]>
git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - block/blk-wbt.c
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
,
105 enum wbt_flags wb_acct
)
107 if (wb_acct
& WBT_KSWAPD
)
108 return &rwb
->rq_wait
[WBT_RWQ_KSWAPD
];
110 return &rwb
->rq_wait
[WBT_RWQ_BG
];
113 static void rwb_wake_all(struct rq_wb
*rwb
)
117 for (i
= 0; i
< WBT_NUM_RWQ
; i
++) {
118 struct rq_wait
*rqw
= &rwb
->rq_wait
[i
];
120 if (wq_has_sleeper(&rqw
->wait
))
121 wake_up_all(&rqw
->wait
);
125 static void wbt_rqw_done(struct rq_wb
*rwb
, struct rq_wait
*rqw
,
126 enum wbt_flags wb_acct
)
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 (wq_has_sleeper(&rqw
->wait
)) {
157 int diff
= limit
- inflight
;
159 if (!inflight
|| diff
>= rwb
->wb_background
/ 2)
160 wake_up_all(&rqw
->wait
);
164 void __wbt_done(struct rq_wb
*rwb
, enum wbt_flags wb_acct
)
168 if (!(wb_acct
& WBT_TRACKED
))
171 rqw
= get_rq_wait(rwb
, wb_acct
);
172 wbt_rqw_done(rwb
, rqw
, wb_acct
);
176 * Called on completion of a request. Note that it's also called when
177 * a request is merged, when the request gets freed.
179 void wbt_done(struct rq_wb
*rwb
, struct blk_issue_stat
*stat
)
184 if (!wbt_is_tracked(stat
)) {
185 if (rwb
->sync_cookie
== stat
) {
187 rwb
->sync_cookie
= NULL
;
190 if (wbt_is_read(stat
))
191 wb_timestamp(rwb
, &rwb
->last_comp
);
193 WARN_ON_ONCE(stat
== rwb
->sync_cookie
);
194 __wbt_done(rwb
, wbt_stat_to_mask(stat
));
196 wbt_clear_state(stat
);
200 * Return true, if we can't increase the depth further by scaling
202 static bool calc_wb_limits(struct rq_wb
*rwb
)
207 if (!rwb
->min_lat_nsec
) {
208 rwb
->wb_max
= rwb
->wb_normal
= rwb
->wb_background
= 0;
213 * For QD=1 devices, this is a special case. It's important for those
214 * to have one request ready when one completes, so force a depth of
215 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
216 * since the device can't have more than that in flight. If we're
217 * scaling down, then keep a setting of 1/1/1.
219 if (rwb
->queue_depth
== 1) {
220 if (rwb
->scale_step
> 0)
221 rwb
->wb_max
= rwb
->wb_normal
= 1;
223 rwb
->wb_max
= rwb
->wb_normal
= 2;
226 rwb
->wb_background
= 1;
229 * scale_step == 0 is our default state. If we have suffered
230 * latency spikes, step will be > 0, and we shrink the
231 * allowed write depths. If step is < 0, we're only doing
232 * writes, and we allow a temporarily higher depth to
233 * increase performance.
235 depth
= min_t(unsigned int, RWB_DEF_DEPTH
, rwb
->queue_depth
);
236 if (rwb
->scale_step
> 0)
237 depth
= 1 + ((depth
- 1) >> min(31, rwb
->scale_step
));
238 else if (rwb
->scale_step
< 0) {
239 unsigned int maxd
= 3 * rwb
->queue_depth
/ 4;
241 depth
= 1 + ((depth
- 1) << -rwb
->scale_step
);
249 * Set our max/normal/bg queue depths based on how far
250 * we have scaled down (->scale_step).
253 rwb
->wb_normal
= (rwb
->wb_max
+ 1) / 2;
254 rwb
->wb_background
= (rwb
->wb_max
+ 3) / 4;
260 static inline bool stat_sample_valid(struct blk_rq_stat
*stat
)
263 * We need at least one read sample, and a minimum of
264 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
265 * that it's writes impacting us, and not just some sole read on
266 * a device that is in a lower power state.
268 return (stat
[READ
].nr_samples
>= 1 &&
269 stat
[WRITE
].nr_samples
>= RWB_MIN_WRITE_SAMPLES
);
272 static u64
rwb_sync_issue_lat(struct rq_wb
*rwb
)
274 u64 now
, issue
= READ_ONCE(rwb
->sync_issue
);
276 if (!issue
|| !rwb
->sync_cookie
)
279 now
= ktime_to_ns(ktime_get());
290 static int latency_exceeded(struct rq_wb
*rwb
, struct blk_rq_stat
*stat
)
292 struct backing_dev_info
*bdi
= rwb
->queue
->backing_dev_info
;
296 * If our stored sync issue exceeds the window size, or it
297 * exceeds our min target AND we haven't logged any entries,
298 * flag the latency as exceeded. wbt works off completion latencies,
299 * but for a flooded device, a single sync IO can take a long time
300 * to complete after being issued. If this time exceeds our
301 * monitoring window AND we didn't see any other completions in that
302 * window, then count that sync IO as a violation of the latency.
304 thislat
= rwb_sync_issue_lat(rwb
);
305 if (thislat
> rwb
->cur_win_nsec
||
306 (thislat
> rwb
->min_lat_nsec
&& !stat
[READ
].nr_samples
)) {
307 trace_wbt_lat(bdi
, thislat
);
312 * No read/write mix, if stat isn't valid
314 if (!stat_sample_valid(stat
)) {
316 * If we had writes in this stat window and the window is
317 * current, we're only doing writes. If a task recently
318 * waited or still has writes in flights, consider us doing
319 * just writes as well.
321 if (stat
[WRITE
].nr_samples
|| wb_recent_wait(rwb
) ||
323 return LAT_UNKNOWN_WRITES
;
328 * If the 'min' latency exceeds our target, step down.
330 if (stat
[READ
].min
> rwb
->min_lat_nsec
) {
331 trace_wbt_lat(bdi
, stat
[READ
].min
);
332 trace_wbt_stat(bdi
, stat
);
337 trace_wbt_stat(bdi
, stat
);
342 static void rwb_trace_step(struct rq_wb
*rwb
, const char *msg
)
344 struct backing_dev_info
*bdi
= rwb
->queue
->backing_dev_info
;
346 trace_wbt_step(bdi
, msg
, rwb
->scale_step
, rwb
->cur_win_nsec
,
347 rwb
->wb_background
, rwb
->wb_normal
, rwb
->wb_max
);
350 static void scale_up(struct rq_wb
*rwb
)
353 * Hit max in previous round, stop here
359 rwb
->unknown_cnt
= 0;
361 rwb
->scaled_max
= calc_wb_limits(rwb
);
365 rwb_trace_step(rwb
, "step up");
369 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
370 * had a latency violation.
372 static void scale_down(struct rq_wb
*rwb
, bool hard_throttle
)
375 * Stop scaling down when we've hit the limit. This also prevents
376 * ->scale_step from going to crazy values, if the device can't
379 if (rwb
->wb_max
== 1)
382 if (rwb
->scale_step
< 0 && hard_throttle
)
387 rwb
->scaled_max
= false;
388 rwb
->unknown_cnt
= 0;
390 rwb_trace_step(rwb
, "step down");
393 static void rwb_arm_timer(struct rq_wb
*rwb
)
395 if (rwb
->scale_step
> 0) {
397 * We should speed this up, using some variant of a fast
398 * integer inverse square root calculation. Since we only do
399 * this for every window expiration, it's not a huge deal,
402 rwb
->cur_win_nsec
= div_u64(rwb
->win_nsec
<< 4,
403 int_sqrt((rwb
->scale_step
+ 1) << 8));
406 * For step < 0, we don't want to increase/decrease the
409 rwb
->cur_win_nsec
= rwb
->win_nsec
;
412 blk_stat_activate_nsecs(rwb
->cb
, rwb
->cur_win_nsec
);
415 static void wb_timer_fn(struct blk_stat_callback
*cb
)
417 struct rq_wb
*rwb
= cb
->data
;
418 unsigned int inflight
= wbt_inflight(rwb
);
421 status
= latency_exceeded(rwb
, cb
->stat
);
423 trace_wbt_timer(rwb
->queue
->backing_dev_info
, status
, rwb
->scale_step
,
427 * If we exceeded the latency target, step down. If we did not,
428 * step one level up. If we don't know enough to say either exceeded
429 * or ok, then don't do anything.
433 scale_down(rwb
, true);
438 case LAT_UNKNOWN_WRITES
:
440 * We started a the center step, but don't have a valid
441 * read/write sample, but we do have writes going on.
442 * Allow step to go negative, to increase write perf.
447 if (++rwb
->unknown_cnt
< RWB_UNKNOWN_BUMP
)
450 * We get here when previously scaled reduced depth, and we
451 * currently don't have a valid read/write sample. For that
452 * case, slowly return to center state (step == 0).
454 if (rwb
->scale_step
> 0)
456 else if (rwb
->scale_step
< 0)
457 scale_down(rwb
, false);
464 * Re-arm timer, if we have IO in flight
466 if (rwb
->scale_step
|| inflight
)
470 void wbt_update_limits(struct rq_wb
*rwb
)
473 rwb
->scaled_max
= false;
479 static bool close_io(struct rq_wb
*rwb
)
481 const unsigned long now
= jiffies
;
483 return time_before(now
, rwb
->last_issue
+ HZ
/ 10) ||
484 time_before(now
, rwb
->last_comp
+ HZ
/ 10);
487 #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
489 static inline unsigned int get_limit(struct rq_wb
*rwb
, unsigned long rw
)
494 * If we got disabled, just return UINT_MAX. This ensures that
495 * we'll properly inc a new IO, and dec+wakeup at the end.
497 if (!rwb_enabled(rwb
))
501 * At this point we know it's a buffered write. If this is
502 * kswapd trying to free memory, or REQ_SYNC is set, then
503 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
504 * that. If the write is marked as a background write, then use
505 * the idle limit, or go to normal if we haven't had competing
508 if ((rw
& REQ_HIPRIO
) || wb_recent_wait(rwb
) || current_is_kswapd())
510 else if ((rw
& REQ_BACKGROUND
) || close_io(rwb
)) {
512 * If less than 100ms since we completed unrelated IO,
513 * limit us to half the depth for background writeback.
515 limit
= rwb
->wb_background
;
517 limit
= rwb
->wb_normal
;
522 struct wbt_wait_data
{
523 struct wait_queue_entry wq
;
524 struct task_struct
*task
;
531 static int wbt_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
532 int wake_flags
, void *key
)
534 struct wbt_wait_data
*data
= container_of(curr
, struct wbt_wait_data
,
538 * If we fail to get a budget, return -1 to interrupt the wake up
539 * loop in __wake_up_common.
541 if (!atomic_inc_below(&data
->rqw
->inflight
, get_limit(data
->rwb
, data
->rw
)))
544 data
->got_token
= true;
545 list_del_init(&curr
->entry
);
546 wake_up_process(data
->task
);
551 * Block if we will exceed our limit, or if we are currently waiting for
552 * the timer to kick off queuing again.
554 static void __wbt_wait(struct rq_wb
*rwb
, enum wbt_flags wb_acct
,
555 unsigned long rw
, spinlock_t
*lock
)
559 struct rq_wait
*rqw
= get_rq_wait(rwb
, wb_acct
);
560 struct wbt_wait_data data
= {
562 .func
= wbt_wake_function
,
563 .entry
= LIST_HEAD_INIT(data
.wq
.entry
),
572 has_sleeper
= wq_has_sleeper(&rqw
->wait
);
573 if (!has_sleeper
&& atomic_inc_below(&rqw
->inflight
, get_limit(rwb
, rw
)))
576 prepare_to_wait_exclusive(&rqw
->wait
, &data
.wq
, TASK_UNINTERRUPTIBLE
);
582 atomic_inc_below(&rqw
->inflight
, get_limit(rwb
, rw
))) {
583 finish_wait(&rqw
->wait
, &data
.wq
);
586 * We raced with wbt_wake_function() getting a token,
587 * which means we now have two. Put our local token
588 * and wake anyone else potentially waiting for one.
591 wbt_rqw_done(rwb
, rqw
, wb_acct
);
596 spin_unlock_irq(lock
);
605 finish_wait(&rqw
->wait
, &data
.wq
);
608 static inline bool wbt_should_throttle(struct rq_wb
*rwb
, struct bio
*bio
)
610 const int op
= bio_op(bio
);
613 * If not a WRITE, do nothing
615 if (op
!= REQ_OP_WRITE
)
619 * Don't throttle WRITE_ODIRECT
621 if ((bio
->bi_opf
& (REQ_SYNC
| REQ_IDLE
)) == (REQ_SYNC
| REQ_IDLE
))
628 * Returns true if the IO request should be accounted, false if not.
629 * May sleep, if we have exceeded the writeback limits. Caller can pass
630 * in an irq held spinlock, if it holds one when calling this function.
631 * If we do sleep, we'll release and re-grab it.
633 enum wbt_flags
wbt_wait(struct rq_wb
*rwb
, struct bio
*bio
, spinlock_t
*lock
)
635 enum wbt_flags ret
= 0;
637 if (!rwb_enabled(rwb
))
640 if (bio_op(bio
) == REQ_OP_READ
)
643 if (!wbt_should_throttle(rwb
, bio
)) {
645 wb_timestamp(rwb
, &rwb
->last_issue
);
649 if (current_is_kswapd())
652 __wbt_wait(rwb
, ret
, bio
->bi_opf
, lock
);
654 if (!blk_stat_is_active(rwb
->cb
))
657 return ret
| WBT_TRACKED
;
660 void wbt_issue(struct rq_wb
*rwb
, struct blk_issue_stat
*stat
)
662 if (!rwb_enabled(rwb
))
666 * Track sync issue, in case it takes a long time to complete. Allows
667 * us to react quicker, if a sync IO takes a long time to complete.
668 * Note that this is just a hint. 'stat' can go away when the
669 * request completes, so it's important we never dereference it. We
670 * only use the address to compare with, which is why we store the
671 * sync_issue time locally.
673 if (wbt_is_read(stat
) && !rwb
->sync_issue
) {
674 rwb
->sync_cookie
= stat
;
675 rwb
->sync_issue
= blk_stat_time(stat
);
679 void wbt_requeue(struct rq_wb
*rwb
, struct blk_issue_stat
*stat
)
681 if (!rwb_enabled(rwb
))
683 if (stat
== rwb
->sync_cookie
) {
685 rwb
->sync_cookie
= NULL
;
689 void wbt_set_queue_depth(struct rq_wb
*rwb
, unsigned int depth
)
692 rwb
->queue_depth
= depth
;
693 wbt_update_limits(rwb
);
697 void wbt_set_write_cache(struct rq_wb
*rwb
, bool write_cache_on
)
700 rwb
->wc
= write_cache_on
;
704 * Disable wbt, if enabled by default.
706 void wbt_disable_default(struct request_queue
*q
)
708 struct rq_wb
*rwb
= q
->rq_wb
;
710 if (rwb
&& rwb
->enable_state
== WBT_STATE_ON_DEFAULT
) {
711 blk_stat_deactivate(rwb
->cb
);
715 EXPORT_SYMBOL_GPL(wbt_disable_default
);
718 * Enable wbt if defaults are configured that way
720 void wbt_enable_default(struct request_queue
*q
)
722 /* Throttling already enabled? */
726 /* Queue not registered? Maybe shutting down... */
727 if (!test_bit(QUEUE_FLAG_REGISTERED
, &q
->queue_flags
))
730 if ((q
->mq_ops
&& IS_ENABLED(CONFIG_BLK_WBT_MQ
)) ||
731 (q
->request_fn
&& IS_ENABLED(CONFIG_BLK_WBT_SQ
)))
734 EXPORT_SYMBOL_GPL(wbt_enable_default
);
736 u64
wbt_default_latency_nsec(struct request_queue
*q
)
739 * We default to 2msec for non-rotational storage, and 75msec
740 * for rotational storage.
742 if (blk_queue_nonrot(q
))
748 static int wbt_data_dir(const struct request
*rq
)
750 const int op
= req_op(rq
);
752 if (op
== REQ_OP_READ
)
754 else if (op
== REQ_OP_WRITE
|| op
== REQ_OP_FLUSH
)
761 int wbt_init(struct request_queue
*q
)
766 BUILD_BUG_ON(WBT_NR_BITS
> BLK_STAT_RES_BITS
);
768 rwb
= kzalloc(sizeof(*rwb
), GFP_KERNEL
);
772 rwb
->cb
= blk_stat_alloc_callback(wb_timer_fn
, wbt_data_dir
, 2, rwb
);
778 for (i
= 0; i
< WBT_NUM_RWQ
; i
++) {
779 atomic_set(&rwb
->rq_wait
[i
].inflight
, 0);
780 init_waitqueue_head(&rwb
->rq_wait
[i
].wait
);
783 rwb
->last_comp
= rwb
->last_issue
= jiffies
;
785 rwb
->win_nsec
= RWB_WINDOW_NSEC
;
786 rwb
->enable_state
= WBT_STATE_ON_DEFAULT
;
787 wbt_update_limits(rwb
);
790 * Assign rwb and add the stats callback.
793 blk_stat_add_callback(q
, rwb
->cb
);
795 rwb
->min_lat_nsec
= wbt_default_latency_nsec(q
);
797 wbt_set_queue_depth(rwb
, blk_queue_depth(q
));
798 wbt_set_write_cache(rwb
, test_bit(QUEUE_FLAG_WC
, &q
->queue_flags
));
803 void wbt_exit(struct request_queue
*q
)
805 struct rq_wb
*rwb
= q
->rq_wb
;
808 blk_stat_remove_callback(q
, rwb
->cb
);
809 blk_stat_free_callback(rwb
->cb
);