[PATCH 3/3] aio-use-cancellation-list-lazily-fix

[Kent Overstreet <koverstreet@google.com>]

The cancellation changes were fubar - we can't cancel a kiocb if it
doesn't actually have a cancellation callback.

The use of xchg() in aio_complete() was right - there we're marking the
kiocb as completed - but we need to use cmpxchg() in kiocb_cancel() - a
lock isn't sufficient since we're synchronizing with aio_complete()
which isn't taking any locks.

Signed-off-by: Kent Overstreet <koverstreet@google.com>
---
 fs/aio.c            | 32 ++++++++++++++++++++++----------
 include/linux/aio.h | 11 +++++++++++
 2 files changed, 33 insertions(+), 10 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 85d1b38..2760180 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -244,28 +244,40 @@ static int aio_setup_ring(struct kioctx *ctx)
 
 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
 {
-	if (!req->ki_list.next) {
-		struct kioctx *ctx = req->ki_ctx;
-		unsigned long flags;
+	struct kioctx *ctx = req->ki_ctx;
+	unsigned long flags;
 
-		spin_lock_irqsave(&ctx->ctx_lock, flags);
+	spin_lock_irqsave(&ctx->ctx_lock, flags);
+
+	if (!req->ki_list.next)
 		list_add(&req->ki_list, &ctx->active_reqs);
-		spin_unlock_irqrestore(&ctx->ctx_lock, flags);
-	}
 
 	req->ki_cancel = cancel;
+
+	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 }
 EXPORT_SYMBOL(kiocb_set_cancel_fn);
 
 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
 			struct io_event *res)
 {
-	kiocb_cancel_fn *cancel;
+	kiocb_cancel_fn *old, *cancel;
 	int ret = -EINVAL;
 
-	cancel = xchg(&kiocb->ki_cancel, KIOCB_CANCELLED);
-	if (!cancel || cancel == KIOCB_CANCELLED)
-		return ret;
+	/*
+	 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
+	 * actually has a cancel function, hence the cmpxchg()
+	 */
+
+	cancel = ACCESS_ONCE(kiocb->ki_cancel);
+	do {
+		if (!cancel || cancel == KIOCB_CANCELLED)
+			return ret;
+
+		BUG();
+		old = cancel;
+		cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
+	} while (cancel != old);
 
 	atomic_inc(&kiocb->ki_users);
 	spin_unlock_irq(&ctx->ctx_lock);
diff --git a/include/linux/aio.h b/include/linux/aio.h
index 8eaa490..cd4aea0 100644
--- a/include/linux/aio.h
+++ b/include/linux/aio.h
@@ -15,6 +15,17 @@ struct batch_complete;
 
 #define KIOCB_KEY		0
 
+/*
+ * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
+ * cancelled or completed (this makes a certain amount of sense because
+ * successful cancellation - io_cancel() - does deliver the completion to
+ * userspace).
+ *
+ * And since most things don't implement kiocb cancellation and we'd really like
+ * kiocb completion to be lockless when possible, we use ki_cancel to
+ * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
+ * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
+ */
 #define KIOCB_CANCELLED		((void *) (~0ULL))
 
 typedef int (kiocb_cancel_fn)(struct kiocb *, struct io_event *);
-- 
1.7.12