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					     bool no_sg_merge)
15{
16	struct bio_vec bv, bvprv = { NULL };
17	int cluster, high, highprv = 1;
18	unsigned int seg_size, nr_phys_segs;
19	struct bio *fbio, *bbio;
20	struct bvec_iter iter;
21
22	if (!bio)
23		return 0;
24
25	/*
26	 * This should probably be returning 0, but blk_add_request_payload()
27	 * (Christoph!!!!)
28	 */
29	if (bio->bi_rw & REQ_DISCARD)
30		return 1;
31
32	if (bio->bi_rw & REQ_WRITE_SAME)
33		return 1;
34
35	fbio = bio;
36	cluster = blk_queue_cluster(q);
37	seg_size = 0;
38	nr_phys_segs = 0;
39	high = 0;
40	for_each_bio(bio) {
41		bio_for_each_segment(bv, bio, iter) {
42			/*
43			 * If SG merging is disabled, each bio vector is
44			 * a segment
45			 */
46			if (no_sg_merge)
47				goto new_segment;
48
49			/*
50			 * the trick here is making sure that a high page is
51			 * never considered part of another segment, since
52			 * that might change with the bounce page.
53			 */
54			high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
55			if (!high && !highprv && cluster) {
56				if (seg_size + bv.bv_len
57				    > queue_max_segment_size(q))
58					goto new_segment;
59				if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
60					goto new_segment;
61				if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
62					goto new_segment;
63
64				seg_size += bv.bv_len;
65				bvprv = bv;
66				continue;
67			}
68new_segment:
69			if (nr_phys_segs == 1 && seg_size >
70			    fbio->bi_seg_front_size)
71				fbio->bi_seg_front_size = seg_size;
72
73			nr_phys_segs++;
74			bvprv = bv;
75			seg_size = bv.bv_len;
76			highprv = high;
77		}
78		bbio = bio;
79	}
80
81	if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
82		fbio->bi_seg_front_size = seg_size;
83	if (seg_size > bbio->bi_seg_back_size)
84		bbio->bi_seg_back_size = seg_size;
85
86	return nr_phys_segs;
87}
88
89void blk_recalc_rq_segments(struct request *rq)
90{
91	bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
92			&rq->q->queue_flags);
93
94	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
95			no_sg_merge);
96}
97
98void blk_recount_segments(struct request_queue *q, struct bio *bio)
99{
100	unsigned short seg_cnt;
101
102	/* estimate segment number by bi_vcnt for non-cloned bio */
103	if (bio_flagged(bio, BIO_CLONED))
104		seg_cnt = bio_segments(bio);
105	else
106		seg_cnt = bio->bi_vcnt;
107
108	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
109			(seg_cnt < queue_max_segments(q)))
110		bio->bi_phys_segments = seg_cnt;
111	else {
112		struct bio *nxt = bio->bi_next;
113
114		bio->bi_next = NULL;
115		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
116		bio->bi_next = nxt;
117	}
118
119	bio->bi_flags |= (1 << BIO_SEG_VALID);
120}
121EXPORT_SYMBOL(blk_recount_segments);
122
123static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
124				   struct bio *nxt)
125{
126	struct bio_vec end_bv = { NULL }, nxt_bv;
127	struct bvec_iter iter;
128
129	if (!blk_queue_cluster(q))
130		return 0;
131
132	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
133	    queue_max_segment_size(q))
134		return 0;
135
136	if (!bio_has_data(bio))
137		return 1;
138
139	bio_for_each_segment(end_bv, bio, iter)
140		if (end_bv.bv_len == iter.bi_size)
141			break;
142
143	nxt_bv = bio_iovec(nxt);
144
145	if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
146		return 0;
147
148	/*
149	 * bio and nxt are contiguous in memory; check if the queue allows
150	 * these two to be merged into one
151	 */
152	if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
153		return 1;
154
155	return 0;
156}
157
158static inline void
159__blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
160		     struct scatterlist *sglist, struct bio_vec *bvprv,
161		     struct scatterlist **sg, int *nsegs, int *cluster)
162{
163
164	int nbytes = bvec->bv_len;
165
166	if (*sg && *cluster) {
167		if ((*sg)->length + nbytes > queue_max_segment_size(q))
168			goto new_segment;
169
170		if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
171			goto new_segment;
172		if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
173			goto new_segment;
174
175		(*sg)->length += nbytes;
176	} else {
177new_segment:
178		if (!*sg)
179			*sg = sglist;
180		else {
181			/*
182			 * If the driver previously mapped a shorter
183			 * list, we could see a termination bit
184			 * prematurely unless it fully inits the sg
185			 * table on each mapping. We KNOW that there
186			 * must be more entries here or the driver
187			 * would be buggy, so force clear the
188			 * termination bit to avoid doing a full
189			 * sg_init_table() in drivers for each command.
190			 */
191			sg_unmark_end(*sg);
192			*sg = sg_next(*sg);
193		}
194
195		sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
196		(*nsegs)++;
197	}
198	*bvprv = *bvec;
199}
200
201static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
202			     struct scatterlist *sglist,
203			     struct scatterlist **sg)
204{
205	struct bio_vec bvec, bvprv = { NULL };
206	struct bvec_iter iter;
207	int nsegs, cluster;
208
209	nsegs = 0;
210	cluster = blk_queue_cluster(q);
211
212	if (bio->bi_rw & REQ_DISCARD) {
213		/*
214		 * This is a hack - drivers should be neither modifying the
215		 * biovec, nor relying on bi_vcnt - but because of
216		 * blk_add_request_payload(), a discard bio may or may not have
217		 * a payload we need to set up here (thank you Christoph) and
218		 * bi_vcnt is really the only way of telling if we need to.
219		 */
220
221		if (bio->bi_vcnt)
222			goto single_segment;
223
224		return 0;
225	}
226
227	if (bio->bi_rw & REQ_WRITE_SAME) {
228single_segment:
229		*sg = sglist;
230		bvec = bio_iovec(bio);
231		sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
232		return 1;
233	}
234
235	for_each_bio(bio)
236		bio_for_each_segment(bvec, bio, iter)
237			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
238					     &nsegs, &cluster);
239
240	return nsegs;
241}
242
243/*
244 * map a request to scatterlist, return number of sg entries setup. Caller
245 * must make sure sg can hold rq->nr_phys_segments entries
246 */
247int blk_rq_map_sg(struct request_queue *q, struct request *rq,
248		  struct scatterlist *sglist)
249{
250	struct scatterlist *sg = NULL;
251	int nsegs = 0;
252
253	if (rq->bio)
254		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
255
256	if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
257	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
258		unsigned int pad_len =
259			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
260
261		sg->length += pad_len;
262		rq->extra_len += pad_len;
263	}
264
265	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
266		if (rq->cmd_flags & REQ_WRITE)
267			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
268
269		sg->page_link &= ~0x02;
270		sg = sg_next(sg);
271		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
272			    q->dma_drain_size,
273			    ((unsigned long)q->dma_drain_buffer) &
274			    (PAGE_SIZE - 1));
275		nsegs++;
276		rq->extra_len += q->dma_drain_size;
277	}
278
279	if (sg)
280		sg_mark_end(sg);
281
282	return nsegs;
283}
284EXPORT_SYMBOL(blk_rq_map_sg);
285
286/**
287 * blk_bio_map_sg - map a bio to a scatterlist
288 * @q: request_queue in question
289 * @bio: bio being mapped
290 * @sglist: scatterlist being mapped
291 *
292 * Note:
293 *    Caller must make sure sg can hold bio->bi_phys_segments entries
294 *
295 * Will return the number of sg entries setup
296 */
297int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
298		   struct scatterlist *sglist)
299{
300	struct scatterlist *sg = NULL;
301	int nsegs;
302	struct bio *next = bio->bi_next;
303	bio->bi_next = NULL;
304
305	nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
306	bio->bi_next = next;
307	if (sg)
308		sg_mark_end(sg);
309
310	BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
311	return nsegs;
312}
313EXPORT_SYMBOL(blk_bio_map_sg);
314
315static inline int ll_new_hw_segment(struct request_queue *q,
316				    struct request *req,
317				    struct bio *bio)
318{
319	int nr_phys_segs = bio_phys_segments(q, bio);
320
321	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
322		goto no_merge;
323
324	if (blk_integrity_merge_bio(q, req, bio) == false)
325		goto no_merge;
326
327	/*
328	 * This will form the start of a new hw segment.  Bump both
329	 * counters.
330	 */
331	req->nr_phys_segments += nr_phys_segs;
332	return 1;
333
334no_merge:
335	req->cmd_flags |= REQ_NOMERGE;
336	if (req == q->last_merge)
337		q->last_merge = NULL;
338	return 0;
339}
340
341int ll_back_merge_fn(struct request_queue *q, struct request *req,
342		     struct bio *bio)
343{
344	if (blk_rq_sectors(req) + bio_sectors(bio) >
345	    blk_rq_get_max_sectors(req)) {
346		req->cmd_flags |= REQ_NOMERGE;
347		if (req == q->last_merge)
348			q->last_merge = NULL;
349		return 0;
350	}
351	if (!bio_flagged(req->biotail, BIO_SEG_VALID))
352		blk_recount_segments(q, req->biotail);
353	if (!bio_flagged(bio, BIO_SEG_VALID))
354		blk_recount_segments(q, bio);
355
356	return ll_new_hw_segment(q, req, bio);
357}
358
359int ll_front_merge_fn(struct request_queue *q, struct request *req,
360		      struct bio *bio)
361{
362	if (blk_rq_sectors(req) + bio_sectors(bio) >
363	    blk_rq_get_max_sectors(req)) {
364		req->cmd_flags |= REQ_NOMERGE;
365		if (req == q->last_merge)
366			q->last_merge = NULL;
367		return 0;
368	}
369	if (!bio_flagged(bio, BIO_SEG_VALID))
370		blk_recount_segments(q, bio);
371	if (!bio_flagged(req->bio, BIO_SEG_VALID))
372		blk_recount_segments(q, req->bio);
373
374	return ll_new_hw_segment(q, req, bio);
375}
376
377/*
378 * blk-mq uses req->special to carry normal driver per-request payload, it
379 * does not indicate a prepared command that we cannot merge with.
380 */
381static bool req_no_special_merge(struct request *req)
382{
383	struct request_queue *q = req->q;
384
385	return !q->mq_ops && req->special;
386}
387
388static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
389				struct request *next)
390{
391	int total_phys_segments;
392	unsigned int seg_size =
393		req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
394
395	/*
396	 * First check if the either of the requests are re-queued
397	 * requests.  Can't merge them if they are.
398	 */
399	if (req_no_special_merge(req) || req_no_special_merge(next))
400		return 0;
401
402	/*
403	 * Will it become too large?
404	 */
405	if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
406	    blk_rq_get_max_sectors(req))
407		return 0;
408
409	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
410	if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
411		if (req->nr_phys_segments == 1)
412			req->bio->bi_seg_front_size = seg_size;
413		if (next->nr_phys_segments == 1)
414			next->biotail->bi_seg_back_size = seg_size;
415		total_phys_segments--;
416	}
417
418	if (total_phys_segments > queue_max_segments(q))
419		return 0;
420
421	if (blk_integrity_merge_rq(q, req, next) == false)
422		return 0;
423
424	/* Merge is OK... */
425	req->nr_phys_segments = total_phys_segments;
426	return 1;
427}
428
429/**
430 * blk_rq_set_mixed_merge - mark a request as mixed merge
431 * @rq: request to mark as mixed merge
432 *
433 * Description:
434 *     @rq is about to be mixed merged.  Make sure the attributes
435 *     which can be mixed are set in each bio and mark @rq as mixed
436 *     merged.
437 */
438void blk_rq_set_mixed_merge(struct request *rq)
439{
440	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
441	struct bio *bio;
442
443	if (rq->cmd_flags & REQ_MIXED_MERGE)
444		return;
445
446	/*
447	 * @rq will no longer represent mixable attributes for all the
448	 * contained bios.  It will just track those of the first one.
449	 * Distributes the attributs to each bio.
450	 */
451	for (bio = rq->bio; bio; bio = bio->bi_next) {
452		WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
453			     (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
454		bio->bi_rw |= ff;
455	}
456	rq->cmd_flags |= REQ_MIXED_MERGE;
457}
458
459static void blk_account_io_merge(struct request *req)
460{
461	if (blk_do_io_stat(req)) {
462		struct hd_struct *part;
463		int cpu;
464
465		cpu = part_stat_lock();
466		part = req->part;
467
468		part_round_stats(cpu, part);
469		part_dec_in_flight(part, rq_data_dir(req));
470
471		hd_struct_put(part);
472		part_stat_unlock();
473	}
474}
475
476/*
477 * Has to be called with the request spinlock acquired
478 */
479static int attempt_merge(struct request_queue *q, struct request *req,
480			  struct request *next)
481{
482	if (!rq_mergeable(req) || !rq_mergeable(next))
483		return 0;
484
485	if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
486		return 0;
487
488	/*
489	 * not contiguous
490	 */
491	if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
492		return 0;
493
494	if (rq_data_dir(req) != rq_data_dir(next)
495	    || req->rq_disk != next->rq_disk
496	    || req_no_special_merge(next))
497		return 0;
498
499	if (req->cmd_flags & REQ_WRITE_SAME &&
500	    !blk_write_same_mergeable(req->bio, next->bio))
501		return 0;
502
503	/*
504	 * If we are allowed to merge, then append bio list
505	 * from next to rq and release next. merge_requests_fn
506	 * will have updated segment counts, update sector
507	 * counts here.
508	 */
509	if (!ll_merge_requests_fn(q, req, next))
510		return 0;
511
512	/*
513	 * If failfast settings disagree or any of the two is already
514	 * a mixed merge, mark both as mixed before proceeding.  This
515	 * makes sure that all involved bios have mixable attributes
516	 * set properly.
517	 */
518	if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
519	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
520	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
521		blk_rq_set_mixed_merge(req);
522		blk_rq_set_mixed_merge(next);
523	}
524
525	/*
526	 * At this point we have either done a back merge
527	 * or front merge. We need the smaller start_time of
528	 * the merged requests to be the current request
529	 * for accounting purposes.
530	 */
531	if (time_after(req->start_time, next->start_time))
532		req->start_time = next->start_time;
533
534	req->biotail->bi_next = next->bio;
535	req->biotail = next->biotail;
536
537	req->__data_len += blk_rq_bytes(next);
538
539	elv_merge_requests(q, req, next);
540
541	/*
542	 * 'next' is going away, so update stats accordingly
543	 */
544	blk_account_io_merge(next);
545
546	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
547	if (blk_rq_cpu_valid(next))
548		req->cpu = next->cpu;
549
550	/* owner-ship of bio passed from next to req */
551	next->bio = NULL;
552	__blk_put_request(q, next);
553	return 1;
554}
555
556int attempt_back_merge(struct request_queue *q, struct request *rq)
557{
558	struct request *next = elv_latter_request(q, rq);
559
560	if (next)
561		return attempt_merge(q, rq, next);
562
563	return 0;
564}
565
566int attempt_front_merge(struct request_queue *q, struct request *rq)
567{
568	struct request *prev = elv_former_request(q, rq);
569
570	if (prev)
571		return attempt_merge(q, prev, rq);
572
573	return 0;
574}
575
576int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
577			  struct request *next)
578{
579	return attempt_merge(q, rq, next);
580}
581
582bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
583{
584	struct request_queue *q = rq->q;
585
586	if (!rq_mergeable(rq) || !bio_mergeable(bio))
587		return false;
588
589	if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
590		return false;
591
592	/* different data direction or already started, don't merge */
593	if (bio_data_dir(bio) != rq_data_dir(rq))
594		return false;
595
596	/* must be same device and not a special request */
597	if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
598		return false;
599
600	/* only merge integrity protected bio into ditto rq */
601	if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
602		return false;
603
604	/* must be using the same buffer */
605	if (rq->cmd_flags & REQ_WRITE_SAME &&
606	    !blk_write_same_mergeable(rq->bio, bio))
607		return false;
608
609	if (q->queue_flags & (1 << QUEUE_FLAG_SG_GAPS)) {
610		struct bio_vec *bprev;
611
612		bprev = &rq->biotail->bi_io_vec[bio->bi_vcnt - 1];
613		if (bvec_gap_to_prev(bprev, bio->bi_io_vec[0].bv_offset))
614			return false;
615	}
616
617	return true;
618}
619
620int blk_try_merge(struct request *rq, struct bio *bio)
621{
622	if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
623		return ELEVATOR_BACK_MERGE;
624	else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
625		return ELEVATOR_FRONT_MERGE;
626	return ELEVATOR_NO_MERGE;
627}
628