lkml.org 
[lkml]   [2021]   [Dec]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
Subject[PATCH] KVM: x86/mmu: Don't advance iterator after restart due to yielding
From
After dropping mmu_lock in the TDP MMU, restart the iterator during
tdp_iter_next() and do not advance the iterator. Advancing the iterator
results in skipping the top-level SPTE and all its children, which is
fatal if any of the skipped SPTEs were not visited before yielding.

When zapping all SPTEs, i.e. when min_level == root_level, restarting the
iter and then invoking tdp_iter_next() is always fatal if the current gfn
has as a valid SPTE, as advancing the iterator results in try_step_side()
skipping the current gfn, which wasn't visited before yielding.

Sprinkle WARNs on iter->yielded being true in various helpers that are
often used in conjunction with yielding, and tag the helper with
__must_check to reduce the probabily of improper usage.

Failing to zap a top-level SPTE manifests in one of two ways. If a valid
SPTE is skipped by both kvm_tdp_mmu_zap_all() and kvm_tdp_mmu_put_root(),
the shadow page will be leaked and KVM will WARN accordingly.

WARNING: CPU: 1 PID: 3509 at arch/x86/kvm/mmu/tdp_mmu.c:46 [kvm]
RIP: 0010:kvm_mmu_uninit_tdp_mmu+0x3e/0x50 [kvm]
Call Trace:
<TASK>
kvm_arch_destroy_vm+0x130/0x1b0 [kvm]
kvm_destroy_vm+0x162/0x2a0 [kvm]
kvm_vcpu_release+0x34/0x60 [kvm]
__fput+0x82/0x240
task_work_run+0x5c/0x90
do_exit+0x364/0xa10
? futex_unqueue+0x38/0x60
do_group_exit+0x33/0xa0
get_signal+0x155/0x850
arch_do_signal_or_restart+0xed/0x750
exit_to_user_mode_prepare+0xc5/0x120
syscall_exit_to_user_mode+0x1d/0x40
do_syscall_64+0x48/0xc0
entry_SYSCALL_64_after_hwframe+0x44/0xae

If kvm_tdp_mmu_zap_all() skips a gfn/SPTE but that SPTE is then zapped by
kvm_tdp_mmu_put_root(), KVM triggers a use-after-free in the form of
marking a struct page as dirty/accessed after it has been put back on the
free list. This directly triggers a WARN due to encountering a page with
page_count() == 0, but it can also lead to data corruption and additional
errors in the kernel.

WARNING: CPU: 7 PID: 1995658 at arch/x86/kvm/../../../virt/kvm/kvm_main.c:171
RIP: 0010:kvm_is_zone_device_pfn.part.0+0x9e/0xd0 [kvm]
Call Trace:
<TASK>
kvm_set_pfn_dirty+0x120/0x1d0 [kvm]
__handle_changed_spte+0x92e/0xca0 [kvm]
__handle_changed_spte+0x63c/0xca0 [kvm]
__handle_changed_spte+0x63c/0xca0 [kvm]
__handle_changed_spte+0x63c/0xca0 [kvm]
zap_gfn_range+0x549/0x620 [kvm]
kvm_tdp_mmu_put_root+0x1b6/0x270 [kvm]
mmu_free_root_page+0x219/0x2c0 [kvm]
kvm_mmu_free_roots+0x1b4/0x4e0 [kvm]
kvm_mmu_unload+0x1c/0xa0 [kvm]
kvm_arch_destroy_vm+0x1f2/0x5c0 [kvm]
kvm_put_kvm+0x3b1/0x8b0 [kvm]
kvm_vcpu_release+0x4e/0x70 [kvm]
__fput+0x1f7/0x8c0
task_work_run+0xf8/0x1a0
do_exit+0x97b/0x2230
do_group_exit+0xda/0x2a0
get_signal+0x3be/0x1e50
arch_do_signal_or_restart+0x244/0x17f0
exit_to_user_mode_prepare+0xcb/0x120
syscall_exit_to_user_mode+0x1d/0x40
do_syscall_64+0x4d/0x90
entry_SYSCALL_64_after_hwframe+0x44/0xae

Note, the underlying bug existed even before commit 1af4a96025b3 ("KVM:
x86/mmu: Yield in TDU MMU iter even if no SPTES changed") moved calls to
tdp_mmu_iter_cond_resched() to the beginning of loops, as KVM could still
incorrectly advance past a top-level entry when yielding on a lower-level
entry. But with respect to leaking shadow pages, the bug was introduced
by yielding before processing the current gfn.

Alternatively, tdp_mmu_iter_cond_resched() could simply fall through, or
callers could jump to their "retry" label. The downside of that approach
is that tdp_mmu_iter_cond_resched() _must_ be called before anything else
in the loop, and there's no easy way to enfornce that requirement.

Ideally, KVM would handling the cond_resched() fully within the iterator
macro (the code is actually quite clean) and avoid this entire class of
bugs, but that is extremely difficult do while also supporting yielding
after tdp_mmu_set_spte_atomic() fails. Yielding after failing to set a
SPTE is very desirable as the "owner" of the REMOVED_SPTE isn't strictly
bounded, e.g. if it's zapping a high-level shadow page, the REMOVED_SPTE
may block operations on the SPTE for a significant amount of time.

Fixes: faaf05b00aec ("kvm: x86/mmu: Support zapping SPTEs in the TDP MMU")
Fixes: 1af4a96025b3 ("KVM: x86/mmu: Yield in TDU MMU iter even if no SPTES changed")
Reported-by: Ignat Korchagin <ignat@cloudflare.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
---
arch/x86/kvm/mmu/tdp_iter.c | 6 ++++++
arch/x86/kvm/mmu/tdp_iter.h | 6 ++++++
arch/x86/kvm/mmu/tdp_mmu.c | 29 ++++++++++++++++-------------
3 files changed, 28 insertions(+), 13 deletions(-)

diff --git a/arch/x86/kvm/mmu/tdp_iter.c b/arch/x86/kvm/mmu/tdp_iter.c
index b3ed302c1a35..caa96c270b95 100644
--- a/arch/x86/kvm/mmu/tdp_iter.c
+++ b/arch/x86/kvm/mmu/tdp_iter.c
@@ -26,6 +26,7 @@ static gfn_t round_gfn_for_level(gfn_t gfn, int level)
*/
void tdp_iter_restart(struct tdp_iter *iter)
{
+ iter->yielded = false;
iter->yielded_gfn = iter->next_last_level_gfn;
iter->level = iter->root_level;

@@ -160,6 +161,11 @@ static bool try_step_up(struct tdp_iter *iter)
*/
void tdp_iter_next(struct tdp_iter *iter)
{
+ if (iter->yielded) {
+ tdp_iter_restart(iter);
+ return;
+ }
+
if (try_step_down(iter))
return;

diff --git a/arch/x86/kvm/mmu/tdp_iter.h b/arch/x86/kvm/mmu/tdp_iter.h
index b1748b988d3a..e19cabbcb65c 100644
--- a/arch/x86/kvm/mmu/tdp_iter.h
+++ b/arch/x86/kvm/mmu/tdp_iter.h
@@ -45,6 +45,12 @@ struct tdp_iter {
* iterator walks off the end of the paging structure.
*/
bool valid;
+ /*
+ * True if KVM dropped mmu_lock and yielded in the middle of a walk, in
+ * which case tdp_iter_next() needs to restart the walk at the root
+ * level instead of advancing to the next entry.
+ */
+ bool yielded;
};

/*
diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
index b69e47e68307..7b1bc816b7c3 100644
--- a/arch/x86/kvm/mmu/tdp_mmu.c
+++ b/arch/x86/kvm/mmu/tdp_mmu.c
@@ -502,6 +502,8 @@ static inline bool tdp_mmu_set_spte_atomic(struct kvm *kvm,
struct tdp_iter *iter,
u64 new_spte)
{
+ WARN_ON_ONCE(iter->yielded);
+
lockdep_assert_held_read(&kvm->mmu_lock);

/*
@@ -575,6 +577,8 @@ static inline void __tdp_mmu_set_spte(struct kvm *kvm, struct tdp_iter *iter,
u64 new_spte, bool record_acc_track,
bool record_dirty_log)
{
+ WARN_ON_ONCE(iter->yielded);
+
lockdep_assert_held_write(&kvm->mmu_lock);

/*
@@ -640,18 +644,19 @@ static inline void tdp_mmu_set_spte_no_dirty_log(struct kvm *kvm,
* If this function should yield and flush is set, it will perform a remote
* TLB flush before yielding.
*
- * If this function yields, it will also reset the tdp_iter's walk over the
- * paging structure and the calling function should skip to the next
- * iteration to allow the iterator to continue its traversal from the
- * paging structure root.
+ * If this function yields, iter->yielded is set and the caller must skip to
+ * the next iteration, where tdp_iter_next() will reset the tdp_iter's walk
+ * over the paging structures to allow the iterator to continue its traversal
+ * from the paging structure root.
*
- * Return true if this function yielded and the iterator's traversal was reset.
- * Return false if a yield was not needed.
+ * Returns true if this function yielded.
*/
-static inline bool tdp_mmu_iter_cond_resched(struct kvm *kvm,
- struct tdp_iter *iter, bool flush,
- bool shared)
+static inline bool __must_check tdp_mmu_iter_cond_resched(struct kvm *kvm,
+ struct tdp_iter *iter,
+ bool flush, bool shared)
{
+ WARN_ON(iter->yielded);
+
/* Ensure forward progress has been made before yielding. */
if (iter->next_last_level_gfn == iter->yielded_gfn)
return false;
@@ -671,12 +676,10 @@ static inline bool tdp_mmu_iter_cond_resched(struct kvm *kvm,

WARN_ON(iter->gfn > iter->next_last_level_gfn);

- tdp_iter_restart(iter);
-
- return true;
+ iter->yielded = true;
}

- return false;
+ return iter->yielded;
}

/*
--
2.34.1.173.g76aa8bc2d0-goog
\
 
 \ /
  Last update: 2021-12-14 04:36    [W:0.080 / U:0.252 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site