On Wed, Nov 04, 2020 at 09:50:30AM +0100, Eric Dumazet wrote: > On 11/4/20 2:16 AM, Rama Nichanamatlu wrote: > >> Thanks for providing the numbers. Do you think that dropping (up to) > >> 7 packets is acceptable? > > > > net.ipv4.tcp_syn_retries = 6 > > > > tcp clients wouldn't even get that far leading to connect establish issues. > > This does not really matter. If host was under memory pressure, > dropping a few packets is really not an issue. > > Please do not add expensive checks in fast path, just to "not drop a packet" > even if the world is collapsing. Right, that was my first patch -- to only recheck if we're about to reuse the page. Do you think that's acceptable, or is that still too close to the fast path? > Also consider that NIC typically have thousands of pre-allocated page/frags > for their RX ring buffers, they might all have pfmemalloc set, so we are speaking > of thousands of packet drops before the RX-ring can be refilled with normal (non pfmemalloc) page/frags. > > If we want to solve this issue more generically, we would have to try > to copy data into a non pfmemalloc frag instead of dropping skb that > had frags allocated minutes ago under memory pressure. I don't think we need to copy anything. We need to figure out if the system is still under memory pressure, and if not, we can clear the pfmemalloc bit on the frag, as in my second patch. The 'least change' way of doing that is to try to allocate a page, but the VM could export a symbol that says "we're not under memory pressure any more". Did you want to move checking that into the networking layer, or do you want to keep it in the pagefrag allocator?
