/block/blk-merge.c

  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	unsigned int phys_size;
 16	struct bio_vec *bv, *bvprv = NULL;
 17	int cluster, i, high, highprv = 1;
 18	unsigned int seg_size, nr_phys_segs;
 19	struct bio *fbio, *bbio;
 20
 21	if (!bio)
 22		return 0;
 23
 24	fbio = bio;
 25	cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
 26	seg_size = 0;
 27	phys_size = nr_phys_segs = 0;
 28	for_each_bio(bio) {
 29		bio_for_each_segment(bv, bio, i) {
 30			/*
 31			 * the trick here is making sure that a high page is
 32			 * never considered part of another segment, since that
 33			 * might change with the bounce page.
 34			 */
 35			high = page_to_pfn(bv->bv_page) > queue_bounce_pfn(q);
 36			if (high || highprv)
 37				goto new_segment;
 38			if (cluster) {
 39				if (seg_size + bv->bv_len
 40				    > queue_max_segment_size(q))
 41					goto new_segment;
 42				if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
 43					goto new_segment;
 44				if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
 45					goto new_segment;
 46
 47				seg_size += bv->bv_len;
 48				bvprv = bv;
 49				continue;
 50			}
 51new_segment:
 52			if (nr_phys_segs == 1 && seg_size >
 53			    fbio->bi_seg_front_size)
 54				fbio->bi_seg_front_size = seg_size;
 55
 56			nr_phys_segs++;
 57			bvprv = bv;
 58			seg_size = bv->bv_len;
 59			highprv = high;
 60		}
 61		bbio = bio;
 62	}
 63
 64	if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
 65		fbio->bi_seg_front_size = seg_size;
 66	if (seg_size > bbio->bi_seg_back_size)
 67		bbio->bi_seg_back_size = seg_size;
 68
 69	return nr_phys_segs;
 70}
 71
 72void blk_recalc_rq_segments(struct request *rq)
 73{
 74	rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
 75}
 76
 77void blk_recount_segments(struct request_queue *q, struct bio *bio)
 78{
 79	struct bio *nxt = bio->bi_next;
 80
 81	bio->bi_next = NULL;
 82	bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
 83	bio->bi_next = nxt;
 84	bio->bi_flags |= (1 << BIO_SEG_VALID);
 85}
 86EXPORT_SYMBOL(blk_recount_segments);
 87
 88static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
 89				   struct bio *nxt)
 90{
 91	if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
 92		return 0;
 93
 94	if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
 95	    queue_max_segment_size(q))
 96		return 0;
 97
 98	if (!bio_has_data(bio))
 99		return 1;
100
101	if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
102		return 0;
103
104	/*
105	 * bio and nxt are contiguous in memory; check if the queue allows
106	 * these two to be merged into one
107	 */
108	if (BIO_SEG_BOUNDARY(q, bio, nxt))
109		return 1;
110
111	return 0;
112}
113
114/*
115 * map a request to scatterlist, return number of sg entries setup. Caller
116 * must make sure sg can hold rq->nr_phys_segments entries
117 */
118int blk_rq_map_sg(struct request_queue *q, struct request *rq,
119		  struct scatterlist *sglist)
120{
121	struct bio_vec *bvec, *bvprv;
122	struct req_iterator iter;
123	struct scatterlist *sg;
124	int nsegs, cluster;
125
126	nsegs = 0;
127	cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
128
129	/*
130	 * for each bio in rq
131	 */
132	bvprv = NULL;
133	sg = NULL;
134	rq_for_each_segment(bvec, rq, iter) {
135		int nbytes = bvec->bv_len;
136
137		if (bvprv && cluster) {
138			if (sg->length + nbytes > queue_max_segment_size(q))
139				goto new_segment;
140
141			if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
142				goto new_segment;
143			if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
144				goto new_segment;
145
146			sg->length += nbytes;
147		} else {
148new_segment:
149			if (!sg)
150				sg = sglist;
151			else {
152				/*
153				 * If the driver previously mapped a shorter
154				 * list, we could see a termination bit
155				 * prematurely unless it fully inits the sg
156				 * table on each mapping. We KNOW that there
157				 * must be more entries here or the driver
158				 * would be buggy, so force clear the
159				 * termination bit to avoid doing a full
160				 * sg_init_table() in drivers for each command.
161				 */
162				sg->page_link &= ~0x02;
163				sg = sg_next(sg);
164			}
165
166			sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
167			nsegs++;
168		}
169		bvprv = bvec;
170	} /* segments in rq */
171
172
173	if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
174	    (blk_rq_bytes(rq) & q->dma_pad_mask)) {
175		unsigned int pad_len =
176			(q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
177
178		sg->length += pad_len;
179		rq->extra_len += pad_len;
180	}
181
182	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
183		if (rq->cmd_flags & REQ_RW)
184			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
185
186		sg->page_link &= ~0x02;
187		sg = sg_next(sg);
188		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
189			    q->dma_drain_size,
190			    ((unsigned long)q->dma_drain_buffer) &
191			    (PAGE_SIZE - 1));
192		nsegs++;
193		rq->extra_len += q->dma_drain_size;
194	}
195
196	if (sg)
197		sg_mark_end(sg);
198
199	return nsegs;
200}
201EXPORT_SYMBOL(blk_rq_map_sg);
202
203static inline int ll_new_hw_segment(struct request_queue *q,
204				    struct request *req,
205				    struct bio *bio)
206{
207	int nr_phys_segs = bio_phys_segments(q, bio);
208
209	if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q)) {
210		req->cmd_flags |= REQ_NOMERGE;
211		if (req == q->last_merge)
212			q->last_merge = NULL;
213		return 0;
214	}
215
216	/*
217	 * This will form the start of a new hw segment.  Bump both
218	 * counters.
219	 */
220	req->nr_phys_segments += nr_phys_segs;
221	return 1;
222}
223
224int ll_back_merge_fn(struct request_queue *q, struct request *req,
225		     struct bio *bio)
226{
227	unsigned short max_sectors;
228
229	if (unlikely(blk_pc_request(req)))
230		max_sectors = queue_max_hw_sectors(q);
231	else
232		max_sectors = queue_max_sectors(q);
233
234	if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
235		req->cmd_flags |= REQ_NOMERGE;
236		if (req == q->last_merge)
237			q->last_merge = NULL;
238		return 0;
239	}
240	if (!bio_flagged(req->biotail, BIO_SEG_VALID))
241		blk_recount_segments(q, req->biotail);
242	if (!bio_flagged(bio, BIO_SEG_VALID))
243		blk_recount_segments(q, bio);
244
245	return ll_new_hw_segment(q, req, bio);
246}
247
248int ll_front_merge_fn(struct request_queue *q, struct request *req,
249		      struct bio *bio)
250{
251	unsigned short max_sectors;
252
253	if (unlikely(blk_pc_request(req)))
254		max_sectors = queue_max_hw_sectors(q);
255	else
256		max_sectors = queue_max_sectors(q);
257
258
259	if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
260		req->cmd_flags |= REQ_NOMERGE;
261		if (req == q->last_merge)
262			q->last_merge = NULL;
263		return 0;
264	}
265	if (!bio_flagged(bio, BIO_SEG_VALID))
266		blk_recount_segments(q, bio);
267	if (!bio_flagged(req->bio, BIO_SEG_VALID))
268		blk_recount_segments(q, req->bio);
269
270	return ll_new_hw_segment(q, req, bio);
271}
272
273static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
274				struct request *next)
275{
276	int total_phys_segments;
277	unsigned int seg_size =
278		req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
279
280	/*
281	 * First check if the either of the requests are re-queued
282	 * requests.  Can't merge them if they are.
283	 */
284	if (req->special || next->special)
285		return 0;
286
287	/*
288	 * Will it become too large?
289	 */
290	if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > queue_max_sectors(q))
291		return 0;
292
293	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
294	if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
295		if (req->nr_phys_segments == 1)
296			req->bio->bi_seg_front_size = seg_size;
297		if (next->nr_phys_segments == 1)
298			next->biotail->bi_seg_back_size = seg_size;
299		total_phys_segments--;
300	}
301
302	if (total_phys_segments > queue_max_segments(q))
303		return 0;
304
305	/* Merge is OK... */
306	req->nr_phys_segments = total_phys_segments;
307	return 1;
308}
309
310/**
311 * blk_rq_set_mixed_merge - mark a request as mixed merge
312 * @rq: request to mark as mixed merge
313 *
314 * Description:
315 *     @rq is about to be mixed merged.  Make sure the attributes
316 *     which can be mixed are set in each bio and mark @rq as mixed
317 *     merged.
318 */
319void blk_rq_set_mixed_merge(struct request *rq)
320{
321	unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
322	struct bio *bio;
323
324	if (rq->cmd_flags & REQ_MIXED_MERGE)
325		return;
326
327	/*
328	 * @rq will no longer represent mixable attributes for all the
329	 * contained bios.  It will just track those of the first one.
330	 * Distributes the attributs to each bio.
331	 */
332	for (bio = rq->bio; bio; bio = bio->bi_next) {
333		WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
334			     (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
335		bio->bi_rw |= ff;
336	}
337	rq->cmd_flags |= REQ_MIXED_MERGE;
338}
339
340static void blk_account_io_merge(struct request *req)
341{
342	if (blk_do_io_stat(req)) {
343		struct hd_struct *part;
344		int cpu;
345
346		cpu = part_stat_lock();
347		part = disk_map_sector_rcu(req->rq_disk, blk_rq_pos(req));
348
349		part_round_stats(cpu, part);
350		part_dec_in_flight(part, rq_data_dir(req));
351
352		part_stat_unlock();
353	}
354}
355
356/*
357 * Has to be called with the request spinlock acquired
358 */
359static int attempt_merge(struct request_queue *q, struct request *req,
360			  struct request *next)
361{
362	if (!rq_mergeable(req) || !rq_mergeable(next))
363		return 0;
364
365	/*
366	 * not contiguous
367	 */
368	if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
369		return 0;
370
371	if (rq_data_dir(req) != rq_data_dir(next)
372	    || req->rq_disk != next->rq_disk
373	    || next->special)
374		return 0;
375
376	if (blk_integrity_rq(req) != blk_integrity_rq(next))
377		return 0;
378
379	/*
380	 * If we are allowed to merge, then append bio list
381	 * from next to rq and release next. merge_requests_fn
382	 * will have updated segment counts, update sector
383	 * counts here.
384	 */
385	if (!ll_merge_requests_fn(q, req, next))
386		return 0;
387
388	/*
389	 * If failfast settings disagree or any of the two is already
390	 * a mixed merge, mark both as mixed before proceeding.  This
391	 * makes sure that all involved bios have mixable attributes
392	 * set properly.
393	 */
394	if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
395	    (req->cmd_flags & REQ_FAILFAST_MASK) !=
396	    (next->cmd_flags & REQ_FAILFAST_MASK)) {
397		blk_rq_set_mixed_merge(req);
398		blk_rq_set_mixed_merge(next);
399	}
400
401	/*
402	 * At this point we have either done a back merge
403	 * or front merge. We need the smaller start_time of
404	 * the merged requests to be the current request
405	 * for accounting purposes.
406	 */
407	if (time_after(req->start_time, next->start_time))
408		req->start_time = next->start_time;
409
410	req->biotail->bi_next = next->bio;
411	req->biotail = next->biotail;
412
413	req->__data_len += blk_rq_bytes(next);
414
415	elv_merge_requests(q, req, next);
416
417	/*
418	 * 'next' is going away, so update stats accordingly
419	 */
420	blk_account_io_merge(next);
421
422	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
423	if (blk_rq_cpu_valid(next))
424		req->cpu = next->cpu;
425
426	/* owner-ship of bio passed from next to req */
427	next->bio = NULL;
428	__blk_put_request(q, next);
429	return 1;
430}
431
432int attempt_back_merge(struct request_queue *q, struct request *rq)
433{
434	struct request *next = elv_latter_request(q, rq);
435
436	if (next)
437		return attempt_merge(q, rq, next);
438
439	return 0;
440}
441
442int attempt_front_merge(struct request_queue *q, struct request *rq)
443{
444	struct request *prev = elv_former_request(q, rq);
445
446	if (prev)
447		return attempt_merge(q, prev, rq);
448
449	return 0;
450}