]>
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 | ||
d9204a0b JA |
104 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
105 | enum wbt_flags wb_acct) | |
e34cbd30 | 106 | { |
d9204a0b JA |
107 | if (wb_acct & WBT_KSWAPD) |
108 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
109 | ||
110 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
111 | } |
112 | ||
113 | static void rwb_wake_all(struct rq_wb *rwb) | |
114 | { | |
115 | int i; | |
116 | ||
117 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
118 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
119 | ||
7e31a9ca | 120 | if (wq_has_sleeper(&rqw->wait)) |
e34cbd30 JA |
121 | wake_up_all(&rqw->wait); |
122 | } | |
123 | } | |
124 | ||
e09b527f JA |
125 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
126 | enum wbt_flags wb_acct) | |
e34cbd30 | 127 | { |
e34cbd30 JA |
128 | int inflight, limit; |
129 | ||
e34cbd30 JA |
130 | inflight = atomic_dec_return(&rqw->inflight); |
131 | ||
132 | /* | |
133 | * wbt got disabled with IO in flight. Wake up any potential | |
134 | * waiters, we don't have to do more than that. | |
135 | */ | |
136 | if (unlikely(!rwb_enabled(rwb))) { | |
137 | rwb_wake_all(rwb); | |
138 | return; | |
139 | } | |
140 | ||
141 | /* | |
142 | * If the device does write back caching, drop further down | |
143 | * before we wake people up. | |
144 | */ | |
145 | if (rwb->wc && !wb_recent_wait(rwb)) | |
146 | limit = 0; | |
147 | else | |
148 | limit = rwb->wb_normal; | |
149 | ||
150 | /* | |
151 | * Don't wake anyone up if we are above the normal limit. | |
152 | */ | |
153 | if (inflight && inflight >= limit) | |
154 | return; | |
155 | ||
7e31a9ca | 156 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
157 | int diff = limit - inflight; |
158 | ||
159 | if (!inflight || diff >= rwb->wb_background / 2) | |
8907e6ec | 160 | wake_up_all(&rqw->wait); |
e34cbd30 JA |
161 | } |
162 | } | |
163 | ||
e09b527f JA |
164 | void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct) |
165 | { | |
166 | struct rq_wait *rqw; | |
167 | ||
168 | if (!(wb_acct & WBT_TRACKED)) | |
169 | return; | |
170 | ||
171 | rqw = get_rq_wait(rwb, wb_acct); | |
172 | wbt_rqw_done(rwb, rqw, wb_acct); | |
173 | } | |
174 | ||
e34cbd30 JA |
175 | /* |
176 | * Called on completion of a request. Note that it's also called when | |
177 | * a request is merged, when the request gets freed. | |
178 | */ | |
179 | void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
180 | { | |
181 | if (!rwb) | |
182 | return; | |
183 | ||
184 | if (!wbt_is_tracked(stat)) { | |
185 | if (rwb->sync_cookie == stat) { | |
186 | rwb->sync_issue = 0; | |
187 | rwb->sync_cookie = NULL; | |
188 | } | |
189 | ||
190 | if (wbt_is_read(stat)) | |
191 | wb_timestamp(rwb, &rwb->last_comp); | |
e34cbd30 JA |
192 | } else { |
193 | WARN_ON_ONCE(stat == rwb->sync_cookie); | |
194 | __wbt_done(rwb, wbt_stat_to_mask(stat)); | |
e34cbd30 | 195 | } |
62d772fa | 196 | wbt_clear_state(stat); |
e34cbd30 JA |
197 | } |
198 | ||
199 | /* | |
200 | * Return true, if we can't increase the depth further by scaling | |
201 | */ | |
202 | static bool calc_wb_limits(struct rq_wb *rwb) | |
203 | { | |
204 | unsigned int depth; | |
205 | bool ret = false; | |
206 | ||
207 | if (!rwb->min_lat_nsec) { | |
208 | rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; | |
209 | return false; | |
210 | } | |
211 | ||
212 | /* | |
213 | * For QD=1 devices, this is a special case. It's important for those | |
214 | * to have one request ready when one completes, so force a depth of | |
215 | * 2 for those devices. On the backend, it'll be a depth of 1 anyway, | |
216 | * since the device can't have more than that in flight. If we're | |
217 | * scaling down, then keep a setting of 1/1/1. | |
218 | */ | |
219 | if (rwb->queue_depth == 1) { | |
220 | if (rwb->scale_step > 0) | |
221 | rwb->wb_max = rwb->wb_normal = 1; | |
222 | else { | |
223 | rwb->wb_max = rwb->wb_normal = 2; | |
224 | ret = true; | |
225 | } | |
226 | rwb->wb_background = 1; | |
227 | } else { | |
228 | /* | |
229 | * scale_step == 0 is our default state. If we have suffered | |
230 | * latency spikes, step will be > 0, and we shrink the | |
231 | * allowed write depths. If step is < 0, we're only doing | |
232 | * writes, and we allow a temporarily higher depth to | |
233 | * increase performance. | |
234 | */ | |
235 | depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth); | |
236 | if (rwb->scale_step > 0) | |
237 | depth = 1 + ((depth - 1) >> min(31, rwb->scale_step)); | |
238 | else if (rwb->scale_step < 0) { | |
239 | unsigned int maxd = 3 * rwb->queue_depth / 4; | |
240 | ||
241 | depth = 1 + ((depth - 1) << -rwb->scale_step); | |
242 | if (depth > maxd) { | |
243 | depth = maxd; | |
244 | ret = true; | |
245 | } | |
246 | } | |
247 | ||
248 | /* | |
249 | * Set our max/normal/bg queue depths based on how far | |
250 | * we have scaled down (->scale_step). | |
251 | */ | |
252 | rwb->wb_max = depth; | |
253 | rwb->wb_normal = (rwb->wb_max + 1) / 2; | |
254 | rwb->wb_background = (rwb->wb_max + 3) / 4; | |
255 | } | |
256 | ||
257 | return ret; | |
258 | } | |
259 | ||
4121d385 | 260 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
261 | { |
262 | /* | |
263 | * We need at least one read sample, and a minimum of | |
264 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
265 | * that it's writes impacting us, and not just some sole read on | |
266 | * a device that is in a lower power state. | |
267 | */ | |
fa2e39cb OS |
268 | return (stat[READ].nr_samples >= 1 && |
269 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
270 | } |
271 | ||
272 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
273 | { | |
6aa7de05 | 274 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
275 | |
276 | if (!issue || !rwb->sync_cookie) | |
277 | return 0; | |
278 | ||
279 | now = ktime_to_ns(ktime_get()); | |
280 | return now - issue; | |
281 | } | |
282 | ||
283 | enum { | |
284 | LAT_OK = 1, | |
285 | LAT_UNKNOWN, | |
286 | LAT_UNKNOWN_WRITES, | |
287 | LAT_EXCEEDED, | |
288 | }; | |
289 | ||
34dbad5d | 290 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 291 | { |
dc3b17cc | 292 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; |
e34cbd30 JA |
293 | u64 thislat; |
294 | ||
295 | /* | |
296 | * If our stored sync issue exceeds the window size, or it | |
297 | * exceeds our min target AND we haven't logged any entries, | |
298 | * flag the latency as exceeded. wbt works off completion latencies, | |
299 | * but for a flooded device, a single sync IO can take a long time | |
300 | * to complete after being issued. If this time exceeds our | |
301 | * monitoring window AND we didn't see any other completions in that | |
302 | * window, then count that sync IO as a violation of the latency. | |
303 | */ | |
304 | thislat = rwb_sync_issue_lat(rwb); | |
305 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 306 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 307 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
308 | return LAT_EXCEEDED; |
309 | } | |
310 | ||
311 | /* | |
312 | * No read/write mix, if stat isn't valid | |
313 | */ | |
314 | if (!stat_sample_valid(stat)) { | |
315 | /* | |
316 | * If we had writes in this stat window and the window is | |
317 | * current, we're only doing writes. If a task recently | |
318 | * waited or still has writes in flights, consider us doing | |
319 | * just writes as well. | |
320 | */ | |
34dbad5d OS |
321 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
322 | wbt_inflight(rwb)) | |
e34cbd30 JA |
323 | return LAT_UNKNOWN_WRITES; |
324 | return LAT_UNKNOWN; | |
325 | } | |
326 | ||
327 | /* | |
328 | * If the 'min' latency exceeds our target, step down. | |
329 | */ | |
fa2e39cb OS |
330 | if (stat[READ].min > rwb->min_lat_nsec) { |
331 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 332 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
333 | return LAT_EXCEEDED; |
334 | } | |
335 | ||
336 | if (rwb->scale_step) | |
d8a0cbfd | 337 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
338 | |
339 | return LAT_OK; | |
340 | } | |
341 | ||
e34cbd30 JA |
342 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
343 | { | |
dc3b17cc | 344 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; |
d8a0cbfd JA |
345 | |
346 | trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec, | |
e34cbd30 JA |
347 | rwb->wb_background, rwb->wb_normal, rwb->wb_max); |
348 | } | |
349 | ||
350 | static void scale_up(struct rq_wb *rwb) | |
351 | { | |
352 | /* | |
353 | * Hit max in previous round, stop here | |
354 | */ | |
355 | if (rwb->scaled_max) | |
356 | return; | |
357 | ||
358 | rwb->scale_step--; | |
359 | rwb->unknown_cnt = 0; | |
e34cbd30 JA |
360 | |
361 | rwb->scaled_max = calc_wb_limits(rwb); | |
362 | ||
363 | rwb_wake_all(rwb); | |
364 | ||
365 | rwb_trace_step(rwb, "step up"); | |
366 | } | |
367 | ||
368 | /* | |
369 | * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we | |
370 | * had a latency violation. | |
371 | */ | |
372 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) | |
373 | { | |
374 | /* | |
375 | * Stop scaling down when we've hit the limit. This also prevents | |
376 | * ->scale_step from going to crazy values, if the device can't | |
377 | * keep up. | |
378 | */ | |
379 | if (rwb->wb_max == 1) | |
380 | return; | |
381 | ||
382 | if (rwb->scale_step < 0 && hard_throttle) | |
383 | rwb->scale_step = 0; | |
384 | else | |
385 | rwb->scale_step++; | |
386 | ||
387 | rwb->scaled_max = false; | |
388 | rwb->unknown_cnt = 0; | |
e34cbd30 JA |
389 | calc_wb_limits(rwb); |
390 | rwb_trace_step(rwb, "step down"); | |
391 | } | |
392 | ||
393 | static void rwb_arm_timer(struct rq_wb *rwb) | |
394 | { | |
e34cbd30 JA |
395 | if (rwb->scale_step > 0) { |
396 | /* | |
397 | * We should speed this up, using some variant of a fast | |
398 | * integer inverse square root calculation. Since we only do | |
399 | * this for every window expiration, it's not a huge deal, | |
400 | * though. | |
401 | */ | |
402 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
403 | int_sqrt((rwb->scale_step + 1) << 8)); | |
404 | } else { | |
405 | /* | |
406 | * For step < 0, we don't want to increase/decrease the | |
407 | * window size. | |
408 | */ | |
409 | rwb->cur_win_nsec = rwb->win_nsec; | |
410 | } | |
411 | ||
34dbad5d | 412 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
413 | } |
414 | ||
34dbad5d | 415 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 416 | { |
34dbad5d | 417 | struct rq_wb *rwb = cb->data; |
e34cbd30 JA |
418 | unsigned int inflight = wbt_inflight(rwb); |
419 | int status; | |
420 | ||
34dbad5d | 421 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 422 | |
dc3b17cc | 423 | trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step, |
d8a0cbfd | 424 | inflight); |
e34cbd30 JA |
425 | |
426 | /* | |
427 | * If we exceeded the latency target, step down. If we did not, | |
428 | * step one level up. If we don't know enough to say either exceeded | |
429 | * or ok, then don't do anything. | |
430 | */ | |
431 | switch (status) { | |
432 | case LAT_EXCEEDED: | |
433 | scale_down(rwb, true); | |
434 | break; | |
435 | case LAT_OK: | |
436 | scale_up(rwb); | |
437 | break; | |
438 | case LAT_UNKNOWN_WRITES: | |
439 | /* | |
440 | * We started a the center step, but don't have a valid | |
441 | * read/write sample, but we do have writes going on. | |
442 | * Allow step to go negative, to increase write perf. | |
443 | */ | |
444 | scale_up(rwb); | |
445 | break; | |
446 | case LAT_UNKNOWN: | |
447 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
448 | break; | |
449 | /* | |
450 | * We get here when previously scaled reduced depth, and we | |
451 | * currently don't have a valid read/write sample. For that | |
452 | * case, slowly return to center state (step == 0). | |
453 | */ | |
454 | if (rwb->scale_step > 0) | |
455 | scale_up(rwb); | |
456 | else if (rwb->scale_step < 0) | |
457 | scale_down(rwb, false); | |
458 | break; | |
459 | default: | |
460 | break; | |
461 | } | |
462 | ||
463 | /* | |
464 | * Re-arm timer, if we have IO in flight | |
465 | */ | |
466 | if (rwb->scale_step || inflight) | |
467 | rwb_arm_timer(rwb); | |
468 | } | |
469 | ||
470 | void wbt_update_limits(struct rq_wb *rwb) | |
471 | { | |
472 | rwb->scale_step = 0; | |
473 | rwb->scaled_max = false; | |
474 | calc_wb_limits(rwb); | |
475 | ||
476 | rwb_wake_all(rwb); | |
477 | } | |
478 | ||
479 | static bool close_io(struct rq_wb *rwb) | |
480 | { | |
481 | const unsigned long now = jiffies; | |
482 | ||
483 | return time_before(now, rwb->last_issue + HZ / 10) || | |
484 | time_before(now, rwb->last_comp + HZ / 10); | |
485 | } | |
486 | ||
487 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
488 | ||
489 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
490 | { | |
491 | unsigned int limit; | |
492 | ||
2fb965cc JA |
493 | /* |
494 | * If we got disabled, just return UINT_MAX. This ensures that | |
495 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
496 | */ | |
497 | if (!rwb_enabled(rwb)) | |
498 | return UINT_MAX; | |
499 | ||
e34cbd30 JA |
500 | /* |
501 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 502 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
503 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
504 | * that. If the write is marked as a background write, then use | |
505 | * the idle limit, or go to normal if we haven't had competing | |
506 | * IO for a bit. | |
507 | */ | |
508 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
509 | limit = rwb->wb_max; | |
510 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { | |
511 | /* | |
512 | * If less than 100ms since we completed unrelated IO, | |
513 | * limit us to half the depth for background writeback. | |
514 | */ | |
515 | limit = rwb->wb_background; | |
516 | } else | |
517 | limit = rwb->wb_normal; | |
518 | ||
519 | return limit; | |
520 | } | |
521 | ||
8907e6ec JA |
522 | struct wbt_wait_data { |
523 | struct wait_queue_entry wq; | |
524 | struct task_struct *task; | |
525 | struct rq_wb *rwb; | |
526 | struct rq_wait *rqw; | |
527 | unsigned long rw; | |
528 | bool got_token; | |
529 | }; | |
530 | ||
531 | static int wbt_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
532 | int wake_flags, void *key) | |
533 | { | |
534 | struct wbt_wait_data *data = container_of(curr, struct wbt_wait_data, | |
535 | wq); | |
536 | ||
537 | /* | |
538 | * If we fail to get a budget, return -1 to interrupt the wake up | |
539 | * loop in __wake_up_common. | |
540 | */ | |
541 | if (!atomic_inc_below(&data->rqw->inflight, get_limit(data->rwb, data->rw))) | |
542 | return -1; | |
543 | ||
544 | data->got_token = true; | |
545 | list_del_init(&curr->entry); | |
546 | wake_up_process(data->task); | |
547 | return 1; | |
548 | } | |
549 | ||
e34cbd30 JA |
550 | /* |
551 | * Block if we will exceed our limit, or if we are currently waiting for | |
552 | * the timer to kick off queuing again. | |
553 | */ | |
d9204a0b JA |
554 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
555 | unsigned long rw, spinlock_t *lock) | |
9eca5350 BVA |
556 | __releases(lock) |
557 | __acquires(lock) | |
e34cbd30 | 558 | { |
d9204a0b | 559 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
8907e6ec JA |
560 | struct wbt_wait_data data = { |
561 | .wq = { | |
562 | .func = wbt_wake_function, | |
563 | .entry = LIST_HEAD_INIT(data.wq.entry), | |
564 | }, | |
565 | .task = current, | |
566 | .rwb = rwb, | |
567 | .rqw = rqw, | |
568 | .rw = rw, | |
569 | }; | |
41d2b768 | 570 | bool has_sleeper; |
e34cbd30 | 571 | |
41d2b768 JA |
572 | has_sleeper = wq_has_sleeper(&rqw->wait); |
573 | if (!has_sleeper && atomic_inc_below(&rqw->inflight, get_limit(rwb, rw))) | |
e6a7d9d8 AA |
574 | return; |
575 | ||
8907e6ec | 576 | prepare_to_wait_exclusive(&rqw->wait, &data.wq, TASK_UNINTERRUPTIBLE); |
e34cbd30 | 577 | do { |
8907e6ec JA |
578 | if (data.got_token) |
579 | break; | |
e6a7d9d8 | 580 | |
8907e6ec JA |
581 | if (!has_sleeper && |
582 | atomic_inc_below(&rqw->inflight, get_limit(rwb, rw))) { | |
583 | finish_wait(&rqw->wait, &data.wq); | |
584 | ||
585 | /* | |
586 | * We raced with wbt_wake_function() getting a token, | |
587 | * which means we now have two. Put our local token | |
588 | * and wake anyone else potentially waiting for one. | |
589 | */ | |
590 | if (data.got_token) | |
591 | wbt_rqw_done(rwb, rqw, wb_acct); | |
e34cbd30 | 592 | break; |
8907e6ec | 593 | } |
e34cbd30 | 594 | |
9eca5350 | 595 | if (lock) { |
e34cbd30 | 596 | spin_unlock_irq(lock); |
9eca5350 | 597 | io_schedule(); |
e34cbd30 | 598 | spin_lock_irq(lock); |
9eca5350 BVA |
599 | } else |
600 | io_schedule(); | |
8907e6ec | 601 | |
41d2b768 | 602 | has_sleeper = false; |
e34cbd30 JA |
603 | } while (1); |
604 | ||
8907e6ec | 605 | finish_wait(&rqw->wait, &data.wq); |
e34cbd30 JA |
606 | } |
607 | ||
608 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
609 | { | |
610 | const int op = bio_op(bio); | |
611 | ||
612 | /* | |
be07e14f | 613 | * If not a WRITE, do nothing |
e34cbd30 | 614 | */ |
be07e14f | 615 | if (op != REQ_OP_WRITE) |
e34cbd30 JA |
616 | return false; |
617 | ||
618 | /* | |
619 | * Don't throttle WRITE_ODIRECT | |
620 | */ | |
621 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE)) | |
622 | return false; | |
623 | ||
624 | return true; | |
625 | } | |
626 | ||
627 | /* | |
628 | * Returns true if the IO request should be accounted, false if not. | |
629 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
630 | * in an irq held spinlock, if it holds one when calling this function. | |
631 | * If we do sleep, we'll release and re-grab it. | |
632 | */ | |
f2e0a0b2 | 633 | enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock) |
e34cbd30 | 634 | { |
d9204a0b | 635 | enum wbt_flags ret = 0; |
e34cbd30 JA |
636 | |
637 | if (!rwb_enabled(rwb)) | |
638 | return 0; | |
639 | ||
640 | if (bio_op(bio) == REQ_OP_READ) | |
641 | ret = WBT_READ; | |
642 | ||
643 | if (!wbt_should_throttle(rwb, bio)) { | |
644 | if (ret & WBT_READ) | |
645 | wb_timestamp(rwb, &rwb->last_issue); | |
646 | return ret; | |
647 | } | |
648 | ||
d9204a0b JA |
649 | if (current_is_kswapd()) |
650 | ret |= WBT_KSWAPD; | |
651 | ||
652 | __wbt_wait(rwb, ret, bio->bi_opf, lock); | |
e34cbd30 | 653 | |
34dbad5d | 654 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 JA |
655 | rwb_arm_timer(rwb); |
656 | ||
e34cbd30 JA |
657 | return ret | WBT_TRACKED; |
658 | } | |
659 | ||
660 | void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
661 | { | |
662 | if (!rwb_enabled(rwb)) | |
663 | return; | |
664 | ||
665 | /* | |
666 | * Track sync issue, in case it takes a long time to complete. Allows | |
667 | * us to react quicker, if a sync IO takes a long time to complete. | |
668 | * Note that this is just a hint. 'stat' can go away when the | |
669 | * request completes, so it's important we never dereference it. We | |
670 | * only use the address to compare with, which is why we store the | |
671 | * sync_issue time locally. | |
672 | */ | |
673 | if (wbt_is_read(stat) && !rwb->sync_issue) { | |
674 | rwb->sync_cookie = stat; | |
675 | rwb->sync_issue = blk_stat_time(stat); | |
676 | } | |
677 | } | |
678 | ||
679 | void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat) | |
680 | { | |
681 | if (!rwb_enabled(rwb)) | |
682 | return; | |
683 | if (stat == rwb->sync_cookie) { | |
684 | rwb->sync_issue = 0; | |
685 | rwb->sync_cookie = NULL; | |
686 | } | |
687 | } | |
688 | ||
689 | void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) | |
690 | { | |
691 | if (rwb) { | |
692 | rwb->queue_depth = depth; | |
693 | wbt_update_limits(rwb); | |
694 | } | |
695 | } | |
696 | ||
697 | void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) | |
698 | { | |
699 | if (rwb) | |
700 | rwb->wc = write_cache_on; | |
701 | } | |
702 | ||
3f19cd23 | 703 | /* |
b5dc5d4d | 704 | * Disable wbt, if enabled by default. |
fa224eed JA |
705 | */ |
706 | void wbt_disable_default(struct request_queue *q) | |
e34cbd30 | 707 | { |
fa224eed JA |
708 | struct rq_wb *rwb = q->rq_wb; |
709 | ||
3f19cd23 JK |
710 | if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) |
711 | wbt_exit(q); | |
e34cbd30 | 712 | } |
fa224eed | 713 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
e34cbd30 | 714 | |
8330cdb0 JK |
715 | /* |
716 | * Enable wbt if defaults are configured that way | |
717 | */ | |
718 | void wbt_enable_default(struct request_queue *q) | |
719 | { | |
720 | /* Throttling already enabled? */ | |
721 | if (q->rq_wb) | |
722 | return; | |
723 | ||
724 | /* Queue not registered? Maybe shutting down... */ | |
725 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
726 | return; | |
727 | ||
728 | if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) || | |
729 | (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ))) | |
730 | wbt_init(q); | |
731 | } | |
732 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
733 | ||
80e091d1 JA |
734 | u64 wbt_default_latency_nsec(struct request_queue *q) |
735 | { | |
736 | /* | |
737 | * We default to 2msec for non-rotational storage, and 75msec | |
738 | * for rotational storage. | |
739 | */ | |
740 | if (blk_queue_nonrot(q)) | |
741 | return 2000000ULL; | |
742 | else | |
743 | return 75000000ULL; | |
744 | } | |
745 | ||
99c749a4 JA |
746 | static int wbt_data_dir(const struct request *rq) |
747 | { | |
79761a35 JA |
748 | const int op = req_op(rq); |
749 | ||
750 | if (op == REQ_OP_READ) | |
751 | return READ; | |
752 | else if (op == REQ_OP_WRITE || op == REQ_OP_FLUSH) | |
753 | return WRITE; | |
754 | ||
755 | /* don't account */ | |
756 | return -1; | |
99c749a4 JA |
757 | } |
758 | ||
8054b89f | 759 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
760 | { |
761 | struct rq_wb *rwb; | |
762 | int i; | |
763 | ||
e34cbd30 JA |
764 | BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS); |
765 | ||
e34cbd30 JA |
766 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
767 | if (!rwb) | |
768 | return -ENOMEM; | |
769 | ||
99c749a4 | 770 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
771 | if (!rwb->cb) { |
772 | kfree(rwb); | |
773 | return -ENOMEM; | |
774 | } | |
775 | ||
e34cbd30 JA |
776 | for (i = 0; i < WBT_NUM_RWQ; i++) { |
777 | atomic_set(&rwb->rq_wait[i].inflight, 0); | |
778 | init_waitqueue_head(&rwb->rq_wait[i].wait); | |
779 | } | |
780 | ||
e34cbd30 | 781 | rwb->last_comp = rwb->last_issue = jiffies; |
d8a0cbfd | 782 | rwb->queue = q; |
e34cbd30 | 783 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 784 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
e34cbd30 JA |
785 | wbt_update_limits(rwb); |
786 | ||
787 | /* | |
34dbad5d | 788 | * Assign rwb and add the stats callback. |
e34cbd30 JA |
789 | */ |
790 | q->rq_wb = rwb; | |
34dbad5d | 791 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 792 | |
80e091d1 | 793 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 JA |
794 | |
795 | wbt_set_queue_depth(rwb, blk_queue_depth(q)); | |
796 | wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
797 | ||
798 | return 0; | |
799 | } | |
800 | ||
801 | void wbt_exit(struct request_queue *q) | |
802 | { | |
803 | struct rq_wb *rwb = q->rq_wb; | |
804 | ||
805 | if (rwb) { | |
34dbad5d OS |
806 | blk_stat_remove_callback(q, rwb->cb); |
807 | blk_stat_free_callback(rwb->cb); | |
e34cbd30 JA |
808 | q->rq_wb = NULL; |
809 | kfree(rwb); | |
810 | } | |
811 | } |