Re: [PATCH] nbd/server: Advertise MULTI_CONN for shared writable exports

[Kevin Wolf]

Am 27.08.2021 um 17:09 hat Eric Blake geschrieben:
> According to the NBD spec, a server advertising
> NBD_FLAG_CAN_MULTI_CONN promises that multiple client connections will
> not see any cache inconsistencies: when properly separated by a single
> flush, actions performed by one client will be visible to another
> client, regardless of which client did the flush.  We satisfy these
> conditions in qemu because our block layer serializes any overlapping
> operations (see bdrv_find_conflicting_request and friends): no matter
> which client performs a flush, parallel requests coming from distinct
> NBD clients will still be well-ordered by the time they are passed on
> to the underlying device, with no caching in qemu proper to allow
> stale results to leak after a flush.
> 
> We don't want to advertise MULTI_CONN when we know that a second
> client can connect (which is the default for qemu-nbd, but not for QMP
> nbd-server-add),

Do you mean when a second client _can't_ connect?

> so it does require a QAPI addition.  But other than
> that, the actual change to advertise the bit for writable servers is
> fairly small.  The harder part of this patch is setting up an iotest
> to demonstrate behavior of multiple NBD clients to a single server.
> It might be possible with parallel qemu-io processes, but concisely
> managing that in shell is painful.

I think it should be fairly straightforward in a Python test case.

Another option is using a single QEMU or QSD instance that has multiple
-blockdev for the same NBD server. For the server these are multiple
clients, even if all connnection come from a single process.

> I found it easier to do by relying
> on the libnbd project's nbdsh, which means this test will be skipped
> on platforms where that is not available.
> 
> Signed-off-by: Eric Blake <eblake@redhat.com>

> diff --git a/docs/interop/nbd.txt b/docs/interop/nbd.txt
> index 10ce098a29bf..d03910f1e9eb 100644
> --- a/docs/interop/nbd.txt
> +++ b/docs/interop/nbd.txt
> @@ -68,3 +68,4 @@ NBD_CMD_BLOCK_STATUS for "qemu:dirty-bitmap:", NBD_CMD_CACHE
>  * 4.2: NBD_FLAG_CAN_MULTI_CONN for shareable read-only exports,
>  NBD_CMD_FLAG_FAST_ZERO
>  * 5.2: NBD_CMD_BLOCK_STATUS for "qemu:allocation-depth"
> +* 6.2: NBD_FLAG_CAN_MULTI_CONN for shareable writable exports
> diff --git a/docs/tools/qemu-nbd.rst b/docs/tools/qemu-nbd.rst
> index 5643da26e982..81be32164a55 100644
> --- a/docs/tools/qemu-nbd.rst
> +++ b/docs/tools/qemu-nbd.rst
> @@ -138,8 +138,7 @@ driver options if ``--image-opts`` is specified.
>  .. option:: -e, --shared=NUM
> 
>    Allow up to *NUM* clients to share the device (default
> -  ``1``), 0 for unlimited. Safe for readers, but for now,
> -  consistency is not guaranteed between multiple writers.
> +  ``1``), 0 for unlimited.
> 
>  .. option:: -t, --persistent

If qemu-nbd supports a maximum number of connections rather than just a
bool...

> diff --git a/qapi/block-export.json b/qapi/block-export.json
> index 0ed63442a819..b2085a9fdd4c 100644
> --- a/qapi/block-export.json
> +++ b/qapi/block-export.json
> @@ -95,11 +95,15 @@
>  #                    the metadata context name "qemu:allocation-depth" to
>  #                    inspect allocation details. (since 5.2)
>  #
> +# @shared: True if the server should advertise that multiple clients may
> +#          connect, default false. (since 6.2)
> +#
>  # Since: 5.2
>  ##
>  { 'struct': 'BlockExportOptionsNbd',
>    'base': 'BlockExportOptionsNbdBase',
> -  'data': { '*bitmaps': ['str'], '*allocation-depth': 'bool' } }
> +  'data': { '*bitmaps': ['str'], '*allocation-depth': 'bool',
> +             '*shared': 'bool' } }

...wouldn't it be better to mirror this in the QAPI interface?

I think eventually we want to add everything missing to the built-in NBD
server and then change qemu-nbd to use it instead of managing the
connections itself. So I'm not sure if diverging here is a good idea.

Kevin