RE: [Xen-ia64-devel] [PATCH] Enable hash vtlb

["Xu, Anthony" <anthony.xu@xxxxxxxxx>]

Hi Tristan,
Thanks for your comments

>From: Tristan Gingold [mailto:Tristan.Gingold@xxxxxxxx]
>Sent: 2006年4月7日 20:11
>To: Xu, Anthony; xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
>Subject: Re: [Xen-ia64-devel] [PATCH] Enable hash vtlb
>
>Le Vendredi 07 Avril 2006 13:41, Xu, Anthony a écrit :
>> This patch is to enable hash vtlb on para domains.
>Hi,
>
>a few formal points:
>
>You'd better to split your patch, because it has 3 parts: xm builder, warnings
>fix and hash vtlb.
>
Accept, some of these modifications come from debugging, I put them together.

>I'd prefer not to have vmx_ calls from vcpu.c and others.  I think it is
>clearer to only vmx_* only for vti.
>
The ideal method is to use same approach to handle para or vti domain as 
possible, 
That will reduce maintenance efforts in future, we are moving in this way,
Eventually the vmx_* will be eliminated.


>My only concern for performance is SMP-g: if we have to add locks,
>performances will decrease.  (You may reply this is my job :-)
>
Seems there are two approaches here.
1. Directly SMP-g and hash-vtlb without collision chain.
2. IPI SMP-g and hash-vtlb with collision chain.
We will let overall performance data on SMP-g to select good one.
Disabling collision chain is not difficult, We can add a option 
to disable collision chain which can accommodate your approach.
We should accommodate multiple methods in the same time, unless one is
Overwhelming.


>And a question:
>Does this patch fixes the gcc segfault bug ?  (if yes this is very good, if no
>this means the bug may not be in this area).
>
I'm not sure, but I never got gcc segfault when I got kernel build performance 
Data. We need further test.

>Tristan.

_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel