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/elevator.h> |
121 | #include <linux/ktime.h> | |
122 | #include <linux/rbtree.h> | |
123 | #include <linux/ioprio.h> | |
124 | #include <linux/sbitmap.h> | |
125 | #include <linux/delay.h> | |
126 | ||
127 | #include "blk.h" | |
128 | #include "blk-mq.h" | |
129 | #include "blk-mq-tag.h" | |
130 | #include "blk-mq-sched.h" | |
ea25da48 | 131 | #include "bfq-iosched.h" |
b5dc5d4d | 132 | #include "blk-wbt.h" |
aee69d78 | 133 | |
ea25da48 PV |
134 | #define BFQ_BFQQ_FNS(name) \ |
135 | void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ | |
136 | { \ | |
137 | __set_bit(BFQQF_##name, &(bfqq)->flags); \ | |
138 | } \ | |
139 | void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ | |
140 | { \ | |
141 | __clear_bit(BFQQF_##name, &(bfqq)->flags); \ | |
142 | } \ | |
143 | int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ | |
144 | { \ | |
145 | return test_bit(BFQQF_##name, &(bfqq)->flags); \ | |
44e44a1b PV |
146 | } |
147 | ||
ea25da48 PV |
148 | BFQ_BFQQ_FNS(just_created); |
149 | BFQ_BFQQ_FNS(busy); | |
150 | BFQ_BFQQ_FNS(wait_request); | |
151 | BFQ_BFQQ_FNS(non_blocking_wait_rq); | |
152 | BFQ_BFQQ_FNS(fifo_expire); | |
d5be3fef | 153 | BFQ_BFQQ_FNS(has_short_ttime); |
ea25da48 PV |
154 | BFQ_BFQQ_FNS(sync); |
155 | BFQ_BFQQ_FNS(IO_bound); | |
156 | BFQ_BFQQ_FNS(in_large_burst); | |
157 | BFQ_BFQQ_FNS(coop); | |
158 | BFQ_BFQQ_FNS(split_coop); | |
159 | BFQ_BFQQ_FNS(softrt_update); | |
13a857a4 | 160 | BFQ_BFQQ_FNS(has_waker); |
ea25da48 | 161 | #undef BFQ_BFQQ_FNS \ |
aee69d78 | 162 | |
ea25da48 PV |
163 | /* Expiration time of sync (0) and async (1) requests, in ns. */ |
164 | static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 }; | |
aee69d78 | 165 | |
ea25da48 PV |
166 | /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */ |
167 | static const int bfq_back_max = 16 * 1024; | |
aee69d78 | 168 | |
ea25da48 PV |
169 | /* Penalty of a backwards seek, in number of sectors. */ |
170 | static const int bfq_back_penalty = 2; | |
e21b7a0b | 171 | |
ea25da48 PV |
172 | /* Idling period duration, in ns. */ |
173 | static u64 bfq_slice_idle = NSEC_PER_SEC / 125; | |
aee69d78 | 174 | |
ea25da48 PV |
175 | /* Minimum number of assigned budgets for which stats are safe to compute. */ |
176 | static const int bfq_stats_min_budgets = 194; | |
aee69d78 | 177 | |
ea25da48 PV |
178 | /* Default maximum budget values, in sectors and number of requests. */ |
179 | static const int bfq_default_max_budget = 16 * 1024; | |
e21b7a0b | 180 | |
ea25da48 | 181 | /* |
d5801088 PV |
182 | * When a sync request is dispatched, the queue that contains that |
183 | * request, and all the ancestor entities of that queue, are charged | |
636b8fe8 | 184 | * with the number of sectors of the request. In contrast, if the |
d5801088 PV |
185 | * request is async, then the queue and its ancestor entities are |
186 | * charged with the number of sectors of the request, multiplied by | |
187 | * the factor below. This throttles the bandwidth for async I/O, | |
188 | * w.r.t. to sync I/O, and it is done to counter the tendency of async | |
189 | * writes to steal I/O throughput to reads. | |
190 | * | |
191 | * The current value of this parameter is the result of a tuning with | |
192 | * several hardware and software configurations. We tried to find the | |
193 | * lowest value for which writes do not cause noticeable problems to | |
194 | * reads. In fact, the lower this parameter, the stabler I/O control, | |
195 | * in the following respect. The lower this parameter is, the less | |
196 | * the bandwidth enjoyed by a group decreases | |
197 | * - when the group does writes, w.r.t. to when it does reads; | |
198 | * - when other groups do reads, w.r.t. to when they do writes. | |
ea25da48 | 199 | */ |
d5801088 | 200 | static const int bfq_async_charge_factor = 3; |
aee69d78 | 201 | |
ea25da48 PV |
202 | /* Default timeout values, in jiffies, approximating CFQ defaults. */ |
203 | const int bfq_timeout = HZ / 8; | |
aee69d78 | 204 | |
7b8fa3b9 PV |
205 | /* |
206 | * Time limit for merging (see comments in bfq_setup_cooperator). Set | |
207 | * to the slowest value that, in our tests, proved to be effective in | |
208 | * removing false positives, while not causing true positives to miss | |
209 | * queue merging. | |
210 | * | |
211 | * As can be deduced from the low time limit below, queue merging, if | |
636b8fe8 | 212 | * successful, happens at the very beginning of the I/O of the involved |
7b8fa3b9 PV |
213 | * cooperating processes, as a consequence of the arrival of the very |
214 | * first requests from each cooperator. After that, there is very | |
215 | * little chance to find cooperators. | |
216 | */ | |
217 | static const unsigned long bfq_merge_time_limit = HZ/10; | |
218 | ||
ea25da48 | 219 | static struct kmem_cache *bfq_pool; |
e21b7a0b | 220 | |
ea25da48 PV |
221 | /* Below this threshold (in ns), we consider thinktime immediate. */ |
222 | #define BFQ_MIN_TT (2 * NSEC_PER_MSEC) | |
e21b7a0b | 223 | |
ea25da48 | 224 | /* hw_tag detection: parallel requests threshold and min samples needed. */ |
a3c92560 | 225 | #define BFQ_HW_QUEUE_THRESHOLD 3 |
ea25da48 | 226 | #define BFQ_HW_QUEUE_SAMPLES 32 |
aee69d78 | 227 | |
ea25da48 PV |
228 | #define BFQQ_SEEK_THR (sector_t)(8 * 100) |
229 | #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32) | |
d87447d8 PV |
230 | #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \ |
231 | (get_sdist(last_pos, rq) > \ | |
232 | BFQQ_SEEK_THR && \ | |
233 | (!blk_queue_nonrot(bfqd->queue) || \ | |
234 | blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT)) | |
ea25da48 | 235 | #define BFQQ_CLOSE_THR (sector_t)(8 * 1024) |
f0ba5ea2 | 236 | #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19) |
7074f076 PV |
237 | /* |
238 | * Sync random I/O is likely to be confused with soft real-time I/O, | |
239 | * because it is characterized by limited throughput and apparently | |
240 | * isochronous arrival pattern. To avoid false positives, queues | |
241 | * containing only random (seeky) I/O are prevented from being tagged | |
242 | * as soft real-time. | |
243 | */ | |
e6feaf21 | 244 | #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1) |
aee69d78 | 245 | |
ea25da48 PV |
246 | /* Min number of samples required to perform peak-rate update */ |
247 | #define BFQ_RATE_MIN_SAMPLES 32 | |
248 | /* Min observation time interval required to perform a peak-rate update (ns) */ | |
249 | #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC) | |
250 | /* Target observation time interval for a peak-rate update (ns) */ | |
251 | #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC | |
aee69d78 | 252 | |
bc56e2ca PV |
253 | /* |
254 | * Shift used for peak-rate fixed precision calculations. | |
255 | * With | |
256 | * - the current shift: 16 positions | |
257 | * - the current type used to store rate: u32 | |
258 | * - the current unit of measure for rate: [sectors/usec], or, more precisely, | |
259 | * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift, | |
260 | * the range of rates that can be stored is | |
261 | * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec = | |
262 | * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec = | |
263 | * [15, 65G] sectors/sec | |
264 | * Which, assuming a sector size of 512B, corresponds to a range of | |
265 | * [7.5K, 33T] B/sec | |
266 | */ | |
ea25da48 | 267 | #define BFQ_RATE_SHIFT 16 |
aee69d78 | 268 | |
ea25da48 | 269 | /* |
4029eef1 PV |
270 | * When configured for computing the duration of the weight-raising |
271 | * for interactive queues automatically (see the comments at the | |
272 | * beginning of this file), BFQ does it using the following formula: | |
e24f1c24 PV |
273 | * duration = (ref_rate / r) * ref_wr_duration, |
274 | * where r is the peak rate of the device, and ref_rate and | |
275 | * ref_wr_duration are two reference parameters. In particular, | |
276 | * ref_rate is the peak rate of the reference storage device (see | |
277 | * below), and ref_wr_duration is about the maximum time needed, with | |
278 | * BFQ and while reading two files in parallel, to load typical large | |
279 | * applications on the reference device (see the comments on | |
280 | * max_service_from_wr below, for more details on how ref_wr_duration | |
281 | * is obtained). In practice, the slower/faster the device at hand | |
282 | * is, the more/less it takes to load applications with respect to the | |
4029eef1 PV |
283 | * reference device. Accordingly, the longer/shorter BFQ grants |
284 | * weight raising to interactive applications. | |
ea25da48 | 285 | * |
e24f1c24 PV |
286 | * BFQ uses two different reference pairs (ref_rate, ref_wr_duration), |
287 | * depending on whether the device is rotational or non-rotational. | |
ea25da48 | 288 | * |
e24f1c24 PV |
289 | * In the following definitions, ref_rate[0] and ref_wr_duration[0] |
290 | * are the reference values for a rotational device, whereas | |
291 | * ref_rate[1] and ref_wr_duration[1] are the reference values for a | |
292 | * non-rotational device. The reference rates are not the actual peak | |
293 | * rates of the devices used as a reference, but slightly lower | |
294 | * values. The reason for using slightly lower values is that the | |
295 | * peak-rate estimator tends to yield slightly lower values than the | |
296 | * actual peak rate (it can yield the actual peak rate only if there | |
297 | * is only one process doing I/O, and the process does sequential | |
298 | * I/O). | |
ea25da48 | 299 | * |
e24f1c24 PV |
300 | * The reference peak rates are measured in sectors/usec, left-shifted |
301 | * by BFQ_RATE_SHIFT. | |
ea25da48 | 302 | */ |
e24f1c24 | 303 | static int ref_rate[2] = {14000, 33000}; |
ea25da48 | 304 | /* |
e24f1c24 PV |
305 | * To improve readability, a conversion function is used to initialize |
306 | * the following array, which entails that the array can be | |
307 | * initialized only in a function. | |
ea25da48 | 308 | */ |
e24f1c24 | 309 | static int ref_wr_duration[2]; |
aee69d78 | 310 | |
8a8747dc PV |
311 | /* |
312 | * BFQ uses the above-detailed, time-based weight-raising mechanism to | |
313 | * privilege interactive tasks. This mechanism is vulnerable to the | |
314 | * following false positives: I/O-bound applications that will go on | |
315 | * doing I/O for much longer than the duration of weight | |
316 | * raising. These applications have basically no benefit from being | |
317 | * weight-raised at the beginning of their I/O. On the opposite end, | |
318 | * while being weight-raised, these applications | |
319 | * a) unjustly steal throughput to applications that may actually need | |
320 | * low latency; | |
321 | * b) make BFQ uselessly perform device idling; device idling results | |
322 | * in loss of device throughput with most flash-based storage, and may | |
323 | * increase latencies when used purposelessly. | |
324 | * | |
325 | * BFQ tries to reduce these problems, by adopting the following | |
326 | * countermeasure. To introduce this countermeasure, we need first to | |
327 | * finish explaining how the duration of weight-raising for | |
328 | * interactive tasks is computed. | |
329 | * | |
330 | * For a bfq_queue deemed as interactive, the duration of weight | |
331 | * raising is dynamically adjusted, as a function of the estimated | |
332 | * peak rate of the device, so as to be equal to the time needed to | |
333 | * execute the 'largest' interactive task we benchmarked so far. By | |
334 | * largest task, we mean the task for which each involved process has | |
335 | * to do more I/O than for any of the other tasks we benchmarked. This | |
336 | * reference interactive task is the start-up of LibreOffice Writer, | |
337 | * and in this task each process/bfq_queue needs to have at most ~110K | |
338 | * sectors transferred. | |
339 | * | |
340 | * This last piece of information enables BFQ to reduce the actual | |
341 | * duration of weight-raising for at least one class of I/O-bound | |
342 | * applications: those doing sequential or quasi-sequential I/O. An | |
343 | * example is file copy. In fact, once started, the main I/O-bound | |
344 | * processes of these applications usually consume the above 110K | |
345 | * sectors in much less time than the processes of an application that | |
346 | * is starting, because these I/O-bound processes will greedily devote | |
347 | * almost all their CPU cycles only to their target, | |
348 | * throughput-friendly I/O operations. This is even more true if BFQ | |
349 | * happens to be underestimating the device peak rate, and thus | |
350 | * overestimating the duration of weight raising. But, according to | |
351 | * our measurements, once transferred 110K sectors, these processes | |
352 | * have no right to be weight-raised any longer. | |
353 | * | |
354 | * Basing on the last consideration, BFQ ends weight-raising for a | |
355 | * bfq_queue if the latter happens to have received an amount of | |
356 | * service at least equal to the following constant. The constant is | |
357 | * set to slightly more than 110K, to have a minimum safety margin. | |
358 | * | |
359 | * This early ending of weight-raising reduces the amount of time | |
360 | * during which interactive false positives cause the two problems | |
361 | * described at the beginning of these comments. | |
362 | */ | |
363 | static const unsigned long max_service_from_wr = 120000; | |
364 | ||
12cd3a2f | 365 | #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0]) |
ea25da48 | 366 | #define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
aee69d78 | 367 | |
ea25da48 | 368 | struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync) |
e21b7a0b | 369 | { |
ea25da48 | 370 | return bic->bfqq[is_sync]; |
aee69d78 PV |
371 | } |
372 | ||
ea25da48 | 373 | void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync) |
aee69d78 | 374 | { |
ea25da48 | 375 | bic->bfqq[is_sync] = bfqq; |
aee69d78 PV |
376 | } |
377 | ||
ea25da48 | 378 | struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
aee69d78 | 379 | { |
ea25da48 | 380 | return bic->icq.q->elevator->elevator_data; |
e21b7a0b | 381 | } |
aee69d78 | 382 | |
ea25da48 PV |
383 | /** |
384 | * icq_to_bic - convert iocontext queue structure to bfq_io_cq. | |
385 | * @icq: the iocontext queue. | |
386 | */ | |
387 | static struct bfq_io_cq *icq_to_bic(struct io_cq *icq) | |
e21b7a0b | 388 | { |
ea25da48 PV |
389 | /* bic->icq is the first member, %NULL will convert to %NULL */ |
390 | return container_of(icq, struct bfq_io_cq, icq); | |
e21b7a0b | 391 | } |
aee69d78 | 392 | |
ea25da48 PV |
393 | /** |
394 | * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. | |
395 | * @bfqd: the lookup key. | |
396 | * @ioc: the io_context of the process doing I/O. | |
397 | * @q: the request queue. | |
398 | */ | |
399 | static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, | |
400 | struct io_context *ioc, | |
401 | struct request_queue *q) | |
e21b7a0b | 402 | { |
ea25da48 PV |
403 | if (ioc) { |
404 | unsigned long flags; | |
405 | struct bfq_io_cq *icq; | |
aee69d78 | 406 | |
0d945c1f | 407 | spin_lock_irqsave(&q->queue_lock, flags); |
ea25da48 | 408 | icq = icq_to_bic(ioc_lookup_icq(ioc, q)); |
0d945c1f | 409 | spin_unlock_irqrestore(&q->queue_lock, flags); |
aee69d78 | 410 | |
ea25da48 | 411 | return icq; |
e21b7a0b | 412 | } |
e21b7a0b | 413 | |
ea25da48 | 414 | return NULL; |
aee69d78 PV |
415 | } |
416 | ||
ea25da48 PV |
417 | /* |
418 | * Scheduler run of queue, if there are requests pending and no one in the | |
419 | * driver that will restart queueing. | |
420 | */ | |
421 | void bfq_schedule_dispatch(struct bfq_data *bfqd) | |
aee69d78 | 422 | { |
ea25da48 PV |
423 | if (bfqd->queued != 0) { |
424 | bfq_log(bfqd, "schedule dispatch"); | |
425 | blk_mq_run_hw_queues(bfqd->queue, true); | |
e21b7a0b | 426 | } |
aee69d78 PV |
427 | } |
428 | ||
429 | #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
430 | #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT) | |
431 | ||
432 | #define bfq_sample_valid(samples) ((samples) > 80) | |
433 | ||
aee69d78 PV |
434 | /* |
435 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
636b8fe8 | 436 | * We choose the request that is closer to the head right now. Distance |
aee69d78 PV |
437 | * behind the head is penalized and only allowed to a certain extent. |
438 | */ | |
439 | static struct request *bfq_choose_req(struct bfq_data *bfqd, | |
440 | struct request *rq1, | |
441 | struct request *rq2, | |
442 | sector_t last) | |
443 | { | |
444 | sector_t s1, s2, d1 = 0, d2 = 0; | |
445 | unsigned long back_max; | |
446 | #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
447 | #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
448 | unsigned int wrap = 0; /* bit mask: requests behind the disk head? */ | |
449 | ||
450 | if (!rq1 || rq1 == rq2) | |
451 | return rq2; | |
452 | if (!rq2) | |
453 | return rq1; | |
454 | ||
455 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) | |
456 | return rq1; | |
457 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
458 | return rq2; | |
459 | if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) | |
460 | return rq1; | |
461 | else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) | |
462 | return rq2; | |
463 | ||
464 | s1 = blk_rq_pos(rq1); | |
465 | s2 = blk_rq_pos(rq2); | |
466 | ||
467 | /* | |
468 | * By definition, 1KiB is 2 sectors. | |
469 | */ | |
470 | back_max = bfqd->bfq_back_max * 2; | |
471 | ||
472 | /* | |
473 | * Strict one way elevator _except_ in the case where we allow | |
474 | * short backward seeks which are biased as twice the cost of a | |
475 | * similar forward seek. | |
476 | */ | |
477 | if (s1 >= last) | |
478 | d1 = s1 - last; | |
479 | else if (s1 + back_max >= last) | |
480 | d1 = (last - s1) * bfqd->bfq_back_penalty; | |
481 | else | |
482 | wrap |= BFQ_RQ1_WRAP; | |
483 | ||
484 | if (s2 >= last) | |
485 | d2 = s2 - last; | |
486 | else if (s2 + back_max >= last) | |
487 | d2 = (last - s2) * bfqd->bfq_back_penalty; | |
488 | else | |
489 | wrap |= BFQ_RQ2_WRAP; | |
490 | ||
491 | /* Found required data */ | |
492 | ||
493 | /* | |
494 | * By doing switch() on the bit mask "wrap" we avoid having to | |
495 | * check two variables for all permutations: --> faster! | |
496 | */ | |
497 | switch (wrap) { | |
498 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
499 | if (d1 < d2) | |
500 | return rq1; | |
501 | else if (d2 < d1) | |
502 | return rq2; | |
503 | ||
504 | if (s1 >= s2) | |
505 | return rq1; | |
506 | else | |
507 | return rq2; | |
508 | ||
509 | case BFQ_RQ2_WRAP: | |
510 | return rq1; | |
511 | case BFQ_RQ1_WRAP: | |
512 | return rq2; | |
513 | case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */ | |
514 | default: | |
515 | /* | |
516 | * Since both rqs are wrapped, | |
517 | * start with the one that's further behind head | |
518 | * (--> only *one* back seek required), | |
519 | * since back seek takes more time than forward. | |
520 | */ | |
521 | if (s1 <= s2) | |
522 | return rq1; | |
523 | else | |
524 | return rq2; | |
525 | } | |
526 | } | |
527 | ||
a52a69ea PV |
528 | /* |
529 | * Async I/O can easily starve sync I/O (both sync reads and sync | |
530 | * writes), by consuming all tags. Similarly, storms of sync writes, | |
531 | * such as those that sync(2) may trigger, can starve sync reads. | |
532 | * Limit depths of async I/O and sync writes so as to counter both | |
533 | * problems. | |
534 | */ | |
535 | static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data) | |
536 | { | |
a52a69ea | 537 | struct bfq_data *bfqd = data->q->elevator->elevator_data; |
a52a69ea PV |
538 | |
539 | if (op_is_sync(op) && !op_is_write(op)) | |
540 | return; | |
541 | ||
a52a69ea PV |
542 | data->shallow_depth = |
543 | bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)]; | |
544 | ||
545 | bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u", | |
546 | __func__, bfqd->wr_busy_queues, op_is_sync(op), | |
547 | data->shallow_depth); | |
548 | } | |
549 | ||
36eca894 AA |
550 | static struct bfq_queue * |
551 | bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, | |
552 | sector_t sector, struct rb_node **ret_parent, | |
553 | struct rb_node ***rb_link) | |
554 | { | |
555 | struct rb_node **p, *parent; | |
556 | struct bfq_queue *bfqq = NULL; | |
557 | ||
558 | parent = NULL; | |
559 | p = &root->rb_node; | |
560 | while (*p) { | |
561 | struct rb_node **n; | |
562 | ||
563 | parent = *p; | |
564 | bfqq = rb_entry(parent, struct bfq_queue, pos_node); | |
565 | ||
566 | /* | |
567 | * Sort strictly based on sector. Smallest to the left, | |
568 | * largest to the right. | |
569 | */ | |
570 | if (sector > blk_rq_pos(bfqq->next_rq)) | |
571 | n = &(*p)->rb_right; | |
572 | else if (sector < blk_rq_pos(bfqq->next_rq)) | |
573 | n = &(*p)->rb_left; | |
574 | else | |
575 | break; | |
576 | p = n; | |
577 | bfqq = NULL; | |
578 | } | |
579 | ||
580 | *ret_parent = parent; | |
581 | if (rb_link) | |
582 | *rb_link = p; | |
583 | ||
584 | bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", | |
585 | (unsigned long long)sector, | |
586 | bfqq ? bfqq->pid : 0); | |
587 | ||
588 | return bfqq; | |
589 | } | |
590 | ||
7b8fa3b9 PV |
591 | static bool bfq_too_late_for_merging(struct bfq_queue *bfqq) |
592 | { | |
593 | return bfqq->service_from_backlogged > 0 && | |
594 | time_is_before_jiffies(bfqq->first_IO_time + | |
595 | bfq_merge_time_limit); | |
596 | } | |
597 | ||
8cacc5ab PV |
598 | /* |
599 | * The following function is not marked as __cold because it is | |
600 | * actually cold, but for the same performance goal described in the | |
601 | * comments on the likely() at the beginning of | |
602 | * bfq_setup_cooperator(). Unexpectedly, to reach an even lower | |
603 | * execution time for the case where this function is not invoked, we | |
604 | * had to add an unlikely() in each involved if(). | |
605 | */ | |
606 | void __cold | |
607 | bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
36eca894 AA |
608 | { |
609 | struct rb_node **p, *parent; | |
610 | struct bfq_queue *__bfqq; | |
611 | ||
612 | if (bfqq->pos_root) { | |
613 | rb_erase(&bfqq->pos_node, bfqq->pos_root); | |
614 | bfqq->pos_root = NULL; | |
615 | } | |
616 | ||
7b8fa3b9 PV |
617 | /* |
618 | * bfqq cannot be merged any longer (see comments in | |
619 | * bfq_setup_cooperator): no point in adding bfqq into the | |
620 | * position tree. | |
621 | */ | |
622 | if (bfq_too_late_for_merging(bfqq)) | |
623 | return; | |
624 | ||
36eca894 AA |
625 | if (bfq_class_idle(bfqq)) |
626 | return; | |
627 | if (!bfqq->next_rq) | |
628 | return; | |
629 | ||
630 | bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree; | |
631 | __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, | |
632 | blk_rq_pos(bfqq->next_rq), &parent, &p); | |
633 | if (!__bfqq) { | |
634 | rb_link_node(&bfqq->pos_node, parent, p); | |
635 | rb_insert_color(&bfqq->pos_node, bfqq->pos_root); | |
636 | } else | |
637 | bfqq->pos_root = NULL; | |
638 | } | |
639 | ||
1de0c4cd | 640 | /* |
fb53ac6c PV |
641 | * The following function returns false either if every active queue |
642 | * must receive the same share of the throughput (symmetric scenario), | |
643 | * or, as a special case, if bfqq must receive a share of the | |
644 | * throughput lower than or equal to the share that every other active | |
645 | * queue must receive. If bfqq does sync I/O, then these are the only | |
646 | * two cases where bfqq happens to be guaranteed its share of the | |
647 | * throughput even if I/O dispatching is not plugged when bfqq remains | |
648 | * temporarily empty (for more details, see the comments in the | |
649 | * function bfq_better_to_idle()). For this reason, the return value | |
650 | * of this function is used to check whether I/O-dispatch plugging can | |
651 | * be avoided. | |
1de0c4cd | 652 | * |
fb53ac6c | 653 | * The above first case (symmetric scenario) occurs when: |
1de0c4cd | 654 | * 1) all active queues have the same weight, |
73d58118 | 655 | * 2) all active queues belong to the same I/O-priority class, |
1de0c4cd | 656 | * 3) all active groups at the same level in the groups tree have the same |
73d58118 PV |
657 | * weight, |
658 | * 4) all active groups at the same level in the groups tree have the same | |
1de0c4cd AA |
659 | * number of children. |
660 | * | |
2d29c9f8 FM |
661 | * Unfortunately, keeping the necessary state for evaluating exactly |
662 | * the last two symmetry sub-conditions above would be quite complex | |
73d58118 PV |
663 | * and time consuming. Therefore this function evaluates, instead, |
664 | * only the following stronger three sub-conditions, for which it is | |
2d29c9f8 | 665 | * much easier to maintain the needed state: |
1de0c4cd | 666 | * 1) all active queues have the same weight, |
73d58118 PV |
667 | * 2) all active queues belong to the same I/O-priority class, |
668 | * 3) there are no active groups. | |
2d29c9f8 FM |
669 | * In particular, the last condition is always true if hierarchical |
670 | * support or the cgroups interface are not enabled, thus no state | |
671 | * needs to be maintained in this case. | |
1de0c4cd | 672 | */ |
fb53ac6c PV |
673 | static bool bfq_asymmetric_scenario(struct bfq_data *bfqd, |
674 | struct bfq_queue *bfqq) | |
1de0c4cd | 675 | { |
fb53ac6c PV |
676 | bool smallest_weight = bfqq && |
677 | bfqq->weight_counter && | |
678 | bfqq->weight_counter == | |
679 | container_of( | |
680 | rb_first_cached(&bfqd->queue_weights_tree), | |
681 | struct bfq_weight_counter, | |
682 | weights_node); | |
683 | ||
73d58118 PV |
684 | /* |
685 | * For queue weights to differ, queue_weights_tree must contain | |
686 | * at least two nodes. | |
687 | */ | |
fb53ac6c PV |
688 | bool varied_queue_weights = !smallest_weight && |
689 | !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) && | |
690 | (bfqd->queue_weights_tree.rb_root.rb_node->rb_left || | |
691 | bfqd->queue_weights_tree.rb_root.rb_node->rb_right); | |
73d58118 PV |
692 | |
693 | bool multiple_classes_busy = | |
694 | (bfqd->busy_queues[0] && bfqd->busy_queues[1]) || | |
695 | (bfqd->busy_queues[0] && bfqd->busy_queues[2]) || | |
696 | (bfqd->busy_queues[1] && bfqd->busy_queues[2]); | |
697 | ||
fb53ac6c | 698 | return varied_queue_weights || multiple_classes_busy |
42b1bd33 | 699 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
73d58118 PV |
700 | || bfqd->num_groups_with_pending_reqs > 0 |
701 | #endif | |
fb53ac6c | 702 | ; |
1de0c4cd AA |
703 | } |
704 | ||
705 | /* | |
706 | * If the weight-counter tree passed as input contains no counter for | |
2d29c9f8 | 707 | * the weight of the input queue, then add that counter; otherwise just |
1de0c4cd AA |
708 | * increment the existing counter. |
709 | * | |
710 | * Note that weight-counter trees contain few nodes in mostly symmetric | |
711 | * scenarios. For example, if all queues have the same weight, then the | |
712 | * weight-counter tree for the queues may contain at most one node. | |
713 | * This holds even if low_latency is on, because weight-raised queues | |
714 | * are not inserted in the tree. | |
715 | * In most scenarios, the rate at which nodes are created/destroyed | |
716 | * should be low too. | |
717 | */ | |
2d29c9f8 | 718 | void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
fb53ac6c | 719 | struct rb_root_cached *root) |
1de0c4cd | 720 | { |
2d29c9f8 | 721 | struct bfq_entity *entity = &bfqq->entity; |
fb53ac6c PV |
722 | struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; |
723 | bool leftmost = true; | |
1de0c4cd AA |
724 | |
725 | /* | |
2d29c9f8 | 726 | * Do not insert if the queue is already associated with a |
1de0c4cd | 727 | * counter, which happens if: |
2d29c9f8 | 728 | * 1) a request arrival has caused the queue to become both |
1de0c4cd AA |
729 | * non-weight-raised, and hence change its weight, and |
730 | * backlogged; in this respect, each of the two events | |
731 | * causes an invocation of this function, | |
2d29c9f8 | 732 | * 2) this is the invocation of this function caused by the |
1de0c4cd AA |
733 | * second event. This second invocation is actually useless, |
734 | * and we handle this fact by exiting immediately. More | |
735 | * efficient or clearer solutions might possibly be adopted. | |
736 | */ | |
2d29c9f8 | 737 | if (bfqq->weight_counter) |
1de0c4cd AA |
738 | return; |
739 | ||
740 | while (*new) { | |
741 | struct bfq_weight_counter *__counter = container_of(*new, | |
742 | struct bfq_weight_counter, | |
743 | weights_node); | |
744 | parent = *new; | |
745 | ||
746 | if (entity->weight == __counter->weight) { | |
2d29c9f8 | 747 | bfqq->weight_counter = __counter; |
1de0c4cd AA |
748 | goto inc_counter; |
749 | } | |
750 | if (entity->weight < __counter->weight) | |
751 | new = &((*new)->rb_left); | |
fb53ac6c | 752 | else { |
1de0c4cd | 753 | new = &((*new)->rb_right); |
fb53ac6c PV |
754 | leftmost = false; |
755 | } | |
1de0c4cd AA |
756 | } |
757 | ||
2d29c9f8 FM |
758 | bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter), |
759 | GFP_ATOMIC); | |
1de0c4cd AA |
760 | |
761 | /* | |
762 | * In the unlucky event of an allocation failure, we just | |
2d29c9f8 | 763 | * exit. This will cause the weight of queue to not be |
fb53ac6c | 764 | * considered in bfq_asymmetric_scenario, which, in its turn, |
73d58118 PV |
765 | * causes the scenario to be deemed wrongly symmetric in case |
766 | * bfqq's weight would have been the only weight making the | |
767 | * scenario asymmetric. On the bright side, no unbalance will | |
768 | * however occur when bfqq becomes inactive again (the | |
769 | * invocation of this function is triggered by an activation | |
770 | * of queue). In fact, bfq_weights_tree_remove does nothing | |
771 | * if !bfqq->weight_counter. | |
1de0c4cd | 772 | */ |
2d29c9f8 | 773 | if (unlikely(!bfqq->weight_counter)) |
1de0c4cd AA |
774 | return; |
775 | ||
2d29c9f8 FM |
776 | bfqq->weight_counter->weight = entity->weight; |
777 | rb_link_node(&bfqq->weight_counter->weights_node, parent, new); | |
fb53ac6c PV |
778 | rb_insert_color_cached(&bfqq->weight_counter->weights_node, root, |
779 | leftmost); | |
1de0c4cd AA |
780 | |
781 | inc_counter: | |
2d29c9f8 | 782 | bfqq->weight_counter->num_active++; |
9dee8b3b | 783 | bfqq->ref++; |
1de0c4cd AA |
784 | } |
785 | ||
786 | /* | |
2d29c9f8 | 787 | * Decrement the weight counter associated with the queue, and, if the |
1de0c4cd AA |
788 | * counter reaches 0, remove the counter from the tree. |
789 | * See the comments to the function bfq_weights_tree_add() for considerations | |
790 | * about overhead. | |
791 | */ | |
0471559c | 792 | void __bfq_weights_tree_remove(struct bfq_data *bfqd, |
2d29c9f8 | 793 | struct bfq_queue *bfqq, |
fb53ac6c | 794 | struct rb_root_cached *root) |
1de0c4cd | 795 | { |
2d29c9f8 | 796 | if (!bfqq->weight_counter) |
1de0c4cd AA |
797 | return; |
798 | ||
2d29c9f8 FM |
799 | bfqq->weight_counter->num_active--; |
800 | if (bfqq->weight_counter->num_active > 0) | |
1de0c4cd AA |
801 | goto reset_entity_pointer; |
802 | ||
fb53ac6c | 803 | rb_erase_cached(&bfqq->weight_counter->weights_node, root); |
2d29c9f8 | 804 | kfree(bfqq->weight_counter); |
1de0c4cd AA |
805 | |
806 | reset_entity_pointer: | |
2d29c9f8 | 807 | bfqq->weight_counter = NULL; |
9dee8b3b | 808 | bfq_put_queue(bfqq); |
1de0c4cd AA |
809 | } |
810 | ||
0471559c | 811 | /* |
2d29c9f8 FM |
812 | * Invoke __bfq_weights_tree_remove on bfqq and decrement the number |
813 | * of active groups for each queue's inactive parent entity. | |
0471559c PV |
814 | */ |
815 | void bfq_weights_tree_remove(struct bfq_data *bfqd, | |
816 | struct bfq_queue *bfqq) | |
817 | { | |
818 | struct bfq_entity *entity = bfqq->entity.parent; | |
819 | ||
0471559c PV |
820 | for_each_entity(entity) { |
821 | struct bfq_sched_data *sd = entity->my_sched_data; | |
822 | ||
823 | if (sd->next_in_service || sd->in_service_entity) { | |
824 | /* | |
825 | * entity is still active, because either | |
826 | * next_in_service or in_service_entity is not | |
827 | * NULL (see the comments on the definition of | |
828 | * next_in_service for details on why | |
829 | * in_service_entity must be checked too). | |
830 | * | |
2d29c9f8 FM |
831 | * As a consequence, its parent entities are |
832 | * active as well, and thus this loop must | |
833 | * stop here. | |
0471559c PV |
834 | */ |
835 | break; | |
836 | } | |
ba7aeae5 PV |
837 | |
838 | /* | |
839 | * The decrement of num_groups_with_pending_reqs is | |
840 | * not performed immediately upon the deactivation of | |
841 | * entity, but it is delayed to when it also happens | |
842 | * that the first leaf descendant bfqq of entity gets | |
843 | * all its pending requests completed. The following | |
844 | * instructions perform this delayed decrement, if | |
845 | * needed. See the comments on | |
846 | * num_groups_with_pending_reqs for details. | |
847 | */ | |
848 | if (entity->in_groups_with_pending_reqs) { | |
849 | entity->in_groups_with_pending_reqs = false; | |
850 | bfqd->num_groups_with_pending_reqs--; | |
851 | } | |
0471559c | 852 | } |
9dee8b3b PV |
853 | |
854 | /* | |
855 | * Next function is invoked last, because it causes bfqq to be | |
856 | * freed if the following holds: bfqq is not in service and | |
857 | * has no dispatched request. DO NOT use bfqq after the next | |
858 | * function invocation. | |
859 | */ | |
860 | __bfq_weights_tree_remove(bfqd, bfqq, | |
861 | &bfqd->queue_weights_tree); | |
0471559c PV |
862 | } |
863 | ||
aee69d78 PV |
864 | /* |
865 | * Return expired entry, or NULL to just start from scratch in rbtree. | |
866 | */ | |
867 | static struct request *bfq_check_fifo(struct bfq_queue *bfqq, | |
868 | struct request *last) | |
869 | { | |
870 | struct request *rq; | |
871 | ||
872 | if (bfq_bfqq_fifo_expire(bfqq)) | |
873 | return NULL; | |
874 | ||
875 | bfq_mark_bfqq_fifo_expire(bfqq); | |
876 | ||
877 | rq = rq_entry_fifo(bfqq->fifo.next); | |
878 | ||
879 | if (rq == last || ktime_get_ns() < rq->fifo_time) | |
880 | return NULL; | |
881 | ||
882 | bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq); | |
883 | return rq; | |
884 | } | |
885 | ||
886 | static struct request *bfq_find_next_rq(struct bfq_data *bfqd, | |
887 | struct bfq_queue *bfqq, | |
888 | struct request *last) | |
889 | { | |
890 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
891 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
892 | struct request *next, *prev = NULL; | |
893 | ||
894 | /* Follow expired path, else get first next available. */ | |
895 | next = bfq_check_fifo(bfqq, last); | |
896 | if (next) | |
897 | return next; | |
898 | ||
899 | if (rbprev) | |
900 | prev = rb_entry_rq(rbprev); | |
901 | ||
902 | if (rbnext) | |
903 | next = rb_entry_rq(rbnext); | |
904 | else { | |
905 | rbnext = rb_first(&bfqq->sort_list); | |
906 | if (rbnext && rbnext != &last->rb_node) | |
907 | next = rb_entry_rq(rbnext); | |
908 | } | |
909 | ||
910 | return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); | |
911 | } | |
912 | ||
c074170e | 913 | /* see the definition of bfq_async_charge_factor for details */ |
aee69d78 PV |
914 | static unsigned long bfq_serv_to_charge(struct request *rq, |
915 | struct bfq_queue *bfqq) | |
916 | { | |
02a6d787 | 917 | if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 || |
fb53ac6c | 918 | bfq_asymmetric_scenario(bfqq->bfqd, bfqq)) |
c074170e PV |
919 | return blk_rq_sectors(rq); |
920 | ||
d5801088 | 921 | return blk_rq_sectors(rq) * bfq_async_charge_factor; |
aee69d78 PV |
922 | } |
923 | ||
924 | /** | |
925 | * bfq_updated_next_req - update the queue after a new next_rq selection. | |
926 | * @bfqd: the device data the queue belongs to. | |
927 | * @bfqq: the queue to update. | |
928 | * | |
929 | * If the first request of a queue changes we make sure that the queue | |
930 | * has enough budget to serve at least its first request (if the | |
931 | * request has grown). We do this because if the queue has not enough | |
932 | * budget for its first request, it has to go through two dispatch | |
933 | * rounds to actually get it dispatched. | |
934 | */ | |
935 | static void bfq_updated_next_req(struct bfq_data *bfqd, | |
936 | struct bfq_queue *bfqq) | |
937 | { | |
938 | struct bfq_entity *entity = &bfqq->entity; | |
939 | struct request *next_rq = bfqq->next_rq; | |
940 | unsigned long new_budget; | |
941 | ||
942 | if (!next_rq) | |
943 | return; | |
944 | ||
945 | if (bfqq == bfqd->in_service_queue) | |
946 | /* | |
947 | * In order not to break guarantees, budgets cannot be | |
948 | * changed after an entity has been selected. | |
949 | */ | |
950 | return; | |
951 | ||
f3218ad8 PV |
952 | new_budget = max_t(unsigned long, |
953 | max_t(unsigned long, bfqq->max_budget, | |
954 | bfq_serv_to_charge(next_rq, bfqq)), | |
955 | entity->service); | |
aee69d78 PV |
956 | if (entity->budget != new_budget) { |
957 | entity->budget = new_budget; | |
958 | bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", | |
959 | new_budget); | |
80294c3b | 960 | bfq_requeue_bfqq(bfqd, bfqq, false); |
aee69d78 PV |
961 | } |
962 | } | |
963 | ||
3e2bdd6d PV |
964 | static unsigned int bfq_wr_duration(struct bfq_data *bfqd) |
965 | { | |
966 | u64 dur; | |
967 | ||
968 | if (bfqd->bfq_wr_max_time > 0) | |
969 | return bfqd->bfq_wr_max_time; | |
970 | ||
e24f1c24 | 971 | dur = bfqd->rate_dur_prod; |
3e2bdd6d PV |
972 | do_div(dur, bfqd->peak_rate); |
973 | ||
974 | /* | |
d450542e DS |
975 | * Limit duration between 3 and 25 seconds. The upper limit |
976 | * has been conservatively set after the following worst case: | |
977 | * on a QEMU/KVM virtual machine | |
978 | * - running in a slow PC | |
979 | * - with a virtual disk stacked on a slow low-end 5400rpm HDD | |
980 | * - serving a heavy I/O workload, such as the sequential reading | |
981 | * of several files | |
982 | * mplayer took 23 seconds to start, if constantly weight-raised. | |
983 | * | |
636b8fe8 | 984 | * As for higher values than that accommodating the above bad |
d450542e DS |
985 | * scenario, tests show that higher values would often yield |
986 | * the opposite of the desired result, i.e., would worsen | |
987 | * responsiveness by allowing non-interactive applications to | |
988 | * preserve weight raising for too long. | |
3e2bdd6d PV |
989 | * |
990 | * On the other end, lower values than 3 seconds make it | |
991 | * difficult for most interactive tasks to complete their jobs | |
992 | * before weight-raising finishes. | |
993 | */ | |
d450542e | 994 | return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000)); |
3e2bdd6d PV |
995 | } |
996 | ||
997 | /* switch back from soft real-time to interactive weight raising */ | |
998 | static void switch_back_to_interactive_wr(struct bfq_queue *bfqq, | |
999 | struct bfq_data *bfqd) | |
1000 | { | |
1001 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1002 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1003 | bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt; | |
1004 | } | |
1005 | ||
36eca894 | 1006 | static void |
13c931bd PV |
1007 | bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd, |
1008 | struct bfq_io_cq *bic, bool bfq_already_existing) | |
36eca894 | 1009 | { |
13c931bd PV |
1010 | unsigned int old_wr_coeff = bfqq->wr_coeff; |
1011 | bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq); | |
1012 | ||
d5be3fef PV |
1013 | if (bic->saved_has_short_ttime) |
1014 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
36eca894 | 1015 | else |
d5be3fef | 1016 | bfq_clear_bfqq_has_short_ttime(bfqq); |
36eca894 AA |
1017 | |
1018 | if (bic->saved_IO_bound) | |
1019 | bfq_mark_bfqq_IO_bound(bfqq); | |
1020 | else | |
1021 | bfq_clear_bfqq_IO_bound(bfqq); | |
1022 | ||
fffca087 | 1023 | bfqq->entity.new_weight = bic->saved_weight; |
36eca894 AA |
1024 | bfqq->ttime = bic->saved_ttime; |
1025 | bfqq->wr_coeff = bic->saved_wr_coeff; | |
1026 | bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt; | |
1027 | bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish; | |
1028 | bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time; | |
1029 | ||
e1b2324d | 1030 | if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) || |
36eca894 | 1031 | time_is_before_jiffies(bfqq->last_wr_start_finish + |
e1b2324d | 1032 | bfqq->wr_cur_max_time))) { |
3e2bdd6d PV |
1033 | if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && |
1034 | !bfq_bfqq_in_large_burst(bfqq) && | |
1035 | time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt + | |
1036 | bfq_wr_duration(bfqd))) { | |
1037 | switch_back_to_interactive_wr(bfqq, bfqd); | |
1038 | } else { | |
1039 | bfqq->wr_coeff = 1; | |
1040 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
1041 | "resume state: switching off wr"); | |
1042 | } | |
36eca894 AA |
1043 | } |
1044 | ||
1045 | /* make sure weight will be updated, however we got here */ | |
1046 | bfqq->entity.prio_changed = 1; | |
13c931bd PV |
1047 | |
1048 | if (likely(!busy)) | |
1049 | return; | |
1050 | ||
1051 | if (old_wr_coeff == 1 && bfqq->wr_coeff > 1) | |
1052 | bfqd->wr_busy_queues++; | |
1053 | else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1) | |
1054 | bfqd->wr_busy_queues--; | |
36eca894 AA |
1055 | } |
1056 | ||
1057 | static int bfqq_process_refs(struct bfq_queue *bfqq) | |
1058 | { | |
9dee8b3b PV |
1059 | return bfqq->ref - bfqq->allocated - bfqq->entity.on_st - |
1060 | (bfqq->weight_counter != NULL); | |
36eca894 AA |
1061 | } |
1062 | ||
e1b2324d AA |
1063 | /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */ |
1064 | static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1065 | { | |
1066 | struct bfq_queue *item; | |
1067 | struct hlist_node *n; | |
1068 | ||
1069 | hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node) | |
1070 | hlist_del_init(&item->burst_list_node); | |
84a74689 PV |
1071 | |
1072 | /* | |
1073 | * Start the creation of a new burst list only if there is no | |
1074 | * active queue. See comments on the conditional invocation of | |
1075 | * bfq_handle_burst(). | |
1076 | */ | |
1077 | if (bfq_tot_busy_queues(bfqd) == 0) { | |
1078 | hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list); | |
1079 | bfqd->burst_size = 1; | |
1080 | } else | |
1081 | bfqd->burst_size = 0; | |
1082 | ||
e1b2324d AA |
1083 | bfqd->burst_parent_entity = bfqq->entity.parent; |
1084 | } | |
1085 | ||
1086 | /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */ | |
1087 | static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1088 | { | |
1089 | /* Increment burst size to take into account also bfqq */ | |
1090 | bfqd->burst_size++; | |
1091 | ||
1092 | if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) { | |
1093 | struct bfq_queue *pos, *bfqq_item; | |
1094 | struct hlist_node *n; | |
1095 | ||
1096 | /* | |
1097 | * Enough queues have been activated shortly after each | |
1098 | * other to consider this burst as large. | |
1099 | */ | |
1100 | bfqd->large_burst = true; | |
1101 | ||
1102 | /* | |
1103 | * We can now mark all queues in the burst list as | |
1104 | * belonging to a large burst. | |
1105 | */ | |
1106 | hlist_for_each_entry(bfqq_item, &bfqd->burst_list, | |
1107 | burst_list_node) | |
1108 | bfq_mark_bfqq_in_large_burst(bfqq_item); | |
1109 | bfq_mark_bfqq_in_large_burst(bfqq); | |
1110 | ||
1111 | /* | |
1112 | * From now on, and until the current burst finishes, any | |
1113 | * new queue being activated shortly after the last queue | |
1114 | * was inserted in the burst can be immediately marked as | |
1115 | * belonging to a large burst. So the burst list is not | |
1116 | * needed any more. Remove it. | |
1117 | */ | |
1118 | hlist_for_each_entry_safe(pos, n, &bfqd->burst_list, | |
1119 | burst_list_node) | |
1120 | hlist_del_init(&pos->burst_list_node); | |
1121 | } else /* | |
1122 | * Burst not yet large: add bfqq to the burst list. Do | |
1123 | * not increment the ref counter for bfqq, because bfqq | |
1124 | * is removed from the burst list before freeing bfqq | |
1125 | * in put_queue. | |
1126 | */ | |
1127 | hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list); | |
1128 | } | |
1129 | ||
1130 | /* | |
1131 | * If many queues belonging to the same group happen to be created | |
1132 | * shortly after each other, then the processes associated with these | |
1133 | * queues have typically a common goal. In particular, bursts of queue | |
1134 | * creations are usually caused by services or applications that spawn | |
1135 | * many parallel threads/processes. Examples are systemd during boot, | |
1136 | * or git grep. To help these processes get their job done as soon as | |
1137 | * possible, it is usually better to not grant either weight-raising | |
84a74689 PV |
1138 | * or device idling to their queues, unless these queues must be |
1139 | * protected from the I/O flowing through other active queues. | |
e1b2324d AA |
1140 | * |
1141 | * In this comment we describe, firstly, the reasons why this fact | |
1142 | * holds, and, secondly, the next function, which implements the main | |
1143 | * steps needed to properly mark these queues so that they can then be | |
1144 | * treated in a different way. | |
1145 | * | |
1146 | * The above services or applications benefit mostly from a high | |
1147 | * throughput: the quicker the requests of the activated queues are | |
1148 | * cumulatively served, the sooner the target job of these queues gets | |
1149 | * completed. As a consequence, weight-raising any of these queues, | |
1150 | * which also implies idling the device for it, is almost always | |
84a74689 PV |
1151 | * counterproductive, unless there are other active queues to isolate |
1152 | * these new queues from. If there no other active queues, then | |
1153 | * weight-raising these new queues just lowers throughput in most | |
1154 | * cases. | |
e1b2324d AA |
1155 | * |
1156 | * On the other hand, a burst of queue creations may be caused also by | |
1157 | * the start of an application that does not consist of a lot of | |
1158 | * parallel I/O-bound threads. In fact, with a complex application, | |
1159 | * several short processes may need to be executed to start-up the | |
1160 | * application. In this respect, to start an application as quickly as | |
1161 | * possible, the best thing to do is in any case to privilege the I/O | |
1162 | * related to the application with respect to all other | |
1163 | * I/O. Therefore, the best strategy to start as quickly as possible | |
1164 | * an application that causes a burst of queue creations is to | |
1165 | * weight-raise all the queues created during the burst. This is the | |
1166 | * exact opposite of the best strategy for the other type of bursts. | |
1167 | * | |
1168 | * In the end, to take the best action for each of the two cases, the | |
1169 | * two types of bursts need to be distinguished. Fortunately, this | |
1170 | * seems relatively easy, by looking at the sizes of the bursts. In | |
1171 | * particular, we found a threshold such that only bursts with a | |
1172 | * larger size than that threshold are apparently caused by | |
1173 | * services or commands such as systemd or git grep. For brevity, | |
1174 | * hereafter we call just 'large' these bursts. BFQ *does not* | |
1175 | * weight-raise queues whose creation occurs in a large burst. In | |
1176 | * addition, for each of these queues BFQ performs or does not perform | |
1177 | * idling depending on which choice boosts the throughput more. The | |
1178 | * exact choice depends on the device and request pattern at | |
1179 | * hand. | |
1180 | * | |
1181 | * Unfortunately, false positives may occur while an interactive task | |
1182 | * is starting (e.g., an application is being started). The | |
1183 | * consequence is that the queues associated with the task do not | |
1184 | * enjoy weight raising as expected. Fortunately these false positives | |
1185 | * are very rare. They typically occur if some service happens to | |
1186 | * start doing I/O exactly when the interactive task starts. | |
1187 | * | |
84a74689 PV |
1188 | * Turning back to the next function, it is invoked only if there are |
1189 | * no active queues (apart from active queues that would belong to the | |
1190 | * same, possible burst bfqq would belong to), and it implements all | |
1191 | * the steps needed to detect the occurrence of a large burst and to | |
1192 | * properly mark all the queues belonging to it (so that they can then | |
1193 | * be treated in a different way). This goal is achieved by | |
1194 | * maintaining a "burst list" that holds, temporarily, the queues that | |
1195 | * belong to the burst in progress. The list is then used to mark | |
1196 | * these queues as belonging to a large burst if the burst does become | |
1197 | * large. The main steps are the following. | |
e1b2324d AA |
1198 | * |
1199 | * . when the very first queue is created, the queue is inserted into the | |
1200 | * list (as it could be the first queue in a possible burst) | |
1201 | * | |
1202 | * . if the current burst has not yet become large, and a queue Q that does | |
1203 | * not yet belong to the burst is activated shortly after the last time | |
1204 | * at which a new queue entered the burst list, then the function appends | |
1205 | * Q to the burst list | |
1206 | * | |
1207 | * . if, as a consequence of the previous step, the burst size reaches | |
1208 | * the large-burst threshold, then | |
1209 | * | |
1210 | * . all the queues in the burst list are marked as belonging to a | |
1211 | * large burst | |
1212 | * | |
1213 | * . the burst list is deleted; in fact, the burst list already served | |
1214 | * its purpose (keeping temporarily track of the queues in a burst, | |
1215 | * so as to be able to mark them as belonging to a large burst in the | |
1216 | * previous sub-step), and now is not needed any more | |
1217 | * | |
1218 | * . the device enters a large-burst mode | |
1219 | * | |
1220 | * . if a queue Q that does not belong to the burst is created while | |
1221 | * the device is in large-burst mode and shortly after the last time | |
1222 | * at which a queue either entered the burst list or was marked as | |
1223 | * belonging to the current large burst, then Q is immediately marked | |
1224 | * as belonging to a large burst. | |
1225 | * | |
1226 | * . if a queue Q that does not belong to the burst is created a while | |
1227 | * later, i.e., not shortly after, than the last time at which a queue | |
1228 | * either entered the burst list or was marked as belonging to the | |
1229 | * current large burst, then the current burst is deemed as finished and: | |
1230 | * | |
1231 | * . the large-burst mode is reset if set | |
1232 | * | |
1233 | * . the burst list is emptied | |
1234 | * | |
1235 | * . Q is inserted in the burst list, as Q may be the first queue | |
1236 | * in a possible new burst (then the burst list contains just Q | |
1237 | * after this step). | |
1238 | */ | |
1239 | static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq) | |
1240 | { | |
1241 | /* | |
1242 | * If bfqq is already in the burst list or is part of a large | |
1243 | * burst, or finally has just been split, then there is | |
1244 | * nothing else to do. | |
1245 | */ | |
1246 | if (!hlist_unhashed(&bfqq->burst_list_node) || | |
1247 | bfq_bfqq_in_large_burst(bfqq) || | |
1248 | time_is_after_eq_jiffies(bfqq->split_time + | |
1249 | msecs_to_jiffies(10))) | |
1250 | return; | |
1251 | ||
1252 | /* | |
1253 | * If bfqq's creation happens late enough, or bfqq belongs to | |
1254 | * a different group than the burst group, then the current | |
1255 | * burst is finished, and related data structures must be | |
1256 | * reset. | |
1257 | * | |
1258 | * In this respect, consider the special case where bfqq is | |
1259 | * the very first queue created after BFQ is selected for this | |
1260 | * device. In this case, last_ins_in_burst and | |
1261 | * burst_parent_entity are not yet significant when we get | |
1262 | * here. But it is easy to verify that, whether or not the | |
1263 | * following condition is true, bfqq will end up being | |
1264 | * inserted into the burst list. In particular the list will | |
1265 | * happen to contain only bfqq. And this is exactly what has | |
1266 | * to happen, as bfqq may be the first queue of the first | |
1267 | * burst. | |
1268 | */ | |
1269 | if (time_is_before_jiffies(bfqd->last_ins_in_burst + | |
1270 | bfqd->bfq_burst_interval) || | |
1271 | bfqq->entity.parent != bfqd->burst_parent_entity) { | |
1272 | bfqd->large_burst = false; | |
1273 | bfq_reset_burst_list(bfqd, bfqq); | |
1274 | goto end; | |
1275 | } | |
1276 | ||
1277 | /* | |
1278 | * If we get here, then bfqq is being activated shortly after the | |
1279 | * last queue. So, if the current burst is also large, we can mark | |
1280 | * bfqq as belonging to this large burst immediately. | |
1281 | */ | |
1282 | if (bfqd->large_burst) { | |
1283 | bfq_mark_bfqq_in_large_burst(bfqq); | |
1284 | goto end; | |
1285 | } | |
1286 | ||
1287 | /* | |
1288 | * If we get here, then a large-burst state has not yet been | |
1289 | * reached, but bfqq is being activated shortly after the last | |
1290 | * queue. Then we add bfqq to the burst. | |
1291 | */ | |
1292 | bfq_add_to_burst(bfqd, bfqq); | |
1293 | end: | |
1294 | /* | |
1295 | * At this point, bfqq either has been added to the current | |
1296 | * burst or has caused the current burst to terminate and a | |
1297 | * possible new burst to start. In particular, in the second | |
1298 | * case, bfqq has become the first queue in the possible new | |
1299 | * burst. In both cases last_ins_in_burst needs to be moved | |
1300 | * forward. | |
1301 | */ | |
1302 | bfqd->last_ins_in_burst = jiffies; | |
1303 | } | |
1304 | ||
aee69d78 PV |
1305 | static int bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
1306 | { | |
1307 | struct bfq_entity *entity = &bfqq->entity; | |
1308 | ||
1309 | return entity->budget - entity->service; | |
1310 | } | |
1311 | ||
1312 | /* | |
1313 | * If enough samples have been computed, return the current max budget | |
1314 | * stored in bfqd, which is dynamically updated according to the | |
1315 | * estimated disk peak rate; otherwise return the default max budget | |
1316 | */ | |
1317 | static int bfq_max_budget(struct bfq_data *bfqd) | |
1318 | { | |
1319 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
1320 | return bfq_default_max_budget; | |
1321 | else | |
1322 | return bfqd->bfq_max_budget; | |
1323 | } | |
1324 | ||
1325 | /* | |
1326 | * Return min budget, which is a fraction of the current or default | |
1327 | * max budget (trying with 1/32) | |
1328 | */ | |
1329 | static int bfq_min_budget(struct bfq_data *bfqd) | |
1330 | { | |
1331 | if (bfqd->budgets_assigned < bfq_stats_min_budgets) | |
1332 | return bfq_default_max_budget / 32; | |
1333 | else | |
1334 | return bfqd->bfq_max_budget / 32; | |
1335 | } | |
1336 | ||
aee69d78 PV |
1337 | /* |
1338 | * The next function, invoked after the input queue bfqq switches from | |
1339 | * idle to busy, updates the budget of bfqq. The function also tells | |
1340 | * whether the in-service queue should be expired, by returning | |
1341 | * true. The purpose of expiring the in-service queue is to give bfqq | |
1342 | * the chance to possibly preempt the in-service queue, and the reason | |
44e44a1b PV |
1343 | * for preempting the in-service queue is to achieve one of the two |
1344 | * goals below. | |
aee69d78 | 1345 | * |
44e44a1b PV |
1346 | * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has |
1347 | * expired because it has remained idle. In particular, bfqq may have | |
1348 | * expired for one of the following two reasons: | |
aee69d78 PV |
1349 | * |
1350 | * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling | |
1351 | * and did not make it to issue a new request before its last | |
1352 | * request was served; | |
1353 | * | |
1354 | * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue | |
1355 | * a new request before the expiration of the idling-time. | |
1356 | * | |
1357 | * Even if bfqq has expired for one of the above reasons, the process | |
1358 | * associated with the queue may be however issuing requests greedily, | |
1359 | * and thus be sensitive to the bandwidth it receives (bfqq may have | |
1360 | * remained idle for other reasons: CPU high load, bfqq not enjoying | |
1361 | * idling, I/O throttling somewhere in the path from the process to | |
1362 | * the I/O scheduler, ...). But if, after every expiration for one of | |
1363 | * the above two reasons, bfqq has to wait for the service of at least | |
1364 | * one full budget of another queue before being served again, then | |
1365 | * bfqq is likely to get a much lower bandwidth or resource time than | |
1366 | * its reserved ones. To address this issue, two countermeasures need | |
1367 | * to be taken. | |
1368 | * | |
1369 | * First, the budget and the timestamps of bfqq need to be updated in | |
1370 | * a special way on bfqq reactivation: they need to be updated as if | |
1371 | * bfqq did not remain idle and did not expire. In fact, if they are | |
1372 | * computed as if bfqq expired and remained idle until reactivation, | |
1373 | * then the process associated with bfqq is treated as if, instead of | |
1374 | * being greedy, it stopped issuing requests when bfqq remained idle, | |
1375 | * and restarts issuing requests only on this reactivation. In other | |
1376 | * words, the scheduler does not help the process recover the "service | |
1377 | * hole" between bfqq expiration and reactivation. As a consequence, | |
1378 | * the process receives a lower bandwidth than its reserved one. In | |
1379 | * contrast, to recover this hole, the budget must be updated as if | |
1380 | * bfqq was not expired at all before this reactivation, i.e., it must | |
1381 | * be set to the value of the remaining budget when bfqq was | |
1382 | * expired. Along the same line, timestamps need to be assigned the | |
1383 | * value they had the last time bfqq was selected for service, i.e., | |
1384 | * before last expiration. Thus timestamps need to be back-shifted | |
1385 | * with respect to their normal computation (see [1] for more details | |
1386 | * on this tricky aspect). | |
1387 | * | |
1388 | * Secondly, to allow the process to recover the hole, the in-service | |
1389 | * queue must be expired too, to give bfqq the chance to preempt it | |
1390 | * immediately. In fact, if bfqq has to wait for a full budget of the | |
1391 | * in-service queue to be completed, then it may become impossible to | |
1392 | * let the process recover the hole, even if the back-shifted | |
1393 | * timestamps of bfqq are lower than those of the in-service queue. If | |
1394 | * this happens for most or all of the holes, then the process may not | |
1395 | * receive its reserved bandwidth. In this respect, it is worth noting | |
1396 | * that, being the service of outstanding requests unpreemptible, a | |
1397 | * little fraction of the holes may however be unrecoverable, thereby | |
1398 | * causing a little loss of bandwidth. | |
1399 | * | |
1400 | * The last important point is detecting whether bfqq does need this | |
1401 | * bandwidth recovery. In this respect, the next function deems the | |
1402 | * process associated with bfqq greedy, and thus allows it to recover | |
1403 | * the hole, if: 1) the process is waiting for the arrival of a new | |
1404 | * request (which implies that bfqq expired for one of the above two | |
1405 | * reasons), and 2) such a request has arrived soon. The first | |
1406 | * condition is controlled through the flag non_blocking_wait_rq, | |
1407 | * while the second through the flag arrived_in_time. If both | |
1408 | * conditions hold, then the function computes the budget in the | |
1409 | * above-described special way, and signals that the in-service queue | |
1410 | * should be expired. Timestamp back-shifting is done later in | |
1411 | * __bfq_activate_entity. | |
44e44a1b PV |
1412 | * |
1413 | * 2. Reduce latency. Even if timestamps are not backshifted to let | |
1414 | * the process associated with bfqq recover a service hole, bfqq may | |
1415 | * however happen to have, after being (re)activated, a lower finish | |
1416 | * timestamp than the in-service queue. That is, the next budget of | |
1417 | * bfqq may have to be completed before the one of the in-service | |
1418 | * queue. If this is the case, then preempting the in-service queue | |
1419 | * allows this goal to be achieved, apart from the unpreemptible, | |
1420 | * outstanding requests mentioned above. | |
1421 | * | |
1422 | * Unfortunately, regardless of which of the above two goals one wants | |
1423 | * to achieve, service trees need first to be updated to know whether | |
1424 | * the in-service queue must be preempted. To have service trees | |
1425 | * correctly updated, the in-service queue must be expired and | |
1426 | * rescheduled, and bfqq must be scheduled too. This is one of the | |
1427 | * most costly operations (in future versions, the scheduling | |
1428 | * mechanism may be re-designed in such a way to make it possible to | |
1429 | * know whether preemption is needed without needing to update service | |
1430 | * trees). In addition, queue preemptions almost always cause random | |
96a291c3 PV |
1431 | * I/O, which may in turn cause loss of throughput. Finally, there may |
1432 | * even be no in-service queue when the next function is invoked (so, | |
1433 | * no queue to compare timestamps with). Because of these facts, the | |
1434 | * next function adopts the following simple scheme to avoid costly | |
1435 | * operations, too frequent preemptions and too many dependencies on | |
1436 | * the state of the scheduler: it requests the expiration of the | |
1437 | * in-service queue (unconditionally) only for queues that need to | |
1438 | * recover a hole. Then it delegates to other parts of the code the | |
1439 | * responsibility of handling the above case 2. | |
aee69d78 PV |
1440 | */ |
1441 | static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd, | |
1442 | struct bfq_queue *bfqq, | |
96a291c3 | 1443 | bool arrived_in_time) |
aee69d78 PV |
1444 | { |
1445 | struct bfq_entity *entity = &bfqq->entity; | |
1446 | ||
218cb897 PV |
1447 | /* |
1448 | * In the next compound condition, we check also whether there | |
1449 | * is some budget left, because otherwise there is no point in | |
1450 | * trying to go on serving bfqq with this same budget: bfqq | |
1451 | * would be expired immediately after being selected for | |
1452 | * service. This would only cause useless overhead. | |
1453 | */ | |
1454 | if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time && | |
1455 | bfq_bfqq_budget_left(bfqq) > 0) { | |
aee69d78 PV |
1456 | /* |
1457 | * We do not clear the flag non_blocking_wait_rq here, as | |
1458 | * the latter is used in bfq_activate_bfqq to signal | |
1459 | * that timestamps need to be back-shifted (and is | |
1460 | * cleared right after). | |
1461 | */ | |
1462 | ||
1463 | /* | |
1464 | * In next assignment we rely on that either | |
1465 | * entity->service or entity->budget are not updated | |
1466 | * on expiration if bfqq is empty (see | |
1467 | * __bfq_bfqq_recalc_budget). Thus both quantities | |
1468 | * remain unchanged after such an expiration, and the | |
1469 | * following statement therefore assigns to | |
1470 | * entity->budget the remaining budget on such an | |
9fae8dd5 | 1471 | * expiration. |
aee69d78 PV |
1472 | */ |
1473 | entity->budget = min_t(unsigned long, | |
1474 | bfq_bfqq_budget_left(bfqq), | |
1475 | bfqq->max_budget); | |
1476 | ||
9fae8dd5 PV |
1477 | /* |
1478 | * At this point, we have used entity->service to get | |
1479 | * the budget left (needed for updating | |
1480 | * entity->budget). Thus we finally can, and have to, | |
1481 | * reset entity->service. The latter must be reset | |
1482 | * because bfqq would otherwise be charged again for | |
1483 | * the service it has received during its previous | |
1484 | * service slot(s). | |
1485 | */ | |
1486 | entity->service = 0; | |
1487 | ||
aee69d78 PV |
1488 | return true; |
1489 | } | |
1490 | ||
9fae8dd5 PV |
1491 | /* |
1492 | * We can finally complete expiration, by setting service to 0. | |
1493 | */ | |
1494 | entity->service = 0; | |
aee69d78 PV |
1495 | entity->budget = max_t(unsigned long, bfqq->max_budget, |
1496 | bfq_serv_to_charge(bfqq->next_rq, bfqq)); | |
1497 | bfq_clear_bfqq_non_blocking_wait_rq(bfqq); | |
96a291c3 | 1498 | return false; |
44e44a1b PV |
1499 | } |
1500 | ||
4baa8bb1 PV |
1501 | /* |
1502 | * Return the farthest past time instant according to jiffies | |
1503 | * macros. | |
1504 | */ | |
1505 | static unsigned long bfq_smallest_from_now(void) | |
1506 | { | |
1507 | return jiffies - MAX_JIFFY_OFFSET; | |
1508 | } | |
1509 | ||
44e44a1b PV |
1510 | static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd, |
1511 | struct bfq_queue *bfqq, | |
1512 | unsigned int old_wr_coeff, | |
1513 | bool wr_or_deserves_wr, | |
77b7dcea | 1514 | bool interactive, |
e1b2324d | 1515 | bool in_burst, |
77b7dcea | 1516 | bool soft_rt) |
44e44a1b PV |
1517 | { |
1518 | if (old_wr_coeff == 1 && wr_or_deserves_wr) { | |
1519 | /* start a weight-raising period */ | |
77b7dcea | 1520 | if (interactive) { |
8a8747dc | 1521 | bfqq->service_from_wr = 0; |
77b7dcea PV |
1522 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; |
1523 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
1524 | } else { | |
4baa8bb1 PV |
1525 | /* |
1526 | * No interactive weight raising in progress | |
1527 | * here: assign minus infinity to | |
1528 | * wr_start_at_switch_to_srt, to make sure | |
1529 | * that, at the end of the soft-real-time | |
1530 | * weight raising periods that is starting | |
1531 | * now, no interactive weight-raising period | |
1532 | * may be wrongly considered as still in | |
1533 | * progress (and thus actually started by | |
1534 | * mistake). | |
1535 | */ | |
1536 | bfqq->wr_start_at_switch_to_srt = | |
1537 | bfq_smallest_from_now(); | |
77b7dcea PV |
1538 | bfqq->wr_coeff = bfqd->bfq_wr_coeff * |
1539 | BFQ_SOFTRT_WEIGHT_FACTOR; | |
1540 | bfqq->wr_cur_max_time = | |
1541 | bfqd->bfq_wr_rt_max_time; | |
1542 | } | |
44e44a1b PV |
1543 | |
1544 | /* | |
1545 | * If needed, further reduce budget to make sure it is | |
1546 | * close to bfqq's backlog, so as to reduce the | |
1547 | * scheduling-error component due to a too large | |
1548 | * budget. Do not care about throughput consequences, | |
1549 | * but only about latency. Finally, do not assign a | |
1550 | * too small budget either, to avoid increasing | |
1551 | * latency by causing too frequent expirations. | |
1552 | */ | |
1553 | bfqq->entity.budget = min_t(unsigned long, | |
1554 | bfqq->entity.budget, | |
1555 | 2 * bfq_min_budget(bfqd)); | |
1556 | } else if (old_wr_coeff > 1) { | |
77b7dcea PV |
1557 | if (interactive) { /* update wr coeff and duration */ |
1558 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
1559 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
e1b2324d AA |
1560 | } else if (in_burst) |
1561 | bfqq->wr_coeff = 1; | |
1562 | else if (soft_rt) { | |
77b7dcea PV |
1563 | /* |
1564 | * The application is now or still meeting the | |
1565 | * requirements for being deemed soft rt. We | |
1566 | * can then correctly and safely (re)charge | |
1567 | * the weight-raising duration for the | |
1568 | * application with the weight-raising | |
1569 | * duration for soft rt applications. | |
1570 | * | |
1571 | * In particular, doing this recharge now, i.e., | |
1572 | * before the weight-raising period for the | |
1573 | * application finishes, reduces the probability | |
1574 | * of the following negative scenario: | |
1575 | * 1) the weight of a soft rt application is | |
1576 | * raised at startup (as for any newly | |
1577 | * created application), | |
1578 | * 2) since the application is not interactive, | |
1579 | * at a certain time weight-raising is | |
1580 | * stopped for the application, | |
1581 | * 3) at that time the application happens to | |
1582 | * still have pending requests, and hence | |
1583 | * is destined to not have a chance to be | |
1584 | * deemed soft rt before these requests are | |
1585 | * completed (see the comments to the | |
1586 | * function bfq_bfqq_softrt_next_start() | |
1587 | * for details on soft rt detection), | |
1588 | * 4) these pending requests experience a high | |
1589 | * latency because the application is not | |
1590 | * weight-raised while they are pending. | |
1591 | */ | |
1592 | if (bfqq->wr_cur_max_time != | |
1593 | bfqd->bfq_wr_rt_max_time) { | |
1594 | bfqq->wr_start_at_switch_to_srt = | |
1595 | bfqq->last_wr_start_finish; | |
1596 | ||
1597 | bfqq->wr_cur_max_time = | |
1598 | bfqd->bfq_wr_rt_max_time; | |
1599 | bfqq->wr_coeff = bfqd->bfq_wr_coeff * | |
1600 | BFQ_SOFTRT_WEIGHT_FACTOR; | |
1601 | } | |
1602 | bfqq->last_wr_start_finish = jiffies; | |
1603 | } | |
44e44a1b PV |
1604 | } |
1605 | } | |
1606 | ||
1607 | static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd, | |
1608 | struct bfq_queue *bfqq) | |
1609 | { | |
1610 | return bfqq->dispatched == 0 && | |
1611 | time_is_before_jiffies( | |
1612 | bfqq->budget_timeout + | |
1613 | bfqd->bfq_wr_min_idle_time); | |
aee69d78 PV |
1614 | } |
1615 | ||
96a291c3 PV |
1616 | |
1617 | /* | |
1618 | * Return true if bfqq is in a higher priority class, or has a higher | |
1619 | * weight than the in-service queue. | |
1620 | */ | |
1621 | static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq, | |
1622 | struct bfq_queue *in_serv_bfqq) | |
1623 | { | |
1624 | int bfqq_weight, in_serv_weight; | |
1625 | ||
1626 | if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class) | |
1627 | return true; | |
1628 | ||
1629 | if (in_serv_bfqq->entity.parent == bfqq->entity.parent) { | |
1630 | bfqq_weight = bfqq->entity.weight; | |
1631 | in_serv_weight = in_serv_bfqq->entity.weight; | |
1632 | } else { | |
1633 | if (bfqq->entity.parent) | |
1634 | bfqq_weight = bfqq->entity.parent->weight; | |
1635 | else | |
1636 | bfqq_weight = bfqq->entity.weight; | |
1637 | if (in_serv_bfqq->entity.parent) | |
1638 | in_serv_weight = in_serv_bfqq->entity.parent->weight; | |
1639 | else | |
1640 | in_serv_weight = in_serv_bfqq->entity.weight; | |
1641 | } | |
1642 | ||
1643 | return bfqq_weight > in_serv_weight; | |
1644 | } | |
1645 | ||
aee69d78 PV |
1646 | static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd, |
1647 | struct bfq_queue *bfqq, | |
44e44a1b PV |
1648 | int old_wr_coeff, |
1649 | struct request *rq, | |
1650 | bool *interactive) | |
aee69d78 | 1651 | { |
e1b2324d AA |
1652 | bool soft_rt, in_burst, wr_or_deserves_wr, |
1653 | bfqq_wants_to_preempt, | |
44e44a1b | 1654 | idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq), |
aee69d78 PV |
1655 | /* |
1656 | * See the comments on | |
1657 | * bfq_bfqq_update_budg_for_activation for | |
1658 | * details on the usage of the next variable. | |
1659 | */ | |
1660 | arrived_in_time = ktime_get_ns() <= | |
1661 | bfqq->ttime.last_end_request + | |
1662 | bfqd->bfq_slice_idle * 3; | |
1663 | ||
e21b7a0b | 1664 | |
aee69d78 | 1665 | /* |
44e44a1b PV |
1666 | * bfqq deserves to be weight-raised if: |
1667 | * - it is sync, | |
e1b2324d | 1668 | * - it does not belong to a large burst, |
36eca894 AA |
1669 | * - it has been idle for enough time or is soft real-time, |
1670 | * - is linked to a bfq_io_cq (it is not shared in any sense). | |
44e44a1b | 1671 | */ |
e1b2324d | 1672 | in_burst = bfq_bfqq_in_large_burst(bfqq); |
77b7dcea | 1673 | soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 && |
7074f076 | 1674 | !BFQQ_TOTALLY_SEEKY(bfqq) && |
e1b2324d | 1675 | !in_burst && |
f6c3ca0e DS |
1676 | time_is_before_jiffies(bfqq->soft_rt_next_start) && |
1677 | bfqq->dispatched == 0; | |
e1b2324d | 1678 | *interactive = !in_burst && idle_for_long_time; |
44e44a1b PV |
1679 | wr_or_deserves_wr = bfqd->low_latency && |
1680 | (bfqq->wr_coeff > 1 || | |
36eca894 AA |
1681 | (bfq_bfqq_sync(bfqq) && |
1682 | bfqq->bic && (*interactive || soft_rt))); | |
44e44a1b PV |
1683 | |
1684 | /* | |
1685 | * Using the last flag, update budget and check whether bfqq | |
1686 | * may want to preempt the in-service queue. | |
aee69d78 PV |
1687 | */ |
1688 | bfqq_wants_to_preempt = | |
1689 | bfq_bfqq_update_budg_for_activation(bfqd, bfqq, | |
96a291c3 | 1690 | arrived_in_time); |
aee69d78 | 1691 | |
e1b2324d AA |
1692 | /* |
1693 | * If bfqq happened to be activated in a burst, but has been | |
1694 | * idle for much more than an interactive queue, then we | |
1695 | * assume that, in the overall I/O initiated in the burst, the | |
1696 | * I/O associated with bfqq is finished. So bfqq does not need | |
1697 | * to be treated as a queue belonging to a burst | |
1698 | * anymore. Accordingly, we reset bfqq's in_large_burst flag | |
1699 | * if set, and remove bfqq from the burst list if it's | |
1700 | * there. We do not decrement burst_size, because the fact | |
1701 | * that bfqq does not need to belong to the burst list any | |
1702 | * more does not invalidate the fact that bfqq was created in | |
1703 | * a burst. | |
1704 | */ | |
1705 | if (likely(!bfq_bfqq_just_created(bfqq)) && | |
1706 | idle_for_long_time && | |
1707 | time_is_before_jiffies( | |
1708 | bfqq->budget_timeout + | |
1709 | msecs_to_jiffies(10000))) { | |
1710 | hlist_del_init(&bfqq->burst_list_node); | |
1711 | bfq_clear_bfqq_in_large_burst(bfqq); | |
1712 | } | |
1713 | ||
1714 | bfq_clear_bfqq_just_created(bfqq); | |
1715 | ||
1716 | ||
aee69d78 PV |
1717 | if (!bfq_bfqq_IO_bound(bfqq)) { |
1718 | if (arrived_in_time) { | |
1719 | bfqq->requests_within_timer++; | |
1720 | if (bfqq->requests_within_timer >= | |
1721 | bfqd->bfq_requests_within_timer) | |
1722 | bfq_mark_bfqq_IO_bound(bfqq); | |
1723 | } else | |
1724 | bfqq->requests_within_timer = 0; | |
1725 | } | |
1726 | ||
44e44a1b | 1727 | if (bfqd->low_latency) { |
36eca894 AA |
1728 | if (unlikely(time_is_after_jiffies(bfqq->split_time))) |
1729 | /* wraparound */ | |
1730 | bfqq->split_time = | |
1731 | jiffies - bfqd->bfq_wr_min_idle_time - 1; | |
1732 | ||
1733 | if (time_is_before_jiffies(bfqq->split_time + | |
1734 | bfqd->bfq_wr_min_idle_time)) { | |
1735 | bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq, | |
1736 | old_wr_coeff, | |
1737 | wr_or_deserves_wr, | |
1738 | *interactive, | |
e1b2324d | 1739 | in_burst, |
36eca894 AA |
1740 | soft_rt); |
1741 | ||
1742 | if (old_wr_coeff != bfqq->wr_coeff) | |
1743 | bfqq->entity.prio_changed = 1; | |
1744 | } | |
44e44a1b PV |
1745 | } |
1746 | ||
77b7dcea PV |
1747 | bfqq->last_idle_bklogged = jiffies; |
1748 | bfqq->service_from_backlogged = 0; | |
1749 | bfq_clear_bfqq_softrt_update(bfqq); | |
1750 | ||
aee69d78 PV |
1751 | bfq_add_bfqq_busy(bfqd, bfqq); |
1752 | ||
1753 | /* | |
1754 | * Expire in-service queue only if preemption may be needed | |
96a291c3 PV |
1755 | * for guarantees. In particular, we care only about two |
1756 | * cases. The first is that bfqq has to recover a service | |
1757 | * hole, as explained in the comments on | |
1758 | * bfq_bfqq_update_budg_for_activation(), i.e., that | |
1759 | * bfqq_wants_to_preempt is true. However, if bfqq does not | |
1760 | * carry time-critical I/O, then bfqq's bandwidth is less | |
1761 | * important than that of queues that carry time-critical I/O. | |
1762 | * So, as a further constraint, we consider this case only if | |
1763 | * bfqq is at least as weight-raised, i.e., at least as time | |
1764 | * critical, as the in-service queue. | |
1765 | * | |
1766 | * The second case is that bfqq is in a higher priority class, | |
1767 | * or has a higher weight than the in-service queue. If this | |
1768 | * condition does not hold, we don't care because, even if | |
1769 | * bfqq does not start to be served immediately, the resulting | |
1770 | * delay for bfqq's I/O is however lower or much lower than | |
1771 | * the ideal completion time to be guaranteed to bfqq's I/O. | |
1772 | * | |
1773 | * In both cases, preemption is needed only if, according to | |
1774 | * the timestamps of both bfqq and of the in-service queue, | |
1775 | * bfqq actually is the next queue to serve. So, to reduce | |
1776 | * useless preemptions, the return value of | |
1777 | * next_queue_may_preempt() is considered in the next compound | |
1778 | * condition too. Yet next_queue_may_preempt() just checks a | |
1779 | * simple, necessary condition for bfqq to be the next queue | |
1780 | * to serve. In fact, to evaluate a sufficient condition, the | |
1781 | * timestamps of the in-service queue would need to be | |
1782 | * updated, and this operation is quite costly (see the | |
1783 | * comments on bfq_bfqq_update_budg_for_activation()). | |
aee69d78 | 1784 | */ |
96a291c3 PV |
1785 | if (bfqd->in_service_queue && |
1786 | ((bfqq_wants_to_preempt && | |
1787 | bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) || | |
1788 | bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue)) && | |
aee69d78 PV |
1789 | next_queue_may_preempt(bfqd)) |
1790 | bfq_bfqq_expire(bfqd, bfqd->in_service_queue, | |
1791 | false, BFQQE_PREEMPTED); | |
1792 | } | |
1793 | ||
766d6141 PV |
1794 | static void bfq_reset_inject_limit(struct bfq_data *bfqd, |
1795 | struct bfq_queue *bfqq) | |
1796 | { | |
1797 | /* invalidate baseline total service time */ | |
1798 | bfqq->last_serv_time_ns = 0; | |
1799 | ||
1800 | /* | |
1801 | * Reset pointer in case we are waiting for | |
1802 | * some request completion. | |
1803 | */ | |
1804 | bfqd->waited_rq = NULL; | |
1805 | ||
1806 | /* | |
1807 | * If bfqq has a short think time, then start by setting the | |
1808 | * inject limit to 0 prudentially, because the service time of | |
1809 | * an injected I/O request may be higher than the think time | |
1810 | * of bfqq, and therefore, if one request was injected when | |
1811 | * bfqq remains empty, this injected request might delay the | |
1812 | * service of the next I/O request for bfqq significantly. In | |
1813 | * case bfqq can actually tolerate some injection, then the | |
1814 | * adaptive update will however raise the limit soon. This | |
1815 | * lucky circumstance holds exactly because bfqq has a short | |
1816 | * think time, and thus, after remaining empty, is likely to | |
1817 | * get new I/O enqueued---and then completed---before being | |
1818 | * expired. This is the very pattern that gives the | |
1819 | * limit-update algorithm the chance to measure the effect of | |
1820 | * injection on request service times, and then to update the | |
1821 | * limit accordingly. | |
1822 | * | |
1823 | * However, in the following special case, the inject limit is | |
1824 | * left to 1 even if the think time is short: bfqq's I/O is | |
1825 | * synchronized with that of some other queue, i.e., bfqq may | |
1826 | * receive new I/O only after the I/O of the other queue is | |
1827 | * completed. Keeping the inject limit to 1 allows the | |
1828 | * blocking I/O to be served while bfqq is in service. And | |
1829 | * this is very convenient both for bfqq and for overall | |
1830 | * throughput, as explained in detail in the comments in | |
1831 | * bfq_update_has_short_ttime(). | |
1832 | * | |
1833 | * On the opposite end, if bfqq has a long think time, then | |
1834 | * start directly by 1, because: | |
1835 | * a) on the bright side, keeping at most one request in | |
1836 | * service in the drive is unlikely to cause any harm to the | |
1837 | * latency of bfqq's requests, as the service time of a single | |
1838 | * request is likely to be lower than the think time of bfqq; | |
1839 | * b) on the downside, after becoming empty, bfqq is likely to | |
1840 | * expire before getting its next request. With this request | |
1841 | * arrival pattern, it is very hard to sample total service | |
1842 | * times and update the inject limit accordingly (see comments | |
1843 | * on bfq_update_inject_limit()). So the limit is likely to be | |
1844 | * never, or at least seldom, updated. As a consequence, by | |
1845 | * setting the limit to 1, we avoid that no injection ever | |
1846 | * occurs with bfqq. On the downside, this proactive step | |
1847 | * further reduces chances to actually compute the baseline | |
1848 | * total service time. Thus it reduces chances to execute the | |
1849 | * limit-update algorithm and possibly raise the limit to more | |
1850 | * than 1. | |
1851 | */ | |
1852 | if (bfq_bfqq_has_short_ttime(bfqq)) | |
1853 | bfqq->inject_limit = 0; | |
1854 | else | |
1855 | bfqq->inject_limit = 1; | |
1856 | ||
1857 | bfqq->decrease_time_jif = jiffies; | |
1858 | } | |
1859 | ||
aee69d78 PV |
1860 | static void bfq_add_request(struct request *rq) |
1861 | { | |
1862 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
1863 | struct bfq_data *bfqd = bfqq->bfqd; | |
1864 | struct request *next_rq, *prev; | |
44e44a1b PV |
1865 | unsigned int old_wr_coeff = bfqq->wr_coeff; |
1866 | bool interactive = false; | |
aee69d78 PV |
1867 | |
1868 | bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq)); | |
1869 | bfqq->queued[rq_is_sync(rq)]++; | |
1870 | bfqd->queued++; | |
1871 | ||
2341d662 | 1872 | if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) { |
13a857a4 PV |
1873 | /* |
1874 | * Detect whether bfqq's I/O seems synchronized with | |
1875 | * that of some other queue, i.e., whether bfqq, after | |
1876 | * remaining empty, happens to receive new I/O only | |
1877 | * right after some I/O request of the other queue has | |
1878 | * been completed. We call waker queue the other | |
1879 | * queue, and we assume, for simplicity, that bfqq may | |
1880 | * have at most one waker queue. | |
1881 | * | |
1882 | * A remarkable throughput boost can be reached by | |
1883 | * unconditionally injecting the I/O of the waker | |
1884 | * queue, every time a new bfq_dispatch_request | |
1885 | * happens to be invoked while I/O is being plugged | |
1886 | * for bfqq. In addition to boosting throughput, this | |
1887 | * unblocks bfqq's I/O, thereby improving bandwidth | |
1888 | * and latency for bfqq. Note that these same results | |
1889 | * may be achieved with the general injection | |
1890 | * mechanism, but less effectively. For details on | |
1891 | * this aspect, see the comments on the choice of the | |
1892 | * queue for injection in bfq_select_queue(). | |
1893 | * | |
1894 | * Turning back to the detection of a waker queue, a | |
1895 | * queue Q is deemed as a waker queue for bfqq if, for | |
1896 | * two consecutive times, bfqq happens to become non | |
1897 | * empty right after a request of Q has been | |
1898 | * completed. In particular, on the first time, Q is | |
1899 | * tentatively set as a candidate waker queue, while | |
1900 | * on the second time, the flag | |
1901 | * bfq_bfqq_has_waker(bfqq) is set to confirm that Q | |
1902 | * is a waker queue for bfqq. These detection steps | |
1903 | * are performed only if bfqq has a long think time, | |
1904 | * so as to make it more likely that bfqq's I/O is | |
1905 | * actually being blocked by a synchronization. This | |
1906 | * last filter, plus the above two-times requirement, | |
1907 | * make false positives less likely. | |
1908 | * | |
1909 | * NOTE | |
1910 | * | |
1911 | * The sooner a waker queue is detected, the sooner | |
1912 | * throughput can be boosted by injecting I/O from the | |
1913 | * waker queue. Fortunately, detection is likely to be | |
1914 | * actually fast, for the following reasons. While | |
1915 | * blocked by synchronization, bfqq has a long think | |
1916 | * time. This implies that bfqq's inject limit is at | |
1917 | * least equal to 1 (see the comments in | |
1918 | * bfq_update_inject_limit()). So, thanks to | |
1919 | * injection, the waker queue is likely to be served | |
1920 | * during the very first I/O-plugging time interval | |
1921 | * for bfqq. This triggers the first step of the | |
1922 | * detection mechanism. Thanks again to injection, the | |
1923 | * candidate waker queue is then likely to be | |
1924 | * confirmed no later than during the next | |
1925 | * I/O-plugging interval for bfqq. | |
1926 | */ | |
1927 | if (!bfq_bfqq_has_short_ttime(bfqq) && | |
1928 | ktime_get_ns() - bfqd->last_completion < | |
1929 | 200 * NSEC_PER_USEC) { | |
1930 | if (bfqd->last_completed_rq_bfqq != bfqq && | |
1931 | bfqd->last_completed_rq_bfqq != | |
1932 | bfqq->waker_bfqq) { | |
1933 | /* | |
1934 | * First synchronization detected with | |
1935 | * a candidate waker queue, or with a | |
1936 | * different candidate waker queue | |
1937 | * from the current one. | |
1938 | */ | |
1939 | bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq; | |
1940 | ||
1941 | /* | |
1942 | * If the waker queue disappears, then | |
1943 | * bfqq->waker_bfqq must be reset. To | |
1944 | * this goal, we maintain in each | |
1945 | * waker queue a list, woken_list, of | |
1946 | * all the queues that reference the | |
1947 | * waker queue through their | |
1948 | * waker_bfqq pointer. When the waker | |
1949 | * queue exits, the waker_bfqq pointer | |
1950 | * of all the queues in the woken_list | |
1951 | * is reset. | |
1952 | * | |
1953 | * In addition, if bfqq is already in | |
1954 | * the woken_list of a waker queue, | |
1955 | * then, before being inserted into | |
1956 | * the woken_list of a new waker | |
1957 | * queue, bfqq must be removed from | |
1958 | * the woken_list of the old waker | |
1959 | * queue. | |
1960 | */ | |
1961 | if (!hlist_unhashed(&bfqq->woken_list_node)) | |
1962 | hlist_del_init(&bfqq->woken_list_node); | |
1963 | hlist_add_head(&bfqq->woken_list_node, | |
1964 | &bfqd->last_completed_rq_bfqq->woken_list); | |
1965 | ||
1966 | bfq_clear_bfqq_has_waker(bfqq); | |
1967 | } else if (bfqd->last_completed_rq_bfqq == | |
1968 | bfqq->waker_bfqq && | |
1969 | !bfq_bfqq_has_waker(bfqq)) { | |
1970 | /* | |
1971 | * synchronization with waker_bfqq | |
1972 | * seen for the second time | |
1973 | */ | |
1974 | bfq_mark_bfqq_has_waker(bfqq); | |
1975 | } | |
1976 | } | |
1977 | ||
2341d662 PV |
1978 | /* |
1979 | * Periodically reset inject limit, to make sure that | |
1980 | * the latter eventually drops in case workload | |
1981 | * changes, see step (3) in the comments on | |
1982 | * bfq_update_inject_limit(). | |
1983 | */ | |
1984 | if (time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
766d6141 PV |
1985 | msecs_to_jiffies(1000))) |
1986 | bfq_reset_inject_limit(bfqd, bfqq); | |
2341d662 PV |
1987 | |
1988 | /* | |
1989 | * The following conditions must hold to setup a new | |
1990 | * sampling of total service time, and then a new | |
1991 | * update of the inject limit: | |
1992 | * - bfqq is in service, because the total service | |
1993 | * time is evaluated only for the I/O requests of | |
1994 | * the queues in service; | |
1995 | * - this is the right occasion to compute or to | |
1996 | * lower the baseline total service time, because | |
1997 | * there are actually no requests in the drive, | |
1998 | * or | |
1999 | * the baseline total service time is available, and | |
2000 | * this is the right occasion to compute the other | |
2001 | * quantity needed to update the inject limit, i.e., | |
2002 | * the total service time caused by the amount of | |
2003 | * injection allowed by the current value of the | |
2004 | * limit. It is the right occasion because injection | |
2005 | * has actually been performed during the service | |
2006 | * hole, and there are still in-flight requests, | |
2007 | * which are very likely to be exactly the injected | |
2008 | * requests, or part of them; | |
2009 | * - the minimum interval for sampling the total | |
2010 | * service time and updating the inject limit has | |
2011 | * elapsed. | |
2012 | */ | |
2013 | if (bfqq == bfqd->in_service_queue && | |
2014 | (bfqd->rq_in_driver == 0 || | |
2015 | (bfqq->last_serv_time_ns > 0 && | |
2016 | bfqd->rqs_injected && bfqd->rq_in_driver > 0)) && | |
2017 | time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
2018 | msecs_to_jiffies(100))) { | |
2019 | bfqd->last_empty_occupied_ns = ktime_get_ns(); | |
2020 | /* | |
2021 | * Start the state machine for measuring the | |
2022 | * total service time of rq: setting | |
2023 | * wait_dispatch will cause bfqd->waited_rq to | |
2024 | * be set when rq will be dispatched. | |
2025 | */ | |
2026 | bfqd->wait_dispatch = true; | |
2027 | bfqd->rqs_injected = false; | |
2028 | } | |
2029 | } | |
2030 | ||
aee69d78 PV |
2031 | elv_rb_add(&bfqq->sort_list, rq); |
2032 | ||
2033 | /* | |
2034 | * Check if this request is a better next-serve candidate. | |
2035 | */ | |
2036 | prev = bfqq->next_rq; | |
2037 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); | |
2038 | bfqq->next_rq = next_rq; | |
2039 | ||
36eca894 AA |
2040 | /* |
2041 | * Adjust priority tree position, if next_rq changes. | |
8cacc5ab | 2042 | * See comments on bfq_pos_tree_add_move() for the unlikely(). |
36eca894 | 2043 | */ |
8cacc5ab | 2044 | if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq)) |
36eca894 AA |
2045 | bfq_pos_tree_add_move(bfqd, bfqq); |
2046 | ||
aee69d78 | 2047 | if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */ |
44e44a1b PV |
2048 | bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff, |
2049 | rq, &interactive); | |
2050 | else { | |
2051 | if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) && | |
2052 | time_is_before_jiffies( | |
2053 | bfqq->last_wr_start_finish + | |
2054 | bfqd->bfq_wr_min_inter_arr_async)) { | |
2055 | bfqq->wr_coeff = bfqd->bfq_wr_coeff; | |
2056 | bfqq->wr_cur_max_time = bfq_wr_duration(bfqd); | |
2057 | ||
cfd69712 | 2058 | bfqd->wr_busy_queues++; |
44e44a1b PV |
2059 | bfqq->entity.prio_changed = 1; |
2060 | } | |
2061 | if (prev != bfqq->next_rq) | |
2062 | bfq_updated_next_req(bfqd, bfqq); | |
2063 | } | |
2064 | ||
2065 | /* | |
2066 | * Assign jiffies to last_wr_start_finish in the following | |
2067 | * cases: | |
2068 | * | |
2069 | * . if bfqq is not going to be weight-raised, because, for | |
2070 | * non weight-raised queues, last_wr_start_finish stores the | |
2071 | * arrival time of the last request; as of now, this piece | |
2072 | * of information is used only for deciding whether to | |
2073 | * weight-raise async queues | |
2074 | * | |
2075 | * . if bfqq is not weight-raised, because, if bfqq is now | |
2076 | * switching to weight-raised, then last_wr_start_finish | |
2077 | * stores the time when weight-raising starts | |
2078 | * | |
2079 | * . if bfqq is interactive, because, regardless of whether | |
2080 | * bfqq is currently weight-raised, the weight-raising | |
2081 | * period must start or restart (this case is considered | |
2082 | * separately because it is not detected by the above | |
2083 | * conditions, if bfqq is already weight-raised) | |
77b7dcea PV |
2084 | * |
2085 | * last_wr_start_finish has to be updated also if bfqq is soft | |
2086 | * real-time, because the weight-raising period is constantly | |
2087 | * restarted on idle-to-busy transitions for these queues, but | |
2088 | * this is already done in bfq_bfqq_handle_idle_busy_switch if | |
2089 | * needed. | |
44e44a1b PV |
2090 | */ |
2091 | if (bfqd->low_latency && | |
2092 | (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive)) | |
2093 | bfqq->last_wr_start_finish = jiffies; | |
aee69d78 PV |
2094 | } |
2095 | ||
2096 | static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, | |
2097 | struct bio *bio, | |
2098 | struct request_queue *q) | |
2099 | { | |
2100 | struct bfq_queue *bfqq = bfqd->bio_bfqq; | |
2101 | ||
2102 | ||
2103 | if (bfqq) | |
2104 | return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); | |
2105 | ||
2106 | return NULL; | |
2107 | } | |
2108 | ||
ab0e43e9 PV |
2109 | static sector_t get_sdist(sector_t last_pos, struct request *rq) |
2110 | { | |
2111 | if (last_pos) | |
2112 | return abs(blk_rq_pos(rq) - last_pos); | |
2113 | ||
2114 | return 0; | |
2115 | } | |
2116 | ||
aee69d78 PV |
2117 | #if 0 /* Still not clear if we can do without next two functions */ |
2118 | static void bfq_activate_request(struct request_queue *q, struct request *rq) | |
2119 | { | |
2120 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2121 | ||
2122 | bfqd->rq_in_driver++; | |
aee69d78 PV |
2123 | } |
2124 | ||
2125 | static void bfq_deactivate_request(struct request_queue *q, struct request *rq) | |
2126 | { | |
2127 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2128 | ||
2129 | bfqd->rq_in_driver--; | |
2130 | } | |
2131 | #endif | |
2132 | ||
2133 | static void bfq_remove_request(struct request_queue *q, | |
2134 | struct request *rq) | |
2135 | { | |
2136 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
2137 | struct bfq_data *bfqd = bfqq->bfqd; | |
2138 | const int sync = rq_is_sync(rq); | |
2139 | ||
2140 | if (bfqq->next_rq == rq) { | |
2141 | bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); | |
2142 | bfq_updated_next_req(bfqd, bfqq); | |
2143 | } | |
2144 | ||
2145 | if (rq->queuelist.prev != &rq->queuelist) | |
2146 | list_del_init(&rq->queuelist); | |
2147 | bfqq->queued[sync]--; | |
2148 | bfqd->queued--; | |
2149 | elv_rb_del(&bfqq->sort_list, rq); | |
2150 | ||
2151 | elv_rqhash_del(q, rq); | |
2152 | if (q->last_merge == rq) | |
2153 | q->last_merge = NULL; | |
2154 | ||
2155 | if (RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
2156 | bfqq->next_rq = NULL; | |
2157 | ||
2158 | if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) { | |
e21b7a0b | 2159 | bfq_del_bfqq_busy(bfqd, bfqq, false); |
aee69d78 PV |
2160 | /* |
2161 | * bfqq emptied. In normal operation, when | |
2162 | * bfqq is empty, bfqq->entity.service and | |
2163 | * bfqq->entity.budget must contain, | |
2164 | * respectively, the service received and the | |
2165 | * budget used last time bfqq emptied. These | |
2166 | * facts do not hold in this case, as at least | |
2167 | * this last removal occurred while bfqq is | |
2168 | * not in service. To avoid inconsistencies, | |
2169 | * reset both bfqq->entity.service and | |
2170 | * bfqq->entity.budget, if bfqq has still a | |
2171 | * process that may issue I/O requests to it. | |
2172 | */ | |
2173 | bfqq->entity.budget = bfqq->entity.service = 0; | |
2174 | } | |
36eca894 AA |
2175 | |
2176 | /* | |
2177 | * Remove queue from request-position tree as it is empty. | |
2178 | */ | |
2179 | if (bfqq->pos_root) { | |
2180 | rb_erase(&bfqq->pos_node, bfqq->pos_root); | |
2181 | bfqq->pos_root = NULL; | |
2182 | } | |
05e90283 | 2183 | } else { |
8cacc5ab PV |
2184 | /* see comments on bfq_pos_tree_add_move() for the unlikely() */ |
2185 | if (unlikely(!bfqd->nonrot_with_queueing)) | |
2186 | bfq_pos_tree_add_move(bfqd, bfqq); | |
aee69d78 PV |
2187 | } |
2188 | ||
2189 | if (rq->cmd_flags & REQ_META) | |
2190 | bfqq->meta_pending--; | |
e21b7a0b | 2191 | |
aee69d78 PV |
2192 | } |
2193 | ||
14ccb66b CH |
2194 | static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio, |
2195 | unsigned int nr_segs) | |
aee69d78 PV |
2196 | { |
2197 | struct request_queue *q = hctx->queue; | |
2198 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2199 | struct request *free = NULL; | |
2200 | /* | |
2201 | * bfq_bic_lookup grabs the queue_lock: invoke it now and | |
2202 | * store its return value for later use, to avoid nesting | |
2203 | * queue_lock inside the bfqd->lock. We assume that the bic | |
2204 | * returned by bfq_bic_lookup does not go away before | |
2205 | * bfqd->lock is taken. | |
2206 | */ | |
2207 | struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q); | |
2208 | bool ret; | |
2209 | ||
2210 | spin_lock_irq(&bfqd->lock); | |
2211 | ||
2212 | if (bic) | |
2213 | bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf)); | |
2214 | else | |
2215 | bfqd->bio_bfqq = NULL; | |
2216 | bfqd->bio_bic = bic; | |
2217 | ||
14ccb66b | 2218 | ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free); |
aee69d78 PV |
2219 | |
2220 | if (free) | |
2221 | blk_mq_free_request(free); | |
2222 | spin_unlock_irq(&bfqd->lock); | |
2223 | ||
2224 | return ret; | |
2225 | } | |
2226 | ||
2227 | static int bfq_request_merge(struct request_queue *q, struct request **req, | |
2228 | struct bio *bio) | |
2229 | { | |
2230 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2231 | struct request *__rq; | |
2232 | ||
2233 | __rq = bfq_find_rq_fmerge(bfqd, bio, q); | |
2234 | if (__rq && elv_bio_merge_ok(__rq, bio)) { | |
2235 | *req = __rq; | |
2236 | return ELEVATOR_FRONT_MERGE; | |
2237 | } | |
2238 | ||
2239 | return ELEVATOR_NO_MERGE; | |
2240 | } | |
2241 | ||
18e5a57d PV |
2242 | static struct bfq_queue *bfq_init_rq(struct request *rq); |
2243 | ||
aee69d78 PV |
2244 | static void bfq_request_merged(struct request_queue *q, struct request *req, |
2245 | enum elv_merge type) | |
2246 | { | |
2247 | if (type == ELEVATOR_FRONT_MERGE && | |
2248 | rb_prev(&req->rb_node) && | |
2249 | blk_rq_pos(req) < | |
2250 | blk_rq_pos(container_of(rb_prev(&req->rb_node), | |
2251 | struct request, rb_node))) { | |
18e5a57d | 2252 | struct bfq_queue *bfqq = bfq_init_rq(req); |
aee69d78 PV |
2253 | struct bfq_data *bfqd = bfqq->bfqd; |
2254 | struct request *prev, *next_rq; | |
2255 | ||
2256 | /* Reposition request in its sort_list */ | |
2257 | elv_rb_del(&bfqq->sort_list, req); | |
2258 | elv_rb_add(&bfqq->sort_list, req); | |
2259 | ||
2260 | /* Choose next request to be served for bfqq */ | |
2261 | prev = bfqq->next_rq; | |
2262 | next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req, | |
2263 | bfqd->last_position); | |
2264 | bfqq->next_rq = next_rq; | |
2265 | /* | |
36eca894 AA |
2266 | * If next_rq changes, update both the queue's budget to |
2267 | * fit the new request and the queue's position in its | |
2268 | * rq_pos_tree. | |
aee69d78 | 2269 | */ |
36eca894 | 2270 | if (prev != bfqq->next_rq) { |
aee69d78 | 2271 | bfq_updated_next_req(bfqd, bfqq); |
8cacc5ab PV |
2272 | /* |
2273 | * See comments on bfq_pos_tree_add_move() for | |
2274 | * the unlikely(). | |
2275 | */ | |
2276 | if (unlikely(!bfqd->nonrot_with_queueing)) | |
2277 | bfq_pos_tree_add_move(bfqd, bfqq); | |
36eca894 | 2278 | } |
aee69d78 PV |
2279 | } |
2280 | } | |
2281 | ||
8abfa4d6 PV |
2282 | /* |
2283 | * This function is called to notify the scheduler that the requests | |
2284 | * rq and 'next' have been merged, with 'next' going away. BFQ | |
2285 | * exploits this hook to address the following issue: if 'next' has a | |
2286 | * fifo_time lower that rq, then the fifo_time of rq must be set to | |
2287 | * the value of 'next', to not forget the greater age of 'next'. | |
8abfa4d6 PV |
2288 | * |
2289 | * NOTE: in this function we assume that rq is in a bfq_queue, basing | |
2290 | * on that rq is picked from the hash table q->elevator->hash, which, | |
2291 | * in its turn, is filled only with I/O requests present in | |
2292 | * bfq_queues, while BFQ is in use for the request queue q. In fact, | |
2293 | * the function that fills this hash table (elv_rqhash_add) is called | |
2294 | * only by bfq_insert_request. | |
2295 | */ | |
aee69d78 PV |
2296 | static void bfq_requests_merged(struct request_queue *q, struct request *rq, |
2297 | struct request *next) | |
2298 | { | |
18e5a57d PV |
2299 | struct bfq_queue *bfqq = bfq_init_rq(rq), |
2300 | *next_bfqq = bfq_init_rq(next); | |
aee69d78 | 2301 | |
aee69d78 PV |
2302 | /* |
2303 | * If next and rq belong to the same bfq_queue and next is older | |
2304 | * than rq, then reposition rq in the fifo (by substituting next | |
2305 | * with rq). Otherwise, if next and rq belong to different | |
2306 | * bfq_queues, never reposition rq: in fact, we would have to | |
2307 | * reposition it with respect to next's position in its own fifo, | |
2308 | * which would most certainly be too expensive with respect to | |
2309 | * the benefits. | |
2310 | */ | |
2311 | if (bfqq == next_bfqq && | |
2312 | !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
2313 | next->fifo_time < rq->fifo_time) { | |
2314 | list_del_init(&rq->queuelist); | |
2315 | list_replace_init(&next->queuelist, &rq->queuelist); | |
2316 | rq->fifo_time = next->fifo_time; | |
2317 | } | |
2318 | ||
2319 | if (bfqq->next_rq == next) | |
2320 | bfqq->next_rq = rq; | |
2321 | ||
e21b7a0b | 2322 | bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags); |
aee69d78 PV |
2323 | } |
2324 | ||
44e44a1b PV |
2325 | /* Must be called with bfqq != NULL */ |
2326 | static void bfq_bfqq_end_wr(struct bfq_queue *bfqq) | |
2327 | { | |
cfd69712 PV |
2328 | if (bfq_bfqq_busy(bfqq)) |
2329 | bfqq->bfqd->wr_busy_queues--; | |
44e44a1b PV |
2330 | bfqq->wr_coeff = 1; |
2331 | bfqq->wr_cur_max_time = 0; | |
77b7dcea | 2332 | bfqq->last_wr_start_finish = jiffies; |
44e44a1b PV |
2333 | /* |
2334 | * Trigger a weight change on the next invocation of | |
2335 | * __bfq_entity_update_weight_prio. | |
2336 | */ | |
2337 | bfqq->entity.prio_changed = 1; | |
2338 | } | |
2339 | ||
ea25da48 PV |
2340 | void bfq_end_wr_async_queues(struct bfq_data *bfqd, |
2341 | struct bfq_group *bfqg) | |
44e44a1b PV |
2342 | { |
2343 | int i, j; | |
2344 | ||
2345 | for (i = 0; i < 2; i++) | |
2346 | for (j = 0; j < IOPRIO_BE_NR; j++) | |
2347 | if (bfqg->async_bfqq[i][j]) | |
2348 | bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]); | |
2349 | if (bfqg->async_idle_bfqq) | |
2350 | bfq_bfqq_end_wr(bfqg->async_idle_bfqq); | |
2351 | } | |
2352 | ||
2353 | static void bfq_end_wr(struct bfq_data *bfqd) | |
2354 | { | |
2355 | struct bfq_queue *bfqq; | |
2356 | ||
2357 | spin_lock_irq(&bfqd->lock); | |
2358 | ||
2359 | list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) | |
2360 | bfq_bfqq_end_wr(bfqq); | |
2361 | list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) | |
2362 | bfq_bfqq_end_wr(bfqq); | |
2363 | bfq_end_wr_async(bfqd); | |
2364 | ||
2365 | spin_unlock_irq(&bfqd->lock); | |
2366 | } | |
2367 | ||
36eca894 AA |
2368 | static sector_t bfq_io_struct_pos(void *io_struct, bool request) |
2369 | { | |
2370 | if (request) | |
2371 | return blk_rq_pos(io_struct); | |
2372 | else | |
2373 | return ((struct bio *)io_struct)->bi_iter.bi_sector; | |
2374 | } | |
2375 | ||
2376 | static int bfq_rq_close_to_sector(void *io_struct, bool request, | |
2377 | sector_t sector) | |
2378 | { | |
2379 | return abs(bfq_io_struct_pos(io_struct, request) - sector) <= | |
2380 | BFQQ_CLOSE_THR; | |
2381 | } | |
2382 | ||
2383 | static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd, | |
2384 | struct bfq_queue *bfqq, | |
2385 | sector_t sector) | |
2386 | { | |
2387 | struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree; | |
2388 | struct rb_node *parent, *node; | |
2389 | struct bfq_queue *__bfqq; | |
2390 | ||
2391 | if (RB_EMPTY_ROOT(root)) | |
2392 | return NULL; | |
2393 | ||
2394 | /* | |
2395 | * First, if we find a request starting at the end of the last | |
2396 | * request, choose it. | |
2397 | */ | |
2398 | __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); | |
2399 | if (__bfqq) | |
2400 | return __bfqq; | |
2401 | ||
2402 | /* | |
2403 | * If the exact sector wasn't found, the parent of the NULL leaf | |
2404 | * will contain the closest sector (rq_pos_tree sorted by | |
2405 | * next_request position). | |
2406 | */ | |
2407 | __bfqq = rb_entry(parent, struct bfq_queue, pos_node); | |
2408 | if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) | |
2409 | return __bfqq; | |
2410 | ||
2411 | if (blk_rq_pos(__bfqq->next_rq) < sector) | |
2412 | node = rb_next(&__bfqq->pos_node); | |
2413 | else | |
2414 | node = rb_prev(&__bfqq->pos_node); | |
2415 | if (!node) | |
2416 | return NULL; | |
2417 | ||
2418 | __bfqq = rb_entry(node, struct bfq_queue, pos_node); | |
2419 | if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) | |
2420 | return __bfqq; | |
2421 | ||
2422 | return NULL; | |
2423 | } | |
2424 | ||
2425 | static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd, | |
2426 | struct bfq_queue *cur_bfqq, | |
2427 | sector_t sector) | |
2428 | { | |
2429 | struct bfq_queue *bfqq; | |
2430 | ||
2431 | /* | |
2432 | * We shall notice if some of the queues are cooperating, | |
2433 | * e.g., working closely on the same area of the device. In | |
2434 | * that case, we can group them together and: 1) don't waste | |
2435 | * time idling, and 2) serve the union of their requests in | |
2436 | * the best possible order for throughput. | |
2437 | */ | |
2438 | bfqq = bfqq_find_close(bfqd, cur_bfqq, sector); | |
2439 | if (!bfqq || bfqq == cur_bfqq) | |
2440 | return NULL; | |
2441 | ||
2442 | return bfqq; | |
2443 | } | |
2444 | ||
2445 | static struct bfq_queue * | |
2446 | bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) | |
2447 | { | |
2448 | int process_refs, new_process_refs; | |
2449 | struct bfq_queue *__bfqq; | |
2450 | ||
2451 | /* | |
2452 | * If there are no process references on the new_bfqq, then it is | |
2453 | * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain | |
2454 | * may have dropped their last reference (not just their last process | |
2455 | * reference). | |
2456 | */ | |
2457 | if (!bfqq_process_refs(new_bfqq)) | |
2458 | return NULL; | |
2459 | ||
2460 | /* Avoid a circular list and skip interim queue merges. */ | |
2461 | while ((__bfqq = new_bfqq->new_bfqq)) { | |
2462 | if (__bfqq == bfqq) | |
2463 | return NULL; | |
2464 | new_bfqq = __bfqq; | |
2465 | } | |
2466 | ||
2467 | process_refs = bfqq_process_refs(bfqq); | |
2468 | new_process_refs = bfqq_process_refs(new_bfqq); | |
2469 | /* | |
2470 | * If the process for the bfqq has gone away, there is no | |
2471 | * sense in merging the queues. | |
2472 | */ | |
2473 | if (process_refs == 0 || new_process_refs == 0) | |
2474 | return NULL; | |
2475 | ||
2476 | bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", | |
2477 | new_bfqq->pid); | |
2478 | ||
2479 | /* | |
2480 | * Merging is just a redirection: the requests of the process | |
2481 | * owning one of the two queues are redirected to the other queue. | |
2482 | * The latter queue, in its turn, is set as shared if this is the | |
2483 | * first time that the requests of some process are redirected to | |
2484 | * it. | |
2485 | * | |
6fa3e8d3 PV |
2486 | * We redirect bfqq to new_bfqq and not the opposite, because |
2487 | * we are in the context of the process owning bfqq, thus we | |
2488 | * have the io_cq of this process. So we can immediately | |
2489 | * configure this io_cq to redirect the requests of the | |
2490 | * process to new_bfqq. In contrast, the io_cq of new_bfqq is | |
2491 | * not available any more (new_bfqq->bic == NULL). | |
36eca894 | 2492 | * |
6fa3e8d3 PV |
2493 | * Anyway, even in case new_bfqq coincides with the in-service |
2494 | * queue, redirecting requests the in-service queue is the | |
2495 | * best option, as we feed the in-service queue with new | |
2496 | * requests close to the last request served and, by doing so, | |
2497 | * are likely to increase the throughput. | |
36eca894 AA |
2498 | */ |
2499 | bfqq->new_bfqq = new_bfqq; | |
2500 | new_bfqq->ref += process_refs; | |
2501 | return new_bfqq; | |
2502 | } | |
2503 | ||
2504 | static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq, | |
2505 | struct bfq_queue *new_bfqq) | |
2506 | { | |
7b8fa3b9 PV |
2507 | if (bfq_too_late_for_merging(new_bfqq)) |
2508 | return false; | |
2509 | ||
36eca894 AA |
2510 | if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) || |
2511 | (bfqq->ioprio_class != new_bfqq->ioprio_class)) | |
2512 | return false; | |
2513 | ||
2514 | /* | |
2515 | * If either of the queues has already been detected as seeky, | |
2516 | * then merging it with the other queue is unlikely to lead to | |
2517 | * sequential I/O. | |
2518 | */ | |
2519 | if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq)) | |
2520 | return false; | |
2521 | ||
2522 | /* | |
2523 | * Interleaved I/O is known to be done by (some) applications | |
2524 | * only for reads, so it does not make sense to merge async | |
2525 | * queues. | |
2526 | */ | |
2527 | if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq)) | |
2528 | return false; | |
2529 | ||
2530 | return true; | |
2531 | } | |
2532 | ||
36eca894 AA |
2533 | /* |
2534 | * Attempt to schedule a merge of bfqq with the currently in-service | |
2535 | * queue or with a close queue among the scheduled queues. Return | |
2536 | * NULL if no merge was scheduled, a pointer to the shared bfq_queue | |
2537 | * structure otherwise. | |
2538 | * | |
2539 | * The OOM queue is not allowed to participate to cooperation: in fact, since | |
2540 | * the requests temporarily redirected to the OOM queue could be redirected | |
2541 | * again to dedicated queues at any time, the state needed to correctly | |
2542 | * handle merging with the OOM queue would be quite complex and expensive | |
2543 | * to maintain. Besides, in such a critical condition as an out of memory, | |
2544 | * the benefits of queue merging may be little relevant, or even negligible. | |
2545 | * | |
36eca894 AA |
2546 | * WARNING: queue merging may impair fairness among non-weight raised |
2547 | * queues, for at least two reasons: 1) the original weight of a | |
2548 | * merged queue may change during the merged state, 2) even being the | |
2549 | * weight the same, a merged queue may be bloated with many more | |
2550 | * requests than the ones produced by its originally-associated | |
2551 | * process. | |
2552 | */ | |
2553 | static struct bfq_queue * | |
2554 | bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
2555 | void *io_struct, bool request) | |
2556 | { | |
2557 | struct bfq_queue *in_service_bfqq, *new_bfqq; | |
2558 | ||
8cacc5ab PV |
2559 | /* |
2560 | * Do not perform queue merging if the device is non | |
2561 | * rotational and performs internal queueing. In fact, such a | |
2562 | * device reaches a high speed through internal parallelism | |
2563 | * and pipelining. This means that, to reach a high | |
2564 | * throughput, it must have many requests enqueued at the same | |
2565 | * time. But, in this configuration, the internal scheduling | |
2566 | * algorithm of the device does exactly the job of queue | |
2567 | * merging: it reorders requests so as to obtain as much as | |
2568 | * possible a sequential I/O pattern. As a consequence, with | |
2569 | * the workload generated by processes doing interleaved I/O, | |
2570 | * the throughput reached by the device is likely to be the | |
2571 | * same, with and without queue merging. | |
2572 | * | |
2573 | * Disabling merging also provides a remarkable benefit in | |
2574 | * terms of throughput. Merging tends to make many workloads | |
2575 | * artificially more uneven, because of shared queues | |
2576 | * remaining non empty for incomparably more time than | |
2577 | * non-merged queues. This may accentuate workload | |
2578 | * asymmetries. For example, if one of the queues in a set of | |
2579 | * merged queues has a higher weight than a normal queue, then | |
2580 | * the shared queue may inherit such a high weight and, by | |
2581 | * staying almost always active, may force BFQ to perform I/O | |
2582 | * plugging most of the time. This evidently makes it harder | |
2583 | * for BFQ to let the device reach a high throughput. | |
2584 | * | |
2585 | * Finally, the likely() macro below is not used because one | |
2586 | * of the two branches is more likely than the other, but to | |
2587 | * have the code path after the following if() executed as | |
2588 | * fast as possible for the case of a non rotational device | |
2589 | * with queueing. We want it because this is the fastest kind | |
2590 | * of device. On the opposite end, the likely() may lengthen | |
2591 | * the execution time of BFQ for the case of slower devices | |
2592 | * (rotational or at least without queueing). But in this case | |
2593 | * the execution time of BFQ matters very little, if not at | |
2594 | * all. | |
2595 | */ | |
2596 | if (likely(bfqd->nonrot_with_queueing)) | |
2597 | return NULL; | |
2598 | ||
7b8fa3b9 PV |
2599 | /* |
2600 | * Prevent bfqq from being merged if it has been created too | |
2601 | * long ago. The idea is that true cooperating processes, and | |
2602 | * thus their associated bfq_queues, are supposed to be | |
2603 | * created shortly after each other. This is the case, e.g., | |
2604 | * for KVM/QEMU and dump I/O threads. Basing on this | |
2605 | * assumption, the following filtering greatly reduces the | |
2606 | * probability that two non-cooperating processes, which just | |
2607 | * happen to do close I/O for some short time interval, have | |
2608 | * their queues merged by mistake. | |
2609 | */ | |
2610 | if (bfq_too_late_for_merging(bfqq)) | |
2611 | return NULL; | |
2612 | ||
36eca894 AA |
2613 | if (bfqq->new_bfqq) |
2614 | return bfqq->new_bfqq; | |
2615 | ||
4403e4e4 | 2616 | if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq)) |
36eca894 AA |
2617 | return NULL; |
2618 | ||
2619 | /* If there is only one backlogged queue, don't search. */ | |
73d58118 | 2620 | if (bfq_tot_busy_queues(bfqd) == 1) |
36eca894 AA |
2621 | return NULL; |
2622 | ||
2623 | in_service_bfqq = bfqd->in_service_queue; | |
2624 | ||
4403e4e4 AR |
2625 | if (in_service_bfqq && in_service_bfqq != bfqq && |
2626 | likely(in_service_bfqq != &bfqd->oom_bfqq) && | |
058fdecc PV |
2627 | bfq_rq_close_to_sector(io_struct, request, |
2628 | bfqd->in_serv_last_pos) && | |
36eca894 AA |
2629 | bfqq->entity.parent == in_service_bfqq->entity.parent && |
2630 | bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) { | |
2631 | new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq); | |
2632 | if (new_bfqq) | |
2633 | return new_bfqq; | |
2634 | } | |
2635 | /* | |
2636 | * Check whether there is a cooperator among currently scheduled | |
2637 | * queues. The only thing we need is that the bio/request is not | |
2638 | * NULL, as we need it to establish whether a cooperator exists. | |
2639 | */ | |
36eca894 AA |
2640 | new_bfqq = bfq_find_close_cooperator(bfqd, bfqq, |
2641 | bfq_io_struct_pos(io_struct, request)); | |
2642 | ||
4403e4e4 | 2643 | if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) && |
36eca894 AA |
2644 | bfq_may_be_close_cooperator(bfqq, new_bfqq)) |
2645 | return bfq_setup_merge(bfqq, new_bfqq); | |
2646 | ||
2647 | return NULL; | |
2648 | } | |
2649 | ||
2650 | static void bfq_bfqq_save_state(struct bfq_queue *bfqq) | |
2651 | { | |
2652 | struct bfq_io_cq *bic = bfqq->bic; | |
2653 | ||
2654 | /* | |
2655 | * If !bfqq->bic, the queue is already shared or its requests | |
2656 | * have already been redirected to a shared queue; both idle window | |
2657 | * and weight raising state have already been saved. Do nothing. | |
2658 | */ | |
2659 | if (!bic) | |
2660 | return; | |
2661 | ||
fffca087 | 2662 | bic->saved_weight = bfqq->entity.orig_weight; |
36eca894 | 2663 | bic->saved_ttime = bfqq->ttime; |
d5be3fef | 2664 | bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq); |
36eca894 | 2665 | bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq); |
e1b2324d AA |
2666 | bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq); |
2667 | bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node); | |
894df937 | 2668 | if (unlikely(bfq_bfqq_just_created(bfqq) && |
1be6e8a9 AR |
2669 | !bfq_bfqq_in_large_burst(bfqq) && |
2670 | bfqq->bfqd->low_latency)) { | |
894df937 PV |
2671 | /* |
2672 | * bfqq being merged right after being created: bfqq | |
2673 | * would have deserved interactive weight raising, but | |
2674 | * did not make it to be set in a weight-raised state, | |
2675 | * because of this early merge. Store directly the | |
2676 | * weight-raising state that would have been assigned | |
2677 | * to bfqq, so that to avoid that bfqq unjustly fails | |
2678 | * to enjoy weight raising if split soon. | |
2679 | */ | |
2680 | bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff; | |
2b50f230 | 2681 | bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now(); |
894df937 PV |
2682 | bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd); |
2683 | bic->saved_last_wr_start_finish = jiffies; | |
2684 | } else { | |
2685 | bic->saved_wr_coeff = bfqq->wr_coeff; | |
2686 | bic->saved_wr_start_at_switch_to_srt = | |
2687 | bfqq->wr_start_at_switch_to_srt; | |
2688 | bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish; | |
2689 | bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time; | |
2690 | } | |
36eca894 AA |
2691 | } |
2692 | ||
36eca894 AA |
2693 | static void |
2694 | bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, | |
2695 | struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) | |
2696 | { | |
2697 | bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", | |
2698 | (unsigned long)new_bfqq->pid); | |
2699 | /* Save weight raising and idle window of the merged queues */ | |
2700 | bfq_bfqq_save_state(bfqq); | |
2701 | bfq_bfqq_save_state(new_bfqq); | |
2702 | if (bfq_bfqq_IO_bound(bfqq)) | |
2703 | bfq_mark_bfqq_IO_bound(new_bfqq); | |
2704 | bfq_clear_bfqq_IO_bound(bfqq); | |
2705 | ||
2706 | /* | |
2707 | * If bfqq is weight-raised, then let new_bfqq inherit | |
2708 | * weight-raising. To reduce false positives, neglect the case | |
2709 | * where bfqq has just been created, but has not yet made it | |
2710 | * to be weight-raised (which may happen because EQM may merge | |
2711 | * bfqq even before bfq_add_request is executed for the first | |
e1b2324d AA |
2712 | * time for bfqq). Handling this case would however be very |
2713 | * easy, thanks to the flag just_created. | |
36eca894 AA |
2714 | */ |
2715 | if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) { | |
2716 | new_bfqq->wr_coeff = bfqq->wr_coeff; | |
2717 | new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time; | |
2718 | new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish; | |
2719 | new_bfqq->wr_start_at_switch_to_srt = | |
2720 | bfqq->wr_start_at_switch_to_srt; | |
2721 | if (bfq_bfqq_busy(new_bfqq)) | |
2722 | bfqd->wr_busy_queues++; | |
2723 | new_bfqq->entity.prio_changed = 1; | |
2724 | } | |
2725 | ||
2726 | if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */ | |
2727 | bfqq->wr_coeff = 1; | |
2728 | bfqq->entity.prio_changed = 1; | |
2729 | if (bfq_bfqq_busy(bfqq)) | |
2730 | bfqd->wr_busy_queues--; | |
2731 | } | |
2732 | ||
2733 | bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d", | |
2734 | bfqd->wr_busy_queues); | |
2735 | ||
36eca894 AA |
2736 | /* |
2737 | * Merge queues (that is, let bic redirect its requests to new_bfqq) | |
2738 | */ | |
2739 | bic_set_bfqq(bic, new_bfqq, 1); | |
2740 | bfq_mark_bfqq_coop(new_bfqq); | |
2741 | /* | |
2742 | * new_bfqq now belongs to at least two bics (it is a shared queue): | |
2743 | * set new_bfqq->bic to NULL. bfqq either: | |
2744 | * - does not belong to any bic any more, and hence bfqq->bic must | |
2745 | * be set to NULL, or | |
2746 | * - is a queue whose owning bics have already been redirected to a | |
2747 | * different queue, hence the queue is destined to not belong to | |
2748 | * any bic soon and bfqq->bic is already NULL (therefore the next | |
2749 | * assignment causes no harm). | |
2750 | */ | |
2751 | new_bfqq->bic = NULL; | |
1e66413c FP |
2752 | /* |
2753 | * If the queue is shared, the pid is the pid of one of the associated | |
2754 | * processes. Which pid depends on the exact sequence of merge events | |
2755 | * the queue underwent. So printing such a pid is useless and confusing | |
2756 | * because it reports a random pid between those of the associated | |
2757 | * processes. | |
2758 | * We mark such a queue with a pid -1, and then print SHARED instead of | |
2759 | * a pid in logging messages. | |
2760 | */ | |
2761 | new_bfqq->pid = -1; | |
36eca894 AA |
2762 | bfqq->bic = NULL; |
2763 | /* release process reference to bfqq */ | |
2764 | bfq_put_queue(bfqq); | |
2765 | } | |
2766 | ||
aee69d78 PV |
2767 | static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq, |
2768 | struct bio *bio) | |
2769 | { | |
2770 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
2771 | bool is_sync = op_is_sync(bio->bi_opf); | |
36eca894 | 2772 | struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq; |
aee69d78 PV |
2773 | |
2774 | /* | |
2775 | * Disallow merge of a sync bio into an async request. | |
2776 | */ | |
2777 | if (is_sync && !rq_is_sync(rq)) | |
2778 | return false; | |
2779 | ||
2780 | /* | |
2781 | * Lookup the bfqq that this bio will be queued with. Allow | |
2782 | * merge only if rq is queued there. | |
2783 | */ | |
2784 | if (!bfqq) | |
2785 | return false; | |
2786 | ||
36eca894 AA |
2787 | /* |
2788 | * We take advantage of this function to perform an early merge | |
2789 | * of the queues of possible cooperating processes. | |
2790 | */ | |
2791 | new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false); | |
2792 | if (new_bfqq) { | |
2793 | /* | |
2794 | * bic still points to bfqq, then it has not yet been | |
2795 | * redirected to some other bfq_queue, and a queue | |
636b8fe8 AR |
2796 | * merge between bfqq and new_bfqq can be safely |
2797 | * fulfilled, i.e., bic can be redirected to new_bfqq | |
36eca894 AA |
2798 | * and bfqq can be put. |
2799 | */ | |
2800 | bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq, | |
2801 | new_bfqq); | |
2802 | /* | |
2803 | * If we get here, bio will be queued into new_queue, | |
2804 | * so use new_bfqq to decide whether bio and rq can be | |
2805 | * merged. | |
2806 | */ | |
2807 | bfqq = new_bfqq; | |
2808 | ||
2809 | /* | |
2810 | * Change also bqfd->bio_bfqq, as | |
2811 | * bfqd->bio_bic now points to new_bfqq, and | |
2812 | * this function may be invoked again (and then may | |
2813 | * use again bqfd->bio_bfqq). | |
2814 | */ | |
2815 | bfqd->bio_bfqq = bfqq; | |
2816 | } | |
2817 | ||
aee69d78 PV |
2818 | return bfqq == RQ_BFQQ(rq); |
2819 | } | |
2820 | ||
44e44a1b PV |
2821 | /* |
2822 | * Set the maximum time for the in-service queue to consume its | |
2823 | * budget. This prevents seeky processes from lowering the throughput. | |
2824 | * In practice, a time-slice service scheme is used with seeky | |
2825 | * processes. | |
2826 | */ | |
2827 | static void bfq_set_budget_timeout(struct bfq_data *bfqd, | |
2828 | struct bfq_queue *bfqq) | |
2829 | { | |
77b7dcea PV |
2830 | unsigned int timeout_coeff; |
2831 | ||
2832 | if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time) | |
2833 | timeout_coeff = 1; | |
2834 | else | |
2835 | timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; | |
2836 | ||
44e44a1b PV |
2837 | bfqd->last_budget_start = ktime_get(); |
2838 | ||
2839 | bfqq->budget_timeout = jiffies + | |
77b7dcea | 2840 | bfqd->bfq_timeout * timeout_coeff; |
44e44a1b PV |
2841 | } |
2842 | ||
aee69d78 PV |
2843 | static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
2844 | struct bfq_queue *bfqq) | |
2845 | { | |
2846 | if (bfqq) { | |
aee69d78 PV |
2847 | bfq_clear_bfqq_fifo_expire(bfqq); |
2848 | ||
2849 | bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8; | |
2850 | ||
77b7dcea PV |
2851 | if (time_is_before_jiffies(bfqq->last_wr_start_finish) && |
2852 | bfqq->wr_coeff > 1 && | |
2853 | bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && | |
2854 | time_is_before_jiffies(bfqq->budget_timeout)) { | |
2855 | /* | |
2856 | * For soft real-time queues, move the start | |
2857 | * of the weight-raising period forward by the | |
2858 | * time the queue has not received any | |
2859 | * service. Otherwise, a relatively long | |
2860 | * service delay is likely to cause the | |
2861 | * weight-raising period of the queue to end, | |
2862 | * because of the short duration of the | |
2863 | * weight-raising period of a soft real-time | |
2864 | * queue. It is worth noting that this move | |
2865 | * is not so dangerous for the other queues, | |
2866 | * because soft real-time queues are not | |
2867 | * greedy. | |
2868 | * | |
2869 | * To not add a further variable, we use the | |
2870 | * overloaded field budget_timeout to | |
2871 | * determine for how long the queue has not | |
2872 | * received service, i.e., how much time has | |
2873 | * elapsed since the queue expired. However, | |
2874 | * this is a little imprecise, because | |
2875 | * budget_timeout is set to jiffies if bfqq | |
2876 | * not only expires, but also remains with no | |
2877 | * request. | |
2878 | */ | |
2879 | if (time_after(bfqq->budget_timeout, | |
2880 | bfqq->last_wr_start_finish)) | |
2881 | bfqq->last_wr_start_finish += | |
2882 | jiffies - bfqq->budget_timeout; | |
2883 | else | |
2884 | bfqq->last_wr_start_finish = jiffies; | |
2885 | } | |
2886 | ||
44e44a1b | 2887 | bfq_set_budget_timeout(bfqd, bfqq); |
aee69d78 PV |
2888 | bfq_log_bfqq(bfqd, bfqq, |
2889 | "set_in_service_queue, cur-budget = %d", | |
2890 | bfqq->entity.budget); | |
2891 | } | |
2892 | ||
2893 | bfqd->in_service_queue = bfqq; | |
2894 | } | |
2895 | ||
2896 | /* | |
2897 | * Get and set a new queue for service. | |
2898 | */ | |
2899 | static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd) | |
2900 | { | |
2901 | struct bfq_queue *bfqq = bfq_get_next_queue(bfqd); | |
2902 | ||
2903 | __bfq_set_in_service_queue(bfqd, bfqq); | |
2904 | return bfqq; | |
2905 | } | |
2906 | ||
aee69d78 PV |
2907 | static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
2908 | { | |
2909 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
aee69d78 PV |
2910 | u32 sl; |
2911 | ||
aee69d78 PV |
2912 | bfq_mark_bfqq_wait_request(bfqq); |
2913 | ||
2914 | /* | |
2915 | * We don't want to idle for seeks, but we do want to allow | |
2916 | * fair distribution of slice time for a process doing back-to-back | |
2917 | * seeks. So allow a little bit of time for him to submit a new rq. | |
2918 | */ | |
2919 | sl = bfqd->bfq_slice_idle; | |
2920 | /* | |
1de0c4cd AA |
2921 | * Unless the queue is being weight-raised or the scenario is |
2922 | * asymmetric, grant only minimum idle time if the queue | |
2923 | * is seeky. A long idling is preserved for a weight-raised | |
2924 | * queue, or, more in general, in an asymmetric scenario, | |
2925 | * because a long idling is needed for guaranteeing to a queue | |
2926 | * its reserved share of the throughput (in particular, it is | |
2927 | * needed if the queue has a higher weight than some other | |
2928 | * queue). | |
aee69d78 | 2929 | */ |
1de0c4cd | 2930 | if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 && |
fb53ac6c | 2931 | !bfq_asymmetric_scenario(bfqd, bfqq)) |
aee69d78 | 2932 | sl = min_t(u64, sl, BFQ_MIN_TT); |
778c02a2 PV |
2933 | else if (bfqq->wr_coeff > 1) |
2934 | sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC); | |
aee69d78 PV |
2935 | |
2936 | bfqd->last_idling_start = ktime_get(); | |
2341d662 PV |
2937 | bfqd->last_idling_start_jiffies = jiffies; |
2938 | ||
aee69d78 PV |
2939 | hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl), |
2940 | HRTIMER_MODE_REL); | |
e21b7a0b | 2941 | bfqg_stats_set_start_idle_time(bfqq_group(bfqq)); |
aee69d78 PV |
2942 | } |
2943 | ||
ab0e43e9 PV |
2944 | /* |
2945 | * In autotuning mode, max_budget is dynamically recomputed as the | |
2946 | * amount of sectors transferred in timeout at the estimated peak | |
2947 | * rate. This enables BFQ to utilize a full timeslice with a full | |
2948 | * budget, even if the in-service queue is served at peak rate. And | |
2949 | * this maximises throughput with sequential workloads. | |
2950 | */ | |
2951 | static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd) | |
2952 | { | |
2953 | return (u64)bfqd->peak_rate * USEC_PER_MSEC * | |
2954 | jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT; | |
2955 | } | |
2956 | ||
44e44a1b PV |
2957 | /* |
2958 | * Update parameters related to throughput and responsiveness, as a | |
2959 | * function of the estimated peak rate. See comments on | |
e24f1c24 | 2960 | * bfq_calc_max_budget(), and on the ref_wr_duration array. |
44e44a1b PV |
2961 | */ |
2962 | static void update_thr_responsiveness_params(struct bfq_data *bfqd) | |
2963 | { | |
e24f1c24 | 2964 | if (bfqd->bfq_user_max_budget == 0) { |
44e44a1b PV |
2965 | bfqd->bfq_max_budget = |
2966 | bfq_calc_max_budget(bfqd); | |
e24f1c24 | 2967 | bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget); |
44e44a1b | 2968 | } |
44e44a1b PV |
2969 | } |
2970 | ||
ab0e43e9 PV |
2971 | static void bfq_reset_rate_computation(struct bfq_data *bfqd, |
2972 | struct request *rq) | |
2973 | { | |
2974 | if (rq != NULL) { /* new rq dispatch now, reset accordingly */ | |
2975 | bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns(); | |
2976 | bfqd->peak_rate_samples = 1; | |
2977 | bfqd->sequential_samples = 0; | |
2978 | bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size = | |
2979 | blk_rq_sectors(rq); | |
2980 | } else /* no new rq dispatched, just reset the number of samples */ | |
2981 | bfqd->peak_rate_samples = 0; /* full re-init on next disp. */ | |
2982 | ||
2983 | bfq_log(bfqd, | |
2984 | "reset_rate_computation at end, sample %u/%u tot_sects %llu", | |
2985 | bfqd->peak_rate_samples, bfqd->sequential_samples, | |
2986 | bfqd->tot_sectors_dispatched); | |
2987 | } | |
2988 | ||
2989 | static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq) | |
2990 | { | |
2991 | u32 rate, weight, divisor; | |
2992 | ||
2993 | /* | |
2994 | * For the convergence property to hold (see comments on | |
2995 | * bfq_update_peak_rate()) and for the assessment to be | |
2996 | * reliable, a minimum number of samples must be present, and | |
2997 | * a minimum amount of time must have elapsed. If not so, do | |
2998 | * not compute new rate. Just reset parameters, to get ready | |
2999 | * for a new evaluation attempt. | |
3000 | */ | |
3001 | if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES || | |
3002 | bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL) | |
3003 | goto reset_computation; | |
3004 | ||
3005 | /* | |
3006 | * If a new request completion has occurred after last | |
3007 | * dispatch, then, to approximate the rate at which requests | |
3008 | * have been served by the device, it is more precise to | |
3009 | * extend the observation interval to the last completion. | |
3010 | */ | |
3011 | bfqd->delta_from_first = | |
3012 | max_t(u64, bfqd->delta_from_first, | |
3013 | bfqd->last_completion - bfqd->first_dispatch); | |
3014 | ||
3015 | /* | |
3016 | * Rate computed in sects/usec, and not sects/nsec, for | |
3017 | * precision issues. | |
3018 | */ | |
3019 | rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT, | |
3020 | div_u64(bfqd->delta_from_first, NSEC_PER_USEC)); | |
3021 | ||
3022 | /* | |
3023 | * Peak rate not updated if: | |
3024 | * - the percentage of sequential dispatches is below 3/4 of the | |
3025 | * total, and rate is below the current estimated peak rate | |
3026 | * - rate is unreasonably high (> 20M sectors/sec) | |
3027 | */ | |
3028 | if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 && | |
3029 | rate <= bfqd->peak_rate) || | |
3030 | rate > 20<<BFQ_RATE_SHIFT) | |
3031 | goto reset_computation; | |
3032 | ||
3033 | /* | |
3034 | * We have to update the peak rate, at last! To this purpose, | |
3035 | * we use a low-pass filter. We compute the smoothing constant | |
3036 | * of the filter as a function of the 'weight' of the new | |
3037 | * measured rate. | |
3038 | * | |
3039 | * As can be seen in next formulas, we define this weight as a | |
3040 | * quantity proportional to how sequential the workload is, | |
3041 | * and to how long the observation time interval is. | |
3042 | * | |
3043 | * The weight runs from 0 to 8. The maximum value of the | |
3044 | * weight, 8, yields the minimum value for the smoothing | |
3045 | * constant. At this minimum value for the smoothing constant, | |
3046 | * the measured rate contributes for half of the next value of | |
3047 | * the estimated peak rate. | |
3048 | * | |
3049 | * So, the first step is to compute the weight as a function | |
3050 | * of how sequential the workload is. Note that the weight | |
3051 | * cannot reach 9, because bfqd->sequential_samples cannot | |
3052 | * become equal to bfqd->peak_rate_samples, which, in its | |
3053 | * turn, holds true because bfqd->sequential_samples is not | |
3054 | * incremented for the first sample. | |
3055 | */ | |
3056 | weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples; | |
3057 | ||
3058 | /* | |
3059 | * Second step: further refine the weight as a function of the | |
3060 | * duration of the observation interval. | |
3061 | */ | |
3062 | weight = min_t(u32, 8, | |
3063 | div_u64(weight * bfqd->delta_from_first, | |
3064 | BFQ_RATE_REF_INTERVAL)); | |
3065 | ||
3066 | /* | |
3067 | * Divisor ranging from 10, for minimum weight, to 2, for | |
3068 | * maximum weight. | |
3069 | */ | |
3070 | divisor = 10 - weight; | |
3071 | ||
3072 | /* | |
3073 | * Finally, update peak rate: | |
3074 | * | |
3075 | * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor | |
3076 | */ | |
3077 | bfqd->peak_rate *= divisor-1; | |
3078 | bfqd->peak_rate /= divisor; | |
3079 | rate /= divisor; /* smoothing constant alpha = 1/divisor */ | |
3080 | ||
3081 | bfqd->peak_rate += rate; | |
bc56e2ca PV |
3082 | |
3083 | /* | |
3084 | * For a very slow device, bfqd->peak_rate can reach 0 (see | |
3085 | * the minimum representable values reported in the comments | |
3086 | * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid | |
3087 | * divisions by zero where bfqd->peak_rate is used as a | |
3088 | * divisor. | |
3089 | */ | |
3090 | bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate); | |
3091 | ||
44e44a1b | 3092 | update_thr_responsiveness_params(bfqd); |
ab0e43e9 PV |
3093 | |
3094 | reset_computation: | |
3095 | bfq_reset_rate_computation(bfqd, rq); | |
3096 | } | |
3097 | ||
3098 | /* | |
3099 | * Update the read/write peak rate (the main quantity used for | |
3100 | * auto-tuning, see update_thr_responsiveness_params()). | |
3101 | * | |
3102 | * It is not trivial to estimate the peak rate (correctly): because of | |
3103 | * the presence of sw and hw queues between the scheduler and the | |
3104 | * device components that finally serve I/O requests, it is hard to | |
3105 | * say exactly when a given dispatched request is served inside the | |
3106 | * device, and for how long. As a consequence, it is hard to know | |
3107 | * precisely at what rate a given set of requests is actually served | |
3108 | * by the device. | |
3109 | * | |
3110 | * On the opposite end, the dispatch time of any request is trivially | |
3111 | * available, and, from this piece of information, the "dispatch rate" | |
3112 | * of requests can be immediately computed. So, the idea in the next | |
3113 | * function is to use what is known, namely request dispatch times | |
3114 | * (plus, when useful, request completion times), to estimate what is | |
3115 | * unknown, namely in-device request service rate. | |
3116 | * | |
3117 | * The main issue is that, because of the above facts, the rate at | |
3118 | * which a certain set of requests is dispatched over a certain time | |
3119 | * interval can vary greatly with respect to the rate at which the | |
3120 | * same requests are then served. But, since the size of any | |
3121 | * intermediate queue is limited, and the service scheme is lossless | |
3122 | * (no request is silently dropped), the following obvious convergence | |
3123 | * property holds: the number of requests dispatched MUST become | |
3124 | * closer and closer to the number of requests completed as the | |
3125 | * observation interval grows. This is the key property used in | |
3126 | * the next function to estimate the peak service rate as a function | |
3127 | * of the observed dispatch rate. The function assumes to be invoked | |
3128 | * on every request dispatch. | |
3129 | */ | |
3130 | static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq) | |
3131 | { | |
3132 | u64 now_ns = ktime_get_ns(); | |
3133 | ||
3134 | if (bfqd->peak_rate_samples == 0) { /* first dispatch */ | |
3135 | bfq_log(bfqd, "update_peak_rate: goto reset, samples %d", | |
3136 | bfqd->peak_rate_samples); | |
3137 | bfq_reset_rate_computation(bfqd, rq); | |
3138 | goto update_last_values; /* will add one sample */ | |
3139 | } | |
3140 | ||
3141 | /* | |
3142 | * Device idle for very long: the observation interval lasting | |
3143 | * up to this dispatch cannot be a valid observation interval | |
3144 | * for computing a new peak rate (similarly to the late- | |
3145 | * completion event in bfq_completed_request()). Go to | |
3146 | * update_rate_and_reset to have the following three steps | |
3147 | * taken: | |
3148 | * - close the observation interval at the last (previous) | |
3149 | * request dispatch or completion | |
3150 | * - compute rate, if possible, for that observation interval | |
3151 | * - start a new observation interval with this dispatch | |
3152 | */ | |
3153 | if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC && | |
3154 | bfqd->rq_in_driver == 0) | |
3155 | goto update_rate_and_reset; | |
3156 | ||
3157 | /* Update sampling information */ | |
3158 | bfqd->peak_rate_samples++; | |
3159 | ||
3160 | if ((bfqd->rq_in_driver > 0 || | |
3161 | now_ns - bfqd->last_completion < BFQ_MIN_TT) | |
d87447d8 | 3162 | && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq)) |
ab0e43e9 PV |
3163 | bfqd->sequential_samples++; |
3164 | ||
3165 | bfqd->tot_sectors_dispatched += blk_rq_sectors(rq); | |
3166 | ||
3167 | /* Reset max observed rq size every 32 dispatches */ | |
3168 | if (likely(bfqd->peak_rate_samples % 32)) | |
3169 | bfqd->last_rq_max_size = | |
3170 | max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size); | |
3171 | else | |
3172 | bfqd->last_rq_max_size = blk_rq_sectors(rq); | |
3173 | ||
3174 | bfqd->delta_from_first = now_ns - bfqd->first_dispatch; | |
3175 | ||
3176 | /* Target observation interval not yet reached, go on sampling */ | |
3177 | if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL) | |
3178 | goto update_last_values; | |
3179 | ||
3180 | update_rate_and_reset: | |
3181 | bfq_update_rate_reset(bfqd, rq); | |
3182 | update_last_values: | |
3183 | bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
058fdecc PV |
3184 | if (RQ_BFQQ(rq) == bfqd->in_service_queue) |
3185 | bfqd->in_serv_last_pos = bfqd->last_position; | |
ab0e43e9 PV |
3186 | bfqd->last_dispatch = now_ns; |
3187 | } | |
3188 | ||
aee69d78 PV |
3189 | /* |
3190 | * Remove request from internal lists. | |
3191 | */ | |
3192 | static void bfq_dispatch_remove(struct request_queue *q, struct request *rq) | |
3193 | { | |
3194 | struct bfq_queue *bfqq = RQ_BFQQ(rq); | |
3195 | ||
3196 | /* | |
3197 | * For consistency, the next instruction should have been | |
3198 | * executed after removing the request from the queue and | |
3199 | * dispatching it. We execute instead this instruction before | |
3200 | * bfq_remove_request() (and hence introduce a temporary | |
3201 | * inconsistency), for efficiency. In fact, should this | |
3202 | * dispatch occur for a non in-service bfqq, this anticipated | |
3203 | * increment prevents two counters related to bfqq->dispatched | |
3204 | * from risking to be, first, uselessly decremented, and then | |
3205 | * incremented again when the (new) value of bfqq->dispatched | |
3206 | * happens to be taken into account. | |
3207 | */ | |
3208 | bfqq->dispatched++; | |
ab0e43e9 | 3209 | bfq_update_peak_rate(q->elevator->elevator_data, rq); |
aee69d78 PV |
3210 | |
3211 | bfq_remove_request(q, rq); | |
3212 | } | |
3213 | ||
3726112e PV |
3214 | /* |
3215 | * There is a case where idling does not have to be performed for | |
3216 | * throughput concerns, but to preserve the throughput share of | |
3217 | * the process associated with bfqq. | |
3218 | * | |
3219 | * To introduce this case, we can note that allowing the drive | |
3220 | * to enqueue more than one request at a time, and hence | |
3221 | * delegating de facto final scheduling decisions to the | |
3222 | * drive's internal scheduler, entails loss of control on the | |
3223 | * actual request service order. In particular, the critical | |
3224 | * situation is when requests from different processes happen | |
3225 | * to be present, at the same time, in the internal queue(s) | |
3226 | * of the drive. In such a situation, the drive, by deciding | |
3227 | * the service order of the internally-queued requests, does | |
3228 | * determine also the actual throughput distribution among | |
3229 | * these processes. But the drive typically has no notion or | |
3230 | * concern about per-process throughput distribution, and | |
3231 | * makes its decisions only on a per-request basis. Therefore, | |
3232 | * the service distribution enforced by the drive's internal | |
3233 | * scheduler is likely to coincide with the desired throughput | |
3234 | * distribution only in a completely symmetric, or favorably | |
3235 | * skewed scenario where: | |
3236 | * (i-a) each of these processes must get the same throughput as | |
3237 | * the others, | |
3238 | * (i-b) in case (i-a) does not hold, it holds that the process | |
3239 | * associated with bfqq must receive a lower or equal | |
3240 | * throughput than any of the other processes; | |
3241 | * (ii) the I/O of each process has the same properties, in | |
3242 | * terms of locality (sequential or random), direction | |
3243 | * (reads or writes), request sizes, greediness | |
3244 | * (from I/O-bound to sporadic), and so on; | |
3245 | ||
3246 | * In fact, in such a scenario, the drive tends to treat the requests | |
3247 | * of each process in about the same way as the requests of the | |
3248 | * others, and thus to provide each of these processes with about the | |
3249 | * same throughput. This is exactly the desired throughput | |
3250 | * distribution if (i-a) holds, or, if (i-b) holds instead, this is an | |
3251 | * even more convenient distribution for (the process associated with) | |
3252 | * bfqq. | |
3253 | * | |
3254 | * In contrast, in any asymmetric or unfavorable scenario, device | |
3255 | * idling (I/O-dispatch plugging) is certainly needed to guarantee | |
3256 | * that bfqq receives its assigned fraction of the device throughput | |
3257 | * (see [1] for details). | |
3258 | * | |
3259 | * The problem is that idling may significantly reduce throughput with | |
3260 | * certain combinations of types of I/O and devices. An important | |
3261 | * example is sync random I/O on flash storage with command | |
3262 | * queueing. So, unless bfqq falls in cases where idling also boosts | |
3263 | * throughput, it is important to check conditions (i-a), i(-b) and | |
3264 | * (ii) accurately, so as to avoid idling when not strictly needed for | |
3265 | * service guarantees. | |
3266 | * | |
3267 | * Unfortunately, it is extremely difficult to thoroughly check | |
3268 | * condition (ii). And, in case there are active groups, it becomes | |
3269 | * very difficult to check conditions (i-a) and (i-b) too. In fact, | |
3270 | * if there are active groups, then, for conditions (i-a) or (i-b) to | |
3271 | * become false 'indirectly', it is enough that an active group | |
3272 | * contains more active processes or sub-groups than some other active | |
3273 | * group. More precisely, for conditions (i-a) or (i-b) to become | |
3274 | * false because of such a group, it is not even necessary that the | |
3275 | * group is (still) active: it is sufficient that, even if the group | |
3276 | * has become inactive, some of its descendant processes still have | |
3277 | * some request already dispatched but still waiting for | |
3278 | * completion. In fact, requests have still to be guaranteed their | |
3279 | * share of the throughput even after being dispatched. In this | |
3280 | * respect, it is easy to show that, if a group frequently becomes | |
3281 | * inactive while still having in-flight requests, and if, when this | |
3282 | * happens, the group is not considered in the calculation of whether | |
3283 | * the scenario is asymmetric, then the group may fail to be | |
3284 | * guaranteed its fair share of the throughput (basically because | |
3285 | * idling may not be performed for the descendant processes of the | |
3286 | * group, but it had to be). We address this issue with the following | |
3287 | * bi-modal behavior, implemented in the function | |
3288 | * bfq_asymmetric_scenario(). | |
3289 | * | |
3290 | * If there are groups with requests waiting for completion | |
3291 | * (as commented above, some of these groups may even be | |
3292 | * already inactive), then the scenario is tagged as | |
3293 | * asymmetric, conservatively, without checking any of the | |
3294 | * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq. | |
3295 | * This behavior matches also the fact that groups are created | |
3296 | * exactly if controlling I/O is a primary concern (to | |
3297 | * preserve bandwidth and latency guarantees). | |
3298 | * | |
3299 | * On the opposite end, if there are no groups with requests waiting | |
3300 | * for completion, then only conditions (i-a) and (i-b) are actually | |
3301 | * controlled, i.e., provided that conditions (i-a) or (i-b) holds, | |
3302 | * idling is not performed, regardless of whether condition (ii) | |
3303 | * holds. In other words, only if conditions (i-a) and (i-b) do not | |
3304 | * hold, then idling is allowed, and the device tends to be prevented | |
3305 | * from queueing many requests, possibly of several processes. Since | |
3306 | * there are no groups with requests waiting for completion, then, to | |
3307 | * control conditions (i-a) and (i-b) it is enough to check just | |
3308 | * whether all the queues with requests waiting for completion also | |
3309 | * have the same weight. | |
3310 | * | |
3311 | * Not checking condition (ii) evidently exposes bfqq to the | |
3312 | * risk of getting less throughput than its fair share. | |
3313 | * However, for queues with the same weight, a further | |
3314 | * mechanism, preemption, mitigates or even eliminates this | |
3315 | * problem. And it does so without consequences on overall | |
3316 | * throughput. This mechanism and its benefits are explained | |
3317 | * in the next three paragraphs. | |
3318 | * | |
3319 | * Even if a queue, say Q, is expired when it remains idle, Q | |
3320 | * can still preempt the new in-service queue if the next | |
3321 | * request of Q arrives soon (see the comments on | |
3322 | * bfq_bfqq_update_budg_for_activation). If all queues and | |
3323 | * groups have the same weight, this form of preemption, | |
3324 | * combined with the hole-recovery heuristic described in the | |
3325 | * comments on function bfq_bfqq_update_budg_for_activation, | |
3326 | * are enough to preserve a correct bandwidth distribution in | |
3327 | * the mid term, even without idling. In fact, even if not | |
3328 | * idling allows the internal queues of the device to contain | |
3329 | * many requests, and thus to reorder requests, we can rather | |
3330 | * safely assume that the internal scheduler still preserves a | |
3331 | * minimum of mid-term fairness. | |
3332 | * | |
3333 | * More precisely, this preemption-based, idleless approach | |
3334 | * provides fairness in terms of IOPS, and not sectors per | |
3335 | * second. This can be seen with a simple example. Suppose | |
3336 | * that there are two queues with the same weight, but that | |
3337 | * the first queue receives requests of 8 sectors, while the | |
3338 | * second queue receives requests of 1024 sectors. In | |
3339 | * addition, suppose that each of the two queues contains at | |
3340 | * most one request at a time, which implies that each queue | |
3341 | * always remains idle after it is served. Finally, after | |
3342 | * remaining idle, each queue receives very quickly a new | |
3343 | * request. It follows that the two queues are served | |
3344 | * alternatively, preempting each other if needed. This | |
3345 | * implies that, although both queues have the same weight, | |
3346 | * the queue with large requests receives a service that is | |
3347 | * 1024/8 times as high as the service received by the other | |
3348 | * queue. | |
3349 | * | |
3350 | * The motivation for using preemption instead of idling (for | |
3351 | * queues with the same weight) is that, by not idling, | |
3352 | * service guarantees are preserved (completely or at least in | |
3353 | * part) without minimally sacrificing throughput. And, if | |
3354 | * there is no active group, then the primary expectation for | |
3355 | * this device is probably a high throughput. | |
3356 | * | |
b5e02b48 PV |
3357 | * We are now left only with explaining the two sub-conditions in the |
3358 | * additional compound condition that is checked below for deciding | |
3359 | * whether the scenario is asymmetric. To explain the first | |
3360 | * sub-condition, we need to add that the function | |
3726112e | 3361 | * bfq_asymmetric_scenario checks the weights of only |
b5e02b48 PV |
3362 | * non-weight-raised queues, for efficiency reasons (see comments on |
3363 | * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised | |
3364 | * is checked explicitly here. More precisely, the compound condition | |
3365 | * below takes into account also the fact that, even if bfqq is being | |
3366 | * weight-raised, the scenario is still symmetric if all queues with | |
3367 | * requests waiting for completion happen to be | |
3368 | * weight-raised. Actually, we should be even more precise here, and | |
3369 | * differentiate between interactive weight raising and soft real-time | |
3370 | * weight raising. | |
3371 | * | |
3372 | * The second sub-condition checked in the compound condition is | |
3373 | * whether there is a fair amount of already in-flight I/O not | |
3374 | * belonging to bfqq. If so, I/O dispatching is to be plugged, for the | |
3375 | * following reason. The drive may decide to serve in-flight | |
3376 | * non-bfqq's I/O requests before bfqq's ones, thereby delaying the | |
3377 | * arrival of new I/O requests for bfqq (recall that bfqq is sync). If | |
3378 | * I/O-dispatching is not plugged, then, while bfqq remains empty, a | |
3379 | * basically uncontrolled amount of I/O from other queues may be | |
3380 | * dispatched too, possibly causing the service of bfqq's I/O to be | |
3381 | * delayed even longer in the drive. This problem gets more and more | |
3382 | * serious as the speed and the queue depth of the drive grow, | |
3383 | * because, as these two quantities grow, the probability to find no | |
3384 | * queue busy but many requests in flight grows too. By contrast, | |
3385 | * plugging I/O dispatching minimizes the delay induced by already | |
3386 | * in-flight I/O, and enables bfqq to recover the bandwidth it may | |
3387 | * lose because of this delay. | |
3726112e PV |
3388 | * |
3389 | * As a side note, it is worth considering that the above | |
b5e02b48 PV |
3390 | * device-idling countermeasures may however fail in the following |
3391 | * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled | |
3392 | * in a time period during which all symmetry sub-conditions hold, and | |
3393 | * therefore the device is allowed to enqueue many requests, but at | |
3394 | * some later point in time some sub-condition stops to hold, then it | |
3395 | * may become impossible to make requests be served in the desired | |
3396 | * order until all the requests already queued in the device have been | |
3397 | * served. The last sub-condition commented above somewhat mitigates | |
3398 | * this problem for weight-raised queues. | |
3726112e PV |
3399 | */ |
3400 | static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd, | |
3401 | struct bfq_queue *bfqq) | |
3402 | { | |
3403 | return (bfqq->wr_coeff > 1 && | |
b5e02b48 PV |
3404 | (bfqd->wr_busy_queues < |
3405 | bfq_tot_busy_queues(bfqd) || | |
3406 | bfqd->rq_in_driver >= | |
3407 | bfqq->dispatched + 4)) || | |
3726112e PV |
3408 | bfq_asymmetric_scenario(bfqd, bfqq); |
3409 | } | |
3410 | ||
3411 | static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
3412 | enum bfqq_expiration reason) | |
aee69d78 | 3413 | { |
36eca894 AA |
3414 | /* |
3415 | * If this bfqq is shared between multiple processes, check | |
3416 | * to make sure that those processes are still issuing I/Os | |
3417 | * within the mean seek distance. If not, it may be time to | |
3418 | * break the queues apart again. | |
3419 | */ | |
3420 | if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) | |
3421 | bfq_mark_bfqq_split_coop(bfqq); | |
3422 | ||
3726112e PV |
3423 | /* |
3424 | * Consider queues with a higher finish virtual time than | |
3425 | * bfqq. If idling_needed_for_service_guarantees(bfqq) returns | |
3426 | * true, then bfqq's bandwidth would be violated if an | |
3427 | * uncontrolled amount of I/O from these queues were | |
3428 | * dispatched while bfqq is waiting for its new I/O to | |
3429 | * arrive. This is exactly what may happen if this is a forced | |
3430 | * expiration caused by a preemption attempt, and if bfqq is | |
3431 | * not re-scheduled. To prevent this from happening, re-queue | |
3432 | * bfqq if it needs I/O-dispatch plugging, even if it is | |
3433 | * empty. By doing so, bfqq is granted to be served before the | |
3434 | * above queues (provided that bfqq is of course eligible). | |
3435 | */ | |
3436 | if (RB_EMPTY_ROOT(&bfqq->sort_list) && | |
3437 | !(reason == BFQQE_PREEMPTED && | |
3438 | idling_needed_for_service_guarantees(bfqd, bfqq))) { | |
44e44a1b PV |
3439 | if (bfqq->dispatched == 0) |
3440 | /* | |
3441 | * Overloading budget_timeout field to store | |
3442 | * the time at which the queue remains with no | |
3443 | * backlog and no outstanding request; used by | |
3444 | * the weight-raising mechanism. | |
3445 | */ | |
3446 | bfqq->budget_timeout = jiffies; | |
3447 | ||
e21b7a0b | 3448 | bfq_del_bfqq_busy(bfqd, bfqq, true); |
36eca894 | 3449 | } else { |
80294c3b | 3450 | bfq_requeue_bfqq(bfqd, bfqq, true); |
36eca894 AA |
3451 | /* |
3452 | * Resort priority tree of potential close cooperators. | |
8cacc5ab | 3453 | * See comments on bfq_pos_tree_add_move() for the unlikely(). |
36eca894 | 3454 | */ |
3726112e PV |
3455 | if (unlikely(!bfqd->nonrot_with_queueing && |
3456 | !RB_EMPTY_ROOT(&bfqq->sort_list))) | |
8cacc5ab | 3457 | bfq_pos_tree_add_move(bfqd, bfqq); |
36eca894 | 3458 | } |
e21b7a0b AA |
3459 | |
3460 | /* | |
3461 | * All in-service entities must have been properly deactivated | |
3462 | * or requeued before executing the next function, which | |
eed47d19 PV |
3463 | * resets all in-service entities as no more in service. This |
3464 | * may cause bfqq to be freed. If this happens, the next | |
3465 | * function returns true. | |
e21b7a0b | 3466 | */ |
eed47d19 | 3467 | return __bfq_bfqd_reset_in_service(bfqd); |
aee69d78 PV |
3468 | } |
3469 | ||
3470 | /** | |
3471 | * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. | |
3472 | * @bfqd: device data. | |
3473 | * @bfqq: queue to update. | |
3474 | * @reason: reason for expiration. | |
3475 | * | |
3476 | * Handle the feedback on @bfqq budget at queue expiration. | |
3477 | * See the body for detailed comments. | |
3478 | */ | |
3479 | static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, | |
3480 | struct bfq_queue *bfqq, | |
3481 | enum bfqq_expiration reason) | |
3482 | { | |
3483 | struct request *next_rq; | |
3484 | int budget, min_budget; | |
3485 | ||
aee69d78 PV |
3486 | min_budget = bfq_min_budget(bfqd); |
3487 | ||
44e44a1b PV |
3488 | if (bfqq->wr_coeff == 1) |
3489 | budget = bfqq->max_budget; | |
3490 | else /* | |
3491 | * Use a constant, low budget for weight-raised queues, | |
3492 | * to help achieve a low latency. Keep it slightly higher | |
3493 | * than the minimum possible budget, to cause a little | |
3494 | * bit fewer expirations. | |
3495 | */ | |
3496 | budget = 2 * min_budget; | |
3497 | ||
aee69d78 PV |
3498 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d", |
3499 | bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); | |
3500 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d", | |
3501 | budget, bfq_min_budget(bfqd)); | |
3502 | bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", | |
3503 | bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); | |
3504 | ||
44e44a1b | 3505 | if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) { |
aee69d78 PV |
3506 | switch (reason) { |
3507 | /* | |
3508 | * Caveat: in all the following cases we trade latency | |
3509 | * for throughput. | |
3510 | */ | |
3511 | case BFQQE_TOO_IDLE: | |
54b60456 PV |
3512 | /* |
3513 | * This is the only case where we may reduce | |
3514 | * the budget: if there is no request of the | |
3515 | * process still waiting for completion, then | |
3516 | * we assume (tentatively) that the timer has | |
3517 | * expired because the batch of requests of | |
3518 | * the process could have been served with a | |
3519 | * smaller budget. Hence, betting that | |
3520 | * process will behave in the same way when it | |
3521 | * becomes backlogged again, we reduce its | |
3522 | * next budget. As long as we guess right, | |
3523 | * this budget cut reduces the latency | |
3524 | * experienced by the process. | |
3525 | * | |
3526 | * However, if there are still outstanding | |
3527 | * requests, then the process may have not yet | |
3528 | * issued its next request just because it is | |
3529 | * still waiting for the completion of some of | |
3530 | * the still outstanding ones. So in this | |
3531 | * subcase we do not reduce its budget, on the | |
3532 | * contrary we increase it to possibly boost | |
3533 | * the throughput, as discussed in the | |
3534 | * comments to the BUDGET_TIMEOUT case. | |
3535 | */ | |
3536 | if (bfqq->dispatched > 0) /* still outstanding reqs */ | |
3537 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
3538 | else { | |
3539 | if (budget > 5 * min_budget) | |
3540 | budget -= 4 * min_budget; | |
3541 | else | |
3542 | budget = min_budget; | |
3543 | } | |
aee69d78 PV |
3544 | break; |
3545 | case BFQQE_BUDGET_TIMEOUT: | |
54b60456 PV |
3546 | /* |
3547 | * We double the budget here because it gives | |
3548 | * the chance to boost the throughput if this | |
3549 | * is not a seeky process (and has bumped into | |
3550 | * this timeout because of, e.g., ZBR). | |
3551 | */ | |
3552 | budget = min(budget * 2, bfqd->bfq_max_budget); | |
aee69d78 PV |
3553 | break; |
3554 | case BFQQE_BUDGET_EXHAUSTED: | |
3555 | /* | |
3556 | * The process still has backlog, and did not | |
3557 | * let either the budget timeout or the disk | |
3558 | * idling timeout expire. Hence it is not | |
3559 | * seeky, has a short thinktime and may be | |
3560 | * happy with a higher budget too. So | |
3561 | * definitely increase the budget of this good | |
3562 | * candidate to boost the disk throughput. | |
3563 | */ | |
54b60456 | 3564 | budget = min(budget * 4, bfqd->bfq_max_budget); |
aee69d78 PV |
3565 | break; |
3566 | case BFQQE_NO_MORE_REQUESTS: | |
3567 | /* | |
3568 | * For queues that expire for this reason, it | |
3569 | * is particularly important to keep the | |
3570 | * budget close to the actual service they | |
3571 | * need. Doing so reduces the timestamp | |
3572 | * misalignment problem described in the | |
3573 | * comments in the body of | |
3574 | * __bfq_activate_entity. In fact, suppose | |
3575 | * that a queue systematically expires for | |
3576 | * BFQQE_NO_MORE_REQUESTS and presents a | |
3577 | * new request in time to enjoy timestamp | |
3578 | * back-shifting. The larger the budget of the | |
3579 | * queue is with respect to the service the | |
3580 | * queue actually requests in each service | |
3581 | * slot, the more times the queue can be | |
3582 | * reactivated with the same virtual finish | |
3583 | * time. It follows that, even if this finish | |
3584 | * time is pushed to the system virtual time | |
3585 | * to reduce the consequent timestamp | |
3586 | * misalignment, the queue unjustly enjoys for | |
3587 | * many re-activations a lower finish time | |
3588 | * than all newly activated queues. | |
3589 | * | |
3590 | * The service needed by bfqq is measured | |
3591 | * quite precisely by bfqq->entity.service. | |
3592 | * Since bfqq does not enjoy device idling, | |
3593 | * bfqq->entity.service is equal to the number | |
3594 | * of sectors that the process associated with | |
3595 | * bfqq requested to read/write before waiting | |
3596 | * for request completions, or blocking for | |
3597 | * other reasons. | |
3598 | */ | |
3599 | budget = max_t(int, bfqq->entity.service, min_budget); | |
3600 | break; | |
3601 | default: | |
3602 | return; | |
3603 | } | |
44e44a1b | 3604 | } else if (!bfq_bfqq_sync(bfqq)) { |
aee69d78 PV |
3605 | /* |
3606 | * Async queues get always the maximum possible | |
3607 | * budget, as for them we do not care about latency | |
3608 | * (in addition, their ability to dispatch is limited | |
3609 | * by the charging factor). | |
3610 | */ | |
3611 | budget = bfqd->bfq_max_budget; | |
3612 | } | |
3613 | ||
3614 | bfqq->max_budget = budget; | |
3615 | ||
3616 | if (bfqd->budgets_assigned >= bfq_stats_min_budgets && | |
3617 | !bfqd->bfq_user_max_budget) | |
3618 | bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget); | |
3619 | ||
3620 | /* | |
3621 | * If there is still backlog, then assign a new budget, making | |
3622 | * sure that it is large enough for the next request. Since | |
3623 | * the finish time of bfqq must be kept in sync with the | |
3624 | * budget, be sure to call __bfq_bfqq_expire() *after* this | |
3625 | * update. | |
3626 | * | |
3627 | * If there is no backlog, then no need to update the budget; | |
3628 | * it will be updated on the arrival of a new request. | |
3629 | */ | |
3630 | next_rq = bfqq->next_rq; | |
3631 | if (next_rq) | |
3632 | bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, | |
3633 | bfq_serv_to_charge(next_rq, bfqq)); | |
3634 | ||
3635 | bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d", | |
3636 | next_rq ? blk_rq_sectors(next_rq) : 0, | |
3637 | bfqq->entity.budget); | |
3638 | } | |
3639 | ||
aee69d78 | 3640 | /* |
ab0e43e9 PV |
3641 | * Return true if the process associated with bfqq is "slow". The slow |
3642 | * flag is used, in addition to the budget timeout, to reduce the | |
3643 | * amount of service provided to seeky processes, and thus reduce | |
3644 | * their chances to lower the throughput. More details in the comments | |
3645 | * on the function bfq_bfqq_expire(). | |
3646 | * | |
3647 | * An important observation is in order: as discussed in the comments | |
3648 | * on the function bfq_update_peak_rate(), with devices with internal | |
3649 | * queues, it is hard if ever possible to know when and for how long | |
3650 | * an I/O request is processed by the device (apart from the trivial | |
3651 | * I/O pattern where a new request is dispatched only after the | |
3652 | * previous one has been completed). This makes it hard to evaluate | |
3653 | * the real rate at which the I/O requests of each bfq_queue are | |
3654 | * served. In fact, for an I/O scheduler like BFQ, serving a | |
3655 | * bfq_queue means just dispatching its requests during its service | |
3656 | * slot (i.e., until the budget of the queue is exhausted, or the | |
3657 | * queue remains idle, or, finally, a timeout fires). But, during the | |
3658 | * service slot of a bfq_queue, around 100 ms at most, the device may | |
3659 | * be even still processing requests of bfq_queues served in previous | |
3660 | * service slots. On the opposite end, the requests of the in-service | |
3661 | * bfq_queue may be completed after the service slot of the queue | |
3662 | * finishes. | |
3663 | * | |
3664 | * Anyway, unless more sophisticated solutions are used | |
3665 | * (where possible), the sum of the sizes of the requests dispatched | |
3666 | * during the service slot of a bfq_queue is probably the only | |
3667 | * approximation available for the service received by the bfq_queue | |
3668 | * during its service slot. And this sum is the quantity used in this | |
3669 | * function to evaluate the I/O speed of a process. | |
aee69d78 | 3670 | */ |
ab0e43e9 PV |
3671 | static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
3672 | bool compensate, enum bfqq_expiration reason, | |
3673 | unsigned long *delta_ms) | |
aee69d78 | 3674 | { |
ab0e43e9 PV |
3675 | ktime_t delta_ktime; |
3676 | u32 delta_usecs; | |
3677 | bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */ | |
aee69d78 | 3678 | |
ab0e43e9 | 3679 | if (!bfq_bfqq_sync(bfqq)) |
aee69d78 PV |
3680 | return false; |
3681 | ||
3682 | if (compensate) | |
ab0e43e9 | 3683 | delta_ktime = bfqd->last_idling_start; |
aee69d78 | 3684 | else |
ab0e43e9 PV |
3685 | delta_ktime = ktime_get(); |
3686 | delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start); | |
3687 | delta_usecs = ktime_to_us(delta_ktime); | |
aee69d78 PV |
3688 | |
3689 | /* don't use too short time intervals */ | |
ab0e43e9 PV |
3690 | if (delta_usecs < 1000) { |
3691 | if (blk_queue_nonrot(bfqd->queue)) | |
3692 | /* | |
3693 | * give same worst-case guarantees as idling | |
3694 | * for seeky | |
3695 | */ | |
3696 | *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC; | |
3697 | else /* charge at least one seek */ | |
3698 | *delta_ms = bfq_slice_idle / NSEC_PER_MSEC; | |
3699 | ||
3700 | return slow; | |
3701 | } | |
aee69d78 | 3702 | |
ab0e43e9 | 3703 | *delta_ms = delta_usecs / USEC_PER_MSEC; |
aee69d78 PV |
3704 | |
3705 | /* | |
ab0e43e9 PV |
3706 | * Use only long (> 20ms) intervals to filter out excessive |
3707 | * spikes in service rate estimation. | |
aee69d78 | 3708 | */ |
ab0e43e9 PV |
3709 | if (delta_usecs > 20000) { |
3710 | /* | |
3711 | * Caveat for rotational devices: processes doing I/O | |
3712 | * in the slower disk zones tend to be slow(er) even | |
3713 | * if not seeky. In this respect, the estimated peak | |
3714 | * rate is likely to be an average over the disk | |
3715 | * surface. Accordingly, to not be too harsh with | |
3716 | * unlucky processes, a process is deemed slow only if | |
3717 | * its rate has been lower than half of the estimated | |
3718 | * peak rate. | |
3719 | */ | |
3720 | slow = bfqq->entity.service < bfqd->bfq_max_budget / 2; | |
aee69d78 PV |
3721 | } |
3722 | ||
ab0e43e9 | 3723 | bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow); |
aee69d78 | 3724 | |
ab0e43e9 | 3725 | return slow; |
aee69d78 PV |
3726 | } |
3727 | ||
77b7dcea PV |
3728 | /* |
3729 | * To be deemed as soft real-time, an application must meet two | |
3730 | * requirements. First, the application must not require an average | |
3731 | * bandwidth higher than the approximate bandwidth required to playback or | |
3732 | * record a compressed high-definition video. | |
3733 | * The next function is invoked on the completion of the last request of a | |
3734 | * batch, to compute the next-start time instant, soft_rt_next_start, such | |
3735 | * that, if the next request of the application does not arrive before | |
3736 | * soft_rt_next_start, then the above requirement on the bandwidth is met. | |
3737 | * | |
3738 | * The second requirement is that the request pattern of the application is | |
3739 | * isochronous, i.e., that, after issuing a request or a batch of requests, | |
3740 | * the application stops issuing new requests until all its pending requests | |
3741 | * have been completed. After that, the application may issue a new batch, | |
3742 | * and so on. | |
3743 | * For this reason the next function is invoked to compute | |
3744 | * soft_rt_next_start only for applications that meet this requirement, | |
3745 | * whereas soft_rt_next_start is set to infinity for applications that do | |
3746 | * not. | |
3747 | * | |
a34b0244 PV |
3748 | * Unfortunately, even a greedy (i.e., I/O-bound) application may |
3749 | * happen to meet, occasionally or systematically, both the above | |
3750 | * bandwidth and isochrony requirements. This may happen at least in | |
3751 | * the following circumstances. First, if the CPU load is high. The | |
3752 | * application may stop issuing requests while the CPUs are busy | |
3753 | * serving other processes, then restart, then stop again for a while, | |
3754 | * and so on. The other circumstances are related to the storage | |
3755 | * device: the storage device is highly loaded or reaches a low-enough | |
3756 | * throughput with the I/O of the application (e.g., because the I/O | |
3757 | * is random and/or the device is slow). In all these cases, the | |
3758 | * I/O of the application may be simply slowed down enough to meet | |
3759 | * the bandwidth and isochrony requirements. To reduce the probability | |
3760 | * that greedy applications are deemed as soft real-time in these | |
3761 | * corner cases, a further rule is used in the computation of | |
3762 | * soft_rt_next_start: the return value of this function is forced to | |
3763 | * be higher than the maximum between the following two quantities. | |
3764 | * | |
3765 | * (a) Current time plus: (1) the maximum time for which the arrival | |
3766 | * of a request is waited for when a sync queue becomes idle, | |
3767 | * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We | |
3768 | * postpone for a moment the reason for adding a few extra | |
3769 | * jiffies; we get back to it after next item (b). Lower-bounding | |
3770 | * the return value of this function with the current time plus | |
3771 | * bfqd->bfq_slice_idle tends to filter out greedy applications, | |
3772 | * because the latter issue their next request as soon as possible | |
3773 | * after the last one has been completed. In contrast, a soft | |
3774 | * real-time application spends some time processing data, after a | |
3775 | * batch of its requests has been completed. | |
3776 | * | |
3777 | * (b) Current value of bfqq->soft_rt_next_start. As pointed out | |
3778 | * above, greedy applications may happen to meet both the | |
3779 | * bandwidth and isochrony requirements under heavy CPU or | |
3780 | * storage-device load. In more detail, in these scenarios, these | |
3781 | * applications happen, only for limited time periods, to do I/O | |
3782 | * slowly enough to meet all the requirements described so far, | |
3783 | * including the filtering in above item (a). These slow-speed | |
3784 | * time intervals are usually interspersed between other time | |
3785 | * intervals during which these applications do I/O at a very high | |
3786 | * speed. Fortunately, exactly because of the high speed of the | |
3787 | * I/O in the high-speed intervals, the values returned by this | |
3788 | * function happen to be so high, near the end of any such | |
3789 | * high-speed interval, to be likely to fall *after* the end of | |
3790 | * the low-speed time interval that follows. These high values are | |
3791 | * stored in bfqq->soft_rt_next_start after each invocation of | |
3792 | * this function. As a consequence, if the last value of | |
3793 | * bfqq->soft_rt_next_start is constantly used to lower-bound the | |
3794 | * next value that this function may return, then, from the very | |
3795 | * beginning of a low-speed interval, bfqq->soft_rt_next_start is | |
3796 | * likely to be constantly kept so high that any I/O request | |
3797 | * issued during the low-speed interval is considered as arriving | |
3798 | * to soon for the application to be deemed as soft | |
3799 | * real-time. Then, in the high-speed interval that follows, the | |
3800 | * application will not be deemed as soft real-time, just because | |
3801 | * it will do I/O at a high speed. And so on. | |
3802 | * | |
3803 | * Getting back to the filtering in item (a), in the following two | |
3804 | * cases this filtering might be easily passed by a greedy | |
3805 | * application, if the reference quantity was just | |
3806 | * bfqd->bfq_slice_idle: | |
3807 | * 1) HZ is so low that the duration of a jiffy is comparable to or | |
3808 | * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow | |
3809 | * devices with HZ=100. The time granularity may be so coarse | |
3810 | * that the approximation, in jiffies, of bfqd->bfq_slice_idle | |
3811 | * is rather lower than the exact value. | |
77b7dcea PV |
3812 | * 2) jiffies, instead of increasing at a constant rate, may stop increasing |
3813 | * for a while, then suddenly 'jump' by several units to recover the lost | |
3814 | * increments. This seems to happen, e.g., inside virtual machines. | |
a34b0244 PV |
3815 | * To address this issue, in the filtering in (a) we do not use as a |
3816 | * reference time interval just bfqd->bfq_slice_idle, but | |
3817 | * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the | |
3818 | * minimum number of jiffies for which the filter seems to be quite | |
3819 | * precise also in embedded systems and KVM/QEMU virtual machines. | |
77b7dcea PV |
3820 | */ |
3821 | static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd, | |
3822 | struct bfq_queue *bfqq) | |
3823 | { | |
a34b0244 PV |
3824 | return max3(bfqq->soft_rt_next_start, |
3825 | bfqq->last_idle_bklogged + | |
3826 | HZ * bfqq->service_from_backlogged / | |
3827 | bfqd->bfq_wr_max_softrt_rate, | |
3828 | jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4); | |
77b7dcea PV |
3829 | } |
3830 | ||
aee69d78 PV |
3831 | /** |
3832 | * bfq_bfqq_expire - expire a queue. | |
3833 | * @bfqd: device owning the queue. | |
3834 | * @bfqq: the queue to expire. | |
3835 | * @compensate: if true, compensate for the time spent idling. | |
3836 | * @reason: the reason causing the expiration. | |
3837 | * | |
c074170e PV |
3838 | * If the process associated with bfqq does slow I/O (e.g., because it |
3839 | * issues random requests), we charge bfqq with the time it has been | |
3840 | * in service instead of the service it has received (see | |
3841 | * bfq_bfqq_charge_time for details on how this goal is achieved). As | |
3842 | * a consequence, bfqq will typically get higher timestamps upon | |
3843 | * reactivation, and hence it will be rescheduled as if it had | |
3844 | * received more service than what it has actually received. In the | |
3845 | * end, bfqq receives less service in proportion to how slowly its | |
3846 | * associated process consumes its budgets (and hence how seriously it | |
3847 | * tends to lower the throughput). In addition, this time-charging | |
3848 | * strategy guarantees time fairness among slow processes. In | |
3849 | * contrast, if the process associated with bfqq is not slow, we | |
3850 | * charge bfqq exactly with the service it has received. | |
aee69d78 | 3851 | * |
c074170e PV |
3852 | * Charging time to the first type of queues and the exact service to |
3853 | * the other has the effect of using the WF2Q+ policy to schedule the | |
3854 | * former on a timeslice basis, without violating service domain | |
3855 | * guarantees among the latter. | |
aee69d78 | 3856 | */ |
ea25da48 PV |
3857 | void bfq_bfqq_expire(struct bfq_data *bfqd, |
3858 | struct bfq_queue *bfqq, | |
3859 | bool compensate, | |
3860 | enum bfqq_expiration reason) | |
aee69d78 PV |
3861 | { |
3862 | bool slow; | |
ab0e43e9 PV |
3863 | unsigned long delta = 0; |
3864 | struct bfq_entity *entity = &bfqq->entity; | |
aee69d78 PV |
3865 | |
3866 | /* | |
ab0e43e9 | 3867 | * Check whether the process is slow (see bfq_bfqq_is_slow). |
aee69d78 | 3868 | */ |
ab0e43e9 | 3869 | slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta); |
aee69d78 PV |
3870 | |
3871 | /* | |
c074170e PV |
3872 | * As above explained, charge slow (typically seeky) and |
3873 | * timed-out queues with the time and not the service | |
3874 | * received, to favor sequential workloads. | |
3875 | * | |
3876 | * Processes doing I/O in the slower disk zones will tend to | |
3877 | * be slow(er) even if not seeky. Therefore, since the | |
3878 | * estimated peak rate is actually an average over the disk | |
3879 | * surface, these processes may timeout just for bad luck. To | |
3880 | * avoid punishing them, do not charge time to processes that | |
3881 | * succeeded in consuming at least 2/3 of their budget. This | |
3882 | * allows BFQ to preserve enough elasticity to still perform | |
3883 | * bandwidth, and not time, distribution with little unlucky | |
3884 | * or quasi-sequential processes. | |
aee69d78 | 3885 | */ |
44e44a1b PV |
3886 | if (bfqq->wr_coeff == 1 && |
3887 | (slow || | |
3888 | (reason == BFQQE_BUDGET_TIMEOUT && | |
3889 | bfq_bfqq_budget_left(bfqq) >= entity->budget / 3))) | |
c074170e | 3890 | bfq_bfqq_charge_time(bfqd, bfqq, delta); |
aee69d78 PV |
3891 | |
3892 | if (reason == BFQQE_TOO_IDLE && | |
ab0e43e9 | 3893 | entity->service <= 2 * entity->budget / 10) |
aee69d78 PV |
3894 | bfq_clear_bfqq_IO_bound(bfqq); |
3895 | ||
44e44a1b PV |
3896 | if (bfqd->low_latency && bfqq->wr_coeff == 1) |
3897 | bfqq->last_wr_start_finish = jiffies; | |
3898 | ||
77b7dcea PV |
3899 | if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 && |
3900 | RB_EMPTY_ROOT(&bfqq->sort_list)) { | |
3901 | /* | |
3902 | * If we get here, and there are no outstanding | |
3903 | * requests, then the request pattern is isochronous | |
3904 | * (see the comments on the function | |
3905 | * bfq_bfqq_softrt_next_start()). Thus we can compute | |
20cd3245 PV |
3906 | * soft_rt_next_start. And we do it, unless bfqq is in |
3907 | * interactive weight raising. We do not do it in the | |
3908 | * latter subcase, for the following reason. bfqq may | |
3909 | * be conveying the I/O needed to load a soft | |
3910 | * real-time application. Such an application will | |
3911 | * actually exhibit a soft real-time I/O pattern after | |
3912 | * it finally starts doing its job. But, if | |
3913 | * soft_rt_next_start is computed here for an | |
3914 | * interactive bfqq, and bfqq had received a lot of | |
3915 | * service before remaining with no outstanding | |
3916 | * request (likely to happen on a fast device), then | |
3917 | * soft_rt_next_start would be assigned such a high | |
3918 | * value that, for a very long time, bfqq would be | |
3919 | * prevented from being possibly considered as soft | |
3920 | * real time. | |
3921 | * | |
3922 | * If, instead, the queue still has outstanding | |
3923 | * requests, then we have to wait for the completion | |
3924 | * of all the outstanding requests to discover whether | |
3925 | * the request pattern is actually isochronous. | |
77b7dcea | 3926 | */ |
20cd3245 PV |
3927 | if (bfqq->dispatched == 0 && |
3928 | bfqq->wr_coeff != bfqd->bfq_wr_coeff) | |
77b7dcea PV |
3929 | bfqq->soft_rt_next_start = |
3930 | bfq_bfqq_softrt_next_start(bfqd, bfqq); | |
20cd3245 | 3931 | else if (bfqq->dispatched > 0) { |
77b7dcea PV |
3932 | /* |
3933 | * Schedule an update of soft_rt_next_start to when | |
3934 | * the task may be discovered to be isochronous. | |
3935 | */ | |
3936 | bfq_mark_bfqq_softrt_update(bfqq); | |
3937 | } | |
3938 | } | |
3939 | ||
aee69d78 | 3940 | bfq_log_bfqq(bfqd, bfqq, |
d5be3fef PV |
3941 | "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason, |
3942 | slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq)); | |
aee69d78 | 3943 | |
2341d662 PV |
3944 | /* |
3945 | * bfqq expired, so no total service time needs to be computed | |
3946 | * any longer: reset state machine for measuring total service | |
3947 | * times. | |
3948 | */ | |
3949 | bfqd->rqs_injected = bfqd->wait_dispatch = false; | |
3950 | bfqd->waited_rq = NULL; | |
3951 | ||
aee69d78 PV |
3952 | /* |
3953 | * Increase, decrease or leave budget unchanged according to | |
3954 | * reason. | |
3955 | */ | |
3956 | __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); | |
3726112e | 3957 | if (__bfq_bfqq_expire(bfqd, bfqq, reason)) |
eed47d19 | 3958 | /* bfqq is gone, no more actions on it */ |
9fae8dd5 PV |
3959 | return; |
3960 | ||
aee69d78 | 3961 | /* mark bfqq as waiting a request only if a bic still points to it */ |
9fae8dd5 | 3962 | if (!bfq_bfqq_busy(bfqq) && |
aee69d78 | 3963 | reason != BFQQE_BUDGET_TIMEOUT && |
9fae8dd5 | 3964 | reason != BFQQE_BUDGET_EXHAUSTED) { |
aee69d78 | 3965 | bfq_mark_bfqq_non_blocking_wait_rq(bfqq); |
9fae8dd5 PV |
3966 | /* |
3967 | * Not setting service to 0, because, if the next rq | |
3968 | * arrives in time, the queue will go on receiving | |
3969 | * service with this same budget (as if it never expired) | |
3970 | */ | |
3971 | } else | |
3972 | entity->service = 0; | |
8a511ba5 PV |
3973 | |
3974 | /* | |
3975 | * Reset the received-service counter for every parent entity. | |
3976 | * Differently from what happens with bfqq->entity.service, | |
3977 | * the resetting of this counter never needs to be postponed | |
3978 | * for parent entities. In fact, in case bfqq may have a | |
3979 | * chance to go on being served using the last, partially | |
3980 | * consumed budget, bfqq->entity.service needs to be kept, | |
3981 | * because if bfqq then actually goes on being served using | |
3982 | * the same budget, the last value of bfqq->entity.service is | |
3983 | * needed to properly decrement bfqq->entity.budget by the | |
3984 | * portion already consumed. In contrast, it is not necessary | |
3985 | * to keep entity->service for parent entities too, because | |
3986 | * the bubble up of the new value of bfqq->entity.budget will | |
3987 | * make sure that the budgets of parent entities are correct, | |
3988 | * even in case bfqq and thus parent entities go on receiving | |
3989 | * service with the same budget. | |
3990 | */ | |
3991 | entity = entity->parent; | |
3992 | for_each_entity(entity) | |
3993 | entity->service = 0; | |
aee69d78 PV |
3994 | } |
3995 | ||
3996 | /* | |
3997 | * Budget timeout is not implemented through a dedicated timer, but | |
3998 | * just checked on request arrivals and completions, as well as on | |
3999 | * idle timer expirations. | |
4000 | */ | |
4001 | static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) | |
4002 | { | |
44e44a1b | 4003 | return time_is_before_eq_jiffies(bfqq->budget_timeout); |
aee69d78 PV |
4004 | } |
4005 | ||
4006 | /* | |
4007 | * If we expire a queue that is actively waiting (i.e., with the | |
4008 | * device idled) for the arrival of a new request, then we may incur | |
4009 | * the timestamp misalignment problem described in the body of the | |
4010 | * function __bfq_activate_entity. Hence we return true only if this | |
4011 | * condition does not hold, or if the queue is slow enough to deserve | |
4012 | * only to be kicked off for preserving a high throughput. | |
4013 | */ | |
4014 | static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) | |
4015 | { | |
4016 | bfq_log_bfqq(bfqq->bfqd, bfqq, | |
4017 | "may_budget_timeout: wait_request %d left %d timeout %d", | |
4018 | bfq_bfqq_wait_request(bfqq), | |
4019 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, | |
4020 | bfq_bfqq_budget_timeout(bfqq)); | |
4021 | ||
4022 | return (!bfq_bfqq_wait_request(bfqq) || | |
4023 | bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) | |
4024 | && | |
4025 | bfq_bfqq_budget_timeout(bfqq); | |
4026 | } | |
4027 | ||
05c2f5c3 PV |
4028 | static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd, |
4029 | struct bfq_queue *bfqq) | |
aee69d78 | 4030 | { |
edaf9428 PV |
4031 | bool rot_without_queueing = |
4032 | !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag, | |
4033 | bfqq_sequential_and_IO_bound, | |
05c2f5c3 | 4034 | idling_boosts_thr; |
d5be3fef | 4035 | |
edaf9428 PV |
4036 | bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) && |
4037 | bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq); | |
4038 | ||
aee69d78 | 4039 | /* |
44e44a1b PV |
4040 | * The next variable takes into account the cases where idling |
4041 | * boosts the throughput. | |
4042 | * | |
e01eff01 PV |
4043 | * The value of the variable is computed considering, first, that |
4044 | * idling is virtually always beneficial for the throughput if: | |
edaf9428 PV |
4045 | * (a) the device is not NCQ-capable and rotational, or |
4046 | * (b) regardless of the presence of NCQ, the device is rotational and | |
4047 | * the request pattern for bfqq is I/O-bound and sequential, or | |
4048 | * (c) regardless of whether it is rotational, the device is | |
4049 | * not NCQ-capable and the request pattern for bfqq is | |
4050 | * I/O-bound and sequential. | |
bf2b79e7 PV |
4051 | * |
4052 | * Secondly, and in contrast to the above item (b), idling an | |
4053 | * NCQ-capable flash-based device would not boost the | |
e01eff01 | 4054 | * throughput even with sequential I/O; rather it would lower |
bf2b79e7 PV |
4055 | * the throughput in proportion to how fast the device |
4056 | * is. Accordingly, the next variable is true if any of the | |
edaf9428 PV |
4057 | * above conditions (a), (b) or (c) is true, and, in |
4058 | * particular, happens to be false if bfqd is an NCQ-capable | |
4059 | * flash-based device. | |
aee69d78 | 4060 | */ |
edaf9428 PV |
4061 | idling_boosts_thr = rot_without_queueing || |
4062 | ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) && | |
4063 | bfqq_sequential_and_IO_bound); | |
aee69d78 | 4064 | |
cfd69712 | 4065 | /* |
05c2f5c3 | 4066 | * The return value of this function is equal to that of |
cfd69712 PV |
4067 | * idling_boosts_thr, unless a special case holds. In this |
4068 | * special case, described below, idling may cause problems to | |
4069 | * weight-raised queues. | |
4070 | * | |
4071 | * When the request pool is saturated (e.g., in the presence | |
4072 | * of write hogs), if the processes associated with | |
4073 | * non-weight-raised queues ask for requests at a lower rate, | |
4074 | * then processes associated with weight-raised queues have a | |
4075 | * higher probability to get a request from the pool | |
4076 | * immediately (or at least soon) when they need one. Thus | |
4077 | * they have a higher probability to actually get a fraction | |
4078 | * of the device throughput proportional to their high | |
4079 | * weight. This is especially true with NCQ-capable drives, | |
4080 | * which enqueue several requests in advance, and further | |
4081 | * reorder internally-queued requests. | |
4082 | * | |
05c2f5c3 PV |
4083 | * For this reason, we force to false the return value if |
4084 | * there are weight-raised busy queues. In this case, and if | |
4085 | * bfqq is not weight-raised, this guarantees that the device | |
4086 | * is not idled for bfqq (if, instead, bfqq is weight-raised, | |
4087 | * then idling will be guaranteed by another variable, see | |
4088 | * below). Combined with the timestamping rules of BFQ (see | |
4089 | * [1] for details), this behavior causes bfqq, and hence any | |
4090 | * sync non-weight-raised queue, to get a lower number of | |
4091 | * requests served, and thus to ask for a lower number of | |
4092 | * requests from the request pool, before the busy | |
4093 | * weight-raised queues get served again. This often mitigates | |
4094 | * starvation problems in the presence of heavy write | |
4095 | * workloads and NCQ, thereby guaranteeing a higher | |
4096 | * application and system responsiveness in these hostile | |
4097 | * scenarios. | |
4098 | */ | |
4099 | return idling_boosts_thr && | |
cfd69712 | 4100 | bfqd->wr_busy_queues == 0; |
05c2f5c3 | 4101 | } |
cfd69712 | 4102 | |
05c2f5c3 PV |
4103 | /* |
4104 | * For a queue that becomes empty, device idling is allowed only if | |
4105 | * this function returns true for that queue. As a consequence, since | |
4106 | * device idling plays a critical role for both throughput boosting | |
4107 | * and service guarantees, the return value of this function plays a | |
4108 | * critical role as well. | |
4109 | * | |
4110 | * In a nutshell, this function returns true only if idling is | |
4111 | * beneficial for throughput or, even if detrimental for throughput, | |
4112 | * idling is however necessary to preserve service guarantees (low | |
4113 | * latency, desired throughput distribution, ...). In particular, on | |
4114 | * NCQ-capable devices, this function tries to return false, so as to | |
4115 | * help keep the drives' internal queues full, whenever this helps the | |
4116 | * device boost the throughput without causing any service-guarantee | |
4117 | * issue. | |
4118 | * | |
4119 | * Most of the issues taken into account to get the return value of | |
4120 | * this function are not trivial. We discuss these issues in the two | |
4121 | * functions providing the main pieces of information needed by this | |
4122 | * function. | |
4123 | */ | |
4124 | static bool bfq_better_to_idle(struct bfq_queue *bfqq) | |
4125 | { | |
4126 | struct bfq_data *bfqd = bfqq->bfqd; | |
4127 | bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar; | |
4128 | ||
4129 | if (unlikely(bfqd->strict_guarantees)) | |
4130 | return true; | |
4131 | ||
4132 | /* | |
4133 | * Idling is performed only if slice_idle > 0. In addition, we | |
4134 | * do not idle if | |
4135 | * (a) bfqq is async | |
4136 | * (b) bfqq is in the idle io prio class: in this case we do | |
4137 | * not idle because we want to minimize the bandwidth that | |
4138 | * queues in this class can steal to higher-priority queues | |
4139 | */ | |
4140 | if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) || | |
4141 | bfq_class_idle(bfqq)) | |
4142 | return false; | |
4143 | ||
4144 | idling_boosts_thr_with_no_issue = | |
4145 | idling_boosts_thr_without_issues(bfqd, bfqq); | |
4146 | ||
4147 | idling_needed_for_service_guar = | |
4148 | idling_needed_for_service_guarantees(bfqd, bfqq); | |
e1b2324d | 4149 | |
44e44a1b | 4150 | /* |
05c2f5c3 | 4151 | * We have now the two components we need to compute the |
d5be3fef PV |
4152 | * return value of the function, which is true only if idling |
4153 | * either boosts the throughput (without issues), or is | |
4154 | * necessary to preserve service guarantees. | |
aee69d78 | 4155 | */ |
05c2f5c3 PV |
4156 | return idling_boosts_thr_with_no_issue || |
4157 | idling_needed_for_service_guar; | |
aee69d78 PV |
4158 | } |
4159 | ||
4160 | /* | |
277a4a9b | 4161 | * If the in-service queue is empty but the function bfq_better_to_idle |
aee69d78 PV |
4162 | * returns true, then: |
4163 | * 1) the queue must remain in service and cannot be expired, and | |
4164 | * 2) the device must be idled to wait for the possible arrival of a new | |
4165 | * request for the queue. | |
277a4a9b | 4166 | * See the comments on the function bfq_better_to_idle for the reasons |
aee69d78 | 4167 | * why performing device idling is the best choice to boost the throughput |
277a4a9b | 4168 | * and preserve service guarantees when bfq_better_to_idle itself |
aee69d78 PV |
4169 | * returns true. |
4170 | */ | |
4171 | static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) | |
4172 | { | |
277a4a9b | 4173 | return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq); |
aee69d78 PV |
4174 | } |
4175 | ||
2341d662 PV |
4176 | /* |
4177 | * This function chooses the queue from which to pick the next extra | |
4178 | * I/O request to inject, if it finds a compatible queue. See the | |
4179 | * comments on bfq_update_inject_limit() for details on the injection | |
4180 | * mechanism, and for the definitions of the quantities mentioned | |
4181 | * below. | |
4182 | */ | |
4183 | static struct bfq_queue * | |
4184 | bfq_choose_bfqq_for_injection(struct bfq_data *bfqd) | |
d0edc247 | 4185 | { |
2341d662 PV |
4186 | struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue; |
4187 | unsigned int limit = in_serv_bfqq->inject_limit; | |
4188 | /* | |
4189 | * If | |
4190 | * - bfqq is not weight-raised and therefore does not carry | |
4191 | * time-critical I/O, | |
4192 | * or | |
4193 | * - regardless of whether bfqq is weight-raised, bfqq has | |
4194 | * however a long think time, during which it can absorb the | |
4195 | * effect of an appropriate number of extra I/O requests | |
4196 | * from other queues (see bfq_update_inject_limit for | |
4197 | * details on the computation of this number); | |
4198 | * then injection can be performed without restrictions. | |
4199 | */ | |
4200 | bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 || | |
4201 | !bfq_bfqq_has_short_ttime(in_serv_bfqq); | |
d0edc247 PV |
4202 | |
4203 | /* | |
2341d662 PV |
4204 | * If |
4205 | * - the baseline total service time could not be sampled yet, | |
4206 | * so the inject limit happens to be still 0, and | |
4207 | * - a lot of time has elapsed since the plugging of I/O | |
4208 | * dispatching started, so drive speed is being wasted | |
4209 | * significantly; | |
4210 | * then temporarily raise inject limit to one request. | |
4211 | */ | |
4212 | if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 && | |
4213 | bfq_bfqq_wait_request(in_serv_bfqq) && | |
4214 | time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies + | |
4215 | bfqd->bfq_slice_idle) | |
4216 | ) | |
4217 | limit = 1; | |
4218 | ||
4219 | if (bfqd->rq_in_driver >= limit) | |
4220 | return NULL; | |
4221 | ||
4222 | /* | |
4223 | * Linear search of the source queue for injection; but, with | |
4224 | * a high probability, very few steps are needed to find a | |
4225 | * candidate queue, i.e., a queue with enough budget left for | |
4226 | * its next request. In fact: | |
d0edc247 PV |
4227 | * - BFQ dynamically updates the budget of every queue so as |
4228 | * to accommodate the expected backlog of the queue; | |
4229 | * - if a queue gets all its requests dispatched as injected | |
4230 | * service, then the queue is removed from the active list | |
2341d662 PV |
4231 | * (and re-added only if it gets new requests, but then it |
4232 | * is assigned again enough budget for its new backlog). | |
d0edc247 PV |
4233 | */ |
4234 | list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) | |
4235 | if (!RB_EMPTY_ROOT(&bfqq->sort_list) && | |
2341d662 | 4236 | (in_serv_always_inject || bfqq->wr_coeff > 1) && |
d0edc247 | 4237 | bfq_serv_to_charge(bfqq->next_rq, bfqq) <= |
2341d662 PV |
4238 | bfq_bfqq_budget_left(bfqq)) { |
4239 | /* | |
4240 | * Allow for only one large in-flight request | |
4241 | * on non-rotational devices, for the | |
4242 | * following reason. On non-rotationl drives, | |
4243 | * large requests take much longer than | |
4244 | * smaller requests to be served. In addition, | |
4245 | * the drive prefers to serve large requests | |
4246 | * w.r.t. to small ones, if it can choose. So, | |
4247 | * having more than one large requests queued | |
4248 | * in the drive may easily make the next first | |
4249 | * request of the in-service queue wait for so | |
4250 | * long to break bfqq's service guarantees. On | |
4251 | * the bright side, large requests let the | |
4252 | * drive reach a very high throughput, even if | |
4253 | * there is only one in-flight large request | |
4254 | * at a time. | |
4255 | */ | |
4256 | if (blk_queue_nonrot(bfqd->queue) && | |
4257 | blk_rq_sectors(bfqq->next_rq) >= | |
4258 | BFQQ_SECT_THR_NONROT) | |
4259 | limit = min_t(unsigned int, 1, limit); | |
4260 | else | |
4261 | limit = in_serv_bfqq->inject_limit; | |
4262 | ||
4263 | if (bfqd->rq_in_driver < limit) { | |
4264 | bfqd->rqs_injected = true; | |
4265 | return bfqq; | |
4266 | } | |
4267 | } | |
d0edc247 PV |
4268 | |
4269 | return NULL; | |
4270 | } | |
4271 | ||
aee69d78 PV |
4272 | /* |
4273 | * Select a queue for service. If we have a current queue in service, | |
4274 | * check whether to continue servicing it, or retrieve and set a new one. | |
4275 | */ | |
4276 | static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) | |
4277 | { | |
4278 | struct bfq_queue *bfqq; | |
4279 | struct request *next_rq; | |
4280 | enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT; | |
4281 | ||
4282 | bfqq = bfqd->in_service_queue; | |
4283 | if (!bfqq) | |
4284 | goto new_queue; | |
4285 | ||
4286 | bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); | |
4287 | ||
4420b095 PV |
4288 | /* |
4289 | * Do not expire bfqq for budget timeout if bfqq may be about | |
4290 | * to enjoy device idling. The reason why, in this case, we | |
4291 | * prevent bfqq from expiring is the same as in the comments | |
4292 | * on the case where bfq_bfqq_must_idle() returns true, in | |
4293 | * bfq_completed_request(). | |
4294 | */ | |
aee69d78 | 4295 | if (bfq_may_expire_for_budg_timeout(bfqq) && |
aee69d78 PV |
4296 | !bfq_bfqq_must_idle(bfqq)) |
4297 | goto expire; | |
4298 | ||
4299 | check_queue: | |
4300 | /* | |
4301 | * This loop is rarely executed more than once. Even when it | |
4302 | * happens, it is much more convenient to re-execute this loop | |
4303 | * than to return NULL and trigger a new dispatch to get a | |
4304 | * request served. | |
4305 | */ | |
4306 | next_rq = bfqq->next_rq; | |
4307 | /* | |
4308 | * If bfqq has requests queued and it has enough budget left to | |
4309 | * serve them, keep the queue, otherwise expire it. | |
4310 | */ | |
4311 | if (next_rq) { | |
4312 | if (bfq_serv_to_charge(next_rq, bfqq) > | |
4313 | bfq_bfqq_budget_left(bfqq)) { | |
4314 | /* | |
4315 | * Expire the queue for budget exhaustion, | |
4316 | * which makes sure that the next budget is | |
4317 | * enough to serve the next request, even if | |
4318 | * it comes from the fifo expired path. | |
4319 | */ | |
4320 | reason = BFQQE_BUDGET_EXHAUSTED; | |
4321 | goto expire; | |
4322 | } else { | |
4323 | /* | |
4324 | * The idle timer may be pending because we may | |
4325 | * not disable disk idling even when a new request | |
4326 | * arrives. | |
4327 | */ | |
4328 | if (bfq_bfqq_wait_request(bfqq)) { | |
4329 | /* | |
4330 | * If we get here: 1) at least a new request | |
4331 | * has arrived but we have not disabled the | |
4332 | * timer because the request was too small, | |
4333 | * 2) then the block layer has unplugged | |
4334 | * the device, causing the dispatch to be | |
4335 | * invoked. | |
4336 | * | |
4337 | * Since the device is unplugged, now the | |
4338 | * requests are probably large enough to | |
4339 | * provide a reasonable throughput. | |
4340 | * So we disable idling. | |
4341 | */ | |
4342 | bfq_clear_bfqq_wait_request(bfqq); | |
4343 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
4344 | } | |
4345 | goto keep_queue; | |
4346 | } | |
4347 | } | |
4348 | ||
4349 | /* | |
4350 | * No requests pending. However, if the in-service queue is idling | |
4351 | * for a new request, or has requests waiting for a completion and | |
4352 | * may idle after their completion, then keep it anyway. | |
d0edc247 | 4353 | * |
2341d662 PV |
4354 | * Yet, inject service from other queues if it boosts |
4355 | * throughput and is possible. | |
aee69d78 PV |
4356 | */ |
4357 | if (bfq_bfqq_wait_request(bfqq) || | |
277a4a9b | 4358 | (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) { |
2341d662 PV |
4359 | struct bfq_queue *async_bfqq = |
4360 | bfqq->bic && bfqq->bic->bfqq[0] && | |
3726112e PV |
4361 | bfq_bfqq_busy(bfqq->bic->bfqq[0]) && |
4362 | bfqq->bic->bfqq[0]->next_rq ? | |
2341d662 PV |
4363 | bfqq->bic->bfqq[0] : NULL; |
4364 | ||
4365 | /* | |
13a857a4 PV |
4366 | * The next three mutually-exclusive ifs decide |
4367 | * whether to try injection, and choose the queue to | |
4368 | * pick an I/O request from. | |
4369 | * | |
4370 | * The first if checks whether the process associated | |
4371 | * with bfqq has also async I/O pending. If so, it | |
4372 | * injects such I/O unconditionally. Injecting async | |
4373 | * I/O from the same process can cause no harm to the | |
4374 | * process. On the contrary, it can only increase | |
4375 | * bandwidth and reduce latency for the process. | |
4376 | * | |
4377 | * The second if checks whether there happens to be a | |
4378 | * non-empty waker queue for bfqq, i.e., a queue whose | |
4379 | * I/O needs to be completed for bfqq to receive new | |
4380 | * I/O. This happens, e.g., if bfqq is associated with | |
4381 | * a process that does some sync. A sync generates | |
4382 | * extra blocking I/O, which must be completed before | |
4383 | * the process associated with bfqq can go on with its | |
4384 | * I/O. If the I/O of the waker queue is not served, | |
4385 | * then bfqq remains empty, and no I/O is dispatched, | |
4386 | * until the idle timeout fires for bfqq. This is | |
4387 | * likely to result in lower bandwidth and higher | |
4388 | * latencies for bfqq, and in a severe loss of total | |
4389 | * throughput. The best action to take is therefore to | |
4390 | * serve the waker queue as soon as possible. So do it | |
4391 | * (without relying on the third alternative below for | |
4392 | * eventually serving waker_bfqq's I/O; see the last | |
4393 | * paragraph for further details). This systematic | |
4394 | * injection of I/O from the waker queue does not | |
4395 | * cause any delay to bfqq's I/O. On the contrary, | |
4396 | * next bfqq's I/O is brought forward dramatically, | |
4397 | * for it is not blocked for milliseconds. | |
4398 | * | |
4399 | * The third if checks whether bfqq is a queue for | |
4400 | * which it is better to avoid injection. It is so if | |
4401 | * bfqq delivers more throughput when served without | |
4402 | * any further I/O from other queues in the middle, or | |
4403 | * if the service times of bfqq's I/O requests both | |
4404 | * count more than overall throughput, and may be | |
4405 | * easily increased by injection (this happens if bfqq | |
4406 | * has a short think time). If none of these | |
4407 | * conditions holds, then a candidate queue for | |
4408 | * injection is looked for through | |
4409 | * bfq_choose_bfqq_for_injection(). Note that the | |
4410 | * latter may return NULL (for example if the inject | |
4411 | * limit for bfqq is currently 0). | |
4412 | * | |
4413 | * NOTE: motivation for the second alternative | |
4414 | * | |
4415 | * Thanks to the way the inject limit is updated in | |
4416 | * bfq_update_has_short_ttime(), it is rather likely | |
4417 | * that, if I/O is being plugged for bfqq and the | |
4418 | * waker queue has pending I/O requests that are | |
4419 | * blocking bfqq's I/O, then the third alternative | |
4420 | * above lets the waker queue get served before the | |
4421 | * I/O-plugging timeout fires. So one may deem the | |
4422 | * second alternative superfluous. It is not, because | |
4423 | * the third alternative may be way less effective in | |
4424 | * case of a synchronization. For two main | |
4425 | * reasons. First, throughput may be low because the | |
4426 | * inject limit may be too low to guarantee the same | |
4427 | * amount of injected I/O, from the waker queue or | |
4428 | * other queues, that the second alternative | |
4429 | * guarantees (the second alternative unconditionally | |
4430 | * injects a pending I/O request of the waker queue | |
4431 | * for each bfq_dispatch_request()). Second, with the | |
4432 | * third alternative, the duration of the plugging, | |
4433 | * i.e., the time before bfqq finally receives new I/O, | |
4434 | * may not be minimized, because the waker queue may | |
4435 | * happen to be served only after other queues. | |
2341d662 PV |
4436 | */ |
4437 | if (async_bfqq && | |
4438 | icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic && | |
4439 | bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <= | |
4440 | bfq_bfqq_budget_left(async_bfqq)) | |
4441 | bfqq = bfqq->bic->bfqq[0]; | |
13a857a4 PV |
4442 | else if (bfq_bfqq_has_waker(bfqq) && |
4443 | bfq_bfqq_busy(bfqq->waker_bfqq) && | |
3726112e | 4444 | bfqq->next_rq && |
13a857a4 PV |
4445 | bfq_serv_to_charge(bfqq->waker_bfqq->next_rq, |
4446 | bfqq->waker_bfqq) <= | |
4447 | bfq_bfqq_budget_left(bfqq->waker_bfqq) | |
4448 | ) | |
4449 | bfqq = bfqq->waker_bfqq; | |
2341d662 PV |
4450 | else if (!idling_boosts_thr_without_issues(bfqd, bfqq) && |
4451 | (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 || | |
4452 | !bfq_bfqq_has_short_ttime(bfqq))) | |
d0edc247 PV |
4453 | bfqq = bfq_choose_bfqq_for_injection(bfqd); |
4454 | else | |
4455 | bfqq = NULL; | |
4456 | ||
aee69d78 PV |
4457 | goto keep_queue; |
4458 | } | |
4459 | ||
4460 | reason = BFQQE_NO_MORE_REQUESTS; | |
4461 | expire: | |
4462 | bfq_bfqq_expire(bfqd, bfqq, false, reason); | |
4463 | new_queue: | |
4464 | bfqq = bfq_set_in_service_queue(bfqd); | |
4465 | if (bfqq) { | |
4466 | bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue"); | |
4467 | goto check_queue; | |
4468 | } | |
4469 | keep_queue: | |
4470 | if (bfqq) | |
4471 | bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue"); | |
4472 | else | |
4473 | bfq_log(bfqd, "select_queue: no queue returned"); | |
4474 | ||
4475 | return bfqq; | |
4476 | } | |
4477 | ||
44e44a1b PV |
4478 | static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
4479 | { | |
4480 | struct bfq_entity *entity = &bfqq->entity; | |
4481 | ||
4482 | if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */ | |
4483 | bfq_log_bfqq(bfqd, bfqq, | |
4484 | "raising period dur %u/%u msec, old coeff %u, w %d(%d)", | |
4485 | jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish), | |
4486 | jiffies_to_msecs(bfqq->wr_cur_max_time), | |
4487 | bfqq->wr_coeff, | |
4488 | bfqq->entity.weight, bfqq->entity.orig_weight); | |
4489 | ||
4490 | if (entity->prio_changed) | |
4491 | bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change"); | |
4492 | ||
4493 | /* | |
e1b2324d AA |
4494 | * If the queue was activated in a burst, or too much |
4495 | * time has elapsed from the beginning of this | |
4496 | * weight-raising period, then end weight raising. | |
44e44a1b | 4497 | */ |
e1b2324d AA |
4498 | if (bfq_bfqq_in_large_burst(bfqq)) |
4499 | bfq_bfqq_end_wr(bfqq); | |
4500 | else if (time_is_before_jiffies(bfqq->last_wr_start_finish + | |
4501 | bfqq->wr_cur_max_time)) { | |
77b7dcea PV |
4502 | if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time || |
4503 | time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt + | |
e1b2324d | 4504 | bfq_wr_duration(bfqd))) |
77b7dcea PV |
4505 | bfq_bfqq_end_wr(bfqq); |
4506 | else { | |
3e2bdd6d | 4507 | switch_back_to_interactive_wr(bfqq, bfqd); |
77b7dcea PV |
4508 | bfqq->entity.prio_changed = 1; |
4509 | } | |
44e44a1b | 4510 | } |
8a8747dc PV |
4511 | if (bfqq->wr_coeff > 1 && |
4512 | bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time && | |
4513 | bfqq->service_from_wr > max_service_from_wr) { | |
4514 | /* see comments on max_service_from_wr */ | |
4515 | bfq_bfqq_end_wr(bfqq); | |
4516 | } | |
44e44a1b | 4517 | } |
431b17f9 PV |
4518 | /* |
4519 | * To improve latency (for this or other queues), immediately | |
4520 | * update weight both if it must be raised and if it must be | |
4521 | * lowered. Since, entity may be on some active tree here, and | |
4522 | * might have a pending change of its ioprio class, invoke | |
4523 | * next function with the last parameter unset (see the | |
4524 | * comments on the function). | |
4525 | */ | |
44e44a1b | 4526 | if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1)) |
431b17f9 PV |
4527 | __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity), |
4528 | entity, false); | |
44e44a1b PV |
4529 | } |
4530 | ||
aee69d78 PV |
4531 | /* |
4532 | * Dispatch next request from bfqq. | |
4533 | */ | |
4534 | static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd, | |
4535 | struct bfq_queue *bfqq) | |
4536 | { | |
4537 | struct request *rq = bfqq->next_rq; | |
4538 | unsigned long service_to_charge; | |
4539 | ||
4540 | service_to_charge = bfq_serv_to_charge(rq, bfqq); | |
4541 | ||
4542 | bfq_bfqq_served(bfqq, service_to_charge); | |
4543 | ||
2341d662 PV |
4544 | if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) { |
4545 | bfqd->wait_dispatch = false; | |
4546 | bfqd->waited_rq = rq; | |
4547 | } | |
aee69d78 | 4548 | |
2341d662 | 4549 | bfq_dispatch_remove(bfqd->queue, rq); |
d0edc247 | 4550 | |
2341d662 | 4551 | if (bfqq != bfqd->in_service_queue) |
d0edc247 | 4552 | goto return_rq; |
d0edc247 | 4553 | |
44e44a1b PV |
4554 | /* |
4555 | * If weight raising has to terminate for bfqq, then next | |
4556 | * function causes an immediate update of bfqq's weight, | |
4557 | * without waiting for next activation. As a consequence, on | |
4558 | * expiration, bfqq will be timestamped as if has never been | |
4559 | * weight-raised during this service slot, even if it has | |
4560 | * received part or even most of the service as a | |
4561 | * weight-raised queue. This inflates bfqq's timestamps, which | |
4562 | * is beneficial, as bfqq is then more willing to leave the | |
4563 | * device immediately to possible other weight-raised queues. | |
4564 | */ | |
4565 | bfq_update_wr_data(bfqd, bfqq); | |
4566 | ||
aee69d78 PV |
4567 | /* |
4568 | * Expire bfqq, pretending that its budget expired, if bfqq | |
4569 | * belongs to CLASS_IDLE and other queues are waiting for | |
4570 | * service. | |
4571 | */ | |
73d58118 | 4572 | if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq))) |
d0edc247 | 4573 | goto return_rq; |
aee69d78 | 4574 | |
aee69d78 | 4575 | bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED); |
d0edc247 PV |
4576 | |
4577 | return_rq: | |
aee69d78 PV |
4578 | return rq; |
4579 | } | |
4580 | ||
4581 | static bool bfq_has_work(struct blk_mq_hw_ctx *hctx) | |
4582 | { | |
4583 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
4584 | ||
4585 | /* | |
4586 | * Avoiding lock: a race on bfqd->busy_queues should cause at | |
4587 | * most a call to dispatch for nothing | |
4588 | */ | |
4589 | return !list_empty_careful(&bfqd->dispatch) || | |
73d58118 | 4590 | bfq_tot_busy_queues(bfqd) > 0; |
aee69d78 PV |
4591 | } |
4592 | ||
4593 | static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) | |
4594 | { | |
4595 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
4596 | struct request *rq = NULL; | |
4597 | struct bfq_queue *bfqq = NULL; | |
4598 | ||
4599 | if (!list_empty(&bfqd->dispatch)) { | |
4600 | rq = list_first_entry(&bfqd->dispatch, struct request, | |
4601 | queuelist); | |
4602 | list_del_init(&rq->queuelist); | |
4603 | ||
4604 | bfqq = RQ_BFQQ(rq); | |
4605 | ||
4606 | if (bfqq) { | |
4607 | /* | |
4608 | * Increment counters here, because this | |
4609 | * dispatch does not follow the standard | |
4610 | * dispatch flow (where counters are | |
4611 | * incremented) | |
4612 | */ | |
4613 | bfqq->dispatched++; | |
4614 | ||
4615 | goto inc_in_driver_start_rq; | |
4616 | } | |
4617 | ||
4618 | /* | |
a7877390 PV |
4619 | * We exploit the bfq_finish_requeue_request hook to |
4620 | * decrement rq_in_driver, but | |
4621 | * bfq_finish_requeue_request will not be invoked on | |
4622 | * this request. So, to avoid unbalance, just start | |
4623 | * this request, without incrementing rq_in_driver. As | |
4624 | * a negative consequence, rq_in_driver is deceptively | |
4625 | * lower than it should be while this request is in | |
4626 | * service. This may cause bfq_schedule_dispatch to be | |
4627 | * invoked uselessly. | |
aee69d78 PV |
4628 | * |
4629 | * As for implementing an exact solution, the | |
a7877390 PV |
4630 | * bfq_finish_requeue_request hook, if defined, is |
4631 | * probably invoked also on this request. So, by | |
4632 | * exploiting this hook, we could 1) increment | |
4633 | * rq_in_driver here, and 2) decrement it in | |
4634 | * bfq_finish_requeue_request. Such a solution would | |
4635 | * let the value of the counter be always accurate, | |
4636 | * but it would entail using an extra interface | |
4637 | * function. This cost seems higher than the benefit, | |
4638 | * being the frequency of non-elevator-private | |
aee69d78 PV |
4639 | * requests very low. |
4640 | */ | |
4641 | goto start_rq; | |
4642 | } | |
4643 | ||
73d58118 PV |
4644 | bfq_log(bfqd, "dispatch requests: %d busy queues", |
4645 | bfq_tot_busy_queues(bfqd)); | |
aee69d78 | 4646 | |
73d58118 | 4647 | if (bfq_tot_busy_queues(bfqd) == 0) |
aee69d78 PV |
4648 | goto exit; |
4649 | ||
4650 | /* | |
4651 | * Force device to serve one request at a time if | |
4652 | * strict_guarantees is true. Forcing this service scheme is | |
4653 | * currently the ONLY way to guarantee that the request | |
4654 | * service order enforced by the scheduler is respected by a | |
4655 | * queueing device. Otherwise the device is free even to make | |
4656 | * some unlucky request wait for as long as the device | |
4657 | * wishes. | |
4658 | * | |
4659 | * Of course, serving one request at at time may cause loss of | |
4660 | * throughput. | |
4661 | */ | |
4662 | if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0) | |
4663 | goto exit; | |
4664 | ||
4665 | bfqq = bfq_select_queue(bfqd); | |
4666 | if (!bfqq) | |
4667 | goto exit; | |
4668 | ||
4669 | rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq); | |
4670 | ||
4671 | if (rq) { | |
4672 | inc_in_driver_start_rq: | |
4673 | bfqd->rq_in_driver++; | |
4674 | start_rq: | |
4675 | rq->rq_flags |= RQF_STARTED; | |
4676 | } | |
4677 | exit: | |
4678 | return rq; | |
4679 | } | |
4680 | ||
8060c47b | 4681 | #ifdef CONFIG_BFQ_CGROUP_DEBUG |
9b25bd03 PV |
4682 | static void bfq_update_dispatch_stats(struct request_queue *q, |
4683 | struct request *rq, | |
4684 | struct bfq_queue *in_serv_queue, | |
4685 | bool idle_timer_disabled) | |
4686 | { | |
4687 | struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL; | |
aee69d78 | 4688 | |
24bfd19b | 4689 | if (!idle_timer_disabled && !bfqq) |
9b25bd03 | 4690 | return; |
24bfd19b PV |
4691 | |
4692 | /* | |
4693 | * rq and bfqq are guaranteed to exist until this function | |
4694 | * ends, for the following reasons. First, rq can be | |
4695 | * dispatched to the device, and then can be completed and | |
4696 | * freed, only after this function ends. Second, rq cannot be | |
4697 | * merged (and thus freed because of a merge) any longer, | |
4698 | * because it has already started. Thus rq cannot be freed | |
4699 | * before this function ends, and, since rq has a reference to | |
4700 | * bfqq, the same guarantee holds for bfqq too. | |
4701 | * | |
4702 | * In addition, the following queue lock guarantees that | |
4703 | * bfqq_group(bfqq) exists as well. | |
4704 | */ | |
0d945c1f | 4705 | spin_lock_irq(&q->queue_lock); |
24bfd19b PV |
4706 | if (idle_timer_disabled) |
4707 | /* | |
4708 | * Since the idle timer has been disabled, | |
4709 | * in_serv_queue contained some request when | |
4710 | * __bfq_dispatch_request was invoked above, which | |
4711 | * implies that rq was picked exactly from | |
4712 | * in_serv_queue. Thus in_serv_queue == bfqq, and is | |
4713 | * therefore guaranteed to exist because of the above | |
4714 | * arguments. | |
4715 | */ | |
4716 | bfqg_stats_update_idle_time(bfqq_group(in_serv_queue)); | |
4717 | if (bfqq) { | |
4718 | struct bfq_group *bfqg = bfqq_group(bfqq); | |
4719 | ||
4720 | bfqg_stats_update_avg_queue_size(bfqg); | |
4721 | bfqg_stats_set_start_empty_time(bfqg); | |
4722 | bfqg_stats_update_io_remove(bfqg, rq->cmd_flags); | |
4723 | } | |
0d945c1f | 4724 | spin_unlock_irq(&q->queue_lock); |
9b25bd03 PV |
4725 | } |
4726 | #else | |
4727 | static inline void bfq_update_dispatch_stats(struct request_queue *q, | |
4728 | struct request *rq, | |
4729 | struct bfq_queue *in_serv_queue, | |
4730 | bool idle_timer_disabled) {} | |
8060c47b | 4731 | #endif /* CONFIG_BFQ_CGROUP_DEBUG */ |
24bfd19b | 4732 | |
9b25bd03 PV |
4733 | static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx) |
4734 | { | |
4735 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
4736 | struct request *rq; | |
4737 | struct bfq_queue *in_serv_queue; | |
4738 | bool waiting_rq, idle_timer_disabled; | |
4739 | ||
4740 | spin_lock_irq(&bfqd->lock); | |
4741 | ||
4742 | in_serv_queue = bfqd->in_service_queue; | |
4743 | waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue); | |
4744 | ||
4745 | rq = __bfq_dispatch_request(hctx); | |
4746 | ||
4747 | idle_timer_disabled = | |
4748 | waiting_rq && !bfq_bfqq_wait_request(in_serv_queue); | |
4749 | ||
4750 | spin_unlock_irq(&bfqd->lock); | |
4751 | ||
4752 | bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue, | |
4753 | idle_timer_disabled); | |
4754 | ||
aee69d78 PV |
4755 | return rq; |
4756 | } | |
4757 | ||
4758 | /* | |
4759 | * Task holds one reference to the queue, dropped when task exits. Each rq | |
4760 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
4761 | * | |
4762 | * Scheduler lock must be held here. Recall not to use bfqq after calling | |
4763 | * this function on it. | |
4764 | */ | |
ea25da48 | 4765 | void bfq_put_queue(struct bfq_queue *bfqq) |
aee69d78 | 4766 | { |
e21b7a0b AA |
4767 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
4768 | struct bfq_group *bfqg = bfqq_group(bfqq); | |
4769 | #endif | |
4770 | ||
aee69d78 PV |
4771 | if (bfqq->bfqd) |
4772 | bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", | |
4773 | bfqq, bfqq->ref); | |
4774 | ||
4775 | bfqq->ref--; | |
4776 | if (bfqq->ref) | |
4777 | return; | |
4778 | ||
99fead8d | 4779 | if (!hlist_unhashed(&bfqq->burst_list_node)) { |
e1b2324d | 4780 | hlist_del_init(&bfqq->burst_list_node); |
99fead8d PV |
4781 | /* |
4782 | * Decrement also burst size after the removal, if the | |
4783 | * process associated with bfqq is exiting, and thus | |
4784 | * does not contribute to the burst any longer. This | |
4785 | * decrement helps filter out false positives of large | |
4786 | * bursts, when some short-lived process (often due to | |
4787 | * the execution of commands by some service) happens | |
4788 | * to start and exit while a complex application is | |
4789 | * starting, and thus spawning several processes that | |
4790 | * do I/O (and that *must not* be treated as a large | |
4791 | * burst, see comments on bfq_handle_burst). | |
4792 | * | |
4793 | * In particular, the decrement is performed only if: | |
4794 | * 1) bfqq is not a merged queue, because, if it is, | |
4795 | * then this free of bfqq is not triggered by the exit | |
4796 | * of the process bfqq is associated with, but exactly | |
4797 | * by the fact that bfqq has just been merged. | |
4798 | * 2) burst_size is greater than 0, to handle | |
4799 | * unbalanced decrements. Unbalanced decrements may | |
4800 | * happen in te following case: bfqq is inserted into | |
4801 | * the current burst list--without incrementing | |
4802 | * bust_size--because of a split, but the current | |
4803 | * burst list is not the burst list bfqq belonged to | |
4804 | * (see comments on the case of a split in | |
4805 | * bfq_set_request). | |
4806 | */ | |
4807 | if (bfqq->bic && bfqq->bfqd->burst_size > 0) | |
4808 | bfqq->bfqd->burst_size--; | |
7cb04004 | 4809 | } |
e21b7a0b | 4810 | |
aee69d78 | 4811 | kmem_cache_free(bfq_pool, bfqq); |
e21b7a0b | 4812 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
8f9bebc3 | 4813 | bfqg_and_blkg_put(bfqg); |
e21b7a0b | 4814 | #endif |
aee69d78 PV |
4815 | } |
4816 | ||
36eca894 AA |
4817 | static void bfq_put_cooperator(struct bfq_queue *bfqq) |
4818 | { | |
4819 | struct bfq_queue *__bfqq, *next; | |
4820 | ||
4821 | /* | |
4822 | * If this queue was scheduled to merge with another queue, be | |
4823 | * sure to drop the reference taken on that queue (and others in | |
4824 | * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. | |
4825 | */ | |
4826 | __bfqq = bfqq->new_bfqq; | |
4827 | while (__bfqq) { | |
4828 | if (__bfqq == bfqq) | |
4829 | break; | |
4830 | next = __bfqq->new_bfqq; | |
4831 | bfq_put_queue(__bfqq); | |
4832 | __bfqq = next; | |
4833 | } | |
4834 | } | |
4835 | ||
aee69d78 PV |
4836 | static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
4837 | { | |
13a857a4 PV |
4838 | struct bfq_queue *item; |
4839 | struct hlist_node *n; | |
4840 | ||
aee69d78 | 4841 | if (bfqq == bfqd->in_service_queue) { |
3726112e | 4842 | __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT); |
aee69d78 PV |
4843 | bfq_schedule_dispatch(bfqd); |
4844 | } | |
4845 | ||
4846 | bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref); | |
4847 | ||
36eca894 AA |
4848 | bfq_put_cooperator(bfqq); |
4849 | ||
13a857a4 PV |
4850 | /* remove bfqq from woken list */ |
4851 | if (!hlist_unhashed(&bfqq->woken_list_node)) | |
4852 | hlist_del_init(&bfqq->woken_list_node); | |
4853 | ||
4854 | /* reset waker for all queues in woken list */ | |
4855 | hlist_for_each_entry_safe(item, n, &bfqq->woken_list, | |
4856 | woken_list_node) { | |
4857 | item->waker_bfqq = NULL; | |
4858 | bfq_clear_bfqq_has_waker(item); | |
4859 | hlist_del_init(&item->woken_list_node); | |
4860 | } | |
4861 | ||
aee69d78 PV |
4862 | bfq_put_queue(bfqq); /* release process reference */ |
4863 | } | |
4864 | ||
4865 | static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync) | |
4866 | { | |
4867 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync); | |
4868 | struct bfq_data *bfqd; | |
4869 | ||
4870 | if (bfqq) | |
4871 | bfqd = bfqq->bfqd; /* NULL if scheduler already exited */ | |
4872 | ||
4873 | if (bfqq && bfqd) { | |
4874 | unsigned long flags; | |
4875 | ||
4876 | spin_lock_irqsave(&bfqd->lock, flags); | |
dbc3117d | 4877 | bfqq->bic = NULL; |
aee69d78 PV |
4878 | bfq_exit_bfqq(bfqd, bfqq); |
4879 | bic_set_bfqq(bic, NULL, is_sync); | |
6fa3e8d3 | 4880 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
4881 | } |
4882 | } | |
4883 | ||
4884 | static void bfq_exit_icq(struct io_cq *icq) | |
4885 | { | |
4886 | struct bfq_io_cq *bic = icq_to_bic(icq); | |
4887 | ||
4888 | bfq_exit_icq_bfqq(bic, true); | |
4889 | bfq_exit_icq_bfqq(bic, false); | |
4890 | } | |
4891 | ||
4892 | /* | |
4893 | * Update the entity prio values; note that the new values will not | |
4894 | * be used until the next (re)activation. | |
4895 | */ | |
4896 | static void | |
4897 | bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) | |
4898 | { | |
4899 | struct task_struct *tsk = current; | |
4900 | int ioprio_class; | |
4901 | struct bfq_data *bfqd = bfqq->bfqd; | |
4902 | ||
4903 | if (!bfqd) | |
4904 | return; | |
4905 | ||
4906 | ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
4907 | switch (ioprio_class) { | |
4908 | default: | |
4909 | dev_err(bfqq->bfqd->queue->backing_dev_info->dev, | |
4910 | "bfq: bad prio class %d\n", ioprio_class); | |
fa393d1b | 4911 | /* fall through */ |
aee69d78 PV |
4912 | case IOPRIO_CLASS_NONE: |
4913 | /* | |
4914 | * No prio set, inherit CPU scheduling settings. | |
4915 | */ | |
4916 | bfqq->new_ioprio = task_nice_ioprio(tsk); | |
4917 | bfqq->new_ioprio_class = task_nice_ioclass(tsk); | |
4918 | break; | |
4919 | case IOPRIO_CLASS_RT: | |
4920 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
4921 | bfqq->new_ioprio_class = IOPRIO_CLASS_RT; | |
4922 | break; | |
4923 | case IOPRIO_CLASS_BE: | |
4924 | bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
4925 | bfqq->new_ioprio_class = IOPRIO_CLASS_BE; | |
4926 | break; | |
4927 | case IOPRIO_CLASS_IDLE: | |
4928 | bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE; | |
4929 | bfqq->new_ioprio = 7; | |
aee69d78 PV |
4930 | break; |
4931 | } | |
4932 | ||
4933 | if (bfqq->new_ioprio >= IOPRIO_BE_NR) { | |
4934 | pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n", | |
4935 | bfqq->new_ioprio); | |
4936 | bfqq->new_ioprio = IOPRIO_BE_NR; | |
4937 | } | |
4938 | ||
4939 | bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio); | |
4940 | bfqq->entity.prio_changed = 1; | |
4941 | } | |
4942 | ||
ea25da48 PV |
4943 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
4944 | struct bio *bio, bool is_sync, | |
4945 | struct bfq_io_cq *bic); | |
4946 | ||
aee69d78 PV |
4947 | static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio) |
4948 | { | |
4949 | struct bfq_data *bfqd = bic_to_bfqd(bic); | |
4950 | struct bfq_queue *bfqq; | |
4951 | int ioprio = bic->icq.ioc->ioprio; | |
4952 | ||
4953 | /* | |
4954 | * This condition may trigger on a newly created bic, be sure to | |
4955 | * drop the lock before returning. | |
4956 | */ | |
4957 | if (unlikely(!bfqd) || likely(bic->ioprio == ioprio)) | |
4958 | return; | |
4959 | ||
4960 | bic->ioprio = ioprio; | |
4961 | ||
4962 | bfqq = bic_to_bfqq(bic, false); | |
4963 | if (bfqq) { | |
4964 | /* release process reference on this queue */ | |
4965 | bfq_put_queue(bfqq); | |
4966 | bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic); | |
4967 | bic_set_bfqq(bic, bfqq, false); | |
4968 | } | |
4969 | ||
4970 | bfqq = bic_to_bfqq(bic, true); | |
4971 | if (bfqq) | |
4972 | bfq_set_next_ioprio_data(bfqq, bic); | |
4973 | } | |
4974 | ||
4975 | static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
4976 | struct bfq_io_cq *bic, pid_t pid, int is_sync) | |
4977 | { | |
4978 | RB_CLEAR_NODE(&bfqq->entity.rb_node); | |
4979 | INIT_LIST_HEAD(&bfqq->fifo); | |
e1b2324d | 4980 | INIT_HLIST_NODE(&bfqq->burst_list_node); |
13a857a4 PV |
4981 | INIT_HLIST_NODE(&bfqq->woken_list_node); |
4982 | INIT_HLIST_HEAD(&bfqq->woken_list); | |
aee69d78 PV |
4983 | |
4984 | bfqq->ref = 0; | |
4985 | bfqq->bfqd = bfqd; | |
4986 | ||
4987 | if (bic) | |
4988 | bfq_set_next_ioprio_data(bfqq, bic); | |
4989 | ||
4990 | if (is_sync) { | |
d5be3fef PV |
4991 | /* |
4992 | * No need to mark as has_short_ttime if in | |
4993 | * idle_class, because no device idling is performed | |
4994 | * for queues in idle class | |
4995 | */ | |
aee69d78 | 4996 | if (!bfq_class_idle(bfqq)) |
d5be3fef PV |
4997 | /* tentatively mark as has_short_ttime */ |
4998 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
aee69d78 | 4999 | bfq_mark_bfqq_sync(bfqq); |
e1b2324d | 5000 | bfq_mark_bfqq_just_created(bfqq); |
aee69d78 PV |
5001 | } else |
5002 | bfq_clear_bfqq_sync(bfqq); | |
5003 | ||
5004 | /* set end request to minus infinity from now */ | |
5005 | bfqq->ttime.last_end_request = ktime_get_ns() + 1; | |
5006 | ||
5007 | bfq_mark_bfqq_IO_bound(bfqq); | |
5008 | ||
5009 | bfqq->pid = pid; | |
5010 | ||
5011 | /* Tentative initial value to trade off between thr and lat */ | |
54b60456 | 5012 | bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
aee69d78 | 5013 | bfqq->budget_timeout = bfq_smallest_from_now(); |
aee69d78 | 5014 | |
44e44a1b | 5015 | bfqq->wr_coeff = 1; |
36eca894 | 5016 | bfqq->last_wr_start_finish = jiffies; |
77b7dcea | 5017 | bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now(); |
36eca894 | 5018 | bfqq->split_time = bfq_smallest_from_now(); |
77b7dcea PV |
5019 | |
5020 | /* | |
a34b0244 PV |
5021 | * To not forget the possibly high bandwidth consumed by a |
5022 | * process/queue in the recent past, | |
5023 | * bfq_bfqq_softrt_next_start() returns a value at least equal | |
5024 | * to the current value of bfqq->soft_rt_next_start (see | |
5025 | * comments on bfq_bfqq_softrt_next_start). Set | |
5026 | * soft_rt_next_start to now, to mean that bfqq has consumed | |
5027 | * no bandwidth so far. | |
77b7dcea | 5028 | */ |
a34b0244 | 5029 | bfqq->soft_rt_next_start = jiffies; |
44e44a1b | 5030 | |
aee69d78 PV |
5031 | /* first request is almost certainly seeky */ |
5032 | bfqq->seek_history = 1; | |
5033 | } | |
5034 | ||
5035 | static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, | |
e21b7a0b | 5036 | struct bfq_group *bfqg, |
aee69d78 PV |
5037 | int ioprio_class, int ioprio) |
5038 | { | |
5039 | switch (ioprio_class) { | |
5040 | case IOPRIO_CLASS_RT: | |
e21b7a0b | 5041 | return &bfqg->async_bfqq[0][ioprio]; |
aee69d78 PV |
5042 | case IOPRIO_CLASS_NONE: |
5043 | ioprio = IOPRIO_NORM; | |
5044 | /* fall through */ | |
5045 | case IOPRIO_CLASS_BE: | |
e21b7a0b | 5046 | return &bfqg->async_bfqq[1][ioprio]; |
aee69d78 | 5047 | case IOPRIO_CLASS_IDLE: |
e21b7a0b | 5048 | return &bfqg->async_idle_bfqq; |
aee69d78 PV |
5049 | default: |
5050 | return NULL; | |
5051 | } | |
5052 | } | |
5053 | ||
5054 | static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, | |
5055 | struct bio *bio, bool is_sync, | |
5056 | struct bfq_io_cq *bic) | |
5057 | { | |
5058 | const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); | |
5059 | const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); | |
5060 | struct bfq_queue **async_bfqq = NULL; | |
5061 | struct bfq_queue *bfqq; | |
e21b7a0b | 5062 | struct bfq_group *bfqg; |
aee69d78 PV |
5063 | |
5064 | rcu_read_lock(); | |
5065 | ||
0fe061b9 | 5066 | bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio)); |
e21b7a0b AA |
5067 | if (!bfqg) { |
5068 | bfqq = &bfqd->oom_bfqq; | |
5069 | goto out; | |
5070 | } | |
5071 | ||
aee69d78 | 5072 | if (!is_sync) { |
e21b7a0b | 5073 | async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
aee69d78 PV |
5074 | ioprio); |
5075 | bfqq = *async_bfqq; | |
5076 | if (bfqq) | |
5077 | goto out; | |
5078 | } | |
5079 | ||
5080 | bfqq = kmem_cache_alloc_node(bfq_pool, | |
5081 | GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN, | |
5082 | bfqd->queue->node); | |
5083 | ||
5084 | if (bfqq) { | |
5085 | bfq_init_bfqq(bfqd, bfqq, bic, current->pid, | |
5086 | is_sync); | |
e21b7a0b | 5087 | bfq_init_entity(&bfqq->entity, bfqg); |
aee69d78 PV |
5088 | bfq_log_bfqq(bfqd, bfqq, "allocated"); |
5089 | } else { | |
5090 | bfqq = &bfqd->oom_bfqq; | |
5091 | bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); | |
5092 | goto out; | |
5093 | } | |
5094 | ||
5095 | /* | |
5096 | * Pin the queue now that it's allocated, scheduler exit will | |
5097 | * prune it. | |
5098 | */ | |
5099 | if (async_bfqq) { | |
e21b7a0b AA |
5100 | bfqq->ref++; /* |
5101 | * Extra group reference, w.r.t. sync | |
5102 | * queue. This extra reference is removed | |
5103 | * only if bfqq->bfqg disappears, to | |
5104 | * guarantee that this queue is not freed | |
5105 | * until its group goes away. | |
5106 | */ | |
5107 | bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", | |
aee69d78 PV |
5108 | bfqq, bfqq->ref); |
5109 | *async_bfqq = bfqq; | |
5110 | } | |
5111 | ||
5112 | out: | |
5113 | bfqq->ref++; /* get a process reference to this queue */ | |
5114 | bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref); | |
5115 | rcu_read_unlock(); | |
5116 | return bfqq; | |
5117 | } | |
5118 | ||
5119 | static void bfq_update_io_thinktime(struct bfq_data *bfqd, | |
5120 | struct bfq_queue *bfqq) | |
5121 | { | |
5122 | struct bfq_ttime *ttime = &bfqq->ttime; | |
5123 | u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request; | |
5124 | ||
5125 | elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle); | |
5126 | ||
5127 | ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8; | |
5128 | ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8); | |
5129 | ttime->ttime_mean = div64_ul(ttime->ttime_total + 128, | |
5130 | ttime->ttime_samples); | |
5131 | } | |
5132 | ||
5133 | static void | |
5134 | bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5135 | struct request *rq) | |
5136 | { | |
aee69d78 | 5137 | bfqq->seek_history <<= 1; |
d87447d8 | 5138 | bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq); |
7074f076 PV |
5139 | |
5140 | if (bfqq->wr_coeff > 1 && | |
5141 | bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time && | |
5142 | BFQQ_TOTALLY_SEEKY(bfqq)) | |
5143 | bfq_bfqq_end_wr(bfqq); | |
aee69d78 PV |
5144 | } |
5145 | ||
d5be3fef PV |
5146 | static void bfq_update_has_short_ttime(struct bfq_data *bfqd, |
5147 | struct bfq_queue *bfqq, | |
5148 | struct bfq_io_cq *bic) | |
aee69d78 | 5149 | { |
766d6141 | 5150 | bool has_short_ttime = true, state_changed; |
aee69d78 | 5151 | |
d5be3fef PV |
5152 | /* |
5153 | * No need to update has_short_ttime if bfqq is async or in | |
5154 | * idle io prio class, or if bfq_slice_idle is zero, because | |
5155 | * no device idling is performed for bfqq in this case. | |
5156 | */ | |
5157 | if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) || | |
5158 | bfqd->bfq_slice_idle == 0) | |
aee69d78 PV |
5159 | return; |
5160 | ||
36eca894 AA |
5161 | /* Idle window just restored, statistics are meaningless. */ |
5162 | if (time_is_after_eq_jiffies(bfqq->split_time + | |
5163 | bfqd->bfq_wr_min_idle_time)) | |
5164 | return; | |
5165 | ||
d5be3fef PV |
5166 | /* Think time is infinite if no process is linked to |
5167 | * bfqq. Otherwise check average think time to | |
5168 | * decide whether to mark as has_short_ttime | |
5169 | */ | |
aee69d78 | 5170 | if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
d5be3fef PV |
5171 | (bfq_sample_valid(bfqq->ttime.ttime_samples) && |
5172 | bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle)) | |
5173 | has_short_ttime = false; | |
5174 | ||
766d6141 | 5175 | state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq); |
aee69d78 | 5176 | |
d5be3fef PV |
5177 | if (has_short_ttime) |
5178 | bfq_mark_bfqq_has_short_ttime(bfqq); | |
aee69d78 | 5179 | else |
d5be3fef | 5180 | bfq_clear_bfqq_has_short_ttime(bfqq); |
766d6141 PV |
5181 | |
5182 | /* | |
5183 | * Until the base value for the total service time gets | |
5184 | * finally computed for bfqq, the inject limit does depend on | |
5185 | * the think-time state (short|long). In particular, the limit | |
5186 | * is 0 or 1 if the think time is deemed, respectively, as | |
5187 | * short or long (details in the comments in | |
5188 | * bfq_update_inject_limit()). Accordingly, the next | |
5189 | * instructions reset the inject limit if the think-time state | |
5190 | * has changed and the above base value is still to be | |
5191 | * computed. | |
5192 | * | |
5193 | * However, the reset is performed only if more than 100 ms | |
5194 | * have elapsed since the last update of the inject limit, or | |
5195 | * (inclusive) if the change is from short to long think | |
5196 | * time. The reason for this waiting is as follows. | |
5197 | * | |
5198 | * bfqq may have a long think time because of a | |
5199 | * synchronization with some other queue, i.e., because the | |
5200 | * I/O of some other queue may need to be completed for bfqq | |
13a857a4 PV |
5201 | * to receive new I/O. Details in the comments on the choice |
5202 | * of the queue for injection in bfq_select_queue(). | |
766d6141 | 5203 | * |
13a857a4 PV |
5204 | * As stressed in those comments, if such a synchronization is |
5205 | * actually in place, then, without injection on bfqq, the | |
5206 | * blocking I/O cannot happen to served while bfqq is in | |
5207 | * service. As a consequence, if bfqq is granted | |
5208 | * I/O-dispatch-plugging, then bfqq remains empty, and no I/O | |
5209 | * is dispatched, until the idle timeout fires. This is likely | |
5210 | * to result in lower bandwidth and higher latencies for bfqq, | |
5211 | * and in a severe loss of total throughput. | |
766d6141 PV |
5212 | * |
5213 | * On the opposite end, a non-zero inject limit may allow the | |
5214 | * I/O that blocks bfqq to be executed soon, and therefore | |
13a857a4 PV |
5215 | * bfqq to receive new I/O soon. |
5216 | * | |
5217 | * But, if the blocking gets actually eliminated, then the | |
5218 | * next think-time sample for bfqq may be very low. This in | |
5219 | * turn may cause bfqq's think time to be deemed | |
5220 | * short. Without the 100 ms barrier, this new state change | |
5221 | * would cause the body of the next if to be executed | |
766d6141 PV |
5222 | * immediately. But this would set to 0 the inject |
5223 | * limit. Without injection, the blocking I/O would cause the | |
5224 | * think time of bfqq to become long again, and therefore the | |
5225 | * inject limit to be raised again, and so on. The only effect | |
5226 | * of such a steady oscillation between the two think-time | |
5227 | * states would be to prevent effective injection on bfqq. | |
5228 | * | |
5229 | * In contrast, if the inject limit is not reset during such a | |
5230 | * long time interval as 100 ms, then the number of short | |
5231 | * think time samples can grow significantly before the reset | |
13a857a4 PV |
5232 | * is performed. As a consequence, the think time state can |
5233 | * become stable before the reset. Therefore there will be no | |
5234 | * state change when the 100 ms elapse, and no reset of the | |
5235 | * inject limit. The inject limit remains steadily equal to 1 | |
5236 | * both during and after the 100 ms. So injection can be | |
766d6141 PV |
5237 | * performed at all times, and throughput gets boosted. |
5238 | * | |
5239 | * An inject limit equal to 1 is however in conflict, in | |
5240 | * general, with the fact that the think time of bfqq is | |
5241 | * short, because injection may be likely to delay bfqq's I/O | |
5242 | * (as explained in the comments in | |
5243 | * bfq_update_inject_limit()). But this does not happen in | |
5244 | * this special case, because bfqq's low think time is due to | |
5245 | * an effective handling of a synchronization, through | |
5246 | * injection. In this special case, bfqq's I/O does not get | |
5247 | * delayed by injection; on the contrary, bfqq's I/O is | |
5248 | * brought forward, because it is not blocked for | |
5249 | * milliseconds. | |
5250 | * | |
13a857a4 PV |
5251 | * In addition, serving the blocking I/O much sooner, and much |
5252 | * more frequently than once per I/O-plugging timeout, makes | |
5253 | * it much quicker to detect a waker queue (the concept of | |
5254 | * waker queue is defined in the comments in | |
5255 | * bfq_add_request()). This makes it possible to start sooner | |
5256 | * to boost throughput more effectively, by injecting the I/O | |
5257 | * of the waker queue unconditionally on every | |
5258 | * bfq_dispatch_request(). | |
5259 | * | |
5260 | * One last, important benefit of not resetting the inject | |
5261 | * limit before 100 ms is that, during this time interval, the | |
5262 | * base value for the total service time is likely to get | |
5263 | * finally computed for bfqq, freeing the inject limit from | |
5264 | * its relation with the think time. | |
766d6141 PV |
5265 | */ |
5266 | if (state_changed && bfqq->last_serv_time_ns == 0 && | |
5267 | (time_is_before_eq_jiffies(bfqq->decrease_time_jif + | |
5268 | msecs_to_jiffies(100)) || | |
5269 | !has_short_ttime)) | |
5270 | bfq_reset_inject_limit(bfqd, bfqq); | |
aee69d78 PV |
5271 | } |
5272 | ||
5273 | /* | |
5274 | * Called when a new fs request (rq) is added to bfqq. Check if there's | |
5275 | * something we should do about it. | |
5276 | */ | |
5277 | static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, | |
5278 | struct request *rq) | |
5279 | { | |
aee69d78 PV |
5280 | if (rq->cmd_flags & REQ_META) |
5281 | bfqq->meta_pending++; | |
5282 | ||
aee69d78 PV |
5283 | bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
5284 | ||
5285 | if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { | |
5286 | bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 && | |
5287 | blk_rq_sectors(rq) < 32; | |
5288 | bool budget_timeout = bfq_bfqq_budget_timeout(bfqq); | |
5289 | ||
5290 | /* | |
ac8b0cb4 PV |
5291 | * There is just this request queued: if |
5292 | * - the request is small, and | |
5293 | * - we are idling to boost throughput, and | |
5294 | * - the queue is not to be expired, | |
5295 | * then just exit. | |
aee69d78 PV |
5296 | * |
5297 | * In this way, if the device is being idled to wait | |
5298 | * for a new request from the in-service queue, we | |
5299 | * avoid unplugging the device and committing the | |
ac8b0cb4 PV |
5300 | * device to serve just a small request. In contrast |
5301 | * we wait for the block layer to decide when to | |
5302 | * unplug the device: hopefully, new requests will be | |
5303 | * merged to this one quickly, then the device will be | |
5304 | * unplugged and larger requests will be dispatched. | |
aee69d78 | 5305 | */ |
ac8b0cb4 PV |
5306 | if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) && |
5307 | !budget_timeout) | |
aee69d78 PV |
5308 | return; |
5309 | ||
5310 | /* | |
ac8b0cb4 PV |
5311 | * A large enough request arrived, or idling is being |
5312 | * performed to preserve service guarantees, or | |
5313 | * finally the queue is to be expired: in all these | |
5314 | * cases disk idling is to be stopped, so clear | |
5315 | * wait_request flag and reset timer. | |
aee69d78 PV |
5316 | */ |
5317 | bfq_clear_bfqq_wait_request(bfqq); | |
5318 | hrtimer_try_to_cancel(&bfqd->idle_slice_timer); | |
5319 | ||
5320 | /* | |
5321 | * The queue is not empty, because a new request just | |
5322 | * arrived. Hence we can safely expire the queue, in | |
5323 | * case of budget timeout, without risking that the | |
5324 | * timestamps of the queue are not updated correctly. | |
5325 | * See [1] for more details. | |
5326 | */ | |
5327 | if (budget_timeout) | |
5328 | bfq_bfqq_expire(bfqd, bfqq, false, | |
5329 | BFQQE_BUDGET_TIMEOUT); | |
5330 | } | |
5331 | } | |
5332 | ||
24bfd19b PV |
5333 | /* returns true if it causes the idle timer to be disabled */ |
5334 | static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq) | |
aee69d78 | 5335 | { |
36eca894 AA |
5336 | struct bfq_queue *bfqq = RQ_BFQQ(rq), |
5337 | *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true); | |
24bfd19b | 5338 | bool waiting, idle_timer_disabled = false; |
36eca894 AA |
5339 | |
5340 | if (new_bfqq) { | |
36eca894 AA |
5341 | /* |
5342 | * Release the request's reference to the old bfqq | |
5343 | * and make sure one is taken to the shared queue. | |
5344 | */ | |
5345 | new_bfqq->allocated++; | |
5346 | bfqq->allocated--; | |
5347 | new_bfqq->ref++; | |
5348 | /* | |
5349 | * If the bic associated with the process | |
5350 | * issuing this request still points to bfqq | |
5351 | * (and thus has not been already redirected | |
5352 | * to new_bfqq or even some other bfq_queue), | |
5353 | * then complete the merge and redirect it to | |
5354 | * new_bfqq. | |
5355 | */ | |
5356 | if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq) | |
5357 | bfq_merge_bfqqs(bfqd, RQ_BIC(rq), | |
5358 | bfqq, new_bfqq); | |
894df937 PV |
5359 | |
5360 | bfq_clear_bfqq_just_created(bfqq); | |
36eca894 AA |
5361 | /* |
5362 | * rq is about to be enqueued into new_bfqq, | |
5363 | * release rq reference on bfqq | |
5364 | */ | |
5365 | bfq_put_queue(bfqq); | |
5366 | rq->elv.priv[1] = new_bfqq; | |
5367 | bfqq = new_bfqq; | |
5368 | } | |
aee69d78 | 5369 | |
a3f9bce3 PV |
5370 | bfq_update_io_thinktime(bfqd, bfqq); |
5371 | bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq)); | |
5372 | bfq_update_io_seektime(bfqd, bfqq, rq); | |
5373 | ||
24bfd19b | 5374 | waiting = bfqq && bfq_bfqq_wait_request(bfqq); |
aee69d78 | 5375 | bfq_add_request(rq); |
24bfd19b | 5376 | idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq); |
aee69d78 PV |
5377 | |
5378 | rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)]; | |
5379 | list_add_tail(&rq->queuelist, &bfqq->fifo); | |
5380 | ||
5381 | bfq_rq_enqueued(bfqd, bfqq, rq); | |
24bfd19b PV |
5382 | |
5383 | return idle_timer_disabled; | |
aee69d78 PV |
5384 | } |
5385 | ||
8060c47b | 5386 | #ifdef CONFIG_BFQ_CGROUP_DEBUG |
9b25bd03 PV |
5387 | static void bfq_update_insert_stats(struct request_queue *q, |
5388 | struct bfq_queue *bfqq, | |
5389 | bool idle_timer_disabled, | |
5390 | unsigned int cmd_flags) | |
5391 | { | |
5392 | if (!bfqq) | |
5393 | return; | |
5394 | ||
5395 | /* | |
5396 | * bfqq still exists, because it can disappear only after | |
5397 | * either it is merged with another queue, or the process it | |
5398 | * is associated with exits. But both actions must be taken by | |
5399 | * the same process currently executing this flow of | |
5400 | * instructions. | |
5401 | * | |
5402 | * In addition, the following queue lock guarantees that | |
5403 | * bfqq_group(bfqq) exists as well. | |
5404 | */ | |
0d945c1f | 5405 | spin_lock_irq(&q->queue_lock); |
9b25bd03 PV |
5406 | bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags); |
5407 | if (idle_timer_disabled) | |
5408 | bfqg_stats_update_idle_time(bfqq_group(bfqq)); | |
0d945c1f | 5409 | spin_unlock_irq(&q->queue_lock); |
9b25bd03 PV |
5410 | } |
5411 | #else | |
5412 | static inline void bfq_update_insert_stats(struct request_queue *q, | |
5413 | struct bfq_queue *bfqq, | |
5414 | bool idle_timer_disabled, | |
5415 | unsigned int cmd_flags) {} | |
8060c47b | 5416 | #endif /* CONFIG_BFQ_CGROUP_DEBUG */ |
9b25bd03 | 5417 | |
aee69d78 PV |
5418 | static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
5419 | bool at_head) | |
5420 | { | |
5421 | struct request_queue *q = hctx->queue; | |
5422 | struct bfq_data *bfqd = q->elevator->elevator_data; | |
18e5a57d | 5423 | struct bfq_queue *bfqq; |
24bfd19b PV |
5424 | bool idle_timer_disabled = false; |
5425 | unsigned int cmd_flags; | |
aee69d78 PV |
5426 | |
5427 | spin_lock_irq(&bfqd->lock); | |
5428 | if (blk_mq_sched_try_insert_merge(q, rq)) { | |
5429 | spin_unlock_irq(&bfqd->lock); | |
5430 | return; | |
5431 | } | |
5432 | ||
5433 | spin_unlock_irq(&bfqd->lock); | |
5434 | ||
5435 | blk_mq_sched_request_inserted(rq); | |
5436 | ||
5437 | spin_lock_irq(&bfqd->lock); | |
18e5a57d | 5438 | bfqq = bfq_init_rq(rq); |
aee69d78 PV |
5439 | if (at_head || blk_rq_is_passthrough(rq)) { |
5440 | if (at_head) | |
5441 | list_add(&rq->queuelist, &bfqd->dispatch); | |
5442 | else | |
5443 | list_add_tail(&rq->queuelist, &bfqd->dispatch); | |
18e5a57d | 5444 | } else { /* bfqq is assumed to be non null here */ |
24bfd19b | 5445 | idle_timer_disabled = __bfq_insert_request(bfqd, rq); |
614822f8 LM |
5446 | /* |
5447 | * Update bfqq, because, if a queue merge has occurred | |
5448 | * in __bfq_insert_request, then rq has been | |
5449 | * redirected into a new queue. | |
5450 | */ | |
5451 | bfqq = RQ_BFQQ(rq); | |
aee69d78 PV |
5452 | |
5453 | if (rq_mergeable(rq)) { | |
5454 | elv_rqhash_add(q, rq); | |
5455 | if (!q->last_merge) | |
5456 | q->last_merge = rq; | |
5457 | } | |
5458 | } | |
5459 | ||
24bfd19b PV |
5460 | /* |
5461 | * Cache cmd_flags before releasing scheduler lock, because rq | |
5462 | * may disappear afterwards (for example, because of a request | |
5463 | * merge). | |
5464 | */ | |
5465 | cmd_flags = rq->cmd_flags; | |
9b25bd03 | 5466 | |
6fa3e8d3 | 5467 | spin_unlock_irq(&bfqd->lock); |
24bfd19b | 5468 | |
9b25bd03 PV |
5469 | bfq_update_insert_stats(q, bfqq, idle_timer_disabled, |
5470 | cmd_flags); | |
aee69d78 PV |
5471 | } |
5472 | ||
5473 | static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx, | |
5474 | struct list_head *list, bool at_head) | |
5475 | { | |
5476 | while (!list_empty(list)) { | |
5477 | struct request *rq; | |
5478 | ||
5479 | rq = list_first_entry(list, struct request, queuelist); | |
5480 | list_del_init(&rq->queuelist); | |
5481 | bfq_insert_request(hctx, rq, at_head); | |
5482 | } | |
5483 | } | |
5484 | ||
5485 | static void bfq_update_hw_tag(struct bfq_data *bfqd) | |
5486 | { | |
b3c34981 PV |
5487 | struct bfq_queue *bfqq = bfqd->in_service_queue; |
5488 | ||
aee69d78 PV |
5489 | bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver, |
5490 | bfqd->rq_in_driver); | |
5491 | ||
5492 | if (bfqd->hw_tag == 1) | |
5493 | return; | |
5494 | ||
5495 | /* | |
5496 | * This sample is valid if the number of outstanding requests | |
5497 | * is large enough to allow a queueing behavior. Note that the | |
5498 | * sum is not exact, as it's not taking into account deactivated | |
5499 | * requests. | |
5500 | */ | |
a3c92560 | 5501 | if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD) |
aee69d78 PV |
5502 | return; |
5503 | ||
b3c34981 PV |
5504 | /* |
5505 | * If active queue hasn't enough requests and can idle, bfq might not | |
5506 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
5507 | * case | |
5508 | */ | |
5509 | if (bfqq && bfq_bfqq_has_short_ttime(bfqq) && | |
5510 | bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] < | |
5511 | BFQ_HW_QUEUE_THRESHOLD && | |
5512 | bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD) | |
5513 | return; | |
5514 | ||
aee69d78 PV |
5515 | if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
5516 | return; | |
5517 | ||
5518 | bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; | |
5519 | bfqd->max_rq_in_driver = 0; | |
5520 | bfqd->hw_tag_samples = 0; | |
8cacc5ab PV |
5521 | |
5522 | bfqd->nonrot_with_queueing = | |
5523 | blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag; | |
aee69d78 PV |
5524 | } |
5525 | ||
5526 | static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd) | |
5527 | { | |
ab0e43e9 PV |
5528 | u64 now_ns; |
5529 | u32 delta_us; | |
5530 | ||
aee69d78 PV |
5531 | bfq_update_hw_tag(bfqd); |
5532 | ||
5533 | bfqd->rq_in_driver--; | |
5534 | bfqq->dispatched--; | |
5535 | ||
44e44a1b PV |
5536 | if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) { |
5537 | /* | |
5538 | * Set budget_timeout (which we overload to store the | |
5539 | * time at which the queue remains with no backlog and | |
5540 | * no outstanding request; used by the weight-raising | |
5541 | * mechanism). | |
5542 | */ | |
5543 | bfqq->budget_timeout = jiffies; | |
1de0c4cd | 5544 | |
0471559c | 5545 | bfq_weights_tree_remove(bfqd, bfqq); |
44e44a1b PV |
5546 | } |
5547 | ||
ab0e43e9 PV |
5548 | now_ns = ktime_get_ns(); |
5549 | ||
5550 | bfqq->ttime.last_end_request = now_ns; | |
5551 | ||
5552 | /* | |
5553 | * Using us instead of ns, to get a reasonable precision in | |
5554 | * computing rate in next check. | |
5555 | */ | |
5556 | delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC); | |
5557 | ||
5558 | /* | |
5559 | * If the request took rather long to complete, and, according | |
5560 | * to the maximum request size recorded, this completion latency | |
5561 | * implies that the request was certainly served at a very low | |
5562 | * rate (less than 1M sectors/sec), then the whole observation | |
5563 | * interval that lasts up to this time instant cannot be a | |
5564 | * valid time interval for computing a new peak rate. Invoke | |
5565 | * bfq_update_rate_reset to have the following three steps | |
5566 | * taken: | |
5567 | * - close the observation interval at the last (previous) | |
5568 | * request dispatch or completion | |
5569 | * - compute rate, if possible, for that observation interval | |
5570 | * - reset to zero samples, which will trigger a proper | |
5571 | * re-initialization of the observation interval on next | |
5572 | * dispatch | |
5573 | */ | |
5574 | if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC && | |
5575 | (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us < | |
5576 | 1UL<<(BFQ_RATE_SHIFT - 10)) | |
5577 | bfq_update_rate_reset(bfqd, NULL); | |
5578 | bfqd->last_completion = now_ns; | |
13a857a4 | 5579 | bfqd->last_completed_rq_bfqq = bfqq; |
aee69d78 | 5580 | |
77b7dcea PV |
5581 | /* |
5582 | * If we are waiting to discover whether the request pattern | |
5583 | * of the task associated with the queue is actually | |
5584 | * isochronous, and both requisites for this condition to hold | |
5585 | * are now satisfied, then compute soft_rt_next_start (see the | |
5586 | * comments on the function bfq_bfqq_softrt_next_start()). We | |
20cd3245 PV |
5587 | * do not compute soft_rt_next_start if bfqq is in interactive |
5588 | * weight raising (see the comments in bfq_bfqq_expire() for | |
5589 | * an explanation). We schedule this delayed update when bfqq | |
5590 | * expires, if it still has in-flight requests. | |
77b7dcea PV |
5591 | */ |
5592 | if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 && | |
20cd3245 PV |
5593 | RB_EMPTY_ROOT(&bfqq->sort_list) && |
5594 | bfqq->wr_coeff != bfqd->bfq_wr_coeff) | |
77b7dcea PV |
5595 | bfqq->soft_rt_next_start = |
5596 | bfq_bfqq_softrt_next_start(bfqd, bfqq); | |
5597 | ||
aee69d78 PV |
5598 | /* |
5599 | * If this is the in-service queue, check if it needs to be expired, | |
5600 | * or if we want to idle in case it has no pending requests. | |
5601 | */ | |
5602 | if (bfqd->in_service_queue == bfqq) { | |
4420b095 PV |
5603 | if (bfq_bfqq_must_idle(bfqq)) { |
5604 | if (bfqq->dispatched == 0) | |
5605 | bfq_arm_slice_timer(bfqd); | |
5606 | /* | |
5607 | * If we get here, we do not expire bfqq, even | |
5608 | * if bfqq was in budget timeout or had no | |
5609 | * more requests (as controlled in the next | |
5610 | * conditional instructions). The reason for | |
5611 | * not expiring bfqq is as follows. | |
5612 | * | |
5613 | * Here bfqq->dispatched > 0 holds, but | |
5614 | * bfq_bfqq_must_idle() returned true. This | |
5615 | * implies that, even if no request arrives | |
5616 | * for bfqq before bfqq->dispatched reaches 0, | |
5617 | * bfqq will, however, not be expired on the | |
5618 | * completion event that causes bfqq->dispatch | |
5619 | * to reach zero. In contrast, on this event, | |
5620 | * bfqq will start enjoying device idling | |
5621 | * (I/O-dispatch plugging). | |
5622 | * | |
5623 | * But, if we expired bfqq here, bfqq would | |
5624 | * not have the chance to enjoy device idling | |
5625 | * when bfqq->dispatched finally reaches | |
5626 | * zero. This would expose bfqq to violation | |
5627 | * of its reserved service guarantees. | |
5628 | */ | |
aee69d78 PV |
5629 | return; |
5630 | } else if (bfq_may_expire_for_budg_timeout(bfqq)) | |
5631 | bfq_bfqq_expire(bfqd, bfqq, false, | |
5632 | BFQQE_BUDGET_TIMEOUT); | |
5633 | else if (RB_EMPTY_ROOT(&bfqq->sort_list) && | |
5634 | (bfqq->dispatched == 0 || | |
277a4a9b | 5635 | !bfq_better_to_idle(bfqq))) |
aee69d78 PV |
5636 | bfq_bfqq_expire(bfqd, bfqq, false, |
5637 | BFQQE_NO_MORE_REQUESTS); | |
5638 | } | |
3f7cb4f4 HT |
5639 | |
5640 | if (!bfqd->rq_in_driver) | |
5641 | bfq_schedule_dispatch(bfqd); | |
aee69d78 PV |
5642 | } |
5643 | ||
a7877390 | 5644 | static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq) |
aee69d78 PV |
5645 | { |
5646 | bfqq->allocated--; | |
5647 | ||
5648 | bfq_put_queue(bfqq); | |
5649 | } | |
5650 | ||
2341d662 PV |
5651 | /* |
5652 | * The processes associated with bfqq may happen to generate their | |
5653 | * cumulative I/O at a lower rate than the rate at which the device | |
5654 | * could serve the same I/O. This is rather probable, e.g., if only | |
5655 | * one process is associated with bfqq and the device is an SSD. It | |
5656 | * results in bfqq becoming often empty while in service. In this | |
5657 | * respect, if BFQ is allowed to switch to another queue when bfqq | |
5658 | * remains empty, then the device goes on being fed with I/O requests, | |
5659 | * and the throughput is not affected. In contrast, if BFQ is not | |
5660 | * allowed to switch to another queue---because bfqq is sync and | |
5661 | * I/O-dispatch needs to be plugged while bfqq is temporarily | |
5662 | * empty---then, during the service of bfqq, there will be frequent | |
5663 | * "service holes", i.e., time intervals during which bfqq gets empty | |
5664 | * and the device can only consume the I/O already queued in its | |
5665 | * hardware queues. During service holes, the device may even get to | |
5666 | * remaining idle. In the end, during the service of bfqq, the device | |
5667 | * is driven at a lower speed than the one it can reach with the kind | |
5668 | * of I/O flowing through bfqq. | |
5669 | * | |
5670 | * To counter this loss of throughput, BFQ implements a "request | |
5671 | * injection mechanism", which tries to fill the above service holes | |
5672 | * with I/O requests taken from other queues. The hard part in this | |
5673 | * mechanism is finding the right amount of I/O to inject, so as to | |
5674 | * both boost throughput and not break bfqq's bandwidth and latency | |
5675 | * guarantees. In this respect, the mechanism maintains a per-queue | |
5676 | * inject limit, computed as below. While bfqq is empty, the injection | |
5677 | * mechanism dispatches extra I/O requests only until the total number | |
5678 | * of I/O requests in flight---i.e., already dispatched but not yet | |
5679 | * completed---remains lower than this limit. | |
5680 | * | |
5681 | * A first definition comes in handy to introduce the algorithm by | |
5682 | * which the inject limit is computed. We define as first request for | |
5683 | * bfqq, an I/O request for bfqq that arrives while bfqq is in | |
5684 | * service, and causes bfqq to switch from empty to non-empty. The | |
5685 | * algorithm updates the limit as a function of the effect of | |
5686 | * injection on the service times of only the first requests of | |
5687 | * bfqq. The reason for this restriction is that these are the | |
5688 | * requests whose service time is affected most, because they are the | |
5689 | * first to arrive after injection possibly occurred. | |
5690 | * | |
5691 | * To evaluate the effect of injection, the algorithm measures the | |
5692 | * "total service time" of first requests. We define as total service | |
5693 | * time of an I/O request, the time that elapses since when the | |
5694 | * request is enqueued into bfqq, to when it is completed. This | |
5695 | * quantity allows the whole effect of injection to be measured. It is | |
5696 | * easy to see why. Suppose that some requests of other queues are | |
5697 | * actually injected while bfqq is empty, and that a new request R | |
5698 | * then arrives for bfqq. If the device does start to serve all or | |
5699 | * part of the injected requests during the service hole, then, | |
5700 | * because of this extra service, it may delay the next invocation of | |
5701 | * the dispatch hook of BFQ. Then, even after R gets eventually | |
5702 | * dispatched, the device may delay the actual service of R if it is | |
5703 | * still busy serving the extra requests, or if it decides to serve, | |
5704 | * before R, some extra request still present in its queues. As a | |
5705 | * conclusion, the cumulative extra delay caused by injection can be | |
5706 | * easily evaluated by just comparing the total service time of first | |
5707 | * requests with and without injection. | |
5708 | * | |
5709 | * The limit-update algorithm works as follows. On the arrival of a | |
5710 | * first request of bfqq, the algorithm measures the total time of the | |
5711 | * request only if one of the three cases below holds, and, for each | |
5712 | * case, it updates the limit as described below: | |
5713 | * | |
5714 | * (1) If there is no in-flight request. This gives a baseline for the | |
5715 | * total service time of the requests of bfqq. If the baseline has | |
5716 | * not been computed yet, then, after computing it, the limit is | |
5717 | * set to 1, to start boosting throughput, and to prepare the | |
5718 | * ground for the next case. If the baseline has already been | |
5719 | * computed, then it is updated, in case it results to be lower | |
5720 | * than the previous value. | |
5721 | * | |
5722 | * (2) If the limit is higher than 0 and there are in-flight | |
5723 | * requests. By comparing the total service time in this case with | |
5724 | * the above baseline, it is possible to know at which extent the | |
5725 | * current value of the limit is inflating the total service | |
5726 | * time. If the inflation is below a certain threshold, then bfqq | |
5727 | * is assumed to be suffering from no perceivable loss of its | |
5728 | * service guarantees, and the limit is even tentatively | |
5729 | * increased. If the inflation is above the threshold, then the | |
5730 | * limit is decreased. Due to the lack of any hysteresis, this | |
5731 | * logic makes the limit oscillate even in steady workload | |
5732 | * conditions. Yet we opted for it, because it is fast in reaching | |
5733 | * the best value for the limit, as a function of the current I/O | |
5734 | * workload. To reduce oscillations, this step is disabled for a | |
5735 | * short time interval after the limit happens to be decreased. | |
5736 | * | |
5737 | * (3) Periodically, after resetting the limit, to make sure that the | |
5738 | * limit eventually drops in case the workload changes. This is | |
5739 | * needed because, after the limit has gone safely up for a | |
5740 | * certain workload, it is impossible to guess whether the | |
5741 | * baseline total service time may have changed, without measuring | |
5742 | * it again without injection. A more effective version of this | |
5743 | * step might be to just sample the baseline, by interrupting | |
5744 | * injection only once, and then to reset/lower the limit only if | |
5745 | * the total service time with the current limit does happen to be | |
5746 | * too large. | |
5747 | * | |
5748 | * More details on each step are provided in the comments on the | |
5749 | * pieces of code that implement these steps: the branch handling the | |
5750 | * transition from empty to non empty in bfq_add_request(), the branch | |
5751 | * handling injection in bfq_select_queue(), and the function | |
5752 | * bfq_choose_bfqq_for_injection(). These comments also explain some | |
5753 | * exceptions, made by the injection mechanism in some special cases. | |
5754 | */ | |
5755 | static void bfq_update_inject_limit(struct bfq_data *bfqd, | |
5756 | struct bfq_queue *bfqq) | |
5757 | { | |
5758 | u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns; | |
5759 | unsigned int old_limit = bfqq->inject_limit; | |
5760 | ||
5761 | if (bfqq->last_serv_time_ns > 0) { | |
5762 | u64 threshold = (bfqq->last_serv_time_ns * 3)>>1; | |
5763 | ||
5764 | if (tot_time_ns >= threshold && old_limit > 0) { | |
5765 | bfqq->inject_limit--; | |
5766 | bfqq->decrease_time_jif = jiffies; | |
5767 | } else if (tot_time_ns < threshold && | |
5768 | old_limit < bfqd->max_rq_in_driver<<1) | |
5769 | bfqq->inject_limit++; | |
5770 | } | |
5771 | ||
5772 | /* | |
5773 | * Either we still have to compute the base value for the | |
5774 | * total service time, and there seem to be the right | |
5775 | * conditions to do it, or we can lower the last base value | |
5776 | * computed. | |
db599f9e PV |
5777 | * |
5778 | * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O | |
5779 | * request in flight, because this function is in the code | |
5780 | * path that handles the completion of a request of bfqq, and, | |
5781 | * in particular, this function is executed before | |
5782 | * bfqd->rq_in_driver is decremented in such a code path. | |
2341d662 | 5783 | */ |
db599f9e | 5784 | if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) || |
2341d662 PV |
5785 | tot_time_ns < bfqq->last_serv_time_ns) { |
5786 | bfqq->last_serv_time_ns = tot_time_ns; | |
5787 | /* | |
5788 | * Now we certainly have a base value: make sure we | |
5789 | * start trying injection. | |
5790 | */ | |
5791 | bfqq->inject_limit = max_t(unsigned int, 1, old_limit); | |
24792ad0 PV |
5792 | } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1) |
5793 | /* | |
5794 | * No I/O injected and no request still in service in | |
5795 | * the drive: these are the exact conditions for | |
5796 | * computing the base value of the total service time | |
5797 | * for bfqq. So let's update this value, because it is | |
5798 | * rather variable. For example, it varies if the size | |
5799 | * or the spatial locality of the I/O requests in bfqq | |
5800 | * change. | |
5801 | */ | |
5802 | bfqq->last_serv_time_ns = tot_time_ns; | |
5803 | ||
2341d662 PV |
5804 | |
5805 | /* update complete, not waiting for any request completion any longer */ | |
5806 | bfqd->waited_rq = NULL; | |
5807 | } | |
5808 | ||
a7877390 PV |
5809 | /* |
5810 | * Handle either a requeue or a finish for rq. The things to do are | |
5811 | * the same in both cases: all references to rq are to be dropped. In | |
5812 | * particular, rq is considered completed from the point of view of | |
5813 | * the scheduler. | |
5814 | */ | |
5815 | static void bfq_finish_requeue_request(struct request *rq) | |
aee69d78 | 5816 | { |
a7877390 | 5817 | struct bfq_queue *bfqq = RQ_BFQQ(rq); |
5bbf4e5a CH |
5818 | struct bfq_data *bfqd; |
5819 | ||
a7877390 PV |
5820 | /* |
5821 | * Requeue and finish hooks are invoked in blk-mq without | |
5822 | * checking whether the involved request is actually still | |
5823 | * referenced in the scheduler. To handle this fact, the | |
5824 | * following two checks make this function exit in case of | |
5825 | * spurious invocations, for which there is nothing to do. | |
5826 | * | |
5827 | * First, check whether rq has nothing to do with an elevator. | |
5828 | */ | |
5829 | if (unlikely(!(rq->rq_flags & RQF_ELVPRIV))) | |
5830 | return; | |
5831 | ||
5832 | /* | |
5833 | * rq either is not associated with any icq, or is an already | |
5834 | * requeued request that has not (yet) been re-inserted into | |
5835 | * a bfq_queue. | |
5836 | */ | |
5837 | if (!rq->elv.icq || !bfqq) | |
5bbf4e5a CH |
5838 | return; |
5839 | ||
5bbf4e5a | 5840 | bfqd = bfqq->bfqd; |
aee69d78 | 5841 | |
e21b7a0b AA |
5842 | if (rq->rq_flags & RQF_STARTED) |
5843 | bfqg_stats_update_completion(bfqq_group(bfqq), | |
522a7775 OS |
5844 | rq->start_time_ns, |
5845 | rq->io_start_time_ns, | |
e21b7a0b | 5846 | rq->cmd_flags); |
aee69d78 PV |
5847 | |
5848 | if (likely(rq->rq_flags & RQF_STARTED)) { | |
5849 | unsigned long flags; | |
5850 | ||
5851 | spin_lock_irqsave(&bfqd->lock, flags); | |
5852 | ||
2341d662 PV |
5853 | if (rq == bfqd->waited_rq) |
5854 | bfq_update_inject_limit(bfqd, bfqq); | |
5855 | ||
aee69d78 | 5856 | bfq_completed_request(bfqq, bfqd); |
a7877390 | 5857 | bfq_finish_requeue_request_body(bfqq); |
aee69d78 | 5858 | |
6fa3e8d3 | 5859 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
5860 | } else { |
5861 | /* | |
5862 | * Request rq may be still/already in the scheduler, | |
a7877390 PV |
5863 | * in which case we need to remove it (this should |
5864 | * never happen in case of requeue). And we cannot | |
aee69d78 PV |
5865 | * defer such a check and removal, to avoid |
5866 | * inconsistencies in the time interval from the end | |
5867 | * of this function to the start of the deferred work. | |
5868 | * This situation seems to occur only in process | |
5869 | * context, as a consequence of a merge. In the | |
5870 | * current version of the code, this implies that the | |
5871 | * lock is held. | |
5872 | */ | |
5873 | ||
614822f8 | 5874 | if (!RB_EMPTY_NODE(&rq->rb_node)) { |
7b9e9361 | 5875 | bfq_remove_request(rq->q, rq); |
614822f8 LM |
5876 | bfqg_stats_update_io_remove(bfqq_group(bfqq), |
5877 | rq->cmd_flags); | |
5878 | } | |
a7877390 | 5879 | bfq_finish_requeue_request_body(bfqq); |
aee69d78 PV |
5880 | } |
5881 | ||
a7877390 PV |
5882 | /* |
5883 | * Reset private fields. In case of a requeue, this allows | |
5884 | * this function to correctly do nothing if it is spuriously | |
5885 | * invoked again on this same request (see the check at the | |
5886 | * beginning of the function). Probably, a better general | |
5887 | * design would be to prevent blk-mq from invoking the requeue | |
5888 | * or finish hooks of an elevator, for a request that is not | |
5889 | * referred by that elevator. | |
5890 | * | |
5891 | * Resetting the following fields would break the | |
5892 | * request-insertion logic if rq is re-inserted into a bfq | |
5893 | * internal queue, without a re-preparation. Here we assume | |
5894 | * that re-insertions of requeued requests, without | |
5895 | * re-preparation, can happen only for pass_through or at_head | |
5896 | * requests (which are not re-inserted into bfq internal | |
5897 | * queues). | |
5898 | */ | |
aee69d78 PV |
5899 | rq->elv.priv[0] = NULL; |
5900 | rq->elv.priv[1] = NULL; | |
5901 | } | |
5902 | ||
36eca894 AA |
5903 | /* |
5904 | * Returns NULL if a new bfqq should be allocated, or the old bfqq if this | |
5905 | * was the last process referring to that bfqq. | |
5906 | */ | |
5907 | static struct bfq_queue * | |
5908 | bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) | |
5909 | { | |
5910 | bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); | |
5911 | ||
5912 | if (bfqq_process_refs(bfqq) == 1) { | |
5913 | bfqq->pid = current->pid; | |
5914 | bfq_clear_bfqq_coop(bfqq); | |
5915 | bfq_clear_bfqq_split_coop(bfqq); | |
5916 | return bfqq; | |
5917 | } | |
5918 | ||
5919 | bic_set_bfqq(bic, NULL, 1); | |
5920 | ||
5921 | bfq_put_cooperator(bfqq); | |
5922 | ||
5923 | bfq_put_queue(bfqq); | |
5924 | return NULL; | |
5925 | } | |
5926 | ||
5927 | static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd, | |
5928 | struct bfq_io_cq *bic, | |
5929 | struct bio *bio, | |
5930 | bool split, bool is_sync, | |
5931 | bool *new_queue) | |
5932 | { | |
5933 | struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync); | |
5934 | ||
5935 | if (likely(bfqq && bfqq != &bfqd->oom_bfqq)) | |
5936 | return bfqq; | |
5937 | ||
5938 | if (new_queue) | |
5939 | *new_queue = true; | |
5940 | ||
5941 | if (bfqq) | |
5942 | bfq_put_queue(bfqq); | |
5943 | bfqq = bfq_get_queue(bfqd, bio, is_sync, bic); | |
5944 | ||
5945 | bic_set_bfqq(bic, bfqq, is_sync); | |
e1b2324d AA |
5946 | if (split && is_sync) { |
5947 | if ((bic->was_in_burst_list && bfqd->large_burst) || | |
5948 | bic->saved_in_large_burst) | |
5949 | bfq_mark_bfqq_in_large_burst(bfqq); | |
5950 | else { | |
5951 | bfq_clear_bfqq_in_large_burst(bfqq); | |
5952 | if (bic->was_in_burst_list) | |
99fead8d PV |
5953 | /* |
5954 | * If bfqq was in the current | |
5955 | * burst list before being | |
5956 | * merged, then we have to add | |
5957 | * it back. And we do not need | |
5958 | * to increase burst_size, as | |
5959 | * we did not decrement | |
5960 | * burst_size when we removed | |
5961 | * bfqq from the burst list as | |
5962 | * a consequence of a merge | |
5963 | * (see comments in | |
5964 | * bfq_put_queue). In this | |
5965 | * respect, it would be rather | |
5966 | * costly to know whether the | |
5967 | * current burst list is still | |
5968 | * the same burst list from | |
5969 | * which bfqq was removed on | |
5970 | * the merge. To avoid this | |
5971 | * cost, if bfqq was in a | |
5972 | * burst list, then we add | |
5973 | * bfqq to the current burst | |
5974 | * list without any further | |
5975 | * check. This can cause | |
5976 | * inappropriate insertions, | |
5977 | * but rarely enough to not | |
5978 | * harm the detection of large | |
5979 | * bursts significantly. | |
5980 | */ | |
e1b2324d AA |
5981 | hlist_add_head(&bfqq->burst_list_node, |
5982 | &bfqd->burst_list); | |
5983 | } | |
36eca894 | 5984 | bfqq->split_time = jiffies; |
e1b2324d | 5985 | } |
36eca894 AA |
5986 | |
5987 | return bfqq; | |
5988 | } | |
5989 | ||
aee69d78 | 5990 | /* |
18e5a57d PV |
5991 | * Only reset private fields. The actual request preparation will be |
5992 | * performed by bfq_init_rq, when rq is either inserted or merged. See | |
5993 | * comments on bfq_init_rq for the reason behind this delayed | |
5994 | * preparation. | |
aee69d78 | 5995 | */ |
5bbf4e5a | 5996 | static void bfq_prepare_request(struct request *rq, struct bio *bio) |
18e5a57d PV |
5997 | { |
5998 | /* | |
5999 | * Regardless of whether we have an icq attached, we have to | |
6000 | * clear the scheduler pointers, as they might point to | |
6001 | * previously allocated bic/bfqq structs. | |
6002 | */ | |
6003 | rq->elv.priv[0] = rq->elv.priv[1] = NULL; | |
6004 | } | |
6005 | ||
6006 | /* | |
6007 | * If needed, init rq, allocate bfq data structures associated with | |
6008 | * rq, and increment reference counters in the destination bfq_queue | |
6009 | * for rq. Return the destination bfq_queue for rq, or NULL is rq is | |
6010 | * not associated with any bfq_queue. | |
6011 | * | |
6012 | * This function is invoked by the functions that perform rq insertion | |
6013 | * or merging. One may have expected the above preparation operations | |
6014 | * to be performed in bfq_prepare_request, and not delayed to when rq | |
6015 | * is inserted or merged. The rationale behind this delayed | |
6016 | * preparation is that, after the prepare_request hook is invoked for | |
6017 | * rq, rq may still be transformed into a request with no icq, i.e., a | |
6018 | * request not associated with any queue. No bfq hook is invoked to | |
636b8fe8 | 6019 | * signal this transformation. As a consequence, should these |
18e5a57d PV |
6020 | * preparation operations be performed when the prepare_request hook |
6021 | * is invoked, and should rq be transformed one moment later, bfq | |
6022 | * would end up in an inconsistent state, because it would have | |
6023 | * incremented some queue counters for an rq destined to | |
6024 | * transformation, without any chance to correctly lower these | |
6025 | * counters back. In contrast, no transformation can still happen for | |
6026 | * rq after rq has been inserted or merged. So, it is safe to execute | |
6027 | * these preparation operations when rq is finally inserted or merged. | |
6028 | */ | |
6029 | static struct bfq_queue *bfq_init_rq(struct request *rq) | |
aee69d78 | 6030 | { |
5bbf4e5a | 6031 | struct request_queue *q = rq->q; |
18e5a57d | 6032 | struct bio *bio = rq->bio; |
aee69d78 | 6033 | struct bfq_data *bfqd = q->elevator->elevator_data; |
9f210738 | 6034 | struct bfq_io_cq *bic; |
aee69d78 PV |
6035 | const int is_sync = rq_is_sync(rq); |
6036 | struct bfq_queue *bfqq; | |
36eca894 | 6037 | bool new_queue = false; |
13c931bd | 6038 | bool bfqq_already_existing = false, split = false; |
aee69d78 | 6039 | |
18e5a57d PV |
6040 | if (unlikely(!rq->elv.icq)) |
6041 | return NULL; | |
6042 | ||
72961c4e | 6043 | /* |
18e5a57d PV |
6044 | * Assuming that elv.priv[1] is set only if everything is set |
6045 | * for this rq. This holds true, because this function is | |
6046 | * invoked only for insertion or merging, and, after such | |
6047 | * events, a request cannot be manipulated any longer before | |
6048 | * being removed from bfq. | |
72961c4e | 6049 | */ |
18e5a57d PV |
6050 | if (rq->elv.priv[1]) |
6051 | return rq->elv.priv[1]; | |
72961c4e | 6052 | |
9f210738 | 6053 | bic = icq_to_bic(rq->elv.icq); |
aee69d78 | 6054 | |
8c9ff1ad CIK |
6055 | bfq_check_ioprio_change(bic, bio); |
6056 | ||
e21b7a0b AA |
6057 | bfq_bic_update_cgroup(bic, bio); |
6058 | ||
36eca894 AA |
6059 | bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync, |
6060 | &new_queue); | |
6061 | ||
6062 | if (likely(!new_queue)) { | |
6063 | /* If the queue was seeky for too long, break it apart. */ | |
6064 | if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { | |
6065 | bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); | |
e1b2324d AA |
6066 | |
6067 | /* Update bic before losing reference to bfqq */ | |
6068 | if (bfq_bfqq_in_large_burst(bfqq)) | |
6069 | bic->saved_in_large_burst = true; | |
6070 | ||
36eca894 | 6071 | bfqq = bfq_split_bfqq(bic, bfqq); |
6fa3e8d3 | 6072 | split = true; |
36eca894 AA |
6073 | |
6074 | if (!bfqq) | |
6075 | bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, | |
6076 | true, is_sync, | |
6077 | NULL); | |
13c931bd PV |
6078 | else |
6079 | bfqq_already_existing = true; | |
36eca894 | 6080 | } |
aee69d78 PV |
6081 | } |
6082 | ||
6083 | bfqq->allocated++; | |
6084 | bfqq->ref++; | |
6085 | bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d", | |
6086 | rq, bfqq, bfqq->ref); | |
6087 | ||
6088 | rq->elv.priv[0] = bic; | |
6089 | rq->elv.priv[1] = bfqq; | |
6090 | ||
36eca894 AA |
6091 | /* |
6092 | * If a bfq_queue has only one process reference, it is owned | |
6093 | * by only this bic: we can then set bfqq->bic = bic. in | |
6094 | * addition, if the queue has also just been split, we have to | |
6095 | * resume its state. | |
6096 | */ | |
6097 | if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) { | |
6098 | bfqq->bic = bic; | |
6fa3e8d3 | 6099 | if (split) { |
36eca894 AA |
6100 | /* |
6101 | * The queue has just been split from a shared | |
6102 | * queue: restore the idle window and the | |
6103 | * possible weight raising period. | |
6104 | */ | |
13c931bd PV |
6105 | bfq_bfqq_resume_state(bfqq, bfqd, bic, |
6106 | bfqq_already_existing); | |
36eca894 AA |
6107 | } |
6108 | } | |
6109 | ||
84a74689 PV |
6110 | /* |
6111 | * Consider bfqq as possibly belonging to a burst of newly | |
6112 | * created queues only if: | |
6113 | * 1) A burst is actually happening (bfqd->burst_size > 0) | |
6114 | * or | |
6115 | * 2) There is no other active queue. In fact, if, in | |
6116 | * contrast, there are active queues not belonging to the | |
6117 | * possible burst bfqq may belong to, then there is no gain | |
6118 | * in considering bfqq as belonging to a burst, and | |
6119 | * therefore in not weight-raising bfqq. See comments on | |
6120 | * bfq_handle_burst(). | |
6121 | * | |
6122 | * This filtering also helps eliminating false positives, | |
6123 | * occurring when bfqq does not belong to an actual large | |
6124 | * burst, but some background task (e.g., a service) happens | |
6125 | * to trigger the creation of new queues very close to when | |
6126 | * bfqq and its possible companion queues are created. See | |
6127 | * comments on bfq_handle_burst() for further details also on | |
6128 | * this issue. | |
6129 | */ | |
6130 | if (unlikely(bfq_bfqq_just_created(bfqq) && | |
6131 | (bfqd->burst_size > 0 || | |
6132 | bfq_tot_busy_queues(bfqd) == 0))) | |
e1b2324d AA |
6133 | bfq_handle_burst(bfqd, bfqq); |
6134 | ||
18e5a57d | 6135 | return bfqq; |
aee69d78 PV |
6136 | } |
6137 | ||
6138 | static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq) | |
6139 | { | |
6140 | struct bfq_data *bfqd = bfqq->bfqd; | |
6141 | enum bfqq_expiration reason; | |
6142 | unsigned long flags; | |
6143 | ||
6144 | spin_lock_irqsave(&bfqd->lock, flags); | |
6145 | bfq_clear_bfqq_wait_request(bfqq); | |
6146 | ||
6147 | if (bfqq != bfqd->in_service_queue) { | |
6148 | spin_unlock_irqrestore(&bfqd->lock, flags); | |
6149 | return; | |
6150 | } | |
6151 | ||
6152 | if (bfq_bfqq_budget_timeout(bfqq)) | |
6153 | /* | |
6154 | * Also here the queue can be safely expired | |
6155 | * for budget timeout without wasting | |
6156 | * guarantees | |
6157 | */ | |
6158 | reason = BFQQE_BUDGET_TIMEOUT; | |
6159 | else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) | |
6160 | /* | |
6161 | * The queue may not be empty upon timer expiration, | |
6162 | * because we may not disable the timer when the | |
6163 | * first request of the in-service queue arrives | |
6164 | * during disk idling. | |
6165 | */ | |
6166 | reason = BFQQE_TOO_IDLE; | |
6167 | else | |
6168 | goto schedule_dispatch; | |
6169 | ||
6170 | bfq_bfqq_expire(bfqd, bfqq, true, reason); | |
6171 | ||
6172 | schedule_dispatch: | |
6fa3e8d3 | 6173 | spin_unlock_irqrestore(&bfqd->lock, flags); |
aee69d78 PV |
6174 | bfq_schedule_dispatch(bfqd); |
6175 | } | |
6176 | ||
6177 | /* | |
6178 | * Handler of the expiration of the timer running if the in-service queue | |
6179 | * is idling inside its time slice. | |
6180 | */ | |
6181 | static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer) | |
6182 | { | |
6183 | struct bfq_data *bfqd = container_of(timer, struct bfq_data, | |
6184 | idle_slice_timer); | |
6185 | struct bfq_queue *bfqq = bfqd->in_service_queue; | |
6186 | ||
6187 | /* | |
6188 | * Theoretical race here: the in-service queue can be NULL or | |
6189 | * different from the queue that was idling if a new request | |
6190 | * arrives for the current queue and there is a full dispatch | |
6191 | * cycle that changes the in-service queue. This can hardly | |
6192 | * happen, but in the worst case we just expire a queue too | |
6193 | * early. | |
6194 | */ | |
6195 | if (bfqq) | |
6196 | bfq_idle_slice_timer_body(bfqq); | |
6197 | ||
6198 | return HRTIMER_NORESTART; | |
6199 | } | |
6200 | ||
6201 | static void __bfq_put_async_bfqq(struct bfq_data *bfqd, | |
6202 | struct bfq_queue **bfqq_ptr) | |
6203 | { | |
6204 | struct bfq_queue *bfqq = *bfqq_ptr; | |
6205 | ||
6206 | bfq_log(bfqd, "put_async_bfqq: %p", bfqq); | |
6207 | if (bfqq) { | |
e21b7a0b AA |
6208 | bfq_bfqq_move(bfqd, bfqq, bfqd->root_group); |
6209 | ||
aee69d78 PV |
6210 | bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
6211 | bfqq, bfqq->ref); | |
6212 | bfq_put_queue(bfqq); | |
6213 | *bfqq_ptr = NULL; | |
6214 | } | |
6215 | } | |
6216 | ||
6217 | /* | |
e21b7a0b AA |
6218 | * Release all the bfqg references to its async queues. If we are |
6219 | * deallocating the group these queues may still contain requests, so | |
6220 | * we reparent them to the root cgroup (i.e., the only one that will | |
6221 | * exist for sure until all the requests on a device are gone). | |
aee69d78 | 6222 | */ |
ea25da48 | 6223 | void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
aee69d78 PV |
6224 | { |
6225 | int i, j; | |
6226 | ||
6227 | for (i = 0; i < 2; i++) | |
6228 | for (j = 0; j < IOPRIO_BE_NR; j++) | |
e21b7a0b | 6229 | __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
aee69d78 | 6230 | |
e21b7a0b | 6231 | __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
aee69d78 PV |
6232 | } |
6233 | ||
f0635b8a JA |
6234 | /* |
6235 | * See the comments on bfq_limit_depth for the purpose of | |
483b7bf2 | 6236 | * the depths set in the function. Return minimum shallow depth we'll use. |
f0635b8a | 6237 | */ |
483b7bf2 JA |
6238 | static unsigned int bfq_update_depths(struct bfq_data *bfqd, |
6239 | struct sbitmap_queue *bt) | |
f0635b8a | 6240 | { |
483b7bf2 JA |
6241 | unsigned int i, j, min_shallow = UINT_MAX; |
6242 | ||
f0635b8a JA |
6243 | /* |
6244 | * In-word depths if no bfq_queue is being weight-raised: | |
6245 | * leaving 25% of tags only for sync reads. | |
6246 | * | |
6247 | * In next formulas, right-shift the value | |
bd7d4ef6 JA |
6248 | * (1U<<bt->sb.shift), instead of computing directly |
6249 | * (1U<<(bt->sb.shift - something)), to be robust against | |
6250 | * any possible value of bt->sb.shift, without having to | |
f0635b8a JA |
6251 | * limit 'something'. |
6252 | */ | |
6253 | /* no more than 50% of tags for async I/O */ | |
bd7d4ef6 | 6254 | bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U); |
f0635b8a JA |
6255 | /* |
6256 | * no more than 75% of tags for sync writes (25% extra tags | |
6257 | * w.r.t. async I/O, to prevent async I/O from starving sync | |
6258 | * writes) | |
6259 | */ | |
bd7d4ef6 | 6260 | bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U); |
f0635b8a JA |
6261 | |
6262 | /* | |
6263 | * In-word depths in case some bfq_queue is being weight- | |
6264 | * raised: leaving ~63% of tags for sync reads. This is the | |
6265 | * highest percentage for which, in our tests, application | |
6266 | * start-up times didn't suffer from any regression due to tag | |
6267 | * shortage. | |
6268 | */ | |
6269 | /* no more than ~18% of tags for async I/O */ | |
bd7d4ef6 | 6270 | bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U); |
f0635b8a | 6271 | /* no more than ~37% of tags for sync writes (~20% extra tags) */ |
bd7d4ef6 | 6272 | bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U); |
483b7bf2 JA |
6273 | |
6274 | for (i = 0; i < 2; i++) | |
6275 | for (j = 0; j < 2; j++) | |
6276 | min_shallow = min(min_shallow, bfqd->word_depths[i][j]); | |
6277 | ||
6278 | return min_shallow; | |
f0635b8a JA |
6279 | } |
6280 | ||
77f1e0a5 | 6281 | static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx) |
f0635b8a JA |
6282 | { |
6283 | struct bfq_data *bfqd = hctx->queue->elevator->elevator_data; | |
6284 | struct blk_mq_tags *tags = hctx->sched_tags; | |
483b7bf2 | 6285 | unsigned int min_shallow; |
f0635b8a | 6286 | |
483b7bf2 JA |
6287 | min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags); |
6288 | sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow); | |
77f1e0a5 JA |
6289 | } |
6290 | ||
6291 | static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index) | |
6292 | { | |
6293 | bfq_depth_updated(hctx); | |
f0635b8a JA |
6294 | return 0; |
6295 | } | |
6296 | ||
aee69d78 PV |
6297 | static void bfq_exit_queue(struct elevator_queue *e) |
6298 | { | |
6299 | struct bfq_data *bfqd = e->elevator_data; | |
6300 | struct bfq_queue *bfqq, *n; | |
6301 | ||
6302 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
6303 | ||
6304 | spin_lock_irq(&bfqd->lock); | |
6305 | list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) | |
e21b7a0b | 6306 | bfq_deactivate_bfqq(bfqd, bfqq, false, false); |
aee69d78 PV |
6307 | spin_unlock_irq(&bfqd->lock); |
6308 | ||
6309 | hrtimer_cancel(&bfqd->idle_slice_timer); | |
6310 | ||
8abef10b | 6311 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
0d52af59 PV |
6312 | /* release oom-queue reference to root group */ |
6313 | bfqg_and_blkg_put(bfqd->root_group); | |
6314 | ||
e21b7a0b AA |
6315 | blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq); |
6316 | #else | |
6317 | spin_lock_irq(&bfqd->lock); | |
6318 | bfq_put_async_queues(bfqd, bfqd->root_group); | |
6319 | kfree(bfqd->root_group); | |
6320 | spin_unlock_irq(&bfqd->lock); | |
6321 | #endif | |
6322 | ||
aee69d78 PV |
6323 | kfree(bfqd); |
6324 | } | |
6325 | ||
e21b7a0b AA |
6326 | static void bfq_init_root_group(struct bfq_group *root_group, |
6327 | struct bfq_data *bfqd) | |
6328 | { | |
6329 | int i; | |
6330 | ||
6331 | #ifdef CONFIG_BFQ_GROUP_IOSCHED | |
6332 | root_group->entity.parent = NULL; | |
6333 | root_group->my_entity = NULL; | |
6334 | root_group->bfqd = bfqd; | |
6335 | #endif | |
36eca894 | 6336 | root_group->rq_pos_tree = RB_ROOT; |
e21b7a0b AA |
6337 | for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
6338 | root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; | |
6339 | root_group->sched_data.bfq_class_idle_last_service = jiffies; | |
6340 | } | |
6341 | ||
aee69d78 PV |
6342 | static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
6343 | { | |
6344 | struct bfq_data *bfqd; | |
6345 | struct elevator_queue *eq; | |
aee69d78 PV |
6346 | |
6347 | eq = elevator_alloc(q, e); | |
6348 | if (!eq) | |
6349 | return -ENOMEM; | |
6350 | ||
6351 | bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); | |
6352 | if (!bfqd) { | |
6353 | kobject_put(&eq->kobj); | |
6354 | return -ENOMEM; | |
6355 | } | |
6356 | eq->elevator_data = bfqd; | |
6357 | ||
0d945c1f | 6358 | spin_lock_irq(&q->queue_lock); |
e21b7a0b | 6359 | q->elevator = eq; |
0d945c1f | 6360 | spin_unlock_irq(&q->queue_lock); |
e21b7a0b | 6361 | |
aee69d78 PV |
6362 | /* |
6363 | * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. | |
6364 | * Grab a permanent reference to it, so that the normal code flow | |
6365 | * will not attempt to free it. | |
6366 | */ | |
6367 | bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0); | |
6368 | bfqd->oom_bfqq.ref++; | |
6369 | bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO; | |
6370 | bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE; | |
6371 | bfqd->oom_bfqq.entity.new_weight = | |
6372 | bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio); | |
e1b2324d AA |
6373 | |
6374 | /* oom_bfqq does not participate to bursts */ | |
6375 | bfq_clear_bfqq_just_created(&bfqd->oom_bfqq); | |
6376 | ||
aee69d78 PV |
6377 | /* |
6378 | * Trigger weight initialization, according to ioprio, at the | |
6379 | * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio | |
6380 | * class won't be changed any more. | |
6381 | */ | |
6382 | bfqd->oom_bfqq.entity.prio_changed = 1; | |
6383 | ||
6384 | bfqd->queue = q; | |
6385 | ||
e21b7a0b | 6386 | INIT_LIST_HEAD(&bfqd->dispatch); |
aee69d78 PV |
6387 | |
6388 | hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC, | |
6389 | HRTIMER_MODE_REL); | |
6390 | bfqd->idle_slice_timer.function = bfq_idle_slice_timer; | |
6391 | ||
fb53ac6c | 6392 | bfqd->queue_weights_tree = RB_ROOT_CACHED; |
ba7aeae5 | 6393 | bfqd->num_groups_with_pending_reqs = 0; |
1de0c4cd | 6394 | |
aee69d78 PV |
6395 | INIT_LIST_HEAD(&bfqd->active_list); |
6396 | INIT_LIST_HEAD(&bfqd->idle_list); | |
e1b2324d | 6397 | INIT_HLIST_HEAD(&bfqd->burst_list); |
aee69d78 PV |
6398 | |
6399 | bfqd->hw_tag = -1; | |
8cacc5ab | 6400 | bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue); |
aee69d78 PV |
6401 | |
6402 | bfqd->bfq_max_budget = bfq_default_max_budget; | |
6403 | ||
6404 | bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; | |
6405 | bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; | |
6406 | bfqd->bfq_back_max = bfq_back_max; | |
6407 | bfqd->bfq_back_penalty = bfq_back_penalty; | |
6408 | bfqd->bfq_slice_idle = bfq_slice_idle; | |
aee69d78 PV |
6409 | bfqd->bfq_timeout = bfq_timeout; |
6410 | ||
6411 | bfqd->bfq_requests_within_timer = 120; | |
6412 | ||
e1b2324d AA |
6413 | bfqd->bfq_large_burst_thresh = 8; |
6414 | bfqd->bfq_burst_interval = msecs_to_jiffies(180); | |
6415 | ||
44e44a1b PV |
6416 | bfqd->low_latency = true; |
6417 | ||
6418 | /* | |
6419 | * Trade-off between responsiveness and fairness. | |
6420 | */ | |
6421 | bfqd->bfq_wr_coeff = 30; | |
77b7dcea | 6422 | bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300); |
44e44a1b PV |
6423 | bfqd->bfq_wr_max_time = 0; |
6424 | bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000); | |
6425 | bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500); | |
77b7dcea PV |
6426 | bfqd->bfq_wr_max_softrt_rate = 7000; /* |
6427 | * Approximate rate required | |
6428 | * to playback or record a | |
6429 | * high-definition compressed | |
6430 | * video. | |
6431 | */ | |
cfd69712 | 6432 | bfqd->wr_busy_queues = 0; |
44e44a1b PV |
6433 | |
6434 | /* | |
e24f1c24 PV |
6435 | * Begin by assuming, optimistically, that the device peak |
6436 | * rate is equal to 2/3 of the highest reference rate. | |
44e44a1b | 6437 | */ |
e24f1c24 PV |
6438 | bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] * |
6439 | ref_wr_duration[blk_queue_nonrot(bfqd->queue)]; | |
6440 | bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3; | |
44e44a1b | 6441 | |
aee69d78 | 6442 | spin_lock_init(&bfqd->lock); |
aee69d78 | 6443 | |
e21b7a0b AA |
6444 | /* |
6445 | * The invocation of the next bfq_create_group_hierarchy | |
6446 | * function is the head of a chain of function calls | |
6447 | * (bfq_create_group_hierarchy->blkcg_activate_policy-> | |
6448 | * blk_mq_freeze_queue) that may lead to the invocation of the | |
6449 | * has_work hook function. For this reason, | |
6450 | * bfq_create_group_hierarchy is invoked only after all | |
6451 | * scheduler data has been initialized, apart from the fields | |
6452 | * that can be initialized only after invoking | |
6453 | * bfq_create_group_hierarchy. This, in particular, enables | |
6454 | * has_work to correctly return false. Of course, to avoid | |
6455 | * other inconsistencies, the blk-mq stack must then refrain | |
6456 | * from invoking further scheduler hooks before this init | |
6457 | * function is finished. | |
6458 | */ | |
6459 | bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node); | |
6460 | if (!bfqd->root_group) | |
6461 | goto out_free; | |
6462 | bfq_init_root_group(bfqd->root_group, bfqd); | |
6463 | bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group); | |
6464 | ||
b5dc5d4d | 6465 | wbt_disable_default(q); |
aee69d78 | 6466 | return 0; |
e21b7a0b AA |
6467 | |
6468 | out_free: | |
6469 | kfree(bfqd); | |
6470 | kobject_put(&eq->kobj); | |
6471 | return -ENOMEM; | |
aee69d78 PV |
6472 | } |
6473 | ||
6474 | static void bfq_slab_kill(void) | |
6475 | { | |
6476 | kmem_cache_destroy(bfq_pool); | |
6477 | } | |
6478 | ||
6479 | static int __init bfq_slab_setup(void) | |
6480 | { | |
6481 | bfq_pool = KMEM_CACHE(bfq_queue, 0); | |
6482 | if (!bfq_pool) | |
6483 | return -ENOMEM; | |
6484 | return 0; | |
6485 | } | |
6486 | ||
6487 | static ssize_t bfq_var_show(unsigned int var, char *page) | |
6488 | { | |
6489 | return sprintf(page, "%u\n", var); | |
6490 | } | |
6491 | ||
2f79136b | 6492 | static int bfq_var_store(unsigned long *var, const char *page) |
aee69d78 PV |
6493 | { |
6494 | unsigned long new_val; | |
6495 | int ret = kstrtoul(page, 10, &new_val); | |
6496 | ||
2f79136b BVA |
6497 | if (ret) |
6498 | return ret; | |
6499 | *var = new_val; | |
6500 | return 0; | |
aee69d78 PV |
6501 | } |
6502 | ||
6503 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
6504 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
6505 | { \ | |
6506 | struct bfq_data *bfqd = e->elevator_data; \ | |
6507 | u64 __data = __VAR; \ | |
6508 | if (__CONV == 1) \ | |
6509 | __data = jiffies_to_msecs(__data); \ | |
6510 | else if (__CONV == 2) \ | |
6511 | __data = div_u64(__data, NSEC_PER_MSEC); \ | |
6512 | return bfq_var_show(__data, (page)); \ | |
6513 | } | |
6514 | SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2); | |
6515 | SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2); | |
6516 | SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); | |
6517 | SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); | |
6518 | SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2); | |
6519 | SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); | |
6520 | SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1); | |
6521 | SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0); | |
44e44a1b | 6522 | SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
aee69d78 PV |
6523 | #undef SHOW_FUNCTION |
6524 | ||
6525 | #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \ | |
6526 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
6527 | { \ | |
6528 | struct bfq_data *bfqd = e->elevator_data; \ | |
6529 | u64 __data = __VAR; \ | |
6530 | __data = div_u64(__data, NSEC_PER_USEC); \ | |
6531 | return bfq_var_show(__data, (page)); \ | |
6532 | } | |
6533 | USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle); | |
6534 | #undef USEC_SHOW_FUNCTION | |
6535 | ||
6536 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
6537 | static ssize_t \ | |
6538 | __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
6539 | { \ | |
6540 | struct bfq_data *bfqd = e->elevator_data; \ | |
1530486c | 6541 | unsigned long __data, __min = (MIN), __max = (MAX); \ |
2f79136b BVA |
6542 | int ret; \ |
6543 | \ | |
6544 | ret = bfq_var_store(&__data, (page)); \ | |
6545 | if (ret) \ | |
6546 | return ret; \ | |
1530486c BVA |
6547 | if (__data < __min) \ |
6548 | __data = __min; \ | |
6549 | else if (__data > __max) \ | |
6550 | __data = __max; \ | |
aee69d78 PV |
6551 | if (__CONV == 1) \ |
6552 | *(__PTR) = msecs_to_jiffies(__data); \ | |
6553 | else if (__CONV == 2) \ | |
6554 | *(__PTR) = (u64)__data * NSEC_PER_MSEC; \ | |
6555 | else \ | |
6556 | *(__PTR) = __data; \ | |
235f8da1 | 6557 | return count; \ |
aee69d78 PV |
6558 | } |
6559 | STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, | |
6560 | INT_MAX, 2); | |
6561 | STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, | |
6562 | INT_MAX, 2); | |
6563 | STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); | |
6564 | STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, | |
6565 | INT_MAX, 0); | |
6566 | STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2); | |
6567 | #undef STORE_FUNCTION | |
6568 | ||
6569 | #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ | |
6570 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\ | |
6571 | { \ | |
6572 | struct bfq_data *bfqd = e->elevator_data; \ | |
1530486c | 6573 | unsigned long __data, __min = (MIN), __max = (MAX); \ |
2f79136b BVA |
6574 | int ret; \ |
6575 | \ | |
6576 | ret = bfq_var_store(&__data, (page)); \ | |
6577 | if (ret) \ | |
6578 | return ret; \ | |
1530486c BVA |
6579 | if (__data < __min) \ |
6580 | __data = __min; \ | |
6581 | else if (__data > __max) \ | |
6582 | __data = __max; \ | |
aee69d78 | 6583 | *(__PTR) = (u64)__data * NSEC_PER_USEC; \ |
235f8da1 | 6584 | return count; \ |
aee69d78 PV |
6585 | } |
6586 | USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0, | |
6587 | UINT_MAX); | |
6588 | #undef USEC_STORE_FUNCTION | |
6589 | ||
aee69d78 PV |
6590 | static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
6591 | const char *page, size_t count) | |
6592 | { | |
6593 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6594 | unsigned long __data; |
6595 | int ret; | |
235f8da1 | 6596 | |
2f79136b BVA |
6597 | ret = bfq_var_store(&__data, (page)); |
6598 | if (ret) | |
6599 | return ret; | |
aee69d78 PV |
6600 | |
6601 | if (__data == 0) | |
ab0e43e9 | 6602 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 PV |
6603 | else { |
6604 | if (__data > INT_MAX) | |
6605 | __data = INT_MAX; | |
6606 | bfqd->bfq_max_budget = __data; | |
6607 | } | |
6608 | ||
6609 | bfqd->bfq_user_max_budget = __data; | |
6610 | ||
235f8da1 | 6611 | return count; |
aee69d78 PV |
6612 | } |
6613 | ||
6614 | /* | |
6615 | * Leaving this name to preserve name compatibility with cfq | |
6616 | * parameters, but this timeout is used for both sync and async. | |
6617 | */ | |
6618 | static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, | |
6619 | const char *page, size_t count) | |
6620 | { | |
6621 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6622 | unsigned long __data; |
6623 | int ret; | |
235f8da1 | 6624 | |
2f79136b BVA |
6625 | ret = bfq_var_store(&__data, (page)); |
6626 | if (ret) | |
6627 | return ret; | |
aee69d78 PV |
6628 | |
6629 | if (__data < 1) | |
6630 | __data = 1; | |
6631 | else if (__data > INT_MAX) | |
6632 | __data = INT_MAX; | |
6633 | ||
6634 | bfqd->bfq_timeout = msecs_to_jiffies(__data); | |
6635 | if (bfqd->bfq_user_max_budget == 0) | |
ab0e43e9 | 6636 | bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd); |
aee69d78 | 6637 | |
235f8da1 | 6638 | return count; |
aee69d78 PV |
6639 | } |
6640 | ||
6641 | static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e, | |
6642 | const char *page, size_t count) | |
6643 | { | |
6644 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6645 | unsigned long __data; |
6646 | int ret; | |
235f8da1 | 6647 | |
2f79136b BVA |
6648 | ret = bfq_var_store(&__data, (page)); |
6649 | if (ret) | |
6650 | return ret; | |
aee69d78 PV |
6651 | |
6652 | if (__data > 1) | |
6653 | __data = 1; | |
6654 | if (!bfqd->strict_guarantees && __data == 1 | |
6655 | && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC) | |
6656 | bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC; | |
6657 | ||
6658 | bfqd->strict_guarantees = __data; | |
6659 | ||
235f8da1 | 6660 | return count; |
aee69d78 PV |
6661 | } |
6662 | ||
44e44a1b PV |
6663 | static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
6664 | const char *page, size_t count) | |
6665 | { | |
6666 | struct bfq_data *bfqd = e->elevator_data; | |
2f79136b BVA |
6667 | unsigned long __data; |
6668 | int ret; | |
235f8da1 | 6669 | |
2f79136b BVA |
6670 | ret = bfq_var_store(&__data, (page)); |
6671 | if (ret) | |
6672 | return ret; | |
44e44a1b PV |
6673 | |
6674 | if (__data > 1) | |
6675 | __data = 1; | |
6676 | if (__data == 0 && bfqd->low_latency != 0) | |
6677 | bfq_end_wr(bfqd); | |
6678 | bfqd->low_latency = __data; | |
6679 | ||
235f8da1 | 6680 | return count; |
44e44a1b PV |
6681 | } |
6682 | ||
aee69d78 PV |
6683 | #define BFQ_ATTR(name) \ |
6684 | __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store) | |
6685 | ||
6686 | static struct elv_fs_entry bfq_attrs[] = { | |
6687 | BFQ_ATTR(fifo_expire_sync), | |
6688 | BFQ_ATTR(fifo_expire_async), | |
6689 | BFQ_ATTR(back_seek_max), | |
6690 | BFQ_ATTR(back_seek_penalty), | |
6691 | BFQ_ATTR(slice_idle), | |
6692 | BFQ_ATTR(slice_idle_us), | |
6693 | BFQ_ATTR(max_budget), | |
6694 | BFQ_ATTR(timeout_sync), | |
6695 | BFQ_ATTR(strict_guarantees), | |
44e44a1b | 6696 | BFQ_ATTR(low_latency), |
aee69d78 PV |
6697 | __ATTR_NULL |
6698 | }; | |
6699 | ||
6700 | static struct elevator_type iosched_bfq_mq = { | |
f9cd4bfe | 6701 | .ops = { |
a52a69ea | 6702 | .limit_depth = bfq_limit_depth, |
5bbf4e5a | 6703 | .prepare_request = bfq_prepare_request, |
a7877390 PV |
6704 | .requeue_request = bfq_finish_requeue_request, |
6705 | .finish_request = bfq_finish_requeue_request, | |
aee69d78 PV |
6706 | .exit_icq = bfq_exit_icq, |
6707 | .insert_requests = bfq_insert_requests, | |
6708 | .dispatch_request = bfq_dispatch_request, | |
6709 | .next_request = elv_rb_latter_request, | |
6710 | .former_request = elv_rb_former_request, | |
6711 | .allow_merge = bfq_allow_bio_merge, | |
6712 | .bio_merge = bfq_bio_merge, | |
6713 | .request_merge = bfq_request_merge, | |
6714 | .requests_merged = bfq_requests_merged, | |
6715 | .request_merged = bfq_request_merged, | |
6716 | .has_work = bfq_has_work, | |
77f1e0a5 | 6717 | .depth_updated = bfq_depth_updated, |
f0635b8a | 6718 | .init_hctx = bfq_init_hctx, |
aee69d78 PV |
6719 | .init_sched = bfq_init_queue, |
6720 | .exit_sched = bfq_exit_queue, | |
6721 | }, | |
6722 | ||
aee69d78 PV |
6723 | .icq_size = sizeof(struct bfq_io_cq), |
6724 | .icq_align = __alignof__(struct bfq_io_cq), | |
6725 | .elevator_attrs = bfq_attrs, | |
6726 | .elevator_name = "bfq", | |
6727 | .elevator_owner = THIS_MODULE, | |
6728 | }; | |
26b4cf24 | 6729 | MODULE_ALIAS("bfq-iosched"); |
aee69d78 PV |
6730 | |
6731 | static int __init bfq_init(void) | |
6732 | { | |
6733 | int ret; | |
6734 | ||
e21b7a0b AA |
6735 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6736 | ret = blkcg_policy_register(&blkcg_policy_bfq); | |
6737 | if (ret) | |
6738 | return ret; | |
6739 | #endif | |
6740 | ||
aee69d78 PV |
6741 | ret = -ENOMEM; |
6742 | if (bfq_slab_setup()) | |
6743 | goto err_pol_unreg; | |
6744 | ||
44e44a1b PV |
6745 | /* |
6746 | * Times to load large popular applications for the typical | |
6747 | * systems installed on the reference devices (see the | |
e24f1c24 PV |
6748 | * comments before the definition of the next |
6749 | * array). Actually, we use slightly lower values, as the | |
44e44a1b PV |
6750 | * estimated peak rate tends to be smaller than the actual |
6751 | * peak rate. The reason for this last fact is that estimates | |
6752 | * are computed over much shorter time intervals than the long | |
6753 | * intervals typically used for benchmarking. Why? First, to | |
6754 | * adapt more quickly to variations. Second, because an I/O | |
6755 | * scheduler cannot rely on a peak-rate-evaluation workload to | |
6756 | * be run for a long time. | |
6757 | */ | |
e24f1c24 PV |
6758 | ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */ |
6759 | ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */ | |
44e44a1b | 6760 | |
aee69d78 PV |
6761 | ret = elv_register(&iosched_bfq_mq); |
6762 | if (ret) | |
37dcd657 | 6763 | goto slab_kill; |
aee69d78 PV |
6764 | |
6765 | return 0; | |
6766 | ||
37dcd657 | 6767 | slab_kill: |
6768 | bfq_slab_kill(); | |
aee69d78 | 6769 | err_pol_unreg: |
e21b7a0b AA |
6770 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6771 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
6772 | #endif | |
aee69d78 PV |
6773 | return ret; |
6774 | } | |
6775 | ||
6776 | static void __exit bfq_exit(void) | |
6777 | { | |
6778 | elv_unregister(&iosched_bfq_mq); | |
e21b7a0b AA |
6779 | #ifdef CONFIG_BFQ_GROUP_IOSCHED |
6780 | blkcg_policy_unregister(&blkcg_policy_bfq); | |
6781 | #endif | |
aee69d78 PV |
6782 | bfq_slab_kill(); |
6783 | } | |
6784 | ||
6785 | module_init(bfq_init); | |
6786 | module_exit(bfq_exit); | |
6787 | ||
6788 | MODULE_AUTHOR("Paolo Valente"); | |
6789 | MODULE_LICENSE("GPL"); | |
6790 | MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler"); |