Skip to contentSkip to conversation

perf/core: Fix data race between perf_event_set_output and perf_mmap_close

PATCH Applied Yang Jihong 3 participants 7 messages Jul 4, 2022 – Jul 14, 2022 perf events

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 perf/core: Fix data race between perf_event_set_output and perf_mmap_close

Data race exists between perf_event_set_output and perf_mmap_close. The scenario is as follows:

                  CPU1                                                       CPU2
                                                                    perf_mmap_close(event2)
                                                                      if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0
                                                                        detach_rest = true;
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
  perf_event_set_output(event1, event2)
                                                                      if (!detach_rest)
                                                                        goto out_put;
                                                                      list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)
                                                                        ring_buffer_attach(event, NULL)
                                                                      // because event1 has not been added to event2->rb->event_list,
                                                                      // event1->rb is not set to NULL in these loops

    ring_buffer_attach(event1, event2->rb)
      list_add_rcu(&event1->rb_entry, &event2->rb->event_list)

The above data race causes a problem, that is, event1->rb is not NULL, but event1->rb->mmap_count is 0. If the perf_mmap interface is invoked for the fd of event1, the kernel keeps in the perf_mmap infinite loop:

again:

        mutex_lock(&event->mmap_mutex);
        if (event->rb) {
<SNIP>
                if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
                        /*
                         * Raced against perf_mmap_close() through
                         * perf_event_set_output(). Try again, hope for better
                         * luck.
                         */
                        mutex_unlock(&event->mmap_mutex);
                        goto again;
                }

<SNIP>

This problem has occurred on the Syzkaller, causing the softlockup. The call stack is as follows:

[ 3666.984385][    C2] watchdog: BUG: soft lockup - CPU#2 stuck for 23s! [syz-executor.2:32404]
[ 3666.986137][    C2] Modules linked in:
[ 3666.989581][    C2] CPU: 2 PID: 32404 Comm: syz-executor.2 Not tainted 5.10.0+ #4
[ 3666.990697][    C2] Hardware name: linux,dummy-virt (DT)
[ 3666.992270][    C2] pstate: 80000005 (Nzcv daif -PAN -UAO -TCO BTYPE=--)
[ 3666.993787][    C2] pc : __kasan_check_write+0x0/0x40
[ 3666.994841][    C2] lr : perf_mmap+0x3c8/0xf80
[ 3666.995661][    C2] sp : ffff00001011f8f0
[ 3666.996598][    C2] x29: ffff00001011f8f0 x28: ffff0000cf644868
[ 3666.998488][    C2] x27: ffff000012cad2c0 x26: 0000000000000000
[ 3666.999888][    C2] x25: 0000000000000001 x24: ffff000012cad298
[ 3667.003511][    C2] x23: 0000000000000000 x22: ffff000012cad000
[ 3667.005504][    C2] x21: ffff0000cf644818 x20: ffff0000cf6d2400
[ 3667.006891][    C2] x19: ffff0000cf6d24c0 x18: 0000000000000000
[ 3667.008295][    C2] x17: 0000000000000000 x16: 0000000000000000
[ 3667.009528][    C2] x15: 0000000000000000 x14: 0000000000000000
[ 3667.010658][    C2] x13: 0000000000000000 x12: ffff800002023f17
[ 3667.012169][    C2] x11: 1fffe00002023f16 x10: ffff800002023f16
[ 3667.013780][    C2] x9 : dfffa00000000000 x8 : ffff00001011f8b7
[ 3667.015265][    C2] x7 : 0000000000000001 x6 : ffff800002023f16
[ 3667.016683][    C2] x5 : ffff0000c0f36400 x4 : 0000000000000000
[ 3667.018078][    C2] x3 : ffffa00010000000 x2 : ffffa000119a0000
[ 3667.019343][    C2] x1 : 0000000000000004 x0 : ffff0000cf6d24c0
[ 3667.021276][    C2] Call trace:
[ 3667.022598][    C2]  __kasan_check_write+0x0/0x40
[ 3667.023666][    C2]  __mmap_region+0x7a4/0xc90
[ 3667.024679][    C2]  __do_mmap_mm+0x600/0xa20
[ 3667.025700][    C2]  do_mmap+0x114/0x384
[ 3667.026583][    C2]  vm_mmap_pgoff+0x138/0x230
[ 3667.027532][    C2]  ksys_mmap_pgoff+0x1d8/0x570
[ 3667.028537][    C2]  __arm64_sys_mmap+0xa4/0xd0
[ 3667.029597][    C2]  el0_svc_common.constprop.0+0xf4/0x414
[ 3667.030682][    C2]  do_el0_svc+0x50/0x11c
[ 3667.031545][    C2]  el0_svc+0x20/0x30
[ 3667.032368][    C2]  el0_sync_handler+0xe4/0x1e0
[ 3667.033305][    C2]  el0_sync+0x148/0x180

perf_mmap_close and ring_buffer_attach involve complicated lock recursion. The solution is to modify ring_buffer_attach function, that is, after new event is added to rb->event_list, check whether rb->mmap_count is 0, if the value is 0, another core performs perf_mmap_close operation on ring buffer during this period. In this case, we set event->rb to NULL.

---

Changes since v1:
- Modify commit message description.

 kernel/events/core.c | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)
kernel/events/core.c +23 -1
index 80782cddb1da..c67c070f7b39 100644
@@ -5900,6 +5900,7 @@ static void ring_buffer_attach(struct perf_event *event,
 			       struct perf_buffer *rb)
 {
 	struct perf_buffer *old_rb = NULL;
+	struct perf_buffer *new_rb = rb;
 	unsigned long flags;
 
 	WARN_ON_ONCE(event->parent);
@@ -5928,6 +5929,20 @@ static void ring_buffer_attach(struct perf_event *event,
 
 		spin_lock_irqsave(&rb->event_lock, flags);
 		list_add_rcu(&event->rb_entry, &rb->event_list);
+
+		/*
+		 * When perf_mmap_close traverses rb->event_list during
+		 * detach all other events, new event may not be added to
+		 * rb->event_list, let's check again, if rb->mmap_count is 0,
+		 * it indicates that perf_mmap_close is executed.
+		 * Manually delete event from rb->event_list and
+		 * set event->rb to null.
+		 */
+		if (!atomic_read(&rb->mmap_count)) {
+			list_del_rcu(&event->rb_entry);
+			new_rb = NULL;
+		}
+
 		spin_unlock_irqrestore(&rb->event_lock, flags);
 	}
 
@@ -5944,7 +5959,7 @@ static void ring_buffer_attach(struct perf_event *event,
 	if (has_aux(event))
 		perf_event_stop(event, 0);
 
-	rcu_assign_pointer(event->rb, rb);
+	rcu_assign_pointer(event->rb, new_rb);
 
 	if (old_rb) {
 		ring_buffer_put(old_rb);
@@ -11883,6 +11898,13 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 			goto unlock;
 	}
 
+	/*
+	 * If rb->mmap_count is 0, perf_mmap_close is being executed,
+	 * the ring buffer is about to be unmapped and cannot be attached.
+	 */
+	if (rb && !atomic_read(&rb->mmap_count))
+		goto unlock;
+
 	ring_buffer_attach(event, rb);
 
 	ret = 0;
Signature · 1 line
2.30.GIT
maintainer
#1 lore raw
Quoted 35 lines · Yang Jihong

Data race exists between perf_event_set_output and perf_mmap_close. The scenario is as follows:

                  CPU1                                                       CPU2
                                                                    perf_mmap_close(event2)
                                                                      if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0
                                                                        detach_rest = true;
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
  perf_event_set_output(event1, event2)
                                                                      if (!detach_rest)
                                                                        goto out_put;
                                                                      list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)
                                                                        ring_buffer_attach(event, NULL)
                                                                      // because event1 has not been added to event2->rb->event_list,
                                                                      // event1->rb is not set to NULL in these loops

    ring_buffer_attach(event1, event2->rb)
      list_add_rcu(&event1->rb_entry, &event2->rb->event_list)

The above data race causes a problem, that is, event1->rb is not NULL, but event1->rb->mmap_count is 0. If the perf_mmap interface is invoked for the fd of event1, the kernel keeps in the perf_mmap infinite loop:

again:

        mutex_lock(&event->mmap_mutex);
        if (event->rb) {
<SNIP>
                if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
                        /*
                         * Raced against perf_mmap_close() through
                         * perf_event_set_output(). Try again, hope for better
                         * luck.
                         */
                        mutex_unlock(&event->mmap_mutex);
                        goto again;
                }
Quoted 1 line

<SNIP>

Too tired, must look again tomorrow, little feeback below.

Quoted 37 lines
 kernel/events/core.c | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)
kernel/events/core.c +15 -0
index 80782cddb1da..c67c070f7b39 100644
@@ -5900,6 +5900,7 @@ static void ring_buffer_attach(struct perf_event *event,
 			       struct perf_buffer *rb)
 {
 	struct perf_buffer *old_rb = NULL;
+	struct perf_buffer *new_rb = rb;
 	unsigned long flags;
 
 	WARN_ON_ONCE(event->parent);
@@ -5928,6 +5929,20 @@ static void ring_buffer_attach(struct perf_event *event,
 
 		spin_lock_irqsave(&rb->event_lock, flags);
 		list_add_rcu(&event->rb_entry, &rb->event_list);
+
+		/*
+		 * When perf_mmap_close traverses rb->event_list during
+		 * detach all other events, new event may not be added to
+		 * rb->event_list, let's check again, if rb->mmap_count is 0,
+		 * it indicates that perf_mmap_close is executed.
+		 * Manually delete event from rb->event_list and
+		 * set event->rb to null.
+		 */
+		if (!atomic_read(&rb->mmap_count)) {
+			list_del_rcu(&event->rb_entry);
+			new_rb = NULL;
+		}
+
 		spin_unlock_irqrestore(&rb->event_lock, flags);
 	}
 

kernel/events/core.c static void ring_buffer_attach(struct perf_event *event, · line 5965

Quoted 8 lines
kernel/events/core.c +1 -1
@@ -5944,7 +5959,7 @@ static void ring_buffer_attach(struct perf_event *event,
 	if (has_aux(event))
 		perf_event_stop(event, 0);
 
-	rcu_assign_pointer(event->rb, rb);
+	rcu_assign_pointer(event->rb, new_rb);
 
 	if (old_rb) {
 		ring_buffer_put(old_rb);

I'm confused by the above hunks; the below will avoid calling ring_buffer_attach() when !rb->mmap_count, so how can the above ever execute?

Quoted 4 lines
kernel/events/core.c +0 -0
@@ -11883,6 +11898,13 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 			goto unlock;
 	}
 

kernel/events/core.c perf_event_set_output(struct perf_event *event, struct perf_event *output_event) · line 11910

Quoted 10 lines
kernel/events/core.c +7 -0
+	/*
+	 * If rb->mmap_count is 0, perf_mmap_close is being executed,
+	 * the ring buffer is about to be unmapped and cannot be attached.
+	 */
+	if (rb && !atomic_read(&rb->mmap_count))
+		goto unlock;
+
 	ring_buffer_attach(event, rb);
 
 	ret = 0;

This is wrong I think, it'll leak ring_buffer_get().

#2 lore raw

Hello,

Quoted 38 lines · Peter Zijlstra
Quoted 35 lines · Yang Jihong

Data race exists between perf_event_set_output and perf_mmap_close. The scenario is as follows:

                   CPU1                                                       CPU2
                                                                     perf_mmap_close(event2)
                                                                       if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0
                                                                         detach_rest = true;
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
   perf_event_set_output(event1, event2)
                                                                       if (!detach_rest)
                                                                         goto out_put;
                                                                       list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)
                                                                         ring_buffer_attach(event, NULL)
                                                                       // because event1 has not been added to event2->rb->event_list,
                                                                       // event1->rb is not set to NULL in these loops

     ring_buffer_attach(event1, event2->rb)
       list_add_rcu(&event1->rb_entry, &event2->rb->event_list)

The above data race causes a problem, that is, event1->rb is not NULL, but event1->rb->mmap_count is 0. If the perf_mmap interface is invoked for the fd of event1, the kernel keeps in the perf_mmap infinite loop:

again:

         mutex_lock(&event->mmap_mutex);
         if (event->rb) {
<SNIP>
                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
                         /*
                          * Raced against perf_mmap_close() through
                          * perf_event_set_output(). Try again, hope for better
                          * luck.
                          */
                         mutex_unlock(&event->mmap_mutex);
                         goto again;
                 }
Quoted 1 line

<SNIP>

Quoted 1 line

Too tired, must look again tomorrow, little feeback below.

Thanks for reviewing this patch. The perf_mmap_close, perf_event_set_output, and perf_mmap involve complex data race and lock relationships. Therefore, this simple fix is proposed.

Quoted 44 lines
Quoted 44 lines
  kernel/events/core.c | 24 +++++++++++++++++++++++-
  1 file changed, 23 insertions(+), 1 deletion(-)
kernel/events/core.c +16 -1
index 80782cddb1da..c67c070f7b39 100644
@@ -5900,6 +5900,7 @@ static void ring_buffer_attach(struct perf_event *event,
  			       struct perf_buffer *rb)
  {
  	struct perf_buffer *old_rb = NULL;
+	struct perf_buffer *new_rb = rb;
  	unsigned long flags;
 
  	WARN_ON_ONCE(event->parent);
@@ -5928,6 +5929,20 @@ static void ring_buffer_attach(struct perf_event *event,
 
  		spin_lock_irqsave(&rb->event_lock, flags);
  		list_add_rcu(&event->rb_entry, &rb->event_list);
+
+		/*
+		 * When perf_mmap_close traverses rb->event_list during
+		 * detach all other events, new event may not be added to
+		 * rb->event_list, let's check again, if rb->mmap_count is 0,
+		 * it indicates that perf_mmap_close is executed.
+		 * Manually delete event from rb->event_list and
+		 * set event->rb to null.
+		 */
+		if (!atomic_read(&rb->mmap_count)) {
+			list_del_rcu(&event->rb_entry);
+			new_rb = NULL;
+		}
+
  		spin_unlock_irqrestore(&rb->event_lock, flags);
  	}
 
@@ -5944,7 +5959,7 @@ static void ring_buffer_attach(struct perf_event *event,
  	if (has_aux(event))
  		perf_event_stop(event, 0);
 
-	rcu_assign_pointer(event->rb, rb);
+	rcu_assign_pointer(event->rb, new_rb);
 
Quoted 6 lines

kernel/events/core.c static void ring_buffer_attach(struct perf_event *event, · line 5965

Quoted 2 lines
kernel/events/core.c +0 -0
  	if (old_rb) {
  		ring_buffer_put(old_rb);

I'm confused by the above hunks; the below will avoid calling ring_buffer_attach() when !rb->mmap_count, so how can the above ever execute?

In this patch, !atomic_read(&rb->mmap_count) is checked before the perf_event_set_output function invokes ring_buffer_attach(event, rb). Therefore, !atomic_read(&rb->mmap_count) does not need to be checked in the ring_buffer_attach function.

Am I right to understand that?

Because there is no lock parallel protection between ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2) and perf_mmap_close(event2), they can be executed in parallel.

The following scenarios may exist:

CPU1
CPU2

perf_mmap_close(event2)

																	   ...
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
   perf_event_set_output(event1, event2)
     ...
     if (rb && !atomic_read(&rb->mmap_count))
	   goto unlock;
	 // Here rb->mmap_count = 1, Keep going.
	 ...
	                                                                  if 

(atomic_dec_and_test(&event2->rb->mmap_count) // mmap_count 1 -> 0

 detach_rest = true;
                                                                      ...

list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)

ring_buffer_attach(event, NULL)

// because event1 has not been added to event2->rb->event_list,

// event1->rb is not set to NULL in these loops

																	  ...
	 ring_buffer_attach(event1, rb)
	   ...
	   list_add_rcu(&event1->rb_entry, &event2->rb->event_list)
	   ...

In this case, the above problems arise.

Quoted 14 lines
Quoted 14 lines
kernel/events/core.c +7 -0
@@ -11883,6 +11898,13 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
  			goto unlock;
  	}
 
+	/*
+	 * If rb->mmap_count is 0, perf_mmap_close is being executed,
+	 * the ring buffer is about to be unmapped and cannot be attached.
+	 */
+	if (rb && !atomic_read(&rb->mmap_count))
+		goto unlock;
+
  	ring_buffer_attach(event, rb);
 
Quoted 3 lines

kernel/events/core.c perf_event_set_output(struct perf_event *event, struct perf_event *output_event) · line 11910

Quoted 1 line
kernel/events/core.c +0 -0
  	ret = 0;

This is wrong I think, it'll leak ring_buffer_get().

Yes, ring_buffer_put(rb) needs to be added before goto unlock. I'll fix in next version.

Thanks,
Yang

Quoted 4 lines

.

maintainer
#3 lore raw
In reply to Peter Zijlstra
Quoted 38 lines · Peter Zijlstra
Quoted 35 lines · Yang Jihong

Data race exists between perf_event_set_output and perf_mmap_close. The scenario is as follows:

                  CPU1                                                       CPU2
                                                                    perf_mmap_close(event2)
                                                                      if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0
                                                                        detach_rest = true;
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
  perf_event_set_output(event1, event2)
                                                                      if (!detach_rest)
                                                                        goto out_put;
                                                                      list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)
                                                                        ring_buffer_attach(event, NULL)
                                                                      // because event1 has not been added to event2->rb->event_list,
                                                                      // event1->rb is not set to NULL in these loops

    ring_buffer_attach(event1, event2->rb)
      list_add_rcu(&event1->rb_entry, &event2->rb->event_list)

The above data race causes a problem, that is, event1->rb is not NULL, but event1->rb->mmap_count is 0. If the perf_mmap interface is invoked for the fd of event1, the kernel keeps in the perf_mmap infinite loop:

again:

        mutex_lock(&event->mmap_mutex);
        if (event->rb) {
<SNIP>
                if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
                        /*
                         * Raced against perf_mmap_close() through
                         * perf_event_set_output(). Try again, hope for better
                         * luck.
                         */
                        mutex_unlock(&event->mmap_mutex);
                        goto again;
                }
Quoted 1 line

<SNIP>

Quoted 1 line

Too tired, must look again tomorrow, little feeback below.

With brain more awake I ended up with the below. Does that work?

---
 kernel/events/core.c | 45 +++++++++++++++++++++++++++++++--------------
 1 file changed, 31 insertions(+), 14 deletions(-)
kernel/events/core.c +31 -14
index 4d8c335a07db..c9d32d4d2e20 100644
@@ -6262,10 +6262,10 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 
 		if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
 			/*
-			 * Raced against perf_mmap_close() through
-			 * perf_event_set_output(). Try again, hope for better
-			 * luck.
+			 * Raced against perf_mmap_close(); remove the
+			 * event and try again.
 			 */
+			ring_buffer_attach(event, NULL);
 			mutex_unlock(&event->mmap_mutex);
 			goto again;
 		}
@@ -11840,14 +11840,25 @@ static int perf_copy_attr(struct perf_event_attr __user *uattr,
 	goto out;
 }
 
+static void mutex_lock_double(struct mutex *a, struct mutex *b)
+{
+	if (b < a)
+		swap(a, b);
+
+	mutex_lock(a);
+	mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
+}
+
 static int
 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 {
 	struct perf_buffer *rb = NULL;
 	int ret = -EINVAL;
 
-	if (!output_event)
+	if (!output_event) {
+		mutex_lock(&event->mmap_mutex);
 		goto set;
+	}
 
 	/* don't allow circular references */
 	if (event == output_event)
@@ -11885,8 +11896,15 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 	    event->pmu != output_event->pmu)
 		goto out;
 
+	/*
+	 * Hold both mmap_mutex to serialize against perf_mmap_close().  Since
+	 * output_event is already on rb->event_list, and the list iteration
+	 * restarts after every removal, it is guaranteed this new event is
+	 * observed *OR* if output_event is already removed, it's guaranteed we
+	 * observe !rb->mmap_count.
+	 */
+	mutex_lock_double(&event->mmap_mutex, &output_event->mmap_mutex);
 set:
-	mutex_lock(&event->mmap_mutex);
 	/* Can't redirect output if we've got an active mmap() */
 	if (atomic_read(&event->mmap_count))
 		goto unlock;
@@ -11896,6 +11914,12 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 		rb = ring_buffer_get(output_event);
 		if (!rb)
 			goto unlock;
+
+		/* did we race against perf_mmap_close() */
+		if (!atomic_read(&rb->mmap_count)) {
+			ring_buffer_put(rb);
+			goto unlock;
+		}
 	}
 
 	ring_buffer_attach(event, rb);
@@ -11903,20 +11927,13 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 	ret = 0;
 unlock:
 	mutex_unlock(&event->mmap_mutex);
+	if (output_event)
+		mutex_unlock(&output_event->mmap_mutex);
 
 out:
 	return ret;
 }
 
-static void mutex_lock_double(struct mutex *a, struct mutex *b)
-{
-	if (b < a)
-		swap(a, b);
-
-	mutex_lock(a);
-	mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
-}
-
 static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id)
 {
 	bool nmi_safe = false;
author Tested-by
#4 lore raw

Hello,

Quoted 39 lines · Peter Zijlstra
Quoted 38 lines · Peter Zijlstra
Quoted 36 lines · Yang Jihong

Data race exists between perf_event_set_output and perf_mmap_close. The scenario is as follows:

                   CPU1                                                       CPU2
                                                                     perf_mmap_close(event2)
                                                                       if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0
                                                                         detach_rest = true;
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
   perf_event_set_output(event1, event2)
                                                                       if (!detach_rest)
                                                                         goto out_put;
                                                                       list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)
                                                                         ring_buffer_attach(event, NULL)
                                                                       // because event1 has not been added to event2->rb->event_list,
                                                                       // event1->rb is not set to NULL in these loops

     ring_buffer_attach(event1, event2->rb)
       list_add_rcu(&event1->rb_entry, &event2->rb->event_list)

The above data race causes a problem, that is, event1->rb is not NULL, but event1->rb->mmap_count is 0. If the perf_mmap interface is invoked for the fd of event1, the kernel keeps in the perf_mmap infinite loop:

again:

         mutex_lock(&event->mmap_mutex);
         if (event->rb) {
<SNIP>
                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
                         /*
                          * Raced against perf_mmap_close() through
                          * perf_event_set_output(). Try again, hope for better
                          * luck.
                          */
                         mutex_unlock(&event->mmap_mutex);
                         goto again;
                 }

<SNIP>

Quoted 3 lines
Quoted 1 line

Too tired, must look again tomorrow, little feeback below.

With brain more awake I ended up with the below. Does that work?

Yes, I apply the patch on kernel versions 5.10 and mainline, and it could fixed the problem.

Thanks,
Yang

#5 lore raw

Hello,

Quoted 55 lines · Yang Jihong

Hello,

Quoted 52 lines · Peter Zijlstra
Quoted 48 lines · Peter Zijlstra
Quoted 32 lines · Yang Jihong

Data race exists between perf_event_set_output and perf_mmap_close. The scenario is as follows:

CPU1                                                       CPU2

perf_mmap_close(event2)

if (atomic_dec_and_test(&event2->rb->mmap_count)  // mmap_count 1 -> 0

detach_rest = true;
ioctl(event1, PERF_EVENT_IOC_SET_OUTPUT, event2)
   perf_event_set_output(event1, event2)

if (!detach_rest)

goto out_put;

list_for_each_entry_rcu(event, &event2->rb->event_list, rb_entry)

ring_buffer_attach(event, NULL)

// because event1 has not been added to event2->rb->event_list,

// event1->rb is not set to NULL in these loops

     ring_buffer_attach(event1, event2->rb) list_add_rcu(&event1->rb_entry, &event2->rb->event_list)

The above data race causes a problem, that is, event1->rb is not NULL, but event1->rb->mmap_count is 0. If the perf_mmap interface is invoked for the fd of event1, the kernel keeps in the perf_mmap infinite loop:

Quoted 14 lines

again:
         mutex_lock(&event->mmap_mutex);
         if (event->rb) {
<SNIP>
                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) { /*
                          * Raced against perf_mmap_close() through
                          * perf_event_set_output(). Try again, hope for better
                          * luck.
                          */
                         mutex_unlock(&event->mmap_mutex);
                         goto again;
                 }
<SNIP>

Quoted 1 line

Too tired, must look again tomorrow, little feeback below.

With brain more awake I ended up with the below. Does that work?

I have verified that this patch can solve the problem.

Do I submit this patch? Or do you submit it?

Thanks,
Yang

Quoted 9 lines

Yes, I apply the patch on kernel versions 5.10 and mainline, and it could fixed the problem.

Thanks,
Yang
.

merge bot Applied
#6 lore raw

[tip: perf/urgent] perf/core: Fix data race between perf_event_set_output() and perf_mmap_close()

In reply to Peter Zijlstra
The following commit has been merged into the perf/urgent branch of tip:

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     68e3c69803dada336893640110cb87221bb01dcf
Gitweb:        https://git.kernel.org/tip/68e3c69803dada336893640110cb87221bb01dcf
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Tue, 05 Jul 2022 15:07:26 +02:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Wed, 13 Jul 2022 11:29:12 +02:00

perf/core: Fix data race between perf_event_set_output() and perf_mmap_close()

Yang Jihing reported a race between perf_event_set_output() and perf_mmap_close():

	CPU1					CPU2

	perf_mmap_close(e2)
	  if (atomic_dec_and_test(&e2->rb->mmap_count)) // 1 - > 0
	    detach_rest = true

						ioctl(e1, IOC_SET_OUTPUT, e2)
						  perf_event_set_output(e1, e2)

	  ...
	  list_for_each_entry_rcu(e, &e2->rb->event_list, rb_entry)
	    ring_buffer_attach(e, NULL);
	    // e1 isn't yet added and
	    // therefore not detached

						    ring_buffer_attach(e1, e2->rb)
						      list_add_rcu(&e1->rb_entry,
								   &e2->rb->event_list)

After this; e1 is attached to an unmapped rb and a subsequent perf_mmap() will loop forever more:

	again:
		mutex_lock(&e->mmap_mutex);
		if (event->rb) {
			...
			if (!atomic_inc_not_zero(&e->rb->mmap_count)) {
				...
				mutex_unlock(&e->mmap_mutex);
				goto again;
			}
		}

The loop in perf_mmap_close() holds e2->mmap_mutex, while the attach in perf_event_set_output() holds e1->mmap_mutex. As such there is no serialization to avoid this race.

Change perf_event_set_output() to take both e1->mmap_mutex and e2->mmap_mutex to alleviate that problem. Additionally, have the loop in perf_mmap() detach the rb directly, this avoids having to wait for the concurrent perf_mmap_close() to get around to doing it to make progress.

---
 kernel/events/core.c | 45 +++++++++++++++++++++++++++++--------------
 1 file changed, 31 insertions(+), 14 deletions(-)
kernel/events/core.c +31 -14
index 80782cd..d2b3549 100644
@@ -6253,10 +6253,10 @@ again:
 
 		if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
 			/*
-			 * Raced against perf_mmap_close() through
-			 * perf_event_set_output(). Try again, hope for better
-			 * luck.
+			 * Raced against perf_mmap_close(); remove the
+			 * event and try again.
 			 */
+			ring_buffer_attach(event, NULL);
 			mutex_unlock(&event->mmap_mutex);
 			goto again;
 		}
@@ -11825,14 +11825,25 @@ err_size:
 	goto out;
 }
 
+static void mutex_lock_double(struct mutex *a, struct mutex *b)
+{
+	if (b < a)
+		swap(a, b);
+
+	mutex_lock(a);
+	mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
+}
+
 static int
 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 {
 	struct perf_buffer *rb = NULL;
 	int ret = -EINVAL;
 
-	if (!output_event)
+	if (!output_event) {
+		mutex_lock(&event->mmap_mutex);
 		goto set;
+	}
 
 	/* don't allow circular references */
 	if (event == output_event)
@@ -11870,8 +11881,15 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 	    event->pmu != output_event->pmu)
 		goto out;
 
+	/*
+	 * Hold both mmap_mutex to serialize against perf_mmap_close().  Since
+	 * output_event is already on rb->event_list, and the list iteration
+	 * restarts after every removal, it is guaranteed this new event is
+	 * observed *OR* if output_event is already removed, it's guaranteed we
+	 * observe !rb->mmap_count.
+	 */
+	mutex_lock_double(&event->mmap_mutex, &output_event->mmap_mutex);
 set:
-	mutex_lock(&event->mmap_mutex);
 	/* Can't redirect output if we've got an active mmap() */
 	if (atomic_read(&event->mmap_count))
 		goto unlock;
@@ -11881,6 +11899,12 @@ set:
 		rb = ring_buffer_get(output_event);
 		if (!rb)
 			goto unlock;
+
+		/* did we race against perf_mmap_close() */
+		if (!atomic_read(&rb->mmap_count)) {
+			ring_buffer_put(rb);
+			goto unlock;
+		}
 	}
 
 	ring_buffer_attach(event, rb);
@@ -11888,20 +11912,13 @@ set:
 	ret = 0;
 unlock:
 	mutex_unlock(&event->mmap_mutex);
+	if (output_event)
+		mutex_unlock(&output_event->mmap_mutex);
 
 out:
 	return ret;
 }
 
-static void mutex_lock_double(struct mutex *a, struct mutex *b)
-{
-	if (b < a)
-		swap(a, b);
-
-	mutex_lock(a);
-	mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
-}
-
 static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id)
 {
 	bool nmi_safe = false;