diff options
| author | Sowmini Varadhan <sowmini.varadhan@oracle.com> | 2018-02-27 09:52:43 -0800 |
|---|---|---|
| committer | David S. Miller <davem@davemloft.net> | 2018-02-27 14:19:11 -0500 |
| commit | 401910db4cd425899832a093539222b6174f92a2 (patch) | |
| tree | 7eb70bb22f23c1b7edb5245e17ad0f8c81bfc333 /net/rds/af_rds.c | |
| parent | 67490e34ba2b1c02fb554a8059cd6ce12b47ccfb (diff) | |
rds: deliver zerocopy completion notification with data
This commit is an optimization over commit 01883eda72bd
("rds: support for zcopy completion notification") for PF_RDS sockets.
RDS applications are predominantly request-response transactions, so
it is more efficient to reduce the number of system calls and have
zerocopy completion notification delivered as ancillary data on the
POLLIN channel.
Cookies are passed up as ancillary data (at level SOL_RDS) in a
struct rds_zcopy_cookies when the returned value of recvmsg() is
greater than, or equal to, 0. A max of RDS_MAX_ZCOOKIES may be passed
with each message.
This commit removes support for zerocopy completion notification on
MSG_ERRQUEUE for PF_RDS sockets.
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'net/rds/af_rds.c')
| -rw-r--r-- | net/rds/af_rds.c | 7 |
1 files changed, 5 insertions, 2 deletions
diff --git a/net/rds/af_rds.c b/net/rds/af_rds.c index a937f18896ae..f7126108a811 100644 --- a/net/rds/af_rds.c +++ b/net/rds/af_rds.c @@ -77,6 +77,7 @@ static int rds_release(struct socket *sock) rds_send_drop_to(rs, NULL); rds_rdma_drop_keys(rs); rds_notify_queue_get(rs, NULL); + __skb_queue_purge(&rs->rs_zcookie_queue); spin_lock_bh(&rds_sock_lock); list_del_init(&rs->rs_item); @@ -144,7 +145,7 @@ static int rds_getname(struct socket *sock, struct sockaddr *uaddr, * - to signal that a previously congested destination may have become * uncongested * - A notification has been queued to the socket (this can be a congestion - * update, or a RDMA completion). + * update, or a RDMA completion, or a MSG_ZEROCOPY completion). * * EPOLLOUT is asserted if there is room on the send queue. This does not mean * however, that the next sendmsg() call will succeed. If the application tries @@ -178,7 +179,8 @@ static __poll_t rds_poll(struct file *file, struct socket *sock, spin_unlock(&rs->rs_lock); } if (!list_empty(&rs->rs_recv_queue) || - !list_empty(&rs->rs_notify_queue)) + !list_empty(&rs->rs_notify_queue) || + !skb_queue_empty(&rs->rs_zcookie_queue)) mask |= (EPOLLIN | EPOLLRDNORM); if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) mask |= (EPOLLOUT | EPOLLWRNORM); @@ -513,6 +515,7 @@ static int __rds_create(struct socket *sock, struct sock *sk, int protocol) INIT_LIST_HEAD(&rs->rs_recv_queue); INIT_LIST_HEAD(&rs->rs_notify_queue); INIT_LIST_HEAD(&rs->rs_cong_list); + skb_queue_head_init(&rs->rs_zcookie_queue); spin_lock_init(&rs->rs_rdma_lock); rs->rs_rdma_keys = RB_ROOT; rs->rs_rx_traces = 0; |