[pve-devel] [PATCH storage] plugin: file size info: use fallback for actual size

[Wolfgang Bumiller]

On Thu, Jan 12, 2023 at 09:33:51AM +0100, Fiona Ebner wrote:
> Am 11.01.23 um 16:46 schrieb Thomas Lamprecht:
> > Am 28/11/2022 um 16:55 schrieb Fiona Ebner:
> >> The actual-size property is an optional property in the QAPI
> >> definition for ImageInfo. If it's not set, simply use the information
> >> from stat() as a fallback. This is essentially the same
> >> raw_get_allocated_file_size() in QEMU does anyways.
> >>
> >> Reported in the community forum:
> >> https://forum.proxmox.com/threads/118443/post-513421
> >>
> >> Signed-off-by: Fiona Ebner <f.ebner at proxmox.com>
> >> ---
> >>
> >> Thanks to Mira for setting up a GlusterFS instance and discussing the
> >> issue with me!
> >>
> >> I'm not sure why QEMU fails here, didn't see much that could go wrong
> >> beside the fstat() call failing. But our stat() call in the beginning
> >> of file_size_info already succeeded at that point :/ The mysteries of
> >> QEMU+GlusterFS.
> >>
> >> Also, it's a bit strange to call qemu-img info regardless of whether
> >> the image is a VM image or not. E.g., this results in the format
> >> property for backups to always be raw, rather than the backup format.
> >> Should we change that (for 8.0)?
> >>
> >>  PVE/Storage/Plugin.pm | 1 +
> >>  1 file changed, 1 insertion(+)
> >>
> >> diff --git a/PVE/Storage/Plugin.pm b/PVE/Storage/Plugin.pm
> >> index 8a41df1..7773ac3 100644
> >> --- a/PVE/Storage/Plugin.pm
> >> +++ b/PVE/Storage/Plugin.pm
> >> @@ -899,6 +899,7 @@ sub file_size_info {
> >>      }
> >>  
> >>      my ($size, $format, $used, $parent) = $info->@{qw(virtual-size format actual-size backing-filename)};
> >> +    $used ||= $st->blocks * 512;
> > 
> > in general OK, but can we really be sure that blocks are always 512 bytes?
> > 
> 
> Initially, I only checked 'man 2 stat' where this is the case:
> >                blkcnt_t  st_blocks;      /* Number of 512B blocks allocated */
> 
> But checking again now, 'perldoc -f stat' (File::stat mentions it uses
> Perl's builtin stat()) actually states:
> >              12 blocks   actual number of system-specific blocks allocated
> >                          on disk (often, but not always, 512 bytes each)
> 
> Trying to decipher the Perl 5 source code, I /think/ it will just use
> stat(2) on Linux (a quick check with strace seems to confirm this) and
> I'd say it would be surprising if not, but I'm not 100% sure.

There aren't too many choices there. It may switch to `statx` at some
point where there's only `stx_blocks` which is specifically documented
as "Number of 512B blocks allocated".
The terminology in the I/O layer around blocks, sectors and "groups of
512 bytes" is quite annoying really.

But hey, at least it's not `statvfs` which fills a struct with
"garbage" on some systems (see `man 3 statvfs` on freebsd14)

As for why glusterfs would fail with our stat() having succeeded, that's
just glusterfs. Don't question it...

> 
> That said, the original issue here was GlusterFS reporting an incorrect
> value (see the forum thread). The new fallback introduced by this patch
> would only help if 'qemu-img info' fails to determine the size for some
> other reason (it also just does st_blocks * 512 in
> raw_get_allocated_file_size()), so I'm not sure the patch is even worth
> it after all.