rc_unmount: add parallel unmounting #811
Conversation
N-R-K
force-pushed
the
rc_unmount
branch
3 times, most recently
from
263d4d6 to
b586bf6
Compare
N-R-K
force-pushed
the
rc_unmount
branch
2 times, most recently
from
d3c6a6c to
e52d16f
Compare
| res = do_exec(&args); | ||
| if (res.pid > 0) | ||
| rc_waitpid(res.pid); |
There was a problem hiding this comment.
The first fuser command to query the pids has a timeout and is async, this one to send the signals on the other hand is synchronous and has no timeout (copied from the rc-mount.sh script).
But this doesn't seem to make sense (?) If we anticipate fuser might take a while then shouldn't we be timing out both of the calls rather than just one of them?
Also the only reason we're doing 2 commands (query pids + send signal) is to avoid killing ourselves if we happen to be using that mount. If we can figure out whether we're using it or not beforehand then it can be changed into a single fuser call which directly sends the signal (async, with timeout).
openrc-run has a (hardcoded) 60 seconds timeout, it a service fails to start/stop within that time, it will be SIGKILL-ed. so a default 60 seconds fuser_timeout is kind of bad. change it to 20seconds which is somewhat more sensible.
this turns mouninfo into a multicall binary which can also do parallel unmounting. most mountinfo code is unchanged; only notable change is find_mounts() now returns the number of mounts found and also takes the "process" function as an argument; so that is_mounted() can be implemented on top of it. do_unmount mostly follows the logic of previous code. some notable changes: - do_unmount is now "lazy" when it comes to retrying failed unmounts. it will greedily keep running unmount as long as it can before it looks at the "waiting" queue. - it will check if the mountpoint is still mounted or not when umount returns non-zero exit code. this is due to the fact that multiple umount calls might race to unmount a shared mount. so if umount fails _but_ the mountpoint is no longer mounted, we assume success. - do_unmount used to fail if fuser did not find any pids using the mount. the new code tries one more time; the rationale being that there's a gap between the umount call and the fuser call, and so whatever was using the mount before might have stopped using it now and so it's worth another attempt. Fixes: OpenRC#662 Closes: OpenRC#698
navi-desu
force-pushed
the
rc_unmount
branch
from
b00524f to
b1dcc9b
Compare
|
rebased and renamed it back to do_unmount, and removed the manpage entry, i'd like to keep it internal for now, like it used to be |
2 participants
this turns mouninfo into a multicall binary which can also do
parallel unmounting. most mountinfo code is unchanged; only
notable change is find_mounts() now returns the number of mounts
found and also takes the "process" function as an argument; so
that is_mounted() can be implemented on top of it.
rc_unmount mostly follows the logic of previous do_unmount. some
notable changes:
unmounts. it will greedily keep running unmount as long as it
can before it looks at the "waiting" queue.
not when umount returns non-zero exit code. this is due to the
fact that multiple umount calls might race to unmount a shared
mount. so if umount fails but the mountpoint is no longer
mounted, we assume success.
the mount. rc_unmount tries one more time; the rationale being
that there's a gap between the umount call and the fuser call,
and so whatever was using the mount before might have stopped
using it now and so it's worth another attempt.
Closes: #662