Commit | Line | Data |
---|---|---|
7caa4715 TH |
1 | /* SPDX-License-Identifier: GPL-2.0 |
2 | * | |
3 | * IO cost model based controller. | |
4 | * | |
5 | * Copyright (C) 2019 Tejun Heo <tj@kernel.org> | |
6 | * Copyright (C) 2019 Andy Newell <newella@fb.com> | |
7 | * Copyright (C) 2019 Facebook | |
8 | * | |
9 | * One challenge of controlling IO resources is the lack of trivially | |
10 | * observable cost metric. This is distinguished from CPU and memory where | |
11 | * wallclock time and the number of bytes can serve as accurate enough | |
12 | * approximations. | |
13 | * | |
14 | * Bandwidth and iops are the most commonly used metrics for IO devices but | |
15 | * depending on the type and specifics of the device, different IO patterns | |
16 | * easily lead to multiple orders of magnitude variations rendering them | |
17 | * useless for the purpose of IO capacity distribution. While on-device | |
18 | * time, with a lot of clutches, could serve as a useful approximation for | |
19 | * non-queued rotational devices, this is no longer viable with modern | |
20 | * devices, even the rotational ones. | |
21 | * | |
22 | * While there is no cost metric we can trivially observe, it isn't a | |
23 | * complete mystery. For example, on a rotational device, seek cost | |
24 | * dominates while a contiguous transfer contributes a smaller amount | |
25 | * proportional to the size. If we can characterize at least the relative | |
26 | * costs of these different types of IOs, it should be possible to | |
27 | * implement a reasonable work-conserving proportional IO resource | |
28 | * distribution. | |
29 | * | |
30 | * 1. IO Cost Model | |
31 | * | |
32 | * IO cost model estimates the cost of an IO given its basic parameters and | |
33 | * history (e.g. the end sector of the last IO). The cost is measured in | |
34 | * device time. If a given IO is estimated to cost 10ms, the device should | |
35 | * be able to process ~100 of those IOs in a second. | |
36 | * | |
37 | * Currently, there's only one builtin cost model - linear. Each IO is | |
38 | * classified as sequential or random and given a base cost accordingly. | |
39 | * On top of that, a size cost proportional to the length of the IO is | |
40 | * added. While simple, this model captures the operational | |
41 | * characteristics of a wide varienty of devices well enough. Default | |
5ba1add2 | 42 | * parameters for several different classes of devices are provided and the |
7caa4715 TH |
43 | * parameters can be configured from userspace via |
44 | * /sys/fs/cgroup/io.cost.model. | |
45 | * | |
46 | * If needed, tools/cgroup/iocost_coef_gen.py can be used to generate | |
47 | * device-specific coefficients. | |
48 | * | |
49 | * 2. Control Strategy | |
50 | * | |
51 | * The device virtual time (vtime) is used as the primary control metric. | |
52 | * The control strategy is composed of the following three parts. | |
53 | * | |
54 | * 2-1. Vtime Distribution | |
55 | * | |
56 | * When a cgroup becomes active in terms of IOs, its hierarchical share is | |
57 | * calculated. Please consider the following hierarchy where the numbers | |
58 | * inside parentheses denote the configured weights. | |
59 | * | |
60 | * root | |
61 | * / \ | |
62 | * A (w:100) B (w:300) | |
63 | * / \ | |
64 | * A0 (w:100) A1 (w:100) | |
65 | * | |
66 | * If B is idle and only A0 and A1 are actively issuing IOs, as the two are | |
67 | * of equal weight, each gets 50% share. If then B starts issuing IOs, B | |
68 | * gets 300/(100+300) or 75% share, and A0 and A1 equally splits the rest, | |
69 | * 12.5% each. The distribution mechanism only cares about these flattened | |
70 | * shares. They're called hweights (hierarchical weights) and always add | |
fe20cdb5 | 71 | * upto 1 (WEIGHT_ONE). |
7caa4715 TH |
72 | * |
73 | * A given cgroup's vtime runs slower in inverse proportion to its hweight. | |
74 | * For example, with 12.5% weight, A0's time runs 8 times slower (100/12.5) | |
75 | * against the device vtime - an IO which takes 10ms on the underlying | |
76 | * device is considered to take 80ms on A0. | |
77 | * | |
78 | * This constitutes the basis of IO capacity distribution. Each cgroup's | |
79 | * vtime is running at a rate determined by its hweight. A cgroup tracks | |
5ba1add2 | 80 | * the vtime consumed by past IOs and can issue a new IO if doing so |
7caa4715 TH |
81 | * wouldn't outrun the current device vtime. Otherwise, the IO is |
82 | * suspended until the vtime has progressed enough to cover it. | |
83 | * | |
84 | * 2-2. Vrate Adjustment | |
85 | * | |
86 | * It's unrealistic to expect the cost model to be perfect. There are too | |
87 | * many devices and even on the same device the overall performance | |
88 | * fluctuates depending on numerous factors such as IO mixture and device | |
89 | * internal garbage collection. The controller needs to adapt dynamically. | |
90 | * | |
91 | * This is achieved by adjusting the overall IO rate according to how busy | |
92 | * the device is. If the device becomes overloaded, we're sending down too | |
93 | * many IOs and should generally slow down. If there are waiting issuers | |
94 | * but the device isn't saturated, we're issuing too few and should | |
95 | * generally speed up. | |
96 | * | |
97 | * To slow down, we lower the vrate - the rate at which the device vtime | |
98 | * passes compared to the wall clock. For example, if the vtime is running | |
99 | * at the vrate of 75%, all cgroups added up would only be able to issue | |
100 | * 750ms worth of IOs per second, and vice-versa for speeding up. | |
101 | * | |
102 | * Device business is determined using two criteria - rq wait and | |
103 | * completion latencies. | |
104 | * | |
105 | * When a device gets saturated, the on-device and then the request queues | |
106 | * fill up and a bio which is ready to be issued has to wait for a request | |
107 | * to become available. When this delay becomes noticeable, it's a clear | |
108 | * indication that the device is saturated and we lower the vrate. This | |
109 | * saturation signal is fairly conservative as it only triggers when both | |
110 | * hardware and software queues are filled up, and is used as the default | |
111 | * busy signal. | |
112 | * | |
113 | * As devices can have deep queues and be unfair in how the queued commands | |
114 | * are executed, soley depending on rq wait may not result in satisfactory | |
115 | * control quality. For a better control quality, completion latency QoS | |
116 | * parameters can be configured so that the device is considered saturated | |
117 | * if N'th percentile completion latency rises above the set point. | |
118 | * | |
119 | * The completion latency requirements are a function of both the | |
120 | * underlying device characteristics and the desired IO latency quality of | |
121 | * service. There is an inherent trade-off - the tighter the latency QoS, | |
122 | * the higher the bandwidth lossage. Latency QoS is disabled by default | |
123 | * and can be set through /sys/fs/cgroup/io.cost.qos. | |
124 | * | |
125 | * 2-3. Work Conservation | |
126 | * | |
127 | * Imagine two cgroups A and B with equal weights. A is issuing a small IO | |
128 | * periodically while B is sending out enough parallel IOs to saturate the | |
129 | * device on its own. Let's say A's usage amounts to 100ms worth of IO | |
130 | * cost per second, i.e., 10% of the device capacity. The naive | |
131 | * distribution of half and half would lead to 60% utilization of the | |
132 | * device, a significant reduction in the total amount of work done | |
133 | * compared to free-for-all competition. This is too high a cost to pay | |
134 | * for IO control. | |
135 | * | |
136 | * To conserve the total amount of work done, we keep track of how much | |
137 | * each active cgroup is actually using and yield part of its weight if | |
138 | * there are other cgroups which can make use of it. In the above case, | |
139 | * A's weight will be lowered so that it hovers above the actual usage and | |
140 | * B would be able to use the rest. | |
141 | * | |
142 | * As we don't want to penalize a cgroup for donating its weight, the | |
143 | * surplus weight adjustment factors in a margin and has an immediate | |
144 | * snapback mechanism in case the cgroup needs more IO vtime for itself. | |
145 | * | |
146 | * Note that adjusting down surplus weights has the same effects as | |
147 | * accelerating vtime for other cgroups and work conservation can also be | |
148 | * implemented by adjusting vrate dynamically. However, squaring who can | |
149 | * donate and should take back how much requires hweight propagations | |
150 | * anyway making it easier to implement and understand as a separate | |
151 | * mechanism. | |
6954ff18 TH |
152 | * |
153 | * 3. Monitoring | |
154 | * | |
155 | * Instead of debugfs or other clumsy monitoring mechanisms, this | |
156 | * controller uses a drgn based monitoring script - | |
157 | * tools/cgroup/iocost_monitor.py. For details on drgn, please see | |
5ba1add2 | 158 | * https://github.com/osandov/drgn. The output looks like the following. |
6954ff18 TH |
159 | * |
160 | * sdb RUN per=300ms cur_per=234.218:v203.695 busy= +1 vrate= 62.12% | |
7c1ee704 TH |
161 | * active weight hweight% inflt% dbt delay usages% |
162 | * test/a * 50/ 50 33.33/ 33.33 27.65 2 0*041 033:033:033 | |
163 | * test/b * 100/ 100 66.67/ 66.67 17.56 0 0*000 066:079:077 | |
6954ff18 TH |
164 | * |
165 | * - per : Timer period | |
166 | * - cur_per : Internal wall and device vtime clock | |
167 | * - vrate : Device virtual time rate against wall clock | |
168 | * - weight : Surplus-adjusted and configured weights | |
169 | * - hweight : Surplus-adjusted and configured hierarchical weights | |
170 | * - inflt : The percentage of in-flight IO cost at the end of last period | |
171 | * - del_ms : Deferred issuer delay induction level and duration | |
172 | * - usages : Usage history | |
7caa4715 TH |
173 | */ |
174 | ||
175 | #include <linux/kernel.h> | |
176 | #include <linux/module.h> | |
177 | #include <linux/timer.h> | |
178 | #include <linux/time64.h> | |
179 | #include <linux/parser.h> | |
180 | #include <linux/sched/signal.h> | |
5e124f74 TH |
181 | #include <asm/local.h> |
182 | #include <asm/local64.h> | |
7caa4715 TH |
183 | #include "blk-rq-qos.h" |
184 | #include "blk-stat.h" | |
185 | #include "blk-wbt.h" | |
672fdcf0 | 186 | #include "blk-cgroup.h" |
7caa4715 TH |
187 | |
188 | #ifdef CONFIG_TRACEPOINTS | |
189 | ||
190 | /* copied from TRACE_CGROUP_PATH, see cgroup-internal.h */ | |
191 | #define TRACE_IOCG_PATH_LEN 1024 | |
192 | static DEFINE_SPINLOCK(trace_iocg_path_lock); | |
193 | static char trace_iocg_path[TRACE_IOCG_PATH_LEN]; | |
194 | ||
195 | #define TRACE_IOCG_PATH(type, iocg, ...) \ | |
196 | do { \ | |
197 | unsigned long flags; \ | |
198 | if (trace_iocost_##type##_enabled()) { \ | |
199 | spin_lock_irqsave(&trace_iocg_path_lock, flags); \ | |
200 | cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup, \ | |
201 | trace_iocg_path, TRACE_IOCG_PATH_LEN); \ | |
202 | trace_iocost_##type(iocg, trace_iocg_path, \ | |
203 | ##__VA_ARGS__); \ | |
204 | spin_unlock_irqrestore(&trace_iocg_path_lock, flags); \ | |
205 | } \ | |
206 | } while (0) | |
207 | ||
208 | #else /* CONFIG_TRACE_POINTS */ | |
209 | #define TRACE_IOCG_PATH(type, iocg, ...) do { } while (0) | |
210 | #endif /* CONFIG_TRACE_POINTS */ | |
211 | ||
212 | enum { | |
213 | MILLION = 1000000, | |
214 | ||
215 | /* timer period is calculated from latency requirements, bound it */ | |
216 | MIN_PERIOD = USEC_PER_MSEC, | |
217 | MAX_PERIOD = USEC_PER_SEC, | |
218 | ||
219 | /* | |
f1de2439 | 220 | * iocg->vtime is targeted at 50% behind the device vtime, which |
7caa4715 TH |
221 | * serves as its IO credit buffer. Surplus weight adjustment is |
222 | * immediately canceled if the vtime margin runs below 10%. | |
223 | */ | |
7ca5b2e6 | 224 | MARGIN_MIN_PCT = 10, |
f1de2439 TH |
225 | MARGIN_LOW_PCT = 20, |
226 | MARGIN_TARGET_PCT = 50, | |
7caa4715 | 227 | |
b0853ab4 TH |
228 | INUSE_ADJ_STEP_PCT = 25, |
229 | ||
7ca5b2e6 TH |
230 | /* Have some play in timer operations */ |
231 | TIMER_SLACK_PCT = 1, | |
7caa4715 | 232 | |
7caa4715 | 233 | /* 1/64k is granular enough and can easily be handled w/ u32 */ |
fe20cdb5 | 234 | WEIGHT_ONE = 1 << 16, |
7caa4715 TH |
235 | |
236 | /* | |
237 | * As vtime is used to calculate the cost of each IO, it needs to | |
238 | * be fairly high precision. For example, it should be able to | |
239 | * represent the cost of a single page worth of discard with | |
240 | * suffificient accuracy. At the same time, it should be able to | |
241 | * represent reasonably long enough durations to be useful and | |
242 | * convenient during operation. | |
243 | * | |
244 | * 1s worth of vtime is 2^37. This gives us both sub-nanosecond | |
245 | * granularity and days of wrap-around time even at extreme vrates. | |
246 | */ | |
247 | VTIME_PER_SEC_SHIFT = 37, | |
248 | VTIME_PER_SEC = 1LLU << VTIME_PER_SEC_SHIFT, | |
249 | VTIME_PER_USEC = VTIME_PER_SEC / USEC_PER_SEC, | |
cd006509 | 250 | VTIME_PER_NSEC = VTIME_PER_SEC / NSEC_PER_SEC, |
7caa4715 TH |
251 | |
252 | /* bound vrate adjustments within two orders of magnitude */ | |
253 | VRATE_MIN_PPM = 10000, /* 1% */ | |
254 | VRATE_MAX_PPM = 100000000, /* 10000% */ | |
255 | ||
256 | VRATE_MIN = VTIME_PER_USEC * VRATE_MIN_PPM / MILLION, | |
257 | VRATE_CLAMP_ADJ_PCT = 4, | |
258 | ||
259 | /* if IOs end up waiting for requests, issue less */ | |
260 | RQ_WAIT_BUSY_PCT = 5, | |
261 | ||
262 | /* unbusy hysterisis */ | |
263 | UNBUSY_THR_PCT = 75, | |
264 | ||
5160a5a5 TH |
265 | /* |
266 | * The effect of delay is indirect and non-linear and a huge amount of | |
267 | * future debt can accumulate abruptly while unthrottled. Linearly scale | |
268 | * up delay as debt is going up and then let it decay exponentially. | |
269 | * This gives us quick ramp ups while delay is accumulating and long | |
270 | * tails which can help reducing the frequency of debt explosions on | |
271 | * unthrottle. The parameters are experimentally determined. | |
272 | * | |
273 | * The delay mechanism provides adequate protection and behavior in many | |
274 | * cases. However, this is far from ideal and falls shorts on both | |
275 | * fronts. The debtors are often throttled too harshly costing a | |
276 | * significant level of fairness and possibly total work while the | |
277 | * protection against their impacts on the system can be choppy and | |
278 | * unreliable. | |
279 | * | |
280 | * The shortcoming primarily stems from the fact that, unlike for page | |
281 | * cache, the kernel doesn't have well-defined back-pressure propagation | |
282 | * mechanism and policies for anonymous memory. Fully addressing this | |
283 | * issue will likely require substantial improvements in the area. | |
284 | */ | |
285 | MIN_DELAY_THR_PCT = 500, | |
286 | MAX_DELAY_THR_PCT = 25000, | |
287 | MIN_DELAY = 250, | |
288 | MAX_DELAY = 250 * USEC_PER_MSEC, | |
289 | ||
c7af2a00 TH |
290 | /* halve debts if avg usage over 100ms is under 50% */ |
291 | DFGV_USAGE_PCT = 50, | |
292 | DFGV_PERIOD = 100 * USEC_PER_MSEC, | |
dda1315f | 293 | |
7caa4715 TH |
294 | /* don't let cmds which take a very long time pin lagging for too long */ |
295 | MAX_LAGGING_PERIODS = 10, | |
296 | ||
7caa4715 TH |
297 | /* switch iff the conditions are met for longer than this */ |
298 | AUTOP_CYCLE_NSEC = 10LLU * NSEC_PER_SEC, | |
299 | ||
300 | /* | |
301 | * Count IO size in 4k pages. The 12bit shift helps keeping | |
302 | * size-proportional components of cost calculation in closer | |
303 | * numbers of digits to per-IO cost components. | |
304 | */ | |
305 | IOC_PAGE_SHIFT = 12, | |
306 | IOC_PAGE_SIZE = 1 << IOC_PAGE_SHIFT, | |
307 | IOC_SECT_TO_PAGE_SHIFT = IOC_PAGE_SHIFT - SECTOR_SHIFT, | |
308 | ||
309 | /* if apart further than 16M, consider randio for linear model */ | |
310 | LCOEF_RANDIO_PAGES = 4096, | |
311 | }; | |
312 | ||
313 | enum ioc_running { | |
314 | IOC_IDLE, | |
315 | IOC_RUNNING, | |
316 | IOC_STOP, | |
317 | }; | |
318 | ||
319 | /* io.cost.qos controls including per-dev enable of the whole controller */ | |
320 | enum { | |
321 | QOS_ENABLE, | |
322 | QOS_CTRL, | |
323 | NR_QOS_CTRL_PARAMS, | |
324 | }; | |
325 | ||
326 | /* io.cost.qos params */ | |
327 | enum { | |
328 | QOS_RPPM, | |
329 | QOS_RLAT, | |
330 | QOS_WPPM, | |
331 | QOS_WLAT, | |
332 | QOS_MIN, | |
333 | QOS_MAX, | |
334 | NR_QOS_PARAMS, | |
335 | }; | |
336 | ||
337 | /* io.cost.model controls */ | |
338 | enum { | |
339 | COST_CTRL, | |
340 | COST_MODEL, | |
341 | NR_COST_CTRL_PARAMS, | |
342 | }; | |
343 | ||
344 | /* builtin linear cost model coefficients */ | |
345 | enum { | |
346 | I_LCOEF_RBPS, | |
347 | I_LCOEF_RSEQIOPS, | |
348 | I_LCOEF_RRANDIOPS, | |
349 | I_LCOEF_WBPS, | |
350 | I_LCOEF_WSEQIOPS, | |
351 | I_LCOEF_WRANDIOPS, | |
352 | NR_I_LCOEFS, | |
353 | }; | |
354 | ||
355 | enum { | |
356 | LCOEF_RPAGE, | |
357 | LCOEF_RSEQIO, | |
358 | LCOEF_RRANDIO, | |
359 | LCOEF_WPAGE, | |
360 | LCOEF_WSEQIO, | |
361 | LCOEF_WRANDIO, | |
362 | NR_LCOEFS, | |
363 | }; | |
364 | ||
365 | enum { | |
366 | AUTOP_INVALID, | |
367 | AUTOP_HDD, | |
368 | AUTOP_SSD_QD1, | |
369 | AUTOP_SSD_DFL, | |
370 | AUTOP_SSD_FAST, | |
371 | }; | |
372 | ||
7caa4715 TH |
373 | struct ioc_params { |
374 | u32 qos[NR_QOS_PARAMS]; | |
375 | u64 i_lcoefs[NR_I_LCOEFS]; | |
376 | u64 lcoefs[NR_LCOEFS]; | |
377 | u32 too_fast_vrate_pct; | |
378 | u32 too_slow_vrate_pct; | |
379 | }; | |
380 | ||
7ca5b2e6 TH |
381 | struct ioc_margins { |
382 | s64 min; | |
f1de2439 TH |
383 | s64 low; |
384 | s64 target; | |
7ca5b2e6 TH |
385 | }; |
386 | ||
7caa4715 | 387 | struct ioc_missed { |
5e124f74 TH |
388 | local_t nr_met; |
389 | local_t nr_missed; | |
7caa4715 TH |
390 | u32 last_met; |
391 | u32 last_missed; | |
392 | }; | |
393 | ||
394 | struct ioc_pcpu_stat { | |
395 | struct ioc_missed missed[2]; | |
396 | ||
5e124f74 | 397 | local64_t rq_wait_ns; |
7caa4715 TH |
398 | u64 last_rq_wait_ns; |
399 | }; | |
400 | ||
401 | /* per device */ | |
402 | struct ioc { | |
403 | struct rq_qos rqos; | |
404 | ||
405 | bool enabled; | |
406 | ||
407 | struct ioc_params params; | |
7ca5b2e6 | 408 | struct ioc_margins margins; |
7caa4715 | 409 | u32 period_us; |
7ca5b2e6 | 410 | u32 timer_slack_ns; |
7caa4715 TH |
411 | u64 vrate_min; |
412 | u64 vrate_max; | |
413 | ||
414 | spinlock_t lock; | |
415 | struct timer_list timer; | |
416 | struct list_head active_iocgs; /* active cgroups */ | |
417 | struct ioc_pcpu_stat __percpu *pcpu_stat; | |
418 | ||
419 | enum ioc_running running; | |
420 | atomic64_t vtime_rate; | |
ac33e91e TH |
421 | u64 vtime_base_rate; |
422 | s64 vtime_err; | |
7caa4715 | 423 | |
67b7b641 | 424 | seqcount_spinlock_t period_seqcount; |
ce95570a | 425 | u64 period_at; /* wallclock starttime */ |
7caa4715 TH |
426 | u64 period_at_vtime; /* vtime starttime */ |
427 | ||
428 | atomic64_t cur_period; /* inc'd each period */ | |
429 | int busy_level; /* saturation history */ | |
430 | ||
7caa4715 TH |
431 | bool weights_updated; |
432 | atomic_t hweight_gen; /* for lazy hweights */ | |
433 | ||
c7af2a00 TH |
434 | /* debt forgivness */ |
435 | u64 dfgv_period_at; | |
436 | u64 dfgv_period_rem; | |
437 | u64 dfgv_usage_us_sum; | |
dda1315f | 438 | |
7caa4715 TH |
439 | u64 autop_too_fast_at; |
440 | u64 autop_too_slow_at; | |
441 | int autop_idx; | |
442 | bool user_qos_params:1; | |
443 | bool user_cost_model:1; | |
444 | }; | |
445 | ||
97eb1975 TH |
446 | struct iocg_pcpu_stat { |
447 | local64_t abs_vusage; | |
448 | }; | |
449 | ||
450 | struct iocg_stat { | |
451 | u64 usage_us; | |
f0bf84a5 TH |
452 | u64 wait_us; |
453 | u64 indebt_us; | |
454 | u64 indelay_us; | |
97eb1975 TH |
455 | }; |
456 | ||
7caa4715 TH |
457 | /* per device-cgroup pair */ |
458 | struct ioc_gq { | |
459 | struct blkg_policy_data pd; | |
460 | struct ioc *ioc; | |
461 | ||
462 | /* | |
463 | * A iocg can get its weight from two sources - an explicit | |
464 | * per-device-cgroup configuration or the default weight of the | |
465 | * cgroup. `cfg_weight` is the explicit per-device-cgroup | |
466 | * configuration. `weight` is the effective considering both | |
467 | * sources. | |
468 | * | |
469 | * When an idle cgroup becomes active its `active` goes from 0 to | |
470 | * `weight`. `inuse` is the surplus adjusted active weight. | |
471 | * `active` and `inuse` are used to calculate `hweight_active` and | |
472 | * `hweight_inuse`. | |
473 | * | |
474 | * `last_inuse` remembers `inuse` while an iocg is idle to persist | |
475 | * surplus adjustments. | |
b0853ab4 TH |
476 | * |
477 | * `inuse` may be adjusted dynamically during period. `saved_*` are used | |
478 | * to determine and track adjustments. | |
7caa4715 TH |
479 | */ |
480 | u32 cfg_weight; | |
481 | u32 weight; | |
482 | u32 active; | |
483 | u32 inuse; | |
b0853ab4 | 484 | |
7caa4715 | 485 | u32 last_inuse; |
b0853ab4 | 486 | s64 saved_margin; |
7caa4715 TH |
487 | |
488 | sector_t cursor; /* to detect randio */ | |
489 | ||
490 | /* | |
491 | * `vtime` is this iocg's vtime cursor which progresses as IOs are | |
492 | * issued. If lagging behind device vtime, the delta represents | |
5ba1add2 | 493 | * the currently available IO budget. If running ahead, the |
7caa4715 TH |
494 | * overage. |
495 | * | |
496 | * `vtime_done` is the same but progressed on completion rather | |
497 | * than issue. The delta behind `vtime` represents the cost of | |
498 | * currently in-flight IOs. | |
7caa4715 TH |
499 | */ |
500 | atomic64_t vtime; | |
501 | atomic64_t done_vtime; | |
0b80f986 | 502 | u64 abs_vdebt; |
7caa4715 | 503 | |
5160a5a5 TH |
504 | /* current delay in effect and when it started */ |
505 | u64 delay; | |
506 | u64 delay_at; | |
507 | ||
7caa4715 TH |
508 | /* |
509 | * The period this iocg was last active in. Used for deactivation | |
510 | * and invalidating `vtime`. | |
511 | */ | |
512 | atomic64_t active_period; | |
513 | struct list_head active_list; | |
514 | ||
00410f1b | 515 | /* see __propagate_weights() and current_hweight() for details */ |
7caa4715 TH |
516 | u64 child_active_sum; |
517 | u64 child_inuse_sum; | |
e08d02aa | 518 | u64 child_adjusted_sum; |
7caa4715 TH |
519 | int hweight_gen; |
520 | u32 hweight_active; | |
521 | u32 hweight_inuse; | |
e08d02aa | 522 | u32 hweight_donating; |
93f7d2db | 523 | u32 hweight_after_donation; |
7caa4715 | 524 | |
97eb1975 | 525 | struct list_head walk_list; |
8692d2db | 526 | struct list_head surplus_list; |
97eb1975 | 527 | |
7caa4715 TH |
528 | struct wait_queue_head waitq; |
529 | struct hrtimer waitq_timer; | |
7caa4715 | 530 | |
1aa50d02 TH |
531 | /* timestamp at the latest activation */ |
532 | u64 activated_at; | |
533 | ||
97eb1975 TH |
534 | /* statistics */ |
535 | struct iocg_pcpu_stat __percpu *pcpu_stat; | |
2a371f7d | 536 | struct iocg_stat stat; |
97eb1975 TH |
537 | struct iocg_stat last_stat; |
538 | u64 last_stat_abs_vusage; | |
f1de2439 | 539 | u64 usage_delta_us; |
f0bf84a5 TH |
540 | u64 wait_since; |
541 | u64 indebt_since; | |
542 | u64 indelay_since; | |
7caa4715 TH |
543 | |
544 | /* this iocg's depth in the hierarchy and ancestors including self */ | |
545 | int level; | |
546 | struct ioc_gq *ancestors[]; | |
547 | }; | |
548 | ||
549 | /* per cgroup */ | |
550 | struct ioc_cgrp { | |
551 | struct blkcg_policy_data cpd; | |
552 | unsigned int dfl_weight; | |
553 | }; | |
554 | ||
555 | struct ioc_now { | |
556 | u64 now_ns; | |
ce95570a | 557 | u64 now; |
7caa4715 TH |
558 | u64 vnow; |
559 | u64 vrate; | |
560 | }; | |
561 | ||
562 | struct iocg_wait { | |
563 | struct wait_queue_entry wait; | |
564 | struct bio *bio; | |
565 | u64 abs_cost; | |
566 | bool committed; | |
567 | }; | |
568 | ||
569 | struct iocg_wake_ctx { | |
570 | struct ioc_gq *iocg; | |
571 | u32 hw_inuse; | |
572 | s64 vbudget; | |
573 | }; | |
574 | ||
575 | static const struct ioc_params autop[] = { | |
576 | [AUTOP_HDD] = { | |
577 | .qos = { | |
7afcccaf TH |
578 | [QOS_RLAT] = 250000, /* 250ms */ |
579 | [QOS_WLAT] = 250000, | |
7caa4715 TH |
580 | [QOS_MIN] = VRATE_MIN_PPM, |
581 | [QOS_MAX] = VRATE_MAX_PPM, | |
582 | }, | |
583 | .i_lcoefs = { | |
584 | [I_LCOEF_RBPS] = 174019176, | |
585 | [I_LCOEF_RSEQIOPS] = 41708, | |
586 | [I_LCOEF_RRANDIOPS] = 370, | |
587 | [I_LCOEF_WBPS] = 178075866, | |
588 | [I_LCOEF_WSEQIOPS] = 42705, | |
589 | [I_LCOEF_WRANDIOPS] = 378, | |
590 | }, | |
591 | }, | |
592 | [AUTOP_SSD_QD1] = { | |
593 | .qos = { | |
594 | [QOS_RLAT] = 25000, /* 25ms */ | |
595 | [QOS_WLAT] = 25000, | |
596 | [QOS_MIN] = VRATE_MIN_PPM, | |
597 | [QOS_MAX] = VRATE_MAX_PPM, | |
598 | }, | |
599 | .i_lcoefs = { | |
600 | [I_LCOEF_RBPS] = 245855193, | |
601 | [I_LCOEF_RSEQIOPS] = 61575, | |
602 | [I_LCOEF_RRANDIOPS] = 6946, | |
603 | [I_LCOEF_WBPS] = 141365009, | |
604 | [I_LCOEF_WSEQIOPS] = 33716, | |
605 | [I_LCOEF_WRANDIOPS] = 26796, | |
606 | }, | |
607 | }, | |
608 | [AUTOP_SSD_DFL] = { | |
609 | .qos = { | |
610 | [QOS_RLAT] = 25000, /* 25ms */ | |
611 | [QOS_WLAT] = 25000, | |
612 | [QOS_MIN] = VRATE_MIN_PPM, | |
613 | [QOS_MAX] = VRATE_MAX_PPM, | |
614 | }, | |
615 | .i_lcoefs = { | |
616 | [I_LCOEF_RBPS] = 488636629, | |
617 | [I_LCOEF_RSEQIOPS] = 8932, | |
618 | [I_LCOEF_RRANDIOPS] = 8518, | |
619 | [I_LCOEF_WBPS] = 427891549, | |
620 | [I_LCOEF_WSEQIOPS] = 28755, | |
621 | [I_LCOEF_WRANDIOPS] = 21940, | |
622 | }, | |
623 | .too_fast_vrate_pct = 500, | |
624 | }, | |
625 | [AUTOP_SSD_FAST] = { | |
626 | .qos = { | |
627 | [QOS_RLAT] = 5000, /* 5ms */ | |
628 | [QOS_WLAT] = 5000, | |
629 | [QOS_MIN] = VRATE_MIN_PPM, | |
630 | [QOS_MAX] = VRATE_MAX_PPM, | |
631 | }, | |
632 | .i_lcoefs = { | |
633 | [I_LCOEF_RBPS] = 3102524156LLU, | |
634 | [I_LCOEF_RSEQIOPS] = 724816, | |
635 | [I_LCOEF_RRANDIOPS] = 778122, | |
636 | [I_LCOEF_WBPS] = 1742780862LLU, | |
637 | [I_LCOEF_WSEQIOPS] = 425702, | |
638 | [I_LCOEF_WRANDIOPS] = 443193, | |
639 | }, | |
640 | .too_slow_vrate_pct = 10, | |
641 | }, | |
642 | }; | |
643 | ||
644 | /* | |
645 | * vrate adjust percentages indexed by ioc->busy_level. We adjust up on | |
646 | * vtime credit shortage and down on device saturation. | |
647 | */ | |
648 | static u32 vrate_adj_pct[] = | |
649 | { 0, 0, 0, 0, | |
650 | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, | |
651 | 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, | |
652 | 4, 4, 4, 4, 4, 4, 4, 4, 8, 8, 8, 8, 8, 8, 8, 8, 16 }; | |
653 | ||
654 | static struct blkcg_policy blkcg_policy_iocost; | |
655 | ||
656 | /* accessors and helpers */ | |
657 | static struct ioc *rqos_to_ioc(struct rq_qos *rqos) | |
658 | { | |
659 | return container_of(rqos, struct ioc, rqos); | |
660 | } | |
661 | ||
662 | static struct ioc *q_to_ioc(struct request_queue *q) | |
663 | { | |
664 | return rqos_to_ioc(rq_qos_id(q, RQ_QOS_COST)); | |
665 | } | |
666 | ||
7caa4715 TH |
667 | static const char __maybe_unused *ioc_name(struct ioc *ioc) |
668 | { | |
9df3e651 CH |
669 | struct gendisk *disk = ioc->rqos.q->disk; |
670 | ||
671 | if (!disk) | |
672 | return "<unknown>"; | |
673 | return disk->disk_name; | |
7caa4715 TH |
674 | } |
675 | ||
676 | static struct ioc_gq *pd_to_iocg(struct blkg_policy_data *pd) | |
677 | { | |
678 | return pd ? container_of(pd, struct ioc_gq, pd) : NULL; | |
679 | } | |
680 | ||
681 | static struct ioc_gq *blkg_to_iocg(struct blkcg_gq *blkg) | |
682 | { | |
683 | return pd_to_iocg(blkg_to_pd(blkg, &blkcg_policy_iocost)); | |
684 | } | |
685 | ||
686 | static struct blkcg_gq *iocg_to_blkg(struct ioc_gq *iocg) | |
687 | { | |
688 | return pd_to_blkg(&iocg->pd); | |
689 | } | |
690 | ||
691 | static struct ioc_cgrp *blkcg_to_iocc(struct blkcg *blkcg) | |
692 | { | |
693 | return container_of(blkcg_to_cpd(blkcg, &blkcg_policy_iocost), | |
694 | struct ioc_cgrp, cpd); | |
695 | } | |
696 | ||
697 | /* | |
698 | * Scale @abs_cost to the inverse of @hw_inuse. The lower the hierarchical | |
36a52481 | 699 | * weight, the more expensive each IO. Must round up. |
7caa4715 TH |
700 | */ |
701 | static u64 abs_cost_to_cost(u64 abs_cost, u32 hw_inuse) | |
702 | { | |
fe20cdb5 | 703 | return DIV64_U64_ROUND_UP(abs_cost * WEIGHT_ONE, hw_inuse); |
7caa4715 TH |
704 | } |
705 | ||
36a52481 TH |
706 | /* |
707 | * The inverse of abs_cost_to_cost(). Must round up. | |
708 | */ | |
709 | static u64 cost_to_abs_cost(u64 cost, u32 hw_inuse) | |
710 | { | |
fe20cdb5 | 711 | return DIV64_U64_ROUND_UP(cost * hw_inuse, WEIGHT_ONE); |
36a52481 TH |
712 | } |
713 | ||
97eb1975 TH |
714 | static void iocg_commit_bio(struct ioc_gq *iocg, struct bio *bio, |
715 | u64 abs_cost, u64 cost) | |
7caa4715 | 716 | { |
97eb1975 TH |
717 | struct iocg_pcpu_stat *gcs; |
718 | ||
7caa4715 TH |
719 | bio->bi_iocost_cost = cost; |
720 | atomic64_add(cost, &iocg->vtime); | |
97eb1975 TH |
721 | |
722 | gcs = get_cpu_ptr(iocg->pcpu_stat); | |
723 | local64_add(abs_cost, &gcs->abs_vusage); | |
724 | put_cpu_ptr(gcs); | |
7caa4715 TH |
725 | } |
726 | ||
da437b95 TH |
727 | static void iocg_lock(struct ioc_gq *iocg, bool lock_ioc, unsigned long *flags) |
728 | { | |
729 | if (lock_ioc) { | |
730 | spin_lock_irqsave(&iocg->ioc->lock, *flags); | |
731 | spin_lock(&iocg->waitq.lock); | |
732 | } else { | |
733 | spin_lock_irqsave(&iocg->waitq.lock, *flags); | |
734 | } | |
735 | } | |
736 | ||
737 | static void iocg_unlock(struct ioc_gq *iocg, bool unlock_ioc, unsigned long *flags) | |
738 | { | |
739 | if (unlock_ioc) { | |
740 | spin_unlock(&iocg->waitq.lock); | |
741 | spin_unlock_irqrestore(&iocg->ioc->lock, *flags); | |
742 | } else { | |
743 | spin_unlock_irqrestore(&iocg->waitq.lock, *flags); | |
744 | } | |
745 | } | |
746 | ||
7caa4715 TH |
747 | #define CREATE_TRACE_POINTS |
748 | #include <trace/events/iocost.h> | |
749 | ||
7ca5b2e6 TH |
750 | static void ioc_refresh_margins(struct ioc *ioc) |
751 | { | |
752 | struct ioc_margins *margins = &ioc->margins; | |
753 | u32 period_us = ioc->period_us; | |
ac33e91e | 754 | u64 vrate = ioc->vtime_base_rate; |
7ca5b2e6 TH |
755 | |
756 | margins->min = (period_us * MARGIN_MIN_PCT / 100) * vrate; | |
f1de2439 TH |
757 | margins->low = (period_us * MARGIN_LOW_PCT / 100) * vrate; |
758 | margins->target = (period_us * MARGIN_TARGET_PCT / 100) * vrate; | |
7ca5b2e6 TH |
759 | } |
760 | ||
7caa4715 TH |
761 | /* latency Qos params changed, update period_us and all the dependent params */ |
762 | static void ioc_refresh_period_us(struct ioc *ioc) | |
763 | { | |
764 | u32 ppm, lat, multi, period_us; | |
765 | ||
766 | lockdep_assert_held(&ioc->lock); | |
767 | ||
768 | /* pick the higher latency target */ | |
769 | if (ioc->params.qos[QOS_RLAT] >= ioc->params.qos[QOS_WLAT]) { | |
770 | ppm = ioc->params.qos[QOS_RPPM]; | |
771 | lat = ioc->params.qos[QOS_RLAT]; | |
772 | } else { | |
773 | ppm = ioc->params.qos[QOS_WPPM]; | |
774 | lat = ioc->params.qos[QOS_WLAT]; | |
775 | } | |
776 | ||
777 | /* | |
778 | * We want the period to be long enough to contain a healthy number | |
779 | * of IOs while short enough for granular control. Define it as a | |
780 | * multiple of the latency target. Ideally, the multiplier should | |
781 | * be scaled according to the percentile so that it would nominally | |
782 | * contain a certain number of requests. Let's be simpler and | |
783 | * scale it linearly so that it's 2x >= pct(90) and 10x at pct(50). | |
784 | */ | |
785 | if (ppm) | |
786 | multi = max_t(u32, (MILLION - ppm) / 50000, 2); | |
787 | else | |
788 | multi = 2; | |
789 | period_us = multi * lat; | |
790 | period_us = clamp_t(u32, period_us, MIN_PERIOD, MAX_PERIOD); | |
791 | ||
792 | /* calculate dependent params */ | |
793 | ioc->period_us = period_us; | |
7ca5b2e6 TH |
794 | ioc->timer_slack_ns = div64_u64( |
795 | (u64)period_us * NSEC_PER_USEC * TIMER_SLACK_PCT, | |
796 | 100); | |
797 | ioc_refresh_margins(ioc); | |
7caa4715 TH |
798 | } |
799 | ||
800 | static int ioc_autop_idx(struct ioc *ioc) | |
801 | { | |
802 | int idx = ioc->autop_idx; | |
803 | const struct ioc_params *p = &autop[idx]; | |
804 | u32 vrate_pct; | |
805 | u64 now_ns; | |
806 | ||
807 | /* rotational? */ | |
808 | if (!blk_queue_nonrot(ioc->rqos.q)) | |
809 | return AUTOP_HDD; | |
810 | ||
811 | /* handle SATA SSDs w/ broken NCQ */ | |
812 | if (blk_queue_depth(ioc->rqos.q) == 1) | |
813 | return AUTOP_SSD_QD1; | |
814 | ||
815 | /* use one of the normal ssd sets */ | |
816 | if (idx < AUTOP_SSD_DFL) | |
817 | return AUTOP_SSD_DFL; | |
818 | ||
819 | /* if user is overriding anything, maintain what was there */ | |
820 | if (ioc->user_qos_params || ioc->user_cost_model) | |
821 | return idx; | |
822 | ||
823 | /* step up/down based on the vrate */ | |
ac33e91e | 824 | vrate_pct = div64_u64(ioc->vtime_base_rate * 100, VTIME_PER_USEC); |
7caa4715 TH |
825 | now_ns = ktime_get_ns(); |
826 | ||
827 | if (p->too_fast_vrate_pct && p->too_fast_vrate_pct <= vrate_pct) { | |
828 | if (!ioc->autop_too_fast_at) | |
829 | ioc->autop_too_fast_at = now_ns; | |
830 | if (now_ns - ioc->autop_too_fast_at >= AUTOP_CYCLE_NSEC) | |
831 | return idx + 1; | |
832 | } else { | |
833 | ioc->autop_too_fast_at = 0; | |
834 | } | |
835 | ||
836 | if (p->too_slow_vrate_pct && p->too_slow_vrate_pct >= vrate_pct) { | |
837 | if (!ioc->autop_too_slow_at) | |
838 | ioc->autop_too_slow_at = now_ns; | |
839 | if (now_ns - ioc->autop_too_slow_at >= AUTOP_CYCLE_NSEC) | |
840 | return idx - 1; | |
841 | } else { | |
842 | ioc->autop_too_slow_at = 0; | |
843 | } | |
844 | ||
845 | return idx; | |
846 | } | |
847 | ||
848 | /* | |
849 | * Take the followings as input | |
850 | * | |
851 | * @bps maximum sequential throughput | |
852 | * @seqiops maximum sequential 4k iops | |
853 | * @randiops maximum random 4k iops | |
854 | * | |
855 | * and calculate the linear model cost coefficients. | |
856 | * | |
857 | * *@page per-page cost 1s / (@bps / 4096) | |
858 | * *@seqio base cost of a seq IO max((1s / @seqiops) - *@page, 0) | |
859 | * @randiops base cost of a rand IO max((1s / @randiops) - *@page, 0) | |
860 | */ | |
861 | static void calc_lcoefs(u64 bps, u64 seqiops, u64 randiops, | |
862 | u64 *page, u64 *seqio, u64 *randio) | |
863 | { | |
864 | u64 v; | |
865 | ||
866 | *page = *seqio = *randio = 0; | |
867 | ||
868 | if (bps) | |
869 | *page = DIV64_U64_ROUND_UP(VTIME_PER_SEC, | |
870 | DIV_ROUND_UP_ULL(bps, IOC_PAGE_SIZE)); | |
871 | ||
872 | if (seqiops) { | |
873 | v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, seqiops); | |
874 | if (v > *page) | |
875 | *seqio = v - *page; | |
876 | } | |
877 | ||
878 | if (randiops) { | |
879 | v = DIV64_U64_ROUND_UP(VTIME_PER_SEC, randiops); | |
880 | if (v > *page) | |
881 | *randio = v - *page; | |
882 | } | |
883 | } | |
884 | ||
885 | static void ioc_refresh_lcoefs(struct ioc *ioc) | |
886 | { | |
887 | u64 *u = ioc->params.i_lcoefs; | |
888 | u64 *c = ioc->params.lcoefs; | |
889 | ||
890 | calc_lcoefs(u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS], | |
891 | &c[LCOEF_RPAGE], &c[LCOEF_RSEQIO], &c[LCOEF_RRANDIO]); | |
892 | calc_lcoefs(u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS], | |
893 | &c[LCOEF_WPAGE], &c[LCOEF_WSEQIO], &c[LCOEF_WRANDIO]); | |
894 | } | |
895 | ||
896 | static bool ioc_refresh_params(struct ioc *ioc, bool force) | |
897 | { | |
898 | const struct ioc_params *p; | |
899 | int idx; | |
900 | ||
901 | lockdep_assert_held(&ioc->lock); | |
902 | ||
903 | idx = ioc_autop_idx(ioc); | |
904 | p = &autop[idx]; | |
905 | ||
906 | if (idx == ioc->autop_idx && !force) | |
907 | return false; | |
908 | ||
909 | if (idx != ioc->autop_idx) | |
910 | atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC); | |
911 | ||
912 | ioc->autop_idx = idx; | |
913 | ioc->autop_too_fast_at = 0; | |
914 | ioc->autop_too_slow_at = 0; | |
915 | ||
916 | if (!ioc->user_qos_params) | |
917 | memcpy(ioc->params.qos, p->qos, sizeof(p->qos)); | |
918 | if (!ioc->user_cost_model) | |
919 | memcpy(ioc->params.i_lcoefs, p->i_lcoefs, sizeof(p->i_lcoefs)); | |
920 | ||
921 | ioc_refresh_period_us(ioc); | |
922 | ioc_refresh_lcoefs(ioc); | |
923 | ||
924 | ioc->vrate_min = DIV64_U64_ROUND_UP((u64)ioc->params.qos[QOS_MIN] * | |
925 | VTIME_PER_USEC, MILLION); | |
926 | ioc->vrate_max = div64_u64((u64)ioc->params.qos[QOS_MAX] * | |
927 | VTIME_PER_USEC, MILLION); | |
928 | ||
929 | return true; | |
930 | } | |
931 | ||
ac33e91e TH |
932 | /* |
933 | * When an iocg accumulates too much vtime or gets deactivated, we throw away | |
934 | * some vtime, which lowers the overall device utilization. As the exact amount | |
935 | * which is being thrown away is known, we can compensate by accelerating the | |
936 | * vrate accordingly so that the extra vtime generated in the current period | |
937 | * matches what got lost. | |
938 | */ | |
939 | static void ioc_refresh_vrate(struct ioc *ioc, struct ioc_now *now) | |
940 | { | |
941 | s64 pleft = ioc->period_at + ioc->period_us - now->now; | |
942 | s64 vperiod = ioc->period_us * ioc->vtime_base_rate; | |
943 | s64 vcomp, vcomp_min, vcomp_max; | |
944 | ||
945 | lockdep_assert_held(&ioc->lock); | |
946 | ||
947 | /* we need some time left in this period */ | |
948 | if (pleft <= 0) | |
949 | goto done; | |
950 | ||
951 | /* | |
952 | * Calculate how much vrate should be adjusted to offset the error. | |
953 | * Limit the amount of adjustment and deduct the adjusted amount from | |
954 | * the error. | |
955 | */ | |
956 | vcomp = -div64_s64(ioc->vtime_err, pleft); | |
957 | vcomp_min = -(ioc->vtime_base_rate >> 1); | |
958 | vcomp_max = ioc->vtime_base_rate; | |
959 | vcomp = clamp(vcomp, vcomp_min, vcomp_max); | |
960 | ||
961 | ioc->vtime_err += vcomp * pleft; | |
962 | ||
963 | atomic64_set(&ioc->vtime_rate, ioc->vtime_base_rate + vcomp); | |
964 | done: | |
965 | /* bound how much error can accumulate */ | |
966 | ioc->vtime_err = clamp(ioc->vtime_err, -vperiod, vperiod); | |
967 | } | |
968 | ||
926f75f6 BW |
969 | static void ioc_adjust_base_vrate(struct ioc *ioc, u32 rq_wait_pct, |
970 | int nr_lagging, int nr_shortages, | |
971 | int prev_busy_level, u32 *missed_ppm) | |
972 | { | |
973 | u64 vrate = ioc->vtime_base_rate; | |
974 | u64 vrate_min = ioc->vrate_min, vrate_max = ioc->vrate_max; | |
975 | ||
976 | if (!ioc->busy_level || (ioc->busy_level < 0 && nr_lagging)) { | |
977 | if (ioc->busy_level != prev_busy_level || nr_lagging) | |
978 | trace_iocost_ioc_vrate_adj(ioc, atomic64_read(&ioc->vtime_rate), | |
979 | missed_ppm, rq_wait_pct, | |
980 | nr_lagging, nr_shortages); | |
981 | ||
982 | return; | |
983 | } | |
984 | ||
926f75f6 BW |
985 | /* |
986 | * If vrate is out of bounds, apply clamp gradually as the | |
987 | * bounds can change abruptly. Otherwise, apply busy_level | |
988 | * based adjustment. | |
989 | */ | |
990 | if (vrate < vrate_min) { | |
991 | vrate = div64_u64(vrate * (100 + VRATE_CLAMP_ADJ_PCT), 100); | |
992 | vrate = min(vrate, vrate_min); | |
993 | } else if (vrate > vrate_max) { | |
994 | vrate = div64_u64(vrate * (100 - VRATE_CLAMP_ADJ_PCT), 100); | |
995 | vrate = max(vrate, vrate_max); | |
996 | } else { | |
997 | int idx = min_t(int, abs(ioc->busy_level), | |
998 | ARRAY_SIZE(vrate_adj_pct) - 1); | |
999 | u32 adj_pct = vrate_adj_pct[idx]; | |
1000 | ||
1001 | if (ioc->busy_level > 0) | |
1002 | adj_pct = 100 - adj_pct; | |
1003 | else | |
1004 | adj_pct = 100 + adj_pct; | |
1005 | ||
1006 | vrate = clamp(DIV64_U64_ROUND_UP(vrate * adj_pct, 100), | |
1007 | vrate_min, vrate_max); | |
1008 | } | |
1009 | ||
1010 | trace_iocost_ioc_vrate_adj(ioc, vrate, missed_ppm, rq_wait_pct, | |
1011 | nr_lagging, nr_shortages); | |
1012 | ||
1013 | ioc->vtime_base_rate = vrate; | |
1014 | ioc_refresh_margins(ioc); | |
1015 | } | |
1016 | ||
7caa4715 TH |
1017 | /* take a snapshot of the current [v]time and vrate */ |
1018 | static void ioc_now(struct ioc *ioc, struct ioc_now *now) | |
1019 | { | |
1020 | unsigned seq; | |
1021 | ||
1022 | now->now_ns = ktime_get(); | |
1023 | now->now = ktime_to_us(now->now_ns); | |
1024 | now->vrate = atomic64_read(&ioc->vtime_rate); | |
1025 | ||
1026 | /* | |
1027 | * The current vtime is | |
1028 | * | |
1029 | * vtime at period start + (wallclock time since the start) * vrate | |
1030 | * | |
1031 | * As a consistent snapshot of `period_at_vtime` and `period_at` is | |
1032 | * needed, they're seqcount protected. | |
1033 | */ | |
1034 | do { | |
1035 | seq = read_seqcount_begin(&ioc->period_seqcount); | |
1036 | now->vnow = ioc->period_at_vtime + | |
1037 | (now->now - ioc->period_at) * now->vrate; | |
1038 | } while (read_seqcount_retry(&ioc->period_seqcount, seq)); | |
1039 | } | |
1040 | ||
1041 | static void ioc_start_period(struct ioc *ioc, struct ioc_now *now) | |
1042 | { | |
7caa4715 TH |
1043 | WARN_ON_ONCE(ioc->running != IOC_RUNNING); |
1044 | ||
1045 | write_seqcount_begin(&ioc->period_seqcount); | |
1046 | ioc->period_at = now->now; | |
1047 | ioc->period_at_vtime = now->vnow; | |
1048 | write_seqcount_end(&ioc->period_seqcount); | |
1049 | ||
1050 | ioc->timer.expires = jiffies + usecs_to_jiffies(ioc->period_us); | |
1051 | add_timer(&ioc->timer); | |
1052 | } | |
1053 | ||
1054 | /* | |
1055 | * Update @iocg's `active` and `inuse` to @active and @inuse, update level | |
b0853ab4 TH |
1056 | * weight sums and propagate upwards accordingly. If @save, the current margin |
1057 | * is saved to be used as reference for later inuse in-period adjustments. | |
7caa4715 | 1058 | */ |
b0853ab4 TH |
1059 | static void __propagate_weights(struct ioc_gq *iocg, u32 active, u32 inuse, |
1060 | bool save, struct ioc_now *now) | |
7caa4715 TH |
1061 | { |
1062 | struct ioc *ioc = iocg->ioc; | |
1063 | int lvl; | |
1064 | ||
1065 | lockdep_assert_held(&ioc->lock); | |
1066 | ||
e9f4eee9 TH |
1067 | /* |
1068 | * For an active leaf node, its inuse shouldn't be zero or exceed | |
1069 | * @active. An active internal node's inuse is solely determined by the | |
1070 | * inuse to active ratio of its children regardless of @inuse. | |
1071 | */ | |
1072 | if (list_empty(&iocg->active_list) && iocg->child_active_sum) { | |
1073 | inuse = DIV64_U64_ROUND_UP(active * iocg->child_inuse_sum, | |
1074 | iocg->child_active_sum); | |
1075 | } else { | |
1076 | inuse = clamp_t(u32, inuse, 1, active); | |
1077 | } | |
db84a72a | 1078 | |
b0853ab4 TH |
1079 | iocg->last_inuse = iocg->inuse; |
1080 | if (save) | |
1081 | iocg->saved_margin = now->vnow - atomic64_read(&iocg->vtime); | |
1082 | ||
db84a72a TH |
1083 | if (active == iocg->active && inuse == iocg->inuse) |
1084 | return; | |
7caa4715 TH |
1085 | |
1086 | for (lvl = iocg->level - 1; lvl >= 0; lvl--) { | |
1087 | struct ioc_gq *parent = iocg->ancestors[lvl]; | |
1088 | struct ioc_gq *child = iocg->ancestors[lvl + 1]; | |
1089 | u32 parent_active = 0, parent_inuse = 0; | |
1090 | ||
1091 | /* update the level sums */ | |
1092 | parent->child_active_sum += (s32)(active - child->active); | |
1093 | parent->child_inuse_sum += (s32)(inuse - child->inuse); | |
e9f4eee9 | 1094 | /* apply the updates */ |
7caa4715 TH |
1095 | child->active = active; |
1096 | child->inuse = inuse; | |
1097 | ||
1098 | /* | |
1099 | * The delta between inuse and active sums indicates that | |
5ba1add2 | 1100 | * much of weight is being given away. Parent's inuse |
7caa4715 TH |
1101 | * and active should reflect the ratio. |
1102 | */ | |
1103 | if (parent->child_active_sum) { | |
1104 | parent_active = parent->weight; | |
1105 | parent_inuse = DIV64_U64_ROUND_UP( | |
1106 | parent_active * parent->child_inuse_sum, | |
1107 | parent->child_active_sum); | |
1108 | } | |
1109 | ||
1110 | /* do we need to keep walking up? */ | |
1111 | if (parent_active == parent->active && | |
1112 | parent_inuse == parent->inuse) | |
1113 | break; | |
1114 | ||
1115 | active = parent_active; | |
1116 | inuse = parent_inuse; | |
1117 | } | |
1118 | ||
1119 | ioc->weights_updated = true; | |
1120 | } | |
1121 | ||
00410f1b | 1122 | static void commit_weights(struct ioc *ioc) |
7caa4715 TH |
1123 | { |
1124 | lockdep_assert_held(&ioc->lock); | |
1125 | ||
1126 | if (ioc->weights_updated) { | |
1127 | /* paired with rmb in current_hweight(), see there */ | |
1128 | smp_wmb(); | |
1129 | atomic_inc(&ioc->hweight_gen); | |
1130 | ioc->weights_updated = false; | |
1131 | } | |
1132 | } | |
1133 | ||
b0853ab4 TH |
1134 | static void propagate_weights(struct ioc_gq *iocg, u32 active, u32 inuse, |
1135 | bool save, struct ioc_now *now) | |
7caa4715 | 1136 | { |
b0853ab4 | 1137 | __propagate_weights(iocg, active, inuse, save, now); |
00410f1b | 1138 | commit_weights(iocg->ioc); |
7caa4715 TH |
1139 | } |
1140 | ||
1141 | static void current_hweight(struct ioc_gq *iocg, u32 *hw_activep, u32 *hw_inusep) | |
1142 | { | |
1143 | struct ioc *ioc = iocg->ioc; | |
1144 | int lvl; | |
1145 | u32 hwa, hwi; | |
1146 | int ioc_gen; | |
1147 | ||
1148 | /* hot path - if uptodate, use cached */ | |
1149 | ioc_gen = atomic_read(&ioc->hweight_gen); | |
1150 | if (ioc_gen == iocg->hweight_gen) | |
1151 | goto out; | |
1152 | ||
1153 | /* | |
00410f1b TH |
1154 | * Paired with wmb in commit_weights(). If we saw the updated |
1155 | * hweight_gen, all the weight updates from __propagate_weights() are | |
1156 | * visible too. | |
7caa4715 TH |
1157 | * |
1158 | * We can race with weight updates during calculation and get it | |
1159 | * wrong. However, hweight_gen would have changed and a future | |
1160 | * reader will recalculate and we're guaranteed to discard the | |
1161 | * wrong result soon. | |
1162 | */ | |
1163 | smp_rmb(); | |
1164 | ||
fe20cdb5 | 1165 | hwa = hwi = WEIGHT_ONE; |
7caa4715 TH |
1166 | for (lvl = 0; lvl <= iocg->level - 1; lvl++) { |
1167 | struct ioc_gq *parent = iocg->ancestors[lvl]; | |
1168 | struct ioc_gq *child = iocg->ancestors[lvl + 1]; | |
bd0adb91 TH |
1169 | u64 active_sum = READ_ONCE(parent->child_active_sum); |
1170 | u64 inuse_sum = READ_ONCE(parent->child_inuse_sum); | |
7caa4715 TH |
1171 | u32 active = READ_ONCE(child->active); |
1172 | u32 inuse = READ_ONCE(child->inuse); | |
1173 | ||
1174 | /* we can race with deactivations and either may read as zero */ | |
1175 | if (!active_sum || !inuse_sum) | |
1176 | continue; | |
1177 | ||
bd0adb91 TH |
1178 | active_sum = max_t(u64, active, active_sum); |
1179 | hwa = div64_u64((u64)hwa * active, active_sum); | |
7caa4715 | 1180 | |
bd0adb91 TH |
1181 | inuse_sum = max_t(u64, inuse, inuse_sum); |
1182 | hwi = div64_u64((u64)hwi * inuse, inuse_sum); | |
7caa4715 TH |
1183 | } |
1184 | ||
1185 | iocg->hweight_active = max_t(u32, hwa, 1); | |
1186 | iocg->hweight_inuse = max_t(u32, hwi, 1); | |
1187 | iocg->hweight_gen = ioc_gen; | |
1188 | out: | |
1189 | if (hw_activep) | |
1190 | *hw_activep = iocg->hweight_active; | |
1191 | if (hw_inusep) | |
1192 | *hw_inusep = iocg->hweight_inuse; | |
1193 | } | |
1194 | ||
93f7d2db TH |
1195 | /* |
1196 | * Calculate the hweight_inuse @iocg would get with max @inuse assuming all the | |
1197 | * other weights stay unchanged. | |
1198 | */ | |
1199 | static u32 current_hweight_max(struct ioc_gq *iocg) | |
1200 | { | |
1201 | u32 hwm = WEIGHT_ONE; | |
1202 | u32 inuse = iocg->active; | |
1203 | u64 child_inuse_sum; | |
1204 | int lvl; | |
1205 | ||
1206 | lockdep_assert_held(&iocg->ioc->lock); | |
1207 | ||
1208 | for (lvl = iocg->level - 1; lvl >= 0; lvl--) { | |
1209 | struct ioc_gq *parent = iocg->ancestors[lvl]; | |
1210 | struct ioc_gq *child = iocg->ancestors[lvl + 1]; | |
1211 | ||
1212 | child_inuse_sum = parent->child_inuse_sum + inuse - child->inuse; | |
1213 | hwm = div64_u64((u64)hwm * inuse, child_inuse_sum); | |
1214 | inuse = DIV64_U64_ROUND_UP(parent->active * child_inuse_sum, | |
1215 | parent->child_active_sum); | |
1216 | } | |
1217 | ||
1218 | return max_t(u32, hwm, 1); | |
1219 | } | |
1220 | ||
b0853ab4 | 1221 | static void weight_updated(struct ioc_gq *iocg, struct ioc_now *now) |
7caa4715 TH |
1222 | { |
1223 | struct ioc *ioc = iocg->ioc; | |
1224 | struct blkcg_gq *blkg = iocg_to_blkg(iocg); | |
1225 | struct ioc_cgrp *iocc = blkcg_to_iocc(blkg->blkcg); | |
1226 | u32 weight; | |
1227 | ||
1228 | lockdep_assert_held(&ioc->lock); | |
1229 | ||
1230 | weight = iocg->cfg_weight ?: iocc->dfl_weight; | |
1231 | if (weight != iocg->weight && iocg->active) | |
b0853ab4 | 1232 | propagate_weights(iocg, weight, iocg->inuse, true, now); |
7caa4715 TH |
1233 | iocg->weight = weight; |
1234 | } | |
1235 | ||
1236 | static bool iocg_activate(struct ioc_gq *iocg, struct ioc_now *now) | |
1237 | { | |
1238 | struct ioc *ioc = iocg->ioc; | |
ac33e91e TH |
1239 | u64 last_period, cur_period; |
1240 | u64 vtime, vtarget; | |
7caa4715 TH |
1241 | int i; |
1242 | ||
1243 | /* | |
1244 | * If seem to be already active, just update the stamp to tell the | |
1245 | * timer that we're still active. We don't mind occassional races. | |
1246 | */ | |
1247 | if (!list_empty(&iocg->active_list)) { | |
1248 | ioc_now(ioc, now); | |
1249 | cur_period = atomic64_read(&ioc->cur_period); | |
1250 | if (atomic64_read(&iocg->active_period) != cur_period) | |
1251 | atomic64_set(&iocg->active_period, cur_period); | |
1252 | return true; | |
1253 | } | |
1254 | ||
1255 | /* racy check on internal node IOs, treat as root level IOs */ | |
1256 | if (iocg->child_active_sum) | |
1257 | return false; | |
1258 | ||
1259 | spin_lock_irq(&ioc->lock); | |
1260 | ||
1261 | ioc_now(ioc, now); | |
1262 | ||
1263 | /* update period */ | |
1264 | cur_period = atomic64_read(&ioc->cur_period); | |
1265 | last_period = atomic64_read(&iocg->active_period); | |
1266 | atomic64_set(&iocg->active_period, cur_period); | |
1267 | ||
1268 | /* already activated or breaking leaf-only constraint? */ | |
8b37bc27 JX |
1269 | if (!list_empty(&iocg->active_list)) |
1270 | goto succeed_unlock; | |
1271 | for (i = iocg->level - 1; i > 0; i--) | |
1272 | if (!list_empty(&iocg->ancestors[i]->active_list)) | |
7caa4715 | 1273 | goto fail_unlock; |
8b37bc27 | 1274 | |
7caa4715 TH |
1275 | if (iocg->child_active_sum) |
1276 | goto fail_unlock; | |
1277 | ||
1278 | /* | |
ac33e91e TH |
1279 | * Always start with the target budget. On deactivation, we throw away |
1280 | * anything above it. | |
7caa4715 | 1281 | */ |
ac33e91e | 1282 | vtarget = now->vnow - ioc->margins.target; |
7caa4715 | 1283 | vtime = atomic64_read(&iocg->vtime); |
7caa4715 | 1284 | |
ac33e91e TH |
1285 | atomic64_add(vtarget - vtime, &iocg->vtime); |
1286 | atomic64_add(vtarget - vtime, &iocg->done_vtime); | |
1287 | vtime = vtarget; | |
7caa4715 TH |
1288 | |
1289 | /* | |
1290 | * Activate, propagate weight and start period timer if not | |
1291 | * running. Reset hweight_gen to avoid accidental match from | |
1292 | * wrapping. | |
1293 | */ | |
1294 | iocg->hweight_gen = atomic_read(&ioc->hweight_gen) - 1; | |
1295 | list_add(&iocg->active_list, &ioc->active_iocgs); | |
b0853ab4 | 1296 | |
00410f1b | 1297 | propagate_weights(iocg, iocg->weight, |
b0853ab4 | 1298 | iocg->last_inuse ?: iocg->weight, true, now); |
7caa4715 TH |
1299 | |
1300 | TRACE_IOCG_PATH(iocg_activate, iocg, now, | |
1301 | last_period, cur_period, vtime); | |
1302 | ||
1aa50d02 | 1303 | iocg->activated_at = now->now; |
7caa4715 TH |
1304 | |
1305 | if (ioc->running == IOC_IDLE) { | |
1306 | ioc->running = IOC_RUNNING; | |
c7af2a00 TH |
1307 | ioc->dfgv_period_at = now->now; |
1308 | ioc->dfgv_period_rem = 0; | |
7caa4715 TH |
1309 | ioc_start_period(ioc, now); |
1310 | } | |
1311 | ||
8b37bc27 | 1312 | succeed_unlock: |
7caa4715 TH |
1313 | spin_unlock_irq(&ioc->lock); |
1314 | return true; | |
1315 | ||
1316 | fail_unlock: | |
1317 | spin_unlock_irq(&ioc->lock); | |
1318 | return false; | |
1319 | } | |
1320 | ||
6ef20f78 TH |
1321 | static bool iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now) |
1322 | { | |
1323 | struct ioc *ioc = iocg->ioc; | |
1324 | struct blkcg_gq *blkg = iocg_to_blkg(iocg); | |
5160a5a5 TH |
1325 | u64 tdelta, delay, new_delay; |
1326 | s64 vover, vover_pct; | |
c421a3eb | 1327 | u32 hwa; |
6ef20f78 TH |
1328 | |
1329 | lockdep_assert_held(&iocg->waitq.lock); | |
1330 | ||
5160a5a5 TH |
1331 | /* calculate the current delay in effect - 1/2 every second */ |
1332 | tdelta = now->now - iocg->delay_at; | |
1333 | if (iocg->delay) | |
1334 | delay = iocg->delay >> div64_u64(tdelta, USEC_PER_SEC); | |
1335 | else | |
1336 | delay = 0; | |
1337 | ||
1338 | /* calculate the new delay from the debt amount */ | |
c421a3eb | 1339 | current_hweight(iocg, &hwa, NULL); |
5160a5a5 TH |
1340 | vover = atomic64_read(&iocg->vtime) + |
1341 | abs_cost_to_cost(iocg->abs_vdebt, hwa) - now->vnow; | |
ac33e91e TH |
1342 | vover_pct = div64_s64(100 * vover, |
1343 | ioc->period_us * ioc->vtime_base_rate); | |
5160a5a5 TH |
1344 | |
1345 | if (vover_pct <= MIN_DELAY_THR_PCT) | |
1346 | new_delay = 0; | |
1347 | else if (vover_pct >= MAX_DELAY_THR_PCT) | |
1348 | new_delay = MAX_DELAY; | |
1349 | else | |
1350 | new_delay = MIN_DELAY + | |
1351 | div_u64((MAX_DELAY - MIN_DELAY) * | |
1352 | (vover_pct - MIN_DELAY_THR_PCT), | |
1353 | MAX_DELAY_THR_PCT - MIN_DELAY_THR_PCT); | |
1354 | ||
1355 | /* pick the higher one and apply */ | |
1356 | if (new_delay > delay) { | |
1357 | iocg->delay = new_delay; | |
1358 | iocg->delay_at = now->now; | |
1359 | delay = new_delay; | |
1360 | } | |
6ef20f78 | 1361 | |
5160a5a5 | 1362 | if (delay >= MIN_DELAY) { |
f0bf84a5 TH |
1363 | if (!iocg->indelay_since) |
1364 | iocg->indelay_since = now->now; | |
5160a5a5 TH |
1365 | blkcg_set_delay(blkg, delay * NSEC_PER_USEC); |
1366 | return true; | |
1367 | } else { | |
f0bf84a5 | 1368 | if (iocg->indelay_since) { |
2a371f7d | 1369 | iocg->stat.indelay_us += now->now - iocg->indelay_since; |
f0bf84a5 TH |
1370 | iocg->indelay_since = 0; |
1371 | } | |
5160a5a5 | 1372 | iocg->delay = 0; |
6ef20f78 TH |
1373 | blkcg_clear_delay(blkg); |
1374 | return false; | |
1375 | } | |
6ef20f78 TH |
1376 | } |
1377 | ||
c421a3eb TH |
1378 | static void iocg_incur_debt(struct ioc_gq *iocg, u64 abs_cost, |
1379 | struct ioc_now *now) | |
1380 | { | |
1381 | struct iocg_pcpu_stat *gcs; | |
1382 | ||
1383 | lockdep_assert_held(&iocg->ioc->lock); | |
1384 | lockdep_assert_held(&iocg->waitq.lock); | |
1385 | WARN_ON_ONCE(list_empty(&iocg->active_list)); | |
1386 | ||
1387 | /* | |
1388 | * Once in debt, debt handling owns inuse. @iocg stays at the minimum | |
1389 | * inuse donating all of it share to others until its debt is paid off. | |
1390 | */ | |
f0bf84a5 TH |
1391 | if (!iocg->abs_vdebt && abs_cost) { |
1392 | iocg->indebt_since = now->now; | |
c421a3eb | 1393 | propagate_weights(iocg, iocg->active, 0, false, now); |
f0bf84a5 | 1394 | } |
c421a3eb TH |
1395 | |
1396 | iocg->abs_vdebt += abs_cost; | |
1397 | ||
1398 | gcs = get_cpu_ptr(iocg->pcpu_stat); | |
1399 | local64_add(abs_cost, &gcs->abs_vusage); | |
1400 | put_cpu_ptr(gcs); | |
1401 | } | |
1402 | ||
1403 | static void iocg_pay_debt(struct ioc_gq *iocg, u64 abs_vpay, | |
1404 | struct ioc_now *now) | |
1405 | { | |
1406 | lockdep_assert_held(&iocg->ioc->lock); | |
1407 | lockdep_assert_held(&iocg->waitq.lock); | |
1408 | ||
1409 | /* make sure that nobody messed with @iocg */ | |
1410 | WARN_ON_ONCE(list_empty(&iocg->active_list)); | |
1411 | WARN_ON_ONCE(iocg->inuse > 1); | |
1412 | ||
1413 | iocg->abs_vdebt -= min(abs_vpay, iocg->abs_vdebt); | |
1414 | ||
1415 | /* if debt is paid in full, restore inuse */ | |
f0bf84a5 | 1416 | if (!iocg->abs_vdebt) { |
2a371f7d | 1417 | iocg->stat.indebt_us += now->now - iocg->indebt_since; |
f0bf84a5 TH |
1418 | iocg->indebt_since = 0; |
1419 | ||
c421a3eb TH |
1420 | propagate_weights(iocg, iocg->active, iocg->last_inuse, |
1421 | false, now); | |
f0bf84a5 | 1422 | } |
c421a3eb TH |
1423 | } |
1424 | ||
7caa4715 TH |
1425 | static int iocg_wake_fn(struct wait_queue_entry *wq_entry, unsigned mode, |
1426 | int flags, void *key) | |
1427 | { | |
1428 | struct iocg_wait *wait = container_of(wq_entry, struct iocg_wait, wait); | |
a7609c68 | 1429 | struct iocg_wake_ctx *ctx = key; |
7caa4715 TH |
1430 | u64 cost = abs_cost_to_cost(wait->abs_cost, ctx->hw_inuse); |
1431 | ||
1432 | ctx->vbudget -= cost; | |
1433 | ||
1434 | if (ctx->vbudget < 0) | |
1435 | return -1; | |
1436 | ||
97eb1975 | 1437 | iocg_commit_bio(ctx->iocg, wait->bio, wait->abs_cost, cost); |
5ab189cf | 1438 | wait->committed = true; |
7caa4715 TH |
1439 | |
1440 | /* | |
1441 | * autoremove_wake_function() removes the wait entry only when it | |
5ab189cf TH |
1442 | * actually changed the task state. We want the wait always removed. |
1443 | * Remove explicitly and use default_wake_function(). Note that the | |
1444 | * order of operations is important as finish_wait() tests whether | |
1445 | * @wq_entry is removed without grabbing the lock. | |
7caa4715 | 1446 | */ |
7caa4715 | 1447 | default_wake_function(wq_entry, mode, flags, key); |
5ab189cf | 1448 | list_del_init_careful(&wq_entry->entry); |
7caa4715 TH |
1449 | return 0; |
1450 | } | |
1451 | ||
da437b95 TH |
1452 | /* |
1453 | * Calculate the accumulated budget, pay debt if @pay_debt and wake up waiters | |
1454 | * accordingly. When @pay_debt is %true, the caller must be holding ioc->lock in | |
1455 | * addition to iocg->waitq.lock. | |
1456 | */ | |
1457 | static void iocg_kick_waitq(struct ioc_gq *iocg, bool pay_debt, | |
1458 | struct ioc_now *now) | |
7caa4715 TH |
1459 | { |
1460 | struct ioc *ioc = iocg->ioc; | |
1461 | struct iocg_wake_ctx ctx = { .iocg = iocg }; | |
da437b95 | 1462 | u64 vshortage, expires, oexpires; |
36a52481 | 1463 | s64 vbudget; |
c421a3eb | 1464 | u32 hwa; |
7caa4715 TH |
1465 | |
1466 | lockdep_assert_held(&iocg->waitq.lock); | |
1467 | ||
c421a3eb | 1468 | current_hweight(iocg, &hwa, NULL); |
36a52481 TH |
1469 | vbudget = now->vnow - atomic64_read(&iocg->vtime); |
1470 | ||
1471 | /* pay off debt */ | |
da437b95 | 1472 | if (pay_debt && iocg->abs_vdebt && vbudget > 0) { |
c421a3eb TH |
1473 | u64 abs_vbudget = cost_to_abs_cost(vbudget, hwa); |
1474 | u64 abs_vpay = min_t(u64, abs_vbudget, iocg->abs_vdebt); | |
1475 | u64 vpay = abs_cost_to_cost(abs_vpay, hwa); | |
36a52481 | 1476 | |
da437b95 TH |
1477 | lockdep_assert_held(&ioc->lock); |
1478 | ||
c421a3eb TH |
1479 | atomic64_add(vpay, &iocg->vtime); |
1480 | atomic64_add(vpay, &iocg->done_vtime); | |
1481 | iocg_pay_debt(iocg, abs_vpay, now); | |
1482 | vbudget -= vpay; | |
5160a5a5 | 1483 | } |
7b84b49e | 1484 | |
5160a5a5 | 1485 | if (iocg->abs_vdebt || iocg->delay) |
7b84b49e | 1486 | iocg_kick_delay(iocg, now); |
36a52481 | 1487 | |
da437b95 TH |
1488 | /* |
1489 | * Debt can still be outstanding if we haven't paid all yet or the | |
1490 | * caller raced and called without @pay_debt. Shouldn't wake up waiters | |
1491 | * under debt. Make sure @vbudget reflects the outstanding amount and is | |
1492 | * not positive. | |
1493 | */ | |
1494 | if (iocg->abs_vdebt) { | |
c421a3eb | 1495 | s64 vdebt = abs_cost_to_cost(iocg->abs_vdebt, hwa); |
da437b95 TH |
1496 | vbudget = min_t(s64, 0, vbudget - vdebt); |
1497 | } | |
1498 | ||
7caa4715 | 1499 | /* |
c421a3eb TH |
1500 | * Wake up the ones which are due and see how much vtime we'll need for |
1501 | * the next one. As paying off debt restores hw_inuse, it must be read | |
1502 | * after the above debt payment. | |
7caa4715 | 1503 | */ |
da437b95 | 1504 | ctx.vbudget = vbudget; |
c421a3eb TH |
1505 | current_hweight(iocg, NULL, &ctx.hw_inuse); |
1506 | ||
7caa4715 | 1507 | __wake_up_locked_key(&iocg->waitq, TASK_NORMAL, &ctx); |
c421a3eb | 1508 | |
f0bf84a5 TH |
1509 | if (!waitqueue_active(&iocg->waitq)) { |
1510 | if (iocg->wait_since) { | |
2a371f7d | 1511 | iocg->stat.wait_us += now->now - iocg->wait_since; |
f0bf84a5 TH |
1512 | iocg->wait_since = 0; |
1513 | } | |
7caa4715 | 1514 | return; |
f0bf84a5 TH |
1515 | } |
1516 | ||
1517 | if (!iocg->wait_since) | |
1518 | iocg->wait_since = now->now; | |
1519 | ||
7caa4715 TH |
1520 | if (WARN_ON_ONCE(ctx.vbudget >= 0)) |
1521 | return; | |
1522 | ||
7ca5b2e6 | 1523 | /* determine next wakeup, add a timer margin to guarantee chunking */ |
7caa4715 TH |
1524 | vshortage = -ctx.vbudget; |
1525 | expires = now->now_ns + | |
ac33e91e TH |
1526 | DIV64_U64_ROUND_UP(vshortage, ioc->vtime_base_rate) * |
1527 | NSEC_PER_USEC; | |
7ca5b2e6 | 1528 | expires += ioc->timer_slack_ns; |
7caa4715 TH |
1529 | |
1530 | /* if already active and close enough, don't bother */ | |
1531 | oexpires = ktime_to_ns(hrtimer_get_softexpires(&iocg->waitq_timer)); | |
1532 | if (hrtimer_is_queued(&iocg->waitq_timer) && | |
7ca5b2e6 | 1533 | abs(oexpires - expires) <= ioc->timer_slack_ns) |
7caa4715 TH |
1534 | return; |
1535 | ||
1536 | hrtimer_start_range_ns(&iocg->waitq_timer, ns_to_ktime(expires), | |
7ca5b2e6 | 1537 | ioc->timer_slack_ns, HRTIMER_MODE_ABS); |
7caa4715 TH |
1538 | } |
1539 | ||
1540 | static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer) | |
1541 | { | |
1542 | struct ioc_gq *iocg = container_of(timer, struct ioc_gq, waitq_timer); | |
da437b95 | 1543 | bool pay_debt = READ_ONCE(iocg->abs_vdebt); |
7caa4715 TH |
1544 | struct ioc_now now; |
1545 | unsigned long flags; | |
1546 | ||
1547 | ioc_now(iocg->ioc, &now); | |
1548 | ||
da437b95 TH |
1549 | iocg_lock(iocg, pay_debt, &flags); |
1550 | iocg_kick_waitq(iocg, pay_debt, &now); | |
1551 | iocg_unlock(iocg, pay_debt, &flags); | |
7caa4715 TH |
1552 | |
1553 | return HRTIMER_NORESTART; | |
1554 | } | |
1555 | ||
7caa4715 TH |
1556 | static void ioc_lat_stat(struct ioc *ioc, u32 *missed_ppm_ar, u32 *rq_wait_pct_p) |
1557 | { | |
1558 | u32 nr_met[2] = { }; | |
1559 | u32 nr_missed[2] = { }; | |
1560 | u64 rq_wait_ns = 0; | |
1561 | int cpu, rw; | |
1562 | ||
1563 | for_each_online_cpu(cpu) { | |
1564 | struct ioc_pcpu_stat *stat = per_cpu_ptr(ioc->pcpu_stat, cpu); | |
1565 | u64 this_rq_wait_ns; | |
1566 | ||
1567 | for (rw = READ; rw <= WRITE; rw++) { | |
5e124f74 TH |
1568 | u32 this_met = local_read(&stat->missed[rw].nr_met); |
1569 | u32 this_missed = local_read(&stat->missed[rw].nr_missed); | |
7caa4715 TH |
1570 | |
1571 | nr_met[rw] += this_met - stat->missed[rw].last_met; | |
1572 | nr_missed[rw] += this_missed - stat->missed[rw].last_missed; | |
1573 | stat->missed[rw].last_met = this_met; | |
1574 | stat->missed[rw].last_missed = this_missed; | |
1575 | } | |
1576 | ||
5e124f74 | 1577 | this_rq_wait_ns = local64_read(&stat->rq_wait_ns); |
7caa4715 TH |
1578 | rq_wait_ns += this_rq_wait_ns - stat->last_rq_wait_ns; |
1579 | stat->last_rq_wait_ns = this_rq_wait_ns; | |
1580 | } | |
1581 | ||
1582 | for (rw = READ; rw <= WRITE; rw++) { | |
1583 | if (nr_met[rw] + nr_missed[rw]) | |
1584 | missed_ppm_ar[rw] = | |
1585 | DIV64_U64_ROUND_UP((u64)nr_missed[rw] * MILLION, | |
1586 | nr_met[rw] + nr_missed[rw]); | |
1587 | else | |
1588 | missed_ppm_ar[rw] = 0; | |
1589 | } | |
1590 | ||
1591 | *rq_wait_pct_p = div64_u64(rq_wait_ns * 100, | |
1592 | ioc->period_us * NSEC_PER_USEC); | |
1593 | } | |
1594 | ||
1595 | /* was iocg idle this period? */ | |
1596 | static bool iocg_is_idle(struct ioc_gq *iocg) | |
1597 | { | |
1598 | struct ioc *ioc = iocg->ioc; | |
1599 | ||
1600 | /* did something get issued this period? */ | |
1601 | if (atomic64_read(&iocg->active_period) == | |
1602 | atomic64_read(&ioc->cur_period)) | |
1603 | return false; | |
1604 | ||
1605 | /* is something in flight? */ | |
dcd6589b | 1606 | if (atomic64_read(&iocg->done_vtime) != atomic64_read(&iocg->vtime)) |
7caa4715 TH |
1607 | return false; |
1608 | ||
1609 | return true; | |
1610 | } | |
1611 | ||
97eb1975 TH |
1612 | /* |
1613 | * Call this function on the target leaf @iocg's to build pre-order traversal | |
1614 | * list of all the ancestors in @inner_walk. The inner nodes are linked through | |
1615 | * ->walk_list and the caller is responsible for dissolving the list after use. | |
1616 | */ | |
1617 | static void iocg_build_inner_walk(struct ioc_gq *iocg, | |
1618 | struct list_head *inner_walk) | |
1619 | { | |
1620 | int lvl; | |
1621 | ||
1622 | WARN_ON_ONCE(!list_empty(&iocg->walk_list)); | |
1623 | ||
1624 | /* find the first ancestor which hasn't been visited yet */ | |
1625 | for (lvl = iocg->level - 1; lvl >= 0; lvl--) { | |
1626 | if (!list_empty(&iocg->ancestors[lvl]->walk_list)) | |
1627 | break; | |
1628 | } | |
1629 | ||
1630 | /* walk down and visit the inner nodes to get pre-order traversal */ | |
1631 | while (++lvl <= iocg->level - 1) { | |
1632 | struct ioc_gq *inner = iocg->ancestors[lvl]; | |
1633 | ||
1634 | /* record traversal order */ | |
1635 | list_add_tail(&inner->walk_list, inner_walk); | |
1636 | } | |
1637 | } | |
1638 | ||
2a371f7d CZ |
1639 | /* propagate the deltas to the parent */ |
1640 | static void iocg_flush_stat_upward(struct ioc_gq *iocg) | |
1641 | { | |
1642 | if (iocg->level > 0) { | |
1643 | struct iocg_stat *parent_stat = | |
1644 | &iocg->ancestors[iocg->level - 1]->stat; | |
1645 | ||
1646 | parent_stat->usage_us += | |
1647 | iocg->stat.usage_us - iocg->last_stat.usage_us; | |
1648 | parent_stat->wait_us += | |
1649 | iocg->stat.wait_us - iocg->last_stat.wait_us; | |
1650 | parent_stat->indebt_us += | |
1651 | iocg->stat.indebt_us - iocg->last_stat.indebt_us; | |
1652 | parent_stat->indelay_us += | |
1653 | iocg->stat.indelay_us - iocg->last_stat.indelay_us; | |
1654 | } | |
1655 | ||
1656 | iocg->last_stat = iocg->stat; | |
1657 | } | |
1658 | ||
97eb1975 | 1659 | /* collect per-cpu counters and propagate the deltas to the parent */ |
2a371f7d | 1660 | static void iocg_flush_stat_leaf(struct ioc_gq *iocg, struct ioc_now *now) |
97eb1975 | 1661 | { |
ac33e91e | 1662 | struct ioc *ioc = iocg->ioc; |
97eb1975 TH |
1663 | u64 abs_vusage = 0; |
1664 | u64 vusage_delta; | |
1665 | int cpu; | |
1666 | ||
1667 | lockdep_assert_held(&iocg->ioc->lock); | |
1668 | ||
1669 | /* collect per-cpu counters */ | |
1670 | for_each_possible_cpu(cpu) { | |
1671 | abs_vusage += local64_read( | |
1672 | per_cpu_ptr(&iocg->pcpu_stat->abs_vusage, cpu)); | |
1673 | } | |
1674 | vusage_delta = abs_vusage - iocg->last_stat_abs_vusage; | |
1675 | iocg->last_stat_abs_vusage = abs_vusage; | |
1676 | ||
ac33e91e | 1677 | iocg->usage_delta_us = div64_u64(vusage_delta, ioc->vtime_base_rate); |
2a371f7d | 1678 | iocg->stat.usage_us += iocg->usage_delta_us; |
97eb1975 | 1679 | |
2a371f7d | 1680 | iocg_flush_stat_upward(iocg); |
97eb1975 TH |
1681 | } |
1682 | ||
1683 | /* get stat counters ready for reading on all active iocgs */ | |
1684 | static void iocg_flush_stat(struct list_head *target_iocgs, struct ioc_now *now) | |
1685 | { | |
1686 | LIST_HEAD(inner_walk); | |
1687 | struct ioc_gq *iocg, *tiocg; | |
1688 | ||
1689 | /* flush leaves and build inner node walk list */ | |
1690 | list_for_each_entry(iocg, target_iocgs, active_list) { | |
2a371f7d | 1691 | iocg_flush_stat_leaf(iocg, now); |
97eb1975 TH |
1692 | iocg_build_inner_walk(iocg, &inner_walk); |
1693 | } | |
1694 | ||
1695 | /* keep flushing upwards by walking the inner list backwards */ | |
1696 | list_for_each_entry_safe_reverse(iocg, tiocg, &inner_walk, walk_list) { | |
2a371f7d | 1697 | iocg_flush_stat_upward(iocg); |
97eb1975 TH |
1698 | list_del_init(&iocg->walk_list); |
1699 | } | |
1700 | } | |
1701 | ||
93f7d2db TH |
1702 | /* |
1703 | * Determine what @iocg's hweight_inuse should be after donating unused | |
1704 | * capacity. @hwm is the upper bound and used to signal no donation. This | |
1705 | * function also throws away @iocg's excess budget. | |
1706 | */ | |
ac33e91e TH |
1707 | static u32 hweight_after_donation(struct ioc_gq *iocg, u32 old_hwi, u32 hwm, |
1708 | u32 usage, struct ioc_now *now) | |
7caa4715 | 1709 | { |
93f7d2db TH |
1710 | struct ioc *ioc = iocg->ioc; |
1711 | u64 vtime = atomic64_read(&iocg->vtime); | |
f1de2439 | 1712 | s64 excess, delta, target, new_hwi; |
93f7d2db | 1713 | |
c421a3eb TH |
1714 | /* debt handling owns inuse for debtors */ |
1715 | if (iocg->abs_vdebt) | |
1716 | return 1; | |
1717 | ||
93f7d2db TH |
1718 | /* see whether minimum margin requirement is met */ |
1719 | if (waitqueue_active(&iocg->waitq) || | |
1720 | time_after64(vtime, now->vnow - ioc->margins.min)) | |
1721 | return hwm; | |
1722 | ||
ac33e91e TH |
1723 | /* throw away excess above target */ |
1724 | excess = now->vnow - vtime - ioc->margins.target; | |
93f7d2db TH |
1725 | if (excess > 0) { |
1726 | atomic64_add(excess, &iocg->vtime); | |
1727 | atomic64_add(excess, &iocg->done_vtime); | |
1728 | vtime += excess; | |
ac33e91e | 1729 | ioc->vtime_err -= div64_u64(excess * old_hwi, WEIGHT_ONE); |
93f7d2db TH |
1730 | } |
1731 | ||
f1de2439 TH |
1732 | /* |
1733 | * Let's say the distance between iocg's and device's vtimes as a | |
1734 | * fraction of period duration is delta. Assuming that the iocg will | |
1735 | * consume the usage determined above, we want to determine new_hwi so | |
1736 | * that delta equals MARGIN_TARGET at the end of the next period. | |
1737 | * | |
1738 | * We need to execute usage worth of IOs while spending the sum of the | |
1739 | * new budget (1 - MARGIN_TARGET) and the leftover from the last period | |
1740 | * (delta): | |
1741 | * | |
1742 | * usage = (1 - MARGIN_TARGET + delta) * new_hwi | |
1743 | * | |
1744 | * Therefore, the new_hwi is: | |
1745 | * | |
1746 | * new_hwi = usage / (1 - MARGIN_TARGET + delta) | |
1747 | */ | |
1748 | delta = div64_s64(WEIGHT_ONE * (now->vnow - vtime), | |
1749 | now->vnow - ioc->period_at_vtime); | |
1750 | target = WEIGHT_ONE * MARGIN_TARGET_PCT / 100; | |
1751 | new_hwi = div64_s64(WEIGHT_ONE * usage, WEIGHT_ONE - target + delta); | |
7caa4715 | 1752 | |
f1de2439 | 1753 | return clamp_t(s64, new_hwi, 1, hwm); |
7caa4715 TH |
1754 | } |
1755 | ||
e08d02aa TH |
1756 | /* |
1757 | * For work-conservation, an iocg which isn't using all of its share should | |
1758 | * donate the leftover to other iocgs. There are two ways to achieve this - 1. | |
1759 | * bumping up vrate accordingly 2. lowering the donating iocg's inuse weight. | |
1760 | * | |
1761 | * #1 is mathematically simpler but has the drawback of requiring synchronous | |
1762 | * global hweight_inuse updates when idle iocg's get activated or inuse weights | |
1763 | * change due to donation snapbacks as it has the possibility of grossly | |
1764 | * overshooting what's allowed by the model and vrate. | |
1765 | * | |
1766 | * #2 is inherently safe with local operations. The donating iocg can easily | |
1767 | * snap back to higher weights when needed without worrying about impacts on | |
1768 | * other nodes as the impacts will be inherently correct. This also makes idle | |
1769 | * iocg activations safe. The only effect activations have is decreasing | |
1770 | * hweight_inuse of others, the right solution to which is for those iocgs to | |
1771 | * snap back to higher weights. | |
1772 | * | |
1773 | * So, we go with #2. The challenge is calculating how each donating iocg's | |
1774 | * inuse should be adjusted to achieve the target donation amounts. This is done | |
1775 | * using Andy's method described in the following pdf. | |
1776 | * | |
1777 | * https://drive.google.com/file/d/1PsJwxPFtjUnwOY1QJ5AeICCcsL7BM3bo | |
1778 | * | |
1779 | * Given the weights and target after-donation hweight_inuse values, Andy's | |
1780 | * method determines how the proportional distribution should look like at each | |
1781 | * sibling level to maintain the relative relationship between all non-donating | |
1782 | * pairs. To roughly summarize, it divides the tree into donating and | |
1783 | * non-donating parts, calculates global donation rate which is used to | |
1784 | * determine the target hweight_inuse for each node, and then derives per-level | |
1785 | * proportions. | |
1786 | * | |
1787 | * The following pdf shows that global distribution calculated this way can be | |
1788 | * achieved by scaling inuse weights of donating leaves and propagating the | |
1789 | * adjustments upwards proportionally. | |
1790 | * | |
1791 | * https://drive.google.com/file/d/1vONz1-fzVO7oY5DXXsLjSxEtYYQbOvsE | |
1792 | * | |
1793 | * Combining the above two, we can determine how each leaf iocg's inuse should | |
1794 | * be adjusted to achieve the target donation. | |
1795 | * | |
1796 | * https://drive.google.com/file/d/1WcrltBOSPN0qXVdBgnKm4mdp9FhuEFQN | |
1797 | * | |
1798 | * The inline comments use symbols from the last pdf. | |
1799 | * | |
1800 | * b is the sum of the absolute budgets in the subtree. 1 for the root node. | |
1801 | * f is the sum of the absolute budgets of non-donating nodes in the subtree. | |
1802 | * t is the sum of the absolute budgets of donating nodes in the subtree. | |
1803 | * w is the weight of the node. w = w_f + w_t | |
1804 | * w_f is the non-donating portion of w. w_f = w * f / b | |
1805 | * w_b is the donating portion of w. w_t = w * t / b | |
1806 | * s is the sum of all sibling weights. s = Sum(w) for siblings | |
1807 | * s_f and s_t are the non-donating and donating portions of s. | |
1808 | * | |
1809 | * Subscript p denotes the parent's counterpart and ' the adjusted value - e.g. | |
1810 | * w_pt is the donating portion of the parent's weight and w'_pt the same value | |
1811 | * after adjustments. Subscript r denotes the root node's values. | |
1812 | */ | |
93f7d2db TH |
1813 | static void transfer_surpluses(struct list_head *surpluses, struct ioc_now *now) |
1814 | { | |
e08d02aa TH |
1815 | LIST_HEAD(over_hwa); |
1816 | LIST_HEAD(inner_walk); | |
1817 | struct ioc_gq *iocg, *tiocg, *root_iocg; | |
1818 | u32 after_sum, over_sum, over_target, gamma; | |
93f7d2db | 1819 | |
e08d02aa TH |
1820 | /* |
1821 | * It's pretty unlikely but possible for the total sum of | |
1822 | * hweight_after_donation's to be higher than WEIGHT_ONE, which will | |
1823 | * confuse the following calculations. If such condition is detected, | |
1824 | * scale down everyone over its full share equally to keep the sum below | |
1825 | * WEIGHT_ONE. | |
1826 | */ | |
1827 | after_sum = 0; | |
1828 | over_sum = 0; | |
93f7d2db | 1829 | list_for_each_entry(iocg, surpluses, surplus_list) { |
e08d02aa | 1830 | u32 hwa; |
93f7d2db | 1831 | |
e08d02aa TH |
1832 | current_hweight(iocg, &hwa, NULL); |
1833 | after_sum += iocg->hweight_after_donation; | |
93f7d2db | 1834 | |
e08d02aa TH |
1835 | if (iocg->hweight_after_donation > hwa) { |
1836 | over_sum += iocg->hweight_after_donation; | |
1837 | list_add(&iocg->walk_list, &over_hwa); | |
1838 | } | |
93f7d2db | 1839 | } |
e08d02aa TH |
1840 | |
1841 | if (after_sum >= WEIGHT_ONE) { | |
1842 | /* | |
1843 | * The delta should be deducted from the over_sum, calculate | |
1844 | * target over_sum value. | |
1845 | */ | |
1846 | u32 over_delta = after_sum - (WEIGHT_ONE - 1); | |
1847 | WARN_ON_ONCE(over_sum <= over_delta); | |
1848 | over_target = over_sum - over_delta; | |
1849 | } else { | |
1850 | over_target = 0; | |
1851 | } | |
1852 | ||
1853 | list_for_each_entry_safe(iocg, tiocg, &over_hwa, walk_list) { | |
1854 | if (over_target) | |
1855 | iocg->hweight_after_donation = | |
1856 | div_u64((u64)iocg->hweight_after_donation * | |
1857 | over_target, over_sum); | |
1858 | list_del_init(&iocg->walk_list); | |
1859 | } | |
1860 | ||
1861 | /* | |
1862 | * Build pre-order inner node walk list and prepare for donation | |
1863 | * adjustment calculations. | |
1864 | */ | |
1865 | list_for_each_entry(iocg, surpluses, surplus_list) { | |
1866 | iocg_build_inner_walk(iocg, &inner_walk); | |
1867 | } | |
1868 | ||
1869 | root_iocg = list_first_entry(&inner_walk, struct ioc_gq, walk_list); | |
1870 | WARN_ON_ONCE(root_iocg->level > 0); | |
1871 | ||
1872 | list_for_each_entry(iocg, &inner_walk, walk_list) { | |
1873 | iocg->child_adjusted_sum = 0; | |
1874 | iocg->hweight_donating = 0; | |
1875 | iocg->hweight_after_donation = 0; | |
1876 | } | |
1877 | ||
1878 | /* | |
1879 | * Propagate the donating budget (b_t) and after donation budget (b'_t) | |
1880 | * up the hierarchy. | |
1881 | */ | |
1882 | list_for_each_entry(iocg, surpluses, surplus_list) { | |
1883 | struct ioc_gq *parent = iocg->ancestors[iocg->level - 1]; | |
1884 | ||
1885 | parent->hweight_donating += iocg->hweight_donating; | |
1886 | parent->hweight_after_donation += iocg->hweight_after_donation; | |
1887 | } | |
1888 | ||
1889 | list_for_each_entry_reverse(iocg, &inner_walk, walk_list) { | |
1890 | if (iocg->level > 0) { | |
1891 | struct ioc_gq *parent = iocg->ancestors[iocg->level - 1]; | |
1892 | ||
1893 | parent->hweight_donating += iocg->hweight_donating; | |
1894 | parent->hweight_after_donation += iocg->hweight_after_donation; | |
1895 | } | |
1896 | } | |
1897 | ||
1898 | /* | |
1899 | * Calculate inner hwa's (b) and make sure the donation values are | |
1900 | * within the accepted ranges as we're doing low res calculations with | |
1901 | * roundups. | |
1902 | */ | |
1903 | list_for_each_entry(iocg, &inner_walk, walk_list) { | |
1904 | if (iocg->level) { | |
1905 | struct ioc_gq *parent = iocg->ancestors[iocg->level - 1]; | |
1906 | ||
1907 | iocg->hweight_active = DIV64_U64_ROUND_UP( | |
1908 | (u64)parent->hweight_active * iocg->active, | |
1909 | parent->child_active_sum); | |
1910 | ||
1911 | } | |
1912 | ||
1913 | iocg->hweight_donating = min(iocg->hweight_donating, | |
1914 | iocg->hweight_active); | |
1915 | iocg->hweight_after_donation = min(iocg->hweight_after_donation, | |
1916 | iocg->hweight_donating - 1); | |
1917 | if (WARN_ON_ONCE(iocg->hweight_active <= 1 || | |
1918 | iocg->hweight_donating <= 1 || | |
1919 | iocg->hweight_after_donation == 0)) { | |
1920 | pr_warn("iocg: invalid donation weights in "); | |
1921 | pr_cont_cgroup_path(iocg_to_blkg(iocg)->blkcg->css.cgroup); | |
1922 | pr_cont(": active=%u donating=%u after=%u\n", | |
1923 | iocg->hweight_active, iocg->hweight_donating, | |
1924 | iocg->hweight_after_donation); | |
1925 | } | |
1926 | } | |
1927 | ||
1928 | /* | |
1929 | * Calculate the global donation rate (gamma) - the rate to adjust | |
769b628d TH |
1930 | * non-donating budgets by. |
1931 | * | |
1932 | * No need to use 64bit multiplication here as the first operand is | |
1933 | * guaranteed to be smaller than WEIGHT_ONE (1<<16). | |
1934 | * | |
1935 | * We know that there are beneficiary nodes and the sum of the donating | |
1936 | * hweights can't be whole; however, due to the round-ups during hweight | |
1937 | * calculations, root_iocg->hweight_donating might still end up equal to | |
1938 | * or greater than whole. Limit the range when calculating the divider. | |
e08d02aa TH |
1939 | * |
1940 | * gamma = (1 - t_r') / (1 - t_r) | |
1941 | */ | |
1942 | gamma = DIV_ROUND_UP( | |
1943 | (WEIGHT_ONE - root_iocg->hweight_after_donation) * WEIGHT_ONE, | |
769b628d | 1944 | WEIGHT_ONE - min_t(u32, root_iocg->hweight_donating, WEIGHT_ONE - 1)); |
e08d02aa TH |
1945 | |
1946 | /* | |
1947 | * Calculate adjusted hwi, child_adjusted_sum and inuse for the inner | |
1948 | * nodes. | |
1949 | */ | |
1950 | list_for_each_entry(iocg, &inner_walk, walk_list) { | |
1951 | struct ioc_gq *parent; | |
1952 | u32 inuse, wpt, wptp; | |
1953 | u64 st, sf; | |
1954 | ||
1955 | if (iocg->level == 0) { | |
1956 | /* adjusted weight sum for 1st level: s' = s * b_pf / b'_pf */ | |
1957 | iocg->child_adjusted_sum = DIV64_U64_ROUND_UP( | |
1958 | iocg->child_active_sum * (WEIGHT_ONE - iocg->hweight_donating), | |
1959 | WEIGHT_ONE - iocg->hweight_after_donation); | |
1960 | continue; | |
1961 | } | |
1962 | ||
1963 | parent = iocg->ancestors[iocg->level - 1]; | |
1964 | ||
1965 | /* b' = gamma * b_f + b_t' */ | |
1966 | iocg->hweight_inuse = DIV64_U64_ROUND_UP( | |
1967 | (u64)gamma * (iocg->hweight_active - iocg->hweight_donating), | |
1968 | WEIGHT_ONE) + iocg->hweight_after_donation; | |
1969 | ||
1970 | /* w' = s' * b' / b'_p */ | |
1971 | inuse = DIV64_U64_ROUND_UP( | |
1972 | (u64)parent->child_adjusted_sum * iocg->hweight_inuse, | |
1973 | parent->hweight_inuse); | |
1974 | ||
1975 | /* adjusted weight sum for children: s' = s_f + s_t * w'_pt / w_pt */ | |
1976 | st = DIV64_U64_ROUND_UP( | |
1977 | iocg->child_active_sum * iocg->hweight_donating, | |
1978 | iocg->hweight_active); | |
1979 | sf = iocg->child_active_sum - st; | |
1980 | wpt = DIV64_U64_ROUND_UP( | |
1981 | (u64)iocg->active * iocg->hweight_donating, | |
1982 | iocg->hweight_active); | |
1983 | wptp = DIV64_U64_ROUND_UP( | |
1984 | (u64)inuse * iocg->hweight_after_donation, | |
1985 | iocg->hweight_inuse); | |
1986 | ||
1987 | iocg->child_adjusted_sum = sf + DIV64_U64_ROUND_UP(st * wptp, wpt); | |
1988 | } | |
1989 | ||
1990 | /* | |
1991 | * All inner nodes now have ->hweight_inuse and ->child_adjusted_sum and | |
1992 | * we can finally determine leaf adjustments. | |
1993 | */ | |
1994 | list_for_each_entry(iocg, surpluses, surplus_list) { | |
1995 | struct ioc_gq *parent = iocg->ancestors[iocg->level - 1]; | |
1996 | u32 inuse; | |
1997 | ||
c421a3eb TH |
1998 | /* |
1999 | * In-debt iocgs participated in the donation calculation with | |
2000 | * the minimum target hweight_inuse. Configuring inuse | |
2001 | * accordingly would work fine but debt handling expects | |
2002 | * @iocg->inuse stay at the minimum and we don't wanna | |
2003 | * interfere. | |
2004 | */ | |
2005 | if (iocg->abs_vdebt) { | |
2006 | WARN_ON_ONCE(iocg->inuse > 1); | |
2007 | continue; | |
2008 | } | |
2009 | ||
e08d02aa TH |
2010 | /* w' = s' * b' / b'_p, note that b' == b'_t for donating leaves */ |
2011 | inuse = DIV64_U64_ROUND_UP( | |
2012 | parent->child_adjusted_sum * iocg->hweight_after_donation, | |
2013 | parent->hweight_inuse); | |
04603755 TH |
2014 | |
2015 | TRACE_IOCG_PATH(inuse_transfer, iocg, now, | |
2016 | iocg->inuse, inuse, | |
2017 | iocg->hweight_inuse, | |
2018 | iocg->hweight_after_donation); | |
2019 | ||
b0853ab4 | 2020 | __propagate_weights(iocg, iocg->active, inuse, true, now); |
e08d02aa TH |
2021 | } |
2022 | ||
2023 | /* walk list should be dissolved after use */ | |
2024 | list_for_each_entry_safe(iocg, tiocg, &inner_walk, walk_list) | |
2025 | list_del_init(&iocg->walk_list); | |
93f7d2db TH |
2026 | } |
2027 | ||
ab8df828 TH |
2028 | /* |
2029 | * A low weight iocg can amass a large amount of debt, for example, when | |
2030 | * anonymous memory gets reclaimed aggressively. If the system has a lot of | |
2031 | * memory paired with a slow IO device, the debt can span multiple seconds or | |
2032 | * more. If there are no other subsequent IO issuers, the in-debt iocg may end | |
2033 | * up blocked paying its debt while the IO device is idle. | |
2034 | * | |
2035 | * The following protects against such cases. If the device has been | |
d9517841 TH |
2036 | * sufficiently idle for a while, the debts are halved and delays are |
2037 | * recalculated. | |
ab8df828 TH |
2038 | */ |
2039 | static void ioc_forgive_debts(struct ioc *ioc, u64 usage_us_sum, int nr_debtors, | |
33a1fe6d | 2040 | struct ioc_now *now) |
ab8df828 | 2041 | { |
c7af2a00 TH |
2042 | struct ioc_gq *iocg; |
2043 | u64 dur, usage_pct, nr_cycles; | |
2044 | ||
2045 | /* if no debtor, reset the cycle */ | |
2046 | if (!nr_debtors) { | |
2047 | ioc->dfgv_period_at = now->now; | |
2048 | ioc->dfgv_period_rem = 0; | |
2049 | ioc->dfgv_usage_us_sum = 0; | |
2050 | return; | |
2051 | } | |
2052 | ||
2053 | /* | |
2054 | * Debtors can pass through a lot of writes choking the device and we | |
2055 | * don't want to be forgiving debts while the device is struggling from | |
2056 | * write bursts. If we're missing latency targets, consider the device | |
2057 | * fully utilized. | |
2058 | */ | |
2059 | if (ioc->busy_level > 0) | |
2060 | usage_us_sum = max_t(u64, usage_us_sum, ioc->period_us); | |
2061 | ||
2062 | ioc->dfgv_usage_us_sum += usage_us_sum; | |
2063 | if (time_before64(now->now, ioc->dfgv_period_at + DFGV_PERIOD)) | |
2064 | return; | |
2065 | ||
2066 | /* | |
2067 | * At least DFGV_PERIOD has passed since the last period. Calculate the | |
2068 | * average usage and reset the period counters. | |
2069 | */ | |
2070 | dur = now->now - ioc->dfgv_period_at; | |
2071 | usage_pct = div64_u64(100 * ioc->dfgv_usage_us_sum, dur); | |
2072 | ||
2073 | ioc->dfgv_period_at = now->now; | |
2074 | ioc->dfgv_usage_us_sum = 0; | |
2075 | ||
2076 | /* if was too busy, reset everything */ | |
2077 | if (usage_pct > DFGV_USAGE_PCT) { | |
2078 | ioc->dfgv_period_rem = 0; | |
2079 | return; | |
2080 | } | |
2081 | ||
2082 | /* | |
2083 | * Usage is lower than threshold. Let's forgive some debts. Debt | |
2084 | * forgiveness runs off of the usual ioc timer but its period usually | |
2085 | * doesn't match ioc's. Compensate the difference by performing the | |
2086 | * reduction as many times as would fit in the duration since the last | |
2087 | * run and carrying over the left-over duration in @ioc->dfgv_period_rem | |
2088 | * - if ioc period is 75% of DFGV_PERIOD, one out of three consecutive | |
2089 | * reductions is doubled. | |
2090 | */ | |
2091 | nr_cycles = dur + ioc->dfgv_period_rem; | |
2092 | ioc->dfgv_period_rem = do_div(nr_cycles, DFGV_PERIOD); | |
2093 | ||
2094 | list_for_each_entry(iocg, &ioc->active_iocgs, active_list) { | |
c5a6561b TH |
2095 | u64 __maybe_unused old_debt, __maybe_unused old_delay; |
2096 | ||
bec02dbb | 2097 | if (!iocg->abs_vdebt && !iocg->delay) |
c7af2a00 | 2098 | continue; |
c5a6561b | 2099 | |
c7af2a00 | 2100 | spin_lock(&iocg->waitq.lock); |
c5a6561b TH |
2101 | |
2102 | old_debt = iocg->abs_vdebt; | |
2103 | old_delay = iocg->delay; | |
2104 | ||
bec02dbb TH |
2105 | if (iocg->abs_vdebt) |
2106 | iocg->abs_vdebt = iocg->abs_vdebt >> nr_cycles ?: 1; | |
2107 | if (iocg->delay) | |
2108 | iocg->delay = iocg->delay >> nr_cycles ?: 1; | |
2109 | ||
c7af2a00 | 2110 | iocg_kick_waitq(iocg, true, now); |
c5a6561b TH |
2111 | |
2112 | TRACE_IOCG_PATH(iocg_forgive_debt, iocg, now, usage_pct, | |
2113 | old_debt, iocg->abs_vdebt, | |
2114 | old_delay, iocg->delay); | |
2115 | ||
c7af2a00 | 2116 | spin_unlock(&iocg->waitq.lock); |
ab8df828 TH |
2117 | } |
2118 | } | |
2119 | ||
2474787a BW |
2120 | /* |
2121 | * Check the active iocgs' state to avoid oversleeping and deactive | |
2122 | * idle iocgs. | |
2123 | * | |
2124 | * Since waiters determine the sleep durations based on the vrate | |
2125 | * they saw at the time of sleep, if vrate has increased, some | |
2126 | * waiters could be sleeping for too long. Wake up tardy waiters | |
2127 | * which should have woken up in the last period and expire idle | |
2128 | * iocgs. | |
2129 | */ | |
2130 | static int ioc_check_iocgs(struct ioc *ioc, struct ioc_now *now) | |
7caa4715 | 2131 | { |
2474787a | 2132 | int nr_debtors = 0; |
7caa4715 | 2133 | struct ioc_gq *iocg, *tiocg; |
7caa4715 | 2134 | |
7caa4715 | 2135 | list_for_each_entry_safe(iocg, tiocg, &ioc->active_iocgs, active_list) { |
d9012a59 | 2136 | if (!waitqueue_active(&iocg->waitq) && !iocg->abs_vdebt && |
5160a5a5 | 2137 | !iocg->delay && !iocg_is_idle(iocg)) |
7caa4715 TH |
2138 | continue; |
2139 | ||
2140 | spin_lock(&iocg->waitq.lock); | |
2141 | ||
f0bf84a5 TH |
2142 | /* flush wait and indebt stat deltas */ |
2143 | if (iocg->wait_since) { | |
2a371f7d | 2144 | iocg->stat.wait_us += now->now - iocg->wait_since; |
2474787a | 2145 | iocg->wait_since = now->now; |
f0bf84a5 TH |
2146 | } |
2147 | if (iocg->indebt_since) { | |
2a371f7d | 2148 | iocg->stat.indebt_us += |
2474787a BW |
2149 | now->now - iocg->indebt_since; |
2150 | iocg->indebt_since = now->now; | |
f0bf84a5 TH |
2151 | } |
2152 | if (iocg->indelay_since) { | |
2a371f7d | 2153 | iocg->stat.indelay_us += |
2474787a BW |
2154 | now->now - iocg->indelay_since; |
2155 | iocg->indelay_since = now->now; | |
f0bf84a5 TH |
2156 | } |
2157 | ||
5160a5a5 TH |
2158 | if (waitqueue_active(&iocg->waitq) || iocg->abs_vdebt || |
2159 | iocg->delay) { | |
7caa4715 | 2160 | /* might be oversleeping vtime / hweight changes, kick */ |
2474787a | 2161 | iocg_kick_waitq(iocg, true, now); |
bec02dbb | 2162 | if (iocg->abs_vdebt || iocg->delay) |
dda1315f | 2163 | nr_debtors++; |
7caa4715 TH |
2164 | } else if (iocg_is_idle(iocg)) { |
2165 | /* no waiter and idle, deactivate */ | |
ac33e91e TH |
2166 | u64 vtime = atomic64_read(&iocg->vtime); |
2167 | s64 excess; | |
2168 | ||
2169 | /* | |
2170 | * @iocg has been inactive for a full duration and will | |
2171 | * have a high budget. Account anything above target as | |
2172 | * error and throw away. On reactivation, it'll start | |
2173 | * with the target budget. | |
2174 | */ | |
2474787a | 2175 | excess = now->vnow - vtime - ioc->margins.target; |
ac33e91e TH |
2176 | if (excess > 0) { |
2177 | u32 old_hwi; | |
2178 | ||
2179 | current_hweight(iocg, NULL, &old_hwi); | |
2180 | ioc->vtime_err -= div64_u64(excess * old_hwi, | |
2181 | WEIGHT_ONE); | |
2182 | } | |
2183 | ||
76efc1c7 BW |
2184 | TRACE_IOCG_PATH(iocg_idle, iocg, now, |
2185 | atomic64_read(&iocg->active_period), | |
2186 | atomic64_read(&ioc->cur_period), vtime); | |
2474787a | 2187 | __propagate_weights(iocg, 0, 0, false, now); |
7caa4715 TH |
2188 | list_del_init(&iocg->active_list); |
2189 | } | |
2190 | ||
2191 | spin_unlock(&iocg->waitq.lock); | |
2192 | } | |
2474787a | 2193 | |
00410f1b | 2194 | commit_weights(ioc); |
2474787a BW |
2195 | return nr_debtors; |
2196 | } | |
2197 | ||
2198 | static void ioc_timer_fn(struct timer_list *timer) | |
2199 | { | |
2200 | struct ioc *ioc = container_of(timer, struct ioc, timer); | |
2201 | struct ioc_gq *iocg, *tiocg; | |
2202 | struct ioc_now now; | |
2203 | LIST_HEAD(surpluses); | |
2204 | int nr_debtors, nr_shortages = 0, nr_lagging = 0; | |
2205 | u64 usage_us_sum = 0; | |
2206 | u32 ppm_rthr = MILLION - ioc->params.qos[QOS_RPPM]; | |
2207 | u32 ppm_wthr = MILLION - ioc->params.qos[QOS_WPPM]; | |
2208 | u32 missed_ppm[2], rq_wait_pct; | |
2209 | u64 period_vtime; | |
2210 | int prev_busy_level; | |
2211 | ||
2212 | /* how were the latencies during the period? */ | |
2213 | ioc_lat_stat(ioc, missed_ppm, &rq_wait_pct); | |
2214 | ||
2215 | /* take care of active iocgs */ | |
2216 | spin_lock_irq(&ioc->lock); | |
2217 | ||
2218 | ioc_now(ioc, &now); | |
2219 | ||
2220 | period_vtime = now.vnow - ioc->period_at_vtime; | |
2221 | if (WARN_ON_ONCE(!period_vtime)) { | |
2222 | spin_unlock_irq(&ioc->lock); | |
2223 | return; | |
2224 | } | |
2225 | ||
2226 | nr_debtors = ioc_check_iocgs(ioc, &now); | |
7caa4715 | 2227 | |
f0bf84a5 TH |
2228 | /* |
2229 | * Wait and indebt stat are flushed above and the donation calculation | |
2230 | * below needs updated usage stat. Let's bring stat up-to-date. | |
2231 | */ | |
2232 | iocg_flush_stat(&ioc->active_iocgs, &now); | |
2233 | ||
f1de2439 | 2234 | /* calc usage and see whether some weights need to be moved around */ |
7caa4715 | 2235 | list_for_each_entry(iocg, &ioc->active_iocgs, active_list) { |
c09245f6 BW |
2236 | u64 vdone, vtime, usage_us; |
2237 | u32 hw_active, hw_inuse; | |
7caa4715 TH |
2238 | |
2239 | /* | |
2240 | * Collect unused and wind vtime closer to vnow to prevent | |
2241 | * iocgs from accumulating a large amount of budget. | |
2242 | */ | |
2243 | vdone = atomic64_read(&iocg->done_vtime); | |
2244 | vtime = atomic64_read(&iocg->vtime); | |
2245 | current_hweight(iocg, &hw_active, &hw_inuse); | |
2246 | ||
2247 | /* | |
2248 | * Latency QoS detection doesn't account for IOs which are | |
2249 | * in-flight for longer than a period. Detect them by | |
2250 | * comparing vdone against period start. If lagging behind | |
2251 | * IOs from past periods, don't increase vrate. | |
2252 | */ | |
7cd806a9 TH |
2253 | if ((ppm_rthr != MILLION || ppm_wthr != MILLION) && |
2254 | !atomic_read(&iocg_to_blkg(iocg)->use_delay) && | |
7caa4715 TH |
2255 | time_after64(vtime, vdone) && |
2256 | time_after64(vtime, now.vnow - | |
2257 | MAX_LAGGING_PERIODS * period_vtime) && | |
2258 | time_before64(vdone, now.vnow - period_vtime)) | |
2259 | nr_lagging++; | |
2260 | ||
7caa4715 | 2261 | /* |
f1de2439 TH |
2262 | * Determine absolute usage factoring in in-flight IOs to avoid |
2263 | * high-latency completions appearing as idle. | |
7caa4715 | 2264 | */ |
1aa50d02 | 2265 | usage_us = iocg->usage_delta_us; |
dda1315f | 2266 | usage_us_sum += usage_us; |
f1de2439 | 2267 | |
7caa4715 | 2268 | /* see whether there's surplus vtime */ |
8692d2db | 2269 | WARN_ON_ONCE(!list_empty(&iocg->surplus_list)); |
93f7d2db TH |
2270 | if (hw_inuse < hw_active || |
2271 | (!waitqueue_active(&iocg->waitq) && | |
f1de2439 | 2272 | time_before64(vtime, now.vnow - ioc->margins.low))) { |
c09245f6 BW |
2273 | u32 hwa, old_hwi, hwm, new_hwi, usage; |
2274 | u64 usage_dur; | |
2275 | ||
2276 | if (vdone != vtime) { | |
2277 | u64 inflight_us = DIV64_U64_ROUND_UP( | |
2278 | cost_to_abs_cost(vtime - vdone, hw_inuse), | |
2279 | ioc->vtime_base_rate); | |
2280 | ||
2281 | usage_us = max(usage_us, inflight_us); | |
2282 | } | |
2283 | ||
2284 | /* convert to hweight based usage ratio */ | |
2285 | if (time_after64(iocg->activated_at, ioc->period_at)) | |
2286 | usage_dur = max_t(u64, now.now - iocg->activated_at, 1); | |
2287 | else | |
2288 | usage_dur = max_t(u64, now.now - ioc->period_at, 1); | |
2289 | ||
2290 | usage = clamp_t(u32, | |
2291 | DIV64_U64_ROUND_UP(usage_us * WEIGHT_ONE, | |
2292 | usage_dur), | |
2293 | 1, WEIGHT_ONE); | |
93f7d2db TH |
2294 | |
2295 | /* | |
2296 | * Already donating or accumulated enough to start. | |
2297 | * Determine the donation amount. | |
2298 | */ | |
ac33e91e | 2299 | current_hweight(iocg, &hwa, &old_hwi); |
93f7d2db | 2300 | hwm = current_hweight_max(iocg); |
ac33e91e TH |
2301 | new_hwi = hweight_after_donation(iocg, old_hwi, hwm, |
2302 | usage, &now); | |
edaa2633 TH |
2303 | /* |
2304 | * Donation calculation assumes hweight_after_donation | |
2305 | * to be positive, a condition that a donor w/ hwa < 2 | |
2306 | * can't meet. Don't bother with donation if hwa is | |
2307 | * below 2. It's not gonna make a meaningful difference | |
2308 | * anyway. | |
2309 | */ | |
2310 | if (new_hwi < hwm && hwa >= 2) { | |
e08d02aa | 2311 | iocg->hweight_donating = hwa; |
93f7d2db | 2312 | iocg->hweight_after_donation = new_hwi; |
8692d2db | 2313 | list_add(&iocg->surplus_list, &surpluses); |
8c936f9e TH |
2314 | } else if (!iocg->abs_vdebt) { |
2315 | /* | |
2316 | * @iocg doesn't have enough to donate. Reset | |
2317 | * its inuse to active. | |
2318 | * | |
2319 | * Don't reset debtors as their inuse's are | |
2320 | * owned by debt handling. This shouldn't affect | |
2321 | * donation calculuation in any meaningful way | |
2322 | * as @iocg doesn't have a meaningful amount of | |
2323 | * share anyway. | |
2324 | */ | |
04603755 TH |
2325 | TRACE_IOCG_PATH(inuse_shortage, iocg, &now, |
2326 | iocg->inuse, iocg->active, | |
2327 | iocg->hweight_inuse, new_hwi); | |
2328 | ||
93f7d2db | 2329 | __propagate_weights(iocg, iocg->active, |
b0853ab4 | 2330 | iocg->active, true, &now); |
93f7d2db | 2331 | nr_shortages++; |
7caa4715 TH |
2332 | } |
2333 | } else { | |
93f7d2db | 2334 | /* genuinely short on vtime */ |
7caa4715 TH |
2335 | nr_shortages++; |
2336 | } | |
2337 | } | |
2338 | ||
93f7d2db TH |
2339 | if (!list_empty(&surpluses) && nr_shortages) |
2340 | transfer_surpluses(&surpluses, &now); | |
7caa4715 | 2341 | |
00410f1b | 2342 | commit_weights(ioc); |
7caa4715 | 2343 | |
8692d2db TH |
2344 | /* surplus list should be dissolved after use */ |
2345 | list_for_each_entry_safe(iocg, tiocg, &surpluses, surplus_list) | |
2346 | list_del_init(&iocg->surplus_list); | |
2347 | ||
7caa4715 TH |
2348 | /* |
2349 | * If q is getting clogged or we're missing too much, we're issuing | |
2350 | * too much IO and should lower vtime rate. If we're not missing | |
2351 | * and experiencing shortages but not surpluses, we're too stingy | |
2352 | * and should increase vtime rate. | |
2353 | */ | |
25d41e4a | 2354 | prev_busy_level = ioc->busy_level; |
7caa4715 TH |
2355 | if (rq_wait_pct > RQ_WAIT_BUSY_PCT || |
2356 | missed_ppm[READ] > ppm_rthr || | |
2357 | missed_ppm[WRITE] > ppm_wthr) { | |
81ca627a | 2358 | /* clearly missing QoS targets, slow down vrate */ |
7caa4715 TH |
2359 | ioc->busy_level = max(ioc->busy_level, 0); |
2360 | ioc->busy_level++; | |
7cd806a9 | 2361 | } else if (rq_wait_pct <= RQ_WAIT_BUSY_PCT * UNBUSY_THR_PCT / 100 && |
7caa4715 TH |
2362 | missed_ppm[READ] <= ppm_rthr * UNBUSY_THR_PCT / 100 && |
2363 | missed_ppm[WRITE] <= ppm_wthr * UNBUSY_THR_PCT / 100) { | |
81ca627a TH |
2364 | /* QoS targets are being met with >25% margin */ |
2365 | if (nr_shortages) { | |
2366 | /* | |
2367 | * We're throttling while the device has spare | |
2368 | * capacity. If vrate was being slowed down, stop. | |
2369 | */ | |
7cd806a9 | 2370 | ioc->busy_level = min(ioc->busy_level, 0); |
81ca627a TH |
2371 | |
2372 | /* | |
2373 | * If there are IOs spanning multiple periods, wait | |
065655c8 | 2374 | * them out before pushing the device harder. |
81ca627a | 2375 | */ |
065655c8 | 2376 | if (!nr_lagging) |
7cd806a9 | 2377 | ioc->busy_level--; |
81ca627a TH |
2378 | } else { |
2379 | /* | |
2380 | * Nobody is being throttled and the users aren't | |
2381 | * issuing enough IOs to saturate the device. We | |
2382 | * simply don't know how close the device is to | |
2383 | * saturation. Coast. | |
2384 | */ | |
2385 | ioc->busy_level = 0; | |
7cd806a9 | 2386 | } |
7caa4715 | 2387 | } else { |
81ca627a | 2388 | /* inside the hysterisis margin, we're good */ |
7caa4715 TH |
2389 | ioc->busy_level = 0; |
2390 | } | |
2391 | ||
2392 | ioc->busy_level = clamp(ioc->busy_level, -1000, 1000); | |
2393 | ||
926f75f6 BW |
2394 | ioc_adjust_base_vrate(ioc, rq_wait_pct, nr_lagging, nr_shortages, |
2395 | prev_busy_level, missed_ppm); | |
7caa4715 TH |
2396 | |
2397 | ioc_refresh_params(ioc, false); | |
2398 | ||
33a1fe6d TH |
2399 | ioc_forgive_debts(ioc, usage_us_sum, nr_debtors, &now); |
2400 | ||
7caa4715 TH |
2401 | /* |
2402 | * This period is done. Move onto the next one. If nothing's | |
2403 | * going on with the device, stop the timer. | |
2404 | */ | |
2405 | atomic64_inc(&ioc->cur_period); | |
2406 | ||
2407 | if (ioc->running != IOC_STOP) { | |
2408 | if (!list_empty(&ioc->active_iocgs)) { | |
2409 | ioc_start_period(ioc, &now); | |
2410 | } else { | |
2411 | ioc->busy_level = 0; | |
ac33e91e | 2412 | ioc->vtime_err = 0; |
7caa4715 TH |
2413 | ioc->running = IOC_IDLE; |
2414 | } | |
ac33e91e TH |
2415 | |
2416 | ioc_refresh_vrate(ioc, &now); | |
7caa4715 TH |
2417 | } |
2418 | ||
2419 | spin_unlock_irq(&ioc->lock); | |
2420 | } | |
2421 | ||
b0853ab4 TH |
2422 | static u64 adjust_inuse_and_calc_cost(struct ioc_gq *iocg, u64 vtime, |
2423 | u64 abs_cost, struct ioc_now *now) | |
2424 | { | |
2425 | struct ioc *ioc = iocg->ioc; | |
2426 | struct ioc_margins *margins = &ioc->margins; | |
04603755 | 2427 | u32 __maybe_unused old_inuse = iocg->inuse, __maybe_unused old_hwi; |
aa67db24 | 2428 | u32 hwi, adj_step; |
b0853ab4 TH |
2429 | s64 margin; |
2430 | u64 cost, new_inuse; | |
2431 | ||
2432 | current_hweight(iocg, NULL, &hwi); | |
04603755 | 2433 | old_hwi = hwi; |
b0853ab4 TH |
2434 | cost = abs_cost_to_cost(abs_cost, hwi); |
2435 | margin = now->vnow - vtime - cost; | |
2436 | ||
c421a3eb TH |
2437 | /* debt handling owns inuse for debtors */ |
2438 | if (iocg->abs_vdebt) | |
2439 | return cost; | |
2440 | ||
b0853ab4 | 2441 | /* |
5ba1add2 | 2442 | * We only increase inuse during period and do so if the margin has |
b0853ab4 TH |
2443 | * deteriorated since the previous adjustment. |
2444 | */ | |
2445 | if (margin >= iocg->saved_margin || margin >= margins->low || | |
2446 | iocg->inuse == iocg->active) | |
2447 | return cost; | |
2448 | ||
2449 | spin_lock_irq(&ioc->lock); | |
2450 | ||
2451 | /* we own inuse only when @iocg is in the normal active state */ | |
c421a3eb | 2452 | if (iocg->abs_vdebt || list_empty(&iocg->active_list)) { |
b0853ab4 TH |
2453 | spin_unlock_irq(&ioc->lock); |
2454 | return cost; | |
2455 | } | |
2456 | ||
aa67db24 TH |
2457 | /* |
2458 | * Bump up inuse till @abs_cost fits in the existing budget. | |
2459 | * adj_step must be determined after acquiring ioc->lock - we might | |
2460 | * have raced and lost to another thread for activation and could | |
2461 | * be reading 0 iocg->active before ioc->lock which will lead to | |
2462 | * infinite loop. | |
2463 | */ | |
b0853ab4 | 2464 | new_inuse = iocg->inuse; |
aa67db24 | 2465 | adj_step = DIV_ROUND_UP(iocg->active * INUSE_ADJ_STEP_PCT, 100); |
b0853ab4 TH |
2466 | do { |
2467 | new_inuse = new_inuse + adj_step; | |
2468 | propagate_weights(iocg, iocg->active, new_inuse, true, now); | |
2469 | current_hweight(iocg, NULL, &hwi); | |
2470 | cost = abs_cost_to_cost(abs_cost, hwi); | |
2471 | } while (time_after64(vtime + cost, now->vnow) && | |
2472 | iocg->inuse != iocg->active); | |
2473 | ||
2474 | spin_unlock_irq(&ioc->lock); | |
04603755 TH |
2475 | |
2476 | TRACE_IOCG_PATH(inuse_adjust, iocg, now, | |
2477 | old_inuse, iocg->inuse, old_hwi, hwi); | |
2478 | ||
b0853ab4 TH |
2479 | return cost; |
2480 | } | |
2481 | ||
7caa4715 TH |
2482 | static void calc_vtime_cost_builtin(struct bio *bio, struct ioc_gq *iocg, |
2483 | bool is_merge, u64 *costp) | |
2484 | { | |
2485 | struct ioc *ioc = iocg->ioc; | |
2486 | u64 coef_seqio, coef_randio, coef_page; | |
2487 | u64 pages = max_t(u64, bio_sectors(bio) >> IOC_SECT_TO_PAGE_SHIFT, 1); | |
2488 | u64 seek_pages = 0; | |
2489 | u64 cost = 0; | |
2490 | ||
2491 | switch (bio_op(bio)) { | |
2492 | case REQ_OP_READ: | |
2493 | coef_seqio = ioc->params.lcoefs[LCOEF_RSEQIO]; | |
2494 | coef_randio = ioc->params.lcoefs[LCOEF_RRANDIO]; | |
2495 | coef_page = ioc->params.lcoefs[LCOEF_RPAGE]; | |
2496 | break; | |
2497 | case REQ_OP_WRITE: | |
2498 | coef_seqio = ioc->params.lcoefs[LCOEF_WSEQIO]; | |
2499 | coef_randio = ioc->params.lcoefs[LCOEF_WRANDIO]; | |
2500 | coef_page = ioc->params.lcoefs[LCOEF_WPAGE]; | |
2501 | break; | |
2502 | default: | |
2503 | goto out; | |
2504 | } | |
2505 | ||
2506 | if (iocg->cursor) { | |
2507 | seek_pages = abs(bio->bi_iter.bi_sector - iocg->cursor); | |
2508 | seek_pages >>= IOC_SECT_TO_PAGE_SHIFT; | |
2509 | } | |
2510 | ||
2511 | if (!is_merge) { | |
2512 | if (seek_pages > LCOEF_RANDIO_PAGES) { | |
2513 | cost += coef_randio; | |
2514 | } else { | |
2515 | cost += coef_seqio; | |
2516 | } | |
2517 | } | |
2518 | cost += pages * coef_page; | |
2519 | out: | |
2520 | *costp = cost; | |
2521 | } | |
2522 | ||
2523 | static u64 calc_vtime_cost(struct bio *bio, struct ioc_gq *iocg, bool is_merge) | |
2524 | { | |
2525 | u64 cost; | |
2526 | ||
2527 | calc_vtime_cost_builtin(bio, iocg, is_merge, &cost); | |
2528 | return cost; | |
2529 | } | |
2530 | ||
cd006509 TH |
2531 | static void calc_size_vtime_cost_builtin(struct request *rq, struct ioc *ioc, |
2532 | u64 *costp) | |
2533 | { | |
2534 | unsigned int pages = blk_rq_stats_sectors(rq) >> IOC_SECT_TO_PAGE_SHIFT; | |
2535 | ||
2536 | switch (req_op(rq)) { | |
2537 | case REQ_OP_READ: | |
2538 | *costp = pages * ioc->params.lcoefs[LCOEF_RPAGE]; | |
2539 | break; | |
2540 | case REQ_OP_WRITE: | |
2541 | *costp = pages * ioc->params.lcoefs[LCOEF_WPAGE]; | |
2542 | break; | |
2543 | default: | |
2544 | *costp = 0; | |
2545 | } | |
2546 | } | |
2547 | ||
2548 | static u64 calc_size_vtime_cost(struct request *rq, struct ioc *ioc) | |
2549 | { | |
2550 | u64 cost; | |
2551 | ||
2552 | calc_size_vtime_cost_builtin(rq, ioc, &cost); | |
2553 | return cost; | |
2554 | } | |
2555 | ||
7caa4715 TH |
2556 | static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio) |
2557 | { | |
2558 | struct blkcg_gq *blkg = bio->bi_blkg; | |
2559 | struct ioc *ioc = rqos_to_ioc(rqos); | |
2560 | struct ioc_gq *iocg = blkg_to_iocg(blkg); | |
2561 | struct ioc_now now; | |
2562 | struct iocg_wait wait; | |
7caa4715 | 2563 | u64 abs_cost, cost, vtime; |
da437b95 TH |
2564 | bool use_debt, ioc_locked; |
2565 | unsigned long flags; | |
7caa4715 | 2566 | |
d16baa3f TH |
2567 | /* bypass IOs if disabled, still initializing, or for root cgroup */ |
2568 | if (!ioc->enabled || !iocg || !iocg->level) | |
7caa4715 TH |
2569 | return; |
2570 | ||
7caa4715 TH |
2571 | /* calculate the absolute vtime cost */ |
2572 | abs_cost = calc_vtime_cost(bio, iocg, false); | |
2573 | if (!abs_cost) | |
2574 | return; | |
2575 | ||
f1de2439 TH |
2576 | if (!iocg_activate(iocg, &now)) |
2577 | return; | |
2578 | ||
7caa4715 | 2579 | iocg->cursor = bio_end_sector(bio); |
7caa4715 | 2580 | vtime = atomic64_read(&iocg->vtime); |
b0853ab4 | 2581 | cost = adjust_inuse_and_calc_cost(iocg, vtime, abs_cost, &now); |
7caa4715 TH |
2582 | |
2583 | /* | |
2584 | * If no one's waiting and within budget, issue right away. The | |
2585 | * tests are racy but the races aren't systemic - we only miss once | |
2586 | * in a while which is fine. | |
2587 | */ | |
0b80f986 | 2588 | if (!waitqueue_active(&iocg->waitq) && !iocg->abs_vdebt && |
7caa4715 | 2589 | time_before_eq64(vtime + cost, now.vnow)) { |
97eb1975 | 2590 | iocg_commit_bio(iocg, bio, abs_cost, cost); |
7caa4715 TH |
2591 | return; |
2592 | } | |
2593 | ||
36a52481 | 2594 | /* |
da437b95 TH |
2595 | * We're over budget. This can be handled in two ways. IOs which may |
2596 | * cause priority inversions are punted to @ioc->aux_iocg and charged as | |
2597 | * debt. Otherwise, the issuer is blocked on @iocg->waitq. Debt handling | |
2598 | * requires @ioc->lock, waitq handling @iocg->waitq.lock. Determine | |
2599 | * whether debt handling is needed and acquire locks accordingly. | |
0b80f986 | 2600 | */ |
da437b95 TH |
2601 | use_debt = bio_issue_as_root_blkg(bio) || fatal_signal_pending(current); |
2602 | ioc_locked = use_debt || READ_ONCE(iocg->abs_vdebt); | |
b0853ab4 | 2603 | retry_lock: |
da437b95 TH |
2604 | iocg_lock(iocg, ioc_locked, &flags); |
2605 | ||
2606 | /* | |
2607 | * @iocg must stay activated for debt and waitq handling. Deactivation | |
2608 | * is synchronized against both ioc->lock and waitq.lock and we won't | |
2609 | * get deactivated as long as we're waiting or has debt, so we're good | |
2610 | * if we're activated here. In the unlikely cases that we aren't, just | |
2611 | * issue the IO. | |
2612 | */ | |
0b80f986 | 2613 | if (unlikely(list_empty(&iocg->active_list))) { |
da437b95 | 2614 | iocg_unlock(iocg, ioc_locked, &flags); |
97eb1975 | 2615 | iocg_commit_bio(iocg, bio, abs_cost, cost); |
0b80f986 TH |
2616 | return; |
2617 | } | |
2618 | ||
2619 | /* | |
2620 | * We're over budget. If @bio has to be issued regardless, remember | |
2621 | * the abs_cost instead of advancing vtime. iocg_kick_waitq() will pay | |
2622 | * off the debt before waking more IOs. | |
2623 | * | |
36a52481 | 2624 | * This way, the debt is continuously paid off each period with the |
0b80f986 TH |
2625 | * actual budget available to the cgroup. If we just wound vtime, we |
2626 | * would incorrectly use the current hw_inuse for the entire amount | |
2627 | * which, for example, can lead to the cgroup staying blocked for a | |
2628 | * long time even with substantially raised hw_inuse. | |
2629 | * | |
2630 | * An iocg with vdebt should stay online so that the timer can keep | |
2631 | * deducting its vdebt and [de]activate use_delay mechanism | |
2632 | * accordingly. We don't want to race against the timer trying to | |
2633 | * clear them and leave @iocg inactive w/ dangling use_delay heavily | |
2634 | * penalizing the cgroup and its descendants. | |
36a52481 | 2635 | */ |
da437b95 | 2636 | if (use_debt) { |
c421a3eb | 2637 | iocg_incur_debt(iocg, abs_cost, &now); |
54c52e10 | 2638 | if (iocg_kick_delay(iocg, &now)) |
de185b56 | 2639 | blkcg_schedule_throttle(rqos->q->disk, |
d7bd15a1 | 2640 | (bio->bi_opf & REQ_SWAP) == REQ_SWAP); |
da437b95 | 2641 | iocg_unlock(iocg, ioc_locked, &flags); |
7caa4715 TH |
2642 | return; |
2643 | } | |
2644 | ||
b0853ab4 | 2645 | /* guarantee that iocgs w/ waiters have maximum inuse */ |
c421a3eb | 2646 | if (!iocg->abs_vdebt && iocg->inuse != iocg->active) { |
b0853ab4 TH |
2647 | if (!ioc_locked) { |
2648 | iocg_unlock(iocg, false, &flags); | |
2649 | ioc_locked = true; | |
2650 | goto retry_lock; | |
2651 | } | |
2652 | propagate_weights(iocg, iocg->active, iocg->active, true, | |
2653 | &now); | |
2654 | } | |
2655 | ||
7caa4715 TH |
2656 | /* |
2657 | * Append self to the waitq and schedule the wakeup timer if we're | |
2658 | * the first waiter. The timer duration is calculated based on the | |
2659 | * current vrate. vtime and hweight changes can make it too short | |
2660 | * or too long. Each wait entry records the absolute cost it's | |
2661 | * waiting for to allow re-evaluation using a custom wait entry. | |
2662 | * | |
2663 | * If too short, the timer simply reschedules itself. If too long, | |
2664 | * the period timer will notice and trigger wakeups. | |
2665 | * | |
2666 | * All waiters are on iocg->waitq and the wait states are | |
2667 | * synchronized using waitq.lock. | |
2668 | */ | |
7caa4715 TH |
2669 | init_waitqueue_func_entry(&wait.wait, iocg_wake_fn); |
2670 | wait.wait.private = current; | |
2671 | wait.bio = bio; | |
2672 | wait.abs_cost = abs_cost; | |
2673 | wait.committed = false; /* will be set true by waker */ | |
2674 | ||
2675 | __add_wait_queue_entry_tail(&iocg->waitq, &wait.wait); | |
da437b95 | 2676 | iocg_kick_waitq(iocg, ioc_locked, &now); |
7caa4715 | 2677 | |
da437b95 | 2678 | iocg_unlock(iocg, ioc_locked, &flags); |
7caa4715 TH |
2679 | |
2680 | while (true) { | |
2681 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2682 | if (wait.committed) | |
2683 | break; | |
2684 | io_schedule(); | |
2685 | } | |
2686 | ||
2687 | /* waker already committed us, proceed */ | |
2688 | finish_wait(&iocg->waitq, &wait.wait); | |
2689 | } | |
2690 | ||
2691 | static void ioc_rqos_merge(struct rq_qos *rqos, struct request *rq, | |
2692 | struct bio *bio) | |
2693 | { | |
2694 | struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg); | |
d16baa3f | 2695 | struct ioc *ioc = rqos_to_ioc(rqos); |
7caa4715 | 2696 | sector_t bio_end = bio_end_sector(bio); |
e1518f63 | 2697 | struct ioc_now now; |
b0853ab4 | 2698 | u64 vtime, abs_cost, cost; |
0b80f986 | 2699 | unsigned long flags; |
7caa4715 | 2700 | |
d16baa3f TH |
2701 | /* bypass if disabled, still initializing, or for root cgroup */ |
2702 | if (!ioc->enabled || !iocg || !iocg->level) | |
7caa4715 TH |
2703 | return; |
2704 | ||
2705 | abs_cost = calc_vtime_cost(bio, iocg, true); | |
2706 | if (!abs_cost) | |
2707 | return; | |
2708 | ||
e1518f63 | 2709 | ioc_now(ioc, &now); |
b0853ab4 TH |
2710 | |
2711 | vtime = atomic64_read(&iocg->vtime); | |
2712 | cost = adjust_inuse_and_calc_cost(iocg, vtime, abs_cost, &now); | |
e1518f63 | 2713 | |
7caa4715 TH |
2714 | /* update cursor if backmerging into the request at the cursor */ |
2715 | if (blk_rq_pos(rq) < bio_end && | |
2716 | blk_rq_pos(rq) + blk_rq_sectors(rq) == iocg->cursor) | |
2717 | iocg->cursor = bio_end; | |
2718 | ||
e1518f63 | 2719 | /* |
0b80f986 TH |
2720 | * Charge if there's enough vtime budget and the existing request has |
2721 | * cost assigned. | |
e1518f63 TH |
2722 | */ |
2723 | if (rq->bio && rq->bio->bi_iocost_cost && | |
0b80f986 | 2724 | time_before_eq64(atomic64_read(&iocg->vtime) + cost, now.vnow)) { |
97eb1975 | 2725 | iocg_commit_bio(iocg, bio, abs_cost, cost); |
0b80f986 TH |
2726 | return; |
2727 | } | |
2728 | ||
2729 | /* | |
2730 | * Otherwise, account it as debt if @iocg is online, which it should | |
2731 | * be for the vast majority of cases. See debt handling in | |
2732 | * ioc_rqos_throttle() for details. | |
2733 | */ | |
c421a3eb TH |
2734 | spin_lock_irqsave(&ioc->lock, flags); |
2735 | spin_lock(&iocg->waitq.lock); | |
2736 | ||
0b80f986 | 2737 | if (likely(!list_empty(&iocg->active_list))) { |
c421a3eb TH |
2738 | iocg_incur_debt(iocg, abs_cost, &now); |
2739 | if (iocg_kick_delay(iocg, &now)) | |
de185b56 | 2740 | blkcg_schedule_throttle(rqos->q->disk, |
c421a3eb | 2741 | (bio->bi_opf & REQ_SWAP) == REQ_SWAP); |
0b80f986 | 2742 | } else { |
97eb1975 | 2743 | iocg_commit_bio(iocg, bio, abs_cost, cost); |
0b80f986 | 2744 | } |
c421a3eb TH |
2745 | |
2746 | spin_unlock(&iocg->waitq.lock); | |
2747 | spin_unlock_irqrestore(&ioc->lock, flags); | |
7caa4715 TH |
2748 | } |
2749 | ||
2750 | static void ioc_rqos_done_bio(struct rq_qos *rqos, struct bio *bio) | |
2751 | { | |
2752 | struct ioc_gq *iocg = blkg_to_iocg(bio->bi_blkg); | |
2753 | ||
2754 | if (iocg && bio->bi_iocost_cost) | |
2755 | atomic64_add(bio->bi_iocost_cost, &iocg->done_vtime); | |
2756 | } | |
2757 | ||
2758 | static void ioc_rqos_done(struct rq_qos *rqos, struct request *rq) | |
2759 | { | |
2760 | struct ioc *ioc = rqos_to_ioc(rqos); | |
5e124f74 | 2761 | struct ioc_pcpu_stat *ccs; |
cd006509 | 2762 | u64 on_q_ns, rq_wait_ns, size_nsec; |
7caa4715 TH |
2763 | int pidx, rw; |
2764 | ||
2765 | if (!ioc->enabled || !rq->alloc_time_ns || !rq->start_time_ns) | |
2766 | return; | |
2767 | ||
62c159a0 | 2768 | switch (req_op(rq)) { |
7caa4715 TH |
2769 | case REQ_OP_READ: |
2770 | pidx = QOS_RLAT; | |
2771 | rw = READ; | |
2772 | break; | |
2773 | case REQ_OP_WRITE: | |
2774 | pidx = QOS_WLAT; | |
2775 | rw = WRITE; | |
2776 | break; | |
2777 | default: | |
2778 | return; | |
2779 | } | |
2780 | ||
2781 | on_q_ns = ktime_get_ns() - rq->alloc_time_ns; | |
2782 | rq_wait_ns = rq->start_time_ns - rq->alloc_time_ns; | |
cd006509 | 2783 | size_nsec = div64_u64(calc_size_vtime_cost(rq, ioc), VTIME_PER_NSEC); |
7caa4715 | 2784 | |
5e124f74 TH |
2785 | ccs = get_cpu_ptr(ioc->pcpu_stat); |
2786 | ||
cd006509 TH |
2787 | if (on_q_ns <= size_nsec || |
2788 | on_q_ns - size_nsec <= ioc->params.qos[pidx] * NSEC_PER_USEC) | |
5e124f74 | 2789 | local_inc(&ccs->missed[rw].nr_met); |
7caa4715 | 2790 | else |
5e124f74 TH |
2791 | local_inc(&ccs->missed[rw].nr_missed); |
2792 | ||
2793 | local64_add(rq_wait_ns, &ccs->rq_wait_ns); | |
7caa4715 | 2794 | |
5e124f74 | 2795 | put_cpu_ptr(ccs); |
7caa4715 TH |
2796 | } |
2797 | ||
2798 | static void ioc_rqos_queue_depth_changed(struct rq_qos *rqos) | |
2799 | { | |
2800 | struct ioc *ioc = rqos_to_ioc(rqos); | |
2801 | ||
2802 | spin_lock_irq(&ioc->lock); | |
2803 | ioc_refresh_params(ioc, false); | |
2804 | spin_unlock_irq(&ioc->lock); | |
2805 | } | |
2806 | ||
2807 | static void ioc_rqos_exit(struct rq_qos *rqos) | |
2808 | { | |
2809 | struct ioc *ioc = rqos_to_ioc(rqos); | |
2810 | ||
2811 | blkcg_deactivate_policy(rqos->q, &blkcg_policy_iocost); | |
2812 | ||
2813 | spin_lock_irq(&ioc->lock); | |
2814 | ioc->running = IOC_STOP; | |
2815 | spin_unlock_irq(&ioc->lock); | |
2816 | ||
2817 | del_timer_sync(&ioc->timer); | |
2818 | free_percpu(ioc->pcpu_stat); | |
2819 | kfree(ioc); | |
2820 | } | |
2821 | ||
2822 | static struct rq_qos_ops ioc_rqos_ops = { | |
2823 | .throttle = ioc_rqos_throttle, | |
2824 | .merge = ioc_rqos_merge, | |
2825 | .done_bio = ioc_rqos_done_bio, | |
2826 | .done = ioc_rqos_done, | |
2827 | .queue_depth_changed = ioc_rqos_queue_depth_changed, | |
2828 | .exit = ioc_rqos_exit, | |
2829 | }; | |
2830 | ||
57b64554 | 2831 | static int blk_iocost_init(struct gendisk *disk) |
7caa4715 | 2832 | { |
57b64554 | 2833 | struct request_queue *q = disk->queue; |
7caa4715 TH |
2834 | struct ioc *ioc; |
2835 | struct rq_qos *rqos; | |
5e124f74 | 2836 | int i, cpu, ret; |
7caa4715 TH |
2837 | |
2838 | ioc = kzalloc(sizeof(*ioc), GFP_KERNEL); | |
2839 | if (!ioc) | |
2840 | return -ENOMEM; | |
2841 | ||
2842 | ioc->pcpu_stat = alloc_percpu(struct ioc_pcpu_stat); | |
2843 | if (!ioc->pcpu_stat) { | |
2844 | kfree(ioc); | |
2845 | return -ENOMEM; | |
2846 | } | |
2847 | ||
5e124f74 TH |
2848 | for_each_possible_cpu(cpu) { |
2849 | struct ioc_pcpu_stat *ccs = per_cpu_ptr(ioc->pcpu_stat, cpu); | |
2850 | ||
2851 | for (i = 0; i < ARRAY_SIZE(ccs->missed); i++) { | |
2852 | local_set(&ccs->missed[i].nr_met, 0); | |
2853 | local_set(&ccs->missed[i].nr_missed, 0); | |
2854 | } | |
2855 | local64_set(&ccs->rq_wait_ns, 0); | |
2856 | } | |
2857 | ||
7caa4715 TH |
2858 | rqos = &ioc->rqos; |
2859 | rqos->id = RQ_QOS_COST; | |
2860 | rqos->ops = &ioc_rqos_ops; | |
2861 | rqos->q = q; | |
2862 | ||
2863 | spin_lock_init(&ioc->lock); | |
2864 | timer_setup(&ioc->timer, ioc_timer_fn, 0); | |
2865 | INIT_LIST_HEAD(&ioc->active_iocgs); | |
2866 | ||
2867 | ioc->running = IOC_IDLE; | |
ac33e91e | 2868 | ioc->vtime_base_rate = VTIME_PER_USEC; |
7caa4715 | 2869 | atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC); |
67b7b641 | 2870 | seqcount_spinlock_init(&ioc->period_seqcount, &ioc->lock); |
7caa4715 TH |
2871 | ioc->period_at = ktime_to_us(ktime_get()); |
2872 | atomic64_set(&ioc->cur_period, 0); | |
2873 | atomic_set(&ioc->hweight_gen, 0); | |
2874 | ||
2875 | spin_lock_irq(&ioc->lock); | |
2876 | ioc->autop_idx = AUTOP_INVALID; | |
2877 | ioc_refresh_params(ioc, true); | |
2878 | spin_unlock_irq(&ioc->lock); | |
2879 | ||
d16baa3f TH |
2880 | /* |
2881 | * rqos must be added before activation to allow iocg_pd_init() to | |
2882 | * lookup the ioc from q. This means that the rqos methods may get | |
2883 | * called before policy activation completion, can't assume that the | |
2884 | * target bio has an iocg associated and need to test for NULL iocg. | |
2885 | */ | |
14a6e2eb JH |
2886 | ret = rq_qos_add(q, rqos); |
2887 | if (ret) | |
2888 | goto err_free_ioc; | |
2889 | ||
7caa4715 | 2890 | ret = blkcg_activate_policy(q, &blkcg_policy_iocost); |
14a6e2eb JH |
2891 | if (ret) |
2892 | goto err_del_qos; | |
7caa4715 | 2893 | return 0; |
14a6e2eb JH |
2894 | |
2895 | err_del_qos: | |
2896 | rq_qos_del(q, rqos); | |
2897 | err_free_ioc: | |
2898 | free_percpu(ioc->pcpu_stat); | |
2899 | kfree(ioc); | |
2900 | return ret; | |
7caa4715 TH |
2901 | } |
2902 | ||
2903 | static struct blkcg_policy_data *ioc_cpd_alloc(gfp_t gfp) | |
2904 | { | |
2905 | struct ioc_cgrp *iocc; | |
2906 | ||
2907 | iocc = kzalloc(sizeof(struct ioc_cgrp), gfp); | |
e916ad29 TH |
2908 | if (!iocc) |
2909 | return NULL; | |
7caa4715 | 2910 | |
bd0adb91 | 2911 | iocc->dfl_weight = CGROUP_WEIGHT_DFL * WEIGHT_ONE; |
7caa4715 TH |
2912 | return &iocc->cpd; |
2913 | } | |
2914 | ||
2915 | static void ioc_cpd_free(struct blkcg_policy_data *cpd) | |
2916 | { | |
2917 | kfree(container_of(cpd, struct ioc_cgrp, cpd)); | |
2918 | } | |
2919 | ||
2920 | static struct blkg_policy_data *ioc_pd_alloc(gfp_t gfp, struct request_queue *q, | |
2921 | struct blkcg *blkcg) | |
2922 | { | |
2923 | int levels = blkcg->css.cgroup->level + 1; | |
2924 | struct ioc_gq *iocg; | |
2925 | ||
f61d6e25 | 2926 | iocg = kzalloc_node(struct_size(iocg, ancestors, levels), gfp, q->node); |
7caa4715 TH |
2927 | if (!iocg) |
2928 | return NULL; | |
2929 | ||
97eb1975 TH |
2930 | iocg->pcpu_stat = alloc_percpu_gfp(struct iocg_pcpu_stat, gfp); |
2931 | if (!iocg->pcpu_stat) { | |
2932 | kfree(iocg); | |
2933 | return NULL; | |
2934 | } | |
2935 | ||
7caa4715 TH |
2936 | return &iocg->pd; |
2937 | } | |
2938 | ||
2939 | static void ioc_pd_init(struct blkg_policy_data *pd) | |
2940 | { | |
2941 | struct ioc_gq *iocg = pd_to_iocg(pd); | |
2942 | struct blkcg_gq *blkg = pd_to_blkg(&iocg->pd); | |
2943 | struct ioc *ioc = q_to_ioc(blkg->q); | |
2944 | struct ioc_now now; | |
2945 | struct blkcg_gq *tblkg; | |
2946 | unsigned long flags; | |
2947 | ||
2948 | ioc_now(ioc, &now); | |
2949 | ||
2950 | iocg->ioc = ioc; | |
2951 | atomic64_set(&iocg->vtime, now.vnow); | |
2952 | atomic64_set(&iocg->done_vtime, now.vnow); | |
2953 | atomic64_set(&iocg->active_period, atomic64_read(&ioc->cur_period)); | |
2954 | INIT_LIST_HEAD(&iocg->active_list); | |
97eb1975 | 2955 | INIT_LIST_HEAD(&iocg->walk_list); |
8692d2db | 2956 | INIT_LIST_HEAD(&iocg->surplus_list); |
fe20cdb5 TH |
2957 | iocg->hweight_active = WEIGHT_ONE; |
2958 | iocg->hweight_inuse = WEIGHT_ONE; | |
7caa4715 TH |
2959 | |
2960 | init_waitqueue_head(&iocg->waitq); | |
2961 | hrtimer_init(&iocg->waitq_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); | |
2962 | iocg->waitq_timer.function = iocg_waitq_timer_fn; | |
7caa4715 TH |
2963 | |
2964 | iocg->level = blkg->blkcg->css.cgroup->level; | |
2965 | ||
2966 | for (tblkg = blkg; tblkg; tblkg = tblkg->parent) { | |
2967 | struct ioc_gq *tiocg = blkg_to_iocg(tblkg); | |
2968 | iocg->ancestors[tiocg->level] = tiocg; | |
2969 | } | |
2970 | ||
2971 | spin_lock_irqsave(&ioc->lock, flags); | |
b0853ab4 | 2972 | weight_updated(iocg, &now); |
7caa4715 TH |
2973 | spin_unlock_irqrestore(&ioc->lock, flags); |
2974 | } | |
2975 | ||
2976 | static void ioc_pd_free(struct blkg_policy_data *pd) | |
2977 | { | |
2978 | struct ioc_gq *iocg = pd_to_iocg(pd); | |
2979 | struct ioc *ioc = iocg->ioc; | |
5aeac7c4 | 2980 | unsigned long flags; |
7caa4715 TH |
2981 | |
2982 | if (ioc) { | |
5aeac7c4 | 2983 | spin_lock_irqsave(&ioc->lock, flags); |
97eb1975 | 2984 | |
7caa4715 | 2985 | if (!list_empty(&iocg->active_list)) { |
b0853ab4 TH |
2986 | struct ioc_now now; |
2987 | ||
2988 | ioc_now(ioc, &now); | |
2989 | propagate_weights(iocg, 0, 0, false, &now); | |
7caa4715 TH |
2990 | list_del_init(&iocg->active_list); |
2991 | } | |
97eb1975 TH |
2992 | |
2993 | WARN_ON_ONCE(!list_empty(&iocg->walk_list)); | |
8692d2db | 2994 | WARN_ON_ONCE(!list_empty(&iocg->surplus_list)); |
97eb1975 | 2995 | |
5aeac7c4 | 2996 | spin_unlock_irqrestore(&ioc->lock, flags); |
e036c4ca TH |
2997 | |
2998 | hrtimer_cancel(&iocg->waitq_timer); | |
7caa4715 | 2999 | } |
97eb1975 | 3000 | free_percpu(iocg->pcpu_stat); |
7caa4715 TH |
3001 | kfree(iocg); |
3002 | } | |
3003 | ||
3607849d | 3004 | static void ioc_pd_stat(struct blkg_policy_data *pd, struct seq_file *s) |
97eb1975 TH |
3005 | { |
3006 | struct ioc_gq *iocg = pd_to_iocg(pd); | |
3007 | struct ioc *ioc = iocg->ioc; | |
97eb1975 TH |
3008 | |
3009 | if (!ioc->enabled) | |
3607849d | 3010 | return; |
97eb1975 TH |
3011 | |
3012 | if (iocg->level == 0) { | |
3013 | unsigned vp10k = DIV64_U64_ROUND_CLOSEST( | |
ac33e91e | 3014 | ioc->vtime_base_rate * 10000, |
97eb1975 | 3015 | VTIME_PER_USEC); |
252c651a | 3016 | seq_printf(s, " cost.vrate=%u.%02u", vp10k / 100, vp10k % 100); |
97eb1975 TH |
3017 | } |
3018 | ||
252c651a | 3019 | seq_printf(s, " cost.usage=%llu", iocg->last_stat.usage_us); |
97eb1975 | 3020 | |
f0bf84a5 | 3021 | if (blkcg_debug_stats) |
252c651a CH |
3022 | seq_printf(s, " cost.wait=%llu cost.indebt=%llu cost.indelay=%llu", |
3023 | iocg->last_stat.wait_us, | |
3024 | iocg->last_stat.indebt_us, | |
3025 | iocg->last_stat.indelay_us); | |
97eb1975 TH |
3026 | } |
3027 | ||
7caa4715 TH |
3028 | static u64 ioc_weight_prfill(struct seq_file *sf, struct blkg_policy_data *pd, |
3029 | int off) | |
3030 | { | |
3031 | const char *dname = blkg_dev_name(pd->blkg); | |
3032 | struct ioc_gq *iocg = pd_to_iocg(pd); | |
3033 | ||
3034 | if (dname && iocg->cfg_weight) | |
bd0adb91 | 3035 | seq_printf(sf, "%s %u\n", dname, iocg->cfg_weight / WEIGHT_ONE); |
7caa4715 TH |
3036 | return 0; |
3037 | } | |
3038 | ||
3039 | ||
3040 | static int ioc_weight_show(struct seq_file *sf, void *v) | |
3041 | { | |
3042 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
3043 | struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg); | |
3044 | ||
bd0adb91 | 3045 | seq_printf(sf, "default %u\n", iocc->dfl_weight / WEIGHT_ONE); |
7caa4715 TH |
3046 | blkcg_print_blkgs(sf, blkcg, ioc_weight_prfill, |
3047 | &blkcg_policy_iocost, seq_cft(sf)->private, false); | |
3048 | return 0; | |
3049 | } | |
3050 | ||
3051 | static ssize_t ioc_weight_write(struct kernfs_open_file *of, char *buf, | |
3052 | size_t nbytes, loff_t off) | |
3053 | { | |
3054 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
3055 | struct ioc_cgrp *iocc = blkcg_to_iocc(blkcg); | |
3056 | struct blkg_conf_ctx ctx; | |
b0853ab4 | 3057 | struct ioc_now now; |
7caa4715 TH |
3058 | struct ioc_gq *iocg; |
3059 | u32 v; | |
3060 | int ret; | |
3061 | ||
3062 | if (!strchr(buf, ':')) { | |
3063 | struct blkcg_gq *blkg; | |
3064 | ||
3065 | if (!sscanf(buf, "default %u", &v) && !sscanf(buf, "%u", &v)) | |
3066 | return -EINVAL; | |
3067 | ||
3068 | if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX) | |
3069 | return -EINVAL; | |
3070 | ||
11431e26 | 3071 | spin_lock_irq(&blkcg->lock); |
bd0adb91 | 3072 | iocc->dfl_weight = v * WEIGHT_ONE; |
7caa4715 TH |
3073 | hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { |
3074 | struct ioc_gq *iocg = blkg_to_iocg(blkg); | |
3075 | ||
3076 | if (iocg) { | |
11431e26 | 3077 | spin_lock(&iocg->ioc->lock); |
b0853ab4 TH |
3078 | ioc_now(iocg->ioc, &now); |
3079 | weight_updated(iocg, &now); | |
11431e26 | 3080 | spin_unlock(&iocg->ioc->lock); |
7caa4715 TH |
3081 | } |
3082 | } | |
11431e26 | 3083 | spin_unlock_irq(&blkcg->lock); |
7caa4715 TH |
3084 | |
3085 | return nbytes; | |
3086 | } | |
3087 | ||
3088 | ret = blkg_conf_prep(blkcg, &blkcg_policy_iocost, buf, &ctx); | |
3089 | if (ret) | |
3090 | return ret; | |
3091 | ||
3092 | iocg = blkg_to_iocg(ctx.blkg); | |
3093 | ||
3094 | if (!strncmp(ctx.body, "default", 7)) { | |
3095 | v = 0; | |
3096 | } else { | |
3097 | if (!sscanf(ctx.body, "%u", &v)) | |
3098 | goto einval; | |
3099 | if (v < CGROUP_WEIGHT_MIN || v > CGROUP_WEIGHT_MAX) | |
3100 | goto einval; | |
3101 | } | |
3102 | ||
41591a51 | 3103 | spin_lock(&iocg->ioc->lock); |
bd0adb91 | 3104 | iocg->cfg_weight = v * WEIGHT_ONE; |
b0853ab4 TH |
3105 | ioc_now(iocg->ioc, &now); |
3106 | weight_updated(iocg, &now); | |
41591a51 | 3107 | spin_unlock(&iocg->ioc->lock); |
7caa4715 TH |
3108 | |
3109 | blkg_conf_finish(&ctx); | |
3110 | return nbytes; | |
3111 | ||
3112 | einval: | |
3113 | blkg_conf_finish(&ctx); | |
3114 | return -EINVAL; | |
3115 | } | |
3116 | ||
3117 | static u64 ioc_qos_prfill(struct seq_file *sf, struct blkg_policy_data *pd, | |
3118 | int off) | |
3119 | { | |
3120 | const char *dname = blkg_dev_name(pd->blkg); | |
3121 | struct ioc *ioc = pd_to_iocg(pd)->ioc; | |
3122 | ||
3123 | if (!dname) | |
3124 | return 0; | |
3125 | ||
3126 | seq_printf(sf, "%s enable=%d ctrl=%s rpct=%u.%02u rlat=%u wpct=%u.%02u wlat=%u min=%u.%02u max=%u.%02u\n", | |
3127 | dname, ioc->enabled, ioc->user_qos_params ? "user" : "auto", | |
3128 | ioc->params.qos[QOS_RPPM] / 10000, | |
3129 | ioc->params.qos[QOS_RPPM] % 10000 / 100, | |
3130 | ioc->params.qos[QOS_RLAT], | |
3131 | ioc->params.qos[QOS_WPPM] / 10000, | |
3132 | ioc->params.qos[QOS_WPPM] % 10000 / 100, | |
3133 | ioc->params.qos[QOS_WLAT], | |
3134 | ioc->params.qos[QOS_MIN] / 10000, | |
3135 | ioc->params.qos[QOS_MIN] % 10000 / 100, | |
3136 | ioc->params.qos[QOS_MAX] / 10000, | |
3137 | ioc->params.qos[QOS_MAX] % 10000 / 100); | |
3138 | return 0; | |
3139 | } | |
3140 | ||
3141 | static int ioc_qos_show(struct seq_file *sf, void *v) | |
3142 | { | |
3143 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
3144 | ||
3145 | blkcg_print_blkgs(sf, blkcg, ioc_qos_prfill, | |
3146 | &blkcg_policy_iocost, seq_cft(sf)->private, false); | |
3147 | return 0; | |
3148 | } | |
3149 | ||
3150 | static const match_table_t qos_ctrl_tokens = { | |
3151 | { QOS_ENABLE, "enable=%u" }, | |
3152 | { QOS_CTRL, "ctrl=%s" }, | |
3153 | { NR_QOS_CTRL_PARAMS, NULL }, | |
3154 | }; | |
3155 | ||
3156 | static const match_table_t qos_tokens = { | |
3157 | { QOS_RPPM, "rpct=%s" }, | |
3158 | { QOS_RLAT, "rlat=%u" }, | |
3159 | { QOS_WPPM, "wpct=%s" }, | |
3160 | { QOS_WLAT, "wlat=%u" }, | |
3161 | { QOS_MIN, "min=%s" }, | |
3162 | { QOS_MAX, "max=%s" }, | |
3163 | { NR_QOS_PARAMS, NULL }, | |
3164 | }; | |
3165 | ||
3166 | static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input, | |
3167 | size_t nbytes, loff_t off) | |
3168 | { | |
22ae8ce8 | 3169 | struct block_device *bdev; |
3657647e | 3170 | struct gendisk *disk; |
7caa4715 TH |
3171 | struct ioc *ioc; |
3172 | u32 qos[NR_QOS_PARAMS]; | |
3173 | bool enable, user; | |
3174 | char *p; | |
3175 | int ret; | |
3176 | ||
22ae8ce8 CH |
3177 | bdev = blkcg_conf_open_bdev(&input); |
3178 | if (IS_ERR(bdev)) | |
3179 | return PTR_ERR(bdev); | |
7caa4715 | 3180 | |
3657647e CH |
3181 | disk = bdev->bd_disk; |
3182 | ioc = q_to_ioc(disk->queue); | |
7caa4715 | 3183 | if (!ioc) { |
3657647e | 3184 | ret = blk_iocost_init(disk); |
7caa4715 TH |
3185 | if (ret) |
3186 | goto err; | |
3657647e | 3187 | ioc = q_to_ioc(disk->queue); |
7caa4715 TH |
3188 | } |
3189 | ||
3190 | spin_lock_irq(&ioc->lock); | |
3191 | memcpy(qos, ioc->params.qos, sizeof(qos)); | |
3192 | enable = ioc->enabled; | |
3193 | user = ioc->user_qos_params; | |
3194 | spin_unlock_irq(&ioc->lock); | |
3195 | ||
3196 | while ((p = strsep(&input, " \t\n"))) { | |
3197 | substring_t args[MAX_OPT_ARGS]; | |
3198 | char buf[32]; | |
3199 | int tok; | |
3200 | s64 v; | |
3201 | ||
3202 | if (!*p) | |
3203 | continue; | |
3204 | ||
3205 | switch (match_token(p, qos_ctrl_tokens, args)) { | |
3206 | case QOS_ENABLE: | |
3207 | match_u64(&args[0], &v); | |
3208 | enable = v; | |
3209 | continue; | |
3210 | case QOS_CTRL: | |
3211 | match_strlcpy(buf, &args[0], sizeof(buf)); | |
3212 | if (!strcmp(buf, "auto")) | |
3213 | user = false; | |
3214 | else if (!strcmp(buf, "user")) | |
3215 | user = true; | |
3216 | else | |
3217 | goto einval; | |
3218 | continue; | |
3219 | } | |
3220 | ||
3221 | tok = match_token(p, qos_tokens, args); | |
3222 | switch (tok) { | |
3223 | case QOS_RPPM: | |
3224 | case QOS_WPPM: | |
3225 | if (match_strlcpy(buf, &args[0], sizeof(buf)) >= | |
3226 | sizeof(buf)) | |
3227 | goto einval; | |
3228 | if (cgroup_parse_float(buf, 2, &v)) | |
3229 | goto einval; | |
3230 | if (v < 0 || v > 10000) | |
3231 | goto einval; | |
3232 | qos[tok] = v * 100; | |
3233 | break; | |
3234 | case QOS_RLAT: | |
3235 | case QOS_WLAT: | |
3236 | if (match_u64(&args[0], &v)) | |
3237 | goto einval; | |
3238 | qos[tok] = v; | |
3239 | break; | |
3240 | case QOS_MIN: | |
3241 | case QOS_MAX: | |
3242 | if (match_strlcpy(buf, &args[0], sizeof(buf)) >= | |
3243 | sizeof(buf)) | |
3244 | goto einval; | |
3245 | if (cgroup_parse_float(buf, 2, &v)) | |
3246 | goto einval; | |
3247 | if (v < 0) | |
3248 | goto einval; | |
3249 | qos[tok] = clamp_t(s64, v * 100, | |
3250 | VRATE_MIN_PPM, VRATE_MAX_PPM); | |
3251 | break; | |
3252 | default: | |
3253 | goto einval; | |
3254 | } | |
3255 | user = true; | |
3256 | } | |
3257 | ||
3258 | if (qos[QOS_MIN] > qos[QOS_MAX]) | |
3259 | goto einval; | |
3260 | ||
3261 | spin_lock_irq(&ioc->lock); | |
3262 | ||
3263 | if (enable) { | |
3657647e CH |
3264 | blk_stat_enable_accounting(disk->queue); |
3265 | blk_queue_flag_set(QUEUE_FLAG_RQ_ALLOC_TIME, disk->queue); | |
7caa4715 TH |
3266 | ioc->enabled = true; |
3267 | } else { | |
3657647e | 3268 | blk_queue_flag_clear(QUEUE_FLAG_RQ_ALLOC_TIME, disk->queue); |
7caa4715 TH |
3269 | ioc->enabled = false; |
3270 | } | |
3271 | ||
3272 | if (user) { | |
3273 | memcpy(ioc->params.qos, qos, sizeof(qos)); | |
3274 | ioc->user_qos_params = true; | |
3275 | } else { | |
3276 | ioc->user_qos_params = false; | |
3277 | } | |
3278 | ||
3279 | ioc_refresh_params(ioc, true); | |
3280 | spin_unlock_irq(&ioc->lock); | |
3281 | ||
22ae8ce8 | 3282 | blkdev_put_no_open(bdev); |
7caa4715 TH |
3283 | return nbytes; |
3284 | einval: | |
3285 | ret = -EINVAL; | |
3286 | err: | |
22ae8ce8 | 3287 | blkdev_put_no_open(bdev); |
7caa4715 TH |
3288 | return ret; |
3289 | } | |
3290 | ||
3291 | static u64 ioc_cost_model_prfill(struct seq_file *sf, | |
3292 | struct blkg_policy_data *pd, int off) | |
3293 | { | |
3294 | const char *dname = blkg_dev_name(pd->blkg); | |
3295 | struct ioc *ioc = pd_to_iocg(pd)->ioc; | |
3296 | u64 *u = ioc->params.i_lcoefs; | |
3297 | ||
3298 | if (!dname) | |
3299 | return 0; | |
3300 | ||
3301 | seq_printf(sf, "%s ctrl=%s model=linear " | |
3302 | "rbps=%llu rseqiops=%llu rrandiops=%llu " | |
3303 | "wbps=%llu wseqiops=%llu wrandiops=%llu\n", | |
3304 | dname, ioc->user_cost_model ? "user" : "auto", | |
3305 | u[I_LCOEF_RBPS], u[I_LCOEF_RSEQIOPS], u[I_LCOEF_RRANDIOPS], | |
3306 | u[I_LCOEF_WBPS], u[I_LCOEF_WSEQIOPS], u[I_LCOEF_WRANDIOPS]); | |
3307 | return 0; | |
3308 | } | |
3309 | ||
3310 | static int ioc_cost_model_show(struct seq_file *sf, void *v) | |
3311 | { | |
3312 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
3313 | ||
3314 | blkcg_print_blkgs(sf, blkcg, ioc_cost_model_prfill, | |
3315 | &blkcg_policy_iocost, seq_cft(sf)->private, false); | |
3316 | return 0; | |
3317 | } | |
3318 | ||
3319 | static const match_table_t cost_ctrl_tokens = { | |
3320 | { COST_CTRL, "ctrl=%s" }, | |
3321 | { COST_MODEL, "model=%s" }, | |
3322 | { NR_COST_CTRL_PARAMS, NULL }, | |
3323 | }; | |
3324 | ||
3325 | static const match_table_t i_lcoef_tokens = { | |
3326 | { I_LCOEF_RBPS, "rbps=%u" }, | |
3327 | { I_LCOEF_RSEQIOPS, "rseqiops=%u" }, | |
3328 | { I_LCOEF_RRANDIOPS, "rrandiops=%u" }, | |
3329 | { I_LCOEF_WBPS, "wbps=%u" }, | |
3330 | { I_LCOEF_WSEQIOPS, "wseqiops=%u" }, | |
3331 | { I_LCOEF_WRANDIOPS, "wrandiops=%u" }, | |
3332 | { NR_I_LCOEFS, NULL }, | |
3333 | }; | |
3334 | ||
3335 | static ssize_t ioc_cost_model_write(struct kernfs_open_file *of, char *input, | |
3336 | size_t nbytes, loff_t off) | |
3337 | { | |
22ae8ce8 | 3338 | struct block_device *bdev; |
7caa4715 TH |
3339 | struct ioc *ioc; |
3340 | u64 u[NR_I_LCOEFS]; | |
3341 | bool user; | |
3342 | char *p; | |
3343 | int ret; | |
3344 | ||
22ae8ce8 CH |
3345 | bdev = blkcg_conf_open_bdev(&input); |
3346 | if (IS_ERR(bdev)) | |
3347 | return PTR_ERR(bdev); | |
7caa4715 | 3348 | |
ed6cddef | 3349 | ioc = q_to_ioc(bdev_get_queue(bdev)); |
7caa4715 | 3350 | if (!ioc) { |
57b64554 | 3351 | ret = blk_iocost_init(bdev->bd_disk); |
7caa4715 TH |
3352 | if (ret) |
3353 | goto err; | |
ed6cddef | 3354 | ioc = q_to_ioc(bdev_get_queue(bdev)); |
7caa4715 TH |
3355 | } |
3356 | ||
3357 | spin_lock_irq(&ioc->lock); | |
3358 | memcpy(u, ioc->params.i_lcoefs, sizeof(u)); | |
3359 | user = ioc->user_cost_model; | |
3360 | spin_unlock_irq(&ioc->lock); | |
3361 | ||
3362 | while ((p = strsep(&input, " \t\n"))) { | |
3363 | substring_t args[MAX_OPT_ARGS]; | |
3364 | char buf[32]; | |
3365 | int tok; | |
3366 | u64 v; | |
3367 | ||
3368 | if (!*p) | |
3369 | continue; | |
3370 | ||
3371 | switch (match_token(p, cost_ctrl_tokens, args)) { | |
3372 | case COST_CTRL: | |
3373 | match_strlcpy(buf, &args[0], sizeof(buf)); | |
3374 | if (!strcmp(buf, "auto")) | |
3375 | user = false; | |
3376 | else if (!strcmp(buf, "user")) | |
3377 | user = true; | |
3378 | else | |
3379 | goto einval; | |
3380 | continue; | |
3381 | case COST_MODEL: | |
3382 | match_strlcpy(buf, &args[0], sizeof(buf)); | |
3383 | if (strcmp(buf, "linear")) | |
3384 | goto einval; | |
3385 | continue; | |
3386 | } | |
3387 | ||
3388 | tok = match_token(p, i_lcoef_tokens, args); | |
3389 | if (tok == NR_I_LCOEFS) | |
3390 | goto einval; | |
3391 | if (match_u64(&args[0], &v)) | |
3392 | goto einval; | |
3393 | u[tok] = v; | |
3394 | user = true; | |
3395 | } | |
3396 | ||
3397 | spin_lock_irq(&ioc->lock); | |
3398 | if (user) { | |
3399 | memcpy(ioc->params.i_lcoefs, u, sizeof(u)); | |
3400 | ioc->user_cost_model = true; | |
3401 | } else { | |
3402 | ioc->user_cost_model = false; | |
3403 | } | |
3404 | ioc_refresh_params(ioc, true); | |
3405 | spin_unlock_irq(&ioc->lock); | |
3406 | ||
22ae8ce8 | 3407 | blkdev_put_no_open(bdev); |
7caa4715 TH |
3408 | return nbytes; |
3409 | ||
3410 | einval: | |
3411 | ret = -EINVAL; | |
3412 | err: | |
22ae8ce8 | 3413 | blkdev_put_no_open(bdev); |
7caa4715 TH |
3414 | return ret; |
3415 | } | |
3416 | ||
3417 | static struct cftype ioc_files[] = { | |
3418 | { | |
3419 | .name = "weight", | |
3420 | .flags = CFTYPE_NOT_ON_ROOT, | |
3421 | .seq_show = ioc_weight_show, | |
3422 | .write = ioc_weight_write, | |
3423 | }, | |
3424 | { | |
3425 | .name = "cost.qos", | |
3426 | .flags = CFTYPE_ONLY_ON_ROOT, | |
3427 | .seq_show = ioc_qos_show, | |
3428 | .write = ioc_qos_write, | |
3429 | }, | |
3430 | { | |
3431 | .name = "cost.model", | |
3432 | .flags = CFTYPE_ONLY_ON_ROOT, | |
3433 | .seq_show = ioc_cost_model_show, | |
3434 | .write = ioc_cost_model_write, | |
3435 | }, | |
3436 | {} | |
3437 | }; | |
3438 | ||
3439 | static struct blkcg_policy blkcg_policy_iocost = { | |
3440 | .dfl_cftypes = ioc_files, | |
3441 | .cpd_alloc_fn = ioc_cpd_alloc, | |
3442 | .cpd_free_fn = ioc_cpd_free, | |
3443 | .pd_alloc_fn = ioc_pd_alloc, | |
3444 | .pd_init_fn = ioc_pd_init, | |
3445 | .pd_free_fn = ioc_pd_free, | |
97eb1975 | 3446 | .pd_stat_fn = ioc_pd_stat, |
7caa4715 TH |
3447 | }; |
3448 | ||
3449 | static int __init ioc_init(void) | |
3450 | { | |
3451 | return blkcg_policy_register(&blkcg_policy_iocost); | |
3452 | } | |
3453 | ||
3454 | static void __exit ioc_exit(void) | |
3455 | { | |
fa1c3eaf | 3456 | blkcg_policy_unregister(&blkcg_policy_iocost); |
7caa4715 TH |
3457 | } |
3458 | ||
3459 | module_init(ioc_init); | |
3460 | module_exit(ioc_exit); |