]>
Commit | Line | Data |
---|---|---|
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->queue->backing_dev_info->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, | |
105 | enum wbt_flags wb_acct) | |
106 | { | |
107 | if (wb_acct & WBT_KSWAPD) | |
108 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
109 | ||
110 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
111 | } | |
112 | ||
113 | static void rwb_wake_all(struct rq_wb *rwb) | |
114 | { | |
115 | int i; | |
116 | ||
117 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
118 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
119 | ||
120 | if (waitqueue_active(&rqw->wait)) | |
121 | wake_up_all(&rqw->wait); | |
122 | } | |
123 | } | |
124 | ||
125 | void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct) | |
126 | { | |
127 | struct rq_wait *rqw; | |
128 | int inflight, limit; | |
129 | ||
130 | if (!(wb_acct & WBT_TRACKED)) | |
131 | return; | |
132 | ||
133 | rqw = get_rq_wait(rwb, wb_acct); | |
134 | inflight = atomic_dec_return(&rqw->inflight); | |
135 | ||
136 | /* | |
137 | * wbt got disabled with IO in flight. Wake up any potential | |
138 | * waiters, we don't have to do more than that. | |
139 | */ | |
140 | if (unlikely(!rwb_enabled(rwb))) { | |
141 | rwb_wake_all(rwb); | |
142 | return; | |
143 | } | |
144 | ||
145 | /* | |
146 | * If the device does write back caching, drop further down | |
147 | * before we wake people up. | |
148 | */ | |
149 | if (rwb->wc && !wb_recent_wait(rwb)) | |
150 | limit = 0; | |
151 | else | |
152 | limit = rwb->wb_normal; | |
153 | ||
154 | /* | |
155 | * Don't wake anyone up if we are above the normal limit. | |
156 | */ | |
157 | if (inflight && inflight >= limit) | |
158 | return; | |
159 | ||
160 | if (waitqueue_active(&rqw->wait)) { | |
161 | int diff = limit - inflight; | |
162 | ||
163 | if (!inflight || diff >= rwb->wb_background / 2) | |
164 | wake_up(&rqw->wait); | |
165 | } | |
166 | } | |
167 | ||
168 | /* | |
169 | * Called on completion of a request. Note that it's also called when | |
170 | * a request is merged, when the request gets freed. | |
171 | */ | |
172 | void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
173 | { | |
174 | if (!rwb) | |
175 | return; | |
176 | ||
177 | if (!wbt_is_tracked(stat)) { | |
178 | if (rwb->sync_cookie == stat) { | |
179 | rwb->sync_issue = 0; | |
180 | rwb->sync_cookie = NULL; | |
181 | } | |
182 | ||
183 | if (wbt_is_read(stat)) | |
184 | wb_timestamp(rwb, &rwb->last_comp); | |
185 | } else { | |
186 | WARN_ON_ONCE(stat == rwb->sync_cookie); | |
187 | __wbt_done(rwb, wbt_stat_to_mask(stat)); | |
188 | } | |
189 | wbt_clear_state(stat); | |
190 | } | |
191 | ||
192 | /* | |
193 | * Return true, if we can't increase the depth further by scaling | |
194 | */ | |
195 | static bool calc_wb_limits(struct rq_wb *rwb) | |
196 | { | |
197 | unsigned int depth; | |
198 | bool ret = false; | |
199 | ||
200 | if (!rwb->min_lat_nsec) { | |
201 | rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; | |
202 | return false; | |
203 | } | |
204 | ||
205 | /* | |
206 | * For QD=1 devices, this is a special case. It's important for those | |
207 | * to have one request ready when one completes, so force a depth of | |
208 | * 2 for those devices. On the backend, it'll be a depth of 1 anyway, | |
209 | * since the device can't have more than that in flight. If we're | |
210 | * scaling down, then keep a setting of 1/1/1. | |
211 | */ | |
212 | if (rwb->queue_depth == 1) { | |
213 | if (rwb->scale_step > 0) | |
214 | rwb->wb_max = rwb->wb_normal = 1; | |
215 | else { | |
216 | rwb->wb_max = rwb->wb_normal = 2; | |
217 | ret = true; | |
218 | } | |
219 | rwb->wb_background = 1; | |
220 | } else { | |
221 | /* | |
222 | * scale_step == 0 is our default state. If we have suffered | |
223 | * latency spikes, step will be > 0, and we shrink the | |
224 | * allowed write depths. If step is < 0, we're only doing | |
225 | * writes, and we allow a temporarily higher depth to | |
226 | * increase performance. | |
227 | */ | |
228 | depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth); | |
229 | if (rwb->scale_step > 0) | |
230 | depth = 1 + ((depth - 1) >> min(31, rwb->scale_step)); | |
231 | else if (rwb->scale_step < 0) { | |
232 | unsigned int maxd = 3 * rwb->queue_depth / 4; | |
233 | ||
234 | depth = 1 + ((depth - 1) << -rwb->scale_step); | |
235 | if (depth > maxd) { | |
236 | depth = maxd; | |
237 | ret = true; | |
238 | } | |
239 | } | |
240 | ||
241 | /* | |
242 | * Set our max/normal/bg queue depths based on how far | |
243 | * we have scaled down (->scale_step). | |
244 | */ | |
245 | rwb->wb_max = depth; | |
246 | rwb->wb_normal = (rwb->wb_max + 1) / 2; | |
247 | rwb->wb_background = (rwb->wb_max + 3) / 4; | |
248 | } | |
249 | ||
250 | return ret; | |
251 | } | |
252 | ||
253 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) | |
254 | { | |
255 | /* | |
256 | * We need at least one read sample, and a minimum of | |
257 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
258 | * that it's writes impacting us, and not just some sole read on | |
259 | * a device that is in a lower power state. | |
260 | */ | |
261 | return (stat[READ].nr_samples >= 1 && | |
262 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
263 | } | |
264 | ||
265 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
266 | { | |
267 | u64 now, issue = READ_ONCE(rwb->sync_issue); | |
268 | ||
269 | if (!issue || !rwb->sync_cookie) | |
270 | return 0; | |
271 | ||
272 | now = ktime_to_ns(ktime_get()); | |
273 | return now - issue; | |
274 | } | |
275 | ||
276 | enum { | |
277 | LAT_OK = 1, | |
278 | LAT_UNKNOWN, | |
279 | LAT_UNKNOWN_WRITES, | |
280 | LAT_EXCEEDED, | |
281 | }; | |
282 | ||
283 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) | |
284 | { | |
285 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; | |
286 | u64 thislat; | |
287 | ||
288 | /* | |
289 | * If our stored sync issue exceeds the window size, or it | |
290 | * exceeds our min target AND we haven't logged any entries, | |
291 | * flag the latency as exceeded. wbt works off completion latencies, | |
292 | * but for a flooded device, a single sync IO can take a long time | |
293 | * to complete after being issued. If this time exceeds our | |
294 | * monitoring window AND we didn't see any other completions in that | |
295 | * window, then count that sync IO as a violation of the latency. | |
296 | */ | |
297 | thislat = rwb_sync_issue_lat(rwb); | |
298 | if (thislat > rwb->cur_win_nsec || | |
299 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { | |
300 | trace_wbt_lat(bdi, thislat); | |
301 | return LAT_EXCEEDED; | |
302 | } | |
303 | ||
304 | /* | |
305 | * No read/write mix, if stat isn't valid | |
306 | */ | |
307 | if (!stat_sample_valid(stat)) { | |
308 | /* | |
309 | * If we had writes in this stat window and the window is | |
310 | * current, we're only doing writes. If a task recently | |
311 | * waited or still has writes in flights, consider us doing | |
312 | * just writes as well. | |
313 | */ | |
314 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || | |
315 | wbt_inflight(rwb)) | |
316 | return LAT_UNKNOWN_WRITES; | |
317 | return LAT_UNKNOWN; | |
318 | } | |
319 | ||
320 | /* | |
321 | * If the 'min' latency exceeds our target, step down. | |
322 | */ | |
323 | if (stat[READ].min > rwb->min_lat_nsec) { | |
324 | trace_wbt_lat(bdi, stat[READ].min); | |
325 | trace_wbt_stat(bdi, stat); | |
326 | return LAT_EXCEEDED; | |
327 | } | |
328 | ||
329 | if (rwb->scale_step) | |
330 | trace_wbt_stat(bdi, stat); | |
331 | ||
332 | return LAT_OK; | |
333 | } | |
334 | ||
335 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) | |
336 | { | |
337 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; | |
338 | ||
339 | trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec, | |
340 | rwb->wb_background, rwb->wb_normal, rwb->wb_max); | |
341 | } | |
342 | ||
343 | static void scale_up(struct rq_wb *rwb) | |
344 | { | |
345 | /* | |
346 | * Hit max in previous round, stop here | |
347 | */ | |
348 | if (rwb->scaled_max) | |
349 | return; | |
350 | ||
351 | rwb->scale_step--; | |
352 | rwb->unknown_cnt = 0; | |
353 | ||
354 | rwb->scaled_max = calc_wb_limits(rwb); | |
355 | ||
356 | rwb_wake_all(rwb); | |
357 | ||
358 | rwb_trace_step(rwb, "step up"); | |
359 | } | |
360 | ||
361 | /* | |
362 | * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we | |
363 | * had a latency violation. | |
364 | */ | |
365 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) | |
366 | { | |
367 | /* | |
368 | * Stop scaling down when we've hit the limit. This also prevents | |
369 | * ->scale_step from going to crazy values, if the device can't | |
370 | * keep up. | |
371 | */ | |
372 | if (rwb->wb_max == 1) | |
373 | return; | |
374 | ||
375 | if (rwb->scale_step < 0 && hard_throttle) | |
376 | rwb->scale_step = 0; | |
377 | else | |
378 | rwb->scale_step++; | |
379 | ||
380 | rwb->scaled_max = false; | |
381 | rwb->unknown_cnt = 0; | |
382 | calc_wb_limits(rwb); | |
383 | rwb_trace_step(rwb, "step down"); | |
384 | } | |
385 | ||
386 | static void rwb_arm_timer(struct rq_wb *rwb) | |
387 | { | |
388 | if (rwb->scale_step > 0) { | |
389 | /* | |
390 | * We should speed this up, using some variant of a fast | |
391 | * integer inverse square root calculation. Since we only do | |
392 | * this for every window expiration, it's not a huge deal, | |
393 | * though. | |
394 | */ | |
395 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
396 | int_sqrt((rwb->scale_step + 1) << 8)); | |
397 | } else { | |
398 | /* | |
399 | * For step < 0, we don't want to increase/decrease the | |
400 | * window size. | |
401 | */ | |
402 | rwb->cur_win_nsec = rwb->win_nsec; | |
403 | } | |
404 | ||
405 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); | |
406 | } | |
407 | ||
408 | static void wb_timer_fn(struct blk_stat_callback *cb) | |
409 | { | |
410 | struct rq_wb *rwb = cb->data; | |
411 | unsigned int inflight = wbt_inflight(rwb); | |
412 | int status; | |
413 | ||
414 | status = latency_exceeded(rwb, cb->stat); | |
415 | ||
416 | trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step, | |
417 | inflight); | |
418 | ||
419 | /* | |
420 | * If we exceeded the latency target, step down. If we did not, | |
421 | * step one level up. If we don't know enough to say either exceeded | |
422 | * or ok, then don't do anything. | |
423 | */ | |
424 | switch (status) { | |
425 | case LAT_EXCEEDED: | |
426 | scale_down(rwb, true); | |
427 | break; | |
428 | case LAT_OK: | |
429 | scale_up(rwb); | |
430 | break; | |
431 | case LAT_UNKNOWN_WRITES: | |
432 | /* | |
433 | * We started a the center step, but don't have a valid | |
434 | * read/write sample, but we do have writes going on. | |
435 | * Allow step to go negative, to increase write perf. | |
436 | */ | |
437 | scale_up(rwb); | |
438 | break; | |
439 | case LAT_UNKNOWN: | |
440 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
441 | break; | |
442 | /* | |
443 | * We get here when previously scaled reduced depth, and we | |
444 | * currently don't have a valid read/write sample. For that | |
445 | * case, slowly return to center state (step == 0). | |
446 | */ | |
447 | if (rwb->scale_step > 0) | |
448 | scale_up(rwb); | |
449 | else if (rwb->scale_step < 0) | |
450 | scale_down(rwb, false); | |
451 | break; | |
452 | default: | |
453 | break; | |
454 | } | |
455 | ||
456 | /* | |
457 | * Re-arm timer, if we have IO in flight | |
458 | */ | |
459 | if (rwb->scale_step || inflight) | |
460 | rwb_arm_timer(rwb); | |
461 | } | |
462 | ||
463 | void wbt_update_limits(struct rq_wb *rwb) | |
464 | { | |
465 | rwb->scale_step = 0; | |
466 | rwb->scaled_max = false; | |
467 | calc_wb_limits(rwb); | |
468 | ||
469 | rwb_wake_all(rwb); | |
470 | } | |
471 | ||
472 | static bool close_io(struct rq_wb *rwb) | |
473 | { | |
474 | const unsigned long now = jiffies; | |
475 | ||
476 | return time_before(now, rwb->last_issue + HZ / 10) || | |
477 | time_before(now, rwb->last_comp + HZ / 10); | |
478 | } | |
479 | ||
480 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
481 | ||
482 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
483 | { | |
484 | unsigned int limit; | |
485 | ||
486 | /* | |
487 | * If we got disabled, just return UINT_MAX. This ensures that | |
488 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
489 | */ | |
490 | if (!rwb_enabled(rwb)) | |
491 | return UINT_MAX; | |
492 | ||
493 | /* | |
494 | * At this point we know it's a buffered write. If this is | |
495 | * kswapd trying to free memory, or REQ_SYNC is set, then | |
496 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for | |
497 | * that. If the write is marked as a background write, then use | |
498 | * the idle limit, or go to normal if we haven't had competing | |
499 | * IO for a bit. | |
500 | */ | |
501 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
502 | limit = rwb->wb_max; | |
503 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { | |
504 | /* | |
505 | * If less than 100ms since we completed unrelated IO, | |
506 | * limit us to half the depth for background writeback. | |
507 | */ | |
508 | limit = rwb->wb_background; | |
509 | } else | |
510 | limit = rwb->wb_normal; | |
511 | ||
512 | return limit; | |
513 | } | |
514 | ||
515 | /* | |
516 | * Block if we will exceed our limit, or if we are currently waiting for | |
517 | * the timer to kick off queuing again. | |
518 | */ | |
519 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, | |
520 | unsigned long rw, spinlock_t *lock) | |
521 | __releases(lock) | |
522 | __acquires(lock) | |
523 | { | |
524 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); | |
525 | DECLARE_WAITQUEUE(wait, current); | |
526 | ||
527 | if (!waitqueue_active(&rqw->wait) | |
528 | && atomic_inc_below(&rqw->inflight, get_limit(rwb, rw))) | |
529 | return; | |
530 | ||
531 | add_wait_queue_exclusive(&rqw->wait, &wait); | |
532 | do { | |
533 | set_current_state(TASK_UNINTERRUPTIBLE); | |
534 | ||
535 | if (atomic_inc_below(&rqw->inflight, get_limit(rwb, rw))) | |
536 | break; | |
537 | ||
538 | if (lock) { | |
539 | spin_unlock_irq(lock); | |
540 | io_schedule(); | |
541 | spin_lock_irq(lock); | |
542 | } else | |
543 | io_schedule(); | |
544 | } while (1); | |
545 | ||
546 | __set_current_state(TASK_RUNNING); | |
547 | remove_wait_queue(&rqw->wait, &wait); | |
548 | } | |
549 | ||
550 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
551 | { | |
552 | const int op = bio_op(bio); | |
553 | ||
554 | /* | |
555 | * If not a WRITE, do nothing | |
556 | */ | |
557 | if (op != REQ_OP_WRITE) | |
558 | return false; | |
559 | ||
560 | /* | |
561 | * Don't throttle WRITE_ODIRECT | |
562 | */ | |
563 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE)) | |
564 | return false; | |
565 | ||
566 | return true; | |
567 | } | |
568 | ||
569 | /* | |
570 | * Returns true if the IO request should be accounted, false if not. | |
571 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
572 | * in an irq held spinlock, if it holds one when calling this function. | |
573 | * If we do sleep, we'll release and re-grab it. | |
574 | */ | |
575 | enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock) | |
576 | { | |
577 | enum wbt_flags ret = 0; | |
578 | ||
579 | if (!rwb_enabled(rwb)) | |
580 | return 0; | |
581 | ||
582 | if (bio_op(bio) == REQ_OP_READ) | |
583 | ret = WBT_READ; | |
584 | ||
585 | if (!wbt_should_throttle(rwb, bio)) { | |
586 | if (ret & WBT_READ) | |
587 | wb_timestamp(rwb, &rwb->last_issue); | |
588 | return ret; | |
589 | } | |
590 | ||
591 | if (current_is_kswapd()) | |
592 | ret |= WBT_KSWAPD; | |
593 | ||
594 | __wbt_wait(rwb, ret, bio->bi_opf, lock); | |
595 | ||
596 | if (!blk_stat_is_active(rwb->cb)) | |
597 | rwb_arm_timer(rwb); | |
598 | ||
599 | return ret | WBT_TRACKED; | |
600 | } | |
601 | ||
602 | void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
603 | { | |
604 | if (!rwb_enabled(rwb)) | |
605 | return; | |
606 | ||
607 | /* | |
608 | * Track sync issue, in case it takes a long time to complete. Allows | |
609 | * us to react quicker, if a sync IO takes a long time to complete. | |
610 | * Note that this is just a hint. 'stat' can go away when the | |
611 | * request completes, so it's important we never dereference it. We | |
612 | * only use the address to compare with, which is why we store the | |
613 | * sync_issue time locally. | |
614 | */ | |
615 | if (wbt_is_read(stat) && !rwb->sync_issue) { | |
616 | rwb->sync_cookie = stat; | |
617 | rwb->sync_issue = blk_stat_time(stat); | |
618 | } | |
619 | } | |
620 | ||
621 | void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
622 | { | |
623 | if (!rwb_enabled(rwb)) | |
624 | return; | |
625 | if (stat == rwb->sync_cookie) { | |
626 | rwb->sync_issue = 0; | |
627 | rwb->sync_cookie = NULL; | |
628 | } | |
629 | } | |
630 | ||
631 | void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) | |
632 | { | |
633 | if (rwb) { | |
634 | rwb->queue_depth = depth; | |
635 | wbt_update_limits(rwb); | |
636 | } | |
637 | } | |
638 | ||
639 | void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) | |
640 | { | |
641 | if (rwb) | |
642 | rwb->wc = write_cache_on; | |
643 | } | |
644 | ||
645 | /* | |
646 | * Disable wbt, if enabled by default. | |
647 | */ | |
648 | void wbt_disable_default(struct request_queue *q) | |
649 | { | |
650 | struct rq_wb *rwb = q->rq_wb; | |
651 | ||
652 | if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) | |
653 | wbt_exit(q); | |
654 | } | |
655 | EXPORT_SYMBOL_GPL(wbt_disable_default); | |
656 | ||
657 | /* | |
658 | * Enable wbt if defaults are configured that way | |
659 | */ | |
660 | void wbt_enable_default(struct request_queue *q) | |
661 | { | |
662 | /* Throttling already enabled? */ | |
663 | if (q->rq_wb) | |
664 | return; | |
665 | ||
666 | /* Queue not registered? Maybe shutting down... */ | |
667 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
668 | return; | |
669 | ||
670 | if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) || | |
671 | (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ))) | |
672 | wbt_init(q); | |
673 | } | |
674 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
675 | ||
676 | u64 wbt_default_latency_nsec(struct request_queue *q) | |
677 | { | |
678 | /* | |
679 | * We default to 2msec for non-rotational storage, and 75msec | |
680 | * for rotational storage. | |
681 | */ | |
682 | if (blk_queue_nonrot(q)) | |
683 | return 2000000ULL; | |
684 | else | |
685 | return 75000000ULL; | |
686 | } | |
687 | ||
688 | static int wbt_data_dir(const struct request *rq) | |
689 | { | |
690 | const int op = req_op(rq); | |
691 | ||
692 | if (op == REQ_OP_READ) | |
693 | return READ; | |
694 | else if (op == REQ_OP_WRITE || op == REQ_OP_FLUSH) | |
695 | return WRITE; | |
696 | ||
697 | /* don't account */ | |
698 | return -1; | |
699 | } | |
700 | ||
701 | int wbt_init(struct request_queue *q) | |
702 | { | |
703 | struct rq_wb *rwb; | |
704 | int i; | |
705 | ||
706 | BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS); | |
707 | ||
708 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); | |
709 | if (!rwb) | |
710 | return -ENOMEM; | |
711 | ||
712 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); | |
713 | if (!rwb->cb) { | |
714 | kfree(rwb); | |
715 | return -ENOMEM; | |
716 | } | |
717 | ||
718 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
719 | atomic_set(&rwb->rq_wait[i].inflight, 0); | |
720 | init_waitqueue_head(&rwb->rq_wait[i].wait); | |
721 | } | |
722 | ||
723 | rwb->last_comp = rwb->last_issue = jiffies; | |
724 | rwb->queue = q; | |
725 | rwb->win_nsec = RWB_WINDOW_NSEC; | |
726 | rwb->enable_state = WBT_STATE_ON_DEFAULT; | |
727 | wbt_update_limits(rwb); | |
728 | ||
729 | /* | |
730 | * Assign rwb and add the stats callback. | |
731 | */ | |
732 | q->rq_wb = rwb; | |
733 | blk_stat_add_callback(q, rwb->cb); | |
734 | ||
735 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); | |
736 | ||
737 | wbt_set_queue_depth(rwb, blk_queue_depth(q)); | |
738 | wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
739 | ||
740 | return 0; | |
741 | } | |
742 | ||
743 | void wbt_exit(struct request_queue *q) | |
744 | { | |
745 | struct rq_wb *rwb = q->rq_wb; | |
746 | ||
747 | if (rwb) { | |
748 | blk_stat_remove_callback(q, rwb->cb); | |
749 | blk_stat_free_callback(rwb->cb); | |
750 | q->rq_wb = NULL; | |
751 | kfree(rwb); | |
752 | } | |
753 | } |