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