Commit | Line | Data |
---|---|---|
e34cbd30 JA |
1 | /* |
2 | * buffered writeback throttling. loosely based on CoDel. We can't drop | |
3 | * packets for IO scheduling, so the logic is something like this: | |
4 | * | |
5 | * - Monitor latencies in a defined window of time. | |
6 | * - If the minimum latency in the above window exceeds some target, increment | |
7 | * scaling step and scale down queue depth by a factor of 2x. The monitoring | |
8 | * window is then shrunk to 100 / sqrt(scaling step + 1). | |
9 | * - For any window where we don't have solid data on what the latencies | |
10 | * look like, retain status quo. | |
11 | * - If latencies look good, decrement scaling step. | |
12 | * - If we're only doing writes, allow the scaling step to go negative. This | |
13 | * will temporarily boost write performance, snapping back to a stable | |
14 | * scaling step of 0 if reads show up or the heavy writers finish. Unlike | |
15 | * positive scaling steps where we shrink the monitoring window, a negative | |
16 | * scaling step retains the default step==0 window size. | |
17 | * | |
18 | * Copyright (C) 2016 Jens Axboe | |
19 | * | |
20 | */ | |
21 | #include <linux/kernel.h> | |
22 | #include <linux/blk_types.h> | |
23 | #include <linux/slab.h> | |
24 | #include <linux/backing-dev.h> | |
25 | #include <linux/swap.h> | |
26 | ||
27 | #include "blk-wbt.h" | |
a7905043 | 28 | #include "blk-rq-qos.h" |
e34cbd30 JA |
29 | |
30 | #define CREATE_TRACE_POINTS | |
31 | #include <trace/events/wbt.h> | |
32 | ||
a8a45941 | 33 | static inline void wbt_clear_state(struct request *rq) |
934031a1 | 34 | { |
544ccc8d | 35 | rq->wbt_flags = 0; |
934031a1 OS |
36 | } |
37 | ||
a8a45941 | 38 | static inline enum wbt_flags wbt_flags(struct request *rq) |
934031a1 | 39 | { |
544ccc8d | 40 | return rq->wbt_flags; |
934031a1 OS |
41 | } |
42 | ||
a8a45941 | 43 | static inline bool wbt_is_tracked(struct request *rq) |
934031a1 | 44 | { |
544ccc8d | 45 | return rq->wbt_flags & WBT_TRACKED; |
934031a1 OS |
46 | } |
47 | ||
a8a45941 | 48 | static inline bool wbt_is_read(struct request *rq) |
934031a1 | 49 | { |
544ccc8d | 50 | return rq->wbt_flags & WBT_READ; |
934031a1 OS |
51 | } |
52 | ||
e34cbd30 JA |
53 | enum { |
54 | /* | |
55 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
56 | * from here depending on device stats | |
57 | */ | |
58 | RWB_DEF_DEPTH = 16, | |
59 | ||
60 | /* | |
61 | * 100msec window | |
62 | */ | |
63 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
64 | ||
65 | /* | |
66 | * Disregard stats, if we don't meet this minimum | |
67 | */ | |
68 | RWB_MIN_WRITE_SAMPLES = 3, | |
69 | ||
70 | /* | |
71 | * If we have this number of consecutive windows with not enough | |
72 | * information to scale up or down, scale up. | |
73 | */ | |
74 | RWB_UNKNOWN_BUMP = 5, | |
75 | }; | |
76 | ||
77 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
78 | { | |
79 | return rwb && rwb->wb_normal != 0; | |
80 | } | |
81 | ||
e34cbd30 JA |
82 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) |
83 | { | |
84 | if (rwb_enabled(rwb)) { | |
85 | const unsigned long cur = jiffies; | |
86 | ||
87 | if (cur != *var) | |
88 | *var = cur; | |
89 | } | |
90 | } | |
91 | ||
92 | /* | |
93 | * If a task was rate throttled in balance_dirty_pages() within the last | |
94 | * second or so, use that to indicate a higher cleaning rate. | |
95 | */ | |
96 | static bool wb_recent_wait(struct rq_wb *rwb) | |
97 | { | |
a7905043 | 98 | struct bdi_writeback *wb = &rwb->rqos.q->backing_dev_info->wb; |
e34cbd30 JA |
99 | |
100 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
101 | } | |
102 | ||
8bea6090 JA |
103 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
104 | enum wbt_flags wb_acct) | |
e34cbd30 | 105 | { |
8bea6090 JA |
106 | if (wb_acct & WBT_KSWAPD) |
107 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
108 | else if (wb_acct & WBT_DISCARD) |
109 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
110 | |
111 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
112 | } |
113 | ||
114 | static void rwb_wake_all(struct rq_wb *rwb) | |
115 | { | |
116 | int i; | |
117 | ||
118 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
119 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
120 | ||
b7882093 | 121 | if (wq_has_sleeper(&rqw->wait)) |
e34cbd30 JA |
122 | wake_up_all(&rqw->wait); |
123 | } | |
124 | } | |
125 | ||
061a5427 JA |
126 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
127 | enum wbt_flags wb_acct) | |
e34cbd30 | 128 | { |
e34cbd30 JA |
129 | int inflight, limit; |
130 | ||
e34cbd30 JA |
131 | inflight = atomic_dec_return(&rqw->inflight); |
132 | ||
133 | /* | |
134 | * wbt got disabled with IO in flight. Wake up any potential | |
135 | * waiters, we don't have to do more than that. | |
136 | */ | |
137 | if (unlikely(!rwb_enabled(rwb))) { | |
138 | rwb_wake_all(rwb); | |
139 | return; | |
140 | } | |
141 | ||
142 | /* | |
782f5697 JA |
143 | * For discards, our limit is always the background. For writes, if |
144 | * the device does write back caching, drop further down before we | |
145 | * wake people up. | |
e34cbd30 | 146 | */ |
782f5697 JA |
147 | if (wb_acct & WBT_DISCARD) |
148 | limit = rwb->wb_background; | |
149 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
150 | limit = 0; |
151 | else | |
152 | limit = rwb->wb_normal; | |
153 | ||
154 | /* | |
155 | * Don't wake anyone up if we are above the normal limit. | |
156 | */ | |
157 | if (inflight && inflight >= limit) | |
158 | return; | |
159 | ||
b7882093 | 160 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
161 | int diff = limit - inflight; |
162 | ||
163 | if (!inflight || diff >= rwb->wb_background / 2) | |
38cfb5a4 | 164 | wake_up_all(&rqw->wait); |
e34cbd30 JA |
165 | } |
166 | } | |
167 | ||
061a5427 JA |
168 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
169 | { | |
170 | struct rq_wb *rwb = RQWB(rqos); | |
171 | struct rq_wait *rqw; | |
172 | ||
173 | if (!(wb_acct & WBT_TRACKED)) | |
174 | return; | |
175 | ||
176 | rqw = get_rq_wait(rwb, wb_acct); | |
177 | wbt_rqw_done(rwb, rqw, wb_acct); | |
178 | } | |
179 | ||
e34cbd30 JA |
180 | /* |
181 | * Called on completion of a request. Note that it's also called when | |
182 | * a request is merged, when the request gets freed. | |
183 | */ | |
a7905043 | 184 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 185 | { |
a7905043 | 186 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 | 187 | |
a8a45941 OS |
188 | if (!wbt_is_tracked(rq)) { |
189 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
190 | rwb->sync_issue = 0; |
191 | rwb->sync_cookie = NULL; | |
192 | } | |
193 | ||
a8a45941 | 194 | if (wbt_is_read(rq)) |
e34cbd30 | 195 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 196 | } else { |
a8a45941 | 197 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
a7905043 | 198 | __wbt_done(rqos, wbt_flags(rq)); |
e34cbd30 | 199 | } |
a8a45941 | 200 | wbt_clear_state(rq); |
e34cbd30 JA |
201 | } |
202 | ||
4121d385 | 203 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
204 | { |
205 | /* | |
206 | * We need at least one read sample, and a minimum of | |
207 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
208 | * that it's writes impacting us, and not just some sole read on | |
209 | * a device that is in a lower power state. | |
210 | */ | |
fa2e39cb OS |
211 | return (stat[READ].nr_samples >= 1 && |
212 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
213 | } |
214 | ||
215 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
216 | { | |
6aa7de05 | 217 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
218 | |
219 | if (!issue || !rwb->sync_cookie) | |
220 | return 0; | |
221 | ||
222 | now = ktime_to_ns(ktime_get()); | |
223 | return now - issue; | |
224 | } | |
225 | ||
226 | enum { | |
227 | LAT_OK = 1, | |
228 | LAT_UNKNOWN, | |
229 | LAT_UNKNOWN_WRITES, | |
230 | LAT_EXCEEDED, | |
231 | }; | |
232 | ||
34dbad5d | 233 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 234 | { |
a7905043 JB |
235 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; |
236 | struct rq_depth *rqd = &rwb->rq_depth; | |
e34cbd30 JA |
237 | u64 thislat; |
238 | ||
239 | /* | |
240 | * If our stored sync issue exceeds the window size, or it | |
241 | * exceeds our min target AND we haven't logged any entries, | |
242 | * flag the latency as exceeded. wbt works off completion latencies, | |
243 | * but for a flooded device, a single sync IO can take a long time | |
244 | * to complete after being issued. If this time exceeds our | |
245 | * monitoring window AND we didn't see any other completions in that | |
246 | * window, then count that sync IO as a violation of the latency. | |
247 | */ | |
248 | thislat = rwb_sync_issue_lat(rwb); | |
249 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 250 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 251 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
252 | return LAT_EXCEEDED; |
253 | } | |
254 | ||
255 | /* | |
256 | * No read/write mix, if stat isn't valid | |
257 | */ | |
258 | if (!stat_sample_valid(stat)) { | |
259 | /* | |
260 | * If we had writes in this stat window and the window is | |
261 | * current, we're only doing writes. If a task recently | |
262 | * waited or still has writes in flights, consider us doing | |
263 | * just writes as well. | |
264 | */ | |
34dbad5d OS |
265 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
266 | wbt_inflight(rwb)) | |
e34cbd30 JA |
267 | return LAT_UNKNOWN_WRITES; |
268 | return LAT_UNKNOWN; | |
269 | } | |
270 | ||
271 | /* | |
272 | * If the 'min' latency exceeds our target, step down. | |
273 | */ | |
fa2e39cb OS |
274 | if (stat[READ].min > rwb->min_lat_nsec) { |
275 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 276 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
277 | return LAT_EXCEEDED; |
278 | } | |
279 | ||
a7905043 | 280 | if (rqd->scale_step) |
d8a0cbfd | 281 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
282 | |
283 | return LAT_OK; | |
284 | } | |
285 | ||
e34cbd30 JA |
286 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
287 | { | |
a7905043 JB |
288 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; |
289 | struct rq_depth *rqd = &rwb->rq_depth; | |
d8a0cbfd | 290 | |
a7905043 JB |
291 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, |
292 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
e34cbd30 JA |
293 | } |
294 | ||
a7905043 | 295 | static void calc_wb_limits(struct rq_wb *rwb) |
e34cbd30 | 296 | { |
a7905043 JB |
297 | if (rwb->min_lat_nsec == 0) { |
298 | rwb->wb_normal = rwb->wb_background = 0; | |
299 | } else if (rwb->rq_depth.max_depth <= 2) { | |
300 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
301 | rwb->wb_background = 1; | |
302 | } else { | |
303 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
304 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
305 | } | |
306 | } | |
e34cbd30 | 307 | |
a7905043 JB |
308 | static void scale_up(struct rq_wb *rwb) |
309 | { | |
310 | rq_depth_scale_up(&rwb->rq_depth); | |
311 | calc_wb_limits(rwb); | |
e34cbd30 | 312 | rwb->unknown_cnt = 0; |
5e65a203 | 313 | rwb_wake_all(rwb); |
a7905043 | 314 | rwb_trace_step(rwb, "scale up"); |
e34cbd30 JA |
315 | } |
316 | ||
e34cbd30 JA |
317 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
318 | { | |
a7905043 | 319 | rq_depth_scale_down(&rwb->rq_depth, hard_throttle); |
e34cbd30 | 320 | calc_wb_limits(rwb); |
a7905043 | 321 | rwb->unknown_cnt = 0; |
a7905043 | 322 | rwb_trace_step(rwb, "scale down"); |
e34cbd30 JA |
323 | } |
324 | ||
325 | static void rwb_arm_timer(struct rq_wb *rwb) | |
326 | { | |
a7905043 JB |
327 | struct rq_depth *rqd = &rwb->rq_depth; |
328 | ||
329 | if (rqd->scale_step > 0) { | |
e34cbd30 JA |
330 | /* |
331 | * We should speed this up, using some variant of a fast | |
332 | * integer inverse square root calculation. Since we only do | |
333 | * this for every window expiration, it's not a huge deal, | |
334 | * though. | |
335 | */ | |
336 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
a7905043 | 337 | int_sqrt((rqd->scale_step + 1) << 8)); |
e34cbd30 JA |
338 | } else { |
339 | /* | |
340 | * For step < 0, we don't want to increase/decrease the | |
341 | * window size. | |
342 | */ | |
343 | rwb->cur_win_nsec = rwb->win_nsec; | |
344 | } | |
345 | ||
34dbad5d | 346 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
347 | } |
348 | ||
34dbad5d | 349 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 350 | { |
34dbad5d | 351 | struct rq_wb *rwb = cb->data; |
a7905043 | 352 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
353 | unsigned int inflight = wbt_inflight(rwb); |
354 | int status; | |
355 | ||
34dbad5d | 356 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 357 | |
a7905043 | 358 | trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step, |
d8a0cbfd | 359 | inflight); |
e34cbd30 JA |
360 | |
361 | /* | |
362 | * If we exceeded the latency target, step down. If we did not, | |
363 | * step one level up. If we don't know enough to say either exceeded | |
364 | * or ok, then don't do anything. | |
365 | */ | |
366 | switch (status) { | |
367 | case LAT_EXCEEDED: | |
368 | scale_down(rwb, true); | |
369 | break; | |
370 | case LAT_OK: | |
371 | scale_up(rwb); | |
372 | break; | |
373 | case LAT_UNKNOWN_WRITES: | |
374 | /* | |
375 | * We started a the center step, but don't have a valid | |
376 | * read/write sample, but we do have writes going on. | |
377 | * Allow step to go negative, to increase write perf. | |
378 | */ | |
379 | scale_up(rwb); | |
380 | break; | |
381 | case LAT_UNKNOWN: | |
382 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
383 | break; | |
384 | /* | |
385 | * We get here when previously scaled reduced depth, and we | |
386 | * currently don't have a valid read/write sample. For that | |
387 | * case, slowly return to center state (step == 0). | |
388 | */ | |
a7905043 | 389 | if (rqd->scale_step > 0) |
e34cbd30 | 390 | scale_up(rwb); |
a7905043 | 391 | else if (rqd->scale_step < 0) |
e34cbd30 JA |
392 | scale_down(rwb, false); |
393 | break; | |
394 | default: | |
395 | break; | |
396 | } | |
397 | ||
398 | /* | |
399 | * Re-arm timer, if we have IO in flight | |
400 | */ | |
a7905043 | 401 | if (rqd->scale_step || inflight) |
e34cbd30 JA |
402 | rwb_arm_timer(rwb); |
403 | } | |
404 | ||
a7905043 | 405 | static void __wbt_update_limits(struct rq_wb *rwb) |
e34cbd30 | 406 | { |
a7905043 JB |
407 | struct rq_depth *rqd = &rwb->rq_depth; |
408 | ||
409 | rqd->scale_step = 0; | |
410 | rqd->scaled_max = false; | |
411 | ||
412 | rq_depth_calc_max_depth(rqd); | |
e34cbd30 JA |
413 | calc_wb_limits(rwb); |
414 | ||
415 | rwb_wake_all(rwb); | |
416 | } | |
417 | ||
a7905043 JB |
418 | void wbt_update_limits(struct request_queue *q) |
419 | { | |
420 | struct rq_qos *rqos = wbt_rq_qos(q); | |
421 | if (!rqos) | |
422 | return; | |
423 | __wbt_update_limits(RQWB(rqos)); | |
424 | } | |
425 | ||
426 | u64 wbt_get_min_lat(struct request_queue *q) | |
427 | { | |
428 | struct rq_qos *rqos = wbt_rq_qos(q); | |
429 | if (!rqos) | |
430 | return 0; | |
431 | return RQWB(rqos)->min_lat_nsec; | |
432 | } | |
433 | ||
434 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
435 | { | |
436 | struct rq_qos *rqos = wbt_rq_qos(q); | |
437 | if (!rqos) | |
438 | return; | |
439 | RQWB(rqos)->min_lat_nsec = val; | |
440 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
441 | __wbt_update_limits(RQWB(rqos)); | |
442 | } | |
443 | ||
444 | ||
e34cbd30 JA |
445 | static bool close_io(struct rq_wb *rwb) |
446 | { | |
447 | const unsigned long now = jiffies; | |
448 | ||
449 | return time_before(now, rwb->last_issue + HZ / 10) || | |
450 | time_before(now, rwb->last_comp + HZ / 10); | |
451 | } | |
452 | ||
453 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
454 | ||
455 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
456 | { | |
457 | unsigned int limit; | |
458 | ||
ffa358dc JA |
459 | /* |
460 | * If we got disabled, just return UINT_MAX. This ensures that | |
461 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
462 | */ | |
463 | if (!rwb_enabled(rwb)) | |
464 | return UINT_MAX; | |
465 | ||
782f5697 JA |
466 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) |
467 | return rwb->wb_background; | |
468 | ||
e34cbd30 JA |
469 | /* |
470 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 471 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
472 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
473 | * that. If the write is marked as a background write, then use | |
474 | * the idle limit, or go to normal if we haven't had competing | |
475 | * IO for a bit. | |
476 | */ | |
477 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
a7905043 | 478 | limit = rwb->rq_depth.max_depth; |
e34cbd30 JA |
479 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { |
480 | /* | |
481 | * If less than 100ms since we completed unrelated IO, | |
482 | * limit us to half the depth for background writeback. | |
483 | */ | |
484 | limit = rwb->wb_background; | |
485 | } else | |
486 | limit = rwb->wb_normal; | |
487 | ||
488 | return limit; | |
489 | } | |
490 | ||
38cfb5a4 JA |
491 | struct wbt_wait_data { |
492 | struct wait_queue_entry wq; | |
493 | struct task_struct *task; | |
494 | struct rq_wb *rwb; | |
495 | struct rq_wait *rqw; | |
496 | unsigned long rw; | |
497 | bool got_token; | |
498 | }; | |
499 | ||
500 | static int wbt_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
501 | int wake_flags, void *key) | |
502 | { | |
503 | struct wbt_wait_data *data = container_of(curr, struct wbt_wait_data, | |
504 | wq); | |
505 | ||
506 | /* | |
507 | * If we fail to get a budget, return -1 to interrupt the wake up | |
508 | * loop in __wake_up_common. | |
509 | */ | |
510 | if (!rq_wait_inc_below(data->rqw, get_limit(data->rwb, data->rw))) | |
511 | return -1; | |
512 | ||
513 | data->got_token = true; | |
514 | list_del_init(&curr->entry); | |
515 | wake_up_process(data->task); | |
516 | return 1; | |
517 | } | |
518 | ||
e34cbd30 JA |
519 | /* |
520 | * Block if we will exceed our limit, or if we are currently waiting for | |
521 | * the timer to kick off queuing again. | |
522 | */ | |
8bea6090 JA |
523 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
524 | unsigned long rw, spinlock_t *lock) | |
9eca5350 BVA |
525 | __releases(lock) |
526 | __acquires(lock) | |
e34cbd30 | 527 | { |
8bea6090 | 528 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
38cfb5a4 JA |
529 | struct wbt_wait_data data = { |
530 | .wq = { | |
531 | .func = wbt_wake_function, | |
532 | .entry = LIST_HEAD_INIT(data.wq.entry), | |
533 | }, | |
534 | .task = current, | |
535 | .rwb = rwb, | |
536 | .rqw = rqw, | |
537 | .rw = rw, | |
538 | }; | |
c45e6a03 | 539 | bool has_sleeper; |
2887e41b | 540 | |
c45e6a03 JA |
541 | has_sleeper = wq_has_sleeper(&rqw->wait); |
542 | if (!has_sleeper && rq_wait_inc_below(rqw, get_limit(rwb, rw))) | |
e34cbd30 JA |
543 | return; |
544 | ||
38cfb5a4 | 545 | prepare_to_wait_exclusive(&rqw->wait, &data.wq, TASK_UNINTERRUPTIBLE); |
e34cbd30 | 546 | do { |
38cfb5a4 JA |
547 | if (data.got_token) |
548 | break; | |
2887e41b | 549 | |
38cfb5a4 JA |
550 | if (!has_sleeper && |
551 | rq_wait_inc_below(rqw, get_limit(rwb, rw))) { | |
552 | finish_wait(&rqw->wait, &data.wq); | |
553 | ||
554 | /* | |
555 | * We raced with wbt_wake_function() getting a token, | |
556 | * which means we now have two. Put our local token | |
557 | * and wake anyone else potentially waiting for one. | |
558 | */ | |
559 | if (data.got_token) | |
560 | wbt_rqw_done(rwb, rqw, wb_acct); | |
e34cbd30 | 561 | break; |
38cfb5a4 | 562 | } |
e34cbd30 | 563 | |
9eca5350 | 564 | if (lock) { |
e34cbd30 | 565 | spin_unlock_irq(lock); |
9eca5350 | 566 | io_schedule(); |
e34cbd30 | 567 | spin_lock_irq(lock); |
9eca5350 BVA |
568 | } else |
569 | io_schedule(); | |
38cfb5a4 | 570 | |
c45e6a03 | 571 | has_sleeper = false; |
e34cbd30 JA |
572 | } while (1); |
573 | ||
38cfb5a4 | 574 | finish_wait(&rqw->wait, &data.wq); |
e34cbd30 JA |
575 | } |
576 | ||
577 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
578 | { | |
782f5697 JA |
579 | switch (bio_op(bio)) { |
580 | case REQ_OP_WRITE: | |
581 | /* | |
582 | * Don't throttle WRITE_ODIRECT | |
583 | */ | |
584 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
585 | (REQ_SYNC | REQ_IDLE)) | |
586 | return false; | |
587 | /* fallthrough */ | |
588 | case REQ_OP_DISCARD: | |
589 | return true; | |
590 | default: | |
e34cbd30 | 591 | return false; |
782f5697 | 592 | } |
e34cbd30 JA |
593 | } |
594 | ||
c1c80384 JB |
595 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) |
596 | { | |
597 | enum wbt_flags flags = 0; | |
598 | ||
c125311d JA |
599 | if (!rwb_enabled(rwb)) |
600 | return 0; | |
601 | ||
c1c80384 JB |
602 | if (bio_op(bio) == REQ_OP_READ) { |
603 | flags = WBT_READ; | |
604 | } else if (wbt_should_throttle(rwb, bio)) { | |
605 | if (current_is_kswapd()) | |
606 | flags |= WBT_KSWAPD; | |
607 | if (bio_op(bio) == REQ_OP_DISCARD) | |
608 | flags |= WBT_DISCARD; | |
609 | flags |= WBT_TRACKED; | |
610 | } | |
611 | return flags; | |
612 | } | |
613 | ||
614 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
615 | { | |
616 | struct rq_wb *rwb = RQWB(rqos); | |
617 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
618 | __wbt_done(rqos, flags); | |
619 | } | |
620 | ||
e34cbd30 JA |
621 | /* |
622 | * Returns true if the IO request should be accounted, false if not. | |
623 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
624 | * in an irq held spinlock, if it holds one when calling this function. | |
625 | * If we do sleep, we'll release and re-grab it. | |
626 | */ | |
c1c80384 | 627 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio, spinlock_t *lock) |
e34cbd30 | 628 | { |
a7905043 | 629 | struct rq_wb *rwb = RQWB(rqos); |
c1c80384 | 630 | enum wbt_flags flags; |
e34cbd30 | 631 | |
c1c80384 | 632 | flags = bio_to_wbt_flags(rwb, bio); |
df60f6e8 | 633 | if (!(flags & WBT_TRACKED)) { |
c1c80384 | 634 | if (flags & WBT_READ) |
e34cbd30 | 635 | wb_timestamp(rwb, &rwb->last_issue); |
c1c80384 | 636 | return; |
e34cbd30 JA |
637 | } |
638 | ||
c1c80384 | 639 | __wbt_wait(rwb, flags, bio->bi_opf, lock); |
e34cbd30 | 640 | |
34dbad5d | 641 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 | 642 | rwb_arm_timer(rwb); |
c1c80384 | 643 | } |
e34cbd30 | 644 | |
c1c80384 JB |
645 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
646 | { | |
647 | struct rq_wb *rwb = RQWB(rqos); | |
648 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
e34cbd30 JA |
649 | } |
650 | ||
a7905043 | 651 | void wbt_issue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 652 | { |
a7905043 JB |
653 | struct rq_wb *rwb = RQWB(rqos); |
654 | ||
e34cbd30 JA |
655 | if (!rwb_enabled(rwb)) |
656 | return; | |
657 | ||
658 | /* | |
a8a45941 OS |
659 | * Track sync issue, in case it takes a long time to complete. Allows us |
660 | * to react quicker, if a sync IO takes a long time to complete. Note | |
661 | * that this is just a hint. The request can go away when it completes, | |
662 | * so it's important we never dereference it. We only use the address to | |
663 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 664 | */ |
a8a45941 OS |
665 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
666 | rwb->sync_cookie = rq; | |
544ccc8d | 667 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
668 | } |
669 | } | |
670 | ||
a7905043 | 671 | void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 672 | { |
a7905043 | 673 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
674 | if (!rwb_enabled(rwb)) |
675 | return; | |
a8a45941 | 676 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
677 | rwb->sync_issue = 0; |
678 | rwb->sync_cookie = NULL; | |
679 | } | |
680 | } | |
681 | ||
a7905043 | 682 | void wbt_set_queue_depth(struct request_queue *q, unsigned int depth) |
e34cbd30 | 683 | { |
a7905043 JB |
684 | struct rq_qos *rqos = wbt_rq_qos(q); |
685 | if (rqos) { | |
686 | RQWB(rqos)->rq_depth.queue_depth = depth; | |
687 | __wbt_update_limits(RQWB(rqos)); | |
e34cbd30 JA |
688 | } |
689 | } | |
690 | ||
a7905043 | 691 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
e34cbd30 | 692 | { |
a7905043 JB |
693 | struct rq_qos *rqos = wbt_rq_qos(q); |
694 | if (rqos) | |
695 | RQWB(rqos)->wc = write_cache_on; | |
e34cbd30 | 696 | } |
e34cbd30 | 697 | |
8330cdb0 JK |
698 | /* |
699 | * Enable wbt if defaults are configured that way | |
700 | */ | |
701 | void wbt_enable_default(struct request_queue *q) | |
702 | { | |
a7905043 | 703 | struct rq_qos *rqos = wbt_rq_qos(q); |
8330cdb0 | 704 | /* Throttling already enabled? */ |
a7905043 | 705 | if (rqos) |
8330cdb0 JK |
706 | return; |
707 | ||
708 | /* Queue not registered? Maybe shutting down... */ | |
709 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
710 | return; | |
711 | ||
712 | if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) || | |
713 | (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ))) | |
714 | wbt_init(q); | |
715 | } | |
716 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
717 | ||
80e091d1 JA |
718 | u64 wbt_default_latency_nsec(struct request_queue *q) |
719 | { | |
720 | /* | |
721 | * We default to 2msec for non-rotational storage, and 75msec | |
722 | * for rotational storage. | |
723 | */ | |
724 | if (blk_queue_nonrot(q)) | |
725 | return 2000000ULL; | |
726 | else | |
727 | return 75000000ULL; | |
728 | } | |
729 | ||
99c749a4 JA |
730 | static int wbt_data_dir(const struct request *rq) |
731 | { | |
5235553d JA |
732 | const int op = req_op(rq); |
733 | ||
734 | if (op == REQ_OP_READ) | |
735 | return READ; | |
825843b0 | 736 | else if (op_is_write(op)) |
5235553d JA |
737 | return WRITE; |
738 | ||
739 | /* don't account */ | |
740 | return -1; | |
99c749a4 JA |
741 | } |
742 | ||
a7905043 JB |
743 | static void wbt_exit(struct rq_qos *rqos) |
744 | { | |
745 | struct rq_wb *rwb = RQWB(rqos); | |
746 | struct request_queue *q = rqos->q; | |
747 | ||
748 | blk_stat_remove_callback(q, rwb->cb); | |
749 | blk_stat_free_callback(rwb->cb); | |
750 | kfree(rwb); | |
751 | } | |
752 | ||
753 | /* | |
754 | * Disable wbt, if enabled by default. | |
755 | */ | |
756 | void wbt_disable_default(struct request_queue *q) | |
757 | { | |
758 | struct rq_qos *rqos = wbt_rq_qos(q); | |
759 | struct rq_wb *rwb; | |
760 | if (!rqos) | |
761 | return; | |
762 | rwb = RQWB(rqos); | |
763 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) | |
764 | rwb->wb_normal = 0; | |
765 | } | |
766 | EXPORT_SYMBOL_GPL(wbt_disable_default); | |
767 | ||
768 | ||
769 | static struct rq_qos_ops wbt_rqos_ops = { | |
770 | .throttle = wbt_wait, | |
771 | .issue = wbt_issue, | |
c1c80384 | 772 | .track = wbt_track, |
a7905043 JB |
773 | .requeue = wbt_requeue, |
774 | .done = wbt_done, | |
c1c80384 | 775 | .cleanup = wbt_cleanup, |
a7905043 JB |
776 | .exit = wbt_exit, |
777 | }; | |
778 | ||
8054b89f | 779 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
780 | { |
781 | struct rq_wb *rwb; | |
782 | int i; | |
783 | ||
e34cbd30 JA |
784 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
785 | if (!rwb) | |
786 | return -ENOMEM; | |
787 | ||
99c749a4 | 788 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
789 | if (!rwb->cb) { |
790 | kfree(rwb); | |
791 | return -ENOMEM; | |
792 | } | |
793 | ||
a7905043 JB |
794 | for (i = 0; i < WBT_NUM_RWQ; i++) |
795 | rq_wait_init(&rwb->rq_wait[i]); | |
e34cbd30 | 796 | |
a7905043 JB |
797 | rwb->rqos.id = RQ_QOS_WBT; |
798 | rwb->rqos.ops = &wbt_rqos_ops; | |
799 | rwb->rqos.q = q; | |
e34cbd30 | 800 | rwb->last_comp = rwb->last_issue = jiffies; |
e34cbd30 | 801 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 802 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
a7905043 JB |
803 | rwb->wc = 1; |
804 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; | |
805 | __wbt_update_limits(rwb); | |
e34cbd30 JA |
806 | |
807 | /* | |
34dbad5d | 808 | * Assign rwb and add the stats callback. |
e34cbd30 | 809 | */ |
a7905043 | 810 | rq_qos_add(q, &rwb->rqos); |
34dbad5d | 811 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 812 | |
80e091d1 | 813 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 | 814 | |
a7905043 JB |
815 | wbt_set_queue_depth(q, blk_queue_depth(q)); |
816 | wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
e34cbd30 JA |
817 | |
818 | return 0; | |
819 | } |