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    struct bio_vec *bv, *bvprv = NULL;
16    int cluster, i, high, highprv = 1;
17    unsigned int seg_size, nr_phys_segs;
18    struct bio *fbio, *bbio;
19
20    if (!bio)
21        return 0;
22
23    fbio = bio;
24    cluster = blk_queue_cluster(q);
25    seg_size = 0;
26    nr_phys_segs = 0;
27    for_each_bio(bio) {
28        bio_for_each_segment(bv, bio, i) {
29            /*
30             * the trick here is making sure that a high page is
31             * never considered part of another segment, since that
32             * might change with the bounce page.
33             */
34            high = page_to_pfn(bv->bv_page) > queue_bounce_pfn(q);
35            if (high || highprv)
36                goto new_segment;
37            if (cluster) {
38                if (seg_size + bv->bv_len
39                    > queue_max_segment_size(q))
40                    goto new_segment;
41                if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
42                    goto new_segment;
43                if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
44                    goto new_segment;
45
46                seg_size += bv->bv_len;
47                bvprv = bv;
48                continue;
49            }
50new_segment:
51            if (nr_phys_segs == 1 && seg_size >
52                fbio->bi_seg_front_size)
53                fbio->bi_seg_front_size = seg_size;
54
55            nr_phys_segs++;
56            bvprv = bv;
57            seg_size = bv->bv_len;
58            highprv = high;
59        }
60        bbio = bio;
61    }
62
63    if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
64        fbio->bi_seg_front_size = seg_size;
65    if (seg_size > bbio->bi_seg_back_size)
66        bbio->bi_seg_back_size = seg_size;
67
68    return nr_phys_segs;
69}
70
71void blk_recalc_rq_segments(struct request *rq)
72{
73    rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
74}
75
76void blk_recount_segments(struct request_queue *q, struct bio *bio)
77{
78    struct bio *nxt = bio->bi_next;
79
80    bio->bi_next = NULL;
81    bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
82    bio->bi_next = nxt;
83    bio->bi_flags |= (1 << BIO_SEG_VALID);
84}
85EXPORT_SYMBOL(blk_recount_segments);
86
87static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
88                   struct bio *nxt)
89{
90    if (!blk_queue_cluster(q))
91        return 0;
92
93    if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
94        queue_max_segment_size(q))
95        return 0;
96
97    if (!bio_has_data(bio))
98        return 1;
99
100    if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
101        return 0;
102
103    /*
104     * bio and nxt are contiguous in memory; check if the queue allows
105     * these two to be merged into one
106     */
107    if (BIO_SEG_BOUNDARY(q, bio, nxt))
108        return 1;
109
110    return 0;
111}
112
113/*
114 * map a request to scatterlist, return number of sg entries setup. Caller
115 * must make sure sg can hold rq->nr_phys_segments entries
116 */
117int blk_rq_map_sg(struct request_queue *q, struct request *rq,
118          struct scatterlist *sglist)
119{
120    struct bio_vec *bvec, *bvprv;
121    struct req_iterator iter;
122    struct scatterlist *sg;
123    int nsegs, cluster;
124
125    nsegs = 0;
126    cluster = blk_queue_cluster(q);
127
128    /*
129     * for each bio in rq
130     */
131    bvprv = NULL;
132    sg = NULL;
133    rq_for_each_segment(bvec, rq, iter) {
134        int nbytes = bvec->bv_len;
135
136        if (bvprv && cluster) {
137            if (sg->length + nbytes > queue_max_segment_size(q))
138                goto new_segment;
139
140            if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
141                goto new_segment;
142            if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
143                goto new_segment;
144
145            sg->length += nbytes;
146        } else {
147new_segment:
148            if (!sg)
149                sg = sglist;
150            else {
151                /*
152                 * If the driver previously mapped a shorter
153                 * list, we could see a termination bit
154                 * prematurely unless it fully inits the sg
155                 * table on each mapping. We KNOW that there
156                 * must be more entries here or the driver
157                 * would be buggy, so force clear the
158                 * termination bit to avoid doing a full
159                 * sg_init_table() in drivers for each command.
160                 */
161                sg->page_link &= ~0x02;
162                sg = sg_next(sg);
163            }
164
165            sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
166            nsegs++;
167        }
168        bvprv = bvec;
169    } /* segments in rq */
170
171
172    if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
173        (blk_rq_bytes(rq) & q->dma_pad_mask)) {
174        unsigned int pad_len =
175            (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
176
177        sg->length += pad_len;
178        rq->extra_len += pad_len;
179    }
180
181    if (q->dma_drain_size && q->dma_drain_needed(rq)) {
182        if (rq->cmd_flags & REQ_WRITE)
183            memset(q->dma_drain_buffer, 0, q->dma_drain_size);
184
185        sg->page_link &= ~0x02;
186        sg = sg_next(sg);
187        sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
188                q->dma_drain_size,
189                ((unsigned long)q->dma_drain_buffer) &
190                (PAGE_SIZE - 1));
191        nsegs++;
192        rq->extra_len += q->dma_drain_size;
193    }
194
195    if (sg)
196        sg_mark_end(sg);
197
198    return nsegs;
199}
200EXPORT_SYMBOL(blk_rq_map_sg);
201
202static inline int ll_new_hw_segment(struct request_queue *q,
203                    struct request *req,
204                    struct bio *bio)
205{
206    int nr_phys_segs = bio_phys_segments(q, bio);
207
208    if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
209        goto no_merge;
210
211    if (bio_integrity(bio) && blk_integrity_merge_bio(q, req, bio))
212        goto no_merge;
213
214    /*
215     * This will form the start of a new hw segment. Bump both
216     * counters.
217     */
218    req->nr_phys_segments += nr_phys_segs;
219    return 1;
220
221no_merge:
222    req->cmd_flags |= REQ_NOMERGE;
223    if (req == q->last_merge)
224        q->last_merge = NULL;
225    return 0;
226}
227
228int ll_back_merge_fn(struct request_queue *q, struct request *req,
229             struct bio *bio)
230{
231    unsigned short max_sectors;
232
233    if (unlikely(req->cmd_type == REQ_TYPE_BLOCK_PC))
234        max_sectors = queue_max_hw_sectors(q);
235    else
236        max_sectors = queue_max_sectors(q);
237
238    if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
239        req->cmd_flags |= REQ_NOMERGE;
240        if (req == q->last_merge)
241            q->last_merge = NULL;
242        return 0;
243    }
244    if (!bio_flagged(req->biotail, BIO_SEG_VALID))
245        blk_recount_segments(q, req->biotail);
246    if (!bio_flagged(bio, BIO_SEG_VALID))
247        blk_recount_segments(q, bio);
248
249    return ll_new_hw_segment(q, req, bio);
250}
251
252int ll_front_merge_fn(struct request_queue *q, struct request *req,
253              struct bio *bio)
254{
255    unsigned short max_sectors;
256
257    if (unlikely(req->cmd_type == REQ_TYPE_BLOCK_PC))
258        max_sectors = queue_max_hw_sectors(q);
259    else
260        max_sectors = queue_max_sectors(q);
261
262
263    if (blk_rq_sectors(req) + bio_sectors(bio) > max_sectors) {
264        req->cmd_flags |= REQ_NOMERGE;
265        if (req == q->last_merge)
266            q->last_merge = NULL;
267        return 0;
268    }
269    if (!bio_flagged(bio, BIO_SEG_VALID))
270        blk_recount_segments(q, bio);
271    if (!bio_flagged(req->bio, BIO_SEG_VALID))
272        blk_recount_segments(q, req->bio);
273
274    return ll_new_hw_segment(q, req, bio);
275}
276
277static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
278                struct request *next)
279{
280    int total_phys_segments;
281    unsigned int seg_size =
282        req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
283
284    /*
285     * First check if the either of the requests are re-queued
286     * requests. Can't merge them if they are.
287     */
288    if (req->special || next->special)
289        return 0;
290
291    /*
292     * Will it become too large?
293     */
294    if ((blk_rq_sectors(req) + blk_rq_sectors(next)) > queue_max_sectors(q))
295        return 0;
296
297    total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
298    if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
299        if (req->nr_phys_segments == 1)
300            req->bio->bi_seg_front_size = seg_size;
301        if (next->nr_phys_segments == 1)
302            next->biotail->bi_seg_back_size = seg_size;
303        total_phys_segments--;
304    }
305
306    if (total_phys_segments > queue_max_segments(q))
307        return 0;
308
309    if (blk_integrity_rq(req) && blk_integrity_merge_rq(q, req, next))
310        return 0;
311
312    /* Merge is OK... */
313    req->nr_phys_segments = total_phys_segments;
314    return 1;
315}
316
317/**
318 * blk_rq_set_mixed_merge - mark a request as mixed merge
319 * @rq: request to mark as mixed merge
320 *
321 * Description:
322 * @rq is about to be mixed merged. Make sure the attributes
323 * which can be mixed are set in each bio and mark @rq as mixed
324 * merged.
325 */
326void blk_rq_set_mixed_merge(struct request *rq)
327{
328    unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
329    struct bio *bio;
330
331    if (rq->cmd_flags & REQ_MIXED_MERGE)
332        return;
333
334    /*
335     * @rq will no longer represent mixable attributes for all the
336     * contained bios. It will just track those of the first one.
337     * Distributes the attributs to each bio.
338     */
339    for (bio = rq->bio; bio; bio = bio->bi_next) {
340        WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
341                 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
342        bio->bi_rw |= ff;
343    }
344    rq->cmd_flags |= REQ_MIXED_MERGE;
345}
346
347static void blk_account_io_merge(struct request *req)
348{
349    if (blk_do_io_stat(req)) {
350        struct hd_struct *part;
351        int cpu;
352
353        cpu = part_stat_lock();
354        part = disk_map_sector_rcu(req->rq_disk, blk_rq_pos(req));
355
356        part_round_stats(cpu, part);
357        part_dec_in_flight(part, rq_data_dir(req));
358
359        part_stat_unlock();
360    }
361}
362
363/*
364 * Has to be called with the request spinlock acquired
365 */
366static int attempt_merge(struct request_queue *q, struct request *req,
367              struct request *next)
368{
369    if (!rq_mergeable(req) || !rq_mergeable(next))
370        return 0;
371
372    /*
373     * Don't merge file system requests and discard requests
374     */
375    if ((req->cmd_flags & REQ_DISCARD) != (next->cmd_flags & REQ_DISCARD))
376        return 0;
377
378    /*
379     * Don't merge discard requests and secure discard requests
380     */
381    if ((req->cmd_flags & REQ_SECURE) != (next->cmd_flags & REQ_SECURE))
382        return 0;
383
384    /*
385     * not contiguous
386     */
387    if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
388        return 0;
389
390    if (rq_data_dir(req) != rq_data_dir(next)
391        || req->rq_disk != next->rq_disk
392        || next->special)
393        return 0;
394
395    /*
396     * If we are allowed to merge, then append bio list
397     * from next to rq and release next. merge_requests_fn
398     * will have updated segment counts, update sector
399     * counts here.
400     */
401    if (!ll_merge_requests_fn(q, req, next))
402        return 0;
403
404    /*
405     * If failfast settings disagree or any of the two is already
406     * a mixed merge, mark both as mixed before proceeding. This
407     * makes sure that all involved bios have mixable attributes
408     * set properly.
409     */
410    if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
411        (req->cmd_flags & REQ_FAILFAST_MASK) !=
412        (next->cmd_flags & REQ_FAILFAST_MASK)) {
413        blk_rq_set_mixed_merge(req);
414        blk_rq_set_mixed_merge(next);
415    }
416
417    /*
418     * At this point we have either done a back merge
419     * or front merge. We need the smaller start_time of
420     * the merged requests to be the current request
421     * for accounting purposes.
422     */
423    if (time_after(req->start_time, next->start_time))
424        req->start_time = next->start_time;
425
426    req->biotail->bi_next = next->bio;
427    req->biotail = next->biotail;
428
429    req->__data_len += blk_rq_bytes(next);
430
431    elv_merge_requests(q, req, next);
432
433    /*
434     * 'next' is going away, so update stats accordingly
435     */
436    blk_account_io_merge(next);
437
438    req->ioprio = ioprio_best(req->ioprio, next->ioprio);
439    if (blk_rq_cpu_valid(next))
440        req->cpu = next->cpu;
441
442    /* owner-ship of bio passed from next to req */
443    next->bio = NULL;
444    __blk_put_request(q, next);
445    return 1;
446}
447
448int attempt_back_merge(struct request_queue *q, struct request *rq)
449{
450    struct request *next = elv_latter_request(q, rq);
451
452    if (next)
453        return attempt_merge(q, rq, next);
454
455    return 0;
456}
457
458int attempt_front_merge(struct request_queue *q, struct request *rq)
459{
460    struct request *prev = elv_former_request(q, rq);
461
462    if (prev)
463        return attempt_merge(q, prev, rq);
464
465    return 0;
466}
467

Archive Download this file



interactive