On Thu, Jan 19, 2023, Huang, Kai wrote:
> On Thu, 2023-01-19 at 15:37 +0000, Sean Christopherson wrote:
> > On Thu, Jan 19, 2023, Huang, Kai wrote:
> > > On Tue, 2023-01-17 at 21:01 +0000, Sean Christopherson wrote:
> > > > On Tue, Jan 17, 2023, Sean Christopherson wrote:
> > > > > On Tue, Jan 17, 2023, Zhi Wang wrote:
> > > > Oh, the other important piece I forgot to mention is that dropping mmu_lock deep
> > > > in KVM's MMU in order to wait isn't always an option. Most flows would play nice
> > > > with dropping mmu_lock and sleeping, but some paths, e.g. from the mmu_notifier,
> > > > (conditionally) disallow sleeping.
> > >
> > > Could we do something similar to tdp_mmu_iter_cond_resched() but not simple busy
> > > retrying "X times", at least at those paths that can release mmu_lock()?
> >
> > That's effectively what happens by unwinding up the stak with an error code.
> > Eventually the page fault handler will get the error and retry the guest.
> >
> > > Basically we treat TDX_OPERAND_BUSY as seamcall_needbreak(), similar to
> > > rwlock_needbreak(). I haven't thought about details though.
> >
> > I am strongly opposed to that approach. I do not want to pollute KVM's MMU code
> > with a bunch of retry logic and error handling just because the TDX module is
> > ultra paranoid and hostile to hypervisors.
>
> Right. But IIUC there's legal cases that SEPT SEAMCALL can return BUSY due to
> multiple threads trying to read/modify SEPT simultaneously in case of TDP MMU.
> For instance, parallel page faults on different vcpus on private pages. I
> believe this is the main reason to retry.
Um, crud. I think there's a bigger issue. KVM always operates on its copy of the
S-EPT tables and assumes the the real S-EPT tables will always be synchronized with
KVM's mirror. That assumption doesn't hold true without serializing SEAMCALLs in
some way. E.g. if a SPTE is zapped and mapped at the same time, we can end up with:
vCPU0 vCPU1
===== =====
mirror[x] = xyz
old_spte = mirror[x]
mirror[x] = REMOVED_SPTE
sept[x] = REMOVED_SPTE
sept[x] = xyz
In other words, when mmu_lock is held for read, KVM relies on atomic SPTE updates.
With the mirror=>s-ept scheme, updates are no longer atomic and everything falls
apart.
Gracefully retrying only papers over the visible failures, the really problematic
scenarios are where multiple updates race and _don't_ trigger conflicts in the TDX
module.
> We previously used spinlock around the SEAMCALLs to avoid, but looks that is
> not preferred.
That doesn't address the race above either. And even if it did, serializing all
S-EPT SEAMCALLs for a VM is not an option, at least not in the long term.
The least invasive idea I have is expand the TDP MMU's concept of "frozen" SPTEs
and freeze (a.k.a. lock) the SPTE (KVM's mirror) until the corresponding S-EPT
update completes.
The other idea is to scrap the mirror concept entirely, though I gotta imagine
that would provide pretty awful performance.
diff --git a/arch/x86/kvm/mmu/spte.h b/arch/x86/kvm/mmu/spte.h
index 0d8deefee66c..bcb398e71475 100644
--- a/arch/x86/kvm/mmu/spte.h
+++ b/arch/x86/kvm/mmu/spte.h
@@ -198,9 +198,9 @@ extern u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
/* Removed SPTEs must not be misconstrued as shadow present PTEs. */
static_assert(!(REMOVED_SPTE & SPTE_MMU_PRESENT_MASK));
-static inline bool is_removed_spte(u64 spte)
+static inline bool is_frozen_spte(u64 spte)
{
- return spte == REMOVED_SPTE;
+ return spte == REMOVED_SPTE || spte & FROZEN_SPTE;
}
/* Get an SPTE's index into its parent's page table (and the spt array). */
diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
index bba33aea0fb0..7f34eccadf98 100644
--- a/arch/x86/kvm/mmu/tdp_mmu.c
+++ b/arch/x86/kvm/mmu/tdp_mmu.c
@@ -651,6 +651,9 @@ static inline int tdp_mmu_set_spte_atomic(struct kvm *kvm,
lockdep_assert_held_read(&kvm->mmu_lock);
+ if (<is TDX> && new_spte != REMOVED_SPTE)
+ new_spte |= FROZEN_SPTE;
+
/*
* Note, fast_pf_fix_direct_spte() can also modify TDP MMU SPTEs and
* does not hold the mmu_lock.
@@ -662,6 +665,9 @@ static inline int tdp_mmu_set_spte_atomic(struct kvm *kvm,
new_spte, iter->level, true);
handle_changed_spte_acc_track(iter->old_spte, new_spte, iter->level);
+ if (<is TDX> && new_spte != REMOVED_SPTE)
+ __kvm_tdp_mmu_write_spte(iter->sptep, new_spte);
+
return 0;
}
