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 *split; 335 336 switch (bio_op(bio)) { 337 case REQ_OP_DISCARD: 338 case REQ_OP_SECURE_ERASE: 339 split = bio_split_discard(bio, q, nr_segs, &q->bio_split); 340 break; 341 case REQ_OP_WRITE_ZEROES: 342 split = bio_split_write_zeroes(bio, q, nr_segs, &q->bio_split); 343 break; 344 default: 345 split = bio_split_rw(bio, q, nr_segs, &q->bio_split); 346 break; 347 } 348 349 if (split) { 350 /* there isn't chance to merge the splitted bio */ 351 split->bi_opf |= REQ_NOMERGE; 352 353 blkcg_bio_issue_init(split); 354 bio_chain(split, bio); 355 trace_block_split(split, bio->bi_iter.bi_sector); 356 submit_bio_noacct(bio); 357 return split; 358 } 359 return bio; 360 } 361 362 /** 363 * bio_split_to_limits - split a bio to fit the queue limits 364 * @bio: bio to be split 365 * 366 * Check if @bio needs splitting based on the queue limits of @bio->bi_bdev, and 367 * if so split off a bio fitting the limits from the beginning of @bio and 368 * return it. @bio is shortened to the remainder and re-submitted. 369 * 370 * The split bio is allocated from @q->bio_split, which is provided by the 371 * block layer. 372 */ 373 struct bio *bio_split_to_limits(struct bio *bio) 374 { 375 struct request_queue *q = bdev_get_queue(bio->bi_bdev); 376 unsigned int nr_segs; 377 378 if (bio_may_exceed_limits(bio, q)) 379 return __bio_split_to_limits(bio, q, &nr_segs); 380 return bio; 381 } 382 EXPORT_SYMBOL(bio_split_to_limits); 383 384 unsigned int blk_recalc_rq_segments(struct request *rq) 385 { 386 unsigned int nr_phys_segs = 0; 387 unsigned int bytes = 0; 388 struct req_iterator iter; 389 struct bio_vec bv; 390 391 if (!rq->bio) 392 return 0; 393 394 switch (bio_op(rq->bio)) { 395 case REQ_OP_DISCARD: 396 case REQ_OP_SECURE_ERASE: 397 if (queue_max_discard_segments(rq->q) > 1) { 398 struct bio *bio = rq->bio; 399 400 for_each_bio(bio) 401 nr_phys_segs++; 402 return nr_phys_segs; 403 } 404 return 1; 405 case REQ_OP_WRITE_ZEROES: 406 return 0; 407 default: 408 break; 409 } 410 411 rq_for_each_bvec(bv, rq, iter) 412 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &bytes, 413 UINT_MAX, UINT_MAX); 414 return nr_phys_segs; 415 } 416 417 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 418 struct scatterlist *sglist) 419 { 420 if (!*sg) 421 return sglist; 422 423 /* 424 * If the driver previously mapped a shorter list, we could see a 425 * termination bit prematurely unless it fully inits the sg table 426 * on each mapping. We KNOW that there must be more entries here 427 * or the driver would be buggy, so force clear the termination bit 428 * to avoid doing a full sg_init_table() in drivers for each command. 429 */ 430 sg_unmark_end(*sg); 431 return sg_next(*sg); 432 } 433 434 static unsigned blk_bvec_map_sg(struct request_queue *q, 435 struct bio_vec *bvec, struct scatterlist *sglist, 436 struct scatterlist **sg) 437 { 438 unsigned nbytes = bvec->bv_len; 439 unsigned nsegs = 0, total = 0; 440 441 while (nbytes > 0) { 442 unsigned offset = bvec->bv_offset + total; 443 unsigned len = min(get_max_segment_size(q, bvec->bv_page, 444 offset), nbytes); 445 struct page *page = bvec->bv_page; 446 447 /* 448 * Unfortunately a fair number of drivers barf on scatterlists 449 * that have an offset larger than PAGE_SIZE, despite other 450 * subsystems dealing with that invariant just fine. For now 451 * stick to the legacy format where we never present those from 452 * the block layer, but the code below should be removed once 453 * these offenders (mostly MMC/SD drivers) are fixed. 454 */ 455 page += (offset >> PAGE_SHIFT); 456 offset &= ~PAGE_MASK; 457 458 *sg = blk_next_sg(sg, sglist); 459 sg_set_page(*sg, page, len, offset); 460 461 total += len; 462 nbytes -= len; 463 nsegs++; 464 } 465 466 return nsegs; 467 } 468 469 static inline int __blk_bvec_map_sg(struct bio_vec bv, 470 struct scatterlist *sglist, struct scatterlist **sg) 471 { 472 *sg = blk_next_sg(sg, sglist); 473 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 474 return 1; 475 } 476 477 /* only try to merge bvecs into one sg if they are from two bios */ 478 static inline bool 479 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec, 480 struct bio_vec *bvprv, struct scatterlist **sg) 481 { 482 483 int nbytes = bvec->bv_len; 484 485 if (!*sg) 486 return false; 487 488 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 489 return false; 490 491 if (!biovec_phys_mergeable(q, bvprv, bvec)) 492 return false; 493 494 (*sg)->length += nbytes; 495 496 return true; 497 } 498 499 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 500 struct scatterlist *sglist, 501 struct scatterlist **sg) 502 { 503 struct bio_vec bvec, bvprv = { NULL }; 504 struct bvec_iter iter; 505 int nsegs = 0; 506 bool new_bio = false; 507 508 for_each_bio(bio) { 509 bio_for_each_bvec(bvec, bio, iter) { 510 /* 511 * Only try to merge bvecs from two bios given we 512 * have done bio internal merge when adding pages 513 * to bio 514 */ 515 if (new_bio && 516 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg)) 517 goto next_bvec; 518 519 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE) 520 nsegs += __blk_bvec_map_sg(bvec, sglist, sg); 521 else 522 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg); 523 next_bvec: 524 new_bio = false; 525 } 526 if (likely(bio->bi_iter.bi_size)) { 527 bvprv = bvec; 528 new_bio = true; 529 } 530 } 531 532 return nsegs; 533 } 534 535 /* 536 * map a request to scatterlist, return number of sg entries setup. Caller 537 * must make sure sg can hold rq->nr_phys_segments entries 538 */ 539 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 540 struct scatterlist *sglist, struct scatterlist **last_sg) 541 { 542 int nsegs = 0; 543 544 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 545 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg); 546 else if (rq->bio) 547 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg); 548 549 if (*last_sg) 550 sg_mark_end(*last_sg); 551 552 /* 553 * Something must have been wrong if the figured number of 554 * segment is bigger than number of req's physical segments 555 */ 556 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 557 558 return nsegs; 559 } 560 EXPORT_SYMBOL(__blk_rq_map_sg); 561 562 static inline unsigned int blk_rq_get_max_segments(struct request *rq) 563 { 564 if (req_op(rq) == REQ_OP_DISCARD) 565 return queue_max_discard_segments(rq->q); 566 return queue_max_segments(rq->q); 567 } 568 569 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 570 sector_t offset) 571 { 572 struct request_queue *q = rq->q; 573 unsigned int max_sectors; 574 575 if (blk_rq_is_passthrough(rq)) 576 return q->limits.max_hw_sectors; 577 578 max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); 579 if (!q->limits.chunk_sectors || 580 req_op(rq) == REQ_OP_DISCARD || 581 req_op(rq) == REQ_OP_SECURE_ERASE) 582 return max_sectors; 583 return min(max_sectors, 584 blk_chunk_sectors_left(offset, q->limits.chunk_sectors)); 585 } 586 587 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 588 unsigned int nr_phys_segs) 589 { 590 if (!blk_cgroup_mergeable(req, bio)) 591 goto no_merge; 592 593 if (blk_integrity_merge_bio(req->q, req, bio) == false) 594 goto no_merge; 595 596 /* discard request merge won't add new segment */ 597 if (req_op(req) == REQ_OP_DISCARD) 598 return 1; 599 600 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 601 goto no_merge; 602 603 /* 604 * This will form the start of a new hw segment. Bump both 605 * counters. 606 */ 607 req->nr_phys_segments += nr_phys_segs; 608 return 1; 609 610 no_merge: 611 req_set_nomerge(req->q, req); 612 return 0; 613 } 614 615 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 616 { 617 if (req_gap_back_merge(req, bio)) 618 return 0; 619 if (blk_integrity_rq(req) && 620 integrity_req_gap_back_merge(req, bio)) 621 return 0; 622 if (!bio_crypt_ctx_back_mergeable(req, bio)) 623 return 0; 624 if (blk_rq_sectors(req) + bio_sectors(bio) > 625 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 626 req_set_nomerge(req->q, req); 627 return 0; 628 } 629 630 return ll_new_hw_segment(req, bio, nr_segs); 631 } 632 633 static int ll_front_merge_fn(struct request *req, struct bio *bio, 634 unsigned int nr_segs) 635 { 636 if (req_gap_front_merge(req, bio)) 637 return 0; 638 if (blk_integrity_rq(req) && 639 integrity_req_gap_front_merge(req, bio)) 640 return 0; 641 if (!bio_crypt_ctx_front_mergeable(req, bio)) 642 return 0; 643 if (blk_rq_sectors(req) + bio_sectors(bio) > 644 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 645 req_set_nomerge(req->q, req); 646 return 0; 647 } 648 649 return ll_new_hw_segment(req, bio, nr_segs); 650 } 651 652 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 653 struct request *next) 654 { 655 unsigned short segments = blk_rq_nr_discard_segments(req); 656 657 if (segments >= queue_max_discard_segments(q)) 658 goto no_merge; 659 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 660 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 661 goto no_merge; 662 663 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 664 return true; 665 no_merge: 666 req_set_nomerge(q, req); 667 return false; 668 } 669 670 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 671 struct request *next) 672 { 673 int total_phys_segments; 674 675 if (req_gap_back_merge(req, next->bio)) 676 return 0; 677 678 /* 679 * Will it become too large? 680 */ 681 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 682 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 683 return 0; 684 685 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 686 if (total_phys_segments > blk_rq_get_max_segments(req)) 687 return 0; 688 689 if (!blk_cgroup_mergeable(req, next->bio)) 690 return 0; 691 692 if (blk_integrity_merge_rq(q, req, next) == false) 693 return 0; 694 695 if (!bio_crypt_ctx_merge_rq(req, next)) 696 return 0; 697 698 /* Merge is OK... */ 699 req->nr_phys_segments = total_phys_segments; 700 return 1; 701 } 702 703 /** 704 * blk_rq_set_mixed_merge - mark a request as mixed merge 705 * @rq: request to mark as mixed merge 706 * 707 * Description: 708 * @rq is about to be mixed merged. Make sure the attributes 709 * which can be mixed are set in each bio and mark @rq as mixed 710 * merged. 711 */ 712 void blk_rq_set_mixed_merge(struct request *rq) 713 { 714 blk_opf_t ff = rq->cmd_flags & REQ_FAILFAST_MASK; 715 struct bio *bio; 716 717 if (rq->rq_flags & RQF_MIXED_MERGE) 718 return; 719 720 /* 721 * @rq will no longer represent mixable attributes for all the 722 * contained bios. It will just track those of the first one. 723 * Distributes the attributs to each bio. 724 */ 725 for (bio = rq->bio; bio; bio = bio->bi_next) { 726 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 727 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 728 bio->bi_opf |= ff; 729 } 730 rq->rq_flags |= RQF_MIXED_MERGE; 731 } 732 733 static void blk_account_io_merge_request(struct request *req) 734 { 735 if (blk_do_io_stat(req)) { 736 part_stat_lock(); 737 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 738 part_stat_unlock(); 739 } 740 } 741 742 static enum elv_merge blk_try_req_merge(struct request *req, 743 struct request *next) 744 { 745 if (blk_discard_mergable(req)) 746 return ELEVATOR_DISCARD_MERGE; 747 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 748 return ELEVATOR_BACK_MERGE; 749 750 return ELEVATOR_NO_MERGE; 751 } 752 753 /* 754 * For non-mq, this has to be called with the request spinlock acquired. 755 * For mq with scheduling, the appropriate queue wide lock should be held. 756 */ 757 static struct request *attempt_merge(struct request_queue *q, 758 struct request *req, struct request *next) 759 { 760 if (!rq_mergeable(req) || !rq_mergeable(next)) 761 return NULL; 762 763 if (req_op(req) != req_op(next)) 764 return NULL; 765 766 if (rq_data_dir(req) != rq_data_dir(next)) 767 return NULL; 768 769 if (req->ioprio != next->ioprio) 770 return NULL; 771 772 /* 773 * If we are allowed to merge, then append bio list 774 * from next to rq and release next. merge_requests_fn 775 * will have updated segment counts, update sector 776 * counts here. Handle DISCARDs separately, as they 777 * have separate settings. 778 */ 779 780 switch (blk_try_req_merge(req, next)) { 781 case ELEVATOR_DISCARD_MERGE: 782 if (!req_attempt_discard_merge(q, req, next)) 783 return NULL; 784 break; 785 case ELEVATOR_BACK_MERGE: 786 if (!ll_merge_requests_fn(q, req, next)) 787 return NULL; 788 break; 789 default: 790 return NULL; 791 } 792 793 /* 794 * If failfast settings disagree or any of the two is already 795 * a mixed merge, mark both as mixed before proceeding. This 796 * makes sure that all involved bios have mixable attributes 797 * set properly. 798 */ 799 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 800 (req->cmd_flags & REQ_FAILFAST_MASK) != 801 (next->cmd_flags & REQ_FAILFAST_MASK)) { 802 blk_rq_set_mixed_merge(req); 803 blk_rq_set_mixed_merge(next); 804 } 805 806 /* 807 * At this point we have either done a back merge or front merge. We 808 * need the smaller start_time_ns of the merged requests to be the 809 * current request for accounting purposes. 810 */ 811 if (next->start_time_ns < req->start_time_ns) 812 req->start_time_ns = next->start_time_ns; 813 814 req->biotail->bi_next = next->bio; 815 req->biotail = next->biotail; 816 817 req->__data_len += blk_rq_bytes(next); 818 819 if (!blk_discard_mergable(req)) 820 elv_merge_requests(q, req, next); 821 822 /* 823 * 'next' is going away, so update stats accordingly 824 */ 825 blk_account_io_merge_request(next); 826 827 trace_block_rq_merge(next); 828 829 /* 830 * ownership of bio passed from next to req, return 'next' for 831 * the caller to free 832 */ 833 next->bio = NULL; 834 return next; 835 } 836 837 static struct request *attempt_back_merge(struct request_queue *q, 838 struct request *rq) 839 { 840 struct request *next = elv_latter_request(q, rq); 841 842 if (next) 843 return attempt_merge(q, rq, next); 844 845 return NULL; 846 } 847 848 static struct request *attempt_front_merge(struct request_queue *q, 849 struct request *rq) 850 { 851 struct request *prev = elv_former_request(q, rq); 852 853 if (prev) 854 return attempt_merge(q, prev, rq); 855 856 return NULL; 857 } 858 859 /* 860 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 861 * otherwise. The caller is responsible for freeing 'next' if the merge 862 * happened. 863 */ 864 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 865 struct request *next) 866 { 867 return attempt_merge(q, rq, next); 868 } 869 870 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 871 { 872 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 873 return false; 874 875 if (req_op(rq) != bio_op(bio)) 876 return false; 877 878 /* different data direction or already started, don't merge */ 879 if (bio_data_dir(bio) != rq_data_dir(rq)) 880 return false; 881 882 /* don't merge across cgroup boundaries */ 883 if (!blk_cgroup_mergeable(rq, bio)) 884 return false; 885 886 /* only merge integrity protected bio into ditto rq */ 887 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 888 return false; 889 890 /* Only merge if the crypt contexts are compatible */ 891 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 892 return false; 893 894 if (rq->ioprio != bio_prio(bio)) 895 return false; 896 897 return true; 898 } 899 900 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 901 { 902 if (blk_discard_mergable(rq)) 903 return ELEVATOR_DISCARD_MERGE; 904 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 905 return ELEVATOR_BACK_MERGE; 906 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 907 return ELEVATOR_FRONT_MERGE; 908 return ELEVATOR_NO_MERGE; 909 } 910 911 static void blk_account_io_merge_bio(struct request *req) 912 { 913 if (!blk_do_io_stat(req)) 914 return; 915 916 part_stat_lock(); 917 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 918 part_stat_unlock(); 919 } 920 921 enum bio_merge_status { 922 BIO_MERGE_OK, 923 BIO_MERGE_NONE, 924 BIO_MERGE_FAILED, 925 }; 926 927 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 928 struct bio *bio, unsigned int nr_segs) 929 { 930 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 931 932 if (!ll_back_merge_fn(req, bio, nr_segs)) 933 return BIO_MERGE_FAILED; 934 935 trace_block_bio_backmerge(bio); 936 rq_qos_merge(req->q, req, bio); 937 938 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 939 blk_rq_set_mixed_merge(req); 940 941 req->biotail->bi_next = bio; 942 req->biotail = bio; 943 req->__data_len += bio->bi_iter.bi_size; 944 945 bio_crypt_free_ctx(bio); 946 947 blk_account_io_merge_bio(req); 948 return BIO_MERGE_OK; 949 } 950 951 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 952 struct bio *bio, unsigned int nr_segs) 953 { 954 const blk_opf_t ff = bio->bi_opf & REQ_FAILFAST_MASK; 955 956 if (!ll_front_merge_fn(req, bio, nr_segs)) 957 return BIO_MERGE_FAILED; 958 959 trace_block_bio_frontmerge(bio); 960 rq_qos_merge(req->q, req, bio); 961 962 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 963 blk_rq_set_mixed_merge(req); 964 965 bio->bi_next = req->bio; 966 req->bio = bio; 967 968 req->__sector = bio->bi_iter.bi_sector; 969 req->__data_len += bio->bi_iter.bi_size; 970 971 bio_crypt_do_front_merge(req, bio); 972 973 blk_account_io_merge_bio(req); 974 return BIO_MERGE_OK; 975 } 976 977 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 978 struct request *req, struct bio *bio) 979 { 980 unsigned short segments = blk_rq_nr_discard_segments(req); 981 982 if (segments >= queue_max_discard_segments(q)) 983 goto no_merge; 984 if (blk_rq_sectors(req) + bio_sectors(bio) > 985 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 986 goto no_merge; 987 988 rq_qos_merge(q, req, bio); 989 990 req->biotail->bi_next = bio; 991 req->biotail = bio; 992 req->__data_len += bio->bi_iter.bi_size; 993 req->nr_phys_segments = segments + 1; 994 995 blk_account_io_merge_bio(req); 996 return BIO_MERGE_OK; 997 no_merge: 998 req_set_nomerge(q, req); 999 return BIO_MERGE_FAILED; 1000 } 1001 1002 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1003 struct request *rq, 1004 struct bio *bio, 1005 unsigned int nr_segs, 1006 bool sched_allow_merge) 1007 { 1008 if (!blk_rq_merge_ok(rq, bio)) 1009 return BIO_MERGE_NONE; 1010 1011 switch (blk_try_merge(rq, bio)) { 1012 case ELEVATOR_BACK_MERGE: 1013 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1014 return bio_attempt_back_merge(rq, bio, nr_segs); 1015 break; 1016 case ELEVATOR_FRONT_MERGE: 1017 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1018 return bio_attempt_front_merge(rq, bio, nr_segs); 1019 break; 1020 case ELEVATOR_DISCARD_MERGE: 1021 return bio_attempt_discard_merge(q, rq, bio); 1022 default: 1023 return BIO_MERGE_NONE; 1024 } 1025 1026 return BIO_MERGE_FAILED; 1027 } 1028 1029 /** 1030 * blk_attempt_plug_merge - try to merge with %current's plugged list 1031 * @q: request_queue new bio is being queued at 1032 * @bio: new bio being queued 1033 * @nr_segs: number of segments in @bio 1034 * from the passed in @q already in the plug list 1035 * 1036 * Determine whether @bio being queued on @q can be merged with the previous 1037 * request on %current's plugged list. Returns %true if merge was successful, 1038 * otherwise %false. 1039 * 1040 * Plugging coalesces IOs from the same issuer for the same purpose without 1041 * going through @q->queue_lock. As such it's more of an issuing mechanism 1042 * than scheduling, and the request, while may have elvpriv data, is not 1043 * added on the elevator at this point. In addition, we don't have 1044 * reliable access to the elevator outside queue lock. Only check basic 1045 * merging parameters without querying the elevator. 1046 * 1047 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1048 */ 1049 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1050 unsigned int nr_segs) 1051 { 1052 struct blk_plug *plug; 1053 struct request *rq; 1054 1055 plug = blk_mq_plug(bio); 1056 if (!plug || rq_list_empty(plug->mq_list)) 1057 return false; 1058 1059 rq_list_for_each(&plug->mq_list, rq) { 1060 if (rq->q == q) { 1061 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1062 BIO_MERGE_OK) 1063 return true; 1064 break; 1065 } 1066 1067 /* 1068 * Only keep iterating plug list for merges if we have multiple 1069 * queues 1070 */ 1071 if (!plug->multiple_queues) 1072 break; 1073 } 1074 return false; 1075 } 1076 1077 /* 1078 * Iterate list of requests and see if we can merge this bio with any 1079 * of them. 1080 */ 1081 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1082 struct bio *bio, unsigned int nr_segs) 1083 { 1084 struct request *rq; 1085 int checked = 8; 1086 1087 list_for_each_entry_reverse(rq, list, queuelist) { 1088 if (!checked--) 1089 break; 1090 1091 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1092 case BIO_MERGE_NONE: 1093 continue; 1094 case BIO_MERGE_OK: 1095 return true; 1096 case BIO_MERGE_FAILED: 1097 return false; 1098 } 1099 1100 } 1101 1102 return false; 1103 } 1104 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1105 1106 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1107 unsigned int nr_segs, struct request **merged_request) 1108 { 1109 struct request *rq; 1110 1111 switch (elv_merge(q, &rq, bio)) { 1112 case ELEVATOR_BACK_MERGE: 1113 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1114 return false; 1115 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1116 return false; 1117 *merged_request = attempt_back_merge(q, rq); 1118 if (!*merged_request) 1119 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1120 return true; 1121 case ELEVATOR_FRONT_MERGE: 1122 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1123 return false; 1124 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1125 return false; 1126 *merged_request = attempt_front_merge(q, rq); 1127 if (!*merged_request) 1128 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1129 return true; 1130 case ELEVATOR_DISCARD_MERGE: 1131 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1132 default: 1133 return false; 1134 } 1135 } 1136 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1137