Skip to contentSkip to conversation

perf_event: add check for group_list if the parent isn't counted

PATCH In review Liming Wang 3 participants 6 messages Dec 30, 2009 – Dec 31, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf_event: add check for group_list if the parent isn't counted

If the parent has no entry in group_list, child_ctx will not be allocated, which will lead dereference of a NULL child_ctx.

---
 kernel/perf_event.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)
kernel/perf_event.c +2 -0
index 5b987b4..3664c4b 100644
@@ -5126,6 +5126,8 @@ int perf_event_init_task(struct task_struct *child)
 	 */
 	mutex_lock(&parent_ctx->mutex);
 
+	if (list_empty(&parent_ctx->group_list))
+		goto exit;
 	/*
 	 * We dont have to disable NMIs - we are only looking at
 	 * the list, not manipulating it:
Signature · 1 line
1.6.0.3
#1 lore raw
Quoted 2 lines · Liming Wang

If the parent has no entry in group_list, child_ctx will not be allocated, which will lead dereference of a NULL child_ctx.

That changelog sucks.

Best I can make of it is that there is a race where the parent gets his context instantiated and we manage to get the mutex before the other thread manages to add the first event.

Then we observe parent_event_ctx but have an empty list.

Is that it?

In that case, would it not be better to change the if (inherited_all) condition to if (inherited_all && child_ctx) ?

Quoted 18 lines
---
 kernel/perf_event.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)
kernel/perf_event.c +2 -0
index 5b987b4..3664c4b 100644
@@ -5126,6 +5126,8 @@ int perf_event_init_task(struct task_struct *child)
 	 */
 	mutex_lock(&parent_ctx->mutex);
 
+	if (list_empty(&parent_ctx->group_list))
+		goto exit;
 	/*
 	 * We dont have to disable NMIs - we are only looking at
 	 * the list, not manipulating it:
#2 lore raw
Quoted 5 lines · Peter Zijlstra
Quoted 2 lines · Liming Wang

If the parent has no entry in group_list, child_ctx will not be allocated, which will lead dereference of a NULL child_ctx.

That changelog sucks.

Sorry the changelog is too simple.

Quoted 8 lines

Best I can make of it is that there is a race where the parent gets his context instantiated and we manage to get the mutex before the other thread manages to add the first event.

Then we observe parent_event_ctx but have an empty list.

Is that it?

I didn't find this case.
In my case, if I perf record a existing process with "--pid" and finish record, and if later the recorded process forks a process, the condition will occur.

Liming Wang

Quoted 25 lines

In that case, would it not be better to change the if (inherited_all) condition to if (inherited_all && child_ctx) ?

Quoted 18 lines
---
 kernel/perf_event.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)
kernel/perf_event.c +2 -0
index 5b987b4..3664c4b 100644
@@ -5126,6 +5126,8 @@ int perf_event_init_task(struct task_struct *child)
 	 */
 	mutex_lock(&parent_ctx->mutex);
 
+	if (list_empty(&parent_ctx->group_list))
+		goto exit;
 	/*
 	 * We dont have to disable NMIs - we are only looking at
 	 * the list, not manipulating it:
#3 lore raw
Quoted 10 lines · Wang Liming
Quoted 7 lines

Best I can make of it is that there is a race where the parent gets his context instantiated and we manage to get the mutex before the other thread manages to add the first event.

Then we observe parent_event_ctx but have an empty list.

Is that it?

I didn't find this case.
In my case, if I perf record a existing process with "--pid" and finish record, and if later the recorded process forks a process, the condition will occur.

Ah, right, that will lead to the same state, since closing the last counter will not remove the context.

Does the below also fix your issue?

---
Subject: perf: Fix NULL deref in inheritance code
From: Peter Zijlstra <[email protected]>
Date: Wed Dec 30 16:00:35 CET 2009

Liming found a NULL deref when a task has a perf context but no counters when it forks.

This can occur in two cases, a race during construction where the fork hits after installing the context but before the first counter gets inserted, or more reproducably, a fork after the last counter is closed (which leaves the context around).

---
 kernel/perf_event.c |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
kernel/perf_event.c +2 -3
Index: linux-2.6/kernel/perf_event.c
===================================================================
@@ -5149,7 +5149,7 @@ int perf_event_init_task(struct task_str
 					    GFP_KERNEL);
 			if (!child_ctx) {
 				ret = -ENOMEM;
-				goto exit;
+				break;
 			}
 
 			__perf_event_init_context(child_ctx, child);
@@ -5165,7 +5165,7 @@ int perf_event_init_task(struct task_str
 		}
 	}
 
-	if (inherited_all) {
+	if (child_ctx && inherited_all) {
 		/*
 		 * Mark the child context as a clone of the parent
 		 * context, or of whatever the parent is a clone of.
@@ -5185,7 +5185,6 @@ int perf_event_init_task(struct task_str
 		get_ctx(child_ctx->parent_ctx);
 	}
 
-exit:
 	mutex_unlock(&parent_ctx->mutex);
 
 	perf_unpin_context(parent_ctx);
#4 lore raw
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Wang Liming
Quoted 7 lines

Best I can make of it is that there is a race where the parent gets his context instantiated and we manage to get the mutex before the other thread manages to add the first event.

Then we observe parent_event_ctx but have an empty list.

Is that it?

Quoted 8 lines
Quoted 3 lines

I didn't find this case.
In my case, if I perf record a existing process with "--pid" and finish record, and if later the recorded process forks a process, the condition will occur.

Ah, right, that will lead to the same state, since closing the last counter will not remove the context.

Does the below also fix your issue?

Yes, it's OK to me.
Thanks a lot!

Liming Wang

Quoted 53 lines

---
Subject: perf: Fix NULL deref in inheritance code
From: Peter Zijlstra <[email protected]>
Date: Wed Dec 30 16:00:35 CET 2009

Liming found a NULL deref when a task has a perf context but no counters when it forks.

This can occur in two cases, a race during construction where the fork hits after installing the context but before the first counter gets inserted, or more reproducably, a fork after the last counter is closed (which leaves the context around).

---
 kernel/perf_event.c |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

Index: linux-2.6/kernel/perf_event.c
===================================================================

kernel/perf_event.c +2 -3
@@ -5149,7 +5149,7 @@ int perf_event_init_task(struct task_str
 					    GFP_KERNEL);
 			if (!child_ctx) {
 				ret = -ENOMEM;
-				goto exit;
+				break;
 			}
 
 			__perf_event_init_context(child_ctx, child);
@@ -5165,7 +5165,7 @@ int perf_event_init_task(struct task_str
 		}
 	}
 
-	if (inherited_all) {
+	if (child_ctx && inherited_all) {
 		/*
 		 * Mark the child context as a clone of the parent
 		 * context, or of whatever the parent is a clone of.
@@ -5185,7 +5185,6 @@ int perf_event_init_task(struct task_str
 		get_ctx(child_ctx->parent_ctx);
 	}
 
-exit:
 	mutex_unlock(&parent_ctx->mutex);
 
 	perf_unpin_context(parent_ctx);
#5 lore raw

[tip:perf/urgent] perf: Fix NULL deref in inheritance code

In reply to Peter Zijlstra
Commit-ID:  05cbaa2853cdfc255fdd04e65a82bfe9208c4e52
Gitweb:     http://git.kernel.org/tip/05cbaa2853cdfc255fdd04e65a82bfe9208c4e52
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Wed, 30 Dec 2009 16:00:35 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 31 Dec 2009 13:11:31 +0100

perf: Fix NULL deref in inheritance code

Liming found a NULL deref when a task has a perf context but no counters when it forks.

This can occur in two cases, a race during construction where the fork hits after installing the context but before the first counter gets inserted, or more reproducably, a fork after the last counter is closed (which leaves the context around).

LKML-Reference: <1262185684.7135.222.camel@laptop>

---
 kernel/perf_event.c |    5 ++---
 1 files changed, 2 insertions(+), 3 deletions(-)
kernel/perf_event.c +2 -3
index 03cc061..58ed1da 100644
@@ -5148,7 +5148,7 @@ int perf_event_init_task(struct task_struct *child)
 					    GFP_KERNEL);
 			if (!child_ctx) {
 				ret = -ENOMEM;
-				goto exit;
+				break;
 			}
 
 			__perf_event_init_context(child_ctx, child);
@@ -5164,7 +5164,7 @@ int perf_event_init_task(struct task_struct *child)
 		}
 	}
 
-	if (inherited_all) {
+	if (child_ctx && inherited_all) {
 		/*
 		 * Mark the child context as a clone of the parent
 		 * context, or of whatever the parent is a clone of.
@@ -5184,7 +5184,6 @@ int perf_event_init_task(struct task_struct *child)
 		get_ctx(child_ctx->parent_ctx);
 	}
 
-exit:
 	mutex_unlock(&parent_ctx->mutex);
 
 	perf_unpin_context(parent_ctx);