[PATCH] mutex subsystem, debugging code
authorIngo Molnar <mingo@elte.hu>
Mon, 9 Jan 2006 23:59:20 +0000 (15:59 -0800)
committerIngo Molnar <mingo@hera.kernel.org>
Mon, 9 Jan 2006 23:59:20 +0000 (15:59 -0800)
mutex implementation - add debugging code.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
include/linux/mutex-debug.h [new file with mode: 0644]
include/linux/sched.h
kernel/Makefile
kernel/fork.c
kernel/mutex-debug.c [new file with mode: 0644]
kernel/mutex-debug.h [new file with mode: 0644]
lib/Kconfig.debug

diff --git a/include/linux/mutex-debug.h b/include/linux/mutex-debug.h
new file mode 100644 (file)
index 0000000..0ccd8f9
--- /dev/null
@@ -0,0 +1,21 @@
+#ifndef __LINUX_MUTEX_DEBUG_H
+#define __LINUX_MUTEX_DEBUG_H
+
+/*
+ * Mutexes - debugging helpers:
+ */
+
+#define __DEBUG_MUTEX_INITIALIZER(lockname) \
+       , .held_list = LIST_HEAD_INIT(lockname.held_list), \
+         .name = #lockname , .magic = &lockname
+
+#define mutex_init(sem)                __mutex_init(sem, __FUNCTION__)
+
+extern void FASTCALL(mutex_destroy(struct mutex *lock));
+
+extern void mutex_debug_show_all_locks(void);
+extern void mutex_debug_show_held_locks(struct task_struct *filter);
+extern void mutex_debug_check_no_locks_held(struct task_struct *task);
+extern void mutex_debug_check_no_locks_freed(const void *from, const void *to);
+
+#endif
index 78eb92ae4d94b34240d457498ffb9c6dda6f34b6..85b53f87c7035ac7d6386b56db89c1abd0355559 100644 (file)
@@ -817,6 +817,11 @@ struct task_struct {
 /* Protection of proc_dentry: nesting proc_lock, dcache_lock, write_lock_irq(&tasklist_lock); */
        spinlock_t proc_lock;
 
+#ifdef CONFIG_DEBUG_MUTEXES
+       /* mutex deadlock detection */
+       struct mutex_waiter *blocked_on;
+#endif
+
 /* journalling filesystem info */
        void *journal_info;
 
index de580b4d54a4787ed35a5e62970de8e06baab1a5..a940bac028379d923ace8f304ed253622c2e0520 100644 (file)
@@ -9,6 +9,7 @@ obj-y     = sched.o fork.o exec_domain.o panic.o printk.o profile.o \
            rcupdate.o intermodule.o extable.o params.o posix-timers.o \
            kthread.o wait.o kfifo.o sys_ni.o posix-cpu-timers.o mutex.o
 
+obj-$(CONFIG_DEBUG_MUTEXES) += mutex-debug.o
 obj-$(CONFIG_FUTEX) += futex.o
 obj-$(CONFIG_GENERIC_ISA_DMA) += dma.o
 obj-$(CONFIG_SMP) += cpu.o spinlock.o
index 72e3252c6763400fea40a4f387a780627d2cbbf7..b18d64554feb7e48f1cbde8e2d58bafd52af1f04 100644 (file)
@@ -979,6 +979,10 @@ static task_t *copy_process(unsigned long clone_flags,
        }
 #endif
 
+#ifdef CONFIG_DEBUG_MUTEXES
+       p->blocked_on = NULL; /* not blocked yet */
+#endif
+
        p->tgid = p->pid;
        if (clone_flags & CLONE_THREAD)
                p->tgid = current->tgid;
diff --git a/kernel/mutex-debug.c b/kernel/mutex-debug.c
new file mode 100644 (file)
index 0000000..4fcb051
--- /dev/null
@@ -0,0 +1,464 @@
+/*
+ * kernel/mutex-debug.c
+ *
+ * Debugging code for mutexes
+ *
+ * Started by Ingo Molnar:
+ *
+ *  Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *
+ * lock debugging, locking tree, deadlock detection started by:
+ *
+ *  Copyright (C) 2004, LynuxWorks, Inc., Igor Manyilov, Bill Huey
+ *  Released under the General Public License (GPL).
+ */
+#include <linux/mutex.h>
+#include <linux/sched.h>
+#include <linux/delay.h>
+#include <linux/module.h>
+#include <linux/spinlock.h>
+#include <linux/kallsyms.h>
+#include <linux/interrupt.h>
+
+#include <asm/mutex.h>
+
+#include "mutex-debug.h"
+
+/*
+ * We need a global lock when we walk through the multi-process
+ * lock tree. Only used in the deadlock-debugging case.
+ */
+DEFINE_SPINLOCK(debug_mutex_lock);
+
+/*
+ * All locks held by all tasks, in a single global list:
+ */
+LIST_HEAD(debug_mutex_held_locks);
+
+/*
+ * In the debug case we carry the caller's instruction pointer into
+ * other functions, but we dont want the function argument overhead
+ * in the nondebug case - hence these macros:
+ */
+#define __IP_DECL__            , unsigned long ip
+#define __IP__                 , ip
+#define __RET_IP__             , (unsigned long)__builtin_return_address(0)
+
+/*
+ * "mutex debugging enabled" flag. We turn it off when we detect
+ * the first problem because we dont want to recurse back
+ * into the tracing code when doing error printk or
+ * executing a BUG():
+ */
+int debug_mutex_on = 1;
+
+static void printk_task(struct task_struct *p)
+{
+       if (p)
+               printk("%16s:%5d [%p, %3d]", p->comm, p->pid, p, p->prio);
+       else
+               printk("<none>");
+}
+
+static void printk_ti(struct thread_info *ti)
+{
+       if (ti)
+               printk_task(ti->task);
+       else
+               printk("<none>");
+}
+
+static void printk_task_short(struct task_struct *p)
+{
+       if (p)
+               printk("%s/%d [%p, %3d]", p->comm, p->pid, p, p->prio);
+       else
+               printk("<none>");
+}
+
+static void printk_lock(struct mutex *lock, int print_owner)
+{
+       printk(" [%p] {%s}\n", lock, lock->name);
+
+       if (print_owner && lock->owner) {
+               printk(".. held by:  ");
+               printk_ti(lock->owner);
+               printk("\n");
+       }
+       if (lock->owner) {
+               printk("... acquired at:               ");
+               print_symbol("%s\n", lock->acquire_ip);
+       }
+}
+
+/*
+ * printk locks held by a task:
+ */
+static void show_task_locks(struct task_struct *p)
+{
+       switch (p->state) {
+       case TASK_RUNNING:              printk("R"); break;
+       case TASK_INTERRUPTIBLE:        printk("S"); break;
+       case TASK_UNINTERRUPTIBLE:      printk("D"); break;
+       case TASK_STOPPED:              printk("T"); break;
+       case EXIT_ZOMBIE:               printk("Z"); break;
+       case EXIT_DEAD:                 printk("X"); break;
+       default:                        printk("?"); break;
+       }
+       printk_task(p);
+       if (p->blocked_on) {
+               struct mutex *lock = p->blocked_on->lock;
+
+               printk(" blocked on mutex:");
+               printk_lock(lock, 1);
+       } else
+               printk(" (not blocked on mutex)\n");
+}
+
+/*
+ * printk all locks held in the system (if filter == NULL),
+ * or all locks belonging to a single task (if filter != NULL):
+ */
+void show_held_locks(struct task_struct *filter)
+{
+       struct list_head *curr, *cursor = NULL;
+       struct mutex *lock;
+       struct thread_info *t;
+       unsigned long flags;
+       int count = 0;
+
+       if (filter) {
+               printk("------------------------------\n");
+               printk("| showing all locks held by: |  (");
+               printk_task_short(filter);
+               printk("):\n");
+               printk("------------------------------\n");
+       } else {
+               printk("---------------------------\n");
+               printk("| showing all locks held: |\n");
+               printk("---------------------------\n");
+       }
+
+       /*
+        * Play safe and acquire the global trace lock. We
+        * cannot printk with that lock held so we iterate
+        * very carefully:
+        */
+next:
+       debug_spin_lock_save(&debug_mutex_lock, flags);
+       list_for_each(curr, &debug_mutex_held_locks) {
+               if (cursor && curr != cursor)
+                       continue;
+               lock = list_entry(curr, struct mutex, held_list);
+               t = lock->owner;
+               if (filter && (t != filter->thread_info))
+                       continue;
+               count++;
+               cursor = curr->next;
+               debug_spin_lock_restore(&debug_mutex_lock, flags);
+
+               printk("\n#%03d:            ", count);
+               printk_lock(lock, filter ? 0 : 1);
+               goto next;
+       }
+       debug_spin_lock_restore(&debug_mutex_lock, flags);
+       printk("\n");
+}
+
+void mutex_debug_show_all_locks(void)
+{
+       struct task_struct *g, *p;
+       int count = 10;
+       int unlock = 1;
+
+       printk("\nShowing all blocking locks in the system:\n");
+
+       /*
+        * Here we try to get the tasklist_lock as hard as possible,
+        * if not successful after 2 seconds we ignore it (but keep
+        * trying). This is to enable a debug printout even if a
+        * tasklist_lock-holding task deadlocks or crashes.
+        */
+retry:
+       if (!read_trylock(&tasklist_lock)) {
+               if (count == 10)
+                       printk("hm, tasklist_lock locked, retrying... ");
+               if (count) {
+                       count--;
+                       printk(" #%d", 10-count);
+                       mdelay(200);
+                       goto retry;
+               }
+               printk(" ignoring it.\n");
+               unlock = 0;
+       }
+       if (count != 10)
+               printk(" locked it.\n");
+
+       do_each_thread(g, p) {
+               show_task_locks(p);
+               if (!unlock)
+                       if (read_trylock(&tasklist_lock))
+                               unlock = 1;
+       } while_each_thread(g, p);
+
+       printk("\n");
+       show_held_locks(NULL);
+       printk("=============================================\n\n");
+
+       if (unlock)
+               read_unlock(&tasklist_lock);
+}
+
+static void report_deadlock(struct task_struct *task, struct mutex *lock,
+                           struct mutex *lockblk, unsigned long ip)
+{
+       printk("\n%s/%d is trying to acquire this lock:\n",
+               current->comm, current->pid);
+       printk_lock(lock, 1);
+       printk("... trying at:                 ");
+       print_symbol("%s\n", ip);
+       show_held_locks(current);
+
+       if (lockblk) {
+               printk("but %s/%d is deadlocking current task %s/%d!\n\n",
+                       task->comm, task->pid, current->comm, current->pid);
+               printk("\n%s/%d is blocked on this lock:\n",
+                       task->comm, task->pid);
+               printk_lock(lockblk, 1);
+
+               show_held_locks(task);
+
+               printk("\n%s/%d's [blocked] stackdump:\n\n",
+                       task->comm, task->pid);
+               show_stack(task, NULL);
+       }
+
+       printk("\n%s/%d's [current] stackdump:\n\n",
+               current->comm, current->pid);
+       dump_stack();
+       mutex_debug_show_all_locks();
+       printk("[ turning off deadlock detection. Please report this. ]\n\n");
+       local_irq_disable();
+}
+
+/*
+ * Recursively check for mutex deadlocks:
+ */
+static int check_deadlock(struct mutex *lock, int depth,
+                         struct thread_info *ti, unsigned long ip)
+{
+       struct mutex *lockblk;
+       struct task_struct *task;
+
+       if (!debug_mutex_on)
+               return 0;
+
+       ti = lock->owner;
+       if (!ti)
+               return 0;
+
+       task = ti->task;
+       lockblk = NULL;
+       if (task->blocked_on)
+               lockblk = task->blocked_on->lock;
+
+       /* Self-deadlock: */
+       if (current == task) {
+               DEBUG_OFF();
+               if (depth)
+                       return 1;
+               printk("\n==========================================\n");
+               printk(  "[ BUG: lock recursion deadlock detected! |\n");
+               printk(  "------------------------------------------\n");
+               report_deadlock(task, lock, NULL, ip);
+               return 0;
+       }
+
+       /* Ugh, something corrupted the lock data structure? */
+       if (depth > 20) {
+               DEBUG_OFF();
+               printk("\n===========================================\n");
+               printk(  "[ BUG: infinite lock dependency detected!? |\n");
+               printk(  "-------------------------------------------\n");
+               report_deadlock(task, lock, lockblk, ip);
+               return 0;
+       }
+
+       /* Recursively check for dependencies: */
+       if (lockblk && check_deadlock(lockblk, depth+1, ti, ip)) {
+               printk("\n============================================\n");
+               printk(  "[ BUG: circular locking deadlock detected! ]\n");
+               printk(  "--------------------------------------------\n");
+               report_deadlock(task, lock, lockblk, ip);
+               return 0;
+       }
+       return 0;
+}
+
+/*
+ * Called when a task exits, this function checks whether the
+ * task is holding any locks, and reports the first one if so:
+ */
+void mutex_debug_check_no_locks_held(struct task_struct *task)
+{
+       struct list_head *curr, *next;
+       struct thread_info *t;
+       unsigned long flags;
+       struct mutex *lock;
+
+       if (!debug_mutex_on)
+               return;
+
+       debug_spin_lock_save(&debug_mutex_lock, flags);
+       list_for_each_safe(curr, next, &debug_mutex_held_locks) {
+               lock = list_entry(curr, struct mutex, held_list);
+               t = lock->owner;
+               if (t != task->thread_info)
+                       continue;
+               list_del_init(curr);
+               DEBUG_OFF();
+               debug_spin_lock_restore(&debug_mutex_lock, flags);
+
+               printk("BUG: %s/%d, lock held at task exit time!\n",
+                       task->comm, task->pid);
+               printk_lock(lock, 1);
+               if (lock->owner != task->thread_info)
+                       printk("exiting task is not even the owner??\n");
+               return;
+       }
+       debug_spin_lock_restore(&debug_mutex_lock, flags);
+}
+
+/*
+ * Called when kernel memory is freed (or unmapped), or if a mutex
+ * is destroyed or reinitialized - this code checks whether there is
+ * any held lock in the memory range of <from> to <to>:
+ */
+void mutex_debug_check_no_locks_freed(const void *from, const void *to)
+{
+       struct list_head *curr, *next;
+       unsigned long flags;
+       struct mutex *lock;
+       void *lock_addr;
+
+       if (!debug_mutex_on)
+               return;
+
+       debug_spin_lock_save(&debug_mutex_lock, flags);
+       list_for_each_safe(curr, next, &debug_mutex_held_locks) {
+               lock = list_entry(curr, struct mutex, held_list);
+               lock_addr = lock;
+               if (lock_addr < from || lock_addr >= to)
+                       continue;
+               list_del_init(curr);
+               DEBUG_OFF();
+               debug_spin_lock_restore(&debug_mutex_lock, flags);
+
+               printk("BUG: %s/%d, active lock [%p(%p-%p)] freed!\n",
+                       current->comm, current->pid, lock, from, to);
+               dump_stack();
+               printk_lock(lock, 1);
+               if (lock->owner != current_thread_info())
+                       printk("freeing task is not even the owner??\n");
+               return;
+       }
+       debug_spin_lock_restore(&debug_mutex_lock, flags);
+}
+
+/*
+ * Must be called with lock->wait_lock held.
+ */
+void debug_mutex_set_owner(struct mutex *lock,
+                          struct thread_info *new_owner __IP_DECL__)
+{
+       lock->owner = new_owner;
+       DEBUG_WARN_ON(!list_empty(&lock->held_list));
+       if (debug_mutex_on) {
+               list_add_tail(&lock->held_list, &debug_mutex_held_locks);
+               lock->acquire_ip = ip;
+       }
+}
+
+void debug_mutex_init_waiter(struct mutex_waiter *waiter)
+{
+       memset(waiter, 0x11, sizeof(*waiter));
+       waiter->magic = waiter;
+       INIT_LIST_HEAD(&waiter->list);
+}
+
+void debug_mutex_wake_waiter(struct mutex *lock, struct mutex_waiter *waiter)
+{
+       SMP_DEBUG_WARN_ON(!spin_is_locked(&lock->wait_lock));
+       DEBUG_WARN_ON(list_empty(&lock->wait_list));
+       DEBUG_WARN_ON(waiter->magic != waiter);
+       DEBUG_WARN_ON(list_empty(&waiter->list));
+}
+
+void debug_mutex_free_waiter(struct mutex_waiter *waiter)
+{
+       DEBUG_WARN_ON(!list_empty(&waiter->list));
+       memset(waiter, 0x22, sizeof(*waiter));
+}
+
+void debug_mutex_add_waiter(struct mutex *lock, struct mutex_waiter *waiter,
+                           struct thread_info *ti __IP_DECL__)
+{
+       SMP_DEBUG_WARN_ON(!spin_is_locked(&lock->wait_lock));
+       check_deadlock(lock, 0, ti, ip);
+       /* Mark the current thread as blocked on the lock: */
+       ti->task->blocked_on = waiter;
+       waiter->lock = lock;
+}
+
+void mutex_remove_waiter(struct mutex *lock, struct mutex_waiter *waiter,
+                        struct thread_info *ti)
+{
+       DEBUG_WARN_ON(list_empty(&waiter->list));
+       DEBUG_WARN_ON(waiter->task != ti->task);
+       DEBUG_WARN_ON(ti->task->blocked_on != waiter);
+       ti->task->blocked_on = NULL;
+
+       list_del_init(&waiter->list);
+       waiter->task = NULL;
+}
+
+void debug_mutex_unlock(struct mutex *lock)
+{
+       DEBUG_WARN_ON(lock->magic != lock);
+       DEBUG_WARN_ON(!lock->wait_list.prev && !lock->wait_list.next);
+       DEBUG_WARN_ON(lock->owner != current_thread_info());
+       if (debug_mutex_on) {
+               DEBUG_WARN_ON(list_empty(&lock->held_list));
+               list_del_init(&lock->held_list);
+       }
+}
+
+void debug_mutex_init(struct mutex *lock, const char *name)
+{
+       /*
+        * Make sure we are not reinitializing a held lock:
+        */
+       mutex_debug_check_no_locks_freed((void *)lock, (void *)(lock + 1));
+       lock->owner = NULL;
+       INIT_LIST_HEAD(&lock->held_list);
+       lock->name = name;
+       lock->magic = lock;
+}
+
+/***
+ * mutex_destroy - mark a mutex unusable
+ * @lock: the mutex to be destroyed
+ *
+ * This function marks the mutex uninitialized, and any subsequent
+ * use of the mutex is forbidden. The mutex must not be locked when
+ * this function is called.
+ */
+void fastcall mutex_destroy(struct mutex *lock)
+{
+       DEBUG_WARN_ON(mutex_is_locked(lock));
+       lock->magic = NULL;
+}
+
+EXPORT_SYMBOL_GPL(mutex_destroy);
+
diff --git a/kernel/mutex-debug.h b/kernel/mutex-debug.h
new file mode 100644 (file)
index 0000000..fd38405
--- /dev/null
@@ -0,0 +1,134 @@
+/*
+ * Mutexes: blocking mutual exclusion locks
+ *
+ * started by Ingo Molnar:
+ *
+ *  Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *
+ * This file contains mutex debugging related internal declarations,
+ * prototypes and inline functions, for the CONFIG_DEBUG_MUTEXES case.
+ * More details are in kernel/mutex-debug.c.
+ */
+
+extern spinlock_t debug_mutex_lock;
+extern struct list_head debug_mutex_held_locks;
+extern int debug_mutex_on;
+
+/*
+ * In the debug case we carry the caller's instruction pointer into
+ * other functions, but we dont want the function argument overhead
+ * in the nondebug case - hence these macros:
+ */
+#define __IP_DECL__            , unsigned long ip
+#define __IP__                 , ip
+#define __RET_IP__             , (unsigned long)__builtin_return_address(0)
+
+/*
+ * This must be called with lock->wait_lock held.
+ */
+extern void debug_mutex_set_owner(struct mutex *lock,
+                                 struct thread_info *new_owner __IP_DECL__);
+
+static inline void debug_mutex_clear_owner(struct mutex *lock)
+{
+       lock->owner = NULL;
+}
+
+extern void debug_mutex_init_waiter(struct mutex_waiter *waiter);
+extern void debug_mutex_wake_waiter(struct mutex *lock,
+                                   struct mutex_waiter *waiter);
+extern void debug_mutex_free_waiter(struct mutex_waiter *waiter);
+extern void debug_mutex_add_waiter(struct mutex *lock,
+                                  struct mutex_waiter *waiter,
+                                  struct thread_info *ti __IP_DECL__);
+extern void mutex_remove_waiter(struct mutex *lock, struct mutex_waiter *waiter,
+                               struct thread_info *ti);
+extern void debug_mutex_unlock(struct mutex *lock);
+extern void debug_mutex_init(struct mutex *lock, const char *name);
+
+#define debug_spin_lock(lock)                          \
+       do {                                            \
+               local_irq_disable();                    \
+               if (debug_mutex_on)                     \
+                       spin_lock(lock);                \
+       } while (0)
+
+#define debug_spin_unlock(lock)                                \
+       do {                                            \
+               if (debug_mutex_on)                     \
+                       spin_unlock(lock);              \
+               local_irq_enable();                     \
+               preempt_check_resched();                \
+       } while (0)
+
+#define debug_spin_lock_save(lock, flags)              \
+       do {                                            \
+               local_irq_save(flags);                  \
+               if (debug_mutex_on)                     \
+                       spin_lock(lock);                \
+       } while (0)
+
+#define debug_spin_lock_restore(lock, flags)           \
+       do {                                            \
+               if (debug_mutex_on)                     \
+                       spin_unlock(lock);              \
+               local_irq_restore(flags);               \
+               preempt_check_resched();                \
+       } while (0)
+
+#define spin_lock_mutex(lock)                          \
+       do {                                            \
+               struct mutex *l = container_of(lock, struct mutex, wait_lock); \
+                                                       \
+               DEBUG_WARN_ON(in_interrupt());          \
+               debug_spin_lock(&debug_mutex_lock);     \
+               spin_lock(lock);                        \
+               DEBUG_WARN_ON(l->magic != l);           \
+       } while (0)
+
+#define spin_unlock_mutex(lock)                                \
+       do {                                            \
+               spin_unlock(lock);                      \
+               debug_spin_unlock(&debug_mutex_lock);   \
+       } while (0)
+
+#define DEBUG_OFF()                                    \
+do {                                                   \
+       if (debug_mutex_on) {                           \
+               debug_mutex_on = 0;                     \
+               console_verbose();                      \
+               if (spin_is_locked(&debug_mutex_lock))  \
+                       spin_unlock(&debug_mutex_lock); \
+       }                                               \
+} while (0)
+
+#define DEBUG_BUG()                                    \
+do {                                                   \
+       if (debug_mutex_on) {                           \
+               DEBUG_OFF();                            \
+               BUG();                                  \
+       }                                               \
+} while (0)
+
+#define DEBUG_WARN_ON(c)                               \
+do {                                                   \
+       if (unlikely(c && debug_mutex_on)) {            \
+               DEBUG_OFF();                            \
+               WARN_ON(1);                             \
+       }                                               \
+} while (0)
+
+# define DEBUG_BUG_ON(c)                               \
+do {                                                   \
+       if (unlikely(c))                                \
+               DEBUG_BUG();                            \
+} while (0)
+
+#ifdef CONFIG_SMP
+# define SMP_DEBUG_WARN_ON(c)                  DEBUG_WARN_ON(c)
+# define SMP_DEBUG_BUG_ON(c)                   DEBUG_BUG_ON(c)
+#else
+# define SMP_DEBUG_WARN_ON(c)                  do { } while (0)
+# define SMP_DEBUG_BUG_ON(c)                   do { } while (0)
+#endif
+
index c48260fb8fd910add057a6d94f5ecc7b14d4fa2f..1fcd856edec1913e463551c8715abb61b70b55b2 100644 (file)
@@ -95,6 +95,14 @@ config DEBUG_PREEMPT
          if kernel code uses it in a preemption-unsafe way. Also, the kernel
          will detect preemption count underflows.
 
+config DEBUG_MUTEXES
+       bool "Mutex debugging, deadlock detection"
+       default y
+       depends on DEBUG_KERNEL
+       help
+        This allows mutex semantics violations and mutex related deadlocks
+        (lockups) to be detected and reported automatically.
+
 config DEBUG_SPINLOCK
        bool "Spinlock debugging"
        depends on DEBUG_KERNEL