| Message ID | 20190114111744.113188-1-anton.nefedov@virtuozzo.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | qcow2: cluster space preallocation | expand |
On 14/1/2019 2:18 PM, Anton Nefedov wrote: > This pull request is to start to improve a few performance points of > qcow2 format: > > 1. non cluster-aligned write requests (to unallocated clusters) explicitly > pad data with zeroes if there is no backing data. > Resulting increase in ops number and potential cluster fragmentation > (on the host file) is already solved by: > ee22a9d qcow2: Merge the writing of the COW regions with the guest data > However, in case of zero COW regions, that can be avoided at all > but the whole clusters are preallocated and zeroed in a single > efficient write_zeroes() operation > > 2. moreover, efficient write_zeroes() operation can be used to preallocate > space megabytes (*configurable number) ahead which gives noticeable > improvement on some storage types (e.g. distributed storage) > where the space allocation operation might be expensive) > (Not included in this patchset since v6). > > 3. this will also allow to enable simultaneous writes to the same unallocated > cluster after the space has been allocated & zeroed but before > the first data is written and the cluster is linked to L2. > (Not included in this patchset). > > Efficient write_zeroes usually implies that the blocks are not actually > written to but only reserved and marked as zeroed by the storage. > In this patchset, file-posix driver is marked as supporting this operation > if it supports (/configured to support) fallocate() operation. > > Existing bdrv_write_zeroes() falls back to writing zero buffers if > write_zeroes is not supported by the driver. > A new flag (BDRV_REQ_ALLOCATE) is introduced to avoid that but return ENOTSUP. > Such allocate requests are also implemented to possibly overlap with the > other requests. No wait is performed but an error returned in such case as well. > So the operation should be considered advisory and a fallback scenario still > handled by the caller (in this case, qcow2 driver). > > simple perf test: > > qemu-img create -f qcow2 test.img 4G && \ > qemu-img bench -c $((1024*1024)) -f qcow2 -n -s 4k -t none -w test.img > > test results (seconds): > > +-----------+-------+------+-------+------+------+ > | file | before | after | gain | > +-----------+-------+------+-------+------+------+ > | ssd | 61.153 | 36.313 | 41% | > | hdd | 112.676 | 122.056 | -8% | > +-----------+--------------+--------------+------+ > ping
new in v12: - patch 9: pre-write overlap check added v11: http://lists.nongnu.org/archive/html/qemu-devel/2018-12/msg04342.html - patch 4, 9: fixed commentary format - patch 4: removed one hunk with a dead check - patch 5: added commentary to BDRV_REQ_ALLOCATE definition - new auxiliary patch 6 for the following patch-7 change: - patch 7: reset BDRV_REQ_ALLOCATE from supported flag if CONFIG_FALLOCATE is false - patch 9: add commentary about missing qcow2_pre_write_overlap_check(). Omit redundant changes in the test 060. v10: http://lists.nongnu.org/archive/html/qemu-devel/2018-12/msg00121.html - patches 1-3,6,7: rebase after REQ_WRITE_UNCHANGED - patch 3: drop supported_zero_flags. My bad, no write_zeroes in quorum. - patch 4: almost trivial rebase. RB-tags not stripped. Choose another constant for BDRV_REQ_ALLOCATE - patch 5: rebase. Instead of marking REQ_ALLOCATE serialising, accompany it with REQ_SERIALISING. - patch 7: add symmetric copy-on-read change - patch 8: trivial rebase. RB-tags not stripped. ---- This pull request is to start to improve a few performance points of qcow2 format: 1. non cluster-aligned write requests (to unallocated clusters) explicitly pad data with zeroes if there is no backing data. Resulting increase in ops number and potential cluster fragmentation (on the host file) is already solved by: ee22a9d qcow2: Merge the writing of the COW regions with the guest data However, in case of zero COW regions, that can be avoided at all but the whole clusters are preallocated and zeroed in a single efficient write_zeroes() operation 2. moreover, efficient write_zeroes() operation can be used to preallocate space megabytes (*configurable number) ahead which gives noticeable improvement on some storage types (e.g. distributed storage) where the space allocation operation might be expensive) (Not included in this patchset since v6). 3. this will also allow to enable simultaneous writes to the same unallocated cluster after the space has been allocated & zeroed but before the first data is written and the cluster is linked to L2. (Not included in this patchset). Efficient write_zeroes usually implies that the blocks are not actually written to but only reserved and marked as zeroed by the storage. In this patchset, file-posix driver is marked as supporting this operation if it supports (/configured to support) fallocate() operation. Existing bdrv_write_zeroes() falls back to writing zero buffers if write_zeroes is not supported by the driver. A new flag (BDRV_REQ_ALLOCATE) is introduced to avoid that but return ENOTSUP. Such allocate requests are also implemented to possibly overlap with the other requests. No wait is performed but an error returned in such case as well. So the operation should be considered advisory and a fallback scenario still handled by the caller (in this case, qcow2 driver). simple perf test: qemu-img create -f qcow2 test.img 4G && \ qemu-img bench -c $((1024*1024)) -f qcow2 -n -s 4k -t none -w test.img test results (seconds): +-----------+-------+------+-------+------+------+ | file | before | after | gain | +-----------+-------+------+-------+------+------+ | ssd | 61.153 | 36.313 | 41% | | hdd | 112.676 | 122.056 | -8% | +-----------+--------------+--------------+------+ Anton Nefedov (10): mirror: inherit supported write/zero flags blkverify: set supported write/zero flags quorum: set supported write flags block: introduce BDRV_REQ_ALLOCATE flag block: treat BDRV_REQ_ALLOCATE as serialising file-posix: reset fallocate-related flags without CONFIG_FALLOCATE* file-posix: support BDRV_REQ_ALLOCATE block: support BDRV_REQ_ALLOCATE in passthrough drivers qcow2: skip writing zero buffers to empty COW areas iotest 134: test cluster-misaligned encrypted write qapi/block-core.json | 4 +- block/qcow2.h | 6 +++ include/block/block.h | 13 +++++- include/block/block_int.h | 3 +- block/blkdebug.c | 2 +- block/blkverify.c | 10 ++++- block/copy-on-read.c | 4 +- block/file-posix.c | 16 +++++-- block/io.c | 45 +++++++++++++++---- block/mirror.c | 8 +++- block/qcow2-cluster.c | 2 +- block/qcow2.c | 91 +++++++++++++++++++++++++++++++++++++- block/quorum.c | 19 +++++++- block/raw-format.c | 2 +- block/trace-events | 1 + tests/qemu-iotests/060 | 7 ++- tests/qemu-iotests/060.out | 5 ++- tests/qemu-iotests/134 | 9 ++++ tests/qemu-iotests/134.out | 10 +++++ 19 files changed, 229 insertions(+), 28 deletions(-)