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 | ||
a8a45941 | 32 | static inline void wbt_clear_state(struct request *rq) |
934031a1 | 33 | { |
544ccc8d | 34 | rq->wbt_flags = 0; |
934031a1 OS |
35 | } |
36 | ||
a8a45941 | 37 | static inline enum wbt_flags wbt_flags(struct request *rq) |
934031a1 | 38 | { |
544ccc8d | 39 | return rq->wbt_flags; |
934031a1 OS |
40 | } |
41 | ||
a8a45941 | 42 | static inline bool wbt_is_tracked(struct request *rq) |
934031a1 | 43 | { |
544ccc8d | 44 | return rq->wbt_flags & WBT_TRACKED; |
934031a1 OS |
45 | } |
46 | ||
a8a45941 | 47 | static inline bool wbt_is_read(struct request *rq) |
934031a1 | 48 | { |
544ccc8d | 49 | return rq->wbt_flags & WBT_READ; |
934031a1 OS |
50 | } |
51 | ||
e34cbd30 JA |
52 | enum { |
53 | /* | |
54 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
55 | * from here depending on device stats | |
56 | */ | |
57 | RWB_DEF_DEPTH = 16, | |
58 | ||
59 | /* | |
60 | * 100msec window | |
61 | */ | |
62 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
63 | ||
64 | /* | |
65 | * Disregard stats, if we don't meet this minimum | |
66 | */ | |
67 | RWB_MIN_WRITE_SAMPLES = 3, | |
68 | ||
69 | /* | |
70 | * If we have this number of consecutive windows with not enough | |
71 | * information to scale up or down, scale up. | |
72 | */ | |
73 | RWB_UNKNOWN_BUMP = 5, | |
74 | }; | |
75 | ||
76 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
77 | { | |
78 | return rwb && rwb->wb_normal != 0; | |
79 | } | |
80 | ||
81 | /* | |
82 | * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded, | |
83 | * false if 'v' + 1 would be bigger than 'below'. | |
84 | */ | |
85 | static bool atomic_inc_below(atomic_t *v, int below) | |
86 | { | |
87 | int cur = atomic_read(v); | |
88 | ||
89 | for (;;) { | |
90 | int old; | |
91 | ||
92 | if (cur >= below) | |
93 | return false; | |
94 | old = atomic_cmpxchg(v, cur, cur + 1); | |
95 | if (old == cur) | |
96 | break; | |
97 | cur = old; | |
98 | } | |
99 | ||
100 | return true; | |
101 | } | |
102 | ||
103 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) | |
104 | { | |
105 | if (rwb_enabled(rwb)) { | |
106 | const unsigned long cur = jiffies; | |
107 | ||
108 | if (cur != *var) | |
109 | *var = cur; | |
110 | } | |
111 | } | |
112 | ||
113 | /* | |
114 | * If a task was rate throttled in balance_dirty_pages() within the last | |
115 | * second or so, use that to indicate a higher cleaning rate. | |
116 | */ | |
117 | static bool wb_recent_wait(struct rq_wb *rwb) | |
118 | { | |
dc3b17cc | 119 | struct bdi_writeback *wb = &rwb->queue->backing_dev_info->wb; |
e34cbd30 JA |
120 | |
121 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
122 | } | |
123 | ||
8bea6090 JA |
124 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
125 | enum wbt_flags wb_acct) | |
e34cbd30 | 126 | { |
8bea6090 JA |
127 | if (wb_acct & WBT_KSWAPD) |
128 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
129 | else if (wb_acct & WBT_DISCARD) |
130 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
131 | |
132 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
133 | } |
134 | ||
135 | static void rwb_wake_all(struct rq_wb *rwb) | |
136 | { | |
137 | int i; | |
138 | ||
139 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
140 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
141 | ||
142 | if (waitqueue_active(&rqw->wait)) | |
143 | wake_up_all(&rqw->wait); | |
144 | } | |
145 | } | |
146 | ||
147 | void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct) | |
148 | { | |
149 | struct rq_wait *rqw; | |
150 | int inflight, limit; | |
151 | ||
152 | if (!(wb_acct & WBT_TRACKED)) | |
153 | return; | |
154 | ||
8bea6090 | 155 | rqw = get_rq_wait(rwb, wb_acct); |
e34cbd30 JA |
156 | inflight = atomic_dec_return(&rqw->inflight); |
157 | ||
158 | /* | |
159 | * wbt got disabled with IO in flight. Wake up any potential | |
160 | * waiters, we don't have to do more than that. | |
161 | */ | |
162 | if (unlikely(!rwb_enabled(rwb))) { | |
163 | rwb_wake_all(rwb); | |
164 | return; | |
165 | } | |
166 | ||
167 | /* | |
782f5697 JA |
168 | * For discards, our limit is always the background. For writes, if |
169 | * the device does write back caching, drop further down before we | |
170 | * wake people up. | |
e34cbd30 | 171 | */ |
782f5697 JA |
172 | if (wb_acct & WBT_DISCARD) |
173 | limit = rwb->wb_background; | |
174 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
175 | limit = 0; |
176 | else | |
177 | limit = rwb->wb_normal; | |
178 | ||
179 | /* | |
180 | * Don't wake anyone up if we are above the normal limit. | |
181 | */ | |
182 | if (inflight && inflight >= limit) | |
183 | return; | |
184 | ||
185 | if (waitqueue_active(&rqw->wait)) { | |
186 | int diff = limit - inflight; | |
187 | ||
188 | if (!inflight || diff >= rwb->wb_background / 2) | |
189 | wake_up_all(&rqw->wait); | |
190 | } | |
191 | } | |
192 | ||
193 | /* | |
194 | * Called on completion of a request. Note that it's also called when | |
195 | * a request is merged, when the request gets freed. | |
196 | */ | |
a8a45941 | 197 | void wbt_done(struct rq_wb *rwb, struct request *rq) |
e34cbd30 JA |
198 | { |
199 | if (!rwb) | |
200 | return; | |
201 | ||
a8a45941 OS |
202 | if (!wbt_is_tracked(rq)) { |
203 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
204 | rwb->sync_issue = 0; |
205 | rwb->sync_cookie = NULL; | |
206 | } | |
207 | ||
a8a45941 | 208 | if (wbt_is_read(rq)) |
e34cbd30 | 209 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 210 | } else { |
a8a45941 OS |
211 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
212 | __wbt_done(rwb, wbt_flags(rq)); | |
e34cbd30 | 213 | } |
a8a45941 | 214 | wbt_clear_state(rq); |
e34cbd30 JA |
215 | } |
216 | ||
217 | /* | |
218 | * Return true, if we can't increase the depth further by scaling | |
219 | */ | |
220 | static bool calc_wb_limits(struct rq_wb *rwb) | |
221 | { | |
222 | unsigned int depth; | |
223 | bool ret = false; | |
224 | ||
225 | if (!rwb->min_lat_nsec) { | |
226 | rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; | |
227 | return false; | |
228 | } | |
229 | ||
230 | /* | |
231 | * For QD=1 devices, this is a special case. It's important for those | |
232 | * to have one request ready when one completes, so force a depth of | |
233 | * 2 for those devices. On the backend, it'll be a depth of 1 anyway, | |
234 | * since the device can't have more than that in flight. If we're | |
235 | * scaling down, then keep a setting of 1/1/1. | |
236 | */ | |
237 | if (rwb->queue_depth == 1) { | |
238 | if (rwb->scale_step > 0) | |
239 | rwb->wb_max = rwb->wb_normal = 1; | |
240 | else { | |
241 | rwb->wb_max = rwb->wb_normal = 2; | |
242 | ret = true; | |
243 | } | |
244 | rwb->wb_background = 1; | |
245 | } else { | |
246 | /* | |
247 | * scale_step == 0 is our default state. If we have suffered | |
248 | * latency spikes, step will be > 0, and we shrink the | |
249 | * allowed write depths. If step is < 0, we're only doing | |
250 | * writes, and we allow a temporarily higher depth to | |
251 | * increase performance. | |
252 | */ | |
253 | depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth); | |
254 | if (rwb->scale_step > 0) | |
255 | depth = 1 + ((depth - 1) >> min(31, rwb->scale_step)); | |
256 | else if (rwb->scale_step < 0) { | |
257 | unsigned int maxd = 3 * rwb->queue_depth / 4; | |
258 | ||
259 | depth = 1 + ((depth - 1) << -rwb->scale_step); | |
260 | if (depth > maxd) { | |
261 | depth = maxd; | |
262 | ret = true; | |
263 | } | |
264 | } | |
265 | ||
266 | /* | |
267 | * Set our max/normal/bg queue depths based on how far | |
268 | * we have scaled down (->scale_step). | |
269 | */ | |
270 | rwb->wb_max = depth; | |
271 | rwb->wb_normal = (rwb->wb_max + 1) / 2; | |
272 | rwb->wb_background = (rwb->wb_max + 3) / 4; | |
273 | } | |
274 | ||
275 | return ret; | |
276 | } | |
277 | ||
4121d385 | 278 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
279 | { |
280 | /* | |
281 | * We need at least one read sample, and a minimum of | |
282 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
283 | * that it's writes impacting us, and not just some sole read on | |
284 | * a device that is in a lower power state. | |
285 | */ | |
fa2e39cb OS |
286 | return (stat[READ].nr_samples >= 1 && |
287 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
288 | } |
289 | ||
290 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
291 | { | |
6aa7de05 | 292 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
293 | |
294 | if (!issue || !rwb->sync_cookie) | |
295 | return 0; | |
296 | ||
297 | now = ktime_to_ns(ktime_get()); | |
298 | return now - issue; | |
299 | } | |
300 | ||
301 | enum { | |
302 | LAT_OK = 1, | |
303 | LAT_UNKNOWN, | |
304 | LAT_UNKNOWN_WRITES, | |
305 | LAT_EXCEEDED, | |
306 | }; | |
307 | ||
34dbad5d | 308 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 309 | { |
dc3b17cc | 310 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; |
e34cbd30 JA |
311 | u64 thislat; |
312 | ||
313 | /* | |
314 | * If our stored sync issue exceeds the window size, or it | |
315 | * exceeds our min target AND we haven't logged any entries, | |
316 | * flag the latency as exceeded. wbt works off completion latencies, | |
317 | * but for a flooded device, a single sync IO can take a long time | |
318 | * to complete after being issued. If this time exceeds our | |
319 | * monitoring window AND we didn't see any other completions in that | |
320 | * window, then count that sync IO as a violation of the latency. | |
321 | */ | |
322 | thislat = rwb_sync_issue_lat(rwb); | |
323 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 324 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 325 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
326 | return LAT_EXCEEDED; |
327 | } | |
328 | ||
329 | /* | |
330 | * No read/write mix, if stat isn't valid | |
331 | */ | |
332 | if (!stat_sample_valid(stat)) { | |
333 | /* | |
334 | * If we had writes in this stat window and the window is | |
335 | * current, we're only doing writes. If a task recently | |
336 | * waited or still has writes in flights, consider us doing | |
337 | * just writes as well. | |
338 | */ | |
34dbad5d OS |
339 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
340 | wbt_inflight(rwb)) | |
e34cbd30 JA |
341 | return LAT_UNKNOWN_WRITES; |
342 | return LAT_UNKNOWN; | |
343 | } | |
344 | ||
345 | /* | |
346 | * If the 'min' latency exceeds our target, step down. | |
347 | */ | |
fa2e39cb OS |
348 | if (stat[READ].min > rwb->min_lat_nsec) { |
349 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 350 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
351 | return LAT_EXCEEDED; |
352 | } | |
353 | ||
354 | if (rwb->scale_step) | |
d8a0cbfd | 355 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
356 | |
357 | return LAT_OK; | |
358 | } | |
359 | ||
e34cbd30 JA |
360 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
361 | { | |
dc3b17cc | 362 | struct backing_dev_info *bdi = rwb->queue->backing_dev_info; |
d8a0cbfd JA |
363 | |
364 | trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec, | |
e34cbd30 JA |
365 | rwb->wb_background, rwb->wb_normal, rwb->wb_max); |
366 | } | |
367 | ||
368 | static void scale_up(struct rq_wb *rwb) | |
369 | { | |
370 | /* | |
371 | * Hit max in previous round, stop here | |
372 | */ | |
373 | if (rwb->scaled_max) | |
374 | return; | |
375 | ||
376 | rwb->scale_step--; | |
377 | rwb->unknown_cnt = 0; | |
e34cbd30 JA |
378 | |
379 | rwb->scaled_max = calc_wb_limits(rwb); | |
380 | ||
381 | rwb_wake_all(rwb); | |
382 | ||
383 | rwb_trace_step(rwb, "step up"); | |
384 | } | |
385 | ||
386 | /* | |
387 | * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we | |
388 | * had a latency violation. | |
389 | */ | |
390 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) | |
391 | { | |
392 | /* | |
393 | * Stop scaling down when we've hit the limit. This also prevents | |
394 | * ->scale_step from going to crazy values, if the device can't | |
395 | * keep up. | |
396 | */ | |
397 | if (rwb->wb_max == 1) | |
398 | return; | |
399 | ||
400 | if (rwb->scale_step < 0 && hard_throttle) | |
401 | rwb->scale_step = 0; | |
402 | else | |
403 | rwb->scale_step++; | |
404 | ||
405 | rwb->scaled_max = false; | |
406 | rwb->unknown_cnt = 0; | |
e34cbd30 JA |
407 | calc_wb_limits(rwb); |
408 | rwb_trace_step(rwb, "step down"); | |
409 | } | |
410 | ||
411 | static void rwb_arm_timer(struct rq_wb *rwb) | |
412 | { | |
e34cbd30 JA |
413 | if (rwb->scale_step > 0) { |
414 | /* | |
415 | * We should speed this up, using some variant of a fast | |
416 | * integer inverse square root calculation. Since we only do | |
417 | * this for every window expiration, it's not a huge deal, | |
418 | * though. | |
419 | */ | |
420 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
421 | int_sqrt((rwb->scale_step + 1) << 8)); | |
422 | } else { | |
423 | /* | |
424 | * For step < 0, we don't want to increase/decrease the | |
425 | * window size. | |
426 | */ | |
427 | rwb->cur_win_nsec = rwb->win_nsec; | |
428 | } | |
429 | ||
34dbad5d | 430 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
431 | } |
432 | ||
34dbad5d | 433 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 434 | { |
34dbad5d | 435 | struct rq_wb *rwb = cb->data; |
e34cbd30 JA |
436 | unsigned int inflight = wbt_inflight(rwb); |
437 | int status; | |
438 | ||
34dbad5d | 439 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 440 | |
dc3b17cc | 441 | trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step, |
d8a0cbfd | 442 | inflight); |
e34cbd30 JA |
443 | |
444 | /* | |
445 | * If we exceeded the latency target, step down. If we did not, | |
446 | * step one level up. If we don't know enough to say either exceeded | |
447 | * or ok, then don't do anything. | |
448 | */ | |
449 | switch (status) { | |
450 | case LAT_EXCEEDED: | |
451 | scale_down(rwb, true); | |
452 | break; | |
453 | case LAT_OK: | |
454 | scale_up(rwb); | |
455 | break; | |
456 | case LAT_UNKNOWN_WRITES: | |
457 | /* | |
458 | * We started a the center step, but don't have a valid | |
459 | * read/write sample, but we do have writes going on. | |
460 | * Allow step to go negative, to increase write perf. | |
461 | */ | |
462 | scale_up(rwb); | |
463 | break; | |
464 | case LAT_UNKNOWN: | |
465 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
466 | break; | |
467 | /* | |
468 | * We get here when previously scaled reduced depth, and we | |
469 | * currently don't have a valid read/write sample. For that | |
470 | * case, slowly return to center state (step == 0). | |
471 | */ | |
472 | if (rwb->scale_step > 0) | |
473 | scale_up(rwb); | |
474 | else if (rwb->scale_step < 0) | |
475 | scale_down(rwb, false); | |
476 | break; | |
477 | default: | |
478 | break; | |
479 | } | |
480 | ||
481 | /* | |
482 | * Re-arm timer, if we have IO in flight | |
483 | */ | |
484 | if (rwb->scale_step || inflight) | |
485 | rwb_arm_timer(rwb); | |
486 | } | |
487 | ||
488 | void wbt_update_limits(struct rq_wb *rwb) | |
489 | { | |
490 | rwb->scale_step = 0; | |
491 | rwb->scaled_max = false; | |
492 | calc_wb_limits(rwb); | |
493 | ||
494 | rwb_wake_all(rwb); | |
495 | } | |
496 | ||
497 | static bool close_io(struct rq_wb *rwb) | |
498 | { | |
499 | const unsigned long now = jiffies; | |
500 | ||
501 | return time_before(now, rwb->last_issue + HZ / 10) || | |
502 | time_before(now, rwb->last_comp + HZ / 10); | |
503 | } | |
504 | ||
505 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
506 | ||
507 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
508 | { | |
509 | unsigned int limit; | |
510 | ||
782f5697 JA |
511 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) |
512 | return rwb->wb_background; | |
513 | ||
e34cbd30 JA |
514 | /* |
515 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 516 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
517 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
518 | * that. If the write is marked as a background write, then use | |
519 | * the idle limit, or go to normal if we haven't had competing | |
520 | * IO for a bit. | |
521 | */ | |
522 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
523 | limit = rwb->wb_max; | |
524 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { | |
525 | /* | |
526 | * If less than 100ms since we completed unrelated IO, | |
527 | * limit us to half the depth for background writeback. | |
528 | */ | |
529 | limit = rwb->wb_background; | |
530 | } else | |
531 | limit = rwb->wb_normal; | |
532 | ||
533 | return limit; | |
534 | } | |
535 | ||
536 | static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw, | |
ac6424b9 | 537 | wait_queue_entry_t *wait, unsigned long rw) |
e34cbd30 JA |
538 | { |
539 | /* | |
540 | * inc it here even if disabled, since we'll dec it at completion. | |
541 | * this only happens if the task was sleeping in __wbt_wait(), | |
542 | * and someone turned it off at the same time. | |
543 | */ | |
544 | if (!rwb_enabled(rwb)) { | |
545 | atomic_inc(&rqw->inflight); | |
546 | return true; | |
547 | } | |
548 | ||
549 | /* | |
550 | * If the waitqueue is already active and we are not the next | |
551 | * in line to be woken up, wait for our turn. | |
552 | */ | |
553 | if (waitqueue_active(&rqw->wait) && | |
2055da97 | 554 | rqw->wait.head.next != &wait->entry) |
e34cbd30 JA |
555 | return false; |
556 | ||
557 | return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw)); | |
558 | } | |
559 | ||
560 | /* | |
561 | * Block if we will exceed our limit, or if we are currently waiting for | |
562 | * the timer to kick off queuing again. | |
563 | */ | |
8bea6090 JA |
564 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
565 | unsigned long rw, spinlock_t *lock) | |
9eca5350 BVA |
566 | __releases(lock) |
567 | __acquires(lock) | |
e34cbd30 | 568 | { |
8bea6090 | 569 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
e34cbd30 JA |
570 | DEFINE_WAIT(wait); |
571 | ||
572 | if (may_queue(rwb, rqw, &wait, rw)) | |
573 | return; | |
574 | ||
575 | do { | |
576 | prepare_to_wait_exclusive(&rqw->wait, &wait, | |
577 | TASK_UNINTERRUPTIBLE); | |
578 | ||
579 | if (may_queue(rwb, rqw, &wait, rw)) | |
580 | break; | |
581 | ||
9eca5350 | 582 | if (lock) { |
e34cbd30 | 583 | spin_unlock_irq(lock); |
9eca5350 | 584 | io_schedule(); |
e34cbd30 | 585 | spin_lock_irq(lock); |
9eca5350 BVA |
586 | } else |
587 | io_schedule(); | |
e34cbd30 JA |
588 | } while (1); |
589 | ||
590 | finish_wait(&rqw->wait, &wait); | |
591 | } | |
592 | ||
593 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
594 | { | |
782f5697 JA |
595 | switch (bio_op(bio)) { |
596 | case REQ_OP_WRITE: | |
597 | /* | |
598 | * Don't throttle WRITE_ODIRECT | |
599 | */ | |
600 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
601 | (REQ_SYNC | REQ_IDLE)) | |
602 | return false; | |
603 | /* fallthrough */ | |
604 | case REQ_OP_DISCARD: | |
605 | return true; | |
606 | default: | |
e34cbd30 | 607 | return false; |
782f5697 | 608 | } |
e34cbd30 JA |
609 | } |
610 | ||
611 | /* | |
612 | * Returns true if the IO request should be accounted, false if not. | |
613 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
614 | * in an irq held spinlock, if it holds one when calling this function. | |
615 | * If we do sleep, we'll release and re-grab it. | |
616 | */ | |
f2e0a0b2 | 617 | enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock) |
e34cbd30 | 618 | { |
8bea6090 | 619 | enum wbt_flags ret = 0; |
e34cbd30 JA |
620 | |
621 | if (!rwb_enabled(rwb)) | |
622 | return 0; | |
623 | ||
624 | if (bio_op(bio) == REQ_OP_READ) | |
625 | ret = WBT_READ; | |
626 | ||
627 | if (!wbt_should_throttle(rwb, bio)) { | |
628 | if (ret & WBT_READ) | |
629 | wb_timestamp(rwb, &rwb->last_issue); | |
630 | return ret; | |
631 | } | |
632 | ||
8bea6090 JA |
633 | if (current_is_kswapd()) |
634 | ret |= WBT_KSWAPD; | |
782f5697 JA |
635 | if (bio_op(bio) == REQ_OP_DISCARD) |
636 | ret |= WBT_DISCARD; | |
8bea6090 JA |
637 | |
638 | __wbt_wait(rwb, ret, bio->bi_opf, lock); | |
e34cbd30 | 639 | |
34dbad5d | 640 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 JA |
641 | rwb_arm_timer(rwb); |
642 | ||
e34cbd30 JA |
643 | return ret | WBT_TRACKED; |
644 | } | |
645 | ||
a8a45941 | 646 | void wbt_issue(struct rq_wb *rwb, struct request *rq) |
e34cbd30 JA |
647 | { |
648 | if (!rwb_enabled(rwb)) | |
649 | return; | |
650 | ||
651 | /* | |
a8a45941 OS |
652 | * Track sync issue, in case it takes a long time to complete. Allows us |
653 | * to react quicker, if a sync IO takes a long time to complete. Note | |
654 | * that this is just a hint. The request can go away when it completes, | |
655 | * so it's important we never dereference it. We only use the address to | |
656 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 657 | */ |
a8a45941 OS |
658 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
659 | rwb->sync_cookie = rq; | |
544ccc8d | 660 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
661 | } |
662 | } | |
663 | ||
a8a45941 | 664 | void wbt_requeue(struct rq_wb *rwb, struct request *rq) |
e34cbd30 JA |
665 | { |
666 | if (!rwb_enabled(rwb)) | |
667 | return; | |
a8a45941 | 668 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
669 | rwb->sync_issue = 0; |
670 | rwb->sync_cookie = NULL; | |
671 | } | |
672 | } | |
673 | ||
674 | void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) | |
675 | { | |
676 | if (rwb) { | |
677 | rwb->queue_depth = depth; | |
678 | wbt_update_limits(rwb); | |
679 | } | |
680 | } | |
681 | ||
682 | void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) | |
683 | { | |
684 | if (rwb) | |
685 | rwb->wc = write_cache_on; | |
686 | } | |
687 | ||
3f19cd23 | 688 | /* |
b5dc5d4d | 689 | * Disable wbt, if enabled by default. |
fa224eed JA |
690 | */ |
691 | void wbt_disable_default(struct request_queue *q) | |
e34cbd30 | 692 | { |
fa224eed JA |
693 | struct rq_wb *rwb = q->rq_wb; |
694 | ||
3f19cd23 JK |
695 | if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) |
696 | wbt_exit(q); | |
e34cbd30 | 697 | } |
fa224eed | 698 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
e34cbd30 | 699 | |
8330cdb0 JK |
700 | /* |
701 | * Enable wbt if defaults are configured that way | |
702 | */ | |
703 | void wbt_enable_default(struct request_queue *q) | |
704 | { | |
705 | /* Throttling already enabled? */ | |
706 | if (q->rq_wb) | |
707 | return; | |
708 | ||
709 | /* Queue not registered? Maybe shutting down... */ | |
710 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
711 | return; | |
712 | ||
713 | if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) || | |
714 | (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ))) | |
715 | wbt_init(q); | |
716 | } | |
717 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
718 | ||
80e091d1 JA |
719 | u64 wbt_default_latency_nsec(struct request_queue *q) |
720 | { | |
721 | /* | |
722 | * We default to 2msec for non-rotational storage, and 75msec | |
723 | * for rotational storage. | |
724 | */ | |
725 | if (blk_queue_nonrot(q)) | |
726 | return 2000000ULL; | |
727 | else | |
728 | return 75000000ULL; | |
729 | } | |
730 | ||
99c749a4 JA |
731 | static int wbt_data_dir(const struct request *rq) |
732 | { | |
5235553d JA |
733 | const int op = req_op(rq); |
734 | ||
735 | if (op == REQ_OP_READ) | |
736 | return READ; | |
825843b0 | 737 | else if (op_is_write(op)) |
5235553d JA |
738 | return WRITE; |
739 | ||
740 | /* don't account */ | |
741 | return -1; | |
99c749a4 JA |
742 | } |
743 | ||
8054b89f | 744 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
745 | { |
746 | struct rq_wb *rwb; | |
747 | int i; | |
748 | ||
e34cbd30 JA |
749 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
750 | if (!rwb) | |
751 | return -ENOMEM; | |
752 | ||
99c749a4 | 753 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
754 | if (!rwb->cb) { |
755 | kfree(rwb); | |
756 | return -ENOMEM; | |
757 | } | |
758 | ||
e34cbd30 JA |
759 | for (i = 0; i < WBT_NUM_RWQ; i++) { |
760 | atomic_set(&rwb->rq_wait[i].inflight, 0); | |
761 | init_waitqueue_head(&rwb->rq_wait[i].wait); | |
762 | } | |
763 | ||
e34cbd30 | 764 | rwb->last_comp = rwb->last_issue = jiffies; |
d8a0cbfd | 765 | rwb->queue = q; |
e34cbd30 | 766 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 767 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
e34cbd30 JA |
768 | wbt_update_limits(rwb); |
769 | ||
770 | /* | |
34dbad5d | 771 | * Assign rwb and add the stats callback. |
e34cbd30 JA |
772 | */ |
773 | q->rq_wb = rwb; | |
34dbad5d | 774 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 775 | |
80e091d1 | 776 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 JA |
777 | |
778 | wbt_set_queue_depth(rwb, blk_queue_depth(q)); | |
779 | wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
780 | ||
781 | return 0; | |
782 | } | |
783 | ||
784 | void wbt_exit(struct request_queue *q) | |
785 | { | |
786 | struct rq_wb *rwb = q->rq_wb; | |
787 | ||
788 | if (rwb) { | |
34dbad5d OS |
789 | blk_stat_remove_callback(q, rwb->cb); |
790 | blk_stat_free_callback(rwb->cb); | |
e34cbd30 JA |
791 | q->rq_wb = NULL; |
792 | kfree(rwb); | |
793 | } | |
794 | } |