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 12 #include <trace/events/block.h> 13 14 #include "blk.h" 15 #include "blk-rq-qos.h" 16 #include "blk-throttle.h" 17 18 static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv) 19 { 20 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 21 } 22 23 static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv) 24 { 25 struct bvec_iter iter = bio->bi_iter; 26 int idx; 27 28 bio_get_first_bvec(bio, bv); 29 if (bv->bv_len == bio->bi_iter.bi_size) 30 return; /* this bio only has a single bvec */ 31 32 bio_advance_iter(bio, &iter, iter.bi_size); 33 34 if (!iter.bi_bvec_done) 35 idx = iter.bi_idx - 1; 36 else /* in the middle of bvec */ 37 idx = iter.bi_idx; 38 39 *bv = bio->bi_io_vec[idx]; 40 41 /* 42 * iter.bi_bvec_done records actual length of the last bvec 43 * if this bio ends in the middle of one io vector 44 */ 45 if (iter.bi_bvec_done) 46 bv->bv_len = iter.bi_bvec_done; 47 } 48 49 static inline bool bio_will_gap(struct request_queue *q, 50 struct request *prev_rq, struct bio *prev, struct bio *next) 51 { 52 struct bio_vec pb, nb; 53 54 if (!bio_has_data(prev) || !queue_virt_boundary(q)) 55 return false; 56 57 /* 58 * Don't merge if the 1st bio starts with non-zero offset, otherwise it 59 * is quite difficult to respect the sg gap limit. We work hard to 60 * merge a huge number of small single bios in case of mkfs. 61 */ 62 if (prev_rq) 63 bio_get_first_bvec(prev_rq->bio, &pb); 64 else 65 bio_get_first_bvec(prev, &pb); 66 if (pb.bv_offset & queue_virt_boundary(q)) 67 return true; 68 69 /* 70 * We don't need to worry about the situation that the merged segment 71 * ends in unaligned virt boundary: 72 * 73 * - if 'pb' ends aligned, the merged segment ends aligned 74 * - if 'pb' ends unaligned, the next bio must include 75 * one single bvec of 'nb', otherwise the 'nb' can't 76 * merge with 'pb' 77 */ 78 bio_get_last_bvec(prev, &pb); 79 bio_get_first_bvec(next, &nb); 80 if (biovec_phys_mergeable(q, &pb, &nb)) 81 return false; 82 return __bvec_gap_to_prev(q, &pb, nb.bv_offset); 83 } 84 85 static inline bool req_gap_back_merge(struct request *req, struct bio *bio) 86 { 87 return bio_will_gap(req->q, req, req->biotail, bio); 88 } 89 90 static inline bool req_gap_front_merge(struct request *req, struct bio *bio) 91 { 92 return bio_will_gap(req->q, NULL, bio, req->bio); 93 } 94 95 static struct bio *blk_bio_discard_split(struct request_queue *q, 96 struct bio *bio, 97 struct bio_set *bs, 98 unsigned *nsegs) 99 { 100 unsigned int max_discard_sectors, granularity; 101 int alignment; 102 sector_t tmp; 103 unsigned split_sectors; 104 105 *nsegs = 1; 106 107 /* Zero-sector (unknown) and one-sector granularities are the same. */ 108 granularity = max(q->limits.discard_granularity >> 9, 1U); 109 110 max_discard_sectors = min(q->limits.max_discard_sectors, 111 bio_allowed_max_sectors(q)); 112 max_discard_sectors -= max_discard_sectors % granularity; 113 114 if (unlikely(!max_discard_sectors)) { 115 /* XXX: warn */ 116 return NULL; 117 } 118 119 if (bio_sectors(bio) <= max_discard_sectors) 120 return NULL; 121 122 split_sectors = max_discard_sectors; 123 124 /* 125 * If the next starting sector would be misaligned, stop the discard at 126 * the previous aligned sector. 127 */ 128 alignment = (q->limits.discard_alignment >> 9) % granularity; 129 130 tmp = bio->bi_iter.bi_sector + split_sectors - alignment; 131 tmp = sector_div(tmp, granularity); 132 133 if (split_sectors > tmp) 134 split_sectors -= tmp; 135 136 return bio_split(bio, split_sectors, GFP_NOIO, bs); 137 } 138 139 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q, 140 struct bio *bio, struct bio_set *bs, unsigned *nsegs) 141 { 142 *nsegs = 0; 143 144 if (!q->limits.max_write_zeroes_sectors) 145 return NULL; 146 147 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) 148 return NULL; 149 150 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); 151 } 152 153 static struct bio *blk_bio_write_same_split(struct request_queue *q, 154 struct bio *bio, 155 struct bio_set *bs, 156 unsigned *nsegs) 157 { 158 *nsegs = 1; 159 160 if (!q->limits.max_write_same_sectors) 161 return NULL; 162 163 if (bio_sectors(bio) <= q->limits.max_write_same_sectors) 164 return NULL; 165 166 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); 167 } 168 169 /* 170 * Return the maximum number of sectors from the start of a bio that may be 171 * submitted as a single request to a block device. If enough sectors remain, 172 * align the end to the physical block size. Otherwise align the end to the 173 * logical block size. This approach minimizes the number of non-aligned 174 * requests that are submitted to a block device if the start of a bio is not 175 * aligned to a physical block boundary. 176 */ 177 static inline unsigned get_max_io_size(struct request_queue *q, 178 struct bio *bio) 179 { 180 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0); 181 unsigned max_sectors = sectors; 182 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT; 183 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT; 184 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1); 185 186 max_sectors += start_offset; 187 max_sectors &= ~(pbs - 1); 188 if (max_sectors > start_offset) 189 return max_sectors - start_offset; 190 191 return sectors & ~(lbs - 1); 192 } 193 194 static inline unsigned get_max_segment_size(const struct request_queue *q, 195 struct page *start_page, 196 unsigned long offset) 197 { 198 unsigned long mask = queue_segment_boundary(q); 199 200 offset = mask & (page_to_phys(start_page) + offset); 201 202 /* 203 * overflow may be triggered in case of zero page physical address 204 * on 32bit arch, use queue's max segment size when that happens. 205 */ 206 return min_not_zero(mask - offset + 1, 207 (unsigned long)queue_max_segment_size(q)); 208 } 209 210 /** 211 * bvec_split_segs - verify whether or not a bvec should be split in the middle 212 * @q: [in] request queue associated with the bio associated with @bv 213 * @bv: [in] bvec to examine 214 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 215 * by the number of segments from @bv that may be appended to that 216 * bio without exceeding @max_segs 217 * @sectors: [in,out] Number of sectors in the bio being built. Incremented 218 * by the number of sectors from @bv that may be appended to that 219 * bio without exceeding @max_sectors 220 * @max_segs: [in] upper bound for *@nsegs 221 * @max_sectors: [in] upper bound for *@sectors 222 * 223 * When splitting a bio, it can happen that a bvec is encountered that is too 224 * big to fit in a single segment and hence that it has to be split in the 225 * middle. This function verifies whether or not that should happen. The value 226 * %true is returned if and only if appending the entire @bv to a bio with 227 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 228 * the block driver. 229 */ 230 static bool bvec_split_segs(const struct request_queue *q, 231 const struct bio_vec *bv, unsigned *nsegs, 232 unsigned *sectors, unsigned max_segs, 233 unsigned max_sectors) 234 { 235 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9; 236 unsigned len = min(bv->bv_len, max_len); 237 unsigned total_len = 0; 238 unsigned seg_size = 0; 239 240 while (len && *nsegs < max_segs) { 241 seg_size = get_max_segment_size(q, bv->bv_page, 242 bv->bv_offset + total_len); 243 seg_size = min(seg_size, len); 244 245 (*nsegs)++; 246 total_len += seg_size; 247 len -= seg_size; 248 249 if ((bv->bv_offset + total_len) & queue_virt_boundary(q)) 250 break; 251 } 252 253 *sectors += total_len >> 9; 254 255 /* tell the caller to split the bvec if it is too big to fit */ 256 return len > 0 || bv->bv_len > max_len; 257 } 258 259 /** 260 * blk_bio_segment_split - split a bio in two bios 261 * @q: [in] request queue pointer 262 * @bio: [in] bio to be split 263 * @bs: [in] bio set to allocate the clone from 264 * @segs: [out] number of segments in the bio with the first half of the sectors 265 * 266 * Clone @bio, update the bi_iter of the clone to represent the first sectors 267 * of @bio and update @bio->bi_iter to represent the remaining sectors. The 268 * following is guaranteed for the cloned bio: 269 * - That it has at most get_max_io_size(@q, @bio) sectors. 270 * - That it has at most queue_max_segments(@q) segments. 271 * 272 * Except for discard requests the cloned bio will point at the bi_io_vec of 273 * the original bio. It is the responsibility of the caller to ensure that the 274 * original bio is not freed before the cloned bio. The caller is also 275 * responsible for ensuring that @bs is only destroyed after processing of the 276 * split bio has finished. 277 */ 278 static struct bio *blk_bio_segment_split(struct request_queue *q, 279 struct bio *bio, 280 struct bio_set *bs, 281 unsigned *segs) 282 { 283 struct bio_vec bv, bvprv, *bvprvp = NULL; 284 struct bvec_iter iter; 285 unsigned nsegs = 0, sectors = 0; 286 const unsigned max_sectors = get_max_io_size(q, bio); 287 const unsigned max_segs = queue_max_segments(q); 288 289 bio_for_each_bvec(bv, bio, iter) { 290 /* 291 * If the queue doesn't support SG gaps and adding this 292 * offset would create a gap, disallow it. 293 */ 294 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) 295 goto split; 296 297 if (nsegs < max_segs && 298 sectors + (bv.bv_len >> 9) <= max_sectors && 299 bv.bv_offset + bv.bv_len <= PAGE_SIZE) { 300 nsegs++; 301 sectors += bv.bv_len >> 9; 302 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs, 303 max_sectors)) { 304 goto split; 305 } 306 307 bvprv = bv; 308 bvprvp = &bvprv; 309 } 310 311 *segs = nsegs; 312 return NULL; 313 split: 314 *segs = nsegs; 315 316 /* 317 * Bio splitting may cause subtle trouble such as hang when doing sync 318 * iopoll in direct IO routine. Given performance gain of iopoll for 319 * big IO can be trival, disable iopoll when split needed. 320 */ 321 bio_clear_polled(bio); 322 return bio_split(bio, sectors, GFP_NOIO, bs); 323 } 324 325 /** 326 * __blk_queue_split - split a bio and submit the second half 327 * @q: [in] request_queue new bio is being queued at 328 * @bio: [in, out] bio to be split 329 * @nr_segs: [out] number of segments in the first bio 330 * 331 * Split a bio into two bios, chain the two bios, submit the second half and 332 * store a pointer to the first half in *@bio. If the second bio is still too 333 * big it will be split by a recursive call to this function. Since this 334 * function may allocate a new bio from q->bio_split, it is the responsibility 335 * of the caller to ensure that q->bio_split is only released after processing 336 * of the split bio has finished. 337 */ 338 void __blk_queue_split(struct request_queue *q, struct bio **bio, 339 unsigned int *nr_segs) 340 { 341 struct bio *split = NULL; 342 343 switch (bio_op(*bio)) { 344 case REQ_OP_DISCARD: 345 case REQ_OP_SECURE_ERASE: 346 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs); 347 break; 348 case REQ_OP_WRITE_ZEROES: 349 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, 350 nr_segs); 351 break; 352 case REQ_OP_WRITE_SAME: 353 split = blk_bio_write_same_split(q, *bio, &q->bio_split, 354 nr_segs); 355 break; 356 default: 357 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs); 358 break; 359 } 360 361 if (split) { 362 /* there isn't chance to merge the splitted bio */ 363 split->bi_opf |= REQ_NOMERGE; 364 365 bio_chain(split, *bio); 366 trace_block_split(split, (*bio)->bi_iter.bi_sector); 367 submit_bio_noacct(*bio); 368 *bio = split; 369 370 blk_throtl_charge_bio_split(*bio); 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 || req->rq_disk != next->rq_disk) 780 return NULL; 781 782 if (req_op(req) == REQ_OP_WRITE_SAME && 783 !blk_write_same_mergeable(req->bio, next->bio)) 784 return NULL; 785 786 /* 787 * Don't allow merge of different write hints, or for a hint with 788 * non-hint IO. 789 */ 790 if (req->write_hint != next->write_hint) 791 return NULL; 792 793 if (req->ioprio != next->ioprio) 794 return NULL; 795 796 /* 797 * If we are allowed to merge, then append bio list 798 * from next to rq and release next. merge_requests_fn 799 * will have updated segment counts, update sector 800 * counts here. Handle DISCARDs separately, as they 801 * have separate settings. 802 */ 803 804 switch (blk_try_req_merge(req, next)) { 805 case ELEVATOR_DISCARD_MERGE: 806 if (!req_attempt_discard_merge(q, req, next)) 807 return NULL; 808 break; 809 case ELEVATOR_BACK_MERGE: 810 if (!ll_merge_requests_fn(q, req, next)) 811 return NULL; 812 break; 813 default: 814 return NULL; 815 } 816 817 /* 818 * If failfast settings disagree or any of the two is already 819 * a mixed merge, mark both as mixed before proceeding. This 820 * makes sure that all involved bios have mixable attributes 821 * set properly. 822 */ 823 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 824 (req->cmd_flags & REQ_FAILFAST_MASK) != 825 (next->cmd_flags & REQ_FAILFAST_MASK)) { 826 blk_rq_set_mixed_merge(req); 827 blk_rq_set_mixed_merge(next); 828 } 829 830 /* 831 * At this point we have either done a back merge or front merge. We 832 * need the smaller start_time_ns of the merged requests to be the 833 * current request for accounting purposes. 834 */ 835 if (next->start_time_ns < req->start_time_ns) 836 req->start_time_ns = next->start_time_ns; 837 838 req->biotail->bi_next = next->bio; 839 req->biotail = next->biotail; 840 841 req->__data_len += blk_rq_bytes(next); 842 843 if (!blk_discard_mergable(req)) 844 elv_merge_requests(q, req, next); 845 846 /* 847 * 'next' is going away, so update stats accordingly 848 */ 849 blk_account_io_merge_request(next); 850 851 trace_block_rq_merge(next); 852 853 /* 854 * ownership of bio passed from next to req, return 'next' for 855 * the caller to free 856 */ 857 next->bio = NULL; 858 return next; 859 } 860 861 static struct request *attempt_back_merge(struct request_queue *q, 862 struct request *rq) 863 { 864 struct request *next = elv_latter_request(q, rq); 865 866 if (next) 867 return attempt_merge(q, rq, next); 868 869 return NULL; 870 } 871 872 static struct request *attempt_front_merge(struct request_queue *q, 873 struct request *rq) 874 { 875 struct request *prev = elv_former_request(q, rq); 876 877 if (prev) 878 return attempt_merge(q, prev, rq); 879 880 return NULL; 881 } 882 883 /* 884 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 885 * otherwise. The caller is responsible for freeing 'next' if the merge 886 * happened. 887 */ 888 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 889 struct request *next) 890 { 891 return attempt_merge(q, rq, next); 892 } 893 894 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 895 { 896 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 897 return false; 898 899 if (req_op(rq) != bio_op(bio)) 900 return false; 901 902 /* different data direction or already started, don't merge */ 903 if (bio_data_dir(bio) != rq_data_dir(rq)) 904 return false; 905 906 /* must be same device */ 907 if (rq->rq_disk != bio->bi_bdev->bd_disk) 908 return false; 909 910 /* only merge integrity protected bio into ditto rq */ 911 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 912 return false; 913 914 /* Only merge if the crypt contexts are compatible */ 915 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 916 return false; 917 918 /* must be using the same buffer */ 919 if (req_op(rq) == REQ_OP_WRITE_SAME && 920 !blk_write_same_mergeable(rq->bio, bio)) 921 return false; 922 923 /* 924 * Don't allow merge of different write hints, or for a hint with 925 * non-hint IO. 926 */ 927 if (rq->write_hint != bio->bi_write_hint) 928 return false; 929 930 if (rq->ioprio != bio_prio(bio)) 931 return false; 932 933 return true; 934 } 935 936 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 937 { 938 if (blk_discard_mergable(rq)) 939 return ELEVATOR_DISCARD_MERGE; 940 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 941 return ELEVATOR_BACK_MERGE; 942 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 943 return ELEVATOR_FRONT_MERGE; 944 return ELEVATOR_NO_MERGE; 945 } 946 947 static void blk_account_io_merge_bio(struct request *req) 948 { 949 if (!blk_do_io_stat(req)) 950 return; 951 952 part_stat_lock(); 953 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 954 part_stat_unlock(); 955 } 956 957 enum bio_merge_status { 958 BIO_MERGE_OK, 959 BIO_MERGE_NONE, 960 BIO_MERGE_FAILED, 961 }; 962 963 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 964 struct bio *bio, unsigned int nr_segs) 965 { 966 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 967 968 if (!ll_back_merge_fn(req, bio, nr_segs)) 969 return BIO_MERGE_FAILED; 970 971 trace_block_bio_backmerge(bio); 972 rq_qos_merge(req->q, req, bio); 973 974 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 975 blk_rq_set_mixed_merge(req); 976 977 req->biotail->bi_next = bio; 978 req->biotail = bio; 979 req->__data_len += bio->bi_iter.bi_size; 980 981 bio_crypt_free_ctx(bio); 982 983 blk_account_io_merge_bio(req); 984 return BIO_MERGE_OK; 985 } 986 987 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 988 struct bio *bio, unsigned int nr_segs) 989 { 990 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 991 992 if (!ll_front_merge_fn(req, bio, nr_segs)) 993 return BIO_MERGE_FAILED; 994 995 trace_block_bio_frontmerge(bio); 996 rq_qos_merge(req->q, req, bio); 997 998 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 999 blk_rq_set_mixed_merge(req); 1000 1001 bio->bi_next = req->bio; 1002 req->bio = bio; 1003 1004 req->__sector = bio->bi_iter.bi_sector; 1005 req->__data_len += bio->bi_iter.bi_size; 1006 1007 bio_crypt_do_front_merge(req, bio); 1008 1009 blk_account_io_merge_bio(req); 1010 return BIO_MERGE_OK; 1011 } 1012 1013 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 1014 struct request *req, struct bio *bio) 1015 { 1016 unsigned short segments = blk_rq_nr_discard_segments(req); 1017 1018 if (segments >= queue_max_discard_segments(q)) 1019 goto no_merge; 1020 if (blk_rq_sectors(req) + bio_sectors(bio) > 1021 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 1022 goto no_merge; 1023 1024 rq_qos_merge(q, req, bio); 1025 1026 req->biotail->bi_next = bio; 1027 req->biotail = bio; 1028 req->__data_len += bio->bi_iter.bi_size; 1029 req->nr_phys_segments = segments + 1; 1030 1031 blk_account_io_merge_bio(req); 1032 return BIO_MERGE_OK; 1033 no_merge: 1034 req_set_nomerge(q, req); 1035 return BIO_MERGE_FAILED; 1036 } 1037 1038 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1039 struct request *rq, 1040 struct bio *bio, 1041 unsigned int nr_segs, 1042 bool sched_allow_merge) 1043 { 1044 if (!blk_rq_merge_ok(rq, bio)) 1045 return BIO_MERGE_NONE; 1046 1047 switch (blk_try_merge(rq, bio)) { 1048 case ELEVATOR_BACK_MERGE: 1049 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1050 return bio_attempt_back_merge(rq, bio, nr_segs); 1051 break; 1052 case ELEVATOR_FRONT_MERGE: 1053 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1054 return bio_attempt_front_merge(rq, bio, nr_segs); 1055 break; 1056 case ELEVATOR_DISCARD_MERGE: 1057 return bio_attempt_discard_merge(q, rq, bio); 1058 default: 1059 return BIO_MERGE_NONE; 1060 } 1061 1062 return BIO_MERGE_FAILED; 1063 } 1064 1065 /** 1066 * blk_attempt_plug_merge - try to merge with %current's plugged list 1067 * @q: request_queue new bio is being queued at 1068 * @bio: new bio being queued 1069 * @nr_segs: number of segments in @bio 1070 * @same_queue_rq: output value, will be true if there's an existing request 1071 * from the passed in @q already in the plug list 1072 * 1073 * Determine whether @bio being queued on @q can be merged with the previous 1074 * request on %current's plugged list. Returns %true if merge was successful, 1075 * otherwise %false. 1076 * 1077 * Plugging coalesces IOs from the same issuer for the same purpose without 1078 * going through @q->queue_lock. As such it's more of an issuing mechanism 1079 * than scheduling, and the request, while may have elvpriv data, is not 1080 * added on the elevator at this point. In addition, we don't have 1081 * reliable access to the elevator outside queue lock. Only check basic 1082 * merging parameters without querying the elevator. 1083 * 1084 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1085 */ 1086 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1087 unsigned int nr_segs, bool *same_queue_rq) 1088 { 1089 struct blk_plug *plug; 1090 struct request *rq; 1091 1092 plug = blk_mq_plug(q, bio); 1093 if (!plug || rq_list_empty(plug->mq_list)) 1094 return false; 1095 1096 /* check the previously added entry for a quick merge attempt */ 1097 rq = rq_list_peek(&plug->mq_list); 1098 if (rq->q == q) { 1099 /* 1100 * Only blk-mq multiple hardware queues case checks the rq in 1101 * the same queue, there should be only one such rq in a queue 1102 */ 1103 *same_queue_rq = true; 1104 } 1105 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == BIO_MERGE_OK) 1106 return true; 1107 return false; 1108 } 1109 1110 /* 1111 * Iterate list of requests and see if we can merge this bio with any 1112 * of them. 1113 */ 1114 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1115 struct bio *bio, unsigned int nr_segs) 1116 { 1117 struct request *rq; 1118 int checked = 8; 1119 1120 list_for_each_entry_reverse(rq, list, queuelist) { 1121 if (!checked--) 1122 break; 1123 1124 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1125 case BIO_MERGE_NONE: 1126 continue; 1127 case BIO_MERGE_OK: 1128 return true; 1129 case BIO_MERGE_FAILED: 1130 return false; 1131 } 1132 1133 } 1134 1135 return false; 1136 } 1137 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1138 1139 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1140 unsigned int nr_segs, struct request **merged_request) 1141 { 1142 struct request *rq; 1143 1144 switch (elv_merge(q, &rq, bio)) { 1145 case ELEVATOR_BACK_MERGE: 1146 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1147 return false; 1148 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1149 return false; 1150 *merged_request = attempt_back_merge(q, rq); 1151 if (!*merged_request) 1152 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1153 return true; 1154 case ELEVATOR_FRONT_MERGE: 1155 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1156 return false; 1157 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1158 return false; 1159 *merged_request = attempt_front_merge(q, rq); 1160 if (!*merged_request) 1161 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1162 return true; 1163 case ELEVATOR_DISCARD_MERGE: 1164 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1165 default: 1166 return false; 1167 } 1168 } 1169 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1170