block: fix nr_phys_segments miscalculation bug
[linux-3.10.git] / 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
12 void blk_recalc_rq_sectors(struct request *rq, int nsect)
13 {
14         if (blk_fs_request(rq) || blk_discard_rq(rq)) {
15                 rq->hard_sector += nsect;
16                 rq->hard_nr_sectors -= nsect;
17
18                 /*
19                  * Move the I/O submission pointers ahead if required.
20                  */
21                 if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
22                     (rq->sector <= rq->hard_sector)) {
23                         rq->sector = rq->hard_sector;
24                         rq->nr_sectors = rq->hard_nr_sectors;
25                         rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
26                         rq->current_nr_sectors = rq->hard_cur_sectors;
27                         rq->buffer = bio_data(rq->bio);
28                 }
29
30                 /*
31                  * if total number of sectors is less than the first segment
32                  * size, something has gone terribly wrong
33                  */
34                 if (rq->nr_sectors < rq->current_nr_sectors) {
35                         printk(KERN_ERR "blk: request botched\n");
36                         rq->nr_sectors = rq->current_nr_sectors;
37                 }
38         }
39 }
40
41 void blk_recalc_rq_segments(struct request *rq)
42 {
43         int nr_phys_segs;
44         unsigned int phys_size;
45         struct bio_vec *bv, *bvprv = NULL;
46         int seg_size;
47         int cluster;
48         struct req_iterator iter;
49         int high, highprv = 1;
50         struct request_queue *q = rq->q;
51
52         if (!rq->bio)
53                 return;
54
55         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
56         seg_size = 0;
57         phys_size = nr_phys_segs = 0;
58         rq_for_each_segment(bv, rq, iter) {
59                 /*
60                  * the trick here is making sure that a high page is never
61                  * considered part of another segment, since that might
62                  * change with the bounce page.
63                  */
64                 high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
65                 if (high || highprv)
66                         goto new_segment;
67                 if (cluster) {
68                         if (seg_size + bv->bv_len > q->max_segment_size)
69                                 goto new_segment;
70                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
71                                 goto new_segment;
72                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
73                                 goto new_segment;
74
75                         seg_size += bv->bv_len;
76                         bvprv = bv;
77                         continue;
78                 }
79 new_segment:
80                 if (nr_phys_segs == 1 && seg_size > rq->bio->bi_seg_front_size)
81                         rq->bio->bi_seg_front_size = seg_size;
82
83                 nr_phys_segs++;
84                 bvprv = bv;
85                 seg_size = bv->bv_len;
86                 highprv = high;
87         }
88
89         if (nr_phys_segs == 1 && seg_size > rq->bio->bi_seg_front_size)
90                 rq->bio->bi_seg_front_size = seg_size;
91         if (seg_size > rq->biotail->bi_seg_back_size)
92                 rq->biotail->bi_seg_back_size = seg_size;
93
94         rq->nr_phys_segments = nr_phys_segs;
95 }
96
97 void blk_recount_segments(struct request_queue *q, struct bio *bio)
98 {
99         struct request rq;
100         struct bio *nxt = bio->bi_next;
101         rq.q = q;
102         rq.bio = rq.biotail = bio;
103         bio->bi_next = NULL;
104         blk_recalc_rq_segments(&rq);
105         bio->bi_next = nxt;
106         bio->bi_phys_segments = rq.nr_phys_segments;
107         bio->bi_flags |= (1 << BIO_SEG_VALID);
108 }
109 EXPORT_SYMBOL(blk_recount_segments);
110
111 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
112                                    struct bio *nxt)
113 {
114         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
115                 return 0;
116
117         if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
118             q->max_segment_size)
119                 return 0;
120
121         if (!bio_has_data(bio))
122                 return 1;
123
124         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
125                 return 0;
126
127         /*
128          * bio and nxt are contiguous in memory; check if the queue allows
129          * these two to be merged into one
130          */
131         if (BIO_SEG_BOUNDARY(q, bio, nxt))
132                 return 1;
133
134         return 0;
135 }
136
137 /*
138  * map a request to scatterlist, return number of sg entries setup. Caller
139  * must make sure sg can hold rq->nr_phys_segments entries
140  */
141 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
142                   struct scatterlist *sglist)
143 {
144         struct bio_vec *bvec, *bvprv;
145         struct req_iterator iter;
146         struct scatterlist *sg;
147         int nsegs, cluster;
148
149         nsegs = 0;
150         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
151
152         /*
153          * for each bio in rq
154          */
155         bvprv = NULL;
156         sg = NULL;
157         rq_for_each_segment(bvec, rq, iter) {
158                 int nbytes = bvec->bv_len;
159
160                 if (bvprv && cluster) {
161                         if (sg->length + nbytes > q->max_segment_size)
162                                 goto new_segment;
163
164                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
165                                 goto new_segment;
166                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
167                                 goto new_segment;
168
169                         sg->length += nbytes;
170                 } else {
171 new_segment:
172                         if (!sg)
173                                 sg = sglist;
174                         else {
175                                 /*
176                                  * If the driver previously mapped a shorter
177                                  * list, we could see a termination bit
178                                  * prematurely unless it fully inits the sg
179                                  * table on each mapping. We KNOW that there
180                                  * must be more entries here or the driver
181                                  * would be buggy, so force clear the
182                                  * termination bit to avoid doing a full
183                                  * sg_init_table() in drivers for each command.
184                                  */
185                                 sg->page_link &= ~0x02;
186                                 sg = sg_next(sg);
187                         }
188
189                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
190                         nsegs++;
191                 }
192                 bvprv = bvec;
193         } /* segments in rq */
194
195
196         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
197             (rq->data_len & q->dma_pad_mask)) {
198                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
199
200                 sg->length += pad_len;
201                 rq->extra_len += pad_len;
202         }
203
204         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
205                 if (rq->cmd_flags & REQ_RW)
206                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
207
208                 sg->page_link &= ~0x02;
209                 sg = sg_next(sg);
210                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
211                             q->dma_drain_size,
212                             ((unsigned long)q->dma_drain_buffer) &
213                             (PAGE_SIZE - 1));
214                 nsegs++;
215                 rq->extra_len += q->dma_drain_size;
216         }
217
218         if (sg)
219                 sg_mark_end(sg);
220
221         return nsegs;
222 }
223 EXPORT_SYMBOL(blk_rq_map_sg);
224
225 static inline int ll_new_mergeable(struct request_queue *q,
226                                    struct request *req,
227                                    struct bio *bio)
228 {
229         int nr_phys_segs = bio_phys_segments(q, bio);
230
231         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
232                 req->cmd_flags |= REQ_NOMERGE;
233                 if (req == q->last_merge)
234                         q->last_merge = NULL;
235                 return 0;
236         }
237
238         /*
239          * A hw segment is just getting larger, bump just the phys
240          * counter.
241          */
242         req->nr_phys_segments += nr_phys_segs;
243         return 1;
244 }
245
246 static inline int ll_new_hw_segment(struct request_queue *q,
247                                     struct request *req,
248                                     struct bio *bio)
249 {
250         int nr_phys_segs = bio_phys_segments(q, bio);
251
252         if (req->nr_phys_segments + nr_phys_segs > q->max_hw_segments
253             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
254                 req->cmd_flags |= REQ_NOMERGE;
255                 if (req == q->last_merge)
256                         q->last_merge = NULL;
257                 return 0;
258         }
259
260         /*
261          * This will form the start of a new hw segment.  Bump both
262          * counters.
263          */
264         req->nr_phys_segments += nr_phys_segs;
265         return 1;
266 }
267
268 int ll_back_merge_fn(struct request_queue *q, struct request *req,
269                      struct bio *bio)
270 {
271         unsigned short max_sectors;
272
273         if (unlikely(blk_pc_request(req)))
274                 max_sectors = q->max_hw_sectors;
275         else
276                 max_sectors = q->max_sectors;
277
278         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
279                 req->cmd_flags |= REQ_NOMERGE;
280                 if (req == q->last_merge)
281                         q->last_merge = NULL;
282                 return 0;
283         }
284         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
285                 blk_recount_segments(q, req->biotail);
286         if (!bio_flagged(bio, BIO_SEG_VALID))
287                 blk_recount_segments(q, bio);
288
289         return ll_new_hw_segment(q, req, bio);
290 }
291
292 int ll_front_merge_fn(struct request_queue *q, struct request *req,
293                       struct bio *bio)
294 {
295         unsigned short max_sectors;
296
297         if (unlikely(blk_pc_request(req)))
298                 max_sectors = q->max_hw_sectors;
299         else
300                 max_sectors = q->max_sectors;
301
302
303         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
304                 req->cmd_flags |= REQ_NOMERGE;
305                 if (req == q->last_merge)
306                         q->last_merge = NULL;
307                 return 0;
308         }
309         if (!bio_flagged(bio, BIO_SEG_VALID))
310                 blk_recount_segments(q, bio);
311         if (!bio_flagged(req->bio, BIO_SEG_VALID))
312                 blk_recount_segments(q, req->bio);
313
314         return ll_new_hw_segment(q, req, bio);
315 }
316
317 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
318                                 struct request *next)
319 {
320         int total_phys_segments;
321         unsigned int seg_size =
322                 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
323
324         /*
325          * First check if the either of the requests are re-queued
326          * requests.  Can't merge them if they are.
327          */
328         if (req->special || next->special)
329                 return 0;
330
331         /*
332          * Will it become too large?
333          */
334         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
335                 return 0;
336
337         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
338         if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
339                 if (req->nr_phys_segments == 1)
340                         req->bio->bi_seg_front_size = seg_size;
341                 if (next->nr_phys_segments == 1)
342                         next->biotail->bi_seg_back_size = seg_size;
343                 total_phys_segments--;
344         }
345
346         if (total_phys_segments > q->max_phys_segments)
347                 return 0;
348
349         if (total_phys_segments > q->max_hw_segments)
350                 return 0;
351
352         /* Merge is OK... */
353         req->nr_phys_segments = total_phys_segments;
354         return 1;
355 }
356
357 /*
358  * Has to be called with the request spinlock acquired
359  */
360 static int attempt_merge(struct request_queue *q, struct request *req,
361                           struct request *next)
362 {
363         if (!rq_mergeable(req) || !rq_mergeable(next))
364                 return 0;
365
366         /*
367          * not contiguous
368          */
369         if (req->sector + req->nr_sectors != next->sector)
370                 return 0;
371
372         if (rq_data_dir(req) != rq_data_dir(next)
373             || req->rq_disk != next->rq_disk
374             || next->special)
375                 return 0;
376
377         if (blk_integrity_rq(req) != blk_integrity_rq(next))
378                 return 0;
379
380         /*
381          * If we are allowed to merge, then append bio list
382          * from next to rq and release next. merge_requests_fn
383          * will have updated segment counts, update sector
384          * counts here.
385          */
386         if (!ll_merge_requests_fn(q, req, next))
387                 return 0;
388
389         /*
390          * At this point we have either done a back merge
391          * or front merge. We need the smaller start_time of
392          * the merged requests to be the current request
393          * for accounting purposes.
394          */
395         if (time_after(req->start_time, next->start_time))
396                 req->start_time = next->start_time;
397
398         req->biotail->bi_next = next->bio;
399         req->biotail = next->biotail;
400
401         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
402
403         elv_merge_requests(q, req, next);
404
405         if (req->rq_disk) {
406                 struct hd_struct *part;
407                 int cpu;
408
409                 cpu = part_stat_lock();
410                 part = disk_map_sector_rcu(req->rq_disk, req->sector);
411
412                 part_round_stats(cpu, part);
413                 part_dec_in_flight(part);
414
415                 part_stat_unlock();
416         }
417
418         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
419         if (blk_rq_cpu_valid(next))
420                 req->cpu = next->cpu;
421
422         __blk_put_request(q, next);
423         return 1;
424 }
425
426 int attempt_back_merge(struct request_queue *q, struct request *rq)
427 {
428         struct request *next = elv_latter_request(q, rq);
429
430         if (next)
431                 return attempt_merge(q, rq, next);
432
433         return 0;
434 }
435
436 int attempt_front_merge(struct request_queue *q, struct request *rq)
437 {
438         struct request *prev = elv_former_request(q, rq);
439
440         if (prev)
441                 return attempt_merge(q, prev, rq);
442
443         return 0;
444 }