On Mon, 25 Nov 2019, Bharata B Rao wrote:
> On PEF-enabled POWER platforms that support running of secure guests,
> secure pages of the guest are represented by device private pages
> in the host. Such pages needn't participate in KSM merging. This is
> achieved by using ksm_madvise() call which need to be exported
> since KVM PPC can be a kernel module.
>
> Signed-off-by: Bharata B Rao <bharata@xxxxxxxxxxxxx>
> Acked-by: Paul Mackerras <paulus@xxxxxxxxxx>
> Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
I can say
Acked-by: Hugh Dickins <hughd@xxxxxxxxxx>
to this one.
But not to your 2/7 which actually makes use of it: because sadly it
needs down_write(&kvm->mm->mmap_sem) for the case when it switches off
VM_MERGEABLE in vma->vm_flags. That's frustrating, since I think it's
the only operation for which down_read() is not good enough.
I have no idea how contended that mmap_sem is likely to be, nor how
many to-be-secured pages that vma is likely to contain: you might find
it okay simply to go with it down_write throughout, or you might want
to start out with it down_read, and only restart with down_write (then
perhaps downgrade_write later) when you see VM_MERGEABLE is set.
The crash you got (thanks for the link): that will be because your
migrate_vma_pages() had already been applied to a page that was
already being shared via KSM.
But if these secure pages are expected to be few and far between,
maybe you'd prefer to keep VM_MERGEABLE, and add per-page checks
of some kind into mm/ksm.c, to skip over these surprising hybrids.
Hugh
> ---
> mm/ksm.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/mm/ksm.c b/mm/ksm.c
> index dbee2eb4dd05..e45b02ad3f0b 100644
> --- a/mm/ksm.c
> +++ b/mm/ksm.c
> @@ -2478,6 +2478,7 @@ int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
>
> return 0;
> }
> +EXPORT_SYMBOL_GPL(ksm_madvise);
>
> int __ksm_enter(struct mm_struct *mm)
> {
> --
> 2.21.0
