Re: [PATCH for-4.2 v2 3/3] block/file-posix: Let post-EOF fallocate serialize

[Vladimir Sementsov-Ogievskiy]

22.08.2020 20:04, Vladimir Sementsov-Ogievskiy wrote:
> 22.08.2020 20:03, Vladimir Sementsov-Ogievskiy wrote:
> > 01.11.2019 18:25, Max Reitz wrote:
> > > The XFS kernel driver has a bug that may cause data corruption for qcow2
> > > images as of qemu commit c8bb23cbdbe32f.  We can work around it by
> > > treating post-EOF fallocates as serializing up until infinity (INT64_MAX
> > > in practice).
> >
> > Hi! I'm doing some investigation, and here is an interesting result:
> >
> > Consider the following test:
> >
> > img=/ssd/x.qcow2; ./qemu-img create -f qcow2 $img 16G; ./qemu-img bench -c 
> > 50000 -d 64 -f qcow2 -o 1k -s 64k -t none -w $img
> >
> >
> > Bisecting results changes between 2.12 and 5.1, I found the following:
> >
> > 2.12: ~20s
> >
> > ....
> >
> > c8bb23cbdbe32 "qcow2: skip writing zero buffers to empty COW areas"  -> becomes 
> > ~12s  [1]
> >
> > ....
> >
> >
> > 292d06b925b27 "block/file-posix: Let post-EOF fallocate serialize"   -> becomes 
> > ~9s  [2]
> >
> > ....
> >
> > v5.1 ~9s
> >
> >
> > And [1] is obvious, it is the main purpose of c8bb23cbdbe32. But [2] is a 
> > surprise for me.. Any ideas?
> >
> > ===
> >
> > just to check: staying at c8bb23cbdbe32 I revert c8bb23cbdbe32 and get again 
> > ~19.7s. So [2] doesn't substitute [1].
>
> Note, it's all ext4.


Let's go further:

If I add -n, to use aio native:
./qemu-img create -f qcow2 $img 16G; ./qemu-img bench -c 50000 -d 64 -f qcow2 
-o 1k -s 64k -t none -n -w $img;

Then this patch doesn't have such effect..

But if consider the test without an offset on hdd, the effect is very 
significant:

./qemu-img create -f qcow2 $img 16G; ./qemu-img bench -c 45000 -d 64 -f qcow2  
-s 16k -t none -n -w $img;

v2.12 ~10.4s
c8bb23cbdbe32^ ~10.6s
c8bb23cbdbe32 qcow2: skip writing zero buffers to empty COW areas: ~16.7s : 
degradation!!!
292d06b925b27^ ~16.4s
292d06b925 block/file-posix: Let post-EOF fallocate serialize: ~7.6s: great 
improvement !

--
Best regards,
Vladimir