Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
e34cbd30 JA |
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" | |
a7905043 | 29 | #include "blk-rq-qos.h" |
e34cbd30 JA |
30 | |
31 | #define CREATE_TRACE_POINTS | |
32 | #include <trace/events/wbt.h> | |
33 | ||
a8a45941 | 34 | static inline void wbt_clear_state(struct request *rq) |
934031a1 | 35 | { |
544ccc8d | 36 | rq->wbt_flags = 0; |
934031a1 OS |
37 | } |
38 | ||
a8a45941 | 39 | static inline enum wbt_flags wbt_flags(struct request *rq) |
934031a1 | 40 | { |
544ccc8d | 41 | return rq->wbt_flags; |
934031a1 OS |
42 | } |
43 | ||
a8a45941 | 44 | static inline bool wbt_is_tracked(struct request *rq) |
934031a1 | 45 | { |
544ccc8d | 46 | return rq->wbt_flags & WBT_TRACKED; |
934031a1 OS |
47 | } |
48 | ||
a8a45941 | 49 | static inline bool wbt_is_read(struct request *rq) |
934031a1 | 50 | { |
544ccc8d | 51 | return rq->wbt_flags & WBT_READ; |
934031a1 OS |
52 | } |
53 | ||
e34cbd30 JA |
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 | { | |
1d0903d6 ZY |
80 | return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT && |
81 | rwb->wb_normal != 0; | |
e34cbd30 JA |
82 | } |
83 | ||
e34cbd30 JA |
84 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) |
85 | { | |
86 | if (rwb_enabled(rwb)) { | |
87 | const unsigned long cur = jiffies; | |
88 | ||
89 | if (cur != *var) | |
90 | *var = cur; | |
91 | } | |
92 | } | |
93 | ||
94 | /* | |
95 | * If a task was rate throttled in balance_dirty_pages() within the last | |
96 | * second or so, use that to indicate a higher cleaning rate. | |
97 | */ | |
98 | static bool wb_recent_wait(struct rq_wb *rwb) | |
99 | { | |
d152c682 | 100 | struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb; |
e34cbd30 JA |
101 | |
102 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
103 | } | |
104 | ||
8bea6090 JA |
105 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
106 | enum wbt_flags wb_acct) | |
e34cbd30 | 107 | { |
8bea6090 JA |
108 | if (wb_acct & WBT_KSWAPD) |
109 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
110 | else if (wb_acct & WBT_DISCARD) |
111 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
112 | |
113 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
114 | } |
115 | ||
116 | static void rwb_wake_all(struct rq_wb *rwb) | |
117 | { | |
118 | int i; | |
119 | ||
120 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
121 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
122 | ||
b7882093 | 123 | if (wq_has_sleeper(&rqw->wait)) |
e34cbd30 JA |
124 | wake_up_all(&rqw->wait); |
125 | } | |
126 | } | |
127 | ||
061a5427 JA |
128 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
129 | enum wbt_flags wb_acct) | |
e34cbd30 | 130 | { |
e34cbd30 JA |
131 | int inflight, limit; |
132 | ||
e34cbd30 JA |
133 | inflight = atomic_dec_return(&rqw->inflight); |
134 | ||
135 | /* | |
136 | * wbt got disabled with IO in flight. Wake up any potential | |
137 | * waiters, we don't have to do more than that. | |
138 | */ | |
139 | if (unlikely(!rwb_enabled(rwb))) { | |
140 | rwb_wake_all(rwb); | |
141 | return; | |
142 | } | |
143 | ||
144 | /* | |
782f5697 JA |
145 | * For discards, our limit is always the background. For writes, if |
146 | * the device does write back caching, drop further down before we | |
147 | * wake people up. | |
e34cbd30 | 148 | */ |
782f5697 JA |
149 | if (wb_acct & WBT_DISCARD) |
150 | limit = rwb->wb_background; | |
151 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
152 | limit = 0; |
153 | else | |
154 | limit = rwb->wb_normal; | |
155 | ||
156 | /* | |
157 | * Don't wake anyone up if we are above the normal limit. | |
158 | */ | |
159 | if (inflight && inflight >= limit) | |
160 | return; | |
161 | ||
b7882093 | 162 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
163 | int diff = limit - inflight; |
164 | ||
165 | if (!inflight || diff >= rwb->wb_background / 2) | |
38cfb5a4 | 166 | wake_up_all(&rqw->wait); |
e34cbd30 JA |
167 | } |
168 | } | |
169 | ||
061a5427 JA |
170 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
171 | { | |
172 | struct rq_wb *rwb = RQWB(rqos); | |
173 | struct rq_wait *rqw; | |
174 | ||
175 | if (!(wb_acct & WBT_TRACKED)) | |
176 | return; | |
177 | ||
178 | rqw = get_rq_wait(rwb, wb_acct); | |
179 | wbt_rqw_done(rwb, rqw, wb_acct); | |
180 | } | |
181 | ||
e34cbd30 JA |
182 | /* |
183 | * Called on completion of a request. Note that it's also called when | |
184 | * a request is merged, when the request gets freed. | |
185 | */ | |
a7905043 | 186 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 187 | { |
a7905043 | 188 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 | 189 | |
a8a45941 OS |
190 | if (!wbt_is_tracked(rq)) { |
191 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
192 | rwb->sync_issue = 0; |
193 | rwb->sync_cookie = NULL; | |
194 | } | |
195 | ||
a8a45941 | 196 | if (wbt_is_read(rq)) |
e34cbd30 | 197 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 198 | } else { |
a8a45941 | 199 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
a7905043 | 200 | __wbt_done(rqos, wbt_flags(rq)); |
e34cbd30 | 201 | } |
a8a45941 | 202 | wbt_clear_state(rq); |
e34cbd30 JA |
203 | } |
204 | ||
4121d385 | 205 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
206 | { |
207 | /* | |
208 | * We need at least one read sample, and a minimum of | |
209 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
210 | * that it's writes impacting us, and not just some sole read on | |
211 | * a device that is in a lower power state. | |
212 | */ | |
fa2e39cb OS |
213 | return (stat[READ].nr_samples >= 1 && |
214 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
215 | } |
216 | ||
217 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
218 | { | |
6aa7de05 | 219 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
220 | |
221 | if (!issue || !rwb->sync_cookie) | |
222 | return 0; | |
223 | ||
224 | now = ktime_to_ns(ktime_get()); | |
225 | return now - issue; | |
226 | } | |
227 | ||
228 | enum { | |
229 | LAT_OK = 1, | |
230 | LAT_UNKNOWN, | |
231 | LAT_UNKNOWN_WRITES, | |
232 | LAT_EXCEEDED, | |
233 | }; | |
234 | ||
34dbad5d | 235 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 236 | { |
d152c682 | 237 | struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi; |
a7905043 | 238 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
239 | u64 thislat; |
240 | ||
241 | /* | |
242 | * If our stored sync issue exceeds the window size, or it | |
243 | * exceeds our min target AND we haven't logged any entries, | |
244 | * flag the latency as exceeded. wbt works off completion latencies, | |
245 | * but for a flooded device, a single sync IO can take a long time | |
246 | * to complete after being issued. If this time exceeds our | |
247 | * monitoring window AND we didn't see any other completions in that | |
248 | * window, then count that sync IO as a violation of the latency. | |
249 | */ | |
250 | thislat = rwb_sync_issue_lat(rwb); | |
251 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 252 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 253 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
254 | return LAT_EXCEEDED; |
255 | } | |
256 | ||
257 | /* | |
258 | * No read/write mix, if stat isn't valid | |
259 | */ | |
260 | if (!stat_sample_valid(stat)) { | |
261 | /* | |
262 | * If we had writes in this stat window and the window is | |
263 | * current, we're only doing writes. If a task recently | |
264 | * waited or still has writes in flights, consider us doing | |
265 | * just writes as well. | |
266 | */ | |
34dbad5d OS |
267 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
268 | wbt_inflight(rwb)) | |
e34cbd30 JA |
269 | return LAT_UNKNOWN_WRITES; |
270 | return LAT_UNKNOWN; | |
271 | } | |
272 | ||
273 | /* | |
274 | * If the 'min' latency exceeds our target, step down. | |
275 | */ | |
fa2e39cb OS |
276 | if (stat[READ].min > rwb->min_lat_nsec) { |
277 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 278 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
279 | return LAT_EXCEEDED; |
280 | } | |
281 | ||
a7905043 | 282 | if (rqd->scale_step) |
d8a0cbfd | 283 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
284 | |
285 | return LAT_OK; | |
286 | } | |
287 | ||
e34cbd30 JA |
288 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
289 | { | |
d152c682 | 290 | struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi; |
a7905043 | 291 | struct rq_depth *rqd = &rwb->rq_depth; |
d8a0cbfd | 292 | |
a7905043 JB |
293 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, |
294 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
e34cbd30 JA |
295 | } |
296 | ||
a7905043 | 297 | static void calc_wb_limits(struct rq_wb *rwb) |
e34cbd30 | 298 | { |
a7905043 JB |
299 | if (rwb->min_lat_nsec == 0) { |
300 | rwb->wb_normal = rwb->wb_background = 0; | |
301 | } else if (rwb->rq_depth.max_depth <= 2) { | |
302 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
303 | rwb->wb_background = 1; | |
304 | } else { | |
305 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
306 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
307 | } | |
308 | } | |
e34cbd30 | 309 | |
a7905043 JB |
310 | static void scale_up(struct rq_wb *rwb) |
311 | { | |
b84477d3 HS |
312 | if (!rq_depth_scale_up(&rwb->rq_depth)) |
313 | return; | |
a7905043 | 314 | calc_wb_limits(rwb); |
e34cbd30 | 315 | rwb->unknown_cnt = 0; |
5e65a203 | 316 | rwb_wake_all(rwb); |
3a89c25d | 317 | rwb_trace_step(rwb, tracepoint_string("scale up")); |
e34cbd30 JA |
318 | } |
319 | ||
e34cbd30 JA |
320 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
321 | { | |
b84477d3 HS |
322 | if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle)) |
323 | return; | |
e34cbd30 | 324 | calc_wb_limits(rwb); |
a7905043 | 325 | rwb->unknown_cnt = 0; |
3a89c25d | 326 | rwb_trace_step(rwb, tracepoint_string("scale down")); |
e34cbd30 JA |
327 | } |
328 | ||
329 | static void rwb_arm_timer(struct rq_wb *rwb) | |
330 | { | |
a7905043 JB |
331 | struct rq_depth *rqd = &rwb->rq_depth; |
332 | ||
333 | if (rqd->scale_step > 0) { | |
e34cbd30 JA |
334 | /* |
335 | * We should speed this up, using some variant of a fast | |
336 | * integer inverse square root calculation. Since we only do | |
337 | * this for every window expiration, it's not a huge deal, | |
338 | * though. | |
339 | */ | |
340 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
a7905043 | 341 | int_sqrt((rqd->scale_step + 1) << 8)); |
e34cbd30 JA |
342 | } else { |
343 | /* | |
344 | * For step < 0, we don't want to increase/decrease the | |
345 | * window size. | |
346 | */ | |
347 | rwb->cur_win_nsec = rwb->win_nsec; | |
348 | } | |
349 | ||
34dbad5d | 350 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
351 | } |
352 | ||
34dbad5d | 353 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 354 | { |
34dbad5d | 355 | struct rq_wb *rwb = cb->data; |
a7905043 | 356 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
357 | unsigned int inflight = wbt_inflight(rwb); |
358 | int status; | |
359 | ||
34dbad5d | 360 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 361 | |
d152c682 CH |
362 | trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step, |
363 | inflight); | |
e34cbd30 JA |
364 | |
365 | /* | |
366 | * If we exceeded the latency target, step down. If we did not, | |
367 | * step one level up. If we don't know enough to say either exceeded | |
368 | * or ok, then don't do anything. | |
369 | */ | |
370 | switch (status) { | |
371 | case LAT_EXCEEDED: | |
372 | scale_down(rwb, true); | |
373 | break; | |
374 | case LAT_OK: | |
375 | scale_up(rwb); | |
376 | break; | |
377 | case LAT_UNKNOWN_WRITES: | |
378 | /* | |
379 | * We started a the center step, but don't have a valid | |
380 | * read/write sample, but we do have writes going on. | |
381 | * Allow step to go negative, to increase write perf. | |
382 | */ | |
383 | scale_up(rwb); | |
384 | break; | |
385 | case LAT_UNKNOWN: | |
386 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
387 | break; | |
388 | /* | |
389 | * We get here when previously scaled reduced depth, and we | |
390 | * currently don't have a valid read/write sample. For that | |
391 | * case, slowly return to center state (step == 0). | |
392 | */ | |
a7905043 | 393 | if (rqd->scale_step > 0) |
e34cbd30 | 394 | scale_up(rwb); |
a7905043 | 395 | else if (rqd->scale_step < 0) |
e34cbd30 JA |
396 | scale_down(rwb, false); |
397 | break; | |
398 | default: | |
399 | break; | |
400 | } | |
401 | ||
402 | /* | |
403 | * Re-arm timer, if we have IO in flight | |
404 | */ | |
a7905043 | 405 | if (rqd->scale_step || inflight) |
e34cbd30 JA |
406 | rwb_arm_timer(rwb); |
407 | } | |
408 | ||
4d89e1d1 | 409 | static void wbt_update_limits(struct rq_wb *rwb) |
e34cbd30 | 410 | { |
a7905043 JB |
411 | struct rq_depth *rqd = &rwb->rq_depth; |
412 | ||
413 | rqd->scale_step = 0; | |
414 | rqd->scaled_max = false; | |
415 | ||
416 | rq_depth_calc_max_depth(rqd); | |
e34cbd30 JA |
417 | calc_wb_limits(rwb); |
418 | ||
419 | rwb_wake_all(rwb); | |
420 | } | |
421 | ||
a7905043 JB |
422 | u64 wbt_get_min_lat(struct request_queue *q) |
423 | { | |
424 | struct rq_qos *rqos = wbt_rq_qos(q); | |
425 | if (!rqos) | |
426 | return 0; | |
427 | return RQWB(rqos)->min_lat_nsec; | |
428 | } | |
429 | ||
430 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
431 | { | |
432 | struct rq_qos *rqos = wbt_rq_qos(q); | |
433 | if (!rqos) | |
434 | return; | |
435 | RQWB(rqos)->min_lat_nsec = val; | |
436 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
4d89e1d1 | 437 | wbt_update_limits(RQWB(rqos)); |
a7905043 JB |
438 | } |
439 | ||
440 | ||
e34cbd30 JA |
441 | static bool close_io(struct rq_wb *rwb) |
442 | { | |
443 | const unsigned long now = jiffies; | |
444 | ||
445 | return time_before(now, rwb->last_issue + HZ / 10) || | |
446 | time_before(now, rwb->last_comp + HZ / 10); | |
447 | } | |
448 | ||
449 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
450 | ||
451 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
452 | { | |
453 | unsigned int limit; | |
454 | ||
ffa358dc JA |
455 | /* |
456 | * If we got disabled, just return UINT_MAX. This ensures that | |
457 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
458 | */ | |
459 | if (!rwb_enabled(rwb)) | |
460 | return UINT_MAX; | |
461 | ||
782f5697 JA |
462 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) |
463 | return rwb->wb_background; | |
464 | ||
e34cbd30 JA |
465 | /* |
466 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 467 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
468 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
469 | * that. If the write is marked as a background write, then use | |
470 | * the idle limit, or go to normal if we haven't had competing | |
471 | * IO for a bit. | |
472 | */ | |
473 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
a7905043 | 474 | limit = rwb->rq_depth.max_depth; |
e34cbd30 JA |
475 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { |
476 | /* | |
477 | * If less than 100ms since we completed unrelated IO, | |
478 | * limit us to half the depth for background writeback. | |
479 | */ | |
480 | limit = rwb->wb_background; | |
481 | } else | |
482 | limit = rwb->wb_normal; | |
483 | ||
484 | return limit; | |
485 | } | |
486 | ||
38cfb5a4 | 487 | struct wbt_wait_data { |
38cfb5a4 | 488 | struct rq_wb *rwb; |
b6c7b58f | 489 | enum wbt_flags wb_acct; |
38cfb5a4 | 490 | unsigned long rw; |
38cfb5a4 JA |
491 | }; |
492 | ||
b6c7b58f | 493 | static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data) |
38cfb5a4 | 494 | { |
b6c7b58f JB |
495 | struct wbt_wait_data *data = private_data; |
496 | return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw)); | |
497 | } | |
38cfb5a4 | 498 | |
b6c7b58f JB |
499 | static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data) |
500 | { | |
501 | struct wbt_wait_data *data = private_data; | |
502 | wbt_rqw_done(data->rwb, rqw, data->wb_acct); | |
38cfb5a4 JA |
503 | } |
504 | ||
e34cbd30 JA |
505 | /* |
506 | * Block if we will exceed our limit, or if we are currently waiting for | |
507 | * the timer to kick off queuing again. | |
508 | */ | |
8bea6090 | 509 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
d5337560 | 510 | unsigned long rw) |
e34cbd30 | 511 | { |
8bea6090 | 512 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
38cfb5a4 | 513 | struct wbt_wait_data data = { |
38cfb5a4 | 514 | .rwb = rwb, |
b6c7b58f | 515 | .wb_acct = wb_acct, |
38cfb5a4 JA |
516 | .rw = rw, |
517 | }; | |
e34cbd30 | 518 | |
b6c7b58f | 519 | rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb); |
e34cbd30 JA |
520 | } |
521 | ||
482e302a | 522 | static inline bool wbt_should_throttle(struct bio *bio) |
e34cbd30 | 523 | { |
782f5697 JA |
524 | switch (bio_op(bio)) { |
525 | case REQ_OP_WRITE: | |
526 | /* | |
527 | * Don't throttle WRITE_ODIRECT | |
528 | */ | |
529 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
530 | (REQ_SYNC | REQ_IDLE)) | |
531 | return false; | |
df561f66 | 532 | fallthrough; |
782f5697 JA |
533 | case REQ_OP_DISCARD: |
534 | return true; | |
535 | default: | |
e34cbd30 | 536 | return false; |
782f5697 | 537 | } |
e34cbd30 JA |
538 | } |
539 | ||
c1c80384 JB |
540 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) |
541 | { | |
542 | enum wbt_flags flags = 0; | |
543 | ||
c125311d JA |
544 | if (!rwb_enabled(rwb)) |
545 | return 0; | |
546 | ||
c1c80384 JB |
547 | if (bio_op(bio) == REQ_OP_READ) { |
548 | flags = WBT_READ; | |
482e302a | 549 | } else if (wbt_should_throttle(bio)) { |
c1c80384 JB |
550 | if (current_is_kswapd()) |
551 | flags |= WBT_KSWAPD; | |
552 | if (bio_op(bio) == REQ_OP_DISCARD) | |
553 | flags |= WBT_DISCARD; | |
554 | flags |= WBT_TRACKED; | |
555 | } | |
556 | return flags; | |
557 | } | |
558 | ||
559 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
560 | { | |
561 | struct rq_wb *rwb = RQWB(rqos); | |
562 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
563 | __wbt_done(rqos, flags); | |
564 | } | |
565 | ||
e34cbd30 | 566 | /* |
e34cbd30 JA |
567 | * May sleep, if we have exceeded the writeback limits. Caller can pass |
568 | * in an irq held spinlock, if it holds one when calling this function. | |
569 | * If we do sleep, we'll release and re-grab it. | |
570 | */ | |
d5337560 | 571 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio) |
e34cbd30 | 572 | { |
a7905043 | 573 | struct rq_wb *rwb = RQWB(rqos); |
c1c80384 | 574 | enum wbt_flags flags; |
e34cbd30 | 575 | |
c1c80384 | 576 | flags = bio_to_wbt_flags(rwb, bio); |
df60f6e8 | 577 | if (!(flags & WBT_TRACKED)) { |
c1c80384 | 578 | if (flags & WBT_READ) |
e34cbd30 | 579 | wb_timestamp(rwb, &rwb->last_issue); |
c1c80384 | 580 | return; |
e34cbd30 JA |
581 | } |
582 | ||
d5337560 | 583 | __wbt_wait(rwb, flags, bio->bi_opf); |
e34cbd30 | 584 | |
34dbad5d | 585 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 | 586 | rwb_arm_timer(rwb); |
c1c80384 | 587 | } |
e34cbd30 | 588 | |
c1c80384 JB |
589 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
590 | { | |
591 | struct rq_wb *rwb = RQWB(rqos); | |
592 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
e34cbd30 JA |
593 | } |
594 | ||
c83f536a | 595 | static void wbt_issue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 596 | { |
a7905043 JB |
597 | struct rq_wb *rwb = RQWB(rqos); |
598 | ||
e34cbd30 JA |
599 | if (!rwb_enabled(rwb)) |
600 | return; | |
601 | ||
602 | /* | |
a8a45941 OS |
603 | * Track sync issue, in case it takes a long time to complete. Allows us |
604 | * to react quicker, if a sync IO takes a long time to complete. Note | |
605 | * that this is just a hint. The request can go away when it completes, | |
606 | * so it's important we never dereference it. We only use the address to | |
607 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 608 | */ |
a8a45941 OS |
609 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
610 | rwb->sync_cookie = rq; | |
544ccc8d | 611 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
612 | } |
613 | } | |
614 | ||
c83f536a | 615 | static void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 616 | { |
a7905043 | 617 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
618 | if (!rwb_enabled(rwb)) |
619 | return; | |
a8a45941 | 620 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
621 | rwb->sync_issue = 0; |
622 | rwb->sync_cookie = NULL; | |
623 | } | |
624 | } | |
625 | ||
a7905043 | 626 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
e34cbd30 | 627 | { |
a7905043 JB |
628 | struct rq_qos *rqos = wbt_rq_qos(q); |
629 | if (rqos) | |
630 | RQWB(rqos)->wc = write_cache_on; | |
e34cbd30 | 631 | } |
e34cbd30 | 632 | |
8330cdb0 JK |
633 | /* |
634 | * Enable wbt if defaults are configured that way | |
635 | */ | |
636 | void wbt_enable_default(struct request_queue *q) | |
637 | { | |
a7905043 | 638 | struct rq_qos *rqos = wbt_rq_qos(q); |
76a80408 | 639 | |
8330cdb0 | 640 | /* Throttling already enabled? */ |
76a80408 ZY |
641 | if (rqos) { |
642 | if (RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT) | |
643 | RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT; | |
8330cdb0 | 644 | return; |
76a80408 | 645 | } |
8330cdb0 JK |
646 | |
647 | /* Queue not registered? Maybe shutting down... */ | |
58c898ba | 648 | if (!blk_queue_registered(q)) |
8330cdb0 JK |
649 | return; |
650 | ||
344e9ffc | 651 | if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ)) |
8330cdb0 JK |
652 | wbt_init(q); |
653 | } | |
654 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
655 | ||
80e091d1 JA |
656 | u64 wbt_default_latency_nsec(struct request_queue *q) |
657 | { | |
658 | /* | |
659 | * We default to 2msec for non-rotational storage, and 75msec | |
660 | * for rotational storage. | |
661 | */ | |
662 | if (blk_queue_nonrot(q)) | |
663 | return 2000000ULL; | |
664 | else | |
665 | return 75000000ULL; | |
666 | } | |
667 | ||
99c749a4 JA |
668 | static int wbt_data_dir(const struct request *rq) |
669 | { | |
5235553d JA |
670 | const int op = req_op(rq); |
671 | ||
672 | if (op == REQ_OP_READ) | |
673 | return READ; | |
825843b0 | 674 | else if (op_is_write(op)) |
5235553d JA |
675 | return WRITE; |
676 | ||
677 | /* don't account */ | |
678 | return -1; | |
99c749a4 JA |
679 | } |
680 | ||
9677a3e0 TH |
681 | static void wbt_queue_depth_changed(struct rq_qos *rqos) |
682 | { | |
683 | RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q); | |
4d89e1d1 | 684 | wbt_update_limits(RQWB(rqos)); |
9677a3e0 TH |
685 | } |
686 | ||
a7905043 JB |
687 | static void wbt_exit(struct rq_qos *rqos) |
688 | { | |
689 | struct rq_wb *rwb = RQWB(rqos); | |
690 | struct request_queue *q = rqos->q; | |
691 | ||
692 | blk_stat_remove_callback(q, rwb->cb); | |
693 | blk_stat_free_callback(rwb->cb); | |
694 | kfree(rwb); | |
695 | } | |
696 | ||
697 | /* | |
698 | * Disable wbt, if enabled by default. | |
699 | */ | |
700 | void wbt_disable_default(struct request_queue *q) | |
701 | { | |
702 | struct rq_qos *rqos = wbt_rq_qos(q); | |
703 | struct rq_wb *rwb; | |
704 | if (!rqos) | |
705 | return; | |
706 | rwb = RQWB(rqos); | |
544fbd16 ML |
707 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) { |
708 | blk_stat_deactivate(rwb->cb); | |
1d0903d6 | 709 | rwb->enable_state = WBT_STATE_OFF_DEFAULT; |
544fbd16 | 710 | } |
a7905043 | 711 | } |
e815f404 | 712 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
a7905043 | 713 | |
d19afebc ML |
714 | #ifdef CONFIG_BLK_DEBUG_FS |
715 | static int wbt_curr_win_nsec_show(void *data, struct seq_file *m) | |
716 | { | |
717 | struct rq_qos *rqos = data; | |
718 | struct rq_wb *rwb = RQWB(rqos); | |
719 | ||
720 | seq_printf(m, "%llu\n", rwb->cur_win_nsec); | |
721 | return 0; | |
722 | } | |
723 | ||
724 | static int wbt_enabled_show(void *data, struct seq_file *m) | |
725 | { | |
726 | struct rq_qos *rqos = data; | |
727 | struct rq_wb *rwb = RQWB(rqos); | |
728 | ||
729 | seq_printf(m, "%d\n", rwb->enable_state); | |
730 | return 0; | |
731 | } | |
732 | ||
733 | static int wbt_id_show(void *data, struct seq_file *m) | |
734 | { | |
735 | struct rq_qos *rqos = data; | |
736 | ||
737 | seq_printf(m, "%u\n", rqos->id); | |
738 | return 0; | |
739 | } | |
740 | ||
741 | static int wbt_inflight_show(void *data, struct seq_file *m) | |
742 | { | |
743 | struct rq_qos *rqos = data; | |
744 | struct rq_wb *rwb = RQWB(rqos); | |
745 | int i; | |
746 | ||
747 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
748 | seq_printf(m, "%d: inflight %d\n", i, | |
749 | atomic_read(&rwb->rq_wait[i].inflight)); | |
750 | return 0; | |
751 | } | |
752 | ||
753 | static int wbt_min_lat_nsec_show(void *data, struct seq_file *m) | |
754 | { | |
755 | struct rq_qos *rqos = data; | |
756 | struct rq_wb *rwb = RQWB(rqos); | |
757 | ||
758 | seq_printf(m, "%lu\n", rwb->min_lat_nsec); | |
759 | return 0; | |
760 | } | |
761 | ||
762 | static int wbt_unknown_cnt_show(void *data, struct seq_file *m) | |
763 | { | |
764 | struct rq_qos *rqos = data; | |
765 | struct rq_wb *rwb = RQWB(rqos); | |
766 | ||
767 | seq_printf(m, "%u\n", rwb->unknown_cnt); | |
768 | return 0; | |
769 | } | |
770 | ||
771 | static int wbt_normal_show(void *data, struct seq_file *m) | |
772 | { | |
773 | struct rq_qos *rqos = data; | |
774 | struct rq_wb *rwb = RQWB(rqos); | |
775 | ||
776 | seq_printf(m, "%u\n", rwb->wb_normal); | |
777 | return 0; | |
778 | } | |
779 | ||
780 | static int wbt_background_show(void *data, struct seq_file *m) | |
781 | { | |
782 | struct rq_qos *rqos = data; | |
783 | struct rq_wb *rwb = RQWB(rqos); | |
784 | ||
785 | seq_printf(m, "%u\n", rwb->wb_background); | |
786 | return 0; | |
787 | } | |
788 | ||
789 | static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = { | |
790 | {"curr_win_nsec", 0400, wbt_curr_win_nsec_show}, | |
791 | {"enabled", 0400, wbt_enabled_show}, | |
792 | {"id", 0400, wbt_id_show}, | |
793 | {"inflight", 0400, wbt_inflight_show}, | |
794 | {"min_lat_nsec", 0400, wbt_min_lat_nsec_show}, | |
795 | {"unknown_cnt", 0400, wbt_unknown_cnt_show}, | |
796 | {"wb_normal", 0400, wbt_normal_show}, | |
797 | {"wb_background", 0400, wbt_background_show}, | |
798 | {}, | |
799 | }; | |
800 | #endif | |
801 | ||
a7905043 JB |
802 | static struct rq_qos_ops wbt_rqos_ops = { |
803 | .throttle = wbt_wait, | |
804 | .issue = wbt_issue, | |
c1c80384 | 805 | .track = wbt_track, |
a7905043 JB |
806 | .requeue = wbt_requeue, |
807 | .done = wbt_done, | |
c1c80384 | 808 | .cleanup = wbt_cleanup, |
9677a3e0 | 809 | .queue_depth_changed = wbt_queue_depth_changed, |
a7905043 | 810 | .exit = wbt_exit, |
d19afebc ML |
811 | #ifdef CONFIG_BLK_DEBUG_FS |
812 | .debugfs_attrs = wbt_debugfs_attrs, | |
813 | #endif | |
a7905043 JB |
814 | }; |
815 | ||
8054b89f | 816 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
817 | { |
818 | struct rq_wb *rwb; | |
819 | int i; | |
820 | ||
e34cbd30 JA |
821 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
822 | if (!rwb) | |
823 | return -ENOMEM; | |
824 | ||
99c749a4 | 825 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
826 | if (!rwb->cb) { |
827 | kfree(rwb); | |
828 | return -ENOMEM; | |
829 | } | |
830 | ||
a7905043 JB |
831 | for (i = 0; i < WBT_NUM_RWQ; i++) |
832 | rq_wait_init(&rwb->rq_wait[i]); | |
e34cbd30 | 833 | |
a7905043 JB |
834 | rwb->rqos.id = RQ_QOS_WBT; |
835 | rwb->rqos.ops = &wbt_rqos_ops; | |
836 | rwb->rqos.q = q; | |
e34cbd30 | 837 | rwb->last_comp = rwb->last_issue = jiffies; |
e34cbd30 | 838 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 839 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
a7905043 JB |
840 | rwb->wc = 1; |
841 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; | |
e34cbd30 JA |
842 | |
843 | /* | |
34dbad5d | 844 | * Assign rwb and add the stats callback. |
e34cbd30 | 845 | */ |
a7905043 | 846 | rq_qos_add(q, &rwb->rqos); |
34dbad5d | 847 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 848 | |
80e091d1 | 849 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 | 850 | |
9677a3e0 | 851 | wbt_queue_depth_changed(&rwb->rqos); |
a7905043 | 852 | wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); |
e34cbd30 JA |
853 | |
854 | return 0; | |
855 | } |