netdev - Re: Question about the sndbuf of the tap interface with vhost-net

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: <CAHJXk3bVBKyP2UnKs4rTypfHLRs2d3=DuDSvM8e_2WMWtgYa8w@mail.gmail.com>
Date:   Thu, 24 Feb 2022 12:36:41 +0800
From:   Harold Huang <baymaxhuang@...il.com>
To:     Jason Wang <jasowang@...hat.com>
Cc:     users@...k.org, Maxime Coquelin <maxime.coquelin@...hat.com>,
        Chenbo Xia <chenbo.xia@...el.com>,
        netdev <netdev@...r.kernel.org>
Subject: Re: Question about the sndbuf of the tap interface with vhost-net

Harold Huang <baymaxhuang@...il.com> 于2022年2月24日周四 12:19写道：
>
> Thanks for Jason's comments.
>
> Jason Wang <jasowang@...hat.com> 于2022年2月24日周四 11:23写道：
> >
> > Adding netdev.
> >
> > On Wed, Feb 23, 2022 at 9:46 PM Harold Huang <baymaxhuang@...il.com> wrote:
> > >
> > >  Sorry. The performance tested by iperf is degraded from 4.5 Gbps to
> > > 750Mbps per flow.
> > >
> > > Harold Huang <baymaxhuang@...il.com> 于2022年2月23日周三 21:13写道：
> > > >
> > > > I see in dpdk virtio-user driver, the TUNSETSNDBUF is initialized with
> > > > INT_MAX, see: https://github.com/DPDK/dpdk/blob/main/drivers/net/virtio/virtio_user/vhost_kernel_tap.c#L169
> >
> > Note that Linux use INT_MAX as default sndbuf for tuntap.
> >
> > > > It is ok because tap driver uses it to support tx baching, see this
> > > > patch: https://github.com/torvalds/linux/commit/0a0be13b8fe2cac11da2063fb03f0f39359b3069
> > > >
> > > > But in tun_xdp_one, napi is not supported and I want to user napi in
> > > > tun_get_user to enable gro.
> >
> > NAPI is not enabled in this path, want to send a patch to do that?
>
> Yes, I have a patch in this path to enable NAPI and it greatly
> improves TCP stream performance, from 4.5Gbsp to 9.2 Gbps per flow. I
> will send it later for comments.
>
> >
> > Btw, NAPI mode is used for kernel networking stack hardening at start,
> > but it would be interesting to see if it helps for the performance.
> >
> > > > As I result, I change the sndbuf to a
> > > > value such as 212992 in /proc/sys/net/core/wmem_default.
> >
> > Can you describe your setup in detail? Where did you run the iperf
> > server and client and where did you change the wmem_default?
>
> I use dpdk-testpmd to test the vhost-net performance, such as:
> dpdk-testpmd -l 0-9  -n 4
> --vdev=virtio_user0,path=/dev/vhost-net,queue_size=1024,mac=00:00:0a:00:00:02
> -a 0000:06:00.1 -- -i  --txd=1024 --rxd=1024
>
> And I have changed the sndbuf in
> https://github.com/DPDK/dpdk/blob/main/drivers/net/virtio/virtio_user/vhost_kernel_tap.c#L169
> to 212992, which is not INT_MAX anymore. I also enable NAPI in the tun
> module.  The iperf server ran in the tap interface on the kernel side,
> which would receive TCP stream from dpdk-testpmd. But the performance
> is greatly degraded,  from 4.5 Gbps to 750Mbps. I am confused about
> the perf result of the cpu core where iperf server ran, which has a
> serious bottleneck: 59.86% cpu on the report_bug and  20.66% on the
> module_find_bug. I use centos 8.2 with a native 4.18.0-193.el8.x86_64
> kernel to test.

BTW, if I change sock_can_batch = false in
https://github.com/torvalds/linux/blob/master/drivers/vhost/net.c#L782
directly and use the default sk.sk_sndbuf size, ie. INT_MAX, the test
result seems ok.

>
> >
> > > > But the
> > > > performance tested by iperf is greatly degraded, from 4.5 Gbps to
> > > > 750Gbps per flow. I see the the iperf server consume 100% cpu core,
> > > > which should be the bottleneck of the this test. The perf top result
> > > > of iperf server cpu core is as follows:
> > > >
> > > > '''
> > > > Samples: 72  of event 'cycles', 4000 Hz, Event count (approx.):
> > > > 22685278 lost: 0/0 drop: 0/0
> > > > Overhead  Shared O  Symbol
> > > >   59.86%  [kernel]  [k] report_bug
> > > >   20.66%  [kernel]  [k] module_find_bug
> > > >    6.51%  [kernel]  [k] common_interrupt
> > > >    2.82%  [kernel]  [k] __slab_free
> > > >    1.48%  [kernel]  [k] copy_user_enhanced_fast_string
> > > >    1.44%  [kernel]  [k] __skb_datagram_iter
> > > >    1.42%  [kernel]  [k] notifier_call_chain
> > > >    1.41%  [kernel]  [k] irq_work_run_list
> > > >    1.41%  [kernel]  [k] update_irq_load_avg
> > > >    1.41%  [kernel]  [k] task_tick_fair
> > > >    1.41%  [kernel]  [k] cmp_ex_search
> > > >    0.16%  [kernel]  [k] __ghes_peek_estatus.isra.12
> > > >    0.02%  [kernel]  [k] acpi_os_read_memory
> > > >    0.00%  [kernel]  [k] native_apic_mem_write
> > > > '''
> > > > I am not clear about the test result. Can we change the sndbuf size in
> > > > dpdk? Is any way to enable vhost_net to use napi without changing the
> > > > tun kernel driver?
> >
> > You can do this by not using INT_MAX as sndbuf.
>
> Just mentioned above, I change the sndbuf value and I met a serious
> performance degradation.
>
> >
> > Thanks
> >
> > >
> >