[PATCH 1/1] rapidio/rio_cm: avoid GFP_KERNEL in atomic context

[Alexandre Bounine]

As reported by Alexey Khoroshilov <khoroshilov@xxxxxxxxx>
(see https://lkml.org/lkml/2016/9/9/737):
riocm_send_close() is called from rio_cm_shutdown() under
spin_lock_bh(idr_lock), but riocm_send_close() uses a GFP_KERNEL
allocation.

Fix by taking riocm_send_close() outside of spinlock protected code.

Reported-by: Alexey Khoroshilov <khoroshilov@xxxxxxxxx>
Cc: Alexey Khoroshilov <khoroshilov@xxxxxxxxx>
Cc: linux-kernel@xxxxxxxxxxxxxxx
Signed-off-by: Alexandre Bounine <alexandre.bounine@xxxxxxx>
---
 drivers/rapidio/rio_cm.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/drivers/rapidio/rio_cm.c b/drivers/rapidio/rio_cm.c
index 3226983..0e91335 100644
--- a/drivers/rapidio/rio_cm.c
+++ b/drivers/rapidio/rio_cm.c
@@ -2242,17 +2242,31 @@ static int rio_cm_shutdown(struct notifier_block *nb, unsigned long code,
 {
 	struct rio_channel *ch;
 	unsigned int i;
+	LIST_HEAD(list);
 
 	riocm_debug(EXIT, ".");
 
+	/*
+	 * If there are any channels left in connected state send
+	 * close notification to the connection partner.
+	 * First build a list of channels that require a closing
+	 * notification because function riocm_send_close() should
+	 * be called outside of spinlock protected code.
+	 */
 	spin_lock_bh(&idr_lock);
 	idr_for_each_entry(&ch_idr, ch, i) {
-		riocm_debug(EXIT, "close ch %d", ch->id);
-		if (ch->state == RIO_CM_CONNECTED)
-			riocm_send_close(ch);
+		if (ch->state == RIO_CM_CONNECTED) {
+			riocm_debug(EXIT, "close ch %d", ch->id);
+			idr_remove(&ch_idr, ch->id);
+			list_add(&ch->ch_node, &list);
+		}
 	}
 	spin_unlock_bh(&idr_lock);
 
+	if (!list_empty(&list))
+		list_for_each_entry(ch, &list, ch_node)
+			riocm_send_close(ch);
+
 	return NOTIFY_DONE;
 }
 
-- 
2.9.0