IB/ehca: Prevent RDMA-related connection failures on some eHCA2 hardware
Some HW revisions of eHCA2 may cause an RC connection to break if they
received RDMA Reads over that connection before. This can be
prevented by assuring that, after the first RDMA Read, the QP receives
a new RDMA Read every few million link packets.
Include code into the driver that inserts an empty (size 0) RDMA Read
into the message stream every now and then if the consumer doesn't
post them frequently enough.
Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
diff --git a/drivers/infiniband/hw/ehca/ehca_reqs.c b/drivers/infiniband/hw/ehca/ehca_reqs.c
index ea91360..3aacc8c 100644
--- a/drivers/infiniband/hw/ehca/ehca_reqs.c
+++ b/drivers/infiniband/hw/ehca/ehca_reqs.c
@@ -50,6 +50,9 @@
#include "hcp_if.h"
#include "hipz_fns.h"
+/* in RC traffic, insert an empty RDMA READ every this many packets */
+#define ACK_CIRC_THRESHOLD 2000000
+
static inline int ehca_write_rwqe(struct ipz_queue *ipz_rqueue,
struct ehca_wqe *wqe_p,
struct ib_recv_wr *recv_wr)
@@ -81,7 +84,7 @@
if (ehca_debug_level) {
ehca_gen_dbg("RECEIVE WQE written into ipz_rqueue=%p",
ipz_rqueue);
- ehca_dmp( wqe_p, 16*(6 + wqe_p->nr_of_data_seg), "recv wqe");
+ ehca_dmp(wqe_p, 16*(6 + wqe_p->nr_of_data_seg), "recv wqe");
}
return 0;
@@ -135,7 +138,8 @@
static inline int ehca_write_swqe(struct ehca_qp *qp,
struct ehca_wqe *wqe_p,
- const struct ib_send_wr *send_wr)
+ const struct ib_send_wr *send_wr,
+ int hidden)
{
u32 idx;
u64 dma_length;
@@ -176,7 +180,9 @@
wqe_p->wr_flag = 0;
- if (send_wr->send_flags & IB_SEND_SIGNALED)
+ if ((send_wr->send_flags & IB_SEND_SIGNALED ||
+ qp->init_attr.sq_sig_type == IB_SIGNAL_ALL_WR)
+ && !hidden)
wqe_p->wr_flag |= WQE_WRFLAG_REQ_SIGNAL_COM;
if (send_wr->opcode == IB_WR_SEND_WITH_IMM ||
@@ -199,7 +205,7 @@
wqe_p->destination_qp_number = send_wr->wr.ud.remote_qpn << 8;
wqe_p->local_ee_context_qkey = remote_qkey;
- if (!send_wr->wr.ud.ah) {
+ if (unlikely(!send_wr->wr.ud.ah)) {
ehca_gen_err("wr.ud.ah is NULL. qp=%p", qp);
return -EINVAL;
}
@@ -255,6 +261,15 @@
} /* eof idx */
wqe_p->u.nud.atomic_1st_op_dma_len = dma_length;
+ /* unsolicited ack circumvention */
+ if (send_wr->opcode == IB_WR_RDMA_READ) {
+ /* on RDMA read, switch on and reset counters */
+ qp->message_count = qp->packet_count = 0;
+ qp->unsol_ack_circ = 1;
+ } else
+ /* else estimate #packets */
+ qp->packet_count += (dma_length >> qp->mtu_shift) + 1;
+
break;
default:
@@ -355,13 +370,49 @@
*wc_status = IB_WC_SUCCESS;
}
+static inline int post_one_send(struct ehca_qp *my_qp,
+ struct ib_send_wr *cur_send_wr,
+ struct ib_send_wr **bad_send_wr,
+ int hidden)
+{
+ struct ehca_wqe *wqe_p;
+ int ret;
+ u64 start_offset = my_qp->ipz_squeue.current_q_offset;
+
+ /* get pointer next to free WQE */
+ wqe_p = ipz_qeit_get_inc(&my_qp->ipz_squeue);
+ if (unlikely(!wqe_p)) {
+ /* too many posted work requests: queue overflow */
+ if (bad_send_wr)
+ *bad_send_wr = cur_send_wr;
+ ehca_err(my_qp->ib_qp.device, "Too many posted WQEs "
+ "qp_num=%x", my_qp->ib_qp.qp_num);
+ return -ENOMEM;
+ }
+ /* write a SEND WQE into the QUEUE */
+ ret = ehca_write_swqe(my_qp, wqe_p, cur_send_wr, hidden);
+ /*
+ * if something failed,
+ * reset the free entry pointer to the start value
+ */
+ if (unlikely(ret)) {
+ my_qp->ipz_squeue.current_q_offset = start_offset;
+ if (bad_send_wr)
+ *bad_send_wr = cur_send_wr;
+ ehca_err(my_qp->ib_qp.device, "Could not write WQE "
+ "qp_num=%x", my_qp->ib_qp.qp_num);
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
int ehca_post_send(struct ib_qp *qp,
struct ib_send_wr *send_wr,
struct ib_send_wr **bad_send_wr)
{
struct ehca_qp *my_qp = container_of(qp, struct ehca_qp, ib_qp);
struct ib_send_wr *cur_send_wr;
- struct ehca_wqe *wqe_p;
int wqe_cnt = 0;
int ret = 0;
unsigned long flags;
@@ -369,37 +420,33 @@
/* LOCK the QUEUE */
spin_lock_irqsave(&my_qp->spinlock_s, flags);
+ /* Send an empty extra RDMA read if:
+ * 1) there has been an RDMA read on this connection before
+ * 2) no RDMA read occurred for ACK_CIRC_THRESHOLD link packets
+ * 3) we can be sure that any previous extra RDMA read has been
+ * processed so we don't overflow the SQ
+ */
+ if (unlikely(my_qp->unsol_ack_circ &&
+ my_qp->packet_count > ACK_CIRC_THRESHOLD &&
+ my_qp->message_count > my_qp->init_attr.cap.max_send_wr)) {
+ /* insert an empty RDMA READ to fix up the remote QP state */
+ struct ib_send_wr circ_wr;
+ memset(&circ_wr, 0, sizeof(circ_wr));
+ circ_wr.opcode = IB_WR_RDMA_READ;
+ post_one_send(my_qp, &circ_wr, NULL, 1); /* ignore retcode */
+ wqe_cnt++;
+ ehca_dbg(qp->device, "posted circ wr qp_num=%x", qp->qp_num);
+ my_qp->message_count = my_qp->packet_count = 0;
+ }
+
/* loop processes list of send reqs */
for (cur_send_wr = send_wr; cur_send_wr != NULL;
cur_send_wr = cur_send_wr->next) {
- u64 start_offset = my_qp->ipz_squeue.current_q_offset;
- /* get pointer next to free WQE */
- wqe_p = ipz_qeit_get_inc(&my_qp->ipz_squeue);
- if (unlikely(!wqe_p)) {
- /* too many posted work requests: queue overflow */
- if (bad_send_wr)
- *bad_send_wr = cur_send_wr;
- if (wqe_cnt == 0) {
- ret = -ENOMEM;
- ehca_err(qp->device, "Too many posted WQEs "
- "qp_num=%x", qp->qp_num);
- }
- goto post_send_exit0;
- }
- /* write a SEND WQE into the QUEUE */
- ret = ehca_write_swqe(my_qp, wqe_p, cur_send_wr);
- /*
- * if something failed,
- * reset the free entry pointer to the start value
- */
+ ret = post_one_send(my_qp, cur_send_wr, bad_send_wr, 0);
if (unlikely(ret)) {
- my_qp->ipz_squeue.current_q_offset = start_offset;
- *bad_send_wr = cur_send_wr;
- if (wqe_cnt == 0) {
- ret = -EINVAL;
- ehca_err(qp->device, "Could not write WQE "
- "qp_num=%x", qp->qp_num);
- }
+ /* if one or more WQEs were successful, don't fail */
+ if (wqe_cnt)
+ ret = 0;
goto post_send_exit0;
}
wqe_cnt++;
@@ -410,6 +457,7 @@
post_send_exit0:
iosync(); /* serialize GAL register access */
hipz_update_sqa(my_qp, wqe_cnt);
+ my_qp->message_count += wqe_cnt;
spin_unlock_irqrestore(&my_qp->spinlock_s, flags);
return ret;
}