Messages in this thread | | | Date | Wed, 28 Jan 2009 13:52:47 -0500 (EST) | From | Steven Rostedt <> | Subject | Re: Buggy IPI and MTRR code on low memory |
| |
On Wed, 28 Jan 2009, Peter Zijlstra wrote:
> On Wed, 2009-01-28 at 12:24 -0500, Steven Rostedt wrote: > > On Wed, 28 Jan 2009, Peter Zijlstra wrote: > > > --- > > > Subject: x86: fix potential deadlock in set_mtrr() > > > From: Peter Zijlstra <a.p.zijlstra@chello.nl> > > > Date: Wed Jan 28 17:17:32 CET 2009 > > > > > > smp_call_function() can fall-back to waiting on completion in case of > > > low memory (GFP_ATOMIC). set_mtrr() relies on the async behaviour of !wait. > > > > > > This would deadlock. > > > > > > Fix this by providing per-cpu csd's and using __smp_call_function_single(). > > > > I applied the patch but it still locked up in the testing NMI > > watchdog code. Not saying your patch is at fault, because we never made it > > to your code. This is probably not an issue, since if we have low memory > > at bootup, we have much bigger problems to deal with. > > > > I'll have skip the NMI test and see if it locks up any place else. > > Right, the NMI code does exactly the same, in that case your patch might > well be the best way forward.
Is there any reason to force the user to have to deal with a wait? We can add this code to handle the failed alloc case, and I can even make it a bit better to only do the copy *d = data; when the RELEASE flag is set.
-- Steve
| |