Line data Source code
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/scatterlist.h>
10 :
11 : #include <trace/events/block.h>
12 :
13 : #include "blk.h"
14 : #include "blk-rq-qos.h"
15 :
16 5486 : static inline bool bio_will_gap(struct request_queue *q,
17 : struct request *prev_rq, struct bio *prev, struct bio *next)
18 : {
19 5486 : struct bio_vec pb, nb;
20 :
21 5486 : if (!bio_has_data(prev) || !queue_virt_boundary(q))
22 : return false;
23 :
24 : /*
25 : * Don't merge if the 1st bio starts with non-zero offset, otherwise it
26 : * is quite difficult to respect the sg gap limit. We work hard to
27 : * merge a huge number of small single bios in case of mkfs.
28 : */
29 0 : if (prev_rq)
30 0 : bio_get_first_bvec(prev_rq->bio, &pb);
31 : else
32 0 : bio_get_first_bvec(prev, &pb);
33 0 : if (pb.bv_offset & queue_virt_boundary(q))
34 : return true;
35 :
36 : /*
37 : * We don't need to worry about the situation that the merged segment
38 : * ends in unaligned virt boundary:
39 : *
40 : * - if 'pb' ends aligned, the merged segment ends aligned
41 : * - if 'pb' ends unaligned, the next bio must include
42 : * one single bvec of 'nb', otherwise the 'nb' can't
43 : * merge with 'pb'
44 : */
45 0 : bio_get_last_bvec(prev, &pb);
46 0 : bio_get_first_bvec(next, &nb);
47 0 : if (biovec_phys_mergeable(q, &pb, &nb))
48 : return false;
49 0 : return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
50 : }
51 :
52 5387 : static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
53 : {
54 5387 : return bio_will_gap(req->q, req, req->biotail, bio);
55 : }
56 :
57 99 : static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
58 : {
59 99 : return bio_will_gap(req->q, NULL, bio, req->bio);
60 : }
61 :
62 0 : static struct bio *blk_bio_discard_split(struct request_queue *q,
63 : struct bio *bio,
64 : struct bio_set *bs,
65 : unsigned *nsegs)
66 : {
67 0 : unsigned int max_discard_sectors, granularity;
68 0 : int alignment;
69 0 : sector_t tmp;
70 0 : unsigned split_sectors;
71 :
72 0 : *nsegs = 1;
73 :
74 : /* Zero-sector (unknown) and one-sector granularities are the same. */
75 0 : granularity = max(q->limits.discard_granularity >> 9, 1U);
76 :
77 0 : max_discard_sectors = min(q->limits.max_discard_sectors,
78 : bio_allowed_max_sectors(q));
79 0 : max_discard_sectors -= max_discard_sectors % granularity;
80 :
81 0 : if (unlikely(!max_discard_sectors)) {
82 : /* XXX: warn */
83 : return NULL;
84 : }
85 :
86 0 : if (bio_sectors(bio) <= max_discard_sectors)
87 : return NULL;
88 :
89 0 : split_sectors = max_discard_sectors;
90 :
91 : /*
92 : * If the next starting sector would be misaligned, stop the discard at
93 : * the previous aligned sector.
94 : */
95 0 : alignment = (q->limits.discard_alignment >> 9) % granularity;
96 :
97 0 : tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
98 0 : tmp = sector_div(tmp, granularity);
99 :
100 0 : if (split_sectors > tmp)
101 0 : split_sectors -= tmp;
102 :
103 0 : return bio_split(bio, split_sectors, GFP_NOIO, bs);
104 : }
105 :
106 0 : static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
107 : struct bio *bio, struct bio_set *bs, unsigned *nsegs)
108 : {
109 0 : *nsegs = 0;
110 :
111 0 : if (!q->limits.max_write_zeroes_sectors)
112 : return NULL;
113 :
114 0 : if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
115 : return NULL;
116 :
117 0 : return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
118 : }
119 :
120 0 : static struct bio *blk_bio_write_same_split(struct request_queue *q,
121 : struct bio *bio,
122 : struct bio_set *bs,
123 : unsigned *nsegs)
124 : {
125 0 : *nsegs = 1;
126 :
127 0 : if (!q->limits.max_write_same_sectors)
128 : return NULL;
129 :
130 0 : if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
131 : return NULL;
132 :
133 0 : return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
134 : }
135 :
136 : /*
137 : * Return the maximum number of sectors from the start of a bio that may be
138 : * submitted as a single request to a block device. If enough sectors remain,
139 : * align the end to the physical block size. Otherwise align the end to the
140 : * logical block size. This approach minimizes the number of non-aligned
141 : * requests that are submitted to a block device if the start of a bio is not
142 : * aligned to a physical block boundary.
143 : */
144 1464 : static inline unsigned get_max_io_size(struct request_queue *q,
145 : struct bio *bio)
146 : {
147 1464 : unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0);
148 1464 : unsigned max_sectors = sectors;
149 1464 : unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
150 1464 : unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
151 1464 : unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
152 :
153 1464 : max_sectors += start_offset;
154 1464 : max_sectors &= ~(pbs - 1);
155 1464 : if (max_sectors > start_offset)
156 1464 : return max_sectors - start_offset;
157 :
158 0 : return sectors & ~(lbs - 1);
159 : }
160 :
161 2660 : static inline unsigned get_max_segment_size(const struct request_queue *q,
162 : struct page *start_page,
163 : unsigned long offset)
164 : {
165 2660 : unsigned long mask = queue_segment_boundary(q);
166 :
167 2660 : offset = mask & (page_to_phys(start_page) + offset);
168 :
169 : /*
170 : * overflow may be triggered in case of zero page physical address
171 : * on 32bit arch, use queue's max segment size when that happens.
172 : */
173 2660 : return min_not_zero(mask - offset + 1,
174 : (unsigned long)queue_max_segment_size(q));
175 : }
176 :
177 : /**
178 : * bvec_split_segs - verify whether or not a bvec should be split in the middle
179 : * @q: [in] request queue associated with the bio associated with @bv
180 : * @bv: [in] bvec to examine
181 : * @nsegs: [in,out] Number of segments in the bio being built. Incremented
182 : * by the number of segments from @bv that may be appended to that
183 : * bio without exceeding @max_segs
184 : * @sectors: [in,out] Number of sectors in the bio being built. Incremented
185 : * by the number of sectors from @bv that may be appended to that
186 : * bio without exceeding @max_sectors
187 : * @max_segs: [in] upper bound for *@nsegs
188 : * @max_sectors: [in] upper bound for *@sectors
189 : *
190 : * When splitting a bio, it can happen that a bvec is encountered that is too
191 : * big to fit in a single segment and hence that it has to be split in the
192 : * middle. This function verifies whether or not that should happen. The value
193 : * %true is returned if and only if appending the entire @bv to a bio with
194 : * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
195 : * the block driver.
196 : */
197 1330 : static bool bvec_split_segs(const struct request_queue *q,
198 : const struct bio_vec *bv, unsigned *nsegs,
199 : unsigned *sectors, unsigned max_segs,
200 : unsigned max_sectors)
201 : {
202 1330 : unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
203 1330 : unsigned len = min(bv->bv_len, max_len);
204 1330 : unsigned total_len = 0;
205 1330 : unsigned seg_size = 0;
206 :
207 2660 : while (len && *nsegs < max_segs) {
208 1330 : seg_size = get_max_segment_size(q, bv->bv_page,
209 1330 : bv->bv_offset + total_len);
210 1330 : seg_size = min(seg_size, len);
211 :
212 1330 : (*nsegs)++;
213 1330 : total_len += seg_size;
214 1330 : len -= seg_size;
215 :
216 1330 : if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
217 : break;
218 : }
219 :
220 1330 : *sectors += total_len >> 9;
221 :
222 : /* tell the caller to split the bvec if it is too big to fit */
223 1330 : return len > 0 || bv->bv_len > max_len;
224 : }
225 :
226 : /**
227 : * blk_bio_segment_split - split a bio in two bios
228 : * @q: [in] request queue pointer
229 : * @bio: [in] bio to be split
230 : * @bs: [in] bio set to allocate the clone from
231 : * @segs: [out] number of segments in the bio with the first half of the sectors
232 : *
233 : * Clone @bio, update the bi_iter of the clone to represent the first sectors
234 : * of @bio and update @bio->bi_iter to represent the remaining sectors. The
235 : * following is guaranteed for the cloned bio:
236 : * - That it has at most get_max_io_size(@q, @bio) sectors.
237 : * - That it has at most queue_max_segments(@q) segments.
238 : *
239 : * Except for discard requests the cloned bio will point at the bi_io_vec of
240 : * the original bio. It is the responsibility of the caller to ensure that the
241 : * original bio is not freed before the cloned bio. The caller is also
242 : * responsible for ensuring that @bs is only destroyed after processing of the
243 : * split bio has finished.
244 : */
245 1464 : static struct bio *blk_bio_segment_split(struct request_queue *q,
246 : struct bio *bio,
247 : struct bio_set *bs,
248 : unsigned *segs)
249 : {
250 1464 : struct bio_vec bv, bvprv, *bvprvp = NULL;
251 1464 : struct bvec_iter iter;
252 1464 : unsigned nsegs = 0, sectors = 0;
253 1464 : const unsigned max_sectors = get_max_io_size(q, bio);
254 1464 : const unsigned max_segs = queue_max_segments(q);
255 :
256 6779 : bio_for_each_bvec(bv, bio, iter) {
257 : /*
258 : * If the queue doesn't support SG gaps and adding this
259 : * offset would create a gap, disallow it.
260 : */
261 5315 : if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
262 0 : goto split;
263 :
264 5315 : if (nsegs < max_segs &&
265 5315 : sectors + (bv.bv_len >> 9) <= max_sectors &&
266 5315 : bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
267 3985 : nsegs++;
268 3985 : sectors += bv.bv_len >> 9;
269 1330 : } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs,
270 : max_sectors)) {
271 0 : goto split;
272 : }
273 :
274 5315 : bvprv = bv;
275 5315 : bvprvp = &bvprv;
276 : }
277 :
278 1464 : *segs = nsegs;
279 1464 : return NULL;
280 0 : split:
281 0 : *segs = nsegs;
282 :
283 : /*
284 : * Bio splitting may cause subtle trouble such as hang when doing sync
285 : * iopoll in direct IO routine. Given performance gain of iopoll for
286 : * big IO can be trival, disable iopoll when split needed.
287 : */
288 0 : bio->bi_opf &= ~REQ_HIPRI;
289 :
290 0 : return bio_split(bio, sectors, GFP_NOIO, bs);
291 : }
292 :
293 : /**
294 : * __blk_queue_split - split a bio and submit the second half
295 : * @bio: [in, out] bio to be split
296 : * @nr_segs: [out] number of segments in the first bio
297 : *
298 : * Split a bio into two bios, chain the two bios, submit the second half and
299 : * store a pointer to the first half in *@bio. If the second bio is still too
300 : * big it will be split by a recursive call to this function. Since this
301 : * function may allocate a new bio from q->bio_split, it is the responsibility
302 : * of the caller to ensure that q->bio_split is only released after processing
303 : * of the split bio has finished.
304 : */
305 8736 : void __blk_queue_split(struct bio **bio, unsigned int *nr_segs)
306 : {
307 8736 : struct request_queue *q = (*bio)->bi_bdev->bd_disk->queue;
308 8736 : struct bio *split = NULL;
309 :
310 8736 : switch (bio_op(*bio)) {
311 0 : case REQ_OP_DISCARD:
312 : case REQ_OP_SECURE_ERASE:
313 0 : split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
314 0 : break;
315 0 : case REQ_OP_WRITE_ZEROES:
316 0 : split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
317 : nr_segs);
318 0 : break;
319 0 : case REQ_OP_WRITE_SAME:
320 0 : split = blk_bio_write_same_split(q, *bio, &q->bio_split,
321 : nr_segs);
322 0 : break;
323 8736 : default:
324 : /*
325 : * All drivers must accept single-segments bios that are <=
326 : * PAGE_SIZE. This is a quick and dirty check that relies on
327 : * the fact that bi_io_vec[0] is always valid if a bio has data.
328 : * The check might lead to occasional false negatives when bios
329 : * are cloned, but compared to the performance impact of cloned
330 : * bios themselves the loop below doesn't matter anyway.
331 : */
332 8736 : if (!q->limits.chunk_sectors &&
333 8736 : (*bio)->bi_vcnt == 1 &&
334 8014 : ((*bio)->bi_io_vec[0].bv_len +
335 8014 : (*bio)->bi_io_vec[0].bv_offset) <= PAGE_SIZE) {
336 7273 : *nr_segs = 1;
337 7273 : break;
338 : }
339 1463 : split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
340 1463 : break;
341 : }
342 :
343 8737 : if (split) {
344 : /* there isn't chance to merge the splitted bio */
345 0 : split->bi_opf |= REQ_NOMERGE;
346 :
347 0 : bio_chain(split, *bio);
348 0 : trace_block_split(split, (*bio)->bi_iter.bi_sector);
349 0 : submit_bio_noacct(*bio);
350 0 : *bio = split;
351 : }
352 8737 : }
353 :
354 : /**
355 : * blk_queue_split - split a bio and submit the second half
356 : * @bio: [in, out] bio to be split
357 : *
358 : * Split a bio into two bios, chains the two bios, submit the second half and
359 : * store a pointer to the first half in *@bio. Since this function may allocate
360 : * a new bio from q->bio_split, it is the responsibility of the caller to ensure
361 : * that q->bio_split is only released after processing of the split bio has
362 : * finished.
363 : */
364 0 : void blk_queue_split(struct bio **bio)
365 : {
366 0 : unsigned int nr_segs;
367 :
368 0 : __blk_queue_split(bio, &nr_segs);
369 0 : }
370 : EXPORT_SYMBOL(blk_queue_split);
371 :
372 0 : unsigned int blk_recalc_rq_segments(struct request *rq)
373 : {
374 0 : unsigned int nr_phys_segs = 0;
375 0 : unsigned int nr_sectors = 0;
376 0 : struct req_iterator iter;
377 0 : struct bio_vec bv;
378 :
379 0 : if (!rq->bio)
380 : return 0;
381 :
382 0 : switch (bio_op(rq->bio)) {
383 : case REQ_OP_DISCARD:
384 : case REQ_OP_SECURE_ERASE:
385 : case REQ_OP_WRITE_ZEROES:
386 : return 0;
387 0 : case REQ_OP_WRITE_SAME:
388 0 : return 1;
389 : }
390 :
391 0 : rq_for_each_bvec(bv, rq, iter)
392 0 : bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
393 : UINT_MAX, UINT_MAX);
394 0 : return nr_phys_segs;
395 : }
396 :
397 10720 : static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
398 : struct scatterlist *sglist)
399 : {
400 10720 : if (!*sg)
401 : return sglist;
402 :
403 : /*
404 : * If the driver previously mapped a shorter list, we could see a
405 : * termination bit prematurely unless it fully inits the sg table
406 : * on each mapping. We KNOW that there must be more entries here
407 : * or the driver would be buggy, so force clear the termination bit
408 : * to avoid doing a full sg_init_table() in drivers for each command.
409 : */
410 7538 : sg_unmark_end(*sg);
411 7538 : return sg_next(*sg);
412 : }
413 :
414 1330 : static unsigned blk_bvec_map_sg(struct request_queue *q,
415 : struct bio_vec *bvec, struct scatterlist *sglist,
416 : struct scatterlist **sg)
417 : {
418 1330 : unsigned nbytes = bvec->bv_len;
419 1330 : unsigned nsegs = 0, total = 0;
420 :
421 2660 : while (nbytes > 0) {
422 1330 : unsigned offset = bvec->bv_offset + total;
423 1330 : unsigned len = min(get_max_segment_size(q, bvec->bv_page,
424 : offset), nbytes);
425 1330 : struct page *page = bvec->bv_page;
426 :
427 : /*
428 : * Unfortunately a fair number of drivers barf on scatterlists
429 : * that have an offset larger than PAGE_SIZE, despite other
430 : * subsystems dealing with that invariant just fine. For now
431 : * stick to the legacy format where we never present those from
432 : * the block layer, but the code below should be removed once
433 : * these offenders (mostly MMC/SD drivers) are fixed.
434 : */
435 1330 : page += (offset >> PAGE_SHIFT);
436 1330 : offset &= ~PAGE_MASK;
437 :
438 1330 : *sg = blk_next_sg(sg, sglist);
439 1330 : sg_set_page(*sg, page, len, offset);
440 :
441 1330 : total += len;
442 1330 : nbytes -= len;
443 1330 : nsegs++;
444 : }
445 :
446 1330 : return nsegs;
447 : }
448 :
449 9390 : static inline int __blk_bvec_map_sg(struct bio_vec bv,
450 : struct scatterlist *sglist, struct scatterlist **sg)
451 : {
452 9390 : *sg = blk_next_sg(sg, sglist);
453 9390 : sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
454 9390 : return 1;
455 : }
456 :
457 : /* only try to merge bvecs into one sg if they are from two bios */
458 : static inline bool
459 5485 : __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
460 : struct bio_vec *bvprv, struct scatterlist **sg)
461 : {
462 :
463 5485 : int nbytes = bvec->bv_len;
464 :
465 5485 : if (!*sg)
466 : return false;
467 :
468 5485 : if ((*sg)->length + nbytes > queue_max_segment_size(q))
469 : return false;
470 :
471 5485 : if (!biovec_phys_mergeable(q, bvprv, bvec))
472 : return false;
473 :
474 1870 : (*sg)->length += nbytes;
475 :
476 1870 : return true;
477 : }
478 :
479 3181 : static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
480 : struct scatterlist *sglist,
481 : struct scatterlist **sg)
482 : {
483 3181 : struct bio_vec bvec, bvprv = { NULL };
484 3181 : struct bvec_iter iter;
485 3181 : int nsegs = 0;
486 3181 : bool new_bio = false;
487 :
488 11848 : for_each_bio(bio) {
489 21257 : bio_for_each_bvec(bvec, bio, iter) {
490 : /*
491 : * Only try to merge bvecs from two bios given we
492 : * have done bio internal merge when adding pages
493 : * to bio
494 : */
495 18075 : if (new_bio &&
496 5485 : __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
497 1870 : goto next_bvec;
498 :
499 10720 : if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
500 9390 : nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
501 : else
502 1330 : nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
503 12590 : next_bvec:
504 12590 : new_bio = false;
505 : }
506 8667 : if (likely(bio->bi_iter.bi_size)) {
507 8667 : bvprv = bvec;
508 8667 : new_bio = true;
509 : }
510 : }
511 :
512 3181 : return nsegs;
513 : }
514 :
515 : /*
516 : * map a request to scatterlist, return number of sg entries setup. Caller
517 : * must make sure sg can hold rq->nr_phys_segments entries
518 : */
519 3372 : int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
520 : struct scatterlist *sglist, struct scatterlist **last_sg)
521 : {
522 3372 : int nsegs = 0;
523 :
524 3372 : if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
525 0 : nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
526 3372 : else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
527 0 : nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, last_sg);
528 3372 : else if (rq->bio)
529 3181 : nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
530 :
531 3372 : if (*last_sg)
532 3181 : sg_mark_end(*last_sg);
533 :
534 : /*
535 : * Something must have been wrong if the figured number of
536 : * segment is bigger than number of req's physical segments
537 : */
538 6744 : WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
539 :
540 3372 : return nsegs;
541 : }
542 : EXPORT_SYMBOL(__blk_rq_map_sg);
543 :
544 5486 : static inline unsigned int blk_rq_get_max_segments(struct request *rq)
545 : {
546 5486 : if (req_op(rq) == REQ_OP_DISCARD)
547 0 : return queue_max_discard_segments(rq->q);
548 5486 : return queue_max_segments(rq->q);
549 : }
550 :
551 5486 : static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
552 : unsigned int nr_phys_segs)
553 : {
554 10972 : if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
555 1 : goto no_merge;
556 :
557 5485 : if (blk_integrity_merge_bio(req->q, req, bio) == false)
558 : goto no_merge;
559 :
560 : /*
561 : * This will form the start of a new hw segment. Bump both
562 : * counters.
563 : */
564 5485 : req->nr_phys_segments += nr_phys_segs;
565 5485 : return 1;
566 :
567 1 : no_merge:
568 1 : req_set_nomerge(req->q, req);
569 : return 0;
570 : }
571 :
572 5387 : int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
573 : {
574 5387 : if (req_gap_back_merge(req, bio))
575 : return 0;
576 5387 : if (blk_integrity_rq(req) &&
577 : integrity_req_gap_back_merge(req, bio))
578 : return 0;
579 5387 : if (!bio_crypt_ctx_back_mergeable(req, bio))
580 : return 0;
581 5387 : if (blk_rq_sectors(req) + bio_sectors(bio) >
582 5387 : blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
583 0 : req_set_nomerge(req->q, req);
584 0 : return 0;
585 : }
586 :
587 5387 : return ll_new_hw_segment(req, bio, nr_segs);
588 : }
589 :
590 99 : static int ll_front_merge_fn(struct request *req, struct bio *bio,
591 : unsigned int nr_segs)
592 : {
593 99 : if (req_gap_front_merge(req, bio))
594 : return 0;
595 99 : if (blk_integrity_rq(req) &&
596 : integrity_req_gap_front_merge(req, bio))
597 : return 0;
598 99 : if (!bio_crypt_ctx_front_mergeable(req, bio))
599 : return 0;
600 99 : if (blk_rq_sectors(req) + bio_sectors(bio) >
601 99 : blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
602 0 : req_set_nomerge(req->q, req);
603 0 : return 0;
604 : }
605 :
606 99 : return ll_new_hw_segment(req, bio, nr_segs);
607 : }
608 :
609 0 : static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
610 : struct request *next)
611 : {
612 0 : unsigned short segments = blk_rq_nr_discard_segments(req);
613 :
614 0 : if (segments >= queue_max_discard_segments(q))
615 0 : goto no_merge;
616 0 : if (blk_rq_sectors(req) + bio_sectors(next->bio) >
617 0 : blk_rq_get_max_sectors(req, blk_rq_pos(req)))
618 0 : goto no_merge;
619 :
620 0 : req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
621 0 : return true;
622 0 : no_merge:
623 0 : req_set_nomerge(q, req);
624 : return false;
625 : }
626 :
627 0 : static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
628 : struct request *next)
629 : {
630 0 : int total_phys_segments;
631 :
632 0 : if (req_gap_back_merge(req, next->bio))
633 : return 0;
634 :
635 : /*
636 : * Will it become too large?
637 : */
638 0 : if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
639 0 : blk_rq_get_max_sectors(req, blk_rq_pos(req)))
640 : return 0;
641 :
642 0 : total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
643 0 : if (total_phys_segments > blk_rq_get_max_segments(req))
644 : return 0;
645 :
646 0 : if (blk_integrity_merge_rq(q, req, next) == false)
647 : return 0;
648 :
649 0 : if (!bio_crypt_ctx_merge_rq(req, next))
650 : return 0;
651 :
652 : /* Merge is OK... */
653 0 : req->nr_phys_segments = total_phys_segments;
654 0 : return 1;
655 : }
656 :
657 : /**
658 : * blk_rq_set_mixed_merge - mark a request as mixed merge
659 : * @rq: request to mark as mixed merge
660 : *
661 : * Description:
662 : * @rq is about to be mixed merged. Make sure the attributes
663 : * which can be mixed are set in each bio and mark @rq as mixed
664 : * merged.
665 : */
666 3101 : void blk_rq_set_mixed_merge(struct request *rq)
667 : {
668 3101 : unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
669 3101 : struct bio *bio;
670 :
671 3101 : if (rq->rq_flags & RQF_MIXED_MERGE)
672 : return;
673 :
674 : /*
675 : * @rq will no longer represent mixable attributes for all the
676 : * contained bios. It will just track those of the first one.
677 : * Distributes the attributs to each bio.
678 : */
679 376 : for (bio = rq->bio; bio; bio = bio->bi_next) {
680 376 : WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
681 : (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
682 188 : bio->bi_opf |= ff;
683 : }
684 188 : rq->rq_flags |= RQF_MIXED_MERGE;
685 : }
686 :
687 0 : static void blk_account_io_merge_request(struct request *req)
688 : {
689 0 : if (blk_do_io_stat(req)) {
690 0 : part_stat_lock();
691 0 : part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
692 0 : part_stat_unlock();
693 : }
694 0 : }
695 :
696 : /*
697 : * Two cases of handling DISCARD merge:
698 : * If max_discard_segments > 1, the driver takes every bio
699 : * as a range and send them to controller together. The ranges
700 : * needn't to be contiguous.
701 : * Otherwise, the bios/requests will be handled as same as
702 : * others which should be contiguous.
703 : */
704 9293 : static inline bool blk_discard_mergable(struct request *req)
705 : {
706 0 : if (req_op(req) == REQ_OP_DISCARD &&
707 0 : queue_max_discard_segments(req->q) > 1)
708 : return true;
709 : return false;
710 : }
711 :
712 0 : static enum elv_merge blk_try_req_merge(struct request *req,
713 : struct request *next)
714 : {
715 0 : if (blk_discard_mergable(req))
716 : return ELEVATOR_DISCARD_MERGE;
717 0 : else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
718 0 : return ELEVATOR_BACK_MERGE;
719 :
720 : return ELEVATOR_NO_MERGE;
721 : }
722 :
723 : /*
724 : * For non-mq, this has to be called with the request spinlock acquired.
725 : * For mq with scheduling, the appropriate queue wide lock should be held.
726 : */
727 0 : static struct request *attempt_merge(struct request_queue *q,
728 : struct request *req, struct request *next)
729 : {
730 0 : if (!rq_mergeable(req) || !rq_mergeable(next))
731 : return NULL;
732 :
733 0 : if (req_op(req) != req_op(next))
734 : return NULL;
735 :
736 0 : if (rq_data_dir(req) != rq_data_dir(next)
737 0 : || req->rq_disk != next->rq_disk)
738 : return NULL;
739 :
740 0 : if (req_op(req) == REQ_OP_WRITE_SAME &&
741 0 : !blk_write_same_mergeable(req->bio, next->bio))
742 : return NULL;
743 :
744 : /*
745 : * Don't allow merge of different write hints, or for a hint with
746 : * non-hint IO.
747 : */
748 0 : if (req->write_hint != next->write_hint)
749 : return NULL;
750 :
751 0 : if (req->ioprio != next->ioprio)
752 : return NULL;
753 :
754 : /*
755 : * If we are allowed to merge, then append bio list
756 : * from next to rq and release next. merge_requests_fn
757 : * will have updated segment counts, update sector
758 : * counts here. Handle DISCARDs separately, as they
759 : * have separate settings.
760 : */
761 :
762 0 : switch (blk_try_req_merge(req, next)) {
763 0 : case ELEVATOR_DISCARD_MERGE:
764 0 : if (!req_attempt_discard_merge(q, req, next))
765 : return NULL;
766 : break;
767 0 : case ELEVATOR_BACK_MERGE:
768 0 : if (!ll_merge_requests_fn(q, req, next))
769 : return NULL;
770 : break;
771 : default:
772 : return NULL;
773 : }
774 :
775 : /*
776 : * If failfast settings disagree or any of the two is already
777 : * a mixed merge, mark both as mixed before proceeding. This
778 : * makes sure that all involved bios have mixable attributes
779 : * set properly.
780 : */
781 0 : if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
782 0 : (req->cmd_flags & REQ_FAILFAST_MASK) !=
783 0 : (next->cmd_flags & REQ_FAILFAST_MASK)) {
784 0 : blk_rq_set_mixed_merge(req);
785 0 : blk_rq_set_mixed_merge(next);
786 : }
787 :
788 : /*
789 : * At this point we have either done a back merge or front merge. We
790 : * need the smaller start_time_ns of the merged requests to be the
791 : * current request for accounting purposes.
792 : */
793 0 : if (next->start_time_ns < req->start_time_ns)
794 0 : req->start_time_ns = next->start_time_ns;
795 :
796 0 : req->biotail->bi_next = next->bio;
797 0 : req->biotail = next->biotail;
798 :
799 0 : req->__data_len += blk_rq_bytes(next);
800 :
801 0 : if (!blk_discard_mergable(req))
802 0 : elv_merge_requests(q, req, next);
803 :
804 : /*
805 : * 'next' is going away, so update stats accordingly
806 : */
807 0 : blk_account_io_merge_request(next);
808 :
809 0 : trace_block_rq_merge(next);
810 :
811 : /*
812 : * ownership of bio passed from next to req, return 'next' for
813 : * the caller to free
814 : */
815 0 : next->bio = NULL;
816 0 : return next;
817 : }
818 :
819 0 : static struct request *attempt_back_merge(struct request_queue *q,
820 : struct request *rq)
821 : {
822 0 : struct request *next = elv_latter_request(q, rq);
823 :
824 0 : if (next)
825 0 : return attempt_merge(q, rq, next);
826 :
827 : return NULL;
828 : }
829 :
830 0 : static struct request *attempt_front_merge(struct request_queue *q,
831 : struct request *rq)
832 : {
833 0 : struct request *prev = elv_former_request(q, rq);
834 :
835 0 : if (prev)
836 0 : return attempt_merge(q, prev, rq);
837 :
838 : return NULL;
839 : }
840 :
841 0 : int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
842 : struct request *next)
843 : {
844 0 : struct request *free;
845 :
846 0 : free = attempt_merge(q, rq, next);
847 0 : if (free) {
848 0 : blk_put_request(free);
849 0 : return 1;
850 : }
851 :
852 : return 0;
853 : }
854 :
855 9293 : bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
856 : {
857 9293 : if (!rq_mergeable(rq) || !bio_mergeable(bio))
858 : return false;
859 :
860 9293 : if (req_op(rq) != bio_op(bio))
861 : return false;
862 :
863 : /* different data direction or already started, don't merge */
864 9293 : if (bio_data_dir(bio) != rq_data_dir(rq))
865 : return false;
866 :
867 : /* must be same device */
868 9293 : if (rq->rq_disk != bio->bi_bdev->bd_disk)
869 : return false;
870 :
871 : /* only merge integrity protected bio into ditto rq */
872 9293 : if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
873 : return false;
874 :
875 : /* Only merge if the crypt contexts are compatible */
876 9293 : if (!bio_crypt_rq_ctx_compatible(rq, bio))
877 : return false;
878 :
879 : /* must be using the same buffer */
880 9293 : if (req_op(rq) == REQ_OP_WRITE_SAME &&
881 0 : !blk_write_same_mergeable(rq->bio, bio))
882 : return false;
883 :
884 : /*
885 : * Don't allow merge of different write hints, or for a hint with
886 : * non-hint IO.
887 : */
888 9293 : if (rq->write_hint != bio->bi_write_hint)
889 : return false;
890 :
891 9293 : if (rq->ioprio != bio_prio(bio))
892 0 : return false;
893 :
894 : return true;
895 : }
896 :
897 9293 : enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
898 : {
899 9293 : if (blk_discard_mergable(rq))
900 : return ELEVATOR_DISCARD_MERGE;
901 9293 : else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
902 : return ELEVATOR_BACK_MERGE;
903 3906 : else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
904 99 : return ELEVATOR_FRONT_MERGE;
905 : return ELEVATOR_NO_MERGE;
906 : }
907 :
908 5485 : static void blk_account_io_merge_bio(struct request *req)
909 : {
910 10970 : if (!blk_do_io_stat(req))
911 : return;
912 :
913 5485 : part_stat_lock();
914 16455 : part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
915 5485 : part_stat_unlock();
916 : }
917 :
918 : enum bio_merge_status {
919 : BIO_MERGE_OK,
920 : BIO_MERGE_NONE,
921 : BIO_MERGE_FAILED,
922 : };
923 :
924 5387 : static enum bio_merge_status bio_attempt_back_merge(struct request *req,
925 : struct bio *bio, unsigned int nr_segs)
926 : {
927 5387 : const int ff = bio->bi_opf & REQ_FAILFAST_MASK;
928 :
929 5387 : if (!ll_back_merge_fn(req, bio, nr_segs))
930 : return BIO_MERGE_FAILED;
931 :
932 5386 : trace_block_bio_backmerge(bio);
933 5386 : rq_qos_merge(req->q, req, bio);
934 :
935 5386 : if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
936 3002 : blk_rq_set_mixed_merge(req);
937 :
938 5386 : req->biotail->bi_next = bio;
939 5386 : req->biotail = bio;
940 5386 : req->__data_len += bio->bi_iter.bi_size;
941 :
942 5386 : bio_crypt_free_ctx(bio);
943 :
944 5386 : blk_account_io_merge_bio(req);
945 5386 : return BIO_MERGE_OK;
946 : }
947 :
948 99 : static enum bio_merge_status bio_attempt_front_merge(struct request *req,
949 : struct bio *bio, unsigned int nr_segs)
950 : {
951 99 : const int ff = bio->bi_opf & REQ_FAILFAST_MASK;
952 :
953 99 : if (!ll_front_merge_fn(req, bio, nr_segs))
954 : return BIO_MERGE_FAILED;
955 :
956 99 : trace_block_bio_frontmerge(bio);
957 99 : rq_qos_merge(req->q, req, bio);
958 :
959 99 : if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
960 99 : blk_rq_set_mixed_merge(req);
961 :
962 99 : bio->bi_next = req->bio;
963 99 : req->bio = bio;
964 :
965 99 : req->__sector = bio->bi_iter.bi_sector;
966 99 : req->__data_len += bio->bi_iter.bi_size;
967 :
968 99 : bio_crypt_do_front_merge(req, bio);
969 :
970 99 : blk_account_io_merge_bio(req);
971 99 : return BIO_MERGE_OK;
972 : }
973 :
974 0 : static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q,
975 : struct request *req, struct bio *bio)
976 : {
977 0 : unsigned short segments = blk_rq_nr_discard_segments(req);
978 :
979 0 : if (segments >= queue_max_discard_segments(q))
980 0 : goto no_merge;
981 0 : if (blk_rq_sectors(req) + bio_sectors(bio) >
982 0 : blk_rq_get_max_sectors(req, blk_rq_pos(req)))
983 0 : goto no_merge;
984 :
985 0 : rq_qos_merge(q, req, bio);
986 :
987 0 : req->biotail->bi_next = bio;
988 0 : req->biotail = bio;
989 0 : req->__data_len += bio->bi_iter.bi_size;
990 0 : req->nr_phys_segments = segments + 1;
991 :
992 0 : blk_account_io_merge_bio(req);
993 0 : return BIO_MERGE_OK;
994 0 : no_merge:
995 0 : req_set_nomerge(q, req);
996 : return BIO_MERGE_FAILED;
997 : }
998 :
999 9293 : static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q,
1000 : struct request *rq,
1001 : struct bio *bio,
1002 : unsigned int nr_segs,
1003 : bool sched_allow_merge)
1004 : {
1005 9293 : if (!blk_rq_merge_ok(rq, bio))
1006 : return BIO_MERGE_NONE;
1007 :
1008 9293 : switch (blk_try_merge(rq, bio)) {
1009 5387 : case ELEVATOR_BACK_MERGE:
1010 5387 : if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1011 5387 : return bio_attempt_back_merge(rq, bio, nr_segs);
1012 : break;
1013 99 : case ELEVATOR_FRONT_MERGE:
1014 99 : if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1015 99 : return bio_attempt_front_merge(rq, bio, nr_segs);
1016 : break;
1017 0 : case ELEVATOR_DISCARD_MERGE:
1018 0 : return bio_attempt_discard_merge(q, rq, bio);
1019 : default:
1020 : return BIO_MERGE_NONE;
1021 : }
1022 :
1023 : return BIO_MERGE_FAILED;
1024 : }
1025 :
1026 : /**
1027 : * blk_attempt_plug_merge - try to merge with %current's plugged list
1028 : * @q: request_queue new bio is being queued at
1029 : * @bio: new bio being queued
1030 : * @nr_segs: number of segments in @bio
1031 : * @same_queue_rq: pointer to &struct request that gets filled in when
1032 : * another request associated with @q is found on the plug list
1033 : * (optional, may be %NULL)
1034 : *
1035 : * Determine whether @bio being queued on @q can be merged with a request
1036 : * on %current's plugged list. Returns %true if merge was successful,
1037 : * otherwise %false.
1038 : *
1039 : * Plugging coalesces IOs from the same issuer for the same purpose without
1040 : * going through @q->queue_lock. As such it's more of an issuing mechanism
1041 : * than scheduling, and the request, while may have elvpriv data, is not
1042 : * added on the elevator at this point. In addition, we don't have
1043 : * reliable access to the elevator outside queue lock. Only check basic
1044 : * merging parameters without querying the elevator.
1045 : *
1046 : * Caller must ensure !blk_queue_nomerges(q) beforehand.
1047 : */
1048 8605 : bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1049 : unsigned int nr_segs, struct request **same_queue_rq)
1050 : {
1051 8605 : struct blk_plug *plug;
1052 8605 : struct request *rq;
1053 8605 : struct list_head *plug_list;
1054 :
1055 8605 : plug = blk_mq_plug(q, bio);
1056 8605 : if (!plug)
1057 : return false;
1058 :
1059 7999 : plug_list = &plug->mq_list;
1060 :
1061 11807 : list_for_each_entry_reverse(rq, plug_list, queuelist) {
1062 9293 : if (rq->q == q && same_queue_rq) {
1063 : /*
1064 : * Only blk-mq multiple hardware queues case checks the
1065 : * rq in the same queue, there should be only one such
1066 : * rq in a queue
1067 : **/
1068 9293 : *same_queue_rq = rq;
1069 : }
1070 :
1071 9293 : if (rq->q != q)
1072 0 : continue;
1073 :
1074 9293 : if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) ==
1075 : BIO_MERGE_OK)
1076 : return true;
1077 : }
1078 :
1079 : return false;
1080 : }
1081 :
1082 : /*
1083 : * Iterate list of requests and see if we can merge this bio with any
1084 : * of them.
1085 : */
1086 0 : bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
1087 : struct bio *bio, unsigned int nr_segs)
1088 : {
1089 0 : struct request *rq;
1090 0 : int checked = 8;
1091 :
1092 0 : list_for_each_entry_reverse(rq, list, queuelist) {
1093 0 : if (!checked--)
1094 : break;
1095 :
1096 0 : switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) {
1097 0 : case BIO_MERGE_NONE:
1098 0 : continue;
1099 : case BIO_MERGE_OK:
1100 : return true;
1101 0 : case BIO_MERGE_FAILED:
1102 0 : return false;
1103 : }
1104 :
1105 : }
1106 :
1107 : return false;
1108 : }
1109 : EXPORT_SYMBOL_GPL(blk_bio_list_merge);
1110 :
1111 0 : bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
1112 : unsigned int nr_segs, struct request **merged_request)
1113 : {
1114 0 : struct request *rq;
1115 :
1116 0 : switch (elv_merge(q, &rq, bio)) {
1117 0 : case ELEVATOR_BACK_MERGE:
1118 0 : if (!blk_mq_sched_allow_merge(q, rq, bio))
1119 : return false;
1120 0 : if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1121 : return false;
1122 0 : *merged_request = attempt_back_merge(q, rq);
1123 0 : if (!*merged_request)
1124 0 : elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
1125 : return true;
1126 0 : case ELEVATOR_FRONT_MERGE:
1127 0 : if (!blk_mq_sched_allow_merge(q, rq, bio))
1128 : return false;
1129 0 : if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1130 : return false;
1131 0 : *merged_request = attempt_front_merge(q, rq);
1132 0 : if (!*merged_request)
1133 0 : elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
1134 : return true;
1135 0 : case ELEVATOR_DISCARD_MERGE:
1136 0 : return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
1137 : default:
1138 : return false;
1139 : }
1140 : }
1141 : EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
|
