Re: [Xen-devel] [PATCH v1 1/2] x86/mm: Add mem access rights to NPT

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 05.06.18 at 16:45, <aisaila@xxxxxxxxxxxxxxx> wrote:
> On Mi, 2018-05-30 at 14:56 +0100, Andrew Cooper wrote:
>> On 30/05/18 12:29, Jan Beulich wrote:
>> >
>> > >
>> > > >
>> > > > >
>> > > > > On 30.05.18 at 13:20, <aisaila@xxxxxxxxxxxxxxx> wrote:
>> > > On Mi, 2018-05-30 at 03:52 -0600, Jan Beulich wrote:
>> > > >
>> > > > >
>> > > > > >
>> > > > > > >
>> > > > > > > On 30.05.18 at 11:04, <aisaila@xxxxxxxxxxxxxxx> wrote:
>> > > > > Sorry for the misunderstanding, I wanted to clarify if the
>> > > > > 59:56
>> > > > > bits
>> > > > > are fully ok to be used or if not then where should I use 4
>> > > > > bits to
>> > > > > store the mem access info?
>> > > > I thought I had sufficiently explained this - you have two
>> > > > options:
>> > > > 1) Make sure (via some prereq patch(es)) bit 59 has no other
>> > > > use, and
>> > > >    then use 59:56.
>> > > > 2) Use another range that's provably having no other use, e.g.
>> > > >    58:55.
>> > > I've checked and bits 40:52 are defined in asm/page.h for page
>> > > flags.
>> > 40:52? Hardly.
>> >
>> > >
>> > > I've tried bits 53:56 and there where some problems with the
>> > > guest not
>> > > starting or the image freezing,
>> > Well, you'll have to explain this (perhaps just to yourself).
>> >
>> > >
>> > > bits 62 and 63 are not free so 59:56 is
>> > > the only space to be used for this purpose and is seems to
>> > > function
>> > > correctly.
>> > Well - as said before, bit 59 is not available for use without some
>> > prereq work.
>> There are no software available bits in the top of an AMD IOMMU PTE.
>> Bits 59:62 are defined, while bits 52:58 are strictly reserved and
>> fault
>> if used.
>>
>> I'm also not convinced of the safety of our current uses of bits 9:11
>> which are software available in the regular pagetables, but have
>> specific meaning in the IOMMU entries.
>>
>> If the code IOMMU code disables page sharing, then lets go one small
>> step further and prohibit its use entirely.  There is no point trying
>> to
>> maintain compatibility for an option which isn't used, especially if
>> it
>> gets in the way of improvements like this in the SVM code.
>>
> Another idea is to save the mem access info in a radix tree like on the
> ARM side and we can store the radix tree root in the p2m_domain.
> 
> I think that this is the fastest and cleanest way to solve the free
> bits problem.

But it adds extra (memory, lookup time) overhead.

> Is this a suitable way to go?

If no other option exists - perhaps. But that's more a question to be
answered by George.

Jan



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel