Re: [RFC PATCH 1/1] Move two pinned pages to non-movable node in kvm.

From: Marcelo Tosatti
Date: Fri Jun 20 2014 - 16:32:41 EST


On Fri, Jun 20, 2014 at 05:26:22PM +0300, Gleb Natapov wrote:
> On Fri, Jun 20, 2014 at 09:53:26AM -0300, Marcelo Tosatti wrote:
> > On Fri, Jun 20, 2014 at 02:15:10PM +0300, Gleb Natapov wrote:
> > > On Thu, Jun 19, 2014 at 04:00:24PM -0300, Marcelo Tosatti wrote:
> > > > On Thu, Jun 19, 2014 at 12:20:32PM +0300, Gleb Natapov wrote:
> > > > > CCing Marcelo,
> > > > >
> > > > > On Wed, Jun 18, 2014 at 02:50:44PM +0800, Tang Chen wrote:
> > > > > > Hi Gleb,
> > > > > >
> > > > > > Thanks for the quick reply. Please see below.
> > > > > >
> > > > > > On 06/18/2014 02:12 PM, Gleb Natapov wrote:
> > > > > > >On Wed, Jun 18, 2014 at 01:50:00PM +0800, Tang Chen wrote:
> > > > > > >>[Questions]
> > > > > > >>And by the way, would you guys please answer the following questions for me ?
> > > > > > >>
> > > > > > >>1. What's the ept identity pagetable for ? Only one page is enough ?
> > > > > > >>
> > > > > > >>2. Is the ept identity pagetable only used in realmode ?
> > > > > > >> Can we free it once the guest is up (vcpu in protect mode)?
> > > > > > >>
> > > > > > >>3. Now, ept identity pagetable is allocated in qemu userspace.
> > > > > > >> Can we allocate it in kernel space ?
> > > > > > >What would be the benefit?
> > > > > >
> > > > > > I think the benefit is we can hot-remove the host memory a kvm guest
> > > > > > is using.
> > > > > >
> > > > > > For now, only memory in ZONE_MOVABLE can be migrated/hot-removed. And the
> > > > > > kernel
> > > > > > will never use ZONE_MOVABLE memory. So if we can allocate these two pages in
> > > > > > kernel space, we can pin them without any trouble. When doing memory
> > > > > > hot-remove,
> > > > > > the kernel will not try to migrate these two pages.
> > > > > But we can do that by other means, no? The patch you've sent for instance.
> > > > >
> > > > > >
> > > > > > >
> > > > > > >>
> > > > > > >>4. If I want to migrate these two pages, what do you think is the best way ?
> > > > > > >>
> > > > > > >I answered most of those here: http://www.mail-archive.com/kvm@xxxxxxxxxxxxxxx/msg103718.html
> > > > > >
> > > > > > I'm sorry I must missed this email.
> > > > > >
> > > > > > Seeing your advice, we can unpin these two pages and repin them in the next
> > > > > > EPT violation.
> > > > > > So about this problem, which solution would you prefer, allocate these two
> > > > > > pages in kernel
> > > > > > space, or migrate them before memory hot-remove ?
> > > > > >
> > > > > > I think the first solution is simpler. But I'm not quite sure if there is
> > > > > > any other pages
> > > > > > pinned in memory. If we have the same problem with other kvm pages, I think
> > > > > > it is better to
> > > > > > solve it in the second way.
> > > > > >
> > > > > > What do you think ?
> > > > > Remove pinning is preferable. In fact looks like for identity pagetable
> > > > > it should be trivial, just don't pin. APIC access page is a little bit
> > > > > more complicated since its physical address needs to be tracked to be
> > > > > updated in VMCS.
> > > >
> > > > Yes, and there are new users of page pinning as well soon (see PEBS
> > > > threads on kvm-devel).
> > > >
> > > > Was thinking of notifiers scheme. Perhaps:
> > > >
> > > > ->begin_page_unpin(struct page *page)
> > > > - Remove any possible access to page.
> > > >
> > > > ->end_page_unpin(struct page *page)
> > > > - Reinstantiate any possible access to page.
> > > >
> > > > For KVM:
> > > >
> > > > ->begin_page_unpin()
> > > > - Remove APIC-access page address from VMCS.
> > > > or
> > > > - Remove spte translation to pinned page.
> > > >
> > > > - Put vcpu in state where no VM-entries are allowed.
> > > >
> > > > ->end_page_unpin()
> > > > - Setup APIC-access page, ...
> > > > - Allow vcpu to VM-entry.
> > > >
> > > I believe that to handle identity page and APIC access page we do not
> > > need any of those.
> > > We can use mmu notifiers to track when page begins
> > > to be moved and we can find new page location on EPT violation.
> >
> > Does page migration hook via mmu notifiers? I don't think so.
> >
> Both identity page and APIC access page are userspace pages which will
> have to be unmap from process address space during migration. At this point
> mmu notifiers will be called.

Right.

> > It won't even attempt page migration because the page count is
> > increased (would have to confirm though). Tang?
> >
> Of course, we should not pin.
>
> > The problem with identity page is this: its location is written into the
> > guest CR3. So you cannot allow it (the page which the guest CR3 points
> > to) to be reused before you remove the reference.
> >
> > Where is the guarantee there will be an EPT violation, allowing a vcpu
> > to execute with guest CR3 pointing to page with random data?
> >
> A guest's physical address is written into CR3 (0xfffbc000 usually),
> not a physical address of an identity page directly. When a guest will
> try to use CR3 KVM will get EPT violation and shadow page code will find
> a page that backs guest's address 0xfffbc000 and will map it into EPT
> table. This is what happens on a first vmentry after vcpu creation.

Right.

> > Same with the APIC access page.
> APIC page is always mapped into guest's APIC base address 0xfee00000.
> The way it works is that when vCPU accesses page at 0xfee00000 the access
> is translated to APIC access page physical address. CPU sees that access
> is for APIC page and generates APIC access exit instead of memory access.
> If address 0xfee00000 is not mapped by EPT then EPT violation exit will
> be generated instead, EPT mapping will be instantiated, access retired
> by a guest and this time will generate APIC access exit.

Right, confused with the other APIC page which the CPU writes (the vAPIC page)
to.

> > > > Because allocating APIC access page from distant NUMA node can
> > > > be a performance problem, i believe.
> > > I do not think this is the case. APIC access page is never written to,
> > > and in fact SDM advice to share it between all vcpus.
> >
> > Right.
> >
> > But the point is not so much relevant as this should be handled for
> > PEBS pages which would be interesting to force to non-movable zones.
> >
> IIRC your shadow page pinning patch series support flushing of ptes
> by mmu notifier by forcing MMU reload and, as a result, faulting in of
> pinned pages during next entry. Your patch series does not pin pages
> by elevating their page count.

No but PEBS series does and its required to stop swap-out
of the page.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/