Re: [PATCH v1 net 3/3] selftest: net: Check wraparounds for sk->sk_rmem_alloc.

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Kuniyuki Iwashima wrote:
> From: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
> Date: Mon, 24 Mar 2025 11:58:06 -0400
> > > +TEST_F(so_rcvbuf, rmem_max)
> > > +{
> > > +	char buf[16] = {};
> > > +	int ret, i;
> > > +
> > > +	create_socketpair(_metadata, self, variant);
> > > +
> > > +	ret = setsockopt(self->server, SOL_SOCKET, SO_RCVBUFFORCE,
> > > +			 &(int){INT_MAX}, sizeof(int));
> > > +	ASSERT_EQ(ret, 0);
> > > +
> > > +	ASSERT_EQ(get_prot_pages(_metadata, variant), 0);
> > > +
> > > +	for (i = 1; ; i++) {
> > > +		ret = send(self->client, buf, sizeof(buf), 0);
> > > +		ASSERT_EQ(ret, sizeof(buf));
> > > +
> > > +		if (i % 10000 == 0) {
> > > +			int pages = get_prot_pages(_metadata, variant);
> > > +
> > > +			/* sk_rmem_alloc wrapped around too much ? */
> > > +			ASSERT_LE(pages, *variant->max_pages);
> > > +
> > > +			if (pages == *variant->max_pages)
> > > +				break;
> > 
> > Does correctness depend here on max_pages being a multiple of 10K?
> 
> 10K may be too conservative, but at least we need to ensure
> that the size of accumulated skbs exceeds 1 PAGE_SIZE to
> fail on the ASSERT_LE(), otherwise we can't detect the multiple
> wraparounds even without patch 1.

Thanks. It took me some time to understand. Without overflow,
the pages counter will saturate at max_pages as the queue fills up.
 
> The later sleep for call_rcu() was dominant than this loop on
> my machine.

Ack.