[PATCH] rds: rds_cong_queue_updates needs to defer the congestion update transmission

[PATCH] rds: rds_cong_queue_updates needs to defer the congestion update transmission

[Sowmini Varadhan]

This patch fixes a sock_lock deadlock in the rds_cong_queue_update path.

We cannot inline the call to rds_send_xmit from rds_cong_queue_update
because
(a) we are already holding the sock_lock in the recv path, and
    will deadlock when tcp_setsockopt/tcp_sendmsg try to get the sock
    lock
(b) cong_queue_update does an irqsave on the rds_cong_lock, and this
    will trigger warnings (for a good reason) from functions called
    out of sock_lock.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
---
 net/rds/cong.c |   16 +++++++++++++++-
 1 files changed, 15 insertions(+), 1 deletions(-)

diff --git a/net/rds/cong.c b/net/rds/cong.c
index e5b65ac..765d18f 100644
--- a/net/rds/cong.c
+++ b/net/rds/cong.c
@@ -221,7 +221,21 @@ void rds_cong_queue_updates(struct rds_cong_map *map)
 	list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
 		if (!test_and_set_bit(0, &conn->c_map_queued)) {
 			rds_stats_inc(s_cong_update_queued);
-			rds_send_xmit(conn);
+			/* We cannot inline the call to rds_send_xmit() here
+			 * for two reasons:
+			 * 1. When we get here from the receive path, we
+			 *    are already holding the sock_lock (held by
+			 *    tcp_v4_rcv()). So inlining calls to
+			 *    tcp_setsockopt and/or tcp_sendmsg will deadlock
+			 *    when it tries to get the sock_lock())
+			 * 2. Interrupts are masked so that we can mark the
+			 *    the port congested from both send and recv paths.
+			 *    (See comment around declaration of rds_cong_lock).
+			 *    An attempt to get the sock_lock() here will
+			 *    therefore trigger warnings.
+			 * Defer the xmit to rds_send_worker() instead.
+			 */
+			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 		}
 	}
 
-- 
1.7.1

Re: [PATCH] rds: rds_cong_queue_updates needs to defer the congestion update transmission

[Chuck Lever]

On Feb 10, 2015, at 9:22 AM, Sowmini Varadhan <sowmini.varadhan@oracle.com> wrote:

> 
> This patch fixes a sock_lock deadlock in the rds_cong_queue_update path.

Note that the deadlock appears to exist only with TCP transports.

> We cannot inline the call to rds_send_xmit from rds_cong_queue_update
> because
> (a) we are already holding the sock_lock in the recv path, and
>    will deadlock when tcp_setsockopt/tcp_sendmsg try to get the sock
>    lock
> (b) cong_queue_update does an irqsave on the rds_cong_lock, and this
>    will trigger warnings (for a good reason) from functions called
>    out of sock_lock.
> 
> Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
> ---
> net/rds/cong.c |   16 +++++++++++++++-
> 1 files changed, 15 insertions(+), 1 deletions(-)
> 
> diff --git a/net/rds/cong.c b/net/rds/cong.c
> index e5b65ac..765d18f 100644
> --- a/net/rds/cong.c
> +++ b/net/rds/cong.c
> @@ -221,7 +221,21 @@ void rds_cong_queue_updates(struct rds_cong_map *map)
> 	list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
> 		if (!test_and_set_bit(0, &conn->c_map_queued)) {
> 			rds_stats_inc(s_cong_update_queued);
> -			rds_send_xmit(conn);
> +			/* We cannot inline the call to rds_send_xmit() here
> +			 * for two reasons:
> +			 * 1. When we get here from the receive path, we
> +			 *    are already holding the sock_lock (held by
> +			 *    tcp_v4_rcv()). So inlining calls to
> +			 *    tcp_setsockopt and/or tcp_sendmsg will deadlock
> +			 *    when it tries to get the sock_lock())
> +			 * 2. Interrupts are masked so that we can mark the
> +			 *    the port congested from both send and recv paths.
> +			 *    (See comment around declaration of rds_cong_lock).
> +			 *    An attempt to get the sock_lock() here will
> +			 *    therefore trigger warnings.
> +			 * Defer the xmit to rds_send_worker() instead.
> +			 */
> +			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
> 		}
> 	}
> 
> -- 
> 1.7.1
> 

--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Re: [PATCH] rds: rds_cong_queue_updates needs to defer the congestion update transmission

[Sowmini Varadhan]

On (02/10/15 11:50), Chuck Lever wrote:
> > 
> > This patch fixes a sock_lock deadlock in the rds_cong_queue_update path.
> 
> Note that the deadlock appears to exist only with TCP transports.
> 

True, but the patch does no harm to IB: this actually only reverts 
the change from commit 2fa57129d.

I could update the comment to note that this is only true for TCP
transports. Do you think that would help?

--Sowmini