Hi,
On Wed, Feb 23, 2022 at 09:56:01PM -0800, Reiji Watanabe wrote:
> Hi Alex,
>
> On Wed, Nov 17, 2021 at 7:37 AM Alexandru Elisei
> <alexandru.elisei@xxxxxxx> wrote:
> >
> > KVM relies on doing dcache maintenance on stage 2 faults to present to a
> > guest running with the MMU off the same view of memory as userspace. For
> > locked memslots, KVM so far has done the dcache maintenance when a memslot
> > is locked, but that leaves KVM in a rather awkward position: what userspace
> > writes to guest memory after the memslot is locked, but before a VCPU is
> > run, might not be visible to the guest.
> >
> > Fix this by deferring the dcache maintenance until the first VCPU is run.
> >
> > Signed-off-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
> > ---
> > arch/arm64/include/asm/kvm_host.h | 7 ++++
> > arch/arm64/include/asm/kvm_mmu.h | 5 +++
> > arch/arm64/kvm/arm.c | 3 ++
> > arch/arm64/kvm/mmu.c | 55 ++++++++++++++++++++++++++++---
> > 4 files changed, 66 insertions(+), 4 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> > index 7fd70ad90c16..3b4839b447c4 100644
> > --- a/arch/arm64/include/asm/kvm_host.h
> > +++ b/arch/arm64/include/asm/kvm_host.h
> > @@ -113,6 +113,10 @@ struct kvm_arch_memory_slot {
> > u32 flags;
> > };
> >
> > +/* kvm->arch.mmu_pending_ops flags */
> > +#define KVM_LOCKED_MEMSLOT_FLUSH_DCACHE 0
> > +#define KVM_MAX_MMU_PENDING_OPS 1
> > +
> > struct kvm_arch {
> > struct kvm_s2_mmu mmu;
> >
> > @@ -136,6 +140,9 @@ struct kvm_arch {
> > */
> > bool return_nisv_io_abort_to_user;
> >
> > + /* Defer MMU operations until a VCPU is run. */
> > + unsigned long mmu_pending_ops;
> > +
> > /*
> > * VM-wide PMU filter, implemented as a bitmap and big enough for
> > * up to 2^10 events (ARMv8.0) or 2^16 events (ARMv8.1+).
> > diff --git a/arch/arm64/include/asm/kvm_mmu.h b/arch/arm64/include/asm/kvm_mmu.h
> > index 2c50734f048d..cbf57c474fea 100644
> > --- a/arch/arm64/include/asm/kvm_mmu.h
> > +++ b/arch/arm64/include/asm/kvm_mmu.h
> > @@ -219,6 +219,11 @@ void kvm_toggle_cache(struct kvm_vcpu *vcpu, bool was_enabled);
> > int kvm_mmu_lock_memslot(struct kvm *kvm, u64 slot, u64 flags);
> > int kvm_mmu_unlock_memslot(struct kvm *kvm, u64 slot, u64 flags);
> >
> > +#define kvm_mmu_has_pending_ops(kvm) \
> > + (!bitmap_empty(&(kvm)->arch.mmu_pending_ops, KVM_MAX_MMU_PENDING_OPS))
> > +
> > +void kvm_mmu_perform_pending_ops(struct kvm *kvm);
> > +
> > static inline unsigned int kvm_get_vmid_bits(void)
> > {
> > int reg = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
> > diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> > index b9b8b43835e3..96ed48455cdd 100644
> > --- a/arch/arm64/kvm/arm.c
> > +++ b/arch/arm64/kvm/arm.c
> > @@ -870,6 +870,9 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
> > if (unlikely(!kvm_vcpu_initialized(vcpu)))
> > return -ENOEXEC;
> >
> > + if (unlikely(kvm_mmu_has_pending_ops(vcpu->kvm)))
> > + kvm_mmu_perform_pending_ops(vcpu->kvm);
> > +
> > ret = kvm_vcpu_first_run_init(vcpu);
> > if (ret)
> > return ret;
> > diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> > index b0a8e61315e4..8e4787019840 100644
> > --- a/arch/arm64/kvm/mmu.c
> > +++ b/arch/arm64/kvm/mmu.c
> > @@ -1305,6 +1305,40 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
> > return ret;
> > }
> >
> > +/*
> > + * It's safe to do the CMOs when the first VCPU is run because:
> > + * - VCPUs cannot run until mmu_cmo_needed is cleared.
>
> What does 'mmu_cmo_needed' mean ? Do you mean 'mmu_pending_ops' instead ?
Yes, I meant mmu_pending_ops here. I used mmu_cmo_needed for the field name
as I was working on it and I forgot to change it. Will fix it.
>
>
> > + * - Memslots cannot be modified because we hold the kvm->slots_lock.
> > + *
> > + * It's safe to periodically release the mmu_lock because:
> > + * - VCPUs cannot run.
> > + * - Any changes to the stage 2 tables triggered by the MMU notifiers also take
> > + * the mmu_lock, which means accesses will be serialized.
> > + * - Stage 2 tables cannot be freed from under us as long as at least one VCPU
> > + * is live, which means that the VM will be live.
> > + */
> > +void kvm_mmu_perform_pending_ops(struct kvm *kvm)
> > +{
> > + struct kvm_memory_slot *memslot;
> > +
> > + mutex_lock(&kvm->slots_lock);
> > + if (!kvm_mmu_has_pending_ops(kvm))
> > + goto out_unlock;
> > +
> > + if (test_bit(KVM_LOCKED_MEMSLOT_FLUSH_DCACHE, &kvm->arch.mmu_pending_ops)) {
> > + kvm_for_each_memslot(memslot, kvm_memslots(kvm)) {
> > + if (!memslot_is_locked(memslot))
> > + continue;
>
> Shouldn't the code hold the mmu_lock to call stage2_flush_memslot() ?
There will be no contention between different VCPUs because the stage 2
translation tables are protected against concurrent accesses with the
kvm->slots_lock mutex above. But stage2_flush_memslot() expects the
mmu_lock to be held and it will be periodically released by
cond_resched_lock() in stage2_apply_range(); if the lock is not held, then
lockdep will complain about it.
Your observation actually explains why I was seeing intermitent warnings
when lockdep was enabled: __cond_resched_lock was complaining the KVM was
trying to release a lock it wasn't holding. Thank you for pointing the
missing lock acquire operation.
I'll change the code to avoid the lockdep warning.
>
> > + stage2_flush_memslot(kvm, memslot);
>
> Since stage2_flush_memslot() won't do anything when stage2_has_fwb()
> returns true, I wonder if it can be checked even before iterating
> memslots (so those iterations can be skipped when not needed).
I think this can be further improved by setting the
KVM_LOCKED_MEMSLOT_FLUSH_DCACHE bit only if FWB is not present.
Thanks,
Alex
>
> Thanks,
> Reiji
>
> > + }
> > + clear_bit(KVM_LOCKED_MEMSLOT_FLUSH_DCACHE, &kvm->arch.mmu_pending_ops);
> > + }
> > +
> > +out_unlock:
> > + mutex_unlock(&kvm->slots_lock);
> > + return;
> > +}
> > +
> > static int try_rlimit_memlock(unsigned long npages)
> > {
> > unsigned long lock_limit;
> > @@ -1345,7 +1379,8 @@ static int lock_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot,
> > struct kvm_memory_slot_page *page_entry;
> > bool writable = flags & KVM_ARM_LOCK_MEM_WRITE;
> > enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R;
> > - struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
> > + struct kvm_pgtable pgt;
> > + struct kvm_pgtable_mm_ops mm_ops;
> > struct vm_area_struct *vma;
> > unsigned long npages = memslot->npages;
> > unsigned int pin_flags = FOLL_LONGTERM;
> > @@ -1363,6 +1398,16 @@ static int lock_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot,
> > pin_flags |= FOLL_WRITE;
> > }
> >
> > + /*
> > + * Make a copy of the stage 2 translation table struct to remove the
> > + * dcache callback so we can postpone the cache maintenance operations
> > + * until the first VCPU is run.
> > + */
> > + mm_ops = *kvm->arch.mmu.pgt->mm_ops;
> > + mm_ops.dcache_clean_inval_poc = NULL;
> > + pgt = *kvm->arch.mmu.pgt;
> > + pgt.mm_ops = &mm_ops;
> > +
> > hva = memslot->userspace_addr;
> > ipa = memslot->base_gfn << PAGE_SHIFT;
> >
> > @@ -1414,13 +1459,13 @@ static int lock_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot,
> > goto out_err;
> > }
> >
> > - ret = kvm_pgtable_stage2_map(pgt, ipa, PAGE_SIZE,
> > + ret = kvm_pgtable_stage2_map(&pgt, ipa, PAGE_SIZE,
> > page_to_phys(page_entry->page),
> > prot, &cache);
> > spin_unlock(&kvm->mmu_lock);
> >
> > if (ret) {
> > - kvm_pgtable_stage2_unmap(pgt, memslot->base_gfn << PAGE_SHIFT,
> > + kvm_pgtable_stage2_unmap(&pgt, memslot->base_gfn << PAGE_SHIFT,
> > i << PAGE_SHIFT);
> > unpin_memslot_pages(memslot, writable);
> > goto out_err;
> > @@ -1439,7 +1484,7 @@ static int lock_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot,
> > */
> > ret = account_locked_vm(current->mm, npages, true);
> > if (ret) {
> > - kvm_pgtable_stage2_unmap(pgt, memslot->base_gfn << PAGE_SHIFT,
> > + kvm_pgtable_stage2_unmap(&pgt, memslot->base_gfn << PAGE_SHIFT,
> > npages << PAGE_SHIFT);
> > unpin_memslot_pages(memslot, writable);
> > goto out_err;
> > @@ -1449,6 +1494,8 @@ static int lock_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot,
> > if (writable)
> > memslot->arch.flags |= KVM_MEMSLOT_LOCK_WRITE;
> >
> > + set_bit(KVM_LOCKED_MEMSLOT_FLUSH_DCACHE, &kvm->arch.mmu_pending_ops);
> > +
> > kvm_mmu_free_memory_cache(&cache);
> >
> > return 0;
> > --
> > 2.33.1
> >
> > _______________________________________________
> > kvmarm mailing list
> > kvmarm@xxxxxxxxxxxxxxxxxxxxx
> > https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
