blob: 68a774d7a7c9c0f84b1e2129353975703fc62a5a [file] [log] [blame]
Christoph Hellwig3dcf60bc2019-04-30 14:42:43 -04001// SPDX-License-Identifier: GPL-2.0
Jens Axboee34cbd32016-11-09 12:36:15 -07002/*
3 * buffered writeback throttling. loosely based on CoDel. We can't drop
4 * packets for IO scheduling, so the logic is something like this:
5 *
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.
18 *
19 * Copyright (C) 2016 Jens Axboe
20 *
21 */
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>
27
28#include "blk-wbt.h"
Josef Bacika7905042018-07-03 09:32:35 -060029#include "blk-rq-qos.h"
Yu Kuai671fae52022-10-19 20:15:18 +080030#include "elevator.h"
Jens Axboee34cbd32016-11-09 12:36:15 -070031
32#define CREATE_TRACE_POINTS
33#include <trace/events/wbt.h>
34
Omar Sandovala8a45942018-05-09 02:08:48 -070035static inline void wbt_clear_state(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070036{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070037 rq->wbt_flags = 0;
Omar Sandoval934031a2018-05-09 02:08:47 -070038}
39
Omar Sandovala8a45942018-05-09 02:08:48 -070040static inline enum wbt_flags wbt_flags(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070041{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070042 return rq->wbt_flags;
Omar Sandoval934031a2018-05-09 02:08:47 -070043}
44
Omar Sandovala8a45942018-05-09 02:08:48 -070045static inline bool wbt_is_tracked(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070046{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070047 return rq->wbt_flags & WBT_TRACKED;
Omar Sandoval934031a2018-05-09 02:08:47 -070048}
49
Omar Sandovala8a45942018-05-09 02:08:48 -070050static inline bool wbt_is_read(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070051{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070052 return rq->wbt_flags & WBT_READ;
Omar Sandoval934031a2018-05-09 02:08:47 -070053}
54
Jens Axboee34cbd32016-11-09 12:36:15 -070055enum {
56 /*
57 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
58 * from here depending on device stats
59 */
60 RWB_DEF_DEPTH = 16,
61
62 /*
63 * 100msec window
64 */
65 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
66
67 /*
68 * Disregard stats, if we don't meet this minimum
69 */
70 RWB_MIN_WRITE_SAMPLES = 3,
71
72 /*
73 * If we have this number of consecutive windows with not enough
74 * information to scale up or down, scale up.
75 */
76 RWB_UNKNOWN_BUMP = 5,
77};
78
79static inline bool rwb_enabled(struct rq_wb *rwb)
80{
Zhang Yi1d0903d2021-06-19 17:36:59 +080081 return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&
82 rwb->wb_normal != 0;
Jens Axboee34cbd32016-11-09 12:36:15 -070083}
84
Jens Axboee34cbd32016-11-09 12:36:15 -070085static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
86{
87 if (rwb_enabled(rwb)) {
88 const unsigned long cur = jiffies;
89
90 if (cur != *var)
91 *var = cur;
92 }
93}
94
95/*
96 * If a task was rate throttled in balance_dirty_pages() within the last
97 * second or so, use that to indicate a higher cleaning rate.
98 */
99static bool wb_recent_wait(struct rq_wb *rwb)
100{
Christoph Hellwigd152c682021-08-16 15:46:24 +0200101 struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb;
Jens Axboee34cbd32016-11-09 12:36:15 -0700102
103 return time_before(jiffies, wb->dirty_sleep + HZ);
104}
105
Jens Axboe8bea6092018-05-07 09:57:08 -0600106static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
107 enum wbt_flags wb_acct)
Jens Axboee34cbd32016-11-09 12:36:15 -0700108{
Jens Axboe8bea6092018-05-07 09:57:08 -0600109 if (wb_acct & WBT_KSWAPD)
110 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
Jens Axboe782f5692018-05-07 10:03:23 -0600111 else if (wb_acct & WBT_DISCARD)
112 return &rwb->rq_wait[WBT_RWQ_DISCARD];
Jens Axboe8bea6092018-05-07 09:57:08 -0600113
114 return &rwb->rq_wait[WBT_RWQ_BG];
Jens Axboee34cbd32016-11-09 12:36:15 -0700115}
116
117static void rwb_wake_all(struct rq_wb *rwb)
118{
119 int i;
120
121 for (i = 0; i < WBT_NUM_RWQ; i++) {
122 struct rq_wait *rqw = &rwb->rq_wait[i];
123
Jens Axboeb7882092018-08-20 13:20:50 -0600124 if (wq_has_sleeper(&rqw->wait))
Jens Axboee34cbd32016-11-09 12:36:15 -0700125 wake_up_all(&rqw->wait);
126 }
127}
128
Jens Axboe061a5422018-08-26 10:09:06 -0600129static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
130 enum wbt_flags wb_acct)
Jens Axboee34cbd32016-11-09 12:36:15 -0700131{
Jens Axboee34cbd32016-11-09 12:36:15 -0700132 int inflight, limit;
133
Jens Axboee34cbd32016-11-09 12:36:15 -0700134 inflight = atomic_dec_return(&rqw->inflight);
135
136 /*
137 * wbt got disabled with IO in flight. Wake up any potential
138 * waiters, we don't have to do more than that.
139 */
140 if (unlikely(!rwb_enabled(rwb))) {
141 rwb_wake_all(rwb);
142 return;
143 }
144
145 /*
Jens Axboe782f5692018-05-07 10:03:23 -0600146 * For discards, our limit is always the background. For writes, if
147 * the device does write back caching, drop further down before we
148 * wake people up.
Jens Axboee34cbd32016-11-09 12:36:15 -0700149 */
Jens Axboe782f5692018-05-07 10:03:23 -0600150 if (wb_acct & WBT_DISCARD)
151 limit = rwb->wb_background;
152 else if (rwb->wc && !wb_recent_wait(rwb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700153 limit = 0;
154 else
155 limit = rwb->wb_normal;
156
157 /*
158 * Don't wake anyone up if we are above the normal limit.
159 */
160 if (inflight && inflight >= limit)
161 return;
162
Jens Axboeb7882092018-08-20 13:20:50 -0600163 if (wq_has_sleeper(&rqw->wait)) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700164 int diff = limit - inflight;
165
166 if (!inflight || diff >= rwb->wb_background / 2)
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600167 wake_up_all(&rqw->wait);
Jens Axboee34cbd32016-11-09 12:36:15 -0700168 }
169}
170
Jens Axboe061a5422018-08-26 10:09:06 -0600171static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
172{
173 struct rq_wb *rwb = RQWB(rqos);
174 struct rq_wait *rqw;
175
176 if (!(wb_acct & WBT_TRACKED))
177 return;
178
179 rqw = get_rq_wait(rwb, wb_acct);
180 wbt_rqw_done(rwb, rqw, wb_acct);
181}
182
Jens Axboee34cbd32016-11-09 12:36:15 -0700183/*
184 * Called on completion of a request. Note that it's also called when
185 * a request is merged, when the request gets freed.
186 */
Josef Bacika7905042018-07-03 09:32:35 -0600187static void wbt_done(struct rq_qos *rqos, struct request *rq)
Jens Axboee34cbd32016-11-09 12:36:15 -0700188{
Josef Bacika7905042018-07-03 09:32:35 -0600189 struct rq_wb *rwb = RQWB(rqos);
Jens Axboee34cbd32016-11-09 12:36:15 -0700190
Omar Sandovala8a45942018-05-09 02:08:48 -0700191 if (!wbt_is_tracked(rq)) {
192 if (rwb->sync_cookie == rq) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700193 rwb->sync_issue = 0;
194 rwb->sync_cookie = NULL;
195 }
196
Omar Sandovala8a45942018-05-09 02:08:48 -0700197 if (wbt_is_read(rq))
Jens Axboee34cbd32016-11-09 12:36:15 -0700198 wb_timestamp(rwb, &rwb->last_comp);
Jens Axboee34cbd32016-11-09 12:36:15 -0700199 } else {
Omar Sandovala8a45942018-05-09 02:08:48 -0700200 WARN_ON_ONCE(rq == rwb->sync_cookie);
Josef Bacika7905042018-07-03 09:32:35 -0600201 __wbt_done(rqos, wbt_flags(rq));
Jens Axboee34cbd32016-11-09 12:36:15 -0700202 }
Omar Sandovala8a45942018-05-09 02:08:48 -0700203 wbt_clear_state(rq);
Jens Axboee34cbd32016-11-09 12:36:15 -0700204}
205
Arnd Bergmann4121d382016-11-16 16:29:57 +0100206static inline bool stat_sample_valid(struct blk_rq_stat *stat)
Jens Axboee34cbd32016-11-09 12:36:15 -0700207{
208 /*
209 * We need at least one read sample, and a minimum of
210 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
211 * that it's writes impacting us, and not just some sole read on
212 * a device that is in a lower power state.
213 */
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700214 return (stat[READ].nr_samples >= 1 &&
215 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
Jens Axboee34cbd32016-11-09 12:36:15 -0700216}
217
218static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
219{
Mark Rutland6aa7de02017-10-23 14:07:29 -0700220 u64 now, issue = READ_ONCE(rwb->sync_issue);
Jens Axboee34cbd32016-11-09 12:36:15 -0700221
222 if (!issue || !rwb->sync_cookie)
223 return 0;
224
225 now = ktime_to_ns(ktime_get());
226 return now - issue;
227}
228
229enum {
230 LAT_OK = 1,
231 LAT_UNKNOWN,
232 LAT_UNKNOWN_WRITES,
233 LAT_EXCEEDED,
234};
235
Omar Sandoval34dbad52017-03-21 08:56:08 -0700236static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
Jens Axboee34cbd32016-11-09 12:36:15 -0700237{
Christoph Hellwigd152c682021-08-16 15:46:24 +0200238 struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
Josef Bacika7905042018-07-03 09:32:35 -0600239 struct rq_depth *rqd = &rwb->rq_depth;
Jens Axboee34cbd32016-11-09 12:36:15 -0700240 u64 thislat;
241
242 /*
243 * If our stored sync issue exceeds the window size, or it
244 * exceeds our min target AND we haven't logged any entries,
245 * flag the latency as exceeded. wbt works off completion latencies,
246 * but for a flooded device, a single sync IO can take a long time
247 * to complete after being issued. If this time exceeds our
248 * monitoring window AND we didn't see any other completions in that
249 * window, then count that sync IO as a violation of the latency.
250 */
251 thislat = rwb_sync_issue_lat(rwb);
252 if (thislat > rwb->cur_win_nsec ||
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700253 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700254 trace_wbt_lat(bdi, thislat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700255 return LAT_EXCEEDED;
256 }
257
258 /*
259 * No read/write mix, if stat isn't valid
260 */
261 if (!stat_sample_valid(stat)) {
262 /*
263 * If we had writes in this stat window and the window is
264 * current, we're only doing writes. If a task recently
265 * waited or still has writes in flights, consider us doing
266 * just writes as well.
267 */
Omar Sandoval34dbad52017-03-21 08:56:08 -0700268 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
269 wbt_inflight(rwb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700270 return LAT_UNKNOWN_WRITES;
271 return LAT_UNKNOWN;
272 }
273
274 /*
275 * If the 'min' latency exceeds our target, step down.
276 */
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700277 if (stat[READ].min > rwb->min_lat_nsec) {
278 trace_wbt_lat(bdi, stat[READ].min);
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700279 trace_wbt_stat(bdi, stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700280 return LAT_EXCEEDED;
281 }
282
Josef Bacika7905042018-07-03 09:32:35 -0600283 if (rqd->scale_step)
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700284 trace_wbt_stat(bdi, stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700285
286 return LAT_OK;
287}
288
Jens Axboee34cbd32016-11-09 12:36:15 -0700289static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
290{
Christoph Hellwigd152c682021-08-16 15:46:24 +0200291 struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
Josef Bacika7905042018-07-03 09:32:35 -0600292 struct rq_depth *rqd = &rwb->rq_depth;
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700293
Josef Bacika7905042018-07-03 09:32:35 -0600294 trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
295 rwb->wb_background, rwb->wb_normal, rqd->max_depth);
296}
297
298static void calc_wb_limits(struct rq_wb *rwb)
299{
300 if (rwb->min_lat_nsec == 0) {
301 rwb->wb_normal = rwb->wb_background = 0;
302 } else if (rwb->rq_depth.max_depth <= 2) {
303 rwb->wb_normal = rwb->rq_depth.max_depth;
304 rwb->wb_background = 1;
305 } else {
306 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
307 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
308 }
Jens Axboee34cbd32016-11-09 12:36:15 -0700309}
310
311static void scale_up(struct rq_wb *rwb)
312{
Harshad Shirwadkarb84477d2019-10-05 11:59:27 -0700313 if (!rq_depth_scale_up(&rwb->rq_depth))
314 return;
Josef Bacika7905042018-07-03 09:32:35 -0600315 calc_wb_limits(rwb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700316 rwb->unknown_cnt = 0;
Josef Bacik5e65a2032018-10-11 15:29:30 -0400317 rwb_wake_all(rwb);
Tommi Rantala3a89c252020-04-17 16:00:22 +0300318 rwb_trace_step(rwb, tracepoint_string("scale up"));
Jens Axboee34cbd32016-11-09 12:36:15 -0700319}
320
Jens Axboee34cbd32016-11-09 12:36:15 -0700321static void scale_down(struct rq_wb *rwb, bool hard_throttle)
322{
Harshad Shirwadkarb84477d2019-10-05 11:59:27 -0700323 if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
324 return;
Jens Axboee34cbd32016-11-09 12:36:15 -0700325 calc_wb_limits(rwb);
Josef Bacika7905042018-07-03 09:32:35 -0600326 rwb->unknown_cnt = 0;
Tommi Rantala3a89c252020-04-17 16:00:22 +0300327 rwb_trace_step(rwb, tracepoint_string("scale down"));
Jens Axboee34cbd32016-11-09 12:36:15 -0700328}
329
330static void rwb_arm_timer(struct rq_wb *rwb)
331{
Josef Bacika7905042018-07-03 09:32:35 -0600332 struct rq_depth *rqd = &rwb->rq_depth;
333
334 if (rqd->scale_step > 0) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700335 /*
336 * We should speed this up, using some variant of a fast
337 * integer inverse square root calculation. Since we only do
338 * this for every window expiration, it's not a huge deal,
339 * though.
340 */
341 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
Josef Bacika7905042018-07-03 09:32:35 -0600342 int_sqrt((rqd->scale_step + 1) << 8));
Jens Axboee34cbd32016-11-09 12:36:15 -0700343 } else {
344 /*
345 * For step < 0, we don't want to increase/decrease the
346 * window size.
347 */
348 rwb->cur_win_nsec = rwb->win_nsec;
349 }
350
Omar Sandoval34dbad52017-03-21 08:56:08 -0700351 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
Jens Axboee34cbd32016-11-09 12:36:15 -0700352}
353
Omar Sandoval34dbad52017-03-21 08:56:08 -0700354static void wb_timer_fn(struct blk_stat_callback *cb)
Jens Axboee34cbd32016-11-09 12:36:15 -0700355{
Omar Sandoval34dbad52017-03-21 08:56:08 -0700356 struct rq_wb *rwb = cb->data;
Josef Bacika7905042018-07-03 09:32:35 -0600357 struct rq_depth *rqd = &rwb->rq_depth;
Jens Axboee34cbd32016-11-09 12:36:15 -0700358 unsigned int inflight = wbt_inflight(rwb);
359 int status;
360
Andrea Righi480d42d2021-10-19 11:20:26 +0200361 if (!rwb->rqos.q->disk)
362 return;
363
Omar Sandoval34dbad52017-03-21 08:56:08 -0700364 status = latency_exceeded(rwb, cb->stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700365
Christoph Hellwigd152c682021-08-16 15:46:24 +0200366 trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step,
367 inflight);
Jens Axboee34cbd32016-11-09 12:36:15 -0700368
369 /*
370 * If we exceeded the latency target, step down. If we did not,
371 * step one level up. If we don't know enough to say either exceeded
372 * or ok, then don't do anything.
373 */
374 switch (status) {
375 case LAT_EXCEEDED:
376 scale_down(rwb, true);
377 break;
378 case LAT_OK:
379 scale_up(rwb);
380 break;
381 case LAT_UNKNOWN_WRITES:
382 /*
383 * We started a the center step, but don't have a valid
384 * read/write sample, but we do have writes going on.
385 * Allow step to go negative, to increase write perf.
386 */
387 scale_up(rwb);
388 break;
389 case LAT_UNKNOWN:
390 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
391 break;
392 /*
393 * We get here when previously scaled reduced depth, and we
394 * currently don't have a valid read/write sample. For that
395 * case, slowly return to center state (step == 0).
396 */
Josef Bacika7905042018-07-03 09:32:35 -0600397 if (rqd->scale_step > 0)
Jens Axboee34cbd32016-11-09 12:36:15 -0700398 scale_up(rwb);
Josef Bacika7905042018-07-03 09:32:35 -0600399 else if (rqd->scale_step < 0)
Jens Axboee34cbd32016-11-09 12:36:15 -0700400 scale_down(rwb, false);
401 break;
402 default:
403 break;
404 }
405
406 /*
407 * Re-arm timer, if we have IO in flight
408 */
Josef Bacika7905042018-07-03 09:32:35 -0600409 if (rqd->scale_step || inflight)
Jens Axboee34cbd32016-11-09 12:36:15 -0700410 rwb_arm_timer(rwb);
411}
412
Guoqing Jiang4d89e1d2020-05-09 00:00:15 +0200413static void wbt_update_limits(struct rq_wb *rwb)
Jens Axboee34cbd32016-11-09 12:36:15 -0700414{
Josef Bacika7905042018-07-03 09:32:35 -0600415 struct rq_depth *rqd = &rwb->rq_depth;
416
417 rqd->scale_step = 0;
418 rqd->scaled_max = false;
419
420 rq_depth_calc_max_depth(rqd);
Jens Axboee34cbd32016-11-09 12:36:15 -0700421 calc_wb_limits(rwb);
422
423 rwb_wake_all(rwb);
424}
425
Yu Kuai3642ef42022-10-19 20:15:16 +0800426bool wbt_disabled(struct request_queue *q)
427{
428 struct rq_qos *rqos = wbt_rq_qos(q);
429
430 return !rqos || RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT ||
431 RQWB(rqos)->enable_state == WBT_STATE_OFF_MANUAL;
432}
433
Josef Bacika7905042018-07-03 09:32:35 -0600434u64 wbt_get_min_lat(struct request_queue *q)
435{
436 struct rq_qos *rqos = wbt_rq_qos(q);
437 if (!rqos)
438 return 0;
439 return RQWB(rqos)->min_lat_nsec;
440}
441
442void wbt_set_min_lat(struct request_queue *q, u64 val)
443{
444 struct rq_qos *rqos = wbt_rq_qos(q);
445 if (!rqos)
446 return;
Yu Kuaia9a236d2022-10-19 20:15:15 +0800447
Josef Bacika7905042018-07-03 09:32:35 -0600448 RQWB(rqos)->min_lat_nsec = val;
Yu Kuaia9a236d2022-10-19 20:15:15 +0800449 if (val)
450 RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
451 else
452 RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL;
453
Guoqing Jiang4d89e1d2020-05-09 00:00:15 +0200454 wbt_update_limits(RQWB(rqos));
Josef Bacika7905042018-07-03 09:32:35 -0600455}
456
457
Jens Axboee34cbd32016-11-09 12:36:15 -0700458static bool close_io(struct rq_wb *rwb)
459{
460 const unsigned long now = jiffies;
461
462 return time_before(now, rwb->last_issue + HZ / 10) ||
463 time_before(now, rwb->last_comp + HZ / 10);
464}
465
466#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
467
Bart Van Assche16458cf2022-07-14 11:06:32 -0700468static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf)
Jens Axboee34cbd32016-11-09 12:36:15 -0700469{
470 unsigned int limit;
471
Jens Axboeffa358d2018-08-20 13:24:25 -0600472 /*
473 * If we got disabled, just return UINT_MAX. This ensures that
474 * we'll properly inc a new IO, and dec+wakeup at the end.
475 */
476 if (!rwb_enabled(rwb))
477 return UINT_MAX;
478
Bart Van Assche16458cf2022-07-14 11:06:32 -0700479 if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD)
Jens Axboe782f5692018-05-07 10:03:23 -0600480 return rwb->wb_background;
481
Jens Axboee34cbd32016-11-09 12:36:15 -0700482 /*
483 * At this point we know it's a buffered write. If this is
weiping zhang3dfbdc42017-11-23 21:40:10 +0800484 * kswapd trying to free memory, or REQ_SYNC is set, then
Jens Axboee34cbd32016-11-09 12:36:15 -0700485 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
486 * that. If the write is marked as a background write, then use
487 * the idle limit, or go to normal if we haven't had competing
488 * IO for a bit.
489 */
Bart Van Assche16458cf2022-07-14 11:06:32 -0700490 if ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
Josef Bacika7905042018-07-03 09:32:35 -0600491 limit = rwb->rq_depth.max_depth;
Bart Van Assche16458cf2022-07-14 11:06:32 -0700492 else if ((opf & REQ_BACKGROUND) || close_io(rwb)) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700493 /*
494 * If less than 100ms since we completed unrelated IO,
495 * limit us to half the depth for background writeback.
496 */
497 limit = rwb->wb_background;
498 } else
499 limit = rwb->wb_normal;
500
501 return limit;
502}
503
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600504struct wbt_wait_data {
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600505 struct rq_wb *rwb;
Josef Bacikb6c7b582018-12-04 12:59:03 -0500506 enum wbt_flags wb_acct;
Bart Van Assche16458cf2022-07-14 11:06:32 -0700507 blk_opf_t opf;
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600508};
509
Josef Bacikb6c7b582018-12-04 12:59:03 -0500510static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600511{
Josef Bacikb6c7b582018-12-04 12:59:03 -0500512 struct wbt_wait_data *data = private_data;
Bart Van Assche16458cf2022-07-14 11:06:32 -0700513 return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf));
Josef Bacikb6c7b582018-12-04 12:59:03 -0500514}
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600515
Josef Bacikb6c7b582018-12-04 12:59:03 -0500516static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
517{
518 struct wbt_wait_data *data = private_data;
519 wbt_rqw_done(data->rwb, rqw, data->wb_acct);
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600520}
521
Jens Axboee34cbd32016-11-09 12:36:15 -0700522/*
523 * Block if we will exceed our limit, or if we are currently waiting for
524 * the timer to kick off queuing again.
525 */
Jens Axboe8bea6092018-05-07 09:57:08 -0600526static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700527 blk_opf_t opf)
Jens Axboee34cbd32016-11-09 12:36:15 -0700528{
Jens Axboe8bea6092018-05-07 09:57:08 -0600529 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600530 struct wbt_wait_data data = {
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600531 .rwb = rwb,
Josef Bacikb6c7b582018-12-04 12:59:03 -0500532 .wb_acct = wb_acct,
Bart Van Assche16458cf2022-07-14 11:06:32 -0700533 .opf = opf,
Jens Axboe38cfb5a2018-08-26 10:10:05 -0600534 };
Jens Axboee34cbd32016-11-09 12:36:15 -0700535
Josef Bacikb6c7b582018-12-04 12:59:03 -0500536 rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700537}
538
Lei Chen482e3022021-01-25 19:27:04 +0800539static inline bool wbt_should_throttle(struct bio *bio)
Jens Axboee34cbd32016-11-09 12:36:15 -0700540{
Jens Axboe782f5692018-05-07 10:03:23 -0600541 switch (bio_op(bio)) {
542 case REQ_OP_WRITE:
543 /*
544 * Don't throttle WRITE_ODIRECT
545 */
546 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
547 (REQ_SYNC | REQ_IDLE))
548 return false;
Gustavo A. R. Silvadf561f662020-08-23 17:36:59 -0500549 fallthrough;
Jens Axboe782f5692018-05-07 10:03:23 -0600550 case REQ_OP_DISCARD:
551 return true;
552 default:
Jens Axboee34cbd32016-11-09 12:36:15 -0700553 return false;
Jens Axboe782f5692018-05-07 10:03:23 -0600554 }
Jens Axboee34cbd32016-11-09 12:36:15 -0700555}
556
Josef Bacikc1c80382018-07-03 11:14:59 -0400557static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
558{
559 enum wbt_flags flags = 0;
560
Jens Axboec1253112018-08-23 09:34:46 -0600561 if (!rwb_enabled(rwb))
562 return 0;
563
Josef Bacikc1c80382018-07-03 11:14:59 -0400564 if (bio_op(bio) == REQ_OP_READ) {
565 flags = WBT_READ;
Lei Chen482e3022021-01-25 19:27:04 +0800566 } else if (wbt_should_throttle(bio)) {
Josef Bacikc1c80382018-07-03 11:14:59 -0400567 if (current_is_kswapd())
568 flags |= WBT_KSWAPD;
569 if (bio_op(bio) == REQ_OP_DISCARD)
570 flags |= WBT_DISCARD;
571 flags |= WBT_TRACKED;
572 }
573 return flags;
574}
575
576static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
577{
578 struct rq_wb *rwb = RQWB(rqos);
579 enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
580 __wbt_done(rqos, flags);
581}
582
Jens Axboee34cbd32016-11-09 12:36:15 -0700583/*
Jens Axboee34cbd32016-11-09 12:36:15 -0700584 * May sleep, if we have exceeded the writeback limits. Caller can pass
585 * in an irq held spinlock, if it holds one when calling this function.
586 * If we do sleep, we'll release and re-grab it.
587 */
Christoph Hellwigd5337562018-11-14 17:02:09 +0100588static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
Jens Axboee34cbd32016-11-09 12:36:15 -0700589{
Josef Bacika7905042018-07-03 09:32:35 -0600590 struct rq_wb *rwb = RQWB(rqos);
Josef Bacikc1c80382018-07-03 11:14:59 -0400591 enum wbt_flags flags;
Jens Axboee34cbd32016-11-09 12:36:15 -0700592
Josef Bacikc1c80382018-07-03 11:14:59 -0400593 flags = bio_to_wbt_flags(rwb, bio);
Ming Leidf60f6e2018-08-14 23:57:49 +0800594 if (!(flags & WBT_TRACKED)) {
Josef Bacikc1c80382018-07-03 11:14:59 -0400595 if (flags & WBT_READ)
Jens Axboee34cbd32016-11-09 12:36:15 -0700596 wb_timestamp(rwb, &rwb->last_issue);
Josef Bacikc1c80382018-07-03 11:14:59 -0400597 return;
Jens Axboee34cbd32016-11-09 12:36:15 -0700598 }
599
Christoph Hellwigd5337562018-11-14 17:02:09 +0100600 __wbt_wait(rwb, flags, bio->bi_opf);
Jens Axboee34cbd32016-11-09 12:36:15 -0700601
Omar Sandoval34dbad52017-03-21 08:56:08 -0700602 if (!blk_stat_is_active(rwb->cb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700603 rwb_arm_timer(rwb);
Josef Bacikc1c80382018-07-03 11:14:59 -0400604}
Jens Axboee34cbd32016-11-09 12:36:15 -0700605
Josef Bacikc1c80382018-07-03 11:14:59 -0400606static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
607{
608 struct rq_wb *rwb = RQWB(rqos);
609 rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
Jens Axboee34cbd32016-11-09 12:36:15 -0700610}
611
Bart Van Asschec83f5362019-01-23 11:05:57 -0800612static void wbt_issue(struct rq_qos *rqos, struct request *rq)
Jens Axboee34cbd32016-11-09 12:36:15 -0700613{
Josef Bacika7905042018-07-03 09:32:35 -0600614 struct rq_wb *rwb = RQWB(rqos);
615
Jens Axboee34cbd32016-11-09 12:36:15 -0700616 if (!rwb_enabled(rwb))
617 return;
618
619 /*
Omar Sandovala8a45942018-05-09 02:08:48 -0700620 * Track sync issue, in case it takes a long time to complete. Allows us
621 * to react quicker, if a sync IO takes a long time to complete. Note
622 * that this is just a hint. The request can go away when it completes,
623 * so it's important we never dereference it. We only use the address to
624 * compare with, which is why we store the sync_issue time locally.
Jens Axboee34cbd32016-11-09 12:36:15 -0700625 */
Omar Sandovala8a45942018-05-09 02:08:48 -0700626 if (wbt_is_read(rq) && !rwb->sync_issue) {
627 rwb->sync_cookie = rq;
Omar Sandoval544ccc8d2018-05-09 02:08:50 -0700628 rwb->sync_issue = rq->io_start_time_ns;
Jens Axboee34cbd32016-11-09 12:36:15 -0700629 }
630}
631
Bart Van Asschec83f5362019-01-23 11:05:57 -0800632static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
Jens Axboee34cbd32016-11-09 12:36:15 -0700633{
Josef Bacika7905042018-07-03 09:32:35 -0600634 struct rq_wb *rwb = RQWB(rqos);
Jens Axboee34cbd32016-11-09 12:36:15 -0700635 if (!rwb_enabled(rwb))
636 return;
Omar Sandovala8a45942018-05-09 02:08:48 -0700637 if (rq == rwb->sync_cookie) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700638 rwb->sync_issue = 0;
639 rwb->sync_cookie = NULL;
640 }
641}
642
Josef Bacika7905042018-07-03 09:32:35 -0600643void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
Jens Axboee34cbd32016-11-09 12:36:15 -0700644{
Josef Bacika7905042018-07-03 09:32:35 -0600645 struct rq_qos *rqos = wbt_rq_qos(q);
646 if (rqos)
647 RQWB(rqos)->wc = write_cache_on;
Jens Axboee34cbd32016-11-09 12:36:15 -0700648}
649
Jan Kara3f19cd22017-04-11 11:29:01 +0200650/*
Jan Kara8330cdb2017-04-19 11:33:27 +0200651 * Enable wbt if defaults are configured that way
652 */
653void wbt_enable_default(struct request_queue *q)
654{
Yu Kuai671fae52022-10-19 20:15:18 +0800655 struct rq_qos *rqos;
656 bool disable_flag = q->elevator &&
657 test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags);
Zhang Yi76a80402021-06-19 17:37:00 +0800658
Jan Kara8330cdb2017-04-19 11:33:27 +0200659 /* Throttling already enabled? */
Yu Kuai671fae52022-10-19 20:15:18 +0800660 rqos = wbt_rq_qos(q);
Zhang Yi76a80402021-06-19 17:37:00 +0800661 if (rqos) {
Yu Kuai671fae52022-10-19 20:15:18 +0800662 if (!disable_flag &&
663 RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)
Zhang Yi76a80402021-06-19 17:37:00 +0800664 RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;
Jan Kara8330cdb2017-04-19 11:33:27 +0200665 return;
Zhang Yi76a80402021-06-19 17:37:00 +0800666 }
Jan Kara8330cdb2017-04-19 11:33:27 +0200667
668 /* Queue not registered? Maybe shutting down... */
Ming Lei58c898b2019-08-27 19:01:47 +0800669 if (!blk_queue_registered(q))
Jan Kara8330cdb2017-04-19 11:33:27 +0200670 return;
671
Yu Kuai671fae52022-10-19 20:15:18 +0800672 if (queue_is_mq(q) && !disable_flag)
Jan Kara8330cdb2017-04-19 11:33:27 +0200673 wbt_init(q);
674}
675EXPORT_SYMBOL_GPL(wbt_enable_default);
676
Jens Axboe80e091d2016-11-28 09:22:47 -0700677u64 wbt_default_latency_nsec(struct request_queue *q)
678{
679 /*
680 * We default to 2msec for non-rotational storage, and 75msec
681 * for rotational storage.
682 */
683 if (blk_queue_nonrot(q))
684 return 2000000ULL;
685 else
686 return 75000000ULL;
687}
688
Jens Axboe99c749a2017-04-21 07:55:42 -0600689static int wbt_data_dir(const struct request *rq)
690{
Bart Van Assche77e7ffd2022-07-14 11:06:28 -0700691 const enum req_op op = req_op(rq);
Jens Axboe5235553d2018-02-05 13:16:56 -0700692
693 if (op == REQ_OP_READ)
694 return READ;
Jens Axboe825843b2018-05-03 09:14:57 -0600695 else if (op_is_write(op))
Jens Axboe5235553d2018-02-05 13:16:56 -0700696 return WRITE;
697
698 /* don't account */
699 return -1;
Jens Axboe99c749a2017-04-21 07:55:42 -0600700}
701
Tejun Heo9677a3e2019-08-28 15:05:55 -0700702static void wbt_queue_depth_changed(struct rq_qos *rqos)
703{
704 RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);
Guoqing Jiang4d89e1d2020-05-09 00:00:15 +0200705 wbt_update_limits(RQWB(rqos));
Tejun Heo9677a3e2019-08-28 15:05:55 -0700706}
707
Josef Bacika7905042018-07-03 09:32:35 -0600708static void wbt_exit(struct rq_qos *rqos)
709{
710 struct rq_wb *rwb = RQWB(rqos);
711 struct request_queue *q = rqos->q;
712
713 blk_stat_remove_callback(q, rwb->cb);
714 blk_stat_free_callback(rwb->cb);
715 kfree(rwb);
716}
717
718/*
719 * Disable wbt, if enabled by default.
720 */
721void wbt_disable_default(struct request_queue *q)
722{
723 struct rq_qos *rqos = wbt_rq_qos(q);
724 struct rq_wb *rwb;
725 if (!rqos)
726 return;
727 rwb = RQWB(rqos);
Ming Lei544fbd12018-12-12 19:44:34 +0800728 if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
729 blk_stat_deactivate(rwb->cb);
Zhang Yi1d0903d2021-06-19 17:36:59 +0800730 rwb->enable_state = WBT_STATE_OFF_DEFAULT;
Ming Lei544fbd12018-12-12 19:44:34 +0800731 }
Josef Bacika7905042018-07-03 09:32:35 -0600732}
Jens Axboee815f402018-11-15 12:31:27 -0700733EXPORT_SYMBOL_GPL(wbt_disable_default);
Josef Bacika7905042018-07-03 09:32:35 -0600734
Ming Leid19afeb2018-12-17 09:46:01 +0800735#ifdef CONFIG_BLK_DEBUG_FS
736static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
737{
738 struct rq_qos *rqos = data;
739 struct rq_wb *rwb = RQWB(rqos);
740
741 seq_printf(m, "%llu\n", rwb->cur_win_nsec);
742 return 0;
743}
744
745static int wbt_enabled_show(void *data, struct seq_file *m)
746{
747 struct rq_qos *rqos = data;
748 struct rq_wb *rwb = RQWB(rqos);
749
750 seq_printf(m, "%d\n", rwb->enable_state);
751 return 0;
752}
753
754static int wbt_id_show(void *data, struct seq_file *m)
755{
756 struct rq_qos *rqos = data;
757
758 seq_printf(m, "%u\n", rqos->id);
759 return 0;
760}
761
762static int wbt_inflight_show(void *data, struct seq_file *m)
763{
764 struct rq_qos *rqos = data;
765 struct rq_wb *rwb = RQWB(rqos);
766 int i;
767
768 for (i = 0; i < WBT_NUM_RWQ; i++)
769 seq_printf(m, "%d: inflight %d\n", i,
770 atomic_read(&rwb->rq_wait[i].inflight));
771 return 0;
772}
773
774static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
775{
776 struct rq_qos *rqos = data;
777 struct rq_wb *rwb = RQWB(rqos);
778
779 seq_printf(m, "%lu\n", rwb->min_lat_nsec);
780 return 0;
781}
782
783static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
784{
785 struct rq_qos *rqos = data;
786 struct rq_wb *rwb = RQWB(rqos);
787
788 seq_printf(m, "%u\n", rwb->unknown_cnt);
789 return 0;
790}
791
792static int wbt_normal_show(void *data, struct seq_file *m)
793{
794 struct rq_qos *rqos = data;
795 struct rq_wb *rwb = RQWB(rqos);
796
797 seq_printf(m, "%u\n", rwb->wb_normal);
798 return 0;
799}
800
801static int wbt_background_show(void *data, struct seq_file *m)
802{
803 struct rq_qos *rqos = data;
804 struct rq_wb *rwb = RQWB(rqos);
805
806 seq_printf(m, "%u\n", rwb->wb_background);
807 return 0;
808}
809
810static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
811 {"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
812 {"enabled", 0400, wbt_enabled_show},
813 {"id", 0400, wbt_id_show},
814 {"inflight", 0400, wbt_inflight_show},
815 {"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
816 {"unknown_cnt", 0400, wbt_unknown_cnt_show},
817 {"wb_normal", 0400, wbt_normal_show},
818 {"wb_background", 0400, wbt_background_show},
819 {},
820};
821#endif
822
Josef Bacika7905042018-07-03 09:32:35 -0600823static struct rq_qos_ops wbt_rqos_ops = {
824 .throttle = wbt_wait,
825 .issue = wbt_issue,
Josef Bacikc1c80382018-07-03 11:14:59 -0400826 .track = wbt_track,
Josef Bacika7905042018-07-03 09:32:35 -0600827 .requeue = wbt_requeue,
828 .done = wbt_done,
Josef Bacikc1c80382018-07-03 11:14:59 -0400829 .cleanup = wbt_cleanup,
Tejun Heo9677a3e2019-08-28 15:05:55 -0700830 .queue_depth_changed = wbt_queue_depth_changed,
Josef Bacika7905042018-07-03 09:32:35 -0600831 .exit = wbt_exit,
Ming Leid19afeb2018-12-17 09:46:01 +0800832#ifdef CONFIG_BLK_DEBUG_FS
833 .debugfs_attrs = wbt_debugfs_attrs,
834#endif
Josef Bacika7905042018-07-03 09:32:35 -0600835};
836
Jens Axboe8054b892016-11-10 21:50:51 -0700837int wbt_init(struct request_queue *q)
Jens Axboee34cbd32016-11-09 12:36:15 -0700838{
839 struct rq_wb *rwb;
840 int i;
Jinke Han14a6e2e2022-07-20 17:36:16 +0800841 int ret;
Jens Axboee34cbd32016-11-09 12:36:15 -0700842
Jens Axboee34cbd32016-11-09 12:36:15 -0700843 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
844 if (!rwb)
845 return -ENOMEM;
846
Jens Axboe99c749a2017-04-21 07:55:42 -0600847 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
Omar Sandoval34dbad52017-03-21 08:56:08 -0700848 if (!rwb->cb) {
849 kfree(rwb);
850 return -ENOMEM;
851 }
852
Josef Bacika7905042018-07-03 09:32:35 -0600853 for (i = 0; i < WBT_NUM_RWQ; i++)
854 rq_wait_init(&rwb->rq_wait[i]);
Jens Axboee34cbd32016-11-09 12:36:15 -0700855
Josef Bacika7905042018-07-03 09:32:35 -0600856 rwb->rqos.id = RQ_QOS_WBT;
857 rwb->rqos.ops = &wbt_rqos_ops;
858 rwb->rqos.q = q;
Jens Axboee34cbd32016-11-09 12:36:15 -0700859 rwb->last_comp = rwb->last_issue = jiffies;
Jens Axboee34cbd32016-11-09 12:36:15 -0700860 rwb->win_nsec = RWB_WINDOW_NSEC;
Jens Axboed62118b2016-11-28 09:40:34 -0700861 rwb->enable_state = WBT_STATE_ON_DEFAULT;
Yu Kuai285feba2022-10-09 18:10:38 +0800862 rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags);
Josef Bacika7905042018-07-03 09:32:35 -0600863 rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
Yu Kuai8c5035d2022-09-13 18:57:49 +0800864 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
865
866 wbt_queue_depth_changed(&rwb->rqos);
Jens Axboee34cbd32016-11-09 12:36:15 -0700867
868 /*
Omar Sandoval34dbad52017-03-21 08:56:08 -0700869 * Assign rwb and add the stats callback.
Jens Axboee34cbd32016-11-09 12:36:15 -0700870 */
Jinke Han14a6e2e2022-07-20 17:36:16 +0800871 ret = rq_qos_add(q, &rwb->rqos);
872 if (ret)
873 goto err_free;
874
Omar Sandoval34dbad52017-03-21 08:56:08 -0700875 blk_stat_add_callback(q, rwb->cb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700876
Jens Axboee34cbd32016-11-09 12:36:15 -0700877 return 0;
Jinke Han14a6e2e2022-07-20 17:36:16 +0800878
879err_free:
880 blk_stat_free_callback(rwb->cb);
881 kfree(rwb);
882 return ret;
883
Jens Axboee34cbd32016-11-09 12:36:15 -0700884}