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 REQ_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->cmd_flags &= ~REQ_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 struct request_queue *q = rq->q; 138 139 blk_mq_add_to_requeue_list(rq, add_front); 140 blk_mq_kick_requeue_list(q); 141 return false; 142 } else { 143 if (add_front) 144 list_add(&rq->queuelist, &rq->q->queue_head); 145 else 146 list_add_tail(&rq->queuelist, &rq->q->queue_head); 147 return true; 148 } 149 } 150 151 /** 152 * blk_flush_complete_seq - complete flush sequence 153 * @rq: FLUSH/FUA request being sequenced 154 * @fq: flush queue 155 * @seq: sequences to complete (mask of %REQ_FSEQ_*, can be zero) 156 * @error: whether an error occurred 157 * 158 * @rq just completed @seq part of its flush sequence, record the 159 * completion and trigger the next step. 160 * 161 * CONTEXT: 162 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock) 163 * 164 * RETURNS: 165 * %true if requests were added to the dispatch queue, %false otherwise. 166 */ 167 static bool blk_flush_complete_seq(struct request *rq, 168 struct blk_flush_queue *fq, 169 unsigned int seq, int error) 170 { 171 struct request_queue *q = rq->q; 172 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx]; 173 bool queued = false, kicked; 174 175 BUG_ON(rq->flush.seq & seq); 176 rq->flush.seq |= seq; 177 178 if (likely(!error)) 179 seq = blk_flush_cur_seq(rq); 180 else 181 seq = REQ_FSEQ_DONE; 182 183 switch (seq) { 184 case REQ_FSEQ_PREFLUSH: 185 case REQ_FSEQ_POSTFLUSH: 186 /* queue for flush */ 187 if (list_empty(pending)) 188 fq->flush_pending_since = jiffies; 189 list_move_tail(&rq->flush.list, pending); 190 break; 191 192 case REQ_FSEQ_DATA: 193 list_move_tail(&rq->flush.list, &fq->flush_data_in_flight); 194 queued = blk_flush_queue_rq(rq, true); 195 break; 196 197 case REQ_FSEQ_DONE: 198 /* 199 * @rq was previously adjusted by blk_flush_issue() for 200 * flush sequencing and may already have gone through the 201 * flush data request completion path. Restore @rq for 202 * normal completion and end it. 203 */ 204 BUG_ON(!list_empty(&rq->queuelist)); 205 list_del_init(&rq->flush.list); 206 blk_flush_restore_request(rq); 207 if (q->mq_ops) 208 blk_mq_end_request(rq, error); 209 else 210 __blk_end_request_all(rq, error); 211 break; 212 213 default: 214 BUG(); 215 } 216 217 kicked = blk_kick_flush(q, fq); 218 return kicked | queued; 219 } 220 221 static void flush_end_io(struct request *flush_rq, int error) 222 { 223 struct request_queue *q = flush_rq->q; 224 struct list_head *running; 225 bool queued = false; 226 struct request *rq, *n; 227 unsigned long flags = 0; 228 struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx); 229 230 if (q->mq_ops) { 231 struct blk_mq_hw_ctx *hctx; 232 233 /* release the tag's ownership to the req cloned from */ 234 spin_lock_irqsave(&fq->mq_flush_lock, flags); 235 hctx = blk_mq_map_queue(q, flush_rq->mq_ctx->cpu); 236 blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq); 237 flush_rq->tag = -1; 238 } 239 240 running = &fq->flush_queue[fq->flush_running_idx]; 241 BUG_ON(fq->flush_pending_idx == fq->flush_running_idx); 242 243 /* account completion of the flush request */ 244 fq->flush_running_idx ^= 1; 245 246 if (!q->mq_ops) 247 elv_completed_request(q, flush_rq); 248 249 /* and push the waiting requests to the next stage */ 250 list_for_each_entry_safe(rq, n, running, flush.list) { 251 unsigned int seq = blk_flush_cur_seq(rq); 252 253 BUG_ON(seq != REQ_FSEQ_PREFLUSH && seq != REQ_FSEQ_POSTFLUSH); 254 queued |= blk_flush_complete_seq(rq, fq, seq, error); 255 } 256 257 /* 258 * Kick the queue to avoid stall for two cases: 259 * 1. Moving a request silently to empty queue_head may stall the 260 * queue. 261 * 2. When flush request is running in non-queueable queue, the 262 * queue is hold. Restart the queue after flush request is finished 263 * to avoid stall. 264 * This function is called from request completion path and calling 265 * directly into request_fn may confuse the driver. Always use 266 * kblockd. 267 */ 268 if (queued || fq->flush_queue_delayed) { 269 WARN_ON(q->mq_ops); 270 blk_run_queue_async(q); 271 } 272 fq->flush_queue_delayed = 0; 273 if (q->mq_ops) 274 spin_unlock_irqrestore(&fq->mq_flush_lock, flags); 275 } 276 277 /** 278 * blk_kick_flush - consider issuing flush request 279 * @q: request_queue being kicked 280 * @fq: flush queue 281 * 282 * Flush related states of @q have changed, consider issuing flush request. 283 * Please read the comment at the top of this file for more info. 284 * 285 * CONTEXT: 286 * spin_lock_irq(q->queue_lock or fq->mq_flush_lock) 287 * 288 * RETURNS: 289 * %true if flush was issued, %false otherwise. 290 */ 291 static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq) 292 { 293 struct list_head *pending = &fq->flush_queue[fq->flush_pending_idx]; 294 struct request *first_rq = 295 list_first_entry(pending, struct request, flush.list); 296 struct request *flush_rq = fq->flush_rq; 297 298 /* C1 described at the top of this file */ 299 if (fq->flush_pending_idx != fq->flush_running_idx || list_empty(pending)) 300 return false; 301 302 /* C2 and C3 */ 303 if (!list_empty(&fq->flush_data_in_flight) && 304 time_before(jiffies, 305 fq->flush_pending_since + FLUSH_PENDING_TIMEOUT)) 306 return false; 307 308 /* 309 * Issue flush and toggle pending_idx. This makes pending_idx 310 * different from running_idx, which means flush is in flight. 311 */ 312 fq->flush_pending_idx ^= 1; 313 314 blk_rq_init(q, flush_rq); 315 316 /* 317 * Borrow tag from the first request since they can't 318 * be in flight at the same time. And acquire the tag's 319 * ownership for flush req. 320 */ 321 if (q->mq_ops) { 322 struct blk_mq_hw_ctx *hctx; 323 324 flush_rq->mq_ctx = first_rq->mq_ctx; 325 flush_rq->tag = first_rq->tag; 326 fq->orig_rq = first_rq; 327 328 hctx = blk_mq_map_queue(q, first_rq->mq_ctx->cpu); 329 blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq); 330 } 331 332 flush_rq->cmd_type = REQ_TYPE_FS; 333 req_set_op_attrs(flush_rq, REQ_OP_FLUSH, WRITE_FLUSH | REQ_FLUSH_SEQ); 334 flush_rq->rq_disk = first_rq->rq_disk; 335 flush_rq->end_io = flush_end_io; 336 337 return blk_flush_queue_rq(flush_rq, false); 338 } 339 340 static void flush_data_end_io(struct request *rq, int error) 341 { 342 struct request_queue *q = rq->q; 343 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL); 344 345 /* 346 * Updating q->in_flight[] here for making this tag usable 347 * early. Because in blk_queue_start_tag(), 348 * q->in_flight[BLK_RW_ASYNC] is used to limit async I/O and 349 * reserve tags for sync I/O. 350 * 351 * More importantly this way can avoid the following I/O 352 * deadlock: 353 * 354 * - suppose there are 40 fua requests comming to flush queue 355 * and queue depth is 31 356 * - 30 rqs are scheduled then blk_queue_start_tag() can't alloc 357 * tag for async I/O any more 358 * - all the 30 rqs are completed before FLUSH_PENDING_TIMEOUT 359 * and flush_data_end_io() is called 360 * - the other rqs still can't go ahead if not updating 361 * q->in_flight[BLK_RW_ASYNC] here, meantime these rqs 362 * are held in flush data queue and make no progress of 363 * handling post flush rq 364 * - only after the post flush rq is handled, all these rqs 365 * can be completed 366 */ 367 368 elv_completed_request(q, rq); 369 370 /* for avoiding double accounting */ 371 rq->cmd_flags &= ~REQ_STARTED; 372 373 /* 374 * After populating an empty queue, kick it to avoid stall. Read 375 * the comment in flush_end_io(). 376 */ 377 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error)) 378 blk_run_queue_async(q); 379 } 380 381 static void mq_flush_data_end_io(struct request *rq, int error) 382 { 383 struct request_queue *q = rq->q; 384 struct blk_mq_hw_ctx *hctx; 385 struct blk_mq_ctx *ctx = rq->mq_ctx; 386 unsigned long flags; 387 struct blk_flush_queue *fq = blk_get_flush_queue(q, ctx); 388 389 hctx = blk_mq_map_queue(q, ctx->cpu); 390 391 /* 392 * After populating an empty queue, kick it to avoid stall. Read 393 * the comment in flush_end_io(). 394 */ 395 spin_lock_irqsave(&fq->mq_flush_lock, flags); 396 if (blk_flush_complete_seq(rq, fq, REQ_FSEQ_DATA, error)) 397 blk_mq_run_hw_queue(hctx, true); 398 spin_unlock_irqrestore(&fq->mq_flush_lock, flags); 399 } 400 401 /** 402 * blk_insert_flush - insert a new FLUSH/FUA request 403 * @rq: request to insert 404 * 405 * To be called from __elv_add_request() for %ELEVATOR_INSERT_FLUSH insertions. 406 * or __blk_mq_run_hw_queue() to dispatch request. 407 * @rq is being submitted. Analyze what needs to be done and put it on the 408 * right queue. 409 * 410 * CONTEXT: 411 * spin_lock_irq(q->queue_lock) in !mq case 412 */ 413 void blk_insert_flush(struct request *rq) 414 { 415 struct request_queue *q = rq->q; 416 unsigned long fflags = q->queue_flags; /* may change, cache */ 417 unsigned int policy = blk_flush_policy(fflags, rq); 418 struct blk_flush_queue *fq = blk_get_flush_queue(q, rq->mq_ctx); 419 420 /* 421 * @policy now records what operations need to be done. Adjust 422 * REQ_PREFLUSH and FUA for the driver. 423 */ 424 rq->cmd_flags &= ~REQ_PREFLUSH; 425 if (!(fflags & (1UL << QUEUE_FLAG_FUA))) 426 rq->cmd_flags &= ~REQ_FUA; 427 428 /* 429 * An empty flush handed down from a stacking driver may 430 * translate into nothing if the underlying device does not 431 * advertise a write-back cache. In this case, simply 432 * complete the request. 433 */ 434 if (!policy) { 435 if (q->mq_ops) 436 blk_mq_end_request(rq, 0); 437 else 438 __blk_end_bidi_request(rq, 0, 0, 0); 439 return; 440 } 441 442 BUG_ON(rq->bio != rq->biotail); /*assumes zero or single bio rq */ 443 444 /* 445 * If there's data but flush is not necessary, the request can be 446 * processed directly without going through flush machinery. Queue 447 * for normal execution. 448 */ 449 if ((policy & REQ_FSEQ_DATA) && 450 !(policy & (REQ_FSEQ_PREFLUSH | REQ_FSEQ_POSTFLUSH))) { 451 if (q->mq_ops) { 452 blk_mq_insert_request(rq, false, false, true); 453 } else 454 list_add_tail(&rq->queuelist, &q->queue_head); 455 return; 456 } 457 458 /* 459 * @rq should go through flush machinery. Mark it part of flush 460 * sequence and submit for further processing. 461 */ 462 memset(&rq->flush, 0, sizeof(rq->flush)); 463 INIT_LIST_HEAD(&rq->flush.list); 464 rq->cmd_flags |= REQ_FLUSH_SEQ; 465 rq->flush.saved_end_io = rq->end_io; /* Usually NULL */ 466 if (q->mq_ops) { 467 rq->end_io = mq_flush_data_end_io; 468 469 spin_lock_irq(&fq->mq_flush_lock); 470 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0); 471 spin_unlock_irq(&fq->mq_flush_lock); 472 return; 473 } 474 rq->end_io = flush_data_end_io; 475 476 blk_flush_complete_seq(rq, fq, REQ_FSEQ_ACTIONS & ~policy, 0); 477 } 478 479 /** 480 * blkdev_issue_flush - queue a flush 481 * @bdev: blockdev to issue flush for 482 * @gfp_mask: memory allocation flags (for bio_alloc) 483 * @error_sector: error sector 484 * 485 * Description: 486 * Issue a flush for the block device in question. Caller can supply 487 * room for storing the error offset in case of a flush error, if they 488 * wish to. If WAIT flag is not passed then caller may check only what 489 * request was pushed in some internal queue for later handling. 490 */ 491 int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask, 492 sector_t *error_sector) 493 { 494 struct request_queue *q; 495 struct bio *bio; 496 int ret = 0; 497 498 if (bdev->bd_disk == NULL) 499 return -ENXIO; 500 501 q = bdev_get_queue(bdev); 502 if (!q) 503 return -ENXIO; 504 505 /* 506 * some block devices may not have their queue correctly set up here 507 * (e.g. loop device without a backing file) and so issuing a flush 508 * here will panic. Ensure there is a request function before issuing 509 * the flush. 510 */ 511 if (!q->make_request_fn) 512 return -ENXIO; 513 514 bio = bio_alloc(gfp_mask, 0); 515 bio->bi_bdev = bdev; 516 bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_FLUSH); 517 518 ret = submit_bio_wait(bio); 519 520 /* 521 * The driver must store the error location in ->bi_sector, if 522 * it supports it. For non-stacked drivers, this should be 523 * copied from blk_rq_pos(rq). 524 */ 525 if (error_sector) 526 *error_sector = bio->bi_iter.bi_sector; 527 528 bio_put(bio); 529 return ret; 530 } 531 EXPORT_SYMBOL(blkdev_issue_flush); 532 533 struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q, 534 int node, int cmd_size) 535 { 536 struct blk_flush_queue *fq; 537 int rq_sz = sizeof(struct request); 538 539 fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node); 540 if (!fq) 541 goto fail; 542 543 if (q->mq_ops) { 544 spin_lock_init(&fq->mq_flush_lock); 545 rq_sz = round_up(rq_sz + cmd_size, cache_line_size()); 546 } 547 548 fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node); 549 if (!fq->flush_rq) 550 goto fail_rq; 551 552 INIT_LIST_HEAD(&fq->flush_queue[0]); 553 INIT_LIST_HEAD(&fq->flush_queue[1]); 554 INIT_LIST_HEAD(&fq->flush_data_in_flight); 555 556 return fq; 557 558 fail_rq: 559 kfree(fq); 560 fail: 561 return NULL; 562 } 563 564 void blk_free_flush_queue(struct blk_flush_queue *fq) 565 { 566 /* bio based request queue hasn't flush queue */ 567 if (!fq) 568 return; 569 570 kfree(fq->flush_rq); 571 kfree(fq); 572 } 573