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>
34 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
35 * from here depending on device stats
42 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
45 * Disregard stats, if we don't meet this minimum
47 RWB_MIN_WRITE_SAMPLES = 3,
50 * If we have this number of consecutive windows with not enough
51 * information to scale up or down, scale up.
56 static inline bool rwb_enabled(struct rq_wb *rwb)
58 return rwb && rwb->wb_normal != 0;
62 * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded,
63 * false if 'v' + 1 would be bigger than 'below'.
65 static bool atomic_inc_below(atomic_t *v, int below)
67 int cur = atomic_read(v);
74 old = atomic_cmpxchg(v, cur, cur + 1);
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->queue->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, bool is_kswapd)
106 return &rwb->rq_wait[is_kswapd];
109 static void rwb_wake_all(struct rq_wb *rwb)
113 for (i = 0; i < WBT_NUM_RWQ; i++) {
114 struct rq_wait *rqw = &rwb->rq_wait[i];
116 if (waitqueue_active(&rqw->wait))
117 wake_up_all(&rqw->wait);
121 void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct)
126 if (!(wb_acct & WBT_TRACKED))
129 rqw = get_rq_wait(rwb, wb_acct & WBT_KSWAPD);
130 inflight = atomic_dec_return(&rqw->inflight);
133 * wbt got disabled with IO in flight. Wake up any potential
134 * waiters, we don't have to do more than that.
136 if (unlikely(!rwb_enabled(rwb))) {
142 * If the device does write back caching, drop further down
143 * before we wake people up.
145 if (rwb->wc && !wb_recent_wait(rwb))
148 limit = rwb->wb_normal;
151 * Don't wake anyone up if we are above the normal limit.
153 if (inflight && inflight >= limit)
156 if (waitqueue_active(&rqw->wait)) {
157 int diff = limit - inflight;
159 if (!inflight || diff >= rwb->wb_background / 2)
160 wake_up_all(&rqw->wait);
165 * Called on completion of a request. Note that it's also called when
166 * a request is merged, when the request gets freed.
168 void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat)
173 if (!wbt_is_tracked(stat)) {
174 if (rwb->sync_cookie == stat) {
176 rwb->sync_cookie = NULL;
179 if (wbt_is_read(stat))
180 wb_timestamp(rwb, &rwb->last_comp);
181 wbt_clear_state(stat);
183 WARN_ON_ONCE(stat == rwb->sync_cookie);
184 __wbt_done(rwb, wbt_stat_to_mask(stat));
185 wbt_clear_state(stat);
190 * Return true, if we can't increase the depth further by scaling
192 static bool calc_wb_limits(struct rq_wb *rwb)
197 if (!rwb->min_lat_nsec) {
198 rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0;
203 * For QD=1 devices, this is a special case. It's important for those
204 * to have one request ready when one completes, so force a depth of
205 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
206 * since the device can't have more than that in flight. If we're
207 * scaling down, then keep a setting of 1/1/1.
209 if (rwb->queue_depth == 1) {
210 if (rwb->scale_step > 0)
211 rwb->wb_max = rwb->wb_normal = 1;
213 rwb->wb_max = rwb->wb_normal = 2;
216 rwb->wb_background = 1;
219 * scale_step == 0 is our default state. If we have suffered
220 * latency spikes, step will be > 0, and we shrink the
221 * allowed write depths. If step is < 0, we're only doing
222 * writes, and we allow a temporarily higher depth to
223 * increase performance.
225 depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth);
226 if (rwb->scale_step > 0)
227 depth = 1 + ((depth - 1) >> min(31, rwb->scale_step));
228 else if (rwb->scale_step < 0) {
229 unsigned int maxd = 3 * rwb->queue_depth / 4;
231 depth = 1 + ((depth - 1) << -rwb->scale_step);
239 * Set our max/normal/bg queue depths based on how far
240 * we have scaled down (->scale_step).
243 rwb->wb_normal = (rwb->wb_max + 1) / 2;
244 rwb->wb_background = (rwb->wb_max + 3) / 4;
250 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
253 * We need at least one read sample, and a minimum of
254 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
255 * that it's writes impacting us, and not just some sole read on
256 * a device that is in a lower power state.
258 return stat[BLK_STAT_READ].nr_samples >= 1 &&
259 stat[BLK_STAT_WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES;
262 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
264 u64 now, issue = ACCESS_ONCE(rwb->sync_issue);
266 if (!issue || !rwb->sync_cookie)
269 now = ktime_to_ns(ktime_get());
280 static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
282 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
286 * If our stored sync issue exceeds the window size, or it
287 * exceeds our min target AND we haven't logged any entries,
288 * flag the latency as exceeded. wbt works off completion latencies,
289 * but for a flooded device, a single sync IO can take a long time
290 * to complete after being issued. If this time exceeds our
291 * monitoring window AND we didn't see any other completions in that
292 * window, then count that sync IO as a violation of the latency.
294 thislat = rwb_sync_issue_lat(rwb);
295 if (thislat > rwb->cur_win_nsec ||
296 (thislat > rwb->min_lat_nsec && !stat[BLK_STAT_READ].nr_samples)) {
297 trace_wbt_lat(bdi, thislat);
302 * No read/write mix, if stat isn't valid
304 if (!stat_sample_valid(stat)) {
306 * If we had writes in this stat window and the window is
307 * current, we're only doing writes. If a task recently
308 * waited or still has writes in flights, consider us doing
309 * just writes as well.
311 if ((stat[BLK_STAT_WRITE].nr_samples && blk_stat_is_current(stat)) ||
312 wb_recent_wait(rwb) || wbt_inflight(rwb))
313 return LAT_UNKNOWN_WRITES;
318 * If the 'min' latency exceeds our target, step down.
320 if (stat[BLK_STAT_READ].min > rwb->min_lat_nsec) {
321 trace_wbt_lat(bdi, stat[BLK_STAT_READ].min);
322 trace_wbt_stat(bdi, stat);
327 trace_wbt_stat(bdi, stat);
332 static int latency_exceeded(struct rq_wb *rwb)
334 struct blk_rq_stat stat[2];
336 blk_queue_stat_get(rwb->queue, stat);
337 return __latency_exceeded(rwb, stat);
340 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
342 struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
344 trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
345 rwb->wb_background, rwb->wb_normal, rwb->wb_max);
348 static void scale_up(struct rq_wb *rwb)
351 * Hit max in previous round, stop here
357 rwb->unknown_cnt = 0;
358 blk_stat_clear(rwb->queue);
360 rwb->scaled_max = calc_wb_limits(rwb);
364 rwb_trace_step(rwb, "step up");
368 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
369 * had a latency violation.
371 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
374 * Stop scaling down when we've hit the limit. This also prevents
375 * ->scale_step from going to crazy values, if the device can't
378 if (rwb->wb_max == 1)
381 if (rwb->scale_step < 0 && hard_throttle)
386 rwb->scaled_max = false;
387 rwb->unknown_cnt = 0;
388 blk_stat_clear(rwb->queue);
390 rwb_trace_step(rwb, "step down");
393 static void rwb_arm_timer(struct rq_wb *rwb)
395 unsigned long expires;
397 if (rwb->scale_step > 0) {
399 * We should speed this up, using some variant of a fast
400 * integer inverse square root calculation. Since we only do
401 * this for every window expiration, it's not a huge deal,
404 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
405 int_sqrt((rwb->scale_step + 1) << 8));
408 * For step < 0, we don't want to increase/decrease the
411 rwb->cur_win_nsec = rwb->win_nsec;
414 expires = jiffies + nsecs_to_jiffies(rwb->cur_win_nsec);
415 mod_timer(&rwb->window_timer, expires);
418 static void wb_timer_fn(unsigned long data)
420 struct rq_wb *rwb = (struct rq_wb *) data;
421 unsigned int inflight = wbt_inflight(rwb);
424 status = latency_exceeded(rwb);
426 trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step,
430 * If we exceeded the latency target, step down. If we did not,
431 * step one level up. If we don't know enough to say either exceeded
432 * or ok, then don't do anything.
436 scale_down(rwb, true);
441 case LAT_UNKNOWN_WRITES:
443 * We started a the center step, but don't have a valid
444 * read/write sample, but we do have writes going on.
445 * Allow step to go negative, to increase write perf.
450 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
453 * We get here when previously scaled reduced depth, and we
454 * currently don't have a valid read/write sample. For that
455 * case, slowly return to center state (step == 0).
457 if (rwb->scale_step > 0)
459 else if (rwb->scale_step < 0)
460 scale_down(rwb, false);
467 * Re-arm timer, if we have IO in flight
469 if (rwb->scale_step || inflight)
473 void wbt_update_limits(struct rq_wb *rwb)
476 rwb->scaled_max = false;
482 static bool close_io(struct rq_wb *rwb)
484 const unsigned long now = jiffies;
486 return time_before(now, rwb->last_issue + HZ / 10) ||
487 time_before(now, rwb->last_comp + HZ / 10);
490 #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
492 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
497 * At this point we know it's a buffered write. If this is
498 * kswapd trying to free memory, or REQ_SYNC is set, set, then
499 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
500 * that. If the write is marked as a background write, then use
501 * the idle limit, or go to normal if we haven't had competing
504 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
506 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
508 * If less than 100ms since we completed unrelated IO,
509 * limit us to half the depth for background writeback.
511 limit = rwb->wb_background;
513 limit = rwb->wb_normal;
518 static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
519 wait_queue_t *wait, unsigned long rw)
522 * inc it here even if disabled, since we'll dec it at completion.
523 * this only happens if the task was sleeping in __wbt_wait(),
524 * and someone turned it off at the same time.
526 if (!rwb_enabled(rwb)) {
527 atomic_inc(&rqw->inflight);
532 * If the waitqueue is already active and we are not the next
533 * in line to be woken up, wait for our turn.
535 if (waitqueue_active(&rqw->wait) &&
536 rqw->wait.task_list.next != &wait->task_list)
539 return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
543 * Block if we will exceed our limit, or if we are currently waiting for
544 * the timer to kick off queuing again.
546 static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock)
550 struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd());
553 if (may_queue(rwb, rqw, &wait, rw))
557 prepare_to_wait_exclusive(&rqw->wait, &wait,
558 TASK_UNINTERRUPTIBLE);
560 if (may_queue(rwb, rqw, &wait, rw))
564 spin_unlock_irq(lock);
571 finish_wait(&rqw->wait, &wait);
574 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
576 const int op = bio_op(bio);
579 * If not a WRITE, do nothing
581 if (op != REQ_OP_WRITE)
585 * Don't throttle WRITE_ODIRECT
587 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE))
594 * Returns true if the IO request should be accounted, false if not.
595 * May sleep, if we have exceeded the writeback limits. Caller can pass
596 * in an irq held spinlock, if it holds one when calling this function.
597 * If we do sleep, we'll release and re-grab it.
599 enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
601 unsigned int ret = 0;
603 if (!rwb_enabled(rwb))
606 if (bio_op(bio) == REQ_OP_READ)
609 if (!wbt_should_throttle(rwb, bio)) {
611 wb_timestamp(rwb, &rwb->last_issue);
615 __wbt_wait(rwb, bio->bi_opf, lock);
617 if (!timer_pending(&rwb->window_timer))
620 if (current_is_kswapd())
623 return ret | WBT_TRACKED;
626 void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat)
628 if (!rwb_enabled(rwb))
632 * Track sync issue, in case it takes a long time to complete. Allows
633 * us to react quicker, if a sync IO takes a long time to complete.
634 * Note that this is just a hint. 'stat' can go away when the
635 * request completes, so it's important we never dereference it. We
636 * only use the address to compare with, which is why we store the
637 * sync_issue time locally.
639 if (wbt_is_read(stat) && !rwb->sync_issue) {
640 rwb->sync_cookie = stat;
641 rwb->sync_issue = blk_stat_time(stat);
645 void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat)
647 if (!rwb_enabled(rwb))
649 if (stat == rwb->sync_cookie) {
651 rwb->sync_cookie = NULL;
655 void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
658 rwb->queue_depth = depth;
659 wbt_update_limits(rwb);
663 void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
666 rwb->wc = write_cache_on;
670 * Disable wbt, if enabled by default. Only called from CFQ, if we have
673 void wbt_disable_default(struct request_queue *q)
675 struct rq_wb *rwb = q->rq_wb;
677 if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) {
678 del_timer_sync(&rwb->window_timer);
679 rwb->win_nsec = rwb->min_lat_nsec = 0;
680 wbt_update_limits(rwb);
683 EXPORT_SYMBOL_GPL(wbt_disable_default);
685 u64 wbt_default_latency_nsec(struct request_queue *q)
688 * We default to 2msec for non-rotational storage, and 75msec
689 * for rotational storage.
691 if (blk_queue_nonrot(q))
697 int wbt_init(struct request_queue *q)
703 * For now, we depend on the stats window being larger than
704 * our monitoring window. Ensure that this isn't inadvertently
707 BUILD_BUG_ON(RWB_WINDOW_NSEC > BLK_STAT_NSEC);
708 BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS);
710 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
714 for (i = 0; i < WBT_NUM_RWQ; i++) {
715 atomic_set(&rwb->rq_wait[i].inflight, 0);
716 init_waitqueue_head(&rwb->rq_wait[i].wait);
719 setup_timer(&rwb->window_timer, wb_timer_fn, (unsigned long) rwb);
721 rwb->queue_depth = RWB_DEF_DEPTH;
722 rwb->last_comp = rwb->last_issue = jiffies;
724 rwb->win_nsec = RWB_WINDOW_NSEC;
725 rwb->enable_state = WBT_STATE_ON_DEFAULT;
726 wbt_update_limits(rwb);
729 * Assign rwb, and turn on stats tracking for this queue
734 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
736 wbt_set_queue_depth(rwb, blk_queue_depth(q));
737 wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
742 void wbt_exit(struct request_queue *q)
744 struct rq_wb *rwb = q->rq_wb;
747 del_timer_sync(&rwb->window_timer);