On 22/5/2019 11:40 PM, Max Reitz wrote:
> On 16.05.19 16:27, Anton Nefedov wrote:
>> If COW areas of the newly allocated clusters are zeroes on the backing
>> image, efficient bdrv_write_zeroes(flags=BDRV_REQ_NO_FALLBACK) 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>
>> ---
>> qapi/block-core.json | 4 +-
>> block/qcow2.h | 6 +++
>> block/qcow2-cluster.c | 2 +-
>> block/qcow2.c | 93 +++++++++++++++++++++++++++++++++++++-
>> block/trace-events | 1 +
>> tests/qemu-iotests/060 | 7 ++-
>> tests/qemu-iotests/060.out | 5 +-
>> 7 files changed, 112 insertions(+), 6 deletions(-)
>
> [...]
>
>> diff --git a/block/qcow2.c b/block/qcow2.c
>> index 8e024007db..e6b1293ddf 100644
>> --- a/block/qcow2.c
>> +++ b/block/qcow2.c
>
> [...]
>
>> @@ -2145,6 +2150,80 @@ 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);
>
> It's a pity that this bdrv_is_allocated() throws away BDRV_BLOCK_ZERO
> information. If something in the backing chain has explicit zero
> clusters there, we could get the information for free, but this will
> consider it allocated.
>
> Wouldn't it make sense to make bdrv_co_block_status_above() public and
> then use that (with want_zero == false)?
>
> (But that can be done later, too, of course.)
>
or something like bdrv_has_non_zero_data(), with the argument
want_zero (maybe call it check_file in this case).
Yes, I'd rather implement it separately.
>> +}
>> +
>> +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 (!(s->data_file->bs->supported_zero_flags & BDRV_REQ_NO_FALLBACK)) {
>> + 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;
>> + }
>> +
>> + /*
>> + * instead of writing zero COW buffers,
>> + * efficiently zero out the whole clusters
>> + */
>> +
>> + ret = qcow2_pre_write_overlap_check(bs, 0, m->alloc_offset,
>> + m->nb_clusters *
>> s->cluster_size,
>> + true);
>> + if (ret < 0) {
>> + return ret;
>> + }
>> +
>> + BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_SPACE);
>> + ret = bdrv_co_pwrite_zeroes(s->data_file, m->alloc_offset,
>> + m->nb_clusters * s->cluster_size,
>> + BDRV_REQ_NO_FALLBACK);
>
> Mostly I'm writing this mail because of this. Zeroing the whole range
> seems inefficient to me for very large requests, and the commit message
> doesn't say anything about the reasoning behind unconditionally zeroing
> everything.
>
> Benchmarking looks like in most cases it is about equal to zeroing head
> and tail separately. But if I pre-filll an image with random data, it
> is much slower:
>
> $ qemu-img create -f qcow2 foo.qcow2 10G
> $ dd if=/dev/urandom of=foo.qcow2 bs=1M seek=1 count=4096
> $ sync
>
> Then doing large unaligned requests on it:
>
> $ ./qemu-img bench -w -t none -s $((400 * 1048576)) \
> -S $((401 * 1048576)) -o 32768 -c 10 -d 2 -f qcow2 foo.qcow2
>
> When zeroing the whole range, this usually takes around 25 s for me;
> just zeroing head and tail takes around 14 s. Without this patch, it
> takes around 14 s, too.
>
> On the other hand, when doing smaller requests on a single cluster
> (which is what this patch is for):
>
> $ ./qemu-img bench -w -t none -s 4096 -S 65536 -o 32768 \
> -f qcow2 foo.qcow2
>
> This takes around 26 s before this patch, 12 s with it, and like 30 to
> 40 when zeroing head and tail separately.
>
>
> Now, such large requests surely are the exception, as is allocating
> blocks in an area of the image that already contains data. However, I
> just want to ask back that zeroing the whole range unconditionally is
> done on purpose. Maybe it makes sense to split head and tail if they
> are like more than, I don't know, 16 MB apart? But the "I don't know"
> is the problem of course. Is there a way to make a good default?
>
It's indeed tricky to find a solution that performs better most
universally. I implied that writes are often shorter than a cluster,
and the bigger the cluster size the more the gain too. I traced a while
and did not even see a write larger than 1Mb from a freshly installed
centos VM.
Zeroing head and tail separately has an extra penalty being separate i/o
ops (probably why it loses/equates in performance anyway according to
your measurements).
So I'd say it's either unconditional, or if we want to protect the rare
cases from degradation too, we might fallback to writing
zero buffers. E.g. if zeroes_length/data_length ratio is too small.
Again, no idea what "too small" is.
Probably no other way to know than to measure with different borderline
ratios.
/Anton
