On 7/8/20 6:32 AM, Richard W.M. Jones wrote:
On Tue, Jul 07, 2020 at 05:22:47PM -0500, Eric Blake wrote:
This reverts commit 2c5aec42cf04c567639bd885cf12a57192229215.
Now that we have an easy way to generate aligned extents from the plugin, where the client will never see an unaligned extent transition, it's time to put it to use.
Signed-off-by: Eric Blake <eblake@redhat.com>
--- Here's what I used on the command line; I need to turn it into a formal test:
$ ./nbdkit -U - eval pread='exit 1' get_size='echo 8' can_extents='exit 0' \ extents='printf "0 3\n3 4 hole,zero\n7 1\n"' --run \ 'qemu-img map -f raw --output=json $uri' $ ./nbdkit -U - --filter=swab eval pread='exit 1' get_size='echo 8' can_extents='exit 0' extents='printf "0 3\n3 4 hole,zero\n7 1\n"' --run 'qemu-img map -f raw --output=json $uri'
The fact that I had to supply can_extents='exit 0' was another bug that I just fixed in the eval plugin.
except it's currently failing: nbdkit: eval[1]: error: swab: requests to this filter must be aligned nbdkit: backend.c:621: backend_extents: Assertion `*err' failed. qemu-img: Could not read file metadata: Input/output error
I would add -v to the command line, as it will show you exactly what requests qemu-img is making. The assert failure seems to be a separate and worrying problem.
Yes, the assert failure is a pre-existing bug in the swab filter, which I just fixed: basically, filters MUST set *err when failing, so the swab.c:is_aligned() helper needed an additional parameter. But I hit the assert because of my bug in patch 2 of using a copy instead of editing through a pointer, such that the client wasn't seeing the edited aligned values after all. With that fixed, I'm now ready to formalize the test, and push this patch and a similar one for the blocksize filter. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org