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/scatterlist.h> 10 11 #include <trace/events/block.h> 12 13 #include "blk.h" 14 15 /* 16 * Check if the two bvecs from two bios can be merged to one segment. If yes, 17 * no need to check gap between the two bios since the 1st bio and the 1st bvec 18 * in the 2nd bio can be handled in one segment. 19 */ 20 static inline bool bios_segs_mergeable(struct request_queue *q, 21 struct bio *prev, struct bio_vec *prev_last_bv, 22 struct bio_vec *next_first_bv) 23 { 24 if (!biovec_phys_mergeable(q, prev_last_bv, next_first_bv)) 25 return false; 26 if (prev->bi_seg_back_size + next_first_bv->bv_len > 27 queue_max_segment_size(q)) 28 return false; 29 return true; 30 } 31 32 static inline bool bio_will_gap(struct request_queue *q, 33 struct request *prev_rq, struct bio *prev, struct bio *next) 34 { 35 struct bio_vec pb, nb; 36 37 if (!bio_has_data(prev) || !queue_virt_boundary(q)) 38 return false; 39 40 /* 41 * Don't merge if the 1st bio starts with non-zero offset, otherwise it 42 * is quite difficult to respect the sg gap limit. We work hard to 43 * merge a huge number of small single bios in case of mkfs. 44 */ 45 if (prev_rq) 46 bio_get_first_bvec(prev_rq->bio, &pb); 47 else 48 bio_get_first_bvec(prev, &pb); 49 if (pb.bv_offset & queue_virt_boundary(q)) 50 return true; 51 52 /* 53 * We don't need to worry about the situation that the merged segment 54 * ends in unaligned virt boundary: 55 * 56 * - if 'pb' ends aligned, the merged segment ends aligned 57 * - if 'pb' ends unaligned, the next bio must include 58 * one single bvec of 'nb', otherwise the 'nb' can't 59 * merge with 'pb' 60 */ 61 bio_get_last_bvec(prev, &pb); 62 bio_get_first_bvec(next, &nb); 63 if (bios_segs_mergeable(q, prev, &pb, &nb)) 64 return false; 65 return __bvec_gap_to_prev(q, &pb, nb.bv_offset); 66 } 67 68 static inline bool req_gap_back_merge(struct request *req, struct bio *bio) 69 { 70 return bio_will_gap(req->q, req, req->biotail, bio); 71 } 72 73 static inline bool req_gap_front_merge(struct request *req, struct bio *bio) 74 { 75 return bio_will_gap(req->q, NULL, bio, req->bio); 76 } 77 78 static struct bio *blk_bio_discard_split(struct request_queue *q, 79 struct bio *bio, 80 struct bio_set *bs, 81 unsigned *nsegs) 82 { 83 unsigned int max_discard_sectors, granularity; 84 int alignment; 85 sector_t tmp; 86 unsigned split_sectors; 87 88 *nsegs = 1; 89 90 /* Zero-sector (unknown) and one-sector granularities are the same. */ 91 granularity = max(q->limits.discard_granularity >> 9, 1U); 92 93 max_discard_sectors = min(q->limits.max_discard_sectors, 94 bio_allowed_max_sectors(q)); 95 max_discard_sectors -= max_discard_sectors % granularity; 96 97 if (unlikely(!max_discard_sectors)) { 98 /* XXX: warn */ 99 return NULL; 100 } 101 102 if (bio_sectors(bio) <= max_discard_sectors) 103 return NULL; 104 105 split_sectors = max_discard_sectors; 106 107 /* 108 * If the next starting sector would be misaligned, stop the discard at 109 * the previous aligned sector. 110 */ 111 alignment = (q->limits.discard_alignment >> 9) % granularity; 112 113 tmp = bio->bi_iter.bi_sector + split_sectors - alignment; 114 tmp = sector_div(tmp, granularity); 115 116 if (split_sectors > tmp) 117 split_sectors -= tmp; 118 119 return bio_split(bio, split_sectors, GFP_NOIO, bs); 120 } 121 122 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q, 123 struct bio *bio, struct bio_set *bs, unsigned *nsegs) 124 { 125 *nsegs = 1; 126 127 if (!q->limits.max_write_zeroes_sectors) 128 return NULL; 129 130 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) 131 return NULL; 132 133 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); 134 } 135 136 static struct bio *blk_bio_write_same_split(struct request_queue *q, 137 struct bio *bio, 138 struct bio_set *bs, 139 unsigned *nsegs) 140 { 141 *nsegs = 1; 142 143 if (!q->limits.max_write_same_sectors) 144 return NULL; 145 146 if (bio_sectors(bio) <= q->limits.max_write_same_sectors) 147 return NULL; 148 149 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); 150 } 151 152 static inline unsigned get_max_io_size(struct request_queue *q, 153 struct bio *bio) 154 { 155 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector); 156 unsigned mask = queue_logical_block_size(q) - 1; 157 158 /* aligned to logical block size */ 159 sectors &= ~(mask >> 9); 160 161 return sectors; 162 } 163 164 static struct bio *blk_bio_segment_split(struct request_queue *q, 165 struct bio *bio, 166 struct bio_set *bs, 167 unsigned *segs) 168 { 169 struct bio_vec bv, bvprv, *bvprvp = NULL; 170 struct bvec_iter iter; 171 unsigned seg_size = 0, nsegs = 0, sectors = 0; 172 unsigned front_seg_size = bio->bi_seg_front_size; 173 bool do_split = true; 174 struct bio *new = NULL; 175 const unsigned max_sectors = get_max_io_size(q, bio); 176 177 bio_for_each_segment(bv, bio, iter) { 178 /* 179 * If the queue doesn't support SG gaps and adding this 180 * offset would create a gap, disallow it. 181 */ 182 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) 183 goto split; 184 185 if (sectors + (bv.bv_len >> 9) > max_sectors) { 186 /* 187 * Consider this a new segment if we're splitting in 188 * the middle of this vector. 189 */ 190 if (nsegs < queue_max_segments(q) && 191 sectors < max_sectors) { 192 nsegs++; 193 sectors = max_sectors; 194 } 195 goto split; 196 } 197 198 if (bvprvp && blk_queue_cluster(q)) { 199 if (seg_size + bv.bv_len > queue_max_segment_size(q)) 200 goto new_segment; 201 if (!biovec_phys_mergeable(q, bvprvp, &bv)) 202 goto new_segment; 203 204 seg_size += bv.bv_len; 205 bvprv = bv; 206 bvprvp = &bvprv; 207 sectors += bv.bv_len >> 9; 208 209 continue; 210 } 211 new_segment: 212 if (nsegs == queue_max_segments(q)) 213 goto split; 214 215 if (nsegs == 1 && seg_size > front_seg_size) 216 front_seg_size = seg_size; 217 218 nsegs++; 219 bvprv = bv; 220 bvprvp = &bvprv; 221 seg_size = bv.bv_len; 222 sectors += bv.bv_len >> 9; 223 224 } 225 226 do_split = false; 227 split: 228 *segs = nsegs; 229 230 if (do_split) { 231 new = bio_split(bio, sectors, GFP_NOIO, bs); 232 if (new) 233 bio = new; 234 } 235 236 if (nsegs == 1 && seg_size > front_seg_size) 237 front_seg_size = seg_size; 238 bio->bi_seg_front_size = front_seg_size; 239 if (seg_size > bio->bi_seg_back_size) 240 bio->bi_seg_back_size = seg_size; 241 242 return do_split ? new : NULL; 243 } 244 245 void blk_queue_split(struct request_queue *q, struct bio **bio) 246 { 247 struct bio *split, *res; 248 unsigned nsegs; 249 250 switch (bio_op(*bio)) { 251 case REQ_OP_DISCARD: 252 case REQ_OP_SECURE_ERASE: 253 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs); 254 break; 255 case REQ_OP_WRITE_ZEROES: 256 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs); 257 break; 258 case REQ_OP_WRITE_SAME: 259 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs); 260 break; 261 default: 262 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs); 263 break; 264 } 265 266 /* physical segments can be figured out during splitting */ 267 res = split ? split : *bio; 268 res->bi_phys_segments = nsegs; 269 bio_set_flag(res, BIO_SEG_VALID); 270 271 if (split) { 272 /* there isn't chance to merge the splitted bio */ 273 split->bi_opf |= REQ_NOMERGE; 274 275 /* 276 * Since we're recursing into make_request here, ensure 277 * that we mark this bio as already having entered the queue. 278 * If not, and the queue is going away, we can get stuck 279 * forever on waiting for the queue reference to drop. But 280 * that will never happen, as we're already holding a 281 * reference to it. 282 */ 283 bio_set_flag(*bio, BIO_QUEUE_ENTERED); 284 285 bio_chain(split, *bio); 286 trace_block_split(q, split, (*bio)->bi_iter.bi_sector); 287 generic_make_request(*bio); 288 *bio = split; 289 } 290 } 291 EXPORT_SYMBOL(blk_queue_split); 292 293 static unsigned int __blk_recalc_rq_segments(struct request_queue *q, 294 struct bio *bio, 295 bool no_sg_merge) 296 { 297 struct bio_vec bv, bvprv = { NULL }; 298 int cluster, prev = 0; 299 unsigned int seg_size, nr_phys_segs; 300 struct bio *fbio, *bbio; 301 struct bvec_iter iter; 302 303 if (!bio) 304 return 0; 305 306 switch (bio_op(bio)) { 307 case REQ_OP_DISCARD: 308 case REQ_OP_SECURE_ERASE: 309 case REQ_OP_WRITE_ZEROES: 310 return 0; 311 case REQ_OP_WRITE_SAME: 312 return 1; 313 } 314 315 fbio = bio; 316 cluster = blk_queue_cluster(q); 317 seg_size = 0; 318 nr_phys_segs = 0; 319 for_each_bio(bio) { 320 bio_for_each_segment(bv, bio, iter) { 321 /* 322 * If SG merging is disabled, each bio vector is 323 * a segment 324 */ 325 if (no_sg_merge) 326 goto new_segment; 327 328 if (prev && cluster) { 329 if (seg_size + bv.bv_len 330 > queue_max_segment_size(q)) 331 goto new_segment; 332 if (!biovec_phys_mergeable(q, &bvprv, &bv)) 333 goto new_segment; 334 335 seg_size += bv.bv_len; 336 bvprv = bv; 337 continue; 338 } 339 new_segment: 340 if (nr_phys_segs == 1 && seg_size > 341 fbio->bi_seg_front_size) 342 fbio->bi_seg_front_size = seg_size; 343 344 nr_phys_segs++; 345 bvprv = bv; 346 prev = 1; 347 seg_size = bv.bv_len; 348 } 349 bbio = bio; 350 } 351 352 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size) 353 fbio->bi_seg_front_size = seg_size; 354 if (seg_size > bbio->bi_seg_back_size) 355 bbio->bi_seg_back_size = seg_size; 356 357 return nr_phys_segs; 358 } 359 360 void blk_recalc_rq_segments(struct request *rq) 361 { 362 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE, 363 &rq->q->queue_flags); 364 365 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio, 366 no_sg_merge); 367 } 368 369 void blk_recount_segments(struct request_queue *q, struct bio *bio) 370 { 371 unsigned short seg_cnt; 372 373 /* estimate segment number by bi_vcnt for non-cloned bio */ 374 if (bio_flagged(bio, BIO_CLONED)) 375 seg_cnt = bio_segments(bio); 376 else 377 seg_cnt = bio->bi_vcnt; 378 379 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) && 380 (seg_cnt < queue_max_segments(q))) 381 bio->bi_phys_segments = seg_cnt; 382 else { 383 struct bio *nxt = bio->bi_next; 384 385 bio->bi_next = NULL; 386 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false); 387 bio->bi_next = nxt; 388 } 389 390 bio_set_flag(bio, BIO_SEG_VALID); 391 } 392 393 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio, 394 struct bio *nxt) 395 { 396 struct bio_vec end_bv = { NULL }, nxt_bv; 397 398 if (!blk_queue_cluster(q)) 399 return 0; 400 401 if (bio->bi_seg_back_size + nxt->bi_seg_front_size > 402 queue_max_segment_size(q)) 403 return 0; 404 405 if (!bio_has_data(bio)) 406 return 1; 407 408 bio_get_last_bvec(bio, &end_bv); 409 bio_get_first_bvec(nxt, &nxt_bv); 410 411 return biovec_phys_mergeable(q, &end_bv, &nxt_bv); 412 } 413 414 static inline void 415 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec, 416 struct scatterlist *sglist, struct bio_vec *bvprv, 417 struct scatterlist **sg, int *nsegs, int *cluster) 418 { 419 420 int nbytes = bvec->bv_len; 421 422 if (*sg && *cluster) { 423 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 424 goto new_segment; 425 if (!biovec_phys_mergeable(q, bvprv, bvec)) 426 goto new_segment; 427 428 (*sg)->length += nbytes; 429 } else { 430 new_segment: 431 if (!*sg) 432 *sg = sglist; 433 else { 434 /* 435 * If the driver previously mapped a shorter 436 * list, we could see a termination bit 437 * prematurely unless it fully inits the sg 438 * table on each mapping. We KNOW that there 439 * must be more entries here or the driver 440 * would be buggy, so force clear the 441 * termination bit to avoid doing a full 442 * sg_init_table() in drivers for each command. 443 */ 444 sg_unmark_end(*sg); 445 *sg = sg_next(*sg); 446 } 447 448 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset); 449 (*nsegs)++; 450 } 451 *bvprv = *bvec; 452 } 453 454 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv, 455 struct scatterlist *sglist, struct scatterlist **sg) 456 { 457 *sg = sglist; 458 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 459 return 1; 460 } 461 462 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 463 struct scatterlist *sglist, 464 struct scatterlist **sg) 465 { 466 struct bio_vec bvec, bvprv = { NULL }; 467 struct bvec_iter iter; 468 int cluster = blk_queue_cluster(q), nsegs = 0; 469 470 for_each_bio(bio) 471 bio_for_each_segment(bvec, bio, iter) 472 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg, 473 &nsegs, &cluster); 474 475 return nsegs; 476 } 477 478 /* 479 * map a request to scatterlist, return number of sg entries setup. Caller 480 * must make sure sg can hold rq->nr_phys_segments entries 481 */ 482 int blk_rq_map_sg(struct request_queue *q, struct request *rq, 483 struct scatterlist *sglist) 484 { 485 struct scatterlist *sg = NULL; 486 int nsegs = 0; 487 488 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 489 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg); 490 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME) 491 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg); 492 else if (rq->bio) 493 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg); 494 495 if (unlikely(rq->rq_flags & RQF_COPY_USER) && 496 (blk_rq_bytes(rq) & q->dma_pad_mask)) { 497 unsigned int pad_len = 498 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1; 499 500 sg->length += pad_len; 501 rq->extra_len += pad_len; 502 } 503 504 if (q->dma_drain_size && q->dma_drain_needed(rq)) { 505 if (op_is_write(req_op(rq))) 506 memset(q->dma_drain_buffer, 0, q->dma_drain_size); 507 508 sg_unmark_end(sg); 509 sg = sg_next(sg); 510 sg_set_page(sg, virt_to_page(q->dma_drain_buffer), 511 q->dma_drain_size, 512 ((unsigned long)q->dma_drain_buffer) & 513 (PAGE_SIZE - 1)); 514 nsegs++; 515 rq->extra_len += q->dma_drain_size; 516 } 517 518 if (sg) 519 sg_mark_end(sg); 520 521 /* 522 * Something must have been wrong if the figured number of 523 * segment is bigger than number of req's physical segments 524 */ 525 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 526 527 return nsegs; 528 } 529 EXPORT_SYMBOL(blk_rq_map_sg); 530 531 static inline int ll_new_hw_segment(struct request_queue *q, 532 struct request *req, 533 struct bio *bio) 534 { 535 int nr_phys_segs = bio_phys_segments(q, bio); 536 537 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q)) 538 goto no_merge; 539 540 if (blk_integrity_merge_bio(q, req, bio) == false) 541 goto no_merge; 542 543 /* 544 * This will form the start of a new hw segment. Bump both 545 * counters. 546 */ 547 req->nr_phys_segments += nr_phys_segs; 548 return 1; 549 550 no_merge: 551 req_set_nomerge(q, req); 552 return 0; 553 } 554 555 int ll_back_merge_fn(struct request_queue *q, struct request *req, 556 struct bio *bio) 557 { 558 if (req_gap_back_merge(req, bio)) 559 return 0; 560 if (blk_integrity_rq(req) && 561 integrity_req_gap_back_merge(req, bio)) 562 return 0; 563 if (blk_rq_sectors(req) + bio_sectors(bio) > 564 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 565 req_set_nomerge(q, req); 566 return 0; 567 } 568 if (!bio_flagged(req->biotail, BIO_SEG_VALID)) 569 blk_recount_segments(q, req->biotail); 570 if (!bio_flagged(bio, BIO_SEG_VALID)) 571 blk_recount_segments(q, bio); 572 573 return ll_new_hw_segment(q, req, bio); 574 } 575 576 int ll_front_merge_fn(struct request_queue *q, struct request *req, 577 struct bio *bio) 578 { 579 580 if (req_gap_front_merge(req, bio)) 581 return 0; 582 if (blk_integrity_rq(req) && 583 integrity_req_gap_front_merge(req, bio)) 584 return 0; 585 if (blk_rq_sectors(req) + bio_sectors(bio) > 586 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 587 req_set_nomerge(q, req); 588 return 0; 589 } 590 if (!bio_flagged(bio, BIO_SEG_VALID)) 591 blk_recount_segments(q, bio); 592 if (!bio_flagged(req->bio, BIO_SEG_VALID)) 593 blk_recount_segments(q, req->bio); 594 595 return ll_new_hw_segment(q, req, bio); 596 } 597 598 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 599 struct request *next) 600 { 601 unsigned short segments = blk_rq_nr_discard_segments(req); 602 603 if (segments >= queue_max_discard_segments(q)) 604 goto no_merge; 605 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 606 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 607 goto no_merge; 608 609 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 610 return true; 611 no_merge: 612 req_set_nomerge(q, req); 613 return false; 614 } 615 616 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 617 struct request *next) 618 { 619 int total_phys_segments; 620 unsigned int seg_size = 621 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size; 622 623 if (req_gap_back_merge(req, next->bio)) 624 return 0; 625 626 /* 627 * Will it become too large? 628 */ 629 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 630 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 631 return 0; 632 633 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 634 if (blk_phys_contig_segment(q, req->biotail, next->bio)) { 635 if (req->nr_phys_segments == 1) 636 req->bio->bi_seg_front_size = seg_size; 637 if (next->nr_phys_segments == 1) 638 next->biotail->bi_seg_back_size = seg_size; 639 total_phys_segments--; 640 } 641 642 if (total_phys_segments > queue_max_segments(q)) 643 return 0; 644 645 if (blk_integrity_merge_rq(q, req, next) == false) 646 return 0; 647 648 /* Merge is OK... */ 649 req->nr_phys_segments = total_phys_segments; 650 return 1; 651 } 652 653 /** 654 * blk_rq_set_mixed_merge - mark a request as mixed merge 655 * @rq: request to mark as mixed merge 656 * 657 * Description: 658 * @rq is about to be mixed merged. Make sure the attributes 659 * which can be mixed are set in each bio and mark @rq as mixed 660 * merged. 661 */ 662 void blk_rq_set_mixed_merge(struct request *rq) 663 { 664 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; 665 struct bio *bio; 666 667 if (rq->rq_flags & RQF_MIXED_MERGE) 668 return; 669 670 /* 671 * @rq will no longer represent mixable attributes for all the 672 * contained bios. It will just track those of the first one. 673 * Distributes the attributs to each bio. 674 */ 675 for (bio = rq->bio; bio; bio = bio->bi_next) { 676 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 677 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 678 bio->bi_opf |= ff; 679 } 680 rq->rq_flags |= RQF_MIXED_MERGE; 681 } 682 683 static void blk_account_io_merge(struct request *req) 684 { 685 if (blk_do_io_stat(req)) { 686 struct hd_struct *part; 687 688 part_stat_lock(); 689 part = req->part; 690 691 part_dec_in_flight(req->q, part, rq_data_dir(req)); 692 693 hd_struct_put(part); 694 part_stat_unlock(); 695 } 696 } 697 /* 698 * Two cases of handling DISCARD merge: 699 * If max_discard_segments > 1, the driver takes every bio 700 * as a range and send them to controller together. The ranges 701 * needn't to be contiguous. 702 * Otherwise, the bios/requests will be handled as same as 703 * others which should be contiguous. 704 */ 705 static inline bool blk_discard_mergable(struct request *req) 706 { 707 if (req_op(req) == REQ_OP_DISCARD && 708 queue_max_discard_segments(req->q) > 1) 709 return true; 710 return false; 711 } 712 713 static enum elv_merge blk_try_req_merge(struct request *req, 714 struct request *next) 715 { 716 if (blk_discard_mergable(req)) 717 return ELEVATOR_DISCARD_MERGE; 718 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 719 return ELEVATOR_BACK_MERGE; 720 721 return ELEVATOR_NO_MERGE; 722 } 723 724 /* 725 * For non-mq, this has to be called with the request spinlock acquired. 726 * For mq with scheduling, the appropriate queue wide lock should be held. 727 */ 728 static struct request *attempt_merge(struct request_queue *q, 729 struct request *req, struct request *next) 730 { 731 if (!rq_mergeable(req) || !rq_mergeable(next)) 732 return NULL; 733 734 if (req_op(req) != req_op(next)) 735 return NULL; 736 737 if (rq_data_dir(req) != rq_data_dir(next) 738 || req->rq_disk != next->rq_disk) 739 return NULL; 740 741 if (req_op(req) == REQ_OP_WRITE_SAME && 742 !blk_write_same_mergeable(req->bio, next->bio)) 743 return NULL; 744 745 /* 746 * Don't allow merge of different write hints, or for a hint with 747 * non-hint IO. 748 */ 749 if (req->write_hint != next->write_hint) 750 return NULL; 751 752 if (req->ioprio != next->ioprio) 753 return NULL; 754 755 /* 756 * If we are allowed to merge, then append bio list 757 * from next to rq and release next. merge_requests_fn 758 * will have updated segment counts, update sector 759 * counts here. Handle DISCARDs separately, as they 760 * have separate settings. 761 */ 762 763 switch (blk_try_req_merge(req, next)) { 764 case ELEVATOR_DISCARD_MERGE: 765 if (!req_attempt_discard_merge(q, req, next)) 766 return NULL; 767 break; 768 case ELEVATOR_BACK_MERGE: 769 if (!ll_merge_requests_fn(q, req, next)) 770 return NULL; 771 break; 772 default: 773 return NULL; 774 } 775 776 /* 777 * If failfast settings disagree or any of the two is already 778 * a mixed merge, mark both as mixed before proceeding. This 779 * makes sure that all involved bios have mixable attributes 780 * set properly. 781 */ 782 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 783 (req->cmd_flags & REQ_FAILFAST_MASK) != 784 (next->cmd_flags & REQ_FAILFAST_MASK)) { 785 blk_rq_set_mixed_merge(req); 786 blk_rq_set_mixed_merge(next); 787 } 788 789 /* 790 * At this point we have either done a back merge or front merge. We 791 * need the smaller start_time_ns of the merged requests to be the 792 * current request for accounting purposes. 793 */ 794 if (next->start_time_ns < req->start_time_ns) 795 req->start_time_ns = next->start_time_ns; 796 797 req->biotail->bi_next = next->bio; 798 req->biotail = next->biotail; 799 800 req->__data_len += blk_rq_bytes(next); 801 802 if (!blk_discard_mergable(req)) 803 elv_merge_requests(q, req, next); 804 805 /* 806 * 'next' is going away, so update stats accordingly 807 */ 808 blk_account_io_merge(next); 809 810 /* 811 * ownership of bio passed from next to req, return 'next' for 812 * the caller to free 813 */ 814 next->bio = NULL; 815 return next; 816 } 817 818 struct request *attempt_back_merge(struct request_queue *q, struct request *rq) 819 { 820 struct request *next = elv_latter_request(q, rq); 821 822 if (next) 823 return attempt_merge(q, rq, next); 824 825 return NULL; 826 } 827 828 struct request *attempt_front_merge(struct request_queue *q, struct request *rq) 829 { 830 struct request *prev = elv_former_request(q, rq); 831 832 if (prev) 833 return attempt_merge(q, prev, rq); 834 835 return NULL; 836 } 837 838 int blk_attempt_req_merge(struct request_queue *q, struct request *rq, 839 struct request *next) 840 { 841 struct request *free; 842 843 free = attempt_merge(q, rq, next); 844 if (free) { 845 blk_put_request(free); 846 return 1; 847 } 848 849 return 0; 850 } 851 852 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 853 { 854 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 855 return false; 856 857 if (req_op(rq) != bio_op(bio)) 858 return false; 859 860 /* different data direction or already started, don't merge */ 861 if (bio_data_dir(bio) != rq_data_dir(rq)) 862 return false; 863 864 /* must be same device */ 865 if (rq->rq_disk != bio->bi_disk) 866 return false; 867 868 /* only merge integrity protected bio into ditto rq */ 869 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 870 return false; 871 872 /* must be using the same buffer */ 873 if (req_op(rq) == REQ_OP_WRITE_SAME && 874 !blk_write_same_mergeable(rq->bio, bio)) 875 return false; 876 877 /* 878 * Don't allow merge of different write hints, or for a hint with 879 * non-hint IO. 880 */ 881 if (rq->write_hint != bio->bi_write_hint) 882 return false; 883 884 if (rq->ioprio != bio_prio(bio)) 885 return false; 886 887 return true; 888 } 889 890 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 891 { 892 if (blk_discard_mergable(rq)) 893 return ELEVATOR_DISCARD_MERGE; 894 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 895 return ELEVATOR_BACK_MERGE; 896 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 897 return ELEVATOR_FRONT_MERGE; 898 return ELEVATOR_NO_MERGE; 899 } 900