blob: 84507d3e9a981d5e4888c9b3783f0155f992adba [file] [log] [blame]
Jens Axboee34cbd32016-11-09 12:36:15 -07001/*
2 * buffered writeback throttling. loosely based on CoDel. We can't drop
3 * packets for IO scheduling, so the logic is something like this:
4 *
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.
17 *
18 * Copyright (C) 2016 Jens Axboe
19 *
20 */
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>
26
27#include "blk-wbt.h"
Josef Bacika7905042018-07-03 09:32:35 -060028#include "blk-rq-qos.h"
Jens Axboee34cbd32016-11-09 12:36:15 -070029
30#define CREATE_TRACE_POINTS
31#include <trace/events/wbt.h>
32
Omar Sandovala8a45942018-05-09 02:08:48 -070033static inline void wbt_clear_state(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070034{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070035 rq->wbt_flags = 0;
Omar Sandoval934031a2018-05-09 02:08:47 -070036}
37
Omar Sandovala8a45942018-05-09 02:08:48 -070038static inline enum wbt_flags wbt_flags(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070039{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070040 return rq->wbt_flags;
Omar Sandoval934031a2018-05-09 02:08:47 -070041}
42
Omar Sandovala8a45942018-05-09 02:08:48 -070043static inline bool wbt_is_tracked(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070044{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070045 return rq->wbt_flags & WBT_TRACKED;
Omar Sandoval934031a2018-05-09 02:08:47 -070046}
47
Omar Sandovala8a45942018-05-09 02:08:48 -070048static inline bool wbt_is_read(struct request *rq)
Omar Sandoval934031a2018-05-09 02:08:47 -070049{
Omar Sandoval544ccc8d2018-05-09 02:08:50 -070050 return rq->wbt_flags & WBT_READ;
Omar Sandoval934031a2018-05-09 02:08:47 -070051}
52
Jens Axboee34cbd32016-11-09 12:36:15 -070053enum {
54 /*
55 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
56 * from here depending on device stats
57 */
58 RWB_DEF_DEPTH = 16,
59
60 /*
61 * 100msec window
62 */
63 RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL,
64
65 /*
66 * Disregard stats, if we don't meet this minimum
67 */
68 RWB_MIN_WRITE_SAMPLES = 3,
69
70 /*
71 * If we have this number of consecutive windows with not enough
72 * information to scale up or down, scale up.
73 */
74 RWB_UNKNOWN_BUMP = 5,
75};
76
77static inline bool rwb_enabled(struct rq_wb *rwb)
78{
79 return rwb && rwb->wb_normal != 0;
80}
81
Jens Axboee34cbd32016-11-09 12:36:15 -070082static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
83{
84 if (rwb_enabled(rwb)) {
85 const unsigned long cur = jiffies;
86
87 if (cur != *var)
88 *var = cur;
89 }
90}
91
92/*
93 * If a task was rate throttled in balance_dirty_pages() within the last
94 * second or so, use that to indicate a higher cleaning rate.
95 */
96static bool wb_recent_wait(struct rq_wb *rwb)
97{
Josef Bacika7905042018-07-03 09:32:35 -060098 struct bdi_writeback *wb = &rwb->rqos.q->backing_dev_info->wb;
Jens Axboee34cbd32016-11-09 12:36:15 -070099
100 return time_before(jiffies, wb->dirty_sleep + HZ);
101}
102
Jens Axboe8bea6092018-05-07 09:57:08 -0600103static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
104 enum wbt_flags wb_acct)
Jens Axboee34cbd32016-11-09 12:36:15 -0700105{
Jens Axboe8bea6092018-05-07 09:57:08 -0600106 if (wb_acct & WBT_KSWAPD)
107 return &rwb->rq_wait[WBT_RWQ_KSWAPD];
Jens Axboe782f5692018-05-07 10:03:23 -0600108 else if (wb_acct & WBT_DISCARD)
109 return &rwb->rq_wait[WBT_RWQ_DISCARD];
Jens Axboe8bea6092018-05-07 09:57:08 -0600110
111 return &rwb->rq_wait[WBT_RWQ_BG];
Jens Axboee34cbd32016-11-09 12:36:15 -0700112}
113
114static void rwb_wake_all(struct rq_wb *rwb)
115{
116 int i;
117
118 for (i = 0; i < WBT_NUM_RWQ; i++) {
119 struct rq_wait *rqw = &rwb->rq_wait[i];
120
Jens Axboeb7882092018-08-20 13:20:50 -0600121 if (wq_has_sleeper(&rqw->wait))
Jens Axboee34cbd32016-11-09 12:36:15 -0700122 wake_up_all(&rqw->wait);
123 }
124}
125
Josef Bacika7905042018-07-03 09:32:35 -0600126static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
Jens Axboee34cbd32016-11-09 12:36:15 -0700127{
Josef Bacika7905042018-07-03 09:32:35 -0600128 struct rq_wb *rwb = RQWB(rqos);
Jens Axboee34cbd32016-11-09 12:36:15 -0700129 struct rq_wait *rqw;
130 int inflight, limit;
131
132 if (!(wb_acct & WBT_TRACKED))
133 return;
134
Jens Axboe8bea6092018-05-07 09:57:08 -0600135 rqw = get_rq_wait(rwb, wb_acct);
Jens Axboee34cbd32016-11-09 12:36:15 -0700136 inflight = atomic_dec_return(&rqw->inflight);
137
138 /*
139 * wbt got disabled with IO in flight. Wake up any potential
140 * waiters, we don't have to do more than that.
141 */
142 if (unlikely(!rwb_enabled(rwb))) {
143 rwb_wake_all(rwb);
144 return;
145 }
146
147 /*
Jens Axboe782f5692018-05-07 10:03:23 -0600148 * For discards, our limit is always the background. For writes, if
149 * the device does write back caching, drop further down before we
150 * wake people up.
Jens Axboee34cbd32016-11-09 12:36:15 -0700151 */
Jens Axboe782f5692018-05-07 10:03:23 -0600152 if (wb_acct & WBT_DISCARD)
153 limit = rwb->wb_background;
154 else if (rwb->wc && !wb_recent_wait(rwb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700155 limit = 0;
156 else
157 limit = rwb->wb_normal;
158
159 /*
160 * Don't wake anyone up if we are above the normal limit.
161 */
162 if (inflight && inflight >= limit)
163 return;
164
Jens Axboeb7882092018-08-20 13:20:50 -0600165 if (wq_has_sleeper(&rqw->wait)) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700166 int diff = limit - inflight;
167
168 if (!inflight || diff >= rwb->wb_background / 2)
Anchal Agarwal2887e412018-08-07 14:40:49 -0600169 wake_up(&rqw->wait);
Jens Axboee34cbd32016-11-09 12:36:15 -0700170 }
171}
172
173/*
174 * Called on completion of a request. Note that it's also called when
175 * a request is merged, when the request gets freed.
176 */
Josef Bacika7905042018-07-03 09:32:35 -0600177static void wbt_done(struct rq_qos *rqos, struct request *rq)
Jens Axboee34cbd32016-11-09 12:36:15 -0700178{
Josef Bacika7905042018-07-03 09:32:35 -0600179 struct rq_wb *rwb = RQWB(rqos);
Jens Axboee34cbd32016-11-09 12:36:15 -0700180
Omar Sandovala8a45942018-05-09 02:08:48 -0700181 if (!wbt_is_tracked(rq)) {
182 if (rwb->sync_cookie == rq) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700183 rwb->sync_issue = 0;
184 rwb->sync_cookie = NULL;
185 }
186
Omar Sandovala8a45942018-05-09 02:08:48 -0700187 if (wbt_is_read(rq))
Jens Axboee34cbd32016-11-09 12:36:15 -0700188 wb_timestamp(rwb, &rwb->last_comp);
Jens Axboee34cbd32016-11-09 12:36:15 -0700189 } else {
Omar Sandovala8a45942018-05-09 02:08:48 -0700190 WARN_ON_ONCE(rq == rwb->sync_cookie);
Josef Bacika7905042018-07-03 09:32:35 -0600191 __wbt_done(rqos, wbt_flags(rq));
Jens Axboee34cbd32016-11-09 12:36:15 -0700192 }
Omar Sandovala8a45942018-05-09 02:08:48 -0700193 wbt_clear_state(rq);
Jens Axboee34cbd32016-11-09 12:36:15 -0700194}
195
Arnd Bergmann4121d382016-11-16 16:29:57 +0100196static inline bool stat_sample_valid(struct blk_rq_stat *stat)
Jens Axboee34cbd32016-11-09 12:36:15 -0700197{
198 /*
199 * We need at least one read sample, and a minimum of
200 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
201 * that it's writes impacting us, and not just some sole read on
202 * a device that is in a lower power state.
203 */
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700204 return (stat[READ].nr_samples >= 1 &&
205 stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
Jens Axboee34cbd32016-11-09 12:36:15 -0700206}
207
208static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
209{
Mark Rutland6aa7de02017-10-23 14:07:29 -0700210 u64 now, issue = READ_ONCE(rwb->sync_issue);
Jens Axboee34cbd32016-11-09 12:36:15 -0700211
212 if (!issue || !rwb->sync_cookie)
213 return 0;
214
215 now = ktime_to_ns(ktime_get());
216 return now - issue;
217}
218
219enum {
220 LAT_OK = 1,
221 LAT_UNKNOWN,
222 LAT_UNKNOWN_WRITES,
223 LAT_EXCEEDED,
224};
225
Omar Sandoval34dbad52017-03-21 08:56:08 -0700226static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
Jens Axboee34cbd32016-11-09 12:36:15 -0700227{
Josef Bacika7905042018-07-03 09:32:35 -0600228 struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
229 struct rq_depth *rqd = &rwb->rq_depth;
Jens Axboee34cbd32016-11-09 12:36:15 -0700230 u64 thislat;
231
232 /*
233 * If our stored sync issue exceeds the window size, or it
234 * exceeds our min target AND we haven't logged any entries,
235 * flag the latency as exceeded. wbt works off completion latencies,
236 * but for a flooded device, a single sync IO can take a long time
237 * to complete after being issued. If this time exceeds our
238 * monitoring window AND we didn't see any other completions in that
239 * window, then count that sync IO as a violation of the latency.
240 */
241 thislat = rwb_sync_issue_lat(rwb);
242 if (thislat > rwb->cur_win_nsec ||
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700243 (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700244 trace_wbt_lat(bdi, thislat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700245 return LAT_EXCEEDED;
246 }
247
248 /*
249 * No read/write mix, if stat isn't valid
250 */
251 if (!stat_sample_valid(stat)) {
252 /*
253 * If we had writes in this stat window and the window is
254 * current, we're only doing writes. If a task recently
255 * waited or still has writes in flights, consider us doing
256 * just writes as well.
257 */
Omar Sandoval34dbad52017-03-21 08:56:08 -0700258 if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
259 wbt_inflight(rwb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700260 return LAT_UNKNOWN_WRITES;
261 return LAT_UNKNOWN;
262 }
263
264 /*
265 * If the 'min' latency exceeds our target, step down.
266 */
Omar Sandovalfa2e39c2017-03-21 08:56:06 -0700267 if (stat[READ].min > rwb->min_lat_nsec) {
268 trace_wbt_lat(bdi, stat[READ].min);
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700269 trace_wbt_stat(bdi, stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700270 return LAT_EXCEEDED;
271 }
272
Josef Bacika7905042018-07-03 09:32:35 -0600273 if (rqd->scale_step)
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700274 trace_wbt_stat(bdi, stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700275
276 return LAT_OK;
277}
278
Jens Axboee34cbd32016-11-09 12:36:15 -0700279static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
280{
Josef Bacika7905042018-07-03 09:32:35 -0600281 struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info;
282 struct rq_depth *rqd = &rwb->rq_depth;
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700283
Josef Bacika7905042018-07-03 09:32:35 -0600284 trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
285 rwb->wb_background, rwb->wb_normal, rqd->max_depth);
286}
287
288static void calc_wb_limits(struct rq_wb *rwb)
289{
290 if (rwb->min_lat_nsec == 0) {
291 rwb->wb_normal = rwb->wb_background = 0;
292 } else if (rwb->rq_depth.max_depth <= 2) {
293 rwb->wb_normal = rwb->rq_depth.max_depth;
294 rwb->wb_background = 1;
295 } else {
296 rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
297 rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
298 }
Jens Axboee34cbd32016-11-09 12:36:15 -0700299}
300
301static void scale_up(struct rq_wb *rwb)
302{
Josef Bacika7905042018-07-03 09:32:35 -0600303 rq_depth_scale_up(&rwb->rq_depth);
304 calc_wb_limits(rwb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700305 rwb->unknown_cnt = 0;
Josef Bacika7905042018-07-03 09:32:35 -0600306 rwb_trace_step(rwb, "scale up");
Jens Axboee34cbd32016-11-09 12:36:15 -0700307}
308
Jens Axboee34cbd32016-11-09 12:36:15 -0700309static void scale_down(struct rq_wb *rwb, bool hard_throttle)
310{
Josef Bacika7905042018-07-03 09:32:35 -0600311 rq_depth_scale_down(&rwb->rq_depth, hard_throttle);
Jens Axboee34cbd32016-11-09 12:36:15 -0700312 calc_wb_limits(rwb);
Josef Bacika7905042018-07-03 09:32:35 -0600313 rwb->unknown_cnt = 0;
314 rwb_wake_all(rwb);
315 rwb_trace_step(rwb, "scale down");
Jens Axboee34cbd32016-11-09 12:36:15 -0700316}
317
318static void rwb_arm_timer(struct rq_wb *rwb)
319{
Josef Bacika7905042018-07-03 09:32:35 -0600320 struct rq_depth *rqd = &rwb->rq_depth;
321
322 if (rqd->scale_step > 0) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700323 /*
324 * We should speed this up, using some variant of a fast
325 * integer inverse square root calculation. Since we only do
326 * this for every window expiration, it's not a huge deal,
327 * though.
328 */
329 rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
Josef Bacika7905042018-07-03 09:32:35 -0600330 int_sqrt((rqd->scale_step + 1) << 8));
Jens Axboee34cbd32016-11-09 12:36:15 -0700331 } else {
332 /*
333 * For step < 0, we don't want to increase/decrease the
334 * window size.
335 */
336 rwb->cur_win_nsec = rwb->win_nsec;
337 }
338
Omar Sandoval34dbad52017-03-21 08:56:08 -0700339 blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
Jens Axboee34cbd32016-11-09 12:36:15 -0700340}
341
Omar Sandoval34dbad52017-03-21 08:56:08 -0700342static void wb_timer_fn(struct blk_stat_callback *cb)
Jens Axboee34cbd32016-11-09 12:36:15 -0700343{
Omar Sandoval34dbad52017-03-21 08:56:08 -0700344 struct rq_wb *rwb = cb->data;
Josef Bacika7905042018-07-03 09:32:35 -0600345 struct rq_depth *rqd = &rwb->rq_depth;
Jens Axboee34cbd32016-11-09 12:36:15 -0700346 unsigned int inflight = wbt_inflight(rwb);
347 int status;
348
Omar Sandoval34dbad52017-03-21 08:56:08 -0700349 status = latency_exceeded(rwb, cb->stat);
Jens Axboee34cbd32016-11-09 12:36:15 -0700350
Josef Bacika7905042018-07-03 09:32:35 -0600351 trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step,
Jens Axboed8a0cbf2016-11-10 21:52:53 -0700352 inflight);
Jens Axboee34cbd32016-11-09 12:36:15 -0700353
354 /*
355 * If we exceeded the latency target, step down. If we did not,
356 * step one level up. If we don't know enough to say either exceeded
357 * or ok, then don't do anything.
358 */
359 switch (status) {
360 case LAT_EXCEEDED:
361 scale_down(rwb, true);
362 break;
363 case LAT_OK:
364 scale_up(rwb);
365 break;
366 case LAT_UNKNOWN_WRITES:
367 /*
368 * We started a the center step, but don't have a valid
369 * read/write sample, but we do have writes going on.
370 * Allow step to go negative, to increase write perf.
371 */
372 scale_up(rwb);
373 break;
374 case LAT_UNKNOWN:
375 if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
376 break;
377 /*
378 * We get here when previously scaled reduced depth, and we
379 * currently don't have a valid read/write sample. For that
380 * case, slowly return to center state (step == 0).
381 */
Josef Bacika7905042018-07-03 09:32:35 -0600382 if (rqd->scale_step > 0)
Jens Axboee34cbd32016-11-09 12:36:15 -0700383 scale_up(rwb);
Josef Bacika7905042018-07-03 09:32:35 -0600384 else if (rqd->scale_step < 0)
Jens Axboee34cbd32016-11-09 12:36:15 -0700385 scale_down(rwb, false);
386 break;
387 default:
388 break;
389 }
390
391 /*
392 * Re-arm timer, if we have IO in flight
393 */
Josef Bacika7905042018-07-03 09:32:35 -0600394 if (rqd->scale_step || inflight)
Jens Axboee34cbd32016-11-09 12:36:15 -0700395 rwb_arm_timer(rwb);
396}
397
Josef Bacika7905042018-07-03 09:32:35 -0600398static void __wbt_update_limits(struct rq_wb *rwb)
Jens Axboee34cbd32016-11-09 12:36:15 -0700399{
Josef Bacika7905042018-07-03 09:32:35 -0600400 struct rq_depth *rqd = &rwb->rq_depth;
401
402 rqd->scale_step = 0;
403 rqd->scaled_max = false;
404
405 rq_depth_calc_max_depth(rqd);
Jens Axboee34cbd32016-11-09 12:36:15 -0700406 calc_wb_limits(rwb);
407
408 rwb_wake_all(rwb);
409}
410
Josef Bacika7905042018-07-03 09:32:35 -0600411void wbt_update_limits(struct request_queue *q)
412{
413 struct rq_qos *rqos = wbt_rq_qos(q);
414 if (!rqos)
415 return;
416 __wbt_update_limits(RQWB(rqos));
417}
418
419u64 wbt_get_min_lat(struct request_queue *q)
420{
421 struct rq_qos *rqos = wbt_rq_qos(q);
422 if (!rqos)
423 return 0;
424 return RQWB(rqos)->min_lat_nsec;
425}
426
427void wbt_set_min_lat(struct request_queue *q, u64 val)
428{
429 struct rq_qos *rqos = wbt_rq_qos(q);
430 if (!rqos)
431 return;
432 RQWB(rqos)->min_lat_nsec = val;
433 RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
434 __wbt_update_limits(RQWB(rqos));
435}
436
437
Jens Axboee34cbd32016-11-09 12:36:15 -0700438static bool close_io(struct rq_wb *rwb)
439{
440 const unsigned long now = jiffies;
441
442 return time_before(now, rwb->last_issue + HZ / 10) ||
443 time_before(now, rwb->last_comp + HZ / 10);
444}
445
446#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO)
447
448static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
449{
450 unsigned int limit;
451
Jens Axboeffa358d2018-08-20 13:24:25 -0600452 /*
453 * If we got disabled, just return UINT_MAX. This ensures that
454 * we'll properly inc a new IO, and dec+wakeup at the end.
455 */
456 if (!rwb_enabled(rwb))
457 return UINT_MAX;
458
Jens Axboe782f5692018-05-07 10:03:23 -0600459 if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD)
460 return rwb->wb_background;
461
Jens Axboee34cbd32016-11-09 12:36:15 -0700462 /*
463 * At this point we know it's a buffered write. If this is
weiping zhang3dfbdc42017-11-23 21:40:10 +0800464 * kswapd trying to free memory, or REQ_SYNC is set, then
Jens Axboee34cbd32016-11-09 12:36:15 -0700465 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
466 * that. If the write is marked as a background write, then use
467 * the idle limit, or go to normal if we haven't had competing
468 * IO for a bit.
469 */
470 if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
Josef Bacika7905042018-07-03 09:32:35 -0600471 limit = rwb->rq_depth.max_depth;
Jens Axboee34cbd32016-11-09 12:36:15 -0700472 else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
473 /*
474 * If less than 100ms since we completed unrelated IO,
475 * limit us to half the depth for background writeback.
476 */
477 limit = rwb->wb_background;
478 } else
479 limit = rwb->wb_normal;
480
481 return limit;
482}
483
Jens Axboee34cbd32016-11-09 12:36:15 -0700484/*
485 * Block if we will exceed our limit, or if we are currently waiting for
486 * the timer to kick off queuing again.
487 */
Jens Axboe8bea6092018-05-07 09:57:08 -0600488static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
489 unsigned long rw, spinlock_t *lock)
Bart Van Assche9eca5352017-01-02 09:48:47 -0700490 __releases(lock)
491 __acquires(lock)
Jens Axboee34cbd32016-11-09 12:36:15 -0700492{
Jens Axboe8bea6092018-05-07 09:57:08 -0600493 struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
Anchal Agarwal2887e412018-08-07 14:40:49 -0600494 DECLARE_WAITQUEUE(wait, current);
Jens Axboec45e6a02018-08-20 13:22:27 -0600495 bool has_sleeper;
Jens Axboee34cbd32016-11-09 12:36:15 -0700496
Jens Axboec45e6a02018-08-20 13:22:27 -0600497 has_sleeper = wq_has_sleeper(&rqw->wait);
498 if (!has_sleeper && rq_wait_inc_below(rqw, get_limit(rwb, rw)))
Jens Axboee34cbd32016-11-09 12:36:15 -0700499 return;
500
Anchal Agarwal2887e412018-08-07 14:40:49 -0600501 add_wait_queue_exclusive(&rqw->wait, &wait);
Jens Axboee34cbd32016-11-09 12:36:15 -0700502 do {
Anchal Agarwal2887e412018-08-07 14:40:49 -0600503 set_current_state(TASK_UNINTERRUPTIBLE);
Jens Axboee34cbd32016-11-09 12:36:15 -0700504
Jens Axboec45e6a02018-08-20 13:22:27 -0600505 if (!has_sleeper && rq_wait_inc_below(rqw, get_limit(rwb, rw)))
Jens Axboee34cbd32016-11-09 12:36:15 -0700506 break;
507
Bart Van Assche9eca5352017-01-02 09:48:47 -0700508 if (lock) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700509 spin_unlock_irq(lock);
Bart Van Assche9eca5352017-01-02 09:48:47 -0700510 io_schedule();
Jens Axboee34cbd32016-11-09 12:36:15 -0700511 spin_lock_irq(lock);
Bart Van Assche9eca5352017-01-02 09:48:47 -0700512 } else
513 io_schedule();
Jens Axboec45e6a02018-08-20 13:22:27 -0600514 has_sleeper = false;
Jens Axboee34cbd32016-11-09 12:36:15 -0700515 } while (1);
516
Anchal Agarwal2887e412018-08-07 14:40:49 -0600517 __set_current_state(TASK_RUNNING);
518 remove_wait_queue(&rqw->wait, &wait);
Jens Axboee34cbd32016-11-09 12:36:15 -0700519}
520
521static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
522{
Jens Axboe782f5692018-05-07 10:03:23 -0600523 switch (bio_op(bio)) {
524 case REQ_OP_WRITE:
525 /*
526 * Don't throttle WRITE_ODIRECT
527 */
528 if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
529 (REQ_SYNC | REQ_IDLE))
530 return false;
531 /* fallthrough */
532 case REQ_OP_DISCARD:
533 return true;
534 default:
Jens Axboee34cbd32016-11-09 12:36:15 -0700535 return false;
Jens Axboe782f5692018-05-07 10:03:23 -0600536 }
Jens Axboee34cbd32016-11-09 12:36:15 -0700537}
538
Josef Bacikc1c80382018-07-03 11:14:59 -0400539static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
540{
541 enum wbt_flags flags = 0;
542
Jens Axboec1253112018-08-23 09:34:46 -0600543 if (!rwb_enabled(rwb))
544 return 0;
545
Josef Bacikc1c80382018-07-03 11:14:59 -0400546 if (bio_op(bio) == REQ_OP_READ) {
547 flags = WBT_READ;
548 } else if (wbt_should_throttle(rwb, bio)) {
549 if (current_is_kswapd())
550 flags |= WBT_KSWAPD;
551 if (bio_op(bio) == REQ_OP_DISCARD)
552 flags |= WBT_DISCARD;
553 flags |= WBT_TRACKED;
554 }
555 return flags;
556}
557
558static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
559{
560 struct rq_wb *rwb = RQWB(rqos);
561 enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
562 __wbt_done(rqos, flags);
563}
564
Jens Axboee34cbd32016-11-09 12:36:15 -0700565/*
566 * Returns true if the IO request should be accounted, false if not.
567 * May sleep, if we have exceeded the writeback limits. Caller can pass
568 * in an irq held spinlock, if it holds one when calling this function.
569 * If we do sleep, we'll release and re-grab it.
570 */
Josef Bacikc1c80382018-07-03 11:14:59 -0400571static void wbt_wait(struct rq_qos *rqos, struct bio *bio, spinlock_t *lock)
Jens Axboee34cbd32016-11-09 12:36:15 -0700572{
Josef Bacika7905042018-07-03 09:32:35 -0600573 struct rq_wb *rwb = RQWB(rqos);
Josef Bacikc1c80382018-07-03 11:14:59 -0400574 enum wbt_flags flags;
Jens Axboee34cbd32016-11-09 12:36:15 -0700575
Josef Bacikc1c80382018-07-03 11:14:59 -0400576 flags = bio_to_wbt_flags(rwb, bio);
Ming Leidf60f6e2018-08-14 23:57:49 +0800577 if (!(flags & WBT_TRACKED)) {
Josef Bacikc1c80382018-07-03 11:14:59 -0400578 if (flags & WBT_READ)
Jens Axboee34cbd32016-11-09 12:36:15 -0700579 wb_timestamp(rwb, &rwb->last_issue);
Josef Bacikc1c80382018-07-03 11:14:59 -0400580 return;
Jens Axboee34cbd32016-11-09 12:36:15 -0700581 }
582
Jens Axboe8bea6092018-05-07 09:57:08 -0600583 if (current_is_kswapd())
Josef Bacikc1c80382018-07-03 11:14:59 -0400584 flags |= WBT_KSWAPD;
Jens Axboe782f5692018-05-07 10:03:23 -0600585 if (bio_op(bio) == REQ_OP_DISCARD)
Josef Bacikc1c80382018-07-03 11:14:59 -0400586 flags |= WBT_DISCARD;
Jens Axboe8bea6092018-05-07 09:57:08 -0600587
Josef Bacikc1c80382018-07-03 11:14:59 -0400588 __wbt_wait(rwb, flags, bio->bi_opf, lock);
Jens Axboee34cbd32016-11-09 12:36:15 -0700589
Omar Sandoval34dbad52017-03-21 08:56:08 -0700590 if (!blk_stat_is_active(rwb->cb))
Jens Axboee34cbd32016-11-09 12:36:15 -0700591 rwb_arm_timer(rwb);
Josef Bacikc1c80382018-07-03 11:14:59 -0400592}
Jens Axboee34cbd32016-11-09 12:36:15 -0700593
Josef Bacikc1c80382018-07-03 11:14:59 -0400594static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
595{
596 struct rq_wb *rwb = RQWB(rqos);
597 rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
Jens Axboee34cbd32016-11-09 12:36:15 -0700598}
599
Josef Bacika7905042018-07-03 09:32:35 -0600600void wbt_issue(struct rq_qos *rqos, struct request *rq)
Jens Axboee34cbd32016-11-09 12:36:15 -0700601{
Josef Bacika7905042018-07-03 09:32:35 -0600602 struct rq_wb *rwb = RQWB(rqos);
603
Jens Axboee34cbd32016-11-09 12:36:15 -0700604 if (!rwb_enabled(rwb))
605 return;
606
607 /*
Omar Sandovala8a45942018-05-09 02:08:48 -0700608 * Track sync issue, in case it takes a long time to complete. Allows us
609 * to react quicker, if a sync IO takes a long time to complete. Note
610 * that this is just a hint. The request can go away when it completes,
611 * so it's important we never dereference it. We only use the address to
612 * compare with, which is why we store the sync_issue time locally.
Jens Axboee34cbd32016-11-09 12:36:15 -0700613 */
Omar Sandovala8a45942018-05-09 02:08:48 -0700614 if (wbt_is_read(rq) && !rwb->sync_issue) {
615 rwb->sync_cookie = rq;
Omar Sandoval544ccc8d2018-05-09 02:08:50 -0700616 rwb->sync_issue = rq->io_start_time_ns;
Jens Axboee34cbd32016-11-09 12:36:15 -0700617 }
618}
619
Josef Bacika7905042018-07-03 09:32:35 -0600620void wbt_requeue(struct rq_qos *rqos, struct request *rq)
Jens Axboee34cbd32016-11-09 12:36:15 -0700621{
Josef Bacika7905042018-07-03 09:32:35 -0600622 struct rq_wb *rwb = RQWB(rqos);
Jens Axboee34cbd32016-11-09 12:36:15 -0700623 if (!rwb_enabled(rwb))
624 return;
Omar Sandovala8a45942018-05-09 02:08:48 -0700625 if (rq == rwb->sync_cookie) {
Jens Axboee34cbd32016-11-09 12:36:15 -0700626 rwb->sync_issue = 0;
627 rwb->sync_cookie = NULL;
628 }
629}
630
Josef Bacika7905042018-07-03 09:32:35 -0600631void wbt_set_queue_depth(struct request_queue *q, unsigned int depth)
Jens Axboee34cbd32016-11-09 12:36:15 -0700632{
Josef Bacika7905042018-07-03 09:32:35 -0600633 struct rq_qos *rqos = wbt_rq_qos(q);
634 if (rqos) {
635 RQWB(rqos)->rq_depth.queue_depth = depth;
636 __wbt_update_limits(RQWB(rqos));
Jens Axboee34cbd32016-11-09 12:36:15 -0700637 }
638}
639
Josef Bacika7905042018-07-03 09:32:35 -0600640void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
Jens Axboee34cbd32016-11-09 12:36:15 -0700641{
Josef Bacika7905042018-07-03 09:32:35 -0600642 struct rq_qos *rqos = wbt_rq_qos(q);
643 if (rqos)
644 RQWB(rqos)->wc = write_cache_on;
Jens Axboee34cbd32016-11-09 12:36:15 -0700645}
646
Jan Kara3f19cd22017-04-11 11:29:01 +0200647/*
Jan Kara8330cdb2017-04-19 11:33:27 +0200648 * Enable wbt if defaults are configured that way
649 */
650void wbt_enable_default(struct request_queue *q)
651{
Josef Bacika7905042018-07-03 09:32:35 -0600652 struct rq_qos *rqos = wbt_rq_qos(q);
Jan Kara8330cdb2017-04-19 11:33:27 +0200653 /* Throttling already enabled? */
Josef Bacika7905042018-07-03 09:32:35 -0600654 if (rqos)
Jan Kara8330cdb2017-04-19 11:33:27 +0200655 return;
656
657 /* Queue not registered? Maybe shutting down... */
658 if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
659 return;
660
661 if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) ||
662 (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ)))
663 wbt_init(q);
664}
665EXPORT_SYMBOL_GPL(wbt_enable_default);
666
Jens Axboe80e091d2016-11-28 09:22:47 -0700667u64 wbt_default_latency_nsec(struct request_queue *q)
668{
669 /*
670 * We default to 2msec for non-rotational storage, and 75msec
671 * for rotational storage.
672 */
673 if (blk_queue_nonrot(q))
674 return 2000000ULL;
675 else
676 return 75000000ULL;
677}
678
Jens Axboe99c749a2017-04-21 07:55:42 -0600679static int wbt_data_dir(const struct request *rq)
680{
Jens Axboe52355532018-02-05 13:16:56 -0700681 const int op = req_op(rq);
682
683 if (op == REQ_OP_READ)
684 return READ;
Jens Axboe825843b2018-05-03 09:14:57 -0600685 else if (op_is_write(op))
Jens Axboe52355532018-02-05 13:16:56 -0700686 return WRITE;
687
688 /* don't account */
689 return -1;
Jens Axboe99c749a2017-04-21 07:55:42 -0600690}
691
Josef Bacika7905042018-07-03 09:32:35 -0600692static void wbt_exit(struct rq_qos *rqos)
693{
694 struct rq_wb *rwb = RQWB(rqos);
695 struct request_queue *q = rqos->q;
696
697 blk_stat_remove_callback(q, rwb->cb);
698 blk_stat_free_callback(rwb->cb);
699 kfree(rwb);
700}
701
702/*
703 * Disable wbt, if enabled by default.
704 */
705void wbt_disable_default(struct request_queue *q)
706{
707 struct rq_qos *rqos = wbt_rq_qos(q);
708 struct rq_wb *rwb;
709 if (!rqos)
710 return;
711 rwb = RQWB(rqos);
712 if (rwb->enable_state == WBT_STATE_ON_DEFAULT)
713 rwb->wb_normal = 0;
714}
715EXPORT_SYMBOL_GPL(wbt_disable_default);
716
717
718static struct rq_qos_ops wbt_rqos_ops = {
719 .throttle = wbt_wait,
720 .issue = wbt_issue,
Josef Bacikc1c80382018-07-03 11:14:59 -0400721 .track = wbt_track,
Josef Bacika7905042018-07-03 09:32:35 -0600722 .requeue = wbt_requeue,
723 .done = wbt_done,
Josef Bacikc1c80382018-07-03 11:14:59 -0400724 .cleanup = wbt_cleanup,
Josef Bacika7905042018-07-03 09:32:35 -0600725 .exit = wbt_exit,
726};
727
Jens Axboe8054b892016-11-10 21:50:51 -0700728int wbt_init(struct request_queue *q)
Jens Axboee34cbd32016-11-09 12:36:15 -0700729{
730 struct rq_wb *rwb;
731 int i;
732
Jens Axboee34cbd32016-11-09 12:36:15 -0700733 rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
734 if (!rwb)
735 return -ENOMEM;
736
Jens Axboe99c749a2017-04-21 07:55:42 -0600737 rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
Omar Sandoval34dbad52017-03-21 08:56:08 -0700738 if (!rwb->cb) {
739 kfree(rwb);
740 return -ENOMEM;
741 }
742
Josef Bacika7905042018-07-03 09:32:35 -0600743 for (i = 0; i < WBT_NUM_RWQ; i++)
744 rq_wait_init(&rwb->rq_wait[i]);
Jens Axboee34cbd32016-11-09 12:36:15 -0700745
Josef Bacika7905042018-07-03 09:32:35 -0600746 rwb->rqos.id = RQ_QOS_WBT;
747 rwb->rqos.ops = &wbt_rqos_ops;
748 rwb->rqos.q = q;
Jens Axboee34cbd32016-11-09 12:36:15 -0700749 rwb->last_comp = rwb->last_issue = jiffies;
Jens Axboee34cbd32016-11-09 12:36:15 -0700750 rwb->win_nsec = RWB_WINDOW_NSEC;
Jens Axboed62118b2016-11-28 09:40:34 -0700751 rwb->enable_state = WBT_STATE_ON_DEFAULT;
Josef Bacika7905042018-07-03 09:32:35 -0600752 rwb->wc = 1;
753 rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
754 __wbt_update_limits(rwb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700755
756 /*
Omar Sandoval34dbad52017-03-21 08:56:08 -0700757 * Assign rwb and add the stats callback.
Jens Axboee34cbd32016-11-09 12:36:15 -0700758 */
Josef Bacika7905042018-07-03 09:32:35 -0600759 rq_qos_add(q, &rwb->rqos);
Omar Sandoval34dbad52017-03-21 08:56:08 -0700760 blk_stat_add_callback(q, rwb->cb);
Jens Axboee34cbd32016-11-09 12:36:15 -0700761
Jens Axboe80e091d2016-11-28 09:22:47 -0700762 rwb->min_lat_nsec = wbt_default_latency_nsec(q);
Jens Axboee34cbd32016-11-09 12:36:15 -0700763
Josef Bacika7905042018-07-03 09:32:35 -0600764 wbt_set_queue_depth(q, blk_queue_depth(q));
765 wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
Jens Axboee34cbd32016-11-09 12:36:15 -0700766
767 return 0;
768}