]>
Commit | Line | Data |
---|---|---|
e34cbd30 JA |
1 | /* |
2 | * buffered writeback throttling. loosely based on CoDel. We can't drop | |
3 | * packets for IO scheduling, so the logic is something like this: | |
4 | * | |
5 | * - Monitor latencies in a defined window of time. | |
6 | * - If the minimum latency in the above window exceeds some target, increment | |
7 | * scaling step and scale down queue depth by a factor of 2x. The monitoring | |
8 | * window is then shrunk to 100 / sqrt(scaling step + 1). | |
9 | * - For any window where we don't have solid data on what the latencies | |
10 | * look like, retain status quo. | |
11 | * - If latencies look good, decrement scaling step. | |
12 | * - If we're only doing writes, allow the scaling step to go negative. This | |
13 | * will temporarily boost write performance, snapping back to a stable | |
14 | * scaling step of 0 if reads show up or the heavy writers finish. Unlike | |
15 | * positive scaling steps where we shrink the monitoring window, a negative | |
16 | * scaling step retains the default step==0 window size. | |
17 | * | |
18 | * Copyright (C) 2016 Jens Axboe | |
19 | * | |
20 | */ | |
21 | #include <linux/kernel.h> | |
22 | #include <linux/blk_types.h> | |
23 | #include <linux/slab.h> | |
24 | #include <linux/backing-dev.h> | |
25 | #include <linux/swap.h> | |
26 | ||
27 | #include "blk-wbt.h" | |
28 | ||
29 | #define CREATE_TRACE_POINTS | |
30 | #include <trace/events/wbt.h> | |
31 | ||
32 | enum { | |
33 | /* | |
34 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
35 | * from here depending on device stats | |
36 | */ | |
37 | RWB_DEF_DEPTH = 16, | |
38 | ||
39 | /* | |
40 | * 100msec window | |
41 | */ | |
42 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
43 | ||
44 | /* | |
45 | * Disregard stats, if we don't meet this minimum | |
46 | */ | |
47 | RWB_MIN_WRITE_SAMPLES = 3, | |
48 | ||
49 | /* | |
50 | * If we have this number of consecutive windows with not enough | |
51 | * information to scale up or down, scale up. | |
52 | */ | |
53 | RWB_UNKNOWN_BUMP = 5, | |
54 | }; | |
55 | ||
56 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
57 | { | |
58 | return rwb && rwb->wb_normal != 0; | |
59 | } | |
60 | ||
61 | /* | |
62 | * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded, | |
63 | * false if 'v' + 1 would be bigger than 'below'. | |
64 | */ | |
65 | static bool atomic_inc_below(atomic_t *v, int below) | |
66 | { | |
67 | int cur = atomic_read(v); | |
68 | ||
69 | for (;;) { | |
70 | int old; | |
71 | ||
72 | if (cur >= below) | |
73 | return false; | |
74 | old = atomic_cmpxchg(v, cur, cur + 1); | |
75 | if (old == cur) | |
76 | break; | |
77 | cur = old; | |
78 | } | |
79 | ||
80 | return true; | |
81 | } | |
82 | ||
83 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) | |
84 | { | |
85 | if (rwb_enabled(rwb)) { | |
86 | const unsigned long cur = jiffies; | |
87 | ||
88 | if (cur != *var) | |
89 | *var = cur; | |
90 | } | |
91 | } | |
92 | ||
93 | /* | |
94 | * If a task was rate throttled in balance_dirty_pages() within the last | |
95 | * second or so, use that to indicate a higher cleaning rate. | |
96 | */ | |
97 | static bool wb_recent_wait(struct rq_wb *rwb) | |
98 | { | |
99 | struct bdi_writeback *wb = &rwb->bdi->wb; | |
100 | ||
101 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
102 | } | |
103 | ||
104 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, bool is_kswapd) | |
105 | { | |
106 | return &rwb->rq_wait[is_kswapd]; | |
107 | } | |
108 | ||
109 | static void rwb_wake_all(struct rq_wb *rwb) | |
110 | { | |
111 | int i; | |
112 | ||
113 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
114 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
115 | ||
116 | if (waitqueue_active(&rqw->wait)) | |
117 | wake_up_all(&rqw->wait); | |
118 | } | |
119 | } | |
120 | ||
121 | void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct) | |
122 | { | |
123 | struct rq_wait *rqw; | |
124 | int inflight, limit; | |
125 | ||
126 | if (!(wb_acct & WBT_TRACKED)) | |
127 | return; | |
128 | ||
129 | rqw = get_rq_wait(rwb, wb_acct & WBT_KSWAPD); | |
130 | inflight = atomic_dec_return(&rqw->inflight); | |
131 | ||
132 | /* | |
133 | * wbt got disabled with IO in flight. Wake up any potential | |
134 | * waiters, we don't have to do more than that. | |
135 | */ | |
136 | if (unlikely(!rwb_enabled(rwb))) { | |
137 | rwb_wake_all(rwb); | |
138 | return; | |
139 | } | |
140 | ||
141 | /* | |
142 | * If the device does write back caching, drop further down | |
143 | * before we wake people up. | |
144 | */ | |
145 | if (rwb->wc && !wb_recent_wait(rwb)) | |
146 | limit = 0; | |
147 | else | |
148 | limit = rwb->wb_normal; | |
149 | ||
150 | /* | |
151 | * Don't wake anyone up if we are above the normal limit. | |
152 | */ | |
153 | if (inflight && inflight >= limit) | |
154 | return; | |
155 | ||
156 | if (waitqueue_active(&rqw->wait)) { | |
157 | int diff = limit - inflight; | |
158 | ||
159 | if (!inflight || diff >= rwb->wb_background / 2) | |
160 | wake_up_all(&rqw->wait); | |
161 | } | |
162 | } | |
163 | ||
164 | /* | |
165 | * Called on completion of a request. Note that it's also called when | |
166 | * a request is merged, when the request gets freed. | |
167 | */ | |
168 | void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
169 | { | |
170 | if (!rwb) | |
171 | return; | |
172 | ||
173 | if (!wbt_is_tracked(stat)) { | |
174 | if (rwb->sync_cookie == stat) { | |
175 | rwb->sync_issue = 0; | |
176 | rwb->sync_cookie = NULL; | |
177 | } | |
178 | ||
179 | if (wbt_is_read(stat)) | |
180 | wb_timestamp(rwb, &rwb->last_comp); | |
181 | wbt_clear_state(stat); | |
182 | } else { | |
183 | WARN_ON_ONCE(stat == rwb->sync_cookie); | |
184 | __wbt_done(rwb, wbt_stat_to_mask(stat)); | |
185 | wbt_clear_state(stat); | |
186 | } | |
187 | } | |
188 | ||
189 | /* | |
190 | * Return true, if we can't increase the depth further by scaling | |
191 | */ | |
192 | static bool calc_wb_limits(struct rq_wb *rwb) | |
193 | { | |
194 | unsigned int depth; | |
195 | bool ret = false; | |
196 | ||
197 | if (!rwb->min_lat_nsec) { | |
198 | rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; | |
199 | return false; | |
200 | } | |
201 | ||
202 | /* | |
203 | * For QD=1 devices, this is a special case. It's important for those | |
204 | * to have one request ready when one completes, so force a depth of | |
205 | * 2 for those devices. On the backend, it'll be a depth of 1 anyway, | |
206 | * since the device can't have more than that in flight. If we're | |
207 | * scaling down, then keep a setting of 1/1/1. | |
208 | */ | |
209 | if (rwb->queue_depth == 1) { | |
210 | if (rwb->scale_step > 0) | |
211 | rwb->wb_max = rwb->wb_normal = 1; | |
212 | else { | |
213 | rwb->wb_max = rwb->wb_normal = 2; | |
214 | ret = true; | |
215 | } | |
216 | rwb->wb_background = 1; | |
217 | } else { | |
218 | /* | |
219 | * scale_step == 0 is our default state. If we have suffered | |
220 | * latency spikes, step will be > 0, and we shrink the | |
221 | * allowed write depths. If step is < 0, we're only doing | |
222 | * writes, and we allow a temporarily higher depth to | |
223 | * increase performance. | |
224 | */ | |
225 | depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth); | |
226 | if (rwb->scale_step > 0) | |
227 | depth = 1 + ((depth - 1) >> min(31, rwb->scale_step)); | |
228 | else if (rwb->scale_step < 0) { | |
229 | unsigned int maxd = 3 * rwb->queue_depth / 4; | |
230 | ||
231 | depth = 1 + ((depth - 1) << -rwb->scale_step); | |
232 | if (depth > maxd) { | |
233 | depth = maxd; | |
234 | ret = true; | |
235 | } | |
236 | } | |
237 | ||
238 | /* | |
239 | * Set our max/normal/bg queue depths based on how far | |
240 | * we have scaled down (->scale_step). | |
241 | */ | |
242 | rwb->wb_max = depth; | |
243 | rwb->wb_normal = (rwb->wb_max + 1) / 2; | |
244 | rwb->wb_background = (rwb->wb_max + 3) / 4; | |
245 | } | |
246 | ||
247 | return ret; | |
248 | } | |
249 | ||
250 | static bool inline stat_sample_valid(struct blk_rq_stat *stat) | |
251 | { | |
252 | /* | |
253 | * We need at least one read sample, and a minimum of | |
254 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
255 | * that it's writes impacting us, and not just some sole read on | |
256 | * a device that is in a lower power state. | |
257 | */ | |
258 | return stat[0].nr_samples >= 1 && | |
259 | stat[1].nr_samples >= RWB_MIN_WRITE_SAMPLES; | |
260 | } | |
261 | ||
262 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
263 | { | |
264 | u64 now, issue = ACCESS_ONCE(rwb->sync_issue); | |
265 | ||
266 | if (!issue || !rwb->sync_cookie) | |
267 | return 0; | |
268 | ||
269 | now = ktime_to_ns(ktime_get()); | |
270 | return now - issue; | |
271 | } | |
272 | ||
273 | enum { | |
274 | LAT_OK = 1, | |
275 | LAT_UNKNOWN, | |
276 | LAT_UNKNOWN_WRITES, | |
277 | LAT_EXCEEDED, | |
278 | }; | |
279 | ||
280 | static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) | |
281 | { | |
282 | u64 thislat; | |
283 | ||
284 | /* | |
285 | * If our stored sync issue exceeds the window size, or it | |
286 | * exceeds our min target AND we haven't logged any entries, | |
287 | * flag the latency as exceeded. wbt works off completion latencies, | |
288 | * but for a flooded device, a single sync IO can take a long time | |
289 | * to complete after being issued. If this time exceeds our | |
290 | * monitoring window AND we didn't see any other completions in that | |
291 | * window, then count that sync IO as a violation of the latency. | |
292 | */ | |
293 | thislat = rwb_sync_issue_lat(rwb); | |
294 | if (thislat > rwb->cur_win_nsec || | |
295 | (thislat > rwb->min_lat_nsec && !stat[0].nr_samples)) { | |
296 | trace_wbt_lat(rwb->bdi, thislat); | |
297 | return LAT_EXCEEDED; | |
298 | } | |
299 | ||
300 | /* | |
301 | * No read/write mix, if stat isn't valid | |
302 | */ | |
303 | if (!stat_sample_valid(stat)) { | |
304 | /* | |
305 | * If we had writes in this stat window and the window is | |
306 | * current, we're only doing writes. If a task recently | |
307 | * waited or still has writes in flights, consider us doing | |
308 | * just writes as well. | |
309 | */ | |
310 | if ((stat[1].nr_samples && rwb->stat_ops->is_current(stat)) || | |
311 | wb_recent_wait(rwb) || wbt_inflight(rwb)) | |
312 | return LAT_UNKNOWN_WRITES; | |
313 | return LAT_UNKNOWN; | |
314 | } | |
315 | ||
316 | /* | |
317 | * If the 'min' latency exceeds our target, step down. | |
318 | */ | |
319 | if (stat[0].min > rwb->min_lat_nsec) { | |
320 | trace_wbt_lat(rwb->bdi, stat[0].min); | |
321 | trace_wbt_stat(rwb->bdi, stat); | |
322 | return LAT_EXCEEDED; | |
323 | } | |
324 | ||
325 | if (rwb->scale_step) | |
326 | trace_wbt_stat(rwb->bdi, stat); | |
327 | ||
328 | return LAT_OK; | |
329 | } | |
330 | ||
331 | static int latency_exceeded(struct rq_wb *rwb) | |
332 | { | |
333 | struct blk_rq_stat stat[2]; | |
334 | ||
335 | rwb->stat_ops->get(rwb->ops_data, stat); | |
336 | return __latency_exceeded(rwb, stat); | |
337 | } | |
338 | ||
339 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) | |
340 | { | |
341 | trace_wbt_step(rwb->bdi, msg, rwb->scale_step, rwb->cur_win_nsec, | |
342 | rwb->wb_background, rwb->wb_normal, rwb->wb_max); | |
343 | } | |
344 | ||
345 | static void scale_up(struct rq_wb *rwb) | |
346 | { | |
347 | /* | |
348 | * Hit max in previous round, stop here | |
349 | */ | |
350 | if (rwb->scaled_max) | |
351 | return; | |
352 | ||
353 | rwb->scale_step--; | |
354 | rwb->unknown_cnt = 0; | |
355 | rwb->stat_ops->clear(rwb->ops_data); | |
356 | ||
357 | rwb->scaled_max = calc_wb_limits(rwb); | |
358 | ||
359 | rwb_wake_all(rwb); | |
360 | ||
361 | rwb_trace_step(rwb, "step up"); | |
362 | } | |
363 | ||
364 | /* | |
365 | * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we | |
366 | * had a latency violation. | |
367 | */ | |
368 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) | |
369 | { | |
370 | /* | |
371 | * Stop scaling down when we've hit the limit. This also prevents | |
372 | * ->scale_step from going to crazy values, if the device can't | |
373 | * keep up. | |
374 | */ | |
375 | if (rwb->wb_max == 1) | |
376 | return; | |
377 | ||
378 | if (rwb->scale_step < 0 && hard_throttle) | |
379 | rwb->scale_step = 0; | |
380 | else | |
381 | rwb->scale_step++; | |
382 | ||
383 | rwb->scaled_max = false; | |
384 | rwb->unknown_cnt = 0; | |
385 | rwb->stat_ops->clear(rwb->ops_data); | |
386 | calc_wb_limits(rwb); | |
387 | rwb_trace_step(rwb, "step down"); | |
388 | } | |
389 | ||
390 | static void rwb_arm_timer(struct rq_wb *rwb) | |
391 | { | |
392 | unsigned long expires; | |
393 | ||
394 | if (rwb->scale_step > 0) { | |
395 | /* | |
396 | * We should speed this up, using some variant of a fast | |
397 | * integer inverse square root calculation. Since we only do | |
398 | * this for every window expiration, it's not a huge deal, | |
399 | * though. | |
400 | */ | |
401 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
402 | int_sqrt((rwb->scale_step + 1) << 8)); | |
403 | } else { | |
404 | /* | |
405 | * For step < 0, we don't want to increase/decrease the | |
406 | * window size. | |
407 | */ | |
408 | rwb->cur_win_nsec = rwb->win_nsec; | |
409 | } | |
410 | ||
411 | expires = jiffies + nsecs_to_jiffies(rwb->cur_win_nsec); | |
412 | mod_timer(&rwb->window_timer, expires); | |
413 | } | |
414 | ||
415 | static void wb_timer_fn(unsigned long data) | |
416 | { | |
417 | struct rq_wb *rwb = (struct rq_wb *) data; | |
418 | unsigned int inflight = wbt_inflight(rwb); | |
419 | int status; | |
420 | ||
421 | status = latency_exceeded(rwb); | |
422 | ||
423 | trace_wbt_timer(rwb->bdi, status, rwb->scale_step, inflight); | |
424 | ||
425 | /* | |
426 | * If we exceeded the latency target, step down. If we did not, | |
427 | * step one level up. If we don't know enough to say either exceeded | |
428 | * or ok, then don't do anything. | |
429 | */ | |
430 | switch (status) { | |
431 | case LAT_EXCEEDED: | |
432 | scale_down(rwb, true); | |
433 | break; | |
434 | case LAT_OK: | |
435 | scale_up(rwb); | |
436 | break; | |
437 | case LAT_UNKNOWN_WRITES: | |
438 | /* | |
439 | * We started a the center step, but don't have a valid | |
440 | * read/write sample, but we do have writes going on. | |
441 | * Allow step to go negative, to increase write perf. | |
442 | */ | |
443 | scale_up(rwb); | |
444 | break; | |
445 | case LAT_UNKNOWN: | |
446 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
447 | break; | |
448 | /* | |
449 | * We get here when previously scaled reduced depth, and we | |
450 | * currently don't have a valid read/write sample. For that | |
451 | * case, slowly return to center state (step == 0). | |
452 | */ | |
453 | if (rwb->scale_step > 0) | |
454 | scale_up(rwb); | |
455 | else if (rwb->scale_step < 0) | |
456 | scale_down(rwb, false); | |
457 | break; | |
458 | default: | |
459 | break; | |
460 | } | |
461 | ||
462 | /* | |
463 | * Re-arm timer, if we have IO in flight | |
464 | */ | |
465 | if (rwb->scale_step || inflight) | |
466 | rwb_arm_timer(rwb); | |
467 | } | |
468 | ||
469 | void wbt_update_limits(struct rq_wb *rwb) | |
470 | { | |
471 | rwb->scale_step = 0; | |
472 | rwb->scaled_max = false; | |
473 | calc_wb_limits(rwb); | |
474 | ||
475 | rwb_wake_all(rwb); | |
476 | } | |
477 | ||
478 | static bool close_io(struct rq_wb *rwb) | |
479 | { | |
480 | const unsigned long now = jiffies; | |
481 | ||
482 | return time_before(now, rwb->last_issue + HZ / 10) || | |
483 | time_before(now, rwb->last_comp + HZ / 10); | |
484 | } | |
485 | ||
486 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
487 | ||
488 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
489 | { | |
490 | unsigned int limit; | |
491 | ||
492 | /* | |
493 | * At this point we know it's a buffered write. If this is | |
494 | * kswapd trying to free memory, or REQ_SYNC is set, set, then | |
495 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for | |
496 | * that. If the write is marked as a background write, then use | |
497 | * the idle limit, or go to normal if we haven't had competing | |
498 | * IO for a bit. | |
499 | */ | |
500 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
501 | limit = rwb->wb_max; | |
502 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { | |
503 | /* | |
504 | * If less than 100ms since we completed unrelated IO, | |
505 | * limit us to half the depth for background writeback. | |
506 | */ | |
507 | limit = rwb->wb_background; | |
508 | } else | |
509 | limit = rwb->wb_normal; | |
510 | ||
511 | return limit; | |
512 | } | |
513 | ||
514 | static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw, | |
515 | wait_queue_t *wait, unsigned long rw) | |
516 | { | |
517 | /* | |
518 | * inc it here even if disabled, since we'll dec it at completion. | |
519 | * this only happens if the task was sleeping in __wbt_wait(), | |
520 | * and someone turned it off at the same time. | |
521 | */ | |
522 | if (!rwb_enabled(rwb)) { | |
523 | atomic_inc(&rqw->inflight); | |
524 | return true; | |
525 | } | |
526 | ||
527 | /* | |
528 | * If the waitqueue is already active and we are not the next | |
529 | * in line to be woken up, wait for our turn. | |
530 | */ | |
531 | if (waitqueue_active(&rqw->wait) && | |
532 | rqw->wait.task_list.next != &wait->task_list) | |
533 | return false; | |
534 | ||
535 | return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw)); | |
536 | } | |
537 | ||
538 | /* | |
539 | * Block if we will exceed our limit, or if we are currently waiting for | |
540 | * the timer to kick off queuing again. | |
541 | */ | |
542 | static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock) | |
543 | { | |
544 | struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd()); | |
545 | DEFINE_WAIT(wait); | |
546 | ||
547 | if (may_queue(rwb, rqw, &wait, rw)) | |
548 | return; | |
549 | ||
550 | do { | |
551 | prepare_to_wait_exclusive(&rqw->wait, &wait, | |
552 | TASK_UNINTERRUPTIBLE); | |
553 | ||
554 | if (may_queue(rwb, rqw, &wait, rw)) | |
555 | break; | |
556 | ||
557 | if (lock) | |
558 | spin_unlock_irq(lock); | |
559 | ||
560 | io_schedule(); | |
561 | ||
562 | if (lock) | |
563 | spin_lock_irq(lock); | |
564 | } while (1); | |
565 | ||
566 | finish_wait(&rqw->wait, &wait); | |
567 | } | |
568 | ||
569 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
570 | { | |
571 | const int op = bio_op(bio); | |
572 | ||
573 | /* | |
574 | * If not a WRITE (or a discard), do nothing | |
575 | */ | |
576 | if (!(op == REQ_OP_WRITE || op == REQ_OP_DISCARD)) | |
577 | return false; | |
578 | ||
579 | /* | |
580 | * Don't throttle WRITE_ODIRECT | |
581 | */ | |
582 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE)) | |
583 | return false; | |
584 | ||
585 | return true; | |
586 | } | |
587 | ||
588 | /* | |
589 | * Returns true if the IO request should be accounted, false if not. | |
590 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
591 | * in an irq held spinlock, if it holds one when calling this function. | |
592 | * If we do sleep, we'll release and re-grab it. | |
593 | */ | |
594 | unsigned int wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock) | |
595 | { | |
596 | unsigned int ret = 0; | |
597 | ||
598 | if (!rwb_enabled(rwb)) | |
599 | return 0; | |
600 | ||
601 | if (bio_op(bio) == REQ_OP_READ) | |
602 | ret = WBT_READ; | |
603 | ||
604 | if (!wbt_should_throttle(rwb, bio)) { | |
605 | if (ret & WBT_READ) | |
606 | wb_timestamp(rwb, &rwb->last_issue); | |
607 | return ret; | |
608 | } | |
609 | ||
610 | __wbt_wait(rwb, bio->bi_opf, lock); | |
611 | ||
612 | if (!timer_pending(&rwb->window_timer)) | |
613 | rwb_arm_timer(rwb); | |
614 | ||
615 | if (current_is_kswapd()) | |
616 | ret |= WBT_KSWAPD; | |
617 | ||
618 | return ret | WBT_TRACKED; | |
619 | } | |
620 | ||
621 | void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
622 | { | |
623 | if (!rwb_enabled(rwb)) | |
624 | return; | |
625 | ||
626 | /* | |
627 | * Track sync issue, in case it takes a long time to complete. Allows | |
628 | * us to react quicker, if a sync IO takes a long time to complete. | |
629 | * Note that this is just a hint. 'stat' can go away when the | |
630 | * request completes, so it's important we never dereference it. We | |
631 | * only use the address to compare with, which is why we store the | |
632 | * sync_issue time locally. | |
633 | */ | |
634 | if (wbt_is_read(stat) && !rwb->sync_issue) { | |
635 | rwb->sync_cookie = stat; | |
636 | rwb->sync_issue = blk_stat_time(stat); | |
637 | } | |
638 | } | |
639 | ||
640 | void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
641 | { | |
642 | if (!rwb_enabled(rwb)) | |
643 | return; | |
644 | if (stat == rwb->sync_cookie) { | |
645 | rwb->sync_issue = 0; | |
646 | rwb->sync_cookie = NULL; | |
647 | } | |
648 | } | |
649 | ||
650 | void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) | |
651 | { | |
652 | if (rwb) { | |
653 | rwb->queue_depth = depth; | |
654 | wbt_update_limits(rwb); | |
655 | } | |
656 | } | |
657 | ||
658 | void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) | |
659 | { | |
660 | if (rwb) | |
661 | rwb->wc = write_cache_on; | |
662 | } | |
663 | ||
664 | void wbt_disable(struct rq_wb *rwb) | |
665 | { | |
666 | if (rwb) { | |
667 | del_timer_sync(&rwb->window_timer); | |
668 | rwb->win_nsec = rwb->min_lat_nsec = 0; | |
669 | wbt_update_limits(rwb); | |
670 | } | |
671 | } | |
672 | EXPORT_SYMBOL_GPL(wbt_disable); | |
673 | ||
674 | int wbt_init(struct request_queue *q, struct wb_stat_ops *ops) | |
675 | { | |
676 | struct rq_wb *rwb; | |
677 | int i; | |
678 | ||
679 | /* | |
680 | * For now, we depend on the stats window being larger than | |
681 | * our monitoring window. Ensure that this isn't inadvertently | |
682 | * violated. | |
683 | */ | |
684 | BUILD_BUG_ON(RWB_WINDOW_NSEC > BLK_STAT_NSEC); | |
685 | BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS); | |
686 | ||
687 | if (!ops->get || !ops->is_current || !ops->clear) | |
688 | return -EINVAL; | |
689 | ||
690 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); | |
691 | if (!rwb) | |
692 | return -ENOMEM; | |
693 | ||
694 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
695 | atomic_set(&rwb->rq_wait[i].inflight, 0); | |
696 | init_waitqueue_head(&rwb->rq_wait[i].wait); | |
697 | } | |
698 | ||
699 | setup_timer(&rwb->window_timer, wb_timer_fn, (unsigned long) rwb); | |
700 | rwb->wc = 1; | |
701 | rwb->queue_depth = RWB_DEF_DEPTH; | |
702 | rwb->last_comp = rwb->last_issue = jiffies; | |
703 | rwb->bdi = &q->backing_dev_info; | |
704 | rwb->win_nsec = RWB_WINDOW_NSEC; | |
705 | rwb->stat_ops = ops; | |
706 | rwb->ops_data = q; | |
707 | wbt_update_limits(rwb); | |
708 | ||
709 | /* | |
710 | * Assign rwb, and turn on stats tracking for this queue | |
711 | */ | |
712 | q->rq_wb = rwb; | |
713 | blk_stat_enable(q); | |
714 | ||
715 | if (blk_queue_nonrot(q)) | |
716 | rwb->min_lat_nsec = 2000000ULL; | |
717 | else | |
718 | rwb->min_lat_nsec = 75000000ULL; | |
719 | ||
720 | wbt_set_queue_depth(rwb, blk_queue_depth(q)); | |
721 | wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
722 | ||
723 | return 0; | |
724 | } | |
725 | ||
726 | void wbt_exit(struct request_queue *q) | |
727 | { | |
728 | struct rq_wb *rwb = q->rq_wb; | |
729 | ||
730 | if (rwb) { | |
731 | del_timer_sync(&rwb->window_timer); | |
732 | q->rq_wb = NULL; | |
733 | kfree(rwb); | |
734 | } | |
735 | } |