dlm: slow down filling up processing queue

If there is a burst of message the receive worker will filling up the
processing queue but where are too slow to process dlm messages. This
patch will slow down the receiver worker to keep the buffer on the
socket layer to tell the sender to backoff. This is done by a threshold
to get the next buffers from the socket after all messages were
processed done by a flush_workqueue(). This however only occurs when we
have a message burst when we e.g. create 1 million locks. If we put more
and more new messages to process in the processqueue we will soon run out
of memory.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>

1 files changed, 12 insertions, 0 deletions

diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index f7bc22e74db2..67f8dd8a05ef 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c
@@ -63,6 +63,7 @@
 #include "config.h"
 
 #define DLM_SHUTDOWN_WAIT_TIMEOUT msecs_to_jiffies(5000)
+#define DLM_MAX_PROCESS_BUFFERS 24
 #define NEEDED_RMEM (4*1024*1024)
 
 struct connection {
@@ -194,6 +195,7 @@ static const struct dlm_proto_ops *dlm_proto_ops;
 #define DLM_IO_END 1
 #define DLM_IO_EOF 2
 #define DLM_IO_RESCHED 3
+#define DLM_IO_FLUSH 4
 
 static void process_recv_sockets(struct work_struct *work);
 static void process_send_sockets(struct work_struct *work);
@@ -202,6 +204,7 @@ static void process_dlm_messages(struct work_struct *work);
 static DECLARE_WORK(process_work, process_dlm_messages);
 static DEFINE_SPINLOCK(processqueue_lock);
 static bool process_dlm_messages_pending;
+static atomic_t processqueue_count;
 static LIST_HEAD(processqueue);
 
 bool dlm_lowcomms_is_running(void)
@@ -874,6 +877,7 @@ static void process_dlm_messages(struct work_struct *work)
 	}
 
 	list_del(&pentry->list);
+	atomic_dec(&processqueue_count);
 	spin_unlock(&processqueue_lock);
 
 	for (;;) {
@@ -891,6 +895,7 @@ static void process_dlm_messages(struct work_struct *work)
 		}
 
 		list_del(&pentry->list);
+		atomic_dec(&processqueue_count);
 		spin_unlock(&processqueue_lock);
 	}
 }
@@ -962,6 +967,7 @@ again:
 		con->rx_leftover);
 
 	spin_lock(&processqueue_lock);
+	ret = atomic_inc_return(&processqueue_count);
 	list_add_tail(&pentry->list, &processqueue);
 	if (!process_dlm_messages_pending) {
 		process_dlm_messages_pending = true;
@@ -969,6 +975,9 @@ again:
 	}
 	spin_unlock(&processqueue_lock);
 
+	if (ret > DLM_MAX_PROCESS_BUFFERS)
+		return DLM_IO_FLUSH;
+
 	return DLM_IO_SUCCESS;
 }
 
@@ -1503,6 +1512,9 @@ static void process_recv_sockets(struct work_struct *work)
 		wake_up(&con->shutdown_wait);
 		/* CF_RECV_PENDING cleared */
 		break;
+	case DLM_IO_FLUSH:
+		flush_workqueue(process_workqueue);
+		fallthrough;
 	case DLM_IO_RESCHED:
 		cond_resched();
 		queue_work(io_workqueue, &con->rwork);