1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Functions related to segment and merge handling 4 */ 5 #include <linux/kernel.h> 6 #include <linux/module.h> 7 #include <linux/bio.h> 8 #include <linux/blkdev.h> 9 #include <linux/blk-integrity.h> 10 #include <linux/scatterlist.h> 11 #include <linux/part_stat.h> 12 #include <linux/blk-cgroup.h> 13 14 #include <trace/events/block.h> 15 16 #include "blk.h" 17 #include "blk-mq-sched.h" 18 #include "blk-rq-qos.h" 19 #include "blk-throttle.h" 20 21 static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv) 22 { 23 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 24 } 25 26 static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv) 27 { 28 struct bvec_iter iter = bio->bi_iter; 29 int idx; 30 31 bio_get_first_bvec(bio, bv); 32 if (bv->bv_len == bio->bi_iter.bi_size) 33 return; /* this bio only has a single bvec */ 34 35 bio_advance_iter(bio, &iter, iter.bi_size); 36 37 if (!iter.bi_bvec_done) 38 idx = iter.bi_idx - 1; 39 else /* in the middle of bvec */ 40 idx = iter.bi_idx; 41 42 *bv = bio->bi_io_vec[idx]; 43 44 /* 45 * iter.bi_bvec_done records actual length of the last bvec 46 * if this bio ends in the middle of one io vector 47 */ 48 if (iter.bi_bvec_done) 49 bv->bv_len = iter.bi_bvec_done; 50 } 51 52 static inline bool bio_will_gap(struct request_queue *q, 53 struct request *prev_rq, struct bio *prev, struct bio *next) 54 { 55 struct bio_vec pb, nb; 56 57 if (!bio_has_data(prev) || !queue_virt_boundary(q)) 58 return false; 59 60 /* 61 * Don't merge if the 1st bio starts with non-zero offset, otherwise it 62 * is quite difficult to respect the sg gap limit. We work hard to 63 * merge a huge number of small single bios in case of mkfs. 64 */ 65 if (prev_rq) 66 bio_get_first_bvec(prev_rq->bio, &pb); 67 else 68 bio_get_first_bvec(prev, &pb); 69 if (pb.bv_offset & queue_virt_boundary(q)) 70 return true; 71 72 /* 73 * We don't need to worry about the situation that the merged segment 74 * ends in unaligned virt boundary: 75 * 76 * - if 'pb' ends aligned, the merged segment ends aligned 77 * - if 'pb' ends unaligned, the next bio must include 78 * one single bvec of 'nb', otherwise the 'nb' can't 79 * merge with 'pb' 80 */ 81 bio_get_last_bvec(prev, &pb); 82 bio_get_first_bvec(next, &nb); 83 if (biovec_phys_mergeable(q, &pb, &nb)) 84 return false; 85 return __bvec_gap_to_prev(q, &pb, nb.bv_offset); 86 } 87 88 static inline bool req_gap_back_merge(struct request *req, struct bio *bio) 89 { 90 return bio_will_gap(req->q, req, req->biotail, bio); 91 } 92 93 static inline bool req_gap_front_merge(struct request *req, struct bio *bio) 94 { 95 return bio_will_gap(req->q, NULL, bio, req->bio); 96 } 97 98 static struct bio *bio_split_discard(struct bio *bio, struct request_queue *q, 99 unsigned *nsegs, struct bio_set *bs) 100 { 101 unsigned int max_discard_sectors, granularity; 102 int alignment; 103 sector_t tmp; 104 unsigned split_sectors; 105 106 *nsegs = 1; 107 108 /* Zero-sector (unknown) and one-sector granularities are the same. */ 109 granularity = max(q->limits.discard_granularity >> 9, 1U); 110 111 max_discard_sectors = min(q->limits.max_discard_sectors, 112 bio_allowed_max_sectors(q)); 113 max_discard_sectors -= max_discard_sectors % granularity; 114 115 if (unlikely(!max_discard_sectors)) { 116 /* XXX: warn */ 117 return NULL; 118 } 119 120 if (bio_sectors(bio) <= max_discard_sectors) 121 return NULL; 122 123 split_sectors = max_discard_sectors; 124 125 /* 126 * If the next starting sector would be misaligned, stop the discard at 127 * the previous aligned sector. 128 */ 129 alignment = (q->limits.discard_alignment >> 9) % granularity; 130 131 tmp = bio->bi_iter.bi_sector + split_sectors - alignment; 132 tmp = sector_div(tmp, granularity); 133 134 if (split_sectors > tmp) 135 split_sectors -= tmp; 136 137 return bio_split(bio, split_sectors, GFP_NOIO, bs); 138 } 139 140 static struct bio *bio_split_write_zeroes(struct bio *bio, 141 struct request_queue *q, unsigned *nsegs, struct bio_set *bs) 142 { 143 *nsegs = 0; 144 145 if (!q->limits.max_write_zeroes_sectors) 146 return NULL; 147 148 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) 149 return NULL; 150 151 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); 152 } 153 154 /* 155 * Return the maximum number of sectors from the start of a bio that may be 156 * submitted as a single request to a block device. If enough sectors remain, 157 * align the end to the physical block size. Otherwise align the end to the 158 * logical block size. This approach minimizes the number of non-aligned 159 * requests that are submitted to a block device if the start of a bio is not 160 * aligned to a physical block boundary. 161 */ 162 static inline unsigned get_max_io_size(struct bio *bio, 163 struct request_queue *q) 164 { 165 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT; 166 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT; 167 unsigned max_sectors = queue_max_sectors(q), start, end; 168 169 if (q->limits.chunk_sectors) { 170 max_sectors = min(max_sectors, 171 blk_chunk_sectors_left(bio->bi_iter.bi_sector, 172 q->limits.chunk_sectors)); 173 } 174 175 start = bio->bi_iter.bi_sector & (pbs - 1); 176 end = (start + max_sectors) & ~(pbs - 1); 177 if (end > start) 178 return end - start; 179 return max_sectors & ~(lbs - 1); 180 } 181 182 static inline unsigned get_max_segment_size(const struct request_queue *q, 183 struct page *start_page, 184 unsigned long offset) 185 { 186 unsigned long mask = queue_segment_boundary(q); 187 188 offset = mask & (page_to_phys(start_page) + offset); 189 190 /* 191 * overflow may be triggered in case of zero page physical address 192 * on 32bit arch, use queue's max segment size when that happens. 193 */ 194 return min_not_zero(mask - offset + 1, 195 (unsigned long)queue_max_segment_size(q)); 196 } 197 198 /** 199 * bvec_split_segs - verify whether or not a bvec should be split in the middle 200 * @q: [in] request queue associated with the bio associated with @bv 201 * @bv: [in] bvec to examine 202 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 203 * by the number of segments from @bv that may be appended to that 204 * bio without exceeding @max_segs 205 * @bytes: [in,out] Number of bytes in the bio being built. Incremented 206 * by the number of bytes from @bv that may be appended to that 207 * bio without exceeding @max_bytes 208 * @max_segs: [in] upper bound for *@nsegs 209 * @max_bytes: [in] upper bound for *@bytes 210 * 211 * When splitting a bio, it can happen that a bvec is encountered that is too 212 * big to fit in a single segment and hence that it has to be split in the 213 * middle. This function verifies whether or not that should happen. The value 214 * %true is returned if and only if appending the entire @bv to a bio with 215 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 216 * the block driver. 217 */ 218 static bool bvec_split_segs(const struct request_queue *q, 219 const struct bio_vec *bv, unsigned *nsegs, 220 unsigned *bytes, unsigned max_segs, 221 unsigned max_bytes) 222 { 223 unsigned max_len = min(max_bytes, UINT_MAX) - *bytes; 224 unsigned len = min(bv->bv_len, max_len); 225 unsigned total_len = 0; 226 unsigned seg_size = 0; 227 228 while (len && *nsegs < max_segs) { 229 seg_size = get_max_segment_size(q, bv->bv_page, 230 bv->bv_offset + total_len); 231 seg_size = min(seg_size, len); 232 233 (*nsegs)++; 234 total_len += seg_size; 235 len -= seg_size; 236 237 if ((bv->bv_offset + total_len) & queue_virt_boundary(q)) 238 break; 239 } 240 241 *bytes += total_len; 242 243 /* tell the caller to split the bvec if it is too big to fit */ 244 return len > 0 || bv->bv_len > max_len; 245 } 246 247 /** 248 * bio_split_rw - split a bio in two bios 249 * @bio: [in] bio to be split 250 * @q: [in] request queue pointer 251 * @segs: [out] number of segments in the bio with the first half of the sectors 252 * @bs: [in] bio set to allocate the clone from 253 * @max_bytes: [in] maximum number of bytes per bio 254 * 255 * Clone @bio, update the bi_iter of the clone to represent the first sectors 256 * of @bio and update @bio->bi_iter to represent the remaining sectors. The 257 * following is guaranteed for the cloned bio: 258 * - That it has at most @max_bytes worth of data 259 * - That it has at most queue_max_segments(@q) segments. 260 * 261 * Except for discard requests the cloned bio will point at the bi_io_vec of 262 * the original bio. It is the responsibility of the caller to ensure that the 263 * original bio is not freed before the cloned bio. The caller is also 264 * responsible for ensuring that @bs is only destroyed after processing of the 265 * split bio has finished. 266 */ 267 static struct bio *bio_split_rw(struct bio *bio, struct request_queue *q, 268 unsigned *segs, struct bio_set *bs, unsigned max_bytes) 269 { 270 struct bio_vec bv, bvprv, *bvprvp = NULL; 271 struct bvec_iter iter; 272 unsigned nsegs = 0, bytes = 0; 273 const unsigned max_segs = queue_max_segments(q); 274 275 bio_for_each_bvec(bv, bio, iter) { 276 /* 277 * If the queue doesn't support SG gaps and adding this 278 * offset would create a gap, disallow it. 279 */ 280 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) 281 goto split; 282 283 if (nsegs < max_segs && 284 bytes + bv.bv_len <= max_bytes && 285 bv.bv_offset + bv.bv_len <= PAGE_SIZE) { 286 nsegs++; 287 bytes += bv.bv_len; 288 } else if (bvec_split_segs(q, &bv, &nsegs, &bytes, max_segs, 289 max_bytes)) { 290 goto split; 291 } 292 293 bvprv = bv; 294 bvprvp = &bvprv; 295 } 296 297 *segs = nsegs; 298 return NULL; 299 split: 300 *segs = nsegs; 301 302 /* 303 * Individual bvecs might not be logical block aligned. Round down the 304 * split size so that each bio is properly block size aligned, even if 305 * we do not use the full hardware limits. 306 */ 307 bytes = ALIGN_DOWN(bytes, queue_logical_block_size(q)); 308 309 /* 310 * Bio splitting may cause subtle trouble such as hang when doing sync 311 * iopoll in direct IO routine. Given performance gain of iopoll for 312 * big IO can be trival, disable iopoll when split needed. 313 */ 314 bio_clear_polled(bio); 315 return bio_split(bio, bytes >> SECTOR_SHIFT, GFP_NOIO, bs); 316 } 317 318 /** 319 * __bio_split_to_limits - split a bio to fit the queue limits 320 * @bio: bio to be split 321 * @q: request_queue new bio is being queued at 322 * @nr_segs: returns the number of segments in the returned bio 323 * 324 * Check if @bio needs splitting based on the queue limits, and if so split off 325 * a bio fitting the limits from the beginning of @bio and return it. @bio is 326 * shortened to the remainder and re-submitted. 327 * 328 * The split bio is allocated from @q->bio_split, which is provided by the 329 * block layer. 330 */ 331 struct bio *__bio_split_to_limits(struct bio *bio, struct request_queue *q, 332 unsigned int *nr_segs) 333 { 334 struct bio_set *bs = &bio->bi_bdev->bd_disk->bio_split; 335 struct bio *split; 336 337 switch (bio_op(bio)) { 338 case REQ_OP_DISCARD: 339 case REQ_OP_SECURE_ERASE: 340 split = bio_split_discard(bio, q, nr_segs, bs); 341 break; 342 case REQ_OP_WRITE_ZEROES: 343 split = bio_split_write_zeroes(bio, q, nr_segs, bs); 344 break; 345 default: 346 split = bio_split_rw(bio, q, nr_segs, bs, 347 get_max_io_size(bio, q) << SECTOR_SHIFT); 348 break; 349 } 350 351 if (split) { 352 /* there isn't chance to merge the splitted bio */ 353 split->bi_opf |= REQ_NOMERGE; 354 355 blkcg_bio_issue_init(split); 356 bio_chain(split, bio); 357 trace_block_split(split, bio->bi_iter.bi_sector); 358 submit_bio_noacct(bio); 359 return split; 360 } 361 return bio; 362 } 363 364 /** 365 * bio_split_to_limits - split a bio to fit the queue limits 366 * @bio: bio to be split 367 * 368 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and 369 * if so split off a bio fitting the limits from the beginning of @bio and 370 * return it. @bio is shortened to the remainder and re-submitted. 371 * 372 * The split bio is allocated from @q->bio_split, which is provided by the 373 * block layer. 374 */ 375 struct bio *bio_split_to_limits(struct bio *bio) 376 { 377 struct request_queue *q = bdev_get_queue(bio->bi_bdev); 378 unsigned int nr_segs; 379 380 if (bio_may_exceed_limits(bio, q)) 381 return __bio_split_to_limits(bio, q, &nr_segs); 382 return bio; 383 } 384 EXPORT_SYMBOL(bio_split_to_limits); 385 386 unsigned int blk_recalc_rq_segments(struct request *rq) 387 { 388 unsigned int nr_phys_segs = 0; 389 unsigned int bytes = 0; 390 struct req_iterator iter; 391 struct bio_vec bv; 392 393 if (!rq->bio) 394 return 0; 395 396 switch (bio_op(rq->bio)) { 397 case REQ_OP_DISCARD: 398 case REQ_OP_SECURE_ERASE: 399 if (queue_max_discard_segments(rq->q) > 1) { 400 struct bio *bio = rq->bio; 401 402 for_each_bio(bio) 403 nr_phys_segs++; 404 return nr_phys_segs; 405 } 406 return 1; 407 case REQ_OP_WRITE_ZEROES: 408 return 0; 409 default: 410 break; 411 } 412 413 rq_for_each_bvec(bv, rq, iter) 414 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &bytes, 415 UINT_MAX, UINT_MAX); 416 return nr_phys_segs; 417 } 418 419 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 420 struct scatterlist *sglist) 421 { 422 if (!*sg) 423 return sglist; 424 425 /* 426 * If the driver previously mapped a shorter list, we could see a 427 * termination bit prematurely unless it fully inits the sg table 428 * on each mapping. We KNOW that there must be more entries here 429 * or the driver would be buggy, so force clear the termination bit 430 * to avoid doing a full sg_init_table() in drivers for each command. 431 */ 432 sg_unmark_end(*sg); 433 return sg_next(*sg); 434 } 435 436 static unsigned blk_bvec_map_sg(struct request_queue *q, 437 struct bio_vec *bvec, struct scatterlist *sglist, 438 struct scatterlist **sg) 439 { 440 unsigned nbytes = bvec->bv_len; 441 unsigned nsegs = 0, total = 0; 442 443 while (nbytes > 0) { 444 unsigned offset = bvec->bv_offset + total; 445 unsigned len = min(get_max_segment_size(q, bvec->bv_page, 446 offset), nbytes); 447 struct page *page = bvec->bv_page; 448 449 /* 450 * Unfortunately a fair number of drivers barf on scatterlists 451 * that have an offset larger than PAGE_SIZE, despite other 452 * subsystems dealing with that invariant just fine. For now 453 * stick to the legacy format where we never present those from 454 * the block layer, but the code below should be removed once 455 * these offenders (mostly MMC/SD drivers) are fixed. 456 */ 457 page += (offset >> PAGE_SHIFT); 458 offset &= ~PAGE_MASK; 459 460 *sg = blk_next_sg(sg, sglist); 461 sg_set_page(*sg, page, len, offset); 462 463 total += len; 464 nbytes -= len; 465 nsegs++; 466 } 467 468 return nsegs; 469 } 470 471 static inline int __blk_bvec_map_sg(struct bio_vec bv, 472 struct scatterlist *sglist, struct scatterlist **sg) 473 { 474 *sg = blk_next_sg(sg, sglist); 475 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 476 return 1; 477 } 478 479 /* only try to merge bvecs into one sg if they are from two bios */ 480 static inline bool 481 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec, 482 struct bio_vec *bvprv, struct scatterlist **sg) 483 { 484 485 int nbytes = bvec->bv_len; 486 487 if (!*sg) 488 return false; 489 490 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 491 return false; 492 493 if (!biovec_phys_mergeable(q, bvprv, bvec)) 494 return false; 495 496 (*sg)->length += nbytes; 497 498 return true; 499 } 500 501 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 502 struct scatterlist *sglist, 503 struct scatterlist **sg) 504 { 505 struct bio_vec bvec, bvprv = { NULL }; 506 struct bvec_iter iter; 507 int nsegs = 0; 508 bool new_bio = false; 509 510 for_each_bio(bio) { 511 bio_for_each_bvec(bvec, bio, iter) { 512 /* 513 * Only try to merge bvecs from two bios given we 514 * have done bio internal merge when adding pages 515 * to bio 516 */ 517 if (new_bio && 518 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg)) 519 goto next_bvec; 520 521 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE) 522 nsegs += __blk_bvec_map_sg(bvec, sglist, sg); 523 else 524 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg); 525 next_bvec: 526 new_bio = false; 527 } 528 if (likely(bio->bi_iter.bi_size)) { 529 bvprv = bvec; 530 new_bio = true; 531 } 532 } 533 534 return nsegs; 535 } 536 537 /* 538 * map a request to scatterlist, return number of sg entries setup. Caller 539 * must make sure sg can hold rq->nr_phys_segments entries 540 */ 541 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 542 struct scatterlist *sglist, struct scatterlist **last_sg) 543 { 544 int nsegs = 0; 545 546 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 547 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg); 548 else if (rq->bio) 549 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg); 550 551 if (*last_sg) 552 sg_mark_end(*last_sg); 553 554 /* 555 * Something must have been wrong if the figured number of 556 * segment is bigger than number of req's physical segments 557 */ 558 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 559 560 return nsegs; 561 } 562 EXPORT_SYMBOL(__blk_rq_map_sg); 563 564 static inline unsigned int blk_rq_get_max_segments(struct request *rq) 565 { 566 if (req_op(rq) == REQ_OP_DISCARD) 567 return queue_max_discard_segments(rq->q); 568 return queue_max_segments(rq->q); 569 } 570 571 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 572 sector_t offset) 573 { 574 struct request_queue *q = rq->q; 575 unsigned int max_sectors; 576 577 if (blk_rq_is_passthrough(rq)) 578 return q->limits.max_hw_sectors; 579 580 max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); 581 if (!q->limits.chunk_sectors || 582 req_op(rq) == REQ_OP_DISCARD || 583 req_op(rq) == REQ_OP_SECURE_ERASE) 584 return max_sectors; 585 return min(max_sectors, 586 blk_chunk_sectors_left(offset, q->limits.chunk_sectors)); 587 } 588 589 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 590 unsigned int nr_phys_segs) 591 { 592 if (!blk_cgroup_mergeable(req, bio)) 593 goto no_merge; 594 595 if (blk_integrity_merge_bio(req->q, req, bio) == false) 596 goto no_merge; 597 598 /* discard request merge won't add new segment */ 599 if (req_op(req) == REQ_OP_DISCARD) 600 return 1; 601 602 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 603 goto no_merge; 604 605 /* 606 * This will form the start of a new hw segment. Bump both 607 * counters. 608 */ 609 req->nr_phys_segments += nr_phys_segs; 610 return 1; 611 612 no_merge: 613 req_set_nomerge(req->q, req); 614 return 0; 615 } 616 617 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 618 { 619 if (req_gap_back_merge(req, bio)) 620 return 0; 621 if (blk_integrity_rq(req) && 622 integrity_req_gap_back_merge(req, bio)) 623 return 0; 624 if (!bio_crypt_ctx_back_mergeable(req, bio)) 625 return 0; 626 if (blk_rq_sectors(req) + bio_sectors(bio) > 627 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 628 req_set_nomerge(req->q, req); 629 return 0; 630 } 631 632 return ll_new_hw_segment(req, bio, nr_segs); 633 } 634 635 static int ll_front_merge_fn(struct request *req, struct bio *bio, 636 unsigned int nr_segs) 637 { 638 if (req_gap_front_merge(req, bio)) 639 return 0; 640 if (blk_integrity_rq(req) && 641 integrity_req_gap_front_merge(req, bio)) 642 return 0; 643 if (!bio_crypt_ctx_front_mergeable(req, bio)) 644 return 0; 645 if (blk_rq_sectors(req) + bio_sectors(bio) > 646 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 647 req_set_nomerge(req->q, req); 648 return 0; 649 } 650 651 return ll_new_hw_segment(req, bio, nr_segs); 652 } 653 654 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 655 struct request *next) 656 { 657 unsigned short segments = blk_rq_nr_discard_segments(req); 658 659 if (segments >= queue_max_discard_segments(q)) 660 goto no_merge; 661 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 662 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 663 goto no_merge; 664 665 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 666 return true; 667 no_merge: 668 req_set_nomerge(q, req); 669 return false; 670 } 671 672 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 673 struct request *next) 674 { 675 int total_phys_segments; 676 677 if (req_gap_back_merge(req, next->bio)) 678 return 0; 679 680 /* 681 * Will it become too large? 682 */ 683 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 684 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 685 return 0; 686 687 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 688 if (total_phys_segments > blk_rq_get_max_segments(req)) 689 return 0; 690 691 if (!blk_cgroup_mergeable(req, next->bio)) 692 return 0; 693 694 if (blk_integrity_merge_rq(q, req, next) == false) 695 return 0; 696 697 if (!bio_crypt_ctx_merge_rq(req, next)) 698 return 0; 699 700 /* Merge is OK... */ 701 req->nr_phys_segments = total_phys_segments; 702 return 1; 703 } 704 705 /** 706 * blk_rq_set_mixed_merge - mark a request as mixed merge 707 * @rq: request to mark as mixed merge 708 * 709 * Description: 710 * @rq is about to be mixed merged. Make sure the attributes 711 * which can be mixed are set in each bio and mark @rq as mixed 712 * merged. 713 */ 714 void blk_rq_set_mixed_merge(struct request *rq) 715 { 716 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK; 717 struct bio *bio; 718 719 if (rq->rq_flags & RQF_MIXED_MERGE) 720 return; 721 722 /* 723 * @rq will no longer represent mixable attributes for all the 724 * contained bios. It will just track those of the first one. 725 * Distributes the attributs to each bio. 726 */ 727 for (bio = rq->bio; bio; bio = bio->bi_next) { 728 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 729 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 730 bio->bi_opf |= ff; 731 } 732 rq->rq_flags |= RQF_MIXED_MERGE; 733 } 734 735 static void blk_account_io_merge_request(struct request *req) 736 { 737 if (blk_do_io_stat(req)) { 738 part_stat_lock(); 739 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 740 part_stat_unlock(); 741 } 742 } 743 744 static enum elv_merge blk_try_req_merge(struct request *req, 745 struct request *next) 746 { 747 if (blk_discard_mergable(req)) 748 return ELEVATOR_DISCARD_MERGE; 749 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 750 return ELEVATOR_BACK_MERGE; 751 752 return ELEVATOR_NO_MERGE; 753 } 754 755 /* 756 * For non-mq, this has to be called with the request spinlock acquired. 757 * For mq with scheduling, the appropriate queue wide lock should be held. 758 */ 759 static struct request *attempt_merge(struct request_queue *q, 760 struct request *req, struct request *next) 761 { 762 if (!rq_mergeable(req) || !rq_mergeable(next)) 763 return NULL; 764 765 if (req_op(req) != req_op(next)) 766 return NULL; 767 768 if (rq_data_dir(req) != rq_data_dir(next)) 769 return NULL; 770 771 if (req->ioprio != next->ioprio) 772 return NULL; 773 774 /* 775 * If we are allowed to merge, then append bio list 776 * from next to rq and release next. merge_requests_fn 777 * will have updated segment counts, update sector 778 * counts here. Handle DISCARDs separately, as they 779 * have separate settings. 780 */ 781 782 switch (blk_try_req_merge(req, next)) { 783 case ELEVATOR_DISCARD_MERGE: 784 if (!req_attempt_discard_merge(q, req, next)) 785 return NULL; 786 break; 787 case ELEVATOR_BACK_MERGE: 788 if (!ll_merge_requests_fn(q, req, next)) 789 return NULL; 790 break; 791 default: 792 return NULL; 793 } 794 795 /* 796 * If failfast settings disagree or any of the two is already 797 * a mixed merge, mark both as mixed before proceeding. This 798 * makes sure that all involved bios have mixable attributes 799 * set properly. 800 */ 801 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 802 (req->cmd_flags & REQ_FAILFAST_MASK) != 803 (next->cmd_flags & REQ_FAILFAST_MASK)) { 804 blk_rq_set_mixed_merge(req); 805 blk_rq_set_mixed_merge(next); 806 } 807 808 /* 809 * At this point we have either done a back merge or front merge. We 810 * need the smaller start_time_ns of the merged requests to be the 811 * current request for accounting purposes. 812 */ 813 if (next->start_time_ns < req->start_time_ns) 814 req->start_time_ns = next->start_time_ns; 815 816 req->biotail->bi_next = next->bio; 817 req->biotail = next->biotail; 818 819 req->__data_len += blk_rq_bytes(next); 820 821 if (!blk_discard_mergable(req)) 822 elv_merge_requests(q, req, next); 823 824 /* 825 * 'next' is going away, so update stats accordingly 826 */ 827 blk_account_io_merge_request(next); 828 829 trace_block_rq_merge(next); 830 831 /* 832 * ownership of bio passed from next to req, return 'next' for 833 * the caller to free 834 */ 835 next->bio = NULL; 836 return next; 837 } 838 839 static struct request *attempt_back_merge(struct request_queue *q, 840 struct request *rq) 841 { 842 struct request *next = elv_latter_request(q, rq); 843 844 if (next) 845 return attempt_merge(q, rq, next); 846 847 return NULL; 848 } 849 850 static struct request *attempt_front_merge(struct request_queue *q, 851 struct request *rq) 852 { 853 struct request *prev = elv_former_request(q, rq); 854 855 if (prev) 856 return attempt_merge(q, prev, rq); 857 858 return NULL; 859 } 860 861 /* 862 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 863 * otherwise. The caller is responsible for freeing 'next' if the merge 864 * happened. 865 */ 866 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 867 struct request *next) 868 { 869 return attempt_merge(q, rq, next); 870 } 871 872 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 873 { 874 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 875 return false; 876 877 if (req_op(rq) != bio_op(bio)) 878 return false; 879 880 /* different data direction or already started, don't merge */ 881 if (bio_data_dir(bio) != rq_data_dir(rq)) 882 return false; 883 884 /* don't merge across cgroup boundaries */ 885 if (!blk_cgroup_mergeable(rq, bio)) 886 return false; 887 888 /* only merge integrity protected bio into ditto rq */ 889 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 890 return false; 891 892 /* Only merge if the crypt contexts are compatible */ 893 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 894 return false; 895 896 if (rq->ioprio != bio_prio(bio)) 897 return false; 898 899 return true; 900 } 901 902 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 903 { 904 if (blk_discard_mergable(rq)) 905 return ELEVATOR_DISCARD_MERGE; 906 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 907 return ELEVATOR_BACK_MERGE; 908 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 909 return ELEVATOR_FRONT_MERGE; 910 return ELEVATOR_NO_MERGE; 911 } 912 913 static void blk_account_io_merge_bio(struct request *req) 914 { 915 if (!blk_do_io_stat(req)) 916 return; 917 918 part_stat_lock(); 919 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 920 part_stat_unlock(); 921 } 922 923 enum bio_merge_status { 924 BIO_MERGE_OK, 925 BIO_MERGE_NONE, 926 BIO_MERGE_FAILED, 927 }; 928 929 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 930 struct bio *bio, unsigned int nr_segs) 931 { 932 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 933 934 if (!ll_back_merge_fn(req, bio, nr_segs)) 935 return BIO_MERGE_FAILED; 936 937 trace_block_bio_backmerge(bio); 938 rq_qos_merge(req->q, req, bio); 939 940 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 941 blk_rq_set_mixed_merge(req); 942 943 req->biotail->bi_next = bio; 944 req->biotail = bio; 945 req->__data_len += bio->bi_iter.bi_size; 946 947 bio_crypt_free_ctx(bio); 948 949 blk_account_io_merge_bio(req); 950 return BIO_MERGE_OK; 951 } 952 953 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 954 struct bio *bio, unsigned int nr_segs) 955 { 956 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 957 958 if (!ll_front_merge_fn(req, bio, nr_segs)) 959 return BIO_MERGE_FAILED; 960 961 trace_block_bio_frontmerge(bio); 962 rq_qos_merge(req->q, req, bio); 963 964 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 965 blk_rq_set_mixed_merge(req); 966 967 bio->bi_next = req->bio; 968 req->bio = bio; 969 970 req->__sector = bio->bi_iter.bi_sector; 971 req->__data_len += bio->bi_iter.bi_size; 972 973 bio_crypt_do_front_merge(req, bio); 974 975 blk_account_io_merge_bio(req); 976 return BIO_MERGE_OK; 977 } 978 979 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 980 struct request *req, struct bio *bio) 981 { 982 unsigned short segments = blk_rq_nr_discard_segments(req); 983 984 if (segments >= queue_max_discard_segments(q)) 985 goto no_merge; 986 if (blk_rq_sectors(req) + bio_sectors(bio) > 987 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 988 goto no_merge; 989 990 rq_qos_merge(q, req, bio); 991 992 req->biotail->bi_next = bio; 993 req->biotail = bio; 994 req->__data_len += bio->bi_iter.bi_size; 995 req->nr_phys_segments = segments + 1; 996 997 blk_account_io_merge_bio(req); 998 return BIO_MERGE_OK; 999 no_merge: 1000 req_set_nomerge(q, req); 1001 return BIO_MERGE_FAILED; 1002 } 1003 1004 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1005 struct request *rq, 1006 struct bio *bio, 1007 unsigned int nr_segs, 1008 bool sched_allow_merge) 1009 { 1010 if (!blk_rq_merge_ok(rq, bio)) 1011 return BIO_MERGE_NONE; 1012 1013 switch (blk_try_merge(rq, bio)) { 1014 case ELEVATOR_BACK_MERGE: 1015 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1016 return bio_attempt_back_merge(rq, bio, nr_segs); 1017 break; 1018 case ELEVATOR_FRONT_MERGE: 1019 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1020 return bio_attempt_front_merge(rq, bio, nr_segs); 1021 break; 1022 case ELEVATOR_DISCARD_MERGE: 1023 return bio_attempt_discard_merge(q, rq, bio); 1024 default: 1025 return BIO_MERGE_NONE; 1026 } 1027 1028 return BIO_MERGE_FAILED; 1029 } 1030 1031 /** 1032 * blk_attempt_plug_merge - try to merge with %current's plugged list 1033 * @q: request_queue new bio is being queued at 1034 * @bio: new bio being queued 1035 * @nr_segs: number of segments in @bio 1036 * from the passed in @q already in the plug list 1037 * 1038 * Determine whether @bio being queued on @q can be merged with the previous 1039 * request on %current's plugged list. Returns %true if merge was successful, 1040 * otherwise %false. 1041 * 1042 * Plugging coalesces IOs from the same issuer for the same purpose without 1043 * going through @q->queue_lock. As such it's more of an issuing mechanism 1044 * than scheduling, and the request, while may have elvpriv data, is not 1045 * added on the elevator at this point. In addition, we don't have 1046 * reliable access to the elevator outside queue lock. Only check basic 1047 * merging parameters without querying the elevator. 1048 * 1049 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1050 */ 1051 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1052 unsigned int nr_segs) 1053 { 1054 struct blk_plug *plug; 1055 struct request *rq; 1056 1057 plug = blk_mq_plug(bio); 1058 if (!plug || rq_list_empty(plug->mq_list)) 1059 return false; 1060 1061 rq_list_for_each(&plug->mq_list, rq) { 1062 if (rq->q == q) { 1063 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1064 BIO_MERGE_OK) 1065 return true; 1066 break; 1067 } 1068 1069 /* 1070 * Only keep iterating plug list for merges if we have multiple 1071 * queues 1072 */ 1073 if (!plug->multiple_queues) 1074 break; 1075 } 1076 return false; 1077 } 1078 1079 /* 1080 * Iterate list of requests and see if we can merge this bio with any 1081 * of them. 1082 */ 1083 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1084 struct bio *bio, unsigned int nr_segs) 1085 { 1086 struct request *rq; 1087 int checked = 8; 1088 1089 list_for_each_entry_reverse(rq, list, queuelist) { 1090 if (!checked--) 1091 break; 1092 1093 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1094 case BIO_MERGE_NONE: 1095 continue; 1096 case BIO_MERGE_OK: 1097 return true; 1098 case BIO_MERGE_FAILED: 1099 return false; 1100 } 1101 1102 } 1103 1104 return false; 1105 } 1106 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1107 1108 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1109 unsigned int nr_segs, struct request **merged_request) 1110 { 1111 struct request *rq; 1112 1113 switch (elv_merge(q, &rq, bio)) { 1114 case ELEVATOR_BACK_MERGE: 1115 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1116 return false; 1117 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1118 return false; 1119 *merged_request = attempt_back_merge(q, rq); 1120 if (!*merged_request) 1121 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1122 return true; 1123 case ELEVATOR_FRONT_MERGE: 1124 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1125 return false; 1126 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1127 return false; 1128 *merged_request = attempt_front_merge(q, rq); 1129 if (!*merged_request) 1130 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1131 return true; 1132 case ELEVATOR_DISCARD_MERGE: 1133 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1134 default: 1135 return false; 1136 } 1137 } 1138 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1139