On Fri, Jul 16, 2021, Brijesh Singh wrote:
>
> On 7/16/21 3:01 PM, Sean Christopherson wrote:
> > I'm having a bit of deja vu... This flow needs to hold kvm->srcu to do a memslot
> > lookup.
> >
> > That said, IMO having KVM do the hva->gpa is not a great ABI. The memslots are
> > completely arbitrary (from a certain point of view) and have no impact on the
> > validity of the memory pinning or PSP command. E.g. a memslot update while this
> > code is in-flight would be all kinds of weird.
> >
> > In other words, make userspace provide both the hva (because it's sadly needed
> > to pin memory) as well as the target gpa. That prevents KVM from having to deal
> > with memslot lookups and also means that userspace can issue the command before
> > configuring the memslots (though I've no idea if that's actually feasible for
> > any userspace VMM).
>
> The operation happen during the guest creation time so I was not sure if
> memslot will be updated while we are executing this command. But I guess
> its possible that a VMM may run different thread which may update
> memslot while another thread calls the encryption. I'll let userspace
> provide both the HVA and GPA as you recommended.
I'm not worried about a well-behaved userspace VMM, I'm worried about the code
KVM has to carry to guard against a misbehaving VMM.
> >> + ret = -EINVAL;
> >> + goto e_unpin;
> >> + }
> >> +
> >> + psize = page_level_size(level);
> >> + pmask = page_level_mask(level);
> > Is there any hope of this path supporting 2mb/1gb pages in the not-too-distant
> > future? If not, then I vote to do away with the indirection and just hardcode
> > 4kg sizes in the flow. I.e. if this works on 4kb chunks, make that obvious.
>
> No plans to do 1g/2mb in this path. I will make that obvious by
> hardcoding it.
>
>
> >> + gpa = gpa & pmask;
> >> +
> >> + /* Transition the page state to pre-guest */
> >> + memset(&e, 0, sizeof(e));
> >> + e.assigned = 1;
> >> + e.gpa = gpa;
> >> + e.asid = sev_get_asid(kvm);
> >> + e.immutable = true;
> >> + e.pagesize = X86_TO_RMP_PG_LEVEL(level);
> >> + ret = rmpupdate(inpages[i], &e);
> > What happens if userspace pulls a stupid and assigns the same page to multiple
> > SNP guests? Does RMPUPDATE fail? Can one RMPUPDATE overwrite another?
>
> The RMPUPDATE is available to the hv and it can call anytime with
> whatever it want. The important thing is the RMPUPDATE + PVALIDATE
> combination is what locks the page. In this case, PSP firmware updates
> the RMP table and also validates the page.
>
> If someone else attempts to issue another RMPUPDATE then Validated bit
> will be cleared and page is no longer used as a private. Access to
> unvalidated page will cause #VC.
Hmm, and there's no indication on success that the previous entry was assigned?
Adding a tracepoint in rmpupdate() to allow tracking transitions is probably a
good idea, otherwise debugging RMP violations and/or unexpected #VC is going to
be painful.
And/or if the kernel/KVM behavior is to never reassign directly and reading an RMP
entry isn't prohibitively expensive, then we could add a sanity check that the RMP
is unassigned and reject rmpupdate() if the page is already assigned. Probably
not worth it if the overhead is noticeable, but it could be nice to have if things
go sideways.
> >> +e_unpin:
> >> + /* Content of memory is updated, mark pages dirty */
> >> + memset(&e, 0, sizeof(e));
> >> + for (i = 0; i < npages; i++) {
> >> + set_page_dirty_lock(inpages[i]);
> >> + mark_page_accessed(inpages[i]);
> >> +
> >> + /*
> >> + * If its an error, then update RMP entry to change page ownership
> >> + * to the hypervisor.
> >> + */
> >> + if (ret)
> >> + rmpupdate(inpages[i], &e);
> > This feels wrong since it's purging _all_ RMP entries, not just those that were
> > successfully modified. And maybe add a RMP "reset" helper, e.g. why is zeroing
> > the RMP entry the correct behavior?
>
> By default all the pages are hypervior owned (i.e zero). If the
> LAUNCH_UPDATE was successful then page should have transition from the
> hypervisor owned to guest valid. By zero'ing it are reverting it back to
> hypevisor owned.
>
> I agree that I optimize it to clear the modified entries only and leave
> everything else as a default.
To be clear, it's not just an optimization. Pages that haven't yet been touched
may be already owned by a different VM (or even this VM). I.e. "reverting" those
pages would actually result in a form of corruption. It's somewhat of a moot point
because assigning a single page to multiple guests is going to be fatal anyways,
but potentially making a bug worse by introducing even more noise/confusion is not
good.
On 7/19/21 3:51 PM, Sean Christopherson wrote:
>
> Hmm, and there's no indication on success that the previous entry was assigned?
> Adding a tracepoint in rmpupdate() to allow tracking transitions is probably a
> good idea, otherwise debugging RMP violations and/or unexpected #VC is going to
> be painful.
>
Absolutely agree. It's in my TODO list for v5. I have been using my
private debug patches with all those trace debug and will try to pull
some of those in v5.
> And/or if the kernel/KVM behavior is to never reassign directly and reading an RMP
> entry isn't prohibitively expensive, then we could add a sanity check that the RMP
> is unassigned and reject rmpupdate() if the page is already assigned. Probably
> not worth it if the overhead is noticeable, but it could be nice to have if things
> go sideways.
>
In later patches you see that during the page-state change, I do try to
read RMP entry to detect some of these condition and warn user about
them. The GHCB specification lets the hypervisor choose how it wants to
handle the case in guest wanting to add the previously validated page.
>
> To be clear, it's not just an optimization. Pages that haven't yet been touched
> may be already owned by a different VM (or even this VM). I.e. "reverting" those
> pages would actually result in a form of corruption. It's somewhat of a moot point
> because assigning a single page to multiple guests is going to be fatal anyways,
> but potentially making a bug worse by introducing even more noise/confusion is not
> good.
>
As you said, if a process is assigning the same page to multiple VMs
then its fatal but I agree that we should do the right thing from the
kernel ioctl handling. I will just clear the RMP entry for the pages
which we touched.
thanks