On 12/12/2013 01:34 PM, Bernhard Voelker wrote: > On 12/11/2013 11:24 AM, Pádraig Brady wrote: >>> The question is which of the tests hung? >>> The input (i.e. skip) or the output (i.e. seek) tests? >> >> I think it was the last dd, though can't be sure as >> the net connection died as I was debugging. > > Sorry, I couldn't reproduce it either, and the test code > looks good to me. Could it have been a shell issue?
Probably a shell bug. So what I've done is avoided the shell writing to the fifo. Also I've added timeouts to be defensive. Also I added a `wait` to avoid overlapping writers to the fifo. I tested this test passes without delay on GNU/Linux, Solaris and FreeBSD. The attached should hopefully be the last change before the release. thanks, Pádraig.
>From c6c10d0d0a1c03cb5fe592f2b499eb9923df02c0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?P=C3=A1draig=20Brady?= <[email protected]> Date: Wed, 11 Dec 2013 04:36:08 +0000 Subject: [PATCH] tests: avoid unlikely deadlock in dd/no-allocate on some shells * test/dd/no-allocate.sh: Use 'wait' to ensure we don't have multiple writers to the fifo, which was seen to trigger a very hard to reproduce deadlock with make -j20 on solaris. Also avoid writing to the fifo with the shell; instead using dd. (check_dd_seek_alloc): A new function refactored from the various cases, which are now constructed from function parameters. --- tests/dd/no-allocate.sh | 53 +++++++++++++++++++++++++++++++---------------- 1 files changed, 35 insertions(+), 18 deletions(-) diff --git a/tests/dd/no-allocate.sh b/tests/dd/no-allocate.sh index dd1a740..8823b88 100755 --- a/tests/dd/no-allocate.sh +++ b/tests/dd/no-allocate.sh @@ -20,34 +20,51 @@ print_ver_ dd require_ulimit_v_ -# count and skip is zero, we don't need to allocate memory +# count and skip are zero, we don't need to allocate memory (ulimit -v 20000; dd bs=30M count=0) || fail=1 (ulimit -v 20000; dd ibs=30M count=0) || fail=1 (ulimit -v 20000; dd obs=30M count=0) || fail=1 +check_dd_seek_alloc() { + local file="$1" + local buf="$2" + test "$file" = 'in' && { dd_file=if; dd_op=skip; } + test "$file" = 'out' && { dd_file=of; dd_op=seek; } + test "$buf" = 'in' && { dd_buf=ibs; } + test "$buf" = 'out' && { dd_buf=obs; } + test "$buf" = 'both' && { dd_buf=bs; } -# Use a fifo for which seek fails, but read does not -if mkfifo tape; then - # for non seekable output we need to allocate buffer when needed - echo 1 > tape& - (ulimit -v 20000; dd bs=30M skip=1 count=0 if=tape) && fail=1 - - echo 1 > tape& - (ulimit -v 20000; dd ibs=30M skip=1 count=0 if=tape) && fail=1 + # Provide input to the "tape" + timeout 10 dd count=1 if=/dev/zero of=tape& - echo 1 > tape& - (ulimit -v 20000; dd obs=30M skip=1 count=0 if=tape) || fail=1 + # Allocate buffer and read from the "tape" + (ulimit -v 20000; timeout 10 dd $dd_buf=30M $dd_op=1 count=0 $dd_file=tape) + local ret=$? + # Be defensive in case the tape reader is blocked for some reason + test $ret = 124 && framework_failure_ - # for non seekable output we need to allocate buffer when needed - echo 1 > tape& - (ulimit -v 20000; dd bs=30M seek=1 count=0 of=tape) && fail=1 + # This should happen without delay, + # and is used to ensure we've not multiple writers to the "tape" + wait - echo 1 > tape& - (ulimit -v 20000; dd obs=30M seek=1 count=0 of=tape) && fail=1 + # We want the "tape" reader to fail iff allocating + # a large buffer corresponding to the file being read + case "$file$buf" in + inout|outin) test $ret = 0;; + *) test $ret != 0;; + esac +} - echo 1 > tape& - (ulimit -v 20000; dd ibs=30M seek=1 count=0 of=tape) || fail=1 +# Use a fifo for which seek fails, but read does not. +# For non seekable output we need to allocate a buffer +# when simulating seeking with a read. +if mkfifo tape; then + for file in 'in' 'out'; do + for buf in 'both' 'in' 'out'; do + check_dd_seek_alloc "$file" "$buf" || fail=1 + done + done fi Exit $fail -- 1.7.7.6
