]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Functions related to segment and merge handling | |
3 | */ | |
4 | #include <linux/kernel.h> | |
5 | #include <linux/module.h> | |
6 | #include <linux/bio.h> | |
7 | #include <linux/blkdev.h> | |
8 | #include <linux/scatterlist.h> | |
9 | ||
10 | #include <trace/events/block.h> | |
11 | ||
12 | #include "blk.h" | |
13 | ||
14 | static struct bio *blk_bio_discard_split(struct request_queue *q, | |
15 | struct bio *bio, | |
16 | struct bio_set *bs, | |
17 | unsigned *nsegs) | |
18 | { | |
19 | unsigned int max_discard_sectors, granularity; | |
20 | int alignment; | |
21 | sector_t tmp; | |
22 | unsigned split_sectors; | |
23 | ||
24 | *nsegs = 1; | |
25 | ||
26 | /* Zero-sector (unknown) and one-sector granularities are the same. */ | |
27 | granularity = max(q->limits.discard_granularity >> 9, 1U); | |
28 | ||
29 | max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9); | |
30 | max_discard_sectors -= max_discard_sectors % granularity; | |
31 | ||
32 | if (unlikely(!max_discard_sectors)) { | |
33 | /* XXX: warn */ | |
34 | return NULL; | |
35 | } | |
36 | ||
37 | if (bio_sectors(bio) <= max_discard_sectors) | |
38 | return NULL; | |
39 | ||
40 | split_sectors = max_discard_sectors; | |
41 | ||
42 | /* | |
43 | * If the next starting sector would be misaligned, stop the discard at | |
44 | * the previous aligned sector. | |
45 | */ | |
46 | alignment = (q->limits.discard_alignment >> 9) % granularity; | |
47 | ||
48 | tmp = bio->bi_iter.bi_sector + split_sectors - alignment; | |
49 | tmp = sector_div(tmp, granularity); | |
50 | ||
51 | if (split_sectors > tmp) | |
52 | split_sectors -= tmp; | |
53 | ||
54 | return bio_split(bio, split_sectors, GFP_NOIO, bs); | |
55 | } | |
56 | ||
57 | static struct bio *blk_bio_write_same_split(struct request_queue *q, | |
58 | struct bio *bio, | |
59 | struct bio_set *bs, | |
60 | unsigned *nsegs) | |
61 | { | |
62 | *nsegs = 1; | |
63 | ||
64 | if (!q->limits.max_write_same_sectors) | |
65 | return NULL; | |
66 | ||
67 | if (bio_sectors(bio) <= q->limits.max_write_same_sectors) | |
68 | return NULL; | |
69 | ||
70 | return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); | |
71 | } | |
72 | ||
73 | static inline unsigned get_max_io_size(struct request_queue *q, | |
74 | struct bio *bio) | |
75 | { | |
76 | unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector); | |
77 | unsigned mask = queue_logical_block_size(q) - 1; | |
78 | ||
79 | /* aligned to logical block size */ | |
80 | sectors &= ~(mask >> 9); | |
81 | ||
82 | return sectors; | |
83 | } | |
84 | ||
85 | static struct bio *blk_bio_segment_split(struct request_queue *q, | |
86 | struct bio *bio, | |
87 | struct bio_set *bs, | |
88 | unsigned *segs) | |
89 | { | |
90 | struct bio_vec bv, bvprv, *bvprvp = NULL; | |
91 | struct bvec_iter iter; | |
92 | unsigned seg_size = 0, nsegs = 0, sectors = 0; | |
93 | unsigned front_seg_size = bio->bi_seg_front_size; | |
94 | bool do_split = true; | |
95 | struct bio *new = NULL; | |
96 | const unsigned max_sectors = get_max_io_size(q, bio); | |
97 | unsigned bvecs = 0; | |
98 | ||
99 | bio_for_each_segment(bv, bio, iter) { | |
100 | /* | |
101 | * With arbitrary bio size, the incoming bio may be very | |
102 | * big. We have to split the bio into small bios so that | |
103 | * each holds at most BIO_MAX_PAGES bvecs because | |
104 | * bio_clone() can fail to allocate big bvecs. | |
105 | * | |
106 | * It should have been better to apply the limit per | |
107 | * request queue in which bio_clone() is involved, | |
108 | * instead of globally. The biggest blocker is the | |
109 | * bio_clone() in bio bounce. | |
110 | * | |
111 | * If bio is splitted by this reason, we should have | |
112 | * allowed to continue bios merging, but don't do | |
113 | * that now for making the change simple. | |
114 | * | |
115 | * TODO: deal with bio bounce's bio_clone() gracefully | |
116 | * and convert the global limit into per-queue limit. | |
117 | */ | |
118 | if (bvecs++ >= BIO_MAX_PAGES) | |
119 | goto split; | |
120 | ||
121 | /* | |
122 | * If the queue doesn't support SG gaps and adding this | |
123 | * offset would create a gap, disallow it. | |
124 | */ | |
125 | if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) | |
126 | goto split; | |
127 | ||
128 | if (sectors + (bv.bv_len >> 9) > max_sectors) { | |
129 | /* | |
130 | * Consider this a new segment if we're splitting in | |
131 | * the middle of this vector. | |
132 | */ | |
133 | if (nsegs < queue_max_segments(q) && | |
134 | sectors < max_sectors) { | |
135 | nsegs++; | |
136 | sectors = max_sectors; | |
137 | } | |
138 | if (sectors) | |
139 | goto split; | |
140 | /* Make this single bvec as the 1st segment */ | |
141 | } | |
142 | ||
143 | if (bvprvp && blk_queue_cluster(q)) { | |
144 | if (seg_size + bv.bv_len > queue_max_segment_size(q)) | |
145 | goto new_segment; | |
146 | if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv)) | |
147 | goto new_segment; | |
148 | if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv)) | |
149 | goto new_segment; | |
150 | ||
151 | seg_size += bv.bv_len; | |
152 | bvprv = bv; | |
153 | bvprvp = &bvprv; | |
154 | sectors += bv.bv_len >> 9; | |
155 | ||
156 | if (nsegs == 1 && seg_size > front_seg_size) | |
157 | front_seg_size = seg_size; | |
158 | continue; | |
159 | } | |
160 | new_segment: | |
161 | if (nsegs == queue_max_segments(q)) | |
162 | goto split; | |
163 | ||
164 | nsegs++; | |
165 | bvprv = bv; | |
166 | bvprvp = &bvprv; | |
167 | seg_size = bv.bv_len; | |
168 | sectors += bv.bv_len >> 9; | |
169 | ||
170 | if (nsegs == 1 && seg_size > front_seg_size) | |
171 | front_seg_size = seg_size; | |
172 | } | |
173 | ||
174 | do_split = false; | |
175 | split: | |
176 | *segs = nsegs; | |
177 | ||
178 | if (do_split) { | |
179 | new = bio_split(bio, sectors, GFP_NOIO, bs); | |
180 | if (new) | |
181 | bio = new; | |
182 | } | |
183 | ||
184 | bio->bi_seg_front_size = front_seg_size; | |
185 | if (seg_size > bio->bi_seg_back_size) | |
186 | bio->bi_seg_back_size = seg_size; | |
187 | ||
188 | return do_split ? new : NULL; | |
189 | } | |
190 | ||
191 | void blk_queue_split(struct request_queue *q, struct bio **bio, | |
192 | struct bio_set *bs) | |
193 | { | |
194 | struct bio *split, *res; | |
195 | unsigned nsegs; | |
196 | ||
197 | switch (bio_op(*bio)) { | |
198 | case REQ_OP_DISCARD: | |
199 | case REQ_OP_SECURE_ERASE: | |
200 | split = blk_bio_discard_split(q, *bio, bs, &nsegs); | |
201 | break; | |
202 | case REQ_OP_WRITE_ZEROES: | |
203 | split = NULL; | |
204 | nsegs = (*bio)->bi_phys_segments; | |
205 | break; | |
206 | case REQ_OP_WRITE_SAME: | |
207 | split = blk_bio_write_same_split(q, *bio, bs, &nsegs); | |
208 | break; | |
209 | default: | |
210 | split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs); | |
211 | break; | |
212 | } | |
213 | ||
214 | /* physical segments can be figured out during splitting */ | |
215 | res = split ? split : *bio; | |
216 | res->bi_phys_segments = nsegs; | |
217 | bio_set_flag(res, BIO_SEG_VALID); | |
218 | ||
219 | if (split) { | |
220 | /* there isn't chance to merge the splitted bio */ | |
221 | split->bi_opf |= REQ_NOMERGE; | |
222 | ||
223 | bio_chain(split, *bio); | |
224 | trace_block_split(q, split, (*bio)->bi_iter.bi_sector); | |
225 | generic_make_request(*bio); | |
226 | *bio = split; | |
227 | } | |
228 | } | |
229 | EXPORT_SYMBOL(blk_queue_split); | |
230 | ||
231 | static unsigned int __blk_recalc_rq_segments(struct request_queue *q, | |
232 | struct bio *bio, | |
233 | bool no_sg_merge) | |
234 | { | |
235 | struct bio_vec bv, bvprv = { NULL }; | |
236 | int cluster, prev = 0; | |
237 | unsigned int seg_size, nr_phys_segs; | |
238 | struct bio *fbio, *bbio; | |
239 | struct bvec_iter iter; | |
240 | ||
241 | if (!bio) | |
242 | return 0; | |
243 | ||
244 | switch (bio_op(bio)) { | |
245 | case REQ_OP_DISCARD: | |
246 | case REQ_OP_SECURE_ERASE: | |
247 | case REQ_OP_WRITE_ZEROES: | |
248 | return 0; | |
249 | case REQ_OP_WRITE_SAME: | |
250 | return 1; | |
251 | } | |
252 | ||
253 | fbio = bio; | |
254 | cluster = blk_queue_cluster(q); | |
255 | seg_size = 0; | |
256 | nr_phys_segs = 0; | |
257 | for_each_bio(bio) { | |
258 | bio_for_each_segment(bv, bio, iter) { | |
259 | /* | |
260 | * If SG merging is disabled, each bio vector is | |
261 | * a segment | |
262 | */ | |
263 | if (no_sg_merge) | |
264 | goto new_segment; | |
265 | ||
266 | if (prev && cluster) { | |
267 | if (seg_size + bv.bv_len | |
268 | > queue_max_segment_size(q)) | |
269 | goto new_segment; | |
270 | if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv)) | |
271 | goto new_segment; | |
272 | if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv)) | |
273 | goto new_segment; | |
274 | ||
275 | seg_size += bv.bv_len; | |
276 | bvprv = bv; | |
277 | continue; | |
278 | } | |
279 | new_segment: | |
280 | if (nr_phys_segs == 1 && seg_size > | |
281 | fbio->bi_seg_front_size) | |
282 | fbio->bi_seg_front_size = seg_size; | |
283 | ||
284 | nr_phys_segs++; | |
285 | bvprv = bv; | |
286 | prev = 1; | |
287 | seg_size = bv.bv_len; | |
288 | } | |
289 | bbio = bio; | |
290 | } | |
291 | ||
292 | if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size) | |
293 | fbio->bi_seg_front_size = seg_size; | |
294 | if (seg_size > bbio->bi_seg_back_size) | |
295 | bbio->bi_seg_back_size = seg_size; | |
296 | ||
297 | return nr_phys_segs; | |
298 | } | |
299 | ||
300 | void blk_recalc_rq_segments(struct request *rq) | |
301 | { | |
302 | bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE, | |
303 | &rq->q->queue_flags); | |
304 | ||
305 | rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio, | |
306 | no_sg_merge); | |
307 | } | |
308 | ||
309 | void blk_recount_segments(struct request_queue *q, struct bio *bio) | |
310 | { | |
311 | unsigned short seg_cnt; | |
312 | ||
313 | /* estimate segment number by bi_vcnt for non-cloned bio */ | |
314 | if (bio_flagged(bio, BIO_CLONED)) | |
315 | seg_cnt = bio_segments(bio); | |
316 | else | |
317 | seg_cnt = bio->bi_vcnt; | |
318 | ||
319 | if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) && | |
320 | (seg_cnt < queue_max_segments(q))) | |
321 | bio->bi_phys_segments = seg_cnt; | |
322 | else { | |
323 | struct bio *nxt = bio->bi_next; | |
324 | ||
325 | bio->bi_next = NULL; | |
326 | bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false); | |
327 | bio->bi_next = nxt; | |
328 | } | |
329 | ||
330 | bio_set_flag(bio, BIO_SEG_VALID); | |
331 | } | |
332 | EXPORT_SYMBOL(blk_recount_segments); | |
333 | ||
334 | static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio, | |
335 | struct bio *nxt) | |
336 | { | |
337 | struct bio_vec end_bv = { NULL }, nxt_bv; | |
338 | ||
339 | if (!blk_queue_cluster(q)) | |
340 | return 0; | |
341 | ||
342 | if (bio->bi_seg_back_size + nxt->bi_seg_front_size > | |
343 | queue_max_segment_size(q)) | |
344 | return 0; | |
345 | ||
346 | if (!bio_has_data(bio)) | |
347 | return 1; | |
348 | ||
349 | bio_get_last_bvec(bio, &end_bv); | |
350 | bio_get_first_bvec(nxt, &nxt_bv); | |
351 | ||
352 | if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv)) | |
353 | return 0; | |
354 | ||
355 | /* | |
356 | * bio and nxt are contiguous in memory; check if the queue allows | |
357 | * these two to be merged into one | |
358 | */ | |
359 | if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv)) | |
360 | return 1; | |
361 | ||
362 | return 0; | |
363 | } | |
364 | ||
365 | static inline void | |
366 | __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec, | |
367 | struct scatterlist *sglist, struct bio_vec *bvprv, | |
368 | struct scatterlist **sg, int *nsegs, int *cluster) | |
369 | { | |
370 | ||
371 | int nbytes = bvec->bv_len; | |
372 | ||
373 | if (*sg && *cluster) { | |
374 | if ((*sg)->length + nbytes > queue_max_segment_size(q)) | |
375 | goto new_segment; | |
376 | ||
377 | if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec)) | |
378 | goto new_segment; | |
379 | if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec)) | |
380 | goto new_segment; | |
381 | ||
382 | (*sg)->length += nbytes; | |
383 | } else { | |
384 | new_segment: | |
385 | if (!*sg) | |
386 | *sg = sglist; | |
387 | else { | |
388 | /* | |
389 | * If the driver previously mapped a shorter | |
390 | * list, we could see a termination bit | |
391 | * prematurely unless it fully inits the sg | |
392 | * table on each mapping. We KNOW that there | |
393 | * must be more entries here or the driver | |
394 | * would be buggy, so force clear the | |
395 | * termination bit to avoid doing a full | |
396 | * sg_init_table() in drivers for each command. | |
397 | */ | |
398 | sg_unmark_end(*sg); | |
399 | *sg = sg_next(*sg); | |
400 | } | |
401 | ||
402 | sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset); | |
403 | (*nsegs)++; | |
404 | } | |
405 | *bvprv = *bvec; | |
406 | } | |
407 | ||
408 | static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv, | |
409 | struct scatterlist *sglist, struct scatterlist **sg) | |
410 | { | |
411 | *sg = sglist; | |
412 | sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); | |
413 | return 1; | |
414 | } | |
415 | ||
416 | static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, | |
417 | struct scatterlist *sglist, | |
418 | struct scatterlist **sg) | |
419 | { | |
420 | struct bio_vec bvec, bvprv = { NULL }; | |
421 | struct bvec_iter iter; | |
422 | int cluster = blk_queue_cluster(q), nsegs = 0; | |
423 | ||
424 | for_each_bio(bio) | |
425 | bio_for_each_segment(bvec, bio, iter) | |
426 | __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg, | |
427 | &nsegs, &cluster); | |
428 | ||
429 | return nsegs; | |
430 | } | |
431 | ||
432 | /* | |
433 | * map a request to scatterlist, return number of sg entries setup. Caller | |
434 | * must make sure sg can hold rq->nr_phys_segments entries | |
435 | */ | |
436 | int blk_rq_map_sg(struct request_queue *q, struct request *rq, | |
437 | struct scatterlist *sglist) | |
438 | { | |
439 | struct scatterlist *sg = NULL; | |
440 | int nsegs = 0; | |
441 | ||
442 | if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) | |
443 | nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg); | |
444 | else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME) | |
445 | nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg); | |
446 | else if (rq->bio) | |
447 | nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg); | |
448 | ||
449 | if (unlikely(rq->rq_flags & RQF_COPY_USER) && | |
450 | (blk_rq_bytes(rq) & q->dma_pad_mask)) { | |
451 | unsigned int pad_len = | |
452 | (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1; | |
453 | ||
454 | sg->length += pad_len; | |
455 | rq->extra_len += pad_len; | |
456 | } | |
457 | ||
458 | if (q->dma_drain_size && q->dma_drain_needed(rq)) { | |
459 | if (op_is_write(req_op(rq))) | |
460 | memset(q->dma_drain_buffer, 0, q->dma_drain_size); | |
461 | ||
462 | sg_unmark_end(sg); | |
463 | sg = sg_next(sg); | |
464 | sg_set_page(sg, virt_to_page(q->dma_drain_buffer), | |
465 | q->dma_drain_size, | |
466 | ((unsigned long)q->dma_drain_buffer) & | |
467 | (PAGE_SIZE - 1)); | |
468 | nsegs++; | |
469 | rq->extra_len += q->dma_drain_size; | |
470 | } | |
471 | ||
472 | if (sg) | |
473 | sg_mark_end(sg); | |
474 | ||
475 | /* | |
476 | * Something must have been wrong if the figured number of | |
477 | * segment is bigger than number of req's physical segments | |
478 | */ | |
479 | WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); | |
480 | ||
481 | return nsegs; | |
482 | } | |
483 | EXPORT_SYMBOL(blk_rq_map_sg); | |
484 | ||
485 | static inline int ll_new_hw_segment(struct request_queue *q, | |
486 | struct request *req, | |
487 | struct bio *bio) | |
488 | { | |
489 | int nr_phys_segs = bio_phys_segments(q, bio); | |
490 | ||
491 | if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q)) | |
492 | goto no_merge; | |
493 | ||
494 | if (blk_integrity_merge_bio(q, req, bio) == false) | |
495 | goto no_merge; | |
496 | ||
497 | /* | |
498 | * This will form the start of a new hw segment. Bump both | |
499 | * counters. | |
500 | */ | |
501 | req->nr_phys_segments += nr_phys_segs; | |
502 | return 1; | |
503 | ||
504 | no_merge: | |
505 | req_set_nomerge(q, req); | |
506 | return 0; | |
507 | } | |
508 | ||
509 | int ll_back_merge_fn(struct request_queue *q, struct request *req, | |
510 | struct bio *bio) | |
511 | { | |
512 | if (req_gap_back_merge(req, bio)) | |
513 | return 0; | |
514 | if (blk_integrity_rq(req) && | |
515 | integrity_req_gap_back_merge(req, bio)) | |
516 | return 0; | |
517 | if (blk_rq_sectors(req) + bio_sectors(bio) > | |
518 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) { | |
519 | req_set_nomerge(q, req); | |
520 | return 0; | |
521 | } | |
522 | if (!bio_flagged(req->biotail, BIO_SEG_VALID)) | |
523 | blk_recount_segments(q, req->biotail); | |
524 | if (!bio_flagged(bio, BIO_SEG_VALID)) | |
525 | blk_recount_segments(q, bio); | |
526 | ||
527 | return ll_new_hw_segment(q, req, bio); | |
528 | } | |
529 | ||
530 | int ll_front_merge_fn(struct request_queue *q, struct request *req, | |
531 | struct bio *bio) | |
532 | { | |
533 | ||
534 | if (req_gap_front_merge(req, bio)) | |
535 | return 0; | |
536 | if (blk_integrity_rq(req) && | |
537 | integrity_req_gap_front_merge(req, bio)) | |
538 | return 0; | |
539 | if (blk_rq_sectors(req) + bio_sectors(bio) > | |
540 | blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { | |
541 | req_set_nomerge(q, req); | |
542 | return 0; | |
543 | } | |
544 | if (!bio_flagged(bio, BIO_SEG_VALID)) | |
545 | blk_recount_segments(q, bio); | |
546 | if (!bio_flagged(req->bio, BIO_SEG_VALID)) | |
547 | blk_recount_segments(q, req->bio); | |
548 | ||
549 | return ll_new_hw_segment(q, req, bio); | |
550 | } | |
551 | ||
552 | /* | |
553 | * blk-mq uses req->special to carry normal driver per-request payload, it | |
554 | * does not indicate a prepared command that we cannot merge with. | |
555 | */ | |
556 | static bool req_no_special_merge(struct request *req) | |
557 | { | |
558 | struct request_queue *q = req->q; | |
559 | ||
560 | return !q->mq_ops && req->special; | |
561 | } | |
562 | ||
563 | static int ll_merge_requests_fn(struct request_queue *q, struct request *req, | |
564 | struct request *next) | |
565 | { | |
566 | int total_phys_segments; | |
567 | unsigned int seg_size = | |
568 | req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size; | |
569 | ||
570 | /* | |
571 | * First check if the either of the requests are re-queued | |
572 | * requests. Can't merge them if they are. | |
573 | */ | |
574 | if (req_no_special_merge(req) || req_no_special_merge(next)) | |
575 | return 0; | |
576 | ||
577 | if (req_gap_back_merge(req, next->bio)) | |
578 | return 0; | |
579 | ||
580 | /* | |
581 | * Will it become too large? | |
582 | */ | |
583 | if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > | |
584 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) | |
585 | return 0; | |
586 | ||
587 | total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; | |
588 | if (blk_phys_contig_segment(q, req->biotail, next->bio)) { | |
589 | if (req->nr_phys_segments == 1) | |
590 | req->bio->bi_seg_front_size = seg_size; | |
591 | if (next->nr_phys_segments == 1) | |
592 | next->biotail->bi_seg_back_size = seg_size; | |
593 | total_phys_segments--; | |
594 | } | |
595 | ||
596 | if (total_phys_segments > queue_max_segments(q)) | |
597 | return 0; | |
598 | ||
599 | if (blk_integrity_merge_rq(q, req, next) == false) | |
600 | return 0; | |
601 | ||
602 | /* Merge is OK... */ | |
603 | req->nr_phys_segments = total_phys_segments; | |
604 | return 1; | |
605 | } | |
606 | ||
607 | /** | |
608 | * blk_rq_set_mixed_merge - mark a request as mixed merge | |
609 | * @rq: request to mark as mixed merge | |
610 | * | |
611 | * Description: | |
612 | * @rq is about to be mixed merged. Make sure the attributes | |
613 | * which can be mixed are set in each bio and mark @rq as mixed | |
614 | * merged. | |
615 | */ | |
616 | void blk_rq_set_mixed_merge(struct request *rq) | |
617 | { | |
618 | unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; | |
619 | struct bio *bio; | |
620 | ||
621 | if (rq->rq_flags & RQF_MIXED_MERGE) | |
622 | return; | |
623 | ||
624 | /* | |
625 | * @rq will no longer represent mixable attributes for all the | |
626 | * contained bios. It will just track those of the first one. | |
627 | * Distributes the attributs to each bio. | |
628 | */ | |
629 | for (bio = rq->bio; bio; bio = bio->bi_next) { | |
630 | WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && | |
631 | (bio->bi_opf & REQ_FAILFAST_MASK) != ff); | |
632 | bio->bi_opf |= ff; | |
633 | } | |
634 | rq->rq_flags |= RQF_MIXED_MERGE; | |
635 | } | |
636 | ||
637 | static void blk_account_io_merge(struct request *req) | |
638 | { | |
639 | if (blk_do_io_stat(req)) { | |
640 | struct hd_struct *part; | |
641 | int cpu; | |
642 | ||
643 | cpu = part_stat_lock(); | |
644 | part = req->part; | |
645 | ||
646 | part_round_stats(cpu, part); | |
647 | part_dec_in_flight(part, rq_data_dir(req)); | |
648 | ||
649 | hd_struct_put(part); | |
650 | part_stat_unlock(); | |
651 | } | |
652 | } | |
653 | ||
654 | /* | |
655 | * For non-mq, this has to be called with the request spinlock acquired. | |
656 | * For mq with scheduling, the appropriate queue wide lock should be held. | |
657 | */ | |
658 | static struct request *attempt_merge(struct request_queue *q, | |
659 | struct request *req, struct request *next) | |
660 | { | |
661 | if (!rq_mergeable(req) || !rq_mergeable(next)) | |
662 | return NULL; | |
663 | ||
664 | if (req_op(req) != req_op(next)) | |
665 | return NULL; | |
666 | ||
667 | /* | |
668 | * not contiguous | |
669 | */ | |
670 | if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next)) | |
671 | return NULL; | |
672 | ||
673 | if (rq_data_dir(req) != rq_data_dir(next) | |
674 | || req->rq_disk != next->rq_disk | |
675 | || req_no_special_merge(next)) | |
676 | return NULL; | |
677 | ||
678 | if (req_op(req) == REQ_OP_WRITE_SAME && | |
679 | !blk_write_same_mergeable(req->bio, next->bio)) | |
680 | return NULL; | |
681 | ||
682 | /* | |
683 | * If we are allowed to merge, then append bio list | |
684 | * from next to rq and release next. merge_requests_fn | |
685 | * will have updated segment counts, update sector | |
686 | * counts here. | |
687 | */ | |
688 | if (!ll_merge_requests_fn(q, req, next)) | |
689 | return NULL; | |
690 | ||
691 | /* | |
692 | * If failfast settings disagree or any of the two is already | |
693 | * a mixed merge, mark both as mixed before proceeding. This | |
694 | * makes sure that all involved bios have mixable attributes | |
695 | * set properly. | |
696 | */ | |
697 | if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) || | |
698 | (req->cmd_flags & REQ_FAILFAST_MASK) != | |
699 | (next->cmd_flags & REQ_FAILFAST_MASK)) { | |
700 | blk_rq_set_mixed_merge(req); | |
701 | blk_rq_set_mixed_merge(next); | |
702 | } | |
703 | ||
704 | /* | |
705 | * At this point we have either done a back merge | |
706 | * or front merge. We need the smaller start_time of | |
707 | * the merged requests to be the current request | |
708 | * for accounting purposes. | |
709 | */ | |
710 | if (time_after(req->start_time, next->start_time)) | |
711 | req->start_time = next->start_time; | |
712 | ||
713 | req->biotail->bi_next = next->bio; | |
714 | req->biotail = next->biotail; | |
715 | ||
716 | req->__data_len += blk_rq_bytes(next); | |
717 | ||
718 | elv_merge_requests(q, req, next); | |
719 | ||
720 | /* | |
721 | * 'next' is going away, so update stats accordingly | |
722 | */ | |
723 | blk_account_io_merge(next); | |
724 | ||
725 | req->ioprio = ioprio_best(req->ioprio, next->ioprio); | |
726 | if (blk_rq_cpu_valid(next)) | |
727 | req->cpu = next->cpu; | |
728 | ||
729 | /* | |
730 | * ownership of bio passed from next to req, return 'next' for | |
731 | * the caller to free | |
732 | */ | |
733 | next->bio = NULL; | |
734 | return next; | |
735 | } | |
736 | ||
737 | struct request *attempt_back_merge(struct request_queue *q, struct request *rq) | |
738 | { | |
739 | struct request *next = elv_latter_request(q, rq); | |
740 | ||
741 | if (next) | |
742 | return attempt_merge(q, rq, next); | |
743 | ||
744 | return NULL; | |
745 | } | |
746 | ||
747 | struct request *attempt_front_merge(struct request_queue *q, struct request *rq) | |
748 | { | |
749 | struct request *prev = elv_former_request(q, rq); | |
750 | ||
751 | if (prev) | |
752 | return attempt_merge(q, prev, rq); | |
753 | ||
754 | return NULL; | |
755 | } | |
756 | ||
757 | int blk_attempt_req_merge(struct request_queue *q, struct request *rq, | |
758 | struct request *next) | |
759 | { | |
760 | struct elevator_queue *e = q->elevator; | |
761 | struct request *free; | |
762 | ||
763 | if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn) | |
764 | if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next)) | |
765 | return 0; | |
766 | ||
767 | free = attempt_merge(q, rq, next); | |
768 | if (free) { | |
769 | __blk_put_request(q, free); | |
770 | return 1; | |
771 | } | |
772 | ||
773 | return 0; | |
774 | } | |
775 | ||
776 | bool blk_rq_merge_ok(struct request *rq, struct bio *bio) | |
777 | { | |
778 | if (!rq_mergeable(rq) || !bio_mergeable(bio)) | |
779 | return false; | |
780 | ||
781 | if (req_op(rq) != bio_op(bio)) | |
782 | return false; | |
783 | ||
784 | /* different data direction or already started, don't merge */ | |
785 | if (bio_data_dir(bio) != rq_data_dir(rq)) | |
786 | return false; | |
787 | ||
788 | /* must be same device and not a special request */ | |
789 | if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq)) | |
790 | return false; | |
791 | ||
792 | /* only merge integrity protected bio into ditto rq */ | |
793 | if (blk_integrity_merge_bio(rq->q, rq, bio) == false) | |
794 | return false; | |
795 | ||
796 | /* must be using the same buffer */ | |
797 | if (req_op(rq) == REQ_OP_WRITE_SAME && | |
798 | !blk_write_same_mergeable(rq->bio, bio)) | |
799 | return false; | |
800 | ||
801 | return true; | |
802 | } | |
803 | ||
804 | enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) | |
805 | { | |
806 | if (req_op(rq) == REQ_OP_DISCARD && | |
807 | queue_max_discard_segments(rq->q) > 1) | |
808 | return ELEVATOR_DISCARD_MERGE; | |
809 | else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) | |
810 | return ELEVATOR_BACK_MERGE; | |
811 | else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) | |
812 | return ELEVATOR_FRONT_MERGE; | |
813 | return ELEVATOR_NO_MERGE; | |
814 | } |