tree a1e04e8149c4beaada883e0b3546aa3d87cf7c87
parent ef3eb5f74022050666168922d96c79531401a16d
author Mike Kasick <mike@kasick.org> 1328542333 -0500
committer Mike Kasick <mike@kasick.org> 1332468701 -0400

init: Safely restart services to avoid race conditions.

Previously, service restarts (either an explicit "restart", or a "stop,
start" pair) exhibited a race condition whereby the new (restarting)
service process was often spawned before the old (stopping) process had
terminated.  This may have resulted in the new service process failing to
acquire a limited resource (file lock, socket bind, etc.) that the old
process had not yet released.

Now, a stopping service remains in the SVC_RUNNING state until its exiting
process has been reaped by waitpid.  This prevents a "stop, start" sequence
from spawning a second service process before resources held by the first
are released.  This enables safe service restarts by stopping the service,
waiting for the old service process to terminate, and (only then) starting
the new service process.  In the event of "restarting" an already stopped
service, the previous behavior is maintained whereby the service is simply
started.

This scenario could be special-cased by the restart command, however, we
have observed instances where services are, unintentionally, stopped and
started "too quickly," and so simultaneous processes for the same service
should never be allowed.

Note that this commit alters the behaviors for explicit restarts of
critical and oneshot services.  Previously these serivces would simply be
restarted, whereas now, an explicit restart of a critical service counts as
a crash (which may result in a recovery reboot) and oneshot services go
into the disabled state.