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->limits, &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 /* 99 * The max size one bio can handle is UINT_MAX becasue bvec_iter.bi_size 100 * is defined as 'unsigned int', meantime it has to be aligned to with the 101 * logical block size, which is the minimum accepted unit by hardware. 102 */ 103 static unsigned int bio_allowed_max_sectors(const struct queue_limits *lim) 104 { 105 return round_down(UINT_MAX, lim->logical_block_size) >> SECTOR_SHIFT; 106 } 107 108 static struct bio *bio_split_discard(struct bio *bio, 109 const struct queue_limits *lim, 110 unsigned *nsegs, struct bio_set *bs) 111 { 112 unsigned int max_discard_sectors, granularity; 113 sector_t tmp; 114 unsigned split_sectors; 115 116 *nsegs = 1; 117 118 /* Zero-sector (unknown) and one-sector granularities are the same. */ 119 granularity = max(lim->discard_granularity >> 9, 1U); 120 121 max_discard_sectors = 122 min(lim->max_discard_sectors, bio_allowed_max_sectors(lim)); 123 max_discard_sectors -= max_discard_sectors % granularity; 124 125 if (unlikely(!max_discard_sectors)) { 126 /* XXX: warn */ 127 return NULL; 128 } 129 130 if (bio_sectors(bio) <= max_discard_sectors) 131 return NULL; 132 133 split_sectors = max_discard_sectors; 134 135 /* 136 * If the next starting sector would be misaligned, stop the discard at 137 * the previous aligned sector. 138 */ 139 tmp = bio->bi_iter.bi_sector + split_sectors - 140 ((lim->discard_alignment >> 9) % granularity); 141 tmp = sector_div(tmp, granularity); 142 143 if (split_sectors > tmp) 144 split_sectors -= tmp; 145 146 return bio_split(bio, split_sectors, GFP_NOIO, bs); 147 } 148 149 static struct bio *bio_split_write_zeroes(struct bio *bio, 150 const struct queue_limits *lim, 151 unsigned *nsegs, struct bio_set *bs) 152 { 153 *nsegs = 0; 154 if (!lim->max_write_zeroes_sectors) 155 return NULL; 156 if (bio_sectors(bio) <= lim->max_write_zeroes_sectors) 157 return NULL; 158 return bio_split(bio, lim->max_write_zeroes_sectors, GFP_NOIO, bs); 159 } 160 161 /* 162 * Return the maximum number of sectors from the start of a bio that may be 163 * submitted as a single request to a block device. If enough sectors remain, 164 * align the end to the physical block size. Otherwise align the end to the 165 * logical block size. This approach minimizes the number of non-aligned 166 * requests that are submitted to a block device if the start of a bio is not 167 * aligned to a physical block boundary. 168 */ 169 static inline unsigned get_max_io_size(struct bio *bio, 170 const struct queue_limits *lim) 171 { 172 unsigned pbs = lim->physical_block_size >> SECTOR_SHIFT; 173 unsigned lbs = lim->logical_block_size >> SECTOR_SHIFT; 174 unsigned max_sectors = lim->max_sectors, start, end; 175 176 if (lim->chunk_sectors) { 177 max_sectors = min(max_sectors, 178 blk_chunk_sectors_left(bio->bi_iter.bi_sector, 179 lim->chunk_sectors)); 180 } 181 182 start = bio->bi_iter.bi_sector & (pbs - 1); 183 end = (start + max_sectors) & ~(pbs - 1); 184 if (end > start) 185 return end - start; 186 return max_sectors & ~(lbs - 1); 187 } 188 189 static inline unsigned get_max_segment_size(const struct queue_limits *lim, 190 struct page *start_page, unsigned long offset) 191 { 192 unsigned long mask = lim->seg_boundary_mask; 193 194 offset = mask & (page_to_phys(start_page) + offset); 195 196 /* 197 * overflow may be triggered in case of zero page physical address 198 * on 32bit arch, use queue's max segment size when that happens. 199 */ 200 return min_not_zero(mask - offset + 1, 201 (unsigned long)lim->max_segment_size); 202 } 203 204 /** 205 * bvec_split_segs - verify whether or not a bvec should be split in the middle 206 * @lim: [in] queue limits to split based on 207 * @bv: [in] bvec to examine 208 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 209 * by the number of segments from @bv that may be appended to that 210 * bio without exceeding @max_segs 211 * @bytes: [in,out] Number of bytes in the bio being built. Incremented 212 * by the number of bytes from @bv that may be appended to that 213 * bio without exceeding @max_bytes 214 * @max_segs: [in] upper bound for *@nsegs 215 * @max_bytes: [in] upper bound for *@bytes 216 * 217 * When splitting a bio, it can happen that a bvec is encountered that is too 218 * big to fit in a single segment and hence that it has to be split in the 219 * middle. This function verifies whether or not that should happen. The value 220 * %true is returned if and only if appending the entire @bv to a bio with 221 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 222 * the block driver. 223 */ 224 static bool bvec_split_segs(const struct queue_limits *lim, 225 const struct bio_vec *bv, unsigned *nsegs, unsigned *bytes, 226 unsigned max_segs, unsigned max_bytes) 227 { 228 unsigned max_len = min(max_bytes, UINT_MAX) - *bytes; 229 unsigned len = min(bv->bv_len, max_len); 230 unsigned total_len = 0; 231 unsigned seg_size = 0; 232 233 while (len && *nsegs < max_segs) { 234 seg_size = get_max_segment_size(lim, bv->bv_page, 235 bv->bv_offset + total_len); 236 seg_size = min(seg_size, len); 237 238 (*nsegs)++; 239 total_len += seg_size; 240 len -= seg_size; 241 242 if ((bv->bv_offset + total_len) & lim->virt_boundary_mask) 243 break; 244 } 245 246 *bytes += total_len; 247 248 /* tell the caller to split the bvec if it is too big to fit */ 249 return len > 0 || bv->bv_len > max_len; 250 } 251 252 /** 253 * bio_split_rw - split a bio in two bios 254 * @bio: [in] bio to be split 255 * @lim: [in] queue limits to split based on 256 * @segs: [out] number of segments in the bio with the first half of the sectors 257 * @bs: [in] bio set to allocate the clone from 258 * @max_bytes: [in] maximum number of bytes per bio 259 * 260 * Clone @bio, update the bi_iter of the clone to represent the first sectors 261 * of @bio and update @bio->bi_iter to represent the remaining sectors. The 262 * following is guaranteed for the cloned bio: 263 * - That it has at most @max_bytes worth of data 264 * - That it has at most queue_max_segments(@q) segments. 265 * 266 * Except for discard requests the cloned bio will point at the bi_io_vec of 267 * the original bio. It is the responsibility of the caller to ensure that the 268 * original bio is not freed before the cloned bio. The caller is also 269 * responsible for ensuring that @bs is only destroyed after processing of the 270 * split bio has finished. 271 */ 272 static struct bio *bio_split_rw(struct bio *bio, const struct queue_limits *lim, 273 unsigned *segs, struct bio_set *bs, unsigned max_bytes) 274 { 275 struct bio_vec bv, bvprv, *bvprvp = NULL; 276 struct bvec_iter iter; 277 unsigned nsegs = 0, bytes = 0; 278 279 bio_for_each_bvec(bv, bio, iter) { 280 /* 281 * If the queue doesn't support SG gaps and adding this 282 * offset would create a gap, disallow it. 283 */ 284 if (bvprvp && bvec_gap_to_prev(lim, bvprvp, bv.bv_offset)) 285 goto split; 286 287 if (nsegs < lim->max_segments && 288 bytes + bv.bv_len <= max_bytes && 289 bv.bv_offset + bv.bv_len <= PAGE_SIZE) { 290 nsegs++; 291 bytes += bv.bv_len; 292 } else { 293 if (bvec_split_segs(lim, &bv, &nsegs, &bytes, 294 lim->max_segments, max_bytes)) 295 goto split; 296 } 297 298 bvprv = bv; 299 bvprvp = &bvprv; 300 } 301 302 *segs = nsegs; 303 return NULL; 304 split: 305 *segs = nsegs; 306 307 /* 308 * Individual bvecs might not be logical block aligned. Round down the 309 * split size so that each bio is properly block size aligned, even if 310 * we do not use the full hardware limits. 311 */ 312 bytes = ALIGN_DOWN(bytes, lim->logical_block_size); 313 314 /* 315 * Bio splitting may cause subtle trouble such as hang when doing sync 316 * iopoll in direct IO routine. Given performance gain of iopoll for 317 * big IO can be trival, disable iopoll when split needed. 318 */ 319 bio_clear_polled(bio); 320 return bio_split(bio, bytes >> SECTOR_SHIFT, GFP_NOIO, bs); 321 } 322 323 /** 324 * __bio_split_to_limits - split a bio to fit the queue limits 325 * @bio: bio to be split 326 * @lim: queue limits to split based on 327 * @nr_segs: returns the number of segments in the returned bio 328 * 329 * Check if @bio needs splitting based on the queue limits, and if so split off 330 * a bio fitting the limits from the beginning of @bio and return it. @bio is 331 * shortened to the remainder and re-submitted. 332 * 333 * The split bio is allocated from @q->bio_split, which is provided by the 334 * block layer. 335 */ 336 struct bio *__bio_split_to_limits(struct bio *bio, 337 const struct queue_limits *lim, 338 unsigned int *nr_segs) 339 { 340 struct bio_set *bs = &bio->bi_bdev->bd_disk->bio_split; 341 struct bio *split; 342 343 switch (bio_op(bio)) { 344 case REQ_OP_DISCARD: 345 case REQ_OP_SECURE_ERASE: 346 split = bio_split_discard(bio, lim, nr_segs, bs); 347 break; 348 case REQ_OP_WRITE_ZEROES: 349 split = bio_split_write_zeroes(bio, lim, nr_segs, bs); 350 break; 351 default: 352 split = bio_split_rw(bio, lim, nr_segs, bs, 353 get_max_io_size(bio, lim) << SECTOR_SHIFT); 354 break; 355 } 356 357 if (split) { 358 /* there isn't chance to merge the splitted bio */ 359 split->bi_opf |= REQ_NOMERGE; 360 361 blkcg_bio_issue_init(split); 362 bio_chain(split, bio); 363 trace_block_split(split, bio->bi_iter.bi_sector); 364 submit_bio_noacct(bio); 365 return split; 366 } 367 return bio; 368 } 369 370 /** 371 * bio_split_to_limits - split a bio to fit the queue limits 372 * @bio: bio to be split 373 * 374 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and 375 * if so split off a bio fitting the limits from the beginning of @bio and 376 * return it. @bio is shortened to the remainder and re-submitted. 377 * 378 * The split bio is allocated from @q->bio_split, which is provided by the 379 * block layer. 380 */ 381 struct bio *bio_split_to_limits(struct bio *bio) 382 { 383 const struct queue_limits *lim = &bdev_get_queue(bio->bi_bdev)->limits; 384 unsigned int nr_segs; 385 386 if (bio_may_exceed_limits(bio, lim)) 387 return __bio_split_to_limits(bio, lim, &nr_segs); 388 return bio; 389 } 390 EXPORT_SYMBOL(bio_split_to_limits); 391 392 unsigned int blk_recalc_rq_segments(struct request *rq) 393 { 394 unsigned int nr_phys_segs = 0; 395 unsigned int bytes = 0; 396 struct req_iterator iter; 397 struct bio_vec bv; 398 399 if (!rq->bio) 400 return 0; 401 402 switch (bio_op(rq->bio)) { 403 case REQ_OP_DISCARD: 404 case REQ_OP_SECURE_ERASE: 405 if (queue_max_discard_segments(rq->q) > 1) { 406 struct bio *bio = rq->bio; 407 408 for_each_bio(bio) 409 nr_phys_segs++; 410 return nr_phys_segs; 411 } 412 return 1; 413 case REQ_OP_WRITE_ZEROES: 414 return 0; 415 default: 416 break; 417 } 418 419 rq_for_each_bvec(bv, rq, iter) 420 bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes, 421 UINT_MAX, UINT_MAX); 422 return nr_phys_segs; 423 } 424 425 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 426 struct scatterlist *sglist) 427 { 428 if (!*sg) 429 return sglist; 430 431 /* 432 * If the driver previously mapped a shorter list, we could see a 433 * termination bit prematurely unless it fully inits the sg table 434 * on each mapping. We KNOW that there must be more entries here 435 * or the driver would be buggy, so force clear the termination bit 436 * to avoid doing a full sg_init_table() in drivers for each command. 437 */ 438 sg_unmark_end(*sg); 439 return sg_next(*sg); 440 } 441 442 static unsigned blk_bvec_map_sg(struct request_queue *q, 443 struct bio_vec *bvec, struct scatterlist *sglist, 444 struct scatterlist **sg) 445 { 446 unsigned nbytes = bvec->bv_len; 447 unsigned nsegs = 0, total = 0; 448 449 while (nbytes > 0) { 450 unsigned offset = bvec->bv_offset + total; 451 unsigned len = min(get_max_segment_size(&q->limits, 452 bvec->bv_page, offset), nbytes); 453 struct page *page = bvec->bv_page; 454 455 /* 456 * Unfortunately a fair number of drivers barf on scatterlists 457 * that have an offset larger than PAGE_SIZE, despite other 458 * subsystems dealing with that invariant just fine. For now 459 * stick to the legacy format where we never present those from 460 * the block layer, but the code below should be removed once 461 * these offenders (mostly MMC/SD drivers) are fixed. 462 */ 463 page += (offset >> PAGE_SHIFT); 464 offset &= ~PAGE_MASK; 465 466 *sg = blk_next_sg(sg, sglist); 467 sg_set_page(*sg, page, len, offset); 468 469 total += len; 470 nbytes -= len; 471 nsegs++; 472 } 473 474 return nsegs; 475 } 476 477 static inline int __blk_bvec_map_sg(struct bio_vec bv, 478 struct scatterlist *sglist, struct scatterlist **sg) 479 { 480 *sg = blk_next_sg(sg, sglist); 481 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 482 return 1; 483 } 484 485 /* only try to merge bvecs into one sg if they are from two bios */ 486 static inline bool 487 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec, 488 struct bio_vec *bvprv, struct scatterlist **sg) 489 { 490 491 int nbytes = bvec->bv_len; 492 493 if (!*sg) 494 return false; 495 496 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 497 return false; 498 499 if (!biovec_phys_mergeable(q, bvprv, bvec)) 500 return false; 501 502 (*sg)->length += nbytes; 503 504 return true; 505 } 506 507 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 508 struct scatterlist *sglist, 509 struct scatterlist **sg) 510 { 511 struct bio_vec bvec, bvprv = { NULL }; 512 struct bvec_iter iter; 513 int nsegs = 0; 514 bool new_bio = false; 515 516 for_each_bio(bio) { 517 bio_for_each_bvec(bvec, bio, iter) { 518 /* 519 * Only try to merge bvecs from two bios given we 520 * have done bio internal merge when adding pages 521 * to bio 522 */ 523 if (new_bio && 524 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg)) 525 goto next_bvec; 526 527 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE) 528 nsegs += __blk_bvec_map_sg(bvec, sglist, sg); 529 else 530 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg); 531 next_bvec: 532 new_bio = false; 533 } 534 if (likely(bio->bi_iter.bi_size)) { 535 bvprv = bvec; 536 new_bio = true; 537 } 538 } 539 540 return nsegs; 541 } 542 543 /* 544 * map a request to scatterlist, return number of sg entries setup. Caller 545 * must make sure sg can hold rq->nr_phys_segments entries 546 */ 547 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 548 struct scatterlist *sglist, struct scatterlist **last_sg) 549 { 550 int nsegs = 0; 551 552 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 553 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg); 554 else if (rq->bio) 555 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg); 556 557 if (*last_sg) 558 sg_mark_end(*last_sg); 559 560 /* 561 * Something must have been wrong if the figured number of 562 * segment is bigger than number of req's physical segments 563 */ 564 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 565 566 return nsegs; 567 } 568 EXPORT_SYMBOL(__blk_rq_map_sg); 569 570 static inline unsigned int blk_rq_get_max_segments(struct request *rq) 571 { 572 if (req_op(rq) == REQ_OP_DISCARD) 573 return queue_max_discard_segments(rq->q); 574 return queue_max_segments(rq->q); 575 } 576 577 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 578 sector_t offset) 579 { 580 struct request_queue *q = rq->q; 581 unsigned int max_sectors; 582 583 if (blk_rq_is_passthrough(rq)) 584 return q->limits.max_hw_sectors; 585 586 max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); 587 if (!q->limits.chunk_sectors || 588 req_op(rq) == REQ_OP_DISCARD || 589 req_op(rq) == REQ_OP_SECURE_ERASE) 590 return max_sectors; 591 return min(max_sectors, 592 blk_chunk_sectors_left(offset, q->limits.chunk_sectors)); 593 } 594 595 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 596 unsigned int nr_phys_segs) 597 { 598 if (!blk_cgroup_mergeable(req, bio)) 599 goto no_merge; 600 601 if (blk_integrity_merge_bio(req->q, req, bio) == false) 602 goto no_merge; 603 604 /* discard request merge won't add new segment */ 605 if (req_op(req) == REQ_OP_DISCARD) 606 return 1; 607 608 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 609 goto no_merge; 610 611 /* 612 * This will form the start of a new hw segment. Bump both 613 * counters. 614 */ 615 req->nr_phys_segments += nr_phys_segs; 616 return 1; 617 618 no_merge: 619 req_set_nomerge(req->q, req); 620 return 0; 621 } 622 623 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 624 { 625 if (req_gap_back_merge(req, bio)) 626 return 0; 627 if (blk_integrity_rq(req) && 628 integrity_req_gap_back_merge(req, bio)) 629 return 0; 630 if (!bio_crypt_ctx_back_mergeable(req, bio)) 631 return 0; 632 if (blk_rq_sectors(req) + bio_sectors(bio) > 633 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 634 req_set_nomerge(req->q, req); 635 return 0; 636 } 637 638 return ll_new_hw_segment(req, bio, nr_segs); 639 } 640 641 static int ll_front_merge_fn(struct request *req, struct bio *bio, 642 unsigned int nr_segs) 643 { 644 if (req_gap_front_merge(req, bio)) 645 return 0; 646 if (blk_integrity_rq(req) && 647 integrity_req_gap_front_merge(req, bio)) 648 return 0; 649 if (!bio_crypt_ctx_front_mergeable(req, bio)) 650 return 0; 651 if (blk_rq_sectors(req) + bio_sectors(bio) > 652 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 653 req_set_nomerge(req->q, req); 654 return 0; 655 } 656 657 return ll_new_hw_segment(req, bio, nr_segs); 658 } 659 660 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 661 struct request *next) 662 { 663 unsigned short segments = blk_rq_nr_discard_segments(req); 664 665 if (segments >= queue_max_discard_segments(q)) 666 goto no_merge; 667 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 668 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 669 goto no_merge; 670 671 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 672 return true; 673 no_merge: 674 req_set_nomerge(q, req); 675 return false; 676 } 677 678 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 679 struct request *next) 680 { 681 int total_phys_segments; 682 683 if (req_gap_back_merge(req, next->bio)) 684 return 0; 685 686 /* 687 * Will it become too large? 688 */ 689 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 690 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 691 return 0; 692 693 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 694 if (total_phys_segments > blk_rq_get_max_segments(req)) 695 return 0; 696 697 if (!blk_cgroup_mergeable(req, next->bio)) 698 return 0; 699 700 if (blk_integrity_merge_rq(q, req, next) == false) 701 return 0; 702 703 if (!bio_crypt_ctx_merge_rq(req, next)) 704 return 0; 705 706 /* Merge is OK... */ 707 req->nr_phys_segments = total_phys_segments; 708 return 1; 709 } 710 711 /** 712 * blk_rq_set_mixed_merge - mark a request as mixed merge 713 * @rq: request to mark as mixed merge 714 * 715 * Description: 716 * @rq is about to be mixed merged. Make sure the attributes 717 * which can be mixed are set in each bio and mark @rq as mixed 718 * merged. 719 */ 720 void blk_rq_set_mixed_merge(struct request *rq) 721 { 722 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK; 723 struct bio *bio; 724 725 if (rq->rq_flags & RQF_MIXED_MERGE) 726 return; 727 728 /* 729 * @rq will no longer represent mixable attributes for all the 730 * contained bios. It will just track those of the first one. 731 * Distributes the attributs to each bio. 732 */ 733 for (bio = rq->bio; bio; bio = bio->bi_next) { 734 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 735 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 736 bio->bi_opf |= ff; 737 } 738 rq->rq_flags |= RQF_MIXED_MERGE; 739 } 740 741 static void blk_account_io_merge_request(struct request *req) 742 { 743 if (blk_do_io_stat(req)) { 744 part_stat_lock(); 745 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 746 part_stat_unlock(); 747 } 748 } 749 750 static enum elv_merge blk_try_req_merge(struct request *req, 751 struct request *next) 752 { 753 if (blk_discard_mergable(req)) 754 return ELEVATOR_DISCARD_MERGE; 755 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 756 return ELEVATOR_BACK_MERGE; 757 758 return ELEVATOR_NO_MERGE; 759 } 760 761 /* 762 * For non-mq, this has to be called with the request spinlock acquired. 763 * For mq with scheduling, the appropriate queue wide lock should be held. 764 */ 765 static struct request *attempt_merge(struct request_queue *q, 766 struct request *req, struct request *next) 767 { 768 if (!rq_mergeable(req) || !rq_mergeable(next)) 769 return NULL; 770 771 if (req_op(req) != req_op(next)) 772 return NULL; 773 774 if (rq_data_dir(req) != rq_data_dir(next)) 775 return NULL; 776 777 if (req->ioprio != next->ioprio) 778 return NULL; 779 780 /* 781 * If we are allowed to merge, then append bio list 782 * from next to rq and release next. merge_requests_fn 783 * will have updated segment counts, update sector 784 * counts here. Handle DISCARDs separately, as they 785 * have separate settings. 786 */ 787 788 switch (blk_try_req_merge(req, next)) { 789 case ELEVATOR_DISCARD_MERGE: 790 if (!req_attempt_discard_merge(q, req, next)) 791 return NULL; 792 break; 793 case ELEVATOR_BACK_MERGE: 794 if (!ll_merge_requests_fn(q, req, next)) 795 return NULL; 796 break; 797 default: 798 return NULL; 799 } 800 801 /* 802 * If failfast settings disagree or any of the two is already 803 * a mixed merge, mark both as mixed before proceeding. This 804 * makes sure that all involved bios have mixable attributes 805 * set properly. 806 */ 807 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 808 (req->cmd_flags & REQ_FAILFAST_MASK) != 809 (next->cmd_flags & REQ_FAILFAST_MASK)) { 810 blk_rq_set_mixed_merge(req); 811 blk_rq_set_mixed_merge(next); 812 } 813 814 /* 815 * At this point we have either done a back merge or front merge. We 816 * need the smaller start_time_ns of the merged requests to be the 817 * current request for accounting purposes. 818 */ 819 if (next->start_time_ns < req->start_time_ns) 820 req->start_time_ns = next->start_time_ns; 821 822 req->biotail->bi_next = next->bio; 823 req->biotail = next->biotail; 824 825 req->__data_len += blk_rq_bytes(next); 826 827 if (!blk_discard_mergable(req)) 828 elv_merge_requests(q, req, next); 829 830 /* 831 * 'next' is going away, so update stats accordingly 832 */ 833 blk_account_io_merge_request(next); 834 835 trace_block_rq_merge(next); 836 837 /* 838 * ownership of bio passed from next to req, return 'next' for 839 * the caller to free 840 */ 841 next->bio = NULL; 842 return next; 843 } 844 845 static struct request *attempt_back_merge(struct request_queue *q, 846 struct request *rq) 847 { 848 struct request *next = elv_latter_request(q, rq); 849 850 if (next) 851 return attempt_merge(q, rq, next); 852 853 return NULL; 854 } 855 856 static struct request *attempt_front_merge(struct request_queue *q, 857 struct request *rq) 858 { 859 struct request *prev = elv_former_request(q, rq); 860 861 if (prev) 862 return attempt_merge(q, prev, rq); 863 864 return NULL; 865 } 866 867 /* 868 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 869 * otherwise. The caller is responsible for freeing 'next' if the merge 870 * happened. 871 */ 872 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 873 struct request *next) 874 { 875 return attempt_merge(q, rq, next); 876 } 877 878 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 879 { 880 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 881 return false; 882 883 if (req_op(rq) != bio_op(bio)) 884 return false; 885 886 /* different data direction or already started, don't merge */ 887 if (bio_data_dir(bio) != rq_data_dir(rq)) 888 return false; 889 890 /* don't merge across cgroup boundaries */ 891 if (!blk_cgroup_mergeable(rq, bio)) 892 return false; 893 894 /* only merge integrity protected bio into ditto rq */ 895 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 896 return false; 897 898 /* Only merge if the crypt contexts are compatible */ 899 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 900 return false; 901 902 if (rq->ioprio != bio_prio(bio)) 903 return false; 904 905 return true; 906 } 907 908 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 909 { 910 if (blk_discard_mergable(rq)) 911 return ELEVATOR_DISCARD_MERGE; 912 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 913 return ELEVATOR_BACK_MERGE; 914 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 915 return ELEVATOR_FRONT_MERGE; 916 return ELEVATOR_NO_MERGE; 917 } 918 919 static void blk_account_io_merge_bio(struct request *req) 920 { 921 if (!blk_do_io_stat(req)) 922 return; 923 924 part_stat_lock(); 925 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 926 part_stat_unlock(); 927 } 928 929 enum bio_merge_status { 930 BIO_MERGE_OK, 931 BIO_MERGE_NONE, 932 BIO_MERGE_FAILED, 933 }; 934 935 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 936 struct bio *bio, unsigned int nr_segs) 937 { 938 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 939 940 if (!ll_back_merge_fn(req, bio, nr_segs)) 941 return BIO_MERGE_FAILED; 942 943 trace_block_bio_backmerge(bio); 944 rq_qos_merge(req->q, req, bio); 945 946 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 947 blk_rq_set_mixed_merge(req); 948 949 req->biotail->bi_next = bio; 950 req->biotail = bio; 951 req->__data_len += bio->bi_iter.bi_size; 952 953 bio_crypt_free_ctx(bio); 954 955 blk_account_io_merge_bio(req); 956 return BIO_MERGE_OK; 957 } 958 959 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 960 struct bio *bio, unsigned int nr_segs) 961 { 962 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 963 964 if (!ll_front_merge_fn(req, bio, nr_segs)) 965 return BIO_MERGE_FAILED; 966 967 trace_block_bio_frontmerge(bio); 968 rq_qos_merge(req->q, req, bio); 969 970 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 971 blk_rq_set_mixed_merge(req); 972 973 bio->bi_next = req->bio; 974 req->bio = bio; 975 976 req->__sector = bio->bi_iter.bi_sector; 977 req->__data_len += bio->bi_iter.bi_size; 978 979 bio_crypt_do_front_merge(req, bio); 980 981 blk_account_io_merge_bio(req); 982 return BIO_MERGE_OK; 983 } 984 985 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 986 struct request *req, struct bio *bio) 987 { 988 unsigned short segments = blk_rq_nr_discard_segments(req); 989 990 if (segments >= queue_max_discard_segments(q)) 991 goto no_merge; 992 if (blk_rq_sectors(req) + bio_sectors(bio) > 993 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 994 goto no_merge; 995 996 rq_qos_merge(q, req, bio); 997 998 req->biotail->bi_next = bio; 999 req->biotail = bio; 1000 req->__data_len += bio->bi_iter.bi_size; 1001 req->nr_phys_segments = segments + 1; 1002 1003 blk_account_io_merge_bio(req); 1004 return BIO_MERGE_OK; 1005 no_merge: 1006 req_set_nomerge(q, req); 1007 return BIO_MERGE_FAILED; 1008 } 1009 1010 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1011 struct request *rq, 1012 struct bio *bio, 1013 unsigned int nr_segs, 1014 bool sched_allow_merge) 1015 { 1016 if (!blk_rq_merge_ok(rq, bio)) 1017 return BIO_MERGE_NONE; 1018 1019 switch (blk_try_merge(rq, bio)) { 1020 case ELEVATOR_BACK_MERGE: 1021 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1022 return bio_attempt_back_merge(rq, bio, nr_segs); 1023 break; 1024 case ELEVATOR_FRONT_MERGE: 1025 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1026 return bio_attempt_front_merge(rq, bio, nr_segs); 1027 break; 1028 case ELEVATOR_DISCARD_MERGE: 1029 return bio_attempt_discard_merge(q, rq, bio); 1030 default: 1031 return BIO_MERGE_NONE; 1032 } 1033 1034 return BIO_MERGE_FAILED; 1035 } 1036 1037 /** 1038 * blk_attempt_plug_merge - try to merge with %current's plugged list 1039 * @q: request_queue new bio is being queued at 1040 * @bio: new bio being queued 1041 * @nr_segs: number of segments in @bio 1042 * from the passed in @q already in the plug list 1043 * 1044 * Determine whether @bio being queued on @q can be merged with the previous 1045 * request on %current's plugged list. Returns %true if merge was successful, 1046 * otherwise %false. 1047 * 1048 * Plugging coalesces IOs from the same issuer for the same purpose without 1049 * going through @q->queue_lock. As such it's more of an issuing mechanism 1050 * than scheduling, and the request, while may have elvpriv data, is not 1051 * added on the elevator at this point. In addition, we don't have 1052 * reliable access to the elevator outside queue lock. Only check basic 1053 * merging parameters without querying the elevator. 1054 * 1055 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1056 */ 1057 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1058 unsigned int nr_segs) 1059 { 1060 struct blk_plug *plug; 1061 struct request *rq; 1062 1063 plug = blk_mq_plug(bio); 1064 if (!plug || rq_list_empty(plug->mq_list)) 1065 return false; 1066 1067 rq_list_for_each(&plug->mq_list, rq) { 1068 if (rq->q == q) { 1069 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1070 BIO_MERGE_OK) 1071 return true; 1072 break; 1073 } 1074 1075 /* 1076 * Only keep iterating plug list for merges if we have multiple 1077 * queues 1078 */ 1079 if (!plug->multiple_queues) 1080 break; 1081 } 1082 return false; 1083 } 1084 1085 /* 1086 * Iterate list of requests and see if we can merge this bio with any 1087 * of them. 1088 */ 1089 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1090 struct bio *bio, unsigned int nr_segs) 1091 { 1092 struct request *rq; 1093 int checked = 8; 1094 1095 list_for_each_entry_reverse(rq, list, queuelist) { 1096 if (!checked--) 1097 break; 1098 1099 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1100 case BIO_MERGE_NONE: 1101 continue; 1102 case BIO_MERGE_OK: 1103 return true; 1104 case BIO_MERGE_FAILED: 1105 return false; 1106 } 1107 1108 } 1109 1110 return false; 1111 } 1112 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1113 1114 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1115 unsigned int nr_segs, struct request **merged_request) 1116 { 1117 struct request *rq; 1118 1119 switch (elv_merge(q, &rq, bio)) { 1120 case ELEVATOR_BACK_MERGE: 1121 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1122 return false; 1123 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1124 return false; 1125 *merged_request = attempt_back_merge(q, rq); 1126 if (!*merged_request) 1127 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1128 return true; 1129 case ELEVATOR_FRONT_MERGE: 1130 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1131 return false; 1132 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1133 return false; 1134 *merged_request = attempt_front_merge(q, rq); 1135 if (!*merged_request) 1136 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1137 return true; 1138 case ELEVATOR_DISCARD_MERGE: 1139 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1140 default: 1141 return false; 1142 } 1143 } 1144 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1145