blob: b0de8574fdc84e40e1bdf87dea4eb1af98769bb4 [file] [log] [blame]
Jens Axboed6d48192008-01-29 14:04:06 +01001/*
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
Jens Axboe1e428072009-02-23 09:03:10 +010012static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
Jens Axboe59247ea2009-03-06 08:55:24 +010013 struct bio *bio)
Jens Axboed6d48192008-01-29 14:04:06 +010014{
Jens Axboed6d48192008-01-29 14:04:06 +010015 unsigned int phys_size;
Jens Axboed6d48192008-01-29 14:04:06 +010016 struct bio_vec *bv, *bvprv = NULL;
Jens Axboe1e428072009-02-23 09:03:10 +010017 int cluster, i, high, highprv = 1;
18 unsigned int seg_size, nr_phys_segs;
Jens Axboe59247ea2009-03-06 08:55:24 +010019 struct bio *fbio, *bbio;
Jens Axboed6d48192008-01-29 14:04:06 +010020
Jens Axboe1e428072009-02-23 09:03:10 +010021 if (!bio)
22 return 0;
Jens Axboed6d48192008-01-29 14:04:06 +010023
Jens Axboe1e428072009-02-23 09:03:10 +010024 fbio = bio;
Nick Piggin75ad23b2008-04-29 14:48:33 +020025 cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
Mikulas Patocka5df97b92008-08-15 10:20:02 +020026 seg_size = 0;
27 phys_size = nr_phys_segs = 0;
Jens Axboe1e428072009-02-23 09:03:10 +010028 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 */
Martin K. Petersenae03bf62009-05-22 17:17:50 -040035 high = page_to_pfn(bv->bv_page) > queue_bounce_pfn(q);
Jens Axboe1e428072009-02-23 09:03:10 +010036 if (high || highprv)
Jens Axboed6d48192008-01-29 14:04:06 +010037 goto new_segment;
Jens Axboe1e428072009-02-23 09:03:10 +010038 if (cluster) {
Martin K. Petersenae03bf62009-05-22 17:17:50 -040039 if (seg_size + bv->bv_len
40 > queue_max_segment_size(q))
Jens Axboe1e428072009-02-23 09:03:10 +010041 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;
Jens Axboed6d48192008-01-29 14:04:06 +010046
Jens Axboe1e428072009-02-23 09:03:10 +010047 seg_size += bv->bv_len;
48 bvprv = bv;
49 continue;
50 }
Jens Axboed6d48192008-01-29 14:04:06 +010051new_segment:
Jens Axboe1e428072009-02-23 09:03:10 +010052 if (nr_phys_segs == 1 && seg_size >
53 fbio->bi_seg_front_size)
54 fbio->bi_seg_front_size = seg_size;
FUJITA Tomonori86771422008-10-13 14:19:05 +020055
Jens Axboe1e428072009-02-23 09:03:10 +010056 nr_phys_segs++;
57 bvprv = bv;
58 seg_size = bv->bv_len;
59 highprv = high;
60 }
Jens Axboe59247ea2009-03-06 08:55:24 +010061 bbio = bio;
Jens Axboed6d48192008-01-29 14:04:06 +010062 }
63
Jens Axboe59247ea2009-03-06 08:55:24 +010064 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;
Jens Axboe1e428072009-02-23 09:03:10 +010068
69 return nr_phys_segs;
70}
71
72void blk_recalc_rq_segments(struct request *rq)
73{
Jens Axboe59247ea2009-03-06 08:55:24 +010074 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
Jens Axboed6d48192008-01-29 14:04:06 +010075}
76
77void blk_recount_segments(struct request_queue *q, struct bio *bio)
78{
Jens Axboed6d48192008-01-29 14:04:06 +010079 struct bio *nxt = bio->bi_next;
Jens Axboe1e428072009-02-23 09:03:10 +010080
Jens Axboed6d48192008-01-29 14:04:06 +010081 bio->bi_next = NULL;
Jens Axboe59247ea2009-03-06 08:55:24 +010082 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
Jens Axboed6d48192008-01-29 14:04:06 +010083 bio->bi_next = nxt;
Jens Axboed6d48192008-01-29 14:04:06 +010084 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{
Nick Piggin75ad23b2008-04-29 14:48:33 +020091 if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
Jens Axboed6d48192008-01-29 14:04:06 +010092 return 0;
93
FUJITA Tomonori86771422008-10-13 14:19:05 +020094 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
Martin K. Petersenae03bf62009-05-22 17:17:50 -040095 queue_max_segment_size(q))
Jens Axboed6d48192008-01-29 14:04:06 +010096 return 0;
97
David Woodhousee17fc0a2008-08-09 16:42:20 +010098 if (!bio_has_data(bio))
99 return 1;
100
101 if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
102 return 0;
103
Jens Axboed6d48192008-01-29 14:04:06 +0100104 /*
David Woodhousee17fc0a2008-08-09 16:42:20 +0100105 * bio and nxt are contiguous in memory; check if the queue allows
Jens Axboed6d48192008-01-29 14:04:06 +0100106 * 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
Jens Axboed6d48192008-01-29 14:04:06 +0100114/*
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;
Nick Piggin75ad23b2008-04-29 14:48:33 +0200127 cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
Jens Axboed6d48192008-01-29 14:04:06 +0100128
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) {
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400138 if (sg->length + nbytes > queue_max_segment_size(q))
Jens Axboed6d48192008-01-29 14:04:06 +0100139 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
FUJITA Tomonorif18573a2008-04-11 12:56:52 +0200172
173 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900174 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
175 unsigned int pad_len =
176 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
FUJITA Tomonorif18573a2008-04-11 12:56:52 +0200177
178 sg->length += pad_len;
179 rq->extra_len += pad_len;
180 }
181
Tejun Heo2fb98e82008-02-19 11:36:53 +0100182 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
Tejun Heodb0a2e02008-02-19 11:36:55 +0100183 if (rq->cmd_flags & REQ_RW)
184 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
185
Jens Axboed6d48192008-01-29 14:04:06 +0100186 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++;
FUJITA Tomonori7a85f882008-03-04 11:17:11 +0100193 rq->extra_len += q->dma_drain_size;
Jens Axboed6d48192008-01-29 14:04:06 +0100194 }
195
196 if (sg)
197 sg_mark_end(sg);
198
199 return nsegs;
200}
Jens Axboed6d48192008-01-29 14:04:06 +0100201EXPORT_SYMBOL(blk_rq_map_sg);
202
Jens Axboed6d48192008-01-29 14:04:06 +0100203static inline int ll_new_hw_segment(struct request_queue *q,
204 struct request *req,
205 struct bio *bio)
206{
Jens Axboed6d48192008-01-29 14:04:06 +0100207 int nr_phys_segs = bio_phys_segments(q, bio);
208
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400209 if (req->nr_phys_segments + nr_phys_segs > queue_max_hw_segments(q) ||
210 req->nr_phys_segments + nr_phys_segs > queue_max_phys_segments(q)) {
Jens Axboed6d48192008-01-29 14:04:06 +0100211 req->cmd_flags |= REQ_NOMERGE;
212 if (req == q->last_merge)
213 q->last_merge = NULL;
214 return 0;
215 }
216
217 /*
218 * This will form the start of a new hw segment. Bump both
219 * counters.
220 */
Jens Axboed6d48192008-01-29 14:04:06 +0100221 req->nr_phys_segments += nr_phys_segs;
222 return 1;
223}
224
225int ll_back_merge_fn(struct request_queue *q, struct request *req,
226 struct bio *bio)
227{
228 unsigned short max_sectors;
Jens Axboed6d48192008-01-29 14:04:06 +0100229
230 if (unlikely(blk_pc_request(req)))
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400231 max_sectors = queue_max_hw_sectors(q);
Jens Axboed6d48192008-01-29 14:04:06 +0100232 else
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400233 max_sectors = queue_max_sectors(q);
Jens Axboed6d48192008-01-29 14:04:06 +0100234
Tejun Heo83096eb2009-05-07 22:24:39 +0900235 if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
Jens Axboed6d48192008-01-29 14:04:06 +0100236 req->cmd_flags |= REQ_NOMERGE;
237 if (req == q->last_merge)
238 q->last_merge = NULL;
239 return 0;
240 }
Jens Axboe2cdf79c2008-05-07 09:33:55 +0200241 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
Jens Axboed6d48192008-01-29 14:04:06 +0100242 blk_recount_segments(q, req->biotail);
Jens Axboe2cdf79c2008-05-07 09:33:55 +0200243 if (!bio_flagged(bio, BIO_SEG_VALID))
Jens Axboed6d48192008-01-29 14:04:06 +0100244 blk_recount_segments(q, bio);
Jens Axboed6d48192008-01-29 14:04:06 +0100245
246 return ll_new_hw_segment(q, req, bio);
247}
248
Jens Axboe6728cb02008-01-31 13:03:55 +0100249int ll_front_merge_fn(struct request_queue *q, struct request *req,
Jens Axboed6d48192008-01-29 14:04:06 +0100250 struct bio *bio)
251{
252 unsigned short max_sectors;
Jens Axboed6d48192008-01-29 14:04:06 +0100253
254 if (unlikely(blk_pc_request(req)))
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400255 max_sectors = queue_max_hw_sectors(q);
Jens Axboed6d48192008-01-29 14:04:06 +0100256 else
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400257 max_sectors = queue_max_sectors(q);
Jens Axboed6d48192008-01-29 14:04:06 +0100258
259
Tejun Heo83096eb2009-05-07 22:24:39 +0900260 if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
Jens Axboed6d48192008-01-29 14:04:06 +0100261 req->cmd_flags |= REQ_NOMERGE;
262 if (req == q->last_merge)
263 q->last_merge = NULL;
264 return 0;
265 }
Jens Axboe2cdf79c2008-05-07 09:33:55 +0200266 if (!bio_flagged(bio, BIO_SEG_VALID))
Jens Axboed6d48192008-01-29 14:04:06 +0100267 blk_recount_segments(q, bio);
Jens Axboe2cdf79c2008-05-07 09:33:55 +0200268 if (!bio_flagged(req->bio, BIO_SEG_VALID))
Jens Axboed6d48192008-01-29 14:04:06 +0100269 blk_recount_segments(q, req->bio);
Jens Axboed6d48192008-01-29 14:04:06 +0100270
271 return ll_new_hw_segment(q, req, bio);
272}
273
274static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
275 struct request *next)
276{
277 int total_phys_segments;
FUJITA Tomonori86771422008-10-13 14:19:05 +0200278 unsigned int seg_size =
279 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
Jens Axboed6d48192008-01-29 14:04:06 +0100280
281 /*
282 * First check if the either of the requests are re-queued
283 * requests. Can't merge them if they are.
284 */
285 if (req->special || next->special)
286 return 0;
287
288 /*
289 * Will it become too large?
290 */
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400291 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > queue_max_sectors(q))
Jens Axboed6d48192008-01-29 14:04:06 +0100292 return 0;
293
294 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
FUJITA Tomonori86771422008-10-13 14:19:05 +0200295 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
296 if (req->nr_phys_segments == 1)
297 req->bio->bi_seg_front_size = seg_size;
298 if (next->nr_phys_segments == 1)
299 next->biotail->bi_seg_back_size = seg_size;
Jens Axboed6d48192008-01-29 14:04:06 +0100300 total_phys_segments--;
FUJITA Tomonori86771422008-10-13 14:19:05 +0200301 }
Jens Axboed6d48192008-01-29 14:04:06 +0100302
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400303 if (total_phys_segments > queue_max_phys_segments(q))
Jens Axboed6d48192008-01-29 14:04:06 +0100304 return 0;
305
Martin K. Petersenae03bf62009-05-22 17:17:50 -0400306 if (total_phys_segments > queue_max_hw_segments(q))
Jens Axboed6d48192008-01-29 14:04:06 +0100307 return 0;
308
309 /* Merge is OK... */
310 req->nr_phys_segments = total_phys_segments;
Jens Axboed6d48192008-01-29 14:04:06 +0100311 return 1;
312}
313
Tejun Heo80a761f2009-07-03 17:48:17 +0900314/**
315 * blk_rq_set_mixed_merge - mark a request as mixed merge
316 * @rq: request to mark as mixed merge
317 *
318 * Description:
319 * @rq is about to be mixed merged. Make sure the attributes
320 * which can be mixed are set in each bio and mark @rq as mixed
321 * merged.
322 */
323void blk_rq_set_mixed_merge(struct request *rq)
324{
325 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
326 struct bio *bio;
327
328 if (rq->cmd_flags & REQ_MIXED_MERGE)
329 return;
330
331 /*
332 * @rq will no longer represent mixable attributes for all the
333 * contained bios. It will just track those of the first one.
334 * Distributes the attributs to each bio.
335 */
336 for (bio = rq->bio; bio; bio = bio->bi_next) {
337 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
338 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
339 bio->bi_rw |= ff;
340 }
341 rq->cmd_flags |= REQ_MIXED_MERGE;
342}
343
Jerome Marchand26308ea2009-03-27 10:31:51 +0100344static void blk_account_io_merge(struct request *req)
345{
346 if (blk_do_io_stat(req)) {
347 struct hd_struct *part;
348 int cpu;
349
350 cpu = part_stat_lock();
Tejun Heo83096eb2009-05-07 22:24:39 +0900351 part = disk_map_sector_rcu(req->rq_disk, blk_rq_pos(req));
Jerome Marchand26308ea2009-03-27 10:31:51 +0100352
353 part_round_stats(cpu, part);
Jens Axboe0f78ab92009-10-04 21:04:38 +0200354 part_dec_in_flight(part);
Jerome Marchand26308ea2009-03-27 10:31:51 +0100355
356 part_stat_unlock();
357 }
358}
359
Jens Axboed6d48192008-01-29 14:04:06 +0100360/*
361 * Has to be called with the request spinlock acquired
362 */
363static int attempt_merge(struct request_queue *q, struct request *req,
364 struct request *next)
365{
366 if (!rq_mergeable(req) || !rq_mergeable(next))
367 return 0;
368
369 /*
370 * not contiguous
371 */
Tejun Heo83096eb2009-05-07 22:24:39 +0900372 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
Jens Axboed6d48192008-01-29 14:04:06 +0100373 return 0;
374
375 if (rq_data_dir(req) != rq_data_dir(next)
376 || req->rq_disk != next->rq_disk
377 || next->special)
378 return 0;
379
Martin K. Petersen7ba1ba12008-06-30 20:04:41 +0200380 if (blk_integrity_rq(req) != blk_integrity_rq(next))
381 return 0;
382
Jens Axboed6d48192008-01-29 14:04:06 +0100383 /*
384 * If we are allowed to merge, then append bio list
385 * from next to rq and release next. merge_requests_fn
386 * will have updated segment counts, update sector
387 * counts here.
388 */
389 if (!ll_merge_requests_fn(q, req, next))
390 return 0;
391
392 /*
Tejun Heo80a761f2009-07-03 17:48:17 +0900393 * If failfast settings disagree or any of the two is already
394 * a mixed merge, mark both as mixed before proceeding. This
395 * makes sure that all involved bios have mixable attributes
396 * set properly.
397 */
398 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
399 (req->cmd_flags & REQ_FAILFAST_MASK) !=
400 (next->cmd_flags & REQ_FAILFAST_MASK)) {
401 blk_rq_set_mixed_merge(req);
402 blk_rq_set_mixed_merge(next);
403 }
404
405 /*
Jens Axboed6d48192008-01-29 14:04:06 +0100406 * At this point we have either done a back merge
407 * or front merge. We need the smaller start_time of
408 * the merged requests to be the current request
409 * for accounting purposes.
410 */
411 if (time_after(req->start_time, next->start_time))
412 req->start_time = next->start_time;
413
414 req->biotail->bi_next = next->bio;
415 req->biotail = next->biotail;
416
Tejun Heoa2dec7b2009-05-07 22:24:44 +0900417 req->__data_len += blk_rq_bytes(next);
Jens Axboed6d48192008-01-29 14:04:06 +0100418
419 elv_merge_requests(q, req, next);
420
Jerome Marchand42dad762009-04-22 14:01:49 +0200421 /*
422 * 'next' is going away, so update stats accordingly
423 */
424 blk_account_io_merge(next);
Jens Axboed6d48192008-01-29 14:04:06 +0100425
426 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
Jens Axboeab780f12008-08-26 10:25:02 +0200427 if (blk_rq_cpu_valid(next))
428 req->cpu = next->cpu;
Jens Axboed6d48192008-01-29 14:04:06 +0100429
Boaz Harrosh1cd96c22009-03-24 12:35:07 +0100430 /* owner-ship of bio passed from next to req */
431 next->bio = NULL;
Jens Axboed6d48192008-01-29 14:04:06 +0100432 __blk_put_request(q, next);
433 return 1;
434}
435
436int attempt_back_merge(struct request_queue *q, struct request *rq)
437{
438 struct request *next = elv_latter_request(q, rq);
439
440 if (next)
441 return attempt_merge(q, rq, next);
442
443 return 0;
444}
445
446int attempt_front_merge(struct request_queue *q, struct request *rq)
447{
448 struct request *prev = elv_former_request(q, rq);
449
450 if (prev)
451 return attempt_merge(q, prev, rq);
452
453 return 0;
454}