blob: 6a9a0f03a67bd9ac9839ba667fb407a03694a304 [file] [log] [blame]
Jens Axboee34cbd32016-11-09 12:36:15 -07001/*
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
32enum {
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
56static 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 */
65static 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
83static 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 */
97static bool wb_recent_wait(struct rq_wb *rwb)
98{
Jan Karadc3b17c2017-02-02 15:56:50 +010099 struct bdi_writeback *wb = &rwb->queue->backing_dev_info->wb;
Jens Axboee34cbd32016-11-09 12:36:15 -0700100
101 return time_before(jiffies, wb->dirty_sleep + HZ);
102}
103
104static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, bool is_kswapd)
105{
106 return &rwb->rq_wait[is_kswapd];
107}
108
109static 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
121void __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 */
168void 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 */
192static 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
Arnd Bergmann4121d382016-11-16 16:29:57 +0100250static inline bool stat_sample_valid(struct blk_rq_stat *stat)
Jens Axboee34cbd32016-11-09 12:36:15 -0700251{
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 */
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700258 return (stat[READ].nr_samples >= 1 &&
259 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
Jens Axboee34cbd32016-11-09 12:36:15 -0700260}
261
262static 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
273enum {
274 LAT_OK = 1,
275 LAT_UNKNOWN,
276 LAT_UNKNOWN_WRITES,
277 LAT_EXCEEDED,
278};
279
Omar Sandoval34dbad52017-03-21 08:56:08 -0700280static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
Jens Axboee34cbd32016-11-09 12:36:15 -0700281{
Jan Karadc3b17c2017-02-02 15:56:50 +0100282 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
Jens Axboee34cbd32016-11-09 12:36:15 -0700283 u64 thislat;
284
285 /*
286 * If our stored sync issue exceeds the window size, or it
287 * exceeds our min target AND we haven't logged any entries,
288 * flag the latency as exceeded. wbt works off completion latencies,
289 * but for a flooded device, a single sync IO can take a long time
290 * to complete after being issued. If this time exceeds our
291 * monitoring window AND we didn't see any other completions in that
292 * window, then count that sync IO as a violation of the latency.
293 */
294 thislat = rwb_sync_issue_lat(rwb);
295 if (thislat > rwb->cur_win_nsec ||
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700296 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700297 trace_wbt_lat(bdi, thislat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700298 return LAT_EXCEEDED;
299 }
300
301 /*
302 * No read/write mix, if stat isn't valid
303 */
304 if (!stat_sample_valid(stat)) {
305 /*
306 * If we had writes in this stat window and the window is
307 * current, we're only doing writes. If a task recently
308 * waited or still has writes in flights, consider us doing
309 * just writes as well.
310 */
Omar Sandoval34dbad52017-03-21 08:56:08 -0700311 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
312 wbt_inflight(rwb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700313 return LAT_UNKNOWN_WRITES;
314 return LAT_UNKNOWN;
315 }
316
317 /*
318 * If the 'min' latency exceeds our target, step down.
319 */
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700320 if (stat[READ].min > rwb->min_lat_nsec) {
321 trace_wbt_lat(bdi, stat[READ].min);
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700322 trace_wbt_stat(bdi, stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700323 return LAT_EXCEEDED;
324 }
325
326 if (rwb->scale_step)
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700327 trace_wbt_stat(bdi, stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700328
329 return LAT_OK;
330}
331
Jens Axboee34cbd32016-11-09 12:36:15 -0700332static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
333{
Jan Karadc3b17c2017-02-02 15:56:50 +0100334 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700335
336 trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
Jens Axboee34cbd32016-11-09 12:36:15 -0700337 rwb->wb_background, rwb->wb_normal, rwb->wb_max);
338}
339
340static void scale_up(struct rq_wb *rwb)
341{
342 /*
343 * Hit max in previous round, stop here
344 */
345 if (rwb->scaled_max)
346 return;
347
348 rwb->scale_step--;
349 rwb->unknown_cnt = 0;
Jens Axboee34cbd32016-11-09 12:36:15 -0700350
351 rwb->scaled_max = calc_wb_limits(rwb);
352
353 rwb_wake_all(rwb);
354
355 rwb_trace_step(rwb, "step up");
356}
357
358/*
359 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
360 * had a latency violation.
361 */
362static void scale_down(struct rq_wb *rwb, bool hard_throttle)
363{
364 /*
365 * Stop scaling down when we've hit the limit. This also prevents
366 * ->scale_step from going to crazy values, if the device can't
367 * keep up.
368 */
369 if (rwb->wb_max == 1)
370 return;
371
372 if (rwb->scale_step < 0 && hard_throttle)
373 rwb->scale_step = 0;
374 else
375 rwb->scale_step++;
376
377 rwb->scaled_max = false;
378 rwb->unknown_cnt = 0;
Jens Axboee34cbd32016-11-09 12:36:15 -0700379 calc_wb_limits(rwb);
380 rwb_trace_step(rwb, "step down");
381}
382
383static void rwb_arm_timer(struct rq_wb *rwb)
384{
Jens Axboee34cbd32016-11-09 12:36:15 -0700385 if (rwb->scale_step > 0) {
386 /*
387 * We should speed this up, using some variant of a fast
388 * integer inverse square root calculation. Since we only do
389 * this for every window expiration, it's not a huge deal,
390 * though.
391 */
392 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
393 int_sqrt((rwb->scale_step + 1) << 8));
394 } else {
395 /*
396 * For step < 0, we don't want to increase/decrease the
397 * window size.
398 */
399 rwb->cur_win_nsec = rwb->win_nsec;
400 }
401
Omar Sandoval34dbad52017-03-21 08:56:08 -0700402 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
Jens Axboee34cbd32016-11-09 12:36:15 -0700403}
404
Omar Sandoval34dbad52017-03-21 08:56:08 -0700405static void wb_timer_fn(struct blk_stat_callback *cb)
Jens Axboee34cbd32016-11-09 12:36:15 -0700406{
Omar Sandoval34dbad52017-03-21 08:56:08 -0700407 struct rq_wb *rwb = cb->data;
Jens Axboee34cbd32016-11-09 12:36:15 -0700408 unsigned int inflight = wbt_inflight(rwb);
409 int status;
410
Omar Sandoval34dbad52017-03-21 08:56:08 -0700411 status = latency_exceeded(rwb, cb->stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700412
Jan Karadc3b17c2017-02-02 15:56:50 +0100413 trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step,
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700414 inflight);
Jens Axboee34cbd32016-11-09 12:36:15 -0700415
416 /*
417 * If we exceeded the latency target, step down. If we did not,
418 * step one level up. If we don't know enough to say either exceeded
419 * or ok, then don't do anything.
420 */
421 switch (status) {
422 case LAT_EXCEEDED:
423 scale_down(rwb, true);
424 break;
425 case LAT_OK:
426 scale_up(rwb);
427 break;
428 case LAT_UNKNOWN_WRITES:
429 /*
430 * We started a the center step, but don't have a valid
431 * read/write sample, but we do have writes going on.
432 * Allow step to go negative, to increase write perf.
433 */
434 scale_up(rwb);
435 break;
436 case LAT_UNKNOWN:
437 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
438 break;
439 /*
440 * We get here when previously scaled reduced depth, and we
441 * currently don't have a valid read/write sample. For that
442 * case, slowly return to center state (step == 0).
443 */
444 if (rwb->scale_step > 0)
445 scale_up(rwb);
446 else if (rwb->scale_step < 0)
447 scale_down(rwb, false);
448 break;
449 default:
450 break;
451 }
452
453 /*
454 * Re-arm timer, if we have IO in flight
455 */
456 if (rwb->scale_step || inflight)
457 rwb_arm_timer(rwb);
458}
459
460void wbt_update_limits(struct rq_wb *rwb)
461{
462 rwb->scale_step = 0;
463 rwb->scaled_max = false;
464 calc_wb_limits(rwb);
465
466 rwb_wake_all(rwb);
467}
468
469static bool close_io(struct rq_wb *rwb)
470{
471 const unsigned long now = jiffies;
472
473 return time_before(now, rwb->last_issue + HZ / 10) ||
474 time_before(now, rwb->last_comp + HZ / 10);
475}
476
477#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
478
479static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
480{
481 unsigned int limit;
482
483 /*
484 * At this point we know it's a buffered write. If this is
485 * kswapd trying to free memory, or REQ_SYNC is set, set, then
486 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
487 * that. If the write is marked as a background write, then use
488 * the idle limit, or go to normal if we haven't had competing
489 * IO for a bit.
490 */
491 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
492 limit = rwb->wb_max;
493 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
494 /*
495 * If less than 100ms since we completed unrelated IO,
496 * limit us to half the depth for background writeback.
497 */
498 limit = rwb->wb_background;
499 } else
500 limit = rwb->wb_normal;
501
502 return limit;
503}
504
505static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
Ingo Molnarac6424b2017-06-20 12:06:13 +0200506 wait_queue_entry_t *wait, unsigned long rw)
Jens Axboee34cbd32016-11-09 12:36:15 -0700507{
508 /*
509 * inc it here even if disabled, since we'll dec it at completion.
510 * this only happens if the task was sleeping in __wbt_wait(),
511 * and someone turned it off at the same time.
512 */
513 if (!rwb_enabled(rwb)) {
514 atomic_inc(&rqw->inflight);
515 return true;
516 }
517
518 /*
519 * If the waitqueue is already active and we are not the next
520 * in line to be woken up, wait for our turn.
521 */
522 if (waitqueue_active(&rqw->wait) &&
Ingo Molnar2055da92017-06-20 12:06:46 +0200523 rqw->wait.head.next != &wait->entry)
Jens Axboee34cbd32016-11-09 12:36:15 -0700524 return false;
525
526 return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
527}
528
529/*
530 * Block if we will exceed our limit, or if we are currently waiting for
531 * the timer to kick off queuing again.
532 */
533static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock)
Bart Van Assche9eca5352017-01-02 09:48:47 -0700534 __releases(lock)
535 __acquires(lock)
Jens Axboee34cbd32016-11-09 12:36:15 -0700536{
537 struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd());
538 DEFINE_WAIT(wait);
539
540 if (may_queue(rwb, rqw, &wait, rw))
541 return;
542
543 do {
544 prepare_to_wait_exclusive(&rqw->wait, &wait,
545 TASK_UNINTERRUPTIBLE);
546
547 if (may_queue(rwb, rqw, &wait, rw))
548 break;
549
Bart Van Assche9eca5352017-01-02 09:48:47 -0700550 if (lock) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700551 spin_unlock_irq(lock);
Bart Van Assche9eca5352017-01-02 09:48:47 -0700552 io_schedule();
Jens Axboee34cbd32016-11-09 12:36:15 -0700553 spin_lock_irq(lock);
Bart Van Assche9eca5352017-01-02 09:48:47 -0700554 } else
555 io_schedule();
Jens Axboee34cbd32016-11-09 12:36:15 -0700556 } while (1);
557
558 finish_wait(&rqw->wait, &wait);
559}
560
561static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
562{
563 const int op = bio_op(bio);
564
565 /*
Christoph Hellwigbe07e142016-12-09 14:19:06 +0100566 * If not a WRITE, do nothing
Jens Axboee34cbd32016-11-09 12:36:15 -0700567 */
Christoph Hellwigbe07e142016-12-09 14:19:06 +0100568 if (op != REQ_OP_WRITE)
Jens Axboee34cbd32016-11-09 12:36:15 -0700569 return false;
570
571 /*
572 * Don't throttle WRITE_ODIRECT
573 */
574 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE))
575 return false;
576
577 return true;
578}
579
580/*
581 * Returns true if the IO request should be accounted, false if not.
582 * May sleep, if we have exceeded the writeback limits. Caller can pass
583 * in an irq held spinlock, if it holds one when calling this function.
584 * If we do sleep, we'll release and re-grab it.
585 */
Bart Van Asschef2e0a0b2017-01-02 09:46:15 -0700586enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
Jens Axboee34cbd32016-11-09 12:36:15 -0700587{
588 unsigned int ret = 0;
589
590 if (!rwb_enabled(rwb))
591 return 0;
592
593 if (bio_op(bio) == REQ_OP_READ)
594 ret = WBT_READ;
595
596 if (!wbt_should_throttle(rwb, bio)) {
597 if (ret & WBT_READ)
598 wb_timestamp(rwb, &rwb->last_issue);
599 return ret;
600 }
601
602 __wbt_wait(rwb, bio->bi_opf, lock);
603
Omar Sandoval34dbad52017-03-21 08:56:08 -0700604 if (!blk_stat_is_active(rwb->cb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700605 rwb_arm_timer(rwb);
606
607 if (current_is_kswapd())
608 ret |= WBT_KSWAPD;
609
610 return ret | WBT_TRACKED;
611}
612
613void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat)
614{
615 if (!rwb_enabled(rwb))
616 return;
617
618 /*
619 * Track sync issue, in case it takes a long time to complete. Allows
620 * us to react quicker, if a sync IO takes a long time to complete.
621 * Note that this is just a hint. 'stat' can go away when the
622 * request completes, so it's important we never dereference it. We
623 * only use the address to compare with, which is why we store the
624 * sync_issue time locally.
625 */
626 if (wbt_is_read(stat) && !rwb->sync_issue) {
627 rwb->sync_cookie = stat;
628 rwb->sync_issue = blk_stat_time(stat);
629 }
630}
631
632void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat)
633{
634 if (!rwb_enabled(rwb))
635 return;
636 if (stat == rwb->sync_cookie) {
637 rwb->sync_issue = 0;
638 rwb->sync_cookie = NULL;
639 }
640}
641
642void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
643{
644 if (rwb) {
645 rwb->queue_depth = depth;
646 wbt_update_limits(rwb);
647 }
648}
649
650void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
651{
652 if (rwb)
653 rwb->wc = write_cache_on;
654}
655
Jan Kara3f19cd22017-04-11 11:29:01 +0200656/*
657 * Disable wbt, if enabled by default. Only called from CFQ.
Jens Axboefa224ee2016-11-28 09:25:50 -0700658 */
659void wbt_disable_default(struct request_queue *q)
Jens Axboee34cbd32016-11-09 12:36:15 -0700660{
Jens Axboefa224ee2016-11-28 09:25:50 -0700661 struct rq_wb *rwb = q->rq_wb;
662
Jan Kara3f19cd22017-04-11 11:29:01 +0200663 if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT)
664 wbt_exit(q);
Jens Axboee34cbd32016-11-09 12:36:15 -0700665}
Jens Axboefa224ee2016-11-28 09:25:50 -0700666EXPORT_SYMBOL_GPL(wbt_disable_default);
Jens Axboee34cbd32016-11-09 12:36:15 -0700667
Jan Kara8330cdb2017-04-19 11:33:27 +0200668/*
669 * Enable wbt if defaults are configured that way
670 */
671void wbt_enable_default(struct request_queue *q)
672{
673 /* Throttling already enabled? */
674 if (q->rq_wb)
675 return;
676
677 /* Queue not registered? Maybe shutting down... */
678 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
679 return;
680
681 if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) ||
682 (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ)))
683 wbt_init(q);
684}
685EXPORT_SYMBOL_GPL(wbt_enable_default);
686
Jens Axboe80e091d2016-11-28 09:22:47 -0700687u64 wbt_default_latency_nsec(struct request_queue *q)
688{
689 /*
690 * We default to 2msec for non-rotational storage, and 75msec
691 * for rotational storage.
692 */
693 if (blk_queue_nonrot(q))
694 return 2000000ULL;
695 else
696 return 75000000ULL;
697}
698
Jens Axboe99c749a2017-04-21 07:55:42 -0600699static int wbt_data_dir(const struct request *rq)
700{
701 return rq_data_dir(rq);
702}
703
Jens Axboe8054b892016-11-10 21:50:51 -0700704int wbt_init(struct request_queue *q)
Jens Axboee34cbd32016-11-09 12:36:15 -0700705{
706 struct rq_wb *rwb;
707 int i;
708
Jens Axboee34cbd32016-11-09 12:36:15 -0700709 BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS);
710
Jens Axboee34cbd32016-11-09 12:36:15 -0700711 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
712 if (!rwb)
713 return -ENOMEM;
714
Jens Axboe99c749a2017-04-21 07:55:42 -0600715 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
Omar Sandoval34dbad52017-03-21 08:56:08 -0700716 if (!rwb->cb) {
717 kfree(rwb);
718 return -ENOMEM;
719 }
720
Jens Axboee34cbd32016-11-09 12:36:15 -0700721 for (i = 0; i < WBT_NUM_RWQ; i++) {
722 atomic_set(&rwb->rq_wait[i].inflight, 0);
723 init_waitqueue_head(&rwb->rq_wait[i].wait);
724 }
725
Jens Axboee34cbd32016-11-09 12:36:15 -0700726 rwb->wc = 1;
727 rwb->queue_depth = RWB_DEF_DEPTH;
728 rwb->last_comp = rwb->last_issue = jiffies;
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700729 rwb->queue = q;
Jens Axboee34cbd32016-11-09 12:36:15 -0700730 rwb->win_nsec = RWB_WINDOW_NSEC;
Jens Axboed62118b2016-11-28 09:40:34 -0700731 rwb->enable_state = WBT_STATE_ON_DEFAULT;
Jens Axboee34cbd32016-11-09 12:36:15 -0700732 wbt_update_limits(rwb);
733
734 /*
Omar Sandoval34dbad52017-03-21 08:56:08 -0700735 * Assign rwb and add the stats callback.
Jens Axboee34cbd32016-11-09 12:36:15 -0700736 */
737 q->rq_wb = rwb;
Omar Sandoval34dbad52017-03-21 08:56:08 -0700738 blk_stat_add_callback(q, rwb->cb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700739
Jens Axboe80e091d2016-11-28 09:22:47 -0700740 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
Jens Axboee34cbd32016-11-09 12:36:15 -0700741
742 wbt_set_queue_depth(rwb, blk_queue_depth(q));
743 wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
744
745 return 0;
746}
747
748void wbt_exit(struct request_queue *q)
749{
750 struct rq_wb *rwb = q->rq_wb;
751
752 if (rwb) {
Omar Sandoval34dbad52017-03-21 08:56:08 -0700753 blk_stat_remove_callback(q, rwb->cb);
754 blk_stat_free_callback(rwb->cb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700755 q->rq_wb = NULL;
756 kfree(rwb);
757 }
758}