Loading...
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
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{
99 struct bdi_writeback *wb = &rwb->queue->backing_dev_info->wb;
100
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 } else {
182 WARN_ON_ONCE(stat == rwb->sync_cookie);
183 __wbt_done(rwb, wbt_stat_to_mask(stat));
184 }
185 wbt_clear_state(stat);
186}
187
188/*
189 * Return true, if we can't increase the depth further by scaling
190 */
191static bool calc_wb_limits(struct rq_wb *rwb)
192{
193 unsigned int depth;
194 bool ret = false;
195
196 if (!rwb->min_lat_nsec) {
197 rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0;
198 return false;
199 }
200
201 /*
202 * For QD=1 devices, this is a special case. It's important for those
203 * to have one request ready when one completes, so force a depth of
204 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
205 * since the device can't have more than that in flight. If we're
206 * scaling down, then keep a setting of 1/1/1.
207 */
208 if (rwb->queue_depth == 1) {
209 if (rwb->scale_step > 0)
210 rwb->wb_max = rwb->wb_normal = 1;
211 else {
212 rwb->wb_max = rwb->wb_normal = 2;
213 ret = true;
214 }
215 rwb->wb_background = 1;
216 } else {
217 /*
218 * scale_step == 0 is our default state. If we have suffered
219 * latency spikes, step will be > 0, and we shrink the
220 * allowed write depths. If step is < 0, we're only doing
221 * writes, and we allow a temporarily higher depth to
222 * increase performance.
223 */
224 depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth);
225 if (rwb->scale_step > 0)
226 depth = 1 + ((depth - 1) >> min(31, rwb->scale_step));
227 else if (rwb->scale_step < 0) {
228 unsigned int maxd = 3 * rwb->queue_depth / 4;
229
230 depth = 1 + ((depth - 1) << -rwb->scale_step);
231 if (depth > maxd) {
232 depth = maxd;
233 ret = true;
234 }
235 }
236
237 /*
238 * Set our max/normal/bg queue depths based on how far
239 * we have scaled down (->scale_step).
240 */
241 rwb->wb_max = depth;
242 rwb->wb_normal = (rwb->wb_max + 1) / 2;
243 rwb->wb_background = (rwb->wb_max + 3) / 4;
244 }
245
246 return ret;
247}
248
249static inline bool stat_sample_valid(struct blk_rq_stat *stat)
250{
251 /*
252 * We need at least one read sample, and a minimum of
253 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
254 * that it's writes impacting us, and not just some sole read on
255 * a device that is in a lower power state.
256 */
257 return (stat[READ].nr_samples >= 1 &&
258 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
259}
260
261static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
262{
263 u64 now, issue = READ_ONCE(rwb->sync_issue);
264
265 if (!issue || !rwb->sync_cookie)
266 return 0;
267
268 now = ktime_to_ns(ktime_get());
269 return now - issue;
270}
271
272enum {
273 LAT_OK = 1,
274 LAT_UNKNOWN,
275 LAT_UNKNOWN_WRITES,
276 LAT_EXCEEDED,
277};
278
279static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
280{
281 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
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[READ].nr_samples)) {
296 trace_wbt_lat(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[WRITE].nr_samples || wb_recent_wait(rwb) ||
311 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[READ].min > rwb->min_lat_nsec) {
320 trace_wbt_lat(bdi, stat[READ].min);
321 trace_wbt_stat(bdi, stat);
322 return LAT_EXCEEDED;
323 }
324
325 if (rwb->scale_step)
326 trace_wbt_stat(bdi, stat);
327
328 return LAT_OK;
329}
330
331static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
332{
333 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
334
335 trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
336 rwb->wb_background, rwb->wb_normal, rwb->wb_max);
337}
338
339static void scale_up(struct rq_wb *rwb)
340{
341 /*
342 * Hit max in previous round, stop here
343 */
344 if (rwb->scaled_max)
345 return;
346
347 rwb->scale_step--;
348 rwb->unknown_cnt = 0;
349
350 rwb->scaled_max = calc_wb_limits(rwb);
351
352 rwb_wake_all(rwb);
353
354 rwb_trace_step(rwb, "step up");
355}
356
357/*
358 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
359 * had a latency violation.
360 */
361static void scale_down(struct rq_wb *rwb, bool hard_throttle)
362{
363 /*
364 * Stop scaling down when we've hit the limit. This also prevents
365 * ->scale_step from going to crazy values, if the device can't
366 * keep up.
367 */
368 if (rwb->wb_max == 1)
369 return;
370
371 if (rwb->scale_step < 0 && hard_throttle)
372 rwb->scale_step = 0;
373 else
374 rwb->scale_step++;
375
376 rwb->scaled_max = false;
377 rwb->unknown_cnt = 0;
378 calc_wb_limits(rwb);
379 rwb_trace_step(rwb, "step down");
380}
381
382static void rwb_arm_timer(struct rq_wb *rwb)
383{
384 if (rwb->scale_step > 0) {
385 /*
386 * We should speed this up, using some variant of a fast
387 * integer inverse square root calculation. Since we only do
388 * this for every window expiration, it's not a huge deal,
389 * though.
390 */
391 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
392 int_sqrt((rwb->scale_step + 1) << 8));
393 } else {
394 /*
395 * For step < 0, we don't want to increase/decrease the
396 * window size.
397 */
398 rwb->cur_win_nsec = rwb->win_nsec;
399 }
400
401 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
402}
403
404static void wb_timer_fn(struct blk_stat_callback *cb)
405{
406 struct rq_wb *rwb = cb->data;
407 unsigned int inflight = wbt_inflight(rwb);
408 int status;
409
410 status = latency_exceeded(rwb, cb->stat);
411
412 trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step,
413 inflight);
414
415 /*
416 * If we exceeded the latency target, step down. If we did not,
417 * step one level up. If we don't know enough to say either exceeded
418 * or ok, then don't do anything.
419 */
420 switch (status) {
421 case LAT_EXCEEDED:
422 scale_down(rwb, true);
423 break;
424 case LAT_OK:
425 scale_up(rwb);
426 break;
427 case LAT_UNKNOWN_WRITES:
428 /*
429 * We started a the center step, but don't have a valid
430 * read/write sample, but we do have writes going on.
431 * Allow step to go negative, to increase write perf.
432 */
433 scale_up(rwb);
434 break;
435 case LAT_UNKNOWN:
436 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
437 break;
438 /*
439 * We get here when previously scaled reduced depth, and we
440 * currently don't have a valid read/write sample. For that
441 * case, slowly return to center state (step == 0).
442 */
443 if (rwb->scale_step > 0)
444 scale_up(rwb);
445 else if (rwb->scale_step < 0)
446 scale_down(rwb, false);
447 break;
448 default:
449 break;
450 }
451
452 /*
453 * Re-arm timer, if we have IO in flight
454 */
455 if (rwb->scale_step || inflight)
456 rwb_arm_timer(rwb);
457}
458
459void wbt_update_limits(struct rq_wb *rwb)
460{
461 rwb->scale_step = 0;
462 rwb->scaled_max = false;
463 calc_wb_limits(rwb);
464
465 rwb_wake_all(rwb);
466}
467
468static bool close_io(struct rq_wb *rwb)
469{
470 const unsigned long now = jiffies;
471
472 return time_before(now, rwb->last_issue + HZ / 10) ||
473 time_before(now, rwb->last_comp + HZ / 10);
474}
475
476#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
477
478static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
479{
480 unsigned int limit;
481
482 /*
483 * At this point we know it's a buffered write. If this is
484 * kswapd trying to free memory, or REQ_SYNC is set, then
485 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
486 * that. If the write is marked as a background write, then use
487 * the idle limit, or go to normal if we haven't had competing
488 * IO for a bit.
489 */
490 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
491 limit = rwb->wb_max;
492 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
493 /*
494 * If less than 100ms since we completed unrelated IO,
495 * limit us to half the depth for background writeback.
496 */
497 limit = rwb->wb_background;
498 } else
499 limit = rwb->wb_normal;
500
501 return limit;
502}
503
504static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
505 wait_queue_entry_t *wait, unsigned long rw)
506{
507 /*
508 * inc it here even if disabled, since we'll dec it at completion.
509 * this only happens if the task was sleeping in __wbt_wait(),
510 * and someone turned it off at the same time.
511 */
512 if (!rwb_enabled(rwb)) {
513 atomic_inc(&rqw->inflight);
514 return true;
515 }
516
517 /*
518 * If the waitqueue is already active and we are not the next
519 * in line to be woken up, wait for our turn.
520 */
521 if (waitqueue_active(&rqw->wait) &&
522 rqw->wait.head.next != &wait->entry)
523 return false;
524
525 return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
526}
527
528/*
529 * Block if we will exceed our limit, or if we are currently waiting for
530 * the timer to kick off queuing again.
531 */
532static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock)
533 __releases(lock)
534 __acquires(lock)
535{
536 struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd());
537 DEFINE_WAIT(wait);
538
539 if (may_queue(rwb, rqw, &wait, rw))
540 return;
541
542 do {
543 prepare_to_wait_exclusive(&rqw->wait, &wait,
544 TASK_UNINTERRUPTIBLE);
545
546 if (may_queue(rwb, rqw, &wait, rw))
547 break;
548
549 if (lock) {
550 spin_unlock_irq(lock);
551 io_schedule();
552 spin_lock_irq(lock);
553 } else
554 io_schedule();
555 } while (1);
556
557 finish_wait(&rqw->wait, &wait);
558}
559
560static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
561{
562 const int op = bio_op(bio);
563
564 /*
565 * If not a WRITE, do nothing
566 */
567 if (op != REQ_OP_WRITE)
568 return false;
569
570 /*
571 * Don't throttle WRITE_ODIRECT
572 */
573 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE))
574 return false;
575
576 return true;
577}
578
579/*
580 * Returns true if the IO request should be accounted, false if not.
581 * May sleep, if we have exceeded the writeback limits. Caller can pass
582 * in an irq held spinlock, if it holds one when calling this function.
583 * If we do sleep, we'll release and re-grab it.
584 */
585enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
586{
587 unsigned int ret = 0;
588
589 if (!rwb_enabled(rwb))
590 return 0;
591
592 if (bio_op(bio) == REQ_OP_READ)
593 ret = WBT_READ;
594
595 if (!wbt_should_throttle(rwb, bio)) {
596 if (ret & WBT_READ)
597 wb_timestamp(rwb, &rwb->last_issue);
598 return ret;
599 }
600
601 __wbt_wait(rwb, bio->bi_opf, lock);
602
603 if (!blk_stat_is_active(rwb->cb))
604 rwb_arm_timer(rwb);
605
606 if (current_is_kswapd())
607 ret |= WBT_KSWAPD;
608
609 return ret | WBT_TRACKED;
610}
611
612void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat)
613{
614 if (!rwb_enabled(rwb))
615 return;
616
617 /*
618 * Track sync issue, in case it takes a long time to complete. Allows
619 * us to react quicker, if a sync IO takes a long time to complete.
620 * Note that this is just a hint. 'stat' can go away when the
621 * request completes, so it's important we never dereference it. We
622 * only use the address to compare with, which is why we store the
623 * sync_issue time locally.
624 */
625 if (wbt_is_read(stat) && !rwb->sync_issue) {
626 rwb->sync_cookie = stat;
627 rwb->sync_issue = blk_stat_time(stat);
628 }
629}
630
631void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat)
632{
633 if (!rwb_enabled(rwb))
634 return;
635 if (stat == rwb->sync_cookie) {
636 rwb->sync_issue = 0;
637 rwb->sync_cookie = NULL;
638 }
639}
640
641void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
642{
643 if (rwb) {
644 rwb->queue_depth = depth;
645 wbt_update_limits(rwb);
646 }
647}
648
649void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
650{
651 if (rwb)
652 rwb->wc = write_cache_on;
653}
654
655/*
656 * Disable wbt, if enabled by default.
657 */
658void wbt_disable_default(struct request_queue *q)
659{
660 struct rq_wb *rwb = q->rq_wb;
661
662 if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT)
663 wbt_exit(q);
664}
665EXPORT_SYMBOL_GPL(wbt_disable_default);
666
667/*
668 * Enable wbt if defaults are configured that way
669 */
670void wbt_enable_default(struct request_queue *q)
671{
672 /* Throttling already enabled? */
673 if (q->rq_wb)
674 return;
675
676 /* Queue not registered? Maybe shutting down... */
677 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
678 return;
679
680 if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) ||
681 (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ)))
682 wbt_init(q);
683}
684EXPORT_SYMBOL_GPL(wbt_enable_default);
685
686u64 wbt_default_latency_nsec(struct request_queue *q)
687{
688 /*
689 * We default to 2msec for non-rotational storage, and 75msec
690 * for rotational storage.
691 */
692 if (blk_queue_nonrot(q))
693 return 2000000ULL;
694 else
695 return 75000000ULL;
696}
697
698static int wbt_data_dir(const struct request *rq)
699{
700 const int op = req_op(rq);
701
702 if (op == REQ_OP_READ)
703 return READ;
704 else if (op == REQ_OP_WRITE || op == REQ_OP_FLUSH)
705 return WRITE;
706
707 /* don't account */
708 return -1;
709}
710
711int wbt_init(struct request_queue *q)
712{
713 struct rq_wb *rwb;
714 int i;
715
716 BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS);
717
718 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
719 if (!rwb)
720 return -ENOMEM;
721
722 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
723 if (!rwb->cb) {
724 kfree(rwb);
725 return -ENOMEM;
726 }
727
728 for (i = 0; i < WBT_NUM_RWQ; i++) {
729 atomic_set(&rwb->rq_wait[i].inflight, 0);
730 init_waitqueue_head(&rwb->rq_wait[i].wait);
731 }
732
733 rwb->last_comp = rwb->last_issue = jiffies;
734 rwb->queue = q;
735 rwb->win_nsec = RWB_WINDOW_NSEC;
736 rwb->enable_state = WBT_STATE_ON_DEFAULT;
737 wbt_update_limits(rwb);
738
739 /*
740 * Assign rwb and add the stats callback.
741 */
742 q->rq_wb = rwb;
743 blk_stat_add_callback(q, rwb->cb);
744
745 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
746
747 wbt_set_queue_depth(rwb, blk_queue_depth(q));
748 wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
749
750 return 0;
751}
752
753void wbt_exit(struct request_queue *q)
754{
755 struct rq_wb *rwb = q->rq_wb;
756
757 if (rwb) {
758 blk_stat_remove_callback(q, rwb->cb);
759 blk_stat_free_callback(rwb->cb);
760 q->rq_wb = NULL;
761 kfree(rwb);
762 }
763}
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
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{
99 struct bdi_writeback *wb = &rwb->queue->backing_dev_info.wb;
100
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
250static inline bool 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[BLK_STAT_READ].nr_samples >= 1 &&
259 stat[BLK_STAT_WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES;
260}
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
280static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
281{
282 struct backing_dev_info *bdi = &rwb->queue->backing_dev_info;
283 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 ||
296 (thislat > rwb->min_lat_nsec && !stat[BLK_STAT_READ].nr_samples)) {
297 trace_wbt_lat(bdi, thislat);
298 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 */
311 if ((stat[BLK_STAT_WRITE].nr_samples && blk_stat_is_current(stat)) ||
312 wb_recent_wait(rwb) || wbt_inflight(rwb))
313 return LAT_UNKNOWN_WRITES;
314 return LAT_UNKNOWN;
315 }
316
317 /*
318 * If the 'min' latency exceeds our target, step down.
319 */
320 if (stat[BLK_STAT_READ].min > rwb->min_lat_nsec) {
321 trace_wbt_lat(bdi, stat[BLK_STAT_READ].min);
322 trace_wbt_stat(bdi, stat);
323 return LAT_EXCEEDED;
324 }
325
326 if (rwb->scale_step)
327 trace_wbt_stat(bdi, stat);
328
329 return LAT_OK;
330}
331
332static int latency_exceeded(struct rq_wb *rwb)
333{
334 struct blk_rq_stat stat[2];
335
336 blk_queue_stat_get(rwb->queue, stat);
337 return __latency_exceeded(rwb, stat);
338}
339
340static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
341{
342 struct backing_dev_info *bdi = &rwb->queue->backing_dev_info;
343
344 trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
345 rwb->wb_background, rwb->wb_normal, rwb->wb_max);
346}
347
348static void scale_up(struct rq_wb *rwb)
349{
350 /*
351 * Hit max in previous round, stop here
352 */
353 if (rwb->scaled_max)
354 return;
355
356 rwb->scale_step--;
357 rwb->unknown_cnt = 0;
358 blk_stat_clear(rwb->queue);
359
360 rwb->scaled_max = calc_wb_limits(rwb);
361
362 rwb_wake_all(rwb);
363
364 rwb_trace_step(rwb, "step up");
365}
366
367/*
368 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
369 * had a latency violation.
370 */
371static void scale_down(struct rq_wb *rwb, bool hard_throttle)
372{
373 /*
374 * Stop scaling down when we've hit the limit. This also prevents
375 * ->scale_step from going to crazy values, if the device can't
376 * keep up.
377 */
378 if (rwb->wb_max == 1)
379 return;
380
381 if (rwb->scale_step < 0 && hard_throttle)
382 rwb->scale_step = 0;
383 else
384 rwb->scale_step++;
385
386 rwb->scaled_max = false;
387 rwb->unknown_cnt = 0;
388 blk_stat_clear(rwb->queue);
389 calc_wb_limits(rwb);
390 rwb_trace_step(rwb, "step down");
391}
392
393static void rwb_arm_timer(struct rq_wb *rwb)
394{
395 unsigned long expires;
396
397 if (rwb->scale_step > 0) {
398 /*
399 * We should speed this up, using some variant of a fast
400 * integer inverse square root calculation. Since we only do
401 * this for every window expiration, it's not a huge deal,
402 * though.
403 */
404 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
405 int_sqrt((rwb->scale_step + 1) << 8));
406 } else {
407 /*
408 * For step < 0, we don't want to increase/decrease the
409 * window size.
410 */
411 rwb->cur_win_nsec = rwb->win_nsec;
412 }
413
414 expires = jiffies + nsecs_to_jiffies(rwb->cur_win_nsec);
415 mod_timer(&rwb->window_timer, expires);
416}
417
418static void wb_timer_fn(unsigned long data)
419{
420 struct rq_wb *rwb = (struct rq_wb *) data;
421 unsigned int inflight = wbt_inflight(rwb);
422 int status;
423
424 status = latency_exceeded(rwb);
425
426 trace_wbt_timer(&rwb->queue->backing_dev_info, status, rwb->scale_step,
427 inflight);
428
429 /*
430 * If we exceeded the latency target, step down. If we did not,
431 * step one level up. If we don't know enough to say either exceeded
432 * or ok, then don't do anything.
433 */
434 switch (status) {
435 case LAT_EXCEEDED:
436 scale_down(rwb, true);
437 break;
438 case LAT_OK:
439 scale_up(rwb);
440 break;
441 case LAT_UNKNOWN_WRITES:
442 /*
443 * We started a the center step, but don't have a valid
444 * read/write sample, but we do have writes going on.
445 * Allow step to go negative, to increase write perf.
446 */
447 scale_up(rwb);
448 break;
449 case LAT_UNKNOWN:
450 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
451 break;
452 /*
453 * We get here when previously scaled reduced depth, and we
454 * currently don't have a valid read/write sample. For that
455 * case, slowly return to center state (step == 0).
456 */
457 if (rwb->scale_step > 0)
458 scale_up(rwb);
459 else if (rwb->scale_step < 0)
460 scale_down(rwb, false);
461 break;
462 default:
463 break;
464 }
465
466 /*
467 * Re-arm timer, if we have IO in flight
468 */
469 if (rwb->scale_step || inflight)
470 rwb_arm_timer(rwb);
471}
472
473void wbt_update_limits(struct rq_wb *rwb)
474{
475 rwb->scale_step = 0;
476 rwb->scaled_max = false;
477 calc_wb_limits(rwb);
478
479 rwb_wake_all(rwb);
480}
481
482static bool close_io(struct rq_wb *rwb)
483{
484 const unsigned long now = jiffies;
485
486 return time_before(now, rwb->last_issue + HZ / 10) ||
487 time_before(now, rwb->last_comp + HZ / 10);
488}
489
490#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
491
492static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
493{
494 unsigned int limit;
495
496 /*
497 * At this point we know it's a buffered write. If this is
498 * kswapd trying to free memory, or REQ_SYNC is set, set, then
499 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
500 * that. If the write is marked as a background write, then use
501 * the idle limit, or go to normal if we haven't had competing
502 * IO for a bit.
503 */
504 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
505 limit = rwb->wb_max;
506 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
507 /*
508 * If less than 100ms since we completed unrelated IO,
509 * limit us to half the depth for background writeback.
510 */
511 limit = rwb->wb_background;
512 } else
513 limit = rwb->wb_normal;
514
515 return limit;
516}
517
518static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
519 wait_queue_t *wait, unsigned long rw)
520{
521 /*
522 * inc it here even if disabled, since we'll dec it at completion.
523 * this only happens if the task was sleeping in __wbt_wait(),
524 * and someone turned it off at the same time.
525 */
526 if (!rwb_enabled(rwb)) {
527 atomic_inc(&rqw->inflight);
528 return true;
529 }
530
531 /*
532 * If the waitqueue is already active and we are not the next
533 * in line to be woken up, wait for our turn.
534 */
535 if (waitqueue_active(&rqw->wait) &&
536 rqw->wait.task_list.next != &wait->task_list)
537 return false;
538
539 return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
540}
541
542/*
543 * Block if we will exceed our limit, or if we are currently waiting for
544 * the timer to kick off queuing again.
545 */
546static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock)
547 __releases(lock)
548 __acquires(lock)
549{
550 struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd());
551 DEFINE_WAIT(wait);
552
553 if (may_queue(rwb, rqw, &wait, rw))
554 return;
555
556 do {
557 prepare_to_wait_exclusive(&rqw->wait, &wait,
558 TASK_UNINTERRUPTIBLE);
559
560 if (may_queue(rwb, rqw, &wait, rw))
561 break;
562
563 if (lock) {
564 spin_unlock_irq(lock);
565 io_schedule();
566 spin_lock_irq(lock);
567 } else
568 io_schedule();
569 } while (1);
570
571 finish_wait(&rqw->wait, &wait);
572}
573
574static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
575{
576 const int op = bio_op(bio);
577
578 /*
579 * If not a WRITE, do nothing
580 */
581 if (op != REQ_OP_WRITE)
582 return false;
583
584 /*
585 * Don't throttle WRITE_ODIRECT
586 */
587 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE))
588 return false;
589
590 return true;
591}
592
593/*
594 * Returns true if the IO request should be accounted, false if not.
595 * May sleep, if we have exceeded the writeback limits. Caller can pass
596 * in an irq held spinlock, if it holds one when calling this function.
597 * If we do sleep, we'll release and re-grab it.
598 */
599enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
600{
601 unsigned int ret = 0;
602
603 if (!rwb_enabled(rwb))
604 return 0;
605
606 if (bio_op(bio) == REQ_OP_READ)
607 ret = WBT_READ;
608
609 if (!wbt_should_throttle(rwb, bio)) {
610 if (ret & WBT_READ)
611 wb_timestamp(rwb, &rwb->last_issue);
612 return ret;
613 }
614
615 __wbt_wait(rwb, bio->bi_opf, lock);
616
617 if (!timer_pending(&rwb->window_timer))
618 rwb_arm_timer(rwb);
619
620 if (current_is_kswapd())
621 ret |= WBT_KSWAPD;
622
623 return ret | WBT_TRACKED;
624}
625
626void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat)
627{
628 if (!rwb_enabled(rwb))
629 return;
630
631 /*
632 * Track sync issue, in case it takes a long time to complete. Allows
633 * us to react quicker, if a sync IO takes a long time to complete.
634 * Note that this is just a hint. 'stat' can go away when the
635 * request completes, so it's important we never dereference it. We
636 * only use the address to compare with, which is why we store the
637 * sync_issue time locally.
638 */
639 if (wbt_is_read(stat) && !rwb->sync_issue) {
640 rwb->sync_cookie = stat;
641 rwb->sync_issue = blk_stat_time(stat);
642 }
643}
644
645void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat)
646{
647 if (!rwb_enabled(rwb))
648 return;
649 if (stat == rwb->sync_cookie) {
650 rwb->sync_issue = 0;
651 rwb->sync_cookie = NULL;
652 }
653}
654
655void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
656{
657 if (rwb) {
658 rwb->queue_depth = depth;
659 wbt_update_limits(rwb);
660 }
661}
662
663void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
664{
665 if (rwb)
666 rwb->wc = write_cache_on;
667}
668
669 /*
670 * Disable wbt, if enabled by default. Only called from CFQ, if we have
671 * cgroups enabled
672 */
673void wbt_disable_default(struct request_queue *q)
674{
675 struct rq_wb *rwb = q->rq_wb;
676
677 if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) {
678 del_timer_sync(&rwb->window_timer);
679 rwb->win_nsec = rwb->min_lat_nsec = 0;
680 wbt_update_limits(rwb);
681 }
682}
683EXPORT_SYMBOL_GPL(wbt_disable_default);
684
685u64 wbt_default_latency_nsec(struct request_queue *q)
686{
687 /*
688 * We default to 2msec for non-rotational storage, and 75msec
689 * for rotational storage.
690 */
691 if (blk_queue_nonrot(q))
692 return 2000000ULL;
693 else
694 return 75000000ULL;
695}
696
697int wbt_init(struct request_queue *q)
698{
699 struct rq_wb *rwb;
700 int i;
701
702 /*
703 * For now, we depend on the stats window being larger than
704 * our monitoring window. Ensure that this isn't inadvertently
705 * violated.
706 */
707 BUILD_BUG_ON(RWB_WINDOW_NSEC > BLK_STAT_NSEC);
708 BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS);
709
710 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
711 if (!rwb)
712 return -ENOMEM;
713
714 for (i = 0; i < WBT_NUM_RWQ; i++) {
715 atomic_set(&rwb->rq_wait[i].inflight, 0);
716 init_waitqueue_head(&rwb->rq_wait[i].wait);
717 }
718
719 setup_timer(&rwb->window_timer, wb_timer_fn, (unsigned long) rwb);
720 rwb->wc = 1;
721 rwb->queue_depth = RWB_DEF_DEPTH;
722 rwb->last_comp = rwb->last_issue = jiffies;
723 rwb->queue = q;
724 rwb->win_nsec = RWB_WINDOW_NSEC;
725 rwb->enable_state = WBT_STATE_ON_DEFAULT;
726 wbt_update_limits(rwb);
727
728 /*
729 * Assign rwb, and turn on stats tracking for this queue
730 */
731 q->rq_wb = rwb;
732 blk_stat_enable(q);
733
734 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
735
736 wbt_set_queue_depth(rwb, blk_queue_depth(q));
737 wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
738
739 return 0;
740}
741
742void wbt_exit(struct request_queue *q)
743{
744 struct rq_wb *rwb = q->rq_wb;
745
746 if (rwb) {
747 del_timer_sync(&rwb->window_timer);
748 q->rq_wb = NULL;
749 kfree(rwb);
750 }
751}