On Tue, Jan 12, 2021, Ben Gardon wrote:
> In order to enable concurrent modifications to the paging structures in
> the TDP MMU, threads must be able to safely remove pages of page table
> memory while other threads are traversing the same memory. To ensure
> threads do not access PT memory after it is freed, protect PT memory
> with RCU.
Normally I like splitting up patches, but the three RCU patches (11-13) probably
need to be combined into a single patch. I assume you introduced the RCU
readers in a separate patch to isolate deadlocks, but it's impossible to give
this patch a proper review without peeking ahead to see how what's actually
being protected with RCU.
The combined changelog should also explain why READING_SHADOW_PAGE_TABLES isn't
a good solution. I suspect the answer is because the longer-running walks would
disable IRQs for too long, but that should be explicitly documented.
> Reviewed-by: Peter Feiner <pfei...@google.com>
>
> Signed-off-by: Ben Gardon <bgar...@google.com>
> ---
> arch/x86/kvm/mmu/tdp_mmu.c | 53 ++++++++++++++++++++++++++++++++++++--
> 1 file changed, 51 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index e8f35cd46b4c..662907d374b3 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -458,11 +458,14 @@ static inline void tdp_mmu_set_spte_no_dirty_log(struct
> kvm *kvm,
> * Return true if this function yielded, the TLBs were flushed, and the
> * iterator's traversal was reset. Return false if a yield was not needed.
> */
> -static bool tdp_mmu_iter_flush_cond_resched(struct kvm *kvm, struct tdp_iter
> *iter)
> +static bool tdp_mmu_iter_flush_cond_resched(struct kvm *kvm,
> + struct tdp_iter *iter)
Unrelated newline.
> {
> if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
> kvm_flush_remote_tlbs(kvm);
> + rcu_read_unlock();
I'm 99% certain rcu_read_unlock() can be moved before the TLB flush. IIUC, RCU
is protecting only the host kernel's software walks; the only true "writer" is
immediately preceded by a remote TLB flush (in patch 13).
kvm_flush_remote_tlbs_with_address(kvm, gfn,
KVM_PAGES_PER_HPAGE(level));
call_rcu(&sp->rcu_head, tdp_mmu_free_sp_rcu_callback);
That also resolves an inconsistency with zap_gfn_range(), which unlocks before
doing the remote flush. Ditto for zap_collapsible_spte_range(), and I think a
few other flows.
> cond_resched_lock(&kvm->mmu_lock);
> + rcu_read_lock();
> tdp_iter_refresh_walk(iter);
> return true;
> } else
> @@ -483,7 +486,9 @@ static bool tdp_mmu_iter_flush_cond_resched(struct kvm
> *kvm, struct tdp_iter *it
> static bool tdp_mmu_iter_cond_resched(struct kvm *kvm, struct tdp_iter *iter)
> {
> if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
> + rcu_read_unlock();
> cond_resched_lock(&kvm->mmu_lock);
> + rcu_read_lock();
> tdp_iter_refresh_walk(iter);
> return true;
> } else
> @@ -508,6 +513,8 @@ static bool zap_gfn_range(struct kvm *kvm, struct
> kvm_mmu_page *root,
> gfn_t last_goal_gfn = start;
> bool flush_needed = false;
>
> + rcu_read_lock();
> +
> tdp_root_for_each_pte(iter, root, start, end) {
> /* Ensure forward progress has been made before yielding. */
> if (can_yield && iter.goal_gfn != last_goal_gfn &&
> @@ -538,6 +545,8 @@ static bool zap_gfn_range(struct kvm *kvm, struct
> kvm_mmu_page *root,
> tdp_mmu_set_spte(kvm, &iter, 0);
> flush_needed = true;
> }
> +
> + rcu_read_unlock();
Unlock before TLB flush. <-------
> return flush_needed;
> }
...
> @@ -844,6 +863,8 @@ static int set_tdp_spte(struct kvm *kvm, struct
> kvm_memory_slot *slot,
> u64 new_spte;
> int need_flush = 0;
>
> + rcu_read_lock();
> +
> WARN_ON(pte_huge(*ptep));
>
> new_pfn = pte_pfn(*ptep);
> @@ -872,6 +893,8 @@ static int set_tdp_spte(struct kvm *kvm, struct
> kvm_memory_slot *slot,
> if (need_flush)
> kvm_flush_remote_tlbs_with_address(kvm, gfn, 1);
>
> + rcu_read_unlock();
Unlock before flush?
> +
> return 0;
> }
>
...
> @@ -1277,10 +1322,14 @@ int kvm_tdp_mmu_get_walk(struct kvm_vcpu *vcpu, u64
> addr, u64 *sptes,
>
> *root_level = vcpu->arch.mmu->shadow_root_level;
>
> + rcu_read_lock();
Hrm, isn't this an existing bug? And also not really the correct fix? mmu_lock
is not held here, so the existing code has no protections. Using
walk_shadow_page_lockless_begin/end() feels more appropriate for this particular
walk.
> +
> tdp_mmu_for_each_pte(iter, mmu, gfn, gfn + 1) {
> leaf = iter.level;
> sptes[leaf] = iter.old_spte;
> }
>
> + rcu_read_unlock();
> +
> return leaf;
> }
> --
> 2.30.0.284.gd98b1dd5eaa7-goog
>
