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

[Fiona Ebner <f.ebner@proxmox.com>]

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@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.

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.