injection on bfqq. | |
6029 | * | |
6030 | * In contrast, if the inject limit is not reset during such a | |
6031 | * long time interval as 100 ms, then the number of short | |
6032 | * think time samples can grow significantly before the reset | |
13a857a4 PV |
6033 | * is performed. As a consequence, the think time state can |
6034 | * become stable before the reset. Therefore there will be no | |
6035 | * state change when the 100 ms elapse, and no reset of the | |
6036 | * inject limit. The inject limit remains steadily equal to 1 | |
6037 | * both during and after the 100 ms. So injection can be | |
766d6141 PV |
6038 | * performed at all times, and throughput gets boosted. |
6039 | * | |
6040 | * An inject limit equal to 1 is however in conflict, in | |
6041 | * general, with the fact that the think time of bfqq is | |
6042 | * short, because injection may be likely to delay bfqq's I/O | |
6043 | * (as explained in the comments in | |
6044 | * bfq_update_inject_limit()). But this does not happen in | |
6045 | * this special case, because bfqq's low think time is due to | |
6046 | * an effective handling of a synchronization, through | |
6047 | * injection. In this special case, bfqq's I/O does not get | |
6048 | * delayed by injection; on the contrary, bfqq's I/O is | |
6049 | * brought forward, because it is not blocked for | |
6050 | * milliseconds. | |
6051 | * | |
13a857a4 PV |
6052 | * In addition, serving the blocking I/O much sooner, and much |
6053 | * more frequently than once per I/O-plugging timeout, makes | |
6054 | * it much quicker to detect a waker queue (the concept of | |
6055 | * waker queue is defined in the comments in | |
6056 | * bfq_add_request()). This makes it possible to start sooner | |
6057 | * to boost throughput more effectively, by injecting the I/O | |
6058 | * of the waker queue unconditionally on every | |
6059 | * bfq_dispatch_request(). | |
6060 | * | |
6061 | * One last, important benefit of not resetting the inject | |
6062 | * limit before 100 ms is that, during this time interval, the | |
6063 | * base value for the total service time is likely to get | |
6064 | * finally computed for bfqq, freeing the inject limit from | |
6065 | * its relation with the think time. | |
766d6141 PV |
6066 | */ |
6067 | if (state_changed && bfqq->last_serv_time_ns == 0 && | |
6068 | (time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
6069 | msecs_to_jiffies(100)) || | |
6070 | !has_short_ttime)) | |
6071 | bfq_reset_inject_limit(bfqd, bfqq); | |
aee69d78 PV |
6072 | } |
6073 | ||
6074 | /* | |
6075 | * Called when a new fs request (rq) is added to bfqq. Check if there's | |
6076 | * something we should do about it. | |
6077 | */ | |
6078 | static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
6079 | struct request *rq) | |
6080 | { | |
aee69d78 PV |
6081 | if (rq->cmd_flags & REQ_META) |
6082 | bfqq->meta_pending++; | |
6083 | ||
aee69d78 PV |
6084 | bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
6085 | ||
6086 | if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { | |
6087 | bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 && | |
6088 | blk_rq_sectors(rq) < 32; | |
6089 | bool budget_timeout = bfq_bfqq_budget_timeout(bfqq); | |
6090 | ||
6091 | /* | |
ac8b0cb4 PV |
6092 | * There is just this request queued: if |
6093 | * - the request is small, and | |
6094 | * - we are idling to boost throughput, and | |
6095 | * - the queue is not to be expired, | |
6096 | * then just exit. | |
aee69d78 PV |
6097 | * |
6098 | * In this way, if the device is being idled to wait | |
6099 | * for a new request from the in-service queue, we | |
6100 | * avoid unplugging the device and committing the | |
ac8b0cb4 PV |
6101 | * device to serve just a small request. In contrast |
6102 | * we wait for the block layer to decide when to | |
6103 | * unplug the device: hopefully, new requests will be | |
6104 | * merged to this one quickly, then the device will be | |
6105 | * unplugged and larger requests will be dispatched. | |
aee69d78 | 6106 | */ |
ac8b0cb4 PV |
6107 | if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) && |
6108 | !budget_timeout) | |
aee69d78 PV |
6109 | return; |
6110 | ||
6111 | /* | |
ac8b0cb4 PV |
6112 | * A large enough request arrived, or idling is being |
6113 | * performed to preserve service guarantees, or | |
6114 | * finally the queue is to be expired: in all these | |
6115 | * cases disk idling is to be stopped, so clear | |
6116 | * wait_request flag and reset timer. | |
aee69d78 PV |
6117 | */ |
6118 | bfq_clear_bfqq_wait_request(bfqq); | |
6119 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
6120 | ||
6121 | /* | |
6122 | * The queue is not empty, because a new request just | |
6123 | * arrived. Hence we can safely expire the queue, in | |
6124 | * case of budget timeout, without risking that the | |
6125 | * timestamps of the queue are not updated correctly. | |
6126 | * See [1] for more details. | |
6127 | */ | |
6128 | if (budget_timeout) | |
6129 | bfq_bfqq_expire(bfqd, bfqq, false, | |
6130 | BFQQE_BUDGET_TIMEOUT); | |
6131 | } | |
6132 | } | |
6133 | ||
98f04499 JK |
6134 | static void bfqq_request_allocated(struct bfq_queue *bfqq) |
6135 | { | |
6136 | struct bfq_entity *entity = &bfqq->entity; | |
6137 | ||
6138 | for_each_entity(entity) | |
6139 | entity->allocated++; | |
6140 | } | |
6141 | ||
6142 | static void bfqq_request_freed(struct bfq_queue *bfqq) | |
6143 | { | |
6144 | struct bfq_entity *entity = &bfqq->entity; | |
6145 | ||
6146 | for_each_entity(entity) | |
6147 | entity->allocated--; | |
6148 | } | |
6149 | ||
24bfd19b PV |
6150 | /* returns true if it causes the idle timer to be disabled */ |
6151 | static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq) | |
aee69d78 | 6152 | { |
36eca894 | 6153 | struct bfq_queue *bfqq = RQ_BFQQ(rq), |
430a67f9 PV |
6154 | *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true, |
6155 | RQ_BIC(rq)); | |
24bfd19b | 6156 | bool waiting, idle_timer_disabled = false; |
36eca894 AA |
6157 | |
6158 | if (new_bfqq) { | |
36eca894 AA |
6159 | /* |
6160 | * Release the request's reference to the old bfqq | |
6161 | * and make sure one is taken to the shared queue. | |
6162 | */ | |
98f04499 JK |
6163 | bfqq_request_allocated(new_bfqq); |
6164 | bfqq_request_freed(bfqq); | |
36eca894 AA |
6165 | new_bfqq->ref++; |
6166 | /* | |
6167 | * If the bic associated with the process | |
6168 | * issuing this request still points to bfqq | |
6169 | * (and thus has not been already redirected | |
6170 | * to new_bfqq or even some other bfq_queue), | |
6171 | * then complete the merge and redirect it to | |
6172 | * new_bfqq. | |
6173 | */ | |
9778369a PV |
6174 | if (bic_to_bfqq(RQ_BIC(rq), true, |
6175 | bfq_actuator_index(bfqd, rq->bio)) == bfqq) | |
36eca894 AA |
6176 | bfq_merge_bfqqs(bfqd, RQ_BIC(rq), |
6177 | bfqq, new_bfqq); | |
894df937 PV |
6178 | |
6179 | bfq_clear_bfqq_just_created(bfqq); | |
36eca894 AA |
6180 | /* |
6181 | * rq is about to be enqueued into new_bfqq, | |
6182 | * release rq reference on bfqq | |
6183 | */ | |
6184 | bfq_put_queue(bfqq); | |
6185 | rq->elv.priv[1] = new_bfqq; | |
6186 | bfqq = new_bfqq; | |
6187 | } | |
aee69d78 | 6188 | |
a3f9bce3 PV |
6189 | bfq_update_io_thinktime(bfqd, bfqq); |
6190 | bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq)); | |
6191 | bfq_update_io_seektime(bfqd, bfqq, rq); | |
6192 | ||
24bfd19b | 6193 | waiting = bfqq && bfq_bfqq_wait_request(bfqq); |
aee69d78 | 6194 | bfq_add_request(rq); |
24bfd19b | 6195 | idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq); |
aee69d78 PV |
6196 | |
6197 | rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)]; | |
6198 | list_add_tail(&rq->queuelist, &bfqq->fifo); | |
6199 | ||
6200 | bfq_rq_enqueued(bfqd, bfqq, rq); | |
24bfd19b PV |
6201 | |
6202 | return idle_timer_disabled; | |
aee69d78 PV |
6203 | } |
6204 | ||
8060c47b | 6205 | #ifdef CONFIG_BFQ_CGROUP_DEBUG |
9b25bd03 PV |
6206 | static void bfq_update_insert_stats(struct request_queue *q, |
6207 | struct bfq_queue *bfqq, | |
6208 | bool idle_timer_disabled, | |
dc469ba2 | 6209 | blk_opf_t cmd_flags) |
9b25bd03 PV |
6210 | { |
6211 | if (!bfqq) | |
6212 | return; | |
6213 | ||
6214 | /* | |
6215 | * bfqq still exists, because it can disappear only after | |
6216 | * either it is merged with another queue, or the process it | |
6217 | * is associated with exits. But both actions must be taken by | |
6218 | * the same process currently executing this flow of | |
6219 | * instructions. | |
6220 | * | |
6221 | * In addition, the following queue lock guarantees that | |
6222 | * bfqq_group(bfqq) exists as well. | |
6223 | */ | |
0d945c1f | 6224 | spin_lock_irq(&q->queue_lock); |
9b25bd03 PV |
6225 | bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags); |
6226 | if (idle_timer_disabled) | |
6227 | bfqg_stats_update_idle_time(bfqq_group(bfqq)); | |
0d945c1f | 6228 | spin_unlock_irq(&q->queue_lock); |
9b25bd03 PV |
6229 | } |
6230 | #else | |
6231 | static inline void bfq_update_insert_stats(struct request_queue *q, | |
6232 | struct bfq_queue *bfqq, | |
6233 | bool idle_timer_disabled, | |
dc469ba2 | 6234 | blk_opf_t cmd_flags) {} |
8060c47b | 6235 | #endif /* CONFIG_BFQ_CGROUP_DEBUG */ |
9b25bd03 | 6236 | |
5f550ede JK |
6237 | static struct bfq_queue *bfq_init_rq(struct request *rq); |
6238 | ||
aee69d78 | 6239 | static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
93fffe16 | 6240 | blk_insert_t flags) |
aee69d78 PV |
6241 | { |
6242 | struct request_queue *q = hctx->queue; | |
6243 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
18e5a57d | 6244 | struct bfq_queue *bfqq; |
24bfd19b | 6245 | bool idle_timer_disabled = false; |
dc469ba2 | 6246 | blk_opf_t cmd_flags; |
fd2ef39c | 6247 | LIST_HEAD(free); |
aee69d78 | 6248 | |
fd41e603 TH |
6249 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6250 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio) | |
6251 | bfqg_stats_update_legacy_io(q, rq); | |
6252 | #endif | |
aee69d78 | 6253 | spin_lock_irq(&bfqd->lock); |
5f550ede | 6254 | bfqq = bfq_init_rq(rq); |
fd2ef39c | 6255 | if (blk_mq_sched_try_insert_merge(q, rq, &free)) { |
aee69d78 | 6256 | spin_unlock_irq(&bfqd->lock); |
fd2ef39c | 6257 | blk_mq_free_requests(&free); |
aee69d78 PV |
6258 | return; |
6259 | } | |
6260 | ||
b357e4a6 | 6261 | trace_block_rq_insert(rq); |
aee69d78 | 6262 | |
93fffe16 CH |
6263 | if (flags & BLK_MQ_INSERT_AT_HEAD) { |
6264 | list_add(&rq->queuelist, &bfqd->dispatch); | |
6265 | } else if (!bfqq) { | |
6266 | list_add_tail(&rq->queuelist, &bfqd->dispatch); | |
fd03177c | 6267 | } else { |
24bfd19b | 6268 | idle_timer_disabled = __bfq_insert_request(bfqd, rq); |
614822f8 LM |
6269 | /* |
6270 | * Update bfqq, because, if a queue merge has occurred | |
6271 | * in __bfq_insert_request, then rq has been | |
6272 | * redirected into a new queue. | |
6273 | */ | |
6274 | bfqq = RQ_BFQQ(rq); | |
aee69d78 PV |
6275 | |
6276 | if (rq_mergeable(rq)) { | |
6277 | elv_rqhash_add(q, rq); | |
6278 | if (!q->last_merge) | |
6279 | q->last_merge = rq; | |
6280 | } | |
6281 | } | |
6282 | ||
24bfd19b PV |
6283 | /* |
6284 | * Cache cmd_flags before releasing scheduler lock, because rq | |
6285 | * may disappear afterwards (for example, because of a request | |
6286 | * merge). | |
6287 | */ | |
6288 | cmd_flags = rq->cmd_flags; | |
6fa3e8d3 | 6289 | spin_unlock_irq(&bfqd->lock); |
24bfd19b | 6290 | |
9b25bd03 PV |
6291 | bfq_update_insert_stats(q, bfqq, idle_timer_disabled, |
6292 | cmd_flags); | |
aee69d78 PV |
6293 | } |
6294 | ||
6295 | static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx, | |
93fffe16 CH |
6296 | struct list_head *list, |
6297 | blk_insert_t flags) | |
aee69d78 PV |
6298 | { |
6299 | while (!list_empty(list)) { | |
6300 | struct request *rq; | |
6301 | ||
6302 | rq = list_first_entry(list, struct request, queuelist); | |
6303 | list_del_init(&rq->queuelist); | |
93fffe16 | 6304 | bfq_insert_request(hctx, rq, flags); |
aee69d78 PV |
6305 | } |
6306 | } | |
6307 | ||
6308 | static void bfq_update_hw_tag(struct bfq_data *bfqd) | |
6309 | { | |
b3c34981 PV |
6310 | struct bfq_queue *bfqq = bfqd->in_service_queue; |
6311 | ||
aee69d78 | 6312 | bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver, |
2d31c684 | 6313 | bfqd->tot_rq_in_driver); |
aee69d78 PV |
6314 | |
6315 | if (bfqd->hw_tag == 1) | |
6316 | return; | |
6317 | ||
6318 | /* | |
6319 | * This sample is valid if the number of outstanding requests | |
6320 | * is large enough to allow a queueing behavior. Note that the | |
6321 | * sum is not exact, as it's not taking into account deactivated | |
6322 | * requests. | |
6323 | */ | |
2d31c684 | 6324 | if (bfqd->tot_rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD) |
aee69d78 PV |
6325 | return; |
6326 | ||
b3c34981 PV |
6327 | /* |
6328 | * If active queue hasn't enough requests and can idle, bfq might not | |
6329 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
6330 | * case | |
6331 | */ | |
6332 | if (bfqq && bfq_bfqq_has_short_ttime(bfqq) && | |
6333 | bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] < | |
6334 | BFQ_HW_QUEUE_THRESHOLD && | |
2d31c684 | 6335 | bfqd->tot_rq_in_driver < BFQ_HW_QUEUE_THRESHOLD) |
b3c34981 PV |
6336 | return; |
6337 | ||
aee69d78 PV |
6338 | if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
6339 | return; | |
6340 | ||
6341 | bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; | |
6342 | bfqd->max_rq_in_driver = 0; | |
6343 | bfqd->hw_tag_samples = 0; | |
8cacc5ab PV |
6344 | |
6345 | bfqd->nonrot_with_queueing = | |
6346 | blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag; | |
aee69d78 PV |
6347 | } |
6348 | ||
6349 | static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd) | |
6350 | { | |
ab0e43e9 PV |
6351 | u64 now_ns; |
6352 | u32 delta_us; | |
6353 | ||
aee69d78 PV |
6354 | bfq_update_hw_tag(bfqd); |
6355 | ||
2d31c684 DZ |
6356 | bfqd->rq_in_driver[bfqq->actuator_idx]--; |
6357 | bfqd->tot_rq_in_driver--; | |
aee69d78 PV |
6358 | bfqq->dispatched--; |
6359 | ||
44e44a1b PV |
6360 | if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) { |
6361 | /* | |
6362 | * Set budget_timeout (which we overload to store the | |
6363 | * time at which the queue remains with no backlog and | |
6364 | * no outstanding request; used by the weight-raising | |
6365 | * mechanism). | |
6366 | */ | |
6367 | bfqq->budget_timeout = jiffies; | |
1de0c4cd | 6368 | |
3d89bd12 | 6369 | bfq_del_bfqq_in_groups_with_pending_reqs(bfqq); |
afdba146 | 6370 | bfq_weights_tree_remove(bfqq); |
44e44a1b PV |
6371 | } |
6372 | ||
ab0e43e9 PV |
6373 | now_ns = ktime_get_ns(); |
6374 | ||
6375 | bfqq->ttime.last_end_request = now_ns; | |
6376 | ||
6377 | /* | |
6378 | * Using us instead of ns, to get a reasonable precision in | |
6379 | * computing rate in next check. | |
6380 | */ | |
6381 | delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC); | |
6382 | ||
6383 | /* | |
6384 | * If the request took rather long to complete, and, according | |
6385 | * to the maximum request size recorded, this completion latency | |
6386 | * implies that the request was certainly served at a very low | |
6387 | * rate (less than 1M sectors/sec), then the whole observation | |
6388 | * interval that lasts up to this time instant cannot be a | |
6389 | * valid time interval for computing a new peak rate. Invoke | |
6390 | * bfq_update_rate_reset to have the following three steps | |
6391 | * taken: | |
6392 | * - close the observation interval at the last (previous) | |
6393 | * request dispatch or completion | |
6394 | * - compute rate, if possible, for that observation interval | |
6395 | * - reset to zero samples, which will trigger a proper | |
6396 | * re-initialization of the observation interval on next | |
6397 | * dispatch | |
6398 | */ | |
6399 | if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC && | |
6400 | (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us < | |
6401 | 1UL<<(BFQ_RATE_SHIFT - 10)) | |
6402 | bfq_update_rate_reset(bfqd, NULL); | |
6403 | bfqd->last_completion = now_ns; | |
85686d0d PV |
6404 | /* |
6405 | * Shared queues are likely to receive I/O at a high | |
6406 | * rate. This may deceptively let them be considered as wakers | |
6407 | * of other queues. But a false waker will unjustly steal | |
6408 | * bandwidth to its supposedly woken queue. So considering | |
6409 | * also shared queues in the waking mechanism may cause more | |
9a2ac41b PV |
6410 | * control troubles than throughput benefits. Then reset |
6411 | * last_completed_rq_bfqq if bfqq is a shared queue. | |
85686d0d PV |
6412 | */ |
6413 | if (!bfq_bfqq_coop(bfqq)) | |
6414 | bfqd->last_completed_rq_bfqq = bfqq; | |
9a2ac41b PV |
6415 | else |
6416 | bfqd->last_completed_rq_bfqq = NULL; | |
aee69d78 | 6417 | |
77b7dcea PV |
6418 | /* |
6419 | * If we are waiting to discover whether the request pattern | |
6420 | * of the task associated with the queue is actually | |
6421 | * isochronous, and both requisites for this condition to hold | |
6422 | * are now satisfied, then compute soft_rt_next_start (see the | |
6423 | * comments on the function bfq_bfqq_softrt_next_start()). We | |
20cd3245 PV |
6424 | * do not compute soft_rt_next_start if bfqq is in interactive |
6425 | * weight raising (see the comments in bfq_bfqq_expire() for | |
6426 | * an explanation). We schedule this delayed update when bfqq | |
6427 | * expires, if it still has in-flight requests. | |
77b7dcea PV |
6428 | */ |
6429 | if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 && | |
20cd3245 PV |
6430 | RB_EMPTY_ROOT(&bfqq->sort_list) && |
6431 | bfqq->wr_coeff != bfqd->bfq_wr_coeff) | |
77b7dcea PV |
6432 | bfqq->soft_rt_next_start = |
6433 | bfq_bfqq_softrt_next_start(bfqd, bfqq); | |
6434 | ||
aee69d78 PV |
6435 | /* |
6436 | * If this is the in-service queue, check if it needs to be expired, | |
6437 | * or if we want to idle in case it has no pending requests. | |
6438 | */ | |
6439 | if (bfqd->in_service_queue == bfqq) { | |
4420b095 PV |
6440 | if (bfq_bfqq_must_idle(bfqq)) { |
6441 | if (bfqq->dispatched == 0) | |
6442 | bfq_arm_slice_timer(bfqd); | |
6443 | /* | |
6444 | * If we get here, we do not expire bfqq, even | |
6445 | * if bfqq was in budget timeout or had no | |
6446 | * more requests (as controlled in the next | |
6447 | * conditional instructions). The reason for | |
6448 | * not expiring bfqq is as follows. | |
6449 | * | |
6450 | * Here bfqq->dispatched > 0 holds, but | |
6451 | * bfq_bfqq_must_idle() returned true. This | |
6452 | * implies that, even if no request arrives | |
6453 | * for bfqq before bfqq->dispatched reaches 0, | |
6454 | * bfqq will, however, not be expired on the | |
6455 | * completion event that causes bfqq->dispatch | |
6456 | * to reach zero. In contrast, on this event, | |
6457 | * bfqq will start enjoying device idling | |
6458 | * (I/O-dispatch plugging). | |
6459 | * | |
6460 | * But, if we expired bfqq here, bfqq would | |
6461 | * not have the chance to enjoy device idling | |
6462 | * when bfqq->dispatched finally reaches | |
6463 | * zero. This would expose bfqq to violation | |
6464 | * of its reserved service guarantees. | |
6465 | */ | |
aee69d78 PV |
6466 | return; |
6467 | } else if (bfq_may_expire_for_budg_timeout(bfqq)) | |
6468 | bfq_bfqq_expire(bfqd, bfqq, false, | |
6469 | BFQQE_BUDGET_TIMEOUT); | |
6470 | else if (RB_EMPTY_ROOT(&bfqq->sort_list) && | |
6471 | (bfqq->dispatched == 0 || | |
277a4a9b | 6472 | !bfq_better_to_idle(bfqq))) |
aee69d78 PV |
6473 | bfq_bfqq_expire(bfqd, bfqq, false, |
6474 | BFQQE_NO_MORE_REQUESTS); | |
6475 | } | |
3f7cb4f4 | 6476 | |
2d31c684 | 6477 | if (!bfqd->tot_rq_in_driver) |
3f7cb4f4 | 6478 | bfq_schedule_dispatch(bfqd); |
aee69d78 PV |
6479 | } |
6480 | ||
2341d662 PV |
6481 | /* |
6482 | * The processes associated with bfqq may happen to generate their | |
6483 | * cumulative I/O at a lower rate than the rate at which the device | |
6484 | * could serve the same I/O. This is rather probable, e.g., if only | |
6485 | * one process is associated with bfqq and the device is an SSD. It | |
6486 | * results in bfqq becoming often empty while in service. In this | |
6487 | * respect, if BFQ is allowed to switch to another queue when bfqq | |
6488 | * remains empty, then the device goes on being fed with I/O requests, | |
6489 | * and the throughput is not affected. In contrast, if BFQ is not | |
6490 | * allowed to switch to another queue---because bfqq is sync and | |
6491 | * I/O-dispatch needs to be plugged while bfqq is temporarily | |
6492 | * empty---then, during the service of bfqq, there will be frequent | |
6493 | * "service holes", i.e., time intervals during which bfqq gets empty | |
6494 | * and the device can only consume the I/O already queued in its | |
6495 | * hardware queues. During service holes, the device may even get to | |
6496 | * remaining idle. In the end, during the service of bfqq, the device | |
6497 | * is driven at a lower speed than the one it can reach with the kind | |
6498 | * of I/O flowing through bfqq. | |
6499 | * | |
6500 | * To counter this loss of throughput, BFQ implements a "request | |
6501 | * injection mechanism", which tries to fill the above service holes | |
6502 | * with I/O requests taken from other queues. The hard part in this | |
6503 | * mechanism is finding the right amount of I/O to inject, so as to | |
6504 | * both boost throughput and not break bfqq's bandwidth and latency | |
6505 | * guarantees. In this respect, the mechanism maintains a per-queue | |
6506 | * inject limit, computed as below. While bfqq is empty, the injection | |
6507 | * mechanism dispatches extra I/O requests only until the total number | |
6508 | * of I/O requests in flight---i.e., already dispatched but not yet | |
6509 | * completed---remains lower than this limit. | |
6510 | * | |
6511 | * A first definition comes in handy to introduce the algorithm by | |
6512 | * which the inject limit is computed. We define as first request for | |
6513 | * bfqq, an I/O request for bfqq that arrives while bfqq is in | |
6514 | * service, and causes bfqq to switch from empty to non-empty. The | |
6515 | * algorithm updates the limit as a function of the effect of | |
6516 | * injection on the service times of only the first requests of | |
6517 | * bfqq. The reason for this restriction is that these are the | |
6518 | * requests whose service time is affected most, because they are the | |
6519 | * first to arrive after injection possibly occurred. | |
6520 | * | |
6521 | * To evaluate the effect of injection, the algorithm measures the | |
6522 | * "total service time" of first requests. We define as total service | |
6523 | * time of an I/O request, the time that elapses since when the | |
6524 | * request is enqueued into bfqq, to when it is completed. This | |
6525 | * quantity allows the whole effect of injection to be measured. It is | |
6526 | * easy to see why. Suppose that some requests of other queues are | |
6527 | * actually injected while bfqq is empty, and that a new request R | |
6528 | * then arrives for bfqq. If the device does start to serve all or | |
6529 | * part of the injected requests during the service hole, then, | |
6530 | * because of this extra service, it may delay the next invocation of | |
6531 | * the dispatch hook of BFQ. Then, even after R gets eventually | |
6532 | * dispatched, the device may delay the actual service of R if it is | |
6533 | * still busy serving the extra requests, or if it decides to serve, | |
6534 | * before R, some extra request still present in its queues. As a | |
6535 | * conclusion, the cumulative extra delay caused by injection can be | |
6536 | * easily evaluated by just comparing the total service time of first | |
6537 | * requests with and without injection. | |
6538 | * | |
6539 | * The limit-update algorithm works as follows. On the arrival of a | |
6540 | * first request of bfqq, the algorithm measures the total time of the | |
6541 | * request only if one of the three cases below holds, and, for each | |
6542 | * case, it updates the limit as described below: | |
6543 | * | |
6544 | * (1) If there is no in-flight request. This gives a baseline for the | |
6545 | * total service time of the requests of bfqq. If the baseline has | |
6546 | * not been computed yet, then, after computing it, the limit is | |
6547 | * set to 1, to start boosting throughput, and to prepare the | |
6548 | * ground for the next case. If the baseline has already been | |
6549 | * computed, then it is updated, in case it results to be lower | |
6550 | * than the previous value. | |
6551 | * | |
6552 | * (2) If the limit is higher than 0 and there are in-flight | |
6553 | * requests. By comparing the total service time in this case with | |
6554 | * the above baseline, it is possible to know at which extent the | |
6555 | * current value of the limit is inflating the total service | |
6556 | * time. If the inflation is below a certain threshold, then bfqq | |
6557 | * is assumed to be suffering from no perceivable loss of its | |
6558 | * service guarantees, and the limit is even tentatively | |
6559 | * increased. If the inflation is above the threshold, then the | |
6560 | * limit is decreased. Due to the lack of any hysteresis, this | |
6561 | * logic makes the limit oscillate even in steady workload | |
6562 | * conditions. Yet we opted for it, because it is fast in reaching | |
6563 | * the best value for the limit, as a function of the current I/O | |
6564 | * workload. To reduce oscillations, this step is disabled for a | |
6565 | * short time interval after the limit happens to be decreased. | |
6566 | * | |
6567 | * (3) Periodically, after resetting the limit, to make sure that the | |
6568 | * limit eventually drops in case the workload changes. This is | |
6569 | * needed because, after the limit has gone safely up for a | |
6570 | * certain workload, it is impossible to guess whether the | |
6571 | * baseline total service time may have changed, without measuring | |
6572 | * it again without injection. A more effective version of this | |
6573 | * step might be to just sample the baseline, by interrupting | |
6574 | * injection only once, and then to reset/lower the limit only if | |
6575 | * the total service time with the current limit does happen to be | |
6576 | * too large. | |
6577 | * | |
6578 | * More details on each step are provided in the comments on the | |
6579 | * pieces of code that implement these steps: the branch handling the | |
6580 | * transition from empty to non empty in bfq_add_request(), the branch | |
6581 | * handling injection in bfq_select_queue(), and the function | |
6582 | * bfq_choose_bfqq_for_injection(). These comments also explain some | |
6583 | * exceptions, made by the injection mechanism in some special cases. | |
6584 | */ | |
6585 | static void bfq_update_inject_limit(struct bfq_data *bfqd, | |
6586 | struct bfq_queue *bfqq) | |
6587 | { | |
6588 | u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns; | |
6589 | unsigned int old_limit = bfqq->inject_limit; | |
6590 | ||
23ed570a | 6591 | if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) { |
2341d662 PV |
6592 | u64 threshold = (bfqq->last_serv_time_ns * 3)>>1; |
6593 | ||
6594 | if (tot_time_ns >= threshold && old_limit > 0) { | |
6595 | bfqq->inject_limit--; | |
6596 | bfqq->decrease_time_jif = jiffies; | |
6597 | } else if (tot_time_ns < threshold && | |
c1e0a182 | 6598 | old_limit <= bfqd->max_rq_in_driver) |
2341d662 PV |
6599 | bfqq->inject_limit++; |
6600 | } | |
6601 | ||
6602 | /* | |
6603 | * Either we still have to compute the base value for the | |
6604 | * total service time, and there seem to be the right | |
6605 | * conditions to do it, or we can lower the last base value | |
6606 | * computed. | |
db599f9e | 6607 | * |
2d31c684 | 6608 | * NOTE: (bfqd->tot_rq_in_driver == 1) means that there is no I/O |
db599f9e PV |
6609 | * request in flight, because this function is in the code |
6610 | * path that handles the completion of a request of bfqq, and, | |
6611 | * in particular, this function is executed before | |
2d31c684 | 6612 | * bfqd->tot_rq_in_driver is decremented in such a code path. |
2341d662 | 6613 | */ |
2d31c684 | 6614 | if ((bfqq->last_serv_time_ns == 0 && bfqd->tot_rq_in_driver == 1) || |
2341d662 | 6615 | tot_time_ns < bfqq->last_serv_time_ns) { |
58494c98 PV |
6616 | if (bfqq->last_serv_time_ns == 0) { |
6617 | /* | |
6618 | * Now we certainly have a base value: make sure we | |
6619 | * start trying injection. | |
6620 | */ | |
6621 | bfqq->inject_limit = max_t(unsigned int, 1, old_limit); | |
6622 | } | |
2341d662 | 6623 | bfqq->last_serv_time_ns = tot_time_ns; |
2d31c684 | 6624 | } else if (!bfqd->rqs_injected && bfqd->tot_rq_in_driver == 1) |
24792ad0 PV |
6625 | /* |
6626 | * No I/O injected and no request still in service in | |
6627 | * the drive: these are the exact conditions for | |
6628 | * computing the base value of the total service time | |
6629 | * for bfqq. So let's update this value, because it is | |
6630 | * rather variable. For example, it varies if the size | |
6631 | * or the spatial locality of the I/O requests in bfqq | |
6632 | * change. | |
6633 | */ | |
6634 | bfqq->last_serv_time_ns = tot_time_ns; | |
6635 | ||
2341d662 PV |
6636 | |
6637 | /* update complete, not waiting for any request completion any longer */ | |
6638 | bfqd->waited_rq = NULL; | |
23ed570a | 6639 | bfqd->rqs_injected = false; |
2341d662 PV |
6640 | } |
6641 | ||
a7877390 PV |
6642 | /* |
6643 | * Handle either a requeue or a finish for rq. The things to do are | |
6644 | * the same in both cases: all references to rq are to be dropped. In | |
6645 | * particular, rq is considered completed from the point of view of | |
6646 | * the scheduler. | |
6647 | */ | |
6648 | static void bfq_finish_requeue_request(struct request *rq) | |
aee69d78 | 6649 | { |
a7877390 | 6650 | struct bfq_queue *bfqq = RQ_BFQQ(rq); |
5bbf4e5a | 6651 | struct bfq_data *bfqd; |
a921c655 | 6652 | unsigned long flags; |
5bbf4e5a | 6653 | |
a7877390 PV |
6654 | /* |
6655 | * rq either is not associated with any icq, or is an already | |
6656 | * requeued request that has not (yet) been re-inserted into | |
6657 | * a bfq_queue. | |
6658 | */ | |
6659 | if (!rq->elv.icq || !bfqq) | |
5bbf4e5a CH |
6660 | return; |
6661 | ||
5bbf4e5a | 6662 | bfqd = bfqq->bfqd; |
aee69d78 | 6663 | |
e21b7a0b AA |
6664 | if (rq->rq_flags & RQF_STARTED) |
6665 | bfqg_stats_update_completion(bfqq_group(bfqq), | |
522a7775 OS |
6666 | rq->start_time_ns, |
6667 | rq->io_start_time_ns, | |
e21b7a0b | 6668 | rq->cmd_flags); |
aee69d78 | 6669 | |
a921c655 | 6670 | spin_lock_irqsave(&bfqd->lock, flags); |
aee69d78 | 6671 | if (likely(rq->rq_flags & RQF_STARTED)) { |
2341d662 PV |
6672 | if (rq == bfqd->waited_rq) |
6673 | bfq_update_inject_limit(bfqd, bfqq); | |
6674 | ||
aee69d78 | 6675 | bfq_completed_request(bfqq, bfqd); |
aee69d78 | 6676 | } |
a249ca7d JK |
6677 | bfqq_request_freed(bfqq); |
6678 | bfq_put_queue(bfqq); | |
f9506673 | 6679 | RQ_BIC(rq)->requests--; |
a921c655 | 6680 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 | 6681 | |
a7877390 PV |
6682 | /* |
6683 | * Reset private fields. In case of a requeue, this allows | |
6684 | * this function to correctly do nothing if it is spuriously | |
6685 | * invoked again on this same request (see the check at the | |
6686 | * beginning of the function). Probably, a better general | |
6687 | * design would be to prevent blk-mq from invoking the requeue | |
6688 | * or finish hooks of an elevator, for a request that is not | |
6689 | * referred by that elevator. | |
6690 | * | |
6691 | * Resetting the following fields would break the | |
6692 | * request-insertion logic if rq is re-inserted into a bfq | |
6693 | * internal queue, without a re-preparation. Here we assume | |
6694 | * that re-insertions of requeued requests, without | |
6695 | * re-preparation, can happen only for pass_through or at_head | |
6696 | * requests (which are not re-inserted into bfq internal | |
6697 | * queues). | |
6698 | */ | |
aee69d78 PV |
6699 | rq->elv.priv[0] = NULL; |
6700 | rq->elv.priv[1] = NULL; | |
6701 | } | |
6702 | ||
222ee581 CH |
6703 | static void bfq_finish_request(struct request *rq) |
6704 | { | |
6705 | bfq_finish_requeue_request(rq); | |
6706 | ||
6707 | if (rq->elv.icq) { | |
6708 | put_io_context(rq->elv.icq->ioc); | |
6709 | rq->elv.icq = NULL; | |
6710 | } | |
6711 | } | |
6712 | ||
36eca894 | 6713 | /* |
c92bddee PV |
6714 | * Removes the association between the current task and bfqq, assuming |
6715 | * that bic points to the bfq iocontext of the task. | |
36eca894 AA |
6716 | * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
6717 | * was the last process referring to that bfqq. | |
6718 | */ | |
6719 | static struct bfq_queue * | |
6720 | bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) | |
6721 | { | |
6722 | bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); | |
6723 | ||
6724 | if (bfqq_process_refs(bfqq) == 1) { | |
6725 | bfqq->pid = current->pid; | |
6726 | bfq_clear_bfqq_coop(bfqq); | |
6727 | bfq_clear_bfqq_split_coop(bfqq); | |
6728 | return bfqq; | |
6729 | } | |
6730 | ||
9778369a | 6731 | bic_set_bfqq(bic, NULL, true, bfqq->actuator_idx); |
36eca894 AA |
6732 | |
6733 | bfq_put_cooperator(bfqq); | |
6734 | ||
478de338 | 6735 | bfq_release_process_ref(bfqq->bfqd, bfqq); |
36eca894 AA |
6736 | return NULL; |
6737 | } | |
6738 | ||
6739 | static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd, | |
6740 | struct bfq_io_cq *bic, | |
6741 | struct bio *bio, | |
6742 | bool split, bool is_sync, | |
6743 | bool *new_queue) | |
6744 | { | |
9778369a PV |
6745 | unsigned int act_idx = bfq_actuator_index(bfqd, bio); |
6746 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, act_idx); | |
fd571df0 | 6747 | struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[act_idx]; |
36eca894 AA |
6748 | |
6749 | if (likely(bfqq && bfqq != &bfqd->oom_bfqq)) | |
6750 | return bfqq; | |
6751 | ||
6752 | if (new_queue) | |
6753 | *new_queue = true; | |
6754 | ||
6755 | if (bfqq) | |
6756 | bfq_put_queue(bfqq); | |
430a67f9 | 6757 | bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split); |
36eca894 | 6758 | |
9778369a | 6759 | bic_set_bfqq(bic, bfqq, is_sync, act_idx); |
e1b2324d | 6760 | if (split && is_sync) { |
a6123047 PV |
6761 | if ((bfqq_data->was_in_burst_list && bfqd->large_burst) || |
6762 | bfqq_data->saved_in_large_burst) | |
e1b2324d AA |
6763 | bfq_mark_bfqq_in_large_burst(bfqq); |
6764 | else { | |
6765 | bfq_clear_bfqq_in_large_burst(bfqq); | |
a6123047 | 6766 | if (bfqq_data->was_in_burst_list) |
99fead8d PV |
6767 | /* |
6768 | * If bfqq was in the current | |
6769 | * burst list before being | |
6770 | * merged, then we have to add | |
6771 | * it back. And we do not need | |
6772 | * to increase burst_size, as | |
6773 | * we did not decrement | |
6774 | * burst_size when we removed | |
6775 | * bfqq from the burst list as | |
6776 | * a consequence of a merge | |
6777 | * (see comments in | |
6778 | * bfq_put_queue). In this | |
6779 | * respect, it would be rather | |
6780 | * costly to know whether the | |
6781 | * current burst list is still | |
6782 | * the same burst list from | |
6783 | * which bfqq was removed on | |
6784 | * the merge. To avoid this | |
6785 | * cost, if bfqq was in a | |
6786 | * burst list, then we add | |
6787 | * bfqq to the current burst | |
6788 | * list without any further | |
6789 | * check. This can cause | |
6790 | * inappropriate insertions, | |
6791 | * but rarely enough to not | |
6792 | * harm the detection of large | |
6793 | * bursts significantly. | |
6794 | */ | |
e1b2324d AA |
6795 | hlist_add_head(&bfqq->burst_list_node, |
6796 | &bfqd->burst_list); | |
6797 | } | |
36eca894 | 6798 | bfqq->split_time = jiffies; |
e1b2324d | 6799 | } |
36eca894 AA |
6800 | |
6801 | return bfqq; | |
6802 | } | |
6803 | ||
aee69d78 | 6804 | /* |
18e5a57d PV |
6805 | * Only reset private fields. The actual request preparation will be |
6806 | * performed by bfq_init_rq, when rq is either inserted or merged. See | |
6807 | * comments on bfq_init_rq for the reason behind this delayed | |
6808 | * preparation. | |
aee69d78 | 6809 | */ |
5d9c305b | 6810 | static void bfq_prepare_request(struct request *rq) |
18e5a57d | 6811 | { |
87dd1d63 | 6812 | rq->elv.icq = ioc_find_get_icq(rq->q); |
5a9d041b | 6813 | |
18e5a57d PV |
6814 | /* |
6815 | * Regardless of whether we have an icq attached, we have to | |
6816 | * clear the scheduler pointers, as they might point to | |
6817 | * previously allocated bic/bfqq structs. | |
6818 | */ | |
6819 | rq->elv.priv[0] = rq->elv.priv[1] = NULL; | |
6820 | } | |
6821 | ||
6822 | /* | |
6823 | * If needed, init rq, allocate bfq data structures associated with | |
6824 | * rq, and increment reference counters in the destination bfq_queue | |
6825 | * for rq. Return the destination bfq_queue for rq, or NULL is rq is | |
6826 | * not associated with any bfq_queue. | |
6827 | * | |
6828 | * This function is invoked by the functions that perform rq insertion | |
6829 | * or merging. One may have expected the above preparation operations | |
6830 | * to be performed in bfq_prepare_request, and not delayed to when rq | |
6831 | * is inserted or merged. The rationale behind this delayed | |
6832 | * preparation is that, after the prepare_request hook is invoked for | |
6833 | * rq, rq may still be transformed into a request with no icq, i.e., a | |
6834 | * request not associated with any queue. No bfq hook is invoked to | |
636b8fe8 | 6835 | * signal this transformation. As a consequence, should these |
18e5a57d PV |
6836 | * preparation operations be performed when the prepare_request hook |
6837 | * is invoked, and should rq be transformed one moment later, bfq | |
6838 | * would end up in an inconsistent state, because it would have | |
6839 | * incremented some queue counters for an rq destined to | |
6840 | * transformation, without any chance to correctly lower these | |
6841 | * counters back. In contrast, no transformation can still happen for | |
6842 | * rq after rq has been inserted or merged. So, it is safe to execute | |
6843 | * these preparation operations when rq is finally inserted or merged. | |
6844 | */ | |
6845 | static struct bfq_queue *bfq_init_rq(struct request *rq) | |
aee69d78 | 6846 | { |
5bbf4e5a | 6847 | struct request_queue *q = rq->q; |
18e5a57d | 6848 | struct bio *bio = rq->bio; |
aee69d78 | 6849 | struct bfq_data *bfqd = q->elevator->elevator_data; |
9f210738 | 6850 | struct bfq_io_cq *bic; |
aee69d78 PV |
6851 | const int is_sync = rq_is_sync(rq); |
6852 | struct bfq_queue *bfqq; | |
36eca894 | 6853 | bool new_queue = false; |
13c931bd | 6854 | bool bfqq_already_existing = false, split = false; |
fd571df0 | 6855 | unsigned int a_idx = bfq_actuator_index(bfqd, bio); |
aee69d78 | 6856 | |
18e5a57d PV |
6857 | if (unlikely(!rq->elv.icq)) |
6858 | return NULL; | |
6859 | ||
72961c4e | 6860 | /* |
8ac2e43c | 6861 | * Assuming that RQ_BFQQ(rq) is set only if everything is set |
18e5a57d PV |
6862 | * for this rq. This holds true, because this function is |
6863 | * invoked only for insertion or merging, and, after such | |
6864 | * events, a request cannot be manipulated any longer before | |
6865 | * being removed from bfq. | |
72961c4e | 6866 | */ |
8ac2e43c KS |
6867 | if (RQ_BFQQ(rq)) |
6868 | return RQ_BFQQ(rq); | |
72961c4e | 6869 | |
9f210738 | 6870 | bic = icq_to_bic(rq->elv.icq); |
aee69d78 | 6871 | |
8c9ff1ad CIK |
6872 | bfq_check_ioprio_change(bic, bio); |
6873 | ||
e21b7a0b AA |
6874 | bfq_bic_update_cgroup(bic, bio); |
6875 | ||
36eca894 AA |
6876 | bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync, |
6877 | &new_queue); | |
6878 | ||
6879 | if (likely(!new_queue)) { | |
6880 | /* If the queue was seeky for too long, break it apart. */ | |
430a67f9 | 6881 | if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) && |
fd571df0 | 6882 | !bic->bfqq_data[a_idx].stably_merged) { |
8ef3fc3a | 6883 | struct bfq_queue *old_bfqq = bfqq; |
e1b2324d AA |
6884 | |
6885 | /* Update bic before losing reference to bfqq */ | |
6886 | if (bfq_bfqq_in_large_burst(bfqq)) | |
fd571df0 PV |
6887 | bic->bfqq_data[a_idx].saved_in_large_burst = |
6888 | true; | |
e1b2324d | 6889 | |
36eca894 | 6890 | bfqq = bfq_split_bfqq(bic, bfqq); |
6fa3e8d3 | 6891 | split = true; |
36eca894 | 6892 | |
8ef3fc3a | 6893 | if (!bfqq) { |
36eca894 AA |
6894 | bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, |
6895 | true, is_sync, | |
6896 | NULL); | |
a1795c2c KK |
6897 | if (unlikely(bfqq == &bfqd->oom_bfqq)) |
6898 | bfqq_already_existing = true; | |
6899 | } else | |
6900 | bfqq_already_existing = true; | |
6901 | ||
6902 | if (!bfqq_already_existing) { | |
8ef3fc3a PV |
6903 | bfqq->waker_bfqq = old_bfqq->waker_bfqq; |
6904 | bfqq->tentative_waker_bfqq = NULL; | |
6905 | ||
6906 | /* | |
6907 | * If the waker queue disappears, then | |
6908 | * new_bfqq->waker_bfqq must be | |
6909 | * reset. So insert new_bfqq into the | |
6910 | * woken_list of the waker. See | |
6911 | * bfq_check_waker for details. | |
6912 | */ | |
6913 | if (bfqq->waker_bfqq) | |
6914 | hlist_add_head(&bfqq->woken_list_node, | |
6915 | &bfqq->waker_bfqq->woken_list); | |
a1795c2c | 6916 | } |
36eca894 | 6917 | } |
aee69d78 PV |
6918 | } |
6919 | ||
98f04499 | 6920 | bfqq_request_allocated(bfqq); |
aee69d78 | 6921 | bfqq->ref++; |
f9506673 | 6922 | bic->requests++; |
aee69d78 PV |
6923 | bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d", |
6924 | rq, bfqq, bfqq->ref); | |
6925 | ||
6926 | rq->elv.priv[0] = bic; | |
6927 | rq->elv.priv[1] = bfqq; | |
6928 | ||
36eca894 AA |
6929 | /* |
6930 | * If a bfq_queue has only one process reference, it is owned | |
6931 | * by only this bic: we can then set bfqq->bic = bic. in | |
6932 | * addition, if the queue has also just been split, we have to | |
6933 | * resume its state. | |
6934 | */ | |
6935 | if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) { | |
6936 | bfqq->bic = bic; | |
6fa3e8d3 | 6937 | if (split) { |
36eca894 AA |
6938 | /* |
6939 | * The queue has just been split from a shared | |
6940 | * queue: restore the idle window and the | |
6941 | * possible weight raising period. | |
6942 | */ | |
13c931bd PV |
6943 | bfq_bfqq_resume_state(bfqq, bfqd, bic, |
6944 | bfqq_already_existing); | |
36eca894 AA |
6945 | } |
6946 | } | |
6947 | ||
84a74689 PV |
6948 | /* |
6949 | * Consider bfqq as possibly belonging to a burst of newly | |
6950 | * created queues only if: | |
6951 | * 1) A burst is actually happening (bfqd->burst_size > 0) | |
6952 | * or | |
6953 | * 2) There is no other active queue. In fact, if, in | |
6954 | * contrast, there are active queues not belonging to the | |
6955 | * possible burst bfqq may belong to, then there is no gain | |
6956 | * in considering bfqq as belonging to a burst, and | |
6957 | * therefore in not weight-raising bfqq. See comments on | |
6958 | * bfq_handle_burst(). | |
6959 | * | |
6960 | * This filtering also helps eliminating false positives, | |
6961 | * occurring when bfqq does not belong to an actual large | |
6962 | * burst, but some background task (e.g., a service) happens | |
6963 | * to trigger the creation of new queues very close to when | |
6964 | * bfqq and its possible companion queues are created. See | |
6965 | * comments on bfq_handle_burst() for further details also on | |
6966 | * this issue. | |
6967 | */ | |
6968 | if (unlikely(bfq_bfqq_just_created(bfqq) && | |
6969 | (bfqd->burst_size > 0 || | |
6970 | bfq_tot_busy_queues(bfqd) == 0))) | |
e1b2324d AA |
6971 | bfq_handle_burst(bfqd, bfqq); |
6972 | ||
18e5a57d | 6973 | return bfqq; |
aee69d78 PV |
6974 | } |
6975 | ||
2f95fa5c ZL |
6976 | static void |
6977 | bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
aee69d78 | 6978 | { |
aee69d78 PV |
6979 | enum bfqq_expiration reason; |
6980 | unsigned long flags; | |
6981 | ||
6982 | spin_lock_irqsave(&bfqd->lock, flags); | |
aee69d78 | 6983 | |
2f95fa5c ZL |
6984 | /* |
6985 | * Considering that bfqq may be in race, we should firstly check | |
6986 | * whether bfqq is in service before doing something on it. If | |
6987 | * the bfqq in race is not in service, it has already been expired | |
6988 | * through __bfq_bfqq_expire func and its wait_request flags has | |
6989 | * been cleared in __bfq_bfqd_reset_in_service func. | |
6990 | */ | |
aee69d78 PV |
6991 | if (bfqq != bfqd->in_service_queue) { |
6992 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
6993 | return; | |
6994 | } | |
6995 | ||
2f95fa5c ZL |
6996 | bfq_clear_bfqq_wait_request(bfqq); |
6997 | ||
aee69d78 PV |
6998 | if (bfq_bfqq_budget_timeout(bfqq)) |
6999 | /* | |
7000 | * Also here the queue can be safely expired | |
7001 | * for budget timeout without wasting | |
7002 | * guarantees | |
7003 | */ | |
7004 | reason = BFQQE_BUDGET_TIMEOUT; | |
7005 | else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) | |
7006 | /* | |
7007 | * The queue may not be empty upon timer expiration, | |
7008 | * because we may not disable the timer when the | |
7009 | * first request of the in-service queue arrives | |
7010 | * during disk idling. | |
7011 | */ | |
7012 | reason = BFQQE_TOO_IDLE; | |
7013 | else | |
7014 | goto schedule_dispatch; | |
7015 | ||
7016 | bfq_bfqq_expire(bfqd, bfqq, true, reason); | |
7017 | ||
7018 | schedule_dispatch: | |
aee69d78 | 7019 | bfq_schedule_dispatch(bfqd); |
181490d5 | 7020 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
7021 | } |
7022 | ||
7023 | /* | |
7024 | * Handler of the expiration of the timer running if the in-service queue | |
7025 | * is idling inside its time slice. | |
7026 | */ | |
7027 | static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer) | |
7028 | { | |
7029 | struct bfq_data *bfqd = container_of(timer, struct bfq_data, | |
7030 | idle_slice_timer); | |
7031 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
7032 | ||
7033 | /* | |
7034 | * Theoretical race here: the in-service queue can be NULL or | |
7035 | * different from the queue that was idling if a new request | |
7036 | * arrives for the current queue and there is a full dispatch | |
7037 | * cycle that changes the in-service queue. This can hardly | |
7038 | * happen, but in the worst case we just expire a queue too | |
7039 | * early. | |
7040 | */ | |
7041 | if (bfqq) | |
2f95fa5c | 7042 | bfq_idle_slice_timer_body(bfqd, bfqq); |
aee69d78 PV |
7043 | |
7044 | return HRTIMER_NORESTART; | |
7045 | } | |
7046 | ||
7047 | static void __bfq_put_async_bfqq(struct bfq_data *bfqd, | |
7048 | struct bfq_queue **bfqq_ptr) | |
7049 | { | |
7050 | struct bfq_queue *bfqq = *bfqq_ptr; | |
7051 | ||
7052 | bfq_log(bfqd, "put_async_bfqq: %p", bfqq); | |
7053 | if (bfqq) { | |
e21b7a0b AA |
7054 | bfq_bfqq_move(bfqd, bfqq, bfqd->root_group); |
7055 | ||
aee69d78 PV |
7056 | bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
7057 | bfqq, bfqq->ref); | |
7058 | bfq_put_queue(bfqq); | |
7059 | *bfqq_ptr = NULL; | |
7060 | } | |
7061 | } | |
7062 | ||
7063 | /* | |
e21b7a0b AA |
7064 | * Release all the bfqg references to its async queues. If we are |
7065 | * deallocating the group these queues may still contain requests, so | |
7066 | * we reparent them to the root cgroup (i.e., the only one that will | |
7067 | * exist for sure until all the requests on a device are gone). | |
aee69d78 | 7068 | */ |
ea25da48 | 7069 | void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
aee69d78 | 7070 | { |
8b7fd741 | 7071 | int i, j, k; |
aee69d78 | 7072 | |
8b7fd741 DZ |
7073 | for (k = 0; k < bfqd->num_actuators; k++) { |
7074 | for (i = 0; i < 2; i++) | |
7075 | for (j = 0; j < IOPRIO_NR_LEVELS; j++) | |
7076 | __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j][k]); | |
aee69d78 | 7077 | |
8b7fd741 DZ |
7078 | __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq[k]); |
7079 | } | |
aee69d78 PV |
7080 | } |
7081 | ||
f0635b8a JA |
7082 | /* |
7083 | * See the comments on bfq_limit_depth for the purpose of | |
483b7bf2 | 7084 | * the depths set in the function. Return minimum shallow depth we'll use. |
f0635b8a | 7085 | */ |
76f1df88 | 7086 | static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt) |
f0635b8a | 7087 | { |
44dfa279 | 7088 | unsigned int depth = 1U << bt->sb.shift; |
483b7bf2 | 7089 | |
44dfa279 | 7090 | bfqd->full_depth_shift = bt->sb.shift; |
f0635b8a JA |
7091 | /* |
7092 | * In-word depths if no bfq_queue is being weight-raised: | |
7093 | * leaving 25% of tags only for sync reads. | |
7094 | * | |
7095 | * In next formulas, right-shift the value | |
bd7d4ef6 JA |
7096 | * (1U<<bt->sb.shift), instead of computing directly |
7097 | * (1U<<(bt->sb.shift - something)), to be robust against | |
7098 | * any possible value of bt->sb.shift, without having to | |
f0635b8a JA |
7099 | * limit 'something'. |
7100 | */ | |
7101 | /* no more than 50% of tags for async I/O */ | |
44dfa279 | 7102 | bfqd->word_depths[0][0] = max(depth >> 1, 1U); |
f0635b8a JA |
7103 | /* |
7104 | * no more than 75% of tags for sync writes (25% extra tags | |
7105 | * w.r.t. async I/O, to prevent async I/O from starving sync | |
7106 | * writes) | |
7107 | */ | |
44dfa279 | 7108 | bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U); |
f0635b8a JA |
7109 | |
7110 | /* | |
7111 | * In-word depths in case some bfq_queue is being weight- | |
7112 | * raised: leaving ~63% of tags for sync reads. This is the | |
7113 | * highest percentage for which, in our tests, application | |
7114 | * start-up times didn't suffer from any regression due to tag | |
7115 | * shortage. | |
7116 | */ | |
7117 | /* no more than ~18% of tags for async I/O */ | |
44dfa279 | 7118 | bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U); |
f0635b8a | 7119 | /* no more than ~37% of tags for sync writes (~20% extra tags) */ |
44dfa279 | 7120 | bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U); |
f0635b8a JA |
7121 | } |
7122 | ||
77f1e0a5 | 7123 | static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx) |
f0635b8a JA |
7124 | { |
7125 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
7126 | struct blk_mq_tags *tags = hctx->sched_tags; | |
7127 | ||
76f1df88 JK |
7128 | bfq_update_depths(bfqd, &tags->bitmap_tags); |
7129 | sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1); | |
77f1e0a5 JA |
7130 | } |
7131 | ||
7132 | static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index) | |
7133 | { | |
7134 | bfq_depth_updated(hctx); | |
f0635b8a JA |
7135 | return 0; |
7136 | } | |
7137 | ||
aee69d78 PV |
7138 | static void bfq_exit_queue(struct elevator_queue *e) |
7139 | { | |
7140 | struct bfq_data *bfqd = e->elevator_data; | |
7141 | struct bfq_queue *bfqq, *n; | |
3e49c1e4 | 7142 | unsigned int actuator; |
aee69d78 PV |
7143 | |
7144 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
7145 | ||
7146 | spin_lock_irq(&bfqd->lock); | |
7147 | list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) | |
e21b7a0b | 7148 | bfq_deactivate_bfqq(bfqd, bfqq, false, false); |
aee69d78 PV |
7149 | spin_unlock_irq(&bfqd->lock); |
7150 | ||
3e49c1e4 BVA |
7151 | for (actuator = 0; actuator < bfqd->num_actuators; actuator++) |
7152 | WARN_ON_ONCE(bfqd->rq_in_driver[actuator]); | |
7153 | WARN_ON_ONCE(bfqd->tot_rq_in_driver); | |
3e49c1e4 | 7154 | |
aee69d78 PV |
7155 | hrtimer_cancel(&bfqd->idle_slice_timer); |
7156 | ||
0d52af59 PV |
7157 | /* release oom-queue reference to root group */ |
7158 | bfqg_and_blkg_put(bfqd->root_group); | |
7159 | ||
4d8340d0 | 7160 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
40e4996e | 7161 | blkcg_deactivate_policy(bfqd->queue->disk, &blkcg_policy_bfq); |
e21b7a0b AA |
7162 | #else |
7163 | spin_lock_irq(&bfqd->lock); | |
7164 | bfq_put_async_queues(bfqd, bfqd->root_group); | |
7165 | kfree(bfqd->root_group); | |
7166 | spin_unlock_irq(&bfqd->lock); | |
7167 | #endif | |
7168 | ||
b96f3cab | 7169 | blk_stat_disable_accounting(bfqd->queue); |
671fae5e | 7170 | clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags); |
04aad37b | 7171 | wbt_enable_default(bfqd->queue->disk); |
e92bc4cd | 7172 | |
aee69d78 PV |
7173 | kfree(bfqd); |
7174 | } | |
7175 | ||
e21b7a0b AA |
7176 | static void bfq_init_root_group(struct bfq_group *root_group, |
7177 | struct bfq_data *bfqd) | |
7178 | { | |
7179 | int i; | |
7180 | ||
7181 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
7182 | root_group->entity.parent = NULL; | |
7183 | root_group->my_entity = NULL; | |
7184 | root_group->bfqd = bfqd; | |
7185 | #endif | |
36eca894 | 7186 | root_group->rq_pos_tree = RB_ROOT; |
e21b7a0b AA |
7187 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
7188 | root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; | |
7189 | root_group->sched_data.bfq_class_idle_last_service = jiffies; | |
7190 | } | |
7191 | ||
aee69d78 PV |
7192 | static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
7193 | { | |
7194 | struct bfq_data *bfqd; | |
7195 | struct elevator_queue *eq; | |
4fdb3b9f FG |
7196 | unsigned int i; |
7197 | struct blk_independent_access_ranges *ia_ranges = q->disk->ia_ranges; | |
aee69d78 PV |
7198 | |
7199 | eq = elevator_alloc(q, e); | |
7200 | if (!eq) | |
7201 | return -ENOMEM; | |
7202 | ||
7203 | bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); | |
7204 | if (!bfqd) { | |
7205 | kobject_put(&eq->kobj); | |
7206 | return -ENOMEM; | |
7207 | } | |
7208 | eq->elevator_data = bfqd; | |
7209 | ||
0d945c1f | 7210 | spin_lock_irq(&q->queue_lock); |
e21b7a0b | 7211 | q->elevator = eq; |
0d945c1f | 7212 | spin_unlock_irq(&q->queue_lock); |
e21b7a0b | 7213 | |
aee69d78 PV |
7214 | /* |
7215 | * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. | |
7216 | * Grab a permanent reference to it, so that the normal code flow | |
7217 | * will not attempt to free it. | |
9778369a PV |
7218 | * Set zero as actuator index: we will pretend that |
7219 | * all I/O requests are for the same actuator. | |
aee69d78 | 7220 | */ |
9778369a | 7221 | bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0, 0); |
aee69d78 PV |
7222 | bfqd->oom_bfqq.ref++; |
7223 | bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO; | |
7224 | bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE; | |
7225 | bfqd->oom_bfqq.entity.new_weight = | |
7226 | bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio); | |
e1b2324d AA |
7227 | |
7228 | /* oom_bfqq does not participate to bursts */ | |
7229 | bfq_clear_bfqq_just_created(&bfqd->oom_bfqq); | |
7230 | ||
aee69d78 PV |
7231 | /* |
7232 | * Trigger weight initialization, according to ioprio, at the | |
7233 | * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio | |
7234 | * class won't be changed any more. | |
7235 | */ | |
7236 | bfqd->oom_bfqq.entity.prio_changed = 1; | |
7237 | ||
7238 | bfqd->queue = q; | |
7239 | ||
4fdb3b9f | 7240 | bfqd->num_actuators = 1; |
9778369a | 7241 | /* |
4fdb3b9f FG |
7242 | * If the disk supports multiple actuators, copy independent |
7243 | * access ranges from the request queue structure. | |
9778369a | 7244 | */ |
4fdb3b9f FG |
7245 | spin_lock_irq(&q->queue_lock); |
7246 | if (ia_ranges) { | |
7247 | /* | |
7248 | * Check if the disk ia_ranges size exceeds the current bfq | |
7249 | * actuator limit. | |
7250 | */ | |
7251 | if (ia_ranges->nr_ia_ranges > BFQ_MAX_ACTUATORS) { | |
7252 | pr_crit("nr_ia_ranges higher than act limit: iars=%d, max=%d.\n", | |
7253 | ia_ranges->nr_ia_ranges, BFQ_MAX_ACTUATORS); | |
7254 | pr_crit("Falling back to single actuator mode.\n"); | |
7255 | } else { | |
7256 | bfqd->num_actuators = ia_ranges->nr_ia_ranges; | |
7257 | ||
7258 | for (i = 0; i < bfqd->num_actuators; i++) { | |
7259 | bfqd->sector[i] = ia_ranges->ia_range[i].sector; | |
7260 | bfqd->nr_sectors[i] = | |
7261 | ia_ranges->ia_range[i].nr_sectors; | |
7262 | } | |
7263 | } | |
7264 | } | |
7265 | ||
7266 | /* Otherwise use single-actuator dev info */ | |
7267 | if (bfqd->num_actuators == 1) { | |
7268 | bfqd->sector[0] = 0; | |
7269 | bfqd->nr_sectors[0] = get_capacity(q->disk); | |
7270 | } | |
7271 | spin_unlock_irq(&q->queue_lock); | |
9778369a | 7272 | |
e21b7a0b | 7273 | INIT_LIST_HEAD(&bfqd->dispatch); |
aee69d78 PV |
7274 | |
7275 | hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC, | |
7276 | HRTIMER_MODE_REL); | |
7277 | bfqd->idle_slice_timer.function = bfq_idle_slice_timer; | |
7278 | ||
fb53ac6c | 7279 | bfqd->queue_weights_tree = RB_ROOT_CACHED; |
1eb20620 | 7280 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
ba7aeae5 | 7281 | bfqd->num_groups_with_pending_reqs = 0; |
1eb20620 | 7282 | #endif |
1de0c4cd | 7283 | |
2d31c684 DZ |
7284 | INIT_LIST_HEAD(&bfqd->active_list[0]); |
7285 | INIT_LIST_HEAD(&bfqd->active_list[1]); | |
aee69d78 | 7286 | INIT_LIST_HEAD(&bfqd->idle_list); |
e1b2324d | 7287 | INIT_HLIST_HEAD(&bfqd->burst_list); |
aee69d78 PV |
7288 | |
7289 | bfqd->hw_tag = -1; | |
8cacc5ab | 7290 | bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue); |
aee69d78 PV |
7291 | |
7292 | bfqd->bfq_max_budget = bfq_default_max_budget; | |
7293 | ||
7294 | bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; | |
7295 | bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; | |
7296 | bfqd->bfq_back_max = bfq_back_max; | |
7297 | bfqd->bfq_back_penalty = bfq_back_penalty; | |
7298 | bfqd->bfq_slice_idle = bfq_slice_idle; | |
aee69d78 PV |
7299 | bfqd->bfq_timeout = bfq_timeout; |
7300 | ||
e1b2324d AA |
7301 | bfqd->bfq_large_burst_thresh = 8; |
7302 | bfqd->bfq_burst_interval = msecs_to_jiffies(180); | |
7303 | ||
44e44a1b PV |
7304 | bfqd->low_latency = true; |
7305 | ||
7306 | /* | |
7307 | * Trade-off between responsiveness and fairness. | |
7308 | */ | |
7309 | bfqd->bfq_wr_coeff = 30; | |
77b7dcea | 7310 | bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300); |
44e44a1b PV |
7311 | bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000); |
7312 | bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500); | |
77b7dcea PV |
7313 | bfqd->bfq_wr_max_softrt_rate = 7000; /* |
7314 | * Approximate rate required | |
7315 | * to playback or record a | |
7316 | * high-definition compressed | |
7317 | * video. | |
7318 | */ | |
cfd69712 | 7319 | bfqd->wr_busy_queues = 0; |
44e44a1b PV |
7320 | |
7321 | /* | |
e24f1c24 PV |
7322 | * Begin by assuming, optimistically, that the device peak |
7323 | * rate is equal to 2/3 of the highest reference rate. | |
44e44a1b | 7324 | */ |
e24f1c24 PV |
7325 | bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] * |
7326 | ref_wr_duration[blk_queue_nonrot(bfqd->queue)]; | |
7327 | bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3; | |
44e44a1b | 7328 | |
2d31c684 DZ |
7329 | /* see comments on the definition of next field inside bfq_data */ |
7330 | bfqd->actuator_load_threshold = 4; | |
7331 | ||
aee69d78 | 7332 | spin_lock_init(&bfqd->lock); |
aee69d78 | 7333 | |
e21b7a0b AA |
7334 | /* |
7335 | * The invocation of the next bfq_create_group_hierarchy | |
7336 | * function is the head of a chain of function calls | |
7337 | * (bfq_create_group_hierarchy->blkcg_activate_policy-> | |
7338 | * blk_mq_freeze_queue) that may lead to the invocation of the | |
7339 | * has_work hook function. For this reason, | |
7340 | * bfq_create_group_hierarchy is invoked only after all | |
7341 | * scheduler data has been initialized, apart from the fields | |
7342 | * that can be initialized only after invoking | |
7343 | * bfq_create_group_hierarchy. This, in particular, enables | |
7344 | * has_work to correctly return false. Of course, to avoid | |
7345 | * other inconsistencies, the blk-mq stack must then refrain | |
7346 | * from invoking further scheduler hooks before this init | |
7347 | * function is finished. | |
7348 | */ | |
7349 | bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node); | |
7350 | if (!bfqd->root_group) | |
7351 | goto out_free; | |
7352 | bfq_init_root_group(bfqd->root_group, bfqd); | |
7353 | bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group); | |
7354 | ||
4d337ceb ML |
7355 | /* We dispatch from request queue wide instead of hw queue */ |
7356 | blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q); | |
7357 | ||
671fae5e | 7358 | set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags); |
04aad37b | 7359 | wbt_disable_default(q->disk); |
b96f3cab BVA |
7360 | blk_stat_enable_accounting(q); |
7361 | ||
aee69d78 | 7362 | return 0; |
e21b7a0b AA |
7363 | |
7364 | out_free: | |
7365 | kfree(bfqd); | |
7366 | kobject_put(&eq->kobj); | |
7367 | return -ENOMEM; | |
aee69d78 PV |
7368 | } |
7369 | ||
7370 | static void bfq_slab_kill(void) | |
7371 | { | |
7372 | kmem_cache_destroy(bfq_pool); | |
7373 | } | |
7374 | ||
7375 | static int __init bfq_slab_setup(void) | |
7376 | { | |
7377 | bfq_pool = KMEM_CACHE(bfq_queue, 0); | |
7378 | if (!bfq_pool) | |
7379 | return -ENOMEM; | |
7380 | return 0; | |
7381 | } | |
7382 | ||
7383 | static ssize_t bfq_var_show(unsigned int var, char *page) | |
7384 | { | |
7385 | return sprintf(page, "%u\n", var); | |
7386 | } | |
7387 | ||
2f79136b | 7388 | static int bfq_var_store(unsigned long *var, const char *page) |
aee69d78 PV |
7389 | { |
7390 | unsigned long new_val; | |
7391 | int ret = kstrtoul(page, 10, &new_val); | |
7392 | ||
2f79136b BVA |
7393 | if (ret) |
7394 | return ret; | |
7395 | *var = new_val; | |
7396 | return 0; | |
aee69d78 PV |
7397 | } |
7398 | ||
7399 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
7400 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
7401 | { \ | |
7402 | struct bfq_data *bfqd = e->elevator_data; \ | |
7403 | u64 __data = __VAR; \ | |
7404 | if (__CONV == 1) \ | |
7405 | __data = jiffies_to_msecs(__data); \ | |
7406 | else if (__CONV == 2) \ | |
7407 | __data = div_u64(__data, NSEC_PER_MSEC); \ | |
7408 | return bfq_var_show(__data, (page)); \ | |
7409 | } | |
7410 | SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2); | |
7411 | SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2); | |
7412 | SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); | |
7413 | SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); | |
7414 | SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2); | |
7415 | SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); | |
7416 | SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1); | |
7417 | SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0); | |
44e44a1b | 7418 | SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
aee69d78 PV |
7419 | #undef SHOW_FUNCTION |
7420 | ||
7421 | #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \ | |
7422 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
7423 | { \ | |
7424 | struct bfq_data *bfqd = e->elevator_data; \ | |
7425 | u64 __data = __VAR; \ | |
7426 | __data = div_u64(__data, NSEC_PER_USEC); \ | |
7427 | return bfq_var_show(__data, (page)); \ | |
7428 | } | |
7429 | USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle); | |
7430 | #undef USEC_SHOW_FUNCTION | |
7431 | ||
7432 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
7433 | static ssize_t \ | |
7434 | __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
7435 | { \ | |
7436 | struct bfq_data *bfqd = e->elevator_data; \ | |
1530486c | 7437 | unsigned long __data, __min = (MIN), __max = (MAX); \ |
2f79136b BVA |
7438 | int ret; \ |
7439 | \ | |
7440 | ret = bfq_var_store(&__data, (page)); \ | |
7441 | if (ret) \ | |
7442 | return ret; \ | |
1530486c BVA |
7443 | if (__data < __min) \ |
7444 | __data = __min; \ | |
7445 | else if (__data > __max) \ | |
7446 | __data = __max; \ | |
aee69d78 PV |
7447 | if (__CONV == 1) \ |
7448 | *(__PTR) = msecs_to_jiffies(__data); \ | |
7449 | else if (__CONV == 2) \ | |
7450 | *(__PTR) = (u64)__data * NSEC_PER_MSEC; \ | |
7451 | else \ | |
7452 | *(__PTR) = __data; \ | |
235f8da1 | 7453 | return count; \ |
aee69d78 PV |
7454 | } |
7455 | STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, | |
7456 | INT_MAX, 2); | |
7457 | STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, | |
7458 | INT_MAX, 2); | |
7459 | STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); | |
7460 | STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, | |
7461 | INT_MAX, 0); | |
7462 | STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2); | |
7463 | #undef STORE_FUNCTION | |
7464 | ||
7465 | #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
7466 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\ | |
7467 | { \ | |
7468 | struct bfq_data *bfqd = e->elevator_data; \ | |
1530486c | 7469 | unsigned long __data, __min = (MIN), __max = (MAX); \ |
2f79136b BVA |
7470 | int ret; \ |
7471 | \ | |
7472 | ret = bfq_var_store(&__data, (page)); \ | |
7473 | if (ret) \ | |
7474 | return ret; \ | |
1530486c BVA |
7475 | if (__data < __min) \ |
7476 | __data = __min; \ | |
7477 | else if (__data > __max) \ | |
7478 | __data = __max; \ | |
aee69d78 | 7479 | *(__PTR) = (u64)__data * NSEC_PER_USEC; \ |
235f8da1 | 7480 | return count; \ |
aee69d78 PV |
7481 | } |
7482 | USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0, | |
7483 | UINT_MAX); | |
7484 | #undef USEC_STORE_FUNCTION | |
7485 | ||
aee69d78 PV |
7486 | static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
7487 | const char *page, size_t count) | |
7488 | { | |
7489 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
7490 | unsigned long __data; |
7491 | int ret; | |
235f8da1 | 7492 | |
2f79136b BVA |
7493 | ret = bfq_var_store(&__data, (page)); |
7494 | if (ret) | |
7495 | return ret; | |
aee69d78 PV |
7496 | |
7497 | if (__data == 0) | |
ab0e43e9 | 7498 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 PV |
7499 | else { |
7500 | if (__data > INT_MAX) | |
7501 | __data = INT_MAX; | |
7502 | bfqd->bfq_max_budget = __data; | |
7503 | } | |
7504 | ||
7505 | bfqd->bfq_user_max_budget = __data; | |
7506 | ||
235f8da1 | 7507 | return count; |
aee69d78 PV |
7508 | } |
7509 | ||
7510 | /* | |
7511 | * Leaving this name to preserve name compatibility with cfq | |
7512 | * parameters, but this timeout is used for both sync and async. | |
7513 | */ | |
7514 | static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, | |
7515 | const char *page, size_t count) | |
7516 | { | |
7517 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
7518 | unsigned long __data; |
7519 | int ret; | |
235f8da1 | 7520 | |
2f79136b BVA |
7521 | ret = bfq_var_store(&__data, (page)); |
7522 | if (ret) | |
7523 | return ret; | |
aee69d78 PV |
7524 | |
7525 | if (__data < 1) | |
7526 | __data = 1; | |
7527 | else if (__data > INT_MAX) | |
7528 | __data = INT_MAX; | |
7529 | ||
7530 | bfqd->bfq_timeout = msecs_to_jiffies(__data); | |
7531 | if (bfqd->bfq_user_max_budget == 0) | |
ab0e43e9 | 7532 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 | 7533 | |
235f8da1 | 7534 | return count; |
aee69d78 PV |
7535 | } |
7536 | ||
7537 | static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e, | |
7538 | const char *page, size_t count) | |
7539 | { | |
7540 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
7541 | unsigned long __data; |
7542 | int ret; | |
235f8da1 | 7543 | |
2f79136b BVA |
7544 | ret = bfq_var_store(&__data, (page)); |
7545 | if (ret) | |
7546 | return ret; | |
aee69d78 PV |
7547 | |
7548 | if (__data > 1) | |
7549 | __data = 1; | |
7550 | if (!bfqd->strict_guarantees && __data == 1 | |
7551 | && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC) | |
7552 | bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC; | |
7553 | ||
7554 | bfqd->strict_guarantees = __data; | |
7555 | ||
235f8da1 | 7556 | return count; |
aee69d78 PV |
7557 | } |
7558 | ||
44e44a1b PV |
7559 | static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
7560 | const char *page, size_t count) | |
7561 | { | |
7562 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
7563 | unsigned long __data; |
7564 | int ret; | |
235f8da1 | 7565 | |
2f79136b BVA |
7566 | ret = bfq_var_store(&__data, (page)); |
7567 | if (ret) | |
7568 | return ret; | |
44e44a1b PV |
7569 | |
7570 | if (__data > 1) | |
7571 | __data = 1; | |
7572 | if (__data == 0 && bfqd->low_latency != 0) | |
7573 | bfq_end_wr(bfqd); | |
7574 | bfqd->low_latency = __data; | |
7575 | ||
235f8da1 | 7576 | return count; |
44e44a1b PV |
7577 | } |
7578 | ||
aee69d78 PV |
7579 | #define BFQ_ATTR(name) \ |
7580 | __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store) | |
7581 | ||
7582 | static struct elv_fs_entry bfq_attrs[] = { | |
7583 | BFQ_ATTR(fifo_expire_sync), | |
7584 | BFQ_ATTR(fifo_expire_async), | |
7585 | BFQ_ATTR(back_seek_max), | |
7586 | BFQ_ATTR(back_seek_penalty), | |
7587 | BFQ_ATTR(slice_idle), | |
7588 | BFQ_ATTR(slice_idle_us), | |
7589 | BFQ_ATTR(max_budget), | |
7590 | BFQ_ATTR(timeout_sync), | |
7591 | BFQ_ATTR(strict_guarantees), | |
44e44a1b | 7592 | BFQ_ATTR(low_latency), |
aee69d78 PV |
7593 | __ATTR_NULL |
7594 | }; | |
7595 | ||
7596 | static struct elevator_type iosched_bfq_mq = { | |
f9cd4bfe | 7597 | .ops = { |
a52a69ea | 7598 | .limit_depth = bfq_limit_depth, |
5bbf4e5a | 7599 | .prepare_request = bfq_prepare_request, |
a7877390 | 7600 | .requeue_request = bfq_finish_requeue_request, |
222ee581 | 7601 | .finish_request = bfq_finish_request, |
aee69d78 PV |
7602 | .exit_icq = bfq_exit_icq, |
7603 | .insert_requests = bfq_insert_requests, | |
7604 | .dispatch_request = bfq_dispatch_request, | |
7605 | .next_request = elv_rb_latter_request, | |
7606 | .former_request = elv_rb_former_request, | |
7607 | .allow_merge = bfq_allow_bio_merge, | |
7608 | .bio_merge = bfq_bio_merge, | |
7609 | .request_merge = bfq_request_merge, | |
7610 | .requests_merged = bfq_requests_merged, | |
7611 | .request_merged = bfq_request_merged, | |
7612 | .has_work = bfq_has_work, | |
77f1e0a5 | 7613 | .depth_updated = bfq_depth_updated, |
f0635b8a | 7614 | .init_hctx = bfq_init_hctx, |
aee69d78 PV |
7615 | .init_sched = bfq_init_queue, |
7616 | .exit_sched = bfq_exit_queue, | |
7617 | }, | |
7618 | ||
aee69d78 PV |
7619 | .icq_size = sizeof(struct bfq_io_cq), |
7620 | .icq_align = __alignof__(struct bfq_io_cq), | |
7621 | .elevator_attrs = bfq_attrs, | |
7622 | .elevator_name = "bfq", | |
7623 | .elevator_owner = THIS_MODULE, | |
7624 | }; | |
26b4cf24 | 7625 | MODULE_ALIAS("bfq-iosched"); |
aee69d78 PV |
7626 | |
7627 | static int __init bfq_init(void) | |
7628 | { | |
7629 | int ret; | |
7630 | ||
e21b7a0b AA |
7631 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
7632 | ret = blkcg_policy_register(&blkcg_policy_bfq); | |
7633 | if (ret) | |
7634 | return ret; | |
7635 | #endif | |
7636 | ||
aee69d78 PV |
7637 | ret = -ENOMEM; |
7638 | if (bfq_slab_setup()) | |
7639 | goto err_pol_unreg; | |
7640 | ||
44e44a1b PV |
7641 | /* |
7642 | * Times to load large popular applications for the typical | |
7643 | * systems installed on the reference devices (see the | |
e24f1c24 PV |
7644 | * comments before the definition of the next |
7645 | * array). Actually, we use slightly lower values, as the | |
44e44a1b PV |
7646 | * estimated peak rate tends to be smaller than the actual |
7647 | * peak rate. The reason for this last fact is that estimates | |
7648 | * are computed over much shorter time intervals than the long | |
7649 | * intervals typically used for benchmarking. Why? First, to | |
7650 | * adapt more quickly to variations. Second, because an I/O | |
7651 | * scheduler cannot rely on a peak-rate-evaluation workload to | |
7652 | * be run for a long time. | |
7653 | */ | |
e24f1c24 PV |
7654 | ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */ |
7655 | ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */ | |
44e44a1b | 7656 | |
aee69d78 PV |
7657 | ret = elv_register(&iosched_bfq_mq); |
7658 | if (ret) | |
37dcd657 | 7659 | goto slab_kill; |
aee69d78 PV |
7660 | |
7661 | return 0; | |
7662 | ||
37dcd657 | 7663 | slab_kill: |
7664 | bfq_slab_kill(); | |
aee69d78 | 7665 | err_pol_unreg: |
e21b7a0b AA |
7666 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
7667 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
7668 | #endif | |
aee69d78 PV |
7669 | return ret; |
7670 | } | |
7671 | ||
7672 | static void __exit bfq_exit(void) | |
7673 | { | |
7674 | elv_unregister(&iosched_bfq_mq); | |
e21b7a0b AA |
7675 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
7676 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
7677 | #endif | |
aee69d78 PV |
7678 | bfq_slab_kill(); |
7679 | } | |
7680 | ||
7681 | module_init(bfq_init); | |
7682 | module_exit(bfq_exit); | |
7683 | ||
7684 | MODULE_AUTHOR("Paolo Valente"); | |
7685 | MODULE_LICENSE("GPL"); | |
7686 | MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler"); |