2 * Functions related to segment and merge handling
4 #include <linux/kernel.h>
5 #include <linux/module.h>
7 #include <linux/blkdev.h>
8 #include <linux/scatterlist.h>
12 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
16 struct bio_vec bv, bvprv = { NULL };
17 int cluster, high, highprv = 1;
18 unsigned int seg_size, nr_phys_segs;
19 struct bio *fbio, *bbio;
20 struct bvec_iter iter;
26 * This should probably be returning 0, but blk_add_request_payload()
29 if (bio->bi_rw & REQ_DISCARD)
32 if (bio->bi_rw & REQ_WRITE_SAME)
36 cluster = blk_queue_cluster(q);
41 bio_for_each_segment(bv, bio, iter) {
43 * If SG merging is disabled, each bio vector is
50 * the trick here is making sure that a high page is
51 * never considered part of another segment, since
52 * that might change with the bounce page.
54 high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
55 if (!high && !highprv && cluster) {
56 if (seg_size + bv.bv_len
57 > queue_max_segment_size(q))
59 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
61 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
64 seg_size += bv.bv_len;
69 if (nr_phys_segs == 1 && seg_size >
70 fbio->bi_seg_front_size)
71 fbio->bi_seg_front_size = seg_size;
81 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
82 fbio->bi_seg_front_size = seg_size;
83 if (seg_size > bbio->bi_seg_back_size)
84 bbio->bi_seg_back_size = seg_size;
89 void blk_recalc_rq_segments(struct request *rq)
91 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
94 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
98 void blk_recount_segments(struct request_queue *q, struct bio *bio)
100 unsigned short seg_cnt;
102 /* estimate segment number by bi_vcnt for non-cloned bio */
103 if (bio_flagged(bio, BIO_CLONED))
104 seg_cnt = bio_segments(bio);
106 seg_cnt = bio->bi_vcnt;
108 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
109 (seg_cnt < queue_max_segments(q)))
110 bio->bi_phys_segments = seg_cnt;
112 struct bio *nxt = bio->bi_next;
115 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
119 bio->bi_flags |= (1 << BIO_SEG_VALID);
121 EXPORT_SYMBOL(blk_recount_segments);
123 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
126 struct bio_vec end_bv = { NULL }, nxt_bv;
127 struct bvec_iter iter;
129 if (!blk_queue_cluster(q))
132 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
133 queue_max_segment_size(q))
136 if (!bio_has_data(bio))
139 bio_for_each_segment(end_bv, bio, iter)
140 if (end_bv.bv_len == iter.bi_size)
143 nxt_bv = bio_iovec(nxt);
145 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
149 * bio and nxt are contiguous in memory; check if the queue allows
150 * these two to be merged into one
152 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
159 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
160 struct scatterlist *sglist, struct bio_vec *bvprv,
161 struct scatterlist **sg, int *nsegs, int *cluster)
164 int nbytes = bvec->bv_len;
166 if (*sg && *cluster) {
167 if ((*sg)->length + nbytes > queue_max_segment_size(q))
170 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
172 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
175 (*sg)->length += nbytes;
182 * If the driver previously mapped a shorter
183 * list, we could see a termination bit
184 * prematurely unless it fully inits the sg
185 * table on each mapping. We KNOW that there
186 * must be more entries here or the driver
187 * would be buggy, so force clear the
188 * termination bit to avoid doing a full
189 * sg_init_table() in drivers for each command.
195 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
201 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
202 struct scatterlist *sglist,
203 struct scatterlist **sg)
205 struct bio_vec bvec, bvprv = { NULL };
206 struct bvec_iter iter;
210 cluster = blk_queue_cluster(q);
212 if (bio->bi_rw & REQ_DISCARD) {
214 * This is a hack - drivers should be neither modifying the
215 * biovec, nor relying on bi_vcnt - but because of
216 * blk_add_request_payload(), a discard bio may or may not have
217 * a payload we need to set up here (thank you Christoph) and
218 * bi_vcnt is really the only way of telling if we need to.
227 if (bio->bi_rw & REQ_WRITE_SAME) {
230 bvec = bio_iovec(bio);
231 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
236 bio_for_each_segment(bvec, bio, iter)
237 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
244 * map a request to scatterlist, return number of sg entries setup. Caller
245 * must make sure sg can hold rq->nr_phys_segments entries
247 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
248 struct scatterlist *sglist)
250 struct scatterlist *sg = NULL;
254 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
256 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
257 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
258 unsigned int pad_len =
259 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
261 sg->length += pad_len;
262 rq->extra_len += pad_len;
265 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
266 if (rq->cmd_flags & REQ_WRITE)
267 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
269 sg->page_link &= ~0x02;
271 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
273 ((unsigned long)q->dma_drain_buffer) &
276 rq->extra_len += q->dma_drain_size;
284 EXPORT_SYMBOL(blk_rq_map_sg);
287 * blk_bio_map_sg - map a bio to a scatterlist
288 * @q: request_queue in question
289 * @bio: bio being mapped
290 * @sglist: scatterlist being mapped
293 * Caller must make sure sg can hold bio->bi_phys_segments entries
295 * Will return the number of sg entries setup
297 int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
298 struct scatterlist *sglist)
300 struct scatterlist *sg = NULL;
302 struct bio *next = bio->bi_next;
305 nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
310 BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
313 EXPORT_SYMBOL(blk_bio_map_sg);
315 static inline int ll_new_hw_segment(struct request_queue *q,
319 int nr_phys_segs = bio_phys_segments(q, bio);
321 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
324 if (blk_integrity_merge_bio(q, req, bio) == false)
328 * This will form the start of a new hw segment. Bump both
331 req->nr_phys_segments += nr_phys_segs;
335 req->cmd_flags |= REQ_NOMERGE;
336 if (req == q->last_merge)
337 q->last_merge = NULL;
341 int ll_back_merge_fn(struct request_queue *q, struct request *req,
344 if (blk_rq_sectors(req) + bio_sectors(bio) >
345 blk_rq_get_max_sectors(req)) {
346 req->cmd_flags |= REQ_NOMERGE;
347 if (req == q->last_merge)
348 q->last_merge = NULL;
351 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
352 blk_recount_segments(q, req->biotail);
353 if (!bio_flagged(bio, BIO_SEG_VALID))
354 blk_recount_segments(q, bio);
356 return ll_new_hw_segment(q, req, bio);
359 int ll_front_merge_fn(struct request_queue *q, struct request *req,
362 if (blk_rq_sectors(req) + bio_sectors(bio) >
363 blk_rq_get_max_sectors(req)) {
364 req->cmd_flags |= REQ_NOMERGE;
365 if (req == q->last_merge)
366 q->last_merge = NULL;
369 if (!bio_flagged(bio, BIO_SEG_VALID))
370 blk_recount_segments(q, bio);
371 if (!bio_flagged(req->bio, BIO_SEG_VALID))
372 blk_recount_segments(q, req->bio);
374 return ll_new_hw_segment(q, req, bio);
378 * blk-mq uses req->special to carry normal driver per-request payload, it
379 * does not indicate a prepared command that we cannot merge with.
381 static bool req_no_special_merge(struct request *req)
383 struct request_queue *q = req->q;
385 return !q->mq_ops && req->special;
388 static int req_gap_to_prev(struct request *req, struct request *next)
390 struct bio *prev = req->biotail;
392 return bvec_gap_to_prev(&prev->bi_io_vec[prev->bi_vcnt - 1],
393 next->bio->bi_io_vec[0].bv_offset);
396 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
397 struct request *next)
399 int total_phys_segments;
400 unsigned int seg_size =
401 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
404 * First check if the either of the requests are re-queued
405 * requests. Can't merge them if they are.
407 if (req_no_special_merge(req) || req_no_special_merge(next))
410 if (test_bit(QUEUE_FLAG_SG_GAPS, &q->queue_flags) &&
411 req_gap_to_prev(req, next))
415 * Will it become too large?
417 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
418 blk_rq_get_max_sectors(req))
421 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
422 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
423 if (req->nr_phys_segments == 1)
424 req->bio->bi_seg_front_size = seg_size;
425 if (next->nr_phys_segments == 1)
426 next->biotail->bi_seg_back_size = seg_size;
427 total_phys_segments--;
430 if (total_phys_segments > queue_max_segments(q))
433 if (blk_integrity_merge_rq(q, req, next) == false)
437 req->nr_phys_segments = total_phys_segments;
442 * blk_rq_set_mixed_merge - mark a request as mixed merge
443 * @rq: request to mark as mixed merge
446 * @rq is about to be mixed merged. Make sure the attributes
447 * which can be mixed are set in each bio and mark @rq as mixed
450 void blk_rq_set_mixed_merge(struct request *rq)
452 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
455 if (rq->cmd_flags & REQ_MIXED_MERGE)
459 * @rq will no longer represent mixable attributes for all the
460 * contained bios. It will just track those of the first one.
461 * Distributes the attributs to each bio.
463 for (bio = rq->bio; bio; bio = bio->bi_next) {
464 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
465 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
468 rq->cmd_flags |= REQ_MIXED_MERGE;
471 static void blk_account_io_merge(struct request *req)
473 if (blk_do_io_stat(req)) {
474 struct hd_struct *part;
477 cpu = part_stat_lock();
480 part_round_stats(cpu, part);
481 part_dec_in_flight(part, rq_data_dir(req));
489 * Has to be called with the request spinlock acquired
491 static int attempt_merge(struct request_queue *q, struct request *req,
492 struct request *next)
494 if (!rq_mergeable(req) || !rq_mergeable(next))
497 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
503 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
506 if (rq_data_dir(req) != rq_data_dir(next)
507 || req->rq_disk != next->rq_disk
508 || req_no_special_merge(next))
511 if (req->cmd_flags & REQ_WRITE_SAME &&
512 !blk_write_same_mergeable(req->bio, next->bio))
516 * If we are allowed to merge, then append bio list
517 * from next to rq and release next. merge_requests_fn
518 * will have updated segment counts, update sector
521 if (!ll_merge_requests_fn(q, req, next))
525 * If failfast settings disagree or any of the two is already
526 * a mixed merge, mark both as mixed before proceeding. This
527 * makes sure that all involved bios have mixable attributes
530 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
531 (req->cmd_flags & REQ_FAILFAST_MASK) !=
532 (next->cmd_flags & REQ_FAILFAST_MASK)) {
533 blk_rq_set_mixed_merge(req);
534 blk_rq_set_mixed_merge(next);
538 * At this point we have either done a back merge
539 * or front merge. We need the smaller start_time of
540 * the merged requests to be the current request
541 * for accounting purposes.
543 if (time_after(req->start_time, next->start_time))
544 req->start_time = next->start_time;
546 req->biotail->bi_next = next->bio;
547 req->biotail = next->biotail;
549 req->__data_len += blk_rq_bytes(next);
551 elv_merge_requests(q, req, next);
554 * 'next' is going away, so update stats accordingly
556 blk_account_io_merge(next);
558 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
559 if (blk_rq_cpu_valid(next))
560 req->cpu = next->cpu;
562 /* owner-ship of bio passed from next to req */
564 __blk_put_request(q, next);
568 int attempt_back_merge(struct request_queue *q, struct request *rq)
570 struct request *next = elv_latter_request(q, rq);
573 return attempt_merge(q, rq, next);
578 int attempt_front_merge(struct request_queue *q, struct request *rq)
580 struct request *prev = elv_former_request(q, rq);
583 return attempt_merge(q, prev, rq);
588 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
589 struct request *next)
591 return attempt_merge(q, rq, next);
594 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
596 struct request_queue *q = rq->q;
598 if (!rq_mergeable(rq) || !bio_mergeable(bio))
601 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
604 /* different data direction or already started, don't merge */
605 if (bio_data_dir(bio) != rq_data_dir(rq))
608 /* must be same device and not a special request */
609 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
612 /* only merge integrity protected bio into ditto rq */
613 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
616 /* must be using the same buffer */
617 if (rq->cmd_flags & REQ_WRITE_SAME &&
618 !blk_write_same_mergeable(rq->bio, bio))
621 if (q->queue_flags & (1 << QUEUE_FLAG_SG_GAPS)) {
622 struct bio_vec *bprev;
624 bprev = &rq->biotail->bi_io_vec[bio->bi_vcnt - 1];
625 if (bvec_gap_to_prev(bprev, bio->bi_io_vec[0].bv_offset))
632 int blk_try_merge(struct request *rq, struct bio *bio)
634 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
635 return ELEVATOR_BACK_MERGE;
636 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
637 return ELEVATOR_FRONT_MERGE;
638 return ELEVATOR_NO_MERGE;