[LTP] [PATCH] [RFC] lib: shell: Fix timeout process races

[Petr Vorel]

Hi Cyril,

Reviewed-by: Petr Vorel <pvorel@suse.cz>
Tested-by: Petr Vorel <pvorel@suse.cz>

> There were actually several races in the shell library timeout handling.

> This commit fixes hopefully all of them by:

> * Reimplementing the backgroud timer in C
+1

> * Making sure that the timer has started before we leave test setup
+1

> The rewrite of the backround timer to C allows us to run all the timeout
> logic in a single process, which simplifies the whole problem greatly
> since previously we had chain of processes that estabilished signal
> handlers to kill it's descendants, which in the end had a few races in
> it.

> The race that caused the problems is, as far as I can tell, in the way
> how shell spawns it's children. I haven't checked the shell code, but I
> guess that when shell runs a process in bacground it does vfork() +
> exec() and because signals are ignored during the operation. If the
> SIGTERM arrives at that point it gets lost.

> That means that we created a race window in the shell library each time
> we started a new process. The rewrite to C simplifies the code but we
> still end up with a single place where this can happen and that is when
> we execute the tst_timeout_kill binary. This is now fixed in the shell
> library by waiting until the background process gets to a sleep state,
> which means that the proces has been executed and waiting for the
> timeout.

> After these fixes I haven't been able to reproduce the hang with:

> cat > debug.sh <<EOF
> #!/bin/sh

> TST_SETUP="setup"
> TST_TESTFUNC="do_test"
> . tst_test.sh

> setup()
> {
>         tst_brk TCONF "quit now!"
> }

> do_test()
> {
>         tst_res TPASS "pass :)"
> }

> tst_run
> EOF

> # while true; do ./debug.sh; done
I verified it's ok on both VM which were previously affected.

After release I might write a test for tst_timeout_kill.c.
Thanks for fixing it!

Kind regards,
Petr