lockdep: Reduce stack_trace usage
authorYong Zhang <yong.zhang@windriver.com>
Tue, 4 May 2010 06:16:48 +0000 (14:16 +0800)
committerIngo Molnar <mingo@elte.hu>
Fri, 7 May 2010 09:27:26 +0000 (11:27 +0200)
When calling check_prevs_add(), if all validations passed
add_lock_to_list() will add new lock to dependency tree and
alloc stack_trace for each list_entry.

But at this time, we are always on the same stack, so stack_trace
for each list_entry has the same value. This is redundant and eats
up lots of memory which could lead to warning on low
MAX_STACK_TRACE_ENTRIES.

Use one copy of stack_trace instead.

V2: As suggested by Peter Zijlstra, move save_trace() from
    check_prevs_add() to check_prev_add().
    Add tracking for trylock dependence which is also redundant.

Signed-off-by: Yong Zhang <yong.zhang0@windriver.com>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20100504065711.GC10784@windriver.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/lockdep.c

index 9cf79858fd821646e5fef74f1ea14fd054392552..51080807dc8c78125bc4afa125792b449daeefbd 100644 (file)
@@ -805,7 +805,8 @@ static struct lock_list *alloc_list_entry(void)
  * Add a new dependency to the head of the list:
  */
 static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
-                           struct list_head *head, unsigned long ip, int distance)
+                           struct list_head *head, unsigned long ip,
+                           int distance, struct stack_trace *trace)
 {
        struct lock_list *entry;
        /*
@@ -816,11 +817,9 @@ static int add_lock_to_list(struct lock_class *class, struct lock_class *this,
        if (!entry)
                return 0;
 
-       if (!save_trace(&entry->trace))
-               return 0;
-
        entry->class = this;
        entry->distance = distance;
+       entry->trace = *trace;
        /*
         * Since we never remove from the dependency list, the list can
         * be walked lockless by other CPUs, it's only allocation
@@ -1622,12 +1621,20 @@ 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)
+              struct held_lock *next, int distance, int trylock_loop)
 {
        struct lock_list *entry;
        int ret;
        struct lock_list this;
        struct lock_list *uninitialized_var(target_entry);
+       /*
+        * Static variable, serialized by the graph_lock().
+        *
+        * We use this static variable to save the stack trace in case
+        * we call into this function multiple times due to encountering
+        * trylocks in the held lock stack.
+        */
+       static struct stack_trace trace;
 
        /*
         * Prove that the new <prev> -> <next> dependency would not
@@ -1675,20 +1682,23 @@ check_prev_add(struct task_struct *curr, struct held_lock *prev,
                }
        }
 
+       if (!trylock_loop && !save_trace(&trace))
+               return 0;
+
        /*
         * Ok, all validations passed, add the new lock
         * to the previous lock's dependency list:
         */
        ret = add_lock_to_list(hlock_class(prev), hlock_class(next),
                               &hlock_class(prev)->locks_after,
-                              next->acquire_ip, distance);
+                              next->acquire_ip, distance, &trace);
 
        if (!ret)
                return 0;
 
        ret = add_lock_to_list(hlock_class(next), hlock_class(prev),
                               &hlock_class(next)->locks_before,
-                              next->acquire_ip, distance);
+                              next->acquire_ip, distance, &trace);
        if (!ret)
                return 0;
 
@@ -1718,6 +1728,7 @@ static int
 check_prevs_add(struct task_struct *curr, struct held_lock *next)
 {
        int depth = curr->lockdep_depth;
+       int trylock_loop = 0;
        struct held_lock *hlock;
 
        /*
@@ -1743,7 +1754,8 @@ check_prevs_add(struct task_struct *curr, struct held_lock *next)
                 * added:
                 */
                if (hlock->read != 2) {
-                       if (!check_prev_add(curr, hlock, next, distance))
+                       if (!check_prev_add(curr, hlock, next,
+                                               distance, trylock_loop))
                                return 0;
                        /*
                         * Stop after the first non-trylock entry,
@@ -1766,6 +1778,7 @@ 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: