[PATCH] lockdep: core
Ingo Molnar [Mon, 3 Jul 2006 07:24:50 +0000 (00:24 -0700)]
Do 'make oldconfig' and accept all the defaults for new config options -
reboot into the kernel and if everything goes well it should boot up fine and
you should have /proc/lockdep and /proc/lockdep_stats files.

Typically if the lock validator finds some problem it will print out
voluminous debug output that begins with "BUG: ..." and which syslog output
can be used by kernel developers to figure out the precise locking scenario.

What does the lock validator do?  It "observes" and maps all locking rules as
they occur dynamically (as triggered by the kernel's natural use of spinlocks,
rwlocks, mutexes and rwsems).  Whenever the lock validator subsystem detects a
new locking scenario, it validates this new rule against the existing set of
rules.  If this new rule is consistent with the existing set of rules then the
new rule is added transparently and the kernel continues as normal.  If the
new rule could create a deadlock scenario then this condition is printed out.

When determining validity of locking, all possible "deadlock scenarios" are
considered: assuming arbitrary number of CPUs, arbitrary irq context and task
context constellations, running arbitrary combinations of all the existing
locking scenarios.  In a typical system this means millions of separate
scenarios.  This is why we call it a "locking correctness" validator - for all
rules that are observed the lock validator proves it with mathematical
certainty that a deadlock could not occur (assuming that the lock validator
implementation itself is correct and its internal data structures are not
corrupted by some other kernel subsystem).  [see more details and conditionals
of this statement in include/linux/lockdep.h and
Documentation/lockdep-design.txt]

Furthermore, this "all possible scenarios" property of the validator also
enables the finding of complex, highly unlikely multi-CPU multi-context races
via single single-context rules, increasing the likelyhood of finding bugs
drastically.  In practical terms: the lock validator already found a bug in
the upstream kernel that could only occur on systems with 3 or more CPUs, and
which needed 3 very unlikely code sequences to occur at once on the 3 CPUs.
That bug was found and reported on a single-CPU system (!).  So in essence a
race will be found "piecemail-wise", triggering all the necessary components
for the race, without having to reproduce the race scenario itself!  In its
short existence the lock validator found and reported many bugs before they
actually caused a real deadlock.

To further increase the efficiency of the validator, the mapping is not per
"lock instance", but per "lock-class".  For example, all struct inode objects
in the kernel have inode->inotify_mutex.  If there are 10,000 inodes cached,
then there are 10,000 lock objects.  But ->inotify_mutex is a single "lock
type", and all locking activities that occur against ->inotify_mutex are
"unified" into this single lock-class.  The advantage of the lock-class
approach is that all historical ->inotify_mutex uses are mapped into a single
(and as narrow as possible) set of locking rules - regardless of how many
different tasks or inode structures it took to build this set of rules.  The
set of rules persist during the lifetime of the kernel.

To see the rough magnitude of checking that the lock validator does, here's a
portion of /proc/lockdep_stats, fresh after bootup:

 lock-classes:                            694 [max: 2048]
 direct dependencies:                  1598 [max: 8192]
 indirect dependencies:               17896
 all direct dependencies:             16206
 dependency chains:                    1910 [max: 8192]
 in-hardirq chains:                      17
 in-softirq chains:                     105
 in-process chains:                    1065
 stack-trace entries:                 38761 [max: 131072]
 combined max dependencies:         2033928
 hardirq-safe locks:                     24
 hardirq-unsafe locks:                  176
 softirq-safe locks:                     53
 softirq-unsafe locks:                  137
 irq-safe locks:                         59
 irq-unsafe locks:                      176

The lock validator has observed 1598 actual single-thread locking patterns,
and has validated all possible 2033928 distinct locking scenarios.

More details about the design of the lock validator can be found in
Documentation/lockdep-design.txt, which can also found at:

   http://redhat.com/~mingo/lockdep-patches/lockdep-design.txt

[bunk@stusta.de: cleanups]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

13 files changed:
include/linux/hardirq.h
include/linux/init_task.h
include/linux/lockdep.h [new file with mode: 0644]
include/linux/sched.h
init/main.c
kernel/Makefile
kernel/fork.c
kernel/irq/manage.c
kernel/lockdep.c [new file with mode: 0644]
kernel/lockdep_internals.h [new file with mode: 0644]
kernel/module.c
lib/Kconfig.debug
lib/locking-selftest.c

index b1d4332..50d8b57 100644 (file)
@@ -3,6 +3,7 @@
 
 #include <linux/preempt.h>
 #include <linux/smp_lock.h>
+#include <linux/lockdep.h>
 #include <asm/hardirq.h>
 #include <asm/system.h>
 
@@ -122,7 +123,7 @@ static inline void account_system_vtime(struct task_struct *tsk)
  */
 extern void irq_exit(void);
 
-#define nmi_enter()            irq_enter()
-#define nmi_exit()             __irq_exit()
+#define nmi_enter()            do { lockdep_off(); irq_enter(); } while (0)
+#define nmi_exit()             do { __irq_exit(); lockdep_on(); } while (0)
 
 #endif /* LINUX_HARDIRQ_H */
index 444a3ae..60aac2c 100644 (file)
@@ -4,6 +4,7 @@
 #include <linux/file.h>
 #include <linux/rcupdate.h>
 #include <linux/irqflags.h>
+#include <linux/lockdep.h>
 
 #define INIT_FDTABLE \
 {                                                      \
@@ -126,6 +127,7 @@ extern struct group_info init_groups;
        .fs_excl        = ATOMIC_INIT(0),                               \
        .pi_lock        = SPIN_LOCK_UNLOCKED,                           \
        INIT_TRACE_IRQFLAGS                                             \
+       INIT_LOCKDEP                                                    \
 }
 
 
diff --git a/include/linux/lockdep.h b/include/linux/lockdep.h
new file mode 100644 (file)
index 0000000..80ec7a4
--- /dev/null
@@ -0,0 +1,347 @@
+/*
+ * Runtime locking correctness validator
+ *
+ *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *
+ * see Documentation/lockdep-design.txt for more details.
+ */
+#ifndef __LINUX_LOCKDEP_H
+#define __LINUX_LOCKDEP_H
+
+#include <linux/linkage.h>
+#include <linux/list.h>
+#include <linux/debug_locks.h>
+#include <linux/stacktrace.h>
+
+#ifdef CONFIG_LOCKDEP
+
+/*
+ * Lock-class usage-state bits:
+ */
+enum lock_usage_bit
+{
+       LOCK_USED = 0,
+       LOCK_USED_IN_HARDIRQ,
+       LOCK_USED_IN_SOFTIRQ,
+       LOCK_ENABLED_SOFTIRQS,
+       LOCK_ENABLED_HARDIRQS,
+       LOCK_USED_IN_HARDIRQ_READ,
+       LOCK_USED_IN_SOFTIRQ_READ,
+       LOCK_ENABLED_SOFTIRQS_READ,
+       LOCK_ENABLED_HARDIRQS_READ,
+       LOCK_USAGE_STATES
+};
+
+/*
+ * Usage-state bitmasks:
+ */
+#define LOCKF_USED                     (1 << LOCK_USED)
+#define LOCKF_USED_IN_HARDIRQ          (1 << LOCK_USED_IN_HARDIRQ)
+#define LOCKF_USED_IN_SOFTIRQ          (1 << LOCK_USED_IN_SOFTIRQ)
+#define LOCKF_ENABLED_HARDIRQS         (1 << LOCK_ENABLED_HARDIRQS)
+#define LOCKF_ENABLED_SOFTIRQS         (1 << LOCK_ENABLED_SOFTIRQS)
+
+#define LOCKF_ENABLED_IRQS (LOCKF_ENABLED_HARDIRQS | LOCKF_ENABLED_SOFTIRQS)
+#define LOCKF_USED_IN_IRQ (LOCKF_USED_IN_HARDIRQ | LOCKF_USED_IN_SOFTIRQ)
+
+#define LOCKF_USED_IN_HARDIRQ_READ     (1 << LOCK_USED_IN_HARDIRQ_READ)
+#define LOCKF_USED_IN_SOFTIRQ_READ     (1 << LOCK_USED_IN_SOFTIRQ_READ)
+#define LOCKF_ENABLED_HARDIRQS_READ    (1 << LOCK_ENABLED_HARDIRQS_READ)
+#define LOCKF_ENABLED_SOFTIRQS_READ    (1 << LOCK_ENABLED_SOFTIRQS_READ)
+
+#define LOCKF_ENABLED_IRQS_READ \
+               (LOCKF_ENABLED_HARDIRQS_READ | LOCKF_ENABLED_SOFTIRQS_READ)
+#define LOCKF_USED_IN_IRQ_READ \
+               (LOCKF_USED_IN_HARDIRQ_READ | LOCKF_USED_IN_SOFTIRQ_READ)
+
+#define MAX_LOCKDEP_SUBCLASSES         8UL
+
+/*
+ * Lock-classes are keyed via unique addresses, by embedding the
+ * lockclass-key into the kernel (or module) .data section. (For
+ * static locks we use the lock address itself as the key.)
+ */
+struct lockdep_subclass_key {
+       char __one_byte;
+} __attribute__ ((__packed__));
+
+struct lock_class_key {
+       struct lockdep_subclass_key     subkeys[MAX_LOCKDEP_SUBCLASSES];
+};
+
+/*
+ * The lock-class itself:
+ */
+struct lock_class {
+       /*
+        * class-hash:
+        */
+       struct list_head                hash_entry;
+
+       /*
+        * global list of all lock-classes:
+        */
+       struct list_head                lock_entry;
+
+       struct lockdep_subclass_key     *key;
+       unsigned int                    subclass;
+
+       /*
+        * IRQ/softirq usage tracking bits:
+        */
+       unsigned long                   usage_mask;
+       struct stack_trace              usage_traces[LOCK_USAGE_STATES];
+
+       /*
+        * These fields represent a directed graph of lock dependencies,
+        * to every node we attach a list of "forward" and a list of
+        * "backward" graph nodes.
+        */
+       struct list_head                locks_after, locks_before;
+
+       /*
+        * Generation counter, when doing certain classes of graph walking,
+        * to ensure that we check one node only once:
+        */
+       unsigned int                    version;
+
+       /*
+        * Statistics counter:
+        */
+       unsigned long                   ops;
+
+       const char                      *name;
+       int                             name_version;
+};
+
+/*
+ * Map the lock object (the lock instance) to the lock-class object.
+ * This is embedded into specific lock instances:
+ */
+struct lockdep_map {
+       struct lock_class_key           *key;
+       struct lock_class               *class[MAX_LOCKDEP_SUBCLASSES];
+       const char                      *name;
+};
+
+/*
+ * Every lock has a list of other locks that were taken after it.
+ * We only grow the list, never remove from it:
+ */
+struct lock_list {
+       struct list_head                entry;
+       struct lock_class               *class;
+       struct stack_trace              trace;
+};
+
+/*
+ * We record lock dependency chains, so that we can cache them:
+ */
+struct lock_chain {
+       struct list_head                entry;
+       u64                             chain_key;
+};
+
+struct held_lock {
+       /*
+        * One-way hash of the dependency chain up to this point. We
+        * hash the hashes step by step as the dependency chain grows.
+        *
+        * We use it for dependency-caching and we skip detection
+        * passes and dependency-updates if there is a cache-hit, so
+        * it is absolutely critical for 100% coverage of the validator
+        * to have a unique key value for every unique dependency path
+        * that can occur in the system, to make a unique hash value
+        * as likely as possible - hence the 64-bit width.
+        *
+        * The task struct holds the current hash value (initialized
+        * with zero), here we store the previous hash value:
+        */
+       u64                             prev_chain_key;
+       struct lock_class               *class;
+       unsigned long                   acquire_ip;
+       struct lockdep_map              *instance;
+
+       /*
+        * The lock-stack is unified in that the lock chains of interrupt
+        * contexts nest ontop of process context chains, but we 'separate'
+        * the hashes by starting with 0 if we cross into an interrupt
+        * context, and we also keep do not add cross-context lock
+        * dependencies - the lock usage graph walking covers that area
+        * anyway, and we'd just unnecessarily increase the number of
+        * dependencies otherwise. [Note: hardirq and softirq contexts
+        * are separated from each other too.]
+        *
+        * The following field is used to detect when we cross into an
+        * interrupt context:
+        */
+       int                             irq_context;
+       int                             trylock;
+       int                             read;
+       int                             check;
+       int                             hardirqs_off;
+};
+
+/*
+ * Initialization, self-test and debugging-output methods:
+ */
+extern void lockdep_init(void);
+extern void lockdep_info(void);
+extern void lockdep_reset(void);
+extern void lockdep_reset_lock(struct lockdep_map *lock);
+extern void lockdep_free_key_range(void *start, unsigned long size);
+
+extern void lockdep_off(void);
+extern void lockdep_on(void);
+extern int lockdep_internal(void);
+
+/*
+ * These methods are used by specific locking variants (spinlocks,
+ * rwlocks, mutexes and rwsems) to pass init/acquire/release events
+ * to lockdep:
+ */
+
+extern void lockdep_init_map(struct lockdep_map *lock, const char *name,
+                            struct lock_class_key *key);
+
+/*
+ * Reinitialize a lock key - for cases where there is special locking or
+ * special initialization of locks so that the validator gets the scope
+ * of dependencies wrong: they are either too broad (they need a class-split)
+ * or they are too narrow (they suffer from a false class-split):
+ */
+#define lockdep_set_class(lock, key) \
+               lockdep_init_map(&(lock)->dep_map, #key, key)
+#define lockdep_set_class_and_name(lock, key, name) \
+               lockdep_init_map(&(lock)->dep_map, name, key)
+
+/*
+ * Acquire a lock.
+ *
+ * Values for "read":
+ *
+ *   0: exclusive (write) acquire
+ *   1: read-acquire (no recursion allowed)
+ *   2: read-acquire with same-instance recursion allowed
+ *
+ * Values for check:
+ *
+ *   0: disabled
+ *   1: simple checks (freeing, held-at-exit-time, etc.)
+ *   2: full validation
+ */
+extern void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
+                        int trylock, int read, int check, unsigned long ip);
+
+extern void lock_release(struct lockdep_map *lock, int nested,
+                        unsigned long ip);
+
+# define INIT_LOCKDEP                          .lockdep_recursion = 0,
+
+#else /* !LOCKDEP */
+
+static inline void lockdep_off(void)
+{
+}
+
+static inline void lockdep_on(void)
+{
+}
+
+static inline int lockdep_internal(void)
+{
+       return 0;
+}
+
+# define lock_acquire(l, s, t, r, c, i)                do { } while (0)
+# define lock_release(l, n, i)                 do { } while (0)
+# define lockdep_init()                                do { } while (0)
+# define lockdep_info()                                do { } while (0)
+# define lockdep_init_map(lock, name, key)     do { (void)(key); } while (0)
+# define lockdep_set_class(lock, key)          do { (void)(key); } while (0)
+# define lockdep_set_class_and_name(lock, key, name) \
+               do { (void)(key); } while (0)
+# define INIT_LOCKDEP
+# define lockdep_reset()               do { debug_locks = 1; } while (0)
+# define lockdep_free_key_range(start, size)   do { } while (0)
+/*
+ * The class key takes no space if lockdep is disabled:
+ */
+struct lock_class_key { };
+#endif /* !LOCKDEP */
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+extern void early_boot_irqs_off(void);
+extern void early_boot_irqs_on(void);
+#else
+# define early_boot_irqs_off()                 do { } while (0)
+# define early_boot_irqs_on()                  do { } while (0)
+#endif
+
+/*
+ * For trivial one-depth nesting of a lock-class, the following
+ * global define can be used. (Subsystems with multiple levels
+ * of nesting should define their own lock-nesting subclasses.)
+ */
+#define SINGLE_DEPTH_NESTING                   1
+
+/*
+ * Map the dependency ops to NOP or to real lockdep ops, depending
+ * on the per lock-class debug mode:
+ */
+
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+# ifdef CONFIG_PROVE_LOCKING
+#  define spin_acquire(l, s, t, i)             lock_acquire(l, s, t, 0, 2, i)
+# else
+#  define spin_acquire(l, s, t, i)             lock_acquire(l, s, t, 0, 1, i)
+# endif
+# define spin_release(l, n, i)                 lock_release(l, n, i)
+#else
+# define spin_acquire(l, s, t, i)              do { } while (0)
+# define spin_release(l, n, i)                 do { } while (0)
+#endif
+
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+# ifdef CONFIG_PROVE_LOCKING
+#  define rwlock_acquire(l, s, t, i)           lock_acquire(l, s, t, 0, 2, i)
+#  define rwlock_acquire_read(l, s, t, i)      lock_acquire(l, s, t, 2, 2, i)
+# else
+#  define rwlock_acquire(l, s, t, i)           lock_acquire(l, s, t, 0, 1, i)
+#  define rwlock_acquire_read(l, s, t, i)      lock_acquire(l, s, t, 2, 1, i)
+# endif
+# define rwlock_release(l, n, i)               lock_release(l, n, i)
+#else
+# define rwlock_acquire(l, s, t, i)            do { } while (0)
+# define rwlock_acquire_read(l, s, t, i)       do { } while (0)
+# define rwlock_release(l, n, i)               do { } while (0)
+#endif
+
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+# ifdef CONFIG_PROVE_LOCKING
+#  define mutex_acquire(l, s, t, i)            lock_acquire(l, s, t, 0, 2, i)
+# else
+#  define mutex_acquire(l, s, t, i)            lock_acquire(l, s, t, 0, 1, i)
+# endif
+# define mutex_release(l, n, i)                        lock_release(l, n, i)
+#else
+# define mutex_acquire(l, s, t, i)             do { } while (0)
+# define mutex_release(l, n, i)                        do { } while (0)
+#endif
+
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+# ifdef CONFIG_PROVE_LOCKING
+#  define rwsem_acquire(l, s, t, i)            lock_acquire(l, s, t, 0, 2, i)
+#  define rwsem_acquire_read(l, s, t, i)       lock_acquire(l, s, t, 1, 2, i)
+# else
+#  define rwsem_acquire(l, s, t, i)            lock_acquire(l, s, t, 0, 1, i)
+#  define rwsem_acquire_read(l, s, t, i)       lock_acquire(l, s, t, 1, 1, i)
+# endif
+# define rwsem_release(l, n, i)                        lock_release(l, n, i)
+#else
+# define rwsem_acquire(l, s, t, i)             do { } while (0)
+# define rwsem_acquire_read(l, s, t, i)                do { } while (0)
+# define rwsem_release(l, n, i)                        do { } while (0)
+#endif
+
+#endif /* __LINUX_LOCKDEP_H */
index ad7a890..8ebddba 100644 (file)
@@ -886,6 +886,13 @@ struct task_struct {
        int hardirq_context;
        int softirq_context;
 #endif
+#ifdef CONFIG_LOCKDEP
+# define MAX_LOCK_DEPTH 30UL
+       u64 curr_chain_key;
+       int lockdep_depth;
+       struct held_lock held_locks[MAX_LOCK_DEPTH];
+       unsigned int lockdep_recursion;
+#endif
 
 /* journalling filesystem info */
        void *journal_info;
index fc73e1c..fc473d4 100644 (file)
@@ -48,6 +48,7 @@
 #include <linux/unwind.h>
 #include <linux/buffer_head.h>
 #include <linux/debug_locks.h>
+#include <linux/lockdep.h>
 
 #include <asm/io.h>
 #include <asm/bugs.h>
@@ -457,6 +458,15 @@ asmlinkage void __init start_kernel(void)
 
        smp_setup_processor_id();
 
+       /*
+        * Need to run as early as possible, to initialize the
+        * lockdep hash:
+        */
+       lockdep_init();
+
+       local_irq_disable();
+       early_boot_irqs_off();
+
 /*
  * Interrupts are still disabled. Do necessary setups, then
  * enable them
@@ -502,6 +512,7 @@ asmlinkage void __init start_kernel(void)
        profile_init();
        if (!irqs_disabled())
                printk("start_kernel(): bug: interrupts were enabled early\n");
+       early_boot_irqs_on();
        local_irq_enable();
 
        /*
@@ -512,6 +523,9 @@ asmlinkage void __init start_kernel(void)
        console_init();
        if (panic_later)
                panic(panic_later, panic_param);
+
+       lockdep_info();
+
        /*
         * Need to run this when irqs are enabled, because it wants
         * to self-test [hard/soft]-irqs on/off lock inversion bugs
index e7fd20e..049aa79 100644 (file)
@@ -13,6 +13,7 @@ obj-y     = sched.o fork.o exec_domain.o panic.o printk.o profile.o \
 obj-$(CONFIG_STACKTRACE) += stacktrace.o
 obj-y += time/
 obj-$(CONFIG_DEBUG_MUTEXES) += mutex-debug.o
+obj-$(CONFIG_LOCKDEP) += lockdep.o
 obj-$(CONFIG_FUTEX) += futex.o
 ifeq ($(CONFIG_COMPAT),y)
 obj-$(CONFIG_FUTEX) += futex_compat.o
index b7db7fb..7f48abd 100644 (file)
@@ -1061,6 +1061,11 @@ static task_t *copy_process(unsigned long clone_flags,
        p->hardirq_context = 0;
        p->softirq_context = 0;
 #endif
+#ifdef CONFIG_LOCKDEP
+       p->lockdep_depth = 0; /* no locks held yet */
+       p->curr_chain_key = 0;
+       p->lockdep_recursion = 0;
+#endif
 
        rt_mutex_init_task(p);
 
index c911c6e..4e46143 100644 (file)
@@ -410,6 +410,12 @@ int request_irq(unsigned int irq,
        struct irqaction *action;
        int retval;
 
+#ifdef CONFIG_LOCKDEP
+       /*
+        * Lockdep wants atomic interrupt handlers:
+        */
+       irqflags |= SA_INTERRUPT;
+#endif
        /*
         * Sanity-check: shared interrupts must pass in a real dev-ID,
         * otherwise we'll have trouble later trying to figure out
diff --git a/kernel/lockdep.c b/kernel/lockdep.c
new file mode 100644 (file)
index 0000000..dd05809
--- /dev/null
@@ -0,0 +1,2703 @@
+/*
+ * kernel/lockdep.c
+ *
+ * Runtime locking correctness validator
+ *
+ * Started by Ingo Molnar:
+ *
+ *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@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:
+ *
+ * - lock inversion scenarios
+ * - circular lock dependencies
+ * - hardirq/softirq safe/unsafe locking bugs
+ *
+ * Bugs are reported even if the current locking scenario does not cause
+ * any deadlock at this point.
+ *
+ * I.e. if anytime in the past two locks were taken in a different order,
+ * even if it happened for another task, even if those were different
+ * locks (but of the same class as this lock), this code will detect it.
+ *
+ * Thanks to Arjan van de Ven for coming up with the initial idea of
+ * mapping lock dependencies runtime.
+ */
+#include <linux/mutex.h>
+#include <linux/sched.h>
+#include <linux/delay.h>
+#include <linux/module.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
+#include <linux/spinlock.h>
+#include <linux/kallsyms.h>
+#include <linux/interrupt.h>
+#include <linux/stacktrace.h>
+#include <linux/debug_locks.h>
+#include <linux/irqflags.h>
+
+#include <asm/sections.h>
+
+#include "lockdep_internals.h"
+
+/*
+ * hash_lock: protects the lockdep hashes and class/list/hash allocators.
+ *
+ * This is one of the rare exceptions where it's justified
+ * to use a raw spinlock - we really dont want the spinlock
+ * code to recurse back into the lockdep code.
+ */
+static raw_spinlock_t hash_lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
+
+static int lockdep_initialized;
+
+unsigned long nr_list_entries;
+static struct lock_list list_entries[MAX_LOCKDEP_ENTRIES];
+
+/*
+ * Allocate a lockdep entry. (assumes hash_lock held, returns
+ * with NULL on failure)
+ */
+static struct lock_list *alloc_list_entry(void)
+{
+       if (nr_list_entries >= MAX_LOCKDEP_ENTRIES) {
+               __raw_spin_unlock(&hash_lock);
+               debug_locks_off();
+               printk("BUG: MAX_LOCKDEP_ENTRIES too low!\n");
+               printk("turning off the locking correctness validator.\n");
+               return NULL;
+       }
+       return list_entries + nr_list_entries++;
+}
+
+/*
+ * All data structures here are protected by the global debug_lock.
+ *
+ * Mutex key structs only get allocated, once during bootup, and never
+ * get freed - this significantly simplifies the debugging code.
+ */
+unsigned long nr_lock_classes;
+static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
+
+/*
+ * We keep a global list of all lock classes. The list only grows,
+ * never shrinks. The list is only accessed with the lockdep
+ * spinlock lock held.
+ */
+LIST_HEAD(all_lock_classes);
+
+/*
+ * The lockdep classes are in a hash-table as well, for fast lookup:
+ */
+#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 classhashentry(key)    (classhash_table + __classhashfn((key)))
+
+static struct list_head classhash_table[CLASSHASH_SIZE];
+
+unsigned long nr_lock_chains;
+static struct lock_chain lock_chains[MAX_LOCKDEP_CHAINS];
+
+/*
+ * We put the lock dependency chains into a hash-table as well, to cache
+ * their existence:
+ */
+#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 chainhashentry(chain)  (chainhash_table + __chainhashfn((chain)))
+
+static struct list_head chainhash_table[CHAINHASH_SIZE];
+
+/*
+ * The hash key of the lock dependency chains is a hash itself too:
+ * it's a hash of all locks taken up to that lock, including that lock.
+ * It's a 64-bit hash, because it's important for the keys to be
+ * unique.
+ */
+#define iterate_chain_key(key1, key2) \
+       (((key1) << MAX_LOCKDEP_KEYS_BITS/2) ^ \
+       ((key1) >> (64-MAX_LOCKDEP_KEYS_BITS/2)) ^ \
+       (key2))
+
+void lockdep_off(void)
+{
+       current->lockdep_recursion++;
+}
+
+EXPORT_SYMBOL(lockdep_off);
+
+void lockdep_on(void)
+{
+       current->lockdep_recursion--;
+}
+
+EXPORT_SYMBOL(lockdep_on);
+
+int lockdep_internal(void)
+{
+       return current->lockdep_recursion != 0;
+}
+
+EXPORT_SYMBOL(lockdep_internal);
+
+/*
+ * Debugging switches:
+ */
+
+#define VERBOSE                        0
+#ifdef VERBOSE
+# define VERY_VERBOSE          0
+#endif
+
+#if VERBOSE
+# define HARDIRQ_VERBOSE       1
+# define SOFTIRQ_VERBOSE       1
+#else
+# define HARDIRQ_VERBOSE       0
+# define SOFTIRQ_VERBOSE       0
+#endif
+
+#if VERBOSE || HARDIRQ_VERBOSE || SOFTIRQ_VERBOSE
+/*
+ * Quick filtering for interesting events:
+ */
+static int class_filter(struct lock_class *class)
+{
+       if (class->name_version == 1 &&
+                       !strcmp(class->name, "&rl->lock"))
+               return 1;
+       if (class->name_version == 1 &&
+                       !strcmp(class->name, "&ni->mrec_lock"))
+               return 1;
+       if (class->name_version == 1 &&
+                       !strcmp(class->name, "mft_ni_runlist_lock"))
+               return 1;
+       if (class->name_version == 1 &&
+                       !strcmp(class->name, "mft_ni_mrec_lock"))
+               return 1;
+       if (class->name_version == 1 &&
+                       !strcmp(class->name, "&vol->lcnbmp_lock"))
+               return 1;
+       return 0;
+}
+#endif
+
+static int verbose(struct lock_class *class)
+{
+#if VERBOSE
+       return class_filter(class);
+#endif
+       return 0;
+}
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+
+static int hardirq_verbose(struct lock_class *class)
+{
+#if HARDIRQ_VERBOSE
+       return class_filter(class);
+#endif
+       return 0;
+}
+
+static int softirq_verbose(struct lock_class *class)
+{
+#if SOFTIRQ_VERBOSE
+       return class_filter(class);
+#endif
+       return 0;
+}
+
+#endif
+
+/*
+ * Stack-trace: tightly packed array of stack backtrace
+ * addresses. Protected by the hash_lock.
+ */
+unsigned long nr_stack_trace_entries;
+static unsigned long stack_trace[MAX_STACK_TRACE_ENTRIES];
+
+static int save_trace(struct stack_trace *trace)
+{
+       trace->nr_entries = 0;
+       trace->max_entries = MAX_STACK_TRACE_ENTRIES - nr_stack_trace_entries;
+       trace->entries = stack_trace + nr_stack_trace_entries;
+
+       save_stack_trace(trace, NULL, 0, 3);
+
+       trace->max_entries = trace->nr_entries;
+
+       nr_stack_trace_entries += trace->nr_entries;
+       if (DEBUG_LOCKS_WARN_ON(nr_stack_trace_entries > MAX_STACK_TRACE_ENTRIES))
+               return 0;
+
+       if (nr_stack_trace_entries == MAX_STACK_TRACE_ENTRIES) {
+               __raw_spin_unlock(&hash_lock);
+               if (debug_locks_off()) {
+                       printk("BUG: MAX_STACK_TRACE_ENTRIES too low!\n");
+                       printk("turning off the locking correctness validator.\n");
+                       dump_stack();
+               }
+               return 0;
+       }
+
+       return 1;
+}
+
+unsigned int nr_hardirq_chains;
+unsigned int nr_softirq_chains;
+unsigned int nr_process_chains;
+unsigned int max_lockdep_depth;
+unsigned int max_recursion_depth;
+
+#ifdef CONFIG_DEBUG_LOCKDEP
+/*
+ * We cannot printk in early bootup code. Not even early_printk()
+ * might work. So we mark any initialization errors and printk
+ * about it later on, in lockdep_info().
+ */
+static int lockdep_init_error;
+
+/*
+ * Various lockdep statistics:
+ */
+atomic_t chain_lookup_hits;
+atomic_t chain_lookup_misses;
+atomic_t hardirqs_on_events;
+atomic_t hardirqs_off_events;
+atomic_t redundant_hardirqs_on;
+atomic_t redundant_hardirqs_off;
+atomic_t softirqs_on_events;
+atomic_t softirqs_off_events;
+atomic_t redundant_softirqs_on;
+atomic_t redundant_softirqs_off;
+atomic_t nr_unused_locks;
+atomic_t nr_cyclic_checks;
+atomic_t nr_cyclic_check_recursions;
+atomic_t nr_find_usage_forwards_checks;
+atomic_t nr_find_usage_forwards_recursions;
+atomic_t nr_find_usage_backwards_checks;
+atomic_t nr_find_usage_backwards_recursions;
+# define debug_atomic_inc(ptr)         atomic_inc(ptr)
+# define debug_atomic_dec(ptr)         atomic_dec(ptr)
+# define debug_atomic_read(ptr)                atomic_read(ptr)
+#else
+# define debug_atomic_inc(ptr)         do { } while (0)
+# define debug_atomic_dec(ptr)         do { } while (0)
+# define debug_atomic_read(ptr)                0
+#endif
+
+/*
+ * Locking printouts:
+ */
+
+static const char *usage_str[] =
+{
+       [LOCK_USED] =                   "initial-use ",
+       [LOCK_USED_IN_HARDIRQ] =        "in-hardirq-W",
+       [LOCK_USED_IN_SOFTIRQ] =        "in-softirq-W",
+       [LOCK_ENABLED_SOFTIRQS] =       "softirq-on-W",
+       [LOCK_ENABLED_HARDIRQS] =       "hardirq-on-W",
+       [LOCK_USED_IN_HARDIRQ_READ] =   "in-hardirq-R",
+       [LOCK_USED_IN_SOFTIRQ_READ] =   "in-softirq-R",
+       [LOCK_ENABLED_SOFTIRQS_READ] =  "softirq-on-R",
+       [LOCK_ENABLED_HARDIRQS_READ] =  "hardirq-on-R",
+};
+
+const char * __get_key_name(struct lockdep_subclass_key *key, char *str)
+{
+       unsigned long offs, size;
+       char *modname;
+
+       return kallsyms_lookup((unsigned long)key, &size, &offs, &modname, str);
+}
+
+void
+get_usage_chars(struct lock_class *class, char *c1, char *c2, char *c3, char *c4)
+{
+       *c1 = '.', *c2 = '.', *c3 = '.', *c4 = '.';
+
+       if (class->usage_mask & LOCKF_USED_IN_HARDIRQ)
+               *c1 = '+';
+       else
+               if (class->usage_mask & LOCKF_ENABLED_HARDIRQS)
+                       *c1 = '-';
+
+       if (class->usage_mask & LOCKF_USED_IN_SOFTIRQ)
+               *c2 = '+';
+       else
+               if (class->usage_mask & LOCKF_ENABLED_SOFTIRQS)
+                       *c2 = '-';
+
+       if (class->usage_mask & LOCKF_ENABLED_HARDIRQS_READ)
+               *c3 = '-';
+       if (class->usage_mask & LOCKF_USED_IN_HARDIRQ_READ) {
+               *c3 = '+';
+               if (class->usage_mask & LOCKF_ENABLED_HARDIRQS_READ)
+                       *c3 = '?';
+       }
+
+       if (class->usage_mask & LOCKF_ENABLED_SOFTIRQS_READ)
+               *c4 = '-';
+       if (class->usage_mask & LOCKF_USED_IN_SOFTIRQ_READ) {
+               *c4 = '+';
+               if (class->usage_mask & LOCKF_ENABLED_SOFTIRQS_READ)
+                       *c4 = '?';
+       }
+}
+
+static void print_lock_name(struct lock_class *class)
+{
+       char str[128], c1, c2, c3, c4;
+       const char *name;
+
+       get_usage_chars(class, &c1, &c2, &c3, &c4);
+
+       name = class->name;
+       if (!name) {
+               name = __get_key_name(class->key, str);
+               printk(" (%s", name);
+       } else {
+               printk(" (%s", name);
+               if (class->name_version > 1)
+                       printk("#%d", class->name_version);
+               if (class->subclass)
+                       printk("/%d", class->subclass);
+       }
+       printk("){%c%c%c%c}", c1, c2, c3, c4);
+}
+
+static void print_lockdep_cache(struct lockdep_map *lock)
+{
+       const char *name;
+       char str[128];
+
+       name = lock->name;
+       if (!name)
+               name = __get_key_name(lock->key->subkeys, str);
+
+       printk("%s", name);
+}
+
+static void print_lock(struct held_lock *hlock)
+{
+       print_lock_name(hlock->class);
+       printk(", at: ");
+       print_ip_sym(hlock->acquire_ip);
+}
+
+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);
+               return;
+       }
+       printk("%d lock%s held by %s/%d:\n",
+               depth, depth > 1 ? "s" : "", curr->comm, curr->pid);
+
+       for (i = 0; i < depth; i++) {
+               printk(" #%d: ", i);
+               print_lock(curr->held_locks + i);
+       }
+}
+/*
+ * Helper to print a nice hierarchy of lock dependencies:
+ */
+static void print_spaces(int nr)
+{
+       int i;
+
+       for (i = 0; i < nr; i++)
+               printk("  ");
+}
+
+static void print_lock_class_header(struct lock_class *class, int depth)
+{
+       int bit;
+
+       print_spaces(depth);
+       printk("->");
+       print_lock_name(class);
+       printk(" ops: %lu", class->ops);
+       printk(" {\n");
+
+       for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
+               if (class->usage_mask & (1 << bit)) {
+                       int len = depth;
+
+                       print_spaces(depth);
+                       len += printk("   %s", usage_str[bit]);
+                       len += printk(" at:\n");
+                       print_stack_trace(class->usage_traces + bit, len);
+               }
+       }
+       print_spaces(depth);
+       printk(" }\n");
+
+       print_spaces(depth);
+       printk(" ... key      at: ");
+       print_ip_sym((unsigned long)class->key);
+}
+
+/*
+ * printk all lock dependencies starting at <entry>:
+ */
+static void print_lock_dependencies(struct lock_class *class, int depth)
+{
+       struct lock_list *entry;
+
+       if (DEBUG_LOCKS_WARN_ON(depth >= 20))
+               return;
+
+       print_lock_class_header(class, depth);
+
+       list_for_each_entry(entry, &class->locks_after, entry) {
+               DEBUG_LOCKS_WARN_ON(!entry->class);
+               print_lock_dependencies(entry->class, depth + 1);
+
+               print_spaces(depth);
+               printk(" ... acquired at:\n");
+               print_stack_trace(&entry->trace, 2);
+               printk("\n");
+       }
+}
+
+/*
+ * 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)
+{
+       struct lock_list *entry;
+       /*
+        * Lock not present yet - get a new dependency struct and
+        * add it to the list:
+        */
+       entry = alloc_list_entry();
+       if (!entry)
+               return 0;
+
+       entry->class = this;
+       save_trace(&entry->trace);
+
+       /*
+        * Since we never remove from the dependency list, the list can
+        * be walked lockless by other CPUs, it's only allocation
+        * that must be protected by the spinlock. But this also means
+        * we must make new entries visible only once writes to the
+        * entry become visible - hence the RCU op:
+        */
+       list_add_tail_rcu(&entry->entry, head);
+
+       return 1;
+}
+
+/*
+ * Recursive, forwards-direction lock-dependency checking, used for
+ * both noncyclic checking and for hardirq-unsafe/softirq-unsafe
+ * checking.
+ *
+ * (to keep the stackframe of the recursive functions small we
+ *  use these global variables, and we also mark various helper
+ *  functions as noinline.)
+ */
+static struct held_lock *check_source, *check_target;
+
+/*
+ * Print a dependency chain entry (this is only done when a deadlock
+ * has been detected):
+ */
+static noinline int
+print_circular_bug_entry(struct lock_list *target, unsigned int depth)
+{
+       if (debug_locks_silent)
+               return 0;
+       printk("\n-> #%u", depth);
+       print_lock_name(target->class);
+       printk(":\n");
+       print_stack_trace(&target->trace, 6);
+
+       return 0;
+}
+
+/*
+ * When a circular dependency is detected, print the
+ * header first:
+ */
+static noinline int
+print_circular_bug_header(struct lock_list *entry, unsigned int depth)
+{
+       struct task_struct *curr = current;
+
+       __raw_spin_unlock(&hash_lock);
+       debug_locks_off();
+       if (debug_locks_silent)
+               return 0;
+
+       printk("\n=======================================================\n");
+       printk(  "[ INFO: possible circular locking dependency detected ]\n");
+       printk(  "-------------------------------------------------------\n");
+       printk("%s/%d is trying to acquire lock:\n",
+               curr->comm, curr->pid);
+       print_lock(check_source);
+       printk("\nbut task is already holding lock:\n");
+       print_lock(check_target);
+       printk("\nwhich lock already depends on the new lock.\n\n");
+       printk("\nthe existing dependency chain (in reverse order) is:\n");
+
+       print_circular_bug_entry(entry, depth);
+
+       return 0;
+}
+
+static noinline int print_circular_bug_tail(void)
+{
+       struct task_struct *curr = current;
+       struct lock_list this;
+
+       if (debug_locks_silent)
+               return 0;
+
+       this.class = check_source->class;
+       save_trace(&this.trace);
+       print_circular_bug_entry(&this, 0);
+
+       printk("\nother info that might help us debug this:\n\n");
+       lockdep_print_held_locks(curr);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+
+       return 0;
+}
+
+static int noinline print_infinite_recursion_bug(void)
+{
+       __raw_spin_unlock(&hash_lock);
+       DEBUG_LOCKS_WARN_ON(1);
+
+       return 0;
+}
+
+/*
+ * Prove that the dependency graph starting at <entry> can not
+ * lead to <target>. Print an error and return 0 if it does.
+ */
+static noinline int
+check_noncircular(struct lock_class *source, unsigned int depth)
+{
+       struct lock_list *entry;
+
+       debug_atomic_inc(&nr_cyclic_check_recursions);
+       if (depth > max_recursion_depth)
+               max_recursion_depth = depth;
+       if (depth >= 20)
+               return print_infinite_recursion_bug();
+       /*
+        * Check this lock's dependency list:
+        */
+       list_for_each_entry(entry, &source->locks_after, entry) {
+               if (entry->class == check_target->class)
+                       return print_circular_bug_header(entry, depth+1);
+               debug_atomic_inc(&nr_cyclic_checks);
+               if (!check_noncircular(entry->class, depth+1))
+                       return print_circular_bug_entry(entry, depth+1);
+       }
+       return 1;
+}
+
+static int very_verbose(struct lock_class *class)
+{
+#if VERY_VERBOSE
+       return class_filter(class);
+#endif
+       return 0;
+}
+#ifdef CONFIG_TRACE_IRQFLAGS
+
+/*
+ * Forwards and backwards subgraph searching, for the purposes of
+ * proving that two subgraphs can be connected by a new dependency
+ * without creating any illegal irq-safe -> irq-unsafe lock dependency.
+ */
+static enum lock_usage_bit find_usage_bit;
+static struct lock_class *forwards_match, *backwards_match;
+
+/*
+ * Find a node in the forwards-direction dependency sub-graph starting
+ * at <source> that matches <find_usage_bit>.
+ *
+ * Return 2 if such a node exists in the subgraph, and put that node
+ * into <forwards_match>.
+ *
+ * Return 1 otherwise and keep <forwards_match> unchanged.
+ * Return 0 on error.
+ */
+static noinline int
+find_usage_forwards(struct lock_class *source, unsigned int depth)
+{
+       struct lock_list *entry;
+       int ret;
+
+       if (depth > max_recursion_depth)
+               max_recursion_depth = depth;
+       if (depth >= 20)
+               return print_infinite_recursion_bug();
+
+       debug_atomic_inc(&nr_find_usage_forwards_checks);
+       if (source->usage_mask & (1 << find_usage_bit)) {
+               forwards_match = source;
+               return 2;
+       }
+
+       /*
+        * Check this lock's dependency list:
+        */
+       list_for_each_entry(entry, &source->locks_after, entry) {
+               debug_atomic_inc(&nr_find_usage_forwards_recursions);
+               ret = find_usage_forwards(entry->class, depth+1);
+               if (ret == 2 || ret == 0)
+                       return ret;
+       }
+       return 1;
+}
+
+/*
+ * Find a node in the backwards-direction dependency sub-graph starting
+ * at <source> that matches <find_usage_bit>.
+ *
+ * Return 2 if such a node exists in the subgraph, and put that node
+ * into <backwards_match>.
+ *
+ * Return 1 otherwise and keep <backwards_match> unchanged.
+ * Return 0 on error.
+ */
+static noinline int
+find_usage_backwards(struct lock_class *source, unsigned int depth)
+{
+       struct lock_list *entry;
+       int ret;
+
+       if (depth > max_recursion_depth)
+               max_recursion_depth = depth;
+       if (depth >= 20)
+               return print_infinite_recursion_bug();
+
+       debug_atomic_inc(&nr_find_usage_backwards_checks);
+       if (source->usage_mask & (1 << find_usage_bit)) {
+               backwards_match = source;
+               return 2;
+       }
+
+       /*
+        * Check this lock's dependency list:
+        */
+       list_for_each_entry(entry, &source->locks_before, entry) {
+               debug_atomic_inc(&nr_find_usage_backwards_recursions);
+               ret = find_usage_backwards(entry->class, depth+1);
+               if (ret == 2 || ret == 0)
+                       return ret;
+       }
+       return 1;
+}
+
+static int
+print_bad_irq_dependency(struct task_struct *curr,
+                        struct held_lock *prev,
+                        struct held_lock *next,
+                        enum lock_usage_bit bit1,
+                        enum lock_usage_bit bit2,
+                        const char *irqclass)
+{
+       __raw_spin_unlock(&hash_lock);
+       debug_locks_off();
+       if (debug_locks_silent)
+               return 0;
+
+       printk("\n======================================================\n");
+       printk(  "[ INFO: %s-safe -> %s-unsafe lock order detected ]\n",
+               irqclass, irqclass);
+       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->hardirq_context, hardirq_count() >> HARDIRQ_SHIFT,
+               curr->softirq_context, softirq_count() >> SOFTIRQ_SHIFT,
+               curr->hardirqs_enabled,
+               curr->softirqs_enabled);
+       print_lock(next);
+
+       printk("\nand this task is already holding:\n");
+       print_lock(prev);
+       printk("which would create a new lock dependency:\n");
+       print_lock_name(prev->class);
+       printk(" ->");
+       print_lock_name(next->class);
+       printk("\n");
+
+       printk("\nbut this new dependency connects a %s-irq-safe lock:\n",
+               irqclass);
+       print_lock_name(backwards_match);
+       printk("\n... which became %s-irq-safe at:\n", irqclass);
+
+       print_stack_trace(backwards_match->usage_traces + bit1, 1);
+
+       printk("\nto a %s-irq-unsafe lock:\n", irqclass);
+       print_lock_name(forwards_match);
+       printk("\n... which became %s-irq-unsafe at:\n", irqclass);
+       printk("...");
+
+       print_stack_trace(forwards_match->usage_traces + bit2, 1);
+
+       printk("\nother info that might help us debug this:\n\n");
+       lockdep_print_held_locks(curr);
+
+       printk("\nthe %s-irq-safe lock's dependencies:\n", irqclass);
+       print_lock_dependencies(backwards_match, 0);
+
+       printk("\nthe %s-irq-unsafe lock's dependencies:\n", irqclass);
+       print_lock_dependencies(forwards_match, 0);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+
+       return 0;
+}
+
+static int
+check_usage(struct task_struct *curr, struct held_lock *prev,
+           struct held_lock *next, enum lock_usage_bit bit_backwards,
+           enum lock_usage_bit bit_forwards, const char *irqclass)
+{
+       int ret;
+
+       find_usage_bit = bit_backwards;
+       /* fills in <backwards_match> */
+       ret = find_usage_backwards(prev->class, 0);
+       if (!ret || ret == 1)
+               return ret;
+
+       find_usage_bit = bit_forwards;
+       ret = find_usage_forwards(next->class, 0);
+       if (!ret || ret == 1)
+               return ret;
+       /* ret == 2 */
+       return print_bad_irq_dependency(curr, prev, next,
+                       bit_backwards, bit_forwards, irqclass);
+}
+
+#endif
+
+static int
+print_deadlock_bug(struct task_struct *curr, struct held_lock *prev,
+                  struct held_lock *next)
+{
+       debug_locks_off();
+       __raw_spin_unlock(&hash_lock);
+       if (debug_locks_silent)
+               return 0;
+
+       printk("\n=============================================\n");
+       printk(  "[ INFO: possible recursive locking detected ]\n");
+       printk(  "---------------------------------------------\n");
+       printk("%s/%d is trying to acquire lock:\n",
+               curr->comm, curr->pid);
+       print_lock(next);
+       printk("\nbut task is already holding lock:\n");
+       print_lock(prev);
+
+       printk("\nother info that might help us debug this:\n");
+       lockdep_print_held_locks(curr);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+
+       return 0;
+}
+
+/*
+ * Check whether we are holding such a class already.
+ *
+ * (Note that this has to be done separately, because the graph cannot
+ * detect such classes of deadlocks.)
+ *
+ * Returns: 0 on deadlock detected, 1 on OK, 2 on recursive read
+ */
+static int
+check_deadlock(struct task_struct *curr, struct held_lock *next,
+              struct lockdep_map *next_instance, int read)
+{
+       struct held_lock *prev;
+       int i;
+
+       for (i = 0; i < curr->lockdep_depth; i++) {
+               prev = curr->held_locks + i;
+               if (prev->class != next->class)
+                       continue;
+               /*
+                * Allow read-after-read recursion of the same
+                * lock instance (i.e. read_lock(lock)+read_lock(lock)):
+                */
+               if ((read == 2) && prev->read &&
+                               (prev->instance == next_instance))
+                       return 2;
+               return print_deadlock_bug(curr, prev, next);
+       }
+       return 1;
+}
+
+/*
+ * There was a chain-cache miss, and we are about to add a new dependency
+ * to a previous lock. We recursively validate the following rules:
+ *
+ *  - would the adding of the <prev> -> <next> dependency create a
+ *    circular dependency in the graph? [== circular deadlock]
+ *
+ *  - does the new prev->next dependency connect any hardirq-safe lock
+ *    (in the full backwards-subgraph starting at <prev>) with any
+ *    hardirq-unsafe lock (in the full forwards-subgraph starting at
+ *    <next>)? [== illegal lock inversion with hardirq contexts]
+ *
+ *  - does the new prev->next dependency connect any softirq-safe lock
+ *    (in the full backwards-subgraph starting at <prev>) with any
+ *    softirq-unsafe lock (in the full forwards-subgraph starting at
+ *    <next>)? [== illegal lock inversion with softirq contexts]
+ *
+ * any of these scenarios could lead to a deadlock.
+ *
+ * Then if all the validations pass, we add the forwards and backwards
+ * dependency.
+ */
+static int
+check_prev_add(struct task_struct *curr, struct held_lock *prev,
+              struct held_lock *next)
+{
+       struct lock_list *entry;
+       int ret;
+
+       /*
+        * Prove that the new <prev> -> <next> dependency would not
+        * create a circular dependency in the graph. (We do this by
+        * forward-recursing into the graph starting at <next>, and
+        * checking whether we can reach <prev>.)
+        *
+        * We are using global variables to control the recursion, to
+        * keep the stackframe size of the recursive functions low:
+        */
+       check_source = next;
+       check_target = prev;
+       if (!(check_noncircular(next->class, 0)))
+               return print_circular_bug_tail();
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+       /*
+        * Prove that the new dependency does not connect a hardirq-safe
+        * lock with a hardirq-unsafe lock - to achieve this we search
+        * the backwards-subgraph starting at <prev>, and the
+        * forwards-subgraph starting at <next>:
+        */
+       if (!check_usage(curr, prev, next, LOCK_USED_IN_HARDIRQ,
+                                       LOCK_ENABLED_HARDIRQS, "hard"))
+               return 0;
+
+       /*
+        * Prove that the new dependency does not connect a hardirq-safe-read
+        * lock with a hardirq-unsafe lock - to achieve this we search
+        * the backwards-subgraph starting at <prev>, and the
+        * forwards-subgraph starting at <next>:
+        */
+       if (!check_usage(curr, prev, next, LOCK_USED_IN_HARDIRQ_READ,
+                                       LOCK_ENABLED_HARDIRQS, "hard-read"))
+               return 0;
+
+       /*
+        * Prove that the new dependency does not connect a softirq-safe
+        * lock with a softirq-unsafe lock - to achieve this we search
+        * the backwards-subgraph starting at <prev>, and the
+        * forwards-subgraph starting at <next>:
+        */
+       if (!check_usage(curr, prev, next, LOCK_USED_IN_SOFTIRQ,
+                                       LOCK_ENABLED_SOFTIRQS, "soft"))
+               return 0;
+       /*
+        * Prove that the new dependency does not connect a softirq-safe-read
+        * lock with a softirq-unsafe lock - to achieve this we search
+        * the backwards-subgraph starting at <prev>, and the
+        * forwards-subgraph starting at <next>:
+        */
+       if (!check_usage(curr, prev, next, LOCK_USED_IN_SOFTIRQ_READ,
+                                       LOCK_ENABLED_SOFTIRQS, "soft"))
+               return 0;
+#endif
+       /*
+        * For recursive read-locks we do all the dependency checks,
+        * but we dont store read-triggered dependencies (only
+        * write-triggered dependencies). This ensures that only the
+        * write-side dependencies matter, and that if for example a
+        * write-lock never takes any other locks, then the reads are
+        * equivalent to a NOP.
+        */
+       if (next->read == 2 || prev->read == 2)
+               return 1;
+       /*
+        * Is the <prev> -> <next> dependency already present?
+        *
+        * (this may occur even though this is a new chain: consider
+        *  e.g. the L1 -> L2 -> L3 -> L4 and the L5 -> L1 -> L2 -> L3
+        *  chains - the second one will be new, but L1 already has
+        *  L2 added to its dependency list, due to the first chain.)
+        */
+       list_for_each_entry(entry, &prev->class->locks_after, entry) {
+               if (entry->class == next->class)
+                       return 2;
+       }
+
+       /*
+        * Ok, all validations passed, add the new lock
+        * to the previous lock's dependency list:
+        */
+       ret = add_lock_to_list(prev->class, next->class,
+                              &prev->class->locks_after, next->acquire_ip);
+       if (!ret)
+               return 0;
+       /*
+        * Return value of 2 signals 'dependency already added',
+        * in that case we dont have to add the backlink either.
+        */
+       if (ret == 2)
+               return 2;
+       ret = add_lock_to_list(next->class, prev->class,
+                              &next->class->locks_before, next->acquire_ip);
+
+       /*
+        * Debugging printouts:
+        */
+       if (verbose(prev->class) || verbose(next->class)) {
+               __raw_spin_unlock(&hash_lock);
+               printk("\n new dependency: ");
+               print_lock_name(prev->class);
+               printk(" => ");
+               print_lock_name(next->class);
+               printk("\n");
+               dump_stack();
+               __raw_spin_lock(&hash_lock);
+       }
+       return 1;
+}
+
+/*
+ * Add the dependency to all directly-previous locks that are 'relevant'.
+ * The ones that are relevant are (in increasing distance from curr):
+ * all consecutive trylock entries and the final non-trylock entry - or
+ * the end of this context's lock-chain - whichever comes first.
+ */
+static int
+check_prevs_add(struct task_struct *curr, struct held_lock *next)
+{
+       int depth = curr->lockdep_depth;
+       struct held_lock *hlock;
+
+       /*
+        * Debugging checks.
+        *
+        * Depth must not be zero for a non-head lock:
+        */
+       if (!depth)
+               goto out_bug;
+       /*
+        * At least two relevant locks must exist for this
+        * to be a head:
+        */
+       if (curr->held_locks[depth].irq_context !=
+                       curr->held_locks[depth-1].irq_context)
+               goto out_bug;
+
+       for (;;) {
+               hlock = curr->held_locks + depth-1;
+               /*
+                * Only non-recursive-read entries get new dependencies
+                * added:
+                */
+               if (hlock->read != 2) {
+                       check_prev_add(curr, hlock, next);
+                       /*
+                        * Stop after the first non-trylock entry,
+                        * as non-trylock entries have added their
+                        * own direct dependencies already, so this
+                        * lock is connected to them indirectly:
+                        */
+                       if (!hlock->trylock)
+                               break;
+               }
+               depth--;
+               /*
+                * End of lock-stack?
+                */
+               if (!depth)
+                       break;
+               /*
+                * Stop the search if we cross into another context:
+                */
+               if (curr->held_locks[depth].irq_context !=
+                               curr->held_locks[depth-1].irq_context)
+                       break;
+       }
+       return 1;
+out_bug:
+       __raw_spin_unlock(&hash_lock);
+       DEBUG_LOCKS_WARN_ON(1);
+
+       return 0;
+}
+
+
+/*
+ * Is this the address of a static object:
+ */
+static int static_obj(void *obj)
+{
+       unsigned long start = (unsigned long) &_stext,
+                     end   = (unsigned long) &_end,
+                     addr  = (unsigned long) obj;
+#ifdef CONFIG_SMP
+       int i;
+#endif
+
+       /*
+        * static variable?
+        */
+       if ((addr >= start) && (addr < end))
+               return 1;
+
+#ifdef CONFIG_SMP
+       /*
+        * percpu var?
+        */
+       for_each_possible_cpu(i) {
+               start = (unsigned long) &__per_cpu_start + per_cpu_offset(i);
+               end   = (unsigned long) &__per_cpu_end   + per_cpu_offset(i);
+
+               if ((addr >= start) && (addr < end))
+                       return 1;
+       }
+#endif
+
+       /*
+        * module var?
+        */
+       return is_module_address(addr);
+}
+
+/*
+ * To make lock name printouts unique, we calculate a unique
+ * class->name_version generation counter:
+ */
+static int count_matching_names(struct lock_class *new_class)
+{
+       struct lock_class *class;
+       int count = 0;
+
+       if (!new_class->name)
+               return 0;
+
+       list_for_each_entry(class, &all_lock_classes, lock_entry) {
+               if (new_class->key - new_class->subclass == class->key)
+                       return class->name_version;
+               if (class->name && !strcmp(class->name, new_class->name))
+                       count = max(count, class->name_version);
+       }
+
+       return count + 1;
+}
+
+extern void __error_too_big_MAX_LOCKDEP_SUBCLASSES(void);
+
+/*
+ * Register a lock's class in the hash-table, if the class is not present
+ * yet. Otherwise we look it up. We cache the result in the lock object
+ * itself, so actual lookup of the hash should be once per lock object.
+ */
+static inline struct lock_class *
+register_lock_class(struct lockdep_map *lock, unsigned int subclass)
+{
+       struct lockdep_subclass_key *key;
+       struct list_head *hash_head;
+       struct lock_class *class;
+
+#ifdef CONFIG_DEBUG_LOCKDEP
+       /*
+        * If the architecture calls into lockdep before initializing
+        * the hashes then we'll warn about it later. (we cannot printk
+        * right now)
+        */
+       if (unlikely(!lockdep_initialized)) {
+               lockdep_init();
+               lockdep_init_error = 1;
+       }
+#endif
+
+       /*
+        * Static locks do not have their class-keys yet - for them the key
+        * is the lock object itself:
+        */
+       if (unlikely(!lock->key))
+               lock->key = (void *)lock;
+
+       /*
+        * NOTE: the class-key must be unique. For dynamic locks, a static
+        * lock_class_key variable is passed in through the mutex_init()
+        * (or spin_lock_init()) call - which acts as the key. For static
+        * locks we use the lock object itself as the key.
+        */
+       if (sizeof(struct lock_class_key) > sizeof(struct lock_class))
+               __error_too_big_MAX_LOCKDEP_SUBCLASSES();
+
+       key = lock->key->subkeys + subclass;
+
+       hash_head = classhashentry(key);
+
+       /*
+        * 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)
+                       goto out_set;
+
+       /*
+        * Debug-check: all keys must be persistent!
+        */
+       if (!static_obj(lock->key)) {
+               debug_locks_off();
+               printk("INFO: trying to register non-static key.\n");
+               printk("the code is fine but needs lockdep annotation.\n");
+               printk("turning off the locking correctness validator.\n");
+               dump_stack();
+
+               return NULL;
+       }
+
+       __raw_spin_lock(&hash_lock);
+       /*
+        * We have to do the hash-walk again, to avoid races
+        * with another CPU:
+        */
+       list_for_each_entry(class, hash_head, hash_entry)
+               if (class->key == key)
+                       goto out_unlock_set;
+       /*
+        * Allocate a new key from the static array, and add it to
+        * the hash:
+        */
+       if (nr_lock_classes >= MAX_LOCKDEP_KEYS) {
+               __raw_spin_unlock(&hash_lock);
+               debug_locks_off();
+               printk("BUG: MAX_LOCKDEP_KEYS too low!\n");
+               printk("turning off the locking correctness validator.\n");
+               return NULL;
+       }
+       class = lock_classes + nr_lock_classes++;
+       debug_atomic_inc(&nr_unused_locks);
+       class->key = key;
+       class->name = lock->name;
+       class->subclass = subclass;
+       INIT_LIST_HEAD(&class->lock_entry);
+       INIT_LIST_HEAD(&class->locks_before);
+       INIT_LIST_HEAD(&class->locks_after);
+       class->name_version = count_matching_names(class);
+       /*
+        * We use RCU's safe list-add method to make
+        * parallel walking of the hash-list safe:
+        */
+       list_add_tail_rcu(&class->hash_entry, hash_head);
+
+       if (verbose(class)) {
+               __raw_spin_unlock(&hash_lock);
+               printk("\nnew class %p: %s", class->key, class->name);
+               if (class->name_version > 1)
+                       printk("#%d", class->name_version);
+               printk("\n");
+               dump_stack();
+               __raw_spin_lock(&hash_lock);
+       }
+out_unlock_set:
+       __raw_spin_unlock(&hash_lock);
+
+out_set:
+       lock->class[subclass] = class;
+
+       DEBUG_LOCKS_WARN_ON(class->subclass != subclass);
+
+       return class;
+}
+
+/*
+ * Look up a dependency chain. If the key is not present yet then
+ * add it and return 0 - in this case the new dependency chain is
+ * validated. If the key is already hashed, return 1.
+ */
+static inline int lookup_chain_cache(u64 chain_key)
+{
+       struct list_head *hash_head = chainhashentry(chain_key);
+       struct lock_chain *chain;
+
+       DEBUG_LOCKS_WARN_ON(!irqs_disabled());
+       /*
+        * We can walk it lock-free, because entries only get added
+        * to the hash:
+        */
+       list_for_each_entry(chain, hash_head, entry) {
+               if (chain->chain_key == chain_key) {
+cache_hit:
+                       debug_atomic_inc(&chain_lookup_hits);
+                       /*
+                        * In the debugging case, force redundant checking
+                        * by returning 1:
+                        */
+#ifdef CONFIG_DEBUG_LOCKDEP
+                       __raw_spin_lock(&hash_lock);
+                       return 1;
+#endif
+                       return 0;
+               }
+       }
+       /*
+        * Allocate a new chain entry from the static array, and add
+        * it to the hash:
+        */
+       __raw_spin_lock(&hash_lock);
+       /*
+        * We have to walk the chain again locked - to avoid duplicates:
+        */
+       list_for_each_entry(chain, hash_head, entry) {
+               if (chain->chain_key == chain_key) {
+                       __raw_spin_unlock(&hash_lock);
+                       goto cache_hit;
+               }
+       }
+       if (unlikely(nr_lock_chains >= MAX_LOCKDEP_CHAINS)) {
+               __raw_spin_unlock(&hash_lock);
+               debug_locks_off();
+               printk("BUG: MAX_LOCKDEP_CHAINS too low!\n");
+               printk("turning off the locking correctness validator.\n");
+               return 0;
+       }
+       chain = lock_chains + nr_lock_chains++;
+       chain->chain_key = chain_key;
+       list_add_tail_rcu(&chain->entry, hash_head);
+       debug_atomic_inc(&chain_lookup_misses);
+#ifdef CONFIG_TRACE_IRQFLAGS
+       if (current->hardirq_context)
+               nr_hardirq_chains++;
+       else {
+               if (current->softirq_context)
+                       nr_softirq_chains++;
+               else
+                       nr_process_chains++;
+       }
+#else
+       nr_process_chains++;
+#endif
+
+       return 1;
+}
+
+/*
+ * We are building curr_chain_key incrementally, so double-check
+ * it from scratch, to make sure that it's done correctly:
+ */
+static void check_chain_key(struct task_struct *curr)
+{
+#ifdef CONFIG_DEBUG_LOCKDEP
+       struct held_lock *hlock, *prev_hlock = NULL;
+       unsigned int i, id;
+       u64 chain_key = 0;
+
+       for (i = 0; i < curr->lockdep_depth; i++) {
+               hlock = curr->held_locks + i;
+               if (chain_key != hlock->prev_chain_key) {
+                       debug_locks_off();
+                       printk("hm#1, depth: %u [%u], %016Lx != %016Lx\n",
+                               curr->lockdep_depth, i,
+                               (unsigned long long)chain_key,
+                               (unsigned long long)hlock->prev_chain_key);
+                       WARN_ON(1);
+                       return;
+               }
+               id = hlock->class - lock_classes;
+               DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS);
+               if (prev_hlock && (prev_hlock->irq_context !=
+                                                       hlock->irq_context))
+                       chain_key = 0;
+               chain_key = iterate_chain_key(chain_key, id);
+               prev_hlock = hlock;
+       }
+       if (chain_key != curr->curr_chain_key) {
+               debug_locks_off();
+               printk("hm#2, depth: %u [%u], %016Lx != %016Lx\n",
+                       curr->lockdep_depth, i,
+                       (unsigned long long)chain_key,
+                       (unsigned long long)curr->curr_chain_key);
+               WARN_ON(1);
+       }
+#endif
+}
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+
+/*
+ * print irq inversion bug:
+ */
+static int
+print_irq_inversion_bug(struct task_struct *curr, struct lock_class *other,
+                       struct held_lock *this, int forwards,
+                       const char *irqclass)
+{
+       __raw_spin_unlock(&hash_lock);
+       debug_locks_off();
+       if (debug_locks_silent)
+               return 0;
+
+       printk("\n=========================================================\n");
+       printk(  "[ INFO: possible irq lock inversion dependency detected ]\n");
+       printk(  "---------------------------------------------------------\n");
+       printk("%s/%d just changed the state of lock:\n",
+               curr->comm, curr->pid);
+       print_lock(this);
+       if (forwards)
+               printk("but this lock took another, %s-irq-unsafe lock in the past:\n", irqclass);
+       else
+               printk("but this lock was taken by another, %s-irq-safe lock in the past:\n", irqclass);
+       print_lock_name(other);
+       printk("\n\nand interrupts could create inverse lock ordering between them.\n\n");
+
+       printk("\nother info that might help us debug this:\n");
+       lockdep_print_held_locks(curr);
+
+       printk("\nthe first lock's dependencies:\n");
+       print_lock_dependencies(this->class, 0);
+
+       printk("\nthe second lock's dependencies:\n");
+       print_lock_dependencies(other, 0);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+
+       return 0;
+}
+
+/*
+ * Prove that in the forwards-direction subgraph starting at <this>
+ * there is no lock matching <mask>:
+ */
+static int
+check_usage_forwards(struct task_struct *curr, struct held_lock *this,
+                    enum lock_usage_bit bit, const char *irqclass)
+{
+       int ret;
+
+       find_usage_bit = bit;
+       /* fills in <forwards_match> */
+       ret = find_usage_forwards(this->class, 0);
+       if (!ret || ret == 1)
+               return ret;
+
+       return print_irq_inversion_bug(curr, forwards_match, this, 1, irqclass);
+}
+
+/*
+ * Prove that in the backwards-direction subgraph starting at <this>
+ * there is no lock matching <mask>:
+ */
+static int
+check_usage_backwards(struct task_struct *curr, struct held_lock *this,
+                     enum lock_usage_bit bit, const char *irqclass)
+{
+       int ret;
+
+       find_usage_bit = bit;
+       /* fills in <backwards_match> */
+       ret = find_usage_backwards(this->class, 0);
+       if (!ret || ret == 1)
+               return ret;
+
+       return print_irq_inversion_bug(curr, backwards_match, this, 0, irqclass);
+}
+
+static inline void print_irqtrace_events(struct task_struct *curr)
+{
+       printk("irq event stamp: %u\n", curr->irq_events);
+       printk("hardirqs last  enabled at (%u): ", curr->hardirq_enable_event);
+       print_ip_sym(curr->hardirq_enable_ip);
+       printk("hardirqs last disabled at (%u): ", curr->hardirq_disable_event);
+       print_ip_sym(curr->hardirq_disable_ip);
+       printk("softirqs last  enabled at (%u): ", curr->softirq_enable_event);
+       print_ip_sym(curr->softirq_enable_ip);
+       printk("softirqs last disabled at (%u): ", curr->softirq_disable_event);
+       print_ip_sym(curr->softirq_disable_ip);
+}
+
+#else
+static inline void print_irqtrace_events(struct task_struct *curr)
+{
+}
+#endif
+
+static int
+print_usage_bug(struct task_struct *curr, struct held_lock *this,
+               enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
+{
+       __raw_spin_unlock(&hash_lock);
+       debug_locks_off();
+       if (debug_locks_silent)
+               return 0;
+
+       printk("\n=================================\n");
+       printk(  "[ INFO: inconsistent lock state ]\n");
+       printk(  "---------------------------------\n");
+
+       printk("inconsistent {%s} -> {%s} usage.\n",
+               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,
+               trace_hardirq_context(curr), hardirq_count() >> HARDIRQ_SHIFT,
+               trace_softirq_context(curr), softirq_count() >> SOFTIRQ_SHIFT,
+               trace_hardirqs_enabled(curr),
+               trace_softirqs_enabled(curr));
+       print_lock(this);
+
+       printk("{%s} state was registered at:\n", usage_str[prev_bit]);
+       print_stack_trace(this->class->usage_traces + prev_bit, 1);
+
+       print_irqtrace_events(curr);
+       printk("\nother info that might help us debug this:\n");
+       lockdep_print_held_locks(curr);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+
+       return 0;
+}
+
+/*
+ * Print out an error if an invalid bit is set:
+ */
+static inline int
+valid_state(struct task_struct *curr, struct held_lock *this,
+           enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
+{
+       if (unlikely(this->class->usage_mask & (1 << bad_bit)))
+               return print_usage_bug(curr, this, bad_bit, new_bit);
+       return 1;
+}
+
+#define STRICT_READ_CHECKS     1
+
+/*
+ * Mark a lock with a usage bit, and validate the state transition:
+ */
+static int mark_lock(struct task_struct *curr, struct held_lock *this,
+                    enum lock_usage_bit new_bit, unsigned long ip)
+{
+       unsigned int new_mask = 1 << new_bit, ret = 1;
+
+       /*
+        * If already set then do not dirty the cacheline,
+        * nor do any checks:
+        */
+       if (likely(this->class->usage_mask & new_mask))
+               return 1;
+
+       __raw_spin_lock(&hash_lock);
+       /*
+        * Make sure we didnt race:
+        */
+       if (unlikely(this->class->usage_mask & new_mask)) {
+               __raw_spin_unlock(&hash_lock);
+               return 1;
+       }
+
+       this->class->usage_mask |= new_mask;
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+       if (new_bit == LOCK_ENABLED_HARDIRQS ||
+                       new_bit == LOCK_ENABLED_HARDIRQS_READ)
+               ip = curr->hardirq_enable_ip;
+       else if (new_bit == LOCK_ENABLED_SOFTIRQS ||
+                       new_bit == LOCK_ENABLED_SOFTIRQS_READ)
+               ip = curr->softirq_enable_ip;
+#endif
+       if (!save_trace(this->class->usage_traces + new_bit))
+               return 0;
+
+       switch (new_bit) {
+#ifdef CONFIG_TRACE_IRQFLAGS
+       case LOCK_USED_IN_HARDIRQ:
+               if (!valid_state(curr, this, new_bit, LOCK_ENABLED_HARDIRQS))
+                       return 0;
+               if (!valid_state(curr, this, new_bit,
+                                LOCK_ENABLED_HARDIRQS_READ))
+                       return 0;
+               /*
+                * just marked it hardirq-safe, check that this lock
+                * took no hardirq-unsafe lock in the past:
+                */
+               if (!check_usage_forwards(curr, this,
+                                         LOCK_ENABLED_HARDIRQS, "hard"))
+                       return 0;
+#if STRICT_READ_CHECKS
+               /*
+                * just marked it hardirq-safe, check that this lock
+                * took no hardirq-unsafe-read lock in the past:
+                */
+               if (!check_usage_forwards(curr, this,
+                               LOCK_ENABLED_HARDIRQS_READ, "hard-read"))
+                       return 0;
+#endif
+               if (hardirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_USED_IN_SOFTIRQ:
+               if (!valid_state(curr, this, new_bit, LOCK_ENABLED_SOFTIRQS))
+                       return 0;
+               if (!valid_state(curr, this, new_bit,
+                                LOCK_ENABLED_SOFTIRQS_READ))
+                       return 0;
+               /*
+                * just marked it softirq-safe, check that this lock
+                * took no softirq-unsafe lock in the past:
+                */
+               if (!check_usage_forwards(curr, this,
+                                         LOCK_ENABLED_SOFTIRQS, "soft"))
+                       return 0;
+#if STRICT_READ_CHECKS
+               /*
+                * just marked it softirq-safe, check that this lock
+                * took no softirq-unsafe-read lock in the past:
+                */
+               if (!check_usage_forwards(curr, this,
+                               LOCK_ENABLED_SOFTIRQS_READ, "soft-read"))
+                       return 0;
+#endif
+               if (softirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_USED_IN_HARDIRQ_READ:
+               if (!valid_state(curr, this, new_bit, LOCK_ENABLED_HARDIRQS))
+                       return 0;
+               /*
+                * just marked it hardirq-read-safe, check that this lock
+                * took no hardirq-unsafe lock in the past:
+                */
+               if (!check_usage_forwards(curr, this,
+                                         LOCK_ENABLED_HARDIRQS, "hard"))
+                       return 0;
+               if (hardirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_USED_IN_SOFTIRQ_READ:
+               if (!valid_state(curr, this, new_bit, LOCK_ENABLED_SOFTIRQS))
+                       return 0;
+               /*
+                * just marked it softirq-read-safe, check that this lock
+                * took no softirq-unsafe lock in the past:
+                */
+               if (!check_usage_forwards(curr, this,
+                                         LOCK_ENABLED_SOFTIRQS, "soft"))
+                       return 0;
+               if (softirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_ENABLED_HARDIRQS:
+               if (!valid_state(curr, this, new_bit, LOCK_USED_IN_HARDIRQ))
+                       return 0;
+               if (!valid_state(curr, this, new_bit,
+                                LOCK_USED_IN_HARDIRQ_READ))
+                       return 0;
+               /*
+                * just marked it hardirq-unsafe, check that no hardirq-safe
+                * lock in the system ever took it in the past:
+                */
+               if (!check_usage_backwards(curr, this,
+                                          LOCK_USED_IN_HARDIRQ, "hard"))
+                       return 0;
+#if STRICT_READ_CHECKS
+               /*
+                * just marked it hardirq-unsafe, check that no
+                * hardirq-safe-read lock in the system ever took
+                * it in the past:
+                */
+               if (!check_usage_backwards(curr, this,
+                                  LOCK_USED_IN_HARDIRQ_READ, "hard-read"))
+                       return 0;
+#endif
+               if (hardirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_ENABLED_SOFTIRQS:
+               if (!valid_state(curr, this, new_bit, LOCK_USED_IN_SOFTIRQ))
+                       return 0;
+               if (!valid_state(curr, this, new_bit,
+                                LOCK_USED_IN_SOFTIRQ_READ))
+                       return 0;
+               /*
+                * just marked it softirq-unsafe, check that no softirq-safe
+                * lock in the system ever took it in the past:
+                */
+               if (!check_usage_backwards(curr, this,
+                                          LOCK_USED_IN_SOFTIRQ, "soft"))
+                       return 0;
+#if STRICT_READ_CHECKS
+               /*
+                * just marked it softirq-unsafe, check that no
+                * softirq-safe-read lock in the system ever took
+                * it in the past:
+                */
+               if (!check_usage_backwards(curr, this,
+                                  LOCK_USED_IN_SOFTIRQ_READ, "soft-read"))
+                       return 0;
+#endif
+               if (softirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_ENABLED_HARDIRQS_READ:
+               if (!valid_state(curr, this, new_bit, LOCK_USED_IN_HARDIRQ))
+                       return 0;
+#if STRICT_READ_CHECKS
+               /*
+                * just marked it hardirq-read-unsafe, check that no
+                * hardirq-safe lock in the system ever took it in the past:
+                */
+               if (!check_usage_backwards(curr, this,
+                                          LOCK_USED_IN_HARDIRQ, "hard"))
+                       return 0;
+#endif
+               if (hardirq_verbose(this->class))
+                       ret = 2;
+               break;
+       case LOCK_ENABLED_SOFTIRQS_READ:
+               if (!valid_state(curr, this, new_bit, LOCK_USED_IN_SOFTIRQ))
+                       return 0;
+#if STRICT_READ_CHECKS
+               /*
+                * just marked it softirq-read-unsafe, check that no
+                * softirq-safe lock in the system ever took it in the past:
+                */
+               if (!check_usage_backwards(curr, this,
+                                          LOCK_USED_IN_SOFTIRQ, "soft"))
+                       return 0;
+#endif
+               if (softirq_verbose(this->class))
+                       ret = 2;
+               break;
+#endif
+       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:
+               debug_locks_off();
+               WARN_ON(1);
+               return 0;
+       }
+
+       __raw_spin_unlock(&hash_lock);
+
+       /*
+        * We must printk outside of the hash_lock:
+        */
+       if (ret == 2) {
+               printk("\nmarked lock as {%s}:\n", usage_str[new_bit]);
+               print_lock(this);
+               print_irqtrace_events(curr);
+               dump_stack();
+       }
+
+       return ret;
+}
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+/*
+ * Mark all held locks with a usage bit:
+ */
+static int
+mark_held_locks(struct task_struct *curr, int hardirq, unsigned long ip)
+{
+       enum lock_usage_bit usage_bit;
+       struct held_lock *hlock;
+       int i;
+
+       for (i = 0; i < curr->lockdep_depth; i++) {
+               hlock = curr->held_locks + i;
+
+               if (hardirq) {
+                       if (hlock->read)
+                               usage_bit = LOCK_ENABLED_HARDIRQS_READ;
+                       else
+                               usage_bit = LOCK_ENABLED_HARDIRQS;
+               } else {
+                       if (hlock->read)
+                               usage_bit = LOCK_ENABLED_SOFTIRQS_READ;
+                       else
+                               usage_bit = LOCK_ENABLED_SOFTIRQS;
+               }
+               if (!mark_lock(curr, hlock, usage_bit, ip))
+                       return 0;
+       }
+
+       return 1;
+}
+
+/*
+ * Debugging helper: via this flag we know that we are in
+ * 'early bootup code', and will warn about any invalid irqs-on event:
+ */
+static int early_boot_irqs_enabled;
+
+void early_boot_irqs_off(void)
+{
+       early_boot_irqs_enabled = 0;
+}
+
+void early_boot_irqs_on(void)
+{
+       early_boot_irqs_enabled = 1;
+}
+
+/*
+ * Hardirqs will be enabled:
+ */
+void trace_hardirqs_on(void)
+{
+       struct task_struct *curr = current;
+       unsigned long ip;
+
+       if (unlikely(!debug_locks || current->lockdep_recursion))
+               return;
+
+       if (DEBUG_LOCKS_WARN_ON(unlikely(!early_boot_irqs_enabled)))
+               return;
+
+       if (unlikely(curr->hardirqs_enabled)) {
+               debug_atomic_inc(&redundant_hardirqs_on);
+               return;
+       }
+       /* we'll do an OFF -> ON transition: */
+       curr->hardirqs_enabled = 1;
+       ip = (unsigned long) __builtin_return_address(0);
+
+       if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
+               return;
+       if (DEBUG_LOCKS_WARN_ON(current->hardirq_context))
+               return;
+       /*
+        * We are going to turn hardirqs on, so set the
+        * usage bit for all held locks:
+        */
+       if (!mark_held_locks(curr, 1, ip))
+               return;
+       /*
+        * If we have softirqs enabled, then set the usage
+        * bit for all held locks. (disabled hardirqs prevented
+        * this bit from being set before)
+        */
+       if (curr->softirqs_enabled)
+               if (!mark_held_locks(curr, 0, ip))
+                       return;
+
+       curr->hardirq_enable_ip = ip;
+       curr->hardirq_enable_event = ++curr->irq_events;
+       debug_atomic_inc(&hardirqs_on_events);
+}
+
+EXPORT_SYMBOL(trace_hardirqs_on);
+
+/*
+ * Hardirqs were disabled:
+ */
+void trace_hardirqs_off(void)
+{
+       struct task_struct *curr = current;
+
+       if (unlikely(!debug_locks || current->lockdep_recursion))
+               return;
+
+       if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
+               return;
+
+       if (curr->hardirqs_enabled) {
+               /*
+                * We have done an ON -> OFF transition:
+                */
+               curr->hardirqs_enabled = 0;
+               curr->hardirq_disable_ip = _RET_IP_;
+               curr->hardirq_disable_event = ++curr->irq_events;
+               debug_atomic_inc(&hardirqs_off_events);
+       } else
+               debug_atomic_inc(&redundant_hardirqs_off);
+}
+
+EXPORT_SYMBOL(trace_hardirqs_off);
+
+/*
+ * Softirqs will be enabled:
+ */
+void trace_softirqs_on(unsigned long ip)
+{
+       struct task_struct *curr = current;
+
+       if (unlikely(!debug_locks))
+               return;
+
+       if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
+               return;
+
+       if (curr->softirqs_enabled) {
+               debug_atomic_inc(&redundant_softirqs_on);
+               return;
+       }
+
+       /*
+        * We'll do an OFF -> ON transition:
+        */
+       curr->softirqs_enabled = 1;
+       curr->softirq_enable_ip = ip;
+       curr->softirq_enable_event = ++curr->irq_events;
+       debug_atomic_inc(&softirqs_on_events);
+       /*
+        * We are going to turn softirqs on, so set the
+        * usage bit for all held locks, if hardirqs are
+        * enabled too:
+        */
+       if (curr->hardirqs_enabled)
+               mark_held_locks(curr, 0, ip);
+}
+
+/*
+ * Softirqs were disabled:
+ */
+void trace_softirqs_off(unsigned long ip)
+{
+       struct task_struct *curr = current;
+
+       if (unlikely(!debug_locks))
+               return;
+
+       if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
+               return;
+
+       if (curr->softirqs_enabled) {
+               /*
+                * We have done an ON -> OFF transition:
+                */
+               curr->softirqs_enabled = 0;
+               curr->softirq_disable_ip = ip;
+               curr->softirq_disable_event = ++curr->irq_events;
+               debug_atomic_inc(&softirqs_off_events);
+               DEBUG_LOCKS_WARN_ON(!softirq_count());
+       } else
+               debug_atomic_inc(&redundant_softirqs_off);
+}
+
+#endif
+
+/*
+ * Initialize a lock instance's lock-class mapping info:
+ */
+void lockdep_init_map(struct lockdep_map *lock, const char *name,
+                     struct lock_class_key *key)
+{
+       if (unlikely(!debug_locks))
+               return;
+
+       if (DEBUG_LOCKS_WARN_ON(!key))
+               return;
+       if (DEBUG_LOCKS_WARN_ON(!name))
+               return;
+       /*
+        * Sanity check, the lock-class key must be persistent:
+        */
+       if (!static_obj(key)) {
+               printk("BUG: key %p not in .data!\n", key);
+               DEBUG_LOCKS_WARN_ON(1);
+               return;
+       }
+       lock->name = name;
+       lock->key = key;
+       memset(lock->class, 0, sizeof(lock->class[0])*MAX_LOCKDEP_SUBCLASSES);
+}
+
+EXPORT_SYMBOL_GPL(lockdep_init_map);
+
+/*
+ * This gets called for every mutex_lock*()/spin_lock*() operation.
+ * We maintain the dependency maps and validate the locking attempt:
+ */
+static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
+                         int trylock, int read, int check, int hardirqs_off,
+                         unsigned long ip)
+{
+       struct task_struct *curr = current;
+       struct held_lock *hlock;
+       struct lock_class *class;
+       unsigned int depth, id;
+       int chain_head = 0;
+       u64 chain_key;
+
+       if (unlikely(!debug_locks))
+               return 0;
+
+       if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
+               return 0;
+
+       if (unlikely(subclass >= MAX_LOCKDEP_SUBCLASSES)) {
+               debug_locks_off();
+               printk("BUG: MAX_LOCKDEP_SUBCLASSES too low!\n");
+               printk("turning off the locking correctness validator.\n");
+               return 0;
+       }
+
+       class = lock->class[subclass];
+       /* not cached yet? */
+       if (unlikely(!class)) {
+               class = register_lock_class(lock, subclass);
+               if (!class)
+                       return 0;
+       }
+       debug_atomic_inc((atomic_t *)&class->ops);
+       if (very_verbose(class)) {
+               printk("\nacquire class [%p] %s", class->key, class->name);
+               if (class->name_version > 1)
+                       printk("#%d", class->name_version);
+               printk("\n");
+               dump_stack();
+       }
+
+       /*
+        * Add the lock to the list of currently held locks.
+        * (we dont increase the depth just yet, up until the
+        * dependency checks are done)
+        */
+       depth = curr->lockdep_depth;
+       if (DEBUG_LOCKS_WARN_ON(depth >= MAX_LOCK_DEPTH))
+               return 0;
+
+       hlock = curr->held_locks + depth;
+
+       hlock->class = class;
+       hlock->acquire_ip = ip;
+       hlock->instance = lock;
+       hlock->trylock = trylock;
+       hlock->read = read;
+       hlock->check = check;
+       hlock->hardirqs_off = hardirqs_off;
+
+       if (check != 2)
+               goto out_calc_hash;
+#ifdef CONFIG_TRACE_IRQFLAGS
+       /*
+        * If non-trylock use in a hardirq or softirq context, then
+        * mark the lock as used in these contexts:
+        */
+       if (!trylock) {
+               if (read) {
+                       if (curr->hardirq_context)
+                               if (!mark_lock(curr, hlock,
+                                               LOCK_USED_IN_HARDIRQ_READ, ip))
+                                       return 0;
+                       if (curr->softirq_context)
+                               if (!mark_lock(curr, hlock,
+                                               LOCK_USED_IN_SOFTIRQ_READ, ip))
+                                       return 0;
+               } else {
+                       if (curr->hardirq_context)
+                               if (!mark_lock(curr, hlock, LOCK_USED_IN_HARDIRQ, ip))
+                                       return 0;
+                       if (curr->softirq_context)
+                               if (!mark_lock(curr, hlock, LOCK_USED_IN_SOFTIRQ, ip))
+                                       return 0;
+               }
+       }
+       if (!hardirqs_off) {
+               if (read) {
+                       if (!mark_lock(curr, hlock,
+                                       LOCK_ENABLED_HARDIRQS_READ, ip))
+                               return 0;
+                       if (curr->softirqs_enabled)
+                               if (!mark_lock(curr, hlock,
+                                               LOCK_ENABLED_SOFTIRQS_READ, ip))
+                                       return 0;
+               } else {
+                       if (!mark_lock(curr, hlock,
+                                       LOCK_ENABLED_HARDIRQS, ip))
+                               return 0;
+                       if (curr->softirqs_enabled)
+                               if (!mark_lock(curr, hlock,
+                                               LOCK_ENABLED_SOFTIRQS, ip))
+                                       return 0;
+               }
+       }
+#endif
+       /* mark it as used: */
+       if (!mark_lock(curr, hlock, LOCK_USED, ip))
+               return 0;
+out_calc_hash:
+       /*
+        * Calculate the chain hash: it's the combined has 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
+        * results.
+        *
+        * The 'key ID' is what is the most compact key value to drive
+        * the hash, not class->key.
+        */
+       id = class - lock_classes;
+       if (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP_KEYS))
+               return 0;
+
+       chain_key = curr->curr_chain_key;
+       if (!depth) {
+               if (DEBUG_LOCKS_WARN_ON(chain_key != 0))
+                       return 0;
+               chain_head = 1;
+       }
+
+       hlock->prev_chain_key = chain_key;
+
+#ifdef CONFIG_TRACE_IRQFLAGS
+       /*
+        * Keep track of points where we cross into an interrupt context:
+        */
+       hlock->irq_context = 2*(curr->hardirq_context ? 1 : 0) +
+                               curr->softirq_context;
+       if (depth) {
+               struct held_lock *prev_hlock;
+
+               prev_hlock = curr->held_locks + depth-1;
+               /*
+                * If we cross into another context, reset the
+                * hash key (this also prevents the checking and the
+                * adding of the dependency to 'prev'):
+                */
+               if (prev_hlock->irq_context != hlock->irq_context) {
+                       chain_key = 0;
+                       chain_head = 1;
+               }
+       }
+#endif
+       chain_key = iterate_chain_key(chain_key, id);
+       curr->curr_chain_key = chain_key;
+
+       /*
+        * Trylock needs to maintain the stack of held locks, but it
+        * does not add new dependencies, because trylock can be done
+        * in any order.
+        *
+        * We look up the chain_key and do the O(N^2) check and update of
+        * the dependencies only if this is a new dependency chain.
+        * (If lookup_chain_cache() returns with 1 it acquires
+        * hash_lock for us)
+        */
+       if (!trylock && (check == 2) && lookup_chain_cache(chain_key)) {
+               /*
+                * Check whether last held lock:
+                *
+                * - is irq-safe, if this lock is irq-unsafe
+                * - is softirq-safe, if this lock is hardirq-unsafe
+                *
+                * And check whether the new lock's dependency graph
+                * could lead back to the previous lock.
+                *
+                * any of these scenarios could lead to a deadlock. If
+                * All validations
+                */
+               int ret = check_deadlock(curr, hlock, lock, read);
+
+               if (!ret)
+                       return 0;
+               /*
+                * Mark recursive read, as we jump over it when
+                * building dependencies (just like we jump over
+                * trylock entries):
+                */
+               if (ret == 2)
+                       hlock->read = 2;
+               /*
+                * Add dependency only if this lock is not the head
+                * of the chain, and if it's not a secondary read-lock:
+                */
+               if (!chain_head && ret != 2)
+                       if (!check_prevs_add(curr, hlock))
+                               return 0;
+               __raw_spin_unlock(&hash_lock);
+       }
+       curr->lockdep_depth++;
+       check_chain_key(curr);
+       if (unlikely(curr->lockdep_depth >= MAX_LOCK_DEPTH)) {
+               debug_locks_off();
+               printk("BUG: MAX_LOCK_DEPTH too low!\n");
+               printk("turning off the locking correctness validator.\n");
+               return 0;
+       }
+       if (unlikely(curr->lockdep_depth > max_lockdep_depth))
+               max_lockdep_depth = curr->lockdep_depth;
+
+       return 1;
+}
+
+static int
+print_unlock_inbalance_bug(struct task_struct *curr, struct lockdep_map *lock,
+                          unsigned long ip)
+{
+       if (!debug_locks_off())
+               return 0;
+       if (debug_locks_silent)
+               return 0;
+
+       printk("\n=====================================\n");
+       printk(  "[ BUG: bad unlock balance detected! ]\n");
+       printk(  "-------------------------------------\n");
+       printk("%s/%d is trying to release lock (",
+               curr->comm, curr->pid);
+       print_lockdep_cache(lock);
+       printk(") at:\n");
+       print_ip_sym(ip);
+       printk("but there are no more locks to release!\n");
+       printk("\nother info that might help us debug this:\n");
+       lockdep_print_held_locks(curr);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+
+       return 0;
+}
+
+/*
+ * Common debugging checks for both nested and non-nested unlock:
+ */
+static int check_unlock(struct task_struct *curr, struct lockdep_map *lock,
+                       unsigned long ip)
+{
+       if (unlikely(!debug_locks))
+               return 0;
+       if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
+               return 0;
+
+       if (curr->lockdep_depth <= 0)
+               return print_unlock_inbalance_bug(curr, lock, ip);
+
+       return 1;
+}
+
+/*
+ * Remove the lock to the list of currently held locks in a
+ * potentially non-nested (out of order) manner. This is a
+ * relatively rare operation, as all the unlock APIs default
+ * to nested mode (which uses lock_release()):
+ */
+static int
+lock_release_non_nested(struct task_struct *curr,
+                       struct lockdep_map *lock, unsigned long ip)
+{
+       struct held_lock *hlock, *prev_hlock;
+       unsigned int depth;
+       int i;
+
+       /*
+        * Check whether the lock exists in the current stack
+        * of held locks:
+        */
+       depth = curr->lockdep_depth;
+       if (DEBUG_LOCKS_WARN_ON(!depth))
+               return 0;
+
+       prev_hlock = NULL;
+       for (i = depth-1; i >= 0; i--) {
+               hlock = curr->held_locks + i;
+               /*
+                * We must not cross into another context:
+                */
+               if (prev_hlock && prev_hlock->irq_context != hlock->irq_context)
+                       break;
+               if (hlock->instance == lock)
+                       goto found_it;
+               prev_hlock = hlock;
+       }
+       return print_unlock_inbalance_bug(curr, lock, ip);
+
+found_it:
+       /*
+        * We have the right lock to unlock, 'hlock' points to it.
+        * Now we remove it from the stack, and add back the other
+        * entries (if any), recalculating the hash along the way:
+        */
+       curr->lockdep_depth = i;
+       curr->curr_chain_key = hlock->prev_chain_key;
+
+       for (i++; i < depth; i++) {
+               hlock = curr->held_locks + i;
+               if (!__lock_acquire(hlock->instance,
+                       hlock->class->subclass, hlock->trylock,
+                               hlock->read, hlock->check, hlock->hardirqs_off,
+                               hlock->acquire_ip))
+                       return 0;
+       }
+
+       if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1))
+               return 0;
+       return 1;
+}
+
+/*
+ * Remove the lock to the list of currently held locks - this gets
+ * called on mutex_unlock()/spin_unlock*() (or on a failed
+ * mutex_lock_interruptible()). This is done for unlocks that nest
+ * perfectly. (i.e. the current top of the lock-stack is unlocked)
+ */
+static int lock_release_nested(struct task_struct *curr,
+                              struct lockdep_map *lock, unsigned long ip)
+{
+       struct held_lock *hlock;
+       unsigned int depth;
+
+       /*
+        * Pop off the top of the lock stack:
+        */
+       depth = curr->lockdep_depth - 1;
+       hlock = curr->held_locks + depth;
+
+       /*
+        * Is the unlock non-nested:
+        */
+       if (hlock->instance != lock)
+               return lock_release_non_nested(curr, lock, ip);
+       curr->lockdep_depth--;
+
+       if (DEBUG_LOCKS_WARN_ON(!depth && (hlock->prev_chain_key != 0)))
+               return 0;
+
+       curr->curr_chain_key = hlock->prev_chain_key;
+
+#ifdef CONFIG_DEBUG_LOCKDEP
+       hlock->prev_chain_key = 0;
+       hlock->class = NULL;
+       hlock->acquire_ip = 0;
+       hlock->irq_context = 0;
+#endif
+       return 1;
+}
+
+/*
+ * Remove the lock to the list of currently held locks - this gets
+ * called on mutex_unlock()/spin_unlock*() (or on a failed
+ * mutex_lock_interruptible()). This is done for unlocks that nest
+ * perfectly. (i.e. the current top of the lock-stack is unlocked)
+ */
+static void
+__lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
+{
+       struct task_struct *curr = current;
+
+       if (!check_unlock(curr, lock, ip))
+               return;
+
+       if (nested) {
+               if (!lock_release_nested(curr, lock, ip))
+                       return;
+       } else {
+               if (!lock_release_non_nested(curr, lock, ip))
+                       return;
+       }
+
+       check_chain_key(curr);
+}
+
+/*
+ * Check whether we follow the irq-flags state precisely:
+ */
+static void check_flags(unsigned long flags)
+{
+#if defined(CONFIG_DEBUG_LOCKDEP) && defined(CONFIG_TRACE_IRQFLAGS)
+       if (!debug_locks)
+               return;
+
+       if (irqs_disabled_flags(flags))
+               DEBUG_LOCKS_WARN_ON(current->hardirqs_enabled);
+       else
+               DEBUG_LOCKS_WARN_ON(!current->hardirqs_enabled);
+
+       /*
+        * We dont accurately track softirq state in e.g.
+        * hardirq contexts (such as on 4KSTACKS), so only
+        * check if not in hardirq contexts:
+        */
+       if (!hardirq_count()) {
+               if (softirq_count())
+                       DEBUG_LOCKS_WARN_ON(current->softirqs_enabled);
+               else
+                       DEBUG_LOCKS_WARN_ON(!current->softirqs_enabled);
+       }
+
+       if (!debug_locks)
+               print_irqtrace_events(current);
+#endif
+}
+
+/*
+ * We are not always called with irqs disabled - do that here,
+ * and also avoid lockdep recursion:
+ */
+void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
+                 int trylock, int read, int check, unsigned long ip)
+{
+       unsigned long flags;
+
+       if (unlikely(current->lockdep_recursion))
+               return;
+
+       raw_local_irq_save(flags);
+       check_flags(flags);
+
+       current->lockdep_recursion = 1;
+       __lock_acquire(lock, subclass, trylock, read, check,
+                      irqs_disabled_flags(flags), ip);
+       current->lockdep_recursion = 0;
+       raw_local_irq_restore(flags);
+}
+
+EXPORT_SYMBOL_GPL(lock_acquire);
+
+void lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
+{
+       unsigned long flags;
+
+       if (unlikely(current->lockdep_recursion))
+               return;
+
+       raw_local_irq_save(flags);
+       check_flags(flags);
+       current->lockdep_recursion = 1;
+       __lock_release(lock, nested, ip);
+       current->lockdep_recursion = 0;
+       raw_local_irq_restore(flags);
+}
+
+EXPORT_SYMBOL_GPL(lock_release);
+
+/*
+ * Used by the testsuite, sanitize the validator state
+ * after a simulated failure:
+ */
+
+void lockdep_reset(void)
+{
+       unsigned long flags;
+
+       raw_local_irq_save(flags);
+       current->curr_chain_key = 0;
+       current->lockdep_depth = 0;
+       current->lockdep_recursion = 0;
+       memset(current->held_locks, 0, MAX_LOCK_DEPTH*sizeof(struct held_lock));
+       nr_hardirq_chains = 0;
+       nr_softirq_chains = 0;
+       nr_process_chains = 0;
+       debug_locks = 1;
+       raw_local_irq_restore(flags);
+}
+
+static void zap_class(struct lock_class *class)
+{
+       int i;
+
+       /*
+        * Remove all dependencies this lock is
+        * involved in:
+        */
+       for (i = 0; i < nr_list_entries; i++) {
+               if (list_entries[i].class == class)
+                       list_del_rcu(&list_entries[i].entry);
+       }
+       /*
+        * Unhash the class and remove it from the all_lock_classes list:
+        */
+       list_del_rcu(&class->hash_entry);
+       list_del_rcu(&class->lock_entry);
+
+}
+
+static inline int within(void *addr, void *start, unsigned long size)
+{
+       return addr >= start && addr < start + size;
+}
+
+void lockdep_free_key_range(void *start, unsigned long size)
+{
+       struct lock_class *class, *next;
+       struct list_head *head;
+       unsigned long flags;
+       int i;
+
+       raw_local_irq_save(flags);
+       __raw_spin_lock(&hash_lock);
+
+       /*
+        * Unhash all classes that were created by this module:
+        */
+       for (i = 0; i < CLASSHASH_SIZE; i++) {
+               head = classhash_table + i;
+               if (list_empty(head))
+                       continue;
+               list_for_each_entry_safe(class, next, head, hash_entry)
+                       if (within(class->key, start, size))
+                               zap_class(class);
+       }
+
+       __raw_spin_unlock(&hash_lock);
+       raw_local_irq_restore(flags);
+}
+
+void lockdep_reset_lock(struct lockdep_map *lock)
+{
+       struct lock_class *class, *next, *entry;
+       struct list_head *head;
+       unsigned long flags;
+       int i, j;
+
+       raw_local_irq_save(flags);
+       __raw_spin_lock(&hash_lock);
+
+       /*
+        * Remove all classes this lock has:
+        */
+       for (i = 0; i < CLASSHASH_SIZE; i++) {
+               head = classhash_table + i;
+               if (list_empty(head))
+                       continue;
+               list_for_each_entry_safe(class, next, head, hash_entry) {
+                       for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
+                               entry = lock->class[j];
+                               if (class == entry) {
+                                       zap_class(class);
+                                       lock->class[j] = NULL;
+                                       break;
+                               }
+                       }
+               }
+       }
+
+       /*
+        * Debug check: in the end all mapped classes should
+        * be gone.
+        */
+       for (j = 0; j < MAX_LOCKDEP_SUBCLASSES; j++) {
+               entry = lock->class[j];
+               if (!entry)
+                       continue;
+               __raw_spin_unlock(&hash_lock);
+               DEBUG_LOCKS_WARN_ON(1);
+               raw_local_irq_restore(flags);
+               return;
+       }
+
+       __raw_spin_unlock(&hash_lock);
+       raw_local_irq_restore(flags);
+}
+
+void __init lockdep_init(void)
+{
+       int i;
+
+       /*
+        * Some architectures have their own start_kernel()
+        * code which calls lockdep_init(), while we also
+        * call lockdep_init() from the start_kernel() itself,
+        * and we want to initialize the hashes only once:
+        */
+       if (lockdep_initialized)
+               return;
+
+       for (i = 0; i < CLASSHASH_SIZE; i++)
+               INIT_LIST_HEAD(classhash_table + i);
+
+       for (i = 0; i < CHAINHASH_SIZE; i++)
+               INIT_LIST_HEAD(chainhash_table + i);
+
+       lockdep_initialized = 1;
+}
+
+void __init lockdep_info(void)
+{
+       printk("Lock dependency validator: Copyright (c) 2006 Red Hat, Inc., Ingo Molnar\n");
+
+       printk("... MAX_LOCKDEP_SUBCLASSES:    %lu\n", MAX_LOCKDEP_SUBCLASSES);
+       printk("... MAX_LOCK_DEPTH:          %lu\n", MAX_LOCK_DEPTH);
+       printk("... MAX_LOCKDEP_KEYS:        %lu\n", MAX_LOCKDEP_KEYS);
+       printk("... CLASSHASH_SIZE:           %lu\n", CLASSHASH_SIZE);
+       printk("... MAX_LOCKDEP_ENTRIES:     %lu\n", MAX_LOCKDEP_ENTRIES);
+       printk("... MAX_LOCKDEP_CHAINS:      %lu\n", MAX_LOCKDEP_CHAINS);
+       printk("... CHAINHASH_SIZE:          %lu\n", CHAINHASH_SIZE);
+
+       printk(" memory used by lock dependency info: %lu kB\n",
+               (sizeof(struct lock_class) * MAX_LOCKDEP_KEYS +
+               sizeof(struct list_head) * CLASSHASH_SIZE +
+               sizeof(struct lock_list) * MAX_LOCKDEP_ENTRIES +
+               sizeof(struct lock_chain) * MAX_LOCKDEP_CHAINS +
+               sizeof(struct list_head) * CHAINHASH_SIZE) / 1024);
+
+       printk(" per task-struct memory footprint: %lu bytes\n",
+               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");
+#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)
+{
+       if (!debug_locks_off())
+               return;
+       if (debug_locks_silent)
+               return;
+
+       printk("\n=========================\n");
+       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);
+       lockdep_print_held_locks(curr);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+}
+
+/*
+ * Called when kernel memory is freed (or unmapped), or if a lock
+ * is destroyed or reinitialized - this code checks whether there is
+ * any held lock in the memory range of <from> to <to>:
+ */
+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;
+       int i;
+
+       if (unlikely(!debug_locks))
+               return;
+
+       local_irq_save(flags);
+       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))
+                       continue;
+
+               print_freed_lock_bug(curr, mem_from, mem_to);
+               break;
+       }
+       local_irq_restore(flags);
+}
+
+static void print_held_locks_bug(struct task_struct *curr)
+{
+       if (!debug_locks_off())
+               return;
+       if (debug_locks_silent)
+               return;
+
+       printk("\n=====================================\n");
+       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);
+       lockdep_print_held_locks(curr);
+
+       printk("\nstack backtrace:\n");
+       dump_stack();
+}
+
+void debug_check_no_locks_held(struct task_struct *task)
+{
+       if (unlikely(task->lockdep_depth > 0))
+               print_held_locks_bug(task);
+}
+
+void debug_show_all_locks(void)
+{
+       struct task_struct *g, *p;
+       int count = 10;
+       int unlock = 1;
+
+       printk("\nShowing all locks held 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) {
+               if (p->lockdep_depth)
+                       lockdep_print_held_locks(p);
+               if (!unlock)
+                       if (read_trylock(&tasklist_lock))
+                               unlock = 1;
+       } while_each_thread(g, p);
+
+       printk("\n");
+       printk("=============================================\n\n");
+
+       if (unlock)
+               read_unlock(&tasklist_lock);
+}
+
+EXPORT_SYMBOL_GPL(debug_show_all_locks);
+
+void debug_show_held_locks(struct task_struct *task)
+{
+       lockdep_print_held_locks(task);
+}
+
+EXPORT_SYMBOL_GPL(debug_show_held_locks);
+
diff --git a/kernel/lockdep_internals.h b/kernel/lockdep_internals.h
new file mode 100644 (file)
index 0000000..0d355f2
--- /dev/null
@@ -0,0 +1,78 @@
+/*
+ * kernel/lockdep_internals.h
+ *
+ * Runtime locking correctness validator
+ *
+ * lockdep subsystem internal functions and variables.
+ */
+
+/*
+ * MAX_LOCKDEP_ENTRIES is the maximum number of lock dependencies
+ * we track.
+ *
+ * We use the per-lock dependency maps in two ways: we grow it by adding
+ * every to-be-taken lock to all currently held lock's own dependency
+ * table (if it's not there yet), and we check it for lock order
+ * conflicts and deadlocks.
+ */
+#define MAX_LOCKDEP_ENTRIES    8192UL
+
+#define MAX_LOCKDEP_KEYS_BITS  11
+#define MAX_LOCKDEP_KEYS       (1UL << MAX_LOCKDEP_KEYS_BITS)
+
+#define MAX_LOCKDEP_CHAINS_BITS        13
+#define MAX_LOCKDEP_CHAINS     (1UL << MAX_LOCKDEP_CHAINS_BITS)
+
+/*
+ * Stack-trace: tightly packed array of stack backtrace
+ * addresses. Protected by the hash_lock.
+ */
+#define MAX_STACK_TRACE_ENTRIES        131072UL
+
+extern struct list_head all_lock_classes;
+
+extern void
+get_usage_chars(struct lock_class *class, char *c1, char *c2, char *c3, char *c4);
+
+extern const char * __get_key_name(struct lockdep_subclass_key *key, char *str);
+
+extern unsigned long nr_lock_classes;
+extern unsigned long nr_list_entries;
+extern unsigned long nr_lock_chains;
+extern unsigned long nr_stack_trace_entries;
+
+extern unsigned int nr_hardirq_chains;
+extern unsigned int nr_softirq_chains;
+extern unsigned int nr_process_chains;
+extern unsigned int max_lockdep_depth;
+extern unsigned int max_recursion_depth;
+
+#ifdef CONFIG_DEBUG_LOCKDEP
+/*
+ * Various lockdep statistics:
+ */
+extern atomic_t chain_lookup_hits;
+extern atomic_t chain_lookup_misses;
+extern atomic_t hardirqs_on_events;
+extern atomic_t hardirqs_off_events;
+extern atomic_t redundant_hardirqs_on;
+extern atomic_t redundant_hardirqs_off;
+extern atomic_t softirqs_on_events;
+extern atomic_t softirqs_off_events;
+extern atomic_t redundant_softirqs_on;
+extern atomic_t redundant_softirqs_off;
+extern atomic_t nr_unused_locks;
+extern atomic_t nr_cyclic_checks;
+extern atomic_t nr_cyclic_check_recursions;
+extern atomic_t nr_find_usage_forwards_checks;
+extern atomic_t nr_find_usage_forwards_recursions;
+extern atomic_t nr_find_usage_backwards_checks;
+extern atomic_t nr_find_usage_backwards_recursions;
+# define debug_atomic_inc(ptr)         atomic_inc(ptr)
+# define debug_atomic_dec(ptr)         atomic_dec(ptr)
+# define debug_atomic_read(ptr)                atomic_read(ptr)
+#else
+# define debug_atomic_inc(ptr)         do { } while (0)
+# define debug_atomic_dec(ptr)         do { } while (0)
+# define debug_atomic_read(ptr)                0
+#endif
index 0351625..35e1b1f 100644 (file)
@@ -1121,6 +1121,9 @@ static void free_module(struct module *mod)
        if (mod->percpu)
                percpu_modfree(mod->percpu);
 
+       /* Free lock-classes: */
+       lockdep_free_key_range(mod->module_core, mod->core_size);
+
        /* Finally, free the core (containing the module structure) */
        module_free(mod, mod->module_core);
 }
index 16021b0..16c2e98 100644 (file)
@@ -48,7 +48,7 @@ config DEBUG_KERNEL
 config LOG_BUF_SHIFT
        int "Kernel log buffer size (16 => 64KB, 17 => 128KB)" if DEBUG_KERNEL
        range 12 21
-       default 17 if S390
+       default 17 if S390 || LOCKDEP
        default 16 if X86_NUMAQ || IA64
        default 15 if SMP
        default 14
index 5cd05f2..9bdc8d4 100644 (file)
@@ -15,6 +15,7 @@
 #include <linux/sched.h>
 #include <linux/delay.h>
 #include <linux/module.h>
+#include <linux/lockdep.h>
 #include <linux/spinlock.h>
 #include <linux/kallsyms.h>
 #include <linux/interrupt.h>
@@ -889,9 +890,6 @@ GENERATE_PERMUTATIONS_3_EVENTS(irq_read_recursion_soft)
 #include "locking-selftest-softirq.h"
 // GENERATE_PERMUTATIONS_3_EVENTS(irq_read_recursion2_soft)
 
-#define lockdep_reset()
-#define lockdep_reset_lock(x)
-
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 # define I_SPINLOCK(x) lockdep_reset_lock(&lock_##x.dep_map)
 # define I_RWLOCK(x)   lockdep_reset_lock(&rwlock_##x.dep_map)