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 /* 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(struct request_queue *q) 104 { 105 return round_down(UINT_MAX, queue_logical_block_size(q)) >> 9; 106 } 107 108 static struct bio *bio_split_discard(struct bio *bio, struct request_queue *q, 109 unsigned *nsegs, struct bio_set *bs) 110 { 111 unsigned int max_discard_sectors, granularity; 112 int alignment; 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(q->limits.discard_granularity >> 9, 1U); 120 121 max_discard_sectors = min(q->limits.max_discard_sectors, 122 bio_allowed_max_sectors(q)); 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 alignment = (q->limits.discard_alignment >> 9) % granularity; 140 141 tmp = bio->bi_iter.bi_sector + split_sectors - alignment; 142 tmp = sector_div(tmp, granularity); 143 144 if (split_sectors > tmp) 145 split_sectors -= tmp; 146 147 return bio_split(bio, split_sectors, GFP_NOIO, bs); 148 } 149 150 static struct bio *bio_split_write_zeroes(struct bio *bio, 151 struct request_queue *q, unsigned *nsegs, struct bio_set *bs) 152 { 153 *nsegs = 0; 154 155 if (!q->limits.max_write_zeroes_sectors) 156 return NULL; 157 158 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) 159 return NULL; 160 161 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); 162 } 163 164 /* 165 * Return the maximum number of sectors from the start of a bio that may be 166 * submitted as a single request to a block device. If enough sectors remain, 167 * align the end to the physical block size. Otherwise align the end to the 168 * logical block size. This approach minimizes the number of non-aligned 169 * requests that are submitted to a block device if the start of a bio is not 170 * aligned to a physical block boundary. 171 */ 172 static inline unsigned get_max_io_size(struct bio *bio, 173 struct request_queue *q) 174 { 175 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT; 176 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT; 177 unsigned max_sectors = queue_max_sectors(q), start, end; 178 179 if (q->limits.chunk_sectors) { 180 max_sectors = min(max_sectors, 181 blk_chunk_sectors_left(bio->bi_iter.bi_sector, 182 q->limits.chunk_sectors)); 183 } 184 185 start = bio->bi_iter.bi_sector & (pbs - 1); 186 end = (start + max_sectors) & ~(pbs - 1); 187 if (end > start) 188 return end - start; 189 return max_sectors & ~(lbs - 1); 190 } 191 192 static inline unsigned get_max_segment_size(const struct request_queue *q, 193 struct page *start_page, 194 unsigned long offset) 195 { 196 unsigned long mask = queue_segment_boundary(q); 197 198 offset = mask & (page_to_phys(start_page) + offset); 199 200 /* 201 * overflow may be triggered in case of zero page physical address 202 * on 32bit arch, use queue's max segment size when that happens. 203 */ 204 return min_not_zero(mask - offset + 1, 205 (unsigned long)queue_max_segment_size(q)); 206 } 207 208 /** 209 * bvec_split_segs - verify whether or not a bvec should be split in the middle 210 * @q: [in] request queue associated with the bio associated with @bv 211 * @bv: [in] bvec to examine 212 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 213 * by the number of segments from @bv that may be appended to that 214 * bio without exceeding @max_segs 215 * @bytes: [in,out] Number of bytes in the bio being built. Incremented 216 * by the number of bytes from @bv that may be appended to that 217 * bio without exceeding @max_bytes 218 * @max_segs: [in] upper bound for *@nsegs 219 * @max_bytes: [in] upper bound for *@bytes 220 * 221 * When splitting a bio, it can happen that a bvec is encountered that is too 222 * big to fit in a single segment and hence that it has to be split in the 223 * middle. This function verifies whether or not that should happen. The value 224 * %true is returned if and only if appending the entire @bv to a bio with 225 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 226 * the block driver. 227 */ 228 static bool bvec_split_segs(const struct request_queue *q, 229 const struct bio_vec *bv, unsigned *nsegs, 230 unsigned *bytes, unsigned max_segs, 231 unsigned max_bytes) 232 { 233 unsigned max_len = min(max_bytes, UINT_MAX) - *bytes; 234 unsigned len = min(bv->bv_len, max_len); 235 unsigned total_len = 0; 236 unsigned seg_size = 0; 237 238 while (len && *nsegs < max_segs) { 239 seg_size = get_max_segment_size(q, bv->bv_page, 240 bv->bv_offset + total_len); 241 seg_size = min(seg_size, len); 242 243 (*nsegs)++; 244 total_len += seg_size; 245 len -= seg_size; 246 247 if ((bv->bv_offset + total_len) & queue_virt_boundary(q)) 248 break; 249 } 250 251 *bytes += total_len; 252 253 /* tell the caller to split the bvec if it is too big to fit */ 254 return len > 0 || bv->bv_len > max_len; 255 } 256 257 /** 258 * bio_split_rw - split a bio in two bios 259 * @bio: [in] bio to be split 260 * @q: [in] request queue pointer 261 * @segs: [out] number of segments in the bio with the first half of the sectors 262 * @bs: [in] bio set to allocate the clone from 263 * @max_bytes: [in] maximum number of bytes per bio 264 * 265 * Clone @bio, update the bi_iter of the clone to represent the first sectors 266 * of @bio and update @bio->bi_iter to represent the remaining sectors. The 267 * following is guaranteed for the cloned bio: 268 * - That it has at most @max_bytes worth of data 269 * - That it has at most queue_max_segments(@q) segments. 270 * 271 * Except for discard requests the cloned bio will point at the bi_io_vec of 272 * the original bio. It is the responsibility of the caller to ensure that the 273 * original bio is not freed before the cloned bio. The caller is also 274 * responsible for ensuring that @bs is only destroyed after processing of the 275 * split bio has finished. 276 */ 277 static struct bio *bio_split_rw(struct bio *bio, struct request_queue *q, 278 unsigned *segs, struct bio_set *bs, unsigned max_bytes) 279 { 280 struct bio_vec bv, bvprv, *bvprvp = NULL; 281 struct bvec_iter iter; 282 unsigned nsegs = 0, bytes = 0; 283 const unsigned max_segs = queue_max_segments(q); 284 285 bio_for_each_bvec(bv, bio, iter) { 286 /* 287 * If the queue doesn't support SG gaps and adding this 288 * offset would create a gap, disallow it. 289 */ 290 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) 291 goto split; 292 293 if (nsegs < max_segs && 294 bytes + bv.bv_len <= max_bytes && 295 bv.bv_offset + bv.bv_len <= PAGE_SIZE) { 296 nsegs++; 297 bytes += bv.bv_len; 298 } else if (bvec_split_segs(q, &bv, &nsegs, &bytes, max_segs, 299 max_bytes)) { 300 goto split; 301 } 302 303 bvprv = bv; 304 bvprvp = &bvprv; 305 } 306 307 *segs = nsegs; 308 return NULL; 309 split: 310 *segs = nsegs; 311 312 /* 313 * Individual bvecs might not be logical block aligned. Round down the 314 * split size so that each bio is properly block size aligned, even if 315 * we do not use the full hardware limits. 316 */ 317 bytes = ALIGN_DOWN(bytes, queue_logical_block_size(q)); 318 319 /* 320 * Bio splitting may cause subtle trouble such as hang when doing sync 321 * iopoll in direct IO routine. Given performance gain of iopoll for 322 * big IO can be trival, disable iopoll when split needed. 323 */ 324 bio_clear_polled(bio); 325 return bio_split(bio, bytes >> SECTOR_SHIFT, GFP_NOIO, bs); 326 } 327 328 /** 329 * __bio_split_to_limits - split a bio to fit the queue limits 330 * @bio: bio to be split 331 * @q: request_queue new bio is being queued at 332 * @nr_segs: returns the number of segments in the returned bio 333 * 334 * Check if @bio needs splitting based on the queue limits, and if so split off 335 * a bio fitting the limits from the beginning of @bio and return it. @bio is 336 * shortened to the remainder and re-submitted. 337 * 338 * The split bio is allocated from @q->bio_split, which is provided by the 339 * block layer. 340 */ 341 struct bio *__bio_split_to_limits(struct bio *bio, struct request_queue *q, 342 unsigned int *nr_segs) 343 { 344 struct bio_set *bs = &bio->bi_bdev->bd_disk->bio_split; 345 struct bio *split; 346 347 switch (bio_op(bio)) { 348 case REQ_OP_DISCARD: 349 case REQ_OP_SECURE_ERASE: 350 split = bio_split_discard(bio, q, nr_segs, bs); 351 break; 352 case REQ_OP_WRITE_ZEROES: 353 split = bio_split_write_zeroes(bio, q, nr_segs, bs); 354 break; 355 default: 356 split = bio_split_rw(bio, q, nr_segs, bs, 357 get_max_io_size(bio, q) << SECTOR_SHIFT); 358 break; 359 } 360 361 if (split) { 362 /* there isn't chance to merge the splitted bio */ 363 split->bi_opf |= REQ_NOMERGE; 364 365 blkcg_bio_issue_init(split); 366 bio_chain(split, bio); 367 trace_block_split(split, bio->bi_iter.bi_sector); 368 submit_bio_noacct(bio); 369 return split; 370 } 371 return bio; 372 } 373 374 /** 375 * bio_split_to_limits - split a bio to fit the queue limits 376 * @bio: bio to be split 377 * 378 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and 379 * if so split off a bio fitting the limits from the beginning of @bio and 380 * return it. @bio is shortened to the remainder and re-submitted. 381 * 382 * The split bio is allocated from @q->bio_split, which is provided by the 383 * block layer. 384 */ 385 struct bio *bio_split_to_limits(struct bio *bio) 386 { 387 struct request_queue *q = bdev_get_queue(bio->bi_bdev); 388 unsigned int nr_segs; 389 390 if (bio_may_exceed_limits(bio, q)) 391 return __bio_split_to_limits(bio, q, &nr_segs); 392 return bio; 393 } 394 EXPORT_SYMBOL(bio_split_to_limits); 395 396 unsigned int blk_recalc_rq_segments(struct request *rq) 397 { 398 unsigned int nr_phys_segs = 0; 399 unsigned int bytes = 0; 400 struct req_iterator iter; 401 struct bio_vec bv; 402 403 if (!rq->bio) 404 return 0; 405 406 switch (bio_op(rq->bio)) { 407 case REQ_OP_DISCARD: 408 case REQ_OP_SECURE_ERASE: 409 if (queue_max_discard_segments(rq->q) > 1) { 410 struct bio *bio = rq->bio; 411 412 for_each_bio(bio) 413 nr_phys_segs++; 414 return nr_phys_segs; 415 } 416 return 1; 417 case REQ_OP_WRITE_ZEROES: 418 return 0; 419 default: 420 break; 421 } 422 423 rq_for_each_bvec(bv, rq, iter) 424 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &bytes, 425 UINT_MAX, UINT_MAX); 426 return nr_phys_segs; 427 } 428 429 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 430 struct scatterlist *sglist) 431 { 432 if (!*sg) 433 return sglist; 434 435 /* 436 * If the driver previously mapped a shorter list, we could see a 437 * termination bit prematurely unless it fully inits the sg table 438 * on each mapping. We KNOW that there must be more entries here 439 * or the driver would be buggy, so force clear the termination bit 440 * to avoid doing a full sg_init_table() in drivers for each command. 441 */ 442 sg_unmark_end(*sg); 443 return sg_next(*sg); 444 } 445 446 static unsigned blk_bvec_map_sg(struct request_queue *q, 447 struct bio_vec *bvec, struct scatterlist *sglist, 448 struct scatterlist **sg) 449 { 450 unsigned nbytes = bvec->bv_len; 451 unsigned nsegs = 0, total = 0; 452 453 while (nbytes > 0) { 454 unsigned offset = bvec->bv_offset + total; 455 unsigned len = min(get_max_segment_size(q, bvec->bv_page, 456 offset), nbytes); 457 struct page *page = bvec->bv_page; 458 459 /* 460 * Unfortunately a fair number of drivers barf on scatterlists 461 * that have an offset larger than PAGE_SIZE, despite other 462 * subsystems dealing with that invariant just fine. For now 463 * stick to the legacy format where we never present those from 464 * the block layer, but the code below should be removed once 465 * these offenders (mostly MMC/SD drivers) are fixed. 466 */ 467 page += (offset >> PAGE_SHIFT); 468 offset &= ~PAGE_MASK; 469 470 *sg = blk_next_sg(sg, sglist); 471 sg_set_page(*sg, page, len, offset); 472 473 total += len; 474 nbytes -= len; 475 nsegs++; 476 } 477 478 return nsegs; 479 } 480 481 static inline int __blk_bvec_map_sg(struct bio_vec bv, 482 struct scatterlist *sglist, struct scatterlist **sg) 483 { 484 *sg = blk_next_sg(sg, sglist); 485 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 486 return 1; 487 } 488 489 /* only try to merge bvecs into one sg if they are from two bios */ 490 static inline bool 491 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec, 492 struct bio_vec *bvprv, struct scatterlist **sg) 493 { 494 495 int nbytes = bvec->bv_len; 496 497 if (!*sg) 498 return false; 499 500 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 501 return false; 502 503 if (!biovec_phys_mergeable(q, bvprv, bvec)) 504 return false; 505 506 (*sg)->length += nbytes; 507 508 return true; 509 } 510 511 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 512 struct scatterlist *sglist, 513 struct scatterlist **sg) 514 { 515 struct bio_vec bvec, bvprv = { NULL }; 516 struct bvec_iter iter; 517 int nsegs = 0; 518 bool new_bio = false; 519 520 for_each_bio(bio) { 521 bio_for_each_bvec(bvec, bio, iter) { 522 /* 523 * Only try to merge bvecs from two bios given we 524 * have done bio internal merge when adding pages 525 * to bio 526 */ 527 if (new_bio && 528 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg)) 529 goto next_bvec; 530 531 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE) 532 nsegs += __blk_bvec_map_sg(bvec, sglist, sg); 533 else 534 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg); 535 next_bvec: 536 new_bio = false; 537 } 538 if (likely(bio->bi_iter.bi_size)) { 539 bvprv = bvec; 540 new_bio = true; 541 } 542 } 543 544 return nsegs; 545 } 546 547 /* 548 * map a request to scatterlist, return number of sg entries setup. Caller 549 * must make sure sg can hold rq->nr_phys_segments entries 550 */ 551 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 552 struct scatterlist *sglist, struct scatterlist **last_sg) 553 { 554 int nsegs = 0; 555 556 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 557 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg); 558 else if (rq->bio) 559 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg); 560 561 if (*last_sg) 562 sg_mark_end(*last_sg); 563 564 /* 565 * Something must have been wrong if the figured number of 566 * segment is bigger than number of req's physical segments 567 */ 568 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 569 570 return nsegs; 571 } 572 EXPORT_SYMBOL(__blk_rq_map_sg); 573 574 static inline unsigned int blk_rq_get_max_segments(struct request *rq) 575 { 576 if (req_op(rq) == REQ_OP_DISCARD) 577 return queue_max_discard_segments(rq->q); 578 return queue_max_segments(rq->q); 579 } 580 581 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 582 sector_t offset) 583 { 584 struct request_queue *q = rq->q; 585 unsigned int max_sectors; 586 587 if (blk_rq_is_passthrough(rq)) 588 return q->limits.max_hw_sectors; 589 590 max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); 591 if (!q->limits.chunk_sectors || 592 req_op(rq) == REQ_OP_DISCARD || 593 req_op(rq) == REQ_OP_SECURE_ERASE) 594 return max_sectors; 595 return min(max_sectors, 596 blk_chunk_sectors_left(offset, q->limits.chunk_sectors)); 597 } 598 599 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 600 unsigned int nr_phys_segs) 601 { 602 if (!blk_cgroup_mergeable(req, bio)) 603 goto no_merge; 604 605 if (blk_integrity_merge_bio(req->q, req, bio) == false) 606 goto no_merge; 607 608 /* discard request merge won't add new segment */ 609 if (req_op(req) == REQ_OP_DISCARD) 610 return 1; 611 612 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 613 goto no_merge; 614 615 /* 616 * This will form the start of a new hw segment. Bump both 617 * counters. 618 */ 619 req->nr_phys_segments += nr_phys_segs; 620 return 1; 621 622 no_merge: 623 req_set_nomerge(req->q, req); 624 return 0; 625 } 626 627 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 628 { 629 if (req_gap_back_merge(req, bio)) 630 return 0; 631 if (blk_integrity_rq(req) && 632 integrity_req_gap_back_merge(req, bio)) 633 return 0; 634 if (!bio_crypt_ctx_back_mergeable(req, bio)) 635 return 0; 636 if (blk_rq_sectors(req) + bio_sectors(bio) > 637 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 638 req_set_nomerge(req->q, req); 639 return 0; 640 } 641 642 return ll_new_hw_segment(req, bio, nr_segs); 643 } 644 645 static int ll_front_merge_fn(struct request *req, struct bio *bio, 646 unsigned int nr_segs) 647 { 648 if (req_gap_front_merge(req, bio)) 649 return 0; 650 if (blk_integrity_rq(req) && 651 integrity_req_gap_front_merge(req, bio)) 652 return 0; 653 if (!bio_crypt_ctx_front_mergeable(req, bio)) 654 return 0; 655 if (blk_rq_sectors(req) + bio_sectors(bio) > 656 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 657 req_set_nomerge(req->q, req); 658 return 0; 659 } 660 661 return ll_new_hw_segment(req, bio, nr_segs); 662 } 663 664 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 665 struct request *next) 666 { 667 unsigned short segments = blk_rq_nr_discard_segments(req); 668 669 if (segments >= queue_max_discard_segments(q)) 670 goto no_merge; 671 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 672 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 673 goto no_merge; 674 675 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 676 return true; 677 no_merge: 678 req_set_nomerge(q, req); 679 return false; 680 } 681 682 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 683 struct request *next) 684 { 685 int total_phys_segments; 686 687 if (req_gap_back_merge(req, next->bio)) 688 return 0; 689 690 /* 691 * Will it become too large? 692 */ 693 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 694 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 695 return 0; 696 697 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 698 if (total_phys_segments > blk_rq_get_max_segments(req)) 699 return 0; 700 701 if (!blk_cgroup_mergeable(req, next->bio)) 702 return 0; 703 704 if (blk_integrity_merge_rq(q, req, next) == false) 705 return 0; 706 707 if (!bio_crypt_ctx_merge_rq(req, next)) 708 return 0; 709 710 /* Merge is OK... */ 711 req->nr_phys_segments = total_phys_segments; 712 return 1; 713 } 714 715 /** 716 * blk_rq_set_mixed_merge - mark a request as mixed merge 717 * @rq: request to mark as mixed merge 718 * 719 * Description: 720 * @rq is about to be mixed merged. Make sure the attributes 721 * which can be mixed are set in each bio and mark @rq as mixed 722 * merged. 723 */ 724 void blk_rq_set_mixed_merge(struct request *rq) 725 { 726 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK; 727 struct bio *bio; 728 729 if (rq->rq_flags & RQF_MIXED_MERGE) 730 return; 731 732 /* 733 * @rq will no longer represent mixable attributes for all the 734 * contained bios. It will just track those of the first one. 735 * Distributes the attributs to each bio. 736 */ 737 for (bio = rq->bio; bio; bio = bio->bi_next) { 738 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 739 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 740 bio->bi_opf |= ff; 741 } 742 rq->rq_flags |= RQF_MIXED_MERGE; 743 } 744 745 static void blk_account_io_merge_request(struct request *req) 746 { 747 if (blk_do_io_stat(req)) { 748 part_stat_lock(); 749 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 750 part_stat_unlock(); 751 } 752 } 753 754 static enum elv_merge blk_try_req_merge(struct request *req, 755 struct request *next) 756 { 757 if (blk_discard_mergable(req)) 758 return ELEVATOR_DISCARD_MERGE; 759 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 760 return ELEVATOR_BACK_MERGE; 761 762 return ELEVATOR_NO_MERGE; 763 } 764 765 /* 766 * For non-mq, this has to be called with the request spinlock acquired. 767 * For mq with scheduling, the appropriate queue wide lock should be held. 768 */ 769 static struct request *attempt_merge(struct request_queue *q, 770 struct request *req, struct request *next) 771 { 772 if (!rq_mergeable(req) || !rq_mergeable(next)) 773 return NULL; 774 775 if (req_op(req) != req_op(next)) 776 return NULL; 777 778 if (rq_data_dir(req) != rq_data_dir(next)) 779 return NULL; 780 781 if (req->ioprio != next->ioprio) 782 return NULL; 783 784 /* 785 * If we are allowed to merge, then append bio list 786 * from next to rq and release next. merge_requests_fn 787 * will have updated segment counts, update sector 788 * counts here. Handle DISCARDs separately, as they 789 * have separate settings. 790 */ 791 792 switch (blk_try_req_merge(req, next)) { 793 case ELEVATOR_DISCARD_MERGE: 794 if (!req_attempt_discard_merge(q, req, next)) 795 return NULL; 796 break; 797 case ELEVATOR_BACK_MERGE: 798 if (!ll_merge_requests_fn(q, req, next)) 799 return NULL; 800 break; 801 default: 802 return NULL; 803 } 804 805 /* 806 * If failfast settings disagree or any of the two is already 807 * a mixed merge, mark both as mixed before proceeding. This 808 * makes sure that all involved bios have mixable attributes 809 * set properly. 810 */ 811 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 812 (req->cmd_flags & REQ_FAILFAST_MASK) != 813 (next->cmd_flags & REQ_FAILFAST_MASK)) { 814 blk_rq_set_mixed_merge(req); 815 blk_rq_set_mixed_merge(next); 816 } 817 818 /* 819 * At this point we have either done a back merge or front merge. We 820 * need the smaller start_time_ns of the merged requests to be the 821 * current request for accounting purposes. 822 */ 823 if (next->start_time_ns < req->start_time_ns) 824 req->start_time_ns = next->start_time_ns; 825 826 req->biotail->bi_next = next->bio; 827 req->biotail = next->biotail; 828 829 req->__data_len += blk_rq_bytes(next); 830 831 if (!blk_discard_mergable(req)) 832 elv_merge_requests(q, req, next); 833 834 /* 835 * 'next' is going away, so update stats accordingly 836 */ 837 blk_account_io_merge_request(next); 838 839 trace_block_rq_merge(next); 840 841 /* 842 * ownership of bio passed from next to req, return 'next' for 843 * the caller to free 844 */ 845 next->bio = NULL; 846 return next; 847 } 848 849 static struct request *attempt_back_merge(struct request_queue *q, 850 struct request *rq) 851 { 852 struct request *next = elv_latter_request(q, rq); 853 854 if (next) 855 return attempt_merge(q, rq, next); 856 857 return NULL; 858 } 859 860 static struct request *attempt_front_merge(struct request_queue *q, 861 struct request *rq) 862 { 863 struct request *prev = elv_former_request(q, rq); 864 865 if (prev) 866 return attempt_merge(q, prev, rq); 867 868 return NULL; 869 } 870 871 /* 872 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 873 * otherwise. The caller is responsible for freeing 'next' if the merge 874 * happened. 875 */ 876 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 877 struct request *next) 878 { 879 return attempt_merge(q, rq, next); 880 } 881 882 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 883 { 884 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 885 return false; 886 887 if (req_op(rq) != bio_op(bio)) 888 return false; 889 890 /* different data direction or already started, don't merge */ 891 if (bio_data_dir(bio) != rq_data_dir(rq)) 892 return false; 893 894 /* don't merge across cgroup boundaries */ 895 if (!blk_cgroup_mergeable(rq, bio)) 896 return false; 897 898 /* only merge integrity protected bio into ditto rq */ 899 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 900 return false; 901 902 /* Only merge if the crypt contexts are compatible */ 903 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 904 return false; 905 906 if (rq->ioprio != bio_prio(bio)) 907 return false; 908 909 return true; 910 } 911 912 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 913 { 914 if (blk_discard_mergable(rq)) 915 return ELEVATOR_DISCARD_MERGE; 916 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 917 return ELEVATOR_BACK_MERGE; 918 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 919 return ELEVATOR_FRONT_MERGE; 920 return ELEVATOR_NO_MERGE; 921 } 922 923 static void blk_account_io_merge_bio(struct request *req) 924 { 925 if (!blk_do_io_stat(req)) 926 return; 927 928 part_stat_lock(); 929 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 930 part_stat_unlock(); 931 } 932 933 enum bio_merge_status { 934 BIO_MERGE_OK, 935 BIO_MERGE_NONE, 936 BIO_MERGE_FAILED, 937 }; 938 939 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 940 struct bio *bio, unsigned int nr_segs) 941 { 942 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 943 944 if (!ll_back_merge_fn(req, bio, nr_segs)) 945 return BIO_MERGE_FAILED; 946 947 trace_block_bio_backmerge(bio); 948 rq_qos_merge(req->q, req, bio); 949 950 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 951 blk_rq_set_mixed_merge(req); 952 953 req->biotail->bi_next = bio; 954 req->biotail = bio; 955 req->__data_len += bio->bi_iter.bi_size; 956 957 bio_crypt_free_ctx(bio); 958 959 blk_account_io_merge_bio(req); 960 return BIO_MERGE_OK; 961 } 962 963 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 964 struct bio *bio, unsigned int nr_segs) 965 { 966 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 967 968 if (!ll_front_merge_fn(req, bio, nr_segs)) 969 return BIO_MERGE_FAILED; 970 971 trace_block_bio_frontmerge(bio); 972 rq_qos_merge(req->q, req, bio); 973 974 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 975 blk_rq_set_mixed_merge(req); 976 977 bio->bi_next = req->bio; 978 req->bio = bio; 979 980 req->__sector = bio->bi_iter.bi_sector; 981 req->__data_len += bio->bi_iter.bi_size; 982 983 bio_crypt_do_front_merge(req, bio); 984 985 blk_account_io_merge_bio(req); 986 return BIO_MERGE_OK; 987 } 988 989 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 990 struct request *req, struct bio *bio) 991 { 992 unsigned short segments = blk_rq_nr_discard_segments(req); 993 994 if (segments >= queue_max_discard_segments(q)) 995 goto no_merge; 996 if (blk_rq_sectors(req) + bio_sectors(bio) > 997 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 998 goto no_merge; 999 1000 rq_qos_merge(q, req, bio); 1001 1002 req->biotail->bi_next = bio; 1003 req->biotail = bio; 1004 req->__data_len += bio->bi_iter.bi_size; 1005 req->nr_phys_segments = segments + 1; 1006 1007 blk_account_io_merge_bio(req); 1008 return BIO_MERGE_OK; 1009 no_merge: 1010 req_set_nomerge(q, req); 1011 return BIO_MERGE_FAILED; 1012 } 1013 1014 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1015 struct request *rq, 1016 struct bio *bio, 1017 unsigned int nr_segs, 1018 bool sched_allow_merge) 1019 { 1020 if (!blk_rq_merge_ok(rq, bio)) 1021 return BIO_MERGE_NONE; 1022 1023 switch (blk_try_merge(rq, bio)) { 1024 case ELEVATOR_BACK_MERGE: 1025 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1026 return bio_attempt_back_merge(rq, bio, nr_segs); 1027 break; 1028 case ELEVATOR_FRONT_MERGE: 1029 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1030 return bio_attempt_front_merge(rq, bio, nr_segs); 1031 break; 1032 case ELEVATOR_DISCARD_MERGE: 1033 return bio_attempt_discard_merge(q, rq, bio); 1034 default: 1035 return BIO_MERGE_NONE; 1036 } 1037 1038 return BIO_MERGE_FAILED; 1039 } 1040 1041 /** 1042 * blk_attempt_plug_merge - try to merge with %current's plugged list 1043 * @q: request_queue new bio is being queued at 1044 * @bio: new bio being queued 1045 * @nr_segs: number of segments in @bio 1046 * from the passed in @q already in the plug list 1047 * 1048 * Determine whether @bio being queued on @q can be merged with the previous 1049 * request on %current's plugged list. Returns %true if merge was successful, 1050 * otherwise %false. 1051 * 1052 * Plugging coalesces IOs from the same issuer for the same purpose without 1053 * going through @q->queue_lock. As such it's more of an issuing mechanism 1054 * than scheduling, and the request, while may have elvpriv data, is not 1055 * added on the elevator at this point. In addition, we don't have 1056 * reliable access to the elevator outside queue lock. Only check basic 1057 * merging parameters without querying the elevator. 1058 * 1059 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1060 */ 1061 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1062 unsigned int nr_segs) 1063 { 1064 struct blk_plug *plug; 1065 struct request *rq; 1066 1067 plug = blk_mq_plug(bio); 1068 if (!plug || rq_list_empty(plug->mq_list)) 1069 return false; 1070 1071 rq_list_for_each(&plug->mq_list, rq) { 1072 if (rq->q == q) { 1073 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1074 BIO_MERGE_OK) 1075 return true; 1076 break; 1077 } 1078 1079 /* 1080 * Only keep iterating plug list for merges if we have multiple 1081 * queues 1082 */ 1083 if (!plug->multiple_queues) 1084 break; 1085 } 1086 return false; 1087 } 1088 1089 /* 1090 * Iterate list of requests and see if we can merge this bio with any 1091 * of them. 1092 */ 1093 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1094 struct bio *bio, unsigned int nr_segs) 1095 { 1096 struct request *rq; 1097 int checked = 8; 1098 1099 list_for_each_entry_reverse(rq, list, queuelist) { 1100 if (!checked--) 1101 break; 1102 1103 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1104 case BIO_MERGE_NONE: 1105 continue; 1106 case BIO_MERGE_OK: 1107 return true; 1108 case BIO_MERGE_FAILED: 1109 return false; 1110 } 1111 1112 } 1113 1114 return false; 1115 } 1116 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1117 1118 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1119 unsigned int nr_segs, struct request **merged_request) 1120 { 1121 struct request *rq; 1122 1123 switch (elv_merge(q, &rq, bio)) { 1124 case ELEVATOR_BACK_MERGE: 1125 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1126 return false; 1127 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1128 return false; 1129 *merged_request = attempt_back_merge(q, rq); 1130 if (!*merged_request) 1131 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1132 return true; 1133 case ELEVATOR_FRONT_MERGE: 1134 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1135 return false; 1136 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1137 return false; 1138 *merged_request = attempt_front_merge(q, rq); 1139 if (!*merged_request) 1140 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1141 return true; 1142 case ELEVATOR_DISCARD_MERGE: 1143 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1144 default: 1145 return false; 1146 } 1147 } 1148 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1149