[PATCH] eventfd: convert global percpu eventfd_wake_count to ctx percpu eventfd_wake_count

[qiang.zhang@xxxxxxxxxxxxx]

From: Zqiang <qiang.zhang@xxxxxxxxxxxxx>

In RT system, the spinlock_irq be replaced by rt_mutex, when
call eventfd_signal(), if the current task is preempted after
increasing the current CPU eventfd_wake_count, when other task
run on this CPU and  call eventfd_signal(), find this CPU
eventfd_wake_count is not zero, will trigger warning and direct
return, miss wakeup.
In no-RT system, even if the eventfd_signal() call is nested, if
if it's different eventfd_ctx object, it is not happen deadlock.

Fixes: b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth")
Signed-off-by: Zqiang <qiang.zhang@xxxxxxxxxxxxx>
---
 fs/aio.c                |  2 +-
 fs/eventfd.c            | 21 +++++++++++++++++----
 include/linux/eventfd.h |  9 ++-------
 3 files changed, 20 insertions(+), 12 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 76ce0cc3ee4e..b45983d5d35a 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1695,7 +1695,7 @@ static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
 		list_del(&iocb->ki_list);
 		iocb->ki_res.res = mangle_poll(mask);
 		req->done = true;
-		if (iocb->ki_eventfd && eventfd_signal_count()) {
+		if (iocb->ki_eventfd && eventfd_signal_count(iocb->ki_eventfd)) {
 			iocb = NULL;
 			INIT_WORK(&req->work, aio_poll_put_work);
 			schedule_work(&req->work);
diff --git a/fs/eventfd.c b/fs/eventfd.c
index e265b6dd4f34..ef92d3dedde8 100644
--- a/fs/eventfd.c
+++ b/fs/eventfd.c
@@ -25,7 +25,6 @@
 #include <linux/idr.h>
 #include <linux/uio.h>
 
-DEFINE_PER_CPU(int, eventfd_wake_count);
 
 static DEFINE_IDA(eventfd_ida);
 
@@ -43,8 +42,15 @@ struct eventfd_ctx {
 	__u64 count;
 	unsigned int flags;
 	int id;
+	int __percpu *eventfd_wake_count;
 };
 
+inline bool eventfd_signal_count(struct eventfd_ctx *ctx)
+{
+	return this_cpu_read(*ctx->eventfd_wake_count);
+}
+EXPORT_SYMBOL_GPL(eventfd_signal_count);
+
 /**
  * eventfd_signal - Adds @n to the eventfd counter.
  * @ctx: [in] Pointer to the eventfd context.
@@ -71,17 +77,17 @@ __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n)
 	 * it returns true, the eventfd_signal() call should be deferred to a
 	 * safe context.
 	 */
-	if (WARN_ON_ONCE(this_cpu_read(eventfd_wake_count)))
+	if (WARN_ON_ONCE(this_cpu_read(*ctx->eventfd_wake_count)))
 		return 0;
 
 	spin_lock_irqsave(&ctx->wqh.lock, flags);
-	this_cpu_inc(eventfd_wake_count);
+	this_cpu_inc(*ctx->eventfd_wake_count);
 	if (ULLONG_MAX - ctx->count < n)
 		n = ULLONG_MAX - ctx->count;
 	ctx->count += n;
 	if (waitqueue_active(&ctx->wqh))
 		wake_up_locked_poll(&ctx->wqh, EPOLLIN);
-	this_cpu_dec(eventfd_wake_count);
+	this_cpu_dec(*ctx->eventfd_wake_count);
 	spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 
 	return n;
@@ -92,6 +98,9 @@ static void eventfd_free_ctx(struct eventfd_ctx *ctx)
 {
 	if (ctx->id >= 0)
 		ida_simple_remove(&eventfd_ida, ctx->id);
+
+	if (ctx->eventfd_wake_count)
+		free_percpu(ctx->eventfd_wake_count);
 	kfree(ctx);
 }
 
@@ -421,6 +430,10 @@ static int do_eventfd(unsigned int count, int flags)
 	if (!ctx)
 		return -ENOMEM;
 
+	ctx->eventfd_wake_count = alloc_percpu(int);
+	if (!ctx->eventfd_wake_count)
+		goto err;
+
 	kref_init(&ctx->kref);
 	init_waitqueue_head(&ctx->wqh);
 	ctx->count = count;
diff --git a/include/linux/eventfd.h b/include/linux/eventfd.h
index fa0a524baed0..1deda815ef1b 100644
--- a/include/linux/eventfd.h
+++ b/include/linux/eventfd.h
@@ -43,12 +43,7 @@ int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *w
 				  __u64 *cnt);
 void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt);
 
-DECLARE_PER_CPU(int, eventfd_wake_count);
-
-static inline bool eventfd_signal_count(void)
-{
-	return this_cpu_read(eventfd_wake_count);
-}
+inline bool eventfd_signal_count(struct eventfd_ctx *ctx);
 
 #else /* CONFIG_EVENTFD */
 
@@ -78,7 +73,7 @@ static inline int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx,
 	return -ENOSYS;
 }
 
-static inline bool eventfd_signal_count(void)
+static inline bool eventfd_signal_count(struct eventfd_ctx *ctx)
 {
 	return false;
 }
-- 
2.17.1