Re: [PATCH][2.5] replace flush_map() in arch/i386/mm/pageattr.c w ith flush_tlb_all()

[Thomas Schlichter <schlicht@uni-mannheim.de>]

[-- Attachment #1: signed data --]
[-- Type: text/plain, Size: 1773 bytes --]

Yes, you are right. I was just looking for this preempt-problem where a 
flush_tlb* was done, but there are many other places where this problem 
occours, too... Nearly everywhere where smp_call_function() is used!

I found a function in the file mm/slab.c called smp_call_function_all_cpus() 
which tries to do the thing we want, but I think not even this function is 
preempt-safe...!

But here I think I better get off my fingers, I just wanted to help a bit with 
this issue, but I don't think I have the time and knowledge to solve it 
completely in the mentioned good way...

Perhaps even the semantic of the function smp_call_function() could be changed 
to call the function on every CPU? Just an idea...

Best regards
  Thomas Schlichter


On Fri, 21 Feb 2003 15:20, Dave Jones wrote:
> On Fri, Feb 21, 2003 at 01:42:12PM +0100, Thomas Schlichter wrote:
>  > > No.  All that does is make sure that the cpu you start out on is
>  > > flushed, once or twice, and the cpu you end up on may be missed.
>  > > Use preempt_disable and preempt_enable.
>  >
>  > Oh, you are right! I think I am totally stupid this morning...!
>  > Now finally I hope this is the correct patch...
>
> That would appear to do what you want, but its an ugly construct to
> be repeating everywhere that wants to call a function on all CPUs.
> It would probably clean things up a lot if we had a function to do..
>
> static inline void on_each_cpu(void *func)
> {
> #ifdef CONFIG_SMP
> 	preempt_disable();
> 	smp_call_function(func, NULL, 1, 1);
> 	func(NULL);
> 	preempt_enable();
> #else
> 	func(NULL);
> #endif
> }
>
> Bluesmoke and agpgart could both use this to cleanup some mess,
> and no doubt there are others
>
> Comments?
>
> 		Dave

[-- Attachment #2: signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]