Hi Joerg,

On Tue, May 11, 2021 at 1:52 PM Joerg Vehlow [off-list ref] wrote:

Hi Li,

first of all thanks for fixing my patchset and getting it merged.

On 5/8/2021 7:51 AM, Li Wang wrote:

quoted

We have to guarantee _tst_kill_test alive for a while to check if
the target test eixst or not, so ignore SIGINT in _tst_kill_test
is necessary, otherwise it will be stopped by the SIGINT sending
by itself.

The timeout03.sh verify this mechanism proccess well in output:

   timeout03 1 TBROK: Test timeouted, sending SIGINT! If you are running on slow machine, try exporting LTP_TIMEOUT_MUL > 1
   timeout03 1 TBROK: test interrupted or timed out
   timeout03 1 TPASS: test run cleanup after timeout
   timeout03 1 TINFO: Test is still running, waiting 10s
   timeout03 1 TINFO: Test is still running, waiting 9s
   timeout03 1 TINFO: Test is still running, waiting 8s
   timeout03 1 TINFO: Test is still running, waiting 7s
   timeout03 1 TINFO: Test is still running, waiting 6s
   timeout03 1 TINFO: Test is still running, waiting 5s
   timeout03 1 TINFO: Test is still running, waiting 4s
   timeout03 1 TINFO: Test is still running, waiting 3s
   timeout03 1 TINFO: Test is still running, waiting 2s
   timeout03 1 TINFO: Test is still running, waiting 1s
   timeout03 1 TBROK: Test still running, sending SIGKILL
   Killed

At first I did bot understand the problem you found, because I tried
with dash, busybox sh and zsh.
All three shells had no problem here. But then I tried with bash and it
failed.

I wonder if this is a bug in bash or in the other shells. I guess
sending the signal to the whole
process group should also send it to the process running _tst_kill_test.

I did some digging into this while writing this (see conclusion below
for results only):
1. All shells have their own implementation of kill (compare <SHELL> -c
kill with /usr/bin kill)
2. When replacing "just" kill in the script with /usr/bin/kill, it still
only fails in bash.
3. zsh seems to ignore SIGINT, but it can be caught using trap. busybox
sh, and dash can't even get it when trapped
4. zsh disables SIGINT by callling "trap '' INT" internally somehow.
When resetting SIGINT to default behavior, it is the same as bash.

For zsh this seems to be default behavior for background processes,
probably to prevent keyboard interruption by CTRL+C:
zsh -c "trap&"
trap -- '' INT
trap -- '' QUIT

zsh -c "trap"
# No output



To conclude:
- bash does not seem to care about SIGINT delivery to background
processes, but can be blocked using trap
- zsh ignores SIGINT for background processes by default, but can be
allowed using trap
- dash and busybox sh ignore the signal to background processes, and
this cannot be changed with trap

I tried with the following snippets:
<SHELL> -c 'trap "echo trap;" INT; (sleep 2; echo end sub) & sleep 1;
kill -INT -$$; echo end main'
<SHELL> -c 'trap "echo trap;" INT; (trap - SIGINT sleep 2; echo end sub)
& sleep 1; kill -INT -$$; echo end main'
<SHELL> -c 'trap "echo trap;" INT; (trap "exit" SIGINT sleep 2; echo end
sub) & sleep 1; kill -INT -$$; echo end main'

Thanks for the demos above, it shows the difference clearly.

SIGINT handling for child processes is strange. This might have some
implication for the shell tests,
because it is possible, that SIGINT is not delivered to all processes
and some may reside as orphans.
Since this can happen only in case of timeouts, I guess there is no real
Problem.

Yes.

Looks like the behaviors on signal 'SIGINT' are not unify in background
processes handling for different SHELL. So as you said that using SIGINT
seems NOT a good idea to stop the process in timeout.

A possible fix could be using SIGTERM instead of SIGINT. This signal
does not seem to have some "intelligent" handling for background processes.

I agree. Can you make a patch to replace that INT?

(and this is only a timeout issue, so patch merging may be delayed due
to LTP new release)

I do not know why LTP used SIGINT in the first place. My first thought
would have been to use SIGTERM.  It is the way to "politely ask
processes to terminate"

Yes, but that not strange to me, the possible reason is just to
stop(ctrl ^c) the LTP test manually for debugging, so we went
too far for using SIGINT but forget the original purpose :).

-- 
Regards,
Li Wang