Change in behaviour between 4.2.53 and 4.3 series

[Dmitry Mikhin]

Hello everyone.

Lately we experienced some strange termination issues in our system. The 
problem was traced down to bash "wait" not waiting for process groups.

The original implementation has been done under bash 4.2 series, and 
apparently works there, but not under 4.3 variants that we tried.

I'm not sure if our approach is indeed correct, and any alternative 
suggestions for achieving the same goals in a compliant way are welcome.

To the problem: roughly, we want to start some software (a test 
instance, as an example) in a separate process group, and then terminate 
it all at once. Something along the lines:

timeout -s TERM 1000 our_server_script & script_pid=$!

.....

later on:

kill -s TERM -- -$script_pid
wait $script_pid

The "kill" above sends the signal to the entire process group. In bash 
4.2, the "wait" will not return until all processes in the given process 
group terminate. In bash 4.3, however, the wait returns immediately. 
(Termination in our case would take a few seconds, and the delay is 
visible. As some process in the group are still running, they corrupt data.)

A small test is attached. It reproduces the pattern above. The test 
builds a small C executable using gcc to trace and report the signals 
received. The test checks for the last line of output from running 
main.sh script. It is a race condition, so, has to be run a few times. 
On Ubuntu 14.04 (bash --version gives "GNU bash, version 
4.3.11(1)-release (x86_64-pc-linux-gnu)"), the test produces something like:
$ ( for i in $( seq 10 ) ; do ./test ; done ) | sort -u
last="./baz.sh: finally exiting"
last="./foo.sh: finally exiting"
last="main terminated"

On the same system with a custom old version of bash (bash --version 
gives "GNU bash, version 4.2.53(1)-release (x86_64-unknown-linux-gnu)"):
$ ( export PATH="$HOME/usr/bash-4.2.53/bin:$PATH"; for i in $( seq 10 ) 
; do ./test ; done ) | sort -u
last="main terminated"

I.e., in the older bash the last line is always from the main, i.e., the 
wait indeed waited until all subprocesses are over, while on 4.3 some of 
the processes in the group return after the wait completed. The 4.2 
behaviour checked for 10000 repetitions of the test, and seems to be 
always the same.

Questions:

1. are our expectations for waiting for the entire process group 
correct? If not, is this a bug in 4.3?

2. if our expectations are wrong, what would be the recommended ways to 
achieve the same effect, i.e., wait for the process group to complete?

Best regards,
Dmitry Mikhin

test.04.tar.gz
Description: application/gzip