Re: [Qemu-devel] [RFC] Safely reopening image files by stashing fds

[Kevin Wolf <kwolf@redhat.com>]

Am 16.08.2011 21:18, schrieb Supriya Kannery:
> On 08/09/2011 03:02 PM, supriya kannery wrote:
>  > Kevin Wolf wrote:
>  >> Am 09.08.2011 11:22, schrieb supriya kannery:
>  >>> Kevin Wolf wrote:
>  >>
>  >> What I meant is that in the end, with a generic bdrv_reopen(), we can
>  >> have raw-posix only call dup() and fcntl() instead of doing a
>  >> close()/open() sequence if it can satisfy the new flags this way. But
>  >> this would be an implementation detail and not be visible in the
>  >> interface.
>  >>
>  >> Kevin
>  >
>  > ok
>  > - thanks, Supriya
>  >
> 
> Though I started the RFC patch with defining BDRVReopenState, ended up
> in enhancing struct BlockDriver with .bdrv_reopen. bdrv_reopen 
> mplementation specific to respective driver is assigned to this
> function pointer.
> 
> Please find the implementation of O_DIRECT flag change, done in
> raw-posix.c. Similar implementation can be done for vmdk (with
> bdrv_reopen_commit and bdrv_reopen_abort as internal functions in
> vmdk.c for opening split files), sheepdog, nbd etc..

I haven't looked at the code yet, but this can't work. I have thought of
BDRVReopenState for a reason. If you have multiple files like VMDK can
have, then you want an all-or-nothing semantics of reopen.

With only .bdrv_reopen, I can't see a way to guarantee this. If
reopening the first file succeeds and the second one goes wrong, you
have the new flags on the first file, but the old flags on the second
one. You would have to re-reopen the first file with the old flags, but
then this is not guaranteed to succeed.

Kevin