Skip to contentSkip to conversation

2.6.15-rt4-sr1

DISCUSSION Steven Rostedt 2 participants 2 messages Jan 11, 2006

lore
Share & tools
lore
#0 lore raw

I've just uploaded my new maintenance release.

fixes:

- light soft lockup - currently it doesn't detect the light soft lockups because the counter was being reset in the wrong place. Not sure if this was done earlier with some other issue. But I needed this fix to get the low priority lockup to be detected.

- posix_timers deadlock - There's a loop in the posix_timeres code that is entered if the current process is a higher priority than the softirqd thread, and it spins until the softirqd thread is finished. But since the thread is of a higher priority than the softirqd, it deadlocks.

- read_trylock_rt - added this for cases that have a loop on read_trylock. Since in RT a read lock is a mutex, a read_trylock will fail even if the owner has it only for reading. But since the reads are not protected with irqsaves, this can deadlock when a high priority process preempts the owner of the lock and spins till it's released. (this was seen in signal.c: send_group_sigqueue, there may be others)

Still at the normal location:

http://home.stny.rr.com/rostedt/patches/patch-2.6.15-rt4-sr1

Here is the program that showed all these nice deadlocks ;)

http://www.kihontech.com/tests/rt/timer_stress.c

(Run with -P to get the posix deadlocks)

-- Steve

#1 lore raw
Quoted 5 lines · Steven Rostedt

- posix_timers deadlock - There's a loop in the posix_timeres code that is entered if the current process is a higher priority than the softirqd thread, and it spins until the softirqd thread is finished. But since the thread is of a higher priority than the softirqd, it deadlocks.

Good catch, thanks

tglx