Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
e34cbd30 JA |
2 | /* |
3 | * buffered writeback throttling. loosely based on CoDel. We can't drop | |
4 | * packets for IO scheduling, so the logic is something like this: | |
5 | * | |
6 | * - Monitor latencies in a defined window of time. | |
7 | * - If the minimum latency in the above window exceeds some target, increment | |
8 | * scaling step and scale down queue depth by a factor of 2x. The monitoring | |
9 | * window is then shrunk to 100 / sqrt(scaling step + 1). | |
10 | * - For any window where we don't have solid data on what the latencies | |
11 | * look like, retain status quo. | |
12 | * - If latencies look good, decrement scaling step. | |
13 | * - If we're only doing writes, allow the scaling step to go negative. This | |
14 | * will temporarily boost write performance, snapping back to a stable | |
15 | * scaling step of 0 if reads show up or the heavy writers finish. Unlike | |
16 | * positive scaling steps where we shrink the monitoring window, a negative | |
17 | * scaling step retains the default step==0 window size. | |
18 | * | |
19 | * Copyright (C) 2016 Jens Axboe | |
20 | * | |
21 | */ | |
22 | #include <linux/kernel.h> | |
23 | #include <linux/blk_types.h> | |
24 | #include <linux/slab.h> | |
25 | #include <linux/backing-dev.h> | |
26 | #include <linux/swap.h> | |
27 | ||
28 | #include "blk-wbt.h" | |
a7905043 | 29 | #include "blk-rq-qos.h" |
e34cbd30 JA |
30 | |
31 | #define CREATE_TRACE_POINTS | |
32 | #include <trace/events/wbt.h> | |
33 | ||
a8a45941 | 34 | static inline void wbt_clear_state(struct request *rq) |
934031a1 | 35 | { |
544ccc8d | 36 | rq->wbt_flags = 0; |
934031a1 OS |
37 | } |
38 | ||
a8a45941 | 39 | static inline enum wbt_flags wbt_flags(struct request *rq) |
934031a1 | 40 | { |
544ccc8d | 41 | return rq->wbt_flags; |
934031a1 OS |
42 | } |
43 | ||
a8a45941 | 44 | static inline bool wbt_is_tracked(struct request *rq) |
934031a1 | 45 | { |
544ccc8d | 46 | return rq->wbt_flags & WBT_TRACKED; |
934031a1 OS |
47 | } |
48 | ||
a8a45941 | 49 | static inline bool wbt_is_read(struct request *rq) |
934031a1 | 50 | { |
544ccc8d | 51 | return rq->wbt_flags & WBT_READ; |
934031a1 OS |
52 | } |
53 | ||
e34cbd30 JA |
54 | enum { |
55 | /* | |
56 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) | |
57 | * from here depending on device stats | |
58 | */ | |
59 | RWB_DEF_DEPTH = 16, | |
60 | ||
61 | /* | |
62 | * 100msec window | |
63 | */ | |
64 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, | |
65 | ||
66 | /* | |
67 | * Disregard stats, if we don't meet this minimum | |
68 | */ | |
69 | RWB_MIN_WRITE_SAMPLES = 3, | |
70 | ||
71 | /* | |
72 | * If we have this number of consecutive windows with not enough | |
73 | * information to scale up or down, scale up. | |
74 | */ | |
75 | RWB_UNKNOWN_BUMP = 5, | |
76 | }; | |
77 | ||
78 | static inline bool rwb_enabled(struct rq_wb *rwb) | |
79 | { | |
80 | return rwb && rwb->wb_normal != 0; | |
81 | } | |
82 | ||
e34cbd30 JA |
83 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) |
84 | { | |
85 | if (rwb_enabled(rwb)) { | |
86 | const unsigned long cur = jiffies; | |
87 | ||
88 | if (cur != *var) | |
89 | *var = cur; | |
90 | } | |
91 | } | |
92 | ||
93 | /* | |
94 | * If a task was rate throttled in balance_dirty_pages() within the last | |
95 | * second or so, use that to indicate a higher cleaning rate. | |
96 | */ | |
97 | static bool wb_recent_wait(struct rq_wb *rwb) | |
98 | { | |
a7905043 | 99 | struct bdi_writeback *wb = &rwb->rqos.q->backing_dev_info->wb; |
e34cbd30 JA |
100 | |
101 | return time_before(jiffies, wb->dirty_sleep + HZ); | |
102 | } | |
103 | ||
8bea6090 JA |
104 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
105 | enum wbt_flags wb_acct) | |
e34cbd30 | 106 | { |
8bea6090 JA |
107 | if (wb_acct & WBT_KSWAPD) |
108 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; | |
782f5697 JA |
109 | else if (wb_acct & WBT_DISCARD) |
110 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; | |
8bea6090 JA |
111 | |
112 | return &rwb->rq_wait[WBT_RWQ_BG]; | |
e34cbd30 JA |
113 | } |
114 | ||
115 | static void rwb_wake_all(struct rq_wb *rwb) | |
116 | { | |
117 | int i; | |
118 | ||
119 | for (i = 0; i < WBT_NUM_RWQ; i++) { | |
120 | struct rq_wait *rqw = &rwb->rq_wait[i]; | |
121 | ||
b7882093 | 122 | if (wq_has_sleeper(&rqw->wait)) |
e34cbd30 JA |
123 | wake_up_all(&rqw->wait); |
124 | } | |
125 | } | |
126 | ||
061a5427 JA |
127 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
128 | enum wbt_flags wb_acct) | |
e34cbd30 | 129 | { |
e34cbd30 JA |
130 | int inflight, limit; |
131 | ||
e34cbd30 JA |
132 | inflight = atomic_dec_return(&rqw->inflight); |
133 | ||
134 | /* | |
135 | * wbt got disabled with IO in flight. Wake up any potential | |
136 | * waiters, we don't have to do more than that. | |
137 | */ | |
138 | if (unlikely(!rwb_enabled(rwb))) { | |
139 | rwb_wake_all(rwb); | |
140 | return; | |
141 | } | |
142 | ||
143 | /* | |
782f5697 JA |
144 | * For discards, our limit is always the background. For writes, if |
145 | * the device does write back caching, drop further down before we | |
146 | * wake people up. | |
e34cbd30 | 147 | */ |
782f5697 JA |
148 | if (wb_acct & WBT_DISCARD) |
149 | limit = rwb->wb_background; | |
150 | else if (rwb->wc && !wb_recent_wait(rwb)) | |
e34cbd30 JA |
151 | limit = 0; |
152 | else | |
153 | limit = rwb->wb_normal; | |
154 | ||
155 | /* | |
156 | * Don't wake anyone up if we are above the normal limit. | |
157 | */ | |
158 | if (inflight && inflight >= limit) | |
159 | return; | |
160 | ||
b7882093 | 161 | if (wq_has_sleeper(&rqw->wait)) { |
e34cbd30 JA |
162 | int diff = limit - inflight; |
163 | ||
164 | if (!inflight || diff >= rwb->wb_background / 2) | |
38cfb5a4 | 165 | wake_up_all(&rqw->wait); |
e34cbd30 JA |
166 | } |
167 | } | |
168 | ||
061a5427 JA |
169 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
170 | { | |
171 | struct rq_wb *rwb = RQWB(rqos); | |
172 | struct rq_wait *rqw; | |
173 | ||
174 | if (!(wb_acct & WBT_TRACKED)) | |
175 | return; | |
176 | ||
177 | rqw = get_rq_wait(rwb, wb_acct); | |
178 | wbt_rqw_done(rwb, rqw, wb_acct); | |
179 | } | |
180 | ||
e34cbd30 JA |
181 | /* |
182 | * Called on completion of a request. Note that it's also called when | |
183 | * a request is merged, when the request gets freed. | |
184 | */ | |
a7905043 | 185 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 186 | { |
a7905043 | 187 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 | 188 | |
a8a45941 OS |
189 | if (!wbt_is_tracked(rq)) { |
190 | if (rwb->sync_cookie == rq) { | |
e34cbd30 JA |
191 | rwb->sync_issue = 0; |
192 | rwb->sync_cookie = NULL; | |
193 | } | |
194 | ||
a8a45941 | 195 | if (wbt_is_read(rq)) |
e34cbd30 | 196 | wb_timestamp(rwb, &rwb->last_comp); |
e34cbd30 | 197 | } else { |
a8a45941 | 198 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
a7905043 | 199 | __wbt_done(rqos, wbt_flags(rq)); |
e34cbd30 | 200 | } |
a8a45941 | 201 | wbt_clear_state(rq); |
e34cbd30 JA |
202 | } |
203 | ||
4121d385 | 204 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
e34cbd30 JA |
205 | { |
206 | /* | |
207 | * We need at least one read sample, and a minimum of | |
208 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know | |
209 | * that it's writes impacting us, and not just some sole read on | |
210 | * a device that is in a lower power state. | |
211 | */ | |
fa2e39cb OS |
212 | return (stat[READ].nr_samples >= 1 && |
213 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); | |
e34cbd30 JA |
214 | } |
215 | ||
216 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) | |
217 | { | |
6aa7de05 | 218 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
e34cbd30 JA |
219 | |
220 | if (!issue || !rwb->sync_cookie) | |
221 | return 0; | |
222 | ||
223 | now = ktime_to_ns(ktime_get()); | |
224 | return now - issue; | |
225 | } | |
226 | ||
227 | enum { | |
228 | LAT_OK = 1, | |
229 | LAT_UNKNOWN, | |
230 | LAT_UNKNOWN_WRITES, | |
231 | LAT_EXCEEDED, | |
232 | }; | |
233 | ||
34dbad5d | 234 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
e34cbd30 | 235 | { |
a7905043 JB |
236 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; |
237 | struct rq_depth *rqd = &rwb->rq_depth; | |
e34cbd30 JA |
238 | u64 thislat; |
239 | ||
240 | /* | |
241 | * If our stored sync issue exceeds the window size, or it | |
242 | * exceeds our min target AND we haven't logged any entries, | |
243 | * flag the latency as exceeded. wbt works off completion latencies, | |
244 | * but for a flooded device, a single sync IO can take a long time | |
245 | * to complete after being issued. If this time exceeds our | |
246 | * monitoring window AND we didn't see any other completions in that | |
247 | * window, then count that sync IO as a violation of the latency. | |
248 | */ | |
249 | thislat = rwb_sync_issue_lat(rwb); | |
250 | if (thislat > rwb->cur_win_nsec || | |
fa2e39cb | 251 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
d8a0cbfd | 252 | trace_wbt_lat(bdi, thislat); |
e34cbd30 JA |
253 | return LAT_EXCEEDED; |
254 | } | |
255 | ||
256 | /* | |
257 | * No read/write mix, if stat isn't valid | |
258 | */ | |
259 | if (!stat_sample_valid(stat)) { | |
260 | /* | |
261 | * If we had writes in this stat window and the window is | |
262 | * current, we're only doing writes. If a task recently | |
263 | * waited or still has writes in flights, consider us doing | |
264 | * just writes as well. | |
265 | */ | |
34dbad5d OS |
266 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
267 | wbt_inflight(rwb)) | |
e34cbd30 JA |
268 | return LAT_UNKNOWN_WRITES; |
269 | return LAT_UNKNOWN; | |
270 | } | |
271 | ||
272 | /* | |
273 | * If the 'min' latency exceeds our target, step down. | |
274 | */ | |
fa2e39cb OS |
275 | if (stat[READ].min > rwb->min_lat_nsec) { |
276 | trace_wbt_lat(bdi, stat[READ].min); | |
d8a0cbfd | 277 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
278 | return LAT_EXCEEDED; |
279 | } | |
280 | ||
a7905043 | 281 | if (rqd->scale_step) |
d8a0cbfd | 282 | trace_wbt_stat(bdi, stat); |
e34cbd30 JA |
283 | |
284 | return LAT_OK; | |
285 | } | |
286 | ||
e34cbd30 JA |
287 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
288 | { | |
a7905043 JB |
289 | struct backing_dev_info *bdi = rwb->rqos.q->backing_dev_info; |
290 | struct rq_depth *rqd = &rwb->rq_depth; | |
d8a0cbfd | 291 | |
a7905043 JB |
292 | trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec, |
293 | rwb->wb_background, rwb->wb_normal, rqd->max_depth); | |
e34cbd30 JA |
294 | } |
295 | ||
a7905043 | 296 | static void calc_wb_limits(struct rq_wb *rwb) |
e34cbd30 | 297 | { |
a7905043 JB |
298 | if (rwb->min_lat_nsec == 0) { |
299 | rwb->wb_normal = rwb->wb_background = 0; | |
300 | } else if (rwb->rq_depth.max_depth <= 2) { | |
301 | rwb->wb_normal = rwb->rq_depth.max_depth; | |
302 | rwb->wb_background = 1; | |
303 | } else { | |
304 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; | |
305 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; | |
306 | } | |
307 | } | |
e34cbd30 | 308 | |
a7905043 JB |
309 | static void scale_up(struct rq_wb *rwb) |
310 | { | |
b84477d3 HS |
311 | if (!rq_depth_scale_up(&rwb->rq_depth)) |
312 | return; | |
a7905043 | 313 | calc_wb_limits(rwb); |
e34cbd30 | 314 | rwb->unknown_cnt = 0; |
5e65a203 | 315 | rwb_wake_all(rwb); |
3a89c25d | 316 | rwb_trace_step(rwb, tracepoint_string("scale up")); |
e34cbd30 JA |
317 | } |
318 | ||
e34cbd30 JA |
319 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
320 | { | |
b84477d3 HS |
321 | if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle)) |
322 | return; | |
e34cbd30 | 323 | calc_wb_limits(rwb); |
a7905043 | 324 | rwb->unknown_cnt = 0; |
3a89c25d | 325 | rwb_trace_step(rwb, tracepoint_string("scale down")); |
e34cbd30 JA |
326 | } |
327 | ||
328 | static void rwb_arm_timer(struct rq_wb *rwb) | |
329 | { | |
a7905043 JB |
330 | struct rq_depth *rqd = &rwb->rq_depth; |
331 | ||
332 | if (rqd->scale_step > 0) { | |
e34cbd30 JA |
333 | /* |
334 | * We should speed this up, using some variant of a fast | |
335 | * integer inverse square root calculation. Since we only do | |
336 | * this for every window expiration, it's not a huge deal, | |
337 | * though. | |
338 | */ | |
339 | rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, | |
a7905043 | 340 | int_sqrt((rqd->scale_step + 1) << 8)); |
e34cbd30 JA |
341 | } else { |
342 | /* | |
343 | * For step < 0, we don't want to increase/decrease the | |
344 | * window size. | |
345 | */ | |
346 | rwb->cur_win_nsec = rwb->win_nsec; | |
347 | } | |
348 | ||
34dbad5d | 349 | blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec); |
e34cbd30 JA |
350 | } |
351 | ||
34dbad5d | 352 | static void wb_timer_fn(struct blk_stat_callback *cb) |
e34cbd30 | 353 | { |
34dbad5d | 354 | struct rq_wb *rwb = cb->data; |
a7905043 | 355 | struct rq_depth *rqd = &rwb->rq_depth; |
e34cbd30 JA |
356 | unsigned int inflight = wbt_inflight(rwb); |
357 | int status; | |
358 | ||
34dbad5d | 359 | status = latency_exceeded(rwb, cb->stat); |
e34cbd30 | 360 | |
a7905043 | 361 | trace_wbt_timer(rwb->rqos.q->backing_dev_info, status, rqd->scale_step, |
d8a0cbfd | 362 | inflight); |
e34cbd30 JA |
363 | |
364 | /* | |
365 | * If we exceeded the latency target, step down. If we did not, | |
366 | * step one level up. If we don't know enough to say either exceeded | |
367 | * or ok, then don't do anything. | |
368 | */ | |
369 | switch (status) { | |
370 | case LAT_EXCEEDED: | |
371 | scale_down(rwb, true); | |
372 | break; | |
373 | case LAT_OK: | |
374 | scale_up(rwb); | |
375 | break; | |
376 | case LAT_UNKNOWN_WRITES: | |
377 | /* | |
378 | * We started a the center step, but don't have a valid | |
379 | * read/write sample, but we do have writes going on. | |
380 | * Allow step to go negative, to increase write perf. | |
381 | */ | |
382 | scale_up(rwb); | |
383 | break; | |
384 | case LAT_UNKNOWN: | |
385 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) | |
386 | break; | |
387 | /* | |
388 | * We get here when previously scaled reduced depth, and we | |
389 | * currently don't have a valid read/write sample. For that | |
390 | * case, slowly return to center state (step == 0). | |
391 | */ | |
a7905043 | 392 | if (rqd->scale_step > 0) |
e34cbd30 | 393 | scale_up(rwb); |
a7905043 | 394 | else if (rqd->scale_step < 0) |
e34cbd30 JA |
395 | scale_down(rwb, false); |
396 | break; | |
397 | default: | |
398 | break; | |
399 | } | |
400 | ||
401 | /* | |
402 | * Re-arm timer, if we have IO in flight | |
403 | */ | |
a7905043 | 404 | if (rqd->scale_step || inflight) |
e34cbd30 JA |
405 | rwb_arm_timer(rwb); |
406 | } | |
407 | ||
a7905043 | 408 | static void __wbt_update_limits(struct rq_wb *rwb) |
e34cbd30 | 409 | { |
a7905043 JB |
410 | struct rq_depth *rqd = &rwb->rq_depth; |
411 | ||
412 | rqd->scale_step = 0; | |
413 | rqd->scaled_max = false; | |
414 | ||
415 | rq_depth_calc_max_depth(rqd); | |
e34cbd30 JA |
416 | calc_wb_limits(rwb); |
417 | ||
418 | rwb_wake_all(rwb); | |
419 | } | |
420 | ||
a7905043 JB |
421 | u64 wbt_get_min_lat(struct request_queue *q) |
422 | { | |
423 | struct rq_qos *rqos = wbt_rq_qos(q); | |
424 | if (!rqos) | |
425 | return 0; | |
426 | return RQWB(rqos)->min_lat_nsec; | |
427 | } | |
428 | ||
429 | void wbt_set_min_lat(struct request_queue *q, u64 val) | |
430 | { | |
431 | struct rq_qos *rqos = wbt_rq_qos(q); | |
432 | if (!rqos) | |
433 | return; | |
434 | RQWB(rqos)->min_lat_nsec = val; | |
435 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; | |
436 | __wbt_update_limits(RQWB(rqos)); | |
437 | } | |
438 | ||
439 | ||
e34cbd30 JA |
440 | static bool close_io(struct rq_wb *rwb) |
441 | { | |
442 | const unsigned long now = jiffies; | |
443 | ||
444 | return time_before(now, rwb->last_issue + HZ / 10) || | |
445 | time_before(now, rwb->last_comp + HZ / 10); | |
446 | } | |
447 | ||
448 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) | |
449 | ||
450 | static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) | |
451 | { | |
452 | unsigned int limit; | |
453 | ||
ffa358dc JA |
454 | /* |
455 | * If we got disabled, just return UINT_MAX. This ensures that | |
456 | * we'll properly inc a new IO, and dec+wakeup at the end. | |
457 | */ | |
458 | if (!rwb_enabled(rwb)) | |
459 | return UINT_MAX; | |
460 | ||
782f5697 JA |
461 | if ((rw & REQ_OP_MASK) == REQ_OP_DISCARD) |
462 | return rwb->wb_background; | |
463 | ||
e34cbd30 JA |
464 | /* |
465 | * At this point we know it's a buffered write. If this is | |
3dfbdc44 | 466 | * kswapd trying to free memory, or REQ_SYNC is set, then |
e34cbd30 JA |
467 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
468 | * that. If the write is marked as a background write, then use | |
469 | * the idle limit, or go to normal if we haven't had competing | |
470 | * IO for a bit. | |
471 | */ | |
472 | if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) | |
a7905043 | 473 | limit = rwb->rq_depth.max_depth; |
e34cbd30 JA |
474 | else if ((rw & REQ_BACKGROUND) || close_io(rwb)) { |
475 | /* | |
476 | * If less than 100ms since we completed unrelated IO, | |
477 | * limit us to half the depth for background writeback. | |
478 | */ | |
479 | limit = rwb->wb_background; | |
480 | } else | |
481 | limit = rwb->wb_normal; | |
482 | ||
483 | return limit; | |
484 | } | |
485 | ||
38cfb5a4 | 486 | struct wbt_wait_data { |
38cfb5a4 | 487 | struct rq_wb *rwb; |
b6c7b58f | 488 | enum wbt_flags wb_acct; |
38cfb5a4 | 489 | unsigned long rw; |
38cfb5a4 JA |
490 | }; |
491 | ||
b6c7b58f | 492 | static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data) |
38cfb5a4 | 493 | { |
b6c7b58f JB |
494 | struct wbt_wait_data *data = private_data; |
495 | return rq_wait_inc_below(rqw, get_limit(data->rwb, data->rw)); | |
496 | } | |
38cfb5a4 | 497 | |
b6c7b58f JB |
498 | static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data) |
499 | { | |
500 | struct wbt_wait_data *data = private_data; | |
501 | wbt_rqw_done(data->rwb, rqw, data->wb_acct); | |
38cfb5a4 JA |
502 | } |
503 | ||
e34cbd30 JA |
504 | /* |
505 | * Block if we will exceed our limit, or if we are currently waiting for | |
506 | * the timer to kick off queuing again. | |
507 | */ | |
8bea6090 | 508 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
d5337560 | 509 | unsigned long rw) |
e34cbd30 | 510 | { |
8bea6090 | 511 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
38cfb5a4 | 512 | struct wbt_wait_data data = { |
38cfb5a4 | 513 | .rwb = rwb, |
b6c7b58f | 514 | .wb_acct = wb_acct, |
38cfb5a4 JA |
515 | .rw = rw, |
516 | }; | |
e34cbd30 | 517 | |
b6c7b58f | 518 | rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb); |
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 | ||
c125311d JA |
543 | if (!rwb_enabled(rwb)) |
544 | return 0; | |
545 | ||
c1c80384 JB |
546 | if (bio_op(bio) == REQ_OP_READ) { |
547 | flags = WBT_READ; | |
548 | } else if (wbt_should_throttle(rwb, bio)) { | |
549 | if (current_is_kswapd()) | |
550 | flags |= WBT_KSWAPD; | |
551 | if (bio_op(bio) == REQ_OP_DISCARD) | |
552 | flags |= WBT_DISCARD; | |
553 | flags |= WBT_TRACKED; | |
554 | } | |
555 | return flags; | |
556 | } | |
557 | ||
558 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) | |
559 | { | |
560 | struct rq_wb *rwb = RQWB(rqos); | |
561 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); | |
562 | __wbt_done(rqos, flags); | |
563 | } | |
564 | ||
e34cbd30 JA |
565 | /* |
566 | * Returns true if the IO request should be accounted, false if not. | |
567 | * May sleep, if we have exceeded the writeback limits. Caller can pass | |
568 | * in an irq held spinlock, if it holds one when calling this function. | |
569 | * If we do sleep, we'll release and re-grab it. | |
570 | */ | |
d5337560 | 571 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio) |
e34cbd30 | 572 | { |
a7905043 | 573 | struct rq_wb *rwb = RQWB(rqos); |
c1c80384 | 574 | enum wbt_flags flags; |
e34cbd30 | 575 | |
c1c80384 | 576 | flags = bio_to_wbt_flags(rwb, bio); |
df60f6e8 | 577 | if (!(flags & WBT_TRACKED)) { |
c1c80384 | 578 | if (flags & WBT_READ) |
e34cbd30 | 579 | wb_timestamp(rwb, &rwb->last_issue); |
c1c80384 | 580 | return; |
e34cbd30 JA |
581 | } |
582 | ||
d5337560 | 583 | __wbt_wait(rwb, flags, bio->bi_opf); |
e34cbd30 | 584 | |
34dbad5d | 585 | if (!blk_stat_is_active(rwb->cb)) |
e34cbd30 | 586 | rwb_arm_timer(rwb); |
c1c80384 | 587 | } |
e34cbd30 | 588 | |
c1c80384 JB |
589 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
590 | { | |
591 | struct rq_wb *rwb = RQWB(rqos); | |
592 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); | |
e34cbd30 JA |
593 | } |
594 | ||
c83f536a | 595 | static void wbt_issue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 596 | { |
a7905043 JB |
597 | struct rq_wb *rwb = RQWB(rqos); |
598 | ||
e34cbd30 JA |
599 | if (!rwb_enabled(rwb)) |
600 | return; | |
601 | ||
602 | /* | |
a8a45941 OS |
603 | * Track sync issue, in case it takes a long time to complete. Allows us |
604 | * to react quicker, if a sync IO takes a long time to complete. Note | |
605 | * that this is just a hint. The request can go away when it completes, | |
606 | * so it's important we never dereference it. We only use the address to | |
607 | * compare with, which is why we store the sync_issue time locally. | |
e34cbd30 | 608 | */ |
a8a45941 OS |
609 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
610 | rwb->sync_cookie = rq; | |
544ccc8d | 611 | rwb->sync_issue = rq->io_start_time_ns; |
e34cbd30 JA |
612 | } |
613 | } | |
614 | ||
c83f536a | 615 | static void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
e34cbd30 | 616 | { |
a7905043 | 617 | struct rq_wb *rwb = RQWB(rqos); |
e34cbd30 JA |
618 | if (!rwb_enabled(rwb)) |
619 | return; | |
a8a45941 | 620 | if (rq == rwb->sync_cookie) { |
e34cbd30 JA |
621 | rwb->sync_issue = 0; |
622 | rwb->sync_cookie = NULL; | |
623 | } | |
624 | } | |
625 | ||
a7905043 | 626 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
e34cbd30 | 627 | { |
a7905043 JB |
628 | struct rq_qos *rqos = wbt_rq_qos(q); |
629 | if (rqos) | |
630 | RQWB(rqos)->wc = write_cache_on; | |
e34cbd30 | 631 | } |
e34cbd30 | 632 | |
8330cdb0 JK |
633 | /* |
634 | * Enable wbt if defaults are configured that way | |
635 | */ | |
636 | void wbt_enable_default(struct request_queue *q) | |
637 | { | |
a7905043 | 638 | struct rq_qos *rqos = wbt_rq_qos(q); |
8330cdb0 | 639 | /* Throttling already enabled? */ |
a7905043 | 640 | if (rqos) |
8330cdb0 JK |
641 | return; |
642 | ||
643 | /* Queue not registered? Maybe shutting down... */ | |
58c898ba | 644 | if (!blk_queue_registered(q)) |
8330cdb0 JK |
645 | return; |
646 | ||
344e9ffc | 647 | if (queue_is_mq(q) && IS_ENABLED(CONFIG_BLK_WBT_MQ)) |
8330cdb0 JK |
648 | wbt_init(q); |
649 | } | |
650 | EXPORT_SYMBOL_GPL(wbt_enable_default); | |
651 | ||
80e091d1 JA |
652 | u64 wbt_default_latency_nsec(struct request_queue *q) |
653 | { | |
654 | /* | |
655 | * We default to 2msec for non-rotational storage, and 75msec | |
656 | * for rotational storage. | |
657 | */ | |
658 | if (blk_queue_nonrot(q)) | |
659 | return 2000000ULL; | |
660 | else | |
661 | return 75000000ULL; | |
662 | } | |
663 | ||
99c749a4 JA |
664 | static int wbt_data_dir(const struct request *rq) |
665 | { | |
5235553d JA |
666 | const int op = req_op(rq); |
667 | ||
668 | if (op == REQ_OP_READ) | |
669 | return READ; | |
825843b0 | 670 | else if (op_is_write(op)) |
5235553d JA |
671 | return WRITE; |
672 | ||
673 | /* don't account */ | |
674 | return -1; | |
99c749a4 JA |
675 | } |
676 | ||
9677a3e0 TH |
677 | static void wbt_queue_depth_changed(struct rq_qos *rqos) |
678 | { | |
679 | RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q); | |
680 | __wbt_update_limits(RQWB(rqos)); | |
681 | } | |
682 | ||
a7905043 JB |
683 | static void wbt_exit(struct rq_qos *rqos) |
684 | { | |
685 | struct rq_wb *rwb = RQWB(rqos); | |
686 | struct request_queue *q = rqos->q; | |
687 | ||
688 | blk_stat_remove_callback(q, rwb->cb); | |
689 | blk_stat_free_callback(rwb->cb); | |
690 | kfree(rwb); | |
691 | } | |
692 | ||
693 | /* | |
694 | * Disable wbt, if enabled by default. | |
695 | */ | |
696 | void wbt_disable_default(struct request_queue *q) | |
697 | { | |
698 | struct rq_qos *rqos = wbt_rq_qos(q); | |
699 | struct rq_wb *rwb; | |
700 | if (!rqos) | |
701 | return; | |
702 | rwb = RQWB(rqos); | |
544fbd16 ML |
703 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) { |
704 | blk_stat_deactivate(rwb->cb); | |
a7905043 | 705 | rwb->wb_normal = 0; |
544fbd16 | 706 | } |
a7905043 | 707 | } |
e815f404 | 708 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
a7905043 | 709 | |
d19afebc ML |
710 | #ifdef CONFIG_BLK_DEBUG_FS |
711 | static int wbt_curr_win_nsec_show(void *data, struct seq_file *m) | |
712 | { | |
713 | struct rq_qos *rqos = data; | |
714 | struct rq_wb *rwb = RQWB(rqos); | |
715 | ||
716 | seq_printf(m, "%llu\n", rwb->cur_win_nsec); | |
717 | return 0; | |
718 | } | |
719 | ||
720 | static int wbt_enabled_show(void *data, struct seq_file *m) | |
721 | { | |
722 | struct rq_qos *rqos = data; | |
723 | struct rq_wb *rwb = RQWB(rqos); | |
724 | ||
725 | seq_printf(m, "%d\n", rwb->enable_state); | |
726 | return 0; | |
727 | } | |
728 | ||
729 | static int wbt_id_show(void *data, struct seq_file *m) | |
730 | { | |
731 | struct rq_qos *rqos = data; | |
732 | ||
733 | seq_printf(m, "%u\n", rqos->id); | |
734 | return 0; | |
735 | } | |
736 | ||
737 | static int wbt_inflight_show(void *data, struct seq_file *m) | |
738 | { | |
739 | struct rq_qos *rqos = data; | |
740 | struct rq_wb *rwb = RQWB(rqos); | |
741 | int i; | |
742 | ||
743 | for (i = 0; i < WBT_NUM_RWQ; i++) | |
744 | seq_printf(m, "%d: inflight %d\n", i, | |
745 | atomic_read(&rwb->rq_wait[i].inflight)); | |
746 | return 0; | |
747 | } | |
748 | ||
749 | static int wbt_min_lat_nsec_show(void *data, struct seq_file *m) | |
750 | { | |
751 | struct rq_qos *rqos = data; | |
752 | struct rq_wb *rwb = RQWB(rqos); | |
753 | ||
754 | seq_printf(m, "%lu\n", rwb->min_lat_nsec); | |
755 | return 0; | |
756 | } | |
757 | ||
758 | static int wbt_unknown_cnt_show(void *data, struct seq_file *m) | |
759 | { | |
760 | struct rq_qos *rqos = data; | |
761 | struct rq_wb *rwb = RQWB(rqos); | |
762 | ||
763 | seq_printf(m, "%u\n", rwb->unknown_cnt); | |
764 | return 0; | |
765 | } | |
766 | ||
767 | static int wbt_normal_show(void *data, struct seq_file *m) | |
768 | { | |
769 | struct rq_qos *rqos = data; | |
770 | struct rq_wb *rwb = RQWB(rqos); | |
771 | ||
772 | seq_printf(m, "%u\n", rwb->wb_normal); | |
773 | return 0; | |
774 | } | |
775 | ||
776 | static int wbt_background_show(void *data, struct seq_file *m) | |
777 | { | |
778 | struct rq_qos *rqos = data; | |
779 | struct rq_wb *rwb = RQWB(rqos); | |
780 | ||
781 | seq_printf(m, "%u\n", rwb->wb_background); | |
782 | return 0; | |
783 | } | |
784 | ||
785 | static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = { | |
786 | {"curr_win_nsec", 0400, wbt_curr_win_nsec_show}, | |
787 | {"enabled", 0400, wbt_enabled_show}, | |
788 | {"id", 0400, wbt_id_show}, | |
789 | {"inflight", 0400, wbt_inflight_show}, | |
790 | {"min_lat_nsec", 0400, wbt_min_lat_nsec_show}, | |
791 | {"unknown_cnt", 0400, wbt_unknown_cnt_show}, | |
792 | {"wb_normal", 0400, wbt_normal_show}, | |
793 | {"wb_background", 0400, wbt_background_show}, | |
794 | {}, | |
795 | }; | |
796 | #endif | |
797 | ||
a7905043 JB |
798 | static struct rq_qos_ops wbt_rqos_ops = { |
799 | .throttle = wbt_wait, | |
800 | .issue = wbt_issue, | |
c1c80384 | 801 | .track = wbt_track, |
a7905043 JB |
802 | .requeue = wbt_requeue, |
803 | .done = wbt_done, | |
c1c80384 | 804 | .cleanup = wbt_cleanup, |
9677a3e0 | 805 | .queue_depth_changed = wbt_queue_depth_changed, |
a7905043 | 806 | .exit = wbt_exit, |
d19afebc ML |
807 | #ifdef CONFIG_BLK_DEBUG_FS |
808 | .debugfs_attrs = wbt_debugfs_attrs, | |
809 | #endif | |
a7905043 JB |
810 | }; |
811 | ||
8054b89f | 812 | int wbt_init(struct request_queue *q) |
e34cbd30 JA |
813 | { |
814 | struct rq_wb *rwb; | |
815 | int i; | |
816 | ||
e34cbd30 JA |
817 | rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); |
818 | if (!rwb) | |
819 | return -ENOMEM; | |
820 | ||
99c749a4 | 821 | rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb); |
34dbad5d OS |
822 | if (!rwb->cb) { |
823 | kfree(rwb); | |
824 | return -ENOMEM; | |
825 | } | |
826 | ||
a7905043 JB |
827 | for (i = 0; i < WBT_NUM_RWQ; i++) |
828 | rq_wait_init(&rwb->rq_wait[i]); | |
e34cbd30 | 829 | |
a7905043 JB |
830 | rwb->rqos.id = RQ_QOS_WBT; |
831 | rwb->rqos.ops = &wbt_rqos_ops; | |
832 | rwb->rqos.q = q; | |
e34cbd30 | 833 | rwb->last_comp = rwb->last_issue = jiffies; |
e34cbd30 | 834 | rwb->win_nsec = RWB_WINDOW_NSEC; |
d62118b6 | 835 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
a7905043 JB |
836 | rwb->wc = 1; |
837 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; | |
838 | __wbt_update_limits(rwb); | |
e34cbd30 JA |
839 | |
840 | /* | |
34dbad5d | 841 | * Assign rwb and add the stats callback. |
e34cbd30 | 842 | */ |
a7905043 | 843 | rq_qos_add(q, &rwb->rqos); |
34dbad5d | 844 | blk_stat_add_callback(q, rwb->cb); |
e34cbd30 | 845 | |
80e091d1 | 846 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
e34cbd30 | 847 | |
9677a3e0 | 848 | wbt_queue_depth_changed(&rwb->rqos); |
a7905043 | 849 | wbt_set_write_cache(q, test_bit(QUEUE_FLAG_WC, &q->queue_flags)); |
e34cbd30 JA |
850 | |
851 | return 0; | |
852 | } |