Message-ID: <Y9gnqNnkiPEw+Pp8@qwirkle>
Date:   Mon, 30 Jan 2023 20:25:12 +0000
From:   Andrei Gherzan <andrei.gherzan@...onical.com>
To:     Willem de Bruijn <willemb@...gle.com>
Cc:     Paolo Abeni <pabeni@...hat.com>,
        "David S. Miller" <davem@...emloft.net>,
        Eric Dumazet <edumazet@...gle.com>,
        Jakub Kicinski <kuba@...nel.org>,
        Shuah Khan <shuah@...nel.org>, netdev@...r.kernel.org,
        linux-kselftest@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] selftests: net: udpgso_bench_tx: Introduce exponential
 back-off retries

On 23/01/30 02:57PM, Willem de Bruijn wrote:
> On Mon, Jan 30, 2023 at 2:51 PM Andrei Gherzan
> <andrei.gherzan@...onical.com> wrote:
> >
> > On 23/01/30 06:24PM, Andrei Gherzan wrote:
> > > On 23/01/30 12:35PM, Willem de Bruijn wrote:
> > > > On Mon, Jan 30, 2023 at 12:31 PM Andrei Gherzan
> > > > <andrei.gherzan@...onical.com> wrote:
> > > > >
> > > > > On 23/01/30 11:29AM, Willem de Bruijn wrote:
> > > > > > On Mon, Jan 30, 2023 at 11:23 AM Andrei Gherzan
> > > > > > <andrei.gherzan@...onical.com> wrote:
> > > > > > >
> > > > > > > On 23/01/30 04:15PM, Andrei Gherzan wrote:
> > > > > > > > On 23/01/30 11:03AM, Willem de Bruijn wrote:
> > > > > > > > > On Mon, Jan 30, 2023 at 9:28 AM Andrei Gherzan
> > > > > > > > > <andrei.gherzan@...onical.com> wrote:
> > > > > > > > > >
> > > > > > > > > > On 23/01/30 08:35AM, Willem de Bruijn wrote:
> > > > > > > > > > > On Mon, Jan 30, 2023 at 7:51 AM Andrei Gherzan
> > > > > > > > > > > <andrei.gherzan@...onical.com> wrote:
> > > > > > > > > > > >
> > > > > > > > > > > > On 23/01/30 09:26AM, Paolo Abeni wrote:
> > > > > > > > > > > > > On Fri, 2023-01-27 at 17:03 -0500, Willem de Bruijn wrote:
> > > > > > > > > > > > > > On Fri, Jan 27, 2023 at 1:16 PM Andrei Gherzan
> > > > > > > > > > > > > > <andrei.gherzan@...onical.com> wrote:
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > The tx and rx test programs are used in a couple of test scripts including
> > > > > > > > > > > > > > > "udpgro_bench.sh". Taking this as an example, when the rx/tx programs
> > > > > > > > > > > > > > > are invoked subsequently, there is a chance that the rx one is not ready to
> > > > > > > > > > > > > > > accept socket connections. This racing bug could fail the test with at
> > > > > > > > > > > > > > > least one of the following:
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > ./udpgso_bench_tx: connect: Connection refused
> > > > > > > > > > > > > > > ./udpgso_bench_tx: sendmsg: Connection refused
> > > > > > > > > > > > > > > ./udpgso_bench_tx: write: Connection refused
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > This change addresses this by adding routines that retry the socket
> > > > > > > > > > > > > > > operations with an exponential back off algorithm from 100ms to 2s.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > Fixes: 3a687bef148d ("selftests: udp gso benchmark")
> > > > > > > > > > > > > > > Signed-off-by: Andrei Gherzan <andrei.gherzan@...onical.com>
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Synchronizing the two processes is indeed tricky.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Perhaps more robust is opening an initial TCP connection, with
> > > > > > > > > > > > > > SO_RCVTIMEO to bound the waiting time. That covers all tests in one
> > > > > > > > > > > > > > go.
> > > > > > > > > > > > >
> > > > > > > > > > > > > Another option would be waiting for the listener(tcp)/receiver(udp)
> > > > > > > > > > > > > socket to show up in 'ss' output before firing-up the client - quite
> > > > > > > > > > > > > alike what mptcp self-tests are doing.
> > > > > > > > > > > >
> > > > > > > > > > > > I like this idea. I have tested it and it works as expected with the
> > > > > > > > > > > > exeception of:
> > > > > > > > > > > >
> > > > > > > > > > > > ./udpgso_bench_tx: sendmsg: No buffer space available
> > > > > > > > > > > >
> > > > > > > > > > > > Any ideas on how to handle this? I could retry and that works.
> > > > > > > > > > >
> > > > > > > > > > > This happens (also) without the zerocopy flag, right? That
> > > > > > > > > > >
> > > > > > > > > > > It might mean reaching the sndbuf limit, which can be adjusted with
> > > > > > > > > > > SO_SNDBUF (or SO_SNDBUFFORCE if CAP_NET_ADMIN). Though I would not
> > > > > > > > > > > expect this test to bump up against that limit.
> > > > > > > > > > >
> > > > > > > > > > > A few zerocopy specific reasons are captured in
> > > > > > > > > > > https://www.kernel.org/doc/html/latest/networking/msg_zerocopy.html#transmission.
> > > > > > > > > >
> > > > > > > > > > I have dug a bit more into this, and it does look like your hint was in
> > > > > > > > > > the right direction. The fails I'm seeing are only with the zerocopy
> > > > > > > > > > flag.
> > > > > > > > > >
> > > > > > > > > > From the reasons (doc) above I can only assume optmem limit as I've
> > > > > > > > > > reproduced it with unlimited locked pages and the fails are transient.
> > > > > > > > > > That leaves optmem limit. Bumping the value I have by default (20480) to
> > > > > > > > > > (2048000) made the sendmsg succeed as expected. On the other hand, the
> > > > > > > > > > tests started to fail with something like:
> > > > > > > > > >
> > > > > > > > > > ./udpgso_bench_tx: Unexpected number of Zerocopy completions:    774783
> > > > > > > > > > expected    773707 received
> > > > > > > > >
> > > > > > > > > More zerocopy completions than number of sends. I have not seen this before.
> > > > > > > > >
> > > > > > > > > The completions are ranges of IDs, one per send call for datagram sockets.
> > > > > > > > >
> > > > > > > > > Even with segmentation offload, the counter increases per call, not per segment.
> > > > > > > > >
> > > > > > > > > Do you experience this without any other changes to udpgso_bench_tx.c.
> > > > > > > > > Or are there perhaps additional sendmsg calls somewhere (during
> > > > > > > > > initial sync) that are not accounted to num_sends?
> > > > > > > >
> > > > > > > > Indeed, that looks off. No, I have run into this without any changes in
> > > > > > > > the tests (besides the retry routine in the shell script that waits for
> > > > > > > > rx to come up). Also, as a data point.
> > > > > > >
> > > > > > > Actually wait. I don't think that is the case here. "expected" is the
> > > > > > > number of sends. In this case we sent 1076 more messages than
> > > > > > > completions. Am I missing something obvious?
> > > > > >
> > > > > > Oh indeed.
> > > > > >
> > > > > > Receiving fewer completions than transmission is more likely.
> > > > >
> > > > > Exactly, yes.
> > > > >
> > > > > > This should be the result of datagrams still being somewhere in the
> > > > > > system. In a qdisc, or waiting for the network interface to return a
> > > > > > completion notification, say.
> > > > > >
> > > > > > Does this remain if adding a longer wait before the final flush_errqueue?
> > > > >
> > > > > Yes and no. But not realiably unless I go overboard.
> > > > >
> > > > > > Or, really, the right fix is to keep polling there until the two are
> > > > > > equal, up to some timeout. Currently flush_errqueue calls poll only
> > > > > > once.
> > > > >
> > > > > That makes sense. I have implemented a retry and this ran for a good
> > > > > while now.
> > > > >
> > > > > -               flush_errqueue(fd, true);
> > > > > +               while (true) {
> > > > > +                       flush_errqueue(fd, true);
> > > > > +                       if ((stat_zcopies == num_sends) || (delay >= MAX_DELAY))
> > > > > +                               break;
> > > > > +                       usleep(delay);
> > > > > +                       delay *= 2;
> > > > > +               }
> > > > >
> > > > > What do you think?
> > > >
> > > > Thanks for running experiments.
> > > >
> > > > We can avoid the unconditional sleep, as the poll() inside
> > > > flush_errqueue already takes a timeout.
> > > >
> > > > One option is to use start_time = clock_gettime(..) or gettimeofday
> > > > before poll, and restart poll until either the exit condition or
> > > > timeout is reached, with timeout = orig_time - elapsed_time.
> > >
> > > Yes, this was more of a quick draft. I was thinking to move it into the
> > > flush function (while making it aware of num_sends via a parameter):
> > >
> > > if (do_poll) {
> > >   struct pollfd fds = {0};
> > >   int ret;
> > >   unsigned long tnow, tstop;
> > >
> > >   fds.fd = fd;
> > >   tnow = gettimeofday_ms();
> > >   tstop = tnow + POLL_LOOP_TIMEOUT_MS;
> > >   while ((stat_zcopies != num_sends) && (tnow < tstop)) {
> 
> The new condition to loop until stat_zcopies == num_sends should only
> be tested on the final call. This likely needs to become a separate
> boolean. Or a separate flush_errqueue_last() function, and leave the
> existing one as is.

Wouldn't a do/while be enough here?

> 
> We can probably merge the outer for and inner while loops
> 
> > >     ret = poll(&fds, 1, 500);
> 
> Instead of 500, this becomes tstop - tnow.

Right. Missed this one.

> 
> > >     if (ret == 0) {
> > >       if (cfg_verbose)
> > >         fprintf(stderr, "poll timeout\n");
> 
> Poll timeouts are now expected to an extent. Only report once at the
> end of the function if the poll was only called once and timed out.

I had to think about this a bit but now I see your point and it makes
sense.

> > >       } else if (ret < 0) {
> > >         error(1, errno, "poll");
> > >     }
> > >     tnow = gettimeofday_ms();
> > >   }
> > > }
> > >
> > > Does this make more sense?
> >
> > Obviously, this should be a do/while. Anyway, this works as expected
> > after leaving it for a around two hours.
> 
> Great to hear you found the cause.

Hats off for hints.

-- 
Andrei Gherzan

Powered by blists - more mailing lists