Changing os_idle_sleep() to use pause() (I accidentally described
it as an empty select() in the commit log because I had changed it
from that to pause() in a later revision) exposed a race condition
in the idle code. The following can happen:
timer_settime(0, 0, {it_interval={tv_sec=0, tv_nsec=0}, it_value={tv_sec=0, tv_nsec=624017}}, NULL) = 0
...
<SIGALRM is delivered but we're already on the way to idle>
pause()
and we now hang forever. This was previously possible as well, but
it could never cause UML to hang for more than a second since we
could only sleep for that much, so at most you'd notice a "hiccup"
in the UML. Obviously, any sort of external interrupt also "saves"
it and interrupts pause().
Fix this by properly handling the race, rather than papering over
it again:
- first, block SIGALRM, and obtain the old signal set
- check the timer
- suspend, waiting for any signal out of the old set, if, and only
if, the timer will fire in the future
- restore the old signal mask
This ensures race-free operation: as it's blocked, the signal won't
be delivered while we're looking at the timer even if it were to be
triggered right _after_ we've returned from timer_gettime() with a
non-zero value (telling us the timer will trigger). Thus, despite
getting to sigsuspend() because timer_gettime() told us we're still
waiting, we'll not hang because sigsuspend() will return immediately
due to the pending signal.
Fixes: 49da38a3ef ("um: Simplify os_idle_sleep() and sleep longer")
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Acked-By: Anton Ivanov <anton.ivanov@cambridgegreys.com>
Signed-off-by: Richard Weinberger <richard@nod.at>
9868c2081d