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 13 #include <trace/events/block.h> 14 15 #include "blk.h" 16 #include "blk-mq-sched.h" 17 #include "blk-rq-qos.h" 18 #include "blk-throttle.h" 19 20 static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv) 21 { 22 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 23 } 24 25 static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv) 26 { 27 struct bvec_iter iter = bio->bi_iter; 28 int idx; 29 30 bio_get_first_bvec(bio, bv); 31 if (bv->bv_len == bio->bi_iter.bi_size) 32 return; /* this bio only has a single bvec */ 33 34 bio_advance_iter(bio, &iter, iter.bi_size); 35 36 if (!iter.bi_bvec_done) 37 idx = iter.bi_idx - 1; 38 else /* in the middle of bvec */ 39 idx = iter.bi_idx; 40 41 *bv = bio->bi_io_vec[idx]; 42 43 /* 44 * iter.bi_bvec_done records actual length of the last bvec 45 * if this bio ends in the middle of one io vector 46 */ 47 if (iter.bi_bvec_done) 48 bv->bv_len = iter.bi_bvec_done; 49 } 50 51 static inline bool bio_will_gap(struct request_queue *q, 52 struct request *prev_rq, struct bio *prev, struct bio *next) 53 { 54 struct bio_vec pb, nb; 55 56 if (!bio_has_data(prev) || !queue_virt_boundary(q)) 57 return false; 58 59 /* 60 * Don't merge if the 1st bio starts with non-zero offset, otherwise it 61 * is quite difficult to respect the sg gap limit. We work hard to 62 * merge a huge number of small single bios in case of mkfs. 63 */ 64 if (prev_rq) 65 bio_get_first_bvec(prev_rq->bio, &pb); 66 else 67 bio_get_first_bvec(prev, &pb); 68 if (pb.bv_offset & queue_virt_boundary(q)) 69 return true; 70 71 /* 72 * We don't need to worry about the situation that the merged segment 73 * ends in unaligned virt boundary: 74 * 75 * - if 'pb' ends aligned, the merged segment ends aligned 76 * - if 'pb' ends unaligned, the next bio must include 77 * one single bvec of 'nb', otherwise the 'nb' can't 78 * merge with 'pb' 79 */ 80 bio_get_last_bvec(prev, &pb); 81 bio_get_first_bvec(next, &nb); 82 if (biovec_phys_mergeable(q, &pb, &nb)) 83 return false; 84 return __bvec_gap_to_prev(q, &pb, nb.bv_offset); 85 } 86 87 static inline bool req_gap_back_merge(struct request *req, struct bio *bio) 88 { 89 return bio_will_gap(req->q, req, req->biotail, bio); 90 } 91 92 static inline bool req_gap_front_merge(struct request *req, struct bio *bio) 93 { 94 return bio_will_gap(req->q, NULL, bio, req->bio); 95 } 96 97 static struct bio *blk_bio_discard_split(struct request_queue *q, 98 struct bio *bio, 99 struct bio_set *bs, 100 unsigned *nsegs) 101 { 102 unsigned int max_discard_sectors, granularity; 103 int alignment; 104 sector_t tmp; 105 unsigned split_sectors; 106 107 *nsegs = 1; 108 109 /* Zero-sector (unknown) and one-sector granularities are the same. */ 110 granularity = max(q->limits.discard_granularity >> 9, 1U); 111 112 max_discard_sectors = min(q->limits.max_discard_sectors, 113 bio_allowed_max_sectors(q)); 114 max_discard_sectors -= max_discard_sectors % granularity; 115 116 if (unlikely(!max_discard_sectors)) { 117 /* XXX: warn */ 118 return NULL; 119 } 120 121 if (bio_sectors(bio) <= max_discard_sectors) 122 return NULL; 123 124 split_sectors = max_discard_sectors; 125 126 /* 127 * If the next starting sector would be misaligned, stop the discard at 128 * the previous aligned sector. 129 */ 130 alignment = (q->limits.discard_alignment >> 9) % granularity; 131 132 tmp = bio->bi_iter.bi_sector + split_sectors - alignment; 133 tmp = sector_div(tmp, granularity); 134 135 if (split_sectors > tmp) 136 split_sectors -= tmp; 137 138 return bio_split(bio, split_sectors, GFP_NOIO, bs); 139 } 140 141 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q, 142 struct bio *bio, struct bio_set *bs, unsigned *nsegs) 143 { 144 *nsegs = 0; 145 146 if (!q->limits.max_write_zeroes_sectors) 147 return NULL; 148 149 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) 150 return NULL; 151 152 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); 153 } 154 155 static struct bio *blk_bio_write_same_split(struct request_queue *q, 156 struct bio *bio, 157 struct bio_set *bs, 158 unsigned *nsegs) 159 { 160 *nsegs = 1; 161 162 if (!q->limits.max_write_same_sectors) 163 return NULL; 164 165 if (bio_sectors(bio) <= q->limits.max_write_same_sectors) 166 return NULL; 167 168 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); 169 } 170 171 /* 172 * Return the maximum number of sectors from the start of a bio that may be 173 * submitted as a single request to a block device. If enough sectors remain, 174 * align the end to the physical block size. Otherwise align the end to the 175 * logical block size. This approach minimizes the number of non-aligned 176 * requests that are submitted to a block device if the start of a bio is not 177 * aligned to a physical block boundary. 178 */ 179 static inline unsigned get_max_io_size(struct request_queue *q, 180 struct bio *bio) 181 { 182 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0); 183 unsigned max_sectors = sectors; 184 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT; 185 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT; 186 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1); 187 188 max_sectors += start_offset; 189 max_sectors &= ~(pbs - 1); 190 if (max_sectors > start_offset) 191 return max_sectors - start_offset; 192 193 return sectors & ~(lbs - 1); 194 } 195 196 static inline unsigned get_max_segment_size(const struct request_queue *q, 197 struct page *start_page, 198 unsigned long offset) 199 { 200 unsigned long mask = queue_segment_boundary(q); 201 202 offset = mask & (page_to_phys(start_page) + offset); 203 204 /* 205 * overflow may be triggered in case of zero page physical address 206 * on 32bit arch, use queue's max segment size when that happens. 207 */ 208 return min_not_zero(mask - offset + 1, 209 (unsigned long)queue_max_segment_size(q)); 210 } 211 212 /** 213 * bvec_split_segs - verify whether or not a bvec should be split in the middle 214 * @q: [in] request queue associated with the bio associated with @bv 215 * @bv: [in] bvec to examine 216 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 217 * by the number of segments from @bv that may be appended to that 218 * bio without exceeding @max_segs 219 * @sectors: [in,out] Number of sectors in the bio being built. Incremented 220 * by the number of sectors from @bv that may be appended to that 221 * bio without exceeding @max_sectors 222 * @max_segs: [in] upper bound for *@nsegs 223 * @max_sectors: [in] upper bound for *@sectors 224 * 225 * When splitting a bio, it can happen that a bvec is encountered that is too 226 * big to fit in a single segment and hence that it has to be split in the 227 * middle. This function verifies whether or not that should happen. The value 228 * %true is returned if and only if appending the entire @bv to a bio with 229 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 230 * the block driver. 231 */ 232 static bool bvec_split_segs(const struct request_queue *q, 233 const struct bio_vec *bv, unsigned *nsegs, 234 unsigned *sectors, unsigned max_segs, 235 unsigned max_sectors) 236 { 237 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9; 238 unsigned len = min(bv->bv_len, max_len); 239 unsigned total_len = 0; 240 unsigned seg_size = 0; 241 242 while (len && *nsegs < max_segs) { 243 seg_size = get_max_segment_size(q, bv->bv_page, 244 bv->bv_offset + total_len); 245 seg_size = min(seg_size, len); 246 247 (*nsegs)++; 248 total_len += seg_size; 249 len -= seg_size; 250 251 if ((bv->bv_offset + total_len) & queue_virt_boundary(q)) 252 break; 253 } 254 255 *sectors += total_len >> 9; 256 257 /* tell the caller to split the bvec if it is too big to fit */ 258 return len > 0 || bv->bv_len > max_len; 259 } 260 261 /** 262 * blk_bio_segment_split - split a bio in two bios 263 * @q: [in] request queue pointer 264 * @bio: [in] bio to be split 265 * @bs: [in] bio set to allocate the clone from 266 * @segs: [out] number of segments in the bio with the first half of the sectors 267 * 268 * Clone @bio, update the bi_iter of the clone to represent the first sectors 269 * of @bio and update @bio->bi_iter to represent the remaining sectors. The 270 * following is guaranteed for the cloned bio: 271 * - That it has at most get_max_io_size(@q, @bio) sectors. 272 * - That it has at most queue_max_segments(@q) segments. 273 * 274 * Except for discard requests the cloned bio will point at the bi_io_vec of 275 * the original bio. It is the responsibility of the caller to ensure that the 276 * original bio is not freed before the cloned bio. The caller is also 277 * responsible for ensuring that @bs is only destroyed after processing of the 278 * split bio has finished. 279 */ 280 static struct bio *blk_bio_segment_split(struct request_queue *q, 281 struct bio *bio, 282 struct bio_set *bs, 283 unsigned *segs) 284 { 285 struct bio_vec bv, bvprv, *bvprvp = NULL; 286 struct bvec_iter iter; 287 unsigned nsegs = 0, sectors = 0; 288 const unsigned max_sectors = get_max_io_size(q, bio); 289 const unsigned max_segs = queue_max_segments(q); 290 291 bio_for_each_bvec(bv, bio, iter) { 292 /* 293 * If the queue doesn't support SG gaps and adding this 294 * offset would create a gap, disallow it. 295 */ 296 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) 297 goto split; 298 299 if (nsegs < max_segs && 300 sectors + (bv.bv_len >> 9) <= max_sectors && 301 bv.bv_offset + bv.bv_len <= PAGE_SIZE) { 302 nsegs++; 303 sectors += bv.bv_len >> 9; 304 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs, 305 max_sectors)) { 306 goto split; 307 } 308 309 bvprv = bv; 310 bvprvp = &bvprv; 311 } 312 313 *segs = nsegs; 314 return NULL; 315 split: 316 *segs = nsegs; 317 318 /* 319 * Bio splitting may cause subtle trouble such as hang when doing sync 320 * iopoll in direct IO routine. Given performance gain of iopoll for 321 * big IO can be trival, disable iopoll when split needed. 322 */ 323 bio_clear_polled(bio); 324 return bio_split(bio, sectors, GFP_NOIO, bs); 325 } 326 327 /** 328 * __blk_queue_split - split a bio and submit the second half 329 * @q: [in] request_queue new bio is being queued at 330 * @bio: [in, out] bio to be split 331 * @nr_segs: [out] number of segments in the first bio 332 * 333 * Split a bio into two bios, chain the two bios, submit the second half and 334 * store a pointer to the first half in *@bio. If the second bio is still too 335 * big it will be split by a recursive call to this function. Since this 336 * function may allocate a new bio from q->bio_split, it is the responsibility 337 * of the caller to ensure that q->bio_split is only released after processing 338 * of the split bio has finished. 339 */ 340 void __blk_queue_split(struct request_queue *q, struct bio **bio, 341 unsigned int *nr_segs) 342 { 343 struct bio *split = NULL; 344 345 switch (bio_op(*bio)) { 346 case REQ_OP_DISCARD: 347 case REQ_OP_SECURE_ERASE: 348 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs); 349 break; 350 case REQ_OP_WRITE_ZEROES: 351 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, 352 nr_segs); 353 break; 354 case REQ_OP_WRITE_SAME: 355 split = blk_bio_write_same_split(q, *bio, &q->bio_split, 356 nr_segs); 357 break; 358 default: 359 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs); 360 break; 361 } 362 363 if (split) { 364 /* there isn't chance to merge the splitted bio */ 365 split->bi_opf |= REQ_NOMERGE; 366 367 bio_chain(split, *bio); 368 trace_block_split(split, (*bio)->bi_iter.bi_sector); 369 submit_bio_noacct(*bio); 370 *bio = split; 371 } 372 } 373 374 /** 375 * blk_queue_split - split a bio and submit the second half 376 * @bio: [in, out] bio to be split 377 * 378 * Split a bio into two bios, chains the two bios, submit the second half and 379 * store a pointer to the first half in *@bio. Since this function may allocate 380 * a new bio from q->bio_split, it is the responsibility of the caller to ensure 381 * that q->bio_split is only released after processing of the split bio has 382 * finished. 383 */ 384 void blk_queue_split(struct bio **bio) 385 { 386 struct request_queue *q = bdev_get_queue((*bio)->bi_bdev); 387 unsigned int nr_segs; 388 389 if (blk_may_split(q, *bio)) 390 __blk_queue_split(q, bio, &nr_segs); 391 } 392 EXPORT_SYMBOL(blk_queue_split); 393 394 unsigned int blk_recalc_rq_segments(struct request *rq) 395 { 396 unsigned int nr_phys_segs = 0; 397 unsigned int nr_sectors = 0; 398 struct req_iterator iter; 399 struct bio_vec bv; 400 401 if (!rq->bio) 402 return 0; 403 404 switch (bio_op(rq->bio)) { 405 case REQ_OP_DISCARD: 406 case REQ_OP_SECURE_ERASE: 407 if (queue_max_discard_segments(rq->q) > 1) { 408 struct bio *bio = rq->bio; 409 410 for_each_bio(bio) 411 nr_phys_segs++; 412 return nr_phys_segs; 413 } 414 return 1; 415 case REQ_OP_WRITE_ZEROES: 416 return 0; 417 case REQ_OP_WRITE_SAME: 418 return 1; 419 } 420 421 rq_for_each_bvec(bv, rq, iter) 422 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors, 423 UINT_MAX, UINT_MAX); 424 return nr_phys_segs; 425 } 426 427 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 428 struct scatterlist *sglist) 429 { 430 if (!*sg) 431 return sglist; 432 433 /* 434 * If the driver previously mapped a shorter list, we could see a 435 * termination bit prematurely unless it fully inits the sg table 436 * on each mapping. We KNOW that there must be more entries here 437 * or the driver would be buggy, so force clear the termination bit 438 * to avoid doing a full sg_init_table() in drivers for each command. 439 */ 440 sg_unmark_end(*sg); 441 return sg_next(*sg); 442 } 443 444 static unsigned blk_bvec_map_sg(struct request_queue *q, 445 struct bio_vec *bvec, struct scatterlist *sglist, 446 struct scatterlist **sg) 447 { 448 unsigned nbytes = bvec->bv_len; 449 unsigned nsegs = 0, total = 0; 450 451 while (nbytes > 0) { 452 unsigned offset = bvec->bv_offset + total; 453 unsigned len = min(get_max_segment_size(q, bvec->bv_page, 454 offset), nbytes); 455 struct page *page = bvec->bv_page; 456 457 /* 458 * Unfortunately a fair number of drivers barf on scatterlists 459 * that have an offset larger than PAGE_SIZE, despite other 460 * subsystems dealing with that invariant just fine. For now 461 * stick to the legacy format where we never present those from 462 * the block layer, but the code below should be removed once 463 * these offenders (mostly MMC/SD drivers) are fixed. 464 */ 465 page += (offset >> PAGE_SHIFT); 466 offset &= ~PAGE_MASK; 467 468 *sg = blk_next_sg(sg, sglist); 469 sg_set_page(*sg, page, len, offset); 470 471 total += len; 472 nbytes -= len; 473 nsegs++; 474 } 475 476 return nsegs; 477 } 478 479 static inline int __blk_bvec_map_sg(struct bio_vec bv, 480 struct scatterlist *sglist, struct scatterlist **sg) 481 { 482 *sg = blk_next_sg(sg, sglist); 483 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 484 return 1; 485 } 486 487 /* only try to merge bvecs into one sg if they are from two bios */ 488 static inline bool 489 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec, 490 struct bio_vec *bvprv, struct scatterlist **sg) 491 { 492 493 int nbytes = bvec->bv_len; 494 495 if (!*sg) 496 return false; 497 498 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 499 return false; 500 501 if (!biovec_phys_mergeable(q, bvprv, bvec)) 502 return false; 503 504 (*sg)->length += nbytes; 505 506 return true; 507 } 508 509 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 510 struct scatterlist *sglist, 511 struct scatterlist **sg) 512 { 513 struct bio_vec bvec, bvprv = { NULL }; 514 struct bvec_iter iter; 515 int nsegs = 0; 516 bool new_bio = false; 517 518 for_each_bio(bio) { 519 bio_for_each_bvec(bvec, bio, iter) { 520 /* 521 * Only try to merge bvecs from two bios given we 522 * have done bio internal merge when adding pages 523 * to bio 524 */ 525 if (new_bio && 526 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg)) 527 goto next_bvec; 528 529 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE) 530 nsegs += __blk_bvec_map_sg(bvec, sglist, sg); 531 else 532 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg); 533 next_bvec: 534 new_bio = false; 535 } 536 if (likely(bio->bi_iter.bi_size)) { 537 bvprv = bvec; 538 new_bio = true; 539 } 540 } 541 542 return nsegs; 543 } 544 545 /* 546 * map a request to scatterlist, return number of sg entries setup. Caller 547 * must make sure sg can hold rq->nr_phys_segments entries 548 */ 549 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 550 struct scatterlist *sglist, struct scatterlist **last_sg) 551 { 552 int nsegs = 0; 553 554 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 555 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg); 556 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME) 557 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, last_sg); 558 else if (rq->bio) 559 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg); 560 561 if (*last_sg) 562 sg_mark_end(*last_sg); 563 564 /* 565 * Something must have been wrong if the figured number of 566 * segment is bigger than number of req's physical segments 567 */ 568 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 569 570 return nsegs; 571 } 572 EXPORT_SYMBOL(__blk_rq_map_sg); 573 574 static inline unsigned int blk_rq_get_max_segments(struct request *rq) 575 { 576 if (req_op(rq) == REQ_OP_DISCARD) 577 return queue_max_discard_segments(rq->q); 578 return queue_max_segments(rq->q); 579 } 580 581 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 582 sector_t offset) 583 { 584 struct request_queue *q = rq->q; 585 586 if (blk_rq_is_passthrough(rq)) 587 return q->limits.max_hw_sectors; 588 589 if (!q->limits.chunk_sectors || 590 req_op(rq) == REQ_OP_DISCARD || 591 req_op(rq) == REQ_OP_SECURE_ERASE) 592 return blk_queue_get_max_sectors(q, req_op(rq)); 593 594 return min(blk_max_size_offset(q, offset, 0), 595 blk_queue_get_max_sectors(q, req_op(rq))); 596 } 597 598 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 599 unsigned int nr_phys_segs) 600 { 601 if (blk_integrity_merge_bio(req->q, req, bio) == false) 602 goto no_merge; 603 604 /* discard request merge won't add new segment */ 605 if (req_op(req) == REQ_OP_DISCARD) 606 return 1; 607 608 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 609 goto no_merge; 610 611 /* 612 * This will form the start of a new hw segment. Bump both 613 * counters. 614 */ 615 req->nr_phys_segments += nr_phys_segs; 616 return 1; 617 618 no_merge: 619 req_set_nomerge(req->q, req); 620 return 0; 621 } 622 623 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 624 { 625 if (req_gap_back_merge(req, bio)) 626 return 0; 627 if (blk_integrity_rq(req) && 628 integrity_req_gap_back_merge(req, bio)) 629 return 0; 630 if (!bio_crypt_ctx_back_mergeable(req, bio)) 631 return 0; 632 if (blk_rq_sectors(req) + bio_sectors(bio) > 633 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 634 req_set_nomerge(req->q, req); 635 return 0; 636 } 637 638 return ll_new_hw_segment(req, bio, nr_segs); 639 } 640 641 static int ll_front_merge_fn(struct request *req, struct bio *bio, 642 unsigned int nr_segs) 643 { 644 if (req_gap_front_merge(req, bio)) 645 return 0; 646 if (blk_integrity_rq(req) && 647 integrity_req_gap_front_merge(req, bio)) 648 return 0; 649 if (!bio_crypt_ctx_front_mergeable(req, bio)) 650 return 0; 651 if (blk_rq_sectors(req) + bio_sectors(bio) > 652 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 653 req_set_nomerge(req->q, req); 654 return 0; 655 } 656 657 return ll_new_hw_segment(req, bio, nr_segs); 658 } 659 660 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 661 struct request *next) 662 { 663 unsigned short segments = blk_rq_nr_discard_segments(req); 664 665 if (segments >= queue_max_discard_segments(q)) 666 goto no_merge; 667 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 668 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 669 goto no_merge; 670 671 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 672 return true; 673 no_merge: 674 req_set_nomerge(q, req); 675 return false; 676 } 677 678 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 679 struct request *next) 680 { 681 int total_phys_segments; 682 683 if (req_gap_back_merge(req, next->bio)) 684 return 0; 685 686 /* 687 * Will it become too large? 688 */ 689 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 690 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 691 return 0; 692 693 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 694 if (total_phys_segments > blk_rq_get_max_segments(req)) 695 return 0; 696 697 if (blk_integrity_merge_rq(q, req, next) == false) 698 return 0; 699 700 if (!bio_crypt_ctx_merge_rq(req, next)) 701 return 0; 702 703 /* Merge is OK... */ 704 req->nr_phys_segments = total_phys_segments; 705 return 1; 706 } 707 708 /** 709 * blk_rq_set_mixed_merge - mark a request as mixed merge 710 * @rq: request to mark as mixed merge 711 * 712 * Description: 713 * @rq is about to be mixed merged. Make sure the attributes 714 * which can be mixed are set in each bio and mark @rq as mixed 715 * merged. 716 */ 717 void blk_rq_set_mixed_merge(struct request *rq) 718 { 719 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; 720 struct bio *bio; 721 722 if (rq->rq_flags & RQF_MIXED_MERGE) 723 return; 724 725 /* 726 * @rq will no longer represent mixable attributes for all the 727 * contained bios. It will just track those of the first one. 728 * Distributes the attributs to each bio. 729 */ 730 for (bio = rq->bio; bio; bio = bio->bi_next) { 731 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 732 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 733 bio->bi_opf |= ff; 734 } 735 rq->rq_flags |= RQF_MIXED_MERGE; 736 } 737 738 static void blk_account_io_merge_request(struct request *req) 739 { 740 if (blk_do_io_stat(req)) { 741 part_stat_lock(); 742 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 743 part_stat_unlock(); 744 } 745 } 746 747 static enum elv_merge blk_try_req_merge(struct request *req, 748 struct request *next) 749 { 750 if (blk_discard_mergable(req)) 751 return ELEVATOR_DISCARD_MERGE; 752 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 753 return ELEVATOR_BACK_MERGE; 754 755 return ELEVATOR_NO_MERGE; 756 } 757 758 static inline bool blk_write_same_mergeable(struct bio *a, struct bio *b) 759 { 760 if (bio_page(a) == bio_page(b) && bio_offset(a) == bio_offset(b)) 761 return true; 762 return false; 763 } 764 765 /* 766 * For non-mq, this has to be called with the request spinlock acquired. 767 * For mq with scheduling, the appropriate queue wide lock should be held. 768 */ 769 static struct request *attempt_merge(struct request_queue *q, 770 struct request *req, struct request *next) 771 { 772 if (!rq_mergeable(req) || !rq_mergeable(next)) 773 return NULL; 774 775 if (req_op(req) != req_op(next)) 776 return NULL; 777 778 if (rq_data_dir(req) != rq_data_dir(next)) 779 return NULL; 780 781 if (req_op(req) == REQ_OP_WRITE_SAME && 782 !blk_write_same_mergeable(req->bio, next->bio)) 783 return NULL; 784 785 /* 786 * Don't allow merge of different write hints, or for a hint with 787 * non-hint IO. 788 */ 789 if (req->write_hint != next->write_hint) 790 return NULL; 791 792 if (req->ioprio != next->ioprio) 793 return NULL; 794 795 /* 796 * If we are allowed to merge, then append bio list 797 * from next to rq and release next. merge_requests_fn 798 * will have updated segment counts, update sector 799 * counts here. Handle DISCARDs separately, as they 800 * have separate settings. 801 */ 802 803 switch (blk_try_req_merge(req, next)) { 804 case ELEVATOR_DISCARD_MERGE: 805 if (!req_attempt_discard_merge(q, req, next)) 806 return NULL; 807 break; 808 case ELEVATOR_BACK_MERGE: 809 if (!ll_merge_requests_fn(q, req, next)) 810 return NULL; 811 break; 812 default: 813 return NULL; 814 } 815 816 /* 817 * If failfast settings disagree or any of the two is already 818 * a mixed merge, mark both as mixed before proceeding. This 819 * makes sure that all involved bios have mixable attributes 820 * set properly. 821 */ 822 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 823 (req->cmd_flags & REQ_FAILFAST_MASK) != 824 (next->cmd_flags & REQ_FAILFAST_MASK)) { 825 blk_rq_set_mixed_merge(req); 826 blk_rq_set_mixed_merge(next); 827 } 828 829 /* 830 * At this point we have either done a back merge or front merge. We 831 * need the smaller start_time_ns of the merged requests to be the 832 * current request for accounting purposes. 833 */ 834 if (next->start_time_ns < req->start_time_ns) 835 req->start_time_ns = next->start_time_ns; 836 837 req->biotail->bi_next = next->bio; 838 req->biotail = next->biotail; 839 840 req->__data_len += blk_rq_bytes(next); 841 842 if (!blk_discard_mergable(req)) 843 elv_merge_requests(q, req, next); 844 845 /* 846 * 'next' is going away, so update stats accordingly 847 */ 848 blk_account_io_merge_request(next); 849 850 trace_block_rq_merge(next); 851 852 /* 853 * ownership of bio passed from next to req, return 'next' for 854 * the caller to free 855 */ 856 next->bio = NULL; 857 return next; 858 } 859 860 static struct request *attempt_back_merge(struct request_queue *q, 861 struct request *rq) 862 { 863 struct request *next = elv_latter_request(q, rq); 864 865 if (next) 866 return attempt_merge(q, rq, next); 867 868 return NULL; 869 } 870 871 static struct request *attempt_front_merge(struct request_queue *q, 872 struct request *rq) 873 { 874 struct request *prev = elv_former_request(q, rq); 875 876 if (prev) 877 return attempt_merge(q, prev, rq); 878 879 return NULL; 880 } 881 882 /* 883 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 884 * otherwise. The caller is responsible for freeing 'next' if the merge 885 * happened. 886 */ 887 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 888 struct request *next) 889 { 890 return attempt_merge(q, rq, next); 891 } 892 893 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 894 { 895 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 896 return false; 897 898 if (req_op(rq) != bio_op(bio)) 899 return false; 900 901 /* different data direction or already started, don't merge */ 902 if (bio_data_dir(bio) != rq_data_dir(rq)) 903 return false; 904 905 /* only merge integrity protected bio into ditto rq */ 906 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 907 return false; 908 909 /* Only merge if the crypt contexts are compatible */ 910 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 911 return false; 912 913 /* must be using the same buffer */ 914 if (req_op(rq) == REQ_OP_WRITE_SAME && 915 !blk_write_same_mergeable(rq->bio, bio)) 916 return false; 917 918 /* 919 * Don't allow merge of different write hints, or for a hint with 920 * non-hint IO. 921 */ 922 if (rq->write_hint != bio->bi_write_hint) 923 return false; 924 925 if (rq->ioprio != bio_prio(bio)) 926 return false; 927 928 return true; 929 } 930 931 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 932 { 933 if (blk_discard_mergable(rq)) 934 return ELEVATOR_DISCARD_MERGE; 935 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 936 return ELEVATOR_BACK_MERGE; 937 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 938 return ELEVATOR_FRONT_MERGE; 939 return ELEVATOR_NO_MERGE; 940 } 941 942 static void blk_account_io_merge_bio(struct request *req) 943 { 944 if (!blk_do_io_stat(req)) 945 return; 946 947 part_stat_lock(); 948 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 949 part_stat_unlock(); 950 } 951 952 enum bio_merge_status { 953 BIO_MERGE_OK, 954 BIO_MERGE_NONE, 955 BIO_MERGE_FAILED, 956 }; 957 958 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 959 struct bio *bio, unsigned int nr_segs) 960 { 961 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 962 963 if (!ll_back_merge_fn(req, bio, nr_segs)) 964 return BIO_MERGE_FAILED; 965 966 trace_block_bio_backmerge(bio); 967 rq_qos_merge(req->q, req, bio); 968 969 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 970 blk_rq_set_mixed_merge(req); 971 972 req->biotail->bi_next = bio; 973 req->biotail = bio; 974 req->__data_len += bio->bi_iter.bi_size; 975 976 bio_crypt_free_ctx(bio); 977 978 blk_account_io_merge_bio(req); 979 return BIO_MERGE_OK; 980 } 981 982 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 983 struct bio *bio, unsigned int nr_segs) 984 { 985 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 986 987 if (!ll_front_merge_fn(req, bio, nr_segs)) 988 return BIO_MERGE_FAILED; 989 990 trace_block_bio_frontmerge(bio); 991 rq_qos_merge(req->q, req, bio); 992 993 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 994 blk_rq_set_mixed_merge(req); 995 996 bio->bi_next = req->bio; 997 req->bio = bio; 998 999 req->__sector = bio->bi_iter.bi_sector; 1000 req->__data_len += bio->bi_iter.bi_size; 1001 1002 bio_crypt_do_front_merge(req, bio); 1003 1004 blk_account_io_merge_bio(req); 1005 return BIO_MERGE_OK; 1006 } 1007 1008 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 1009 struct request *req, struct bio *bio) 1010 { 1011 unsigned short segments = blk_rq_nr_discard_segments(req); 1012 1013 if (segments >= queue_max_discard_segments(q)) 1014 goto no_merge; 1015 if (blk_rq_sectors(req) + bio_sectors(bio) > 1016 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 1017 goto no_merge; 1018 1019 rq_qos_merge(q, req, bio); 1020 1021 req->biotail->bi_next = bio; 1022 req->biotail = bio; 1023 req->__data_len += bio->bi_iter.bi_size; 1024 req->nr_phys_segments = segments + 1; 1025 1026 blk_account_io_merge_bio(req); 1027 return BIO_MERGE_OK; 1028 no_merge: 1029 req_set_nomerge(q, req); 1030 return BIO_MERGE_FAILED; 1031 } 1032 1033 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1034 struct request *rq, 1035 struct bio *bio, 1036 unsigned int nr_segs, 1037 bool sched_allow_merge) 1038 { 1039 if (!blk_rq_merge_ok(rq, bio)) 1040 return BIO_MERGE_NONE; 1041 1042 switch (blk_try_merge(rq, bio)) { 1043 case ELEVATOR_BACK_MERGE: 1044 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1045 return bio_attempt_back_merge(rq, bio, nr_segs); 1046 break; 1047 case ELEVATOR_FRONT_MERGE: 1048 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1049 return bio_attempt_front_merge(rq, bio, nr_segs); 1050 break; 1051 case ELEVATOR_DISCARD_MERGE: 1052 return bio_attempt_discard_merge(q, rq, bio); 1053 default: 1054 return BIO_MERGE_NONE; 1055 } 1056 1057 return BIO_MERGE_FAILED; 1058 } 1059 1060 /** 1061 * blk_attempt_plug_merge - try to merge with %current's plugged list 1062 * @q: request_queue new bio is being queued at 1063 * @bio: new bio being queued 1064 * @nr_segs: number of segments in @bio 1065 * from the passed in @q already in the plug list 1066 * 1067 * Determine whether @bio being queued on @q can be merged with the previous 1068 * request on %current's plugged list. Returns %true if merge was successful, 1069 * otherwise %false. 1070 * 1071 * Plugging coalesces IOs from the same issuer for the same purpose without 1072 * going through @q->queue_lock. As such it's more of an issuing mechanism 1073 * than scheduling, and the request, while may have elvpriv data, is not 1074 * added on the elevator at this point. In addition, we don't have 1075 * reliable access to the elevator outside queue lock. Only check basic 1076 * merging parameters without querying the elevator. 1077 * 1078 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1079 */ 1080 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1081 unsigned int nr_segs) 1082 { 1083 struct blk_plug *plug; 1084 struct request *rq; 1085 1086 plug = blk_mq_plug(q, bio); 1087 if (!plug || rq_list_empty(plug->mq_list)) 1088 return false; 1089 1090 /* check the previously added entry for a quick merge attempt */ 1091 rq = rq_list_peek(&plug->mq_list); 1092 if (rq->q == q) { 1093 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1094 BIO_MERGE_OK) 1095 return true; 1096 } 1097 return false; 1098 } 1099 1100 /* 1101 * Iterate list of requests and see if we can merge this bio with any 1102 * of them. 1103 */ 1104 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1105 struct bio *bio, unsigned int nr_segs) 1106 { 1107 struct request *rq; 1108 int checked = 8; 1109 1110 list_for_each_entry_reverse(rq, list, queuelist) { 1111 if (!checked--) 1112 break; 1113 1114 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1115 case BIO_MERGE_NONE: 1116 continue; 1117 case BIO_MERGE_OK: 1118 return true; 1119 case BIO_MERGE_FAILED: 1120 return false; 1121 } 1122 1123 } 1124 1125 return false; 1126 } 1127 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1128 1129 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1130 unsigned int nr_segs, struct request **merged_request) 1131 { 1132 struct request *rq; 1133 1134 switch (elv_merge(q, &rq, bio)) { 1135 case ELEVATOR_BACK_MERGE: 1136 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1137 return false; 1138 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1139 return false; 1140 *merged_request = attempt_back_merge(q, rq); 1141 if (!*merged_request) 1142 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1143 return true; 1144 case ELEVATOR_FRONT_MERGE: 1145 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1146 return false; 1147 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1148 return false; 1149 *merged_request = attempt_front_merge(q, rq); 1150 if (!*merged_request) 1151 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1152 return true; 1153 case ELEVATOR_DISCARD_MERGE: 1154 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1155 default: 1156 return false; 1157 } 1158 } 1159 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1160