Loading...
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 "blk.h"
11
12static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
13 struct bio *bio)
14{
15 struct bio_vec *bv, *bvprv = NULL;
16 int cluster, i, high, highprv = 1;
17 unsigned int seg_size, nr_phys_segs;
18 struct bio *fbio, *bbio;
19
20 if (!bio)
21 return 0;
22
23 fbio = bio;
24 cluster = blk_queue_cluster(q);
25 seg_size = 0;
26 nr_phys_segs = 0;
27 for_each_bio(bio) {
28 bio_for_each_segment(bv, bio, i) {
29 /*
30 * the trick here is making sure that a high page is
31 * never considered part of another segment, since that
32 * might change with the bounce page.
33 */
34 high = page_to_pfn(bv->bv_page) > queue_bounce_pfn(q);
35 if (high || highprv)
36 goto new_segment;
37 if (cluster) {
38 if (seg_size + bv->bv_len
39 > queue_max_segment_size(q))
40 goto new_segment;
41 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
42 goto new_segment;
43 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
44 goto new_segment;
45
46 seg_size += bv->bv_len;
47 bvprv = bv;
48 continue;
49 }
50new_segment:
51 if (nr_phys_segs == 1 && seg_size >
52 fbio->bi_seg_front_size)
53 fbio->bi_seg_front_size = seg_size;
54
55 nr_phys_segs++;
56 bvprv = bv;
57 seg_size = bv->bv_len;
58 highprv = high;
59 }
60 bbio = bio;
61 }
62
63 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
64 fbio->bi_seg_front_size = seg_size;
65 if (seg_size > bbio->bi_seg_back_size)
66 bbio->bi_seg_back_size = seg_size;
67
68 return nr_phys_segs;
69}
70
71void blk_recalc_rq_segments(struct request *rq)
72{
73 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
74}
75
76void blk_recount_segments(struct request_queue *q, struct bio *bio)
77{
78 struct bio *nxt = bio->bi_next;
79
80 bio->bi_next = NULL;
81 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
82 bio->bi_next = nxt;
83 bio->bi_flags |= (1 << BIO_SEG_VALID);
84}
85EXPORT_SYMBOL(blk_recount_segments);
86
87static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
88 struct bio *nxt)
89{
90 if (!blk_queue_cluster(q))
91 return 0;
92
93 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
94 queue_max_segment_size(q))
95 return 0;
96
97 if (!bio_has_data(bio))
98 return 1;
99
100 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
101 return 0;
102
103 /*
104 * bio and nxt are contiguous in memory; check if the queue allows
105 * these two to be merged into one
106 */
107 if (BIO_SEG_BOUNDARY(q, bio, nxt))
108 return 1;
109
110 return 0;
111}
112
113/*
114 * map a request to scatterlist, return number of sg entries setup. Caller
115 * must make sure sg can hold rq->nr_phys_segments entries
116 */
117int blk_rq_map_sg(struct request_queue *q, struct request *rq,
118 struct scatterlist *sglist)
119{
120 struct bio_vec *bvec, *bvprv;
121 struct req_iterator iter;
122 struct scatterlist *sg;
123 int nsegs, cluster;
124
125 nsegs = 0;
126 cluster = blk_queue_cluster(q);
127
128 /*
129 * for each bio in rq
130 */
131 bvprv = NULL;
132 sg = NULL;
133 rq_for_each_segment(bvec, rq, iter) {
134 int nbytes = bvec->bv_len;
135
136 if (bvprv && cluster) {
137 if (sg->length + nbytes > queue_max_segment_size(q))
138 goto new_segment;
139
140 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
141 goto new_segment;
142 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
143 goto new_segment;
144
145 sg->length += nbytes;
146 } else {
147new_segment:
148 if (!sg)
149 sg = sglist;
150 else {
151 /*
152 * If the driver previously mapped a shorter
153 * list, we could see a termination bit
154 * prematurely unless it fully inits the sg
155 * table on each mapping. We KNOW that there
156 * must be more entries here or the driver
157 * would be buggy, so force clear the
158 * termination bit to avoid doing a full
159 * sg_init_table() in drivers for each command.
160 */
161 sg->page_link &= ~0x02;
162 sg = sg_next(sg);
163 }
164
165 sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
166 nsegs++;
167 }
168 bvprv = bvec;
169 } /* segments in rq */
170
171
172 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
173 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
174 unsigned int pad_len =
175 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
176
177 sg->length += pad_len;
178 rq->extra_len += pad_len;
179 }
180
181 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
182 if (rq->cmd_flags & REQ_WRITE)
183 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
184
185 sg->page_link &= ~0x02;
186 sg = sg_next(sg);
187 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
188 q->dma_drain_size,
189 ((unsigned long)q->dma_drain_buffer) &
190 (PAGE_SIZE - 1));
191 nsegs++;
192 rq->extra_len += q->dma_drain_size;
193 }
194
195 if (sg)
196 sg_mark_end(sg);
197
198 return nsegs;
199}
200EXPORT_SYMBOL(blk_rq_map_sg);
201
202static inline int ll_new_hw_segment(struct request_queue *q,
203 struct request *req,
204 struct bio *bio)
205{
206 int nr_phys_segs = bio_phys_segments(q, bio);
207
208 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
209 goto no_merge;
210
211 if (bio_integrity(bio) && blk_integrity_merge_bio(q, req, bio))
212 goto no_merge;
213
214 /*
215 * This will form the start of a new hw segment. Bump both
216 * counters.
217 */
218 req->nr_phys_segments += nr_phys_segs;
219 return 1;
220
221no_merge:
222 req->cmd_flags |= REQ_NOMERGE;
223 if (req == q->last_merge)
224 q->last_merge = NULL;
225 return 0;
226}
227
228int ll_back_merge_fn(struct request_queue *q, struct request *req,
229 struct bio *bio)
230{
231 unsigned short max_sectors;
232
233 if (unlikely(req->cmd_type == REQ_TYPE_BLOCK_PC))
234 max_sectors = queue_max_hw_sectors(q);
235 else
236 max_sectors = queue_max_sectors(q);
237
238 if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
239 req->cmd_flags |= REQ_NOMERGE;
240 if (req == q->last_merge)
241 q->last_merge = NULL;
242 return 0;
243 }
244 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
245 blk_recount_segments(q, req->biotail);
246 if (!bio_flagged(bio, BIO_SEG_VALID))
247 blk_recount_segments(q, bio);
248
249 return ll_new_hw_segment(q, req, bio);
250}
251
252int ll_front_merge_fn(struct request_queue *q, struct request *req,
253 struct bio *bio)
254{
255 unsigned short max_sectors;
256
257 if (unlikely(req->cmd_type == REQ_TYPE_BLOCK_PC))
258 max_sectors = queue_max_hw_sectors(q);
259 else
260 max_sectors = queue_max_sectors(q);
261
262
263 if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
264 req->cmd_flags |= REQ_NOMERGE;
265 if (req == q->last_merge)
266 q->last_merge = NULL;
267 return 0;
268 }
269 if (!bio_flagged(bio, BIO_SEG_VALID))
270 blk_recount_segments(q, bio);
271 if (!bio_flagged(req->bio, BIO_SEG_VALID))
272 blk_recount_segments(q, req->bio);
273
274 return ll_new_hw_segment(q, req, bio);
275}
276
277static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
278 struct request *next)
279{
280 int total_phys_segments;
281 unsigned int seg_size =
282 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
283
284 /*
285 * First check if the either of the requests are re-queued
286 * requests. Can't merge them if they are.
287 */
288 if (req->special || next->special)
289 return 0;
290
291 /*
292 * Will it become too large?
293 */
294 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > queue_max_sectors(q))
295 return 0;
296
297 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
298 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
299 if (req->nr_phys_segments == 1)
300 req->bio->bi_seg_front_size = seg_size;
301 if (next->nr_phys_segments == 1)
302 next->biotail->bi_seg_back_size = seg_size;
303 total_phys_segments--;
304 }
305
306 if (total_phys_segments > queue_max_segments(q))
307 return 0;
308
309 if (blk_integrity_rq(req) && blk_integrity_merge_rq(q, req, next))
310 return 0;
311
312 /* Merge is OK... */
313 req->nr_phys_segments = total_phys_segments;
314 return 1;
315}
316
317/**
318 * blk_rq_set_mixed_merge - mark a request as mixed merge
319 * @rq: request to mark as mixed merge
320 *
321 * Description:
322 * @rq is about to be mixed merged. Make sure the attributes
323 * which can be mixed are set in each bio and mark @rq as mixed
324 * merged.
325 */
326void blk_rq_set_mixed_merge(struct request *rq)
327{
328 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
329 struct bio *bio;
330
331 if (rq->cmd_flags & REQ_MIXED_MERGE)
332 return;
333
334 /*
335 * @rq will no longer represent mixable attributes for all the
336 * contained bios. It will just track those of the first one.
337 * Distributes the attributs to each bio.
338 */
339 for (bio = rq->bio; bio; bio = bio->bi_next) {
340 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
341 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
342 bio->bi_rw |= ff;
343 }
344 rq->cmd_flags |= REQ_MIXED_MERGE;
345}
346
347static void blk_account_io_merge(struct request *req)
348{
349 if (blk_do_io_stat(req)) {
350 struct hd_struct *part;
351 int cpu;
352
353 cpu = part_stat_lock();
354 part = req->part;
355
356 part_round_stats(cpu, part);
357 part_dec_in_flight(part, rq_data_dir(req));
358
359 hd_struct_put(part);
360 part_stat_unlock();
361 }
362}
363
364/*
365 * Has to be called with the request spinlock acquired
366 */
367static int attempt_merge(struct request_queue *q, struct request *req,
368 struct request *next)
369{
370 if (!rq_mergeable(req) || !rq_mergeable(next))
371 return 0;
372
373 /*
374 * Don't merge file system requests and discard requests
375 */
376 if ((req->cmd_flags & REQ_DISCARD) != (next->cmd_flags & REQ_DISCARD))
377 return 0;
378
379 /*
380 * Don't merge discard requests and secure discard requests
381 */
382 if ((req->cmd_flags & REQ_SECURE) != (next->cmd_flags & REQ_SECURE))
383 return 0;
384
385 /*
386 * not contiguous
387 */
388 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
389 return 0;
390
391 if (rq_data_dir(req) != rq_data_dir(next)
392 || req->rq_disk != next->rq_disk
393 || next->special)
394 return 0;
395
396 /*
397 * If we are allowed to merge, then append bio list
398 * from next to rq and release next. merge_requests_fn
399 * will have updated segment counts, update sector
400 * counts here.
401 */
402 if (!ll_merge_requests_fn(q, req, next))
403 return 0;
404
405 /*
406 * If failfast settings disagree or any of the two is already
407 * a mixed merge, mark both as mixed before proceeding. This
408 * makes sure that all involved bios have mixable attributes
409 * set properly.
410 */
411 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
412 (req->cmd_flags & REQ_FAILFAST_MASK) !=
413 (next->cmd_flags & REQ_FAILFAST_MASK)) {
414 blk_rq_set_mixed_merge(req);
415 blk_rq_set_mixed_merge(next);
416 }
417
418 /*
419 * At this point we have either done a back merge
420 * or front merge. We need the smaller start_time of
421 * the merged requests to be the current request
422 * for accounting purposes.
423 */
424 if (time_after(req->start_time, next->start_time))
425 req->start_time = next->start_time;
426
427 req->biotail->bi_next = next->bio;
428 req->biotail = next->biotail;
429
430 req->__data_len += blk_rq_bytes(next);
431
432 elv_merge_requests(q, req, next);
433
434 /*
435 * 'next' is going away, so update stats accordingly
436 */
437 blk_account_io_merge(next);
438
439 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
440 if (blk_rq_cpu_valid(next))
441 req->cpu = next->cpu;
442
443 /* owner-ship of bio passed from next to req */
444 next->bio = NULL;
445 __blk_put_request(q, next);
446 return 1;
447}
448
449int attempt_back_merge(struct request_queue *q, struct request *rq)
450{
451 struct request *next = elv_latter_request(q, rq);
452
453 if (next)
454 return attempt_merge(q, rq, next);
455
456 return 0;
457}
458
459int attempt_front_merge(struct request_queue *q, struct request *rq)
460{
461 struct request *prev = elv_former_request(q, rq);
462
463 if (prev)
464 return attempt_merge(q, prev, rq);
465
466 return 0;
467}
468
469int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
470 struct request *next)
471{
472 return attempt_merge(q, rq, next);
473}
474
475bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
476{
477 if (!rq_mergeable(rq))
478 return false;
479
480 /* don't merge file system requests and discard requests */
481 if ((bio->bi_rw & REQ_DISCARD) != (rq->bio->bi_rw & REQ_DISCARD))
482 return false;
483
484 /* don't merge discard requests and secure discard requests */
485 if ((bio->bi_rw & REQ_SECURE) != (rq->bio->bi_rw & REQ_SECURE))
486 return false;
487
488 /* different data direction or already started, don't merge */
489 if (bio_data_dir(bio) != rq_data_dir(rq))
490 return false;
491
492 /* must be same device and not a special request */
493 if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
494 return false;
495
496 /* only merge integrity protected bio into ditto rq */
497 if (bio_integrity(bio) != blk_integrity_rq(rq))
498 return false;
499
500 return true;
501}
502
503int blk_try_merge(struct request *rq, struct bio *bio)
504{
505 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_sector)
506 return ELEVATOR_BACK_MERGE;
507 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_sector)
508 return ELEVATOR_FRONT_MERGE;
509 return ELEVATOR_NO_MERGE;
510}
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
14static 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
57static 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
73static 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
85static 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 }
160new_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;
175split:
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
191void 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}
229EXPORT_SYMBOL(blk_queue_split);
230
231static 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 }
279new_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
300void 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
309void 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}
332EXPORT_SYMBOL(blk_recount_segments);
333
334static 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
365static 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 {
384new_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
408static 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
416static 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 */
436int 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}
483EXPORT_SYMBOL(blk_rq_map_sg);
484
485static void req_set_nomerge(struct request_queue *q, struct request *req)
486{
487 req->cmd_flags |= REQ_NOMERGE;
488 if (req == q->last_merge)
489 q->last_merge = NULL;
490}
491
492static inline int ll_new_hw_segment(struct request_queue *q,
493 struct request *req,
494 struct bio *bio)
495{
496 int nr_phys_segs = bio_phys_segments(q, bio);
497
498 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
499 goto no_merge;
500
501 if (blk_integrity_merge_bio(q, req, bio) == false)
502 goto no_merge;
503
504 /*
505 * This will form the start of a new hw segment. Bump both
506 * counters.
507 */
508 req->nr_phys_segments += nr_phys_segs;
509 return 1;
510
511no_merge:
512 req_set_nomerge(q, req);
513 return 0;
514}
515
516int ll_back_merge_fn(struct request_queue *q, struct request *req,
517 struct bio *bio)
518{
519 if (req_gap_back_merge(req, bio))
520 return 0;
521 if (blk_integrity_rq(req) &&
522 integrity_req_gap_back_merge(req, bio))
523 return 0;
524 if (blk_rq_sectors(req) + bio_sectors(bio) >
525 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
526 req_set_nomerge(q, req);
527 return 0;
528 }
529 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
530 blk_recount_segments(q, req->biotail);
531 if (!bio_flagged(bio, BIO_SEG_VALID))
532 blk_recount_segments(q, bio);
533
534 return ll_new_hw_segment(q, req, bio);
535}
536
537int ll_front_merge_fn(struct request_queue *q, struct request *req,
538 struct bio *bio)
539{
540
541 if (req_gap_front_merge(req, bio))
542 return 0;
543 if (blk_integrity_rq(req) &&
544 integrity_req_gap_front_merge(req, bio))
545 return 0;
546 if (blk_rq_sectors(req) + bio_sectors(bio) >
547 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
548 req_set_nomerge(q, req);
549 return 0;
550 }
551 if (!bio_flagged(bio, BIO_SEG_VALID))
552 blk_recount_segments(q, bio);
553 if (!bio_flagged(req->bio, BIO_SEG_VALID))
554 blk_recount_segments(q, req->bio);
555
556 return ll_new_hw_segment(q, req, bio);
557}
558
559/*
560 * blk-mq uses req->special to carry normal driver per-request payload, it
561 * does not indicate a prepared command that we cannot merge with.
562 */
563static bool req_no_special_merge(struct request *req)
564{
565 struct request_queue *q = req->q;
566
567 return !q->mq_ops && req->special;
568}
569
570static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
571 struct request *next)
572{
573 int total_phys_segments;
574 unsigned int seg_size =
575 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
576
577 /*
578 * First check if the either of the requests are re-queued
579 * requests. Can't merge them if they are.
580 */
581 if (req_no_special_merge(req) || req_no_special_merge(next))
582 return 0;
583
584 if (req_gap_back_merge(req, next->bio))
585 return 0;
586
587 /*
588 * Will it become too large?
589 */
590 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
591 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
592 return 0;
593
594 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
595 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
596 if (req->nr_phys_segments == 1)
597 req->bio->bi_seg_front_size = seg_size;
598 if (next->nr_phys_segments == 1)
599 next->biotail->bi_seg_back_size = seg_size;
600 total_phys_segments--;
601 }
602
603 if (total_phys_segments > queue_max_segments(q))
604 return 0;
605
606 if (blk_integrity_merge_rq(q, req, next) == false)
607 return 0;
608
609 /* Merge is OK... */
610 req->nr_phys_segments = total_phys_segments;
611 return 1;
612}
613
614/**
615 * blk_rq_set_mixed_merge - mark a request as mixed merge
616 * @rq: request to mark as mixed merge
617 *
618 * Description:
619 * @rq is about to be mixed merged. Make sure the attributes
620 * which can be mixed are set in each bio and mark @rq as mixed
621 * merged.
622 */
623void blk_rq_set_mixed_merge(struct request *rq)
624{
625 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
626 struct bio *bio;
627
628 if (rq->rq_flags & RQF_MIXED_MERGE)
629 return;
630
631 /*
632 * @rq will no longer represent mixable attributes for all the
633 * contained bios. It will just track those of the first one.
634 * Distributes the attributs to each bio.
635 */
636 for (bio = rq->bio; bio; bio = bio->bi_next) {
637 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
638 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
639 bio->bi_opf |= ff;
640 }
641 rq->rq_flags |= RQF_MIXED_MERGE;
642}
643
644static void blk_account_io_merge(struct request *req)
645{
646 if (blk_do_io_stat(req)) {
647 struct hd_struct *part;
648 int cpu;
649
650 cpu = part_stat_lock();
651 part = req->part;
652
653 part_round_stats(cpu, part);
654 part_dec_in_flight(part, rq_data_dir(req));
655
656 hd_struct_put(part);
657 part_stat_unlock();
658 }
659}
660
661/*
662 * Has to be called with the request spinlock acquired
663 */
664static int attempt_merge(struct request_queue *q, struct request *req,
665 struct request *next)
666{
667 if (!rq_mergeable(req) || !rq_mergeable(next))
668 return 0;
669
670 if (req_op(req) != req_op(next))
671 return 0;
672
673 /*
674 * not contiguous
675 */
676 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
677 return 0;
678
679 if (rq_data_dir(req) != rq_data_dir(next)
680 || req->rq_disk != next->rq_disk
681 || req_no_special_merge(next))
682 return 0;
683
684 if (req_op(req) == REQ_OP_WRITE_SAME &&
685 !blk_write_same_mergeable(req->bio, next->bio))
686 return 0;
687
688 /*
689 * If we are allowed to merge, then append bio list
690 * from next to rq and release next. merge_requests_fn
691 * will have updated segment counts, update sector
692 * counts here.
693 */
694 if (!ll_merge_requests_fn(q, req, next))
695 return 0;
696
697 /*
698 * If failfast settings disagree or any of the two is already
699 * a mixed merge, mark both as mixed before proceeding. This
700 * makes sure that all involved bios have mixable attributes
701 * set properly.
702 */
703 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
704 (req->cmd_flags & REQ_FAILFAST_MASK) !=
705 (next->cmd_flags & REQ_FAILFAST_MASK)) {
706 blk_rq_set_mixed_merge(req);
707 blk_rq_set_mixed_merge(next);
708 }
709
710 /*
711 * At this point we have either done a back merge
712 * or front merge. We need the smaller start_time of
713 * the merged requests to be the current request
714 * for accounting purposes.
715 */
716 if (time_after(req->start_time, next->start_time))
717 req->start_time = next->start_time;
718
719 req->biotail->bi_next = next->bio;
720 req->biotail = next->biotail;
721
722 req->__data_len += blk_rq_bytes(next);
723
724 elv_merge_requests(q, req, next);
725
726 /*
727 * 'next' is going away, so update stats accordingly
728 */
729 blk_account_io_merge(next);
730
731 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
732 if (blk_rq_cpu_valid(next))
733 req->cpu = next->cpu;
734
735 /* owner-ship of bio passed from next to req */
736 next->bio = NULL;
737 __blk_put_request(q, next);
738 return 1;
739}
740
741int attempt_back_merge(struct request_queue *q, struct request *rq)
742{
743 struct request *next = elv_latter_request(q, rq);
744
745 if (next)
746 return attempt_merge(q, rq, next);
747
748 return 0;
749}
750
751int attempt_front_merge(struct request_queue *q, struct request *rq)
752{
753 struct request *prev = elv_former_request(q, rq);
754
755 if (prev)
756 return attempt_merge(q, prev, rq);
757
758 return 0;
759}
760
761int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
762 struct request *next)
763{
764 struct elevator_queue *e = q->elevator;
765
766 if (e->type->ops.elevator_allow_rq_merge_fn)
767 if (!e->type->ops.elevator_allow_rq_merge_fn(q, rq, next))
768 return 0;
769
770 return attempt_merge(q, rq, next);
771}
772
773bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
774{
775 if (!rq_mergeable(rq) || !bio_mergeable(bio))
776 return false;
777
778 if (req_op(rq) != bio_op(bio))
779 return false;
780
781 /* different data direction or already started, don't merge */
782 if (bio_data_dir(bio) != rq_data_dir(rq))
783 return false;
784
785 /* must be same device and not a special request */
786 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
787 return false;
788
789 /* only merge integrity protected bio into ditto rq */
790 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
791 return false;
792
793 /* must be using the same buffer */
794 if (req_op(rq) == REQ_OP_WRITE_SAME &&
795 !blk_write_same_mergeable(rq->bio, bio))
796 return false;
797
798 return true;
799}
800
801int blk_try_merge(struct request *rq, struct bio *bio)
802{
803 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
804 return ELEVATOR_BACK_MERGE;
805 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
806 return ELEVATOR_FRONT_MERGE;
807 return ELEVATOR_NO_MERGE;
808}