Add wait_on_atomic_t() and wake_up_atomic_t()

Add wait_on_atomic_t() and wake_up_atomic_t() to indicate became-zero events on
atomic_t types.  This uses the bit-wake waitqueue table.  The key is set to a
value outside of the number of bits in a long so that wait_on_bit() won't be
woken up accidentally.

What I'm using this for is: in a following patch I add a counter to struct
fscache_cookie to count the number of outstanding operations that need access
to netfs data.  The way this works is:

 (1) When a cookie is allocated, the counter is initialised to 1.

 (2) When an operation wants to access netfs data, it calls atomic_inc_unless()
     to increment the counter before it does so.  If it was 0, then the counter
     isn't incremented, the operation isn't permitted to access the netfs data
     (which might by this point no longer exist) and the operation aborts in
     some appropriate manner.

 (3) When an operation finishes with the netfs data, it decrements the counter
     and if it reaches 0, calls wake_up_atomic_t() on it - the assumption being
     that it was the last blocker.

 (4) When a cookie is released, the counter is decremented and the releaser
     uses wait_on_atomic_t() to wait for the counter to become 0 - which should
     indicate no one is using the netfs data any longer.  The netfs data can
     then be destroyed.

There are some alternatives that I have thought of and that have been suggested
by Tejun Heo:

 (A) Using wait_on_bit() to wait on a bit in the counter.  This doesn't work
     because if that bit happens to be 0 then the wait won't happen - even if
     the counter is non-zero.

 (B) Using wait_on_bit() to wait on a flag elsewhere which is cleared when the
     counter reaches 0.  Such a flag would be redundant and would add
     complexity.

 (C) Adding a waitqueue to fscache_cookie - this would expand that struct by
     several words for an event that happens just once in each cookie's
     lifetime.  Further, cookies are generally per-file so there are likely to
     be a lot of them.

 (D) Similar to (C), but add a pointer to a waitqueue in the cookie instead of
     a waitqueue.  This would add single word per cookie and so would be less
     of an expansion - but still an expansion.

 (E) Adding a static waitqueue to the fscache module.  Generally this would be
     fine, but under certain circumstances many cookies will all get added at
     the same time (eg. NFS umount, cache withdrawal) thereby presenting
     scaling issues.  Note that the wait may be significant as disk I/O may be
     in progress.

So, I think reusing the wait_on_bit() waitqueue set is reasonable.  I don't
make much use of the waitqueue I need on a per-cookie basis, but sometimes I
have a huge flood of the cookies to deal with.

I also don't want to add a whole new set of global waitqueue tables
specifically for the dec-to-0 event if I can reuse the bit tables.

Signed-off-by: David Howells <dhowells@redhat.com>
Tested-By: Milosz Tanski <milosz@adfin.com>
Acked-by: Jeff Layton <jlayton@redhat.com>

include/linux/wait.h

@@ -23,6 +23,7 @@ struct __wait_queue {
 struct wait_bit_key {
 	void *flags;
 	int bit_nr;
+#define WAIT_ATOMIC_T_BIT_NR -1
 };
 
 struct wait_bit_queue {
@@ -60,6 +61,9 @@ struct task_struct;
 #define __WAIT_BIT_KEY_INITIALIZER(word, bit)				\
 	{ .flags = word, .bit_nr = bit, }
 
+#define __WAIT_ATOMIC_T_KEY_INITIALIZER(p)				\
+	{ .flags = p, .bit_nr = WAIT_ATOMIC_T_BIT_NR, }
+
 extern void __init_waitqueue_head(wait_queue_head_t *q, const char *name, struct lock_class_key *);
 
 #define init_waitqueue_head(q)				\
@@ -146,8 +150,10 @@ void __wake_up_bit(wait_queue_head_t *, void *, int);
 int __wait_on_bit(wait_queue_head_t *, struct wait_bit_queue *, int (*)(void *), unsigned);
 int __wait_on_bit_lock(wait_queue_head_t *, struct wait_bit_queue *, int (*)(void *), unsigned);
 void wake_up_bit(void *, int);
+void wake_up_atomic_t(atomic_t *);
 int out_of_line_wait_on_bit(void *, int, int (*)(void *), unsigned);
 int out_of_line_wait_on_bit_lock(void *, int, int (*)(void *), unsigned);
+int out_of_line_wait_on_atomic_t(atomic_t *, int (*)(atomic_t *), unsigned);
 wait_queue_head_t *bit_waitqueue(void *, int);
 
 #define wake_up(x)			__wake_up(x, TASK_NORMAL, 1, NULL)
@@ -896,5 +902,23 @@ static inline int wait_on_bit_lock(void *word, int bit,
 		return 0;
 	return out_of_line_wait_on_bit_lock(word, bit, action, mode);
 }
+
+/**
+ * wait_on_atomic_t - Wait for an atomic_t to become 0
+ * @val: The atomic value being waited on, a kernel virtual address
+ * @action: the function used to sleep, which may take special actions
+ * @mode: the task state to sleep in
+ *
+ * Wait for an atomic_t to become 0.  We abuse the bit-wait waitqueue table for
+ * the purpose of getting a waitqueue, but we set the key to a bit number
+ * outside of the target 'word'.
+ */
+static inline
+int wait_on_atomic_t(atomic_t *val, int (*action)(atomic_t *), unsigned mode)
+{
+	if (atomic_read(val) == 0)
+		return 0;
+	return out_of_line_wait_on_atomic_t(val, action, mode);
+}
 	
 #endif

kernel/wait.c

@@ -287,3 +287,91 @@ wait_queue_head_t *bit_waitqueue(void *word, int bit)
 	return &zone->wait_table[hash_long(val, zone->wait_table_bits)];
 }
 EXPORT_SYMBOL(bit_waitqueue);
+
+/*
+ * Manipulate the atomic_t address to produce a better bit waitqueue table hash
+ * index (we're keying off bit -1, but that would produce a horrible hash
+ * value).
+ */
+static inline wait_queue_head_t *atomic_t_waitqueue(atomic_t *p)
+{
+	if (BITS_PER_LONG == 64) {
+		unsigned long q = (unsigned long)p;
+		return bit_waitqueue((void *)(q & ~1), q & 1);
+	}
+	return bit_waitqueue(p, 0);
+}
+
+static int wake_atomic_t_function(wait_queue_t *wait, unsigned mode, int sync,
+				  void *arg)
+{
+	struct wait_bit_key *key = arg;
+	struct wait_bit_queue *wait_bit
+		= container_of(wait, struct wait_bit_queue, wait);
+	atomic_t *val = key->flags;
+
+	if (wait_bit->key.flags != key->flags ||
+	    wait_bit->key.bit_nr != key->bit_nr ||
+	    atomic_read(val) != 0)
+		return 0;
+	return autoremove_wake_function(wait, mode, sync, key);
+}
+
+/*
+ * To allow interruptible waiting and asynchronous (i.e. nonblocking) waiting,
+ * the actions of __wait_on_atomic_t() are permitted return codes.  Nonzero
+ * return codes halt waiting and return.
+ */
+static __sched
+int __wait_on_atomic_t(wait_queue_head_t *wq, struct wait_bit_queue *q,
+		       int (*action)(atomic_t *), unsigned mode)
+{
+	atomic_t *val;
+	int ret = 0;
+
+	do {
+		prepare_to_wait(wq, &q->wait, mode);
+		val = q->key.flags;
+		if (atomic_read(val) == 0)
+			ret = (*action)(val);
+	} while (!ret && atomic_read(val) != 0);
+	finish_wait(wq, &q->wait);
+	return ret;
+}
+
+#define DEFINE_WAIT_ATOMIC_T(name, p)					\
+	struct wait_bit_queue name = {					\
+		.key = __WAIT_ATOMIC_T_KEY_INITIALIZER(p),		\
+		.wait	= {						\
+			.private	= current,			\
+			.func		= wake_atomic_t_function,	\
+			.task_list	=				\
+				LIST_HEAD_INIT((name).wait.task_list),	\
+		},							\
+	}
+
+__sched int out_of_line_wait_on_atomic_t(atomic_t *p, int (*action)(atomic_t *),
+					 unsigned mode)
+{
+	wait_queue_head_t *wq = atomic_t_waitqueue(p);
+	DEFINE_WAIT_ATOMIC_T(wait, p);
+
+	return __wait_on_atomic_t(wq, &wait, action, mode);
+}
+EXPORT_SYMBOL(out_of_line_wait_on_atomic_t);
+
+/**
+ * wake_up_atomic_t - Wake up a waiter on a atomic_t
+ * @word: The word being waited on, a kernel virtual address
+ * @bit: The bit of the word being waited on
+ *
+ * Wake up anyone waiting for the atomic_t to go to zero.
+ *
+ * Abuse the bit-waker function and its waitqueue hash table set (the atomic_t
+ * check is done by the waiter's wake function, not the by the waker itself).
+ */
+void wake_up_atomic_t(atomic_t *p)
+{
+	__wake_up_bit(atomic_t_waitqueue(p), p, WAIT_ATOMIC_T_BIT_NR);
+}
+EXPORT_SYMBOL(wake_up_atomic_t);