RT and Signals

RT and Signals

[Daniel Walker]

	I'm not sure if this has changed at all in recent RT patches, but I've
noticed several issues popping up that are related to the timer
interrupt sending signals , one in particular is the fact that
send_sig() calls into __cache_alloc() which has it's interrupt disable
protections removed in RT . I've observed slab corruption due to this
while running lmbench and LTP .

	Another issue was a livelock related to the timer interrupt calling
send_sig which locks tasklist_lock and siglock , which are both mutexes
(deadlock detect was on , but didn't trigger).. 

	LTP and lmbench seem to bring all these issues to the surface, but they
are all different depending on the architecture. I've been treating the
symptoms , but not the disease .. Ultimately , we need some protections,
in signal deliver, to stop the timer interrupt .. 

Daniel

Re: RT and Signals

[john cooper]

Daniel Walker wrote:
> 	I'm not sure if this has changed at all in recent RT patches, but I've
> noticed several issues popping up that are related to the timer
> interrupt sending signals...

I've also seen BUG asserts kicking in on PPC (40-04-ish) in
signal delivery [actual receipt] paths.  These have only been
under fairly heavy load conditions and presumably is hitting
an infrequent path in force_sig_info() IIRC.  Haven't had the
time yet to resolve these but they are on the list.

-john


-- 
john.cooper@timesys.com

Re: RT and Signals

[Daniel Walker]

On Mon, 2005-04-11 at 11:03, john cooper wrote:
> Daniel Walker wrote:
> > 	I'm not sure if this has changed at all in recent RT patches, but I've
> > noticed several issues popping up that are related to the timer
> > interrupt sending signals...
> 
> I've also seen BUG asserts kicking in on PPC (40-04-ish) in
> signal delivery [actual receipt] paths.  These have only been
> under fairly heavy load conditions and presumably is hitting
> an infrequent path in force_sig_info() IIRC.  Haven't had the
> time yet to resolve these but they are on the list.

	Ingo's solution was to remove the signal delivery from the timer
interrupt altogether. Which I don't know if that's acceptable. It seems
like there is a never ending stream of bugs related to this..

Daniel