Xen devel - Feb 2008 - is there any way to avoid lazy updates to page attribute bits?

I''m finding that regardless of which update method I use - writeable
page tables, HYPERVISOR_update_va_mapping, or HYPERVISOR_mmu_update
that xen will frequently only update the address bits in the page
table at the time of the hypercall - presumably setting the attribute
bits some time later when the mapping is accessed. This is fairly
problematic for me in that freebsd relies on the page attribute bits
in the page table accurately reflecting the attributes that were set
for a mapping. Is it now necessary for PV guests to maintain page
attribute data outside of the page tables?


-Kip

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

On 20/2/08 02:34, "Kip Macy" <kip.macy@gmail.com> wrote:
> I''m finding that regardless of which update method I use -
writeable
> page tables, HYPERVISOR_update_va_mapping, or HYPERVISOR_mmu_update
> that xen will frequently only update the address bits in the page
> table at the time of the hypercall - presumably setting the attribute
> bits some time later when the mapping is accessed. This is fairly
> problematic for me in that freebsd relies on the page attribute bits
> in the page table accurately reflecting the attributes that were set
> for a mapping. Is it now necessary for PV guests to maintain page
> attribute data outside of the page tables?
Attributes are always updated synchronously at the same time as the address
bits. Xen may modify the attributes a bit (e.g., fiddling with PAGE_GLOBAL
and PAGE_USER for 64-bit guests) but this is done deterministically and
synchronously; never lazily.

 -- Keir




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

On Wed, Feb 20, 2008 at 12:10 AM, Keir Fraser <Keir.Fraser@cl.cam.ac.uk>
wrote:
>
> On 20/2/08 02:34, "Kip Macy" <kip.macy@gmail.com> wrote:
>
>  > I''m finding that regardless of which update method I use -
writeable
>  > page tables, HYPERVISOR_update_va_mapping, or HYPERVISOR_mmu_update
>  > that xen will frequently only update the address bits in the page
>  > table at the time of the hypercall - presumably setting the attribute
>  > bits some time later when the mapping is accessed. This is fairly
>  > problematic for me in that freebsd relies on the page attribute bits
>  > in the page table accurately reflecting the attributes that were set
>  > for a mapping. Is it now necessary for PV guests to maintain page
>  > attribute data outside of the page tables?
>
>  Attributes are always updated synchronously at the same time as the
address
>  bits. Xen may modify the attributes a bit (e.g., fiddling with PAGE_GLOBAL
>  and PAGE_USER for 64-bit guests) but this is done deterministically and
>  synchronously; never lazily.
Ok. That isn''t what I''m seeing on 3.1 and 3.2, I''ll
instrument xen and
try to track it down.

Also, I removed the use of recursive page tables so that I could use
writeable page tables - not realizing that wr pts were a dud. I''ve
disabled writeable page tables and have tried re-enabling the use of
recursive page tables. It appears that that doesn''t work at all any
more - is that correct?


 -Kip

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

> Also, I removed the use of recursive page tables so that I could use
> writeable page tables - not realizing that wr pts were a dud. I''ve
> disabled writeable page tables and have tried re-enabling the use of
> recursive page tables. It appears that that doesn''t work at all
any
> more - is that correct?
I''m not sure what you mean. You can write directly to pagetables, the
write will just be emulated. It was only the ''unhooking''
optimization
(known as wr_pt) that was removed in 3.1, there''s no guest-visible API
change. You can of course used the batched mmu_update hypercall
interface or the update_va_mapping hypercall.

Recursive pagetables have been observed to work for a number of
different guest types (including 64b and PAE) so I''ve no reason to
believe they''re broken.

Ian



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