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_submit_split(struct bio *bio, int split_sectors) 109 { 110 if (unlikely(split_sectors < 0)) 111 goto error; 112 113 if (split_sectors) { 114 struct bio *split; 115 116 split = bio_split(bio, split_sectors, GFP_NOIO, 117 &bio->bi_bdev->bd_disk->bio_split); 118 if (IS_ERR(split)) { 119 split_sectors = PTR_ERR(split); 120 goto error; 121 } 122 split->bi_opf |= REQ_NOMERGE; 123 blkcg_bio_issue_init(split); 124 bio_chain(split, bio); 125 trace_block_split(split, bio->bi_iter.bi_sector); 126 WARN_ON_ONCE(bio_zone_write_plugging(bio)); 127 submit_bio_noacct(bio); 128 return split; 129 } 130 131 return bio; 132 error: 133 bio->bi_status = errno_to_blk_status(split_sectors); 134 bio_endio(bio); 135 return NULL; 136 } 137 138 struct bio *bio_split_discard(struct bio *bio, const struct queue_limits *lim, 139 unsigned *nsegs) 140 { 141 unsigned int max_discard_sectors, granularity; 142 sector_t tmp; 143 unsigned split_sectors; 144 145 *nsegs = 1; 146 147 granularity = max(lim->discard_granularity >> 9, 1U); 148 149 max_discard_sectors = 150 min(lim->max_discard_sectors, bio_allowed_max_sectors(lim)); 151 max_discard_sectors -= max_discard_sectors % granularity; 152 if (unlikely(!max_discard_sectors)) 153 return bio; 154 155 if (bio_sectors(bio) <= max_discard_sectors) 156 return bio; 157 158 split_sectors = max_discard_sectors; 159 160 /* 161 * If the next starting sector would be misaligned, stop the discard at 162 * the previous aligned sector. 163 */ 164 tmp = bio->bi_iter.bi_sector + split_sectors - 165 ((lim->discard_alignment >> 9) % granularity); 166 tmp = sector_div(tmp, granularity); 167 168 if (split_sectors > tmp) 169 split_sectors -= tmp; 170 171 return bio_submit_split(bio, split_sectors); 172 } 173 174 static inline unsigned int blk_boundary_sectors(const struct queue_limits *lim, 175 bool is_atomic) 176 { 177 /* 178 * chunk_sectors must be a multiple of atomic_write_boundary_sectors if 179 * both non-zero. 180 */ 181 if (is_atomic && lim->atomic_write_boundary_sectors) 182 return lim->atomic_write_boundary_sectors; 183 184 return lim->chunk_sectors; 185 } 186 187 /* 188 * Return the maximum number of sectors from the start of a bio that may be 189 * submitted as a single request to a block device. If enough sectors remain, 190 * align the end to the physical block size. Otherwise align the end to the 191 * logical block size. This approach minimizes the number of non-aligned 192 * requests that are submitted to a block device if the start of a bio is not 193 * aligned to a physical block boundary. 194 */ 195 static inline unsigned get_max_io_size(struct bio *bio, 196 const struct queue_limits *lim) 197 { 198 unsigned pbs = lim->physical_block_size >> SECTOR_SHIFT; 199 unsigned lbs = lim->logical_block_size >> SECTOR_SHIFT; 200 bool is_atomic = bio->bi_opf & REQ_ATOMIC; 201 unsigned boundary_sectors = blk_boundary_sectors(lim, is_atomic); 202 unsigned max_sectors, start, end; 203 204 /* 205 * We ignore lim->max_sectors for atomic writes because it may less 206 * than the actual bio size, which we cannot tolerate. 207 */ 208 if (bio_op(bio) == REQ_OP_WRITE_ZEROES) 209 max_sectors = lim->max_write_zeroes_sectors; 210 else if (is_atomic) 211 max_sectors = lim->atomic_write_max_sectors; 212 else 213 max_sectors = lim->max_sectors; 214 215 if (boundary_sectors) { 216 max_sectors = min(max_sectors, 217 blk_boundary_sectors_left(bio->bi_iter.bi_sector, 218 boundary_sectors)); 219 } 220 221 start = bio->bi_iter.bi_sector & (pbs - 1); 222 end = (start + max_sectors) & ~(pbs - 1); 223 if (end > start) 224 return end - start; 225 return max_sectors & ~(lbs - 1); 226 } 227 228 /** 229 * get_max_segment_size() - maximum number of bytes to add as a single segment 230 * @lim: Request queue limits. 231 * @paddr: address of the range to add 232 * @len: maximum length available to add at @paddr 233 * 234 * Returns the maximum number of bytes of the range starting at @paddr that can 235 * be added to a single segment. 236 */ 237 static inline unsigned get_max_segment_size(const struct queue_limits *lim, 238 phys_addr_t paddr, unsigned int len) 239 { 240 /* 241 * Prevent an overflow if mask = ULONG_MAX and offset = 0 by adding 1 242 * after having calculated the minimum. 243 */ 244 return min_t(unsigned long, len, 245 min(lim->seg_boundary_mask - (lim->seg_boundary_mask & paddr), 246 (unsigned long)lim->max_segment_size - 1) + 1); 247 } 248 249 /** 250 * bvec_split_segs - verify whether or not a bvec should be split in the middle 251 * @lim: [in] queue limits to split based on 252 * @bv: [in] bvec to examine 253 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 254 * by the number of segments from @bv that may be appended to that 255 * bio without exceeding @max_segs 256 * @bytes: [in,out] Number of bytes in the bio being built. Incremented 257 * by the number of bytes from @bv that may be appended to that 258 * bio without exceeding @max_bytes 259 * @max_segs: [in] upper bound for *@nsegs 260 * @max_bytes: [in] upper bound for *@bytes 261 * 262 * When splitting a bio, it can happen that a bvec is encountered that is too 263 * big to fit in a single segment and hence that it has to be split in the 264 * middle. This function verifies whether or not that should happen. The value 265 * %true is returned if and only if appending the entire @bv to a bio with 266 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 267 * the block driver. 268 */ 269 static bool bvec_split_segs(const struct queue_limits *lim, 270 const struct bio_vec *bv, unsigned *nsegs, unsigned *bytes, 271 unsigned max_segs, unsigned max_bytes) 272 { 273 unsigned max_len = max_bytes - *bytes; 274 unsigned len = min(bv->bv_len, max_len); 275 unsigned total_len = 0; 276 unsigned seg_size = 0; 277 278 while (len && *nsegs < max_segs) { 279 seg_size = get_max_segment_size(lim, bvec_phys(bv) + total_len, len); 280 281 (*nsegs)++; 282 total_len += seg_size; 283 len -= seg_size; 284 285 if ((bv->bv_offset + total_len) & lim->virt_boundary_mask) 286 break; 287 } 288 289 *bytes += total_len; 290 291 /* tell the caller to split the bvec if it is too big to fit */ 292 return len > 0 || bv->bv_len > max_len; 293 } 294 295 static unsigned int bio_split_alignment(struct bio *bio, 296 const struct queue_limits *lim) 297 { 298 if (op_is_write(bio_op(bio)) && lim->zone_write_granularity) 299 return lim->zone_write_granularity; 300 return lim->logical_block_size; 301 } 302 303 /** 304 * bio_split_rw_at - check if and where to split a read/write bio 305 * @bio: [in] bio to be split 306 * @lim: [in] queue limits to split based on 307 * @segs: [out] number of segments in the bio with the first half of the sectors 308 * @max_bytes: [in] maximum number of bytes per bio 309 * 310 * Find out if @bio needs to be split to fit the queue limits in @lim and a 311 * maximum size of @max_bytes. Returns a negative error number if @bio can't be 312 * split, 0 if the bio doesn't have to be split, or a positive sector offset if 313 * @bio needs to be split. 314 */ 315 int bio_split_rw_at(struct bio *bio, const struct queue_limits *lim, 316 unsigned *segs, unsigned max_bytes) 317 { 318 struct bio_vec bv, bvprv, *bvprvp = NULL; 319 struct bvec_iter iter; 320 unsigned nsegs = 0, bytes = 0; 321 322 bio_for_each_bvec(bv, bio, iter) { 323 /* 324 * If the queue doesn't support SG gaps and adding this 325 * offset would create a gap, disallow it. 326 */ 327 if (bvprvp && bvec_gap_to_prev(lim, bvprvp, bv.bv_offset)) 328 goto split; 329 330 if (nsegs < lim->max_segments && 331 bytes + bv.bv_len <= max_bytes && 332 bv.bv_offset + bv.bv_len <= lim->min_segment_size) { 333 nsegs++; 334 bytes += bv.bv_len; 335 } else { 336 if (bvec_split_segs(lim, &bv, &nsegs, &bytes, 337 lim->max_segments, max_bytes)) 338 goto split; 339 } 340 341 bvprv = bv; 342 bvprvp = &bvprv; 343 } 344 345 *segs = nsegs; 346 return 0; 347 split: 348 if (bio->bi_opf & REQ_ATOMIC) 349 return -EINVAL; 350 351 /* 352 * We can't sanely support splitting for a REQ_NOWAIT bio. End it 353 * with EAGAIN if splitting is required and return an error pointer. 354 */ 355 if (bio->bi_opf & REQ_NOWAIT) 356 return -EAGAIN; 357 358 *segs = nsegs; 359 360 /* 361 * Individual bvecs might not be logical block aligned. Round down the 362 * split size so that each bio is properly block size aligned, even if 363 * we do not use the full hardware limits. 364 */ 365 bytes = ALIGN_DOWN(bytes, bio_split_alignment(bio, lim)); 366 367 /* 368 * Bio splitting may cause subtle trouble such as hang when doing sync 369 * iopoll in direct IO routine. Given performance gain of iopoll for 370 * big IO can be trival, disable iopoll when split needed. 371 */ 372 bio_clear_polled(bio); 373 return bytes >> SECTOR_SHIFT; 374 } 375 EXPORT_SYMBOL_GPL(bio_split_rw_at); 376 377 struct bio *bio_split_rw(struct bio *bio, const struct queue_limits *lim, 378 unsigned *nr_segs) 379 { 380 return bio_submit_split(bio, 381 bio_split_rw_at(bio, lim, nr_segs, 382 get_max_io_size(bio, lim) << SECTOR_SHIFT)); 383 } 384 385 /* 386 * REQ_OP_ZONE_APPEND bios must never be split by the block layer. 387 * 388 * But we want the nr_segs calculation provided by bio_split_rw_at, and having 389 * a good sanity check that the submitter built the bio correctly is nice to 390 * have as well. 391 */ 392 struct bio *bio_split_zone_append(struct bio *bio, 393 const struct queue_limits *lim, unsigned *nr_segs) 394 { 395 int split_sectors; 396 397 split_sectors = bio_split_rw_at(bio, lim, nr_segs, 398 lim->max_zone_append_sectors << SECTOR_SHIFT); 399 if (WARN_ON_ONCE(split_sectors > 0)) 400 split_sectors = -EINVAL; 401 return bio_submit_split(bio, split_sectors); 402 } 403 404 struct bio *bio_split_write_zeroes(struct bio *bio, 405 const struct queue_limits *lim, unsigned *nsegs) 406 { 407 unsigned int max_sectors = get_max_io_size(bio, lim); 408 409 *nsegs = 0; 410 411 /* 412 * An unset limit should normally not happen, as bio submission is keyed 413 * off having a non-zero limit. But SCSI can clear the limit in the 414 * I/O completion handler, and we can race and see this. Splitting to a 415 * zero limit obviously doesn't make sense, so band-aid it here. 416 */ 417 if (!max_sectors) 418 return bio; 419 if (bio_sectors(bio) <= max_sectors) 420 return bio; 421 return bio_submit_split(bio, max_sectors); 422 } 423 424 /** 425 * bio_split_to_limits - split a bio to fit the queue limits 426 * @bio: bio to be split 427 * 428 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and 429 * if so split off a bio fitting the limits from the beginning of @bio and 430 * return it. @bio is shortened to the remainder and re-submitted. 431 * 432 * The split bio is allocated from @q->bio_split, which is provided by the 433 * block layer. 434 */ 435 struct bio *bio_split_to_limits(struct bio *bio) 436 { 437 unsigned int nr_segs; 438 439 return __bio_split_to_limits(bio, bdev_limits(bio->bi_bdev), &nr_segs); 440 } 441 EXPORT_SYMBOL(bio_split_to_limits); 442 443 unsigned int blk_recalc_rq_segments(struct request *rq) 444 { 445 unsigned int nr_phys_segs = 0; 446 unsigned int bytes = 0; 447 struct req_iterator iter; 448 struct bio_vec bv; 449 450 if (!rq->bio) 451 return 0; 452 453 switch (bio_op(rq->bio)) { 454 case REQ_OP_DISCARD: 455 case REQ_OP_SECURE_ERASE: 456 if (queue_max_discard_segments(rq->q) > 1) { 457 struct bio *bio = rq->bio; 458 459 for_each_bio(bio) 460 nr_phys_segs++; 461 return nr_phys_segs; 462 } 463 return 1; 464 case REQ_OP_WRITE_ZEROES: 465 return 0; 466 default: 467 break; 468 } 469 470 rq_for_each_bvec(bv, rq, iter) 471 bvec_split_segs(&rq->q->limits, &bv, &nr_phys_segs, &bytes, 472 UINT_MAX, UINT_MAX); 473 return nr_phys_segs; 474 } 475 476 struct phys_vec { 477 phys_addr_t paddr; 478 u32 len; 479 }; 480 481 static bool blk_map_iter_next(struct request *req, 482 struct req_iterator *iter, struct phys_vec *vec) 483 { 484 unsigned int max_size; 485 struct bio_vec bv; 486 487 if (req->rq_flags & RQF_SPECIAL_PAYLOAD) { 488 if (!iter->bio) 489 return false; 490 vec->paddr = bvec_phys(&req->special_vec); 491 vec->len = req->special_vec.bv_len; 492 iter->bio = NULL; 493 return true; 494 } 495 496 if (!iter->iter.bi_size) 497 return false; 498 499 bv = mp_bvec_iter_bvec(iter->bio->bi_io_vec, iter->iter); 500 vec->paddr = bvec_phys(&bv); 501 max_size = get_max_segment_size(&req->q->limits, vec->paddr, UINT_MAX); 502 bv.bv_len = min(bv.bv_len, max_size); 503 bio_advance_iter_single(iter->bio, &iter->iter, bv.bv_len); 504 505 /* 506 * If we are entirely done with this bi_io_vec entry, check if the next 507 * one could be merged into it. This typically happens when moving to 508 * the next bio, but some callers also don't pack bvecs tight. 509 */ 510 while (!iter->iter.bi_size || !iter->iter.bi_bvec_done) { 511 struct bio_vec next; 512 513 if (!iter->iter.bi_size) { 514 if (!iter->bio->bi_next) 515 break; 516 iter->bio = iter->bio->bi_next; 517 iter->iter = iter->bio->bi_iter; 518 } 519 520 next = mp_bvec_iter_bvec(iter->bio->bi_io_vec, iter->iter); 521 if (bv.bv_len + next.bv_len > max_size || 522 !biovec_phys_mergeable(req->q, &bv, &next)) 523 break; 524 525 bv.bv_len += next.bv_len; 526 bio_advance_iter_single(iter->bio, &iter->iter, next.bv_len); 527 } 528 529 vec->len = bv.bv_len; 530 return true; 531 } 532 533 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 534 struct scatterlist *sglist) 535 { 536 if (!*sg) 537 return sglist; 538 539 /* 540 * If the driver previously mapped a shorter list, we could see a 541 * termination bit prematurely unless it fully inits the sg table 542 * on each mapping. We KNOW that there must be more entries here 543 * or the driver would be buggy, so force clear the termination bit 544 * to avoid doing a full sg_init_table() in drivers for each command. 545 */ 546 sg_unmark_end(*sg); 547 return sg_next(*sg); 548 } 549 550 /* 551 * Map a request to scatterlist, return number of sg entries setup. Caller 552 * must make sure sg can hold rq->nr_phys_segments entries. 553 */ 554 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 555 struct scatterlist *sglist, struct scatterlist **last_sg) 556 { 557 struct req_iterator iter = { 558 .bio = rq->bio, 559 }; 560 struct phys_vec vec; 561 int nsegs = 0; 562 563 /* the internal flush request may not have bio attached */ 564 if (iter.bio) 565 iter.iter = iter.bio->bi_iter; 566 567 while (blk_map_iter_next(rq, &iter, &vec)) { 568 *last_sg = blk_next_sg(last_sg, sglist); 569 sg_set_page(*last_sg, phys_to_page(vec.paddr), vec.len, 570 offset_in_page(vec.paddr)); 571 nsegs++; 572 } 573 574 if (*last_sg) 575 sg_mark_end(*last_sg); 576 577 /* 578 * Something must have been wrong if the figured number of 579 * segment is bigger than number of req's physical segments 580 */ 581 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 582 583 return nsegs; 584 } 585 EXPORT_SYMBOL(__blk_rq_map_sg); 586 587 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 588 sector_t offset) 589 { 590 struct request_queue *q = rq->q; 591 struct queue_limits *lim = &q->limits; 592 unsigned int max_sectors, boundary_sectors; 593 bool is_atomic = rq->cmd_flags & REQ_ATOMIC; 594 595 if (blk_rq_is_passthrough(rq)) 596 return q->limits.max_hw_sectors; 597 598 boundary_sectors = blk_boundary_sectors(lim, is_atomic); 599 max_sectors = blk_queue_get_max_sectors(rq); 600 601 if (!boundary_sectors || 602 req_op(rq) == REQ_OP_DISCARD || 603 req_op(rq) == REQ_OP_SECURE_ERASE) 604 return max_sectors; 605 return min(max_sectors, 606 blk_boundary_sectors_left(offset, boundary_sectors)); 607 } 608 609 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 610 unsigned int nr_phys_segs) 611 { 612 if (!blk_cgroup_mergeable(req, bio)) 613 goto no_merge; 614 615 if (blk_integrity_merge_bio(req->q, req, bio) == false) 616 goto no_merge; 617 618 /* discard request merge won't add new segment */ 619 if (req_op(req) == REQ_OP_DISCARD) 620 return 1; 621 622 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 623 goto no_merge; 624 625 /* 626 * This will form the start of a new hw segment. Bump both 627 * counters. 628 */ 629 req->nr_phys_segments += nr_phys_segs; 630 if (bio_integrity(bio)) 631 req->nr_integrity_segments += blk_rq_count_integrity_sg(req->q, 632 bio); 633 return 1; 634 635 no_merge: 636 req_set_nomerge(req->q, req); 637 return 0; 638 } 639 640 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 641 { 642 if (req_gap_back_merge(req, bio)) 643 return 0; 644 if (blk_integrity_rq(req) && 645 integrity_req_gap_back_merge(req, bio)) 646 return 0; 647 if (!bio_crypt_ctx_back_mergeable(req, bio)) 648 return 0; 649 if (blk_rq_sectors(req) + bio_sectors(bio) > 650 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 651 req_set_nomerge(req->q, req); 652 return 0; 653 } 654 655 return ll_new_hw_segment(req, bio, nr_segs); 656 } 657 658 static int ll_front_merge_fn(struct request *req, struct bio *bio, 659 unsigned int nr_segs) 660 { 661 if (req_gap_front_merge(req, bio)) 662 return 0; 663 if (blk_integrity_rq(req) && 664 integrity_req_gap_front_merge(req, bio)) 665 return 0; 666 if (!bio_crypt_ctx_front_mergeable(req, bio)) 667 return 0; 668 if (blk_rq_sectors(req) + bio_sectors(bio) > 669 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 670 req_set_nomerge(req->q, req); 671 return 0; 672 } 673 674 return ll_new_hw_segment(req, bio, nr_segs); 675 } 676 677 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 678 struct request *next) 679 { 680 unsigned short segments = blk_rq_nr_discard_segments(req); 681 682 if (segments >= queue_max_discard_segments(q)) 683 goto no_merge; 684 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 685 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 686 goto no_merge; 687 688 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 689 return true; 690 no_merge: 691 req_set_nomerge(q, req); 692 return false; 693 } 694 695 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 696 struct request *next) 697 { 698 int total_phys_segments; 699 700 if (req_gap_back_merge(req, next->bio)) 701 return 0; 702 703 /* 704 * Will it become too large? 705 */ 706 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 707 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 708 return 0; 709 710 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 711 if (total_phys_segments > blk_rq_get_max_segments(req)) 712 return 0; 713 714 if (!blk_cgroup_mergeable(req, next->bio)) 715 return 0; 716 717 if (blk_integrity_merge_rq(q, req, next) == false) 718 return 0; 719 720 if (!bio_crypt_ctx_merge_rq(req, next)) 721 return 0; 722 723 /* Merge is OK... */ 724 req->nr_phys_segments = total_phys_segments; 725 req->nr_integrity_segments += next->nr_integrity_segments; 726 return 1; 727 } 728 729 /** 730 * blk_rq_set_mixed_merge - mark a request as mixed merge 731 * @rq: request to mark as mixed merge 732 * 733 * Description: 734 * @rq is about to be mixed merged. Make sure the attributes 735 * which can be mixed are set in each bio and mark @rq as mixed 736 * merged. 737 */ 738 static void blk_rq_set_mixed_merge(struct request *rq) 739 { 740 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK; 741 struct bio *bio; 742 743 if (rq->rq_flags & RQF_MIXED_MERGE) 744 return; 745 746 /* 747 * @rq will no longer represent mixable attributes for all the 748 * contained bios. It will just track those of the first one. 749 * Distributes the attributs to each bio. 750 */ 751 for (bio = rq->bio; bio; bio = bio->bi_next) { 752 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 753 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 754 bio->bi_opf |= ff; 755 } 756 rq->rq_flags |= RQF_MIXED_MERGE; 757 } 758 759 static inline blk_opf_t bio_failfast(const struct bio *bio) 760 { 761 if (bio->bi_opf & REQ_RAHEAD) 762 return REQ_FAILFAST_MASK; 763 764 return bio->bi_opf & REQ_FAILFAST_MASK; 765 } 766 767 /* 768 * After we are marked as MIXED_MERGE, any new RA bio has to be updated 769 * as failfast, and request's failfast has to be updated in case of 770 * front merge. 771 */ 772 static inline void blk_update_mixed_merge(struct request *req, 773 struct bio *bio, bool front_merge) 774 { 775 if (req->rq_flags & RQF_MIXED_MERGE) { 776 if (bio->bi_opf & REQ_RAHEAD) 777 bio->bi_opf |= REQ_FAILFAST_MASK; 778 779 if (front_merge) { 780 req->cmd_flags &= ~REQ_FAILFAST_MASK; 781 req->cmd_flags |= bio->bi_opf & REQ_FAILFAST_MASK; 782 } 783 } 784 } 785 786 static void blk_account_io_merge_request(struct request *req) 787 { 788 if (req->rq_flags & RQF_IO_STAT) { 789 part_stat_lock(); 790 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 791 part_stat_local_dec(req->part, 792 in_flight[op_is_write(req_op(req))]); 793 part_stat_unlock(); 794 } 795 } 796 797 static enum elv_merge blk_try_req_merge(struct request *req, 798 struct request *next) 799 { 800 if (blk_discard_mergable(req)) 801 return ELEVATOR_DISCARD_MERGE; 802 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 803 return ELEVATOR_BACK_MERGE; 804 805 return ELEVATOR_NO_MERGE; 806 } 807 808 static bool blk_atomic_write_mergeable_rq_bio(struct request *rq, 809 struct bio *bio) 810 { 811 return (rq->cmd_flags & REQ_ATOMIC) == (bio->bi_opf & REQ_ATOMIC); 812 } 813 814 static bool blk_atomic_write_mergeable_rqs(struct request *rq, 815 struct request *next) 816 { 817 return (rq->cmd_flags & REQ_ATOMIC) == (next->cmd_flags & REQ_ATOMIC); 818 } 819 820 /* 821 * For non-mq, this has to be called with the request spinlock acquired. 822 * For mq with scheduling, the appropriate queue wide lock should be held. 823 */ 824 static struct request *attempt_merge(struct request_queue *q, 825 struct request *req, struct request *next) 826 { 827 if (!rq_mergeable(req) || !rq_mergeable(next)) 828 return NULL; 829 830 if (req_op(req) != req_op(next)) 831 return NULL; 832 833 if (req->bio->bi_write_hint != next->bio->bi_write_hint) 834 return NULL; 835 if (req->bio->bi_ioprio != next->bio->bi_ioprio) 836 return NULL; 837 if (!blk_atomic_write_mergeable_rqs(req, next)) 838 return NULL; 839 840 /* 841 * If we are allowed to merge, then append bio list 842 * from next to rq and release next. merge_requests_fn 843 * will have updated segment counts, update sector 844 * counts here. Handle DISCARDs separately, as they 845 * have separate settings. 846 */ 847 848 switch (blk_try_req_merge(req, next)) { 849 case ELEVATOR_DISCARD_MERGE: 850 if (!req_attempt_discard_merge(q, req, next)) 851 return NULL; 852 break; 853 case ELEVATOR_BACK_MERGE: 854 if (!ll_merge_requests_fn(q, req, next)) 855 return NULL; 856 break; 857 default: 858 return NULL; 859 } 860 861 /* 862 * If failfast settings disagree or any of the two is already 863 * a mixed merge, mark both as mixed before proceeding. This 864 * makes sure that all involved bios have mixable attributes 865 * set properly. 866 */ 867 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 868 (req->cmd_flags & REQ_FAILFAST_MASK) != 869 (next->cmd_flags & REQ_FAILFAST_MASK)) { 870 blk_rq_set_mixed_merge(req); 871 blk_rq_set_mixed_merge(next); 872 } 873 874 /* 875 * At this point we have either done a back merge or front merge. We 876 * need the smaller start_time_ns of the merged requests to be the 877 * current request for accounting purposes. 878 */ 879 if (next->start_time_ns < req->start_time_ns) 880 req->start_time_ns = next->start_time_ns; 881 882 req->biotail->bi_next = next->bio; 883 req->biotail = next->biotail; 884 885 req->__data_len += blk_rq_bytes(next); 886 887 if (!blk_discard_mergable(req)) 888 elv_merge_requests(q, req, next); 889 890 blk_crypto_rq_put_keyslot(next); 891 892 /* 893 * 'next' is going away, so update stats accordingly 894 */ 895 blk_account_io_merge_request(next); 896 897 trace_block_rq_merge(next); 898 899 /* 900 * ownership of bio passed from next to req, return 'next' for 901 * the caller to free 902 */ 903 next->bio = NULL; 904 return next; 905 } 906 907 static struct request *attempt_back_merge(struct request_queue *q, 908 struct request *rq) 909 { 910 struct request *next = elv_latter_request(q, rq); 911 912 if (next) 913 return attempt_merge(q, rq, next); 914 915 return NULL; 916 } 917 918 static struct request *attempt_front_merge(struct request_queue *q, 919 struct request *rq) 920 { 921 struct request *prev = elv_former_request(q, rq); 922 923 if (prev) 924 return attempt_merge(q, prev, rq); 925 926 return NULL; 927 } 928 929 /* 930 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 931 * otherwise. The caller is responsible for freeing 'next' if the merge 932 * happened. 933 */ 934 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 935 struct request *next) 936 { 937 return attempt_merge(q, rq, next); 938 } 939 940 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 941 { 942 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 943 return false; 944 945 if (req_op(rq) != bio_op(bio)) 946 return false; 947 948 if (!blk_cgroup_mergeable(rq, bio)) 949 return false; 950 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 951 return false; 952 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 953 return false; 954 if (rq->bio->bi_write_hint != bio->bi_write_hint) 955 return false; 956 if (rq->bio->bi_ioprio != bio->bi_ioprio) 957 return false; 958 if (blk_atomic_write_mergeable_rq_bio(rq, bio) == false) 959 return false; 960 961 return true; 962 } 963 964 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 965 { 966 if (blk_discard_mergable(rq)) 967 return ELEVATOR_DISCARD_MERGE; 968 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 969 return ELEVATOR_BACK_MERGE; 970 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 971 return ELEVATOR_FRONT_MERGE; 972 return ELEVATOR_NO_MERGE; 973 } 974 975 static void blk_account_io_merge_bio(struct request *req) 976 { 977 if (req->rq_flags & RQF_IO_STAT) { 978 part_stat_lock(); 979 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 980 part_stat_unlock(); 981 } 982 } 983 984 enum bio_merge_status bio_attempt_back_merge(struct request *req, 985 struct bio *bio, unsigned int nr_segs) 986 { 987 const blk_opf_t ff = bio_failfast(bio); 988 989 if (!ll_back_merge_fn(req, bio, nr_segs)) 990 return BIO_MERGE_FAILED; 991 992 trace_block_bio_backmerge(bio); 993 rq_qos_merge(req->q, req, bio); 994 995 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 996 blk_rq_set_mixed_merge(req); 997 998 blk_update_mixed_merge(req, bio, false); 999 1000 if (req->rq_flags & RQF_ZONE_WRITE_PLUGGING) 1001 blk_zone_write_plug_bio_merged(bio); 1002 1003 req->biotail->bi_next = bio; 1004 req->biotail = bio; 1005 req->__data_len += bio->bi_iter.bi_size; 1006 1007 bio_crypt_free_ctx(bio); 1008 1009 blk_account_io_merge_bio(req); 1010 return BIO_MERGE_OK; 1011 } 1012 1013 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 1014 struct bio *bio, unsigned int nr_segs) 1015 { 1016 const blk_opf_t ff = bio_failfast(bio); 1017 1018 /* 1019 * A front merge for writes to sequential zones of a zoned block device 1020 * can happen only if the user submitted writes out of order. Do not 1021 * merge such write to let it fail. 1022 */ 1023 if (req->rq_flags & RQF_ZONE_WRITE_PLUGGING) 1024 return BIO_MERGE_FAILED; 1025 1026 if (!ll_front_merge_fn(req, bio, nr_segs)) 1027 return BIO_MERGE_FAILED; 1028 1029 trace_block_bio_frontmerge(bio); 1030 rq_qos_merge(req->q, req, bio); 1031 1032 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 1033 blk_rq_set_mixed_merge(req); 1034 1035 blk_update_mixed_merge(req, bio, true); 1036 1037 bio->bi_next = req->bio; 1038 req->bio = bio; 1039 1040 req->__sector = bio->bi_iter.bi_sector; 1041 req->__data_len += bio->bi_iter.bi_size; 1042 1043 bio_crypt_do_front_merge(req, bio); 1044 1045 blk_account_io_merge_bio(req); 1046 return BIO_MERGE_OK; 1047 } 1048 1049 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 1050 struct request *req, struct bio *bio) 1051 { 1052 unsigned short segments = blk_rq_nr_discard_segments(req); 1053 1054 if (segments >= queue_max_discard_segments(q)) 1055 goto no_merge; 1056 if (blk_rq_sectors(req) + bio_sectors(bio) > 1057 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 1058 goto no_merge; 1059 1060 rq_qos_merge(q, req, bio); 1061 1062 req->biotail->bi_next = bio; 1063 req->biotail = bio; 1064 req->__data_len += bio->bi_iter.bi_size; 1065 req->nr_phys_segments = segments + 1; 1066 1067 blk_account_io_merge_bio(req); 1068 return BIO_MERGE_OK; 1069 no_merge: 1070 req_set_nomerge(q, req); 1071 return BIO_MERGE_FAILED; 1072 } 1073 1074 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1075 struct request *rq, 1076 struct bio *bio, 1077 unsigned int nr_segs, 1078 bool sched_allow_merge) 1079 { 1080 if (!blk_rq_merge_ok(rq, bio)) 1081 return BIO_MERGE_NONE; 1082 1083 switch (blk_try_merge(rq, bio)) { 1084 case ELEVATOR_BACK_MERGE: 1085 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1086 return bio_attempt_back_merge(rq, bio, nr_segs); 1087 break; 1088 case ELEVATOR_FRONT_MERGE: 1089 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1090 return bio_attempt_front_merge(rq, bio, nr_segs); 1091 break; 1092 case ELEVATOR_DISCARD_MERGE: 1093 return bio_attempt_discard_merge(q, rq, bio); 1094 default: 1095 return BIO_MERGE_NONE; 1096 } 1097 1098 return BIO_MERGE_FAILED; 1099 } 1100 1101 /** 1102 * blk_attempt_plug_merge - try to merge with %current's plugged list 1103 * @q: request_queue new bio is being queued at 1104 * @bio: new bio being queued 1105 * @nr_segs: number of segments in @bio 1106 * from the passed in @q already in the plug list 1107 * 1108 * Determine whether @bio being queued on @q can be merged with the previous 1109 * request on %current's plugged list. Returns %true if merge was successful, 1110 * otherwise %false. 1111 * 1112 * Plugging coalesces IOs from the same issuer for the same purpose without 1113 * going through @q->queue_lock. As such it's more of an issuing mechanism 1114 * than scheduling, and the request, while may have elvpriv data, is not 1115 * added on the elevator at this point. In addition, we don't have 1116 * reliable access to the elevator outside queue lock. Only check basic 1117 * merging parameters without querying the elevator. 1118 * 1119 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1120 */ 1121 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1122 unsigned int nr_segs) 1123 { 1124 struct blk_plug *plug = current->plug; 1125 struct request *rq; 1126 1127 if (!plug || rq_list_empty(&plug->mq_list)) 1128 return false; 1129 1130 rq_list_for_each(&plug->mq_list, rq) { 1131 if (rq->q == q) { 1132 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1133 BIO_MERGE_OK) 1134 return true; 1135 break; 1136 } 1137 1138 /* 1139 * Only keep iterating plug list for merges if we have multiple 1140 * queues 1141 */ 1142 if (!plug->multiple_queues) 1143 break; 1144 } 1145 return false; 1146 } 1147 1148 /* 1149 * Iterate list of requests and see if we can merge this bio with any 1150 * of them. 1151 */ 1152 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1153 struct bio *bio, unsigned int nr_segs) 1154 { 1155 struct request *rq; 1156 int checked = 8; 1157 1158 list_for_each_entry_reverse(rq, list, queuelist) { 1159 if (!checked--) 1160 break; 1161 1162 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1163 case BIO_MERGE_NONE: 1164 continue; 1165 case BIO_MERGE_OK: 1166 return true; 1167 case BIO_MERGE_FAILED: 1168 return false; 1169 } 1170 1171 } 1172 1173 return false; 1174 } 1175 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1176 1177 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1178 unsigned int nr_segs, struct request **merged_request) 1179 { 1180 struct request *rq; 1181 1182 switch (elv_merge(q, &rq, bio)) { 1183 case ELEVATOR_BACK_MERGE: 1184 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1185 return false; 1186 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1187 return false; 1188 *merged_request = attempt_back_merge(q, rq); 1189 if (!*merged_request) 1190 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1191 return true; 1192 case ELEVATOR_FRONT_MERGE: 1193 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1194 return false; 1195 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1196 return false; 1197 *merged_request = attempt_front_merge(q, rq); 1198 if (!*merged_request) 1199 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1200 return true; 1201 case ELEVATOR_DISCARD_MERGE: 1202 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1203 default: 1204 return false; 1205 } 1206 } 1207 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1208