Re: [PATCH v2 net] net: enetc: survive memory pressure without crashing

[Vladimir Oltean]

On Thu, Oct 27, 2022 at 07:03:31AM +0000, Claudiu Manoil wrote:
> How do you trigger this "extreme case of memory pressure" where no enetc buffer
> can be allocated? Do you simulate it?

As far as I understand, making dev_alloc_page() predictably fail in some
particular spot is hard and probabilistic (but possible given enough tries).

The reason I stubled upon this particularly bad handling of low memory
in the enetc driver is because with AF_XDP zero-copy sockets, memory
for RX buffers comes directly from user space, which may simply opt to
not put any buffers in the fill queue (see "xdpsock --txonly" for
example).

The fix for the case where the buffers come from the kernel's page
allocator is simply analogous to that. It would be shady to add this
exact same patch only as part of the XSK work, when it's clear that
existing code suffers from this problem too, even if it's not easy to
trigger it.