Re: [PATCH 3/4] Move CONFIG_XFS handling to meson.build

[Thomas Huth <thuth@redhat.com>]

On 10/12/2021 11.10, Paolo Bonzini wrote:
> On 12/10/21 09:46, Thomas Huth wrote:
>>>
>>> platform_test_xfs_fd() is only used to decide whether to invoke 
>>> XFS_IOC_DIOINFO; but failures of XFS_IOC_DIOINFO are ignored anyway, so 
>>> we can get rid of is_xfs in BDRVRawState, too.
>>
>> After staring at the code for a while, I wonder why we're not simply using 
>> fstat() here instead to get the st_blksize value... wouldn't that be 
>> better anyway since it also works with other file system types?
> 
> The value that XFS_IOC_DIOINFO returns is the logical sector size of the 
> underlying device; it should be 512 or 4096, but more likely 512.  It can be 
> smaller than st_blksize, because often it will be if it is 512 but the 
> st_blksize is usually 4096.
> 
> If it is wrong, QEMU will do unnecessary read/modify/write operations for 
> disk writes that are not 4K-aligned.

Ok, true, I've checked it and XFS_IOC_DIOINFO return 512 on my laptop 
indeed, while fstat->st_blksize is 4096 instead. So it's not the same :-/

  Thomas