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" |
671fae5e | 30 | #include "elevator.h" |
e34cbd30 JA |
31 | |
32 | #define CREATE_TRACE_POINTS | |
33 | #include <trace/events/wbt.h> | |
34 | ||
a8a45941 | 35 | static inline void wbt_clear_state(struct request *rq) |
934031a1 | 36 | { |
544ccc8d | 37 | rq->wbt_flags = 0; |
934031a1 OS |
38 | } |
39 | ||
a8a45941 | 40 | static inline enum wbt_flags wbt_flags(struct request *rq) |
934031a1 | 41 | { |
544ccc8d | 42 | return rq->wbt_flags; |
934031a1 OS |
43 | } |
44 | ||
a8a45941 | 45 | static inline bool wbt_is_tracked(struct request *rq) |
934031a1 | 46 | { |
544ccc8d | 47 | return rq->wbt_flags & WBT_TRACKED; |
934031a1 OS |
48 | } |
49 | ||
a8a45941 | 50 | static inline bool wbt_is_read(struct request *rq) |
934031a1 | 51 | { |
544ccc8d | 52 | return rq->wbt_flags & WBT_READ; |
934031a1 OS |
53 | } |
54 | ||
e34cbd30 JA |
55 | enum { |
56 | /* | |
57 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
58 | * from here depending on device stats | |
59 | */ | |
60 | RWB_DEF_DEPTH = 16, | |
61 | ||
62 | /* | |
63 | * 100msec window | |
64 | */ | |
65 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
66 | ||
67 | /* | |
68 | * Disregard stats, if we don't meet this minimum | |
69 | */ | |
70 | RWB_MIN_WRITE_SAMPLES = 3, | |
71 | ||
72 | /* | |
73 | * If we have this number of consecutive windows with not enough | |
74 | * information to scale up or down, scale up. | |
75 | */ | |
76 | RWB_UNKNOWN_BUMP = 5, | |
77 | }; | |
78 | ||
79 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
80 | { | |
1d0903d6 ZY |
81 | return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT && |
82 | rwb->wb_normal != 0; | |
e34cbd30 JA |
83 | } |
84 | ||
e34cbd30 JA |
85 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) |
86 | { | |
87 | if (rwb_enabled(rwb)) { | |
88 | const unsigned long cur = jiffies; | |
89 | ||
90 | if (cur != *var) | |
91 | *var = cur; | |
92 | } | |
93 | } | |
94 | ||
95 | /* | |
96 | * If a task was rate throttled in balance_dirty_pages() within the last | |
97 | * second or so, use that to indicate a higher cleaning rate. | |
98 | */ | |
99 | static bool wb_recent_wait(struct rq_wb *rwb) | |
100 | { | |
d152c682 | 101 | struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb; |
e34cbd30 JA |
102 | |
103 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
104 | } | |
105 | ||
8bea6090 JA |
106 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
107 | enum wbt_flags wb_acct) | |
e34cbd30 | 108 | { |
8bea6090 JA |
109 | if (wb_acct & WBT_KSWAPD) |
110 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
111 | else if (wb_acct & WBT_DISCARD) |
112 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
113 | |
114 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
115 | } |
116 | ||
117 | static void rwb_wake_all(struct rq_wb *rwb) | |
118 | { | |
119 | int i; | |
120 | ||
121 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
122 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
123 | ||
b7882093 | 124 | if (wq_has_sleeper(&rqw->wait)) |
e34cbd30 JA |
125 | wake_up_all(&rqw->wait); |
126 | } | |
127 | } | |
128 | ||
061a5427 JA |
129 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
130 | enum wbt_flags wb_acct) | |
e34cbd30 | 131 | { |
e34cbd30 JA |
132 | int inflight, limit; |
133 | ||
e34cbd30 JA |
134 | inflight = atomic_dec_return(&rqw->inflight); |
135 | ||
136 | /* | |
137 | * wbt got disabled with IO in flight. Wake up any potential | |
138 | * waiters, we don't have to do more than that. | |
139 | */ | |
140 | if (unlikely(!rwb_enabled(rwb))) { | |
141 | rwb_wake_all(rwb); | |
142 | return; | |
143 | } | |
144 | ||
145 | /* | |
782f5697 JA |
146 | * For discards, our limit is always the background. For writes, if |
147 | * the device does write back caching, drop further down before we | |
148 | * wake people up. | |
e34cbd30 | 149 | */ |
782f5697 JA |
150 | if (wb_acct & WBT_DISCARD) |
151 | limit = rwb->wb_background; | |
152 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
153 | limit = 0; |
154 | else | |
155 | limit = rwb->wb_normal; | |
156 | ||
157 | /* | |
158 | * Don't wake anyone up if we are above the normal limit. | |
159 | */ | |
160 | if (inflight && inflight >= limit) | |
161 | return; | |
162 | ||
b7882093 | 163 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
164 | int diff = limit - inflight; |
165 | ||
166 | if (!inflight || diff >= rwb->wb_background / 2) | |
38cfb5a4 | 167 | wake_up_all(&rqw->wait); |
e34cbd30 JA |
168 | } |
169 | } | |
170 | ||
061a5427 JA |
171 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
172 | { | |
173 | struct rq_wb *rwb = RQWB(rqos); | |
174 | struct rq_wait *rqw; | |
175 | ||
176 | if (!(wb_acct & WBT_TRACKED)) | |
177 | return; | |
178 | ||
179 | rqw = get_rq_wait(rwb, wb_acct); | |
180 | wbt_rqw_done(rwb, rqw, wb_acct); | |
181 | } | |
182 | ||
e34cbd30 JA |
183 | /* |
184 | * Called on completion of a request. Note that it's also called when | |
185 | * a request is merged, when the request gets freed. | |
186 | */ | |
a7905043 | 187 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 188 | { |
a7905043 | 189 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 | 190 | |
a8a45941 OS |
191 | if (!wbt_is_tracked(rq)) { |
192 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
193 | rwb->sync_issue = 0; |
194 | rwb->sync_cookie = NULL; | |
195 | } | |
196 | ||
a8a45941 | 197 | if (wbt_is_read(rq)) |
e34cbd30 | 198 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 199 | } else { |
a8a45941 | 200 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
a7905043 | 201 | __wbt_done(rqos, wbt_flags(rq)); |
e34cbd30 | 202 | } |
a8a45941 | 203 | wbt_clear_state(rq); |
e34cbd30 JA |
204 | } |
205 | ||
4121d385 | 206 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
207 | { |
208 | /* | |
209 | * We need at least one read sample, and a minimum of | |
210 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
211 | * that it's writes impacting us, and not just some sole read on | |
212 | * a device that is in a lower power state. | |
213 | */ | |
fa2e39cb OS |
214 | return (stat[READ].nr_samples >= 1 && |
215 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
216 | } |
217 | ||
218 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
219 | { | |
6aa7de05 | 220 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
221 | |
222 | if (!issue || !rwb->sync_cookie) | |
223 | return 0; | |
224 | ||
225 | now = ktime_to_ns(ktime_get()); | |
226 | return now - issue; | |
227 | } | |
228 | ||
229 | enum { | |
230 | LAT_OK = 1, | |
231 | LAT_UNKNOWN, | |
232 | LAT_UNKNOWN_WRITES, | |
233 | LAT_EXCEEDED, | |
234 | }; | |
235 | ||
34dbad5d | 236 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 237 | { |
d152c682 | 238 | struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi; |
a7905043 | 239 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
240 | u64 thislat; |
241 | ||
242 | /* | |
243 | * If our stored sync issue exceeds the window size, or it | |
244 | * exceeds our min target AND we haven't logged any entries, | |
245 | * flag the latency as exceeded. wbt works off completion latencies, | |
246 | * but for a flooded device, a single sync IO can take a long time | |
247 | * to complete after being issued. If this time exceeds our | |
248 | * monitoring window AND we didn't see any other completions in that | |
249 | * window, then count that sync IO as a violation of the latency. | |
250 | */ | |
251 | thislat = rwb_sync_issue_lat(rwb); | |
252 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 253 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 254 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
255 | return LAT_EXCEEDED; |
256 | } | |
257 | ||
258 | /* | |
259 | * No read/write mix, if stat isn't valid | |
260 | */ | |
261 | if (!stat_sample_valid(stat)) { | |
262 | /* | |
263 | * If we had writes in this stat window and the window is | |
264 | * current, we're only doing writes. If a task recently | |
265 | * waited or still has writes in flights, consider us doing | |
266 | * just writes as well. | |
267 | */ | |
34dbad5d OS |
268 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
269 | wbt_inflight(rwb)) | |
e34cbd30 JA |
270 | return LAT_UNKNOWN_WRITES; |
271 | return LAT_UNKNOWN; | |
272 | } | |
273 | ||
274 | /* | |
275 | * If the 'min' latency exceeds our target, step down. | |
276 | */ | |
fa2e39cb OS |
277 | if (stat[READ].min > rwb->min_lat_nsec) { |
278 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 279 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
280 | return LAT_EXCEEDED; |
281 | } | |
282 | ||
a7905043 | 283 | if (rqd->scale_step) |
d8a0cbfd | 284 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
285 | |
286 | return LAT_OK; | |
287 | } | |
288 | ||
e34cbd30 JA |
289 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
290 | { | |
d152c682 | 291 | struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi; |
a7905043 | 292 | struct rq_depth *rqd = &rwb->rq_depth; |
d8a0cbfd | 293 | |
a7905043 JB |
294 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, |
295 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
e34cbd30 JA |
296 | } |
297 | ||
a7905043 | 298 | static void calc_wb_limits(struct rq_wb *rwb) |
e34cbd30 | 299 | { |
a7905043 JB |
300 | if (rwb->min_lat_nsec == 0) { |
301 | rwb->wb_normal = rwb->wb_background = 0; | |
302 | } else if (rwb->rq_depth.max_depth <= 2) { | |
303 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
304 | rwb->wb_background = 1; | |
305 | } else { | |
306 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
307 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
308 | } | |
309 | } | |
e34cbd30 | 310 | |
a7905043 JB |
311 | static void scale_up(struct rq_wb *rwb) |
312 | { | |
b84477d3 HS |
313 | if (!rq_depth_scale_up(&rwb->rq_depth)) |
314 | return; | |
a7905043 | 315 | calc_wb_limits(rwb); |
e34cbd30 | 316 | rwb->unknown_cnt = 0; |
5e65a203 | 317 | rwb_wake_all(rwb); |
3a89c25d | 318 | rwb_trace_step(rwb, tracepoint_string("scale up")); |
e34cbd30 JA |
319 | } |
320 | ||
e34cbd30 JA |
321 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
322 | { | |
b84477d3 HS |
323 | if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle)) |
324 | return; | |
e34cbd30 | 325 | calc_wb_limits(rwb); |
a7905043 | 326 | rwb->unknown_cnt = 0; |
3a89c25d | 327 | rwb_trace_step(rwb, tracepoint_string("scale down")); |
e34cbd30 JA |
328 | } |
329 | ||
330 | static void rwb_arm_timer(struct rq_wb *rwb) | |
331 | { | |
a7905043 JB |
332 | struct rq_depth *rqd = &rwb->rq_depth; |
333 | ||
334 | if (rqd->scale_step > 0) { | |
e34cbd30 JA |
335 | /* |
336 | * We should speed this up, using some variant of a fast | |
337 | * integer inverse square root calculation. Since we only do | |
338 | * this for every window expiration, it's not a huge deal, | |
339 | * though. | |
340 | */ | |
341 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
a7905043 | 342 | int_sqrt((rqd->scale_step + 1) << 8)); |
e34cbd30 JA |
343 | } else { |
344 | /* | |
345 | * For step < 0, we don't want to increase/decrease the | |
346 | * window size. | |
347 | */ | |
348 | rwb->cur_win_nsec = rwb->win_nsec; | |
349 | } | |
350 | ||
34dbad5d | 351 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
352 | } |
353 | ||
34dbad5d | 354 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 355 | { |
34dbad5d | 356 | struct rq_wb *rwb = cb->data; |
a7905043 | 357 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
358 | unsigned int inflight = wbt_inflight(rwb); |
359 | int status; | |
360 | ||
480d42dc AR |
361 | if (!rwb->rqos.q->disk) |
362 | return; | |
363 | ||
34dbad5d | 364 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 365 | |
d152c682 CH |
366 | trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step, |
367 | inflight); | |
e34cbd30 JA |
368 | |
369 | /* | |
370 | * If we exceeded the latency target, step down. If we did not, | |
371 | * step one level up. If we don't know enough to say either exceeded | |
372 | * or ok, then don't do anything. | |
373 | */ | |
374 | switch (status) { | |
375 | case LAT_EXCEEDED: | |
376 | scale_down(rwb, true); | |
377 | break; | |
378 | case LAT_OK: | |
379 | scale_up(rwb); | |
380 | break; | |
381 | case LAT_UNKNOWN_WRITES: | |
382 | /* | |
383 | * We started a the center step, but don't have a valid | |
384 | * read/write sample, but we do have writes going on. | |
385 | * Allow step to go negative, to increase write perf. | |
386 | */ | |
387 | scale_up(rwb); | |
388 | break; | |
389 | case LAT_UNKNOWN: | |
390 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
391 | break; | |
392 | /* | |
393 | * We get here when previously scaled reduced depth, and we | |
394 | * currently don't have a valid read/write sample. For that | |
395 | * case, slowly return to center state (step == 0). | |
396 | */ | |
a7905043 | 397 | if (rqd->scale_step > 0) |
e34cbd30 | 398 | scale_up(rwb); |
a7905043 | 399 | else if (rqd->scale_step < 0) |
e34cbd30 JA |
400 | scale_down(rwb, false); |
401 | break; | |
402 | default: | |
403 | break; | |
404 | } | |
405 | ||
406 | /* | |
407 | * Re-arm timer, if we have IO in flight | |
408 | */ | |
a7905043 | 409 | if (rqd->scale_step || inflight) |
e34cbd30 JA |
410 | rwb_arm_timer(rwb); |
411 | } | |
412 | ||
4d89e1d1 | 413 | static void wbt_update_limits(struct rq_wb *rwb) |
e34cbd30 | 414 | { |
a7905043 JB |
415 | struct rq_depth *rqd = &rwb->rq_depth; |
416 | ||
417 | rqd->scale_step = 0; | |
418 | rqd->scaled_max = false; | |
419 | ||
420 | rq_depth_calc_max_depth(rqd); | |
e34cbd30 JA |
421 | calc_wb_limits(rwb); |
422 | ||
423 | rwb_wake_all(rwb); | |
424 | } | |
425 | ||
3642ef4d YK |
426 | bool wbt_disabled(struct request_queue *q) |
427 | { | |
428 | struct rq_qos *rqos = wbt_rq_qos(q); | |
429 | ||
430 | return !rqos || RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT || | |
431 | RQWB(rqos)->enable_state == WBT_STATE_OFF_MANUAL; | |
432 | } | |
433 | ||
a7905043 JB |
434 | u64 wbt_get_min_lat(struct request_queue *q) |
435 | { | |
436 | struct rq_qos *rqos = wbt_rq_qos(q); | |
437 | if (!rqos) | |
438 | return 0; | |
439 | return RQWB(rqos)->min_lat_nsec; | |
440 | } | |
441 | ||
442 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
443 | { | |
444 | struct rq_qos *rqos = wbt_rq_qos(q); | |
445 | if (!rqos) | |
446 | return; | |
a9a236d2 | 447 | |
a7905043 | 448 | RQWB(rqos)->min_lat_nsec = val; |
a9a236d2 YK |
449 | if (val) |
450 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
451 | else | |
452 | RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL; | |
453 | ||
4d89e1d1 | 454 | wbt_update_limits(RQWB(rqos)); |
a7905043 JB |
455 | } |
456 | ||
457 | ||
e34cbd30 JA |
458 | static bool close_io(struct rq_wb *rwb) |
459 | { | |
460 | const unsigned long now = jiffies; | |
461 | ||
462 | return time_before(now, rwb->last_issue + HZ / 10) || | |
463 | time_before(now, rwb->last_comp + HZ / 10); | |
464 | } | |
465 | ||
466 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
467 | ||
16458cf3 | 468 | static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf) |
e34cbd30 JA |
469 | { |
470 | unsigned int limit; | |
471 | ||
ffa358dc JA |
472 | /* |
473 | * If we got disabled, just return UINT_MAX. This ensures that | |
474 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
475 | */ | |
476 | if (!rwb_enabled(rwb)) | |
477 | return UINT_MAX; | |
478 | ||
16458cf3 | 479 | if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD) |
782f5697 JA |
480 | return rwb->wb_background; |
481 | ||
e34cbd30 JA |
482 | /* |
483 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 484 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
485 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
486 | * that. If the write is marked as a background write, then use | |
487 | * the idle limit, or go to normal if we haven't had competing | |
488 | * IO for a bit. | |
489 | */ | |
16458cf3 | 490 | if ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) |
a7905043 | 491 | limit = rwb->rq_depth.max_depth; |
16458cf3 | 492 | else if ((opf & REQ_BACKGROUND) || close_io(rwb)) { |
e34cbd30 JA |
493 | /* |
494 | * If less than 100ms since we completed unrelated IO, | |
495 | * limit us to half the depth for background writeback. | |
496 | */ | |
497 | limit = rwb->wb_background; | |
498 | } else | |
499 | limit = rwb->wb_normal; | |
500 | ||
501 | return limit; | |
502 | } | |
503 | ||
38cfb5a4 | 504 | struct wbt_wait_data { |
38cfb5a4 | 505 | struct rq_wb *rwb; |
b6c7b58f | 506 | enum wbt_flags wb_acct; |
16458cf3 | 507 | blk_opf_t opf; |
38cfb5a4 JA |
508 | }; |
509 | ||
b6c7b58f | 510 | static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data) |
38cfb5a4 | 511 | { |
b6c7b58f | 512 | struct wbt_wait_data *data = private_data; |
16458cf3 | 513 | return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf)); |
b6c7b58f | 514 | } |
38cfb5a4 | 515 | |
b6c7b58f JB |
516 | static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data) |
517 | { | |
518 | struct wbt_wait_data *data = private_data; | |
519 | wbt_rqw_done(data->rwb, rqw, data->wb_acct); | |
38cfb5a4 JA |
520 | } |
521 | ||
e34cbd30 JA |
522 | /* |
523 | * Block if we will exceed our limit, or if we are currently waiting for | |
524 | * the timer to kick off queuing again. | |
525 | */ | |
8bea6090 | 526 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
16458cf3 | 527 | blk_opf_t opf) |
e34cbd30 | 528 | { |
8bea6090 | 529 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
38cfb5a4 | 530 | struct wbt_wait_data data = { |
38cfb5a4 | 531 | .rwb = rwb, |
b6c7b58f | 532 | .wb_acct = wb_acct, |
16458cf3 | 533 | .opf = opf, |
38cfb5a4 | 534 | }; |
e34cbd30 | 535 | |
b6c7b58f | 536 | rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb); |
e34cbd30 JA |
537 | } |
538 | ||
482e302a | 539 | static inline bool wbt_should_throttle(struct bio *bio) |
e34cbd30 | 540 | { |
782f5697 JA |
541 | switch (bio_op(bio)) { |
542 | case REQ_OP_WRITE: | |
543 | /* | |
544 | * Don't throttle WRITE_ODIRECT | |
545 | */ | |
546 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
547 | (REQ_SYNC | REQ_IDLE)) | |
548 | return false; | |
df561f66 | 549 | fallthrough; |
782f5697 JA |
550 | case REQ_OP_DISCARD: |
551 | return true; | |
552 | default: | |
e34cbd30 | 553 | return false; |
782f5697 | 554 | } |
e34cbd30 JA |
555 | } |
556 | ||
c1c80384 JB |
557 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) |
558 | { | |
559 | enum wbt_flags flags = 0; | |
560 | ||
c125311d JA |
561 | if (!rwb_enabled(rwb)) |
562 | return 0; | |
563 | ||
c1c80384 JB |
564 | if (bio_op(bio) == REQ_OP_READ) { |
565 | flags = WBT_READ; | |
482e302a | 566 | } else if (wbt_should_throttle(bio)) { |
c1c80384 JB |
567 | if (current_is_kswapd()) |
568 | flags |= WBT_KSWAPD; | |
569 | if (bio_op(bio) == REQ_OP_DISCARD) | |
570 | flags |= WBT_DISCARD; | |
571 | flags |= WBT_TRACKED; | |
572 | } | |
573 | return flags; | |
574 | } | |
575 | ||
576 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
577 | { | |
578 | struct rq_wb *rwb = RQWB(rqos); | |
579 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
580 | __wbt_done(rqos, flags); | |
581 | } | |
582 | ||
e34cbd30 | 583 | /* |
e34cbd30 JA |
584 | * May sleep, if we have exceeded the writeback limits. Caller can pass |
585 | * in an irq held spinlock, if it holds one when calling this function. | |
586 | * If we do sleep, we'll release and re-grab it. | |
587 | */ | |
d5337560 | 588 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio) |
e34cbd30 | 589 | { |
a7905043 | 590 | struct rq_wb *rwb = RQWB(rqos); |
c1c80384 | 591 | enum wbt_flags flags; |
e34cbd30 | 592 | |
c1c80384 | 593 | flags = bio_to_wbt_flags(rwb, bio); |
df60f6e8 | 594 | if (!(flags & WBT_TRACKED)) { |
c1c80384 | 595 | if (flags & WBT_READ) |
e34cbd30 | 596 | wb_timestamp(rwb, &rwb->last_issue); |
c1c80384 | 597 | return; |
e34cbd30 JA |
598 | } |
599 | ||
d5337560 | 600 | __wbt_wait(rwb, flags, bio->bi_opf); |
e34cbd30 | 601 | |
34dbad5d | 602 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 | 603 | rwb_arm_timer(rwb); |
c1c80384 | 604 | } |
e34cbd30 | 605 | |
c1c80384 JB |
606 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
607 | { | |
608 | struct rq_wb *rwb = RQWB(rqos); | |
609 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
e34cbd30 JA |
610 | } |
611 | ||
c83f536a | 612 | static void wbt_issue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 613 | { |
a7905043 JB |
614 | struct rq_wb *rwb = RQWB(rqos); |
615 | ||
e34cbd30 JA |
616 | if (!rwb_enabled(rwb)) |
617 | return; | |
618 | ||
619 | /* | |
a8a45941 OS |
620 | * Track sync issue, in case it takes a long time to complete. Allows us |
621 | * to react quicker, if a sync IO takes a long time to complete. Note | |
622 | * that this is just a hint. The request can go away when it completes, | |
623 | * so it's important we never dereference it. We only use the address to | |
624 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 625 | */ |
a8a45941 OS |
626 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
627 | rwb->sync_cookie = rq; | |
544ccc8d | 628 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
629 | } |
630 | } | |
631 | ||
c83f536a | 632 | static void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 633 | { |
a7905043 | 634 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
635 | if (!rwb_enabled(rwb)) |
636 | return; | |
a8a45941 | 637 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
638 | rwb->sync_issue = 0; |
639 | rwb->sync_cookie = NULL; | |
640 | } | |
641 | } | |
642 | ||
a7905043 | 643 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
e34cbd30 | 644 | { |
a7905043 JB |
645 | struct rq_qos *rqos = wbt_rq_qos(q); |
646 | if (rqos) | |
647 | RQWB(rqos)->wc = write_cache_on; | |
e34cbd30 | 648 | } |
e34cbd30 | 649 | |
8330cdb0 JK |
650 | /* |
651 | * Enable wbt if defaults are configured that way | |
652 | */ | |
653 | void wbt_enable_default(struct request_queue *q) | |
654 | { | |
671fae5e YK |
655 | struct rq_qos *rqos; |
656 | bool disable_flag = q->elevator && | |
657 | test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags); | |
76a80408 | 658 | |
8330cdb0 | 659 | /* Throttling already enabled? */ |
671fae5e | 660 | rqos = wbt_rq_qos(q); |
76a80408 | 661 | if (rqos) { |
671fae5e YK |
662 | if (!disable_flag && |
663 | RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT) | |
76a80408 | 664 | RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT; |
8330cdb0 | 665 | return; |
76a80408 | 666 | } |
8330cdb0 JK |
667 | |
668 | /* Queue not registered? Maybe shutting down... */ | |
58c898ba | 669 | if (!blk_queue_registered(q)) |
8330cdb0 JK |
670 | return; |
671 | ||
671fae5e | 672 | if (queue_is_mq(q) && !disable_flag) |
8330cdb0 JK |
673 | wbt_init(q); |
674 | } | |
675 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
676 | ||
80e091d1 JA |
677 | u64 wbt_default_latency_nsec(struct request_queue *q) |
678 | { | |
679 | /* | |
680 | * We default to 2msec for non-rotational storage, and 75msec | |
681 | * for rotational storage. | |
682 | */ | |
683 | if (blk_queue_nonrot(q)) | |
684 | return 2000000ULL; | |
685 | else | |
686 | return 75000000ULL; | |
687 | } | |
688 | ||
99c749a4 JA |
689 | static int wbt_data_dir(const struct request *rq) |
690 | { | |
77e7ffd7 | 691 | const enum req_op op = req_op(rq); |
5235553d JA |
692 | |
693 | if (op == REQ_OP_READ) | |
694 | return READ; | |
825843b0 | 695 | else if (op_is_write(op)) |
5235553d JA |
696 | return WRITE; |
697 | ||
698 | /* don't account */ | |
699 | return -1; | |
99c749a4 JA |
700 | } |
701 | ||
9677a3e0 TH |
702 | static void wbt_queue_depth_changed(struct rq_qos *rqos) |
703 | { | |
704 | RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q); | |
4d89e1d1 | 705 | wbt_update_limits(RQWB(rqos)); |
9677a3e0 TH |
706 | } |
707 | ||
a7905043 JB |
708 | static void wbt_exit(struct rq_qos *rqos) |
709 | { | |
710 | struct rq_wb *rwb = RQWB(rqos); | |
711 | struct request_queue *q = rqos->q; | |
712 | ||
713 | blk_stat_remove_callback(q, rwb->cb); | |
714 | blk_stat_free_callback(rwb->cb); | |
715 | kfree(rwb); | |
716 | } | |
717 | ||
718 | /* | |
719 | * Disable wbt, if enabled by default. | |
720 | */ | |
721 | void wbt_disable_default(struct request_queue *q) | |
722 | { | |
723 | struct rq_qos *rqos = wbt_rq_qos(q); | |
724 | struct rq_wb *rwb; | |
725 | if (!rqos) | |
726 | return; | |
727 | rwb = RQWB(rqos); | |
544fbd16 ML |
728 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) { |
729 | blk_stat_deactivate(rwb->cb); | |
1d0903d6 | 730 | rwb->enable_state = WBT_STATE_OFF_DEFAULT; |
544fbd16 | 731 | } |
a7905043 | 732 | } |
e815f404 | 733 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
a7905043 | 734 | |
d19afebc ML |
735 | #ifdef CONFIG_BLK_DEBUG_FS |
736 | static int wbt_curr_win_nsec_show(void *data, struct seq_file *m) | |
737 | { | |
738 | struct rq_qos *rqos = data; | |
739 | struct rq_wb *rwb = RQWB(rqos); | |
740 | ||
741 | seq_printf(m, "%llu\n", rwb->cur_win_nsec); | |
742 | return 0; | |
743 | } | |
744 | ||
745 | static int wbt_enabled_show(void *data, struct seq_file *m) | |
746 | { | |
747 | struct rq_qos *rqos = data; | |
748 | struct rq_wb *rwb = RQWB(rqos); | |
749 | ||
750 | seq_printf(m, "%d\n", rwb->enable_state); | |
751 | return 0; | |
752 | } | |
753 | ||
754 | static int wbt_id_show(void *data, struct seq_file *m) | |
755 | { | |
756 | struct rq_qos *rqos = data; | |
757 | ||
758 | seq_printf(m, "%u\n", rqos->id); | |
759 | return 0; | |
760 | } | |
761 | ||
762 | static int wbt_inflight_show(void *data, struct seq_file *m) | |
763 | { | |
764 | struct rq_qos *rqos = data; | |
765 | struct rq_wb *rwb = RQWB(rqos); | |
766 | int i; | |
767 | ||
768 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
769 | seq_printf(m, "%d: inflight %d\n", i, | |
770 | atomic_read(&rwb->rq_wait[i].inflight)); | |
771 | return 0; | |
772 | } | |
773 | ||
774 | static int wbt_min_lat_nsec_show(void *data, struct seq_file *m) | |
775 | { | |
776 | struct rq_qos *rqos = data; | |
777 | struct rq_wb *rwb = RQWB(rqos); | |
778 | ||
779 | seq_printf(m, "%lu\n", rwb->min_lat_nsec); | |
780 | return 0; | |
781 | } | |
782 | ||
783 | static int wbt_unknown_cnt_show(void *data, struct seq_file *m) | |
784 | { | |
785 | struct rq_qos *rqos = data; | |
786 | struct rq_wb *rwb = RQWB(rqos); | |
787 | ||
788 | seq_printf(m, "%u\n", rwb->unknown_cnt); | |
789 | return 0; | |
790 | } | |
791 | ||
792 | static int wbt_normal_show(void *data, struct seq_file *m) | |
793 | { | |
794 | struct rq_qos *rqos = data; | |
795 | struct rq_wb *rwb = RQWB(rqos); | |
796 | ||
797 | seq_printf(m, "%u\n", rwb->wb_normal); | |
798 | return 0; | |
799 | } | |
800 | ||
801 | static int wbt_background_show(void *data, struct seq_file *m) | |
802 | { | |
803 | struct rq_qos *rqos = data; | |
804 | struct rq_wb *rwb = RQWB(rqos); | |
805 | ||
806 | seq_printf(m, "%u\n", rwb->wb_background); | |
807 | return 0; | |
808 | } | |
809 | ||
810 | static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = { | |
811 | {"curr_win_nsec", 0400, wbt_curr_win_nsec_show}, | |
812 | {"enabled", 0400, wbt_enabled_show}, | |
813 | {"id", 0400, wbt_id_show}, | |
814 | {"inflight", 0400, wbt_inflight_show}, | |
815 | {"min_lat_nsec", 0400, wbt_min_lat_nsec_show}, | |
816 | {"unknown_cnt", 0400, wbt_unknown_cnt_show}, | |
817 | {"wb_normal", 0400, wbt_normal_show}, | |
818 | {"wb_background", 0400, wbt_background_show}, | |
819 | {}, | |
820 | }; | |
821 | #endif | |
822 | ||
a7905043 JB |
823 | static struct rq_qos_ops wbt_rqos_ops = { |
824 | .throttle = wbt_wait, | |
825 | .issue = wbt_issue, | |
c1c80384 | 826 | .track = wbt_track, |
a7905043 JB |
827 | .requeue = wbt_requeue, |
828 | .done = wbt_done, | |
c1c80384 | 829 | .cleanup = wbt_cleanup, |
9677a3e0 | 830 | .queue_depth_changed = wbt_queue_depth_changed, |
a7905043 | 831 | .exit = wbt_exit, |
d19afebc ML |
832 | #ifdef CONFIG_BLK_DEBUG_FS |
833 | .debugfs_attrs = wbt_debugfs_attrs, | |
834 | #endif | |
a7905043 JB |
835 | }; |
836 | ||
8054b89f | 837 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
838 | { |
839 | struct rq_wb *rwb; | |
840 | int i; | |
14a6e2eb | 841 | int ret; |
e34cbd30 | 842 | |
e34cbd30 JA |
843 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
844 | if (!rwb) | |
845 | return -ENOMEM; | |
846 | ||
99c749a4 | 847 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
848 | if (!rwb->cb) { |
849 | kfree(rwb); | |
850 | return -ENOMEM; | |
851 | } | |
852 | ||
a7905043 JB |
853 | for (i = 0; i < WBT_NUM_RWQ; i++) |
854 | rq_wait_init(&rwb->rq_wait[i]); | |
e34cbd30 | 855 | |
a7905043 JB |
856 | rwb->rqos.id = RQ_QOS_WBT; |
857 | rwb->rqos.ops = &wbt_rqos_ops; | |
858 | rwb->rqos.q = q; | |
e34cbd30 | 859 | rwb->last_comp = rwb->last_issue = jiffies; |
e34cbd30 | 860 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 861 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
285febab | 862 | rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags); |
a7905043 | 863 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; |
8c5035df YK |
864 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
865 | ||
866 | wbt_queue_depth_changed(&rwb->rqos); | |
e34cbd30 JA |
867 | |
868 | /* | |
34dbad5d | 869 | * Assign rwb and add the stats callback. |
e34cbd30 | 870 | */ |
14a6e2eb JH |
871 | ret = rq_qos_add(q, &rwb->rqos); |
872 | if (ret) | |
873 | goto err_free; | |
874 | ||
34dbad5d | 875 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 876 | |
e34cbd30 | 877 | return 0; |
14a6e2eb JH |
878 | |
879 | err_free: | |
880 | blk_stat_free_callback(rwb->cb); | |
881 | kfree(rwb); | |
882 | return ret; | |
883 | ||
e34cbd30 | 884 | } |