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 * 9 * REQ_{FLUSH|FUA} requests are decomposed to sequences consisted of three 10 * optional steps - PREFLUSH, DATA and POSTFLUSH - according to the request 11 * properties and hardware capability. 12 * 13 * If a request doesn't have data, only REQ_PREFLUSH makes sense, which 14 * indicates a simple flush request. If there is data, REQ_PREFLUSH indicates 15 * that the device cache should be flushed before the data is executed, and 16 * REQ_FUA means that the data must be on non-volatile media on request 17 * completion. 18 * 19 * If the device doesn't have writeback cache, FLUSH and FUA don't make any 20 * difference. The requests are either completed immediately if there's no 21 * data or executed as normal requests otherwise. 22 * 23 * If the device has writeback cache and supports FUA, REQ_PREFLUSH is 24 * translated to PREFLUSH but REQ_FUA is passed down directly with DATA. 25 * 26 * If the device has writeback cache and doesn't support FUA, REQ_PREFLUSH 27 * is translated to PREFLUSH and REQ_FUA to POSTFLUSH. 28 * 29 * The actual execution of flush is double buffered. Whenever a request 30 * needs to execute PRE or POSTFLUSH, it queues at 31 * fq->flush_queue[fq->flush_pending_idx]. Once certain criteria are met, a 32 * REQ_OP_FLUSH is issued and the pending_idx is toggled. When the flush 33 * completes, all the requests which were pending are proceeded to the next 34 * step. This allows arbitrary merging of different types of FLUSH/FUA 35 * requests. 36 * 37 * Currently, the following conditions are used to determine when to issue 38 * flush. 39 * 40 * C1. At any given time, only one flush shall be in progress. This makes 41 * double buffering sufficient. 42 * 43 * C2. Flush is deferred if any request is executing DATA of its sequence. 44 * This avoids issuing separate POSTFLUSHes for requests which shared 45 * PREFLUSH. 46 * 47 * C3. The second condition is ignored if there is a request which has 48 * waited longer than FLUSH_PENDING_TIMEOUT. This is to avoid 49 * starvation in the unlikely case where there are continuous stream of 50 * FUA (without FLUSH) requests. 51 * 52 * For devices which support FUA, it isn't clear whether C2 (and thus C3) 53 * is beneficial. 54 * 55 * Note that a sequenced FLUSH/FUA request with DATA is completed twice. 56 * Once while executing DATA and again after the whole sequence is 57 * complete. The first completion updates the contained bio but doesn't 58 * finish it so that the bio submitter is notified only after the whole 59 * sequence is complete. This is implemented by testing RQF_FLUSH_SEQ in 60 * req_bio_endio(). 61 * 62 * The above peculiarity requires that each FLUSH/FUA request has only one 63 * bio attached to it, which is guaranteed as they aren't allowed to be 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> 73 74 #include "blk.h" 75 #include "blk-mq.h" 76 #include "blk-mq-tag.h" 77 78 /* FLUSH/FUA sequences */ 79 enum { 80 REQ_FSEQ_PREFLUSH = (1 << 0), /* pre-flushing in progress */ 81 REQ_FSEQ_DATA = (1 << 1), /* data write in progress */ 82 REQ_FSEQ_POSTFLUSH = (1 << 2), /* post-flushing in progress */ 83 REQ_FSEQ_DONE = (1 << 3), 84 85 REQ_FSEQ_ACTIONS = REQ_FSEQ_PREFLUSH | REQ_FSEQ_DATA | 86 REQ_FSEQ_POSTFLUSH, 87 88 /* 89 * If flush has been pending longer than the following timeout, 90 * it's issued even if flush_data requests are still in flight. 91 */ 92 FLUSH_PENDING_TIMEOUT = 5 * HZ, 93 }; 94 95 static bool blk_kick_flush(struct request_queue *q, 96 struct blk_flush_queue *fq); 97 98 static unsigned int blk_flush_policy(unsigned long fflags, struct request *rq) 99 { 100 unsigned int policy = 0; 101 102 if (blk_rq_sectors(rq)) 103 policy |= REQ_FSEQ_DATA; 104 105 if (fflags & (1UL << QUEUE_FLAG_WC)) { 106 if (rq->cmd_flags & REQ_PREFLUSH) 107 policy |= REQ_FSEQ_PREFLUSH; 108 if (!(fflags & (1UL << QUEUE_FLAG_FUA)) && 109 (rq->cmd_flags & REQ_FUA)) 110 policy |= REQ_FSEQ_POSTFLUSH; 111 } 112 return policy; 113 } 114 115 static unsigned int blk_flush_cur_seq(struct request *rq) 116 { 117 return 1 << ffz(rq->flush.seq); 118 } 119 120 static void blk_flush_restore_request(struct request *rq) 121 { 122 /* 123 * After flush data completion, @rq->bio is %NULL but we need to 124 * complete the bio again. @rq->biotail is guaranteed to equal the 125 * original @rq->bio. Restore it. 126 */ 127 rq->bio = rq->biotail; 128 129 /* make @rq a normal request */ 130 rq->rq_flags &= ~RQF_FLUSH_SEQ; 131 rq->end_io = rq->flush.saved_end_io; 132 } 133 134 static bool blk_flush_queue_rq(struct request *rq, bool add_front) 135 { 136 if (rq->q->mq_ops) { 137 blk_mq_add_to_requeue_list(rq, add_front, true); 138 return false; 139 } else { 140 if (add_front) 141 list_add(&rq->queuelist, &rq->q->queue_head); 142 else 143 list_add_tail(&rq->queuelist, &rq->q->queue_head); 144 return true; 145 } 146 } 147 148 /** 149 * blk_flush_complete_seq - complete flush sequence 150 * @rq: FLUSH/FUA request being sequenced 151 * @fq: flush queue 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: 159 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock) 160 * 161 * RETURNS: 162 * %true if requests were added to the dispatch queue, %false otherwise. 163 */ 164 static bool blk_flush_complete_seq(struct request *rq, 165 struct blk_flush_queue *fq, 166 unsigned int seq, int error) 167 { 168 struct request_queue *q = rq->q; 169 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx]; 170 bool queued = false, kicked; 171 172 BUG_ON(rq->flush.seq & seq); 173 rq->flush.seq |= seq; 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)) 185 fq->flush_pending_since = jiffies; 186 list_move_tail(&rq->flush.list, pending); 187 break; 188 189 case REQ_FSEQ_DATA: 190 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight); 191 queued = blk_flush_queue_rq(rq, true); 192 break; 193 194 case REQ_FSEQ_DONE: 195 /* 196 * @rq was previously adjusted by blk_flush_issue() for 197 * flush sequencing and may already have gone through the 198 * flush data request completion path. Restore @rq for 199 * normal completion and end it. 200 */ 201 BUG_ON(!list_empty(&rq->queuelist)); 202 list_del_init(&rq->flush.list); 203 blk_flush_restore_request(rq); 204 if (q->mq_ops) 205 blk_mq_end_request(rq, error); 206 else 207 __blk_end_request_all(rq, error); 208 break; 209 210 default: 211 BUG(); 212 } 213 214 kicked = blk_kick_flush(q, fq); 215 return kicked | queued; 216 } 217 218 static void flush_end_io(struct request *flush_rq, int error) 219 { 220 struct request_queue *q = flush_rq->q; 221 struct list_head *running; 222 bool queued = false; 223 struct request *rq, *n; 224 unsigned long flags = 0; 225 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx); 226 227 if (q->mq_ops) { 228 struct blk_mq_hw_ctx *hctx; 229 230 /* release the tag's ownership to the req cloned from */ 231 spin_lock_irqsave(&fq->mq_flush_lock, flags); 232 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu); 233 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq); 234 flush_rq->tag = -1; 235 } 236 237 running = &fq->flush_queue[fq->flush_running_idx]; 238 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx); 239 240 /* account completion of the flush request */ 241 fq->flush_running_idx ^= 1; 242 243 if (!q->mq_ops) 244 elv_completed_request(q, flush_rq); 245 246 /* 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, fq, 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 */ 265 if (queued || fq->flush_queue_delayed) { 266 WARN_ON(q->mq_ops); 267 blk_run_queue_async(q); 268 } 269 fq->flush_queue_delayed = 0; 270 if (q->mq_ops) 271 spin_unlock_irqrestore(&fq->mq_flush_lock, flags); 272 } 273 274 /** 275 * blk_kick_flush - consider issuing flush request 276 * @q: request_queue being kicked 277 * @fq: flush queue 278 * 279 * Flush related states of @q have changed, consider issuing flush request. 280 * Please read the comment at the top of this file for more info. 281 * 282 * CONTEXT: 283 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock) 284 * 285 * RETURNS: 286 * %true if flush was issued, %false otherwise. 287 */ 288 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq) 289 { 290 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx]; 291 struct request *first_rq = 292 list_first_entry(pending, struct request, flush.list); 293 struct request *flush_rq = fq->flush_rq; 294 295 /* C1 described at the top of this file */ 296 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending)) 297 return false; 298 299 /* C2 and C3 */ 300 if (!list_empty(&fq->flush_data_in_flight) && 301 time_before(jiffies, 302 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT)) 303 return false; 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 */ 309 fq->flush_pending_idx ^= 1; 310 311 blk_rq_init(q, flush_rq); 312 313 /* 314 * Borrow tag from the first request since they can't 315 * be in flight at the same time. And acquire the tag's 316 * ownership for flush req. 317 */ 318 if (q->mq_ops) { 319 struct blk_mq_hw_ctx *hctx; 320 321 flush_rq->mq_ctx = first_rq->mq_ctx; 322 flush_rq->tag = first_rq->tag; 323 fq->orig_rq = first_rq; 324 325 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu); 326 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq); 327 } 328 329 flush_rq->cmd_type = REQ_TYPE_FS; 330 flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH; 331 flush_rq->rq_flags |= RQF_FLUSH_SEQ; 332 flush_rq->rq_disk = first_rq->rq_disk; 333 flush_rq->end_io = flush_end_io; 334 335 return blk_flush_queue_rq(flush_rq, false); 336 } 337 338 static void flush_data_end_io(struct request *rq, int error) 339 { 340 struct request_queue *q = rq->q; 341 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL); 342 343 /* 344 * Updating q->in_flight[] here for making this tag usable 345 * early. Because in blk_queue_start_tag(), 346 * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and 347 * reserve tags for sync I/O. 348 * 349 * More importantly this way can avoid the following I/O 350 * deadlock: 351 * 352 * - suppose there are 40 fua requests comming to flush queue 353 * and queue depth is 31 354 * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc 355 * tag for async I/O any more 356 * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT 357 * and flush_data_end_io() is called 358 * - the other rqs still can't go ahead if not updating 359 * q->in_flight[BLK_RW_ASYNC] here, meantime these rqs 360 * are held in flush data queue and make no progress of 361 * handling post flush rq 362 * - only after the post flush rq is handled, all these rqs 363 * can be completed 364 */ 365 366 elv_completed_request(q, rq); 367 368 /* for avoiding double accounting */ 369 rq->rq_flags &= ~RQF_STARTED; 370 371 /* 372 * After populating an empty queue, kick it to avoid stall. Read 373 * the comment in flush_end_io(). 374 */ 375 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error)) 376 blk_run_queue_async(q); 377 } 378 379 static void mq_flush_data_end_io(struct request *rq, int error) 380 { 381 struct request_queue *q = rq->q; 382 struct blk_mq_hw_ctx *hctx; 383 struct blk_mq_ctx *ctx = rq->mq_ctx; 384 unsigned long flags; 385 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx); 386 387 hctx = blk_mq_map_queue(q, ctx->cpu); 388 389 /* 390 * After populating an empty queue, kick it to avoid stall. Read 391 * the comment in flush_end_io(). 392 */ 393 spin_lock_irqsave(&fq->mq_flush_lock, flags); 394 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error)) 395 blk_mq_run_hw_queue(hctx, true); 396 spin_unlock_irqrestore(&fq->mq_flush_lock, flags); 397 } 398 399 /** 400 * blk_insert_flush - insert a new FLUSH/FUA request 401 * @rq: request to insert 402 * 403 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions. 404 * or __blk_mq_run_hw_queue() to dispatch request. 405 * @rq is being submitted. Analyze what needs to be done and put it on the 406 * right queue. 407 * 408 * CONTEXT: 409 * spin_lock_irq(q->queue_lock) in !mq case 410 */ 411 void blk_insert_flush(struct request *rq) 412 { 413 struct request_queue *q = rq->q; 414 unsigned long fflags = q->queue_flags; /* may change, cache */ 415 unsigned int policy = blk_flush_policy(fflags, rq); 416 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx); 417 418 /* 419 * @policy now records what operations need to be done. Adjust 420 * REQ_PREFLUSH and FUA for the driver. 421 */ 422 rq->cmd_flags &= ~REQ_PREFLUSH; 423 if (!(fflags & (1UL << QUEUE_FLAG_FUA))) 424 rq->cmd_flags &= ~REQ_FUA; 425 426 /* 427 * REQ_PREFLUSH|REQ_FUA implies REQ_SYNC, so if we clear any 428 * of those flags, we have to set REQ_SYNC to avoid skewing 429 * the request accounting. 430 */ 431 rq->cmd_flags |= REQ_SYNC; 432 433 /* 434 * An empty flush handed down from a stacking driver may 435 * translate into nothing if the underlying device does not 436 * advertise a write-back cache. In this case, simply 437 * complete the request. 438 */ 439 if (!policy) { 440 if (q->mq_ops) 441 blk_mq_end_request(rq, 0); 442 else 443 __blk_end_bidi_request(rq, 0, 0, 0); 444 return; 445 } 446 447 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */ 448 449 /* 450 * If there's data but flush is not necessary, the request can be 451 * processed directly without going through flush machinery. Queue 452 * for normal execution. 453 */ 454 if ((policy & REQ_FSEQ_DATA) && 455 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) { 456 if (q->mq_ops) { 457 blk_mq_insert_request(rq, false, true, false); 458 } else 459 list_add_tail(&rq->queuelist, &q->queue_head); 460 return; 461 } 462 463 /* 464 * @rq should go through flush machinery. Mark it part of flush 465 * sequence and submit for further processing. 466 */ 467 memset(&rq->flush, 0, sizeof(rq->flush)); 468 INIT_LIST_HEAD(&rq->flush.list); 469 rq->rq_flags |= RQF_FLUSH_SEQ; 470 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */ 471 if (q->mq_ops) { 472 rq->end_io = mq_flush_data_end_io; 473 474 spin_lock_irq(&fq->mq_flush_lock); 475 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0); 476 spin_unlock_irq(&fq->mq_flush_lock); 477 return; 478 } 479 rq->end_io = flush_data_end_io; 480 481 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0); 482 } 483 484 /** 485 * blkdev_issue_flush - queue a flush 486 * @bdev: blockdev to issue flush for 487 * @gfp_mask: memory allocation flags (for bio_alloc) 488 * @error_sector: error sector 489 * 490 * Description: 491 * Issue a flush for the block device in question. Caller can supply 492 * room for storing the error offset in case of a flush error, if they 493 * wish to. If WAIT flag is not passed then caller may check only what 494 * request was pushed in some internal queue for later handling. 495 */ 496 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask, 497 sector_t *error_sector) 498 { 499 struct request_queue *q; 500 struct bio *bio; 501 int ret = 0; 502 503 if (bdev->bd_disk == NULL) 504 return -ENXIO; 505 506 q = bdev_get_queue(bdev); 507 if (!q) 508 return -ENXIO; 509 510 /* 511 * some block devices may not have their queue correctly set up here 512 * (e.g. loop device without a backing file) and so issuing a flush 513 * here will panic. Ensure there is a request function before issuing 514 * the flush. 515 */ 516 if (!q->make_request_fn) 517 return -ENXIO; 518 519 bio = bio_alloc(gfp_mask, 0); 520 bio->bi_bdev = bdev; 521 bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH; 522 523 ret = submit_bio_wait(bio); 524 525 /* 526 * The driver must store the error location in ->bi_sector, if 527 * it supports it. For non-stacked drivers, this should be 528 * copied from blk_rq_pos(rq). 529 */ 530 if (error_sector) 531 *error_sector = bio->bi_iter.bi_sector; 532 533 bio_put(bio); 534 return ret; 535 } 536 EXPORT_SYMBOL(blkdev_issue_flush); 537 538 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q, 539 int node, int cmd_size) 540 { 541 struct blk_flush_queue *fq; 542 int rq_sz = sizeof(struct request); 543 544 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node); 545 if (!fq) 546 goto fail; 547 548 if (q->mq_ops) { 549 spin_lock_init(&fq->mq_flush_lock); 550 rq_sz = round_up(rq_sz + cmd_size, cache_line_size()); 551 } 552 553 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node); 554 if (!fq->flush_rq) 555 goto fail_rq; 556 557 INIT_LIST_HEAD(&fq->flush_queue[0]); 558 INIT_LIST_HEAD(&fq->flush_queue[1]); 559 INIT_LIST_HEAD(&fq->flush_data_in_flight); 560 561 return fq; 562 563 fail_rq: 564 kfree(fq); 565 fail: 566 return NULL; 567 } 568 569 void blk_free_flush_queue(struct blk_flush_queue *fq) 570 { 571 /* bio based request queue hasn't flush queue */ 572 if (!fq) 573 return; 574 575 kfree(fq->flush_rq); 576 kfree(fq); 577 } 578