Date: Thu, 12 Oct 2023 09:58:26 +1000
From: "Nicholas Piggin" <npiggin@...il.com>
To: "Aaron Conole" <aconole@...hat.com>
Cc: "Eelco Chaudron" <echaudro@...hat.com>, <netdev@...r.kernel.org>,
 <dev@...nvswitch.org>, "Ilya Maximets" <imaximet@...hat.com>, "Flavio
 Leitner" <fbl@...hat.com>
Subject: Re: [ovs-dev] [RFC PATCH 4/7] net: openvswitch: ovs_vport_receive
 reduce stack usage

On Wed Oct 11, 2023 at 11:34 PM AEST, Aaron Conole wrote:
> "Nicholas Piggin" <npiggin@...il.com> writes:
>
> > On Fri Sep 29, 2023 at 6:38 PM AEST, Eelco Chaudron wrote:
> >>
> >>
> >> On 29 Sep 2023, at 9:00, Nicholas Piggin wrote:
> >>
> >> > On Fri Sep 29, 2023 at 1:26 AM AEST, Aaron Conole wrote:
> >> >> Nicholas Piggin <npiggin@...il.com> writes:
> >> >>
> >> >>> Dynamically allocating the sw_flow_key reduces stack usage of
> >> >>> ovs_vport_receive from 544 bytes to 64 bytes at the cost of
> >> >>> another GFP_ATOMIC allocation in the receive path.
> >> >>>
> >> >>> XXX: is this a problem with memory reserves if ovs is in a
> >> >>> memory reclaim path, or since we have a skb allocated, is it
> >> >>> okay to use some GFP_ATOMIC reserves?
> >> >>>
> >> >>> Signed-off-by: Nicholas Piggin <npiggin@...il.com>
> >> >>> ---
> >> >>
> >> >> This represents a fairly large performance hit.  Just my own quick
> >> >> testing on a system using two netns, iperf3, and simple forwarding rules
> >> >> shows between 2.5% and 4% performance reduction on x86-64.  Note that it
> >> >> is a simple case, and doesn't involve a more involved scenario like
> >> >> multiple bridges, tunnels, and internal ports.  I suspect such cases
> >> >> will see even bigger hit.
> >> >>
> >> >> I don't know the impact of the other changes, but just an FYI that the
> >> >> performance impact of this change is extremely noticeable on x86
> >> >> platform.
> >> >
> >> > Thanks for the numbers. This patch is probably the biggest perf cost,
> >> > but unfortunately it's also about the biggest saving. I might have an
> >> > idea to improve it.
> >>
> >> Also, were you able to figure out why we do not see this problem on
> >> x86 and arm64? Is the stack usage so much larger, or is there some
> >> other root cause? Is there a simple replicator, as this might help
> >> you profile the differences between the architectures?
> >
> > I found some snippets of equivalent call chain (this is for 4.18 RHEL8
> > kernels, but it's just to give a general idea of stack overhead
> > differences in C code). Frame size annotated on the right hand side:
> >
> > [c0000007ffdba980] do_execute_actions     496
> > [c0000007ffdbab70] ovs_execute_actions    128
> > [c0000007ffdbabf0] ovs_dp_process_packet  208
> > [c0000007ffdbacc0] clone_execute          176
> > [c0000007ffdbad70] do_execute_actions     496
> > [c0000007ffdbaf60] ovs_execute_actions    128
> > [c0000007ffdbafe0] ovs_dp_process_packet  208
> > [c0000007ffdbb0b0] ovs_vport_receive      528
> > [c0000007ffdbb2c0] internal_dev_xmit
> >                                  total = 2368
> > [ff49b6d4065a3628] do_execute_actions     416
> > [ff49b6d4065a37c8] ovs_execute_actions     48
> > [ff49b6d4065a37f8] ovs_dp_process_packet  112
> > [ff49b6d4065a3868] clone_execute           64
> > [ff49b6d4065a38a8] do_execute_actions     416
> > [ff49b6d4065a3a48] ovs_execute_actions     48
> > [ff49b6d4065a3a78] ovs_dp_process_packet  112
> > [ff49b6d4065a3ae8] ovs_vport_receive      496
> > [ff49b6d4065a3cd8] netdev_frame_hook
> >                                  total = 1712
> >
> > That's more significant than I thought, nearly 40% more stack usage for
> > ppc even with 3 frames having large local variables that can't be
> > avoided for either arch.
> >
> > So, x86_64 could be quite safe with its 16kB stack for the same
> > workload, explaining why same overflow has not been seen there.
>
> This is interesting - is it possible that we could resolve this without
> needing to change the kernel - or at least without changing how OVS
> works?

Not really.

To be clear I don't say ovs is the one and only problem, so it could be
resolved if stack was larger or if other things did not use so much,
etc.

Maybe other things could be changed too, but ovs uses several K of stack
that it doesn't need to, and since it is also causing recursion it needs
to be as tight as possible with its stack use.

> Why are these so different?  Maybe there's some bloat in some of
> the ppc data structures that can be addressed?  For example,
> ovs_execute_actions shouldn't really be that different, but I wonder if
> the way the per-cpu infra works, or the deferred action processing gets
> inlined would be causing stack bloat?

Most other stack usage is not due to Linux powerpc arch defining certain
types and structures to be larger (most are the same size as other
64-bit archs). Rather due to C and GCC. I have asked powerpc GCC people
about stack size and no easy option to reduce it, if it were possible to
improve in new version of GCC then we still need to deal with old.

Powerpc has a larger minimum stack frame size (32 bytes) and larger
alignment (32 bytes vs 16 IIRC). It also has more non-volatile registers
and probably uses them more which requires saving to stack. So some of
it is fundamental.

In some cases I can't really see why GCC on ppc uses so much. AFAIKS
ovs_execute_actions could be using 96 bytes, but it's possible I miss
an alignment requirement.

Thanks,
Nick

Powered by blists - more mailing lists