From: Nadav Amit <[email protected]>
The text poke mechanism claims to have an RCU-like behavior, but it does
not appear that there is any quiescent state to ensure that nobody holds
reference to desc. As a result, the following race appears to be
possible, which can lead to memory corruption.
CPU0 CPU1
---- ----
text_poke_bp_batch()
-> smp_store_release(&bp_desc, &desc)
[ notice that desc is on
the stack ]
poke_int3_handler()
[ int3 might be kprobe's
so sync events are do not
help ]
-> try_get_desc(descp=&bp_desc)
desc = __READ_ONCE(bp_desc)
if (!desc) [false, success]
WRITE_ONCE(bp_desc, NULL);
atomic_dec_and_test(&desc.refs)
[ success, desc space on the stack
is being reused and might have
non-zero value. ]
arch_atomic_inc_not_zero(&desc->refs)
[ might succeed since desc points to
stack memory that was freed and might
be reused. ]
I encountered some occasional crashes of poke_int3_handler() when
kprobes are set, while accessing desc->vec. The analysis has been done
offline and I did not corroborate the cause of the crashes. Yet, it
seems that this race might be the root cause.
Fix this issue with small backportable patch. Instead of trying to make
RCU-like behavior for bp_desc, just eliminate the unnecessary level of
indirection of bp_desc, and hold the whole descriptor on the stack.
Anyhow, there is only a single descriptor at any given moment.
Fixes: 1f676247f36a4 ("x86/alternatives: Implement a better poke_int3_handler() completion scheme"
Cc: [email protected]
Cc: "Peter Zijlstra (Intel)" <[email protected]>
Cc: "Steven Rostedt (Google)" <[email protected]>
Signed-off-by: Nadav Amit <[email protected]>
---
arch/x86/kernel/alternative.c | 42 +++++++++++++++++------------------
1 file changed, 20 insertions(+), 22 deletions(-)
diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
index 62f6b8b7c4a5..4265c9426374 100644
--- a/arch/x86/kernel/alternative.c
+++ b/arch/x86/kernel/alternative.c
@@ -1319,18 +1319,15 @@ struct bp_patching_desc {
atomic_t refs;
};
-static struct bp_patching_desc *bp_desc;
+static struct bp_patching_desc bp_desc;
static __always_inline
-struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
+struct bp_patching_desc *try_get_desc(void)
{
- /* rcu_dereference */
- struct bp_patching_desc *desc = __READ_ONCE(*descp);
-
- if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
+ if (!arch_atomic_inc_not_zero(&bp_desc.refs))
return NULL;
- return desc;
+ return &bp_desc;
}
static __always_inline void put_desc(struct bp_patching_desc *desc)
@@ -1367,15 +1364,15 @@ noinstr int poke_int3_handler(struct pt_regs *regs)
/*
* Having observed our INT3 instruction, we now must observe
- * bp_desc:
+ * bp_desc with non-zero refcount:
*
- * bp_desc = desc INT3
+ * bp_desc.refs = 1 INT3
* WMB RMB
- * write INT3 if (desc)
+ * write INT3 if (bp_desc.refs != 0)
*/
smp_rmb();
- desc = try_get_desc(&bp_desc);
+ desc = try_get_desc();
if (!desc)
return 0;
@@ -1460,18 +1457,21 @@ static int tp_vec_nr;
*/
static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
{
- struct bp_patching_desc desc = {
- .vec = tp,
- .nr_entries = nr_entries,
- .refs = ATOMIC_INIT(1),
- };
unsigned char int3 = INT3_INSN_OPCODE;
unsigned int i;
int do_sync;
lockdep_assert_held(&text_mutex);
- smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
+ bp_desc.vec = tp;
+ bp_desc.nr_entries = nr_entries;
+
+ /*
+ * Corresponds to the implicit memory barrier in try_get_desc() to
+ * ensure reading a non-zero refcount provides up to date bp_desc data.
+ */
+ smp_wmb();
+ atomic_set(&bp_desc.refs, 1);
/*
* Corresponding read barrier in int3 notifier for making sure the
@@ -1559,12 +1559,10 @@ static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries
text_poke_sync();
/*
- * Remove and synchronize_rcu(), except we have a very primitive
- * refcount based completion.
+ * Remove and wait for refs to be zero.
*/
- WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
- if (!atomic_dec_and_test(&desc.refs))
- atomic_cond_read_acquire(&desc.refs, !VAL);
+ if (!atomic_dec_and_test(&bp_desc.refs))
+ atomic_cond_read_acquire(&bp_desc.refs, !VAL);
}
static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
--
2.34.1
On Sep 20, 2022, at 3:47 PM, Nadav Amit <[email protected]> wrote:
> From: Nadav Amit <[email protected]>
>
> The text poke mechanism claims to have an RCU-like behavior, but it does
> not appear that there is any quiescent state to ensure that nobody holds
> reference to desc. As a result, the following race appears to be
> possible, which can lead to memory corruption.
>
> CPU0 CPU1
> ---- ----
> text_poke_bp_batch()
> -> smp_store_release(&bp_desc, &desc)
>
> [ notice that desc is on
> the stack ]
>
> poke_int3_handler()
>
> [ int3 might be kprobe's
> so sync events are do not
> help ]
>
> -> try_get_desc(descp=&bp_desc)
> desc = __READ_ONCE(bp_desc)
>
> if (!desc) [false, success]
> WRITE_ONCE(bp_desc, NULL);
> atomic_dec_and_test(&desc.refs)
>
> [ success, desc space on the stack
> is being reused and might have
> non-zero value. ]
> arch_atomic_inc_not_zero(&desc->refs)
>
> [ might succeed since desc points to
> stack memory that was freed and might
> be reused. ]
>
> I encountered some occasional crashes of poke_int3_handler() when
> kprobes are set, while accessing desc->vec. The analysis has been done
> offline and I did not corroborate the cause of the crashes. Yet, it
> seems that this race might be the root cause.
>
> Fix this issue with small backportable patch. Instead of trying to make
> RCU-like behavior for bp_desc, just eliminate the unnecessary level of
> indirection of bp_desc, and hold the whole descriptor on the stack.
> Anyhow, there is only a single descriptor at any given moment.
>
> Fixes: 1f676247f36a4 ("x86/alternatives: Implement a better poke_int3_handler() completion scheme"
A bracket is mistakenly missing after the patch title. Sorry.
On Tue, Sep 20, 2022 at 10:47:43PM +0000, Nadav Amit wrote:
> Fix this issue with small backportable patch. Instead of trying to make
> RCU-like behavior for bp_desc, just eliminate the unnecessary level of
> indirection of bp_desc, and hold the whole descriptor on the stack.
> Anyhow, there is only a single descriptor at any given moment.
Because of text_mutex; indeed. No idea why I put that thing on the
stack.
I've done a few minor edits to your patch, but it otherwise looks good
to me.
---
Subject: x86/alternative: Fix race in try_get_desc()
From: Nadav Amit <[email protected]>
Date: Tue, 20 Sep 2022 22:47:43 +0000
From: Nadav Amit <[email protected]>
The text poke mechanism claims to have an RCU-like behavior, but it does
not appear that there is any quiescent state to ensure that nobody holds
reference to desc. As a result, the following race appears to be
possible, which can lead to memory corruption.
CPU0 CPU1
---- ----
text_poke_bp_batch()
-> smp_store_release(&bp_desc, &desc)
[ notice that desc is on
the stack ]
poke_int3_handler()
[ int3 might be kprobe's
so sync events are do not
help ]
-> try_get_desc(descp=&bp_desc)
desc = __READ_ONCE(bp_desc)
if (!desc) [false, success]
WRITE_ONCE(bp_desc, NULL);
atomic_dec_and_test(&desc.refs)
[ success, desc space on the stack
is being reused and might have
non-zero value. ]
arch_atomic_inc_not_zero(&desc->refs)
[ might succeed since desc points to
stack memory that was freed and might
be reused. ]
I encountered some occasional crashes of poke_int3_handler() when
kprobes are set, while accessing desc->vec. The analysis has been done
offline and I did not corroborate the cause of the crashes. Yet, it
seems that this race might be the root cause.
Fix this issue with small backportable patch. Instead of trying to make
RCU-like behavior for bp_desc, just eliminate the unnecessary level of
indirection of bp_desc, and hold the whole descriptor on the stack.
Anyhow, there is only a single descriptor at any given moment.
Fixes: 1f676247f36a4 ("x86/alternatives: Implement a better poke_int3_handler() completion scheme")
Signed-off-by: Nadav Amit <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Cc: [email protected]
Link: https://lkml.kernel.org/r/[email protected]
---
arch/x86/kernel/alternative.c | 45 +++++++++++++++++++++---------------------
1 file changed, 23 insertions(+), 22 deletions(-)
--- a/arch/x86/kernel/alternative.c
+++ b/arch/x86/kernel/alternative.c
@@ -1319,22 +1319,23 @@ struct bp_patching_desc {
atomic_t refs;
};
-static struct bp_patching_desc *bp_desc;
+static struct bp_patching_desc bp_desc;
static __always_inline
-struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
+struct bp_patching_desc *try_get_desc(void)
{
- /* rcu_dereference */
- struct bp_patching_desc *desc = __READ_ONCE(*descp);
+ struct bp_patching_desc *desc = &bp_desc;
- if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
+ if (!arch_atomic_inc_not_zero(&desc->refs))
return NULL;
return desc;
}
-static __always_inline void put_desc(struct bp_patching_desc *desc)
+static __always_inline void put_desc(void)
{
+ struct bp_patching_desc *desc = &bp_desc;
+
smp_mb__before_atomic();
arch_atomic_dec(&desc->refs);
}
@@ -1367,15 +1368,15 @@ noinstr int poke_int3_handler(struct pt_
/*
* Having observed our INT3 instruction, we now must observe
- * bp_desc:
+ * bp_desc with non-zero refcount:
*
- * bp_desc = desc INT3
+ * bp_desc.refs = 1 INT3
* WMB RMB
- * write INT3 if (desc)
+ * write INT3 if (bp_desc.refs != 0)
*/
smp_rmb();
- desc = try_get_desc(&bp_desc);
+ desc = try_get_desc();
if (!desc)
return 0;
@@ -1429,7 +1430,7 @@ noinstr int poke_int3_handler(struct pt_
ret = 1;
out_put:
- put_desc(desc);
+ put_desc();
return ret;
}
@@ -1460,18 +1461,20 @@ static int tp_vec_nr;
*/
static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
{
- struct bp_patching_desc desc = {
- .vec = tp,
- .nr_entries = nr_entries,
- .refs = ATOMIC_INIT(1),
- };
unsigned char int3 = INT3_INSN_OPCODE;
unsigned int i;
int do_sync;
lockdep_assert_held(&text_mutex);
- smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
+ bp_desc.vec = tp;
+ bp_desc.nr_entries = nr_entries;
+
+ /*
+ * Corresponds to the implicit memory barrier in try_get_desc() to
+ * ensure reading a non-zero refcount provides up to date bp_desc data.
+ */
+ atomic_set_release(&bp_desc.refs, 1);
/*
* Corresponding read barrier in int3 notifier for making sure the
@@ -1559,12 +1562,10 @@ static void text_poke_bp_batch(struct te
text_poke_sync();
/*
- * Remove and synchronize_rcu(), except we have a very primitive
- * refcount based completion.
+ * Remove and wait for refs to be zero.
*/
- WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
- if (!atomic_dec_and_test(&desc.refs))
- atomic_cond_read_acquire(&desc.refs, !VAL);
+ if (!atomic_dec_and_test(&bp_desc.refs))
+ atomic_cond_read_acquire(&bp_desc.refs, !VAL);
}
static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
On Sep 21, 2022, at 12:52 AM, Peter Zijlstra <[email protected]> wrote:
> ⚠ External Email
>
> On Tue, Sep 20, 2022 at 10:47:43PM +0000, Nadav Amit wrote:
>
>> Fix this issue with small backportable patch. Instead of trying to make
>> RCU-like behavior for bp_desc, just eliminate the unnecessary level of
>> indirection of bp_desc, and hold the whole descriptor on the stack.
>> Anyhow, there is only a single descriptor at any given moment.
>
> Because of text_mutex; indeed. No idea why I put that thing on the
> stack.
>
> I've done a few minor edits to your patch, but it otherwise looks good
> to me.
>
> ---
> Subject: x86/alternative: Fix race in try_get_desc()
> From: Nadav Amit <[email protected]>
> Date: Tue, 20 Sep 2022 22:47:43 +0000
>
> From: Nadav Amit <[email protected]>
>
> The text poke mechanism claims to have an RCU-like behavior, but it does
> not appear that there is any quiescent state to ensure that nobody holds
> reference to desc. As a result, the following race appears to be
> possible, which can lead to memory corruption.
>
> CPU0 CPU1
> ---- ----
> text_poke_bp_batch()
> -> smp_store_release(&bp_desc, &desc)
>
> [ notice that desc is on
> the stack ]
>
> poke_int3_handler()
>
> [ int3 might be kprobe's
> so sync events are do not
> help ]
>
> -> try_get_desc(descp=&bp_desc)
> desc = __READ_ONCE(bp_desc)
>
> if (!desc) [false, success]
> WRITE_ONCE(bp_desc, NULL);
> atomic_dec_and_test(&desc.refs)
>
> [ success, desc space on the stack
> is being reused and might have
> non-zero value. ]
> arch_atomic_inc_not_zero(&desc->refs)
>
> [ might succeed since desc points to
> stack memory that was freed and might
> be reused. ]
>
> I encountered some occasional crashes of poke_int3_handler() when
> kprobes are set, while accessing desc->vec. The analysis has been done
> offline and I did not corroborate the cause of the crashes. Yet, it
> seems that this race might be the root cause.
>
> Fix this issue with small backportable patch. Instead of trying to make
> RCU-like behavior for bp_desc, just eliminate the unnecessary level of
> indirection of bp_desc, and hold the whole descriptor on the stack.
Looks good to me. Thanks for improving my patch.
I just made one small mistake in commit message. It should say “hold
the whole descriptor as a global” in the line above.
I will send v2 with your changes and the updated commit message.
Thanks again.
The following commit has been merged into the x86/urgent branch of tip:
Commit-ID: efd608fa7403ba106412b437f873929e2c862e28
Gitweb: https://git.kernel.org/tip/efd608fa7403ba106412b437f873929e2c862e28
Author: Nadav Amit <[email protected]>
AuthorDate: Wed, 21 Sep 2022 18:09:32
Committer: Peter Zijlstra <[email protected]>
CommitterDate: Tue, 27 Sep 2022 22:50:26 +02:00
x86/alternative: Fix race in try_get_desc()
I encountered some occasional crashes of poke_int3_handler() when
kprobes are set, while accessing desc->vec.
The text poke mechanism claims to have an RCU-like behavior, but it
does not appear that there is any quiescent state to ensure that
nobody holds reference to desc. As a result, the following race
appears to be possible, which can lead to memory corruption.
CPU0 CPU1
---- ----
text_poke_bp_batch()
-> smp_store_release(&bp_desc, &desc)
[ notice that desc is on
the stack ]
poke_int3_handler()
[ int3 might be kprobe's
so sync events are do not
help ]
-> try_get_desc(descp=&bp_desc)
desc = __READ_ONCE(bp_desc)
if (!desc) [false, success]
WRITE_ONCE(bp_desc, NULL);
atomic_dec_and_test(&desc.refs)
[ success, desc space on the stack
is being reused and might have
non-zero value. ]
arch_atomic_inc_not_zero(&desc->refs)
[ might succeed since desc points to
stack memory that was freed and might
be reused. ]
Fix this issue with small backportable patch. Instead of trying to
make RCU-like behavior for bp_desc, just eliminate the unnecessary
level of indirection of bp_desc, and hold the whole descriptor as a
global. Anyhow, there is only a single descriptor at any given
moment.
Fixes: 1f676247f36a4 ("x86/alternatives: Implement a better poke_int3_handler() completion scheme")
Signed-off-by: Nadav Amit <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Cc: [email protected]
Link: https://lkml.kernel.org/r/[email protected]
---
arch/x86/kernel/alternative.c | 45 +++++++++++++++++-----------------
1 file changed, 23 insertions(+), 22 deletions(-)
diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
index 62f6b8b..4f32043 100644
--- a/arch/x86/kernel/alternative.c
+++ b/arch/x86/kernel/alternative.c
@@ -1319,22 +1319,23 @@ struct bp_patching_desc {
atomic_t refs;
};
-static struct bp_patching_desc *bp_desc;
+static struct bp_patching_desc bp_desc;
static __always_inline
-struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
+struct bp_patching_desc *try_get_desc(void)
{
- /* rcu_dereference */
- struct bp_patching_desc *desc = __READ_ONCE(*descp);
+ struct bp_patching_desc *desc = &bp_desc;
- if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
+ if (!arch_atomic_inc_not_zero(&desc->refs))
return NULL;
return desc;
}
-static __always_inline void put_desc(struct bp_patching_desc *desc)
+static __always_inline void put_desc(void)
{
+ struct bp_patching_desc *desc = &bp_desc;
+
smp_mb__before_atomic();
arch_atomic_dec(&desc->refs);
}
@@ -1367,15 +1368,15 @@ noinstr int poke_int3_handler(struct pt_regs *regs)
/*
* Having observed our INT3 instruction, we now must observe
- * bp_desc:
+ * bp_desc with non-zero refcount:
*
- * bp_desc = desc INT3
+ * bp_desc.refs = 1 INT3
* WMB RMB
- * write INT3 if (desc)
+ * write INT3 if (bp_desc.refs != 0)
*/
smp_rmb();
- desc = try_get_desc(&bp_desc);
+ desc = try_get_desc();
if (!desc)
return 0;
@@ -1429,7 +1430,7 @@ noinstr int poke_int3_handler(struct pt_regs *regs)
ret = 1;
out_put:
- put_desc(desc);
+ put_desc();
return ret;
}
@@ -1460,18 +1461,20 @@ static int tp_vec_nr;
*/
static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
{
- struct bp_patching_desc desc = {
- .vec = tp,
- .nr_entries = nr_entries,
- .refs = ATOMIC_INIT(1),
- };
unsigned char int3 = INT3_INSN_OPCODE;
unsigned int i;
int do_sync;
lockdep_assert_held(&text_mutex);
- smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
+ bp_desc.vec = tp;
+ bp_desc.nr_entries = nr_entries;
+
+ /*
+ * Corresponds to the implicit memory barrier in try_get_desc() to
+ * ensure reading a non-zero refcount provides up to date bp_desc data.
+ */
+ atomic_set_release(&bp_desc.refs, 1);
/*
* Corresponding read barrier in int3 notifier for making sure the
@@ -1559,12 +1562,10 @@ static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries
text_poke_sync();
/*
- * Remove and synchronize_rcu(), except we have a very primitive
- * refcount based completion.
+ * Remove and wait for refs to be zero.
*/
- WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
- if (!atomic_dec_and_test(&desc.refs))
- atomic_cond_read_acquire(&desc.refs, !VAL);
+ if (!atomic_dec_and_test(&bp_desc.refs))
+ atomic_cond_read_acquire(&bp_desc.refs, !VAL);
}
static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,