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->enable_state != WBT_STATE_OFF_DEFAULT &&
84 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
86 if (rwb_enabled(rwb)) {
87 const unsigned long cur = jiffies;
95 * If a task was rate throttled in balance_dirty_pages() within the last
96 * second or so, use that to indicate a higher cleaning rate.
98 static bool wb_recent_wait(struct rq_wb *rwb)
100 struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb;
102 return time_before(jiffies, wb->dirty_sleep + HZ);
105 static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
106 enum wbt_flags wb_acct)
108 if (wb_acct & WBT_KSWAPD)
109 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
110 else if (wb_acct & WBT_DISCARD)
111 return &rwb->rq_wait[WBT_RWQ_DISCARD];
113 return &rwb->rq_wait[WBT_RWQ_BG];
116 static void rwb_wake_all(struct rq_wb *rwb)
120 for (i = 0; i < WBT_NUM_RWQ; i++) {
121 struct rq_wait *rqw = &rwb->rq_wait[i];
123 if (wq_has_sleeper(&rqw->wait))
124 wake_up_all(&rqw->wait);
128 static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
129 enum wbt_flags wb_acct)
133 inflight = atomic_dec_return(&rqw->inflight);
136 * wbt got disabled with IO in flight. Wake up any potential
137 * waiters, we don't have to do more than that.
139 if (unlikely(!rwb_enabled(rwb))) {
145 * For discards, our limit is always the background. For writes, if
146 * the device does write back caching, drop further down before we
149 if (wb_acct & WBT_DISCARD)
150 limit = rwb->wb_background;
151 else if (rwb->wc && !wb_recent_wait(rwb))
154 limit = rwb->wb_normal;
157 * Don't wake anyone up if we are above the normal limit.
159 if (inflight && inflight >= limit)
162 if (wq_has_sleeper(&rqw->wait)) {
163 int diff = limit - inflight;
165 if (!inflight || diff >= rwb->wb_background / 2)
166 wake_up_all(&rqw->wait);
170 static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
172 struct rq_wb *rwb = RQWB(rqos);
175 if (!(wb_acct & WBT_TRACKED))
178 rqw = get_rq_wait(rwb, wb_acct);
179 wbt_rqw_done(rwb, rqw, wb_acct);
183 * Called on completion of a request. Note that it's also called when
184 * a request is merged, when the request gets freed.
186 static void wbt_done(struct rq_qos *rqos, struct request *rq)
188 struct rq_wb *rwb = RQWB(rqos);
190 if (!wbt_is_tracked(rq)) {
191 if (rwb->sync_cookie == rq) {
193 rwb->sync_cookie = NULL;
197 wb_timestamp(rwb, &rwb->last_comp);
199 WARN_ON_ONCE(rq == rwb->sync_cookie);
200 __wbt_done(rqos, wbt_flags(rq));
205 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
208 * We need at least one read sample, and a minimum of
209 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
210 * that it's writes impacting us, and not just some sole read on
211 * a device that is in a lower power state.
213 return (stat[READ].nr_samples >= 1 &&
214 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
217 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
219 u64 now, issue = READ_ONCE(rwb->sync_issue);
221 if (!issue || !rwb->sync_cookie)
224 now = ktime_to_ns(ktime_get());
235 static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
237 struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
238 struct rq_depth *rqd = &rwb->rq_depth;
242 * If our stored sync issue exceeds the window size, or it
243 * exceeds our min target AND we haven't logged any entries,
244 * flag the latency as exceeded. wbt works off completion latencies,
245 * but for a flooded device, a single sync IO can take a long time
246 * to complete after being issued. If this time exceeds our
247 * monitoring window AND we didn't see any other completions in that
248 * window, then count that sync IO as a violation of the latency.
250 thislat = rwb_sync_issue_lat(rwb);
251 if (thislat > rwb->cur_win_nsec ||
252 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
253 trace_wbt_lat(bdi, thislat);
258 * No read/write mix, if stat isn't valid
260 if (!stat_sample_valid(stat)) {
262 * If we had writes in this stat window and the window is
263 * current, we're only doing writes. If a task recently
264 * waited or still has writes in flights, consider us doing
265 * just writes as well.
267 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
269 return LAT_UNKNOWN_WRITES;
274 * If the 'min' latency exceeds our target, step down.
276 if (stat[READ].min > rwb->min_lat_nsec) {
277 trace_wbt_lat(bdi, stat[READ].min);
278 trace_wbt_stat(bdi, stat);
283 trace_wbt_stat(bdi, stat);
288 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
290 struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
291 struct rq_depth *rqd = &rwb->rq_depth;
293 trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
294 rwb->wb_background, rwb->wb_normal, rqd->max_depth);
297 static void calc_wb_limits(struct rq_wb *rwb)
299 if (rwb->min_lat_nsec == 0) {
300 rwb->wb_normal = rwb->wb_background = 0;
301 } else if (rwb->rq_depth.max_depth <= 2) {
302 rwb->wb_normal = rwb->rq_depth.max_depth;
303 rwb->wb_background = 1;
305 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
306 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
310 static void scale_up(struct rq_wb *rwb)
312 if (!rq_depth_scale_up(&rwb->rq_depth))
315 rwb->unknown_cnt = 0;
317 rwb_trace_step(rwb, tracepoint_string("scale up"));
320 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
322 if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
325 rwb->unknown_cnt = 0;
326 rwb_trace_step(rwb, tracepoint_string("scale down"));
329 static void rwb_arm_timer(struct rq_wb *rwb)
331 struct rq_depth *rqd = &rwb->rq_depth;
333 if (rqd->scale_step > 0) {
335 * We should speed this up, using some variant of a fast
336 * integer inverse square root calculation. Since we only do
337 * this for every window expiration, it's not a huge deal,
340 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
341 int_sqrt((rqd->scale_step + 1) << 8));
344 * For step < 0, we don't want to increase/decrease the
347 rwb->cur_win_nsec = rwb->win_nsec;
350 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
353 static void wb_timer_fn(struct blk_stat_callback *cb)
355 struct rq_wb *rwb = cb->data;
356 struct rq_depth *rqd = &rwb->rq_depth;
357 unsigned int inflight = wbt_inflight(rwb);
360 if (!rwb->rqos.q->disk)
363 status = latency_exceeded(rwb, cb->stat);
365 trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step,
369 * If we exceeded the latency target, step down. If we did not,
370 * step one level up. If we don't know enough to say either exceeded
371 * or ok, then don't do anything.
375 scale_down(rwb, true);
380 case LAT_UNKNOWN_WRITES:
382 * We started a the center step, but don't have a valid
383 * read/write sample, but we do have writes going on.
384 * Allow step to go negative, to increase write perf.
389 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
392 * We get here when previously scaled reduced depth, and we
393 * currently don't have a valid read/write sample. For that
394 * case, slowly return to center state (step == 0).
396 if (rqd->scale_step > 0)
398 else if (rqd->scale_step < 0)
399 scale_down(rwb, false);
406 * Re-arm timer, if we have IO in flight
408 if (rqd->scale_step || inflight)
412 static void wbt_update_limits(struct rq_wb *rwb)
414 struct rq_depth *rqd = &rwb->rq_depth;
417 rqd->scaled_max = false;
419 rq_depth_calc_max_depth(rqd);
425 u64 wbt_get_min_lat(struct request_queue *q)
427 struct rq_qos *rqos = wbt_rq_qos(q);
430 return RQWB(rqos)->min_lat_nsec;
433 void wbt_set_min_lat(struct request_queue *q, u64 val)
435 struct rq_qos *rqos = wbt_rq_qos(q);
438 RQWB(rqos)->min_lat_nsec = val;
439 RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
440 wbt_update_limits(RQWB(rqos));
444 static bool close_io(struct rq_wb *rwb)
446 const unsigned long now = jiffies;
448 return time_before(now, rwb->last_issue + HZ / 10) ||
449 time_before(now, rwb->last_comp + HZ / 10);
452 #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
454 static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf)
459 * If we got disabled, just return UINT_MAX. This ensures that
460 * we'll properly inc a new IO, and dec+wakeup at the end.
462 if (!rwb_enabled(rwb))
465 if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD)
466 return rwb->wb_background;
469 * At this point we know it's a buffered write. If this is
470 * kswapd trying to free memory, or REQ_SYNC is set, then
471 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
472 * that. If the write is marked as a background write, then use
473 * the idle limit, or go to normal if we haven't had competing
476 if ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
477 limit = rwb->rq_depth.max_depth;
478 else if ((opf & REQ_BACKGROUND) || close_io(rwb)) {
480 * If less than 100ms since we completed unrelated IO,
481 * limit us to half the depth for background writeback.
483 limit = rwb->wb_background;
485 limit = rwb->wb_normal;
490 struct wbt_wait_data {
492 enum wbt_flags wb_acct;
496 static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
498 struct wbt_wait_data *data = private_data;
499 return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf));
502 static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
504 struct wbt_wait_data *data = private_data;
505 wbt_rqw_done(data->rwb, rqw, data->wb_acct);
509 * Block if we will exceed our limit, or if we are currently waiting for
510 * the timer to kick off queuing again.
512 static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
515 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
516 struct wbt_wait_data data = {
522 rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
525 static inline bool wbt_should_throttle(struct bio *bio)
527 switch (bio_op(bio)) {
530 * Don't throttle WRITE_ODIRECT
532 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
533 (REQ_SYNC | REQ_IDLE))
543 static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
545 enum wbt_flags flags = 0;
547 if (!rwb_enabled(rwb))
550 if (bio_op(bio) == REQ_OP_READ) {
552 } else if (wbt_should_throttle(bio)) {
553 if (current_is_kswapd())
555 if (bio_op(bio) == REQ_OP_DISCARD)
556 flags |= WBT_DISCARD;
557 flags |= WBT_TRACKED;
562 static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
564 struct rq_wb *rwb = RQWB(rqos);
565 enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
566 __wbt_done(rqos, flags);
570 * May sleep, if we have exceeded the writeback limits. Caller can pass
571 * in an irq held spinlock, if it holds one when calling this function.
572 * If we do sleep, we'll release and re-grab it.
574 static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
576 struct rq_wb *rwb = RQWB(rqos);
577 enum wbt_flags flags;
579 flags = bio_to_wbt_flags(rwb, bio);
580 if (!(flags & WBT_TRACKED)) {
581 if (flags & WBT_READ)
582 wb_timestamp(rwb, &rwb->last_issue);
586 __wbt_wait(rwb, flags, bio->bi_opf);
588 if (!blk_stat_is_active(rwb->cb))
592 static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
594 struct rq_wb *rwb = RQWB(rqos);
595 rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
598 static void wbt_issue(struct rq_qos *rqos, struct request *rq)
600 struct rq_wb *rwb = RQWB(rqos);
602 if (!rwb_enabled(rwb))
606 * Track sync issue, in case it takes a long time to complete. Allows us
607 * to react quicker, if a sync IO takes a long time to complete. Note
608 * that this is just a hint. The request can go away when it completes,
609 * so it's important we never dereference it. We only use the address to
610 * compare with, which is why we store the sync_issue time locally.
612 if (wbt_is_read(rq) && !rwb->sync_issue) {
613 rwb->sync_cookie = rq;
614 rwb->sync_issue = rq->io_start_time_ns;
618 static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
620 struct rq_wb *rwb = RQWB(rqos);
621 if (!rwb_enabled(rwb))
623 if (rq == rwb->sync_cookie) {
625 rwb->sync_cookie = NULL;
629 void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
631 struct rq_qos *rqos = wbt_rq_qos(q);
633 RQWB(rqos)->wc = write_cache_on;
637 * Enable wbt if defaults are configured that way
639 void wbt_enable_default(struct request_queue *q)
641 struct rq_qos *rqos = wbt_rq_qos(q);
643 /* Throttling already enabled? */
645 if (RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)
646 RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;
650 /* Queue not registered? Maybe shutting down... */
651 if (!blk_queue_registered(q))
654 if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ))
657 EXPORT_SYMBOL_GPL(wbt_enable_default);
659 u64 wbt_default_latency_nsec(struct request_queue *q)
662 * We default to 2msec for non-rotational storage, and 75msec
663 * for rotational storage.
665 if (blk_queue_nonrot(q))
671 static int wbt_data_dir(const struct request *rq)
673 const enum req_op op = req_op(rq);
675 if (op == REQ_OP_READ)
677 else if (op_is_write(op))
684 static void wbt_queue_depth_changed(struct rq_qos *rqos)
686 RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);
687 wbt_update_limits(RQWB(rqos));
690 static void wbt_exit(struct rq_qos *rqos)
692 struct rq_wb *rwb = RQWB(rqos);
693 struct request_queue *q = rqos->q;
695 blk_stat_remove_callback(q, rwb->cb);
696 blk_stat_free_callback(rwb->cb);
701 * Disable wbt, if enabled by default.
703 void wbt_disable_default(struct request_queue *q)
705 struct rq_qos *rqos = wbt_rq_qos(q);
710 if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
711 blk_stat_deactivate(rwb->cb);
712 rwb->enable_state = WBT_STATE_OFF_DEFAULT;
715 EXPORT_SYMBOL_GPL(wbt_disable_default);
717 #ifdef CONFIG_BLK_DEBUG_FS
718 static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
720 struct rq_qos *rqos = data;
721 struct rq_wb *rwb = RQWB(rqos);
723 seq_printf(m, "%llu\n", rwb->cur_win_nsec);
727 static int wbt_enabled_show(void *data, struct seq_file *m)
729 struct rq_qos *rqos = data;
730 struct rq_wb *rwb = RQWB(rqos);
732 seq_printf(m, "%d\n", rwb->enable_state);
736 static int wbt_id_show(void *data, struct seq_file *m)
738 struct rq_qos *rqos = data;
740 seq_printf(m, "%u\n", rqos->id);
744 static int wbt_inflight_show(void *data, struct seq_file *m)
746 struct rq_qos *rqos = data;
747 struct rq_wb *rwb = RQWB(rqos);
750 for (i = 0; i < WBT_NUM_RWQ; i++)
751 seq_printf(m, "%d: inflight %d\n", i,
752 atomic_read(&rwb->rq_wait[i].inflight));
756 static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
758 struct rq_qos *rqos = data;
759 struct rq_wb *rwb = RQWB(rqos);
761 seq_printf(m, "%lu\n", rwb->min_lat_nsec);
765 static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
767 struct rq_qos *rqos = data;
768 struct rq_wb *rwb = RQWB(rqos);
770 seq_printf(m, "%u\n", rwb->unknown_cnt);
774 static int wbt_normal_show(void *data, struct seq_file *m)
776 struct rq_qos *rqos = data;
777 struct rq_wb *rwb = RQWB(rqos);
779 seq_printf(m, "%u\n", rwb->wb_normal);
783 static int wbt_background_show(void *data, struct seq_file *m)
785 struct rq_qos *rqos = data;
786 struct rq_wb *rwb = RQWB(rqos);
788 seq_printf(m, "%u\n", rwb->wb_background);
792 static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
793 {"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
794 {"enabled", 0400, wbt_enabled_show},
795 {"id", 0400, wbt_id_show},
796 {"inflight", 0400, wbt_inflight_show},
797 {"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
798 {"unknown_cnt", 0400, wbt_unknown_cnt_show},
799 {"wb_normal", 0400, wbt_normal_show},
800 {"wb_background", 0400, wbt_background_show},
805 static struct rq_qos_ops wbt_rqos_ops = {
806 .throttle = wbt_wait,
809 .requeue = wbt_requeue,
811 .cleanup = wbt_cleanup,
812 .queue_depth_changed = wbt_queue_depth_changed,
814 #ifdef CONFIG_BLK_DEBUG_FS
815 .debugfs_attrs = wbt_debugfs_attrs,
819 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;
844 rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags);
845 rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
846 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
848 wbt_queue_depth_changed(&rwb->rqos);
851 * Assign rwb and add the stats callback.
853 ret = rq_qos_add(q, &rwb->rqos);
857 blk_stat_add_callback(q, rwb->cb);
862 blk_stat_free_callback(rwb->cb);