Commit | Line | Data |
---|---|---|
8c16567d | 1 | // SPDX-License-Identifier: GPL-2.0 |
86db1e29 | 2 | /* |
3140c3cf | 3 | * Functions to sequence PREFLUSH and FUA writes. |
ae1b1539 TH |
4 | * |
5 | * Copyright (C) 2011 Max Planck Institute for Gravitational Physics | |
6 | * Copyright (C) 2011 Tejun Heo <tj@kernel.org> | |
7 | * | |
3140c3cf | 8 | * REQ_{PREFLUSH|FUA} requests are decomposed to sequences consisted of three |
ae1b1539 TH |
9 | * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request |
10 | * properties and hardware capability. | |
11 | * | |
28a8f0d3 MC |
12 | * If a request doesn't have data, only REQ_PREFLUSH makes sense, which |
13 | * indicates a simple flush request. If there is data, REQ_PREFLUSH indicates | |
ae1b1539 TH |
14 | * that the device cache should be flushed before the data is executed, and |
15 | * REQ_FUA means that the data must be on non-volatile media on request | |
16 | * completion. | |
17 | * | |
3140c3cf OS |
18 | * If the device doesn't have writeback cache, PREFLUSH and FUA don't make any |
19 | * difference. The requests are either completed immediately if there's no data | |
20 | * or executed as normal requests otherwise. | |
ae1b1539 | 21 | * |
28a8f0d3 | 22 | * If the device has writeback cache and supports FUA, REQ_PREFLUSH is |
ae1b1539 TH |
23 | * translated to PREFLUSH but REQ_FUA is passed down directly with DATA. |
24 | * | |
28a8f0d3 MC |
25 | * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH |
26 | * is translated to PREFLUSH and REQ_FUA to POSTFLUSH. | |
ae1b1539 TH |
27 | * |
28 | * The actual execution of flush is double buffered. Whenever a request | |
29 | * needs to execute PRE or POSTFLUSH, it queues at | |
7c94e1c1 | 30 | * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a |
3a5e02ce | 31 | * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush |
ae1b1539 | 32 | * completes, all the requests which were pending are proceeded to the next |
3140c3cf | 33 | * step. This allows arbitrary merging of different types of PREFLUSH/FUA |
ae1b1539 TH |
34 | * requests. |
35 | * | |
36 | * Currently, the following conditions are used to determine when to issue | |
37 | * flush. | |
38 | * | |
39 | * C1. At any given time, only one flush shall be in progress. This makes | |
40 | * double buffering sufficient. | |
41 | * | |
42 | * C2. Flush is deferred if any request is executing DATA of its sequence. | |
43 | * This avoids issuing separate POSTFLUSHes for requests which shared | |
44 | * PREFLUSH. | |
45 | * | |
46 | * C3. The second condition is ignored if there is a request which has | |
47 | * waited longer than FLUSH_PENDING_TIMEOUT. This is to avoid | |
48 | * starvation in the unlikely case where there are continuous stream of | |
3140c3cf | 49 | * FUA (without PREFLUSH) requests. |
ae1b1539 TH |
50 | * |
51 | * For devices which support FUA, it isn't clear whether C2 (and thus C3) | |
52 | * is beneficial. | |
53 | * | |
3140c3cf | 54 | * Note that a sequenced PREFLUSH/FUA request with DATA is completed twice. |
ae1b1539 TH |
55 | * Once while executing DATA and again after the whole sequence is |
56 | * complete. The first completion updates the contained bio but doesn't | |
57 | * finish it so that the bio submitter is notified only after the whole | |
e8064021 | 58 | * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in |
ae1b1539 TH |
59 | * req_bio_endio(). |
60 | * | |
3140c3cf | 61 | * The above peculiarity requires that each PREFLUSH/FUA request has only one |
ae1b1539 TH |
62 | * bio attached to it, which is guaranteed as they aren't allowed to be |
63 | * merged in the usual way. | |
86db1e29 | 64 | */ |
ae1b1539 | 65 | |
86db1e29 JA |
66 | #include <linux/kernel.h> |
67 | #include <linux/module.h> | |
68 | #include <linux/bio.h> | |
69 | #include <linux/blkdev.h> | |
5a0e3ad6 | 70 | #include <linux/gfp.h> |
82d981d4 | 71 | #include <linux/part_stat.h> |
86db1e29 JA |
72 | |
73 | #include "blk.h" | |
320ae51f | 74 | #include "blk-mq.h" |
bd166ef1 | 75 | #include "blk-mq-sched.h" |
86db1e29 | 76 | |
3140c3cf | 77 | /* PREFLUSH/FUA sequences */ |
4fed947c | 78 | enum { |
ae1b1539 TH |
79 | REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */ |
80 | REQ_FSEQ_DATA = (1 << 1), /* data write in progress */ | |
81 | REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */ | |
82 | REQ_FSEQ_DONE = (1 << 3), | |
83 | ||
84 | REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA | | |
85 | REQ_FSEQ_POSTFLUSH, | |
86 | ||
87 | /* | |
88 | * If flush has been pending longer than the following timeout, | |
89 | * it's issued even if flush_data requests are still in flight. | |
90 | */ | |
91 | FLUSH_PENDING_TIMEOUT = 5 * HZ, | |
4fed947c TH |
92 | }; |
93 | ||
404b8f5a | 94 | static void blk_kick_flush(struct request_queue *q, |
16458cf3 | 95 | struct blk_flush_queue *fq, blk_opf_t flags); |
28e7d184 | 96 | |
0281ed3c CH |
97 | static inline struct blk_flush_queue * |
98 | blk_get_flush_queue(struct request_queue *q, struct blk_mq_ctx *ctx) | |
99 | { | |
100 | return blk_mq_map_queue(q, REQ_OP_FLUSH, ctx)->fq; | |
101 | } | |
102 | ||
c888a8f9 | 103 | static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq) |
86db1e29 | 104 | { |
ae1b1539 | 105 | unsigned int policy = 0; |
86db1e29 | 106 | |
fa1bf42f JM |
107 | if (blk_rq_sectors(rq)) |
108 | policy |= REQ_FSEQ_DATA; | |
109 | ||
c888a8f9 | 110 | if (fflags & (1UL << QUEUE_FLAG_WC)) { |
28a8f0d3 | 111 | if (rq->cmd_flags & REQ_PREFLUSH) |
ae1b1539 | 112 | policy |= REQ_FSEQ_PREFLUSH; |
c888a8f9 JA |
113 | if (!(fflags & (1UL << QUEUE_FLAG_FUA)) && |
114 | (rq->cmd_flags & REQ_FUA)) | |
ae1b1539 | 115 | policy |= REQ_FSEQ_POSTFLUSH; |
28e7d184 | 116 | } |
ae1b1539 | 117 | return policy; |
86db1e29 JA |
118 | } |
119 | ||
ae1b1539 | 120 | static unsigned int blk_flush_cur_seq(struct request *rq) |
47f70d5a | 121 | { |
ae1b1539 TH |
122 | return 1 << ffz(rq->flush.seq); |
123 | } | |
47f70d5a | 124 | |
ae1b1539 TH |
125 | static void blk_flush_restore_request(struct request *rq) |
126 | { | |
47f70d5a | 127 | /* |
ae1b1539 TH |
128 | * After flush data completion, @rq->bio is %NULL but we need to |
129 | * complete the bio again. @rq->biotail is guaranteed to equal the | |
130 | * original @rq->bio. Restore it. | |
47f70d5a | 131 | */ |
ae1b1539 TH |
132 | rq->bio = rq->biotail; |
133 | ||
134 | /* make @rq a normal request */ | |
e8064021 | 135 | rq->rq_flags &= ~RQF_FLUSH_SEQ; |
4853abaa | 136 | rq->end_io = rq->flush.saved_end_io; |
320ae51f JA |
137 | } |
138 | ||
b6866318 KK |
139 | static void blk_account_io_flush(struct request *rq) |
140 | { | |
f3fa33ac | 141 | struct block_device *part = rq->q->disk->part0; |
b6866318 KK |
142 | |
143 | part_stat_lock(); | |
144 | part_stat_inc(part, ios[STAT_FLUSH]); | |
145 | part_stat_add(part, nsecs[STAT_FLUSH], | |
146 | ktime_get_ns() - rq->start_time_ns); | |
147 | part_stat_unlock(); | |
148 | } | |
149 | ||
ae1b1539 TH |
150 | /** |
151 | * blk_flush_complete_seq - complete flush sequence | |
3140c3cf | 152 | * @rq: PREFLUSH/FUA request being sequenced |
0bae352d | 153 | * @fq: flush queue |
ae1b1539 TH |
154 | * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero) |
155 | * @error: whether an error occurred | |
156 | * | |
157 | * @rq just completed @seq part of its flush sequence, record the | |
158 | * completion and trigger the next step. | |
159 | * | |
160 | * CONTEXT: | |
9809b4ee | 161 | * spin_lock_irq(fq->mq_flush_lock) |
ae1b1539 | 162 | */ |
404b8f5a | 163 | static void blk_flush_complete_seq(struct request *rq, |
0bae352d | 164 | struct blk_flush_queue *fq, |
2a842aca | 165 | unsigned int seq, blk_status_t error) |
86db1e29 | 166 | { |
ae1b1539 | 167 | struct request_queue *q = rq->q; |
7c94e1c1 | 168 | struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx]; |
16458cf3 | 169 | blk_opf_t cmd_flags; |
ae1b1539 TH |
170 | |
171 | BUG_ON(rq->flush.seq & seq); | |
172 | rq->flush.seq |= seq; | |
190b02ed | 173 | cmd_flags = rq->cmd_flags; |
ae1b1539 TH |
174 | |
175 | if (likely(!error)) | |
176 | seq = blk_flush_cur_seq(rq); | |
177 | else | |
178 | seq = REQ_FSEQ_DONE; | |
179 | ||
180 | switch (seq) { | |
181 | case REQ_FSEQ_PREFLUSH: | |
182 | case REQ_FSEQ_POSTFLUSH: | |
183 | /* queue for flush */ | |
184 | if (list_empty(pending)) | |
7c94e1c1 | 185 | fq->flush_pending_since = jiffies; |
ae1b1539 TH |
186 | list_move_tail(&rq->flush.list, pending); |
187 | break; | |
188 | ||
189 | case REQ_FSEQ_DATA: | |
7c94e1c1 | 190 | list_move_tail(&rq->flush.list, &fq->flush_data_in_flight); |
b12e5c6c | 191 | blk_mq_add_to_requeue_list(rq, BLK_MQ_INSERT_AT_HEAD); |
214a4418 | 192 | blk_mq_kick_requeue_list(q); |
ae1b1539 TH |
193 | break; |
194 | ||
195 | case REQ_FSEQ_DONE: | |
196 | /* | |
b6866318 | 197 | * @rq was previously adjusted by blk_insert_flush() for |
ae1b1539 TH |
198 | * flush sequencing and may already have gone through the |
199 | * flush data request completion path. Restore @rq for | |
200 | * normal completion and end it. | |
201 | */ | |
ae1b1539 TH |
202 | list_del_init(&rq->flush.list); |
203 | blk_flush_restore_request(rq); | |
7e992f84 | 204 | blk_mq_end_request(rq, error); |
ae1b1539 TH |
205 | break; |
206 | ||
207 | default: | |
208 | BUG(); | |
209 | } | |
210 | ||
404b8f5a | 211 | blk_kick_flush(q, fq, cmd_flags); |
86db1e29 JA |
212 | } |
213 | ||
de671d61 JA |
214 | static enum rq_end_io_ret flush_end_io(struct request *flush_rq, |
215 | blk_status_t error) | |
86db1e29 | 216 | { |
ae1b1539 | 217 | struct request_queue *q = flush_rq->q; |
320ae51f | 218 | struct list_head *running; |
ae1b1539 | 219 | struct request *rq, *n; |
320ae51f | 220 | unsigned long flags = 0; |
e97c293c | 221 | struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx); |
ae1b1539 | 222 | |
7e992f84 JA |
223 | /* release the tag's ownership to the req cloned from */ |
224 | spin_lock_irqsave(&fq->mq_flush_lock, flags); | |
8d699663 | 225 | |
0a467d0f | 226 | if (!req_ref_put_and_test(flush_rq)) { |
8d699663 YY |
227 | fq->rq_status = error; |
228 | spin_unlock_irqrestore(&fq->mq_flush_lock, flags); | |
de671d61 | 229 | return RQ_END_IO_NONE; |
8d699663 YY |
230 | } |
231 | ||
84da7acc | 232 | blk_account_io_flush(flush_rq); |
9f16a667 ML |
233 | /* |
234 | * Flush request has to be marked as IDLE when it is really ended | |
235 | * because its .end_io() is called from timeout code path too for | |
236 | * avoiding use-after-free. | |
237 | */ | |
238 | WRITE_ONCE(flush_rq->state, MQ_RQ_IDLE); | |
8a751893 | 239 | if (fq->rq_status != BLK_STS_OK) { |
8d699663 | 240 | error = fq->rq_status; |
8a751893 YB |
241 | fq->rq_status = BLK_STS_OK; |
242 | } | |
8d699663 | 243 | |
4e2f62e5 | 244 | if (!q->elevator) { |
568f2700 | 245 | flush_rq->tag = BLK_MQ_NO_TAG; |
4e2f62e5 JA |
246 | } else { |
247 | blk_mq_put_driver_tag(flush_rq); | |
568f2700 | 248 | flush_rq->internal_tag = BLK_MQ_NO_TAG; |
4e2f62e5 | 249 | } |
18741986 | 250 | |
7c94e1c1 ML |
251 | running = &fq->flush_queue[fq->flush_running_idx]; |
252 | BUG_ON(fq->flush_pending_idx == fq->flush_running_idx); | |
ae1b1539 TH |
253 | |
254 | /* account completion of the flush request */ | |
7c94e1c1 | 255 | fq->flush_running_idx ^= 1; |
320ae51f | 256 | |
ae1b1539 TH |
257 | /* and push the waiting requests to the next stage */ |
258 | list_for_each_entry_safe(rq, n, running, flush.list) { | |
259 | unsigned int seq = blk_flush_cur_seq(rq); | |
260 | ||
261 | BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH); | |
404b8f5a | 262 | blk_flush_complete_seq(rq, fq, seq, error); |
ae1b1539 TH |
263 | } |
264 | ||
7e992f84 | 265 | spin_unlock_irqrestore(&fq->mq_flush_lock, flags); |
de671d61 | 266 | return RQ_END_IO_NONE; |
320ae51f JA |
267 | } |
268 | ||
a9ed27a7 ML |
269 | bool is_flush_rq(struct request *rq) |
270 | { | |
271 | return rq->end_io == flush_end_io; | |
272 | } | |
273 | ||
ae1b1539 TH |
274 | /** |
275 | * blk_kick_flush - consider issuing flush request | |
276 | * @q: request_queue being kicked | |
0bae352d | 277 | * @fq: flush queue |
84fca1b0 | 278 | * @flags: cmd_flags of the original request |
ae1b1539 TH |
279 | * |
280 | * Flush related states of @q have changed, consider issuing flush request. | |
281 | * Please read the comment at the top of this file for more info. | |
282 | * | |
283 | * CONTEXT: | |
9809b4ee | 284 | * spin_lock_irq(fq->mq_flush_lock) |
ae1b1539 | 285 | * |
ae1b1539 | 286 | */ |
404b8f5a | 287 | static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq, |
16458cf3 | 288 | blk_opf_t flags) |
86db1e29 | 289 | { |
7c94e1c1 | 290 | struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx]; |
ae1b1539 TH |
291 | struct request *first_rq = |
292 | list_first_entry(pending, struct request, flush.list); | |
7c94e1c1 | 293 | struct request *flush_rq = fq->flush_rq; |
ae1b1539 TH |
294 | |
295 | /* C1 described at the top of this file */ | |
7c94e1c1 | 296 | if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending)) |
404b8f5a | 297 | return; |
ae1b1539 | 298 | |
b5718d6c YY |
299 | /* C2 and C3 */ |
300 | if (!list_empty(&fq->flush_data_in_flight) && | |
ae1b1539 | 301 | time_before(jiffies, |
7c94e1c1 | 302 | fq->flush_pending_since + FLUSH_PENDING_TIMEOUT)) |
404b8f5a | 303 | return; |
ae1b1539 TH |
304 | |
305 | /* | |
306 | * Issue flush and toggle pending_idx. This makes pending_idx | |
307 | * different from running_idx, which means flush is in flight. | |
308 | */ | |
7c94e1c1 | 309 | fq->flush_pending_idx ^= 1; |
18741986 | 310 | |
7ddab5de | 311 | blk_rq_init(q, flush_rq); |
f70ced09 ML |
312 | |
313 | /* | |
923218f6 ML |
314 | * In case of none scheduler, borrow tag from the first request |
315 | * since they can't be in flight at the same time. And acquire | |
316 | * the tag's ownership for flush req. | |
317 | * | |
318 | * In case of IO scheduler, flush rq need to borrow scheduler tag | |
319 | * just for cheating put/get driver tag. | |
f70ced09 | 320 | */ |
7e992f84 | 321 | flush_rq->mq_ctx = first_rq->mq_ctx; |
ea4f995e | 322 | flush_rq->mq_hctx = first_rq->mq_hctx; |
7e992f84 | 323 | |
c1e2b842 | 324 | if (!q->elevator) { |
7e992f84 | 325 | flush_rq->tag = first_rq->tag; |
c1e2b842 ML |
326 | |
327 | /* | |
328 | * We borrow data request's driver tag, so have to mark | |
329 | * this flush request as INFLIGHT for avoiding double | |
330 | * account of this driver tag | |
331 | */ | |
332 | flush_rq->rq_flags |= RQF_MQ_INFLIGHT; | |
333 | } else | |
7e992f84 | 334 | flush_rq->internal_tag = first_rq->internal_tag; |
320ae51f | 335 | |
70fd7614 | 336 | flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH; |
84fca1b0 | 337 | flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK); |
e8064021 | 338 | flush_rq->rq_flags |= RQF_FLUSH_SEQ; |
7ddab5de | 339 | flush_rq->end_io = flush_end_io; |
c2da19ed ML |
340 | /* |
341 | * Order WRITE ->end_io and WRITE rq->ref, and its pair is the one | |
342 | * implied in refcount_inc_not_zero() called from | |
343 | * blk_mq_find_and_get_req(), which orders WRITE/READ flush_rq->ref | |
344 | * and READ flush_rq->end_io | |
345 | */ | |
346 | smp_wmb(); | |
0a467d0f | 347 | req_ref_set(flush_rq, 1); |
ae1b1539 | 348 | |
26a42b61 | 349 | blk_mq_add_to_requeue_list(flush_rq, 0); |
214a4418 | 350 | blk_mq_kick_requeue_list(q); |
86db1e29 JA |
351 | } |
352 | ||
de671d61 JA |
353 | static enum rq_end_io_ret mq_flush_data_end_io(struct request *rq, |
354 | blk_status_t error) | |
320ae51f JA |
355 | { |
356 | struct request_queue *q = rq->q; | |
ea4f995e | 357 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
e97c293c | 358 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
320ae51f | 359 | unsigned long flags; |
e97c293c | 360 | struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx); |
320ae51f | 361 | |
4e2f62e5 JA |
362 | if (q->elevator) { |
363 | WARN_ON(rq->tag < 0); | |
364 | blk_mq_put_driver_tag(rq); | |
365 | } | |
366 | ||
320ae51f JA |
367 | /* |
368 | * After populating an empty queue, kick it to avoid stall. Read | |
369 | * the comment in flush_end_io(). | |
370 | */ | |
7c94e1c1 | 371 | spin_lock_irqsave(&fq->mq_flush_lock, flags); |
bd166ef1 | 372 | blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error); |
7c94e1c1 | 373 | spin_unlock_irqrestore(&fq->mq_flush_lock, flags); |
bd166ef1 | 374 | |
85bd6e61 | 375 | blk_mq_sched_restart(hctx); |
de671d61 | 376 | return RQ_END_IO_NONE; |
320ae51f JA |
377 | } |
378 | ||
ae1b1539 | 379 | /** |
3140c3cf | 380 | * blk_insert_flush - insert a new PREFLUSH/FUA request |
ae1b1539 TH |
381 | * @rq: request to insert |
382 | * | |
b710a480 | 383 | * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions. |
320ae51f | 384 | * or __blk_mq_run_hw_queue() to dispatch request. |
ae1b1539 TH |
385 | * @rq is being submitted. Analyze what needs to be done and put it on the |
386 | * right queue. | |
ae1b1539 | 387 | */ |
2b504bd4 | 388 | void blk_insert_flush(struct request *rq) |
86db1e29 | 389 | { |
ae1b1539 | 390 | struct request_queue *q = rq->q; |
c888a8f9 | 391 | unsigned long fflags = q->queue_flags; /* may change, cache */ |
ae1b1539 | 392 | unsigned int policy = blk_flush_policy(fflags, rq); |
e97c293c | 393 | struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx); |
2394395c | 394 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
86db1e29 | 395 | |
ae1b1539 TH |
396 | /* |
397 | * @policy now records what operations need to be done. Adjust | |
28a8f0d3 | 398 | * REQ_PREFLUSH and FUA for the driver. |
ae1b1539 | 399 | */ |
28a8f0d3 | 400 | rq->cmd_flags &= ~REQ_PREFLUSH; |
c888a8f9 | 401 | if (!(fflags & (1UL << QUEUE_FLAG_FUA))) |
ae1b1539 TH |
402 | rq->cmd_flags &= ~REQ_FUA; |
403 | ||
ae5b2ec8 JA |
404 | /* |
405 | * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any | |
406 | * of those flags, we have to set REQ_SYNC to avoid skewing | |
407 | * the request accounting. | |
408 | */ | |
409 | rq->cmd_flags |= REQ_SYNC; | |
410 | ||
4853abaa JM |
411 | /* |
412 | * An empty flush handed down from a stacking driver may | |
413 | * translate into nothing if the underlying device does not | |
414 | * advertise a write-back cache. In this case, simply | |
415 | * complete the request. | |
416 | */ | |
417 | if (!policy) { | |
7e992f84 | 418 | blk_mq_end_request(rq, 0); |
2b504bd4 | 419 | return; |
4853abaa JM |
420 | } |
421 | ||
834f9f61 | 422 | BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */ |
4853abaa | 423 | |
ae1b1539 TH |
424 | /* |
425 | * If there's data but flush is not necessary, the request can be | |
426 | * processed directly without going through flush machinery. Queue | |
427 | * for normal execution. | |
428 | */ | |
429 | if ((policy & REQ_FSEQ_DATA) && | |
2b504bd4 | 430 | !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) { |
2b597613 | 431 | blk_mq_request_bypass_insert(rq, 0); |
2394395c | 432 | blk_mq_run_hw_queue(hctx, false); |
2b504bd4 ML |
433 | return; |
434 | } | |
cde4c406 | 435 | |
ae1b1539 TH |
436 | /* |
437 | * @rq should go through flush machinery. Mark it part of flush | |
438 | * sequence and submit for further processing. | |
439 | */ | |
440 | memset(&rq->flush, 0, sizeof(rq->flush)); | |
441 | INIT_LIST_HEAD(&rq->flush.list); | |
e8064021 | 442 | rq->rq_flags |= RQF_FLUSH_SEQ; |
4853abaa | 443 | rq->flush.saved_end_io = rq->end_io; /* Usually NULL */ |
320ae51f | 444 | |
7e992f84 | 445 | rq->end_io = mq_flush_data_end_io; |
ae1b1539 | 446 | |
7e992f84 | 447 | spin_lock_irq(&fq->mq_flush_lock); |
0bae352d | 448 | blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0); |
7e992f84 | 449 | spin_unlock_irq(&fq->mq_flush_lock); |
86db1e29 JA |
450 | } |
451 | ||
86db1e29 JA |
452 | /** |
453 | * blkdev_issue_flush - queue a flush | |
454 | * @bdev: blockdev to issue flush for | |
86db1e29 JA |
455 | * |
456 | * Description: | |
9398554f | 457 | * Issue a flush for the block device in question. |
86db1e29 | 458 | */ |
c6bf3f0e | 459 | int blkdev_issue_flush(struct block_device *bdev) |
86db1e29 | 460 | { |
c6bf3f0e | 461 | struct bio bio; |
86db1e29 | 462 | |
49add496 | 463 | bio_init(&bio, bdev, NULL, 0, REQ_OP_WRITE | REQ_PREFLUSH); |
c6bf3f0e | 464 | return submit_bio_wait(&bio); |
86db1e29 | 465 | } |
86db1e29 | 466 | EXPORT_SYMBOL(blkdev_issue_flush); |
320ae51f | 467 | |
754a1572 GJ |
468 | struct blk_flush_queue *blk_alloc_flush_queue(int node, int cmd_size, |
469 | gfp_t flags) | |
320ae51f | 470 | { |
7c94e1c1 ML |
471 | struct blk_flush_queue *fq; |
472 | int rq_sz = sizeof(struct request); | |
1bcb1ead | 473 | |
5b202853 | 474 | fq = kzalloc_node(sizeof(*fq), flags, node); |
7c94e1c1 ML |
475 | if (!fq) |
476 | goto fail; | |
1bcb1ead | 477 | |
7e992f84 | 478 | spin_lock_init(&fq->mq_flush_lock); |
7c94e1c1 | 479 | |
6d247d7f | 480 | rq_sz = round_up(rq_sz + cmd_size, cache_line_size()); |
5b202853 | 481 | fq->flush_rq = kzalloc_node(rq_sz, flags, node); |
7c94e1c1 ML |
482 | if (!fq->flush_rq) |
483 | goto fail_rq; | |
484 | ||
485 | INIT_LIST_HEAD(&fq->flush_queue[0]); | |
486 | INIT_LIST_HEAD(&fq->flush_queue[1]); | |
487 | INIT_LIST_HEAD(&fq->flush_data_in_flight); | |
488 | ||
489 | return fq; | |
490 | ||
491 | fail_rq: | |
492 | kfree(fq); | |
493 | fail: | |
494 | return NULL; | |
320ae51f | 495 | } |
f3552655 | 496 | |
ba483388 | 497 | void blk_free_flush_queue(struct blk_flush_queue *fq) |
f3552655 | 498 | { |
7c94e1c1 ML |
499 | /* bio based request queue hasn't flush queue */ |
500 | if (!fq) | |
501 | return; | |
3c09676c | 502 | |
7c94e1c1 ML |
503 | kfree(fq->flush_rq); |
504 | kfree(fq); | |
505 | } | |
fb01a293 ML |
506 | |
507 | /* | |
508 | * Allow driver to set its own lock class to fq->mq_flush_lock for | |
509 | * avoiding lockdep complaint. | |
510 | * | |
511 | * flush_end_io() may be called recursively from some driver, such as | |
512 | * nvme-loop, so lockdep may complain 'possible recursive locking' because | |
513 | * all 'struct blk_flush_queue' instance share same mq_flush_lock lock class | |
514 | * key. We need to assign different lock class for these driver's | |
515 | * fq->mq_flush_lock for avoiding the lockdep warning. | |
516 | * | |
517 | * Use dynamically allocated lock class key for each 'blk_flush_queue' | |
518 | * instance is over-kill, and more worse it introduces horrible boot delay | |
519 | * issue because synchronize_rcu() is implied in lockdep_unregister_key which | |
520 | * is called for each hctx release. SCSI probing may synchronously create and | |
521 | * destroy lots of MQ request_queues for non-existent devices, and some robot | |
522 | * test kernel always enable lockdep option. It is observed that more than half | |
523 | * an hour is taken during SCSI MQ probe with per-fq lock class. | |
524 | */ | |
525 | void blk_mq_hctx_set_fq_lock_class(struct blk_mq_hw_ctx *hctx, | |
526 | struct lock_class_key *key) | |
527 | { | |
528 | lockdep_set_class(&hctx->fq->mq_flush_lock, key); | |
529 | } | |
530 | EXPORT_SYMBOL_GPL(blk_mq_hctx_set_fq_lock_class); |