]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0 | |
2 | /* | |
3 | * buffered writeback throttling. loosely based on CoDel. We can't drop | |
4 | * packets for IO scheduling, so the logic is something like this: | |
5 | * | |
6 | * - Monitor latencies in a defined window of time. | |
7 | * - If the minimum latency in the above window exceeds some target, increment | |
8 | * scaling step and scale down queue depth by a factor of 2x. The monitoring | |
9 | * window is then shrunk to 100 / sqrt(scaling step + 1). | |
10 | * - For any window where we don't have solid data on what the latencies | |
11 | * look like, retain status quo. | |
12 | * - If latencies look good, decrement scaling step. | |
13 | * - If we're only doing writes, allow the scaling step to go negative. This | |
14 | * will temporarily boost write performance, snapping back to a stable | |
15 | * scaling step of 0 if reads show up or the heavy writers finish. Unlike | |
16 | * positive scaling steps where we shrink the monitoring window, a negative | |
17 | * scaling step retains the default step==0 window size. | |
18 | * | |
19 | * Copyright (C) 2016 Jens Axboe | |
20 | * | |
21 | */ | |
22 | #include <linux/kernel.h> | |
23 | #include <linux/blk_types.h> | |
24 | #include <linux/slab.h> | |
25 | #include <linux/backing-dev.h> | |
26 | #include <linux/swap.h> | |
27 | ||
28 | #include "blk-wbt.h" | |
29 | #include "blk-rq-qos.h" | |
30 | ||
31 | #define CREATE_TRACE_POINTS | |
32 | #include <trace/events/wbt.h> | |
33 | ||
34 | static inline void wbt_clear_state(struct request *rq) | |
35 | { | |
36 | rq->wbt_flags = 0; | |
37 | } | |
38 | ||
39 | static inline enum wbt_flags wbt_flags(struct request *rq) | |
40 | { | |
41 | return rq->wbt_flags; | |
42 | } | |
43 | ||
44 | static inline bool wbt_is_tracked(struct request *rq) | |
45 | { | |
46 | return rq->wbt_flags & WBT_TRACKED; | |
47 | } | |
48 | ||
49 | static inline bool wbt_is_read(struct request *rq) | |
50 | { | |
51 | return rq->wbt_flags & WBT_READ; | |
52 | } | |
53 | ||
54 | enum { | |
55 | /* | |
56 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
57 | * from here depending on device stats | |
58 | */ | |
59 | RWB_DEF_DEPTH = 16, | |
60 | ||
61 | /* | |
62 | * 100msec window | |
63 | */ | |
64 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
65 | ||
66 | /* | |
67 | * Disregard stats, if we don't meet this minimum | |
68 | */ | |
69 | RWB_MIN_WRITE_SAMPLES = 3, | |
70 | ||
71 | /* | |
72 | * If we have this number of consecutive windows with not enough | |
73 | * information to scale up or down, scale up. | |
74 | */ | |
75 | RWB_UNKNOWN_BUMP = 5, | |
76 | }; | |
77 | ||
78 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
79 | { | |
80 | return rwb && rwb->wb_normal != 0; | |
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 | { | |
99 | struct bdi_writeback *wb = &rwb->rqos.q->backing_dev_info->wb; | |
100 | ||
101 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
102 | } | |
103 | ||
104 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, | |
105 | enum wbt_flags wb_acct) | |
106 | { | |
107 | if (wb_acct & WBT_KSWAPD) | |
108 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
109 | else if (wb_acct & WBT_DISCARD) | |
110 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
111 | ||
112 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
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 (wq_has_sleeper(&rqw->wait)) | |
123 | wake_up_all(&rqw->wait); | |
124 | } | |
125 | } | |
126 | ||
127 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, | |
128 | enum wbt_flags wb_acct) | |
129 | { | |
130 | int inflight, limit; | |
131 | ||
132 | inflight = atomic_dec_return(&rqw->inflight); | |
133 | ||
134 | /* | |
135 | * wbt got disabled with IO in flight. Wake up any potential | |
136 | * waiters, we don't have to do more than that. | |
137 | */ | |
138 | if (unlikely(!rwb_enabled(rwb))) { | |
139 | rwb_wake_all(rwb); | |
140 | return; | |
141 | } | |
142 | ||
143 | /* | |
144 | * For discards, our limit is always the background. For writes, if | |
145 | * the device does write back caching, drop further down before we | |
146 | * wake people up. | |
147 | */ | |
148 | if (wb_acct & WBT_DISCARD) | |
149 | limit = rwb->wb_background; | |
150 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
151 | limit = 0; | |
152 | else | |
153 | limit = rwb->wb_normal; | |
154 | ||
155 | /* | |
156 | * Don't wake anyone up if we are above the normal limit. | |
157 | */ | |
158 | if (inflight && inflight >= limit) | |
159 | return; | |
160 | ||
161 | if (wq_has_sleeper(&rqw->wait)) { | |
162 | int diff = limit - inflight; | |
163 | ||
164 | if (!inflight || diff >= rwb->wb_background / 2) | |
165 | wake_up_all(&rqw->wait); | |
166 | } | |
167 | } | |
168 | ||
169 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) | |
170 | { | |
171 | struct rq_wb *rwb = RQWB(rqos); | |
172 | struct rq_wait *rqw; | |
173 | ||
174 | if (!(wb_acct & WBT_TRACKED)) | |
175 | return; | |
176 | ||
177 | rqw = get_rq_wait(rwb, wb_acct); | |
178 | wbt_rqw_done(rwb, rqw, wb_acct); | |
179 | } | |
180 | ||
181 | /* | |
182 | * Called on completion of a request. Note that it's also called when | |
183 | * a request is merged, when the request gets freed. | |
184 | */ | |
185 | static void wbt_done(struct rq_qos *rqos, struct request *rq) | |
186 | { | |
187 | struct rq_wb *rwb = RQWB(rqos); | |
188 | ||
189 | if (!wbt_is_tracked(rq)) { | |
190 | if (rwb->sync_cookie == rq) { | |
191 | rwb->sync_issue = 0; | |
192 | rwb->sync_cookie = NULL; | |
193 | } | |
194 | ||
195 | if (wbt_is_read(rq)) | |
196 | wb_timestamp(rwb, &rwb->last_comp); | |
197 | } else { | |
198 | WARN_ON_ONCE(rq == rwb->sync_cookie); | |
199 | __wbt_done(rqos, wbt_flags(rq)); | |
200 | } | |
201 | wbt_clear_state(rq); | |
202 | } | |
203 | ||
204 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) | |
205 | { | |
206 | /* | |
207 | * We need at least one read sample, and a minimum of | |
208 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
209 | * that it's writes impacting us, and not just some sole read on | |
210 | * a device that is in a lower power state. | |
211 | */ | |
212 | return (stat[READ].nr_samples >= 1 && | |
213 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
214 | } | |
215 | ||
216 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
217 | { | |
218 | u64 now, issue = READ_ONCE(rwb->sync_issue); | |
219 | ||
220 | if (!issue || !rwb->sync_cookie) | |
221 | return 0; | |
222 | ||
223 | now = ktime_to_ns(ktime_get()); | |
224 | return now - issue; | |
225 | } | |
226 | ||
227 | enum { | |
228 | LAT_OK = 1, | |
229 | LAT_UNKNOWN, | |
230 | LAT_UNKNOWN_WRITES, | |
231 | LAT_EXCEEDED, | |
232 | }; | |
233 | ||
234 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) | |
235 | { | |
236 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; | |
237 | struct rq_depth *rqd = &rwb->rq_depth; | |
238 | u64 thislat; | |
239 | ||
240 | /* | |
241 | * If our stored sync issue exceeds the window size, or it | |
242 | * exceeds our min target AND we haven't logged any entries, | |
243 | * flag the latency as exceeded. wbt works off completion latencies, | |
244 | * but for a flooded device, a single sync IO can take a long time | |
245 | * to complete after being issued. If this time exceeds our | |
246 | * monitoring window AND we didn't see any other completions in that | |
247 | * window, then count that sync IO as a violation of the latency. | |
248 | */ | |
249 | thislat = rwb_sync_issue_lat(rwb); | |
250 | if (thislat > rwb->cur_win_nsec || | |
251 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { | |
252 | trace_wbt_lat(bdi, thislat); | |
253 | return LAT_EXCEEDED; | |
254 | } | |
255 | ||
256 | /* | |
257 | * No read/write mix, if stat isn't valid | |
258 | */ | |
259 | if (!stat_sample_valid(stat)) { | |
260 | /* | |
261 | * If we had writes in this stat window and the window is | |
262 | * current, we're only doing writes. If a task recently | |
263 | * waited or still has writes in flights, consider us doing | |
264 | * just writes as well. | |
265 | */ | |
266 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || | |
267 | wbt_inflight(rwb)) | |
268 | return LAT_UNKNOWN_WRITES; | |
269 | return LAT_UNKNOWN; | |
270 | } | |
271 | ||
272 | /* | |
273 | * If the 'min' latency exceeds our target, step down. | |
274 | */ | |
275 | if (stat[READ].min > rwb->min_lat_nsec) { | |
276 | trace_wbt_lat(bdi, stat[READ].min); | |
277 | trace_wbt_stat(bdi, stat); | |
278 | return LAT_EXCEEDED; | |
279 | } | |
280 | ||
281 | if (rqd->scale_step) | |
282 | trace_wbt_stat(bdi, stat); | |
283 | ||
284 | return LAT_OK; | |
285 | } | |
286 | ||
287 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) | |
288 | { | |
289 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; | |
290 | struct rq_depth *rqd = &rwb->rq_depth; | |
291 | ||
292 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, | |
293 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
294 | } | |
295 | ||
296 | static void calc_wb_limits(struct rq_wb *rwb) | |
297 | { | |
298 | if (rwb->min_lat_nsec == 0) { | |
299 | rwb->wb_normal = rwb->wb_background = 0; | |
300 | } else if (rwb->rq_depth.max_depth <= 2) { | |
301 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
302 | rwb->wb_background = 1; | |
303 | } else { | |
304 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
305 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
306 | } | |
307 | } | |
308 | ||
309 | static void scale_up(struct rq_wb *rwb) | |
310 | { | |
311 | rq_depth_scale_up(&rwb->rq_depth); | |
312 | calc_wb_limits(rwb); | |
313 | rwb->unknown_cnt = 0; | |
314 | rwb_wake_all(rwb); | |
315 | rwb_trace_step(rwb, "scale up"); | |
316 | } | |
317 | ||
318 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) | |
319 | { | |
320 | rq_depth_scale_down(&rwb->rq_depth, hard_throttle); | |
321 | calc_wb_limits(rwb); | |
322 | rwb->unknown_cnt = 0; | |
323 | rwb_trace_step(rwb, "scale down"); | |
324 | } | |
325 | ||
326 | static void rwb_arm_timer(struct rq_wb *rwb) | |
327 | { | |
328 | struct rq_depth *rqd = &rwb->rq_depth; | |
329 | ||
330 | if (rqd->scale_step > 0) { | |
331 | /* | |
332 | * We should speed this up, using some variant of a fast | |
333 | * integer inverse square root calculation. Since we only do | |
334 | * this for every window expiration, it's not a huge deal, | |
335 | * though. | |
336 | */ | |
337 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
338 | int_sqrt((rqd->scale_step + 1) << 8)); | |
339 | } else { | |
340 | /* | |
341 | * For step < 0, we don't want to increase/decrease the | |
342 | * window size. | |
343 | */ | |
344 | rwb->cur_win_nsec = rwb->win_nsec; | |
345 | } | |
346 | ||
347 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); | |
348 | } | |
349 | ||
350 | static void wb_timer_fn(struct blk_stat_callback *cb) | |
351 | { | |
352 | struct rq_wb *rwb = cb->data; | |
353 | struct rq_depth *rqd = &rwb->rq_depth; | |
354 | unsigned int inflight = wbt_inflight(rwb); | |
355 | int status; | |
356 | ||
357 | status = latency_exceeded(rwb, cb->stat); | |
358 | ||
359 | trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step, | |
360 | inflight); | |
361 | ||
362 | /* | |
363 | * If we exceeded the latency target, step down. If we did not, | |
364 | * step one level up. If we don't know enough to say either exceeded | |
365 | * or ok, then don't do anything. | |
366 | */ | |
367 | switch (status) { | |
368 | case LAT_EXCEEDED: | |
369 | scale_down(rwb, true); | |
370 | break; | |
371 | case LAT_OK: | |
372 | scale_up(rwb); | |
373 | break; | |
374 | case LAT_UNKNOWN_WRITES: | |
375 | /* | |
376 | * We started a the center step, but don't have a valid | |
377 | * read/write sample, but we do have writes going on. | |
378 | * Allow step to go negative, to increase write perf. | |
379 | */ | |
380 | scale_up(rwb); | |
381 | break; | |
382 | case LAT_UNKNOWN: | |
383 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
384 | break; | |
385 | /* | |
386 | * We get here when previously scaled reduced depth, and we | |
387 | * currently don't have a valid read/write sample. For that | |
388 | * case, slowly return to center state (step == 0). | |
389 | */ | |
390 | if (rqd->scale_step > 0) | |
391 | scale_up(rwb); | |
392 | else if (rqd->scale_step < 0) | |
393 | scale_down(rwb, false); | |
394 | break; | |
395 | default: | |
396 | break; | |
397 | } | |
398 | ||
399 | /* | |
400 | * Re-arm timer, if we have IO in flight | |
401 | */ | |
402 | if (rqd->scale_step || inflight) | |
403 | rwb_arm_timer(rwb); | |
404 | } | |
405 | ||
406 | static void __wbt_update_limits(struct rq_wb *rwb) | |
407 | { | |
408 | struct rq_depth *rqd = &rwb->rq_depth; | |
409 | ||
410 | rqd->scale_step = 0; | |
411 | rqd->scaled_max = false; | |
412 | ||
413 | rq_depth_calc_max_depth(rqd); | |
414 | calc_wb_limits(rwb); | |
415 | ||
416 | rwb_wake_all(rwb); | |
417 | } | |
418 | ||
419 | void wbt_update_limits(struct request_queue *q) | |
420 | { | |
421 | struct rq_qos *rqos = wbt_rq_qos(q); | |
422 | if (!rqos) | |
423 | return; | |
424 | __wbt_update_limits(RQWB(rqos)); | |
425 | } | |
426 | ||
427 | u64 wbt_get_min_lat(struct request_queue *q) | |
428 | { | |
429 | struct rq_qos *rqos = wbt_rq_qos(q); | |
430 | if (!rqos) | |
431 | return 0; | |
432 | return RQWB(rqos)->min_lat_nsec; | |
433 | } | |
434 | ||
435 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
436 | { | |
437 | struct rq_qos *rqos = wbt_rq_qos(q); | |
438 | if (!rqos) | |
439 | return; | |
440 | RQWB(rqos)->min_lat_nsec = val; | |
441 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
442 | __wbt_update_limits(RQWB(rqos)); | |
443 | } | |
444 | ||
445 | ||
446 | static bool close_io(struct rq_wb *rwb) | |
447 | { | |
448 | const unsigned long now = jiffies; | |
449 | ||
450 | return time_before(now, rwb->last_issue + HZ / 10) || | |
451 | time_before(now, rwb->last_comp + HZ / 10); | |
452 | } | |
453 | ||
454 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
455 | ||
456 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
457 | { | |
458 | unsigned int limit; | |
459 | ||
460 | /* | |
461 | * If we got disabled, just return UINT_MAX. This ensures that | |
462 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
463 | */ | |
464 | if (!rwb_enabled(rwb)) | |
465 | return UINT_MAX; | |
466 | ||
467 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) | |
468 | return rwb->wb_background; | |
469 | ||
470 | /* | |
471 | * At this point we know it's a buffered write. If this is | |
472 | * kswapd trying to free memory, or REQ_SYNC is set, then | |
473 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for | |
474 | * that. If the write is marked as a background write, then use | |
475 | * the idle limit, or go to normal if we haven't had competing | |
476 | * IO for a bit. | |
477 | */ | |
478 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
479 | limit = rwb->rq_depth.max_depth; | |
480 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { | |
481 | /* | |
482 | * If less than 100ms since we completed unrelated IO, | |
483 | * limit us to half the depth for background writeback. | |
484 | */ | |
485 | limit = rwb->wb_background; | |
486 | } else | |
487 | limit = rwb->wb_normal; | |
488 | ||
489 | return limit; | |
490 | } | |
491 | ||
492 | struct wbt_wait_data { | |
493 | struct rq_wb *rwb; | |
494 | enum wbt_flags wb_acct; | |
495 | unsigned long rw; | |
496 | }; | |
497 | ||
498 | static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data) | |
499 | { | |
500 | struct wbt_wait_data *data = private_data; | |
501 | return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw)); | |
502 | } | |
503 | ||
504 | static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data) | |
505 | { | |
506 | struct wbt_wait_data *data = private_data; | |
507 | wbt_rqw_done(data->rwb, rqw, data->wb_acct); | |
508 | } | |
509 | ||
510 | /* | |
511 | * Block if we will exceed our limit, or if we are currently waiting for | |
512 | * the timer to kick off queuing again. | |
513 | */ | |
514 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, | |
515 | unsigned long rw) | |
516 | { | |
517 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); | |
518 | struct wbt_wait_data data = { | |
519 | .rwb = rwb, | |
520 | .wb_acct = wb_acct, | |
521 | .rw = rw, | |
522 | }; | |
523 | ||
524 | rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb); | |
525 | } | |
526 | ||
527 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
528 | { | |
529 | switch (bio_op(bio)) { | |
530 | case REQ_OP_WRITE: | |
531 | /* | |
532 | * Don't throttle WRITE_ODIRECT | |
533 | */ | |
534 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
535 | (REQ_SYNC | REQ_IDLE)) | |
536 | return false; | |
537 | /* fallthrough */ | |
538 | case REQ_OP_DISCARD: | |
539 | return true; | |
540 | default: | |
541 | return false; | |
542 | } | |
543 | } | |
544 | ||
545 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) | |
546 | { | |
547 | enum wbt_flags flags = 0; | |
548 | ||
549 | if (!rwb_enabled(rwb)) | |
550 | return 0; | |
551 | ||
552 | if (bio_op(bio) == REQ_OP_READ) { | |
553 | flags = WBT_READ; | |
554 | } else if (wbt_should_throttle(rwb, bio)) { | |
555 | if (current_is_kswapd()) | |
556 | flags |= WBT_KSWAPD; | |
557 | if (bio_op(bio) == REQ_OP_DISCARD) | |
558 | flags |= WBT_DISCARD; | |
559 | flags |= WBT_TRACKED; | |
560 | } | |
561 | return flags; | |
562 | } | |
563 | ||
564 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
565 | { | |
566 | struct rq_wb *rwb = RQWB(rqos); | |
567 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
568 | __wbt_done(rqos, flags); | |
569 | } | |
570 | ||
571 | /* | |
572 | * Returns true if the IO request should be accounted, false if not. | |
573 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
574 | * in an irq held spinlock, if it holds one when calling this function. | |
575 | * If we do sleep, we'll release and re-grab it. | |
576 | */ | |
577 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio) | |
578 | { | |
579 | struct rq_wb *rwb = RQWB(rqos); | |
580 | enum wbt_flags flags; | |
581 | ||
582 | flags = bio_to_wbt_flags(rwb, bio); | |
583 | if (!(flags & WBT_TRACKED)) { | |
584 | if (flags & WBT_READ) | |
585 | wb_timestamp(rwb, &rwb->last_issue); | |
586 | return; | |
587 | } | |
588 | ||
589 | __wbt_wait(rwb, flags, bio->bi_opf); | |
590 | ||
591 | if (!blk_stat_is_active(rwb->cb)) | |
592 | rwb_arm_timer(rwb); | |
593 | } | |
594 | ||
595 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) | |
596 | { | |
597 | struct rq_wb *rwb = RQWB(rqos); | |
598 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
599 | } | |
600 | ||
601 | static void wbt_issue(struct rq_qos *rqos, struct request *rq) | |
602 | { | |
603 | struct rq_wb *rwb = RQWB(rqos); | |
604 | ||
605 | if (!rwb_enabled(rwb)) | |
606 | return; | |
607 | ||
608 | /* | |
609 | * Track sync issue, in case it takes a long time to complete. Allows us | |
610 | * to react quicker, if a sync IO takes a long time to complete. Note | |
611 | * that this is just a hint. The request can go away when it completes, | |
612 | * so it's important we never dereference it. We only use the address to | |
613 | * compare with, which is why we store the sync_issue time locally. | |
614 | */ | |
615 | if (wbt_is_read(rq) && !rwb->sync_issue) { | |
616 | rwb->sync_cookie = rq; | |
617 | rwb->sync_issue = rq->io_start_time_ns; | |
618 | } | |
619 | } | |
620 | ||
621 | static void wbt_requeue(struct rq_qos *rqos, struct request *rq) | |
622 | { | |
623 | struct rq_wb *rwb = RQWB(rqos); | |
624 | if (!rwb_enabled(rwb)) | |
625 | return; | |
626 | if (rq == rwb->sync_cookie) { | |
627 | rwb->sync_issue = 0; | |
628 | rwb->sync_cookie = NULL; | |
629 | } | |
630 | } | |
631 | ||
632 | void wbt_set_queue_depth(struct request_queue *q, unsigned int depth) | |
633 | { | |
634 | struct rq_qos *rqos = wbt_rq_qos(q); | |
635 | if (rqos) { | |
636 | RQWB(rqos)->rq_depth.queue_depth = depth; | |
637 | __wbt_update_limits(RQWB(rqos)); | |
638 | } | |
639 | } | |
640 | ||
641 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) | |
642 | { | |
643 | struct rq_qos *rqos = wbt_rq_qos(q); | |
644 | if (rqos) | |
645 | RQWB(rqos)->wc = write_cache_on; | |
646 | } | |
647 | ||
648 | /* | |
649 | * Enable wbt if defaults are configured that way | |
650 | */ | |
651 | void wbt_enable_default(struct request_queue *q) | |
652 | { | |
653 | struct rq_qos *rqos = wbt_rq_qos(q); | |
654 | /* Throttling already enabled? */ | |
655 | if (rqos) | |
656 | return; | |
657 | ||
658 | /* Queue not registered? Maybe shutting down... */ | |
659 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
660 | return; | |
661 | ||
662 | if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ)) | |
663 | wbt_init(q); | |
664 | } | |
665 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
666 | ||
667 | u64 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 | ||
679 | static int wbt_data_dir(const struct request *rq) | |
680 | { | |
681 | const int op = req_op(rq); | |
682 | ||
683 | if (op == REQ_OP_READ) | |
684 | return READ; | |
685 | else if (op_is_write(op)) | |
686 | return WRITE; | |
687 | ||
688 | /* don't account */ | |
689 | return -1; | |
690 | } | |
691 | ||
692 | static 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 | */ | |
705 | void 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 | blk_stat_deactivate(rwb->cb); | |
714 | rwb->wb_normal = 0; | |
715 | } | |
716 | } | |
717 | EXPORT_SYMBOL_GPL(wbt_disable_default); | |
718 | ||
719 | #ifdef CONFIG_BLK_DEBUG_FS | |
720 | static int wbt_curr_win_nsec_show(void *data, struct seq_file *m) | |
721 | { | |
722 | struct rq_qos *rqos = data; | |
723 | struct rq_wb *rwb = RQWB(rqos); | |
724 | ||
725 | seq_printf(m, "%llu\n", rwb->cur_win_nsec); | |
726 | return 0; | |
727 | } | |
728 | ||
729 | static int wbt_enabled_show(void *data, struct seq_file *m) | |
730 | { | |
731 | struct rq_qos *rqos = data; | |
732 | struct rq_wb *rwb = RQWB(rqos); | |
733 | ||
734 | seq_printf(m, "%d\n", rwb->enable_state); | |
735 | return 0; | |
736 | } | |
737 | ||
738 | static int wbt_id_show(void *data, struct seq_file *m) | |
739 | { | |
740 | struct rq_qos *rqos = data; | |
741 | ||
742 | seq_printf(m, "%u\n", rqos->id); | |
743 | return 0; | |
744 | } | |
745 | ||
746 | static int wbt_inflight_show(void *data, struct seq_file *m) | |
747 | { | |
748 | struct rq_qos *rqos = data; | |
749 | struct rq_wb *rwb = RQWB(rqos); | |
750 | int i; | |
751 | ||
752 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
753 | seq_printf(m, "%d: inflight %d\n", i, | |
754 | atomic_read(&rwb->rq_wait[i].inflight)); | |
755 | return 0; | |
756 | } | |
757 | ||
758 | static int wbt_min_lat_nsec_show(void *data, struct seq_file *m) | |
759 | { | |
760 | struct rq_qos *rqos = data; | |
761 | struct rq_wb *rwb = RQWB(rqos); | |
762 | ||
763 | seq_printf(m, "%lu\n", rwb->min_lat_nsec); | |
764 | return 0; | |
765 | } | |
766 | ||
767 | static int wbt_unknown_cnt_show(void *data, struct seq_file *m) | |
768 | { | |
769 | struct rq_qos *rqos = data; | |
770 | struct rq_wb *rwb = RQWB(rqos); | |
771 | ||
772 | seq_printf(m, "%u\n", rwb->unknown_cnt); | |
773 | return 0; | |
774 | } | |
775 | ||
776 | static int wbt_normal_show(void *data, struct seq_file *m) | |
777 | { | |
778 | struct rq_qos *rqos = data; | |
779 | struct rq_wb *rwb = RQWB(rqos); | |
780 | ||
781 | seq_printf(m, "%u\n", rwb->wb_normal); | |
782 | return 0; | |
783 | } | |
784 | ||
785 | static int wbt_background_show(void *data, struct seq_file *m) | |
786 | { | |
787 | struct rq_qos *rqos = data; | |
788 | struct rq_wb *rwb = RQWB(rqos); | |
789 | ||
790 | seq_printf(m, "%u\n", rwb->wb_background); | |
791 | return 0; | |
792 | } | |
793 | ||
794 | static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = { | |
795 | {"curr_win_nsec", 0400, wbt_curr_win_nsec_show}, | |
796 | {"enabled", 0400, wbt_enabled_show}, | |
797 | {"id", 0400, wbt_id_show}, | |
798 | {"inflight", 0400, wbt_inflight_show}, | |
799 | {"min_lat_nsec", 0400, wbt_min_lat_nsec_show}, | |
800 | {"unknown_cnt", 0400, wbt_unknown_cnt_show}, | |
801 | {"wb_normal", 0400, wbt_normal_show}, | |
802 | {"wb_background", 0400, wbt_background_show}, | |
803 | {}, | |
804 | }; | |
805 | #endif | |
806 | ||
807 | static struct rq_qos_ops wbt_rqos_ops = { | |
808 | .throttle = wbt_wait, | |
809 | .issue = wbt_issue, | |
810 | .track = wbt_track, | |
811 | .requeue = wbt_requeue, | |
812 | .done = wbt_done, | |
813 | .cleanup = wbt_cleanup, | |
814 | .exit = wbt_exit, | |
815 | #ifdef CONFIG_BLK_DEBUG_FS | |
816 | .debugfs_attrs = wbt_debugfs_attrs, | |
817 | #endif | |
818 | }; | |
819 | ||
820 | int wbt_init(struct request_queue *q) | |
821 | { | |
822 | struct rq_wb *rwb; | |
823 | int i; | |
824 | ||
825 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); | |
826 | if (!rwb) | |
827 | return -ENOMEM; | |
828 | ||
829 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); | |
830 | if (!rwb->cb) { | |
831 | kfree(rwb); | |
832 | return -ENOMEM; | |
833 | } | |
834 | ||
835 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
836 | rq_wait_init(&rwb->rq_wait[i]); | |
837 | ||
838 | rwb->rqos.id = RQ_QOS_WBT; | |
839 | rwb->rqos.ops = &wbt_rqos_ops; | |
840 | rwb->rqos.q = q; | |
841 | rwb->last_comp = rwb->last_issue = jiffies; | |
842 | rwb->win_nsec = RWB_WINDOW_NSEC; | |
843 | rwb->enable_state = WBT_STATE_ON_DEFAULT; | |
844 | rwb->wc = 1; | |
845 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; | |
846 | __wbt_update_limits(rwb); | |
847 | ||
848 | /* | |
849 | * Assign rwb and add the stats callback. | |
850 | */ | |
851 | rq_qos_add(q, &rwb->rqos); | |
852 | blk_stat_add_callback(q, rwb->cb); | |
853 | ||
854 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); | |
855 | ||
856 | wbt_set_queue_depth(q, blk_queue_depth(q)); | |
857 | wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
858 | ||
859 | return 0; | |
860 | } |