run-debuginfod-webapi-concurrency.sh (Was: ☠ Buildbot (GNU Toolchain): elfutils - failed test (failure)) (master)

[Mark Wielaard]

Hi,

On Sat, Apr 23, 2022 at 03:31:04AM +0200, Mark Wielaard wrote:
> On Sat, Apr 23, 2022 at 01:19:53AM +0000, builder--- via Elfutils-devel wrote:
> > A new failure has been detected on builder elfutils-debian-ppc64 while building elfutils.
> > 
> > Full details are available at:
> >     https://builder.sourceware.org/buildbot/#builders/63/builds/4
> > 
> > Build state: failed test (failure)
> > Revision: 7b046b7c060acc32c00748ee66ac350f77bc6571
> > Worker: debian-ppc64
> > Build Reason: (unknown)
> > Blamelist: наб via Elfutils-devel <elfutils-devel@sourceware.org>
> > 
> > Steps:
> > [...]
> > - 7: make check ( failure )
> >     Logs:
> >         - stdio: https://builder.sourceware.org/buildbot/#builders/63/builds/4/steps/7/logs/stdio
> >         - test-suite.log: https://builder.sourceware.org/buildbot/#builders/63/builds/4/steps/7/logs/test-suite_log
> 
> Gah. It is run-debuginfod-webapi-concurrency.sh again with
> error_count{libmicrohttpd="Server reached connection limit.
> Closing inbound connection.\n"} 9
> 
> I guess 100 parallel lookups really is too much. I am going to lower
> it to 64.

And that didn't help. As Frank already explained in
https://sourceware.org/bugzilla/show_bug.cgi?id=28708#c11

The issue isn't the parallel lookups, but the concurrent threadpool.
The total number of connections accepted gets divided by the number of
threads in the daemon threadpool.

So one way to fix this is to lower the max number of -C THREADS::

diff --git a/tests/run-debuginfod-webapi-concurrency.sh b/tests/run-debuginfod-webapi-concurrency.sh
index 4928f6d0..12b460a8 100755
--- a/tests/run-debuginfod-webapi-concurrency.sh
+++ b/tests/run-debuginfod-webapi-concurrency.sh
@@ -33,7 +33,7 @@ cp -rvp ${abs_srcdir}/debuginfod-tars Z
 tempfiles Z
 
 
-for Cnum in "" "-C" "-C10" "-C100"
+for Cnum in "" "-C" "-C10" "-C32"
 do
     env LD_LIBRARY_PATH=$ldpath ${abs_builddir}/../debuginfod/debuginfod $VERBOSE $Cnum -d :memory: -Z .tar.xz -Z .tar.bz2=bzcat -p $PORT1 -t0 -g0 -v --fdcache-fds=0 --fdcache-prefetch-fds=0 Z >> vlog$PORT1 2>&1 &
     PID1=$!

This seem to make the test pass even on a system (s390x 2 vcpus, f36)
where it occasionally FAILs.

But there is another way to prevent the "Server reached connection
limit. Closing inbound connection." Pass the MHD_USE_ITC flag to
MHD_start_daemon:

diff --git a/debuginfod/debuginfod.cxx b/debuginfod/debuginfod.cxx
index 9c0217f6..67124771 100644
--- a/debuginfod/debuginfod.cxx
+++ b/debuginfod/debuginfod.cxx
@@ -3910,6 +3910,7 @@ main (int argc, char *argv[])
                                      | MHD_USE_EPOLL
 #endif
                                      | MHD_USE_DUAL_STACK
+                                     | MHD_USE_ITC
                                      | MHD_USE_DEBUG, /* report errors to stderr */
                                      http_port,
                                      NULL, NULL, /* default accept policy */

This has my preference since it will make sure connections are still
accepted (by the OS) instead of being immediately closed once the
connection limit is reached:

    If the connection limit is reached, MHD’s behavior depends a bit
    on other options. If MHD_USE_ITC was given, MHD will stop
    accepting connections on the listen socket. This will cause the
    operating system to queue connections (up to the listen() limit)
    above the connection limit. Those connections will be held until
    MHD is done processing at least one of the active connections. If
    MHD_USE_ITC is not set, then MHD will continue to accept() and
    immediately close() these connections.

But it has some more consequences which I think are OK, but am not
100% sure of:

MHD_USE_ITC

    Force MHD to use a signal inter-thread communication channel to
    notify the event loop (of threads) of our shutdown and other
    events. This is required if an application uses
    MHD_USE_INTERNAL_POLLING_THREAD and then performs
    MHD_quiesce_daemon (which eliminates our ability to signal
    termination via the listen socket). In these modes,
    MHD_quiesce_daemon will fail if this option was not set. Also, use
    of this option is automatic (as in, you do not even have to
    specify it), if MHD_USE_NO_LISTEN_SOCKET is specified. In
    "external" select mode, this option is always simply ignored.

    Using this option also guarantees that MHD will not call
    shutdown() on the listen socket, which means a parent process can
    continue to use the socket.

Opinions on just changing the testcasse to use -C32 as max, or to add
MHD_USE_ITC to the MHD_start_daemon flags?

Cheers,

Mark