Re: Buggy IPI and MTRR code on low memory

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2009-01-28 at 11:56 -0500, Steven Rostedt wrote:
> On Wed, 28 Jan 2009, Peter Zijlstra wrote:
> 
> > On Wed, 2009-01-28 at 11:38 -0500, Steven Rostedt wrote:
> > 
> > > The problem is that if we use the stack, then we must wait for the 
> > > function to finish. But in the mtrr code, the called functions are waiting 
> > > for the caller to do something after the smp_call_function. Thus we 
> > > deadlock!
> > 
> > You'd have to 'fix' the regular fallback paths to use your scheme as
> > well.
> 
> Not sure what you mean by that. I booted just fine with the patch, and 
> forcing data to NULL. Although, I'm not saying that these should not be 
> fixed.

Yeah, I was being confused by:

> @@ -230,6 +242,20 @@ int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
>                         data = kmalloc(sizeof(*data), GFP_ATOMIC);
>                         if (data)
>                                 data->flags = CSD_FLAG_ALLOC;
> +                       else {
> +                               /*
> +                                * There exist callers that call
> +                                * functions that will wait on the caller
> +                                * to do something after calling this.
> +                                * This means we can not wait for the callee
> +                                * function to finish.
> +                                * Use the stack data but have the callee
> +                                * copy it and tell us we can continue
> +                                * before they call the function.
> +                                */
> +                               data = &d;
> +                               data->flags = CSD_FLAG_RELEASE;
> +                       }
>                 }
>                 if (!data) {
>                         data = &d;

that second !data test in there, thinking there was a fallback path
still relying on CSD_FLAG_WAIT.

It appears all is well indeed.