blk-flush.c (27eb2c4b3d3e13f376a359e293c212a2e9407af5) blk-flush.c (320ae51feed5c2f13664aa05a76bec198967e04d)
1/*
2 * Functions to sequence FLUSH and FUA writes.
3 *
4 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *

--- 55 unchanged lines hidden (view full) ---

64 * merged in the usual way.
65 */
66
67#include <linux/kernel.h>
68#include <linux/module.h>
69#include <linux/bio.h>
70#include <linux/blkdev.h>
71#include <linux/gfp.h>
1/*
2 * Functions to sequence FLUSH and FUA writes.
3 *
4 * Copyright (C) 2011 Max Planck Institute for Gravitational Physics
5 * Copyright (C) 2011 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *

--- 55 unchanged lines hidden (view full) ---

64 * merged in the usual way.
65 */
66
67#include <linux/kernel.h>
68#include <linux/module.h>
69#include <linux/bio.h>
70#include <linux/blkdev.h>
71#include <linux/gfp.h>
72#include <linux/blk-mq.h>
72
73#include "blk.h"
73
74#include "blk.h"
75#include "blk-mq.h"
74
75/* FLUSH/FUA sequences */
76enum {
77 REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */
78 REQ_FSEQ_DATA = (1 << 1), /* data write in progress */
79 REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */
80 REQ_FSEQ_DONE = (1 << 3),
81

--- 37 unchanged lines hidden (view full) ---

119 * complete the bio again. @rq->biotail is guaranteed to equal the
120 * original @rq->bio. Restore it.
121 */
122 rq->bio = rq->biotail;
123
124 /* make @rq a normal request */
125 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
126 rq->end_io = rq->flush.saved_end_io;
76
77/* FLUSH/FUA sequences */
78enum {
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

--- 37 unchanged lines hidden (view full) ---

121 * complete the bio again. @rq->biotail is guaranteed to equal the
122 * original @rq->bio. Restore it.
123 */
124 rq->bio = rq->biotail;
125
126 /* make @rq a normal request */
127 rq->cmd_flags &= ~REQ_FLUSH_SEQ;
128 rq->end_io = rq->flush.saved_end_io;
129
130 blk_clear_rq_complete(rq);
127}
128
131}
132
133static void mq_flush_data_run(struct work_struct *work)
134{
135 struct request *rq;
136
137 rq = container_of(work, struct request, mq_flush_data);
138
139 memset(&rq->csd, 0, sizeof(rq->csd));
140 blk_mq_run_request(rq, true, false);
141}
142
143static void blk_mq_flush_data_insert(struct request *rq)
144{
145 INIT_WORK(&rq->mq_flush_data, mq_flush_data_run);
146 kblockd_schedule_work(rq->q, &rq->mq_flush_data);
147}
148
129/**
130 * blk_flush_complete_seq - complete flush sequence
131 * @rq: FLUSH/FUA request being sequenced
132 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
133 * @error: whether an error occurred
134 *
135 * @rq just completed @seq part of its flush sequence, record the
136 * completion and trigger the next step.
137 *
138 * CONTEXT:
149/**
150 * blk_flush_complete_seq - complete flush sequence
151 * @rq: FLUSH/FUA request being sequenced
152 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero)
153 * @error: whether an error occurred
154 *
155 * @rq just completed @seq part of its flush sequence, record the
156 * completion and trigger the next step.
157 *
158 * CONTEXT:
139 * spin_lock_irq(q->queue_lock)
159 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
140 *
141 * RETURNS:
142 * %true if requests were added to the dispatch queue, %false otherwise.
143 */
144static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
145 int error)
146{
147 struct request_queue *q = rq->q;
148 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
160 *
161 * RETURNS:
162 * %true if requests were added to the dispatch queue, %false otherwise.
163 */
164static bool blk_flush_complete_seq(struct request *rq, unsigned int seq,
165 int error)
166{
167 struct request_queue *q = rq->q;
168 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
149 bool queued = false;
169 bool queued = false, kicked;
150
151 BUG_ON(rq->flush.seq & seq);
152 rq->flush.seq |= seq;
153
154 if (likely(!error))
155 seq = blk_flush_cur_seq(rq);
156 else
157 seq = REQ_FSEQ_DONE;

--- 4 unchanged lines hidden (view full) ---

162 /* queue for flush */
163 if (list_empty(pending))
164 q->flush_pending_since = jiffies;
165 list_move_tail(&rq->flush.list, pending);
166 break;
167
168 case REQ_FSEQ_DATA:
169 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
170
171 BUG_ON(rq->flush.seq & seq);
172 rq->flush.seq |= seq;
173
174 if (likely(!error))
175 seq = blk_flush_cur_seq(rq);
176 else
177 seq = REQ_FSEQ_DONE;

--- 4 unchanged lines hidden (view full) ---

182 /* queue for flush */
183 if (list_empty(pending))
184 q->flush_pending_since = jiffies;
185 list_move_tail(&rq->flush.list, pending);
186 break;
187
188 case REQ_FSEQ_DATA:
189 list_move_tail(&rq->flush.list, &q->flush_data_in_flight);
170 list_add(&rq->queuelist, &q->queue_head);
171 queued = true;
190 if (q->mq_ops)
191 blk_mq_flush_data_insert(rq);
192 else {
193 list_add(&rq->queuelist, &q->queue_head);
194 queued = true;
195 }
172 break;
173
174 case REQ_FSEQ_DONE:
175 /*
176 * @rq was previously adjusted by blk_flush_issue() for
177 * flush sequencing and may already have gone through the
178 * flush data request completion path. Restore @rq for
179 * normal completion and end it.
180 */
181 BUG_ON(!list_empty(&rq->queuelist));
182 list_del_init(&rq->flush.list);
183 blk_flush_restore_request(rq);
196 break;
197
198 case REQ_FSEQ_DONE:
199 /*
200 * @rq was previously adjusted by blk_flush_issue() for
201 * flush sequencing and may already have gone through the
202 * flush data request completion path. Restore @rq for
203 * normal completion and end it.
204 */
205 BUG_ON(!list_empty(&rq->queuelist));
206 list_del_init(&rq->flush.list);
207 blk_flush_restore_request(rq);
184 __blk_end_request_all(rq, error);
208 if (q->mq_ops)
209 blk_mq_end_io(rq, error);
210 else
211 __blk_end_request_all(rq, error);
185 break;
186
187 default:
188 BUG();
189 }
190
212 break;
213
214 default:
215 BUG();
216 }
217
191 return blk_kick_flush(q) | queued;
218 kicked = blk_kick_flush(q);
219 /* blk_mq_run_flush will run queue */
220 if (q->mq_ops)
221 return queued;
222 return kicked | queued;
192}
193
194static void flush_end_io(struct request *flush_rq, int error)
195{
196 struct request_queue *q = flush_rq->q;
223}
224
225static void flush_end_io(struct request *flush_rq, int error)
226{
227 struct request_queue *q = flush_rq->q;
197 struct list_head *running = &q->flush_queue[q->flush_running_idx];
228 struct list_head *running;
198 bool queued = false;
199 struct request *rq, *n;
229 bool queued = false;
230 struct request *rq, *n;
231 unsigned long flags = 0;
200
232
233 if (q->mq_ops) {
234 blk_mq_free_request(flush_rq);
235 spin_lock_irqsave(&q->mq_flush_lock, flags);
236 }
237 running = &q->flush_queue[q->flush_running_idx];
201 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
202
203 /* account completion of the flush request */
204 q->flush_running_idx ^= 1;
238 BUG_ON(q->flush_pending_idx == q->flush_running_idx);
239
240 /* account completion of the flush request */
241 q->flush_running_idx ^= 1;
205 elv_completed_request(q, flush_rq);
206
242
243 if (!q->mq_ops)
244 elv_completed_request(q, flush_rq);
245
207 /* and push the waiting requests to the next stage */
208 list_for_each_entry_safe(rq, n, running, flush.list) {
209 unsigned int seq = blk_flush_cur_seq(rq);
210
211 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
212 queued |= blk_flush_complete_seq(rq, seq, error);
213 }
214
215 /*
216 * Kick the queue to avoid stall for two cases:
217 * 1. Moving a request silently to empty queue_head may stall the
218 * queue.
219 * 2. When flush request is running in non-queueable queue, the
220 * queue is hold. Restart the queue after flush request is finished
221 * to avoid stall.
222 * This function is called from request completion path and calling
223 * directly into request_fn may confuse the driver. Always use
224 * kblockd.
225 */
246 /* and push the waiting requests to the next stage */
247 list_for_each_entry_safe(rq, n, running, flush.list) {
248 unsigned int seq = blk_flush_cur_seq(rq);
249
250 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH);
251 queued |= blk_flush_complete_seq(rq, seq, error);
252 }
253
254 /*
255 * Kick the queue to avoid stall for two cases:
256 * 1. Moving a request silently to empty queue_head may stall the
257 * queue.
258 * 2. When flush request is running in non-queueable queue, the
259 * queue is hold. Restart the queue after flush request is finished
260 * to avoid stall.
261 * This function is called from request completion path and calling
262 * directly into request_fn may confuse the driver. Always use
263 * kblockd.
264 */
226 if (queued || q->flush_queue_delayed)
227 blk_run_queue_async(q);
265 if (queued || q->flush_queue_delayed) {
266 if (!q->mq_ops)
267 blk_run_queue_async(q);
268 else
269 /*
270 * This can be optimized to only run queues with requests
271 * queued if necessary.
272 */
273 blk_mq_run_queues(q, true);
274 }
228 q->flush_queue_delayed = 0;
275 q->flush_queue_delayed = 0;
276 if (q->mq_ops)
277 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
229}
230
278}
279
280static void mq_flush_work(struct work_struct *work)
281{
282 struct request_queue *q;
283 struct request *rq;
284
285 q = container_of(work, struct request_queue, mq_flush_work);
286
287 /* We don't need set REQ_FLUSH_SEQ, it's for consistency */
288 rq = blk_mq_alloc_request(q, WRITE_FLUSH|REQ_FLUSH_SEQ,
289 __GFP_WAIT|GFP_ATOMIC);
290 rq->cmd_type = REQ_TYPE_FS;
291 rq->end_io = flush_end_io;
292
293 blk_mq_run_request(rq, true, false);
294}
295
296/*
297 * We can't directly use q->flush_rq, because it doesn't have tag and is not in
298 * hctx->rqs[]. so we must allocate a new request, since we can't sleep here,
299 * so offload the work to workqueue.
300 *
301 * Note: we assume a flush request finished in any hardware queue will flush
302 * the whole disk cache.
303 */
304static void mq_run_flush(struct request_queue *q)
305{
306 kblockd_schedule_work(q, &q->mq_flush_work);
307}
308
231/**
232 * blk_kick_flush - consider issuing flush request
233 * @q: request_queue being kicked
234 *
235 * Flush related states of @q have changed, consider issuing flush request.
236 * Please read the comment at the top of this file for more info.
237 *
238 * CONTEXT:
309/**
310 * blk_kick_flush - consider issuing flush request
311 * @q: request_queue being kicked
312 *
313 * Flush related states of @q have changed, consider issuing flush request.
314 * Please read the comment at the top of this file for more info.
315 *
316 * CONTEXT:
239 * spin_lock_irq(q->queue_lock)
317 * spin_lock_irq(q->queue_lock or q->mq_flush_lock)
240 *
241 * RETURNS:
242 * %true if flush was issued, %false otherwise.
243 */
244static bool blk_kick_flush(struct request_queue *q)
245{
246 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
247 struct request *first_rq =

--- 8 unchanged lines hidden (view full) ---

256 time_before(jiffies,
257 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
258 return false;
259
260 /*
261 * Issue flush and toggle pending_idx. This makes pending_idx
262 * different from running_idx, which means flush is in flight.
263 */
318 *
319 * RETURNS:
320 * %true if flush was issued, %false otherwise.
321 */
322static bool blk_kick_flush(struct request_queue *q)
323{
324 struct list_head *pending = &q->flush_queue[q->flush_pending_idx];
325 struct request *first_rq =

--- 8 unchanged lines hidden (view full) ---

334 time_before(jiffies,
335 q->flush_pending_since + FLUSH_PENDING_TIMEOUT))
336 return false;
337
338 /*
339 * Issue flush and toggle pending_idx. This makes pending_idx
340 * different from running_idx, which means flush is in flight.
341 */
342 q->flush_pending_idx ^= 1;
343 if (q->mq_ops) {
344 mq_run_flush(q);
345 return true;
346 }
347
264 blk_rq_init(q, &q->flush_rq);
265 q->flush_rq.cmd_type = REQ_TYPE_FS;
266 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
267 q->flush_rq.rq_disk = first_rq->rq_disk;
268 q->flush_rq.end_io = flush_end_io;
269
348 blk_rq_init(q, &q->flush_rq);
349 q->flush_rq.cmd_type = REQ_TYPE_FS;
350 q->flush_rq.cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
351 q->flush_rq.rq_disk = first_rq->rq_disk;
352 q->flush_rq.end_io = flush_end_io;
353
270 q->flush_pending_idx ^= 1;
271 list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
272 return true;
273}
274
275static void flush_data_end_io(struct request *rq, int error)
276{
277 struct request_queue *q = rq->q;
278
279 /*
280 * After populating an empty queue, kick it to avoid stall. Read
281 * the comment in flush_end_io().
282 */
283 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
284 blk_run_queue_async(q);
285}
286
354 list_add_tail(&q->flush_rq.queuelist, &q->queue_head);
355 return true;
356}
357
358static void flush_data_end_io(struct request *rq, int error)
359{
360 struct request_queue *q = rq->q;
361
362 /*
363 * After populating an empty queue, kick it to avoid stall. Read
364 * the comment in flush_end_io().
365 */
366 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
367 blk_run_queue_async(q);
368}
369
370static void mq_flush_data_end_io(struct request *rq, int error)
371{
372 struct request_queue *q = rq->q;
373 struct blk_mq_hw_ctx *hctx;
374 struct blk_mq_ctx *ctx;
375 unsigned long flags;
376
377 ctx = rq->mq_ctx;
378 hctx = q->mq_ops->map_queue(q, ctx->cpu);
379
380 /*
381 * After populating an empty queue, kick it to avoid stall. Read
382 * the comment in flush_end_io().
383 */
384 spin_lock_irqsave(&q->mq_flush_lock, flags);
385 if (blk_flush_complete_seq(rq, REQ_FSEQ_DATA, error))
386 blk_mq_run_hw_queue(hctx, true);
387 spin_unlock_irqrestore(&q->mq_flush_lock, flags);
388}
389
287/**
288 * blk_insert_flush - insert a new FLUSH/FUA request
289 * @rq: request to insert
290 *
291 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
390/**
391 * blk_insert_flush - insert a new FLUSH/FUA request
392 * @rq: request to insert
393 *
394 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions.
395 * or __blk_mq_run_hw_queue() to dispatch request.
292 * @rq is being submitted. Analyze what needs to be done and put it on the
293 * right queue.
294 *
295 * CONTEXT:
396 * @rq is being submitted. Analyze what needs to be done and put it on the
397 * right queue.
398 *
399 * CONTEXT:
296 * spin_lock_irq(q->queue_lock)
400 * spin_lock_irq(q->queue_lock) in !mq case
297 */
298void blk_insert_flush(struct request *rq)
299{
300 struct request_queue *q = rq->q;
301 unsigned int fflags = q->flush_flags; /* may change, cache */
302 unsigned int policy = blk_flush_policy(fflags, rq);
303
304 /*

--- 6 unchanged lines hidden (view full) ---

311
312 /*
313 * An empty flush handed down from a stacking driver may
314 * translate into nothing if the underlying device does not
315 * advertise a write-back cache. In this case, simply
316 * complete the request.
317 */
318 if (!policy) {
401 */
402void blk_insert_flush(struct request *rq)
403{
404 struct request_queue *q = rq->q;
405 unsigned int fflags = q->flush_flags; /* may change, cache */
406 unsigned int policy = blk_flush_policy(fflags, rq);
407
408 /*

--- 6 unchanged lines hidden (view full) ---

415
416 /*
417 * An empty flush handed down from a stacking driver may
418 * translate into nothing if the underlying device does not
419 * advertise a write-back cache. In this case, simply
420 * complete the request.
421 */
422 if (!policy) {
319 __blk_end_bidi_request(rq, 0, 0, 0);
423 if (q->mq_ops)
424 blk_mq_end_io(rq, 0);
425 else
426 __blk_end_bidi_request(rq, 0, 0, 0);
320 return;
321 }
322
323 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
324
325 /*
326 * If there's data but flush is not necessary, the request can be
327 * processed directly without going through flush machinery. Queue
328 * for normal execution.
329 */
330 if ((policy & REQ_FSEQ_DATA) &&
331 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
427 return;
428 }
429
430 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */
431
432 /*
433 * If there's data but flush is not necessary, the request can be
434 * processed directly without going through flush machinery. Queue
435 * for normal execution.
436 */
437 if ((policy & REQ_FSEQ_DATA) &&
438 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) {
332 list_add_tail(&rq->queuelist, &q->queue_head);
439 if (q->mq_ops) {
440 blk_mq_run_request(rq, false, true);
441 } else
442 list_add_tail(&rq->queuelist, &q->queue_head);
333 return;
334 }
335
336 /*
337 * @rq should go through flush machinery. Mark it part of flush
338 * sequence and submit for further processing.
339 */
340 memset(&rq->flush, 0, sizeof(rq->flush));
341 INIT_LIST_HEAD(&rq->flush.list);
342 rq->cmd_flags |= REQ_FLUSH_SEQ;
343 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
443 return;
444 }
445
446 /*
447 * @rq should go through flush machinery. Mark it part of flush
448 * sequence and submit for further processing.
449 */
450 memset(&rq->flush, 0, sizeof(rq->flush));
451 INIT_LIST_HEAD(&rq->flush.list);
452 rq->cmd_flags |= REQ_FLUSH_SEQ;
453 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */
454 if (q->mq_ops) {
455 rq->end_io = mq_flush_data_end_io;
456
457 spin_lock_irq(&q->mq_flush_lock);
458 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
459 spin_unlock_irq(&q->mq_flush_lock);
460 return;
461 }
344 rq->end_io = flush_data_end_io;
345
346 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
347}
348
349/**
350 * blk_abort_flushes - @q is being aborted, abort flush requests
351 * @q: request_queue being aborted

--- 96 unchanged lines hidden (view full) ---

448
449 if (!bio_flagged(bio, BIO_UPTODATE))
450 ret = -EIO;
451
452 bio_put(bio);
453 return ret;
454}
455EXPORT_SYMBOL(blkdev_issue_flush);
462 rq->end_io = flush_data_end_io;
463
464 blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
465}
466
467/**
468 * blk_abort_flushes - @q is being aborted, abort flush requests
469 * @q: request_queue being aborted

--- 96 unchanged lines hidden (view full) ---

566
567 if (!bio_flagged(bio, BIO_UPTODATE))
568 ret = -EIO;
569
570 bio_put(bio);
571 return ret;
572}
573EXPORT_SYMBOL(blkdev_issue_flush);
574
575void blk_mq_init_flush(struct request_queue *q)
576{
577 spin_lock_init(&q->mq_flush_lock);
578 INIT_WORK(&q->mq_flush_work, mq_flush_work);
579}