RDS/IB+IW: Move recv processing to a tasklet
authorAndy Grover <andy.grover@oracle.com>
Fri, 30 Oct 2009 08:51:57 +0000 (08:51 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 30 Oct 2009 22:06:39 +0000 (15:06 -0700)
Move receive processing from event handler to a tasklet.
This should help prevent hangcheck timer from going off
when RDS is under heavy load.

Signed-off-by: Andy Grover <andy.grover@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/rds/ib.h
net/rds/ib_cm.c
net/rds/ib_recv.c
net/rds/iw.h
net/rds/iw_cm.c
net/rds/iw_recv.c

index 1378b85..64df4e7 100644 (file)
@@ -98,6 +98,7 @@ struct rds_ib_connection {
        struct rds_ib_send_work *i_sends;
 
        /* rx */
+       struct tasklet_struct   i_recv_tasklet;
        struct mutex            i_recv_mutex;
        struct rds_ib_work_ring i_recv_ring;
        struct rds_ib_incoming  *i_ibinc;
@@ -303,6 +304,7 @@ void rds_ib_inc_free(struct rds_incoming *inc);
 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
                             size_t size);
 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
+void rds_ib_recv_tasklet_fn(unsigned long data);
 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
index c2d372f..9d32069 100644 (file)
@@ -694,6 +694,8 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
                return -ENOMEM;
 
        INIT_LIST_HEAD(&ic->ib_node);
+       tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
+                    (unsigned long) ic);
        mutex_init(&ic->i_recv_mutex);
 #ifndef KERNEL_HAS_ATOMIC64
        spin_lock_init(&ic->i_ack_lock);
index 2f009d3..fe5ab8c 100644 (file)
@@ -825,17 +825,22 @@ void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
 {
        struct rds_connection *conn = context;
        struct rds_ib_connection *ic = conn->c_transport_data;
-       struct ib_wc wc;
-       struct rds_ib_ack_state state = { 0, };
-       struct rds_ib_recv_work *recv;
 
        rdsdebug("conn %p cq %p\n", conn, cq);
 
        rds_ib_stats_inc(s_ib_rx_cq_call);
 
-       ib_req_notify_cq(cq, IB_CQ_SOLICITED);
+       tasklet_schedule(&ic->i_recv_tasklet);
+}
 
-       while (ib_poll_cq(cq, 1, &wc) > 0) {
+static inline void rds_poll_cq(struct rds_ib_connection *ic,
+                              struct rds_ib_ack_state *state)
+{
+       struct rds_connection *conn = ic->conn;
+       struct ib_wc wc;
+       struct rds_ib_recv_work *recv;
+
+       while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
                rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
                         (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
                         be32_to_cpu(wc.ex.imm_data));
@@ -853,7 +858,7 @@ void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
                if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
                        /* We expect errors as the qp is drained during shutdown */
                        if (wc.status == IB_WC_SUCCESS) {
-                               rds_ib_process_recv(conn, recv, wc.byte_len, &state);
+                               rds_ib_process_recv(conn, recv, wc.byte_len, state);
                        } else {
                                rds_ib_conn_error(conn, "recv completion on "
                                       "%pI4 had status %u, disconnecting and "
@@ -864,6 +869,17 @@ void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
 
                rds_ib_ring_free(&ic->i_recv_ring, 1);
        }
+}
+
+void rds_ib_recv_tasklet_fn(unsigned long data)
+{
+       struct rds_ib_connection *ic = (struct rds_ib_connection *) data;
+       struct rds_connection *conn = ic->conn;
+       struct rds_ib_ack_state state = { 0, };
+
+       rds_poll_cq(ic, &state);
+       ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
+       rds_poll_cq(ic, &state);
 
        if (state.ack_next_valid)
                rds_ib_set_ack(ic, state.ack_next, state.ack_required);
index dd72b62..eef2f0c 100644 (file)
@@ -119,6 +119,7 @@ struct rds_iw_connection {
        struct rds_iw_send_work *i_sends;
 
        /* rx */
+       struct tasklet_struct   i_recv_tasklet;
        struct mutex            i_recv_mutex;
        struct rds_iw_work_ring i_recv_ring;
        struct rds_iw_incoming  *i_iwinc;
@@ -330,6 +331,7 @@ void rds_iw_inc_free(struct rds_incoming *inc);
 int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
                             size_t size);
 void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context);
+void rds_iw_recv_tasklet_fn(unsigned long data);
 void rds_iw_recv_init_ring(struct rds_iw_connection *ic);
 void rds_iw_recv_clear_ring(struct rds_iw_connection *ic);
 void rds_iw_recv_init_ack(struct rds_iw_connection *ic);
index a416b0d..394cf6b 100644 (file)
@@ -696,6 +696,8 @@ int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
                return -ENOMEM;
 
        INIT_LIST_HEAD(&ic->iw_node);
+       tasklet_init(&ic->i_recv_tasklet, rds_iw_recv_tasklet_fn,
+                    (unsigned long) ic);
        mutex_init(&ic->i_recv_mutex);
 #ifndef KERNEL_HAS_ATOMIC64
        spin_lock_init(&ic->i_ack_lock);
index 9f98150..24fc53f 100644 (file)
@@ -784,17 +784,22 @@ void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context)
 {
        struct rds_connection *conn = context;
        struct rds_iw_connection *ic = conn->c_transport_data;
-       struct ib_wc wc;
-       struct rds_iw_ack_state state = { 0, };
-       struct rds_iw_recv_work *recv;
 
        rdsdebug("conn %p cq %p\n", conn, cq);
 
        rds_iw_stats_inc(s_iw_rx_cq_call);
 
-       ib_req_notify_cq(cq, IB_CQ_SOLICITED);
+       tasklet_schedule(&ic->i_recv_tasklet);
+}
 
-       while (ib_poll_cq(cq, 1, &wc) > 0) {
+static inline void rds_poll_cq(struct rds_iw_connection *ic,
+                              struct rds_iw_ack_state *state)
+{
+       struct rds_connection *conn = ic->conn;
+       struct ib_wc wc;
+       struct rds_iw_recv_work *recv;
+
+       while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
                rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
                         (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
                         be32_to_cpu(wc.ex.imm_data));
@@ -812,7 +817,7 @@ void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context)
                if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
                        /* We expect errors as the qp is drained during shutdown */
                        if (wc.status == IB_WC_SUCCESS) {
-                               rds_iw_process_recv(conn, recv, wc.byte_len, &state);
+                               rds_iw_process_recv(conn, recv, wc.byte_len, state);
                        } else {
                                rds_iw_conn_error(conn, "recv completion on "
                                       "%pI4 had status %u, disconnecting and "
@@ -823,6 +828,17 @@ void rds_iw_recv_cq_comp_handler(struct ib_cq *cq, void *context)
 
                rds_iw_ring_free(&ic->i_recv_ring, 1);
        }
+}
+
+void rds_iw_recv_tasklet_fn(unsigned long data)
+{
+       struct rds_iw_connection *ic = (struct rds_iw_connection *) data;
+       struct rds_connection *conn = ic->conn;
+       struct rds_iw_ack_state state = { 0, };
+
+       rds_poll_cq(ic, &state);
+       ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
+       rds_poll_cq(ic, &state);
 
        if (state.ack_next_valid)
                rds_iw_set_ack(ic, state.ack_next, state.ack_required);