Re: [Qemu-devel] [Qemu-block] Request for clarification on qemu-img convert behavior zeroing target host_device

[Eric Blake]

On 12/13/18 8:49 AM, Kevin Wolf wrote:

> > > We observe that in Fedora 29 the qemu-img, before imaging the disk, it fully 
> > > zeroes it. Taking into account the disk size, the whole process now takes 35 
> > > minutes instead of 50 seconds. This causes the ironic-python-agent operation to 
> > > time-out. The Fedora 27 qemu-img doesn't do that.
> >
> > Known issue; Nir and Rich have posted a previous thread on the topic, and
> > the conclusion is that we need to make qemu-img smarter about NOT requesting
> > pre-zeroing of devices where that is more expensive than just zeroing as we
> > go.
> > https://lists.gnu.org/archive/html/qemu-devel/2018-11/msg01182.html
>
> Yes, we should be careful to avoid the fallback in this case.
>
> However, how could this ever go from 50 seconds for writing the whole
> image to 35 minutes?! Even if you end up writing the whole image twice
> because you write zeros first and then overwrite them everywhere with
> data, shouldn't the maximum be doubling the time, i.e. 100 seconds?
>
> Why is the write_zeroes fallback _that_ slow? It will also hit guests
> that request write_zeroes, so I feel this is worth investigating a bit
> more nevertheless.
>
> Can you check with strace which operation actually succeeds writing
> zeros to /dev/sda? The first thing we try is fallocate with
> FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE. This should always be fast,
> so I suppose this fails in your case. The next thing is BLKZEROOUT,
> which I think can do a fallback in the kernel. Does this return success?
> Otherwise we have another fallback mechanism inside of QEMU, which would
> use normal pwrite calls with a zeroed buffer.

It may also be a case of poor lseek(SEEK_HOLE) performance on the source 
(a known issue with at least some versions of tmpfs). The way qemu-img 
queries for block status, it ends up repeatedly hammering on lseek(), 
and if lseek() is already O(n) instead of O(1) in behavior, that 
explodes into some O(n^2) scaling because qemu-img isn't caching the 
answers it got previously.

> Once we know which mechanism is used, we can look into why it is so
> abysmally slow.

Indeed, performance traces are important for issues like this.

--
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org