Merge commit '900cfa46191a7d87cf1891924cb90499287fd235'; branches 'timers/nohz',...
[safe/jmp/linux-2.6] / kernel / lockdep.c
index a8dc99d..81a4e4a 100644 (file)
@@ -5,7 +5,8 @@
  *
  * Started by Ingo Molnar:
  *
- *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *  Copyright (C) 2006,2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  *
  * this code maps all the lock dependencies as they occur in a live kernel
  * and will warn about the following classes of locking bugs:
@@ -37,6 +38,7 @@
 #include <linux/debug_locks.h>
 #include <linux/irqflags.h>
 #include <linux/utsname.h>
+#include <linux/hash.h>
 
 #include <asm/sections.h>
 
@@ -175,6 +177,9 @@ struct lock_class_stats lock_stats(struct lock_class *class)
 
                lock_time_add(&pcs->read_holdtime, &stats.read_holdtime);
                lock_time_add(&pcs->write_holdtime, &stats.write_holdtime);
+
+               for (i = 0; i < ARRAY_SIZE(stats.bounces); i++)
+                       stats.bounces[i] += pcs->bounces[i];
        }
 
        return stats;
@@ -238,8 +243,7 @@ LIST_HEAD(all_lock_classes);
  */
 #define CLASSHASH_BITS         (MAX_LOCKDEP_KEYS_BITS - 1)
 #define CLASSHASH_SIZE         (1UL << CLASSHASH_BITS)
-#define CLASSHASH_MASK         (CLASSHASH_SIZE - 1)
-#define __classhashfn(key)     ((((unsigned long)key >> CLASSHASH_BITS) + (unsigned long)key) & CLASSHASH_MASK)
+#define __classhashfn(key)     hash_long((unsigned long)key, CLASSHASH_BITS)
 #define classhashentry(key)    (classhash_table + __classhashfn((key)))
 
 static struct list_head classhash_table[CLASSHASH_SIZE];
@@ -250,9 +254,7 @@ static struct list_head classhash_table[CLASSHASH_SIZE];
  */
 #define CHAINHASH_BITS         (MAX_LOCKDEP_CHAINS_BITS-1)
 #define CHAINHASH_SIZE         (1UL << CHAINHASH_BITS)
-#define CHAINHASH_MASK         (CHAINHASH_SIZE - 1)
-#define __chainhashfn(chain) \
-               (((chain >> CHAINHASH_BITS) + chain) & CHAINHASH_MASK)
+#define __chainhashfn(chain)   hash_long(chain, CHAINHASH_BITS)
 #define chainhashentry(chain)  (chainhash_table + __chainhashfn((chain)))
 
 static struct list_head chainhash_table[CHAINHASH_SIZE];
@@ -373,6 +375,11 @@ unsigned int max_recursion_depth;
  * about it later on, in lockdep_info().
  */
 static int lockdep_init_error;
+static unsigned long lockdep_init_trace_data[20];
+static struct stack_trace lockdep_init_trace = {
+       .max_entries = ARRAY_SIZE(lockdep_init_trace_data),
+       .entries = lockdep_init_trace_data,
+};
 
 /*
  * Various lockdep statistics:
@@ -504,11 +511,11 @@ static void lockdep_print_held_locks(struct task_struct *curr)
        int i, depth = curr->lockdep_depth;
 
        if (!depth) {
-               printk("no locks held by %s/%d.\n", curr->comm, curr->pid);
+               printk("no locks held by %s/%d.\n", curr->comm, task_pid_nr(curr));
                return;
        }
        printk("%d lock%s held by %s/%d:\n",
-               depth, depth > 1 ? "s" : "", curr->comm, curr->pid);
+               depth, depth > 1 ? "s" : "", curr->comm, task_pid_nr(curr));
 
        for (i = 0; i < depth; i++) {
                printk(" #%d: ", i);
@@ -660,6 +667,7 @@ look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
        if (unlikely(!lockdep_initialized)) {
                lockdep_init();
                lockdep_init_error = 1;
+               save_stack_trace(&lockdep_init_trace);
        }
 #endif
 
@@ -676,7 +684,8 @@ look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
         * (or spin_lock_init()) call - which acts as the key. For static
         * locks we use the lock object itself as the key.
         */
-       BUILD_BUG_ON(sizeof(struct lock_class_key) > sizeof(struct lock_class));
+       BUILD_BUG_ON(sizeof(struct lock_class_key) >
+                       sizeof(struct lockdep_map));
 
        key = lock->key->subkeys + subclass;
 
@@ -686,9 +695,12 @@ look_up_lock_class(struct lockdep_map *lock, unsigned int subclass)
         * We can walk the hash lockfree, because the hash only
         * grows, and we are careful when adding entries to the end:
         */
-       list_for_each_entry(class, hash_head, hash_entry)
-               if (class->key == key)
+       list_for_each_entry(class, hash_head, hash_entry) {
+               if (class->key == key) {
+                       WARN_ON_ONCE(class->name != lock->name);
                        return class;
+               }
+       }
 
        return NULL;
 }
@@ -767,6 +779,10 @@ register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
         * parallel walking of the hash-list safe:
         */
        list_add_tail_rcu(&class->hash_entry, hash_head);
+       /*
+        * Add it to the global list of classes:
+        */
+       list_add_tail_rcu(&class->lock_entry, &all_lock_classes);
 
        if (verbose(class)) {
                graph_unlock();
@@ -892,7 +908,7 @@ print_circular_bug_header(struct lock_list *entry, unsigned int depth)
        print_kernel_version();
        printk(  "-------------------------------------------------------\n");
        printk("%s/%d is trying to acquire lock:\n",
-               curr->comm, curr->pid);
+               curr->comm, task_pid_nr(curr));
        print_lock(check_source);
        printk("\nbut task is already holding lock:\n");
        print_lock(check_target);
@@ -1073,7 +1089,7 @@ print_bad_irq_dependency(struct task_struct *curr,
        print_kernel_version();
        printk(  "------------------------------------------------------\n");
        printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] is trying to acquire:\n",
-               curr->comm, curr->pid,
+               curr->comm, task_pid_nr(curr),
                curr->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
                curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
                curr->hardirqs_enabled,
@@ -1225,7 +1241,7 @@ print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
        print_kernel_version();
        printk(  "---------------------------------------------\n");
        printk("%s/%d is trying to acquire lock:\n",
-               curr->comm, curr->pid);
+               curr->comm, task_pid_nr(curr));
        print_lock(next);
        printk("\nbut task is already holding lock:\n");
        print_lock(prev);
@@ -1509,7 +1525,7 @@ cache_hit:
 }
 
 static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
-               struct held_lock *hlock, int chain_head)
+               struct held_lock *hlock, int chain_head, u64 chain_key)
 {
        /*
         * Trylock needs to maintain the stack of held locks, but it
@@ -1522,7 +1538,7 @@ static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
         * graph_lock for us)
         */
        if (!hlock->trylock && (hlock->check == 2) &&
-                       lookup_chain_cache(curr->curr_chain_key, hlock->class)) {
+                       lookup_chain_cache(chain_key, hlock->class)) {
                /*
                 * Check whether last held lock:
                 *
@@ -1564,7 +1580,7 @@ static int validate_chain(struct task_struct *curr, struct lockdep_map *lock,
 #else
 static inline int validate_chain(struct task_struct *curr,
                struct lockdep_map *lock, struct held_lock *hlock,
-               int chain_head)
+               int chain_head, u64 chain_key)
 {
        return 1;
 }
@@ -1629,7 +1645,7 @@ print_usage_bug(struct task_struct *curr, struct held_lock *this,
                usage_str[prev_bit], usage_str[new_bit]);
 
        printk("%s/%d [HC%u[%lu]:SC%u[%lu]:HE%u:SE%u] takes:\n",
-               curr->comm, curr->pid,
+               curr->comm, task_pid_nr(curr),
                trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
                trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
                trace_hardirqs_enabled(curr),
@@ -1682,7 +1698,7 @@ print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
        print_kernel_version();
        printk(  "---------------------------------------------------------\n");
        printk("%s/%d just changed the state of lock:\n",
-               curr->comm, curr->pid);
+               curr->comm, task_pid_nr(curr));
        print_lock(this);
        if (forwards)
                printk("but this lock took another, %s-irq-unsafe lock in the past:\n", irqclass);
@@ -2270,10 +2286,6 @@ static int mark_lock(struct task_struct *curr, struct held_lock *this,
                        return 0;
                break;
        case LOCK_USED:
-               /*
-                * Add it to the global list of classes:
-                */
-               list_add_tail_rcu(&this->class->lock_entry, &all_lock_classes);
                debug_atomic_dec(&nr_unused_locks);
                break;
        default:
@@ -2322,6 +2334,9 @@ void lockdep_init_map(struct lockdep_map *lock, const char *name,
        lock->name = name;
        lock->key = key;
        lock->class_cache = NULL;
+#ifdef CONFIG_LOCK_STAT
+       lock->cpu = raw_smp_processor_id();
+#endif
        if (subclass)
                register_lock_class(lock, subclass, 1);
 }
@@ -2409,7 +2424,7 @@ static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
                return 0;
 
        /*
-        * Calculate the chain hash: it's the combined has of all the
+        * Calculate the chain hash: it's the combined hash of all the
         * lock keys along the dependency chain. We save the hash value
         * at every step so that we can get the current hash easily
         * after unlock. The chain hash is then used to cache dependency
@@ -2435,11 +2450,11 @@ static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
                chain_head = 1;
        }
        chain_key = iterate_chain_key(chain_key, id);
-       curr->curr_chain_key = chain_key;
 
-       if (!validate_chain(curr, lock, hlock, chain_head))
+       if (!validate_chain(curr, lock, hlock, chain_head, chain_key))
                return 0;
 
+       curr->curr_chain_key = chain_key;
        curr->lockdep_depth++;
        check_chain_key(curr);
 #ifdef CONFIG_DEBUG_LOCKDEP
@@ -2472,7 +2487,7 @@ print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
        printk(  "[ BUG: bad unlock balance detected! ]\n");
        printk(  "-------------------------------------\n");
        printk("%s/%d is trying to release lock (",
-               curr->comm, curr->pid);
+               curr->comm, task_pid_nr(curr));
        print_lockdep_cache(lock);
        printk(") at:\n");
        print_ip_sym(ip);
@@ -2639,10 +2654,15 @@ static void check_flags(unsigned long flags)
        if (!debug_locks)
                return;
 
-       if (irqs_disabled_flags(flags))
-               DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled);
-       else
-               DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled);
+       if (irqs_disabled_flags(flags)) {
+               if (DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled)) {
+                       printk("possible reason: unannotated irqs-off.\n");
+               }
+       } else {
+               if (DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled)) {
+                       printk("possible reason: unannotated irqs-on.\n");
+               }
+       }
 
        /*
         * We dont accurately track softirq state in e.g.
@@ -2722,7 +2742,7 @@ print_lock_contention_bug(struct task_struct *curr, struct lockdep_map *lock,
        printk(  "[ BUG: bad contention detected! ]\n");
        printk(  "---------------------------------\n");
        printk("%s/%d is trying to contend lock (",
-               curr->comm, curr->pid);
+               curr->comm, task_pid_nr(curr));
        print_lockdep_cache(lock);
        printk(") at:\n");
        print_ip_sym(ip);
@@ -2772,6 +2792,8 @@ found_it:
        stats = get_lock_stats(hlock->class);
        if (point < ARRAY_SIZE(stats->contention_point))
                stats->contention_point[i]++;
+       if (lock->cpu != smp_processor_id())
+               stats->bounces[bounce_contended + !!hlock->read]++;
        put_lock_stats(stats);
 }
 
@@ -2783,8 +2805,8 @@ __lock_acquired(struct lockdep_map *lock)
        struct lock_class_stats *stats;
        unsigned int depth;
        u64 now;
-       s64 waittime;
-       int i;
+       s64 waittime = 0;
+       int i, cpu;
 
        depth = curr->lockdep_depth;
        if (DEBUG_LOCKS_WARN_ON(!depth))
@@ -2806,19 +2828,25 @@ __lock_acquired(struct lockdep_map *lock)
        return;
 
 found_it:
-       if (!hlock->waittime_stamp)
-               return;
-
-       now = sched_clock();
-       waittime = now - hlock->waittime_stamp;
-       hlock->holdtime_stamp = now;
+       cpu = smp_processor_id();
+       if (hlock->waittime_stamp) {
+               now = sched_clock();
+               waittime = now - hlock->waittime_stamp;
+               hlock->holdtime_stamp = now;
+       }
 
        stats = get_lock_stats(hlock->class);
-       if (hlock->read)
-               lock_time_inc(&stats->read_waittime, waittime);
-       else
-               lock_time_inc(&stats->write_waittime, waittime);
+       if (waittime) {
+               if (hlock->read)
+                       lock_time_inc(&stats->read_waittime, waittime);
+               else
+                       lock_time_inc(&stats->write_waittime, waittime);
+       }
+       if (lock->cpu != cpu)
+               stats->bounces[bounce_acquired + !!hlock->read]++;
        put_lock_stats(stats);
+
+       lock->cpu = cpu;
 }
 
 void lock_contended(struct lockdep_map *lock, unsigned long ip)
@@ -2904,7 +2932,7 @@ static void zap_class(struct lock_class *class)
 
 }
 
-static inline int within(void *addr, void *start, unsigned long size)
+static inline int within(const void *addr, void *start, unsigned long size)
 {
        return addr >= start && addr < start + size;
 }
@@ -2915,9 +2943,10 @@ void lockdep_free_key_range(void *start, unsigned long size)
        struct list_head *head;
        unsigned long flags;
        int i;
+       int locked;
 
        raw_local_irq_save(flags);
-       graph_lock();
+       locked = graph_lock();
 
        /*
         * Unhash all classes that were created by this module:
@@ -2926,12 +2955,16 @@ void lockdep_free_key_range(void *start, unsigned long size)
                head = classhash_table + i;
                if (list_empty(head))
                        continue;
-               list_for_each_entry_safe(class, next, head, hash_entry)
+               list_for_each_entry_safe(class, next, head, hash_entry) {
                        if (within(class->key, start, size))
                                zap_class(class);
+                       else if (within(class->name, start, size))
+                               zap_class(class);
+               }
        }
 
-       graph_unlock();
+       if (locked)
+               graph_unlock();
        raw_local_irq_restore(flags);
 }
 
@@ -2941,6 +2974,7 @@ void lockdep_reset_lock(struct lockdep_map *lock)
        struct list_head *head;
        unsigned long flags;
        int i, j;
+       int locked;
 
        raw_local_irq_save(flags);
 
@@ -2959,7 +2993,7 @@ void lockdep_reset_lock(struct lockdep_map *lock)
         * Debug check: in the end all mapped classes should
         * be gone.
         */
-       graph_lock();
+       locked = graph_lock();
        for (i = 0; i < CLASSHASH_SIZE; i++) {
                head = classhash_table + i;
                if (list_empty(head))
@@ -2972,7 +3006,8 @@ void lockdep_reset_lock(struct lockdep_map *lock)
                        }
                }
        }
-       graph_unlock();
+       if (locked)
+               graph_unlock();
 
 out_restore:
        raw_local_irq_restore(flags);
@@ -3023,16 +3058,14 @@ void __init lockdep_info(void)
                sizeof(struct held_lock) * MAX_LOCK_DEPTH);
 
 #ifdef CONFIG_DEBUG_LOCKDEP
-       if (lockdep_init_error)
-               printk("WARNING: lockdep init error! Arch code didnt call lockdep_init() early enough?\n");
+       if (lockdep_init_error) {
+               printk("WARNING: lockdep init error! Arch code didn't call lockdep_init() early enough?\n");
+               printk("Call stack leading to lockdep invocation was:\n");
+               print_stack_trace(&lockdep_init_trace, 0);
+       }
 #endif
 }
 
-static inline int in_range(const void *start, const void *addr, const void *end)
-{
-       return addr >= start && addr <= end;
-}
-
 static void
 print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
                     const void *mem_to, struct held_lock *hlock)
@@ -3046,7 +3079,7 @@ print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
        printk(  "[ BUG: held lock freed! ]\n");
        printk(  "-------------------------\n");
        printk("%s/%d is freeing memory %p-%p, with a lock still held there!\n",
-               curr->comm, curr->pid, mem_from, mem_to-1);
+               curr->comm, task_pid_nr(curr), mem_from, mem_to-1);
        print_lock(hlock);
        lockdep_print_held_locks(curr);
 
@@ -3054,6 +3087,13 @@ print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
        dump_stack();
 }
 
+static inline int not_in_range(const void* mem_from, unsigned long mem_len,
+                               const void* lock_from, unsigned long lock_len)
+{
+       return lock_from + lock_len <= mem_from ||
+               mem_from + mem_len <= lock_from;
+}
+
 /*
  * Called when kernel memory is freed (or unmapped), or if a lock
  * is destroyed or reinitialized - this code checks whether there is
@@ -3061,7 +3101,6 @@ print_freed_lock_bug(struct task_struct *curr, const void *mem_from,
  */
 void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
 {
-       const void *mem_to = mem_from + mem_len, *lock_from, *lock_to;
        struct task_struct *curr = current;
        struct held_lock *hlock;
        unsigned long flags;
@@ -3074,14 +3113,11 @@ void debug_check_no_locks_freed(const void *mem_from, unsigned long mem_len)
        for (i = 0; i < curr->lockdep_depth; i++) {
                hlock = curr->held_locks + i;
 
-               lock_from = (void *)hlock->instance;
-               lock_to = (void *)(hlock->instance + 1);
-
-               if (!in_range(mem_from, lock_from, mem_to) &&
-                                       !in_range(mem_from, lock_to, mem_to))
+               if (not_in_range(mem_from, mem_len, hlock->instance,
+                                       sizeof(*hlock->instance)))
                        continue;
 
-               print_freed_lock_bug(curr, mem_from, mem_to, hlock);
+               print_freed_lock_bug(curr, mem_from, mem_from + mem_len, hlock);
                break;
        }
        local_irq_restore(flags);
@@ -3099,7 +3135,7 @@ static void print_held_locks_bug(struct task_struct *curr)
        printk(  "[ BUG: lock held at task exit time! ]\n");
        printk(  "-------------------------------------\n");
        printk("%s/%d is exiting with locks still held!\n",
-               curr->comm, curr->pid);
+               curr->comm, task_pid_nr(curr));
        lockdep_print_held_locks(curr);
 
        printk("\nstack backtrace:\n");
@@ -3147,6 +3183,13 @@ retry:
                printk(" locked it.\n");
 
        do_each_thread(g, p) {
+               /*
+                * It's not reliable to print a task's held locks
+                * if it's not sleeping (or if it's not the current
+                * task):
+                */
+               if (p->state == TASK_RUNNING && p != current)
+                       continue;
                if (p->lockdep_depth)
                        lockdep_print_held_locks(p);
                if (!unlock)
@@ -3163,7 +3206,11 @@ retry:
 
 EXPORT_SYMBOL_GPL(debug_show_all_locks);
 
-void debug_show_held_locks(struct task_struct *task)
+/*
+ * Careful: only use this function if you are sure that
+ * the task cannot run in parallel!
+ */
+void __debug_show_held_locks(struct task_struct *task)
 {
        if (unlikely(!debug_locks)) {
                printk("INFO: lockdep is turned off.\n");
@@ -3171,5 +3218,27 @@ void debug_show_held_locks(struct task_struct *task)
        }
        lockdep_print_held_locks(task);
 }
+EXPORT_SYMBOL_GPL(__debug_show_held_locks);
+
+void debug_show_held_locks(struct task_struct *task)
+{
+               __debug_show_held_locks(task);
+}
 
 EXPORT_SYMBOL_GPL(debug_show_held_locks);
+
+void lockdep_sys_exit(void)
+{
+       struct task_struct *curr = current;
+
+       if (unlikely(curr->lockdep_depth)) {
+               if (!debug_locks_off())
+                       return;
+               printk("\n================================================\n");
+               printk(  "[ BUG: lock held when returning to user space! ]\n");
+               printk(  "------------------------------------------------\n");
+               printk("%s/%d is leaving the kernel with locks still held!\n",
+                               curr->comm, curr->pid);
+               lockdep_print_held_locks(curr);
+       }
+}