fedora core 6 1.2949 + vserver 2.2.0
[linux-2.6.git] / drivers / infiniband / hw / ipath / ipath_ud.c
index 3466129..49f1102 100644 (file)
 #include <rdma/ib_smi.h>
 
 #include "ipath_verbs.h"
-#include "ipath_common.h"
+#include "ipath_kernel.h"
+
+static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
+                   u32 *lengthp, struct ipath_sge_state *ss)
+{
+       int user = to_ipd(qp->ibqp.pd)->user;
+       int i, j, ret;
+       struct ib_wc wc;
+
+       *lengthp = 0;
+       for (i = j = 0; i < wqe->num_sge; i++) {
+               if (wqe->sg_list[i].length == 0)
+                       continue;
+               /* Check LKEY */
+               if ((user && wqe->sg_list[i].lkey == 0) ||
+                   !ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
+                                  &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
+                       goto bad_lkey;
+               *lengthp += wqe->sg_list[i].length;
+               j++;
+       }
+       ss->num_sge = j;
+       ret = 1;
+       goto bail;
+
+bad_lkey:
+       wc.wr_id = wqe->wr_id;
+       wc.status = IB_WC_LOC_PROT_ERR;
+       wc.opcode = IB_WC_RECV;
+       wc.vendor_err = 0;
+       wc.byte_len = 0;
+       wc.imm_data = 0;
+       wc.qp_num = qp->ibqp.qp_num;
+       wc.src_qp = 0;
+       wc.wc_flags = 0;
+       wc.pkey_index = 0;
+       wc.slid = 0;
+       wc.sl = 0;
+       wc.dlid_path_bits = 0;
+       wc.port_num = 0;
+       /* Signal solicited completion event. */
+       ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
+       ret = 0;
+bail:
+       return ret;
+}
 
 /**
  * ipath_ud_loopback - handle send on loopback QPs
@@ -46,6 +91,8 @@
  *
  * This is called from ipath_post_ud_send() to forward a WQE addressed
  * to the same HCA.
+ * Note that the receive interrupt handler may be calling ipath_ud_rcv()
+ * while this is being called.
  */
 static void ipath_ud_loopback(struct ipath_qp *sqp,
                              struct ipath_sge_state *ss,
@@ -60,7 +107,11 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
        struct ipath_srq *srq;
        struct ipath_sge_state rsge;
        struct ipath_sge *sge;
+       struct ipath_rwq *wq;
        struct ipath_rwqe *wqe;
+       void (*handler)(struct ib_event *, void *);
+       u32 tail;
+       u32 rlen;
 
        qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
        if (!qp)
@@ -94,6 +145,13 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
                wc->imm_data = 0;
        }
 
+       if (wr->num_sge > 1) {
+               rsge.sg_list = kmalloc((wr->num_sge - 1) *
+                                       sizeof(struct ipath_sge),
+                                      GFP_ATOMIC);
+       } else
+               rsge.sg_list = NULL;
+
        /*
         * Get the next work request entry to find where to put the data.
         * Note that it is safe to drop the lock after changing rq->tail
@@ -101,37 +159,52 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
         */
        if (qp->ibqp.srq) {
                srq = to_isrq(qp->ibqp.srq);
+               handler = srq->ibsrq.event_handler;
                rq = &srq->rq;
        } else {
                srq = NULL;
+               handler = NULL;
                rq = &qp->r_rq;
        }
+
        spin_lock_irqsave(&rq->lock, flags);
-       if (rq->tail == rq->head) {
-               spin_unlock_irqrestore(&rq->lock, flags);
-               dev->n_pkt_drops++;
-               goto done;
+       wq = rq->wq;
+       tail = wq->tail;
+       while (1) {
+               if (unlikely(tail == wq->head)) {
+                       spin_unlock_irqrestore(&rq->lock, flags);
+                       dev->n_pkt_drops++;
+                       goto bail_sge;
+               }
+               wqe = get_rwqe_ptr(rq, tail);
+               if (++tail >= rq->size)
+                       tail = 0;
+               if (init_sge(qp, wqe, &rlen, &rsge))
+                       break;
+               wq->tail = tail;
        }
        /* Silently drop packets which are too big. */
-       wqe = get_rwqe_ptr(rq, rq->tail);
-       if (wc->byte_len > wqe->length) {
+       if (wc->byte_len > rlen) {
                spin_unlock_irqrestore(&rq->lock, flags);
                dev->n_pkt_drops++;
-               goto done;
+               goto bail_sge;
        }
+       wq->tail = tail;
        wc->wr_id = wqe->wr_id;
-       rsge.sge = wqe->sg_list[0];
-       rsge.sg_list = wqe->sg_list + 1;
-       rsge.num_sge = wqe->num_sge;
-       if (++rq->tail >= rq->size)
-               rq->tail = 0;
-       if (srq && srq->ibsrq.event_handler) {
+       if (handler) {
                u32 n;
 
-               if (rq->head < rq->tail)
-                       n = rq->size + rq->head - rq->tail;
+               /*
+                * validate head pointer value and compute
+                * the number of remaining WQEs.
+                */
+               n = wq->head;
+               if (n >= rq->size)
+                       n = 0;
+               if (n < tail)
+                       n += rq->size - tail;
                else
-                       n = rq->head - rq->tail;
+                       n -= tail;
                if (n < srq->limit) {
                        struct ib_event ev;
 
@@ -140,12 +213,12 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
                        ev.device = qp->ibqp.device;
                        ev.element.srq = qp->ibqp.srq;
                        ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
-                       srq->ibsrq.event_handler(&ev,
-                                                srq->ibsrq.srq_context);
+                       handler(&ev, srq->ibsrq.srq_context);
                } else
                        spin_unlock_irqrestore(&rq->lock, flags);
        } else
                spin_unlock_irqrestore(&rq->lock, flags);
+
        ah_attr = &to_iah(wr->wr.ud.ah)->attr;
        if (ah_attr->ah_flags & IB_AH_GRH) {
                ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
@@ -186,7 +259,7 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
        wc->src_qp = sqp->ibqp.qp_num;
        /* XXX do we know which pkey matched? Only needed for GSI. */
        wc->pkey_index = 0;
-       wc->slid = ipath_layer_get_lid(dev->dd) |
+       wc->slid = dev->dd->ipath_lid |
                (ah_attr->src_path_bits &
                 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
        wc->sl = ah_attr->sl;
@@ -196,6 +269,8 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
        ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
                       wr->send_flags & IB_SEND_SOLICITED);
 
+bail_sge:
+       kfree(rsge.sg_list);
 done:
        if (atomic_dec_and_test(&qp->refcount))
                wake_up(&qp->wait);
@@ -266,7 +341,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
 
                if (wr->sg_list[i].length == 0)
                        continue;
-               if (!ipath_lkey_ok(&dev->lk_table, ss.num_sge ?
+               if (!ipath_lkey_ok(qp, ss.num_sge ?
                                   sg_list + ss.num_sge - 1 : &ss.sge,
                                   &wr->sg_list[i], 0)) {
                        ret = -EINVAL;
@@ -276,7 +351,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
                ss.num_sge++;
        }
        /* Check for invalid packet size. */
-       if (len > ipath_layer_get_ibmtu(dev->dd)) {
+       if (len > dev->dd->ipath_ibmtu) {
                ret = -EINVAL;
                goto bail;
        }
@@ -298,7 +373,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
                dev->n_unicast_xmit++;
                lid = ah_attr->dlid &
                        ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
-               if (unlikely(lid == ipath_layer_get_lid(dev->dd))) {
+               if (unlikely(lid == dev->dd->ipath_lid)) {
                        /*
                         * Pass in an uninitialized ib_wc to save stack
                         * space.
@@ -327,7 +402,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
                qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
                        dev->gid_prefix;
                qp->s_hdr.u.l.grh.sgid.global.interface_id =
-                       ipath_layer_get_guid(dev->dd);
+                       dev->dd->ipath_guid;
                qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
                /*
                 * Don't worry about sending to locally attached multicast
@@ -357,7 +432,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
        qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
        qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);  /* DEST LID */
        qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
-       lid = ipath_layer_get_lid(dev->dd);
+       lid = dev->dd->ipath_lid;
        if (lid) {
                lid |= ah_attr->src_path_bits &
                        ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
@@ -368,7 +443,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
                bth0 |= 1 << 23;
        bth0 |= extra_bytes << 20;
        bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
-               ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
+               ipath_get_pkey(dev->dd, qp->s_pkey_index);
        ohdr->bth[0] = cpu_to_be32(bth0);
        /*
         * Use the multicast QP if the destination LID is a multicast LID.
@@ -433,13 +508,9 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
        int opcode;
        u32 hdrsize;
        u32 pad;
-       unsigned long flags;
        struct ib_wc wc;
        u32 qkey;
        u32 src_qp;
-       struct ipath_rq *rq;
-       struct ipath_srq *srq;
-       struct ipath_rwqe *wqe;
        u16 dlid;
        int header_in_data;
 
@@ -458,8 +529,7 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
                 * the eager header buffer size to 56 bytes so the last 12
                 * bytes of the IB header is in the data buffer.
                 */
-               header_in_data =
-                       ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
+               header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
                if (header_in_data) {
                        qkey = be32_to_cpu(((__be32 *) data)[1]);
                        src_qp = be32_to_cpu(((__be32 *) data)[2]);
@@ -547,19 +617,10 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 
        /*
         * Get the next work request entry to find where to put the data.
-        * Note that it is safe to drop the lock after changing rq->tail
-        * since ipath_post_receive() won't fill the empty slot.
         */
-       if (qp->ibqp.srq) {
-               srq = to_isrq(qp->ibqp.srq);
-               rq = &srq->rq;
-       } else {
-               srq = NULL;
-               rq = &qp->r_rq;
-       }
-       spin_lock_irqsave(&rq->lock, flags);
-       if (rq->tail == rq->head) {
-               spin_unlock_irqrestore(&rq->lock, flags);
+       if (qp->r_reuse_sge)
+               qp->r_reuse_sge = 0;
+       else if (!ipath_get_rwqe(qp, 0)) {
                /*
                 * Count VL15 packets dropped due to no receive buffer.
                 * Otherwise, count them as buffer overruns since usually,
@@ -573,39 +634,11 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
                goto bail;
        }
        /* Silently drop packets which are too big. */
-       wqe = get_rwqe_ptr(rq, rq->tail);
-       if (wc.byte_len > wqe->length) {
-               spin_unlock_irqrestore(&rq->lock, flags);
+       if (wc.byte_len > qp->r_len) {
+               qp->r_reuse_sge = 1;
                dev->n_pkt_drops++;
                goto bail;
        }
-       wc.wr_id = wqe->wr_id;
-       qp->r_sge.sge = wqe->sg_list[0];
-       qp->r_sge.sg_list = wqe->sg_list + 1;
-       qp->r_sge.num_sge = wqe->num_sge;
-       if (++rq->tail >= rq->size)
-               rq->tail = 0;
-       if (srq && srq->ibsrq.event_handler) {
-               u32 n;
-
-               if (rq->head < rq->tail)
-                       n = rq->size + rq->head - rq->tail;
-               else
-                       n = rq->head - rq->tail;
-               if (n < srq->limit) {
-                       struct ib_event ev;
-
-                       srq->limit = 0;
-                       spin_unlock_irqrestore(&rq->lock, flags);
-                       ev.device = qp->ibqp.device;
-                       ev.element.srq = qp->ibqp.srq;
-                       ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
-                       srq->ibsrq.event_handler(&ev,
-                                                srq->ibsrq.srq_context);
-               } else
-                       spin_unlock_irqrestore(&rq->lock, flags);
-       } else
-               spin_unlock_irqrestore(&rq->lock, flags);
        if (has_grh) {
                ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
                               sizeof(struct ib_grh));
@@ -614,6 +647,7 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
                ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
        ipath_copy_sge(&qp->r_sge, data,
                       wc.byte_len - sizeof(struct ib_grh));
+       wc.wr_id = qp->r_wr_id;
        wc.status = IB_WC_SUCCESS;
        wc.opcode = IB_WC_RECV;
        wc.vendor_err = 0;