Commit | Line | Data |
---|---|---|
e34cbd30 JA |
1 | /* |
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" | |
28 | ||
29 | #define CREATE_TRACE_POINTS | |
30 | #include <trace/events/wbt.h> | |
31 | ||
32 | enum { | |
33 | /* | |
34 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
35 | * from here depending on device stats | |
36 | */ | |
37 | RWB_DEF_DEPTH = 16, | |
38 | ||
39 | /* | |
40 | * 100msec window | |
41 | */ | |
42 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
43 | ||
44 | /* | |
45 | * Disregard stats, if we don't meet this minimum | |
46 | */ | |
47 | RWB_MIN_WRITE_SAMPLES = 3, | |
48 | ||
49 | /* | |
50 | * If we have this number of consecutive windows with not enough | |
51 | * information to scale up or down, scale up. | |
52 | */ | |
53 | RWB_UNKNOWN_BUMP = 5, | |
54 | }; | |
55 | ||
56 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
57 | { | |
58 | return rwb && rwb->wb_normal != 0; | |
59 | } | |
60 | ||
61 | /* | |
62 | * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded, | |
63 | * false if 'v' + 1 would be bigger than 'below'. | |
64 | */ | |
65 | static bool atomic_inc_below(atomic_t *v, int below) | |
66 | { | |
67 | int cur = atomic_read(v); | |
68 | ||
69 | for (;;) { | |
70 | int old; | |
71 | ||
72 | if (cur >= below) | |
73 | return false; | |
74 | old = atomic_cmpxchg(v, cur, cur + 1); | |
75 | if (old == cur) | |
76 | break; | |
77 | cur = old; | |
78 | } | |
79 | ||
80 | return true; | |
81 | } | |
82 | ||
83 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) | |
84 | { | |
85 | if (rwb_enabled(rwb)) { | |
86 | const unsigned long cur = jiffies; | |
87 | ||
88 | if (cur != *var) | |
89 | *var = cur; | |
90 | } | |
91 | } | |
92 | ||
93 | /* | |
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. | |
96 | */ | |
97 | static bool wb_recent_wait(struct rq_wb *rwb) | |
98 | { | |
dc3b17cc | 99 | struct bdi_writeback *wb = &rwb->queue->backing_dev_info->wb; |
e34cbd30 JA |
100 | |
101 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
102 | } | |
103 | ||
8bea6090 JA |
104 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
105 | enum wbt_flags wb_acct) | |
e34cbd30 | 106 | { |
8bea6090 JA |
107 | if (wb_acct & WBT_KSWAPD) |
108 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
109 | else if (wb_acct & WBT_DISCARD) |
110 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
111 | |
112 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
113 | } |
114 | ||
115 | static void rwb_wake_all(struct rq_wb *rwb) | |
116 | { | |
117 | int i; | |
118 | ||
119 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
120 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
121 | ||
122 | if (waitqueue_active(&rqw->wait)) | |
123 | wake_up_all(&rqw->wait); | |
124 | } | |
125 | } | |
126 | ||
127 | void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct) | |
128 | { | |
129 | struct rq_wait *rqw; | |
130 | int inflight, limit; | |
131 | ||
132 | if (!(wb_acct & WBT_TRACKED)) | |
133 | return; | |
134 | ||
8bea6090 | 135 | rqw = get_rq_wait(rwb, wb_acct); |
e34cbd30 JA |
136 | 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 | /* | |
782f5697 JA |
148 | * 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. | |
e34cbd30 | 151 | */ |
782f5697 JA |
152 | if (wb_acct & WBT_DISCARD) |
153 | limit = rwb->wb_background; | |
154 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
155 | 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 | ||
165 | if (waitqueue_active(&rqw->wait)) { | |
166 | int diff = limit - inflight; | |
167 | ||
168 | if (!inflight || diff >= rwb->wb_background / 2) | |
169 | wake_up_all(&rqw->wait); | |
170 | } | |
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 | */ | |
177 | void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
178 | { | |
179 | if (!rwb) | |
180 | return; | |
181 | ||
182 | if (!wbt_is_tracked(stat)) { | |
183 | if (rwb->sync_cookie == stat) { | |
184 | rwb->sync_issue = 0; | |
185 | rwb->sync_cookie = NULL; | |
186 | } | |
187 | ||
188 | if (wbt_is_read(stat)) | |
189 | wb_timestamp(rwb, &rwb->last_comp); | |
e34cbd30 JA |
190 | } else { |
191 | WARN_ON_ONCE(stat == rwb->sync_cookie); | |
192 | __wbt_done(rwb, wbt_stat_to_mask(stat)); | |
e34cbd30 | 193 | } |
62d772fa | 194 | wbt_clear_state(stat); |
e34cbd30 JA |
195 | } |
196 | ||
197 | /* | |
198 | * Return true, if we can't increase the depth further by scaling | |
199 | */ | |
200 | static bool calc_wb_limits(struct rq_wb *rwb) | |
201 | { | |
202 | unsigned int depth; | |
203 | bool ret = false; | |
204 | ||
205 | if (!rwb->min_lat_nsec) { | |
206 | rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; | |
207 | return false; | |
208 | } | |
209 | ||
210 | /* | |
211 | * For QD=1 devices, this is a special case. It's important for those | |
212 | * to have one request ready when one completes, so force a depth of | |
213 | * 2 for those devices. On the backend, it'll be a depth of 1 anyway, | |
214 | * since the device can't have more than that in flight. If we're | |
215 | * scaling down, then keep a setting of 1/1/1. | |
216 | */ | |
217 | if (rwb->queue_depth == 1) { | |
218 | if (rwb->scale_step > 0) | |
219 | rwb->wb_max = rwb->wb_normal = 1; | |
220 | else { | |
221 | rwb->wb_max = rwb->wb_normal = 2; | |
222 | ret = true; | |
223 | } | |
224 | rwb->wb_background = 1; | |
225 | } else { | |
226 | /* | |
227 | * scale_step == 0 is our default state. If we have suffered | |
228 | * latency spikes, step will be > 0, and we shrink the | |
229 | * allowed write depths. If step is < 0, we're only doing | |
230 | * writes, and we allow a temporarily higher depth to | |
231 | * increase performance. | |
232 | */ | |
233 | depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth); | |
234 | if (rwb->scale_step > 0) | |
235 | depth = 1 + ((depth - 1) >> min(31, rwb->scale_step)); | |
236 | else if (rwb->scale_step < 0) { | |
237 | unsigned int maxd = 3 * rwb->queue_depth / 4; | |
238 | ||
239 | depth = 1 + ((depth - 1) << -rwb->scale_step); | |
240 | if (depth > maxd) { | |
241 | depth = maxd; | |
242 | ret = true; | |
243 | } | |
244 | } | |
245 | ||
246 | /* | |
247 | * Set our max/normal/bg queue depths based on how far | |
248 | * we have scaled down (->scale_step). | |
249 | */ | |
250 | rwb->wb_max = depth; | |
251 | rwb->wb_normal = (rwb->wb_max + 1) / 2; | |
252 | rwb->wb_background = (rwb->wb_max + 3) / 4; | |
253 | } | |
254 | ||
255 | return ret; | |
256 | } | |
257 | ||
4121d385 | 258 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
259 | { |
260 | /* | |
261 | * We need at least one read sample, and a minimum of | |
262 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
263 | * that it's writes impacting us, and not just some sole read on | |
264 | * a device that is in a lower power state. | |
265 | */ | |
fa2e39cb OS |
266 | return (stat[READ].nr_samples >= 1 && |
267 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
268 | } |
269 | ||
270 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
271 | { | |
6aa7de05 | 272 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
273 | |
274 | if (!issue || !rwb->sync_cookie) | |
275 | return 0; | |
276 | ||
277 | now = ktime_to_ns(ktime_get()); | |
278 | return now - issue; | |
279 | } | |
280 | ||
281 | enum { | |
282 | LAT_OK = 1, | |
283 | LAT_UNKNOWN, | |
284 | LAT_UNKNOWN_WRITES, | |
285 | LAT_EXCEEDED, | |
286 | }; | |
287 | ||
34dbad5d | 288 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 289 | { |
dc3b17cc | 290 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; |
e34cbd30 JA |
291 | u64 thislat; |
292 | ||
293 | /* | |
294 | * If our stored sync issue exceeds the window size, or it | |
295 | * exceeds our min target AND we haven't logged any entries, | |
296 | * flag the latency as exceeded. wbt works off completion latencies, | |
297 | * but for a flooded device, a single sync IO can take a long time | |
298 | * to complete after being issued. If this time exceeds our | |
299 | * monitoring window AND we didn't see any other completions in that | |
300 | * window, then count that sync IO as a violation of the latency. | |
301 | */ | |
302 | thislat = rwb_sync_issue_lat(rwb); | |
303 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 304 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 305 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
306 | return LAT_EXCEEDED; |
307 | } | |
308 | ||
309 | /* | |
310 | * No read/write mix, if stat isn't valid | |
311 | */ | |
312 | if (!stat_sample_valid(stat)) { | |
313 | /* | |
314 | * If we had writes in this stat window and the window is | |
315 | * current, we're only doing writes. If a task recently | |
316 | * waited or still has writes in flights, consider us doing | |
317 | * just writes as well. | |
318 | */ | |
34dbad5d OS |
319 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
320 | wbt_inflight(rwb)) | |
e34cbd30 JA |
321 | return LAT_UNKNOWN_WRITES; |
322 | return LAT_UNKNOWN; | |
323 | } | |
324 | ||
325 | /* | |
326 | * If the 'min' latency exceeds our target, step down. | |
327 | */ | |
fa2e39cb OS |
328 | if (stat[READ].min > rwb->min_lat_nsec) { |
329 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 330 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
331 | return LAT_EXCEEDED; |
332 | } | |
333 | ||
334 | if (rwb->scale_step) | |
d8a0cbfd | 335 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
336 | |
337 | return LAT_OK; | |
338 | } | |
339 | ||
e34cbd30 JA |
340 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
341 | { | |
dc3b17cc | 342 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; |
d8a0cbfd JA |
343 | |
344 | trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec, | |
e34cbd30 JA |
345 | rwb->wb_background, rwb->wb_normal, rwb->wb_max); |
346 | } | |
347 | ||
348 | static void scale_up(struct rq_wb *rwb) | |
349 | { | |
350 | /* | |
351 | * Hit max in previous round, stop here | |
352 | */ | |
353 | if (rwb->scaled_max) | |
354 | return; | |
355 | ||
356 | rwb->scale_step--; | |
357 | rwb->unknown_cnt = 0; | |
e34cbd30 JA |
358 | |
359 | rwb->scaled_max = calc_wb_limits(rwb); | |
360 | ||
361 | rwb_wake_all(rwb); | |
362 | ||
363 | rwb_trace_step(rwb, "step up"); | |
364 | } | |
365 | ||
366 | /* | |
367 | * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we | |
368 | * had a latency violation. | |
369 | */ | |
370 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) | |
371 | { | |
372 | /* | |
373 | * Stop scaling down when we've hit the limit. This also prevents | |
374 | * ->scale_step from going to crazy values, if the device can't | |
375 | * keep up. | |
376 | */ | |
377 | if (rwb->wb_max == 1) | |
378 | return; | |
379 | ||
380 | if (rwb->scale_step < 0 && hard_throttle) | |
381 | rwb->scale_step = 0; | |
382 | else | |
383 | rwb->scale_step++; | |
384 | ||
385 | rwb->scaled_max = false; | |
386 | rwb->unknown_cnt = 0; | |
e34cbd30 JA |
387 | calc_wb_limits(rwb); |
388 | rwb_trace_step(rwb, "step down"); | |
389 | } | |
390 | ||
391 | static void rwb_arm_timer(struct rq_wb *rwb) | |
392 | { | |
e34cbd30 JA |
393 | if (rwb->scale_step > 0) { |
394 | /* | |
395 | * We should speed this up, using some variant of a fast | |
396 | * integer inverse square root calculation. Since we only do | |
397 | * this for every window expiration, it's not a huge deal, | |
398 | * though. | |
399 | */ | |
400 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
401 | int_sqrt((rwb->scale_step + 1) << 8)); | |
402 | } else { | |
403 | /* | |
404 | * For step < 0, we don't want to increase/decrease the | |
405 | * window size. | |
406 | */ | |
407 | rwb->cur_win_nsec = rwb->win_nsec; | |
408 | } | |
409 | ||
34dbad5d | 410 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
411 | } |
412 | ||
34dbad5d | 413 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 414 | { |
34dbad5d | 415 | struct rq_wb *rwb = cb->data; |
e34cbd30 JA |
416 | unsigned int inflight = wbt_inflight(rwb); |
417 | int status; | |
418 | ||
34dbad5d | 419 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 420 | |
dc3b17cc | 421 | trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step, |
d8a0cbfd | 422 | inflight); |
e34cbd30 JA |
423 | |
424 | /* | |
425 | * If we exceeded the latency target, step down. If we did not, | |
426 | * step one level up. If we don't know enough to say either exceeded | |
427 | * or ok, then don't do anything. | |
428 | */ | |
429 | switch (status) { | |
430 | case LAT_EXCEEDED: | |
431 | scale_down(rwb, true); | |
432 | break; | |
433 | case LAT_OK: | |
434 | scale_up(rwb); | |
435 | break; | |
436 | case LAT_UNKNOWN_WRITES: | |
437 | /* | |
438 | * We started a the center step, but don't have a valid | |
439 | * read/write sample, but we do have writes going on. | |
440 | * Allow step to go negative, to increase write perf. | |
441 | */ | |
442 | scale_up(rwb); | |
443 | break; | |
444 | case LAT_UNKNOWN: | |
445 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
446 | break; | |
447 | /* | |
448 | * We get here when previously scaled reduced depth, and we | |
449 | * currently don't have a valid read/write sample. For that | |
450 | * case, slowly return to center state (step == 0). | |
451 | */ | |
452 | if (rwb->scale_step > 0) | |
453 | scale_up(rwb); | |
454 | else if (rwb->scale_step < 0) | |
455 | scale_down(rwb, false); | |
456 | break; | |
457 | default: | |
458 | break; | |
459 | } | |
460 | ||
461 | /* | |
462 | * Re-arm timer, if we have IO in flight | |
463 | */ | |
464 | if (rwb->scale_step || inflight) | |
465 | rwb_arm_timer(rwb); | |
466 | } | |
467 | ||
468 | void wbt_update_limits(struct rq_wb *rwb) | |
469 | { | |
470 | rwb->scale_step = 0; | |
471 | rwb->scaled_max = false; | |
472 | calc_wb_limits(rwb); | |
473 | ||
474 | rwb_wake_all(rwb); | |
475 | } | |
476 | ||
477 | static bool close_io(struct rq_wb *rwb) | |
478 | { | |
479 | const unsigned long now = jiffies; | |
480 | ||
481 | return time_before(now, rwb->last_issue + HZ / 10) || | |
482 | time_before(now, rwb->last_comp + HZ / 10); | |
483 | } | |
484 | ||
485 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
486 | ||
487 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
488 | { | |
489 | unsigned int limit; | |
490 | ||
782f5697 JA |
491 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) |
492 | return rwb->wb_background; | |
493 | ||
e34cbd30 JA |
494 | /* |
495 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 496 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
497 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
498 | * that. If the write is marked as a background write, then use | |
499 | * the idle limit, or go to normal if we haven't had competing | |
500 | * IO for a bit. | |
501 | */ | |
502 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
503 | limit = rwb->wb_max; | |
504 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { | |
505 | /* | |
506 | * If less than 100ms since we completed unrelated IO, | |
507 | * limit us to half the depth for background writeback. | |
508 | */ | |
509 | limit = rwb->wb_background; | |
510 | } else | |
511 | limit = rwb->wb_normal; | |
512 | ||
513 | return limit; | |
514 | } | |
515 | ||
516 | static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw, | |
ac6424b9 | 517 | wait_queue_entry_t *wait, unsigned long rw) |
e34cbd30 JA |
518 | { |
519 | /* | |
520 | * inc it here even if disabled, since we'll dec it at completion. | |
521 | * this only happens if the task was sleeping in __wbt_wait(), | |
522 | * and someone turned it off at the same time. | |
523 | */ | |
524 | if (!rwb_enabled(rwb)) { | |
525 | atomic_inc(&rqw->inflight); | |
526 | return true; | |
527 | } | |
528 | ||
529 | /* | |
530 | * If the waitqueue is already active and we are not the next | |
531 | * in line to be woken up, wait for our turn. | |
532 | */ | |
533 | if (waitqueue_active(&rqw->wait) && | |
2055da97 | 534 | rqw->wait.head.next != &wait->entry) |
e34cbd30 JA |
535 | return false; |
536 | ||
537 | return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw)); | |
538 | } | |
539 | ||
540 | /* | |
541 | * Block if we will exceed our limit, or if we are currently waiting for | |
542 | * the timer to kick off queuing again. | |
543 | */ | |
8bea6090 JA |
544 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
545 | unsigned long rw, spinlock_t *lock) | |
9eca5350 BVA |
546 | __releases(lock) |
547 | __acquires(lock) | |
e34cbd30 | 548 | { |
8bea6090 | 549 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
e34cbd30 JA |
550 | DEFINE_WAIT(wait); |
551 | ||
552 | if (may_queue(rwb, rqw, &wait, rw)) | |
553 | return; | |
554 | ||
555 | do { | |
556 | prepare_to_wait_exclusive(&rqw->wait, &wait, | |
557 | TASK_UNINTERRUPTIBLE); | |
558 | ||
559 | if (may_queue(rwb, rqw, &wait, rw)) | |
560 | break; | |
561 | ||
9eca5350 | 562 | if (lock) { |
e34cbd30 | 563 | spin_unlock_irq(lock); |
9eca5350 | 564 | io_schedule(); |
e34cbd30 | 565 | spin_lock_irq(lock); |
9eca5350 BVA |
566 | } else |
567 | io_schedule(); | |
e34cbd30 JA |
568 | } while (1); |
569 | ||
570 | finish_wait(&rqw->wait, &wait); | |
571 | } | |
572 | ||
573 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
574 | { | |
782f5697 JA |
575 | switch (bio_op(bio)) { |
576 | case REQ_OP_WRITE: | |
577 | /* | |
578 | * Don't throttle WRITE_ODIRECT | |
579 | */ | |
580 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
581 | (REQ_SYNC | REQ_IDLE)) | |
582 | return false; | |
583 | /* fallthrough */ | |
584 | case REQ_OP_DISCARD: | |
585 | return true; | |
586 | default: | |
e34cbd30 | 587 | return false; |
782f5697 | 588 | } |
e34cbd30 JA |
589 | } |
590 | ||
591 | /* | |
592 | * Returns true if the IO request should be accounted, false if not. | |
593 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
594 | * in an irq held spinlock, if it holds one when calling this function. | |
595 | * If we do sleep, we'll release and re-grab it. | |
596 | */ | |
f2e0a0b2 | 597 | enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock) |
e34cbd30 | 598 | { |
8bea6090 | 599 | enum wbt_flags ret = 0; |
e34cbd30 JA |
600 | |
601 | if (!rwb_enabled(rwb)) | |
602 | return 0; | |
603 | ||
604 | if (bio_op(bio) == REQ_OP_READ) | |
605 | ret = WBT_READ; | |
606 | ||
607 | if (!wbt_should_throttle(rwb, bio)) { | |
608 | if (ret & WBT_READ) | |
609 | wb_timestamp(rwb, &rwb->last_issue); | |
610 | return ret; | |
611 | } | |
612 | ||
8bea6090 JA |
613 | if (current_is_kswapd()) |
614 | ret |= WBT_KSWAPD; | |
782f5697 JA |
615 | if (bio_op(bio) == REQ_OP_DISCARD) |
616 | ret |= WBT_DISCARD; | |
8bea6090 JA |
617 | |
618 | __wbt_wait(rwb, ret, bio->bi_opf, lock); | |
e34cbd30 | 619 | |
34dbad5d | 620 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 JA |
621 | rwb_arm_timer(rwb); |
622 | ||
e34cbd30 JA |
623 | return ret | WBT_TRACKED; |
624 | } | |
625 | ||
626 | void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
627 | { | |
628 | if (!rwb_enabled(rwb)) | |
629 | return; | |
630 | ||
631 | /* | |
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. | |
638 | */ | |
639 | if (wbt_is_read(stat) && !rwb->sync_issue) { | |
640 | rwb->sync_cookie = stat; | |
641 | rwb->sync_issue = blk_stat_time(stat); | |
642 | } | |
643 | } | |
644 | ||
645 | void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
646 | { | |
647 | if (!rwb_enabled(rwb)) | |
648 | return; | |
649 | if (stat == rwb->sync_cookie) { | |
650 | rwb->sync_issue = 0; | |
651 | rwb->sync_cookie = NULL; | |
652 | } | |
653 | } | |
654 | ||
655 | void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) | |
656 | { | |
657 | if (rwb) { | |
658 | rwb->queue_depth = depth; | |
659 | wbt_update_limits(rwb); | |
660 | } | |
661 | } | |
662 | ||
663 | void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) | |
664 | { | |
665 | if (rwb) | |
666 | rwb->wc = write_cache_on; | |
667 | } | |
668 | ||
3f19cd23 | 669 | /* |
b5dc5d4d | 670 | * Disable wbt, if enabled by default. |
fa224eed JA |
671 | */ |
672 | void wbt_disable_default(struct request_queue *q) | |
e34cbd30 | 673 | { |
fa224eed JA |
674 | struct rq_wb *rwb = q->rq_wb; |
675 | ||
3f19cd23 JK |
676 | if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) |
677 | wbt_exit(q); | |
e34cbd30 | 678 | } |
fa224eed | 679 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
e34cbd30 | 680 | |
8330cdb0 JK |
681 | /* |
682 | * Enable wbt if defaults are configured that way | |
683 | */ | |
684 | void wbt_enable_default(struct request_queue *q) | |
685 | { | |
686 | /* Throttling already enabled? */ | |
687 | if (q->rq_wb) | |
688 | return; | |
689 | ||
690 | /* Queue not registered? Maybe shutting down... */ | |
691 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
692 | return; | |
693 | ||
694 | if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) || | |
695 | (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ))) | |
696 | wbt_init(q); | |
697 | } | |
698 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
699 | ||
80e091d1 JA |
700 | u64 wbt_default_latency_nsec(struct request_queue *q) |
701 | { | |
702 | /* | |
703 | * We default to 2msec for non-rotational storage, and 75msec | |
704 | * for rotational storage. | |
705 | */ | |
706 | if (blk_queue_nonrot(q)) | |
707 | return 2000000ULL; | |
708 | else | |
709 | return 75000000ULL; | |
710 | } | |
711 | ||
99c749a4 JA |
712 | static int wbt_data_dir(const struct request *rq) |
713 | { | |
5235553d JA |
714 | const int op = req_op(rq); |
715 | ||
716 | if (op == REQ_OP_READ) | |
717 | return READ; | |
825843b0 | 718 | else if (op_is_write(op)) |
5235553d JA |
719 | return WRITE; |
720 | ||
721 | /* don't account */ | |
722 | return -1; | |
99c749a4 JA |
723 | } |
724 | ||
8054b89f | 725 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
726 | { |
727 | struct rq_wb *rwb; | |
728 | int i; | |
729 | ||
e34cbd30 JA |
730 | BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS); |
731 | ||
e34cbd30 JA |
732 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
733 | if (!rwb) | |
734 | return -ENOMEM; | |
735 | ||
99c749a4 | 736 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
737 | if (!rwb->cb) { |
738 | kfree(rwb); | |
739 | return -ENOMEM; | |
740 | } | |
741 | ||
e34cbd30 JA |
742 | for (i = 0; i < WBT_NUM_RWQ; i++) { |
743 | atomic_set(&rwb->rq_wait[i].inflight, 0); | |
744 | init_waitqueue_head(&rwb->rq_wait[i].wait); | |
745 | } | |
746 | ||
e34cbd30 | 747 | rwb->last_comp = rwb->last_issue = jiffies; |
d8a0cbfd | 748 | rwb->queue = q; |
e34cbd30 | 749 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 750 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
e34cbd30 JA |
751 | wbt_update_limits(rwb); |
752 | ||
753 | /* | |
34dbad5d | 754 | * Assign rwb and add the stats callback. |
e34cbd30 JA |
755 | */ |
756 | q->rq_wb = rwb; | |
34dbad5d | 757 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 758 | |
80e091d1 | 759 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 JA |
760 | |
761 | wbt_set_queue_depth(rwb, blk_queue_depth(q)); | |
762 | wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
763 | ||
764 | return 0; | |
765 | } | |
766 | ||
767 | void wbt_exit(struct request_queue *q) | |
768 | { | |
769 | struct rq_wb *rwb = q->rq_wb; | |
770 | ||
771 | if (rwb) { | |
34dbad5d OS |
772 | blk_stat_remove_callback(q, rwb->cb); |
773 | blk_stat_free_callback(rwb->cb); | |
e34cbd30 JA |
774 | q->rq_wb = NULL; |
775 | kfree(rwb); | |
776 | } | |
777 | } |