1 // SPDX-License-Identifier: GPL-2.0
3 * buffered writeback throttling. loosely based on CoDel. We can't drop
4 * packets for IO scheduling, so the logic is something like this:
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.
19 * Copyright (C) 2016 Jens Axboe
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>
29 #include "blk-rq-qos.h"
31 #define CREATE_TRACE_POINTS
32 #include <trace/events/wbt.h>
34 static inline void wbt_clear_state(struct request *rq)
39 static inline enum wbt_flags wbt_flags(struct request *rq)
44 static inline bool wbt_is_tracked(struct request *rq)
46 return rq->wbt_flags & WBT_TRACKED;
49 static inline bool wbt_is_read(struct request *rq)
51 return rq->wbt_flags & WBT_READ;
56 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
57 * from here depending on device stats
64 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
67 * Disregard stats, if we don't meet this minimum
69 RWB_MIN_WRITE_SAMPLES = 3,
72 * If we have this number of consecutive windows with not enough
73 * information to scale up or down, scale up.
78 static inline bool rwb_enabled(struct rq_wb *rwb)
80 return rwb && rwb->wb_normal != 0;
83 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
85 if (rwb_enabled(rwb)) {
86 const unsigned long cur = jiffies;
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.
97 static bool wb_recent_wait(struct rq_wb *rwb)
99 struct bdi_writeback *wb = &rwb->rqos.q->backing_dev_info->wb;
101 return time_before(jiffies, wb->dirty_sleep + HZ);
104 static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
105 enum wbt_flags wb_acct)
107 if (wb_acct & WBT_KSWAPD)
108 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
109 else if (wb_acct & WBT_DISCARD)
110 return &rwb->rq_wait[WBT_RWQ_DISCARD];
112 return &rwb->rq_wait[WBT_RWQ_BG];
115 static void rwb_wake_all(struct rq_wb *rwb)
119 for (i = 0; i < WBT_NUM_RWQ; i++) {
120 struct rq_wait *rqw = &rwb->rq_wait[i];
122 if (wq_has_sleeper(&rqw->wait))
123 wake_up_all(&rqw->wait);
127 static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
128 enum wbt_flags wb_acct)
132 inflight = atomic_dec_return(&rqw->inflight);
135 * wbt got disabled with IO in flight. Wake up any potential
136 * waiters, we don't have to do more than that.
138 if (unlikely(!rwb_enabled(rwb))) {
144 * For discards, our limit is always the background. For writes, if
145 * the device does write back caching, drop further down before we
148 if (wb_acct & WBT_DISCARD)
149 limit = rwb->wb_background;
150 else if (rwb->wc && !wb_recent_wait(rwb))
153 limit = rwb->wb_normal;
156 * Don't wake anyone up if we are above the normal limit.
158 if (inflight && inflight >= limit)
161 if (wq_has_sleeper(&rqw->wait)) {
162 int diff = limit - inflight;
164 if (!inflight || diff >= rwb->wb_background / 2)
165 wake_up_all(&rqw->wait);
169 static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
171 struct rq_wb *rwb = RQWB(rqos);
174 if (!(wb_acct & WBT_TRACKED))
177 rqw = get_rq_wait(rwb, wb_acct);
178 wbt_rqw_done(rwb, rqw, wb_acct);
182 * Called on completion of a request. Note that it's also called when
183 * a request is merged, when the request gets freed.
185 static void wbt_done(struct rq_qos *rqos, struct request *rq)
187 struct rq_wb *rwb = RQWB(rqos);
189 if (!wbt_is_tracked(rq)) {
190 if (rwb->sync_cookie == rq) {
192 rwb->sync_cookie = NULL;
196 wb_timestamp(rwb, &rwb->last_comp);
198 WARN_ON_ONCE(rq == rwb->sync_cookie);
199 __wbt_done(rqos, wbt_flags(rq));
204 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
207 * We need at least one read sample, and a minimum of
208 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
209 * that it's writes impacting us, and not just some sole read on
210 * a device that is in a lower power state.
212 return (stat[READ].nr_samples >= 1 &&
213 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
216 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
218 u64 now, issue = READ_ONCE(rwb->sync_issue);
220 if (!issue || !rwb->sync_cookie)
223 now = ktime_to_ns(ktime_get());
234 static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
236 struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
237 struct rq_depth *rqd = &rwb->rq_depth;
241 * If our stored sync issue exceeds the window size, or it
242 * exceeds our min target AND we haven't logged any entries,
243 * flag the latency as exceeded. wbt works off completion latencies,
244 * but for a flooded device, a single sync IO can take a long time
245 * to complete after being issued. If this time exceeds our
246 * monitoring window AND we didn't see any other completions in that
247 * window, then count that sync IO as a violation of the latency.
249 thislat = rwb_sync_issue_lat(rwb);
250 if (thislat > rwb->cur_win_nsec ||
251 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
252 trace_wbt_lat(bdi, thislat);
257 * No read/write mix, if stat isn't valid
259 if (!stat_sample_valid(stat)) {
261 * If we had writes in this stat window and the window is
262 * current, we're only doing writes. If a task recently
263 * waited or still has writes in flights, consider us doing
264 * just writes as well.
266 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
268 return LAT_UNKNOWN_WRITES;
273 * If the 'min' latency exceeds our target, step down.
275 if (stat[READ].min > rwb->min_lat_nsec) {
276 trace_wbt_lat(bdi, stat[READ].min);
277 trace_wbt_stat(bdi, stat);
282 trace_wbt_stat(bdi, stat);
287 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
289 struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
290 struct rq_depth *rqd = &rwb->rq_depth;
292 trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
293 rwb->wb_background, rwb->wb_normal, rqd->max_depth);
296 static void calc_wb_limits(struct rq_wb *rwb)
298 if (rwb->min_lat_nsec == 0) {
299 rwb->wb_normal = rwb->wb_background = 0;
300 } else if (rwb->rq_depth.max_depth <= 2) {
301 rwb->wb_normal = rwb->rq_depth.max_depth;
302 rwb->wb_background = 1;
304 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
305 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
309 static void scale_up(struct rq_wb *rwb)
311 if (!rq_depth_scale_up(&rwb->rq_depth))
314 rwb->unknown_cnt = 0;
316 rwb_trace_step(rwb, "scale up");
319 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
321 if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
324 rwb->unknown_cnt = 0;
325 rwb_trace_step(rwb, "scale down");
328 static void rwb_arm_timer(struct rq_wb *rwb)
330 struct rq_depth *rqd = &rwb->rq_depth;
332 if (rqd->scale_step > 0) {
334 * We should speed this up, using some variant of a fast
335 * integer inverse square root calculation. Since we only do
336 * this for every window expiration, it's not a huge deal,
339 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
340 int_sqrt((rqd->scale_step + 1) << 8));
343 * For step < 0, we don't want to increase/decrease the
346 rwb->cur_win_nsec = rwb->win_nsec;
349 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
352 static void wb_timer_fn(struct blk_stat_callback *cb)
354 struct rq_wb *rwb = cb->data;
355 struct rq_depth *rqd = &rwb->rq_depth;
356 unsigned int inflight = wbt_inflight(rwb);
359 status = latency_exceeded(rwb, cb->stat);
361 trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step,
365 * If we exceeded the latency target, step down. If we did not,
366 * step one level up. If we don't know enough to say either exceeded
367 * or ok, then don't do anything.
371 scale_down(rwb, true);
376 case LAT_UNKNOWN_WRITES:
378 * We started a the center step, but don't have a valid
379 * read/write sample, but we do have writes going on.
380 * Allow step to go negative, to increase write perf.
385 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
388 * We get here when previously scaled reduced depth, and we
389 * currently don't have a valid read/write sample. For that
390 * case, slowly return to center state (step == 0).
392 if (rqd->scale_step > 0)
394 else if (rqd->scale_step < 0)
395 scale_down(rwb, false);
402 * Re-arm timer, if we have IO in flight
404 if (rqd->scale_step || inflight)
408 static void __wbt_update_limits(struct rq_wb *rwb)
410 struct rq_depth *rqd = &rwb->rq_depth;
413 rqd->scaled_max = false;
415 rq_depth_calc_max_depth(rqd);
421 void wbt_update_limits(struct request_queue *q)
423 struct rq_qos *rqos = wbt_rq_qos(q);
426 __wbt_update_limits(RQWB(rqos));
429 u64 wbt_get_min_lat(struct request_queue *q)
431 struct rq_qos *rqos = wbt_rq_qos(q);
434 return RQWB(rqos)->min_lat_nsec;
437 void wbt_set_min_lat(struct request_queue *q, u64 val)
439 struct rq_qos *rqos = wbt_rq_qos(q);
442 RQWB(rqos)->min_lat_nsec = val;
443 RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
444 __wbt_update_limits(RQWB(rqos));
448 static bool close_io(struct rq_wb *rwb)
450 const unsigned long now = jiffies;
452 return time_before(now, rwb->last_issue + HZ / 10) ||
453 time_before(now, rwb->last_comp + HZ / 10);
456 #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
458 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
463 * If we got disabled, just return UINT_MAX. This ensures that
464 * we'll properly inc a new IO, and dec+wakeup at the end.
466 if (!rwb_enabled(rwb))
469 if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD)
470 return rwb->wb_background;
473 * At this point we know it's a buffered write. If this is
474 * kswapd trying to free memory, or REQ_SYNC is set, then
475 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
476 * that. If the write is marked as a background write, then use
477 * the idle limit, or go to normal if we haven't had competing
480 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
481 limit = rwb->rq_depth.max_depth;
482 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
484 * If less than 100ms since we completed unrelated IO,
485 * limit us to half the depth for background writeback.
487 limit = rwb->wb_background;
489 limit = rwb->wb_normal;
494 struct wbt_wait_data {
496 enum wbt_flags wb_acct;
500 static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
502 struct wbt_wait_data *data = private_data;
503 return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw));
506 static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
508 struct wbt_wait_data *data = private_data;
509 wbt_rqw_done(data->rwb, rqw, data->wb_acct);
513 * Block if we will exceed our limit, or if we are currently waiting for
514 * the timer to kick off queuing again.
516 static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
519 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
520 struct wbt_wait_data data = {
526 rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
529 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
531 switch (bio_op(bio)) {
534 * Don't throttle WRITE_ODIRECT
536 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
537 (REQ_SYNC | REQ_IDLE))
547 static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
549 enum wbt_flags flags = 0;
551 if (!rwb_enabled(rwb))
554 if (bio_op(bio) == REQ_OP_READ) {
556 } else if (wbt_should_throttle(rwb, bio)) {
557 if (current_is_kswapd())
559 if (bio_op(bio) == REQ_OP_DISCARD)
560 flags |= WBT_DISCARD;
561 flags |= WBT_TRACKED;
566 static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
568 struct rq_wb *rwb = RQWB(rqos);
569 enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
570 __wbt_done(rqos, flags);
574 * Returns true if the IO request should be accounted, false if not.
575 * May sleep, if we have exceeded the writeback limits. Caller can pass
576 * in an irq held spinlock, if it holds one when calling this function.
577 * If we do sleep, we'll release and re-grab it.
579 static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
581 struct rq_wb *rwb = RQWB(rqos);
582 enum wbt_flags flags;
584 flags = bio_to_wbt_flags(rwb, bio);
585 if (!(flags & WBT_TRACKED)) {
586 if (flags & WBT_READ)
587 wb_timestamp(rwb, &rwb->last_issue);
591 __wbt_wait(rwb, flags, bio->bi_opf);
593 if (!blk_stat_is_active(rwb->cb))
597 static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
599 struct rq_wb *rwb = RQWB(rqos);
600 rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
603 static void wbt_issue(struct rq_qos *rqos, struct request *rq)
605 struct rq_wb *rwb = RQWB(rqos);
607 if (!rwb_enabled(rwb))
611 * Track sync issue, in case it takes a long time to complete. Allows us
612 * to react quicker, if a sync IO takes a long time to complete. Note
613 * that this is just a hint. The request can go away when it completes,
614 * so it's important we never dereference it. We only use the address to
615 * compare with, which is why we store the sync_issue time locally.
617 if (wbt_is_read(rq) && !rwb->sync_issue) {
618 rwb->sync_cookie = rq;
619 rwb->sync_issue = rq->io_start_time_ns;
623 static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
625 struct rq_wb *rwb = RQWB(rqos);
626 if (!rwb_enabled(rwb))
628 if (rq == rwb->sync_cookie) {
630 rwb->sync_cookie = NULL;
634 void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
636 struct rq_qos *rqos = wbt_rq_qos(q);
638 RQWB(rqos)->wc = write_cache_on;
642 * Enable wbt if defaults are configured that way
644 void wbt_enable_default(struct request_queue *q)
646 struct rq_qos *rqos = wbt_rq_qos(q);
647 /* Throttling already enabled? */
651 /* Queue not registered? Maybe shutting down... */
652 if (!blk_queue_registered(q))
655 if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ))
658 EXPORT_SYMBOL_GPL(wbt_enable_default);
660 u64 wbt_default_latency_nsec(struct request_queue *q)
663 * We default to 2msec for non-rotational storage, and 75msec
664 * for rotational storage.
666 if (blk_queue_nonrot(q))
672 static int wbt_data_dir(const struct request *rq)
674 const int op = req_op(rq);
676 if (op == REQ_OP_READ)
678 else if (op_is_write(op))
685 static void wbt_queue_depth_changed(struct rq_qos *rqos)
687 RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);
688 __wbt_update_limits(RQWB(rqos));
691 static void wbt_exit(struct rq_qos *rqos)
693 struct rq_wb *rwb = RQWB(rqos);
694 struct request_queue *q = rqos->q;
696 blk_stat_remove_callback(q, rwb->cb);
697 blk_stat_free_callback(rwb->cb);
702 * Disable wbt, if enabled by default.
704 void wbt_disable_default(struct request_queue *q)
706 struct rq_qos *rqos = wbt_rq_qos(q);
711 if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
712 blk_stat_deactivate(rwb->cb);
716 EXPORT_SYMBOL_GPL(wbt_disable_default);
718 #ifdef CONFIG_BLK_DEBUG_FS
719 static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
721 struct rq_qos *rqos = data;
722 struct rq_wb *rwb = RQWB(rqos);
724 seq_printf(m, "%llu\n", rwb->cur_win_nsec);
728 static int wbt_enabled_show(void *data, struct seq_file *m)
730 struct rq_qos *rqos = data;
731 struct rq_wb *rwb = RQWB(rqos);
733 seq_printf(m, "%d\n", rwb->enable_state);
737 static int wbt_id_show(void *data, struct seq_file *m)
739 struct rq_qos *rqos = data;
741 seq_printf(m, "%u\n", rqos->id);
745 static int wbt_inflight_show(void *data, struct seq_file *m)
747 struct rq_qos *rqos = data;
748 struct rq_wb *rwb = RQWB(rqos);
751 for (i = 0; i < WBT_NUM_RWQ; i++)
752 seq_printf(m, "%d: inflight %d\n", i,
753 atomic_read(&rwb->rq_wait[i].inflight));
757 static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
759 struct rq_qos *rqos = data;
760 struct rq_wb *rwb = RQWB(rqos);
762 seq_printf(m, "%lu\n", rwb->min_lat_nsec);
766 static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
768 struct rq_qos *rqos = data;
769 struct rq_wb *rwb = RQWB(rqos);
771 seq_printf(m, "%u\n", rwb->unknown_cnt);
775 static int wbt_normal_show(void *data, struct seq_file *m)
777 struct rq_qos *rqos = data;
778 struct rq_wb *rwb = RQWB(rqos);
780 seq_printf(m, "%u\n", rwb->wb_normal);
784 static int wbt_background_show(void *data, struct seq_file *m)
786 struct rq_qos *rqos = data;
787 struct rq_wb *rwb = RQWB(rqos);
789 seq_printf(m, "%u\n", rwb->wb_background);
793 static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
794 {"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
795 {"enabled", 0400, wbt_enabled_show},
796 {"id", 0400, wbt_id_show},
797 {"inflight", 0400, wbt_inflight_show},
798 {"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
799 {"unknown_cnt", 0400, wbt_unknown_cnt_show},
800 {"wb_normal", 0400, wbt_normal_show},
801 {"wb_background", 0400, wbt_background_show},
806 static struct rq_qos_ops wbt_rqos_ops = {
807 .throttle = wbt_wait,
810 .requeue = wbt_requeue,
812 .cleanup = wbt_cleanup,
813 .queue_depth_changed = wbt_queue_depth_changed,
815 #ifdef CONFIG_BLK_DEBUG_FS
816 .debugfs_attrs = wbt_debugfs_attrs,
820 int wbt_init(struct request_queue *q)
825 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
829 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
835 for (i = 0; i < WBT_NUM_RWQ; i++)
836 rq_wait_init(&rwb->rq_wait[i]);
838 rwb->rqos.id = RQ_QOS_WBT;
839 rwb->rqos.ops = &wbt_rqos_ops;
841 rwb->last_comp = rwb->last_issue = jiffies;
842 rwb->win_nsec = RWB_WINDOW_NSEC;
843 rwb->enable_state = WBT_STATE_ON_DEFAULT;
845 rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
846 __wbt_update_limits(rwb);
849 * Assign rwb and add the stats callback.
851 rq_qos_add(q, &rwb->rqos);
852 blk_stat_add_callback(q, rwb->cb);
854 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
856 wbt_queue_depth_changed(&rwb->rqos);
857 wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags));