Skip to contentSkip to conversation

lockdep: fix stack trace caching logic

PATCH In review Dmitry Vyukov 3 participants 5 messages Feb 4, 2016 – Feb 9, 2016 Locking primitives

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 lockdep: fix stack trace caching logic

check_prev_add() caches saved stack trace in static trace variable to avoid duplicate save_trace() calls in dependencies involving trylocks. But that caching logic contains a bug. We may not save trace on first iteration due to early return from check_prev_add(). Then on the second iteration when we actually need the trace we don't save it because we think that we've already saved it.

Let check_prev_add() itself control when stack is saved.

There is another bug. Trace variable is protected by graph lock. But we can temporary release graph lock during printing.

Fix this by invalidating cached stack trace when we release graph lock.

---
 kernel/locking/lockdep.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
kernel/locking/lockdep.c +10 -6
index 60ace56..c7710e4 100644
@@ -1822,7 +1822,7 @@ check_deadlock(struct task_struct *curr, struct held_lock *next,
  */
 static int
 check_prev_add(struct task_struct *curr, struct held_lock *prev,
-	       struct held_lock *next, int distance, int trylock_loop)
+	       struct held_lock *next, int distance, int *stack_saved)
 {
 	struct lock_list *entry;
 	int ret;
@@ -1883,8 +1883,11 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
 		}
 	}
 
-	if (!trylock_loop && !save_trace(&trace))
-		return 0;
+	if (!*stack_saved) {
+		if (!save_trace(&trace))
+			return 0;
+		*stack_saved = 1;
+	}
 
 	/*
 	 * Ok, all validations passed, add the new lock
@@ -1907,6 +1910,8 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
 	 * Debugging printouts:
 	 */
 	if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
+		/* We drop graph lock, so another thread can overwrite trace. */
+		*stack_saved = 0;
 		graph_unlock();
 		printk("\n new dependency: ");
 		print_lock_name(hlock_class(prev));
@@ -1929,7 +1934,7 @@ static int
 check_prevs_add(struct task_struct *curr, struct held_lock *next)
 {
 	int depth = curr->lockdep_depth;
-	int trylock_loop = 0;
+	int stack_saved = 0;
 	struct held_lock *hlock;
 
 	/*
@@ -1956,7 +1961,7 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
 		 */
 		if (hlock->read != 2 && hlock->check) {
 			if (!check_prev_add(curr, hlock, next,
-						distance, trylock_loop))
+						distance, &stack_saved))
 				return 0;
 			/*
 			 * Stop after the first non-trylock entry,
@@ -1979,7 +1984,6 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
 		if (curr->held_locks[depth].irq_context !=
 				curr->held_locks[depth-1].irq_context)
 			break;
-		trylock_loop = 1;
 	}
 	return 1;
 out_bug:
Signature · 1 line
2.7.0.rc3.207.g0ac5344
#1 lore raw
Quoted 5 lines · Dmitry Vyukov

check_prev_add() caches saved stack trace in static trace variable to avoid duplicate save_trace() calls in dependencies involving trylocks. But that caching logic contains a bug.

We may not save trace on first iteration due to early return from check_prev_add().

This commit log should identify the role test instrumentation plays in triggering this bug: is it a recursive read lock dependency injected between existing lock dependencies? What test component triggered this?

Quoted 78 lines

Then on the second iteration when we actually need the trace we don't save it because we think that we've already saved it.

Let check_prev_add() itself control when stack is saved.

There is another bug. Trace variable is protected by graph lock. But we can temporary release graph lock during printing.

Fix this by invalidating cached stack trace when we release graph lock.

---
 kernel/locking/lockdep.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
kernel/locking/lockdep.c +10 -6
index 60ace56..c7710e4 100644
@@ -1822,7 +1822,7 @@ check_deadlock(struct task_struct *curr, struct held_lock *next,
  */
 static int
 check_prev_add(struct task_struct *curr, struct held_lock *prev,
-	       struct held_lock *next, int distance, int trylock_loop)
+	       struct held_lock *next, int distance, int *stack_saved)
 {
 	struct lock_list *entry;
 	int ret;
@@ -1883,8 +1883,11 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
 		}
 	}
 
-	if (!trylock_loop && !save_trace(&trace))
-		return 0;
+	if (!*stack_saved) {
+		if (!save_trace(&trace))
+			return 0;
+		*stack_saved = 1;
+	}
 
 	/*
 	 * Ok, all validations passed, add the new lock
@@ -1907,6 +1910,8 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
 	 * Debugging printouts:
 	 */
 	if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
+		/* We drop graph lock, so another thread can overwrite trace. */
+		*stack_saved = 0;
 		graph_unlock();
 		printk("\n new dependency: ");
 		print_lock_name(hlock_class(prev));
@@ -1929,7 +1934,7 @@ static int
 check_prevs_add(struct task_struct *curr, struct held_lock *next)
 {
 	int depth = curr->lockdep_depth;
-	int trylock_loop = 0;
+	int stack_saved = 0;
 	struct held_lock *hlock;
 
 	/*
@@ -1956,7 +1961,7 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
 		 */
 		if (hlock->read != 2 && hlock->check) {
 			if (!check_prev_add(curr, hlock, next,
-						distance, trylock_loop))
+						distance, &stack_saved))
 				return 0;
 			/*
 			 * Stop after the first non-trylock entry,
@@ -1979,7 +1984,6 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
 		if (curr->held_locks[depth].irq_context !=
 				curr->held_locks[depth-1].irq_context)
 			break;
-		trylock_loop = 1;
 	}
 	return 1;
 out_bug:
#2 lore raw
Quoted 10 lines · Peter Hurley
Quoted 5 lines · Dmitry Vyukov

check_prev_add() caches saved stack trace in static trace variable to avoid duplicate save_trace() calls in dependencies involving trylocks. But that caching logic contains a bug.

We may not save trace on first iteration due to early return from check_prev_add().

This commit log should identify the role test instrumentation plays in triggering this bug: is it a recursive read lock dependency injected between existing lock dependencies? What test component triggered this?

I don't think that it has something to do with instrumentation. check_prev_add() has explicitly coded early exits paths which don't save stack trace.
For example, if first pair of locks is: try-read-lock vs read-lock, we will not save stack trace; then next pair is: write-lock vs read-lock, we will use bogus stack trace.

Quoted 79 lines
Quoted 78 lines

Then on the second iteration when we actually need the trace we don't save it because we think that we've already saved it.

Let check_prev_add() itself control when stack is saved.

There is another bug. Trace variable is protected by graph lock. But we can temporary release graph lock during printing.

Fix this by invalidating cached stack trace when we release graph lock.

---
 kernel/locking/lockdep.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
kernel/locking/lockdep.c +10 -6
index 60ace56..c7710e4 100644
@@ -1822,7 +1822,7 @@ check_deadlock(struct task_struct *curr, struct held_lock *next,
  */
 static int
 check_prev_add(struct task_struct *curr, struct held_lock *prev,
-            struct held_lock *next, int distance, int trylock_loop)
+            struct held_lock *next, int distance, int *stack_saved)
 {
      struct lock_list *entry;
      int ret;
@@ -1883,8 +1883,11 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
              }
      }
 
-     if (!trylock_loop && !save_trace(&trace))
-             return 0;
+     if (!*stack_saved) {
+             if (!save_trace(&trace))
+                     return 0;
+             *stack_saved = 1;
+     }
 
      /*
       * Ok, all validations passed, add the new lock
@@ -1907,6 +1910,8 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
       * Debugging printouts:
       */
      if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
+             /* We drop graph lock, so another thread can overwrite trace. */
+             *stack_saved = 0;
              graph_unlock();
              printk("\n new dependency: ");
              print_lock_name(hlock_class(prev));
@@ -1929,7 +1934,7 @@ static int
 check_prevs_add(struct task_struct *curr, struct held_lock *next)
 {
      int depth = curr->lockdep_depth;
-     int trylock_loop = 0;
+     int stack_saved = 0;
      struct held_lock *hlock;
 
      /*
@@ -1956,7 +1961,7 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
               */
              if (hlock->read != 2 && hlock->check) {
                      if (!check_prev_add(curr, hlock, next,
-                                             distance, trylock_loop))
+                                             distance, &stack_saved))
                              return 0;
                      /*
                       * Stop after the first non-trylock entry,
@@ -1979,7 +1984,6 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
              if (curr->held_locks[depth].irq_context !=
                              curr->held_locks[depth-1].irq_context)
                      break;
-             trylock_loop = 1;
      }
      return 1;
 out_bug:
#3 lore raw
Quoted 13 lines · Dmitry Vyukov
Quoted 10 lines · Peter Hurley
Quoted 5 lines · Dmitry Vyukov

check_prev_add() caches saved stack trace in static trace variable to avoid duplicate save_trace() calls in dependencies involving trylocks. But that caching logic contains a bug.

We may not save trace on first iteration due to early return from check_prev_add().

This commit log should identify the role test instrumentation plays in triggering this bug: is it a recursive read lock dependency injected between existing lock dependencies? What test component triggered this?

Quoted 6 lines

I don't think that it has something to do with instrumentation. check_prev_add() has explicitly coded early exits paths which don't save stack trace.
For example, if first pair of locks is: try-read-lock vs read-lock, we will not save stack trace; then next pair is: write-lock vs read-lock, we will use bogus stack trace.

Ok, but that isn't the case with the triggering stack traces we were getting.

But I see now that there's a existing lock dependency that check_prev_add() is skipping. IOW, the other early-out path is being taken; lockdep has already seen:

new dependency: (&tty->ldisc_sem){++++..} => (&buf->lock){+.+...}

So yeah, not related to test instrumentation then.

Quoted 79 lines
Quoted 79 lines
Quoted 78 lines

Then on the second iteration when we actually need the trace we don't save it because we think that we've already saved it.

Let check_prev_add() itself control when stack is saved.

There is another bug. Trace variable is protected by graph lock. But we can temporary release graph lock during printing.

Fix this by invalidating cached stack trace when we release graph lock.

---
 kernel/locking/lockdep.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
kernel/locking/lockdep.c +10 -6
index 60ace56..c7710e4 100644
@@ -1822,7 +1822,7 @@ check_deadlock(struct task_struct *curr, struct held_lock *next,
  */
 static int
 check_prev_add(struct task_struct *curr, struct held_lock *prev,
-            struct held_lock *next, int distance, int trylock_loop)
+            struct held_lock *next, int distance, int *stack_saved)
 {
      struct lock_list *entry;
      int ret;
@@ -1883,8 +1883,11 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
              }
      }
 
-     if (!trylock_loop && !save_trace(&trace))
-             return 0;
+     if (!*stack_saved) {
+             if (!save_trace(&trace))
+                     return 0;
+             *stack_saved = 1;
+     }
 
      /*
       * Ok, all validations passed, add the new lock
@@ -1907,6 +1910,8 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
       * Debugging printouts:
       */
      if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
+             /* We drop graph lock, so another thread can overwrite trace. */
+             *stack_saved = 0;
              graph_unlock();
              printk("\n new dependency: ");
              print_lock_name(hlock_class(prev));
@@ -1929,7 +1934,7 @@ static int
 check_prevs_add(struct task_struct *curr, struct held_lock *next)
 {
      int depth = curr->lockdep_depth;
-     int trylock_loop = 0;
+     int stack_saved = 0;
      struct held_lock *hlock;
 
      /*
@@ -1956,7 +1961,7 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
               */
              if (hlock->read != 2 && hlock->check) {
                      if (!check_prev_add(curr, hlock, next,
-                                             distance, trylock_loop))
+                                             distance, &stack_saved))
                              return 0;
                      /*
                       * Stop after the first non-trylock entry,
@@ -1979,7 +1984,6 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
              if (curr->held_locks[depth].irq_context !=
                              curr->held_locks[depth-1].irq_context)
                      break;
-             trylock_loop = 1;
      }
      return 1;
 out_bug:
#4 lore raw

[tip:locking/core] locking/lockdep: Fix stack trace caching logic

In reply to Dmitry Vyukov
locking/lockdep: Fix stack trace caching logic
Commit-ID:  8a5fd56431fe1682e870bd6ab0c276e74befbeb9
Gitweb:     http://git.kernel.org/tip/8a5fd56431fe1682e870bd6ab0c276e74befbeb9
Author:     Dmitry Vyukov <[email protected]>
AuthorDate: Thu, 4 Feb 2016 14:40:40 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Tue, 9 Feb 2016 11:06:08 +0100

locking/lockdep: Fix stack trace caching logic

check_prev_add() caches saved stack trace in static trace variable to avoid duplicate save_trace() calls in dependencies involving trylocks. But that caching logic contains a bug. We may not save trace on first iteration due to early return from check_prev_add(). Then on the second iteration when we actually need the trace we don't save it because we think that we've already saved it.

Let check_prev_add() itself control when stack is saved.

There is another bug. Trace variable is protected by graph lock. But we can temporary release graph lock during printing.

Fix this by invalidating cached stack trace when we release graph lock.

---
 kernel/locking/lockdep.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)
kernel/locking/lockdep.c +10 -6
index 60ace56..c7710e4 100644
@@ -1822,7 +1822,7 @@ check_deadlock(struct task_struct *curr, struct held_lock *next,
  */
 static int
 check_prev_add(struct task_struct *curr, struct held_lock *prev,
-	       struct held_lock *next, int distance, int trylock_loop)
+	       struct held_lock *next, int distance, int *stack_saved)
 {
 	struct lock_list *entry;
 	int ret;
@@ -1883,8 +1883,11 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
 		}
 	}
 
-	if (!trylock_loop && !save_trace(&trace))
-		return 0;
+	if (!*stack_saved) {
+		if (!save_trace(&trace))
+			return 0;
+		*stack_saved = 1;
+	}
 
 	/*
 	 * Ok, all validations passed, add the new lock
@@ -1907,6 +1910,8 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
 	 * Debugging printouts:
 	 */
 	if (verbose(hlock_class(prev)) || verbose(hlock_class(next))) {
+		/* We drop graph lock, so another thread can overwrite trace. */
+		*stack_saved = 0;
 		graph_unlock();
 		printk("\n new dependency: ");
 		print_lock_name(hlock_class(prev));
@@ -1929,7 +1934,7 @@ static int
 check_prevs_add(struct task_struct *curr, struct held_lock *next)
 {
 	int depth = curr->lockdep_depth;
-	int trylock_loop = 0;
+	int stack_saved = 0;
 	struct held_lock *hlock;
 
 	/*
@@ -1956,7 +1961,7 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
 		 */
 		if (hlock->read != 2 && hlock->check) {
 			if (!check_prev_add(curr, hlock, next,
-						distance, trylock_loop))
+						distance, &stack_saved))
 				return 0;
 			/*
 			 * Stop after the first non-trylock entry,
@@ -1979,7 +1984,6 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
 		if (curr->held_locks[depth].irq_context !=
 				curr->held_locks[depth-1].irq_context)
 			break;
-		trylock_loop = 1;
 	}
 	return 1;
 out_bug: