Root/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_hw_segments(q) ||
210        req->nr_phys_segments + nr_phys_segs > queue_max_phys_segments(q)) {
211        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     */
221    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;
229
230    if (unlikely(blk_pc_request(req)))
231        max_sectors = queue_max_hw_sectors(q);
232    else
233        max_sectors = queue_max_sectors(q);
234
235    if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
236        req->cmd_flags |= REQ_NOMERGE;
237        if (req == q->last_merge)
238            q->last_merge = NULL;
239        return 0;
240    }
241    if (!bio_flagged(req->biotail, BIO_SEG_VALID))
242        blk_recount_segments(q, req->biotail);
243    if (!bio_flagged(bio, BIO_SEG_VALID))
244        blk_recount_segments(q, bio);
245
246    return ll_new_hw_segment(q, req, bio);
247}
248
249int ll_front_merge_fn(struct request_queue *q, struct request *req,
250              struct bio *bio)
251{
252    unsigned short max_sectors;
253
254    if (unlikely(blk_pc_request(req)))
255        max_sectors = queue_max_hw_sectors(q);
256    else
257        max_sectors = queue_max_sectors(q);
258
259
260    if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
261        req->cmd_flags |= REQ_NOMERGE;
262        if (req == q->last_merge)
263            q->last_merge = NULL;
264        return 0;
265    }
266    if (!bio_flagged(bio, BIO_SEG_VALID))
267        blk_recount_segments(q, bio);
268    if (!bio_flagged(req->bio, BIO_SEG_VALID))
269        blk_recount_segments(q, req->bio);
270
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;
278    unsigned int seg_size =
279        req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
280
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     */
291    if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > queue_max_sectors(q))
292        return 0;
293
294    total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
295    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;
300        total_phys_segments--;
301    }
302
303    if (total_phys_segments > queue_max_phys_segments(q))
304        return 0;
305
306    if (total_phys_segments > queue_max_hw_segments(q))
307        return 0;
308
309    /* Merge is OK... */
310    req->nr_phys_segments = total_phys_segments;
311    return 1;
312}
313
314/**
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
344static 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();
351        part = disk_map_sector_rcu(req->rq_disk, blk_rq_pos(req));
352
353        part_round_stats(cpu, part);
354        part_dec_in_flight(part, rq_data_dir(req));
355
356        part_stat_unlock();
357    }
358}
359
360/*
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     */
372    if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
373        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
380    if (blk_integrity_rq(req) != blk_integrity_rq(next))
381        return 0;
382
383    /*
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    /*
393     * 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    /*
406     * 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
417    req->__data_len += blk_rq_bytes(next);
418
419    elv_merge_requests(q, req, next);
420
421    /*
422     * 'next' is going away, so update stats accordingly
423     */
424    blk_account_io_merge(next);
425
426    req->ioprio = ioprio_best(req->ioprio, next->ioprio);
427    if (blk_rq_cpu_valid(next))
428        req->cpu = next->cpu;
429
430    /* owner-ship of bio passed from next to req */
431    next->bio = NULL;
432    __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}
455

Archive Download this file



interactive