static void rds_iw_qp_event_handler(struct ib_event *event, void *data)
{
struct rds_connection *conn = data;
struct rds_iw_connection *ic = conn->c_transport_data;
rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
switch (event->event) {
case IB_EVENT_COMM_EST:
rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
break;
case IB_EVENT_QP_REQ_ERR:
case IB_EVENT_QP_FATAL:
default:
rds_iw_conn_error(conn, "RDS/IW: Fatal QP Event %u - connection %pI4->%pI4...reconnecting\n",
event->event, &conn->c_laddr,
&conn->c_faddr);
break;
}
}
/*
* The _oldest/_free ring operations here race cleanly with the alloc/unalloc
* operations performed in the send path. As the sender allocs and potentially
* unallocs the next free entry in the ring it doesn't alter which is
* the next to be freed, which is what this is concerned with.
*/
void rds_iw_send_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_send_work *send;
u32 completed;
u32 oldest;
u32 i;
int ret;
rdsdebug("cq %p conn %p\n", cq, conn);
rds_iw_stats_inc(s_iw_tx_cq_call);
ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
if (ret)
rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
while (ib_poll_cq(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));
rds_iw_stats_inc(s_iw_tx_cq_event);
if (wc.status != IB_WC_SUCCESS) {
printk(KERN_ERR "WC Error: status = %d opcode = %d\n", wc.status, wc.opcode);
break;
}
if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
ic->i_fastreg_posted = 0;
continue;
}
if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
ic->i_fastreg_posted = 1;
continue;
}
if (wc.wr_id == RDS_IW_ACK_WR_ID) {
if (ic->i_ack_queued + HZ/2 < jiffies)
rds_iw_stats_inc(s_iw_tx_stalled);
rds_iw_ack_send_complete(ic);
continue;
}
oldest = rds_iw_ring_oldest(&ic->i_send_ring);
completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
for (i = 0; i < completed; i++) {
send = &ic->i_sends[oldest];
/* In the error case, wc.opcode sometimes contains garbage */
switch (send->s_wr.opcode) {
case IB_WR_SEND:
if (send->s_rm)
rds_iw_send_unmap_rm(ic, send, wc.status);
break;
case IB_WR_FAST_REG_MR:
case IB_WR_RDMA_WRITE:
case IB_WR_RDMA_READ:
case IB_WR_RDMA_READ_WITH_INV:
/* Nothing to be done - the SG list will be unmapped
* when the SEND completes. */
break;
default:
if (printk_ratelimit())
printk(KERN_NOTICE
"RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
__func__, send->s_wr.opcode);
break;
}
send->s_wr.opcode = 0xdead;
send->s_wr.num_sge = 1;
if (send->s_queued + HZ/2 < jiffies)
rds_iw_stats_inc(s_iw_tx_stalled);
/* If a RDMA operation produced an error, signal this right
* away. If we don't, the subsequent SEND that goes with this
* RDMA will be canceled with ERR_WFLUSH, and the application
* never learn that the RDMA failed. */
if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
struct rds_message *rm;
rm = rds_send_get_message(conn, send->s_op);
if (rm)
rds_iw_send_rdma_complete(rm, wc.status);
}
oldest = (oldest + 1) % ic->i_send_ring.w_nr;
}
rds_iw_ring_free(&ic->i_send_ring, completed);
if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags)
|| test_bit(0, &conn->c_map_queued))
queue_delayed_work(rds_wq, &conn->c_send_w, 0);
/* We expect errors as the qp is drained during shutdown */
if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
rds_iw_conn_error(conn,
"send completion on %u.%u.%u.%u "
"had status %u, disconnecting and reconnecting\n",
NIPQUAD(conn->c_faddr), wc.status);
}
}
}
