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-mq-sched.h" 16 #include "blk-rq-qos.h" 17 #include "blk-throttle.h" 18 19 static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv) 20 { 21 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 22 } 23 24 static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv) 25 { 26 struct bvec_iter iter = bio->bi_iter; 27 int idx; 28 29 bio_get_first_bvec(bio, bv); 30 if (bv->bv_len == bio->bi_iter.bi_size) 31 return; /* this bio only has a single bvec */ 32 33 bio_advance_iter(bio, &iter, iter.bi_size); 34 35 if (!iter.bi_bvec_done) 36 idx = iter.bi_idx - 1; 37 else /* in the middle of bvec */ 38 idx = iter.bi_idx; 39 40 *bv = bio->bi_io_vec[idx]; 41 42 /* 43 * iter.bi_bvec_done records actual length of the last bvec 44 * if this bio ends in the middle of one io vector 45 */ 46 if (iter.bi_bvec_done) 47 bv->bv_len = iter.bi_bvec_done; 48 } 49 50 static inline bool bio_will_gap(struct request_queue *q, 51 struct request *prev_rq, struct bio *prev, struct bio *next) 52 { 53 struct bio_vec pb, nb; 54 55 if (!bio_has_data(prev) || !queue_virt_boundary(q)) 56 return false; 57 58 /* 59 * Don't merge if the 1st bio starts with non-zero offset, otherwise it 60 * is quite difficult to respect the sg gap limit. We work hard to 61 * merge a huge number of small single bios in case of mkfs. 62 */ 63 if (prev_rq) 64 bio_get_first_bvec(prev_rq->bio, &pb); 65 else 66 bio_get_first_bvec(prev, &pb); 67 if (pb.bv_offset & queue_virt_boundary(q)) 68 return true; 69 70 /* 71 * We don't need to worry about the situation that the merged segment 72 * ends in unaligned virt boundary: 73 * 74 * - if 'pb' ends aligned, the merged segment ends aligned 75 * - if 'pb' ends unaligned, the next bio must include 76 * one single bvec of 'nb', otherwise the 'nb' can't 77 * merge with 'pb' 78 */ 79 bio_get_last_bvec(prev, &pb); 80 bio_get_first_bvec(next, &nb); 81 if (biovec_phys_mergeable(q, &pb, &nb)) 82 return false; 83 return __bvec_gap_to_prev(q, &pb, nb.bv_offset); 84 } 85 86 static inline bool req_gap_back_merge(struct request *req, struct bio *bio) 87 { 88 return bio_will_gap(req->q, req, req->biotail, bio); 89 } 90 91 static inline bool req_gap_front_merge(struct request *req, struct bio *bio) 92 { 93 return bio_will_gap(req->q, NULL, bio, req->bio); 94 } 95 96 static struct bio *blk_bio_discard_split(struct request_queue *q, 97 struct bio *bio, 98 struct bio_set *bs, 99 unsigned *nsegs) 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 *blk_bio_write_zeroes_split(struct request_queue *q, 141 struct bio *bio, struct bio_set *bs, unsigned *nsegs) 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 static struct bio *blk_bio_write_same_split(struct request_queue *q, 155 struct bio *bio, 156 struct bio_set *bs, 157 unsigned *nsegs) 158 { 159 *nsegs = 1; 160 161 if (!q->limits.max_write_same_sectors) 162 return NULL; 163 164 if (bio_sectors(bio) <= q->limits.max_write_same_sectors) 165 return NULL; 166 167 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); 168 } 169 170 /* 171 * Return the maximum number of sectors from the start of a bio that may be 172 * submitted as a single request to a block device. If enough sectors remain, 173 * align the end to the physical block size. Otherwise align the end to the 174 * logical block size. This approach minimizes the number of non-aligned 175 * requests that are submitted to a block device if the start of a bio is not 176 * aligned to a physical block boundary. 177 */ 178 static inline unsigned get_max_io_size(struct request_queue *q, 179 struct bio *bio) 180 { 181 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0); 182 unsigned max_sectors = sectors; 183 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT; 184 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT; 185 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1); 186 187 max_sectors += start_offset; 188 max_sectors &= ~(pbs - 1); 189 if (max_sectors > start_offset) 190 return max_sectors - start_offset; 191 192 return sectors & ~(lbs - 1); 193 } 194 195 static inline unsigned get_max_segment_size(const struct request_queue *q, 196 struct page *start_page, 197 unsigned long offset) 198 { 199 unsigned long mask = queue_segment_boundary(q); 200 201 offset = mask & (page_to_phys(start_page) + offset); 202 203 /* 204 * overflow may be triggered in case of zero page physical address 205 * on 32bit arch, use queue's max segment size when that happens. 206 */ 207 return min_not_zero(mask - offset + 1, 208 (unsigned long)queue_max_segment_size(q)); 209 } 210 211 /** 212 * bvec_split_segs - verify whether or not a bvec should be split in the middle 213 * @q: [in] request queue associated with the bio associated with @bv 214 * @bv: [in] bvec to examine 215 * @nsegs: [in,out] Number of segments in the bio being built. Incremented 216 * by the number of segments from @bv that may be appended to that 217 * bio without exceeding @max_segs 218 * @sectors: [in,out] Number of sectors in the bio being built. Incremented 219 * by the number of sectors from @bv that may be appended to that 220 * bio without exceeding @max_sectors 221 * @max_segs: [in] upper bound for *@nsegs 222 * @max_sectors: [in] upper bound for *@sectors 223 * 224 * When splitting a bio, it can happen that a bvec is encountered that is too 225 * big to fit in a single segment and hence that it has to be split in the 226 * middle. This function verifies whether or not that should happen. The value 227 * %true is returned if and only if appending the entire @bv to a bio with 228 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for 229 * the block driver. 230 */ 231 static bool bvec_split_segs(const struct request_queue *q, 232 const struct bio_vec *bv, unsigned *nsegs, 233 unsigned *sectors, unsigned max_segs, 234 unsigned max_sectors) 235 { 236 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9; 237 unsigned len = min(bv->bv_len, max_len); 238 unsigned total_len = 0; 239 unsigned seg_size = 0; 240 241 while (len && *nsegs < max_segs) { 242 seg_size = get_max_segment_size(q, bv->bv_page, 243 bv->bv_offset + total_len); 244 seg_size = min(seg_size, len); 245 246 (*nsegs)++; 247 total_len += seg_size; 248 len -= seg_size; 249 250 if ((bv->bv_offset + total_len) & queue_virt_boundary(q)) 251 break; 252 } 253 254 *sectors += total_len >> 9; 255 256 /* tell the caller to split the bvec if it is too big to fit */ 257 return len > 0 || bv->bv_len > max_len; 258 } 259 260 /** 261 * blk_bio_segment_split - split a bio in two bios 262 * @q: [in] request queue pointer 263 * @bio: [in] bio to be split 264 * @bs: [in] bio set to allocate the clone from 265 * @segs: [out] number of segments in the bio with the first half of the sectors 266 * 267 * Clone @bio, update the bi_iter of the clone to represent the first sectors 268 * of @bio and update @bio->bi_iter to represent the remaining sectors. The 269 * following is guaranteed for the cloned bio: 270 * - That it has at most get_max_io_size(@q, @bio) sectors. 271 * - That it has at most queue_max_segments(@q) segments. 272 * 273 * Except for discard requests the cloned bio will point at the bi_io_vec of 274 * the original bio. It is the responsibility of the caller to ensure that the 275 * original bio is not freed before the cloned bio. The caller is also 276 * responsible for ensuring that @bs is only destroyed after processing of the 277 * split bio has finished. 278 */ 279 static struct bio *blk_bio_segment_split(struct request_queue *q, 280 struct bio *bio, 281 struct bio_set *bs, 282 unsigned *segs) 283 { 284 struct bio_vec bv, bvprv, *bvprvp = NULL; 285 struct bvec_iter iter; 286 unsigned nsegs = 0, sectors = 0; 287 const unsigned max_sectors = get_max_io_size(q, bio); 288 const unsigned max_segs = queue_max_segments(q); 289 290 bio_for_each_bvec(bv, bio, iter) { 291 /* 292 * If the queue doesn't support SG gaps and adding this 293 * offset would create a gap, disallow it. 294 */ 295 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) 296 goto split; 297 298 if (nsegs < max_segs && 299 sectors + (bv.bv_len >> 9) <= max_sectors && 300 bv.bv_offset + bv.bv_len <= PAGE_SIZE) { 301 nsegs++; 302 sectors += bv.bv_len >> 9; 303 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs, 304 max_sectors)) { 305 goto split; 306 } 307 308 bvprv = bv; 309 bvprvp = &bvprv; 310 } 311 312 *segs = nsegs; 313 return NULL; 314 split: 315 *segs = nsegs; 316 317 /* 318 * Bio splitting may cause subtle trouble such as hang when doing sync 319 * iopoll in direct IO routine. Given performance gain of iopoll for 320 * big IO can be trival, disable iopoll when split needed. 321 */ 322 bio_clear_polled(bio); 323 return bio_split(bio, sectors, GFP_NOIO, bs); 324 } 325 326 /** 327 * __blk_queue_split - split a bio and submit the second half 328 * @q: [in] request_queue new bio is being queued at 329 * @bio: [in, out] bio to be split 330 * @nr_segs: [out] number of segments in the first bio 331 * 332 * Split a bio into two bios, chain the two bios, submit the second half and 333 * store a pointer to the first half in *@bio. If the second bio is still too 334 * big it will be split by a recursive call to this function. Since this 335 * function may allocate a new bio from q->bio_split, it is the responsibility 336 * of the caller to ensure that q->bio_split is only released after processing 337 * of the split bio has finished. 338 */ 339 void __blk_queue_split(struct request_queue *q, struct bio **bio, 340 unsigned int *nr_segs) 341 { 342 struct bio *split = NULL; 343 344 switch (bio_op(*bio)) { 345 case REQ_OP_DISCARD: 346 case REQ_OP_SECURE_ERASE: 347 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs); 348 break; 349 case REQ_OP_WRITE_ZEROES: 350 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, 351 nr_segs); 352 break; 353 case REQ_OP_WRITE_SAME: 354 split = blk_bio_write_same_split(q, *bio, &q->bio_split, 355 nr_segs); 356 break; 357 default: 358 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs); 359 break; 360 } 361 362 if (split) { 363 /* there isn't chance to merge the splitted bio */ 364 split->bi_opf |= REQ_NOMERGE; 365 366 bio_chain(split, *bio); 367 trace_block_split(split, (*bio)->bi_iter.bi_sector); 368 submit_bio_noacct(*bio); 369 *bio = split; 370 371 blk_throtl_charge_bio_split(*bio); 372 } 373 } 374 375 /** 376 * blk_queue_split - split a bio and submit the second half 377 * @bio: [in, out] bio to be split 378 * 379 * Split a bio into two bios, chains the two bios, submit the second half and 380 * store a pointer to the first half in *@bio. Since this function may allocate 381 * a new bio from q->bio_split, it is the responsibility of the caller to ensure 382 * that q->bio_split is only released after processing of the split bio has 383 * finished. 384 */ 385 void blk_queue_split(struct bio **bio) 386 { 387 struct request_queue *q = bdev_get_queue((*bio)->bi_bdev); 388 unsigned int nr_segs; 389 390 if (blk_may_split(q, *bio)) 391 __blk_queue_split(q, bio, &nr_segs); 392 } 393 EXPORT_SYMBOL(blk_queue_split); 394 395 unsigned int blk_recalc_rq_segments(struct request *rq) 396 { 397 unsigned int nr_phys_segs = 0; 398 unsigned int nr_sectors = 0; 399 struct req_iterator iter; 400 struct bio_vec bv; 401 402 if (!rq->bio) 403 return 0; 404 405 switch (bio_op(rq->bio)) { 406 case REQ_OP_DISCARD: 407 case REQ_OP_SECURE_ERASE: 408 if (queue_max_discard_segments(rq->q) > 1) { 409 struct bio *bio = rq->bio; 410 411 for_each_bio(bio) 412 nr_phys_segs++; 413 return nr_phys_segs; 414 } 415 return 1; 416 case REQ_OP_WRITE_ZEROES: 417 return 0; 418 case REQ_OP_WRITE_SAME: 419 return 1; 420 } 421 422 rq_for_each_bvec(bv, rq, iter) 423 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors, 424 UINT_MAX, UINT_MAX); 425 return nr_phys_segs; 426 } 427 428 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg, 429 struct scatterlist *sglist) 430 { 431 if (!*sg) 432 return sglist; 433 434 /* 435 * If the driver previously mapped a shorter list, we could see a 436 * termination bit prematurely unless it fully inits the sg table 437 * on each mapping. We KNOW that there must be more entries here 438 * or the driver would be buggy, so force clear the termination bit 439 * to avoid doing a full sg_init_table() in drivers for each command. 440 */ 441 sg_unmark_end(*sg); 442 return sg_next(*sg); 443 } 444 445 static unsigned blk_bvec_map_sg(struct request_queue *q, 446 struct bio_vec *bvec, struct scatterlist *sglist, 447 struct scatterlist **sg) 448 { 449 unsigned nbytes = bvec->bv_len; 450 unsigned nsegs = 0, total = 0; 451 452 while (nbytes > 0) { 453 unsigned offset = bvec->bv_offset + total; 454 unsigned len = min(get_max_segment_size(q, bvec->bv_page, 455 offset), nbytes); 456 struct page *page = bvec->bv_page; 457 458 /* 459 * Unfortunately a fair number of drivers barf on scatterlists 460 * that have an offset larger than PAGE_SIZE, despite other 461 * subsystems dealing with that invariant just fine. For now 462 * stick to the legacy format where we never present those from 463 * the block layer, but the code below should be removed once 464 * these offenders (mostly MMC/SD drivers) are fixed. 465 */ 466 page += (offset >> PAGE_SHIFT); 467 offset &= ~PAGE_MASK; 468 469 *sg = blk_next_sg(sg, sglist); 470 sg_set_page(*sg, page, len, offset); 471 472 total += len; 473 nbytes -= len; 474 nsegs++; 475 } 476 477 return nsegs; 478 } 479 480 static inline int __blk_bvec_map_sg(struct bio_vec bv, 481 struct scatterlist *sglist, struct scatterlist **sg) 482 { 483 *sg = blk_next_sg(sg, sglist); 484 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); 485 return 1; 486 } 487 488 /* only try to merge bvecs into one sg if they are from two bios */ 489 static inline bool 490 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec, 491 struct bio_vec *bvprv, struct scatterlist **sg) 492 { 493 494 int nbytes = bvec->bv_len; 495 496 if (!*sg) 497 return false; 498 499 if ((*sg)->length + nbytes > queue_max_segment_size(q)) 500 return false; 501 502 if (!biovec_phys_mergeable(q, bvprv, bvec)) 503 return false; 504 505 (*sg)->length += nbytes; 506 507 return true; 508 } 509 510 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, 511 struct scatterlist *sglist, 512 struct scatterlist **sg) 513 { 514 struct bio_vec bvec, bvprv = { NULL }; 515 struct bvec_iter iter; 516 int nsegs = 0; 517 bool new_bio = false; 518 519 for_each_bio(bio) { 520 bio_for_each_bvec(bvec, bio, iter) { 521 /* 522 * Only try to merge bvecs from two bios given we 523 * have done bio internal merge when adding pages 524 * to bio 525 */ 526 if (new_bio && 527 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg)) 528 goto next_bvec; 529 530 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE) 531 nsegs += __blk_bvec_map_sg(bvec, sglist, sg); 532 else 533 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg); 534 next_bvec: 535 new_bio = false; 536 } 537 if (likely(bio->bi_iter.bi_size)) { 538 bvprv = bvec; 539 new_bio = true; 540 } 541 } 542 543 return nsegs; 544 } 545 546 /* 547 * map a request to scatterlist, return number of sg entries setup. Caller 548 * must make sure sg can hold rq->nr_phys_segments entries 549 */ 550 int __blk_rq_map_sg(struct request_queue *q, struct request *rq, 551 struct scatterlist *sglist, struct scatterlist **last_sg) 552 { 553 int nsegs = 0; 554 555 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) 556 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg); 557 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME) 558 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, last_sg); 559 else if (rq->bio) 560 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg); 561 562 if (*last_sg) 563 sg_mark_end(*last_sg); 564 565 /* 566 * Something must have been wrong if the figured number of 567 * segment is bigger than number of req's physical segments 568 */ 569 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); 570 571 return nsegs; 572 } 573 EXPORT_SYMBOL(__blk_rq_map_sg); 574 575 static inline unsigned int blk_rq_get_max_segments(struct request *rq) 576 { 577 if (req_op(rq) == REQ_OP_DISCARD) 578 return queue_max_discard_segments(rq->q); 579 return queue_max_segments(rq->q); 580 } 581 582 static inline unsigned int blk_rq_get_max_sectors(struct request *rq, 583 sector_t offset) 584 { 585 struct request_queue *q = rq->q; 586 587 if (blk_rq_is_passthrough(rq)) 588 return q->limits.max_hw_sectors; 589 590 if (!q->limits.chunk_sectors || 591 req_op(rq) == REQ_OP_DISCARD || 592 req_op(rq) == REQ_OP_SECURE_ERASE) 593 return blk_queue_get_max_sectors(q, req_op(rq)); 594 595 return min(blk_max_size_offset(q, offset, 0), 596 blk_queue_get_max_sectors(q, req_op(rq))); 597 } 598 599 static inline int ll_new_hw_segment(struct request *req, struct bio *bio, 600 unsigned int nr_phys_segs) 601 { 602 if (blk_integrity_merge_bio(req->q, req, bio) == false) 603 goto no_merge; 604 605 /* discard request merge won't add new segment */ 606 if (req_op(req) == REQ_OP_DISCARD) 607 return 1; 608 609 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)) 610 goto no_merge; 611 612 /* 613 * This will form the start of a new hw segment. Bump both 614 * counters. 615 */ 616 req->nr_phys_segments += nr_phys_segs; 617 return 1; 618 619 no_merge: 620 req_set_nomerge(req->q, req); 621 return 0; 622 } 623 624 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs) 625 { 626 if (req_gap_back_merge(req, bio)) 627 return 0; 628 if (blk_integrity_rq(req) && 629 integrity_req_gap_back_merge(req, bio)) 630 return 0; 631 if (!bio_crypt_ctx_back_mergeable(req, bio)) 632 return 0; 633 if (blk_rq_sectors(req) + bio_sectors(bio) > 634 blk_rq_get_max_sectors(req, blk_rq_pos(req))) { 635 req_set_nomerge(req->q, req); 636 return 0; 637 } 638 639 return ll_new_hw_segment(req, bio, nr_segs); 640 } 641 642 static int ll_front_merge_fn(struct request *req, struct bio *bio, 643 unsigned int nr_segs) 644 { 645 if (req_gap_front_merge(req, bio)) 646 return 0; 647 if (blk_integrity_rq(req) && 648 integrity_req_gap_front_merge(req, bio)) 649 return 0; 650 if (!bio_crypt_ctx_front_mergeable(req, bio)) 651 return 0; 652 if (blk_rq_sectors(req) + bio_sectors(bio) > 653 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { 654 req_set_nomerge(req->q, req); 655 return 0; 656 } 657 658 return ll_new_hw_segment(req, bio, nr_segs); 659 } 660 661 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, 662 struct request *next) 663 { 664 unsigned short segments = blk_rq_nr_discard_segments(req); 665 666 if (segments >= queue_max_discard_segments(q)) 667 goto no_merge; 668 if (blk_rq_sectors(req) + bio_sectors(next->bio) > 669 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 670 goto no_merge; 671 672 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); 673 return true; 674 no_merge: 675 req_set_nomerge(q, req); 676 return false; 677 } 678 679 static int ll_merge_requests_fn(struct request_queue *q, struct request *req, 680 struct request *next) 681 { 682 int total_phys_segments; 683 684 if (req_gap_back_merge(req, next->bio)) 685 return 0; 686 687 /* 688 * Will it become too large? 689 */ 690 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > 691 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 692 return 0; 693 694 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; 695 if (total_phys_segments > blk_rq_get_max_segments(req)) 696 return 0; 697 698 if (blk_integrity_merge_rq(q, req, next) == false) 699 return 0; 700 701 if (!bio_crypt_ctx_merge_rq(req, next)) 702 return 0; 703 704 /* Merge is OK... */ 705 req->nr_phys_segments = total_phys_segments; 706 return 1; 707 } 708 709 /** 710 * blk_rq_set_mixed_merge - mark a request as mixed merge 711 * @rq: request to mark as mixed merge 712 * 713 * Description: 714 * @rq is about to be mixed merged. Make sure the attributes 715 * which can be mixed are set in each bio and mark @rq as mixed 716 * merged. 717 */ 718 void blk_rq_set_mixed_merge(struct request *rq) 719 { 720 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; 721 struct bio *bio; 722 723 if (rq->rq_flags & RQF_MIXED_MERGE) 724 return; 725 726 /* 727 * @rq will no longer represent mixable attributes for all the 728 * contained bios. It will just track those of the first one. 729 * Distributes the attributs to each bio. 730 */ 731 for (bio = rq->bio; bio; bio = bio->bi_next) { 732 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && 733 (bio->bi_opf & REQ_FAILFAST_MASK) != ff); 734 bio->bi_opf |= ff; 735 } 736 rq->rq_flags |= RQF_MIXED_MERGE; 737 } 738 739 static void blk_account_io_merge_request(struct request *req) 740 { 741 if (blk_do_io_stat(req)) { 742 part_stat_lock(); 743 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 744 part_stat_unlock(); 745 } 746 } 747 748 static enum elv_merge blk_try_req_merge(struct request *req, 749 struct request *next) 750 { 751 if (blk_discard_mergable(req)) 752 return ELEVATOR_DISCARD_MERGE; 753 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next)) 754 return ELEVATOR_BACK_MERGE; 755 756 return ELEVATOR_NO_MERGE; 757 } 758 759 static inline bool blk_write_same_mergeable(struct bio *a, struct bio *b) 760 { 761 if (bio_page(a) == bio_page(b) && bio_offset(a) == bio_offset(b)) 762 return true; 763 return false; 764 } 765 766 /* 767 * For non-mq, this has to be called with the request spinlock acquired. 768 * For mq with scheduling, the appropriate queue wide lock should be held. 769 */ 770 static struct request *attempt_merge(struct request_queue *q, 771 struct request *req, struct request *next) 772 { 773 if (!rq_mergeable(req) || !rq_mergeable(next)) 774 return NULL; 775 776 if (req_op(req) != req_op(next)) 777 return NULL; 778 779 if (rq_data_dir(req) != rq_data_dir(next) 780 || req->rq_disk != next->rq_disk) 781 return NULL; 782 783 if (req_op(req) == REQ_OP_WRITE_SAME && 784 !blk_write_same_mergeable(req->bio, next->bio)) 785 return NULL; 786 787 /* 788 * Don't allow merge of different write hints, or for a hint with 789 * non-hint IO. 790 */ 791 if (req->write_hint != next->write_hint) 792 return NULL; 793 794 if (req->ioprio != next->ioprio) 795 return NULL; 796 797 /* 798 * If we are allowed to merge, then append bio list 799 * from next to rq and release next. merge_requests_fn 800 * will have updated segment counts, update sector 801 * counts here. Handle DISCARDs separately, as they 802 * have separate settings. 803 */ 804 805 switch (blk_try_req_merge(req, next)) { 806 case ELEVATOR_DISCARD_MERGE: 807 if (!req_attempt_discard_merge(q, req, next)) 808 return NULL; 809 break; 810 case ELEVATOR_BACK_MERGE: 811 if (!ll_merge_requests_fn(q, req, next)) 812 return NULL; 813 break; 814 default: 815 return NULL; 816 } 817 818 /* 819 * If failfast settings disagree or any of the two is already 820 * a mixed merge, mark both as mixed before proceeding. This 821 * makes sure that all involved bios have mixable attributes 822 * set properly. 823 */ 824 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || 825 (req->cmd_flags & REQ_FAILFAST_MASK) != 826 (next->cmd_flags & REQ_FAILFAST_MASK)) { 827 blk_rq_set_mixed_merge(req); 828 blk_rq_set_mixed_merge(next); 829 } 830 831 /* 832 * At this point we have either done a back merge or front merge. We 833 * need the smaller start_time_ns of the merged requests to be the 834 * current request for accounting purposes. 835 */ 836 if (next->start_time_ns < req->start_time_ns) 837 req->start_time_ns = next->start_time_ns; 838 839 req->biotail->bi_next = next->bio; 840 req->biotail = next->biotail; 841 842 req->__data_len += blk_rq_bytes(next); 843 844 if (!blk_discard_mergable(req)) 845 elv_merge_requests(q, req, next); 846 847 /* 848 * 'next' is going away, so update stats accordingly 849 */ 850 blk_account_io_merge_request(next); 851 852 trace_block_rq_merge(next); 853 854 /* 855 * ownership of bio passed from next to req, return 'next' for 856 * the caller to free 857 */ 858 next->bio = NULL; 859 return next; 860 } 861 862 static struct request *attempt_back_merge(struct request_queue *q, 863 struct request *rq) 864 { 865 struct request *next = elv_latter_request(q, rq); 866 867 if (next) 868 return attempt_merge(q, rq, next); 869 870 return NULL; 871 } 872 873 static struct request *attempt_front_merge(struct request_queue *q, 874 struct request *rq) 875 { 876 struct request *prev = elv_former_request(q, rq); 877 878 if (prev) 879 return attempt_merge(q, prev, rq); 880 881 return NULL; 882 } 883 884 /* 885 * Try to merge 'next' into 'rq'. Return true if the merge happened, false 886 * otherwise. The caller is responsible for freeing 'next' if the merge 887 * happened. 888 */ 889 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq, 890 struct request *next) 891 { 892 return attempt_merge(q, rq, next); 893 } 894 895 bool blk_rq_merge_ok(struct request *rq, struct bio *bio) 896 { 897 if (!rq_mergeable(rq) || !bio_mergeable(bio)) 898 return false; 899 900 if (req_op(rq) != bio_op(bio)) 901 return false; 902 903 /* different data direction or already started, don't merge */ 904 if (bio_data_dir(bio) != rq_data_dir(rq)) 905 return false; 906 907 /* must be same device */ 908 if (rq->rq_disk != bio->bi_bdev->bd_disk) 909 return false; 910 911 /* only merge integrity protected bio into ditto rq */ 912 if (blk_integrity_merge_bio(rq->q, rq, bio) == false) 913 return false; 914 915 /* Only merge if the crypt contexts are compatible */ 916 if (!bio_crypt_rq_ctx_compatible(rq, bio)) 917 return false; 918 919 /* must be using the same buffer */ 920 if (req_op(rq) == REQ_OP_WRITE_SAME && 921 !blk_write_same_mergeable(rq->bio, bio)) 922 return false; 923 924 /* 925 * Don't allow merge of different write hints, or for a hint with 926 * non-hint IO. 927 */ 928 if (rq->write_hint != bio->bi_write_hint) 929 return false; 930 931 if (rq->ioprio != bio_prio(bio)) 932 return false; 933 934 return true; 935 } 936 937 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) 938 { 939 if (blk_discard_mergable(rq)) 940 return ELEVATOR_DISCARD_MERGE; 941 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) 942 return ELEVATOR_BACK_MERGE; 943 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) 944 return ELEVATOR_FRONT_MERGE; 945 return ELEVATOR_NO_MERGE; 946 } 947 948 static void blk_account_io_merge_bio(struct request *req) 949 { 950 if (!blk_do_io_stat(req)) 951 return; 952 953 part_stat_lock(); 954 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]); 955 part_stat_unlock(); 956 } 957 958 enum bio_merge_status { 959 BIO_MERGE_OK, 960 BIO_MERGE_NONE, 961 BIO_MERGE_FAILED, 962 }; 963 964 static enum bio_merge_status bio_attempt_back_merge(struct request *req, 965 struct bio *bio, unsigned int nr_segs) 966 { 967 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 968 969 if (!ll_back_merge_fn(req, bio, nr_segs)) 970 return BIO_MERGE_FAILED; 971 972 trace_block_bio_backmerge(bio); 973 rq_qos_merge(req->q, req, bio); 974 975 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 976 blk_rq_set_mixed_merge(req); 977 978 req->biotail->bi_next = bio; 979 req->biotail = bio; 980 req->__data_len += bio->bi_iter.bi_size; 981 982 bio_crypt_free_ctx(bio); 983 984 blk_account_io_merge_bio(req); 985 return BIO_MERGE_OK; 986 } 987 988 static enum bio_merge_status bio_attempt_front_merge(struct request *req, 989 struct bio *bio, unsigned int nr_segs) 990 { 991 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 992 993 if (!ll_front_merge_fn(req, bio, nr_segs)) 994 return BIO_MERGE_FAILED; 995 996 trace_block_bio_frontmerge(bio); 997 rq_qos_merge(req->q, req, bio); 998 999 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 1000 blk_rq_set_mixed_merge(req); 1001 1002 bio->bi_next = req->bio; 1003 req->bio = bio; 1004 1005 req->__sector = bio->bi_iter.bi_sector; 1006 req->__data_len += bio->bi_iter.bi_size; 1007 1008 bio_crypt_do_front_merge(req, bio); 1009 1010 blk_account_io_merge_bio(req); 1011 return BIO_MERGE_OK; 1012 } 1013 1014 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q, 1015 struct request *req, struct bio *bio) 1016 { 1017 unsigned short segments = blk_rq_nr_discard_segments(req); 1018 1019 if (segments >= queue_max_discard_segments(q)) 1020 goto no_merge; 1021 if (blk_rq_sectors(req) + bio_sectors(bio) > 1022 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 1023 goto no_merge; 1024 1025 rq_qos_merge(q, req, bio); 1026 1027 req->biotail->bi_next = bio; 1028 req->biotail = bio; 1029 req->__data_len += bio->bi_iter.bi_size; 1030 req->nr_phys_segments = segments + 1; 1031 1032 blk_account_io_merge_bio(req); 1033 return BIO_MERGE_OK; 1034 no_merge: 1035 req_set_nomerge(q, req); 1036 return BIO_MERGE_FAILED; 1037 } 1038 1039 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q, 1040 struct request *rq, 1041 struct bio *bio, 1042 unsigned int nr_segs, 1043 bool sched_allow_merge) 1044 { 1045 if (!blk_rq_merge_ok(rq, bio)) 1046 return BIO_MERGE_NONE; 1047 1048 switch (blk_try_merge(rq, bio)) { 1049 case ELEVATOR_BACK_MERGE: 1050 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1051 return bio_attempt_back_merge(rq, bio, nr_segs); 1052 break; 1053 case ELEVATOR_FRONT_MERGE: 1054 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio)) 1055 return bio_attempt_front_merge(rq, bio, nr_segs); 1056 break; 1057 case ELEVATOR_DISCARD_MERGE: 1058 return bio_attempt_discard_merge(q, rq, bio); 1059 default: 1060 return BIO_MERGE_NONE; 1061 } 1062 1063 return BIO_MERGE_FAILED; 1064 } 1065 1066 /** 1067 * blk_attempt_plug_merge - try to merge with %current's plugged list 1068 * @q: request_queue new bio is being queued at 1069 * @bio: new bio being queued 1070 * @nr_segs: number of segments in @bio 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) 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 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) == 1100 BIO_MERGE_OK) 1101 return true; 1102 } 1103 return false; 1104 } 1105 1106 /* 1107 * Iterate list of requests and see if we can merge this bio with any 1108 * of them. 1109 */ 1110 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list, 1111 struct bio *bio, unsigned int nr_segs) 1112 { 1113 struct request *rq; 1114 int checked = 8; 1115 1116 list_for_each_entry_reverse(rq, list, queuelist) { 1117 if (!checked--) 1118 break; 1119 1120 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) { 1121 case BIO_MERGE_NONE: 1122 continue; 1123 case BIO_MERGE_OK: 1124 return true; 1125 case BIO_MERGE_FAILED: 1126 return false; 1127 } 1128 1129 } 1130 1131 return false; 1132 } 1133 EXPORT_SYMBOL_GPL(blk_bio_list_merge); 1134 1135 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio, 1136 unsigned int nr_segs, struct request **merged_request) 1137 { 1138 struct request *rq; 1139 1140 switch (elv_merge(q, &rq, bio)) { 1141 case ELEVATOR_BACK_MERGE: 1142 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1143 return false; 1144 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1145 return false; 1146 *merged_request = attempt_back_merge(q, rq); 1147 if (!*merged_request) 1148 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE); 1149 return true; 1150 case ELEVATOR_FRONT_MERGE: 1151 if (!blk_mq_sched_allow_merge(q, rq, bio)) 1152 return false; 1153 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK) 1154 return false; 1155 *merged_request = attempt_front_merge(q, rq); 1156 if (!*merged_request) 1157 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE); 1158 return true; 1159 case ELEVATOR_DISCARD_MERGE: 1160 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK; 1161 default: 1162 return false; 1163 } 1164 } 1165 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge); 1166