2 * buffered writeback throttling. loosely based on CoDel. We can't drop
3 * packets for IO scheduling, so the logic is something like this:
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.
18 * Copyright (C) 2016 Jens Axboe
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>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/wbt.h>
32 static inline void wbt_clear_state(struct request *rq)
37 static inline enum wbt_flags wbt_flags(struct request *rq)
42 static inline bool wbt_is_tracked(struct request *rq)
44 return rq->wbt_flags & WBT_TRACKED;
47 static inline bool wbt_is_read(struct request *rq)
49 return rq->wbt_flags & WBT_READ;
54 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
55 * from here depending on device stats
62 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
65 * Disregard stats, if we don't meet this minimum
67 RWB_MIN_WRITE_SAMPLES = 3,
70 * If we have this number of consecutive windows with not enough
71 * information to scale up or down, scale up.
76 static inline bool rwb_enabled(struct rq_wb *rwb)
78 return rwb && rwb->wb_normal != 0;
82 * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded,
83 * false if 'v' + 1 would be bigger than 'below'.
85 static bool atomic_inc_below(atomic_t *v, int below)
87 int cur = atomic_read(v);
94 old = atomic_cmpxchg(v, cur, cur + 1);
103 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
105 if (rwb_enabled(rwb)) {
106 const unsigned long cur = jiffies;
114 * If a task was rate throttled in balance_dirty_pages() within the last
115 * second or so, use that to indicate a higher cleaning rate.
117 static bool wb_recent_wait(struct rq_wb *rwb)
119 struct bdi_writeback *wb = &rwb->queue->backing_dev_info->wb;
121 return time_before(jiffies, wb->dirty_sleep + HZ);
124 static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
125 enum wbt_flags wb_acct)
127 if (wb_acct & WBT_KSWAPD)
128 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
129 else if (wb_acct & WBT_DISCARD)
130 return &rwb->rq_wait[WBT_RWQ_DISCARD];
132 return &rwb->rq_wait[WBT_RWQ_BG];
135 static void rwb_wake_all(struct rq_wb *rwb)
139 for (i = 0; i < WBT_NUM_RWQ; i++) {
140 struct rq_wait *rqw = &rwb->rq_wait[i];
142 if (waitqueue_active(&rqw->wait))
143 wake_up_all(&rqw->wait);
147 void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct)
152 if (!(wb_acct & WBT_TRACKED))
155 rqw = get_rq_wait(rwb, wb_acct);
156 inflight = atomic_dec_return(&rqw->inflight);
159 * wbt got disabled with IO in flight. Wake up any potential
160 * waiters, we don't have to do more than that.
162 if (unlikely(!rwb_enabled(rwb))) {
168 * For discards, our limit is always the background. For writes, if
169 * the device does write back caching, drop further down before we
172 if (wb_acct & WBT_DISCARD)
173 limit = rwb->wb_background;
174 else if (rwb->wc && !wb_recent_wait(rwb))
177 limit = rwb->wb_normal;
180 * Don't wake anyone up if we are above the normal limit.
182 if (inflight && inflight >= limit)
185 if (waitqueue_active(&rqw->wait)) {
186 int diff = limit - inflight;
188 if (!inflight || diff >= rwb->wb_background / 2)
189 wake_up_all(&rqw->wait);
194 * Called on completion of a request. Note that it's also called when
195 * a request is merged, when the request gets freed.
197 void wbt_done(struct rq_wb *rwb, struct request *rq)
202 if (!wbt_is_tracked(rq)) {
203 if (rwb->sync_cookie == rq) {
205 rwb->sync_cookie = NULL;
209 wb_timestamp(rwb, &rwb->last_comp);
211 WARN_ON_ONCE(rq == rwb->sync_cookie);
212 __wbt_done(rwb, wbt_flags(rq));
218 * Return true, if we can't increase the depth further by scaling
220 static bool calc_wb_limits(struct rq_wb *rwb)
225 if (!rwb->min_lat_nsec) {
226 rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0;
231 * For QD=1 devices, this is a special case. It's important for those
232 * to have one request ready when one completes, so force a depth of
233 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
234 * since the device can't have more than that in flight. If we're
235 * scaling down, then keep a setting of 1/1/1.
237 if (rwb->queue_depth == 1) {
238 if (rwb->scale_step > 0)
239 rwb->wb_max = rwb->wb_normal = 1;
241 rwb->wb_max = rwb->wb_normal = 2;
244 rwb->wb_background = 1;
247 * scale_step == 0 is our default state. If we have suffered
248 * latency spikes, step will be > 0, and we shrink the
249 * allowed write depths. If step is < 0, we're only doing
250 * writes, and we allow a temporarily higher depth to
251 * increase performance.
253 depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth);
254 if (rwb->scale_step > 0)
255 depth = 1 + ((depth - 1) >> min(31, rwb->scale_step));
256 else if (rwb->scale_step < 0) {
257 unsigned int maxd = 3 * rwb->queue_depth / 4;
259 depth = 1 + ((depth - 1) << -rwb->scale_step);
267 * Set our max/normal/bg queue depths based on how far
268 * we have scaled down (->scale_step).
271 rwb->wb_normal = (rwb->wb_max + 1) / 2;
272 rwb->wb_background = (rwb->wb_max + 3) / 4;
278 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
281 * We need at least one read sample, and a minimum of
282 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
283 * that it's writes impacting us, and not just some sole read on
284 * a device that is in a lower power state.
286 return (stat[READ].nr_samples >= 1 &&
287 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
290 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
292 u64 now, issue = READ_ONCE(rwb->sync_issue);
294 if (!issue || !rwb->sync_cookie)
297 now = ktime_to_ns(ktime_get());
308 static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
310 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
314 * If our stored sync issue exceeds the window size, or it
315 * exceeds our min target AND we haven't logged any entries,
316 * flag the latency as exceeded. wbt works off completion latencies,
317 * but for a flooded device, a single sync IO can take a long time
318 * to complete after being issued. If this time exceeds our
319 * monitoring window AND we didn't see any other completions in that
320 * window, then count that sync IO as a violation of the latency.
322 thislat = rwb_sync_issue_lat(rwb);
323 if (thislat > rwb->cur_win_nsec ||
324 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
325 trace_wbt_lat(bdi, thislat);
330 * No read/write mix, if stat isn't valid
332 if (!stat_sample_valid(stat)) {
334 * If we had writes in this stat window and the window is
335 * current, we're only doing writes. If a task recently
336 * waited or still has writes in flights, consider us doing
337 * just writes as well.
339 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
341 return LAT_UNKNOWN_WRITES;
346 * If the 'min' latency exceeds our target, step down.
348 if (stat[READ].min > rwb->min_lat_nsec) {
349 trace_wbt_lat(bdi, stat[READ].min);
350 trace_wbt_stat(bdi, stat);
355 trace_wbt_stat(bdi, stat);
360 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
362 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
364 trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
365 rwb->wb_background, rwb->wb_normal, rwb->wb_max);
368 static void scale_up(struct rq_wb *rwb)
371 * Hit max in previous round, stop here
377 rwb->unknown_cnt = 0;
379 rwb->scaled_max = calc_wb_limits(rwb);
383 rwb_trace_step(rwb, "step up");
387 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
388 * had a latency violation.
390 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
393 * Stop scaling down when we've hit the limit. This also prevents
394 * ->scale_step from going to crazy values, if the device can't
397 if (rwb->wb_max == 1)
400 if (rwb->scale_step < 0 && hard_throttle)
405 rwb->scaled_max = false;
406 rwb->unknown_cnt = 0;
408 rwb_trace_step(rwb, "step down");
411 static void rwb_arm_timer(struct rq_wb *rwb)
413 if (rwb->scale_step > 0) {
415 * We should speed this up, using some variant of a fast
416 * integer inverse square root calculation. Since we only do
417 * this for every window expiration, it's not a huge deal,
420 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
421 int_sqrt((rwb->scale_step + 1) << 8));
424 * For step < 0, we don't want to increase/decrease the
427 rwb->cur_win_nsec = rwb->win_nsec;
430 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
433 static void wb_timer_fn(struct blk_stat_callback *cb)
435 struct rq_wb *rwb = cb->data;
436 unsigned int inflight = wbt_inflight(rwb);
439 status = latency_exceeded(rwb, cb->stat);
441 trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step,
445 * If we exceeded the latency target, step down. If we did not,
446 * step one level up. If we don't know enough to say either exceeded
447 * or ok, then don't do anything.
451 scale_down(rwb, true);
456 case LAT_UNKNOWN_WRITES:
458 * We started a the center step, but don't have a valid
459 * read/write sample, but we do have writes going on.
460 * Allow step to go negative, to increase write perf.
465 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
468 * We get here when previously scaled reduced depth, and we
469 * currently don't have a valid read/write sample. For that
470 * case, slowly return to center state (step == 0).
472 if (rwb->scale_step > 0)
474 else if (rwb->scale_step < 0)
475 scale_down(rwb, false);
482 * Re-arm timer, if we have IO in flight
484 if (rwb->scale_step || inflight)
488 void wbt_update_limits(struct rq_wb *rwb)
491 rwb->scaled_max = false;
497 static bool close_io(struct rq_wb *rwb)
499 const unsigned long now = jiffies;
501 return time_before(now, rwb->last_issue + HZ / 10) ||
502 time_before(now, rwb->last_comp + HZ / 10);
505 #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
507 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
511 if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD)
512 return rwb->wb_background;
515 * At this point we know it's a buffered write. If this is
516 * kswapd trying to free memory, or REQ_SYNC is set, then
517 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
518 * that. If the write is marked as a background write, then use
519 * the idle limit, or go to normal if we haven't had competing
522 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
524 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
526 * If less than 100ms since we completed unrelated IO,
527 * limit us to half the depth for background writeback.
529 limit = rwb->wb_background;
531 limit = rwb->wb_normal;
536 static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
537 wait_queue_entry_t *wait, unsigned long rw)
540 * inc it here even if disabled, since we'll dec it at completion.
541 * this only happens if the task was sleeping in __wbt_wait(),
542 * and someone turned it off at the same time.
544 if (!rwb_enabled(rwb)) {
545 atomic_inc(&rqw->inflight);
550 * If the waitqueue is already active and we are not the next
551 * in line to be woken up, wait for our turn.
553 if (waitqueue_active(&rqw->wait) &&
554 rqw->wait.head.next != &wait->entry)
557 return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
561 * Block if we will exceed our limit, or if we are currently waiting for
562 * the timer to kick off queuing again.
564 static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
565 unsigned long rw, spinlock_t *lock)
569 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
572 if (may_queue(rwb, rqw, &wait, rw))
576 prepare_to_wait_exclusive(&rqw->wait, &wait,
577 TASK_UNINTERRUPTIBLE);
579 if (may_queue(rwb, rqw, &wait, rw))
583 spin_unlock_irq(lock);
590 finish_wait(&rqw->wait, &wait);
593 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
595 switch (bio_op(bio)) {
598 * Don't throttle WRITE_ODIRECT
600 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
601 (REQ_SYNC | REQ_IDLE))
612 * Returns true if the IO request should be accounted, false if not.
613 * May sleep, if we have exceeded the writeback limits. Caller can pass
614 * in an irq held spinlock, if it holds one when calling this function.
615 * If we do sleep, we'll release and re-grab it.
617 enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
619 enum wbt_flags ret = 0;
621 if (!rwb_enabled(rwb))
624 if (bio_op(bio) == REQ_OP_READ)
627 if (!wbt_should_throttle(rwb, bio)) {
629 wb_timestamp(rwb, &rwb->last_issue);
633 if (current_is_kswapd())
635 if (bio_op(bio) == REQ_OP_DISCARD)
638 __wbt_wait(rwb, ret, bio->bi_opf, lock);
640 if (!blk_stat_is_active(rwb->cb))
643 return ret | WBT_TRACKED;
646 void wbt_issue(struct rq_wb *rwb, struct request *rq)
648 if (!rwb_enabled(rwb))
652 * Track sync issue, in case it takes a long time to complete. Allows us
653 * to react quicker, if a sync IO takes a long time to complete. Note
654 * that this is just a hint. The request can go away when it completes,
655 * so it's important we never dereference it. We only use the address to
656 * compare with, which is why we store the sync_issue time locally.
658 if (wbt_is_read(rq) && !rwb->sync_issue) {
659 rwb->sync_cookie = rq;
660 rwb->sync_issue = rq->io_start_time_ns;
664 void wbt_requeue(struct rq_wb *rwb, struct request *rq)
666 if (!rwb_enabled(rwb))
668 if (rq == rwb->sync_cookie) {
670 rwb->sync_cookie = NULL;
674 void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
677 rwb->queue_depth = depth;
678 wbt_update_limits(rwb);
682 void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
685 rwb->wc = write_cache_on;
689 * Disable wbt, if enabled by default.
691 void wbt_disable_default(struct request_queue *q)
693 struct rq_wb *rwb = q->rq_wb;
695 if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT)
698 EXPORT_SYMBOL_GPL(wbt_disable_default);
701 * Enable wbt if defaults are configured that way
703 void wbt_enable_default(struct request_queue *q)
705 /* Throttling already enabled? */
709 /* Queue not registered? Maybe shutting down... */
710 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
713 if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) ||
714 (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ)))
717 EXPORT_SYMBOL_GPL(wbt_enable_default);
719 u64 wbt_default_latency_nsec(struct request_queue *q)
722 * We default to 2msec for non-rotational storage, and 75msec
723 * for rotational storage.
725 if (blk_queue_nonrot(q))
731 static int wbt_data_dir(const struct request *rq)
733 const int op = req_op(rq);
735 if (op == REQ_OP_READ)
737 else if (op_is_write(op))
744 int wbt_init(struct request_queue *q)
749 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
753 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
759 for (i = 0; i < WBT_NUM_RWQ; i++) {
760 atomic_set(&rwb->rq_wait[i].inflight, 0);
761 init_waitqueue_head(&rwb->rq_wait[i].wait);
764 rwb->last_comp = rwb->last_issue = jiffies;
766 rwb->win_nsec = RWB_WINDOW_NSEC;
767 rwb->enable_state = WBT_STATE_ON_DEFAULT;
768 wbt_update_limits(rwb);
771 * Assign rwb and add the stats callback.
774 blk_stat_add_callback(q, rwb->cb);
776 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
778 wbt_set_queue_depth(rwb, blk_queue_depth(q));
779 wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
784 void wbt_exit(struct request_queue *q)
786 struct rq_wb *rwb = q->rq_wb;
789 blk_stat_remove_callback(q, rwb->cb);
790 blk_stat_free_callback(rwb->cb);