On 2019-12-13 11:14, Christoffer Dall wrote:
On Fri, Dec 13, 2019 at 09:28:59AM +0000, Marc Zyngier wrote:
Hi Christoffer,
On 2019-12-13 08:29, Christoffer Dall wrote:
> Hi Marc,
>
> On Wed, Dec 11, 2019 at 04:56:48PM +0000, Marc Zyngier wrote:
> > A device mapping is normally always mapped at Stage-2, since
there
> > is very little gain in having it faulted in.
>
> It is actually becoming less clear to me what the real benefits of
> pre-populating the stage 2 page table are, especially given that
we can
> provoke a situation where they're faulted in anyhow. Do you
recall if
> we had any specific case that motivated us to pre-fault in the
pages?
It's only a minor performance optimization that was introduced by
Ard in
8eef91239e57d. Which makes sense for platform devices that have a
single
fixed location in memory. It makes slightly less sense for PCI,
where
you can move things around.
User space could still decide to move things around in its VA map
even
if the device is fixed.
Anyway, I was thinking more if there was some sort of device, like a
frambuffer, which for example crosses page boundaries and where it
would
be visible to the user that there's a sudden performance drop while
operating the device over page boundaries. Anything like that?
> > Nonetheless, it is possible to end-up in a situation where the
> > device
> > mapping has been removed from Stage-2 (userspace munmaped the
VFIO
> > region, and the MMU notifier did its job), but present in a
> > userspace
> > mapping (userpace has mapped it back at the same address). In
such
> > a situation, the device mapping will be demand-paged as the
guest
> > performs memory accesses.
> >
> > This requires to be careful when dealing with mapping size,
cache
> > management, and to handle potential execution of a device
mapping.
> >
> > Cc: stable@xxxxxxxxxxxxxxx
> > Reported-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
> > Signed-off-by: Marc Zyngier <maz@xxxxxxxxxx>
> > ---
> > virt/kvm/arm/mmu.c | 21 +++++++++++++++++----
> > 1 file changed, 17 insertions(+), 4 deletions(-)
> >
> > diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> > index a48994af70b8..0b32a904a1bb 100644
> > --- a/virt/kvm/arm/mmu.c
> > +++ b/virt/kvm/arm/mmu.c
> > @@ -38,6 +38,11 @@ static unsigned long io_map_base;
> > #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
> > #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
> >
> > +static bool is_iomap(unsigned long flags)
> > +{
> > + return flags & KVM_S2PTE_FLAG_IS_IOMAP;
> > +}
> > +
>
> nit: I'm not really sure this indirection makes the code more
readable,
> but I guess that's a matter of taste.
>
> > static bool memslot_is_logging(struct kvm_memory_slot *memslot)
> > {
> > return memslot->dirty_bitmap && !(memslot->flags &
> > KVM_MEM_READONLY);
> > @@ -1698,6 +1703,7 @@ static int user_mem_abort(struct kvm_vcpu
> > *vcpu, phys_addr_t fault_ipa,
> >
> > vma_pagesize = vma_kernel_pagesize(vma);
> > if (logging_active ||
> > + (vma->vm_flags & VM_PFNMAP) ||
>
> WHat is actually the rationale for this?
>
> Why is a huge mapping not permitted to device memory?
>
> Are we guaranteed that VM_PFNMAP on the vma results in device
mappings?
> I'm not convinced this is the case, and it would be better if we
can
> stick to a single primitive (either kvm_is_device_pfn, or
VM_PFNMAP) to
> detect device mappings.
For now, I've tried to keep the two paths that deal with mapping
devices
(or rather, things that we interpret as devices) as close as
possible.
If we drop the "eager" mapping, then we're at liberty to restructure
this in creative ways.
This includes potential huge mappings, but I'm not sure the rest of
the
kernel uses them for devices anyway (I need to find out).
> As a subsequent patch, I'd like to make sure that at the very
least our
> memslot prepare function follows the exact same logic for mapping
device
> memory as a fault-in approach does, or that we simply always fault
pages
> in.
As far as I can see, the two approach are now identical. Am I
missing
something?
And yes, getting rid of the eager mapping works for me.
As far as I can tell, our user_mem_abort() uses gfn_to_pfn_prot()
which
goes doesn a long trail which ends up at hva_to_pfn_remapped(), which
might result in doing the same offset calculation that we do in
kvm_arch_prepare_memory_region(), but it also considers other
scenarios.
Even if we analyze all that and convince oursleves it's always all
the
same on arm64, the two code paths could change, leading to really
hard
to debug differing behavior, and nobody will actively keep the two
paths
in sync. I'd be fine with keeping the performance optimization if we
have good grounds for that though, and using the same translation
mechanism for VM_PFNMAP as user_mem_abort.
Am I missing something?
I'm not disputing any of the above. I'm only trying to keep this patch
minimal so that we can easily backport it (although it is arguable that
deleting code isn't that big a deal).
[...]
> I can't seem to decide for myself if I think there's a sematic
> difference between trying to execute from somewhere the VMM has
> explicitly told us is device memory and from somewhere which we
happen
> to have mapped with VM_PFNMAP from user space. But I also can't
seem to
> really fault it (pun intended). Thoughts?
The issue is that the VMM never really tells us whether something is
a
device mapping or not (the only exception being the GICv2 cpuif).
Even
with PFNMAP, we guess it (it could well be memory that lives outside
of the linear mapping). I don't see a way to lift this ambiguity.
Ideally, faulting on executing a non-mapping should be offloaded to
userspace for emulation, in line with your patches that offload
non-emulated data accesses. That'd be a new ABI, and I can't imagine
anyone willing to deal with it.
So what I was asking was if it makes sense to report the Prefetch
Abort
in the case where the VMM has already told us that it doesn't want to
register anything backing the IPA (no memslot), and instead return an
error to user space, so that it can make a decision (for example
inject
an external abort, which may have been the right thing to do in the
former case as well, but that could be considered ABI now, so let's
not
kick that hornet's nest).
In any case, no strong feelings here, I just have a vague feeling
that
injecting more prefetch aborts on execute-from-some-device is not
necessarily the right thing to do.
The ARMv8 ARM has the following stuff in B2.7.2 (Device Memory):
<quote>
Hardware does not prevent speculative instruction fetches from a memory
location with any of the Device
memory attributes unless the memory location is also marked as
Execute-never for all Exception levels.
Note
This means that to prevent speculative instruction fetches from memory
locations with Device memory
attributes, any location that is assigned any Device memory type must
also be marked as Execute-never for
all Exception levels. Failure to mark a memory location with any Device
memory attribute as Execute-never
for all Exception levels is a programming error.
</quote>
and
<quote>
For instruction fetches, if branches cause the program counter to point
to an area of memory with the Device
attribute which is not marked as Execute-never for the current
Exception level, an implementation can either:
- Treat the instruction fetch as if it were to a memory location with
the Normal Non-cacheable attribute.
- Take a Permission fault.
</quote>
My reading here is that a prefetch abort is the right thing to do.
What we don't do correctly is that we qualify it as an external abort
instead of a permission fault (which is annoying as it requires us
to find out about the S1 translation level).
Happy to revisit this once we get a S1 PTW.
M.
--
Jazz is not dead. It just smells funny...
