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}
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
15static inline bool bio_will_gap(struct request_queue *q,
16 struct request *prev_rq, struct bio *prev, struct bio *next)
17{
18 struct bio_vec pb, nb;
19
20 if (!bio_has_data(prev) || !queue_virt_boundary(q))
21 return false;
22
23 /*
24 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
25 * is quite difficult to respect the sg gap limit. We work hard to
26 * merge a huge number of small single bios in case of mkfs.
27 */
28 if (prev_rq)
29 bio_get_first_bvec(prev_rq->bio, &pb);
30 else
31 bio_get_first_bvec(prev, &pb);
32 if (pb.bv_offset & queue_virt_boundary(q))
33 return true;
34
35 /*
36 * We don't need to worry about the situation that the merged segment
37 * ends in unaligned virt boundary:
38 *
39 * - if 'pb' ends aligned, the merged segment ends aligned
40 * - if 'pb' ends unaligned, the next bio must include
41 * one single bvec of 'nb', otherwise the 'nb' can't
42 * merge with 'pb'
43 */
44 bio_get_last_bvec(prev, &pb);
45 bio_get_first_bvec(next, &nb);
46 if (biovec_phys_mergeable(q, &pb, &nb))
47 return false;
48 return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
49}
50
51static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
52{
53 return bio_will_gap(req->q, req, req->biotail, bio);
54}
55
56static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
57{
58 return bio_will_gap(req->q, NULL, bio, req->bio);
59}
60
61static struct bio *blk_bio_discard_split(struct request_queue *q,
62 struct bio *bio,
63 struct bio_set *bs,
64 unsigned *nsegs)
65{
66 unsigned int max_discard_sectors, granularity;
67 int alignment;
68 sector_t tmp;
69 unsigned split_sectors;
70
71 *nsegs = 1;
72
73 /* Zero-sector (unknown) and one-sector granularities are the same. */
74 granularity = max(q->limits.discard_granularity >> 9, 1U);
75
76 max_discard_sectors = min(q->limits.max_discard_sectors,
77 bio_allowed_max_sectors(q));
78 max_discard_sectors -= max_discard_sectors % granularity;
79
80 if (unlikely(!max_discard_sectors)) {
81 /* XXX: warn */
82 return NULL;
83 }
84
85 if (bio_sectors(bio) <= max_discard_sectors)
86 return NULL;
87
88 split_sectors = max_discard_sectors;
89
90 /*
91 * If the next starting sector would be misaligned, stop the discard at
92 * the previous aligned sector.
93 */
94 alignment = (q->limits.discard_alignment >> 9) % granularity;
95
96 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
97 tmp = sector_div(tmp, granularity);
98
99 if (split_sectors > tmp)
100 split_sectors -= tmp;
101
102 return bio_split(bio, split_sectors, GFP_NOIO, bs);
103}
104
105static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
106 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
107{
108 *nsegs = 0;
109
110 if (!q->limits.max_write_zeroes_sectors)
111 return NULL;
112
113 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
114 return NULL;
115
116 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
117}
118
119static struct bio *blk_bio_write_same_split(struct request_queue *q,
120 struct bio *bio,
121 struct bio_set *bs,
122 unsigned *nsegs)
123{
124 *nsegs = 1;
125
126 if (!q->limits.max_write_same_sectors)
127 return NULL;
128
129 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
130 return NULL;
131
132 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
133}
134
135/*
136 * Return the maximum number of sectors from the start of a bio that may be
137 * submitted as a single request to a block device. If enough sectors remain,
138 * align the end to the physical block size. Otherwise align the end to the
139 * logical block size. This approach minimizes the number of non-aligned
140 * requests that are submitted to a block device if the start of a bio is not
141 * aligned to a physical block boundary.
142 */
143static inline unsigned get_max_io_size(struct request_queue *q,
144 struct bio *bio)
145{
146 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
147 unsigned max_sectors = sectors;
148 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
149 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
150 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
151
152 max_sectors += start_offset;
153 max_sectors &= ~(pbs - 1);
154 if (max_sectors > start_offset)
155 return max_sectors - start_offset;
156
157 return sectors & (lbs - 1);
158}
159
160static unsigned get_max_segment_size(const struct request_queue *q,
161 unsigned offset)
162{
163 unsigned long mask = queue_segment_boundary(q);
164
165 /* default segment boundary mask means no boundary limit */
166 if (mask == BLK_SEG_BOUNDARY_MASK)
167 return queue_max_segment_size(q);
168
169 return min_t(unsigned long, mask - (mask & offset) + 1,
170 queue_max_segment_size(q));
171}
172
173/**
174 * bvec_split_segs - verify whether or not a bvec should be split in the middle
175 * @q: [in] request queue associated with the bio associated with @bv
176 * @bv: [in] bvec to examine
177 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
178 * by the number of segments from @bv that may be appended to that
179 * bio without exceeding @max_segs
180 * @sectors: [in,out] Number of sectors in the bio being built. Incremented
181 * by the number of sectors from @bv that may be appended to that
182 * bio without exceeding @max_sectors
183 * @max_segs: [in] upper bound for *@nsegs
184 * @max_sectors: [in] upper bound for *@sectors
185 *
186 * When splitting a bio, it can happen that a bvec is encountered that is too
187 * big to fit in a single segment and hence that it has to be split in the
188 * middle. This function verifies whether or not that should happen. The value
189 * %true is returned if and only if appending the entire @bv to a bio with
190 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
191 * the block driver.
192 */
193static bool bvec_split_segs(const struct request_queue *q,
194 const struct bio_vec *bv, unsigned *nsegs,
195 unsigned *sectors, unsigned max_segs,
196 unsigned max_sectors)
197{
198 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
199 unsigned len = min(bv->bv_len, max_len);
200 unsigned total_len = 0;
201 unsigned seg_size = 0;
202
203 while (len && *nsegs < max_segs) {
204 seg_size = get_max_segment_size(q, bv->bv_offset + total_len);
205 seg_size = min(seg_size, len);
206
207 (*nsegs)++;
208 total_len += seg_size;
209 len -= seg_size;
210
211 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
212 break;
213 }
214
215 *sectors += total_len >> 9;
216
217 /* tell the caller to split the bvec if it is too big to fit */
218 return len > 0 || bv->bv_len > max_len;
219}
220
221/**
222 * blk_bio_segment_split - split a bio in two bios
223 * @q: [in] request queue pointer
224 * @bio: [in] bio to be split
225 * @bs: [in] bio set to allocate the clone from
226 * @segs: [out] number of segments in the bio with the first half of the sectors
227 *
228 * Clone @bio, update the bi_iter of the clone to represent the first sectors
229 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
230 * following is guaranteed for the cloned bio:
231 * - That it has at most get_max_io_size(@q, @bio) sectors.
232 * - That it has at most queue_max_segments(@q) segments.
233 *
234 * Except for discard requests the cloned bio will point at the bi_io_vec of
235 * the original bio. It is the responsibility of the caller to ensure that the
236 * original bio is not freed before the cloned bio. The caller is also
237 * responsible for ensuring that @bs is only destroyed after processing of the
238 * split bio has finished.
239 */
240static struct bio *blk_bio_segment_split(struct request_queue *q,
241 struct bio *bio,
242 struct bio_set *bs,
243 unsigned *segs)
244{
245 struct bio_vec bv, bvprv, *bvprvp = NULL;
246 struct bvec_iter iter;
247 unsigned nsegs = 0, sectors = 0;
248 const unsigned max_sectors = get_max_io_size(q, bio);
249 const unsigned max_segs = queue_max_segments(q);
250
251 bio_for_each_bvec(bv, bio, iter) {
252 /*
253 * If the queue doesn't support SG gaps and adding this
254 * offset would create a gap, disallow it.
255 */
256 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
257 goto split;
258
259 if (nsegs < max_segs &&
260 sectors + (bv.bv_len >> 9) <= max_sectors &&
261 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
262 nsegs++;
263 sectors += bv.bv_len >> 9;
264 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs,
265 max_sectors)) {
266 goto split;
267 }
268
269 bvprv = bv;
270 bvprvp = &bvprv;
271 }
272
273 *segs = nsegs;
274 return NULL;
275split:
276 *segs = nsegs;
277 return bio_split(bio, sectors, GFP_NOIO, bs);
278}
279
280/**
281 * __blk_queue_split - split a bio and submit the second half
282 * @q: [in] request queue pointer
283 * @bio: [in, out] bio to be split
284 * @nr_segs: [out] number of segments in the first bio
285 *
286 * Split a bio into two bios, chain the two bios, submit the second half and
287 * store a pointer to the first half in *@bio. If the second bio is still too
288 * big it will be split by a recursive call to this function. Since this
289 * function may allocate a new bio from @q->bio_split, it is the responsibility
290 * of the caller to ensure that @q is only released after processing of the
291 * split bio has finished.
292 */
293void __blk_queue_split(struct request_queue *q, struct bio **bio,
294 unsigned int *nr_segs)
295{
296 struct bio *split;
297
298 switch (bio_op(*bio)) {
299 case REQ_OP_DISCARD:
300 case REQ_OP_SECURE_ERASE:
301 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
302 break;
303 case REQ_OP_WRITE_ZEROES:
304 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
305 nr_segs);
306 break;
307 case REQ_OP_WRITE_SAME:
308 split = blk_bio_write_same_split(q, *bio, &q->bio_split,
309 nr_segs);
310 break;
311 default:
312 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
313 break;
314 }
315
316 if (split) {
317 /* there isn't chance to merge the splitted bio */
318 split->bi_opf |= REQ_NOMERGE;
319
320 /*
321 * Since we're recursing into make_request here, ensure
322 * that we mark this bio as already having entered the queue.
323 * If not, and the queue is going away, we can get stuck
324 * forever on waiting for the queue reference to drop. But
325 * that will never happen, as we're already holding a
326 * reference to it.
327 */
328 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
329
330 bio_chain(split, *bio);
331 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
332 generic_make_request(*bio);
333 *bio = split;
334 }
335}
336
337/**
338 * blk_queue_split - split a bio and submit the second half
339 * @q: [in] request queue pointer
340 * @bio: [in, out] bio to be split
341 *
342 * Split a bio into two bios, chains the two bios, submit the second half and
343 * store a pointer to the first half in *@bio. Since this function may allocate
344 * a new bio from @q->bio_split, it is the responsibility of the caller to
345 * ensure that @q is only released after processing of the split bio has
346 * finished.
347 */
348void blk_queue_split(struct request_queue *q, struct bio **bio)
349{
350 unsigned int nr_segs;
351
352 __blk_queue_split(q, bio, &nr_segs);
353}
354EXPORT_SYMBOL(blk_queue_split);
355
356unsigned int blk_recalc_rq_segments(struct request *rq)
357{
358 unsigned int nr_phys_segs = 0;
359 unsigned int nr_sectors = 0;
360 struct req_iterator iter;
361 struct bio_vec bv;
362
363 if (!rq->bio)
364 return 0;
365
366 switch (bio_op(rq->bio)) {
367 case REQ_OP_DISCARD:
368 case REQ_OP_SECURE_ERASE:
369 case REQ_OP_WRITE_ZEROES:
370 return 0;
371 case REQ_OP_WRITE_SAME:
372 return 1;
373 }
374
375 rq_for_each_bvec(bv, rq, iter)
376 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
377 UINT_MAX, UINT_MAX);
378 return nr_phys_segs;
379}
380
381static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
382 struct scatterlist *sglist)
383{
384 if (!*sg)
385 return sglist;
386
387 /*
388 * If the driver previously mapped a shorter list, we could see a
389 * termination bit prematurely unless it fully inits the sg table
390 * on each mapping. We KNOW that there must be more entries here
391 * or the driver would be buggy, so force clear the termination bit
392 * to avoid doing a full sg_init_table() in drivers for each command.
393 */
394 sg_unmark_end(*sg);
395 return sg_next(*sg);
396}
397
398static unsigned blk_bvec_map_sg(struct request_queue *q,
399 struct bio_vec *bvec, struct scatterlist *sglist,
400 struct scatterlist **sg)
401{
402 unsigned nbytes = bvec->bv_len;
403 unsigned nsegs = 0, total = 0;
404
405 while (nbytes > 0) {
406 unsigned offset = bvec->bv_offset + total;
407 unsigned len = min(get_max_segment_size(q, offset), nbytes);
408 struct page *page = bvec->bv_page;
409
410 /*
411 * Unfortunately a fair number of drivers barf on scatterlists
412 * that have an offset larger than PAGE_SIZE, despite other
413 * subsystems dealing with that invariant just fine. For now
414 * stick to the legacy format where we never present those from
415 * the block layer, but the code below should be removed once
416 * these offenders (mostly MMC/SD drivers) are fixed.
417 */
418 page += (offset >> PAGE_SHIFT);
419 offset &= ~PAGE_MASK;
420
421 *sg = blk_next_sg(sg, sglist);
422 sg_set_page(*sg, page, len, offset);
423
424 total += len;
425 nbytes -= len;
426 nsegs++;
427 }
428
429 return nsegs;
430}
431
432static inline int __blk_bvec_map_sg(struct bio_vec bv,
433 struct scatterlist *sglist, struct scatterlist **sg)
434{
435 *sg = blk_next_sg(sg, sglist);
436 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
437 return 1;
438}
439
440/* only try to merge bvecs into one sg if they are from two bios */
441static inline bool
442__blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
443 struct bio_vec *bvprv, struct scatterlist **sg)
444{
445
446 int nbytes = bvec->bv_len;
447
448 if (!*sg)
449 return false;
450
451 if ((*sg)->length + nbytes > queue_max_segment_size(q))
452 return false;
453
454 if (!biovec_phys_mergeable(q, bvprv, bvec))
455 return false;
456
457 (*sg)->length += nbytes;
458
459 return true;
460}
461
462static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
463 struct scatterlist *sglist,
464 struct scatterlist **sg)
465{
466 struct bio_vec uninitialized_var(bvec), bvprv = { NULL };
467 struct bvec_iter iter;
468 int nsegs = 0;
469 bool new_bio = false;
470
471 for_each_bio(bio) {
472 bio_for_each_bvec(bvec, bio, iter) {
473 /*
474 * Only try to merge bvecs from two bios given we
475 * have done bio internal merge when adding pages
476 * to bio
477 */
478 if (new_bio &&
479 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
480 goto next_bvec;
481
482 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
483 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
484 else
485 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
486 next_bvec:
487 new_bio = false;
488 }
489 if (likely(bio->bi_iter.bi_size)) {
490 bvprv = bvec;
491 new_bio = true;
492 }
493 }
494
495 return nsegs;
496}
497
498/*
499 * map a request to scatterlist, return number of sg entries setup. Caller
500 * must make sure sg can hold rq->nr_phys_segments entries
501 */
502int blk_rq_map_sg(struct request_queue *q, struct request *rq,
503 struct scatterlist *sglist)
504{
505 struct scatterlist *sg = NULL;
506 int nsegs = 0;
507
508 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
509 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, &sg);
510 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
511 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, &sg);
512 else if (rq->bio)
513 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
514
515 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
516 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
517 unsigned int pad_len =
518 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
519
520 sg->length += pad_len;
521 rq->extra_len += pad_len;
522 }
523
524 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
525 if (op_is_write(req_op(rq)))
526 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
527
528 sg_unmark_end(sg);
529 sg = sg_next(sg);
530 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
531 q->dma_drain_size,
532 ((unsigned long)q->dma_drain_buffer) &
533 (PAGE_SIZE - 1));
534 nsegs++;
535 rq->extra_len += q->dma_drain_size;
536 }
537
538 if (sg)
539 sg_mark_end(sg);
540
541 /*
542 * Something must have been wrong if the figured number of
543 * segment is bigger than number of req's physical segments
544 */
545 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
546
547 return nsegs;
548}
549EXPORT_SYMBOL(blk_rq_map_sg);
550
551static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
552 unsigned int nr_phys_segs)
553{
554 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(req->q))
555 goto no_merge;
556
557 if (blk_integrity_merge_bio(req->q, req, bio) == false)
558 goto no_merge;
559
560 /*
561 * This will form the start of a new hw segment. Bump both
562 * counters.
563 */
564 req->nr_phys_segments += nr_phys_segs;
565 return 1;
566
567no_merge:
568 req_set_nomerge(req->q, req);
569 return 0;
570}
571
572int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
573{
574 if (req_gap_back_merge(req, bio))
575 return 0;
576 if (blk_integrity_rq(req) &&
577 integrity_req_gap_back_merge(req, bio))
578 return 0;
579 if (blk_rq_sectors(req) + bio_sectors(bio) >
580 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
581 req_set_nomerge(req->q, req);
582 return 0;
583 }
584
585 return ll_new_hw_segment(req, bio, nr_segs);
586}
587
588int ll_front_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
589{
590 if (req_gap_front_merge(req, bio))
591 return 0;
592 if (blk_integrity_rq(req) &&
593 integrity_req_gap_front_merge(req, bio))
594 return 0;
595 if (blk_rq_sectors(req) + bio_sectors(bio) >
596 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
597 req_set_nomerge(req->q, req);
598 return 0;
599 }
600
601 return ll_new_hw_segment(req, bio, nr_segs);
602}
603
604static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
605 struct request *next)
606{
607 unsigned short segments = blk_rq_nr_discard_segments(req);
608
609 if (segments >= queue_max_discard_segments(q))
610 goto no_merge;
611 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
612 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
613 goto no_merge;
614
615 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
616 return true;
617no_merge:
618 req_set_nomerge(q, req);
619 return false;
620}
621
622static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
623 struct request *next)
624{
625 int total_phys_segments;
626
627 if (req_gap_back_merge(req, next->bio))
628 return 0;
629
630 /*
631 * Will it become too large?
632 */
633 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
634 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
635 return 0;
636
637 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
638 if (total_phys_segments > queue_max_segments(q))
639 return 0;
640
641 if (blk_integrity_merge_rq(q, req, next) == false)
642 return 0;
643
644 /* Merge is OK... */
645 req->nr_phys_segments = total_phys_segments;
646 return 1;
647}
648
649/**
650 * blk_rq_set_mixed_merge - mark a request as mixed merge
651 * @rq: request to mark as mixed merge
652 *
653 * Description:
654 * @rq is about to be mixed merged. Make sure the attributes
655 * which can be mixed are set in each bio and mark @rq as mixed
656 * merged.
657 */
658void blk_rq_set_mixed_merge(struct request *rq)
659{
660 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
661 struct bio *bio;
662
663 if (rq->rq_flags & RQF_MIXED_MERGE)
664 return;
665
666 /*
667 * @rq will no longer represent mixable attributes for all the
668 * contained bios. It will just track those of the first one.
669 * Distributes the attributs to each bio.
670 */
671 for (bio = rq->bio; bio; bio = bio->bi_next) {
672 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
673 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
674 bio->bi_opf |= ff;
675 }
676 rq->rq_flags |= RQF_MIXED_MERGE;
677}
678
679static void blk_account_io_merge(struct request *req)
680{
681 if (blk_do_io_stat(req)) {
682 struct hd_struct *part;
683
684 part_stat_lock();
685 part = req->part;
686
687 part_dec_in_flight(req->q, part, rq_data_dir(req));
688
689 hd_struct_put(part);
690 part_stat_unlock();
691 }
692}
693/*
694 * Two cases of handling DISCARD merge:
695 * If max_discard_segments > 1, the driver takes every bio
696 * as a range and send them to controller together. The ranges
697 * needn't to be contiguous.
698 * Otherwise, the bios/requests will be handled as same as
699 * others which should be contiguous.
700 */
701static inline bool blk_discard_mergable(struct request *req)
702{
703 if (req_op(req) == REQ_OP_DISCARD &&
704 queue_max_discard_segments(req->q) > 1)
705 return true;
706 return false;
707}
708
709static enum elv_merge blk_try_req_merge(struct request *req,
710 struct request *next)
711{
712 if (blk_discard_mergable(req))
713 return ELEVATOR_DISCARD_MERGE;
714 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
715 return ELEVATOR_BACK_MERGE;
716
717 return ELEVATOR_NO_MERGE;
718}
719
720/*
721 * For non-mq, this has to be called with the request spinlock acquired.
722 * For mq with scheduling, the appropriate queue wide lock should be held.
723 */
724static struct request *attempt_merge(struct request_queue *q,
725 struct request *req, struct request *next)
726{
727 if (!rq_mergeable(req) || !rq_mergeable(next))
728 return NULL;
729
730 if (req_op(req) != req_op(next))
731 return NULL;
732
733 if (rq_data_dir(req) != rq_data_dir(next)
734 || req->rq_disk != next->rq_disk)
735 return NULL;
736
737 if (req_op(req) == REQ_OP_WRITE_SAME &&
738 !blk_write_same_mergeable(req->bio, next->bio))
739 return NULL;
740
741 /*
742 * Don't allow merge of different write hints, or for a hint with
743 * non-hint IO.
744 */
745 if (req->write_hint != next->write_hint)
746 return NULL;
747
748 if (req->ioprio != next->ioprio)
749 return NULL;
750
751 /*
752 * If we are allowed to merge, then append bio list
753 * from next to rq and release next. merge_requests_fn
754 * will have updated segment counts, update sector
755 * counts here. Handle DISCARDs separately, as they
756 * have separate settings.
757 */
758
759 switch (blk_try_req_merge(req, next)) {
760 case ELEVATOR_DISCARD_MERGE:
761 if (!req_attempt_discard_merge(q, req, next))
762 return NULL;
763 break;
764 case ELEVATOR_BACK_MERGE:
765 if (!ll_merge_requests_fn(q, req, next))
766 return NULL;
767 break;
768 default:
769 return NULL;
770 }
771
772 /*
773 * If failfast settings disagree or any of the two is already
774 * a mixed merge, mark both as mixed before proceeding. This
775 * makes sure that all involved bios have mixable attributes
776 * set properly.
777 */
778 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
779 (req->cmd_flags & REQ_FAILFAST_MASK) !=
780 (next->cmd_flags & REQ_FAILFAST_MASK)) {
781 blk_rq_set_mixed_merge(req);
782 blk_rq_set_mixed_merge(next);
783 }
784
785 /*
786 * At this point we have either done a back merge or front merge. We
787 * need the smaller start_time_ns of the merged requests to be the
788 * current request for accounting purposes.
789 */
790 if (next->start_time_ns < req->start_time_ns)
791 req->start_time_ns = next->start_time_ns;
792
793 req->biotail->bi_next = next->bio;
794 req->biotail = next->biotail;
795
796 req->__data_len += blk_rq_bytes(next);
797
798 if (!blk_discard_mergable(req))
799 elv_merge_requests(q, req, next);
800
801 /*
802 * 'next' is going away, so update stats accordingly
803 */
804 blk_account_io_merge(next);
805
806 /*
807 * ownership of bio passed from next to req, return 'next' for
808 * the caller to free
809 */
810 next->bio = NULL;
811 return next;
812}
813
814struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
815{
816 struct request *next = elv_latter_request(q, rq);
817
818 if (next)
819 return attempt_merge(q, rq, next);
820
821 return NULL;
822}
823
824struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
825{
826 struct request *prev = elv_former_request(q, rq);
827
828 if (prev)
829 return attempt_merge(q, prev, rq);
830
831 return NULL;
832}
833
834int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
835 struct request *next)
836{
837 struct request *free;
838
839 free = attempt_merge(q, rq, next);
840 if (free) {
841 blk_put_request(free);
842 return 1;
843 }
844
845 return 0;
846}
847
848bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
849{
850 if (!rq_mergeable(rq) || !bio_mergeable(bio))
851 return false;
852
853 if (req_op(rq) != bio_op(bio))
854 return false;
855
856 /* different data direction or already started, don't merge */
857 if (bio_data_dir(bio) != rq_data_dir(rq))
858 return false;
859
860 /* must be same device */
861 if (rq->rq_disk != bio->bi_disk)
862 return false;
863
864 /* only merge integrity protected bio into ditto rq */
865 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
866 return false;
867
868 /* must be using the same buffer */
869 if (req_op(rq) == REQ_OP_WRITE_SAME &&
870 !blk_write_same_mergeable(rq->bio, bio))
871 return false;
872
873 /*
874 * Don't allow merge of different write hints, or for a hint with
875 * non-hint IO.
876 */
877 if (rq->write_hint != bio->bi_write_hint)
878 return false;
879
880 if (rq->ioprio != bio_prio(bio))
881 return false;
882
883 return true;
884}
885
886enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
887{
888 if (blk_discard_mergable(rq))
889 return ELEVATOR_DISCARD_MERGE;
890 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
891 return ELEVATOR_BACK_MERGE;
892 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
893 return ELEVATOR_FRONT_MERGE;
894 return ELEVATOR_NO_MERGE;
895}