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 | ||
0bc65bd4 | 28 | #include "blk-stat.h" |
e34cbd30 | 29 | #include "blk-wbt.h" |
a7905043 | 30 | #include "blk-rq-qos.h" |
671fae5e | 31 | #include "elevator.h" |
e34cbd30 JA |
32 | |
33 | #define CREATE_TRACE_POINTS | |
34 | #include <trace/events/wbt.h> | |
35 | ||
0bc65bd4 CH |
36 | enum wbt_flags { |
37 | WBT_TRACKED = 1, /* write, tracked for throttling */ | |
38 | WBT_READ = 2, /* read */ | |
39 | WBT_KSWAPD = 4, /* write, from kswapd */ | |
40 | WBT_DISCARD = 8, /* discard */ | |
41 | ||
42 | WBT_NR_BITS = 4, /* number of bits */ | |
43 | }; | |
44 | ||
45 | enum { | |
46 | WBT_RWQ_BG = 0, | |
47 | WBT_RWQ_KSWAPD, | |
48 | WBT_RWQ_DISCARD, | |
49 | WBT_NUM_RWQ, | |
50 | }; | |
51 | ||
52 | /* | |
53 | * If current state is WBT_STATE_ON/OFF_DEFAULT, it can be covered to any other | |
54 | * state, if current state is WBT_STATE_ON/OFF_MANUAL, it can only be covered | |
55 | * to WBT_STATE_OFF/ON_MANUAL. | |
56 | */ | |
57 | enum { | |
58 | WBT_STATE_ON_DEFAULT = 1, /* on by default */ | |
59 | WBT_STATE_ON_MANUAL = 2, /* on manually by sysfs */ | |
60 | WBT_STATE_OFF_DEFAULT = 3, /* off by default */ | |
61 | WBT_STATE_OFF_MANUAL = 4, /* off manually by sysfs */ | |
62 | }; | |
63 | ||
64 | struct rq_wb { | |
65 | /* | |
66 | * Settings that govern how we throttle | |
67 | */ | |
68 | unsigned int wb_background; /* background writeback */ | |
69 | unsigned int wb_normal; /* normal writeback */ | |
70 | ||
71 | short enable_state; /* WBT_STATE_* */ | |
72 | ||
73 | /* | |
74 | * Number of consecutive periods where we don't have enough | |
75 | * information to make a firm scale up/down decision. | |
76 | */ | |
77 | unsigned int unknown_cnt; | |
78 | ||
79 | u64 win_nsec; /* default window size */ | |
80 | u64 cur_win_nsec; /* current window size */ | |
81 | ||
82 | struct blk_stat_callback *cb; | |
83 | ||
84 | u64 sync_issue; | |
85 | void *sync_cookie; | |
86 | ||
87 | unsigned int wc; | |
88 | ||
89 | unsigned long last_issue; /* last non-throttled issue */ | |
90 | unsigned long last_comp; /* last non-throttled comp */ | |
91 | unsigned long min_lat_nsec; | |
92 | struct rq_qos rqos; | |
93 | struct rq_wait rq_wait[WBT_NUM_RWQ]; | |
94 | struct rq_depth rq_depth; | |
95 | }; | |
96 | ||
97 | static inline struct rq_wb *RQWB(struct rq_qos *rqos) | |
98 | { | |
99 | return container_of(rqos, struct rq_wb, rqos); | |
100 | } | |
101 | ||
a8a45941 | 102 | static inline void wbt_clear_state(struct request *rq) |
934031a1 | 103 | { |
544ccc8d | 104 | rq->wbt_flags = 0; |
934031a1 OS |
105 | } |
106 | ||
a8a45941 | 107 | static inline enum wbt_flags wbt_flags(struct request *rq) |
934031a1 | 108 | { |
544ccc8d | 109 | return rq->wbt_flags; |
934031a1 OS |
110 | } |
111 | ||
a8a45941 | 112 | static inline bool wbt_is_tracked(struct request *rq) |
934031a1 | 113 | { |
544ccc8d | 114 | return rq->wbt_flags & WBT_TRACKED; |
934031a1 OS |
115 | } |
116 | ||
a8a45941 | 117 | static inline bool wbt_is_read(struct request *rq) |
934031a1 | 118 | { |
544ccc8d | 119 | return rq->wbt_flags & WBT_READ; |
934031a1 OS |
120 | } |
121 | ||
e34cbd30 JA |
122 | enum { |
123 | /* | |
124 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
125 | * from here depending on device stats | |
126 | */ | |
127 | RWB_DEF_DEPTH = 16, | |
128 | ||
129 | /* | |
130 | * 100msec window | |
131 | */ | |
132 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
133 | ||
134 | /* | |
135 | * Disregard stats, if we don't meet this minimum | |
136 | */ | |
137 | RWB_MIN_WRITE_SAMPLES = 3, | |
138 | ||
139 | /* | |
140 | * If we have this number of consecutive windows with not enough | |
141 | * information to scale up or down, scale up. | |
142 | */ | |
143 | RWB_UNKNOWN_BUMP = 5, | |
144 | }; | |
145 | ||
146 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
147 | { | |
1d0903d6 | 148 | return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT && |
06257fda | 149 | rwb->enable_state != WBT_STATE_OFF_MANUAL; |
e34cbd30 JA |
150 | } |
151 | ||
e34cbd30 JA |
152 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) |
153 | { | |
154 | if (rwb_enabled(rwb)) { | |
155 | const unsigned long cur = jiffies; | |
156 | ||
157 | if (cur != *var) | |
158 | *var = cur; | |
159 | } | |
160 | } | |
161 | ||
162 | /* | |
163 | * If a task was rate throttled in balance_dirty_pages() within the last | |
164 | * second or so, use that to indicate a higher cleaning rate. | |
165 | */ | |
166 | static bool wb_recent_wait(struct rq_wb *rwb) | |
167 | { | |
ba91c849 | 168 | struct bdi_writeback *wb = &rwb->rqos.disk->bdi->wb; |
e34cbd30 JA |
169 | |
170 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
171 | } | |
172 | ||
8bea6090 JA |
173 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
174 | enum wbt_flags wb_acct) | |
e34cbd30 | 175 | { |
8bea6090 JA |
176 | if (wb_acct & WBT_KSWAPD) |
177 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
178 | else if (wb_acct & WBT_DISCARD) |
179 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
180 | |
181 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
182 | } |
183 | ||
184 | static void rwb_wake_all(struct rq_wb *rwb) | |
185 | { | |
186 | int i; | |
187 | ||
188 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
189 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
190 | ||
b7882093 | 191 | if (wq_has_sleeper(&rqw->wait)) |
e34cbd30 JA |
192 | wake_up_all(&rqw->wait); |
193 | } | |
194 | } | |
195 | ||
061a5427 JA |
196 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
197 | enum wbt_flags wb_acct) | |
e34cbd30 | 198 | { |
e34cbd30 JA |
199 | int inflight, limit; |
200 | ||
e34cbd30 JA |
201 | inflight = atomic_dec_return(&rqw->inflight); |
202 | ||
e34cbd30 | 203 | /* |
782f5697 JA |
204 | * For discards, our limit is always the background. For writes, if |
205 | * the device does write back caching, drop further down before we | |
206 | * wake people up. | |
e34cbd30 | 207 | */ |
782f5697 JA |
208 | if (wb_acct & WBT_DISCARD) |
209 | limit = rwb->wb_background; | |
210 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
211 | limit = 0; |
212 | else | |
213 | limit = rwb->wb_normal; | |
214 | ||
215 | /* | |
216 | * Don't wake anyone up if we are above the normal limit. | |
217 | */ | |
218 | if (inflight && inflight >= limit) | |
219 | return; | |
220 | ||
b7882093 | 221 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
222 | int diff = limit - inflight; |
223 | ||
224 | if (!inflight || diff >= rwb->wb_background / 2) | |
38cfb5a4 | 225 | wake_up_all(&rqw->wait); |
e34cbd30 JA |
226 | } |
227 | } | |
228 | ||
061a5427 JA |
229 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
230 | { | |
231 | struct rq_wb *rwb = RQWB(rqos); | |
232 | struct rq_wait *rqw; | |
233 | ||
234 | if (!(wb_acct & WBT_TRACKED)) | |
235 | return; | |
236 | ||
237 | rqw = get_rq_wait(rwb, wb_acct); | |
238 | wbt_rqw_done(rwb, rqw, wb_acct); | |
239 | } | |
240 | ||
e34cbd30 JA |
241 | /* |
242 | * Called on completion of a request. Note that it's also called when | |
243 | * a request is merged, when the request gets freed. | |
244 | */ | |
a7905043 | 245 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 246 | { |
a7905043 | 247 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 | 248 | |
a8a45941 OS |
249 | if (!wbt_is_tracked(rq)) { |
250 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
251 | rwb->sync_issue = 0; |
252 | rwb->sync_cookie = NULL; | |
253 | } | |
254 | ||
a8a45941 | 255 | if (wbt_is_read(rq)) |
e34cbd30 | 256 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 257 | } else { |
a8a45941 | 258 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
a7905043 | 259 | __wbt_done(rqos, wbt_flags(rq)); |
e34cbd30 | 260 | } |
a8a45941 | 261 | wbt_clear_state(rq); |
e34cbd30 JA |
262 | } |
263 | ||
4121d385 | 264 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
265 | { |
266 | /* | |
267 | * We need at least one read sample, and a minimum of | |
268 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
269 | * that it's writes impacting us, and not just some sole read on | |
270 | * a device that is in a lower power state. | |
271 | */ | |
fa2e39cb OS |
272 | return (stat[READ].nr_samples >= 1 && |
273 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
274 | } |
275 | ||
276 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
277 | { | |
6aa7de05 | 278 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
279 | |
280 | if (!issue || !rwb->sync_cookie) | |
281 | return 0; | |
282 | ||
283 | now = ktime_to_ns(ktime_get()); | |
284 | return now - issue; | |
285 | } | |
286 | ||
0bc65bd4 CH |
287 | static inline unsigned int wbt_inflight(struct rq_wb *rwb) |
288 | { | |
289 | unsigned int i, ret = 0; | |
290 | ||
291 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
292 | ret += atomic_read(&rwb->rq_wait[i].inflight); | |
293 | ||
294 | return ret; | |
295 | } | |
296 | ||
e34cbd30 JA |
297 | enum { |
298 | LAT_OK = 1, | |
299 | LAT_UNKNOWN, | |
300 | LAT_UNKNOWN_WRITES, | |
301 | LAT_EXCEEDED, | |
302 | }; | |
303 | ||
34dbad5d | 304 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 305 | { |
ba91c849 | 306 | struct backing_dev_info *bdi = rwb->rqos.disk->bdi; |
a7905043 | 307 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
308 | u64 thislat; |
309 | ||
310 | /* | |
311 | * If our stored sync issue exceeds the window size, or it | |
312 | * exceeds our min target AND we haven't logged any entries, | |
313 | * flag the latency as exceeded. wbt works off completion latencies, | |
314 | * but for a flooded device, a single sync IO can take a long time | |
315 | * to complete after being issued. If this time exceeds our | |
316 | * monitoring window AND we didn't see any other completions in that | |
317 | * window, then count that sync IO as a violation of the latency. | |
318 | */ | |
319 | thislat = rwb_sync_issue_lat(rwb); | |
320 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 321 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 322 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
323 | return LAT_EXCEEDED; |
324 | } | |
325 | ||
326 | /* | |
327 | * No read/write mix, if stat isn't valid | |
328 | */ | |
329 | if (!stat_sample_valid(stat)) { | |
330 | /* | |
331 | * If we had writes in this stat window and the window is | |
332 | * current, we're only doing writes. If a task recently | |
333 | * waited or still has writes in flights, consider us doing | |
334 | * just writes as well. | |
335 | */ | |
34dbad5d OS |
336 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
337 | wbt_inflight(rwb)) | |
e34cbd30 JA |
338 | return LAT_UNKNOWN_WRITES; |
339 | return LAT_UNKNOWN; | |
340 | } | |
341 | ||
342 | /* | |
343 | * If the 'min' latency exceeds our target, step down. | |
344 | */ | |
fa2e39cb OS |
345 | if (stat[READ].min > rwb->min_lat_nsec) { |
346 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 347 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
348 | return LAT_EXCEEDED; |
349 | } | |
350 | ||
a7905043 | 351 | if (rqd->scale_step) |
d8a0cbfd | 352 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
353 | |
354 | return LAT_OK; | |
355 | } | |
356 | ||
e34cbd30 JA |
357 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
358 | { | |
ba91c849 | 359 | struct backing_dev_info *bdi = rwb->rqos.disk->bdi; |
a7905043 | 360 | struct rq_depth *rqd = &rwb->rq_depth; |
d8a0cbfd | 361 | |
a7905043 JB |
362 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, |
363 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
e34cbd30 JA |
364 | } |
365 | ||
a7905043 | 366 | static void calc_wb_limits(struct rq_wb *rwb) |
e34cbd30 | 367 | { |
a7905043 JB |
368 | if (rwb->min_lat_nsec == 0) { |
369 | rwb->wb_normal = rwb->wb_background = 0; | |
370 | } else if (rwb->rq_depth.max_depth <= 2) { | |
371 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
372 | rwb->wb_background = 1; | |
373 | } else { | |
374 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
375 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
376 | } | |
377 | } | |
e34cbd30 | 378 | |
a7905043 JB |
379 | static void scale_up(struct rq_wb *rwb) |
380 | { | |
b84477d3 HS |
381 | if (!rq_depth_scale_up(&rwb->rq_depth)) |
382 | return; | |
a7905043 | 383 | calc_wb_limits(rwb); |
e34cbd30 | 384 | rwb->unknown_cnt = 0; |
5e65a203 | 385 | rwb_wake_all(rwb); |
3a89c25d | 386 | rwb_trace_step(rwb, tracepoint_string("scale up")); |
e34cbd30 JA |
387 | } |
388 | ||
e34cbd30 JA |
389 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
390 | { | |
b84477d3 HS |
391 | if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle)) |
392 | return; | |
e34cbd30 | 393 | calc_wb_limits(rwb); |
a7905043 | 394 | rwb->unknown_cnt = 0; |
3a89c25d | 395 | rwb_trace_step(rwb, tracepoint_string("scale down")); |
e34cbd30 JA |
396 | } |
397 | ||
398 | static void rwb_arm_timer(struct rq_wb *rwb) | |
399 | { | |
a7905043 JB |
400 | struct rq_depth *rqd = &rwb->rq_depth; |
401 | ||
402 | if (rqd->scale_step > 0) { | |
e34cbd30 JA |
403 | /* |
404 | * We should speed this up, using some variant of a fast | |
405 | * integer inverse square root calculation. Since we only do | |
406 | * this for every window expiration, it's not a huge deal, | |
407 | * though. | |
408 | */ | |
409 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
a7905043 | 410 | int_sqrt((rqd->scale_step + 1) << 8)); |
e34cbd30 JA |
411 | } else { |
412 | /* | |
413 | * For step < 0, we don't want to increase/decrease the | |
414 | * window size. | |
415 | */ | |
416 | rwb->cur_win_nsec = rwb->win_nsec; | |
417 | } | |
418 | ||
34dbad5d | 419 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
420 | } |
421 | ||
34dbad5d | 422 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 423 | { |
34dbad5d | 424 | struct rq_wb *rwb = cb->data; |
a7905043 | 425 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
426 | unsigned int inflight = wbt_inflight(rwb); |
427 | int status; | |
428 | ||
ba91c849 | 429 | if (!rwb->rqos.disk) |
480d42dc AR |
430 | return; |
431 | ||
34dbad5d | 432 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 433 | |
ba91c849 | 434 | trace_wbt_timer(rwb->rqos.disk->bdi, status, rqd->scale_step, inflight); |
e34cbd30 JA |
435 | |
436 | /* | |
437 | * If we exceeded the latency target, step down. If we did not, | |
438 | * step one level up. If we don't know enough to say either exceeded | |
439 | * or ok, then don't do anything. | |
440 | */ | |
441 | switch (status) { | |
442 | case LAT_EXCEEDED: | |
443 | scale_down(rwb, true); | |
444 | break; | |
445 | case LAT_OK: | |
446 | scale_up(rwb); | |
447 | break; | |
448 | case LAT_UNKNOWN_WRITES: | |
449 | /* | |
450 | * We started a the center step, but don't have a valid | |
451 | * read/write sample, but we do have writes going on. | |
452 | * Allow step to go negative, to increase write perf. | |
453 | */ | |
454 | scale_up(rwb); | |
455 | break; | |
456 | case LAT_UNKNOWN: | |
457 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
458 | break; | |
459 | /* | |
460 | * We get here when previously scaled reduced depth, and we | |
461 | * currently don't have a valid read/write sample. For that | |
462 | * case, slowly return to center state (step == 0). | |
463 | */ | |
a7905043 | 464 | if (rqd->scale_step > 0) |
e34cbd30 | 465 | scale_up(rwb); |
a7905043 | 466 | else if (rqd->scale_step < 0) |
e34cbd30 JA |
467 | scale_down(rwb, false); |
468 | break; | |
469 | default: | |
470 | break; | |
471 | } | |
472 | ||
473 | /* | |
474 | * Re-arm timer, if we have IO in flight | |
475 | */ | |
a7905043 | 476 | if (rqd->scale_step || inflight) |
e34cbd30 JA |
477 | rwb_arm_timer(rwb); |
478 | } | |
479 | ||
4d89e1d1 | 480 | static void wbt_update_limits(struct rq_wb *rwb) |
e34cbd30 | 481 | { |
a7905043 JB |
482 | struct rq_depth *rqd = &rwb->rq_depth; |
483 | ||
484 | rqd->scale_step = 0; | |
485 | rqd->scaled_max = false; | |
486 | ||
487 | rq_depth_calc_max_depth(rqd); | |
e34cbd30 JA |
488 | calc_wb_limits(rwb); |
489 | ||
490 | rwb_wake_all(rwb); | |
491 | } | |
492 | ||
3642ef4d YK |
493 | bool wbt_disabled(struct request_queue *q) |
494 | { | |
495 | struct rq_qos *rqos = wbt_rq_qos(q); | |
496 | ||
06257fda | 497 | return !rqos || !rwb_enabled(RQWB(rqos)); |
3642ef4d YK |
498 | } |
499 | ||
a7905043 JB |
500 | u64 wbt_get_min_lat(struct request_queue *q) |
501 | { | |
502 | struct rq_qos *rqos = wbt_rq_qos(q); | |
503 | if (!rqos) | |
504 | return 0; | |
505 | return RQWB(rqos)->min_lat_nsec; | |
506 | } | |
507 | ||
508 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
509 | { | |
510 | struct rq_qos *rqos = wbt_rq_qos(q); | |
511 | if (!rqos) | |
512 | return; | |
a9a236d2 | 513 | |
a7905043 | 514 | RQWB(rqos)->min_lat_nsec = val; |
a9a236d2 YK |
515 | if (val) |
516 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
517 | else | |
518 | RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL; | |
519 | ||
4d89e1d1 | 520 | wbt_update_limits(RQWB(rqos)); |
a7905043 JB |
521 | } |
522 | ||
523 | ||
e34cbd30 JA |
524 | static bool close_io(struct rq_wb *rwb) |
525 | { | |
526 | const unsigned long now = jiffies; | |
527 | ||
528 | return time_before(now, rwb->last_issue + HZ / 10) || | |
529 | time_before(now, rwb->last_comp + HZ / 10); | |
530 | } | |
531 | ||
532 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
533 | ||
16458cf3 | 534 | static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf) |
e34cbd30 JA |
535 | { |
536 | unsigned int limit; | |
537 | ||
16458cf3 | 538 | if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD) |
782f5697 JA |
539 | return rwb->wb_background; |
540 | ||
e34cbd30 JA |
541 | /* |
542 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 543 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
544 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
545 | * that. If the write is marked as a background write, then use | |
546 | * the idle limit, or go to normal if we haven't had competing | |
547 | * IO for a bit. | |
548 | */ | |
16458cf3 | 549 | if ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) |
a7905043 | 550 | limit = rwb->rq_depth.max_depth; |
16458cf3 | 551 | else if ((opf & REQ_BACKGROUND) || close_io(rwb)) { |
e34cbd30 JA |
552 | /* |
553 | * If less than 100ms since we completed unrelated IO, | |
554 | * limit us to half the depth for background writeback. | |
555 | */ | |
556 | limit = rwb->wb_background; | |
557 | } else | |
558 | limit = rwb->wb_normal; | |
559 | ||
560 | return limit; | |
561 | } | |
562 | ||
38cfb5a4 | 563 | struct wbt_wait_data { |
38cfb5a4 | 564 | struct rq_wb *rwb; |
b6c7b58f | 565 | enum wbt_flags wb_acct; |
16458cf3 | 566 | blk_opf_t opf; |
38cfb5a4 JA |
567 | }; |
568 | ||
b6c7b58f | 569 | static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data) |
38cfb5a4 | 570 | { |
b6c7b58f | 571 | struct wbt_wait_data *data = private_data; |
16458cf3 | 572 | return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf)); |
b6c7b58f | 573 | } |
38cfb5a4 | 574 | |
b6c7b58f JB |
575 | static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data) |
576 | { | |
577 | struct wbt_wait_data *data = private_data; | |
578 | wbt_rqw_done(data->rwb, rqw, data->wb_acct); | |
38cfb5a4 JA |
579 | } |
580 | ||
e34cbd30 JA |
581 | /* |
582 | * Block if we will exceed our limit, or if we are currently waiting for | |
583 | * the timer to kick off queuing again. | |
584 | */ | |
8bea6090 | 585 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
16458cf3 | 586 | blk_opf_t opf) |
e34cbd30 | 587 | { |
8bea6090 | 588 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
38cfb5a4 | 589 | struct wbt_wait_data data = { |
38cfb5a4 | 590 | .rwb = rwb, |
b6c7b58f | 591 | .wb_acct = wb_acct, |
16458cf3 | 592 | .opf = opf, |
38cfb5a4 | 593 | }; |
e34cbd30 | 594 | |
b6c7b58f | 595 | rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb); |
e34cbd30 JA |
596 | } |
597 | ||
482e302a | 598 | static inline bool wbt_should_throttle(struct bio *bio) |
e34cbd30 | 599 | { |
782f5697 JA |
600 | switch (bio_op(bio)) { |
601 | case REQ_OP_WRITE: | |
602 | /* | |
603 | * Don't throttle WRITE_ODIRECT | |
604 | */ | |
605 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
606 | (REQ_SYNC | REQ_IDLE)) | |
607 | return false; | |
df561f66 | 608 | fallthrough; |
782f5697 JA |
609 | case REQ_OP_DISCARD: |
610 | return true; | |
611 | default: | |
e34cbd30 | 612 | return false; |
782f5697 | 613 | } |
e34cbd30 JA |
614 | } |
615 | ||
c1c80384 JB |
616 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) |
617 | { | |
618 | enum wbt_flags flags = 0; | |
619 | ||
c125311d JA |
620 | if (!rwb_enabled(rwb)) |
621 | return 0; | |
622 | ||
c1c80384 JB |
623 | if (bio_op(bio) == REQ_OP_READ) { |
624 | flags = WBT_READ; | |
482e302a | 625 | } else if (wbt_should_throttle(bio)) { |
c1c80384 JB |
626 | if (current_is_kswapd()) |
627 | flags |= WBT_KSWAPD; | |
628 | if (bio_op(bio) == REQ_OP_DISCARD) | |
629 | flags |= WBT_DISCARD; | |
630 | flags |= WBT_TRACKED; | |
631 | } | |
632 | return flags; | |
633 | } | |
634 | ||
635 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
636 | { | |
637 | struct rq_wb *rwb = RQWB(rqos); | |
638 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
639 | __wbt_done(rqos, flags); | |
640 | } | |
641 | ||
e34cbd30 | 642 | /* |
e34cbd30 JA |
643 | * May sleep, if we have exceeded the writeback limits. Caller can pass |
644 | * in an irq held spinlock, if it holds one when calling this function. | |
645 | * If we do sleep, we'll release and re-grab it. | |
646 | */ | |
d5337560 | 647 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio) |
e34cbd30 | 648 | { |
a7905043 | 649 | struct rq_wb *rwb = RQWB(rqos); |
c1c80384 | 650 | enum wbt_flags flags; |
e34cbd30 | 651 | |
c1c80384 | 652 | flags = bio_to_wbt_flags(rwb, bio); |
df60f6e8 | 653 | if (!(flags & WBT_TRACKED)) { |
c1c80384 | 654 | if (flags & WBT_READ) |
e34cbd30 | 655 | wb_timestamp(rwb, &rwb->last_issue); |
c1c80384 | 656 | return; |
e34cbd30 JA |
657 | } |
658 | ||
d5337560 | 659 | __wbt_wait(rwb, flags, bio->bi_opf); |
e34cbd30 | 660 | |
34dbad5d | 661 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 | 662 | rwb_arm_timer(rwb); |
c1c80384 | 663 | } |
e34cbd30 | 664 | |
c1c80384 JB |
665 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
666 | { | |
667 | struct rq_wb *rwb = RQWB(rqos); | |
668 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
e34cbd30 JA |
669 | } |
670 | ||
c83f536a | 671 | static void wbt_issue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 672 | { |
a7905043 JB |
673 | struct rq_wb *rwb = RQWB(rqos); |
674 | ||
e34cbd30 JA |
675 | if (!rwb_enabled(rwb)) |
676 | return; | |
677 | ||
678 | /* | |
a8a45941 OS |
679 | * Track sync issue, in case it takes a long time to complete. Allows us |
680 | * to react quicker, if a sync IO takes a long time to complete. Note | |
681 | * that this is just a hint. The request can go away when it completes, | |
682 | * so it's important we never dereference it. We only use the address to | |
683 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 684 | */ |
a8a45941 OS |
685 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
686 | rwb->sync_cookie = rq; | |
544ccc8d | 687 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
688 | } |
689 | } | |
690 | ||
c83f536a | 691 | static void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 692 | { |
a7905043 | 693 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
694 | if (!rwb_enabled(rwb)) |
695 | return; | |
a8a45941 | 696 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
697 | rwb->sync_issue = 0; |
698 | rwb->sync_cookie = NULL; | |
699 | } | |
700 | } | |
701 | ||
a7905043 | 702 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
e34cbd30 | 703 | { |
a7905043 JB |
704 | struct rq_qos *rqos = wbt_rq_qos(q); |
705 | if (rqos) | |
706 | RQWB(rqos)->wc = write_cache_on; | |
e34cbd30 | 707 | } |
e34cbd30 | 708 | |
8330cdb0 JK |
709 | /* |
710 | * Enable wbt if defaults are configured that way | |
711 | */ | |
04aad37b | 712 | void wbt_enable_default(struct gendisk *disk) |
8330cdb0 | 713 | { |
04aad37b | 714 | struct request_queue *q = disk->queue; |
671fae5e | 715 | struct rq_qos *rqos; |
8a2b20a9 YK |
716 | bool enable = IS_ENABLED(CONFIG_BLK_WBT_MQ); |
717 | ||
718 | if (q->elevator && | |
719 | test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags)) | |
720 | enable = false; | |
76a80408 | 721 | |
8330cdb0 | 722 | /* Throttling already enabled? */ |
671fae5e | 723 | rqos = wbt_rq_qos(q); |
76a80408 | 724 | if (rqos) { |
8a2b20a9 | 725 | if (enable && RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT) |
76a80408 | 726 | RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT; |
8330cdb0 | 727 | return; |
76a80408 | 728 | } |
8330cdb0 JK |
729 | |
730 | /* Queue not registered? Maybe shutting down... */ | |
58c898ba | 731 | if (!blk_queue_registered(q)) |
8330cdb0 JK |
732 | return; |
733 | ||
8a2b20a9 | 734 | if (queue_is_mq(q) && enable) |
958f2965 | 735 | wbt_init(disk); |
8330cdb0 JK |
736 | } |
737 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
738 | ||
80e091d1 JA |
739 | u64 wbt_default_latency_nsec(struct request_queue *q) |
740 | { | |
741 | /* | |
742 | * We default to 2msec for non-rotational storage, and 75msec | |
743 | * for rotational storage. | |
744 | */ | |
745 | if (blk_queue_nonrot(q)) | |
746 | return 2000000ULL; | |
747 | else | |
748 | return 75000000ULL; | |
749 | } | |
750 | ||
99c749a4 JA |
751 | static int wbt_data_dir(const struct request *rq) |
752 | { | |
77e7ffd7 | 753 | const enum req_op op = req_op(rq); |
5235553d JA |
754 | |
755 | if (op == REQ_OP_READ) | |
756 | return READ; | |
825843b0 | 757 | else if (op_is_write(op)) |
5235553d JA |
758 | return WRITE; |
759 | ||
760 | /* don't account */ | |
761 | return -1; | |
99c749a4 JA |
762 | } |
763 | ||
9677a3e0 TH |
764 | static void wbt_queue_depth_changed(struct rq_qos *rqos) |
765 | { | |
ba91c849 | 766 | RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->disk->queue); |
4d89e1d1 | 767 | wbt_update_limits(RQWB(rqos)); |
9677a3e0 TH |
768 | } |
769 | ||
a7905043 JB |
770 | static void wbt_exit(struct rq_qos *rqos) |
771 | { | |
772 | struct rq_wb *rwb = RQWB(rqos); | |
a7905043 | 773 | |
ba91c849 | 774 | blk_stat_remove_callback(rqos->disk->queue, rwb->cb); |
a7905043 JB |
775 | blk_stat_free_callback(rwb->cb); |
776 | kfree(rwb); | |
777 | } | |
778 | ||
779 | /* | |
780 | * Disable wbt, if enabled by default. | |
781 | */ | |
04aad37b | 782 | void wbt_disable_default(struct gendisk *disk) |
a7905043 | 783 | { |
04aad37b | 784 | struct rq_qos *rqos = wbt_rq_qos(disk->queue); |
a7905043 JB |
785 | struct rq_wb *rwb; |
786 | if (!rqos) | |
787 | return; | |
788 | rwb = RQWB(rqos); | |
544fbd16 ML |
789 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) { |
790 | blk_stat_deactivate(rwb->cb); | |
1d0903d6 | 791 | rwb->enable_state = WBT_STATE_OFF_DEFAULT; |
544fbd16 | 792 | } |
a7905043 | 793 | } |
e815f404 | 794 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
a7905043 | 795 | |
d19afebc ML |
796 | #ifdef CONFIG_BLK_DEBUG_FS |
797 | static int wbt_curr_win_nsec_show(void *data, struct seq_file *m) | |
798 | { | |
799 | struct rq_qos *rqos = data; | |
800 | struct rq_wb *rwb = RQWB(rqos); | |
801 | ||
802 | seq_printf(m, "%llu\n", rwb->cur_win_nsec); | |
803 | return 0; | |
804 | } | |
805 | ||
806 | static int wbt_enabled_show(void *data, struct seq_file *m) | |
807 | { | |
808 | struct rq_qos *rqos = data; | |
809 | struct rq_wb *rwb = RQWB(rqos); | |
810 | ||
811 | seq_printf(m, "%d\n", rwb->enable_state); | |
812 | return 0; | |
813 | } | |
814 | ||
815 | static int wbt_id_show(void *data, struct seq_file *m) | |
816 | { | |
817 | struct rq_qos *rqos = data; | |
818 | ||
819 | seq_printf(m, "%u\n", rqos->id); | |
820 | return 0; | |
821 | } | |
822 | ||
823 | static int wbt_inflight_show(void *data, struct seq_file *m) | |
824 | { | |
825 | struct rq_qos *rqos = data; | |
826 | struct rq_wb *rwb = RQWB(rqos); | |
827 | int i; | |
828 | ||
829 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
830 | seq_printf(m, "%d: inflight %d\n", i, | |
831 | atomic_read(&rwb->rq_wait[i].inflight)); | |
832 | return 0; | |
833 | } | |
834 | ||
835 | static int wbt_min_lat_nsec_show(void *data, struct seq_file *m) | |
836 | { | |
837 | struct rq_qos *rqos = data; | |
838 | struct rq_wb *rwb = RQWB(rqos); | |
839 | ||
840 | seq_printf(m, "%lu\n", rwb->min_lat_nsec); | |
841 | return 0; | |
842 | } | |
843 | ||
844 | static int wbt_unknown_cnt_show(void *data, struct seq_file *m) | |
845 | { | |
846 | struct rq_qos *rqos = data; | |
847 | struct rq_wb *rwb = RQWB(rqos); | |
848 | ||
849 | seq_printf(m, "%u\n", rwb->unknown_cnt); | |
850 | return 0; | |
851 | } | |
852 | ||
853 | static int wbt_normal_show(void *data, struct seq_file *m) | |
854 | { | |
855 | struct rq_qos *rqos = data; | |
856 | struct rq_wb *rwb = RQWB(rqos); | |
857 | ||
858 | seq_printf(m, "%u\n", rwb->wb_normal); | |
859 | return 0; | |
860 | } | |
861 | ||
862 | static int wbt_background_show(void *data, struct seq_file *m) | |
863 | { | |
864 | struct rq_qos *rqos = data; | |
865 | struct rq_wb *rwb = RQWB(rqos); | |
866 | ||
867 | seq_printf(m, "%u\n", rwb->wb_background); | |
868 | return 0; | |
869 | } | |
870 | ||
871 | static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = { | |
872 | {"curr_win_nsec", 0400, wbt_curr_win_nsec_show}, | |
873 | {"enabled", 0400, wbt_enabled_show}, | |
874 | {"id", 0400, wbt_id_show}, | |
875 | {"inflight", 0400, wbt_inflight_show}, | |
876 | {"min_lat_nsec", 0400, wbt_min_lat_nsec_show}, | |
877 | {"unknown_cnt", 0400, wbt_unknown_cnt_show}, | |
878 | {"wb_normal", 0400, wbt_normal_show}, | |
879 | {"wb_background", 0400, wbt_background_show}, | |
880 | {}, | |
881 | }; | |
882 | #endif | |
883 | ||
3963d84d | 884 | static const struct rq_qos_ops wbt_rqos_ops = { |
a7905043 JB |
885 | .throttle = wbt_wait, |
886 | .issue = wbt_issue, | |
c1c80384 | 887 | .track = wbt_track, |
a7905043 JB |
888 | .requeue = wbt_requeue, |
889 | .done = wbt_done, | |
c1c80384 | 890 | .cleanup = wbt_cleanup, |
9677a3e0 | 891 | .queue_depth_changed = wbt_queue_depth_changed, |
a7905043 | 892 | .exit = wbt_exit, |
d19afebc ML |
893 | #ifdef CONFIG_BLK_DEBUG_FS |
894 | .debugfs_attrs = wbt_debugfs_attrs, | |
895 | #endif | |
a7905043 JB |
896 | }; |
897 | ||
958f2965 | 898 | int wbt_init(struct gendisk *disk) |
e34cbd30 | 899 | { |
958f2965 | 900 | struct request_queue *q = disk->queue; |
e34cbd30 JA |
901 | struct rq_wb *rwb; |
902 | int i; | |
14a6e2eb | 903 | int ret; |
e34cbd30 | 904 | |
e34cbd30 JA |
905 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
906 | if (!rwb) | |
907 | return -ENOMEM; | |
908 | ||
99c749a4 | 909 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
910 | if (!rwb->cb) { |
911 | kfree(rwb); | |
912 | return -ENOMEM; | |
913 | } | |
914 | ||
a7905043 JB |
915 | for (i = 0; i < WBT_NUM_RWQ; i++) |
916 | rq_wait_init(&rwb->rq_wait[i]); | |
e34cbd30 | 917 | |
e34cbd30 | 918 | rwb->last_comp = rwb->last_issue = jiffies; |
e34cbd30 | 919 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 920 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
285febab | 921 | rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags); |
a7905043 | 922 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; |
8c5035df | 923 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
4e1d91ae CH |
924 | rwb->rq_depth.queue_depth = blk_queue_depth(q); |
925 | wbt_update_limits(rwb); | |
e34cbd30 JA |
926 | |
927 | /* | |
34dbad5d | 928 | * Assign rwb and add the stats callback. |
e34cbd30 | 929 | */ |
a13bd91b | 930 | mutex_lock(&q->rq_qos_mutex); |
ce57b558 | 931 | ret = rq_qos_add(&rwb->rqos, disk, RQ_QOS_WBT, &wbt_rqos_ops); |
a13bd91b | 932 | mutex_unlock(&q->rq_qos_mutex); |
14a6e2eb JH |
933 | if (ret) |
934 | goto err_free; | |
935 | ||
34dbad5d | 936 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 937 | |
e34cbd30 | 938 | return 0; |
14a6e2eb JH |
939 | |
940 | err_free: | |
941 | blk_stat_free_callback(rwb->cb); | |
942 | kfree(rwb); | |
943 | return ret; | |
944 | ||
e34cbd30 | 945 | } |