Re: [PATCH v4 10/12] hostmem: add a new memory backend based on POSIX shm_open()

[Markus Armbruster <armbru@redhat.com>]

Stefano Garzarella <sgarzare@redhat.com> writes:

> shm_open() creates and opens a new POSIX shared memory object.
> A POSIX shared memory object allows creating memory backend with an
> associated file descriptor that can be shared with external processes
> (e.g. vhost-user).
>
> The new `memory-backend-shm` can be used as an alternative when
> `memory-backend-memfd` is not available (Linux only), since shm_open()
> should be provided by any POSIX-compliant operating system.
>
> This backend mimics memfd, allocating memory that is practically
> anonymous. In theory shm_open() requires a name, but this is allocated
> for a short time interval and shm_unlink() is called right after
> shm_open(). After that, only fd is shared with external processes
> (e.g., vhost-user) as if it were associated with anonymous memory.
>
> In the future we may also allow the user to specify the name to be
> passed to shm_open(), but for now we keep the backend simple, mimicking
> anonymous memory such as memfd.
>
> Acked-by: David Hildenbrand <david@redhat.com>
> Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
> ---
> v4
> - fail if we find "share=off" in shm_backend_memory_alloc() [David]
> v3
> - enriched commit message and documentation to highlight that we
>   want to mimic memfd (David)
> ---
>  docs/system/devices/vhost-user.rst |   5 +-
>  qapi/qom.json                      |  17 ++++
>  backends/hostmem-shm.c             | 123 +++++++++++++++++++++++++++++
>  backends/meson.build               |   1 +
>  qemu-options.hx                    |  13 +++
>  5 files changed, 157 insertions(+), 2 deletions(-)
>  create mode 100644 backends/hostmem-shm.c
>
> diff --git a/docs/system/devices/vhost-user.rst b/docs/system/devices/vhost-user.rst
> index 9b2da106ce..35259d8ec7 100644
> --- a/docs/system/devices/vhost-user.rst
> +++ b/docs/system/devices/vhost-user.rst
> @@ -98,8 +98,9 @@ Shared memory object
>  
>  In order for the daemon to access the VirtIO queues to process the
>  requests it needs access to the guest's address space. This is
> -achieved via the ``memory-backend-file`` or ``memory-backend-memfd``
> -objects. A reference to a file-descriptor which can access this object
> +achieved via the ``memory-backend-file``, ``memory-backend-memfd``, or
> +``memory-backend-shm`` objects.
> +A reference to a file-descriptor which can access this object
>  will be passed via the socket as part of the protocol negotiation.
>  
>  Currently the shared memory object needs to match the size of the main
> diff --git a/qapi/qom.json b/qapi/qom.json
> index 38dde6d785..52df052df8 100644
> --- a/qapi/qom.json
> +++ b/qapi/qom.json
> @@ -721,6 +721,19 @@
>              '*hugetlbsize': 'size',
>              '*seal': 'bool' } }
>  
> +##
> +# @MemoryBackendShmProperties:
> +#
> +# Properties for memory-backend-shm objects.
> +#
> +# The @share boolean option is true by default with shm.

This contradicts the doc comment for @share:

   # @share: if false, the memory is private to QEMU; if true, it is
   #     shared (default: false)

Your intention is to override that text.  But that's less than clear.
Moreover, the documentation of @share is pretty far from this override.
John Snow is working on patches that'll pull it closer.

Hmm, MemoryBackendMemfdProperties has the same override.

I think we should change the doc comment for @share to something like

   # @share: if false, the memory is private to QEMU; if true, it is
   #     shared (default depends on the backend type)

and then document the actual default with each backend type.

> +#
> +# Since: 9.1
> +##
> +{ 'struct': 'MemoryBackendShmProperties',
> +  'base': 'MemoryBackendProperties',
> +  'data': { } }

Let's add 'if': 'CONFIG_POSIX' here.

> +
>  ##
>  # @MemoryBackendEpcProperties:
>  #
> @@ -985,6 +998,8 @@
>      { 'name': 'memory-backend-memfd',
>        'if': 'CONFIG_LINUX' },
>      'memory-backend-ram',
> +    { 'name': 'memory-backend-shm',
> +      'if': 'CONFIG_POSIX' },
>      'pef-guest',
>      { 'name': 'pr-manager-helper',
>        'if': 'CONFIG_LINUX' },
> @@ -1056,6 +1071,8 @@
>        'memory-backend-memfd':       { 'type': 'MemoryBackendMemfdProperties',
>                                        'if': 'CONFIG_LINUX' },
>        'memory-backend-ram':         'MemoryBackendProperties',
> +      'memory-backend-shm':         { 'type': 'MemoryBackendShmProperties',
> +                                      'if': 'CONFIG_POSIX' },
>        'pr-manager-helper':          { 'type': 'PrManagerHelperProperties',
>                                        'if': 'CONFIG_LINUX' },
>        'qtest':                      'QtestProperties',

[...]

> diff --git a/qemu-options.hx b/qemu-options.hx
> index cf61f6b863..2226172fb0 100644
> --- a/qemu-options.hx
> +++ b/qemu-options.hx
> @@ -5227,6 +5227,19 @@ SRST
>  
>          The ``share`` boolean option is on by default with memfd.
>  
> +    ``-object memory-backend-shm,id=id,merge=on|off,dump=on|off,share=on|off,prealloc=on|off,size=size,host-nodes=host-nodes,policy=default|preferred|bind|interleave``
> +        Creates a POSIX shared memory backend object, which allows
> +        QEMU to share the memory with an external process (e.g. when
> +        using vhost-user). This backend mimics memfd, allocating memory that is
> +        practically anonymous. This is useful when memfd is not available.

This actually explains the purpose, unlike the doc comment in qom.json.
Same for the existing memory backends; can't fault you for doing your
new one the same way.  We ought to fix them all.  I'm not demanding you
do it.

The text could perhaps a bit clearer.  What does "practically anonymous"
mean?  As far as I understand, memory-backend-shm is a more portable and
less featureful version of memory-backend-memfd.  Say that?

> +
> +        Please refer to ``memory-backend-file`` for a description of the
> +        options.
> +
> +        The ``share`` boolean option is on by default with shm. Setting it to
> +        off will cause a failure during allocation because it is not supported
> +        by this backend.
> +

The "will case a failure" part is documented only here, and not in
qom.json.

Not this patch's fault: documentation for -object memory-backend-epc is
missing.

>      ``-object iommufd,id=id[,fd=fd]``
>          Creates an iommufd backend which allows control of DMA mapping
>          through the ``/dev/iommu`` device.