Re: [PATCH net-next v2 02/12] selftests: ncdevmem: Separate out dmabuf provider

[Mina Almasry <almasrymina@google.com>]

On Thu, Oct 3, 2024 at 3:08 PM Stanislav Fomichev <stfomichev@gmail.com> wrote:
>
> > > @@ -464,14 +515,11 @@ int do_server(void)
> > >  void run_devmem_tests(void)
> > >  {
> > >         struct netdev_queue_id *queues;
> > > -       int devfd, memfd, buf;
> > > +       struct memory_buffer *mem;
> > >         struct ynl_sock *ys;
> > > -       size_t dmabuf_size;
> > >         size_t i = 0;
> > >
> > > -       dmabuf_size = getpagesize() * NUM_PAGES;
> > > -
> > > -       create_udmabuf(&devfd, &memfd, &buf, dmabuf_size);
> > > +       mem = provider->alloc(getpagesize() * NUM_PAGES);
> > >
> >
> > There is some weirdness here where run_devmem_tests() allocates its
> > own provider, but do_server() uses a provider allocated in main(). Any
> > reason these are not symmetric? I would marginally prefer do_server()
> > to allocate its own provider just like run_devmem_tests(), or at least
> > make them both match, if possible.
>
> I wanted to keep them separate in case we end up adding more to
> the selftest part. For example, not sure what would happen now if we pass
> a udmabuf with just one page? Do we need some test for the drivers
> to make sure they handle this case?

The size of the udmabuf (or more generically, any dmabuf) is the
address space for the page pool; if the dmabuf is too small, like 1
page, the rest of the pp allocations will return -ENOMEM.

-- 
Thanks,
Mina