Pádraig Brady wrote: > On 02/08/2013 05:09 PM, Bernhard Voelker wrote: >> On February 8, 2013 at 5:30 PM "Pádraig Brady" <p...@draigbrady.com> wrote: >>> On 02/08/2013 02:53 PM, Bernhard Voelker wrote: >>>> On February 7, 2013 at 8:57 PM "Pádraig Brady" <p...@draigbrady.com> wrote: >>>> * SLES-10.3 (i586): >>>> gcc (GCC) 4.1.2 20070115 (SUSE Linux) >>>> >>>> FAIL: tests/tail-2/inotify-rotate.sh >>>> NFS issue during cleanup_ >>>> reproduced: 2x out of 2 tries. >>> >>> If you put a "wait" before the "Exit" at the end of that test, does it help. >>> As a less desirable solution we could put require_local_dir_ at the top >>> of this test. >> >> Unfortunately not: > > I think I see what's happening. With /bin/sh -> bash, > bash will create a redundant subshell for the : && timeout ... construct. > I.E. this prints bash rather than timeout: > > : && timeout 5 sleep 1 & readlink /proc/$!/exe > > So if you had dash installed, this would probably work: > > make check SHELL=dash TESTS="tests/tail-2/inotify-rotate.sh" \ > SUBDIRS=. VERBOSE=yes RUN_EXPENSIVE_TESTS=yes > > The reason the subshell causes an issue is that it won't > auto send the SIGHUP to timeout when it gets a SIGTERM > as it's a non interactive script, and so the timeout > processes and thus the tail processes accessing the files > hang around until the 40s timeout. > > In any case the attached should avoid the subshell > and hopefully fix the issue.
Oooh. That's subtle. Nice analysis and fix. > BTW I found a related signal race in bash when looking at this, > where it will ignore the SIGTERM before it execs the background > command, but I'm fairly sure we're not hitting this here > and so will only send details of that to the bash list. Now that you've described it, I'm sure I'm not the only one whose interest you've piqued. Here's the link: http://thread.gmane.org/gmane.comp.shells.bash.bugs/19561
