seqcount: Add lockdep functionality to seqcount/seqlock structures
authorJohn Stultz <john.stultz@linaro.org>
Mon, 7 Oct 2013 22:51:59 +0000 (15:51 -0700)
committerIngo Molnar <mingo@kernel.org>
Wed, 6 Nov 2013 11:40:26 +0000 (12:40 +0100)
Currently seqlocks and seqcounts don't support lockdep.

After running across a seqcount related deadlock in the timekeeping
code, I used a less-refined and more focused variant of this patch
to narrow down the cause of the issue.

This is a first-pass attempt to properly enable lockdep functionality
on seqlocks and seqcounts.

Since seqcounts are used in the vdso gettimeofday code, I've provided
non-lockdep accessors for those needs.

I've also handled one case where there were nested seqlock writers
and there may be more edge cases.

Comments and feedback would be appreciated!

Signed-off-by: John Stultz <john.stultz@linaro.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Li Zefan <lizefan@huawei.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: netdev@vger.kernel.org
Link: http://lkml.kernel.org/r/1381186321-4906-3-git-send-email-john.stultz@linaro.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/vdso/vclock_gettime.c
fs/dcache.c
fs/fs_struct.c
include/linux/init_task.h
include/linux/lockdep.h
include/linux/seqlock.h
mm/filemap_xip.c

index 72074d5..2ada505 100644 (file)
@@ -178,7 +178,7 @@ notrace static int __always_inline do_realtime(struct timespec *ts)
 
        ts->tv_nsec = 0;
        do {
-               seq = read_seqcount_begin(&gtod->seq);
+               seq = read_seqcount_begin_no_lockdep(&gtod->seq);
                mode = gtod->clock.vclock_mode;
                ts->tv_sec = gtod->wall_time_sec;
                ns = gtod->wall_time_snsec;
@@ -198,7 +198,7 @@ notrace static int do_monotonic(struct timespec *ts)
 
        ts->tv_nsec = 0;
        do {
-               seq = read_seqcount_begin(&gtod->seq);
+               seq = read_seqcount_begin_no_lockdep(&gtod->seq);
                mode = gtod->clock.vclock_mode;
                ts->tv_sec = gtod->monotonic_time_sec;
                ns = gtod->monotonic_time_snsec;
@@ -214,7 +214,7 @@ notrace static int do_realtime_coarse(struct timespec *ts)
 {
        unsigned long seq;
        do {
-               seq = read_seqcount_begin(&gtod->seq);
+               seq = read_seqcount_begin_no_lockdep(&gtod->seq);
                ts->tv_sec = gtod->wall_time_coarse.tv_sec;
                ts->tv_nsec = gtod->wall_time_coarse.tv_nsec;
        } while (unlikely(read_seqcount_retry(&gtod->seq, seq)));
@@ -225,7 +225,7 @@ notrace static int do_monotonic_coarse(struct timespec *ts)
 {
        unsigned long seq;
        do {
-               seq = read_seqcount_begin(&gtod->seq);
+               seq = read_seqcount_begin_no_lockdep(&gtod->seq);
                ts->tv_sec = gtod->monotonic_time_coarse.tv_sec;
                ts->tv_nsec = gtod->monotonic_time_coarse.tv_nsec;
        } while (unlikely(read_seqcount_retry(&gtod->seq, seq)));
index ae6ebb8..f750be2 100644 (file)
@@ -2574,7 +2574,7 @@ static void __d_move(struct dentry * dentry, struct dentry * target)
        dentry_lock_for_move(dentry, target);
 
        write_seqcount_begin(&dentry->d_seq);
-       write_seqcount_begin(&target->d_seq);
+       write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
 
        /* __d_drop does write_seqcount_barrier, but they're OK to nest. */
 
@@ -2706,7 +2706,7 @@ static void __d_materialise_dentry(struct dentry *dentry, struct dentry *anon)
        dentry_lock_for_move(anon, dentry);
 
        write_seqcount_begin(&dentry->d_seq);
-       write_seqcount_begin(&anon->d_seq);
+       write_seqcount_begin_nested(&anon->d_seq, DENTRY_D_LOCK_NESTED);
 
        dparent = dentry->d_parent;
 
index d8ac61d..7dca743 100644 (file)
@@ -161,6 +161,6 @@ EXPORT_SYMBOL(current_umask);
 struct fs_struct init_fs = {
        .users          = 1,
        .lock           = __SPIN_LOCK_UNLOCKED(init_fs.lock),
-       .seq            = SEQCNT_ZERO,
+       .seq            = SEQCNT_ZERO(init_fs.seq),
        .umask          = 0022,
 };
index 5cd0f09..b0ed422 100644 (file)
@@ -32,10 +32,10 @@ extern struct fs_struct init_fs;
 #endif
 
 #ifdef CONFIG_CPUSETS
-#define INIT_CPUSET_SEQ                                                        \
-       .mems_allowed_seq = SEQCNT_ZERO,
+#define INIT_CPUSET_SEQ(tsk)                                                   \
+       .mems_allowed_seq = SEQCNT_ZERO(tsk.mems_allowed_seq),
 #else
-#define INIT_CPUSET_SEQ
+#define INIT_CPUSET_SEQ(tsk)
 #endif
 
 #define INIT_SIGNALS(sig) {                                            \
@@ -220,7 +220,7 @@ extern struct task_group root_task_group;
        INIT_FTRACE_GRAPH                                               \
        INIT_TRACE_RECURSION                                            \
        INIT_TASK_RCU_PREEMPT(tsk)                                      \
-       INIT_CPUSET_SEQ                                                 \
+       INIT_CPUSET_SEQ(tsk)                                            \
        INIT_VTIME(tsk)                                                 \
 }
 
index cfc2f11..92b1bfc 100644 (file)
@@ -497,6 +497,10 @@ static inline void print_irqtrace_events(struct task_struct *curr)
 #define rwlock_acquire_read(l, s, t, i)                lock_acquire_shared_recursive(l, s, t, NULL, i)
 #define rwlock_release(l, n, i)                        lock_release(l, n, i)
 
+#define seqcount_acquire(l, s, t, i)           lock_acquire_exclusive(l, s, t, NULL, i)
+#define seqcount_acquire_read(l, s, t, i)      lock_acquire_shared_recursive(l, s, t, NULL, i)
+#define seqcount_release(l, n, i)              lock_release(l, n, i)
+
 #define mutex_acquire(l, s, t, i)              lock_acquire_exclusive(l, s, t, NULL, i)
 #define mutex_acquire_nest(l, s, t, n, i)      lock_acquire_exclusive(l, s, t, n, i)
 #define mutex_release(l, n, i)                 lock_release(l, n, i)
@@ -504,11 +508,11 @@ static inline void print_irqtrace_events(struct task_struct *curr)
 #define rwsem_acquire(l, s, t, i)              lock_acquire_exclusive(l, s, t, NULL, i)
 #define rwsem_acquire_nest(l, s, t, n, i)      lock_acquire_exclusive(l, s, t, n, i)
 #define rwsem_acquire_read(l, s, t, i)         lock_acquire_shared(l, s, t, NULL, i)
-# define rwsem_release(l, n, i)                        lock_release(l, n, i)
+#define rwsem_release(l, n, i)                 lock_release(l, n, i)
 
 #define lock_map_acquire(l)                    lock_acquire_exclusive(l, 0, 0, NULL, _THIS_IP_)
 #define lock_map_acquire_read(l)               lock_acquire_shared_recursive(l, 0, 0, NULL, _THIS_IP_)
-# define lock_map_release(l)                   lock_release(l, 1, _THIS_IP_)
+#define lock_map_release(l)                    lock_release(l, 1, _THIS_IP_)
 
 #ifdef CONFIG_PROVE_LOCKING
 # define might_lock(lock)                                              \
index 21a2093..1e8a8b6 100644 (file)
@@ -34,6 +34,7 @@
 
 #include <linux/spinlock.h>
 #include <linux/preempt.h>
+#include <linux/lockdep.h>
 #include <asm/processor.h>
 
 /*
  */
 typedef struct seqcount {
        unsigned sequence;
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+       struct lockdep_map dep_map;
+#endif
 } seqcount_t;
 
-#define SEQCNT_ZERO { 0 }
-#define seqcount_init(x)       do { *(x) = (seqcount_t) SEQCNT_ZERO; } while (0)
+static inline void __seqcount_init(seqcount_t *s, const char *name,
+                                         struct lock_class_key *key)
+{
+       /*
+        * Make sure we are not reinitializing a held lock:
+        */
+       lockdep_init_map(&s->dep_map, name, key, 0);
+       s->sequence = 0;
+}
+
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+# define SEQCOUNT_DEP_MAP_INIT(lockname) \
+               .dep_map = { .name = #lockname } \
+
+# define seqcount_init(s)                              \
+       do {                                            \
+               static struct lock_class_key __key;     \
+               __seqcount_init((s), #s, &__key);       \
+       } while (0)
+
+static inline void seqcount_lockdep_reader_access(const seqcount_t *s)
+{
+       seqcount_t *l = (seqcount_t *)s;
+       unsigned long flags;
+
+       local_irq_save(flags);
+       seqcount_acquire_read(&l->dep_map, 0, 0, _RET_IP_);
+       seqcount_release(&l->dep_map, 1, _RET_IP_);
+       local_irq_restore(flags);
+}
+
+#else
+# define SEQCOUNT_DEP_MAP_INIT(lockname)
+# define seqcount_init(s) __seqcount_init(s, NULL, NULL)
+# define seqcount_lockdep_reader_access(x)
+#endif
+
+#define SEQCNT_ZERO(lockname) { .sequence = 0, SEQCOUNT_DEP_MAP_INIT(lockname)}
+
 
 /**
  * __read_seqcount_begin - begin a seq-read critical section (without barrier)
@@ -76,6 +117,22 @@ repeat:
 }
 
 /**
+ * read_seqcount_begin_no_lockdep - start seq-read critical section w/o lockdep
+ * @s: pointer to seqcount_t
+ * Returns: count to be passed to read_seqcount_retry
+ *
+ * read_seqcount_begin_no_lockdep opens a read critical section of the given
+ * seqcount, but without any lockdep checking. Validity of the critical
+ * section is tested by checking read_seqcount_retry function.
+ */
+static inline unsigned read_seqcount_begin_no_lockdep(const seqcount_t *s)
+{
+       unsigned ret = __read_seqcount_begin(s);
+       smp_rmb();
+       return ret;
+}
+
+/**
  * read_seqcount_begin - begin a seq-read critical section
  * @s: pointer to seqcount_t
  * Returns: count to be passed to read_seqcount_retry
@@ -86,9 +143,8 @@ repeat:
  */
 static inline unsigned read_seqcount_begin(const seqcount_t *s)
 {
-       unsigned ret = __read_seqcount_begin(s);
-       smp_rmb();
-       return ret;
+       seqcount_lockdep_reader_access(s);
+       return read_seqcount_begin_no_lockdep(s);
 }
 
 /**
@@ -108,6 +164,8 @@ static inline unsigned read_seqcount_begin(const seqcount_t *s)
 static inline unsigned raw_seqcount_begin(const seqcount_t *s)
 {
        unsigned ret = ACCESS_ONCE(s->sequence);
+
+       seqcount_lockdep_reader_access(s);
        smp_rmb();
        return ret & ~1;
 }
@@ -152,14 +210,21 @@ static inline int read_seqcount_retry(const seqcount_t *s, unsigned start)
  * Sequence counter only version assumes that callers are using their
  * own mutexing.
  */
-static inline void write_seqcount_begin(seqcount_t *s)
+static inline void write_seqcount_begin_nested(seqcount_t *s, int subclass)
 {
        s->sequence++;
        smp_wmb();
+       seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);
+}
+
+static inline void write_seqcount_begin(seqcount_t *s)
+{
+       write_seqcount_begin_nested(s, 0);
 }
 
 static inline void write_seqcount_end(seqcount_t *s)
 {
+       seqcount_release(&s->dep_map, 1, _RET_IP_);
        smp_wmb();
        s->sequence++;
 }
@@ -188,7 +253,7 @@ typedef struct {
  */
 #define __SEQLOCK_UNLOCKED(lockname)                   \
        {                                               \
-               .seqcount = SEQCNT_ZERO,                \
+               .seqcount = SEQCNT_ZERO(lockname),      \
                .lock = __SPIN_LOCK_UNLOCKED(lockname)  \
        }
 
index 28fe26b..d8d9fe3 100644 (file)
@@ -26,7 +26,7 @@
  * of ZERO_PAGE(), such as /dev/zero
  */
 static DEFINE_MUTEX(xip_sparse_mutex);
-static seqcount_t xip_sparse_seq = SEQCNT_ZERO;
+static seqcount_t xip_sparse_seq = SEQCNT_ZERO(xip_sparse_seq);
 static struct page *__xip_sparse_page;
 
 /* called under xip_sparse_mutex */