]>
Commit | Line | Data |
---|---|---|
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 | ||
15 | static struct bio *blk_bio_discard_split(struct request_queue *q, | |
16 | struct bio *bio, | |
17 | struct bio_set *bs, | |
18 | unsigned *nsegs) | |
19 | { | |
20 | unsigned int max_discard_sectors, granularity; | |
21 | int alignment; | |
22 | sector_t tmp; | |
23 | unsigned split_sectors; | |
24 | ||
25 | *nsegs = 1; | |
26 | ||
27 | /* Zero-sector (unknown) and one-sector granularities are the same. */ | |
28 | granularity = max(q->limits.discard_granularity >> 9, 1U); | |
29 | ||
30 | max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9); | |
31 | max_discard_sectors -= max_discard_sectors % granularity; | |
32 | ||
33 | if (unlikely(!max_discard_sectors)) { | |
34 | /* XXX: warn */ | |
35 | return NULL; | |
36 | } | |
37 | ||
38 | if (bio_sectors(bio) <= max_discard_sectors) | |
39 | return NULL; | |
40 | ||
41 | split_sectors = max_discard_sectors; | |
42 | ||
43 | /* | |
44 | * If the next starting sector would be misaligned, stop the discard at | |
45 | * the previous aligned sector. | |
46 | */ | |
47 | alignment = (q->limits.discard_alignment >> 9) % granularity; | |
48 | ||
49 | tmp = bio->bi_iter.bi_sector + split_sectors - alignment; | |
50 | tmp = sector_div(tmp, granularity); | |
51 | ||
52 | if (split_sectors > tmp) | |
53 | split_sectors -= tmp; | |
54 | ||
55 | return bio_split(bio, split_sectors, GFP_NOIO, bs); | |
56 | } | |
57 | ||
58 | static struct bio *blk_bio_write_zeroes_split(struct request_queue *q, | |
59 | struct bio *bio, struct bio_set *bs, unsigned *nsegs) | |
60 | { | |
61 | *nsegs = 1; | |
62 | ||
63 | if (!q->limits.max_write_zeroes_sectors) | |
64 | return NULL; | |
65 | ||
66 | if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) | |
67 | return NULL; | |
68 | ||
69 | return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); | |
70 | } | |
71 | ||
72 | static struct bio *blk_bio_write_same_split(struct request_queue *q, | |
73 | struct bio *bio, | |
74 | struct bio_set *bs, | |
75 | unsigned *nsegs) | |
76 | { | |
77 | *nsegs = 1; | |
78 | ||
79 | if (!q->limits.max_write_same_sectors) | |
80 | return NULL; | |
81 | ||
82 | if (bio_sectors(bio) <= q->limits.max_write_same_sectors) | |
83 | return NULL; | |
84 | ||
85 | return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); | |
86 | } | |
87 | ||
88 | static inline unsigned get_max_io_size(struct request_queue *q, | |
89 | struct bio *bio) | |
90 | { | |
91 | unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector); | |
92 | unsigned mask = queue_logical_block_size(q) - 1; | |
93 | ||
94 | /* aligned to logical block size */ | |
95 | sectors &= ~(mask >> 9); | |
96 | ||
97 | return sectors; | |
98 | } | |
99 | ||
100 | static struct bio *blk_bio_segment_split(struct request_queue *q, | |
101 | struct bio *bio, | |
102 | struct bio_set *bs, | |
103 | unsigned *segs) | |
104 | { | |
105 | struct bio_vec bv, bvprv, *bvprvp = NULL; | |
106 | struct bvec_iter iter; | |
107 | unsigned seg_size = 0, nsegs = 0, sectors = 0; | |
108 | unsigned front_seg_size = bio->bi_seg_front_size; | |
109 | bool do_split = true; | |
110 | struct bio *new = NULL; | |
111 | const unsigned max_sectors = get_max_io_size(q, bio); | |
112 | ||
113 | bio_for_each_segment(bv, bio, iter) { | |
114 | /* | |
115 | * If the queue doesn't support SG gaps and adding this | |
116 | * offset would create a gap, disallow it. | |
117 | */ | |
118 | if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) | |
119 | goto split; | |
120 | ||
121 | if (sectors + (bv.bv_len >> 9) > max_sectors) { | |
122 | /* | |
123 | * Consider this a new segment if we're splitting in | |
124 | * the middle of this vector. | |
125 | */ | |
126 | if (nsegs < queue_max_segments(q) && | |
127 | sectors < max_sectors) { | |
128 | nsegs++; | |
129 | sectors = max_sectors; | |
130 | } | |
131 | if (sectors) | |
132 | goto split; | |
133 | /* Make this single bvec as the 1st segment */ | |
134 | } | |
135 | ||
136 | if (bvprvp && blk_queue_cluster(q)) { | |
137 | if (seg_size + bv.bv_len > queue_max_segment_size(q)) | |
138 | goto new_segment; | |
139 | if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv)) | |
140 | goto new_segment; | |
141 | if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv)) | |
142 | goto new_segment; | |
143 | ||
144 | seg_size += bv.bv_len; | |
145 | bvprv = bv; | |
146 | bvprvp = &bvprv; | |
147 | sectors += bv.bv_len >> 9; | |
148 | ||
149 | if (nsegs == 1 && seg_size > front_seg_size) | |
150 | front_seg_size = seg_size; | |
151 | continue; | |
152 | } | |
153 | new_segment: | |
154 | if (nsegs == queue_max_segments(q)) | |
155 | goto split; | |
156 | ||
157 | nsegs++; | |
158 | bvprv = bv; | |
159 | bvprvp = &bvprv; | |
160 | seg_size = bv.bv_len; | |
161 | sectors += bv.bv_len >> 9; | |
162 | ||
163 | if (nsegs == 1 && seg_size > front_seg_size) | |
164 | front_seg_size = seg_size; | |
165 | } | |
166 | ||
167 | do_split = false; | |
168 | split: | |
169 | *segs = nsegs; | |
170 | ||
171 | if (do_split) { | |
172 | new = bio_split(bio, sectors, GFP_NOIO, bs); | |
173 | if (new) | |
174 | bio = new; | |
175 | } | |
176 | ||
177 | bio->bi_seg_front_size = front_seg_size; | |
178 | if (seg_size > bio->bi_seg_back_size) | |
179 | bio->bi_seg_back_size = seg_size; | |
180 | ||
181 | return do_split ? new : NULL; | |
182 | } | |
183 | ||
184 | void blk_queue_split(struct request_queue *q, struct bio **bio) | |
185 | { | |
186 | struct bio *split, *res; | |
187 | unsigned nsegs; | |
188 | ||
189 | switch (bio_op(*bio)) { | |
190 | case REQ_OP_DISCARD: | |
191 | case REQ_OP_SECURE_ERASE: | |
192 | split = blk_bio_discard_split(q, *bio, q->bio_split, &nsegs); | |
193 | break; | |
194 | case REQ_OP_WRITE_ZEROES: | |
195 | split = blk_bio_write_zeroes_split(q, *bio, q->bio_split, &nsegs); | |
196 | break; | |
197 | case REQ_OP_WRITE_SAME: | |
198 | split = blk_bio_write_same_split(q, *bio, q->bio_split, &nsegs); | |
199 | break; | |
200 | default: | |
201 | split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs); | |
202 | break; | |
203 | } | |
204 | ||
205 | /* physical segments can be figured out during splitting */ | |
206 | res = split ? split : *bio; | |
207 | res->bi_phys_segments = nsegs; | |
208 | bio_set_flag(res, BIO_SEG_VALID); | |
209 | ||
210 | if (split) { | |
211 | /* there isn't chance to merge the splitted bio */ | |
212 | split->bi_opf |= REQ_NOMERGE; | |
213 | ||
214 | bio_chain(split, *bio); | |
215 | trace_block_split(q, split, (*bio)->bi_iter.bi_sector); | |
216 | generic_make_request(*bio); | |
217 | *bio = split; | |
218 | } | |
219 | } | |
220 | EXPORT_SYMBOL(blk_queue_split); | |
221 | ||
222 | static unsigned int __blk_recalc_rq_segments(struct request_queue *q, | |
223 | struct bio *bio, | |
224 | bool no_sg_merge) | |
225 | { | |
226 | struct bio_vec bv, bvprv = { NULL }; | |
227 | int cluster, prev = 0; | |
228 | unsigned int seg_size, nr_phys_segs; | |
229 | struct bio *fbio, *bbio; | |
230 | struct bvec_iter iter; | |
231 | ||
232 | if (!bio) | |
233 | return 0; | |
234 | ||
235 | switch (bio_op(bio)) { | |
236 | case REQ_OP_DISCARD: | |
237 | case REQ_OP_SECURE_ERASE: | |
238 | case REQ_OP_WRITE_ZEROES: | |
239 | return 0; | |
240 | case REQ_OP_WRITE_SAME: | |
241 | return 1; | |
242 | } | |
243 | ||
244 | fbio = bio; | |
245 | cluster = blk_queue_cluster(q); | |
246 | seg_size = 0; | |
247 | nr_phys_segs = 0; | |
248 | for_each_bio(bio) { | |
249 | bio_for_each_segment(bv, bio, iter) { | |
250 | /* | |
251 | * If SG merging is disabled, each bio vector is | |
252 | * a segment | |
253 | */ | |
254 | if (no_sg_merge) | |
255 | goto new_segment; | |
256 | ||
257 | if (prev && cluster) { | |
258 | if (seg_size + bv.bv_len | |
259 | > queue_max_segment_size(q)) | |
260 | goto new_segment; | |
261 | if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv)) | |
262 | goto new_segment; | |
263 | if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv)) | |
264 | goto new_segment; | |
265 | ||
266 | seg_size += bv.bv_len; | |
267 | bvprv = bv; | |
268 | continue; | |
269 | } | |
270 | new_segment: | |
271 | if (nr_phys_segs == 1 && seg_size > | |
272 | fbio->bi_seg_front_size) | |
273 | fbio->bi_seg_front_size = seg_size; | |
274 | ||
275 | nr_phys_segs++; | |
276 | bvprv = bv; | |
277 | prev = 1; | |
278 | seg_size = bv.bv_len; | |
279 | } | |
280 | bbio = bio; | |
281 | } | |
282 | ||
283 | if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size) | |
284 | fbio->bi_seg_front_size = seg_size; | |
285 | if (seg_size > bbio->bi_seg_back_size) | |
286 | bbio->bi_seg_back_size = seg_size; | |
287 | ||
288 | return nr_phys_segs; | |
289 | } | |
290 | ||
291 | void blk_recalc_rq_segments(struct request *rq) | |
292 | { | |
293 | bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE, | |
294 | &rq->q->queue_flags); | |
295 | ||
296 | rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio, | |
297 | no_sg_merge); | |
298 | } | |
299 | ||
300 | void blk_recount_segments(struct request_queue *q, struct bio *bio) | |
301 | { | |
302 | unsigned short seg_cnt; | |
303 | ||
304 | /* estimate segment number by bi_vcnt for non-cloned bio */ | |
305 | if (bio_flagged(bio, BIO_CLONED)) | |
306 | seg_cnt = bio_segments(bio); | |
307 | else | |
308 | seg_cnt = bio->bi_vcnt; | |
309 | ||
310 | if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) && | |
311 | (seg_cnt < queue_max_segments(q))) | |
312 | bio->bi_phys_segments = seg_cnt; | |
313 | else { | |
314 | struct bio *nxt = bio->bi_next; | |
315 | ||
316 | bio->bi_next = NULL; | |
317 | bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false); | |
318 | bio->bi_next = nxt; | |
319 | } | |
320 | ||
321 | bio_set_flag(bio, BIO_SEG_VALID); | |
322 | } | |
323 | EXPORT_SYMBOL(blk_recount_segments); | |
324 | ||
325 | static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio, | |
326 | struct bio *nxt) | |
327 | { | |
328 | struct bio_vec end_bv = { NULL }, nxt_bv; | |
329 | ||
330 | if (!blk_queue_cluster(q)) | |
331 | return 0; | |
332 | ||
333 | if (bio->bi_seg_back_size + nxt->bi_seg_front_size > | |
334 | queue_max_segment_size(q)) | |
335 | return 0; | |
336 | ||
337 | if (!bio_has_data(bio)) | |
338 | return 1; | |
339 | ||
340 | bio_get_last_bvec(bio, &end_bv); | |
341 | bio_get_first_bvec(nxt, &nxt_bv); | |
342 | ||
343 | if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv)) | |
344 | return 0; | |
345 | ||
346 | /* | |
347 | * bio and nxt are contiguous in memory; check if the queue allows | |
348 | * these two to be merged into one | |
349 | */ | |
350 | if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv)) | |
351 | return 1; | |
352 | ||
353 | return 0; | |
354 | } | |
355 | ||
356 | static inline void | |
357 | __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec, | |
358 | struct scatterlist *sglist, struct bio_vec *bvprv, | |
359 | struct scatterlist **sg, int *nsegs, int *cluster) | |
360 | { | |
361 | ||
362 | int nbytes = bvec->bv_len; | |
363 | ||
364 | if (*sg && *cluster) { | |
365 | if ((*sg)->length + nbytes > queue_max_segment_size(q)) | |
366 | goto new_segment; | |
367 | ||
368 | if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec)) | |
369 | goto new_segment; | |
370 | if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec)) | |
371 | goto new_segment; | |
372 | ||
373 | (*sg)->length += nbytes; | |
374 | } else { | |
375 | new_segment: | |
376 | if (!*sg) | |
377 | *sg = sglist; | |
378 | else { | |
379 | /* | |
380 | * If the driver previously mapped a shorter | |
381 | * list, we could see a termination bit | |
382 | * prematurely unless it fully inits the sg | |
383 | * table on each mapping. We KNOW that there | |
384 | * must be more entries here or the driver | |
385 | * would be buggy, so force clear the | |
386 | * termination bit to avoid doing a full | |
387 | * sg_init_table() in drivers for each command. | |
388 | */ | |
389 | sg_unmark_end(*sg); | |
390 | *sg = sg_next(*sg); | |
391 | } | |
392 | ||
393 | sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset); | |
394 | (*nsegs)++; | |
395 | } | |
396 | *bvprv = *bvec; | |
397 | } | |
398 | ||
399 | static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv, | |
400 | struct scatterlist *sglist, struct scatterlist **sg) | |
401 | { | |
402 | *sg = sglist; | |
403 | sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); | |
404 | return 1; | |
405 | } | |
406 | ||
407 | static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, | |
408 | struct scatterlist *sglist, | |
409 | struct scatterlist **sg) | |
410 | { | |
411 | struct bio_vec bvec, bvprv = { NULL }; | |
412 | struct bvec_iter iter; | |
413 | int cluster = blk_queue_cluster(q), nsegs = 0; | |
414 | ||
415 | for_each_bio(bio) | |
416 | bio_for_each_segment(bvec, bio, iter) | |
417 | __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg, | |
418 | &nsegs, &cluster); | |
419 | ||
420 | return nsegs; | |
421 | } | |
422 | ||
423 | /* | |
424 | * map a request to scatterlist, return number of sg entries setup. Caller | |
425 | * must make sure sg can hold rq->nr_phys_segments entries | |
426 | */ | |
427 | int blk_rq_map_sg(struct request_queue *q, struct request *rq, | |
428 | struct scatterlist *sglist) | |
429 | { | |
430 | struct scatterlist *sg = NULL; | |
431 | int nsegs = 0; | |
432 | ||
433 | if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) | |
434 | nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg); | |
435 | else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME) | |
436 | nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg); | |
437 | else if (rq->bio) | |
438 | nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg); | |
439 | ||
440 | if (unlikely(rq->rq_flags & RQF_COPY_USER) && | |
441 | (blk_rq_bytes(rq) & q->dma_pad_mask)) { | |
442 | unsigned int pad_len = | |
443 | (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1; | |
444 | ||
445 | sg->length += pad_len; | |
446 | rq->extra_len += pad_len; | |
447 | } | |
448 | ||
449 | if (q->dma_drain_size && q->dma_drain_needed(rq)) { | |
450 | if (op_is_write(req_op(rq))) | |
451 | memset(q->dma_drain_buffer, 0, q->dma_drain_size); | |
452 | ||
453 | sg_unmark_end(sg); | |
454 | sg = sg_next(sg); | |
455 | sg_set_page(sg, virt_to_page(q->dma_drain_buffer), | |
456 | q->dma_drain_size, | |
457 | ((unsigned long)q->dma_drain_buffer) & | |
458 | (PAGE_SIZE - 1)); | |
459 | nsegs++; | |
460 | rq->extra_len += q->dma_drain_size; | |
461 | } | |
462 | ||
463 | if (sg) | |
464 | sg_mark_end(sg); | |
465 | ||
466 | /* | |
467 | * Something must have been wrong if the figured number of | |
468 | * segment is bigger than number of req's physical segments | |
469 | */ | |
470 | WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); | |
471 | ||
472 | return nsegs; | |
473 | } | |
474 | EXPORT_SYMBOL(blk_rq_map_sg); | |
475 | ||
476 | static inline int ll_new_hw_segment(struct request_queue *q, | |
477 | struct request *req, | |
478 | struct bio *bio) | |
479 | { | |
480 | int nr_phys_segs = bio_phys_segments(q, bio); | |
481 | ||
482 | if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q)) | |
483 | goto no_merge; | |
484 | ||
485 | if (blk_integrity_merge_bio(q, req, bio) == false) | |
486 | goto no_merge; | |
487 | ||
488 | /* | |
489 | * This will form the start of a new hw segment. Bump both | |
490 | * counters. | |
491 | */ | |
492 | req->nr_phys_segments += nr_phys_segs; | |
493 | return 1; | |
494 | ||
495 | no_merge: | |
496 | req_set_nomerge(q, req); | |
497 | return 0; | |
498 | } | |
499 | ||
500 | int ll_back_merge_fn(struct request_queue *q, struct request *req, | |
501 | struct bio *bio) | |
502 | { | |
503 | if (req_gap_back_merge(req, bio)) | |
504 | return 0; | |
505 | if (blk_integrity_rq(req) && | |
506 | integrity_req_gap_back_merge(req, bio)) | |
507 | return 0; | |
508 | if (blk_rq_sectors(req) + bio_sectors(bio) > | |
509 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) { | |
510 | req_set_nomerge(q, req); | |
511 | return 0; | |
512 | } | |
513 | if (!bio_flagged(req->biotail, BIO_SEG_VALID)) | |
514 | blk_recount_segments(q, req->biotail); | |
515 | if (!bio_flagged(bio, BIO_SEG_VALID)) | |
516 | blk_recount_segments(q, bio); | |
517 | ||
518 | return ll_new_hw_segment(q, req, bio); | |
519 | } | |
520 | ||
521 | int ll_front_merge_fn(struct request_queue *q, struct request *req, | |
522 | struct bio *bio) | |
523 | { | |
524 | ||
525 | if (req_gap_front_merge(req, bio)) | |
526 | return 0; | |
527 | if (blk_integrity_rq(req) && | |
528 | integrity_req_gap_front_merge(req, bio)) | |
529 | return 0; | |
530 | if (blk_rq_sectors(req) + bio_sectors(bio) > | |
531 | blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { | |
532 | req_set_nomerge(q, req); | |
533 | return 0; | |
534 | } | |
535 | if (!bio_flagged(bio, BIO_SEG_VALID)) | |
536 | blk_recount_segments(q, bio); | |
537 | if (!bio_flagged(req->bio, BIO_SEG_VALID)) | |
538 | blk_recount_segments(q, req->bio); | |
539 | ||
540 | return ll_new_hw_segment(q, req, bio); | |
541 | } | |
542 | ||
543 | /* | |
544 | * blk-mq uses req->special to carry normal driver per-request payload, it | |
545 | * does not indicate a prepared command that we cannot merge with. | |
546 | */ | |
547 | static bool req_no_special_merge(struct request *req) | |
548 | { | |
549 | struct request_queue *q = req->q; | |
550 | ||
551 | return !q->mq_ops && req->special; | |
552 | } | |
553 | ||
554 | static bool req_attempt_discard_merge(struct request_queue *q, struct request *req, | |
555 | struct request *next) | |
556 | { | |
557 | unsigned short segments = blk_rq_nr_discard_segments(req); | |
558 | ||
559 | if (segments >= queue_max_discard_segments(q)) | |
560 | goto no_merge; | |
561 | if (blk_rq_sectors(req) + bio_sectors(next->bio) > | |
562 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) | |
563 | goto no_merge; | |
564 | ||
565 | req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next); | |
566 | return true; | |
567 | no_merge: | |
568 | req_set_nomerge(q, req); | |
569 | return false; | |
570 | } | |
571 | ||
572 | static int ll_merge_requests_fn(struct request_queue *q, struct request *req, | |
573 | struct request *next) | |
574 | { | |
575 | int total_phys_segments; | |
576 | unsigned int seg_size = | |
577 | req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size; | |
578 | ||
579 | /* | |
580 | * First check if the either of the requests are re-queued | |
581 | * requests. Can't merge them if they are. | |
582 | */ | |
583 | if (req_no_special_merge(req) || req_no_special_merge(next)) | |
584 | return 0; | |
585 | ||
586 | if (req_gap_back_merge(req, next->bio)) | |
587 | return 0; | |
588 | ||
589 | /* | |
590 | * Will it become too large? | |
591 | */ | |
592 | if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > | |
593 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) | |
594 | return 0; | |
595 | ||
596 | total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; | |
597 | if (blk_phys_contig_segment(q, req->biotail, next->bio)) { | |
598 | if (req->nr_phys_segments == 1) | |
599 | req->bio->bi_seg_front_size = seg_size; | |
600 | if (next->nr_phys_segments == 1) | |
601 | next->biotail->bi_seg_back_size = seg_size; | |
602 | total_phys_segments--; | |
603 | } | |
604 | ||
605 | if (total_phys_segments > queue_max_segments(q)) | |
606 | return 0; | |
607 | ||
608 | if (blk_integrity_merge_rq(q, req, next) == false) | |
609 | return 0; | |
610 | ||
611 | /* Merge is OK... */ | |
612 | req->nr_phys_segments = total_phys_segments; | |
613 | return 1; | |
614 | } | |
615 | ||
616 | /** | |
617 | * blk_rq_set_mixed_merge - mark a request as mixed merge | |
618 | * @rq: request to mark as mixed merge | |
619 | * | |
620 | * Description: | |
621 | * @rq is about to be mixed merged. Make sure the attributes | |
622 | * which can be mixed are set in each bio and mark @rq as mixed | |
623 | * merged. | |
624 | */ | |
625 | void blk_rq_set_mixed_merge(struct request *rq) | |
626 | { | |
627 | unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; | |
628 | struct bio *bio; | |
629 | ||
630 | if (rq->rq_flags & RQF_MIXED_MERGE) | |
631 | return; | |
632 | ||
633 | /* | |
634 | * @rq will no longer represent mixable attributes for all the | |
635 | * contained bios. It will just track those of the first one. | |
636 | * Distributes the attributs to each bio. | |
637 | */ | |
638 | for (bio = rq->bio; bio; bio = bio->bi_next) { | |
639 | WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && | |
640 | (bio->bi_opf & REQ_FAILFAST_MASK) != ff); | |
641 | bio->bi_opf |= ff; | |
642 | } | |
643 | rq->rq_flags |= RQF_MIXED_MERGE; | |
644 | } | |
645 | ||
646 | static void blk_account_io_merge(struct request *req) | |
647 | { | |
648 | if (blk_do_io_stat(req)) { | |
649 | struct hd_struct *part; | |
650 | int cpu; | |
651 | ||
652 | cpu = part_stat_lock(); | |
653 | part = req->part; | |
654 | ||
655 | part_round_stats(req->q, cpu, part); | |
656 | part_dec_in_flight(req->q, part, rq_data_dir(req)); | |
657 | ||
658 | hd_struct_put(part); | |
659 | part_stat_unlock(); | |
660 | } | |
661 | } | |
662 | ||
663 | /* | |
664 | * For non-mq, this has to be called with the request spinlock acquired. | |
665 | * For mq with scheduling, the appropriate queue wide lock should be held. | |
666 | */ | |
667 | static struct request *attempt_merge(struct request_queue *q, | |
668 | struct request *req, struct request *next) | |
669 | { | |
670 | if (!q->mq_ops) | |
671 | lockdep_assert_held(q->queue_lock); | |
672 | ||
673 | if (!rq_mergeable(req) || !rq_mergeable(next)) | |
674 | return NULL; | |
675 | ||
676 | if (req_op(req) != req_op(next)) | |
677 | return NULL; | |
678 | ||
679 | /* | |
680 | * not contiguous | |
681 | */ | |
682 | if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next)) | |
683 | return NULL; | |
684 | ||
685 | if (rq_data_dir(req) != rq_data_dir(next) | |
686 | || req->rq_disk != next->rq_disk | |
687 | || req_no_special_merge(next)) | |
688 | return NULL; | |
689 | ||
690 | if (req_op(req) == REQ_OP_WRITE_SAME && | |
691 | !blk_write_same_mergeable(req->bio, next->bio)) | |
692 | return NULL; | |
693 | ||
694 | /* | |
695 | * Don't allow merge of different write hints, or for a hint with | |
696 | * non-hint IO. | |
697 | */ | |
698 | if (req->write_hint != next->write_hint) | |
699 | return NULL; | |
700 | ||
701 | /* | |
702 | * If we are allowed to merge, then append bio list | |
703 | * from next to rq and release next. merge_requests_fn | |
704 | * will have updated segment counts, update sector | |
705 | * counts here. Handle DISCARDs separately, as they | |
706 | * have separate settings. | |
707 | */ | |
708 | if (req_op(req) == REQ_OP_DISCARD) { | |
709 | if (!req_attempt_discard_merge(q, req, next)) | |
710 | return NULL; | |
711 | } else if (!ll_merge_requests_fn(q, req, next)) | |
712 | return NULL; | |
713 | ||
714 | /* | |
715 | * If failfast settings disagree or any of the two is already | |
716 | * a mixed merge, mark both as mixed before proceeding. This | |
717 | * makes sure that all involved bios have mixable attributes | |
718 | * set properly. | |
719 | */ | |
720 | if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || | |
721 | (req->cmd_flags & REQ_FAILFAST_MASK) != | |
722 | (next->cmd_flags & REQ_FAILFAST_MASK)) { | |
723 | blk_rq_set_mixed_merge(req); | |
724 | blk_rq_set_mixed_merge(next); | |
725 | } | |
726 | ||
727 | /* | |
728 | * At this point we have either done a back merge | |
729 | * or front merge. We need the smaller start_time of | |
730 | * the merged requests to be the current request | |
731 | * for accounting purposes. | |
732 | */ | |
733 | if (time_after(req->start_time, next->start_time)) | |
734 | req->start_time = next->start_time; | |
735 | ||
736 | req->biotail->bi_next = next->bio; | |
737 | req->biotail = next->biotail; | |
738 | ||
739 | req->__data_len += blk_rq_bytes(next); | |
740 | ||
741 | if (req_op(req) != REQ_OP_DISCARD) | |
742 | elv_merge_requests(q, req, next); | |
743 | ||
744 | /* | |
745 | * 'next' is going away, so update stats accordingly | |
746 | */ | |
747 | blk_account_io_merge(next); | |
748 | ||
749 | req->ioprio = ioprio_best(req->ioprio, next->ioprio); | |
750 | if (blk_rq_cpu_valid(next)) | |
751 | req->cpu = next->cpu; | |
752 | ||
753 | /* | |
754 | * ownership of bio passed from next to req, return 'next' for | |
755 | * the caller to free | |
756 | */ | |
757 | next->bio = NULL; | |
758 | return next; | |
759 | } | |
760 | ||
761 | struct request *attempt_back_merge(struct request_queue *q, struct request *rq) | |
762 | { | |
763 | struct request *next = elv_latter_request(q, rq); | |
764 | ||
765 | if (next) | |
766 | return attempt_merge(q, rq, next); | |
767 | ||
768 | return NULL; | |
769 | } | |
770 | ||
771 | struct request *attempt_front_merge(struct request_queue *q, struct request *rq) | |
772 | { | |
773 | struct request *prev = elv_former_request(q, rq); | |
774 | ||
775 | if (prev) | |
776 | return attempt_merge(q, prev, rq); | |
777 | ||
778 | return NULL; | |
779 | } | |
780 | ||
781 | int blk_attempt_req_merge(struct request_queue *q, struct request *rq, | |
782 | struct request *next) | |
783 | { | |
784 | struct elevator_queue *e = q->elevator; | |
785 | struct request *free; | |
786 | ||
787 | if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn) | |
788 | if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next)) | |
789 | return 0; | |
790 | ||
791 | free = attempt_merge(q, rq, next); | |
792 | if (free) { | |
793 | __blk_put_request(q, free); | |
794 | return 1; | |
795 | } | |
796 | ||
797 | return 0; | |
798 | } | |
799 | ||
800 | bool blk_rq_merge_ok(struct request *rq, struct bio *bio) | |
801 | { | |
802 | if (!rq_mergeable(rq) || !bio_mergeable(bio)) | |
803 | return false; | |
804 | ||
805 | if (req_op(rq) != bio_op(bio)) | |
806 | return false; | |
807 | ||
808 | /* different data direction or already started, don't merge */ | |
809 | if (bio_data_dir(bio) != rq_data_dir(rq)) | |
810 | return false; | |
811 | ||
812 | /* must be same device and not a special request */ | |
813 | if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq)) | |
814 | return false; | |
815 | ||
816 | /* only merge integrity protected bio into ditto rq */ | |
817 | if (blk_integrity_merge_bio(rq->q, rq, bio) == false) | |
818 | return false; | |
819 | ||
820 | /* must be using the same buffer */ | |
821 | if (req_op(rq) == REQ_OP_WRITE_SAME && | |
822 | !blk_write_same_mergeable(rq->bio, bio)) | |
823 | return false; | |
824 | ||
825 | /* | |
826 | * Don't allow merge of different write hints, or for a hint with | |
827 | * non-hint IO. | |
828 | */ | |
829 | if (rq->write_hint != bio->bi_write_hint) | |
830 | return false; | |
831 | ||
832 | return true; | |
833 | } | |
834 | ||
835 | enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) | |
836 | { | |
837 | if (req_op(rq) == REQ_OP_DISCARD && | |
838 | queue_max_discard_segments(rq->q) > 1) | |
839 | return ELEVATOR_DISCARD_MERGE; | |
840 | else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) | |
841 | return ELEVATOR_BACK_MERGE; | |
842 | else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) | |
843 | return ELEVATOR_FRONT_MERGE; | |
844 | return ELEVATOR_NO_MERGE; | |
845 | } |