netdev - Re: [PATCH] net: skip RPS if packet is already on target CPU

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CADUfDZrSUNu7nym9dC1_yFUqhC8tUPYjv-ZKHofU9Q8Uv4Jvhw@mail.gmail.com>
Date: Tue, 29 Oct 2024 13:38:37 -0700
From: Caleb Sander <csander@...estorage.com>
To: Eric Dumazet <edumazet@...gle.com>
Cc: "David S. Miller" <davem@...emloft.net>, Jakub Kicinski <kuba@...nel.org>, 
	Paolo Abeni <pabeni@...hat.com>, Simon Horman <horms@...nel.org>, netdev@...r.kernel.org, 
	linux-kernel@...r.kernel.org
Subject: Re: [PATCH] net: skip RPS if packet is already on target CPU

On Tue, Oct 29, 2024 at 12:02 PM Eric Dumazet <edumazet@...gle.com> wrote:
>
> On Tue, Oct 29, 2024 at 7:27 PM Caleb Sander Mateos
> <csander@...estorage.com> wrote:
> >
> > If RPS is enabled, all packets with a CPU flow hint are enqueued to the
> > target CPU's input_pkt_queue and process_backlog() is scheduled on that
> > CPU to dequeue and process the packets. If ARFS has already steered the
> > packets to the correct CPU, this additional queuing is unnecessary and
> > the spinlocks involved incur significant CPU overhead.
> >
> > In netif_receive_skb_internal() and netif_receive_skb_list_internal(),
> > check if the CPU flow hint get_rps_cpu() returns is the current CPU. If
> > so, bypass input_pkt_queue and immediately process the packet(s) on the
> > current CPU.
> >
> > Signed-off-by: Caleb Sander Mateos <csander@...estorage.com>
>
> Current implementation was a conscious choice. This has been discussed
> several times.
>
> By processing packets inline, you are actually increasing latencies of
> packets queued to other cpus.

Sorry, I wasn't aware of these prior discussions. I take it you are
referring to threads like
https://lore.kernel.org/netdev/20230322072142.32751-1-xu.xin16@zte.com.cn/T/
? I see what you mean about the latency penalty for packets that do
require cross-CPU steering.

Do you have an alternate suggestion for how to avoid the overhead of
acquiring a spinlock for every packet? The atomic instruction in
rps_lock_irq_disable() called from process_backlog() is consuming 5%
of our CPU time. For our use case, we don't really want software RPS;
we are expecting ARFS to steer all high-bandwidth traffic to the
desired CPUs. We would happily turn off software RPS entirely if we
could, which seems like it would avoid the concerns about higher
latency for packets that need to be steering to a different CPU. But
my understanding is that using ARFS requires RPS to be enabled
(rps_sock_flow_entries set globally and rps_flow_cnt set on each
queue), which enables these rps_needed static branches. Is that
correct? If so, would you be open to adding a sysctl that disables
software RPS and relies upon ARFS to do the packet steering?

Thanks,
Caleb