Re: [Qemu-devel] [PATCH 1/1] block: fix inability to start VM with native AIO

[Stefan Hajnoczi <stefanha@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2550 bytes --]

On Tue, Dec 22, 2015 at 09:59:46AM +0300, Denis V. Lunev wrote:
> error: Failed to start domain rhel7
> error: internal error: process exited while connecting to monitor:
>     2015-12-22T06:55:18.812637Z qemu-system-x86_64:
>     -drive file=/var/lib/libvirt/images/rhel7.qcow2,if=none,
>         id=drive-scsi0-0-0-0,format=qcow2,cache=none,aio=native:
>     aio=native was specified, but it requires cache.direct=on,
>     which was not specified.
> 
> cache=none option was specified as seen above while the VM is unable to
> start. The patch properly passed BDRV_O_NOCACHE to underlying layer.
> 
> The problem is revealed with
>     commit d657c0c289e944fc22289f5c318f48da87d79dcb
>     Author: Kevin Wolf <kwolf@redhat.com>
>     Date:   Tue Dec 15 11:35:36 2015 +0100
> 
>         raw-posix: Make aio=native option binding
> 
> Signed-off-by: Denis V. Lunev <den@openvz.org>
> CC: Kevin Wolf <kwolf@redhat.com>
> ---
>  block.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/block.c b/block.c
> index 411edbf..fe0fbbc 100644
> --- a/block.c
> +++ b/block.c
> @@ -990,6 +990,7 @@ static int bdrv_open_common(BlockDriverState *bs, BdrvChild *file,
>      bs->opaque = g_malloc0(drv->instance_size);
>  
>      /* Apply cache mode options */
> +    update_flags_from_options(&open_flags, opts);
>      update_flags_from_options(&bs->open_flags, opts);
>      bdrv_set_enable_write_cache(bs, bs->open_flags & BDRV_O_CACHE_WB);

I tried to review this patch and failed to understand block.c flags
handling.  Perhaps there is a larger problem here because I see:

1. .bdrv_open() and friends take an int flags argument in addition to
   the already open bs node, which has bs->open_flags.  Some of the code
   that manipulates the flags argument in block.c updates bs->open_flags
   to keep them in sync.  Some code does not (e.g. BDRV_O_NO_BACKING).
   Is this a bug?

   Should int flags be removed in favor of just bs->open_flags?

2. The bdrv_open_common() open_flags local variable contains different
   values from bs->open_flags (i.e. BDRV_O_PROTOCOL).  I'm not sure
   whether this is intentional or a bug.

   Your patch syncs them but I wonder if it would be cleaner to remove
   the open_flags local (avoiding similar problems in the future)?

3. block/qcow2.c stashes .bdrv_open(flags) in s->flags.  I'm not sure
   if bs->open_flags can be used instead of s->flags.  That would be
   simpler.

Maybe someone can explain how this is all supposed to work.

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 473 bytes --]