lkml.org 
[lkml]   [2020]   [Jun]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: Packet gets stuck in NOLOCK pfifo_fast qdisc
Date
> From: Jonas Bonn <jonas.bonn@netrounds.com>
> To: Paolo Abeni <pabeni@redhat.com>,
> "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
> LKML <linux-kernel@vger.kernel.org>,
> "David S . Miller" <davem@davemloft.net>,
> John Fastabend <john.fastabend@gmail.com>
> Subject: Re: Packet gets stuck in NOLOCK pfifo_fast qdisc
> Date: Fri, 11 Oct 2019 02:39:48 +0200
> Message-ID: <465a540e-5296-32e7-f6a6-79942dfe2618@netrounds.com> (raw)
> In-Reply-To: <95c5a697932e19ebd6577b5dac4d7052fe8c4255.camel@redhat.com>
>
> Hi Paolo,
>
> On 09/10/2019 21:14, Paolo Abeni wrote:
> > Something alike the following code - completely untested - can possibly
> > address the issue, but it's a bit rough and I would prefer not adding
> > additonal complexity to the lockless qdiscs, can you please have a spin
> > a it?
>
> We've tested a couple of variants of this patch today, but unfortunately
> it doesn't fix the problem of packets getting stuck in the queue.
>
> A couple of comments:
>
> i) On 5.4, there is the BYPASS path that also needs the same treatment
> as it's essentially replicating the behavour of qdisc_run, just without
> the queue/dequeue steps
>
> ii) We are working a lot with the 4.19 kernel so I backported to the
> patch to this version and tested there. Here the solution would seem to
> be more robust as the BYPASS path does not exist.
>
> Unfortunately, in both cases we continue to see the issue of the "last
> packet" getting stuck in the queue.
>
> /Jonas

Hello Jonas, Paolo,

We have observed the same problem with pfifo_fast qdisc when sending periodic small
packets on a TCP flow with multiple simultaneous connections on a 4.19.75
kernel. We've been able to catch it in action using perf probes (see trace
below). For qdisc = 0xffff900d7c247c00, skb = 0xffff900b72c334f0,
it takes 200270us to traverse the networking stack on a system that's not otherwise busy.
qdisc only resumes processing when another enqueued packet comes in,
so the packet could have been stuck indefinitely.

proc-19902 19902 [032] 580644.045480: probe:pfifo_fast_dequeue_end: (ffffffff9b69d99d) qdisc=0xffff900d7c247c00 skb=0xffff900bfc294af0 band=2 atomic_qlen=0
proc-19902 19902 [032] 580644.045480: probe:pfifo_fast_dequeue: (ffffffff9b69d8c0) qdisc=0xffff900d7c247c00 skb=0xffffffff9b69d8c0 band=2
proc-19927 19927 [014] 580644.045480: probe:tcp_transmit_skb2: (ffffffff9b6dc4e5) skb=0xffff900b72c334f0 sk=0xffff900d62958040 source=0x4b4e dest=0x9abe
proc-19902 19902 [032] 580644.045480: probe:pfifo_fast_dequeue_end: (ffffffff9b69d99d) qdisc=0xffff900d7c247c00 skb=0x0 band=3 atomic_qlen=0
proc-19927 19927 [014] 580644.045481: probe:ip_finish_output2: (ffffffff9b6bc650) net=0xffffffff9c107c80 sk=0xffff900d62958040 skb=0xffff900b72c334f0 __func__=0x0
proc-19902 19902 [032] 580644.045481: probe:sch_direct_xmit: (ffffffff9b69e570) skb=0xffff900bfc294af0 q=0xffff900d7c247c00 dev=0xffff900d6a140000 txq=0xffff900d6a181180 root_lock=0x0 validate=1 ret=-1 again=155
proc-19927 19927 [014] 580644.045481: net:net_dev_queue: dev=eth0 skbaddr=0xffff900b72c334f0 len=115
proc-19902 19902 [032] 580644.045482: probe:pfifo_fast_dequeue: (ffffffff9b69d8c0) qdisc=0xffff900d7c247c00 skb=0xffffffff9b69d8c0 band=1
proc-19927 19927 [014] 580644.045483: probe:pfifo_fast_enqueue: (ffffffff9b69d9f0) skb=0xffff900b72c334f0 qdisc=0xffff900d7c247c00 to_free=18446622925407304000
proc-19902 19902 [032] 580644.045483: probe:pfifo_fast_dequeue_end: (ffffffff9b69d99d) qdisc=0xffff900d7c247c00 skb=0x0 band=3 atomic_qlen=0
proc-19927 19927 [014] 580644.045483: probe:pfifo_fast_enqueue_end: (ffffffff9b69da9f) skb=0xffff900b72c334f0 qdisc=0xffff900d7c247c00 to_free=0xffff91d0f67ab940 atomic_qlen=1
proc-19902 19902 [032] 580644.045484: probe:__qdisc_run_2: (ffffffff9b69ea5a) q=0xffff900d7c247c00 packets=1
proc-19927 19927 [014] 580644.245745: probe:pfifo_fast_enqueue: (ffffffff9b69d9f0) skb=0xffff900d98fdf6f0 qdisc=0xffff900d7c247c00 to_free=18446622925407304000
proc-19927 19927 [014] 580644.245745: probe:pfifo_fast_enqueue_end: (ffffffff9b69da9f) skb=0xffff900d98fdf6f0 qdisc=0xffff900d7c247c00 to_free=0xffff91d0f67ab940 atomic_qlen=2
proc-19927 19927 [014] 580644.245746: probe:pfifo_fast_dequeue: (ffffffff9b69d8c0) qdisc=0xffff900d7c247c00 skb=0xffffffff9b69d8c0 band=0
proc-19927 19927 [014] 580644.245746: probe:pfifo_fast_dequeue_end: (ffffffff9b69d99d) qdisc=0xffff900d7c247c00 skb=0xffff900b72c334f0 band=2 atomic_qlen=1
proc-19927 19927 [014] 580644.245747: probe:pfifo_fast_dequeue: (ffffffff9b69d8c0) qdisc=0xffff900d7c247c00 skb=0xffffffff9b69d8c0 band=2
proc-19927 19927 [014] 580644.245747: probe:pfifo_fast_dequeue_end: (ffffffff9b69d99d) qdisc=0xffff900d7c247c00 skb=0xffff900d98fdf6f0 band=2 atomic_qlen=0
proc-19927 19927 [014] 580644.245748: probe:pfifo_fast_dequeue: (ffffffff9b69d8c0) qdisc=0xffff900d7c247c00 skb=0xffffffff9b69d8c0 band=2
proc-19927 19927 [014] 580644.245748: probe:pfifo_fast_dequeue_end: (ffffffff9b69d99d) qdisc=0xffff900d7c247c00 skb=0x0 band=3 atomic_qlen=0
proc-19927 19927 [014] 580644.245749: qdisc:qdisc_dequeue: dequeue ifindex=5 qdisc handle=0x0 parent=0xF txq_state=0x0 packets=2 skbaddr=0xffff900b72c334f0
proc-19927 19927 [014] 580644.245749: probe:sch_direct_xmit: (ffffffff9b69e570) skb=0xffff900b72c334f0 q=0xffff900d7c247c00 dev=0xffff900d6a140000 txq=0xffff900d6a181180 root_lock=0x0 validate=1 ret=-1 again=155
proc-19927 19927 [014] 580644.245750: net:net_dev_start_xmit: dev=eth0 queue_mapping=14 skbaddr=0xffff900b72c334f0 vlan_tagged=0 vlan_proto=0x0000 vlan_tci=0x0000 protocol=0x0800 ip_summed=3 len=115 data_len=0 network_offset=14 transport_offset_valid=1 transport_offset=34 tx_flags=0 gso_size=0 gso_segs=1 gso_type=0x1

I was wondering if you had any more luck in finding a solution or workaround for this problem
(that is, aside from switching to a different qdisc)?

Thanks,
~ Michael

\
 
 \ /
  Last update: 2020-06-23 16:08    [W:0.139 / U:1.196 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site