xref: /openbmc/linux/block/blk-merge.c (revision 455f9726)
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 
12 static 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, high, highprv = 1, no_sg_merge;
17 	unsigned int seg_size, nr_phys_segs;
18 	struct bio *fbio, *bbio;
19 	struct bvec_iter iter;
20 
21 	if (!bio)
22 		return 0;
23 
24 	/*
25 	 * This should probably be returning 0, but blk_add_request_payload()
26 	 * (Christoph!!!!)
27 	 */
28 	if (bio->bi_rw & REQ_DISCARD)
29 		return 1;
30 
31 	if (bio->bi_rw & REQ_WRITE_SAME)
32 		return 1;
33 
34 	fbio = bio;
35 	cluster = blk_queue_cluster(q);
36 	seg_size = 0;
37 	nr_phys_segs = 0;
38 	no_sg_merge = test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags);
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 			}
68 new_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 
89 void blk_recalc_rq_segments(struct request *rq)
90 {
91 	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
92 }
93 
94 void blk_recount_segments(struct request_queue *q, struct bio *bio)
95 {
96 	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags))
97 		bio->bi_phys_segments = bio->bi_vcnt;
98 	else {
99 		struct bio *nxt = bio->bi_next;
100 
101 		bio->bi_next = NULL;
102 		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
103 		bio->bi_next = nxt;
104 	}
105 
106 	bio->bi_flags |= (1 << BIO_SEG_VALID);
107 }
108 EXPORT_SYMBOL(blk_recount_segments);
109 
110 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
111 				   struct bio *nxt)
112 {
113 	struct bio_vec end_bv = { NULL }, nxt_bv;
114 	struct bvec_iter iter;
115 
116 	if (!blk_queue_cluster(q))
117 		return 0;
118 
119 	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
120 	    queue_max_segment_size(q))
121 		return 0;
122 
123 	if (!bio_has_data(bio))
124 		return 1;
125 
126 	bio_for_each_segment(end_bv, bio, iter)
127 		if (end_bv.bv_len == iter.bi_size)
128 			break;
129 
130 	nxt_bv = bio_iovec(nxt);
131 
132 	if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
133 		return 0;
134 
135 	/*
136 	 * bio and nxt are contiguous in memory; check if the queue allows
137 	 * these two to be merged into one
138 	 */
139 	if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
140 		return 1;
141 
142 	return 0;
143 }
144 
145 static inline void
146 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
147 		     struct scatterlist *sglist, struct bio_vec *bvprv,
148 		     struct scatterlist **sg, int *nsegs, int *cluster)
149 {
150 
151 	int nbytes = bvec->bv_len;
152 
153 	if (*sg && *cluster) {
154 		if ((*sg)->length + nbytes > queue_max_segment_size(q))
155 			goto new_segment;
156 
157 		if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
158 			goto new_segment;
159 		if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
160 			goto new_segment;
161 
162 		(*sg)->length += nbytes;
163 	} else {
164 new_segment:
165 		if (!*sg)
166 			*sg = sglist;
167 		else {
168 			/*
169 			 * If the driver previously mapped a shorter
170 			 * list, we could see a termination bit
171 			 * prematurely unless it fully inits the sg
172 			 * table on each mapping. We KNOW that there
173 			 * must be more entries here or the driver
174 			 * would be buggy, so force clear the
175 			 * termination bit to avoid doing a full
176 			 * sg_init_table() in drivers for each command.
177 			 */
178 			sg_unmark_end(*sg);
179 			*sg = sg_next(*sg);
180 		}
181 
182 		sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
183 		(*nsegs)++;
184 	}
185 	*bvprv = *bvec;
186 }
187 
188 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
189 			     struct scatterlist *sglist,
190 			     struct scatterlist **sg)
191 {
192 	struct bio_vec bvec, bvprv = { NULL };
193 	struct bvec_iter iter;
194 	int nsegs, cluster;
195 
196 	nsegs = 0;
197 	cluster = blk_queue_cluster(q);
198 
199 	if (bio->bi_rw & REQ_DISCARD) {
200 		/*
201 		 * This is a hack - drivers should be neither modifying the
202 		 * biovec, nor relying on bi_vcnt - but because of
203 		 * blk_add_request_payload(), a discard bio may or may not have
204 		 * a payload we need to set up here (thank you Christoph) and
205 		 * bi_vcnt is really the only way of telling if we need to.
206 		 */
207 
208 		if (bio->bi_vcnt)
209 			goto single_segment;
210 
211 		return 0;
212 	}
213 
214 	if (bio->bi_rw & REQ_WRITE_SAME) {
215 single_segment:
216 		*sg = sglist;
217 		bvec = bio_iovec(bio);
218 		sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
219 		return 1;
220 	}
221 
222 	for_each_bio(bio)
223 		bio_for_each_segment(bvec, bio, iter)
224 			__blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
225 					     &nsegs, &cluster);
226 
227 	return nsegs;
228 }
229 
230 /*
231  * map a request to scatterlist, return number of sg entries setup. Caller
232  * must make sure sg can hold rq->nr_phys_segments entries
233  */
234 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
235 		  struct scatterlist *sglist)
236 {
237 	struct scatterlist *sg = NULL;
238 	int nsegs = 0;
239 
240 	if (rq->bio)
241 		nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
242 
243 	if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
244 	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
245 		unsigned int pad_len =
246 			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
247 
248 		sg->length += pad_len;
249 		rq->extra_len += pad_len;
250 	}
251 
252 	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
253 		if (rq->cmd_flags & REQ_WRITE)
254 			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
255 
256 		sg->page_link &= ~0x02;
257 		sg = sg_next(sg);
258 		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
259 			    q->dma_drain_size,
260 			    ((unsigned long)q->dma_drain_buffer) &
261 			    (PAGE_SIZE - 1));
262 		nsegs++;
263 		rq->extra_len += q->dma_drain_size;
264 	}
265 
266 	if (sg)
267 		sg_mark_end(sg);
268 
269 	return nsegs;
270 }
271 EXPORT_SYMBOL(blk_rq_map_sg);
272 
273 /**
274  * blk_bio_map_sg - map a bio to a scatterlist
275  * @q: request_queue in question
276  * @bio: bio being mapped
277  * @sglist: scatterlist being mapped
278  *
279  * Note:
280  *    Caller must make sure sg can hold bio->bi_phys_segments entries
281  *
282  * Will return the number of sg entries setup
283  */
284 int blk_bio_map_sg(struct request_queue *q, struct bio *bio,
285 		   struct scatterlist *sglist)
286 {
287 	struct scatterlist *sg = NULL;
288 	int nsegs;
289 	struct bio *next = bio->bi_next;
290 	bio->bi_next = NULL;
291 
292 	nsegs = __blk_bios_map_sg(q, bio, sglist, &sg);
293 	bio->bi_next = next;
294 	if (sg)
295 		sg_mark_end(sg);
296 
297 	BUG_ON(bio->bi_phys_segments && nsegs > bio->bi_phys_segments);
298 	return nsegs;
299 }
300 EXPORT_SYMBOL(blk_bio_map_sg);
301 
302 static inline int ll_new_hw_segment(struct request_queue *q,
303 				    struct request *req,
304 				    struct bio *bio)
305 {
306 	int nr_phys_segs = bio_phys_segments(q, bio);
307 
308 	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
309 		goto no_merge;
310 
311 	if (bio_integrity(bio) && blk_integrity_merge_bio(q, req, bio))
312 		goto no_merge;
313 
314 	/*
315 	 * This will form the start of a new hw segment.  Bump both
316 	 * counters.
317 	 */
318 	req->nr_phys_segments += nr_phys_segs;
319 	return 1;
320 
321 no_merge:
322 	req->cmd_flags |= REQ_NOMERGE;
323 	if (req == q->last_merge)
324 		q->last_merge = NULL;
325 	return 0;
326 }
327 
328 int ll_back_merge_fn(struct request_queue *q, struct request *req,
329 		     struct bio *bio)
330 {
331 	if (blk_rq_sectors(req) + bio_sectors(bio) >
332 	    blk_rq_get_max_sectors(req)) {
333 		req->cmd_flags |= REQ_NOMERGE;
334 		if (req == q->last_merge)
335 			q->last_merge = NULL;
336 		return 0;
337 	}
338 	if (!bio_flagged(req->biotail, BIO_SEG_VALID))
339 		blk_recount_segments(q, req->biotail);
340 	if (!bio_flagged(bio, BIO_SEG_VALID))
341 		blk_recount_segments(q, bio);
342 
343 	return ll_new_hw_segment(q, req, bio);
344 }
345 
346 int ll_front_merge_fn(struct request_queue *q, struct request *req,
347 		      struct bio *bio)
348 {
349 	if (blk_rq_sectors(req) + bio_sectors(bio) >
350 	    blk_rq_get_max_sectors(req)) {
351 		req->cmd_flags |= REQ_NOMERGE;
352 		if (req == q->last_merge)
353 			q->last_merge = NULL;
354 		return 0;
355 	}
356 	if (!bio_flagged(bio, BIO_SEG_VALID))
357 		blk_recount_segments(q, bio);
358 	if (!bio_flagged(req->bio, BIO_SEG_VALID))
359 		blk_recount_segments(q, req->bio);
360 
361 	return ll_new_hw_segment(q, req, bio);
362 }
363 
364 /*
365  * blk-mq uses req->special to carry normal driver per-request payload, it
366  * does not indicate a prepared command that we cannot merge with.
367  */
368 static bool req_no_special_merge(struct request *req)
369 {
370 	struct request_queue *q = req->q;
371 
372 	return !q->mq_ops && req->special;
373 }
374 
375 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
376 				struct request *next)
377 {
378 	int total_phys_segments;
379 	unsigned int seg_size =
380 		req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
381 
382 	/*
383 	 * First check if the either of the requests are re-queued
384 	 * requests.  Can't merge them if they are.
385 	 */
386 	if (req_no_special_merge(req) || req_no_special_merge(next))
387 		return 0;
388 
389 	/*
390 	 * Will it become too large?
391 	 */
392 	if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
393 	    blk_rq_get_max_sectors(req))
394 		return 0;
395 
396 	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
397 	if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
398 		if (req->nr_phys_segments == 1)
399 			req->bio->bi_seg_front_size = seg_size;
400 		if (next->nr_phys_segments == 1)
401 			next->biotail->bi_seg_back_size = seg_size;
402 		total_phys_segments--;
403 	}
404 
405 	if (total_phys_segments > queue_max_segments(q))
406 		return 0;
407 
408 	if (blk_integrity_rq(req) && blk_integrity_merge_rq(q, req, next))
409 		return 0;
410 
411 	/* Merge is OK... */
412 	req->nr_phys_segments = total_phys_segments;
413 	return 1;
414 }
415 
416 /**
417  * blk_rq_set_mixed_merge - mark a request as mixed merge
418  * @rq: request to mark as mixed merge
419  *
420  * Description:
421  *     @rq is about to be mixed merged.  Make sure the attributes
422  *     which can be mixed are set in each bio and mark @rq as mixed
423  *     merged.
424  */
425 void blk_rq_set_mixed_merge(struct request *rq)
426 {
427 	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
428 	struct bio *bio;
429 
430 	if (rq->cmd_flags & REQ_MIXED_MERGE)
431 		return;
432 
433 	/*
434 	 * @rq will no longer represent mixable attributes for all the
435 	 * contained bios.  It will just track those of the first one.
436 	 * Distributes the attributs to each bio.
437 	 */
438 	for (bio = rq->bio; bio; bio = bio->bi_next) {
439 		WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
440 			     (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
441 		bio->bi_rw |= ff;
442 	}
443 	rq->cmd_flags |= REQ_MIXED_MERGE;
444 }
445 
446 static void blk_account_io_merge(struct request *req)
447 {
448 	if (blk_do_io_stat(req)) {
449 		struct hd_struct *part;
450 		int cpu;
451 
452 		cpu = part_stat_lock();
453 		part = req->part;
454 
455 		part_round_stats(cpu, part);
456 		part_dec_in_flight(part, rq_data_dir(req));
457 
458 		hd_struct_put(part);
459 		part_stat_unlock();
460 	}
461 }
462 
463 /*
464  * Has to be called with the request spinlock acquired
465  */
466 static int attempt_merge(struct request_queue *q, struct request *req,
467 			  struct request *next)
468 {
469 	if (!rq_mergeable(req) || !rq_mergeable(next))
470 		return 0;
471 
472 	if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
473 		return 0;
474 
475 	/*
476 	 * not contiguous
477 	 */
478 	if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
479 		return 0;
480 
481 	if (rq_data_dir(req) != rq_data_dir(next)
482 	    || req->rq_disk != next->rq_disk
483 	    || req_no_special_merge(next))
484 		return 0;
485 
486 	if (req->cmd_flags & REQ_WRITE_SAME &&
487 	    !blk_write_same_mergeable(req->bio, next->bio))
488 		return 0;
489 
490 	/*
491 	 * If we are allowed to merge, then append bio list
492 	 * from next to rq and release next. merge_requests_fn
493 	 * will have updated segment counts, update sector
494 	 * counts here.
495 	 */
496 	if (!ll_merge_requests_fn(q, req, next))
497 		return 0;
498 
499 	/*
500 	 * If failfast settings disagree or any of the two is already
501 	 * a mixed merge, mark both as mixed before proceeding.  This
502 	 * makes sure that all involved bios have mixable attributes
503 	 * set properly.
504 	 */
505 	if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
506 	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
507 	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
508 		blk_rq_set_mixed_merge(req);
509 		blk_rq_set_mixed_merge(next);
510 	}
511 
512 	/*
513 	 * At this point we have either done a back merge
514 	 * or front merge. We need the smaller start_time of
515 	 * the merged requests to be the current request
516 	 * for accounting purposes.
517 	 */
518 	if (time_after(req->start_time, next->start_time))
519 		req->start_time = next->start_time;
520 
521 	req->biotail->bi_next = next->bio;
522 	req->biotail = next->biotail;
523 
524 	req->__data_len += blk_rq_bytes(next);
525 
526 	elv_merge_requests(q, req, next);
527 
528 	/*
529 	 * 'next' is going away, so update stats accordingly
530 	 */
531 	blk_account_io_merge(next);
532 
533 	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
534 	if (blk_rq_cpu_valid(next))
535 		req->cpu = next->cpu;
536 
537 	/* owner-ship of bio passed from next to req */
538 	next->bio = NULL;
539 	__blk_put_request(q, next);
540 	return 1;
541 }
542 
543 int attempt_back_merge(struct request_queue *q, struct request *rq)
544 {
545 	struct request *next = elv_latter_request(q, rq);
546 
547 	if (next)
548 		return attempt_merge(q, rq, next);
549 
550 	return 0;
551 }
552 
553 int attempt_front_merge(struct request_queue *q, struct request *rq)
554 {
555 	struct request *prev = elv_former_request(q, rq);
556 
557 	if (prev)
558 		return attempt_merge(q, prev, rq);
559 
560 	return 0;
561 }
562 
563 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
564 			  struct request *next)
565 {
566 	return attempt_merge(q, rq, next);
567 }
568 
569 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
570 {
571 	struct request_queue *q = rq->q;
572 
573 	if (!rq_mergeable(rq) || !bio_mergeable(bio))
574 		return false;
575 
576 	if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
577 		return false;
578 
579 	/* different data direction or already started, don't merge */
580 	if (bio_data_dir(bio) != rq_data_dir(rq))
581 		return false;
582 
583 	/* must be same device and not a special request */
584 	if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
585 		return false;
586 
587 	/* only merge integrity protected bio into ditto rq */
588 	if (bio_integrity(bio) != blk_integrity_rq(rq))
589 		return false;
590 
591 	/* must be using the same buffer */
592 	if (rq->cmd_flags & REQ_WRITE_SAME &&
593 	    !blk_write_same_mergeable(rq->bio, bio))
594 		return false;
595 
596 	if (q->queue_flags & (1 << QUEUE_FLAG_SG_GAPS)) {
597 		struct bio_vec *bprev;
598 
599 		bprev = &rq->biotail->bi_io_vec[bio->bi_vcnt - 1];
600 		if (bvec_gap_to_prev(bprev, bio->bi_io_vec[0].bv_offset))
601 			return false;
602 	}
603 
604 	return true;
605 }
606 
607 int blk_try_merge(struct request *rq, struct bio *bio)
608 {
609 	if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
610 		return ELEVATOR_BACK_MERGE;
611 	else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
612 		return ELEVATOR_FRONT_MERGE;
613 	return ELEVATOR_NO_MERGE;
614 }
615