On 08/16/2017 02:23 PM, Andrew Cooper wrote:
> They are only referenced by physical address (either the HSA MSR, or via
> VMSAVE/VMLOAD which take a physical operand). Allocating xenheap hages and
> storing their virtual address is wasteful.
>
> Allocate them with domheap pages instead, taking the opportunity to suitably
> NUMA-position them. This avoids Xen needing to perform a virt to phys
> translation on every context switch.
>
> Signed-off-by: Andrew Cooper <andrew.coop...@citrix.com>
Reviewed-by: Boris Ostrovsky <boris.ostrov...@oracle.com>
> ---
> CC: Jan Beulich <jbeul...@suse.com>
> CC: Boris Ostrovsky <boris.ostrov...@oracle.com>
> CC: Suravee Suthikulpanit <suravee.suthikulpa...@amd.com>
>
> TODO at some other point: Figure out why svm_cpu_up_prepare() is reliably
> called twice for every CPU.
That's because it is called by BSP via PREPARE_CPU notifier and then by
the ASP during svm_cpu_up().
I think
diff --git a/xen/arch/x86/hvm/svm/svm.c b/xen/arch/x86/hvm/svm/svm.c
index 0dc9442..3e7b9fc 100644
--- a/xen/arch/x86/hvm/svm/svm.c
+++ b/xen/arch/x86/hvm/svm/svm.c
@@ -1538,7 +1538,7 @@ static int _svm_cpu_up(bool bsp)
return -EINVAL;
}
- if ( (rc = svm_cpu_up_prepare(cpu)) != 0 )
+ if ( bsp && (rc = svm_cpu_up_prepare(cpu)) != 0 )
return rc;
write_efer(read_efer() | EFER_SVME);
should take care of this. I only had a quick look at intel and seems
they may have the same problem.
-boris
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel
