Re: [PATCH V1 17/26] machine: memfd-alloc option

[Peter Xu <peterx@redhat.com>]

On Wed, May 29, 2024 at 01:31:38PM -0400, Steven Sistare wrote:
> > > diff --git a/system/memory.c b/system/memory.c
> > > index 49f1cb2..ca04a0e 100644
> > > --- a/system/memory.c
> > > +++ b/system/memory.c
> > > @@ -1552,8 +1552,9 @@ bool memory_region_init_ram_nomigrate(MemoryRegion *mr,
> > >                                         uint64_t size,
> > >                                         Error **errp)
> > >   {
> > > +    uint32_t flags = current_machine->memfd_alloc ? RAM_SHARED : 0;
> > 
> > If there's a machine option to "use memfd for allocations", then it's
> > shared mem... Hmm..
> > 
> > It is a bit confusing to me in quite a few levels:
> > 
> >    - Why memory allocation method will be defined by a machine property,
> >      even if we have memory-backend-* which should cover everything?
> 
> Some memory regions are implicitly created, and have no explicit representation
> on the qemu command line.  memfd-alloc affects those.
> 
> More generally, memfd-alloc affects all ramblock allocations that are
> not explicitly represented by memory-backend object.  Thus the simple
> command line "qemu -m 1G" does not explicitly describe an object, so it
> goes through the anonymous allocation path, and is affected by memfd-alloc.

Can we simply now allow "qemu -m 1G" to work for cpr-exec?  AFAIU that's
what we do with cpr-reboot: we ask the user to specify the right things to
make other thing work.  Otherwise it won't.

> 
> Internally, create_default_memdev does create a memory-backend object.
> That is what my doc comment above refers to:
>   Any associated memory-backend objects are created with share=on
> 
> An explicit "qemu -object memory-backend-*" is not affected by memfd-alloc.
> 
> The qapi comments in patch "migration: cpr-exec mode" attempt to say all that:
> 
> +#     Memory backend objects must have the share=on attribute, and
> +#     must be mmap'able in the new QEMU process.  For example,
> +#     memory-backend-file is acceptable, but memory-backend-ram is
> +#     not.
> +#
> +#     The VM must be started with the '-machine memfd-alloc=on'
> +#     option.  This causes implicit ram blocks -- those not explicitly
> +#     described by a memory-backend object -- to be allocated by
> +#     mmap'ing a memfd.  Examples include VGA, ROM, and even guest
> +#     RAM when it is specified without a memory-backend object.

VGA is IIRC 16MB chunk, ROM is even smaller.  If the user specifies -object
memory-backend-file,share=on propertly, these should be the only outliers?

Are these important enough for the downtime?  Can we put them into the
migrated image alongside with the rest device states?

> 
> >    - Even if we have such a machine property, why setting "memfd" will
> >      always imply shared?  why not private?  After all it's not called
> >      "memfd-shared-alloc", and we can create private mappings using
> >      e.g. memory-backend-memfd,share=off.
> 
> There is no use case for memfd-alloc with share=off, so no point IMO in
> making the option more verbose.

Unfortunately this fact doesn't make the property easier to understand. :-(

> For cpr, the mapping with all its modifications must be visible to new
> qemu when qemu mmaps it.

So this might be the important part - do you mean migrating
VGA/ROM/... small ramblocks won't work (besides any performance concerns)?
Could you elaborate?

Cpr-reboot already introduced lots of tricky knobs to QEMU.  We may need to
restrict that specialty to minimal, making the interfacing as clear as
possible, or (at least migration) maintainers will start to be soon scared
and running away, if such proposal was not shot down.

In short, I hope when we introduce new knobs for cpr, we shouldn't always
keep cpr-* modes in mind, but consider whenever the user can use it without
cpr-*.  I'm not sure whether it'll be always possible, but we should try.

Thanks,

-- 
Peter Xu