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