IB/ipath: Use PIO buffer for RC ACKs
authorRalph Campbell <ralph.campbell@qlogic.com>
Thu, 17 Apr 2008 04:09:27 +0000 (21:09 -0700)
committerRoland Dreier <rolandd@cisco.com>
Thu, 17 Apr 2008 04:09:27 +0000 (21:09 -0700)
This reduces the latency for RC ACKs when a PIO buffer is available.

Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_rc.c

index 40f3e37..f765d48 100644 (file)
@@ -31,6 +31,8 @@
  * SOFTWARE.
  */
 
+#include <linux/io.h>
+
 #include "ipath_verbs.h"
 #include "ipath_kernel.h"
 
@@ -585,19 +587,39 @@ bail:
 static void send_rc_ack(struct ipath_qp *qp)
 {
        struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+       struct ipath_devdata *dd;
        u16 lrh0;
        u32 bth0;
        u32 hwords;
+       u32 __iomem *piobuf;
        struct ipath_ib_header hdr;
        struct ipath_other_headers *ohdr;
        unsigned long flags;
 
+       spin_lock_irqsave(&qp->s_lock, flags);
+
        /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
        if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
            (qp->s_flags & IPATH_S_ACK_PENDING) ||
            qp->s_ack_state != OP(ACKNOWLEDGE))
                goto queue_ack;
 
+       spin_unlock_irqrestore(&qp->s_lock, flags);
+
+       dd = dev->dd;
+       piobuf = ipath_getpiobuf(dd, 0, NULL);
+       if (!piobuf) {
+               /*
+                * We are out of PIO buffers at the moment.
+                * Pass responsibility for sending the ACK to the
+                * send tasklet so that when a PIO buffer becomes
+                * available, the ACK is sent ahead of other outgoing
+                * packets.
+                */
+               spin_lock_irqsave(&qp->s_lock, flags);
+               goto queue_ack;
+       }
+
        /* Construct the header. */
        ohdr = &hdr.u.oth;
        lrh0 = IPATH_LRH_BTH;
@@ -611,7 +633,7 @@ static void send_rc_ack(struct ipath_qp *qp)
                lrh0 = IPATH_LRH_GRH;
        }
        /* read pkey_index w/o lock (its atomic) */
-       bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
+       bth0 = ipath_get_pkey(dd, qp->s_pkey_index) |
                (OP(ACKNOWLEDGE) << 24) | (1 << 22);
        if (qp->r_nak_state)
                ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
@@ -623,30 +645,29 @@ static void send_rc_ack(struct ipath_qp *qp)
        hdr.lrh[0] = cpu_to_be16(lrh0);
        hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
        hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
-       hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
+       hdr.lrh[3] = cpu_to_be16(dd->ipath_lid);
        ohdr->bth[0] = cpu_to_be32(bth0);
        ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
        ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
 
-       /*
-        * If we can send the ACK, clear the ACK state.
-        */
-       if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
-               dev->n_unicast_xmit++;
-               goto done;
-       }
+       writeq(hwords + 1, piobuf);
 
-       /*
-        * We are out of PIO buffers at the moment.
-        * Pass responsibility for sending the ACK to the
-        * send tasklet so that when a PIO buffer becomes
-        * available, the ACK is sent ahead of other outgoing
-        * packets.
-        */
-       dev->n_rc_qacks++;
+       if (dd->ipath_flags & IPATH_PIO_FLUSH_WC) {
+               u32 *hdrp = (u32 *) &hdr;
+
+               ipath_flush_wc();
+               __iowrite32_copy(piobuf + 2, hdrp, hwords - 1);
+               ipath_flush_wc();
+               __raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
+       } else
+               __iowrite32_copy(piobuf + 2, (u32 *) &hdr, hwords);
+
+       ipath_flush_wc();
+
+       dev->n_unicast_xmit++;
+       goto done;
 
 queue_ack:
-       spin_lock_irqsave(&qp->s_lock, flags);
        dev->n_rc_qacks++;
        qp->s_flags |= IPATH_S_ACK_PENDING;
        qp->s_nak_state = qp->r_nak_state;