On 5/12/2018 5:01 PM, Vladimir Sementsov-Ogievskiy wrote: > 03.12.2018 13:14, Anton Nefedov wrote: >> If COW areas of the newly allocated clusters are zeroes on the backing image, >> efficient bdrv_write_zeroes(flags=BDRV_REQ_ALLOCATE) can be used on the whole >> cluster instead of writing explicit zero buffers later in perform_cow(). >> >> iotest 060: >> write to the discarded cluster does not trigger COW anymore. >> Use a backing image instead. >> >> Signed-off-by: Anton Nefedov <anton.nefe...@virtuozzo.com> >> Reviewed-by: Alberto Garcia <be...@igalia.com> >> --- >> qapi/block-core.json | 4 +- >> block/qcow2.h | 6 +++ >> block/qcow2-cluster.c | 2 +- >> block/qcow2.c | 80 +++++++++++++++++++++++++++++++++++++- >> block/trace-events | 1 + >> tests/qemu-iotests/060 | 26 ++++++++----- >> tests/qemu-iotests/060.out | 5 ++- >> 7 files changed, 109 insertions(+), 15 deletions(-) >> >> diff --git a/qapi/block-core.json b/qapi/block-core.json >> index d4fe710836..50598aa8fe 100644 >> --- a/qapi/block-core.json >> +++ b/qapi/block-core.json >> @@ -3004,6 +3004,8 @@ >> # >> # @cor_write: a write due to copy-on-read (since 2.11) >> # >> +# @cluster_alloc_space: an allocation of file space for a cluster (since >> 4.0) >> +# >> # Since: 2.9 >> ## >> { 'enum': 'BlkdebugEvent', 'prefix': 'BLKDBG', >> @@ -3022,7 +3024,7 @@ >> 'pwritev_rmw_tail', 'pwritev_rmw_after_tail', 'pwritev', >> 'pwritev_zero', 'pwritev_done', 'empty_image_prepare', >> 'l1_shrink_write_table', 'l1_shrink_free_l2_clusters', >> - 'cor_write'] } >> + 'cor_write', 'cluster_alloc_space'] } >> >> ## >> # @BlkdebugInjectErrorOptions: >> diff --git a/block/qcow2.h b/block/qcow2.h >> index 8662b68575..8a64077897 100644 >> --- a/block/qcow2.h >> +++ b/block/qcow2.h >> @@ -389,6 +389,12 @@ typedef struct QCowL2Meta >> */ >> Qcow2COWRegion cow_end; >> >> + /** >> + * Indicates that COW regions are already handled and do not require >> + * any more processing. >> + */ >> + bool skip_cow; >> + >> /** >> * The I/O vector with the data from the actual guest write request. >> * If non-NULL, this is meant to be merged together with the data >> diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c >> index d37fe08b3d..3685c5f67e 100644 >> --- a/block/qcow2-cluster.c >> +++ b/block/qcow2-cluster.c >> @@ -806,7 +806,7 @@ static int perform_cow(BlockDriverState *bs, QCowL2Meta >> *m) >> assert(start->offset + start->nb_bytes <= end->offset); >> assert(!m->data_qiov || m->data_qiov->size == data_bytes); >> >> - if (start->nb_bytes == 0 && end->nb_bytes == 0) { >> + if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) { >> return 0; >> } >> >> diff --git a/block/qcow2.c b/block/qcow2.c >> index 991d6ac91b..027188a1a3 100644 >> --- a/block/qcow2.c >> +++ b/block/qcow2.c >> @@ -2015,6 +2015,11 @@ static bool merge_cow(uint64_t offset, unsigned bytes, >> continue; >> } >> >> + /* If COW regions are handled already, skip this too */ >> + if (m->skip_cow) { >> + continue; >> + } >> + >> /* The data (middle) region must be immediately after the >> * start region */ >> if (l2meta_cow_start(m) + m->cow_start.nb_bytes != offset) { >> @@ -2040,6 +2045,68 @@ static bool merge_cow(uint64_t offset, unsigned bytes, >> return false; >> } >> >> +static bool is_unallocated(BlockDriverState *bs, int64_t offset, int64_t >> bytes) >> +{ >> + int64_t nr; >> + return !bytes || >> + (!bdrv_is_allocated_above(bs, NULL, offset, bytes, &nr) && nr == >> bytes); > > bdrv_is_allocated_above may return error < 0 >
Probably I just took is_zero() as an example. But somewhere there's even a rationale (bdrv_co_do_copy_on_readv): ret = bdrv_is_allocated(bs, cluster_offset, MIN(cluster_bytes, max_transfer), &pnum); if (ret < 0) { /* Safe to treat errors in querying allocation as if * unallocated; we'll probably fail again soon on the * read, but at least that will set a decent errno. */ pnum = MIN(cluster_bytes, max_transfer); } >> +} >> + >> +static bool is_zero_cow(BlockDriverState *bs, QCowL2Meta *m) >> +{ >> + /* This check is designed for optimization shortcut so it must be >> + * efficient. >> + * Instead of is_zero(), use is_unallocated() as it is faster (but not >> + * as accurate and can result in false negatives). */ >> + return is_unallocated(bs, m->offset + m->cow_start.offset, >> + m->cow_start.nb_bytes) && >> + is_unallocated(bs, m->offset + m->cow_end.offset, >> + m->cow_end.nb_bytes); >> +} >> + >> +static int handle_alloc_space(BlockDriverState *bs, QCowL2Meta *l2meta) >> +{ >> + BDRVQcow2State *s = bs->opaque; >> + QCowL2Meta *m; >> + >> + if (!(bs->file->bs->supported_zero_flags & BDRV_REQ_ALLOCATE)) { >> + return 0; >> + } >> + >> + if (bs->encrypted) { >> + return 0; >> + } >> + >> + for (m = l2meta; m != NULL; m = m->next) { >> + int ret; >> + >> + if (!m->cow_start.nb_bytes && !m->cow_end.nb_bytes) { >> + continue; >> + } >> + >> + if (!is_zero_cow(bs, m)) { >> + continue; >> + } >> + >> + BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_SPACE); >> + /* instead of writing zero COW buffers, >> + efficiently zero out the whole clusters */ >> + ret = bdrv_co_pwrite_zeroes(bs->file, m->alloc_offset, >> + m->nb_clusters * s->cluster_size, >> + BDRV_REQ_ALLOCATE); >> + if (ret < 0) { >> + if (ret != -ENOTSUP && ret != -EAGAIN) { > > why do you ignore ENOTSUP? > This might particularly happen after first pwrite_zeroes() to file-posix. It generally reports BDRV_REQ_ALLOCATE on open, but then it might reset the flag if fallocate() returns ENOTSUP. Nevermind file-posix; with any file driver what can we possibly do here: this operation is non-mandatory and the error is not critical. Never trying again doesn't look quite right, the file driver can reset BDRV_REQ_ALLOCATE if it likes and we won't bother it anymore. > And also, hmm, should not we have problems with serializing? If already > started unaligned request, > it is serializing.. And now, we are going to start nested serializing > request, which will definitely > return EAGAIN.. > The request we're processing here is on qcow2 level yet, it won't affect the bs->file serialization queue. [..]