block: split discard into aligned requests

When a disk has large discard_granularity and small max_discard_sectors,
discards are not split with optimal alignment.  In the limit case of
discard_granularity == max_discard_sectors, no request could be aligned
correctly, so in fact you might end up with no discarded logical blocks
at all.

Another example that helps showing the condition in the patch is with
discard_granularity == 64, max_discard_sectors == 128.  A request that is
submitted for 256 sectors 2..257 will be split in two: 2..129, 130..257.
However, only 2 aligned blocks out of 3 are included in the request;
128..191 may be left intact and not discarded.  With this patch, the
first request will be truncated to ensure good alignment of what's left,
and the split will be 2..127, 128..255, 256..257.  The patch will also
take into account the discard_alignment.

At most one extra request will be introduced, because the first request
will be reduced by at most granularity-1 sectors, and granularity
must be less than max_discard_sectors.  Subsequent requests will run
on round_down(max_discard_sectors, granularity) sectors, as in the
current code.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Tested-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/block/blk-lib.c b/block/blk-lib.c
index 16b06f62e68ca187a05cf46fe8232c4acda02390..19cc761cacb2a4b71fe9d4579226025324ffddf5 100644 (file)
--- a/block/blk-lib.c
+++ b/block/blk-lib.c
@@ -44,7 +44,7 @@ int blkdev_issue_discard(struct block_device *bdev, sector_t sector,
        struct request_queue *q = bdev_get_queue(bdev);
        int type = REQ_WRITE | REQ_DISCARD;
        unsigned int max_discard_sectors;
-       unsigned int granularity;
+       unsigned int granularity, alignment, mask;
        struct bio_batch bb;
        struct bio *bio;
        int ret = 0;
@@ -57,10 +57,12 @@ int blkdev_issue_discard(struct block_device *bdev, sector_t sector,
 
        /* Zero-sector (unknown) and one-sector granularities are the same.  */
        granularity = max(q->limits.discard_granularity >> 9, 1U);
+       mask = granularity - 1;
+       alignment = (bdev_discard_alignment(bdev) >> 9) & mask;
 
        /*
         * Ensure that max_discard_sectors is of the proper
-        * granularity
+        * granularity, so that requests stay aligned after a split.
         */
        max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
        max_discard_sectors = round_down(max_discard_sectors, granularity);
@@ -80,25 +82,37 @@ int blkdev_issue_discard(struct block_device *bdev, sector_t sector,
        bb.wait = &wait;
 
        while (nr_sects) {
+               unsigned int req_sects;
+               sector_t end_sect;
+
                bio = bio_alloc(gfp_mask, 1);
                if (!bio) {
                        ret = -ENOMEM;
                        break;
                }
 
+               req_sects = min_t(sector_t, nr_sects, max_discard_sectors);
+
+               /*
+                * If splitting a request, and the next starting sector would be
+                * misaligned, stop the discard at the previous aligned sector.
+                */
+               end_sect = sector + req_sects;
+               if (req_sects < nr_sects && (end_sect & mask) != alignment) {
+                       end_sect =
+                               round_down(end_sect - alignment, granularity)
+                               + alignment;
+                       req_sects = end_sect - sector;
+               }
+
                bio->bi_sector = sector;
                bio->bi_end_io = bio_batch_end_io;
                bio->bi_bdev = bdev;
                bio->bi_private = &bb;
 
-               if (nr_sects > max_discard_sectors) {
-                       bio->bi_size = max_discard_sectors << 9;
-                       nr_sects -= max_discard_sectors;
-                       sector += max_discard_sectors;
-               } else {
-                       bio->bi_size = nr_sects << 9;
-                       nr_sects = 0;
-               }
+               bio->bi_size = req_sects << 9;
+               nr_sects -= req_sects;
+               sector = end_sect;
 
                atomic_inc(&bb.done);
                submit_bio(type, bio);

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 4e72a9d48232d513b5b2fe44d973de8dfeb1e9e4..281516ae8b4ef37256af453e00c1c098d315f769 100644 (file)
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -1139,6 +1139,16 @@ static inline int queue_limit_discard_alignment(struct queue_limits *lim, sector
                & (lim->discard_granularity - 1);
 }
 
+static inline int bdev_discard_alignment(struct block_device *bdev)
+{
+       struct request_queue *q = bdev_get_queue(bdev);
+
+       if (bdev != bdev->bd_contains)
+               return bdev->bd_part->discard_alignment;
+
+       return q->limits.discard_alignment;
+}
+
 static inline unsigned int queue_discard_zeroes_data(struct request_queue *q)
 {
        if (q->limits.max_discard_sectors && q->limits.discard_zeroes_data == 1)