Linux Audio

Check our new training course

Loading...
v4.6
  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
 98	bio_for_each_segment(bv, bio, iter) {
 99		/*
100		 * If the queue doesn't support SG gaps and adding this
101		 * offset would create a gap, disallow it.
102		 */
103		if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
104			goto split;
105
106		if (sectors + (bv.bv_len >> 9) > max_sectors) {
107			/*
108			 * Consider this a new segment if we're splitting in
109			 * the middle of this vector.
110			 */
111			if (nsegs < queue_max_segments(q) &&
112			    sectors < max_sectors) {
113				nsegs++;
114				sectors = max_sectors;
115			}
116			if (sectors)
117				goto split;
118			/* Make this single bvec as the 1st segment */
119		}
120
121		if (bvprvp && blk_queue_cluster(q)) {
122			if (seg_size + bv.bv_len > queue_max_segment_size(q))
123				goto new_segment;
124			if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
125				goto new_segment;
126			if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
127				goto new_segment;
128
129			seg_size += bv.bv_len;
130			bvprv = bv;
131			bvprvp = &bvprv;
132			sectors += bv.bv_len >> 9;
133
134			if (nsegs == 1 && seg_size > front_seg_size)
135				front_seg_size = seg_size;
136			continue;
137		}
138new_segment:
139		if (nsegs == queue_max_segments(q))
140			goto split;
141
142		nsegs++;
143		bvprv = bv;
144		bvprvp = &bvprv;
145		seg_size = bv.bv_len;
146		sectors += bv.bv_len >> 9;
147
148		if (nsegs == 1 && seg_size > front_seg_size)
149			front_seg_size = seg_size;
150	}
151
152	do_split = false;
153split:
154	*segs = nsegs;
155
156	if (do_split) {
157		new = bio_split(bio, sectors, GFP_NOIO, bs);
158		if (new)
159			bio = new;
160	}
161
162	bio->bi_seg_front_size = front_seg_size;
163	if (seg_size > bio->bi_seg_back_size)
164		bio->bi_seg_back_size = seg_size;
165
166	return do_split ? new : NULL;
167}
168
169void blk_queue_split(struct request_queue *q, struct bio **bio,
170		     struct bio_set *bs)
171{
172	struct bio *split, *res;
173	unsigned nsegs;
174
175	if ((*bio)->bi_rw & REQ_DISCARD)
176		split = blk_bio_discard_split(q, *bio, bs, &nsegs);
177	else if ((*bio)->bi_rw & REQ_WRITE_SAME)
178		split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
179	else
180		split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
181
182	/* physical segments can be figured out during splitting */
183	res = split ? split : *bio;
184	res->bi_phys_segments = nsegs;
185	bio_set_flag(res, BIO_SEG_VALID);
186
187	if (split) {
188		/* there isn't chance to merge the splitted bio */
189		split->bi_rw |= REQ_NOMERGE;
190
191		bio_chain(split, *bio);
192		trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
193		generic_make_request(*bio);
194		*bio = split;
195	}
196}
197EXPORT_SYMBOL(blk_queue_split);
198
199static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
200					     struct bio *bio,
201					     bool no_sg_merge)
202{
203	struct bio_vec bv, bvprv = { NULL };
204	int cluster, prev = 0;
205	unsigned int seg_size, nr_phys_segs;
206	struct bio *fbio, *bbio;
207	struct bvec_iter iter;
208
209	if (!bio)
210		return 0;
211
212	/*
213	 * This should probably be returning 0, but blk_add_request_payload()
214	 * (Christoph!!!!)
215	 */
216	if (bio->bi_rw & REQ_DISCARD)
217		return 1;
218
219	if (bio->bi_rw & REQ_WRITE_SAME)
220		return 1;
221
222	fbio = bio;
223	cluster = blk_queue_cluster(q);
224	seg_size = 0;
225	nr_phys_segs = 0;
226	for_each_bio(bio) {
227		bio_for_each_segment(bv, bio, iter) {
228			/*
229			 * If SG merging is disabled, each bio vector is
230			 * a segment
 
231			 */
232			if (no_sg_merge)
 
233				goto new_segment;
234
235			if (prev && cluster) {
236				if (seg_size + bv.bv_len
237				    > queue_max_segment_size(q))
238					goto new_segment;
239				if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
240					goto new_segment;
241				if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
242					goto new_segment;
243
244				seg_size += bv.bv_len;
245				bvprv = bv;
246				continue;
247			}
248new_segment:
249			if (nr_phys_segs == 1 && seg_size >
250			    fbio->bi_seg_front_size)
251				fbio->bi_seg_front_size = seg_size;
252
253			nr_phys_segs++;
254			bvprv = bv;
255			prev = 1;
256			seg_size = bv.bv_len;
257		}
258		bbio = bio;
259	}
260
261	if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
262		fbio->bi_seg_front_size = seg_size;
263	if (seg_size > bbio->bi_seg_back_size)
264		bbio->bi_seg_back_size = seg_size;
265
266	return nr_phys_segs;
267}
268
269void blk_recalc_rq_segments(struct request *rq)
270{
271	bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
272			&rq->q->queue_flags);
273
274	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
275			no_sg_merge);
276}
277
278void blk_recount_segments(struct request_queue *q, struct bio *bio)
279{
280	unsigned short seg_cnt;
281
282	/* estimate segment number by bi_vcnt for non-cloned bio */
283	if (bio_flagged(bio, BIO_CLONED))
284		seg_cnt = bio_segments(bio);
285	else
286		seg_cnt = bio->bi_vcnt;
287
288	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
289			(seg_cnt < queue_max_segments(q)))
290		bio->bi_phys_segments = seg_cnt;
291	else {
292		struct bio *nxt = bio->bi_next;
293
294		bio->bi_next = NULL;
295		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
296		bio->bi_next = nxt;
297	}
298
299	bio_set_flag(bio, BIO_SEG_VALID);
300}
301EXPORT_SYMBOL(blk_recount_segments);
302
303static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
304				   struct bio *nxt)
305{
306	struct bio_vec end_bv = { NULL }, nxt_bv;
307
308	if (!blk_queue_cluster(q))
309		return 0;
310
311	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
312	    queue_max_segment_size(q))
313		return 0;
314
315	if (!bio_has_data(bio))
316		return 1;
317
318	bio_get_last_bvec(bio, &end_bv);
319	bio_get_first_bvec(nxt, &nxt_bv);
320
321	if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
322		return 0;
323
324	/*
325	 * bio and nxt are contiguous in memory; check if the queue allows
326	 * these two to be merged into one
327	 */
328	if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
329		return 1;
330
331	return 0;
332}
333
334static inline void
335__blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
336		     struct scatterlist *sglist, struct bio_vec *bvprv,
337		     struct scatterlist **sg, int *nsegs, int *cluster)
338{
339
340	int nbytes = bvec->bv_len;
341
342	if (*sg && *cluster) {
343		if ((*sg)->length + nbytes > queue_max_segment_size(q))
344			goto new_segment;
345
346		if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
347			goto new_segment;
348		if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
349			goto new_segment;
350
351		(*sg)->length += nbytes;
352	} else {
353new_segment:
354		if (!*sg)
355			*sg = sglist;
356		else {
357			/*
358			 * If the driver previously mapped a shorter
359			 * list, we could see a termination bit
360			 * prematurely unless it fully inits the sg
361			 * table on each mapping. We KNOW that there
362			 * must be more entries here or the driver
363			 * would be buggy, so force clear the
364			 * termination bit to avoid doing a full
365			 * sg_init_table() in drivers for each command.
366			 */
367			sg_unmark_end(*sg);
368			*sg = sg_next(*sg);
369		}
370
371		sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
372		(*nsegs)++;
373	}
374	*bvprv = *bvec;
375}
376
377static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
378			     struct scatterlist *sglist,
379			     struct scatterlist **sg)
380{
381	struct bio_vec bvec, bvprv = { NULL };
382	struct bvec_iter iter;
 
383	int nsegs, cluster;
384
385	nsegs = 0;
386	cluster = blk_queue_cluster(q);
387
388	if (bio->bi_rw & REQ_DISCARD) {
389		/*
390		 * This is a hack - drivers should be neither modifying the
391		 * biovec, nor relying on bi_vcnt - but because of
392		 * blk_add_request_payload(), a discard bio may or may not have
393		 * a payload we need to set up here (thank you Christoph) and
394		 * bi_vcnt is really the only way of telling if we need to.
395		 */
396
397		if (bio->bi_vcnt)
398			goto single_segment;
399
400		return 0;
401	}
402
403	if (bio->bi_rw & REQ_WRITE_SAME) {
404single_segment:
405		*sg = sglist;
406		bvec = bio_iovec(bio);
407		sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
408		return 1;
409	}
410
411	for_each_bio(bio)
412		bio_for_each_segment(bvec, bio, iter)
413			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
414					     &nsegs, &cluster);
415
416	return nsegs;
417}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
418
419/*
420 * map a request to scatterlist, return number of sg entries setup. Caller
421 * must make sure sg can hold rq->nr_phys_segments entries
422 */
423int blk_rq_map_sg(struct request_queue *q, struct request *rq,
424		  struct scatterlist *sglist)
425{
426	struct scatterlist *sg = NULL;
427	int nsegs = 0;
428
429	if (rq->bio)
430		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
431
432	if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
433	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
434		unsigned int pad_len =
435			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
436
437		sg->length += pad_len;
438		rq->extra_len += pad_len;
439	}
440
441	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
442		if (rq->cmd_flags & REQ_WRITE)
443			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
444
445		sg_unmark_end(sg);
446		sg = sg_next(sg);
447		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
448			    q->dma_drain_size,
449			    ((unsigned long)q->dma_drain_buffer) &
450			    (PAGE_SIZE - 1));
451		nsegs++;
452		rq->extra_len += q->dma_drain_size;
453	}
454
455	if (sg)
456		sg_mark_end(sg);
457
458	/*
459	 * Something must have been wrong if the figured number of
460	 * segment is bigger than number of req's physical segments
461	 */
462	WARN_ON(nsegs > rq->nr_phys_segments);
463
464	return nsegs;
465}
466EXPORT_SYMBOL(blk_rq_map_sg);
467
468static inline int ll_new_hw_segment(struct request_queue *q,
469				    struct request *req,
470				    struct bio *bio)
471{
472	int nr_phys_segs = bio_phys_segments(q, bio);
473
474	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
475		goto no_merge;
476
477	if (blk_integrity_merge_bio(q, req, bio) == false)
478		goto no_merge;
479
480	/*
481	 * This will form the start of a new hw segment.  Bump both
482	 * counters.
483	 */
484	req->nr_phys_segments += nr_phys_segs;
485	return 1;
486
487no_merge:
488	req->cmd_flags |= REQ_NOMERGE;
489	if (req == q->last_merge)
490		q->last_merge = NULL;
491	return 0;
492}
493
494int ll_back_merge_fn(struct request_queue *q, struct request *req,
495		     struct bio *bio)
496{
497	if (req_gap_back_merge(req, bio))
498		return 0;
499	if (blk_integrity_rq(req) &&
500	    integrity_req_gap_back_merge(req, bio))
501		return 0;
502	if (blk_rq_sectors(req) + bio_sectors(bio) >
503	    blk_rq_get_max_sectors(req)) {
 
504		req->cmd_flags |= REQ_NOMERGE;
505		if (req == q->last_merge)
506			q->last_merge = NULL;
507		return 0;
508	}
509	if (!bio_flagged(req->biotail, BIO_SEG_VALID))
510		blk_recount_segments(q, req->biotail);
511	if (!bio_flagged(bio, BIO_SEG_VALID))
512		blk_recount_segments(q, bio);
513
514	return ll_new_hw_segment(q, req, bio);
515}
516
517int ll_front_merge_fn(struct request_queue *q, struct request *req,
518		      struct bio *bio)
519{
 
520
521	if (req_gap_front_merge(req, bio))
522		return 0;
523	if (blk_integrity_rq(req) &&
524	    integrity_req_gap_front_merge(req, bio))
525		return 0;
526	if (blk_rq_sectors(req) + bio_sectors(bio) >
527	    blk_rq_get_max_sectors(req)) {
528		req->cmd_flags |= REQ_NOMERGE;
529		if (req == q->last_merge)
530			q->last_merge = NULL;
531		return 0;
532	}
533	if (!bio_flagged(bio, BIO_SEG_VALID))
534		blk_recount_segments(q, bio);
535	if (!bio_flagged(req->bio, BIO_SEG_VALID))
536		blk_recount_segments(q, req->bio);
537
538	return ll_new_hw_segment(q, req, bio);
539}
540
541/*
542 * blk-mq uses req->special to carry normal driver per-request payload, it
543 * does not indicate a prepared command that we cannot merge with.
544 */
545static bool req_no_special_merge(struct request *req)
546{
547	struct request_queue *q = req->q;
548
549	return !q->mq_ops && req->special;
550}
551
552static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
553				struct request *next)
554{
555	int total_phys_segments;
556	unsigned int seg_size =
557		req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
558
559	/*
560	 * First check if the either of the requests are re-queued
561	 * requests.  Can't merge them if they are.
562	 */
563	if (req_no_special_merge(req) || req_no_special_merge(next))
564		return 0;
565
566	if (req_gap_back_merge(req, next->bio))
567		return 0;
568
569	/*
570	 * Will it become too large?
571	 */
572	if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
573	    blk_rq_get_max_sectors(req))
574		return 0;
575
576	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
577	if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
578		if (req->nr_phys_segments == 1)
579			req->bio->bi_seg_front_size = seg_size;
580		if (next->nr_phys_segments == 1)
581			next->biotail->bi_seg_back_size = seg_size;
582		total_phys_segments--;
583	}
584
585	if (total_phys_segments > queue_max_segments(q))
586		return 0;
587
588	if (blk_integrity_merge_rq(q, req, next) == false)
589		return 0;
590
591	/* Merge is OK... */
592	req->nr_phys_segments = total_phys_segments;
593	return 1;
594}
595
596/**
597 * blk_rq_set_mixed_merge - mark a request as mixed merge
598 * @rq: request to mark as mixed merge
599 *
600 * Description:
601 *     @rq is about to be mixed merged.  Make sure the attributes
602 *     which can be mixed are set in each bio and mark @rq as mixed
603 *     merged.
604 */
605void blk_rq_set_mixed_merge(struct request *rq)
606{
607	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
608	struct bio *bio;
609
610	if (rq->cmd_flags & REQ_MIXED_MERGE)
611		return;
612
613	/*
614	 * @rq will no longer represent mixable attributes for all the
615	 * contained bios.  It will just track those of the first one.
616	 * Distributes the attributs to each bio.
617	 */
618	for (bio = rq->bio; bio; bio = bio->bi_next) {
619		WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
620			     (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
621		bio->bi_rw |= ff;
622	}
623	rq->cmd_flags |= REQ_MIXED_MERGE;
624}
625
626static void blk_account_io_merge(struct request *req)
627{
628	if (blk_do_io_stat(req)) {
629		struct hd_struct *part;
630		int cpu;
631
632		cpu = part_stat_lock();
633		part = req->part;
634
635		part_round_stats(cpu, part);
636		part_dec_in_flight(part, rq_data_dir(req));
637
638		hd_struct_put(part);
639		part_stat_unlock();
640	}
641}
642
643/*
644 * Has to be called with the request spinlock acquired
645 */
646static int attempt_merge(struct request_queue *q, struct request *req,
647			  struct request *next)
648{
649	if (!rq_mergeable(req) || !rq_mergeable(next))
650		return 0;
651
652	if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
 
 
 
 
 
 
 
 
 
653		return 0;
654
655	/*
656	 * not contiguous
657	 */
658	if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
659		return 0;
660
661	if (rq_data_dir(req) != rq_data_dir(next)
662	    || req->rq_disk != next->rq_disk
663	    || req_no_special_merge(next))
664		return 0;
665
666	if (req->cmd_flags & REQ_WRITE_SAME &&
667	    !blk_write_same_mergeable(req->bio, next->bio))
668		return 0;
669
670	/*
671	 * If we are allowed to merge, then append bio list
672	 * from next to rq and release next. merge_requests_fn
673	 * will have updated segment counts, update sector
674	 * counts here.
675	 */
676	if (!ll_merge_requests_fn(q, req, next))
677		return 0;
678
679	/*
680	 * If failfast settings disagree or any of the two is already
681	 * a mixed merge, mark both as mixed before proceeding.  This
682	 * makes sure that all involved bios have mixable attributes
683	 * set properly.
684	 */
685	if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
686	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
687	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
688		blk_rq_set_mixed_merge(req);
689		blk_rq_set_mixed_merge(next);
690	}
691
692	/*
693	 * At this point we have either done a back merge
694	 * or front merge. We need the smaller start_time of
695	 * the merged requests to be the current request
696	 * for accounting purposes.
697	 */
698	if (time_after(req->start_time, next->start_time))
699		req->start_time = next->start_time;
700
701	req->biotail->bi_next = next->bio;
702	req->biotail = next->biotail;
703
704	req->__data_len += blk_rq_bytes(next);
705
706	elv_merge_requests(q, req, next);
707
708	/*
709	 * 'next' is going away, so update stats accordingly
710	 */
711	blk_account_io_merge(next);
712
713	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
714	if (blk_rq_cpu_valid(next))
715		req->cpu = next->cpu;
716
717	/* owner-ship of bio passed from next to req */
718	next->bio = NULL;
719	__blk_put_request(q, next);
720	return 1;
721}
722
723int attempt_back_merge(struct request_queue *q, struct request *rq)
724{
725	struct request *next = elv_latter_request(q, rq);
726
727	if (next)
728		return attempt_merge(q, rq, next);
729
730	return 0;
731}
732
733int attempt_front_merge(struct request_queue *q, struct request *rq)
734{
735	struct request *prev = elv_former_request(q, rq);
736
737	if (prev)
738		return attempt_merge(q, prev, rq);
739
740	return 0;
741}
742
743int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
744			  struct request *next)
745{
746	return attempt_merge(q, rq, next);
747}
748
749bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
750{
751	if (!rq_mergeable(rq) || !bio_mergeable(bio))
752		return false;
753
754	if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
 
 
 
 
 
755		return false;
756
757	/* different data direction or already started, don't merge */
758	if (bio_data_dir(bio) != rq_data_dir(rq))
759		return false;
760
761	/* must be same device and not a special request */
762	if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
763		return false;
764
765	/* only merge integrity protected bio into ditto rq */
766	if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
767		return false;
768
769	/* must be using the same buffer */
770	if (rq->cmd_flags & REQ_WRITE_SAME &&
771	    !blk_write_same_mergeable(rq->bio, bio))
772		return false;
773
774	return true;
775}
776
777int blk_try_merge(struct request *rq, struct bio *bio)
778{
779	if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
780		return ELEVATOR_BACK_MERGE;
781	else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
782		return ELEVATOR_FRONT_MERGE;
783	return ELEVATOR_NO_MERGE;
784}
v3.5.6
  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}