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 | ||
a7905043 | 126 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
e34cbd30 | 127 | { |
a7905043 | 128 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
129 | struct rq_wait *rqw; |
130 | int inflight, limit; | |
131 | ||
132 | if (!(wb_acct & WBT_TRACKED)) | |
133 | return; | |
134 | ||
8bea6090 | 135 | rqw = get_rq_wait(rwb, wb_acct); |
e34cbd30 JA |
136 | inflight = atomic_dec_return(&rqw->inflight); |
137 | ||
138 | /* | |
139 | * wbt got disabled with IO in flight. Wake up any potential | |
140 | * waiters, we don't have to do more than that. | |
141 | */ | |
142 | if (unlikely(!rwb_enabled(rwb))) { | |
143 | rwb_wake_all(rwb); | |
144 | return; | |
145 | } | |
146 | ||
147 | /* | |
782f5697 JA |
148 | * For discards, our limit is always the background. For writes, if |
149 | * the device does write back caching, drop further down before we | |
150 | * wake people up. | |
e34cbd30 | 151 | */ |
782f5697 JA |
152 | if (wb_acct & WBT_DISCARD) |
153 | limit = rwb->wb_background; | |
154 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
155 | limit = 0; |
156 | else | |
157 | limit = rwb->wb_normal; | |
158 | ||
159 | /* | |
160 | * Don't wake anyone up if we are above the normal limit. | |
161 | */ | |
162 | if (inflight && inflight >= limit) | |
163 | return; | |
164 | ||
b7882093 | 165 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
166 | int diff = limit - inflight; |
167 | ||
168 | if (!inflight || diff >= rwb->wb_background / 2) | |
2887e41b | 169 | wake_up(&rqw->wait); |
e34cbd30 JA |
170 | } |
171 | } | |
172 | ||
173 | /* | |
174 | * Called on completion of a request. Note that it's also called when | |
175 | * a request is merged, when the request gets freed. | |
176 | */ | |
a7905043 | 177 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 178 | { |
a7905043 | 179 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 | 180 | |
a8a45941 OS |
181 | if (!wbt_is_tracked(rq)) { |
182 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
183 | rwb->sync_issue = 0; |
184 | rwb->sync_cookie = NULL; | |
185 | } | |
186 | ||
a8a45941 | 187 | if (wbt_is_read(rq)) |
e34cbd30 | 188 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 189 | } else { |
a8a45941 | 190 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
a7905043 | 191 | __wbt_done(rqos, wbt_flags(rq)); |
e34cbd30 | 192 | } |
a8a45941 | 193 | wbt_clear_state(rq); |
e34cbd30 JA |
194 | } |
195 | ||
4121d385 | 196 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
197 | { |
198 | /* | |
199 | * We need at least one read sample, and a minimum of | |
200 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
201 | * that it's writes impacting us, and not just some sole read on | |
202 | * a device that is in a lower power state. | |
203 | */ | |
fa2e39cb OS |
204 | return (stat[READ].nr_samples >= 1 && |
205 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
206 | } |
207 | ||
208 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
209 | { | |
6aa7de05 | 210 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
211 | |
212 | if (!issue || !rwb->sync_cookie) | |
213 | return 0; | |
214 | ||
215 | now = ktime_to_ns(ktime_get()); | |
216 | return now - issue; | |
217 | } | |
218 | ||
219 | enum { | |
220 | LAT_OK = 1, | |
221 | LAT_UNKNOWN, | |
222 | LAT_UNKNOWN_WRITES, | |
223 | LAT_EXCEEDED, | |
224 | }; | |
225 | ||
34dbad5d | 226 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 227 | { |
a7905043 JB |
228 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; |
229 | struct rq_depth *rqd = &rwb->rq_depth; | |
e34cbd30 JA |
230 | u64 thislat; |
231 | ||
232 | /* | |
233 | * If our stored sync issue exceeds the window size, or it | |
234 | * exceeds our min target AND we haven't logged any entries, | |
235 | * flag the latency as exceeded. wbt works off completion latencies, | |
236 | * but for a flooded device, a single sync IO can take a long time | |
237 | * to complete after being issued. If this time exceeds our | |
238 | * monitoring window AND we didn't see any other completions in that | |
239 | * window, then count that sync IO as a violation of the latency. | |
240 | */ | |
241 | thislat = rwb_sync_issue_lat(rwb); | |
242 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 243 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 244 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
245 | return LAT_EXCEEDED; |
246 | } | |
247 | ||
248 | /* | |
249 | * No read/write mix, if stat isn't valid | |
250 | */ | |
251 | if (!stat_sample_valid(stat)) { | |
252 | /* | |
253 | * If we had writes in this stat window and the window is | |
254 | * current, we're only doing writes. If a task recently | |
255 | * waited or still has writes in flights, consider us doing | |
256 | * just writes as well. | |
257 | */ | |
34dbad5d OS |
258 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
259 | wbt_inflight(rwb)) | |
e34cbd30 JA |
260 | return LAT_UNKNOWN_WRITES; |
261 | return LAT_UNKNOWN; | |
262 | } | |
263 | ||
264 | /* | |
265 | * If the 'min' latency exceeds our target, step down. | |
266 | */ | |
fa2e39cb OS |
267 | if (stat[READ].min > rwb->min_lat_nsec) { |
268 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 269 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
270 | return LAT_EXCEEDED; |
271 | } | |
272 | ||
a7905043 | 273 | if (rqd->scale_step) |
d8a0cbfd | 274 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
275 | |
276 | return LAT_OK; | |
277 | } | |
278 | ||
e34cbd30 JA |
279 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
280 | { | |
a7905043 JB |
281 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; |
282 | struct rq_depth *rqd = &rwb->rq_depth; | |
d8a0cbfd | 283 | |
a7905043 JB |
284 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, |
285 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
e34cbd30 JA |
286 | } |
287 | ||
a7905043 | 288 | static void calc_wb_limits(struct rq_wb *rwb) |
e34cbd30 | 289 | { |
a7905043 JB |
290 | if (rwb->min_lat_nsec == 0) { |
291 | rwb->wb_normal = rwb->wb_background = 0; | |
292 | } else if (rwb->rq_depth.max_depth <= 2) { | |
293 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
294 | rwb->wb_background = 1; | |
295 | } else { | |
296 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
297 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
298 | } | |
299 | } | |
e34cbd30 | 300 | |
a7905043 JB |
301 | static void scale_up(struct rq_wb *rwb) |
302 | { | |
303 | rq_depth_scale_up(&rwb->rq_depth); | |
304 | calc_wb_limits(rwb); | |
e34cbd30 | 305 | rwb->unknown_cnt = 0; |
a7905043 | 306 | rwb_trace_step(rwb, "scale up"); |
e34cbd30 JA |
307 | } |
308 | ||
e34cbd30 JA |
309 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
310 | { | |
a7905043 | 311 | rq_depth_scale_down(&rwb->rq_depth, hard_throttle); |
e34cbd30 | 312 | calc_wb_limits(rwb); |
a7905043 JB |
313 | rwb->unknown_cnt = 0; |
314 | rwb_wake_all(rwb); | |
315 | rwb_trace_step(rwb, "scale down"); | |
e34cbd30 JA |
316 | } |
317 | ||
318 | static void rwb_arm_timer(struct rq_wb *rwb) | |
319 | { | |
a7905043 JB |
320 | struct rq_depth *rqd = &rwb->rq_depth; |
321 | ||
322 | if (rqd->scale_step > 0) { | |
e34cbd30 JA |
323 | /* |
324 | * We should speed this up, using some variant of a fast | |
325 | * integer inverse square root calculation. Since we only do | |
326 | * this for every window expiration, it's not a huge deal, | |
327 | * though. | |
328 | */ | |
329 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
a7905043 | 330 | int_sqrt((rqd->scale_step + 1) << 8)); |
e34cbd30 JA |
331 | } else { |
332 | /* | |
333 | * For step < 0, we don't want to increase/decrease the | |
334 | * window size. | |
335 | */ | |
336 | rwb->cur_win_nsec = rwb->win_nsec; | |
337 | } | |
338 | ||
34dbad5d | 339 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
340 | } |
341 | ||
34dbad5d | 342 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 343 | { |
34dbad5d | 344 | struct rq_wb *rwb = cb->data; |
a7905043 | 345 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
346 | unsigned int inflight = wbt_inflight(rwb); |
347 | int status; | |
348 | ||
34dbad5d | 349 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 350 | |
a7905043 | 351 | trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step, |
d8a0cbfd | 352 | inflight); |
e34cbd30 JA |
353 | |
354 | /* | |
355 | * If we exceeded the latency target, step down. If we did not, | |
356 | * step one level up. If we don't know enough to say either exceeded | |
357 | * or ok, then don't do anything. | |
358 | */ | |
359 | switch (status) { | |
360 | case LAT_EXCEEDED: | |
361 | scale_down(rwb, true); | |
362 | break; | |
363 | case LAT_OK: | |
364 | scale_up(rwb); | |
365 | break; | |
366 | case LAT_UNKNOWN_WRITES: | |
367 | /* | |
368 | * We started a the center step, but don't have a valid | |
369 | * read/write sample, but we do have writes going on. | |
370 | * Allow step to go negative, to increase write perf. | |
371 | */ | |
372 | scale_up(rwb); | |
373 | break; | |
374 | case LAT_UNKNOWN: | |
375 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
376 | break; | |
377 | /* | |
378 | * We get here when previously scaled reduced depth, and we | |
379 | * currently don't have a valid read/write sample. For that | |
380 | * case, slowly return to center state (step == 0). | |
381 | */ | |
a7905043 | 382 | if (rqd->scale_step > 0) |
e34cbd30 | 383 | scale_up(rwb); |
a7905043 | 384 | else if (rqd->scale_step < 0) |
e34cbd30 JA |
385 | scale_down(rwb, false); |
386 | break; | |
387 | default: | |
388 | break; | |
389 | } | |
390 | ||
391 | /* | |
392 | * Re-arm timer, if we have IO in flight | |
393 | */ | |
a7905043 | 394 | if (rqd->scale_step || inflight) |
e34cbd30 JA |
395 | rwb_arm_timer(rwb); |
396 | } | |
397 | ||
a7905043 | 398 | static void __wbt_update_limits(struct rq_wb *rwb) |
e34cbd30 | 399 | { |
a7905043 JB |
400 | struct rq_depth *rqd = &rwb->rq_depth; |
401 | ||
402 | rqd->scale_step = 0; | |
403 | rqd->scaled_max = false; | |
404 | ||
405 | rq_depth_calc_max_depth(rqd); | |
e34cbd30 JA |
406 | calc_wb_limits(rwb); |
407 | ||
408 | rwb_wake_all(rwb); | |
409 | } | |
410 | ||
a7905043 JB |
411 | void wbt_update_limits(struct request_queue *q) |
412 | { | |
413 | struct rq_qos *rqos = wbt_rq_qos(q); | |
414 | if (!rqos) | |
415 | return; | |
416 | __wbt_update_limits(RQWB(rqos)); | |
417 | } | |
418 | ||
419 | u64 wbt_get_min_lat(struct request_queue *q) | |
420 | { | |
421 | struct rq_qos *rqos = wbt_rq_qos(q); | |
422 | if (!rqos) | |
423 | return 0; | |
424 | return RQWB(rqos)->min_lat_nsec; | |
425 | } | |
426 | ||
427 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
428 | { | |
429 | struct rq_qos *rqos = wbt_rq_qos(q); | |
430 | if (!rqos) | |
431 | return; | |
432 | RQWB(rqos)->min_lat_nsec = val; | |
433 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
434 | __wbt_update_limits(RQWB(rqos)); | |
435 | } | |
436 | ||
437 | ||
e34cbd30 JA |
438 | static bool close_io(struct rq_wb *rwb) |
439 | { | |
440 | const unsigned long now = jiffies; | |
441 | ||
442 | return time_before(now, rwb->last_issue + HZ / 10) || | |
443 | time_before(now, rwb->last_comp + HZ / 10); | |
444 | } | |
445 | ||
446 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
447 | ||
448 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
449 | { | |
450 | unsigned int limit; | |
451 | ||
ffa358dc JA |
452 | /* |
453 | * If we got disabled, just return UINT_MAX. This ensures that | |
454 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
455 | */ | |
456 | if (!rwb_enabled(rwb)) | |
457 | return UINT_MAX; | |
458 | ||
782f5697 JA |
459 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) |
460 | return rwb->wb_background; | |
461 | ||
e34cbd30 JA |
462 | /* |
463 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 464 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
465 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
466 | * that. If the write is marked as a background write, then use | |
467 | * the idle limit, or go to normal if we haven't had competing | |
468 | * IO for a bit. | |
469 | */ | |
470 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
a7905043 | 471 | limit = rwb->rq_depth.max_depth; |
e34cbd30 JA |
472 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { |
473 | /* | |
474 | * If less than 100ms since we completed unrelated IO, | |
475 | * limit us to half the depth for background writeback. | |
476 | */ | |
477 | limit = rwb->wb_background; | |
478 | } else | |
479 | limit = rwb->wb_normal; | |
480 | ||
481 | return limit; | |
482 | } | |
483 | ||
e34cbd30 JA |
484 | /* |
485 | * Block if we will exceed our limit, or if we are currently waiting for | |
486 | * the timer to kick off queuing again. | |
487 | */ | |
8bea6090 JA |
488 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
489 | unsigned long rw, spinlock_t *lock) | |
9eca5350 BVA |
490 | __releases(lock) |
491 | __acquires(lock) | |
e34cbd30 | 492 | { |
8bea6090 | 493 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
2887e41b | 494 | DECLARE_WAITQUEUE(wait, current); |
c45e6a03 | 495 | bool has_sleeper; |
2887e41b | 496 | |
c45e6a03 JA |
497 | has_sleeper = wq_has_sleeper(&rqw->wait); |
498 | if (!has_sleeper && rq_wait_inc_below(rqw, get_limit(rwb, rw))) | |
e34cbd30 JA |
499 | return; |
500 | ||
2887e41b | 501 | add_wait_queue_exclusive(&rqw->wait, &wait); |
e34cbd30 | 502 | do { |
2887e41b AA |
503 | set_current_state(TASK_UNINTERRUPTIBLE); |
504 | ||
c45e6a03 | 505 | if (!has_sleeper && rq_wait_inc_below(rqw, get_limit(rwb, rw))) |
e34cbd30 JA |
506 | break; |
507 | ||
9eca5350 | 508 | if (lock) { |
e34cbd30 | 509 | spin_unlock_irq(lock); |
9eca5350 | 510 | io_schedule(); |
e34cbd30 | 511 | spin_lock_irq(lock); |
9eca5350 BVA |
512 | } else |
513 | io_schedule(); | |
c45e6a03 | 514 | has_sleeper = false; |
e34cbd30 JA |
515 | } while (1); |
516 | ||
2887e41b AA |
517 | __set_current_state(TASK_RUNNING); |
518 | remove_wait_queue(&rqw->wait, &wait); | |
e34cbd30 JA |
519 | } |
520 | ||
521 | static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio) | |
522 | { | |
782f5697 JA |
523 | switch (bio_op(bio)) { |
524 | case REQ_OP_WRITE: | |
525 | /* | |
526 | * Don't throttle WRITE_ODIRECT | |
527 | */ | |
528 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == | |
529 | (REQ_SYNC | REQ_IDLE)) | |
530 | return false; | |
531 | /* fallthrough */ | |
532 | case REQ_OP_DISCARD: | |
533 | return true; | |
534 | default: | |
e34cbd30 | 535 | return false; |
782f5697 | 536 | } |
e34cbd30 JA |
537 | } |
538 | ||
c1c80384 JB |
539 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) |
540 | { | |
541 | enum wbt_flags flags = 0; | |
542 | ||
543 | if (bio_op(bio) == REQ_OP_READ) { | |
544 | flags = WBT_READ; | |
545 | } else if (wbt_should_throttle(rwb, bio)) { | |
546 | if (current_is_kswapd()) | |
547 | flags |= WBT_KSWAPD; | |
548 | if (bio_op(bio) == REQ_OP_DISCARD) | |
549 | flags |= WBT_DISCARD; | |
550 | flags |= WBT_TRACKED; | |
551 | } | |
552 | return flags; | |
553 | } | |
554 | ||
555 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
556 | { | |
557 | struct rq_wb *rwb = RQWB(rqos); | |
558 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
559 | __wbt_done(rqos, flags); | |
560 | } | |
561 | ||
e34cbd30 JA |
562 | /* |
563 | * Returns true if the IO request should be accounted, false if not. | |
564 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
565 | * in an irq held spinlock, if it holds one when calling this function. | |
566 | * If we do sleep, we'll release and re-grab it. | |
567 | */ | |
c1c80384 | 568 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio, spinlock_t *lock) |
e34cbd30 | 569 | { |
a7905043 | 570 | struct rq_wb *rwb = RQWB(rqos); |
c1c80384 | 571 | enum wbt_flags flags; |
e34cbd30 | 572 | |
c1c80384 | 573 | flags = bio_to_wbt_flags(rwb, bio); |
df60f6e8 | 574 | if (!(flags & WBT_TRACKED)) { |
c1c80384 | 575 | if (flags & WBT_READ) |
e34cbd30 | 576 | wb_timestamp(rwb, &rwb->last_issue); |
c1c80384 | 577 | return; |
e34cbd30 JA |
578 | } |
579 | ||
8bea6090 | 580 | if (current_is_kswapd()) |
c1c80384 | 581 | flags |= WBT_KSWAPD; |
782f5697 | 582 | if (bio_op(bio) == REQ_OP_DISCARD) |
c1c80384 | 583 | flags |= WBT_DISCARD; |
8bea6090 | 584 | |
c1c80384 | 585 | __wbt_wait(rwb, flags, bio->bi_opf, lock); |
e34cbd30 | 586 | |
34dbad5d | 587 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 | 588 | rwb_arm_timer(rwb); |
c1c80384 | 589 | } |
e34cbd30 | 590 | |
c1c80384 JB |
591 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
592 | { | |
593 | struct rq_wb *rwb = RQWB(rqos); | |
594 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
e34cbd30 JA |
595 | } |
596 | ||
a7905043 | 597 | void wbt_issue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 598 | { |
a7905043 JB |
599 | struct rq_wb *rwb = RQWB(rqos); |
600 | ||
e34cbd30 JA |
601 | if (!rwb_enabled(rwb)) |
602 | return; | |
603 | ||
604 | /* | |
a8a45941 OS |
605 | * Track sync issue, in case it takes a long time to complete. Allows us |
606 | * to react quicker, if a sync IO takes a long time to complete. Note | |
607 | * that this is just a hint. The request can go away when it completes, | |
608 | * so it's important we never dereference it. We only use the address to | |
609 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 610 | */ |
a8a45941 OS |
611 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
612 | rwb->sync_cookie = rq; | |
544ccc8d | 613 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
614 | } |
615 | } | |
616 | ||
a7905043 | 617 | void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 618 | { |
a7905043 | 619 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
620 | if (!rwb_enabled(rwb)) |
621 | return; | |
a8a45941 | 622 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
623 | rwb->sync_issue = 0; |
624 | rwb->sync_cookie = NULL; | |
625 | } | |
626 | } | |
627 | ||
a7905043 | 628 | void wbt_set_queue_depth(struct request_queue *q, unsigned int depth) |
e34cbd30 | 629 | { |
a7905043 JB |
630 | struct rq_qos *rqos = wbt_rq_qos(q); |
631 | if (rqos) { | |
632 | RQWB(rqos)->rq_depth.queue_depth = depth; | |
633 | __wbt_update_limits(RQWB(rqos)); | |
e34cbd30 JA |
634 | } |
635 | } | |
636 | ||
a7905043 | 637 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
e34cbd30 | 638 | { |
a7905043 JB |
639 | struct rq_qos *rqos = wbt_rq_qos(q); |
640 | if (rqos) | |
641 | RQWB(rqos)->wc = write_cache_on; | |
e34cbd30 | 642 | } |
e34cbd30 | 643 | |
8330cdb0 JK |
644 | /* |
645 | * Enable wbt if defaults are configured that way | |
646 | */ | |
647 | void wbt_enable_default(struct request_queue *q) | |
648 | { | |
a7905043 | 649 | struct rq_qos *rqos = wbt_rq_qos(q); |
8330cdb0 | 650 | /* Throttling already enabled? */ |
a7905043 | 651 | if (rqos) |
8330cdb0 JK |
652 | return; |
653 | ||
654 | /* Queue not registered? Maybe shutting down... */ | |
655 | if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags)) | |
656 | return; | |
657 | ||
658 | if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) || | |
659 | (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ))) | |
660 | wbt_init(q); | |
661 | } | |
662 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
663 | ||
80e091d1 JA |
664 | u64 wbt_default_latency_nsec(struct request_queue *q) |
665 | { | |
666 | /* | |
667 | * We default to 2msec for non-rotational storage, and 75msec | |
668 | * for rotational storage. | |
669 | */ | |
670 | if (blk_queue_nonrot(q)) | |
671 | return 2000000ULL; | |
672 | else | |
673 | return 75000000ULL; | |
674 | } | |
675 | ||
99c749a4 JA |
676 | static int wbt_data_dir(const struct request *rq) |
677 | { | |
5235553d JA |
678 | const int op = req_op(rq); |
679 | ||
680 | if (op == REQ_OP_READ) | |
681 | return READ; | |
825843b0 | 682 | else if (op_is_write(op)) |
5235553d JA |
683 | return WRITE; |
684 | ||
685 | /* don't account */ | |
686 | return -1; | |
99c749a4 JA |
687 | } |
688 | ||
a7905043 JB |
689 | static void wbt_exit(struct rq_qos *rqos) |
690 | { | |
691 | struct rq_wb *rwb = RQWB(rqos); | |
692 | struct request_queue *q = rqos->q; | |
693 | ||
694 | blk_stat_remove_callback(q, rwb->cb); | |
695 | blk_stat_free_callback(rwb->cb); | |
696 | kfree(rwb); | |
697 | } | |
698 | ||
699 | /* | |
700 | * Disable wbt, if enabled by default. | |
701 | */ | |
702 | void wbt_disable_default(struct request_queue *q) | |
703 | { | |
704 | struct rq_qos *rqos = wbt_rq_qos(q); | |
705 | struct rq_wb *rwb; | |
706 | if (!rqos) | |
707 | return; | |
708 | rwb = RQWB(rqos); | |
709 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) | |
710 | rwb->wb_normal = 0; | |
711 | } | |
712 | EXPORT_SYMBOL_GPL(wbt_disable_default); | |
713 | ||
714 | ||
715 | static struct rq_qos_ops wbt_rqos_ops = { | |
716 | .throttle = wbt_wait, | |
717 | .issue = wbt_issue, | |
c1c80384 | 718 | .track = wbt_track, |
a7905043 JB |
719 | .requeue = wbt_requeue, |
720 | .done = wbt_done, | |
c1c80384 | 721 | .cleanup = wbt_cleanup, |
a7905043 JB |
722 | .exit = wbt_exit, |
723 | }; | |
724 | ||
8054b89f | 725 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
726 | { |
727 | struct rq_wb *rwb; | |
728 | int i; | |
729 | ||
e34cbd30 JA |
730 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
731 | if (!rwb) | |
732 | return -ENOMEM; | |
733 | ||
99c749a4 | 734 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
735 | if (!rwb->cb) { |
736 | kfree(rwb); | |
737 | return -ENOMEM; | |
738 | } | |
739 | ||
a7905043 JB |
740 | for (i = 0; i < WBT_NUM_RWQ; i++) |
741 | rq_wait_init(&rwb->rq_wait[i]); | |
e34cbd30 | 742 | |
a7905043 JB |
743 | rwb->rqos.id = RQ_QOS_WBT; |
744 | rwb->rqos.ops = &wbt_rqos_ops; | |
745 | rwb->rqos.q = q; | |
e34cbd30 | 746 | rwb->last_comp = rwb->last_issue = jiffies; |
e34cbd30 | 747 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 748 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
a7905043 JB |
749 | rwb->wc = 1; |
750 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; | |
751 | __wbt_update_limits(rwb); | |
e34cbd30 JA |
752 | |
753 | /* | |
34dbad5d | 754 | * Assign rwb and add the stats callback. |
e34cbd30 | 755 | */ |
a7905043 | 756 | rq_qos_add(q, &rwb->rqos); |
34dbad5d | 757 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 758 | |
80e091d1 | 759 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 | 760 | |
a7905043 JB |
761 | wbt_set_queue_depth(q, blk_queue_depth(q)); |
762 | wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); | |
e34cbd30 JA |
763 | |
764 | return 0; | |
765 | } |