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