When a disk has a large discard_granularity, discards are not split with
optimal alignment; the pessimization gets bigger as discard_granularity
and max_discard_sectors become closer.

Take the limit case of discard_granularity == max_discard_sectors == 64.
Then, if a request is submitted for 256 sectors 2..257 it will be split
like this: 2..65, 66..129, 130..193, 194..257. None of these requests
is aligned, so in fact you might end up with no discarded logical blocks
at all. With this patch, the split will be 2..63, 64..127, 128..191,
192..255, 256..257. The patches also take the discard_alignment into
consideration.

For ease of debugging, patch 1 adds a sysfs entry for discard_alignment.
Patch 2 adjusts the computation of the granularity-adjusted
max_discard_sectors so that it prepares for the new code in patch 3,
which 3 actually adjusts the split.

v1->v2: added patch 1, fixed line length

Paolo Bonzini (3):
block: add sysfs entry for discard_alignment
block: reorganize rounding of max_discard_sectors
block: split discard into aligned requests

block/blk-lib.c | 41 ++++++++++++++++++++++++++++-------------
block/blk-sysfs.c | 11 +++++++++++
2 files changed, 39 insertions(+), 13 deletions(-)

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel

When a disk has a large discard_granularity, discards are not split with
optimal alignment; the pessimization gets bigger as discard_granularity
and max_discard_sectors become closer.

Take the limit case of discard_granularity == max_discard_sectors == 64.
Then, if a request is submitted for 256 sectors 2..257 it will be split
like this: 2..65, 66..129, 130..193, 194..257. None of these requests
is aligned, so in fact you might end up with no discarded logical blocks
at all. With this patch, the split will be 2..63, 64..127, 128..191,
192..255, 256..257. The patches also take the discard_alignment into
consideration.

Patch 1 adjusts the computation of the granularity-adjusted
max_discard_sectors so that it prepares for the new code in patch 2,
which actually adjusts the split.

v2->v3: drop addition of queue/discard_alignment to sysfs, use
correct alignment for partitions

Paolo Bonzini (2):
block: reorganize rounding of max_discard_sectors
block: split discard into aligned requests

block/blk-lib.c | 41 ++++++++++++++++++++++++++++-------------
include/linux/blkdev.h | 10 ++++++++++
2 files changed, 38 insertions(+), 13 deletions(-)

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel