Medium: Filesystem: faster stop procedure when large number of proces…

[dmuhamedagic]

…ses using the filesystem are found

[krig]

Hm, why not do the mass-kill even with fewer processes? My guess is that it's to kill the processes in a predictable order. But if that is a real concern, it seems like it could be really difficult to debug problems caused by the switch to a different method at an arbitrary limit.

[dmuhamedagic]

On Tue, Oct 24, 2017 at 03:47:33AM -0700, Kristoffer Grönlund wrote: Hm, why not do the mass-kill even with fewer processes?

I guess because we want to log the processes which are holding the filesystem. It is assumed that such a situation is unexpected and such information could be useful to the administrators. Otherwise, I don't know :)

My guess is that it's to kill the processes in a predictable order. But if that is a real concern, it seems like it could be really difficult to debug problems caused by the switch to a different method at an arbitrary limit.

No, we don't really care about the order. It's not possible to establish one anyway (i.e. the pid generation can wrap).

[krig]

If it's just for the logging, perhaps something like this would work?

ps -f $pids | tail -n +2 | while read line; do
  ocf_log info "sending signal $sig to: $line"
done

and then just killing them all in a single command all the time.

[dmuhamedagic]

On Tue, Oct 24, 2017 at 04:35:12AM -0700, Kristoffer Grönlund wrote: If it's just for the logging, perhaps something like this would work? ``` ps -f $pids | tail -n +2 | while read line; do ocf_log info "sending signal $sig to: $line" done ``` and then just killing them all in a single command all the time.

Apparently, it is ocf_log that is the bottleneck, which is probably true. Somebody recently complained about it hence this change. BTW, please don't merge this yet, it's untested and I asked the poster to do some testing.

[krig]

Ah, I see.. yeah, that does make sense that it would be.

[dmuhamedagic]

Did you manage to do some further testing? This PR starting to rot a bit 😉

[oalbrigt]

Our QE suggests simply dropping the entire for-loop.

E.g.:

pidcnt=$(echo "$pids" | wc -w)
ocf_log info "sending signal $sig to $pidcnt processes"
if [ $pidcnt -le 10 ]; then
	ocf_log info $(echo "$pids" | xargs ps -f)
fi
echo "$pids" | xargs kill -s $sig

or

pidcnt=$(echo "$pids" | wc -w)
ocf_log info "sending signal $sig to $pidcnt processes:\n$(echo '$pids' | xargs ps -f)"
echo "$pids" | xargs kill -s $sig

[dmuhamedagic]

The latter seems to be the best way forward. Thank you for your work and thanks to the QE for providing the code!

[oalbrigt]

Sadly it doesnt seem to work with our improved stop-action logic:
#1878