[PATCH] x86: bau_msg_head points to a_queue_first in uv_table_bases_finish()

[PATCH] x86: bau_msg_head points to a_queue_first in uv_table_bases_finish()

[Roel Kluin]

bau_msg_head should point to bau_msg_head

Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
---
Is this correct?

diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c
index 364d015..3d66331 100644
--- a/arch/x86/kernel/tlb_uv.c
+++ b/arch/x86/kernel/tlb_uv.c
@@ -691,7 +691,7 @@ uv_table_bases_finish(int blade,
 			continue;
 
 		bcp = (struct bau_control *)&per_cpu(bau_control, cpu);
-		bcp->bau_msg_head	= bau_tablesp->va_queue_first;
+		bcp->bau_msg_head	= bau_tablesp->bau_msg_head;
 		bcp->va_queue_first	= bau_tablesp->va_queue_first;
 		bcp->va_queue_last	= bau_tablesp->va_queue_last;
 		bcp->msg_statuses	= bau_tablesp->msg_statuses;

Re: [PATCH] x86: bau_msg_head points to a_queue_first in uv_table_bases_finish()

[Ingo Molnar]

* Roel Kluin <roel.kluin@gmail.com> wrote:

> bau_msg_head should point to bau_msg_head
> 
> Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
> ---
> Is this correct?
> 
> diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c
> index 364d015..3d66331 100644
> --- a/arch/x86/kernel/tlb_uv.c
> +++ b/arch/x86/kernel/tlb_uv.c
> @@ -691,7 +691,7 @@ uv_table_bases_finish(int blade,
>  			continue;
>  
>  		bcp = (struct bau_control *)&per_cpu(bau_control, cpu);
> -		bcp->bau_msg_head	= bau_tablesp->va_queue_first;
> +		bcp->bau_msg_head	= bau_tablesp->bau_msg_head;
>  		bcp->va_queue_first	= bau_tablesp->va_queue_first;
>  		bcp->va_queue_last	= bau_tablesp->va_queue_last;
>  		bcp->msg_statuses	= bau_tablesp->msg_statuses;

Seems like you either caught a real bug - or that there's a somewhat quirky 
initialization sequence here which should be commented upon in the source.

Cliff, Robin, what's your take on Roel's patch?

Thanks,

	Ingo

Re: [PATCH] x86: bau_msg_head points to a_queue_first in uv_table_bases_finish()

[Robin Holt]

On Sun, Jan 17, 2010 at 08:12:27AM +0100, Ingo Molnar wrote:
> 
> * Roel Kluin <roel.kluin@gmail.com> wrote:
> 
> > bau_msg_head should point to bau_msg_head
> > 
> > Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
> > ---
> > Is this correct?
> > 
> > diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c
> > index 364d015..3d66331 100644
> > --- a/arch/x86/kernel/tlb_uv.c
> > +++ b/arch/x86/kernel/tlb_uv.c
> > @@ -691,7 +691,7 @@ uv_table_bases_finish(int blade,
> >  			continue;
> >  
> >  		bcp = (struct bau_control *)&per_cpu(bau_control, cpu);
> > -		bcp->bau_msg_head	= bau_tablesp->va_queue_first;
> > +		bcp->bau_msg_head	= bau_tablesp->bau_msg_head;
> >  		bcp->va_queue_first	= bau_tablesp->va_queue_first;
> >  		bcp->va_queue_last	= bau_tablesp->va_queue_last;
> >  		bcp->msg_statuses	= bau_tablesp->msg_statuses;
> 
> Seems like you either caught a real bug - or that there's a somewhat quirky 
> initialization sequence here which should be commented upon in the source.
> 
> Cliff, Robin, what's your take on Roel's patch?

I think I will work with Cliff on rewriting uv_payload_queue_init().

Currently, that does not set up bau_msg_head so the code above is
functional.  In my opinion, it should set up bau_msg_head.  I am sure
I can be convinced otherwise.

Some of the other cleanups:

1)  Use the standard alignment macros instead of hand coding our own.
2)  Replace the kmalloc with a kzalloc and remove the memset which
    currently is zeroing the unaligned allocation instead of the aligned
    allocation so it would miss the last (up to 31) bytes of the last
    payload.

Thanks,
Robin

Re: [PATCH] x86: bau_msg_head points to a_queue_first in uv_table_bases_finish()

[Cliff Wickman]

Hi Ingo,

On Sun, Jan 17, 2010 at 08:12:27AM +0100, Ingo Molnar wrote:
> 
> * Roel Kluin <roel.kluin@gmail.com> wrote:
> 
> > bau_msg_head should point to bau_msg_head
> > 
> > Signed-off-by: Roel Kluin <roel.kluin@gmail.com>
> > ---
> > Is this correct?
> > 
> > diff --git a/arch/x86/kernel/tlb_uv.c b/arch/x86/kernel/tlb_uv.c
> > index 364d015..3d66331 100644
> > --- a/arch/x86/kernel/tlb_uv.c
> > +++ b/arch/x86/kernel/tlb_uv.c
> > @@ -691,7 +691,7 @@ uv_table_bases_finish(int blade,
> >  			continue;
> >  
> >  		bcp = (struct bau_control *)&per_cpu(bau_control, cpu);
> > -		bcp->bau_msg_head	= bau_tablesp->va_queue_first;
> > +		bcp->bau_msg_head	= bau_tablesp->bau_msg_head;
> >  		bcp->va_queue_first	= bau_tablesp->va_queue_first;
> >  		bcp->va_queue_last	= bau_tablesp->va_queue_last;
> >  		bcp->msg_statuses	= bau_tablesp->msg_statuses;
> 
> Seems like you either caught a real bug - or that there's a somewhat quirky 
> initialization sequence here which should be commented upon in the source.
> 
> Cliff, Robin, what's your take on Roel's patch?
> 
> Thanks,
> 
> 	Ingo

The bau_tablesp->bau_msg_head doesn't seem to be initialized, so do not
change the code this way.

bcp->bau_msg_head is correctly initialized to the head of the queue as
it stands.

However I do agree that it is confusing, as there are redundant control
structures, and needs improvement.  I intend to submit a new version
of tlb_uv.c soon.

-Cliff

-- 
Cliff Wickman
SGI
cpw@sgi.com
(651) 683-3824