Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * buffered writeback throttling. loosely based on CoDel. We can't drop
4 * packets for IO scheduling, so the logic is something like this:
5 *
6 * - Monitor latencies in a defined window of time.
7 * - If the minimum latency in the above window exceeds some target, increment
8 * scaling step and scale down queue depth by a factor of 2x. The monitoring
9 * window is then shrunk to 100 / sqrt(scaling step + 1).
10 * - For any window where we don't have solid data on what the latencies
11 * look like, retain status quo.
12 * - If latencies look good, decrement scaling step.
13 * - If we're only doing writes, allow the scaling step to go negative. This
14 * will temporarily boost write performance, snapping back to a stable
15 * scaling step of 0 if reads show up or the heavy writers finish. Unlike
16 * positive scaling steps where we shrink the monitoring window, a negative
17 * scaling step retains the default step==0 window size.
18 *
19 * Copyright (C) 2016 Jens Axboe
20 *
21 */
22#include <linux/kernel.h>
23#include <linux/blk_types.h>
24#include <linux/slab.h>
25#include <linux/backing-dev.h>
26#include <linux/swap.h>
27
28#include "blk-wbt.h"
29#include "blk-rq-qos.h"
30#include "elevator.h"
31
32#define CREATE_TRACE_POINTS
33#include <trace/events/wbt.h>
34
35static inline void wbt_clear_state(struct request *rq)
36{
37 rq->wbt_flags = 0;
38}
39
40static inline enum wbt_flags wbt_flags(struct request *rq)
41{
42 return rq->wbt_flags;
43}
44
45static inline bool wbt_is_tracked(struct request *rq)
46{
47 return rq->wbt_flags & WBT_TRACKED;
48}
49
50static inline bool wbt_is_read(struct request *rq)
51{
52 return rq->wbt_flags & WBT_READ;
53}
54
55enum {
56 /*
57 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
58 * from here depending on device stats
59 */
60 RWB_DEF_DEPTH = 16,
61
62 /*
63 * 100msec window
64 */
65 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
66
67 /*
68 * Disregard stats, if we don't meet this minimum
69 */
70 RWB_MIN_WRITE_SAMPLES = 3,
71
72 /*
73 * If we have this number of consecutive windows with not enough
74 * information to scale up or down, scale up.
75 */
76 RWB_UNKNOWN_BUMP = 5,
77};
78
79static inline bool rwb_enabled(struct rq_wb *rwb)
80{
81 return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&
82 rwb->wb_normal != 0;
83}
84
85static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
86{
87 if (rwb_enabled(rwb)) {
88 const unsigned long cur = jiffies;
89
90 if (cur != *var)
91 *var = cur;
92 }
93}
94
95/*
96 * If a task was rate throttled in balance_dirty_pages() within the last
97 * second or so, use that to indicate a higher cleaning rate.
98 */
99static bool wb_recent_wait(struct rq_wb *rwb)
100{
101 struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb;
102
103 return time_before(jiffies, wb->dirty_sleep + HZ);
104}
105
106static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
107 enum wbt_flags wb_acct)
108{
109 if (wb_acct & WBT_KSWAPD)
110 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
111 else if (wb_acct & WBT_DISCARD)
112 return &rwb->rq_wait[WBT_RWQ_DISCARD];
113
114 return &rwb->rq_wait[WBT_RWQ_BG];
115}
116
117static void rwb_wake_all(struct rq_wb *rwb)
118{
119 int i;
120
121 for (i = 0; i < WBT_NUM_RWQ; i++) {
122 struct rq_wait *rqw = &rwb->rq_wait[i];
123
124 if (wq_has_sleeper(&rqw->wait))
125 wake_up_all(&rqw->wait);
126 }
127}
128
129static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
130 enum wbt_flags wb_acct)
131{
132 int inflight, limit;
133
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 * For discards, our limit is always the background. For writes, if
147 * the device does write back caching, drop further down before we
148 * wake people up.
149 */
150 if (wb_acct & WBT_DISCARD)
151 limit = rwb->wb_background;
152 else if (rwb->wc && !wb_recent_wait(rwb))
153 limit = 0;
154 else
155 limit = rwb->wb_normal;
156
157 /*
158 * Don't wake anyone up if we are above the normal limit.
159 */
160 if (inflight && inflight >= limit)
161 return;
162
163 if (wq_has_sleeper(&rqw->wait)) {
164 int diff = limit - inflight;
165
166 if (!inflight || diff >= rwb->wb_background / 2)
167 wake_up_all(&rqw->wait);
168 }
169}
170
171static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
172{
173 struct rq_wb *rwb = RQWB(rqos);
174 struct rq_wait *rqw;
175
176 if (!(wb_acct & WBT_TRACKED))
177 return;
178
179 rqw = get_rq_wait(rwb, wb_acct);
180 wbt_rqw_done(rwb, rqw, wb_acct);
181}
182
183/*
184 * Called on completion of a request. Note that it's also called when
185 * a request is merged, when the request gets freed.
186 */
187static void wbt_done(struct rq_qos *rqos, struct request *rq)
188{
189 struct rq_wb *rwb = RQWB(rqos);
190
191 if (!wbt_is_tracked(rq)) {
192 if (rwb->sync_cookie == rq) {
193 rwb->sync_issue = 0;
194 rwb->sync_cookie = NULL;
195 }
196
197 if (wbt_is_read(rq))
198 wb_timestamp(rwb, &rwb->last_comp);
199 } else {
200 WARN_ON_ONCE(rq == rwb->sync_cookie);
201 __wbt_done(rqos, wbt_flags(rq));
202 }
203 wbt_clear_state(rq);
204}
205
206static inline bool stat_sample_valid(struct blk_rq_stat *stat)
207{
208 /*
209 * We need at least one read sample, and a minimum of
210 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
211 * that it's writes impacting us, and not just some sole read on
212 * a device that is in a lower power state.
213 */
214 return (stat[READ].nr_samples >= 1 &&
215 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
216}
217
218static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
219{
220 u64 now, issue = READ_ONCE(rwb->sync_issue);
221
222 if (!issue || !rwb->sync_cookie)
223 return 0;
224
225 now = ktime_to_ns(ktime_get());
226 return now - issue;
227}
228
229enum {
230 LAT_OK = 1,
231 LAT_UNKNOWN,
232 LAT_UNKNOWN_WRITES,
233 LAT_EXCEEDED,
234};
235
236static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
237{
238 struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
239 struct rq_depth *rqd = &rwb->rq_depth;
240 u64 thislat;
241
242 /*
243 * If our stored sync issue exceeds the window size, or it
244 * exceeds our min target AND we haven't logged any entries,
245 * flag the latency as exceeded. wbt works off completion latencies,
246 * but for a flooded device, a single sync IO can take a long time
247 * to complete after being issued. If this time exceeds our
248 * monitoring window AND we didn't see any other completions in that
249 * window, then count that sync IO as a violation of the latency.
250 */
251 thislat = rwb_sync_issue_lat(rwb);
252 if (thislat > rwb->cur_win_nsec ||
253 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
254 trace_wbt_lat(bdi, thislat);
255 return LAT_EXCEEDED;
256 }
257
258 /*
259 * No read/write mix, if stat isn't valid
260 */
261 if (!stat_sample_valid(stat)) {
262 /*
263 * If we had writes in this stat window and the window is
264 * current, we're only doing writes. If a task recently
265 * waited or still has writes in flights, consider us doing
266 * just writes as well.
267 */
268 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
269 wbt_inflight(rwb))
270 return LAT_UNKNOWN_WRITES;
271 return LAT_UNKNOWN;
272 }
273
274 /*
275 * If the 'min' latency exceeds our target, step down.
276 */
277 if (stat[READ].min > rwb->min_lat_nsec) {
278 trace_wbt_lat(bdi, stat[READ].min);
279 trace_wbt_stat(bdi, stat);
280 return LAT_EXCEEDED;
281 }
282
283 if (rqd->scale_step)
284 trace_wbt_stat(bdi, stat);
285
286 return LAT_OK;
287}
288
289static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
290{
291 struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
292 struct rq_depth *rqd = &rwb->rq_depth;
293
294 trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
295 rwb->wb_background, rwb->wb_normal, rqd->max_depth);
296}
297
298static void calc_wb_limits(struct rq_wb *rwb)
299{
300 if (rwb->min_lat_nsec == 0) {
301 rwb->wb_normal = rwb->wb_background = 0;
302 } else if (rwb->rq_depth.max_depth <= 2) {
303 rwb->wb_normal = rwb->rq_depth.max_depth;
304 rwb->wb_background = 1;
305 } else {
306 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
307 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
308 }
309}
310
311static void scale_up(struct rq_wb *rwb)
312{
313 if (!rq_depth_scale_up(&rwb->rq_depth))
314 return;
315 calc_wb_limits(rwb);
316 rwb->unknown_cnt = 0;
317 rwb_wake_all(rwb);
318 rwb_trace_step(rwb, tracepoint_string("scale up"));
319}
320
321static void scale_down(struct rq_wb *rwb, bool hard_throttle)
322{
323 if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
324 return;
325 calc_wb_limits(rwb);
326 rwb->unknown_cnt = 0;
327 rwb_trace_step(rwb, tracepoint_string("scale down"));
328}
329
330static void rwb_arm_timer(struct rq_wb *rwb)
331{
332 struct rq_depth *rqd = &rwb->rq_depth;
333
334 if (rqd->scale_step > 0) {
335 /*
336 * We should speed this up, using some variant of a fast
337 * integer inverse square root calculation. Since we only do
338 * this for every window expiration, it's not a huge deal,
339 * though.
340 */
341 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
342 int_sqrt((rqd->scale_step + 1) << 8));
343 } else {
344 /*
345 * For step < 0, we don't want to increase/decrease the
346 * window size.
347 */
348 rwb->cur_win_nsec = rwb->win_nsec;
349 }
350
351 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
352}
353
354static void wb_timer_fn(struct blk_stat_callback *cb)
355{
356 struct rq_wb *rwb = cb->data;
357 struct rq_depth *rqd = &rwb->rq_depth;
358 unsigned int inflight = wbt_inflight(rwb);
359 int status;
360
361 if (!rwb->rqos.q->disk)
362 return;
363
364 status = latency_exceeded(rwb, cb->stat);
365
366 trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step,
367 inflight);
368
369 /*
370 * If we exceeded the latency target, step down. If we did not,
371 * step one level up. If we don't know enough to say either exceeded
372 * or ok, then don't do anything.
373 */
374 switch (status) {
375 case LAT_EXCEEDED:
376 scale_down(rwb, true);
377 break;
378 case LAT_OK:
379 scale_up(rwb);
380 break;
381 case LAT_UNKNOWN_WRITES:
382 /*
383 * We started a the center step, but don't have a valid
384 * read/write sample, but we do have writes going on.
385 * Allow step to go negative, to increase write perf.
386 */
387 scale_up(rwb);
388 break;
389 case LAT_UNKNOWN:
390 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
391 break;
392 /*
393 * We get here when previously scaled reduced depth, and we
394 * currently don't have a valid read/write sample. For that
395 * case, slowly return to center state (step == 0).
396 */
397 if (rqd->scale_step > 0)
398 scale_up(rwb);
399 else if (rqd->scale_step < 0)
400 scale_down(rwb, false);
401 break;
402 default:
403 break;
404 }
405
406 /*
407 * Re-arm timer, if we have IO in flight
408 */
409 if (rqd->scale_step || inflight)
410 rwb_arm_timer(rwb);
411}
412
413static void wbt_update_limits(struct rq_wb *rwb)
414{
415 struct rq_depth *rqd = &rwb->rq_depth;
416
417 rqd->scale_step = 0;
418 rqd->scaled_max = false;
419
420 rq_depth_calc_max_depth(rqd);
421 calc_wb_limits(rwb);
422
423 rwb_wake_all(rwb);
424}
425
426bool wbt_disabled(struct request_queue *q)
427{
428 struct rq_qos *rqos = wbt_rq_qos(q);
429
430 return !rqos || RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT ||
431 RQWB(rqos)->enable_state == WBT_STATE_OFF_MANUAL;
432}
433
434u64 wbt_get_min_lat(struct request_queue *q)
435{
436 struct rq_qos *rqos = wbt_rq_qos(q);
437 if (!rqos)
438 return 0;
439 return RQWB(rqos)->min_lat_nsec;
440}
441
442void wbt_set_min_lat(struct request_queue *q, u64 val)
443{
444 struct rq_qos *rqos = wbt_rq_qos(q);
445 if (!rqos)
446 return;
447
448 RQWB(rqos)->min_lat_nsec = val;
449 if (val)
450 RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
451 else
452 RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL;
453
454 wbt_update_limits(RQWB(rqos));
455}
456
457
458static bool close_io(struct rq_wb *rwb)
459{
460 const unsigned long now = jiffies;
461
462 return time_before(now, rwb->last_issue + HZ / 10) ||
463 time_before(now, rwb->last_comp + HZ / 10);
464}
465
466#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
467
468static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf)
469{
470 unsigned int limit;
471
472 /*
473 * If we got disabled, just return UINT_MAX. This ensures that
474 * we'll properly inc a new IO, and dec+wakeup at the end.
475 */
476 if (!rwb_enabled(rwb))
477 return UINT_MAX;
478
479 if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD)
480 return rwb->wb_background;
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 ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
491 limit = rwb->rq_depth.max_depth;
492 else if ((opf & 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
504struct wbt_wait_data {
505 struct rq_wb *rwb;
506 enum wbt_flags wb_acct;
507 blk_opf_t opf;
508};
509
510static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
511{
512 struct wbt_wait_data *data = private_data;
513 return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf));
514}
515
516static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
517{
518 struct wbt_wait_data *data = private_data;
519 wbt_rqw_done(data->rwb, rqw, data->wb_acct);
520}
521
522/*
523 * Block if we will exceed our limit, or if we are currently waiting for
524 * the timer to kick off queuing again.
525 */
526static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
527 blk_opf_t opf)
528{
529 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
530 struct wbt_wait_data data = {
531 .rwb = rwb,
532 .wb_acct = wb_acct,
533 .opf = opf,
534 };
535
536 rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
537}
538
539static inline bool wbt_should_throttle(struct bio *bio)
540{
541 switch (bio_op(bio)) {
542 case REQ_OP_WRITE:
543 /*
544 * Don't throttle WRITE_ODIRECT
545 */
546 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
547 (REQ_SYNC | REQ_IDLE))
548 return false;
549 fallthrough;
550 case REQ_OP_DISCARD:
551 return true;
552 default:
553 return false;
554 }
555}
556
557static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
558{
559 enum wbt_flags flags = 0;
560
561 if (!rwb_enabled(rwb))
562 return 0;
563
564 if (bio_op(bio) == REQ_OP_READ) {
565 flags = WBT_READ;
566 } else if (wbt_should_throttle(bio)) {
567 if (current_is_kswapd())
568 flags |= WBT_KSWAPD;
569 if (bio_op(bio) == REQ_OP_DISCARD)
570 flags |= WBT_DISCARD;
571 flags |= WBT_TRACKED;
572 }
573 return flags;
574}
575
576static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
577{
578 struct rq_wb *rwb = RQWB(rqos);
579 enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
580 __wbt_done(rqos, flags);
581}
582
583/*
584 * May sleep, if we have exceeded the writeback limits. Caller can pass
585 * in an irq held spinlock, if it holds one when calling this function.
586 * If we do sleep, we'll release and re-grab it.
587 */
588static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
589{
590 struct rq_wb *rwb = RQWB(rqos);
591 enum wbt_flags flags;
592
593 flags = bio_to_wbt_flags(rwb, bio);
594 if (!(flags & WBT_TRACKED)) {
595 if (flags & WBT_READ)
596 wb_timestamp(rwb, &rwb->last_issue);
597 return;
598 }
599
600 __wbt_wait(rwb, flags, bio->bi_opf);
601
602 if (!blk_stat_is_active(rwb->cb))
603 rwb_arm_timer(rwb);
604}
605
606static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
607{
608 struct rq_wb *rwb = RQWB(rqos);
609 rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
610}
611
612static void wbt_issue(struct rq_qos *rqos, struct request *rq)
613{
614 struct rq_wb *rwb = RQWB(rqos);
615
616 if (!rwb_enabled(rwb))
617 return;
618
619 /*
620 * Track sync issue, in case it takes a long time to complete. Allows us
621 * to react quicker, if a sync IO takes a long time to complete. Note
622 * that this is just a hint. The request can go away when it completes,
623 * so it's important we never dereference it. We only use the address to
624 * compare with, which is why we store the sync_issue time locally.
625 */
626 if (wbt_is_read(rq) && !rwb->sync_issue) {
627 rwb->sync_cookie = rq;
628 rwb->sync_issue = rq->io_start_time_ns;
629 }
630}
631
632static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
633{
634 struct rq_wb *rwb = RQWB(rqos);
635 if (!rwb_enabled(rwb))
636 return;
637 if (rq == rwb->sync_cookie) {
638 rwb->sync_issue = 0;
639 rwb->sync_cookie = NULL;
640 }
641}
642
643void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
644{
645 struct rq_qos *rqos = wbt_rq_qos(q);
646 if (rqos)
647 RQWB(rqos)->wc = write_cache_on;
648}
649
650/*
651 * Enable wbt if defaults are configured that way
652 */
653void wbt_enable_default(struct request_queue *q)
654{
655 struct rq_qos *rqos;
656 bool disable_flag = q->elevator &&
657 test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags);
658
659 /* Throttling already enabled? */
660 rqos = wbt_rq_qos(q);
661 if (rqos) {
662 if (!disable_flag &&
663 RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)
664 RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;
665 return;
666 }
667
668 /* Queue not registered? Maybe shutting down... */
669 if (!blk_queue_registered(q))
670 return;
671
672 if (queue_is_mq(q) && !disable_flag)
673 wbt_init(q);
674}
675EXPORT_SYMBOL_GPL(wbt_enable_default);
676
677u64 wbt_default_latency_nsec(struct request_queue *q)
678{
679 /*
680 * We default to 2msec for non-rotational storage, and 75msec
681 * for rotational storage.
682 */
683 if (blk_queue_nonrot(q))
684 return 2000000ULL;
685 else
686 return 75000000ULL;
687}
688
689static int wbt_data_dir(const struct request *rq)
690{
691 const enum req_op op = req_op(rq);
692
693 if (op == REQ_OP_READ)
694 return READ;
695 else if (op_is_write(op))
696 return WRITE;
697
698 /* don't account */
699 return -1;
700}
701
702static void wbt_queue_depth_changed(struct rq_qos *rqos)
703{
704 RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);
705 wbt_update_limits(RQWB(rqos));
706}
707
708static void wbt_exit(struct rq_qos *rqos)
709{
710 struct rq_wb *rwb = RQWB(rqos);
711 struct request_queue *q = rqos->q;
712
713 blk_stat_remove_callback(q, rwb->cb);
714 blk_stat_free_callback(rwb->cb);
715 kfree(rwb);
716}
717
718/*
719 * Disable wbt, if enabled by default.
720 */
721void wbt_disable_default(struct request_queue *q)
722{
723 struct rq_qos *rqos = wbt_rq_qos(q);
724 struct rq_wb *rwb;
725 if (!rqos)
726 return;
727 rwb = RQWB(rqos);
728 if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
729 blk_stat_deactivate(rwb->cb);
730 rwb->enable_state = WBT_STATE_OFF_DEFAULT;
731 }
732}
733EXPORT_SYMBOL_GPL(wbt_disable_default);
734
735#ifdef CONFIG_BLK_DEBUG_FS
736static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
737{
738 struct rq_qos *rqos = data;
739 struct rq_wb *rwb = RQWB(rqos);
740
741 seq_printf(m, "%llu\n", rwb->cur_win_nsec);
742 return 0;
743}
744
745static int wbt_enabled_show(void *data, struct seq_file *m)
746{
747 struct rq_qos *rqos = data;
748 struct rq_wb *rwb = RQWB(rqos);
749
750 seq_printf(m, "%d\n", rwb->enable_state);
751 return 0;
752}
753
754static int wbt_id_show(void *data, struct seq_file *m)
755{
756 struct rq_qos *rqos = data;
757
758 seq_printf(m, "%u\n", rqos->id);
759 return 0;
760}
761
762static int wbt_inflight_show(void *data, struct seq_file *m)
763{
764 struct rq_qos *rqos = data;
765 struct rq_wb *rwb = RQWB(rqos);
766 int i;
767
768 for (i = 0; i < WBT_NUM_RWQ; i++)
769 seq_printf(m, "%d: inflight %d\n", i,
770 atomic_read(&rwb->rq_wait[i].inflight));
771 return 0;
772}
773
774static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
775{
776 struct rq_qos *rqos = data;
777 struct rq_wb *rwb = RQWB(rqos);
778
779 seq_printf(m, "%lu\n", rwb->min_lat_nsec);
780 return 0;
781}
782
783static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
784{
785 struct rq_qos *rqos = data;
786 struct rq_wb *rwb = RQWB(rqos);
787
788 seq_printf(m, "%u\n", rwb->unknown_cnt);
789 return 0;
790}
791
792static int wbt_normal_show(void *data, struct seq_file *m)
793{
794 struct rq_qos *rqos = data;
795 struct rq_wb *rwb = RQWB(rqos);
796
797 seq_printf(m, "%u\n", rwb->wb_normal);
798 return 0;
799}
800
801static int wbt_background_show(void *data, struct seq_file *m)
802{
803 struct rq_qos *rqos = data;
804 struct rq_wb *rwb = RQWB(rqos);
805
806 seq_printf(m, "%u\n", rwb->wb_background);
807 return 0;
808}
809
810static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
811 {"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
812 {"enabled", 0400, wbt_enabled_show},
813 {"id", 0400, wbt_id_show},
814 {"inflight", 0400, wbt_inflight_show},
815 {"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
816 {"unknown_cnt", 0400, wbt_unknown_cnt_show},
817 {"wb_normal", 0400, wbt_normal_show},
818 {"wb_background", 0400, wbt_background_show},
819 {},
820};
821#endif
822
823static struct rq_qos_ops wbt_rqos_ops = {
824 .throttle = wbt_wait,
825 .issue = wbt_issue,
826 .track = wbt_track,
827 .requeue = wbt_requeue,
828 .done = wbt_done,
829 .cleanup = wbt_cleanup,
830 .queue_depth_changed = wbt_queue_depth_changed,
831 .exit = wbt_exit,
832#ifdef CONFIG_BLK_DEBUG_FS
833 .debugfs_attrs = wbt_debugfs_attrs,
834#endif
835};
836
837int wbt_init(struct request_queue *q)
838{
839 struct rq_wb *rwb;
840 int i;
841 int ret;
842
843 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
844 if (!rwb)
845 return -ENOMEM;
846
847 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
848 if (!rwb->cb) {
849 kfree(rwb);
850 return -ENOMEM;
851 }
852
853 for (i = 0; i < WBT_NUM_RWQ; i++)
854 rq_wait_init(&rwb->rq_wait[i]);
855
856 rwb->rqos.id = RQ_QOS_WBT;
857 rwb->rqos.ops = &wbt_rqos_ops;
858 rwb->rqos.q = q;
859 rwb->last_comp = rwb->last_issue = jiffies;
860 rwb->win_nsec = RWB_WINDOW_NSEC;
861 rwb->enable_state = WBT_STATE_ON_DEFAULT;
862 rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags);
863 rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
864 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
865
866 wbt_queue_depth_changed(&rwb->rqos);
867
868 /*
869 * Assign rwb and add the stats callback.
870 */
871 ret = rq_qos_add(q, &rwb->rqos);
872 if (ret)
873 goto err_free;
874
875 blk_stat_add_callback(q, rwb->cb);
876
877 return 0;
878
879err_free:
880 blk_stat_free_callback(rwb->cb);
881 kfree(rwb);
882 return ret;
883
884}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * buffered writeback throttling. loosely based on CoDel. We can't drop
4 * packets for IO scheduling, so the logic is something like this:
5 *
6 * - Monitor latencies in a defined window of time.
7 * - If the minimum latency in the above window exceeds some target, increment
8 * scaling step and scale down queue depth by a factor of 2x. The monitoring
9 * window is then shrunk to 100 / sqrt(scaling step + 1).
10 * - For any window where we don't have solid data on what the latencies
11 * look like, retain status quo.
12 * - If latencies look good, decrement scaling step.
13 * - If we're only doing writes, allow the scaling step to go negative. This
14 * will temporarily boost write performance, snapping back to a stable
15 * scaling step of 0 if reads show up or the heavy writers finish. Unlike
16 * positive scaling steps where we shrink the monitoring window, a negative
17 * scaling step retains the default step==0 window size.
18 *
19 * Copyright (C) 2016 Jens Axboe
20 *
21 */
22#include <linux/kernel.h>
23#include <linux/blk_types.h>
24#include <linux/slab.h>
25#include <linux/backing-dev.h>
26#include <linux/swap.h>
27
28#include "blk-stat.h"
29#include "blk-wbt.h"
30#include "blk-rq-qos.h"
31#include "elevator.h"
32#include "blk.h"
33
34#define CREATE_TRACE_POINTS
35#include <trace/events/wbt.h>
36
37enum wbt_flags {
38 WBT_TRACKED = 1, /* write, tracked for throttling */
39 WBT_READ = 2, /* read */
40 WBT_KSWAPD = 4, /* write, from kswapd */
41 WBT_DISCARD = 8, /* discard */
42
43 WBT_NR_BITS = 4, /* number of bits */
44};
45
46enum {
47 WBT_RWQ_BG = 0,
48 WBT_RWQ_KSWAPD,
49 WBT_RWQ_DISCARD,
50 WBT_NUM_RWQ,
51};
52
53/*
54 * If current state is WBT_STATE_ON/OFF_DEFAULT, it can be covered to any other
55 * state, if current state is WBT_STATE_ON/OFF_MANUAL, it can only be covered
56 * to WBT_STATE_OFF/ON_MANUAL.
57 */
58enum {
59 WBT_STATE_ON_DEFAULT = 1, /* on by default */
60 WBT_STATE_ON_MANUAL = 2, /* on manually by sysfs */
61 WBT_STATE_OFF_DEFAULT = 3, /* off by default */
62 WBT_STATE_OFF_MANUAL = 4, /* off manually by sysfs */
63};
64
65struct rq_wb {
66 /*
67 * Settings that govern how we throttle
68 */
69 unsigned int wb_background; /* background writeback */
70 unsigned int wb_normal; /* normal writeback */
71
72 short enable_state; /* WBT_STATE_* */
73
74 /*
75 * Number of consecutive periods where we don't have enough
76 * information to make a firm scale up/down decision.
77 */
78 unsigned int unknown_cnt;
79
80 u64 win_nsec; /* default window size */
81 u64 cur_win_nsec; /* current window size */
82
83 struct blk_stat_callback *cb;
84
85 u64 sync_issue;
86 void *sync_cookie;
87
88 unsigned long last_issue; /* last non-throttled issue */
89 unsigned long last_comp; /* last non-throttled comp */
90 unsigned long min_lat_nsec;
91 struct rq_qos rqos;
92 struct rq_wait rq_wait[WBT_NUM_RWQ];
93 struct rq_depth rq_depth;
94};
95
96static inline struct rq_wb *RQWB(struct rq_qos *rqos)
97{
98 return container_of(rqos, struct rq_wb, rqos);
99}
100
101static inline void wbt_clear_state(struct request *rq)
102{
103 rq->wbt_flags = 0;
104}
105
106static inline enum wbt_flags wbt_flags(struct request *rq)
107{
108 return rq->wbt_flags;
109}
110
111static inline bool wbt_is_tracked(struct request *rq)
112{
113 return rq->wbt_flags & WBT_TRACKED;
114}
115
116static inline bool wbt_is_read(struct request *rq)
117{
118 return rq->wbt_flags & WBT_READ;
119}
120
121enum {
122 /*
123 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
124 * from here depending on device stats
125 */
126 RWB_DEF_DEPTH = 16,
127
128 /*
129 * 100msec window
130 */
131 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
132
133 /*
134 * Disregard stats, if we don't meet this minimum
135 */
136 RWB_MIN_WRITE_SAMPLES = 3,
137
138 /*
139 * If we have this number of consecutive windows with not enough
140 * information to scale up or down, scale up.
141 */
142 RWB_UNKNOWN_BUMP = 5,
143};
144
145static inline bool rwb_enabled(struct rq_wb *rwb)
146{
147 return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&
148 rwb->enable_state != WBT_STATE_OFF_MANUAL;
149}
150
151static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
152{
153 if (rwb_enabled(rwb)) {
154 const unsigned long cur = jiffies;
155
156 if (cur != *var)
157 *var = cur;
158 }
159}
160
161/*
162 * If a task was rate throttled in balance_dirty_pages() within the last
163 * second or so, use that to indicate a higher cleaning rate.
164 */
165static bool wb_recent_wait(struct rq_wb *rwb)
166{
167 struct backing_dev_info *bdi = rwb->rqos.disk->bdi;
168
169 return time_before(jiffies, bdi->last_bdp_sleep + HZ);
170}
171
172static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
173 enum wbt_flags wb_acct)
174{
175 if (wb_acct & WBT_KSWAPD)
176 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
177 else if (wb_acct & WBT_DISCARD)
178 return &rwb->rq_wait[WBT_RWQ_DISCARD];
179
180 return &rwb->rq_wait[WBT_RWQ_BG];
181}
182
183static void rwb_wake_all(struct rq_wb *rwb)
184{
185 int i;
186
187 for (i = 0; i < WBT_NUM_RWQ; i++) {
188 struct rq_wait *rqw = &rwb->rq_wait[i];
189
190 if (wq_has_sleeper(&rqw->wait))
191 wake_up_all(&rqw->wait);
192 }
193}
194
195static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
196 enum wbt_flags wb_acct)
197{
198 int inflight, limit;
199
200 inflight = atomic_dec_return(&rqw->inflight);
201
202 /*
203 * For discards, our limit is always the background. For writes, if
204 * the device does write back caching, drop further down before we
205 * wake people up.
206 */
207 if (wb_acct & WBT_DISCARD)
208 limit = rwb->wb_background;
209 else if (test_bit(QUEUE_FLAG_WC, &rwb->rqos.disk->queue->queue_flags) &&
210 !wb_recent_wait(rwb))
211 limit = 0;
212 else
213 limit = rwb->wb_normal;
214
215 /*
216 * Don't wake anyone up if we are above the normal limit.
217 */
218 if (inflight && inflight >= limit)
219 return;
220
221 if (wq_has_sleeper(&rqw->wait)) {
222 int diff = limit - inflight;
223
224 if (!inflight || diff >= rwb->wb_background / 2)
225 wake_up_all(&rqw->wait);
226 }
227}
228
229static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
230{
231 struct rq_wb *rwb = RQWB(rqos);
232 struct rq_wait *rqw;
233
234 if (!(wb_acct & WBT_TRACKED))
235 return;
236
237 rqw = get_rq_wait(rwb, wb_acct);
238 wbt_rqw_done(rwb, rqw, wb_acct);
239}
240
241/*
242 * Called on completion of a request. Note that it's also called when
243 * a request is merged, when the request gets freed.
244 */
245static void wbt_done(struct rq_qos *rqos, struct request *rq)
246{
247 struct rq_wb *rwb = RQWB(rqos);
248
249 if (!wbt_is_tracked(rq)) {
250 if (rwb->sync_cookie == rq) {
251 rwb->sync_issue = 0;
252 rwb->sync_cookie = NULL;
253 }
254
255 if (wbt_is_read(rq))
256 wb_timestamp(rwb, &rwb->last_comp);
257 } else {
258 WARN_ON_ONCE(rq == rwb->sync_cookie);
259 __wbt_done(rqos, wbt_flags(rq));
260 }
261 wbt_clear_state(rq);
262}
263
264static inline bool stat_sample_valid(struct blk_rq_stat *stat)
265{
266 /*
267 * We need at least one read sample, and a minimum of
268 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
269 * that it's writes impacting us, and not just some sole read on
270 * a device that is in a lower power state.
271 */
272 return (stat[READ].nr_samples >= 1 &&
273 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
274}
275
276static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
277{
278 u64 issue = READ_ONCE(rwb->sync_issue);
279
280 if (!issue || !rwb->sync_cookie)
281 return 0;
282
283 return blk_time_get_ns() - issue;
284}
285
286static inline unsigned int wbt_inflight(struct rq_wb *rwb)
287{
288 unsigned int i, ret = 0;
289
290 for (i = 0; i < WBT_NUM_RWQ; i++)
291 ret += atomic_read(&rwb->rq_wait[i].inflight);
292
293 return ret;
294}
295
296enum {
297 LAT_OK = 1,
298 LAT_UNKNOWN,
299 LAT_UNKNOWN_WRITES,
300 LAT_EXCEEDED,
301};
302
303static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
304{
305 struct backing_dev_info *bdi = rwb->rqos.disk->bdi;
306 struct rq_depth *rqd = &rwb->rq_depth;
307 u64 thislat;
308
309 /*
310 * If our stored sync issue exceeds the window size, or it
311 * exceeds our min target AND we haven't logged any entries,
312 * flag the latency as exceeded. wbt works off completion latencies,
313 * but for a flooded device, a single sync IO can take a long time
314 * to complete after being issued. If this time exceeds our
315 * monitoring window AND we didn't see any other completions in that
316 * window, then count that sync IO as a violation of the latency.
317 */
318 thislat = rwb_sync_issue_lat(rwb);
319 if (thislat > rwb->cur_win_nsec ||
320 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
321 trace_wbt_lat(bdi, thislat);
322 return LAT_EXCEEDED;
323 }
324
325 /*
326 * No read/write mix, if stat isn't valid
327 */
328 if (!stat_sample_valid(stat)) {
329 /*
330 * If we had writes in this stat window and the window is
331 * current, we're only doing writes. If a task recently
332 * waited or still has writes in flights, consider us doing
333 * just writes as well.
334 */
335 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
336 wbt_inflight(rwb))
337 return LAT_UNKNOWN_WRITES;
338 return LAT_UNKNOWN;
339 }
340
341 /*
342 * If the 'min' latency exceeds our target, step down.
343 */
344 if (stat[READ].min > rwb->min_lat_nsec) {
345 trace_wbt_lat(bdi, stat[READ].min);
346 trace_wbt_stat(bdi, stat);
347 return LAT_EXCEEDED;
348 }
349
350 if (rqd->scale_step)
351 trace_wbt_stat(bdi, stat);
352
353 return LAT_OK;
354}
355
356static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
357{
358 struct backing_dev_info *bdi = rwb->rqos.disk->bdi;
359 struct rq_depth *rqd = &rwb->rq_depth;
360
361 trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
362 rwb->wb_background, rwb->wb_normal, rqd->max_depth);
363}
364
365static void calc_wb_limits(struct rq_wb *rwb)
366{
367 if (rwb->min_lat_nsec == 0) {
368 rwb->wb_normal = rwb->wb_background = 0;
369 } else if (rwb->rq_depth.max_depth <= 2) {
370 rwb->wb_normal = rwb->rq_depth.max_depth;
371 rwb->wb_background = 1;
372 } else {
373 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
374 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
375 }
376}
377
378static void scale_up(struct rq_wb *rwb)
379{
380 if (!rq_depth_scale_up(&rwb->rq_depth))
381 return;
382 calc_wb_limits(rwb);
383 rwb->unknown_cnt = 0;
384 rwb_wake_all(rwb);
385 rwb_trace_step(rwb, tracepoint_string("scale up"));
386}
387
388static void scale_down(struct rq_wb *rwb, bool hard_throttle)
389{
390 if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
391 return;
392 calc_wb_limits(rwb);
393 rwb->unknown_cnt = 0;
394 rwb_trace_step(rwb, tracepoint_string("scale down"));
395}
396
397static void rwb_arm_timer(struct rq_wb *rwb)
398{
399 struct rq_depth *rqd = &rwb->rq_depth;
400
401 if (rqd->scale_step > 0) {
402 /*
403 * We should speed this up, using some variant of a fast
404 * integer inverse square root calculation. Since we only do
405 * this for every window expiration, it's not a huge deal,
406 * though.
407 */
408 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
409 int_sqrt((rqd->scale_step + 1) << 8));
410 } else {
411 /*
412 * For step < 0, we don't want to increase/decrease the
413 * window size.
414 */
415 rwb->cur_win_nsec = rwb->win_nsec;
416 }
417
418 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
419}
420
421static void wb_timer_fn(struct blk_stat_callback *cb)
422{
423 struct rq_wb *rwb = cb->data;
424 struct rq_depth *rqd = &rwb->rq_depth;
425 unsigned int inflight = wbt_inflight(rwb);
426 int status;
427
428 if (!rwb->rqos.disk)
429 return;
430
431 status = latency_exceeded(rwb, cb->stat);
432
433 trace_wbt_timer(rwb->rqos.disk->bdi, status, rqd->scale_step, inflight);
434
435 /*
436 * If we exceeded the latency target, step down. If we did not,
437 * step one level up. If we don't know enough to say either exceeded
438 * or ok, then don't do anything.
439 */
440 switch (status) {
441 case LAT_EXCEEDED:
442 scale_down(rwb, true);
443 break;
444 case LAT_OK:
445 scale_up(rwb);
446 break;
447 case LAT_UNKNOWN_WRITES:
448 /*
449 * We started a the center step, but don't have a valid
450 * read/write sample, but we do have writes going on.
451 * Allow step to go negative, to increase write perf.
452 */
453 scale_up(rwb);
454 break;
455 case LAT_UNKNOWN:
456 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
457 break;
458 /*
459 * We get here when previously scaled reduced depth, and we
460 * currently don't have a valid read/write sample. For that
461 * case, slowly return to center state (step == 0).
462 */
463 if (rqd->scale_step > 0)
464 scale_up(rwb);
465 else if (rqd->scale_step < 0)
466 scale_down(rwb, false);
467 break;
468 default:
469 break;
470 }
471
472 /*
473 * Re-arm timer, if we have IO in flight
474 */
475 if (rqd->scale_step || inflight)
476 rwb_arm_timer(rwb);
477}
478
479static void wbt_update_limits(struct rq_wb *rwb)
480{
481 struct rq_depth *rqd = &rwb->rq_depth;
482
483 rqd->scale_step = 0;
484 rqd->scaled_max = false;
485
486 rq_depth_calc_max_depth(rqd);
487 calc_wb_limits(rwb);
488
489 rwb_wake_all(rwb);
490}
491
492bool wbt_disabled(struct request_queue *q)
493{
494 struct rq_qos *rqos = wbt_rq_qos(q);
495
496 return !rqos || !rwb_enabled(RQWB(rqos));
497}
498
499u64 wbt_get_min_lat(struct request_queue *q)
500{
501 struct rq_qos *rqos = wbt_rq_qos(q);
502 if (!rqos)
503 return 0;
504 return RQWB(rqos)->min_lat_nsec;
505}
506
507void wbt_set_min_lat(struct request_queue *q, u64 val)
508{
509 struct rq_qos *rqos = wbt_rq_qos(q);
510 if (!rqos)
511 return;
512
513 RQWB(rqos)->min_lat_nsec = val;
514 if (val)
515 RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
516 else
517 RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL;
518
519 wbt_update_limits(RQWB(rqos));
520}
521
522
523static bool close_io(struct rq_wb *rwb)
524{
525 const unsigned long now = jiffies;
526
527 return time_before(now, rwb->last_issue + HZ / 10) ||
528 time_before(now, rwb->last_comp + HZ / 10);
529}
530
531#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
532
533static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf)
534{
535 unsigned int limit;
536
537 if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD)
538 return rwb->wb_background;
539
540 /*
541 * At this point we know it's a buffered write. If this is
542 * kswapd trying to free memory, or REQ_SYNC is set, then
543 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
544 * that. If the write is marked as a background write, then use
545 * the idle limit, or go to normal if we haven't had competing
546 * IO for a bit.
547 */
548 if ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
549 limit = rwb->rq_depth.max_depth;
550 else if ((opf & REQ_BACKGROUND) || close_io(rwb)) {
551 /*
552 * If less than 100ms since we completed unrelated IO,
553 * limit us to half the depth for background writeback.
554 */
555 limit = rwb->wb_background;
556 } else
557 limit = rwb->wb_normal;
558
559 return limit;
560}
561
562struct wbt_wait_data {
563 struct rq_wb *rwb;
564 enum wbt_flags wb_acct;
565 blk_opf_t opf;
566};
567
568static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
569{
570 struct wbt_wait_data *data = private_data;
571 return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf));
572}
573
574static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
575{
576 struct wbt_wait_data *data = private_data;
577 wbt_rqw_done(data->rwb, rqw, data->wb_acct);
578}
579
580/*
581 * Block if we will exceed our limit, or if we are currently waiting for
582 * the timer to kick off queuing again.
583 */
584static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
585 blk_opf_t opf)
586{
587 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
588 struct wbt_wait_data data = {
589 .rwb = rwb,
590 .wb_acct = wb_acct,
591 .opf = opf,
592 };
593
594 rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
595}
596
597static inline bool wbt_should_throttle(struct bio *bio)
598{
599 switch (bio_op(bio)) {
600 case REQ_OP_WRITE:
601 /*
602 * Don't throttle WRITE_ODIRECT
603 */
604 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
605 (REQ_SYNC | REQ_IDLE))
606 return false;
607 fallthrough;
608 case REQ_OP_DISCARD:
609 return true;
610 default:
611 return false;
612 }
613}
614
615static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
616{
617 enum wbt_flags flags = 0;
618
619 if (!rwb_enabled(rwb))
620 return 0;
621
622 if (bio_op(bio) == REQ_OP_READ) {
623 flags = WBT_READ;
624 } else if (wbt_should_throttle(bio)) {
625 if (current_is_kswapd())
626 flags |= WBT_KSWAPD;
627 if (bio_op(bio) == REQ_OP_DISCARD)
628 flags |= WBT_DISCARD;
629 flags |= WBT_TRACKED;
630 }
631 return flags;
632}
633
634static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
635{
636 struct rq_wb *rwb = RQWB(rqos);
637 enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
638 __wbt_done(rqos, flags);
639}
640
641/*
642 * May sleep, if we have exceeded the writeback limits. Caller can pass
643 * in an irq held spinlock, if it holds one when calling this function.
644 * If we do sleep, we'll release and re-grab it.
645 */
646static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
647{
648 struct rq_wb *rwb = RQWB(rqos);
649 enum wbt_flags flags;
650
651 flags = bio_to_wbt_flags(rwb, bio);
652 if (!(flags & WBT_TRACKED)) {
653 if (flags & WBT_READ)
654 wb_timestamp(rwb, &rwb->last_issue);
655 return;
656 }
657
658 __wbt_wait(rwb, flags, bio->bi_opf);
659
660 if (!blk_stat_is_active(rwb->cb))
661 rwb_arm_timer(rwb);
662}
663
664static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
665{
666 struct rq_wb *rwb = RQWB(rqos);
667 rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
668}
669
670static void wbt_issue(struct rq_qos *rqos, struct request *rq)
671{
672 struct rq_wb *rwb = RQWB(rqos);
673
674 if (!rwb_enabled(rwb))
675 return;
676
677 /*
678 * Track sync issue, in case it takes a long time to complete. Allows us
679 * to react quicker, if a sync IO takes a long time to complete. Note
680 * that this is just a hint. The request can go away when it completes,
681 * so it's important we never dereference it. We only use the address to
682 * compare with, which is why we store the sync_issue time locally.
683 */
684 if (wbt_is_read(rq) && !rwb->sync_issue) {
685 rwb->sync_cookie = rq;
686 rwb->sync_issue = rq->io_start_time_ns;
687 }
688}
689
690static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
691{
692 struct rq_wb *rwb = RQWB(rqos);
693 if (!rwb_enabled(rwb))
694 return;
695 if (rq == rwb->sync_cookie) {
696 rwb->sync_issue = 0;
697 rwb->sync_cookie = NULL;
698 }
699}
700
701/*
702 * Enable wbt if defaults are configured that way
703 */
704void wbt_enable_default(struct gendisk *disk)
705{
706 struct request_queue *q = disk->queue;
707 struct rq_qos *rqos;
708 bool enable = IS_ENABLED(CONFIG_BLK_WBT_MQ);
709
710 if (q->elevator &&
711 test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags))
712 enable = false;
713
714 /* Throttling already enabled? */
715 rqos = wbt_rq_qos(q);
716 if (rqos) {
717 if (enable && RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)
718 RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;
719 return;
720 }
721
722 /* Queue not registered? Maybe shutting down... */
723 if (!blk_queue_registered(q))
724 return;
725
726 if (queue_is_mq(q) && enable)
727 wbt_init(disk);
728}
729EXPORT_SYMBOL_GPL(wbt_enable_default);
730
731u64 wbt_default_latency_nsec(struct request_queue *q)
732{
733 /*
734 * We default to 2msec for non-rotational storage, and 75msec
735 * for rotational storage.
736 */
737 if (blk_queue_nonrot(q))
738 return 2000000ULL;
739 else
740 return 75000000ULL;
741}
742
743static int wbt_data_dir(const struct request *rq)
744{
745 const enum req_op op = req_op(rq);
746
747 if (op == REQ_OP_READ)
748 return READ;
749 else if (op_is_write(op))
750 return WRITE;
751
752 /* don't account */
753 return -1;
754}
755
756static void wbt_queue_depth_changed(struct rq_qos *rqos)
757{
758 RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->disk->queue);
759 wbt_update_limits(RQWB(rqos));
760}
761
762static void wbt_exit(struct rq_qos *rqos)
763{
764 struct rq_wb *rwb = RQWB(rqos);
765
766 blk_stat_remove_callback(rqos->disk->queue, rwb->cb);
767 blk_stat_free_callback(rwb->cb);
768 kfree(rwb);
769}
770
771/*
772 * Disable wbt, if enabled by default.
773 */
774void wbt_disable_default(struct gendisk *disk)
775{
776 struct rq_qos *rqos = wbt_rq_qos(disk->queue);
777 struct rq_wb *rwb;
778 if (!rqos)
779 return;
780 rwb = RQWB(rqos);
781 if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
782 blk_stat_deactivate(rwb->cb);
783 rwb->enable_state = WBT_STATE_OFF_DEFAULT;
784 }
785}
786EXPORT_SYMBOL_GPL(wbt_disable_default);
787
788#ifdef CONFIG_BLK_DEBUG_FS
789static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
790{
791 struct rq_qos *rqos = data;
792 struct rq_wb *rwb = RQWB(rqos);
793
794 seq_printf(m, "%llu\n", rwb->cur_win_nsec);
795 return 0;
796}
797
798static int wbt_enabled_show(void *data, struct seq_file *m)
799{
800 struct rq_qos *rqos = data;
801 struct rq_wb *rwb = RQWB(rqos);
802
803 seq_printf(m, "%d\n", rwb->enable_state);
804 return 0;
805}
806
807static int wbt_id_show(void *data, struct seq_file *m)
808{
809 struct rq_qos *rqos = data;
810
811 seq_printf(m, "%u\n", rqos->id);
812 return 0;
813}
814
815static int wbt_inflight_show(void *data, struct seq_file *m)
816{
817 struct rq_qos *rqos = data;
818 struct rq_wb *rwb = RQWB(rqos);
819 int i;
820
821 for (i = 0; i < WBT_NUM_RWQ; i++)
822 seq_printf(m, "%d: inflight %d\n", i,
823 atomic_read(&rwb->rq_wait[i].inflight));
824 return 0;
825}
826
827static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
828{
829 struct rq_qos *rqos = data;
830 struct rq_wb *rwb = RQWB(rqos);
831
832 seq_printf(m, "%lu\n", rwb->min_lat_nsec);
833 return 0;
834}
835
836static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
837{
838 struct rq_qos *rqos = data;
839 struct rq_wb *rwb = RQWB(rqos);
840
841 seq_printf(m, "%u\n", rwb->unknown_cnt);
842 return 0;
843}
844
845static int wbt_normal_show(void *data, struct seq_file *m)
846{
847 struct rq_qos *rqos = data;
848 struct rq_wb *rwb = RQWB(rqos);
849
850 seq_printf(m, "%u\n", rwb->wb_normal);
851 return 0;
852}
853
854static int wbt_background_show(void *data, struct seq_file *m)
855{
856 struct rq_qos *rqos = data;
857 struct rq_wb *rwb = RQWB(rqos);
858
859 seq_printf(m, "%u\n", rwb->wb_background);
860 return 0;
861}
862
863static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
864 {"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
865 {"enabled", 0400, wbt_enabled_show},
866 {"id", 0400, wbt_id_show},
867 {"inflight", 0400, wbt_inflight_show},
868 {"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
869 {"unknown_cnt", 0400, wbt_unknown_cnt_show},
870 {"wb_normal", 0400, wbt_normal_show},
871 {"wb_background", 0400, wbt_background_show},
872 {},
873};
874#endif
875
876static const struct rq_qos_ops wbt_rqos_ops = {
877 .throttle = wbt_wait,
878 .issue = wbt_issue,
879 .track = wbt_track,
880 .requeue = wbt_requeue,
881 .done = wbt_done,
882 .cleanup = wbt_cleanup,
883 .queue_depth_changed = wbt_queue_depth_changed,
884 .exit = wbt_exit,
885#ifdef CONFIG_BLK_DEBUG_FS
886 .debugfs_attrs = wbt_debugfs_attrs,
887#endif
888};
889
890int wbt_init(struct gendisk *disk)
891{
892 struct request_queue *q = disk->queue;
893 struct rq_wb *rwb;
894 int i;
895 int ret;
896
897 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
898 if (!rwb)
899 return -ENOMEM;
900
901 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
902 if (!rwb->cb) {
903 kfree(rwb);
904 return -ENOMEM;
905 }
906
907 for (i = 0; i < WBT_NUM_RWQ; i++)
908 rq_wait_init(&rwb->rq_wait[i]);
909
910 rwb->last_comp = rwb->last_issue = jiffies;
911 rwb->win_nsec = RWB_WINDOW_NSEC;
912 rwb->enable_state = WBT_STATE_ON_DEFAULT;
913 rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
914 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
915 rwb->rq_depth.queue_depth = blk_queue_depth(q);
916 wbt_update_limits(rwb);
917
918 /*
919 * Assign rwb and add the stats callback.
920 */
921 mutex_lock(&q->rq_qos_mutex);
922 ret = rq_qos_add(&rwb->rqos, disk, RQ_QOS_WBT, &wbt_rqos_ops);
923 mutex_unlock(&q->rq_qos_mutex);
924 if (ret)
925 goto err_free;
926
927 blk_stat_add_callback(q, rwb->cb);
928
929 return 0;
930
931err_free:
932 blk_stat_free_callback(rwb->cb);
933 kfree(rwb);
934 return ret;
935
936}