]>
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_zeroes_split(struct request_queue *q, | |
58 | struct bio *bio, struct bio_set *bs, unsigned *nsegs) | |
59 | { | |
60 | *nsegs = 1; | |
61 | ||
62 | if (!q->limits.max_write_zeroes_sectors) | |
63 | return NULL; | |
64 | ||
65 | if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors) | |
66 | return NULL; | |
67 | ||
68 | return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs); | |
69 | } | |
70 | ||
71 | static struct bio *blk_bio_write_same_split(struct request_queue *q, | |
72 | struct bio *bio, | |
73 | struct bio_set *bs, | |
74 | unsigned *nsegs) | |
75 | { | |
76 | *nsegs = 1; | |
77 | ||
78 | if (!q->limits.max_write_same_sectors) | |
79 | return NULL; | |
80 | ||
81 | if (bio_sectors(bio) <= q->limits.max_write_same_sectors) | |
82 | return NULL; | |
83 | ||
84 | return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs); | |
85 | } | |
86 | ||
87 | static inline unsigned get_max_io_size(struct request_queue *q, | |
88 | struct bio *bio) | |
89 | { | |
90 | unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector); | |
91 | unsigned mask = queue_logical_block_size(q) - 1; | |
92 | ||
93 | /* aligned to logical block size */ | |
94 | sectors &= ~(mask >> 9); | |
95 | ||
96 | return sectors; | |
97 | } | |
98 | ||
99 | static struct bio *blk_bio_segment_split(struct request_queue *q, | |
100 | struct bio *bio, | |
101 | struct bio_set *bs, | |
102 | unsigned *segs) | |
103 | { | |
104 | struct bio_vec bv, bvprv, *bvprvp = NULL; | |
105 | struct bvec_iter iter; | |
106 | unsigned seg_size = 0, nsegs = 0, sectors = 0; | |
107 | unsigned front_seg_size = bio->bi_seg_front_size; | |
108 | bool do_split = true; | |
109 | struct bio *new = NULL; | |
110 | const unsigned max_sectors = get_max_io_size(q, bio); | |
111 | ||
112 | bio_for_each_segment(bv, bio, iter) { | |
113 | /* | |
114 | * If the queue doesn't support SG gaps and adding this | |
115 | * offset would create a gap, disallow it. | |
116 | */ | |
117 | if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset)) | |
118 | goto split; | |
119 | ||
120 | if (sectors + (bv.bv_len >> 9) > max_sectors) { | |
121 | /* | |
122 | * Consider this a new segment if we're splitting in | |
123 | * the middle of this vector. | |
124 | */ | |
125 | if (nsegs < queue_max_segments(q) && | |
126 | sectors < max_sectors) { | |
127 | nsegs++; | |
128 | sectors = max_sectors; | |
129 | } | |
130 | if (sectors) | |
131 | goto split; | |
132 | /* Make this single bvec as the 1st segment */ | |
133 | } | |
134 | ||
135 | if (bvprvp && blk_queue_cluster(q)) { | |
136 | if (seg_size + bv.bv_len > queue_max_segment_size(q)) | |
137 | goto new_segment; | |
138 | if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv)) | |
139 | goto new_segment; | |
140 | if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv)) | |
141 | goto new_segment; | |
142 | ||
143 | seg_size += bv.bv_len; | |
144 | bvprv = bv; | |
145 | bvprvp = &bvprv; | |
146 | sectors += bv.bv_len >> 9; | |
147 | ||
148 | if (nsegs == 1 && seg_size > front_seg_size) | |
149 | front_seg_size = seg_size; | |
150 | continue; | |
151 | } | |
152 | new_segment: | |
153 | if (nsegs == queue_max_segments(q)) | |
154 | goto split; | |
155 | ||
156 | nsegs++; | |
157 | bvprv = bv; | |
158 | bvprvp = &bvprv; | |
159 | seg_size = bv.bv_len; | |
160 | sectors += bv.bv_len >> 9; | |
161 | ||
162 | if (nsegs == 1 && seg_size > front_seg_size) | |
163 | front_seg_size = seg_size; | |
164 | } | |
165 | ||
166 | do_split = false; | |
167 | split: | |
168 | *segs = nsegs; | |
169 | ||
170 | if (do_split) { | |
171 | new = bio_split(bio, sectors, GFP_NOIO, bs); | |
172 | if (new) | |
173 | bio = new; | |
174 | } | |
175 | ||
176 | bio->bi_seg_front_size = front_seg_size; | |
177 | if (seg_size > bio->bi_seg_back_size) | |
178 | bio->bi_seg_back_size = seg_size; | |
179 | ||
180 | return do_split ? new : NULL; | |
181 | } | |
182 | ||
183 | void blk_queue_split(struct request_queue *q, struct bio **bio) | |
184 | { | |
185 | struct bio *split, *res; | |
186 | unsigned nsegs; | |
187 | ||
188 | switch (bio_op(*bio)) { | |
189 | case REQ_OP_DISCARD: | |
190 | case REQ_OP_SECURE_ERASE: | |
191 | split = blk_bio_discard_split(q, *bio, q->bio_split, &nsegs); | |
192 | break; | |
193 | case REQ_OP_WRITE_ZEROES: | |
194 | split = blk_bio_write_zeroes_split(q, *bio, q->bio_split, &nsegs); | |
195 | break; | |
196 | case REQ_OP_WRITE_SAME: | |
197 | split = blk_bio_write_same_split(q, *bio, q->bio_split, &nsegs); | |
198 | break; | |
199 | default: | |
200 | split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs); | |
201 | break; | |
202 | } | |
203 | ||
204 | /* physical segments can be figured out during splitting */ | |
205 | res = split ? split : *bio; | |
206 | res->bi_phys_segments = nsegs; | |
207 | bio_set_flag(res, BIO_SEG_VALID); | |
208 | ||
209 | if (split) { | |
210 | /* there isn't chance to merge the splitted bio */ | |
211 | split->bi_opf |= REQ_NOMERGE; | |
212 | ||
213 | bio_chain(split, *bio); | |
214 | trace_block_split(q, split, (*bio)->bi_iter.bi_sector); | |
215 | generic_make_request(*bio); | |
216 | *bio = split; | |
217 | } | |
218 | } | |
219 | EXPORT_SYMBOL(blk_queue_split); | |
220 | ||
221 | static unsigned int __blk_recalc_rq_segments(struct request_queue *q, | |
222 | struct bio *bio, | |
223 | bool no_sg_merge) | |
224 | { | |
225 | struct bio_vec bv, bvprv = { NULL }; | |
226 | int cluster, prev = 0; | |
227 | unsigned int seg_size, nr_phys_segs; | |
228 | struct bio *fbio, *bbio; | |
229 | struct bvec_iter iter; | |
230 | ||
231 | if (!bio) | |
232 | return 0; | |
233 | ||
234 | switch (bio_op(bio)) { | |
235 | case REQ_OP_DISCARD: | |
236 | case REQ_OP_SECURE_ERASE: | |
237 | case REQ_OP_WRITE_ZEROES: | |
238 | return 0; | |
239 | case REQ_OP_WRITE_SAME: | |
240 | return 1; | |
241 | } | |
242 | ||
243 | fbio = bio; | |
244 | cluster = blk_queue_cluster(q); | |
245 | seg_size = 0; | |
246 | nr_phys_segs = 0; | |
247 | for_each_bio(bio) { | |
248 | bio_for_each_segment(bv, bio, iter) { | |
249 | /* | |
250 | * If SG merging is disabled, each bio vector is | |
251 | * a segment | |
252 | */ | |
253 | if (no_sg_merge) | |
254 | goto new_segment; | |
255 | ||
256 | if (prev && cluster) { | |
257 | if (seg_size + bv.bv_len | |
258 | > queue_max_segment_size(q)) | |
259 | goto new_segment; | |
260 | if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv)) | |
261 | goto new_segment; | |
262 | if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv)) | |
263 | goto new_segment; | |
264 | ||
265 | seg_size += bv.bv_len; | |
266 | bvprv = bv; | |
267 | continue; | |
268 | } | |
269 | new_segment: | |
270 | if (nr_phys_segs == 1 && seg_size > | |
271 | fbio->bi_seg_front_size) | |
272 | fbio->bi_seg_front_size = seg_size; | |
273 | ||
274 | nr_phys_segs++; | |
275 | bvprv = bv; | |
276 | prev = 1; | |
277 | seg_size = bv.bv_len; | |
278 | } | |
279 | bbio = bio; | |
280 | } | |
281 | ||
282 | if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size) | |
283 | fbio->bi_seg_front_size = seg_size; | |
284 | if (seg_size > bbio->bi_seg_back_size) | |
285 | bbio->bi_seg_back_size = seg_size; | |
286 | ||
287 | return nr_phys_segs; | |
288 | } | |
289 | ||
290 | void blk_recalc_rq_segments(struct request *rq) | |
291 | { | |
292 | bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE, | |
293 | &rq->q->queue_flags); | |
294 | ||
295 | rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio, | |
296 | no_sg_merge); | |
297 | } | |
298 | ||
299 | void blk_recount_segments(struct request_queue *q, struct bio *bio) | |
300 | { | |
301 | unsigned short seg_cnt; | |
302 | ||
303 | /* estimate segment number by bi_vcnt for non-cloned bio */ | |
304 | if (bio_flagged(bio, BIO_CLONED)) | |
305 | seg_cnt = bio_segments(bio); | |
306 | else | |
307 | seg_cnt = bio->bi_vcnt; | |
308 | ||
309 | if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) && | |
310 | (seg_cnt < queue_max_segments(q))) | |
311 | bio->bi_phys_segments = seg_cnt; | |
312 | else { | |
313 | struct bio *nxt = bio->bi_next; | |
314 | ||
315 | bio->bi_next = NULL; | |
316 | bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false); | |
317 | bio->bi_next = nxt; | |
318 | } | |
319 | ||
320 | bio_set_flag(bio, BIO_SEG_VALID); | |
321 | } | |
322 | EXPORT_SYMBOL(blk_recount_segments); | |
323 | ||
324 | static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio, | |
325 | struct bio *nxt) | |
326 | { | |
327 | struct bio_vec end_bv = { NULL }, nxt_bv; | |
328 | ||
329 | if (!blk_queue_cluster(q)) | |
330 | return 0; | |
331 | ||
332 | if (bio->bi_seg_back_size + nxt->bi_seg_front_size > | |
333 | queue_max_segment_size(q)) | |
334 | return 0; | |
335 | ||
336 | if (!bio_has_data(bio)) | |
337 | return 1; | |
338 | ||
339 | bio_get_last_bvec(bio, &end_bv); | |
340 | bio_get_first_bvec(nxt, &nxt_bv); | |
341 | ||
342 | if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv)) | |
343 | return 0; | |
344 | ||
345 | /* | |
346 | * bio and nxt are contiguous in memory; check if the queue allows | |
347 | * these two to be merged into one | |
348 | */ | |
349 | if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv)) | |
350 | return 1; | |
351 | ||
352 | return 0; | |
353 | } | |
354 | ||
355 | static inline void | |
356 | __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec, | |
357 | struct scatterlist *sglist, struct bio_vec *bvprv, | |
358 | struct scatterlist **sg, int *nsegs, int *cluster) | |
359 | { | |
360 | ||
361 | int nbytes = bvec->bv_len; | |
362 | ||
363 | if (*sg && *cluster) { | |
364 | if ((*sg)->length + nbytes > queue_max_segment_size(q)) | |
365 | goto new_segment; | |
366 | ||
367 | if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec)) | |
368 | goto new_segment; | |
369 | if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec)) | |
370 | goto new_segment; | |
371 | ||
372 | (*sg)->length += nbytes; | |
373 | } else { | |
374 | new_segment: | |
375 | if (!*sg) | |
376 | *sg = sglist; | |
377 | else { | |
378 | /* | |
379 | * If the driver previously mapped a shorter | |
380 | * list, we could see a termination bit | |
381 | * prematurely unless it fully inits the sg | |
382 | * table on each mapping. We KNOW that there | |
383 | * must be more entries here or the driver | |
384 | * would be buggy, so force clear the | |
385 | * termination bit to avoid doing a full | |
386 | * sg_init_table() in drivers for each command. | |
387 | */ | |
388 | sg_unmark_end(*sg); | |
389 | *sg = sg_next(*sg); | |
390 | } | |
391 | ||
392 | sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset); | |
393 | (*nsegs)++; | |
394 | } | |
395 | *bvprv = *bvec; | |
396 | } | |
397 | ||
398 | static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv, | |
399 | struct scatterlist *sglist, struct scatterlist **sg) | |
400 | { | |
401 | *sg = sglist; | |
402 | sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset); | |
403 | return 1; | |
404 | } | |
405 | ||
406 | static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio, | |
407 | struct scatterlist *sglist, | |
408 | struct scatterlist **sg) | |
409 | { | |
410 | struct bio_vec bvec, bvprv = { NULL }; | |
411 | struct bvec_iter iter; | |
412 | int cluster = blk_queue_cluster(q), nsegs = 0; | |
413 | ||
414 | for_each_bio(bio) | |
415 | bio_for_each_segment(bvec, bio, iter) | |
416 | __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg, | |
417 | &nsegs, &cluster); | |
418 | ||
419 | return nsegs; | |
420 | } | |
421 | ||
422 | /* | |
423 | * map a request to scatterlist, return number of sg entries setup. Caller | |
424 | * must make sure sg can hold rq->nr_phys_segments entries | |
425 | */ | |
426 | int blk_rq_map_sg(struct request_queue *q, struct request *rq, | |
427 | struct scatterlist *sglist) | |
428 | { | |
429 | struct scatterlist *sg = NULL; | |
430 | int nsegs = 0; | |
431 | ||
432 | if (rq->rq_flags & RQF_SPECIAL_PAYLOAD) | |
433 | nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg); | |
434 | else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME) | |
435 | nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg); | |
436 | else if (rq->bio) | |
437 | nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg); | |
438 | ||
439 | if (unlikely(rq->rq_flags & RQF_COPY_USER) && | |
440 | (blk_rq_bytes(rq) & q->dma_pad_mask)) { | |
441 | unsigned int pad_len = | |
442 | (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1; | |
443 | ||
444 | sg->length += pad_len; | |
445 | rq->extra_len += pad_len; | |
446 | } | |
447 | ||
448 | if (q->dma_drain_size && q->dma_drain_needed(rq)) { | |
449 | if (op_is_write(req_op(rq))) | |
450 | memset(q->dma_drain_buffer, 0, q->dma_drain_size); | |
451 | ||
452 | sg_unmark_end(sg); | |
453 | sg = sg_next(sg); | |
454 | sg_set_page(sg, virt_to_page(q->dma_drain_buffer), | |
455 | q->dma_drain_size, | |
456 | ((unsigned long)q->dma_drain_buffer) & | |
457 | (PAGE_SIZE - 1)); | |
458 | nsegs++; | |
459 | rq->extra_len += q->dma_drain_size; | |
460 | } | |
461 | ||
462 | if (sg) | |
463 | sg_mark_end(sg); | |
464 | ||
465 | /* | |
466 | * Something must have been wrong if the figured number of | |
467 | * segment is bigger than number of req's physical segments | |
468 | */ | |
469 | WARN_ON(nsegs > blk_rq_nr_phys_segments(rq)); | |
470 | ||
471 | return nsegs; | |
472 | } | |
473 | EXPORT_SYMBOL(blk_rq_map_sg); | |
474 | ||
475 | static inline int ll_new_hw_segment(struct request_queue *q, | |
476 | struct request *req, | |
477 | struct bio *bio) | |
478 | { | |
479 | int nr_phys_segs = bio_phys_segments(q, bio); | |
480 | ||
481 | if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q)) | |
482 | goto no_merge; | |
483 | ||
484 | if (blk_integrity_merge_bio(q, req, bio) == false) | |
485 | goto no_merge; | |
486 | ||
487 | /* | |
488 | * This will form the start of a new hw segment. Bump both | |
489 | * counters. | |
490 | */ | |
491 | req->nr_phys_segments += nr_phys_segs; | |
492 | return 1; | |
493 | ||
494 | no_merge: | |
495 | req_set_nomerge(q, req); | |
496 | return 0; | |
497 | } | |
498 | ||
499 | int ll_back_merge_fn(struct request_queue *q, struct request *req, | |
500 | struct bio *bio) | |
501 | { | |
502 | if (req_gap_back_merge(req, bio)) | |
503 | return 0; | |
504 | if (blk_integrity_rq(req) && | |
505 | integrity_req_gap_back_merge(req, bio)) | |
506 | return 0; | |
507 | if (blk_rq_sectors(req) + bio_sectors(bio) > | |
508 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) { | |
509 | req_set_nomerge(q, req); | |
510 | return 0; | |
511 | } | |
512 | if (!bio_flagged(req->biotail, BIO_SEG_VALID)) | |
513 | blk_recount_segments(q, req->biotail); | |
514 | if (!bio_flagged(bio, BIO_SEG_VALID)) | |
515 | blk_recount_segments(q, bio); | |
516 | ||
517 | return ll_new_hw_segment(q, req, bio); | |
518 | } | |
519 | ||
520 | int ll_front_merge_fn(struct request_queue *q, struct request *req, | |
521 | struct bio *bio) | |
522 | { | |
523 | ||
524 | if (req_gap_front_merge(req, bio)) | |
525 | return 0; | |
526 | if (blk_integrity_rq(req) && | |
527 | integrity_req_gap_front_merge(req, bio)) | |
528 | return 0; | |
529 | if (blk_rq_sectors(req) + bio_sectors(bio) > | |
530 | blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) { | |
531 | req_set_nomerge(q, req); | |
532 | return 0; | |
533 | } | |
534 | if (!bio_flagged(bio, BIO_SEG_VALID)) | |
535 | blk_recount_segments(q, bio); | |
536 | if (!bio_flagged(req->bio, BIO_SEG_VALID)) | |
537 | blk_recount_segments(q, req->bio); | |
538 | ||
539 | return ll_new_hw_segment(q, req, bio); | |
540 | } | |
541 | ||
542 | /* | |
543 | * blk-mq uses req->special to carry normal driver per-request payload, it | |
544 | * does not indicate a prepared command that we cannot merge with. | |
545 | */ | |
546 | static bool req_no_special_merge(struct request *req) | |
547 | { | |
548 | struct request_queue *q = req->q; | |
549 | ||
550 | return !q->mq_ops && req->special; | |
551 | } | |
552 | ||
553 | static int ll_merge_requests_fn(struct request_queue *q, struct request *req, | |
554 | struct request *next) | |
555 | { | |
556 | int total_phys_segments; | |
557 | unsigned int seg_size = | |
558 | req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size; | |
559 | ||
560 | /* | |
561 | * First check if the either of the requests are re-queued | |
562 | * requests. Can't merge them if they are. | |
563 | */ | |
564 | if (req_no_special_merge(req) || req_no_special_merge(next)) | |
565 | return 0; | |
566 | ||
567 | if (req_gap_back_merge(req, next->bio)) | |
568 | return 0; | |
569 | ||
570 | /* | |
571 | * Will it become too large? | |
572 | */ | |
573 | if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > | |
574 | blk_rq_get_max_sectors(req, blk_rq_pos(req))) | |
575 | return 0; | |
576 | ||
577 | total_phys_segments = req->nr_phys_segments + next->nr_phys_segments; | |
578 | if (blk_phys_contig_segment(q, req->biotail, next->bio)) { | |
579 | if (req->nr_phys_segments == 1) | |
580 | req->bio->bi_seg_front_size = seg_size; | |
581 | if (next->nr_phys_segments == 1) | |
582 | next->biotail->bi_seg_back_size = seg_size; | |
583 | total_phys_segments--; | |
584 | } | |
585 | ||
586 | if (total_phys_segments > queue_max_segments(q)) | |
587 | return 0; | |
588 | ||
589 | if (blk_integrity_merge_rq(q, req, next) == false) | |
590 | return 0; | |
591 | ||
592 | /* Merge is OK... */ | |
593 | req->nr_phys_segments = total_phys_segments; | |
594 | return 1; | |
595 | } | |
596 | ||
597 | /** | |
598 | * blk_rq_set_mixed_merge - mark a request as mixed merge | |
599 | * @rq: request to mark as mixed merge | |
600 | * | |
601 | * Description: | |
602 | * @rq is about to be mixed merged. Make sure the attributes | |
603 | * which can be mixed are set in each bio and mark @rq as mixed | |
604 | * merged. | |
605 | */ | |
606 | void blk_rq_set_mixed_merge(struct request *rq) | |
607 | { | |
608 | unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; | |
609 | struct bio *bio; | |
610 | ||
611 | if (rq->rq_flags & RQF_MIXED_MERGE) | |
612 | return; | |
613 | ||
614 | /* | |
615 | * @rq will no longer represent mixable attributes for all the | |
616 | * contained bios. It will just track those of the first one. | |
617 | * Distributes the attributs to each bio. | |
618 | */ | |
619 | for (bio = rq->bio; bio; bio = bio->bi_next) { | |
620 | WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) && | |
621 | (bio->bi_opf & REQ_FAILFAST_MASK) != ff); | |
622 | bio->bi_opf |= ff; | |
623 | } | |
624 | rq->rq_flags |= RQF_MIXED_MERGE; | |
625 | } | |
626 | ||
627 | static void blk_account_io_merge(struct request *req) | |
628 | { | |
629 | if (blk_do_io_stat(req)) { | |
630 | struct hd_struct *part; | |
631 | int cpu; | |
632 | ||
633 | cpu = part_stat_lock(); | |
634 | part = req->part; | |
635 | ||
636 | part_round_stats(cpu, part); | |
637 | part_dec_in_flight(part, rq_data_dir(req)); | |
638 | ||
639 | hd_struct_put(part); | |
640 | part_stat_unlock(); | |
641 | } | |
642 | } | |
643 | ||
644 | /* | |
645 | * For non-mq, this has to be called with the request spinlock acquired. | |
646 | * For mq with scheduling, the appropriate queue wide lock should be held. | |
647 | */ | |
648 | static struct request *attempt_merge(struct request_queue *q, | |
649 | struct request *req, struct request *next) | |
650 | { | |
651 | if (!q->mq_ops) | |
652 | lockdep_assert_held(q->queue_lock); | |
653 | ||
654 | if (!rq_mergeable(req) || !rq_mergeable(next)) | |
655 | return NULL; | |
656 | ||
657 | if (req_op(req) != req_op(next)) | |
658 | return NULL; | |
659 | ||
660 | /* | |
661 | * not contiguous | |
662 | */ | |
663 | if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next)) | |
664 | return NULL; | |
665 | ||
666 | if (rq_data_dir(req) != rq_data_dir(next) | |
667 | || req->rq_disk != next->rq_disk | |
668 | || req_no_special_merge(next)) | |
669 | return NULL; | |
670 | ||
671 | if (req_op(req) == REQ_OP_WRITE_SAME && | |
672 | !blk_write_same_mergeable(req->bio, next->bio)) | |
673 | return NULL; | |
674 | ||
675 | /* | |
676 | * Don't allow merge of different write hints, or for a hint with | |
677 | * non-hint IO. | |
678 | */ | |
679 | if (req->write_hint != next->write_hint) | |
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 | /* | |
802 | * Don't allow merge of different write hints, or for a hint with | |
803 | * non-hint IO. | |
804 | */ | |
805 | if (rq->write_hint != bio->bi_write_hint) | |
806 | return false; | |
807 | ||
808 | return true; | |
809 | } | |
810 | ||
811 | enum elv_merge blk_try_merge(struct request *rq, struct bio *bio) | |
812 | { | |
813 | if (req_op(rq) == REQ_OP_DISCARD && | |
814 | queue_max_discard_segments(rq->q) > 1) | |
815 | return ELEVATOR_DISCARD_MERGE; | |
816 | else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector) | |
817 | return ELEVATOR_BACK_MERGE; | |
818 | else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector) | |
819 | return ELEVATOR_FRONT_MERGE; | |
820 | return ELEVATOR_NO_MERGE; | |
821 | } |