block: fix disk->part[] dereferencing race
[linux-2.6.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                 nr_phys_segs++;
81                 bvprv = bv;
82                 seg_size = bv->bv_len;
83                 highprv = high;
84         }
85
86         rq->nr_phys_segments = nr_phys_segs;
87 }
88
89 void blk_recount_segments(struct request_queue *q, struct bio *bio)
90 {
91         struct request rq;
92         struct bio *nxt = bio->bi_next;
93         rq.q = q;
94         rq.bio = rq.biotail = bio;
95         bio->bi_next = NULL;
96         blk_recalc_rq_segments(&rq);
97         bio->bi_next = nxt;
98         bio->bi_phys_segments = rq.nr_phys_segments;
99         bio->bi_flags |= (1 << BIO_SEG_VALID);
100 }
101 EXPORT_SYMBOL(blk_recount_segments);
102
103 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
104                                    struct bio *nxt)
105 {
106         if (!test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags))
107                 return 0;
108
109         if (bio->bi_size + nxt->bi_size > q->max_segment_size)
110                 return 0;
111
112         if (!bio_has_data(bio))
113                 return 1;
114
115         if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
116                 return 0;
117
118         /*
119          * bio and nxt are contiguous in memory; check if the queue allows
120          * these two to be merged into one
121          */
122         if (BIO_SEG_BOUNDARY(q, bio, nxt))
123                 return 1;
124
125         return 0;
126 }
127
128 /*
129  * map a request to scatterlist, return number of sg entries setup. Caller
130  * must make sure sg can hold rq->nr_phys_segments entries
131  */
132 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
133                   struct scatterlist *sglist)
134 {
135         struct bio_vec *bvec, *bvprv;
136         struct req_iterator iter;
137         struct scatterlist *sg;
138         int nsegs, cluster;
139
140         nsegs = 0;
141         cluster = test_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
142
143         /*
144          * for each bio in rq
145          */
146         bvprv = NULL;
147         sg = NULL;
148         rq_for_each_segment(bvec, rq, iter) {
149                 int nbytes = bvec->bv_len;
150
151                 if (bvprv && cluster) {
152                         if (sg->length + nbytes > q->max_segment_size)
153                                 goto new_segment;
154
155                         if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
156                                 goto new_segment;
157                         if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
158                                 goto new_segment;
159
160                         sg->length += nbytes;
161                 } else {
162 new_segment:
163                         if (!sg)
164                                 sg = sglist;
165                         else {
166                                 /*
167                                  * If the driver previously mapped a shorter
168                                  * list, we could see a termination bit
169                                  * prematurely unless it fully inits the sg
170                                  * table on each mapping. We KNOW that there
171                                  * must be more entries here or the driver
172                                  * would be buggy, so force clear the
173                                  * termination bit to avoid doing a full
174                                  * sg_init_table() in drivers for each command.
175                                  */
176                                 sg->page_link &= ~0x02;
177                                 sg = sg_next(sg);
178                         }
179
180                         sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
181                         nsegs++;
182                 }
183                 bvprv = bvec;
184         } /* segments in rq */
185
186
187         if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
188             (rq->data_len & q->dma_pad_mask)) {
189                 unsigned int pad_len = (q->dma_pad_mask & ~rq->data_len) + 1;
190
191                 sg->length += pad_len;
192                 rq->extra_len += pad_len;
193         }
194
195         if (q->dma_drain_size && q->dma_drain_needed(rq)) {
196                 if (rq->cmd_flags & REQ_RW)
197                         memset(q->dma_drain_buffer, 0, q->dma_drain_size);
198
199                 sg->page_link &= ~0x02;
200                 sg = sg_next(sg);
201                 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
202                             q->dma_drain_size,
203                             ((unsigned long)q->dma_drain_buffer) &
204                             (PAGE_SIZE - 1));
205                 nsegs++;
206                 rq->extra_len += q->dma_drain_size;
207         }
208
209         if (sg)
210                 sg_mark_end(sg);
211
212         return nsegs;
213 }
214 EXPORT_SYMBOL(blk_rq_map_sg);
215
216 static inline int ll_new_mergeable(struct request_queue *q,
217                                    struct request *req,
218                                    struct bio *bio)
219 {
220         int nr_phys_segs = bio_phys_segments(q, bio);
221
222         if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
223                 req->cmd_flags |= REQ_NOMERGE;
224                 if (req == q->last_merge)
225                         q->last_merge = NULL;
226                 return 0;
227         }
228
229         /*
230          * A hw segment is just getting larger, bump just the phys
231          * counter.
232          */
233         req->nr_phys_segments += nr_phys_segs;
234         return 1;
235 }
236
237 static inline int ll_new_hw_segment(struct request_queue *q,
238                                     struct request *req,
239                                     struct bio *bio)
240 {
241         int nr_phys_segs = bio_phys_segments(q, bio);
242
243         if (req->nr_phys_segments + nr_phys_segs > q->max_hw_segments
244             || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
245                 req->cmd_flags |= REQ_NOMERGE;
246                 if (req == q->last_merge)
247                         q->last_merge = NULL;
248                 return 0;
249         }
250
251         /*
252          * This will form the start of a new hw segment.  Bump both
253          * counters.
254          */
255         req->nr_phys_segments += nr_phys_segs;
256         return 1;
257 }
258
259 int ll_back_merge_fn(struct request_queue *q, struct request *req,
260                      struct bio *bio)
261 {
262         unsigned short max_sectors;
263
264         if (unlikely(blk_pc_request(req)))
265                 max_sectors = q->max_hw_sectors;
266         else
267                 max_sectors = q->max_sectors;
268
269         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
270                 req->cmd_flags |= REQ_NOMERGE;
271                 if (req == q->last_merge)
272                         q->last_merge = NULL;
273                 return 0;
274         }
275         if (!bio_flagged(req->biotail, BIO_SEG_VALID))
276                 blk_recount_segments(q, req->biotail);
277         if (!bio_flagged(bio, BIO_SEG_VALID))
278                 blk_recount_segments(q, bio);
279
280         return ll_new_hw_segment(q, req, bio);
281 }
282
283 int ll_front_merge_fn(struct request_queue *q, struct request *req,
284                       struct bio *bio)
285 {
286         unsigned short max_sectors;
287
288         if (unlikely(blk_pc_request(req)))
289                 max_sectors = q->max_hw_sectors;
290         else
291                 max_sectors = q->max_sectors;
292
293
294         if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
295                 req->cmd_flags |= REQ_NOMERGE;
296                 if (req == q->last_merge)
297                         q->last_merge = NULL;
298                 return 0;
299         }
300         if (!bio_flagged(bio, BIO_SEG_VALID))
301                 blk_recount_segments(q, bio);
302         if (!bio_flagged(req->bio, BIO_SEG_VALID))
303                 blk_recount_segments(q, req->bio);
304
305         return ll_new_hw_segment(q, req, bio);
306 }
307
308 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
309                                 struct request *next)
310 {
311         int total_phys_segments;
312
313         /*
314          * First check if the either of the requests are re-queued
315          * requests.  Can't merge them if they are.
316          */
317         if (req->special || next->special)
318                 return 0;
319
320         /*
321          * Will it become too large?
322          */
323         if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
324                 return 0;
325
326         total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
327         if (blk_phys_contig_segment(q, req->biotail, next->bio))
328                 total_phys_segments--;
329
330         if (total_phys_segments > q->max_phys_segments)
331                 return 0;
332
333         if (total_phys_segments > q->max_hw_segments)
334                 return 0;
335
336         /* Merge is OK... */
337         req->nr_phys_segments = total_phys_segments;
338         return 1;
339 }
340
341 /*
342  * Has to be called with the request spinlock acquired
343  */
344 static int attempt_merge(struct request_queue *q, struct request *req,
345                           struct request *next)
346 {
347         if (!rq_mergeable(req) || !rq_mergeable(next))
348                 return 0;
349
350         /*
351          * not contiguous
352          */
353         if (req->sector + req->nr_sectors != next->sector)
354                 return 0;
355
356         if (rq_data_dir(req) != rq_data_dir(next)
357             || req->rq_disk != next->rq_disk
358             || next->special)
359                 return 0;
360
361         if (blk_integrity_rq(req) != blk_integrity_rq(next))
362                 return 0;
363
364         /*
365          * If we are allowed to merge, then append bio list
366          * from next to rq and release next. merge_requests_fn
367          * will have updated segment counts, update sector
368          * counts here.
369          */
370         if (!ll_merge_requests_fn(q, req, next))
371                 return 0;
372
373         /*
374          * At this point we have either done a back merge
375          * or front merge. We need the smaller start_time of
376          * the merged requests to be the current request
377          * for accounting purposes.
378          */
379         if (time_after(req->start_time, next->start_time))
380                 req->start_time = next->start_time;
381
382         req->biotail->bi_next = next->bio;
383         req->biotail = next->biotail;
384
385         req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
386
387         elv_merge_requests(q, req, next);
388
389         if (req->rq_disk) {
390                 struct hd_struct *part;
391
392                 rcu_read_lock();
393
394                 part = disk_map_sector_rcu(req->rq_disk, req->sector);
395                 disk_round_stats(req->rq_disk);
396                 req->rq_disk->in_flight--;
397                 if (part) {
398                         part_round_stats(part);
399                         part->in_flight--;
400                 }
401
402                 rcu_read_unlock();
403         }
404
405         req->ioprio = ioprio_best(req->ioprio, next->ioprio);
406
407         __blk_put_request(q, next);
408         return 1;
409 }
410
411 int attempt_back_merge(struct request_queue *q, struct request *rq)
412 {
413         struct request *next = elv_latter_request(q, rq);
414
415         if (next)
416                 return attempt_merge(q, rq, next);
417
418         return 0;
419 }
420
421 int attempt_front_merge(struct request_queue *q, struct request *rq)
422 {
423         struct request *prev = elv_former_request(q, rq);
424
425         if (prev)
426                 return attempt_merge(q, prev, rq);
427
428         return 0;
429 }