skip if timeout not installed

Florian Weimer writes:

 Christian Grothoff:

> I'm seeing some _very_ odd behavior with processes hanging on exit (?)
> with GNU libc 2.28-6 on Debian (amd64 threadripper).  This seems to
> happen at random (for random tests, with very low frequency!) in the
> GNUnet (Git master) testsuite when a child process is about to exit.

It looks like you call exit from a signal handler, see
src/util/scheduler.c:

/**
 * Signal handler called for signals that should cause us to shutdown.
 */
static void
sighandler_shutdown ()
{
  static char c;
  int old_errno = errno;        /* backup errno */

  if (getpid () != my_pid)
    exit (1);                   /* we have fork'ed since the signal handler was created,
                                 * ignore the signal, see https://gnunet.org/vfork discussion */
  GNUNET_DISK_file_write (GNUNET_DISK_pipe_handle
                          (shutdown_pipe_handle, GNUNET_DISK_PIPE_END_WRITE),
                          &c, sizeof (c));
  errno = old_errno;
}

In general, this results in undefined behavior because exit (unlike
_exit) is not an async-signal-safe function.

I suspect you either call the exit function while a fork is in progress,
or since you register this signal handler multiple times for different
signals:

  sh->shc_int = GNUNET_SIGNAL_handler_install (SIGINT,
                                               &sighandler_shutdown);
  sh->shc_term = GNUNET_SIGNAL_handler_install (SIGTERM,
                                                &sighandler_shutdown);

one call to exit might interrupt another call to exit if both signals
are delivered to the process.

The deadlock you see was introduced in commit
27761a1042daf01987e7d79636d0c41511c6df3c ("Refactor atfork handlers"),
first released in glibc 2.28.  The fork deadlock will be gone (in the
single-threaded case) if Debian updates to the current
release/2.28/master branch because we backported commit
60f80624257ef84eacfd9b400bda1b5a5e8e7816 ("nptl: Avoid fork handler lock
for async-signal-safe fork [BZ #24161]") there.

But this will not help you.  Even without the deadlock, I expect you
still experience some random corruption during exit, but it's going to
be difficult to spot.

Thanks,
Florian

fix another linker error

Merge branch 'master' of git+ssh://gnunet.org/gnunet

fix assertion failure reported in #5578

more logger

more logging

gnunet_testing.py.in: first set of logging

add debug output to py.in

add logging

cast no longer required for modern curl

make clang shut up about #5573

Merge branch 'master' of gnunet.org:gnunet

text-only variant of svg logo

fix some compiler warnings

fix dist build

fix macos build of dist tarball

Merge branch 'master' of gnunet.org:gnunet

Fix bug reported by schanzen@ where bootstrap is not enough for our python2.7 script.

die dv.pc

fix #5571

fix nat-auto port conflict

skip instead of fail hard on testbed link failures in cadet tests

fix build ats

fix HAVE_MEMRCHR

README: we are on python3.7 now

Signed-off-by: ng0 <ng0@n0.is>

Merge branch 'master' of gnunet.org:gnunet

Switch to python3.7 (integration-tests incomplete), continue using python2.7 for gnunet-qr with an incredible annoying workaround for autotools inability to deal with 2 major python versions at the same time

Signed-off-by: ng0 <ng0@n0.is>

DCE

fix #3329

mystery solved?

fix uninit e_hash issue

add missing ignores

use disjoint UNIXPATHs and PORT for NAT

trying to improve things for #5560

ignore generated files, one mroe

ignore generated files, one per line

ignore generated files

clarifying namestore api (#5458), fixing code duplication and a memory leak while at it

Merge branch 'master' of gnunet.org:gnunet

attempt to fix 5560, not fixed yet. see log at https://gnunet.org/bugs/view.php?id=5560

Signed-off-by: ng0 <ng0@n0.is>

add handle_address_consider_verify skeleton

implementing libgnunettransportaddress

undo accidental commit

trying to fix #5532

Fix dht python test failure with python3.7

pogen

dv service is dead

add more missing EXTRA_DISTs and remove dv (to be integrated with transport in TNG)

fixing 'make dist' issues

fixing 'make dist' issues

fixing 'make dist' issues

fixing EXTRA_DIST issues'

fix #5460

bug links

tox.ini: ignore bak files

Signed-off-by: ng0 <ng0@n0.is>

integration-tests: futurize gnunet_pyexpect

Signed-off-by: ng0 <ng0@n0.is>

integration-tests: futurize test_reconnect_nat

integration-tests: futurize test_reconnect

integration-tests: futurize test_integration_disconnect_nat

integration-tests: futurize test_integration_disconnect

integration-tests: futurize test_integration_bootstrap_and_connect

integration-tests: futurize test_integration_clique

Merge branch 'master' of gnunet.org:gnunet

util: futurize revocation/test_local_revocation

Signed-off-by: ng0 <ng0@n0.is>

util: futurize fs/test_gnunet_fs_psd

Signed-off-by: ng0 <ng0@n0.is>

util: futurize consensus/consensus-simulation

Signed-off-by: ng0 <ng0@n0.is>

util: futurize contrib/scripts/terminate

Signed-off-by: ng0 <ng0@n0.is>

util: futurize contrib/scripts/pydiffer

Signed-off-by: ng0 <ng0@n0.is>

util: futurize contrib/scripts/gnunet_pyexpect

Signed-off-by: ng0 <ng0@n0.is>

util: futurize contrib/scripts/gnunet-chk

Signed-off-by: ng0 <ng0@n0.is>

util: futurize gnunet-qr

Signed-off-by: ng0 <ng0@n0.is>

another issue with the #5511 patch

use idn2.h or idna.h depending on HAVE_LIBIDN/HAVE_LIBIDN

add API for address injection

update ignore file

externalizing secushare logic

fix test failure due to bad SPDX line

make test skip

add some logging

fix service shutdown regression

rename to reclaimID

more css files needed

try use manual.css instead of docstyle.css

untested fix for #5511

gnunet-tutorial: macro for older makeinfo

Signed-off-by: ng0 <ng0@n0.is>

fix #5554

rps profiler does not need zonemaster monitor

doc/handbook: Fix incorrect texinfo menus and formating, also do not use '@' signs

Signed-off-by: ng0 <ng0@n0.is>

Fix warning in texinfo

Signed-off-by: ng0 <ng0@n0.is>

Fix warning in texinfo

Signed-off-by: ng0 <ng0@n0.is>

Merge branch 'feature/texinfo-fix'

Fix typo in configure.ac

Signed-off-by: ng0 <ng0@n0.is>

configure: do not produce the word "warning" twice

Signed-off-by: ng0 <ng0@n0.is>

avoid assertions from adding headers that MHD deals with internally

doc/tutorial/Makefile: fix leftover

Signed-off-by: ng0 <ng0@n0.is>

doc/man/Makefile: rename section7 leftover to new name

Signed-off-by: ng0 <ng0@n0.is>

doc/handboc: backport a newer makeinfo macro

Signed-off-by: ng0 <ng0@n0.is>

configure: don't check for makeinfo version. rename section7 to texi2mdoc-generation

Signed-off-by: ng0 <ng0@n0.is>

Fix logging.

Give GNUNET_FORCE_LOG priority over GNUNET_LOG.