Re: [PATCH 9/9] hostmem-file: support POSIX shm_open()

[Stefano Garzarella]

On Wed, Feb 28, 2024 at 01:01:55PM +0100, David Hildenbrand wrote:
> On 28.02.24 12:47, Stefano Garzarella wrote:
> > Add a new `shm` bool option for `-object memory-backend-file`.
> >
> > When this option is set to true, the POSIX shm_open(3) is used instead
> > of open(2).
> >
> > So a file will not be created in the filesystem, but a "POSIX shared
> > memory object" will be instantiated. In Linux this turns into a file
> > in /dev/shm, but in other OSes this may not happen (for example in
> > macOS or FreeBSD nothing is shown in any filesystem).
> >
> > This new feature is useful when we need to share guest memory with
> > another process (e.g. vhost-user backend), but we don't have
> > memfd_create() or any special filesystems (e.g. /dev/shm) available
> > as in macOS.
> >
> > Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
> > ---
> > I am not sure this is the best way to support shm_open() in QEMU.
> >
> > Other solutions I had in mind were:
> >
> > - create a new memory-backend-shm
> >
> > - extend memory-backend-memfd to use shm_open() on systems where memfd is
> > not available (problem: shm_open wants a name to assign to the object, but
> > we can do a workaround by using a random name and do the unlink right away)
> >
> > Any preference/suggestion?
>
> Both sound like reasonable options, and IMHO better than hostmem-file 
> with things that are not necessarily "real" files.

Yeah, I see.

> Regarding memory-backend-memfd, we similarly have to pass a name to 
> memfd_create(), although for different purpose: "The  name  supplied 
> in name is used as a filename and will be displayed as the target of 
> the corresponding symbolic link in the directory /proc/self/fd/".
>
> So we simply pass TYPE_MEMORY_BACKEND_MEMFD.

Okay, so I guess it must be unique only in the process, while for 
shm_open() it is global.

> Likely, memory-backend-shm that directly maps to shm_open() and only 
> provides properties reasonable for shm_open() is the cleanest 
> approach. So that would currently be my preference :)

Thank you for your thoughts, I think I will go toward this direction 
(memory-backend-shm).

It was also my first choice, but in order to have a working RFC right 
away, I modified memory-backend-file.

Stefano