+/*
+ * Convert IB rate to delay multiplier.
+ */
+unsigned ipath_ib_rate_to_mult(enum ib_rate rate)
+{
+ switch (rate) {
+ case IB_RATE_2_5_GBPS: return 8;
+ case IB_RATE_5_GBPS: return 4;
+ case IB_RATE_10_GBPS: return 2;
+ case IB_RATE_20_GBPS: return 1;
+ default: return 0;
+ }
+}
+
+/*
+ * Convert delay multiplier to IB rate
+ */
+static enum ib_rate ipath_mult_to_ib_rate(unsigned mult)
+{
+ switch (mult) {
+ case 8: return IB_RATE_2_5_GBPS;
+ case 4: return IB_RATE_5_GBPS;
+ case 2: return IB_RATE_10_GBPS;
+ case 1: return IB_RATE_20_GBPS;
+ default: return IB_RATE_PORT_CURRENT;
+ }
+}
+
+static inline struct ipath_verbs_txreq *get_txreq(struct ipath_ibdev *dev)
+{
+ struct ipath_verbs_txreq *tx = NULL;
+ unsigned long flags;
+
+ spin_lock_irqsave(&dev->pending_lock, flags);
+ if (!list_empty(&dev->txreq_free)) {
+ struct list_head *l = dev->txreq_free.next;
+
+ list_del(l);
+ tx = list_entry(l, struct ipath_verbs_txreq, txreq.list);
+ }
+ spin_unlock_irqrestore(&dev->pending_lock, flags);
+ return tx;
+}
+
+static inline void put_txreq(struct ipath_ibdev *dev,
+ struct ipath_verbs_txreq *tx)
+{
+ unsigned long flags;
+
+ spin_lock_irqsave(&dev->pending_lock, flags);
+ list_add(&tx->txreq.list, &dev->txreq_free);
+ spin_unlock_irqrestore(&dev->pending_lock, flags);
+}
+
+static void sdma_complete(void *cookie, int status)
+{
+ struct ipath_verbs_txreq *tx = cookie;
+ struct ipath_qp *qp = tx->qp;
+ struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+
+ /* Generate a completion queue entry if needed */
+ if (qp->ibqp.qp_type != IB_QPT_RC && tx->wqe) {
+ enum ib_wc_status ibs = status == IPATH_SDMA_TXREQ_S_OK ?
+ IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR;
+
+ ipath_send_complete(qp, tx->wqe, ibs);
+ }
+
+ if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
+ kfree(tx->txreq.map_addr);
+ put_txreq(dev, tx);
+
+ if (atomic_dec_and_test(&qp->refcount))
+ wake_up(&qp->wait);
+}
+
+/*
+ * Compute the number of clock cycles of delay before sending the next packet.
+ * The multipliers reflect the number of clocks for the fastest rate so
+ * one tick at 4xDDR is 8 ticks at 1xSDR.
+ * If the destination port will take longer to receive a packet than
+ * the outgoing link can send it, we need to delay sending the next packet
+ * by the difference in time it takes the receiver to receive and the sender
+ * to send this packet.
+ * Note that this delay is always correct for UC and RC but not always
+ * optimal for UD. For UD, the destination HCA can be different for each
+ * packet, in which case, we could send packets to a different destination
+ * while "waiting" for the delay. The overhead for doing this without
+ * HW support is more than just paying the cost of delaying some packets
+ * unnecessarily.
+ */
+static inline unsigned ipath_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult)
+{
+ return (rcv_mult > snd_mult) ?
+ (plen * (rcv_mult - snd_mult) + 1) >> 1 : 0;
+}
+
+static int ipath_verbs_send_dma(struct ipath_qp *qp,
+ struct ipath_ib_header *hdr, u32 hdrwords,
+ struct ipath_sge_state *ss, u32 len,
+ u32 plen, u32 dwords)
+{
+ struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
+ struct ipath_devdata *dd = dev->dd;
+ struct ipath_verbs_txreq *tx;
+ u32 *piobuf;
+ u32 control;
+ u32 ndesc;
+ int ret;
+
+ tx = qp->s_tx;
+ if (tx) {
+ qp->s_tx = NULL;
+ /* resend previously constructed packet */
+ ret = ipath_sdma_verbs_send(dd, tx->ss, tx->len, tx);
+ if (ret)
+ qp->s_tx = tx;
+ goto bail;
+ }
+
+ tx = get_txreq(dev);
+ if (!tx) {
+ ret = -EBUSY;
+ goto bail;
+ }
+
+ /*
+ * Get the saved delay count we computed for the previous packet
+ * and save the delay count for this packet to be used next time
+ * we get here.
+ */
+ control = qp->s_pkt_delay;
+ qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
+
+ tx->qp = qp;
+ atomic_inc(&qp->refcount);
+ tx->wqe = qp->s_wqe;
+ tx->txreq.callback = sdma_complete;
+ tx->txreq.callback_cookie = tx;
+ tx->txreq.flags = IPATH_SDMA_TXREQ_F_HEADTOHOST |
+ IPATH_SDMA_TXREQ_F_INTREQ | IPATH_SDMA_TXREQ_F_FREEDESC;
+ if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
+ tx->txreq.flags |= IPATH_SDMA_TXREQ_F_USELARGEBUF;
+
+ /* VL15 packets bypass credit check */
+ if ((be16_to_cpu(hdr->lrh[0]) >> 12) == 15) {
+ control |= 1ULL << 31;
+ tx->txreq.flags |= IPATH_SDMA_TXREQ_F_VL15;
+ }
+
+ if (len) {
+ /*
+ * Don't try to DMA if it takes more descriptors than
+ * the queue holds.
+ */
+ ndesc = ipath_count_sge(ss, len);
+ if (ndesc >= dd->ipath_sdma_descq_cnt)
+ ndesc = 0;
+ } else
+ ndesc = 1;
+ if (ndesc) {
+ tx->hdr.pbc[0] = cpu_to_le32(plen);
+ tx->hdr.pbc[1] = cpu_to_le32(control);
+ memcpy(&tx->hdr.hdr, hdr, hdrwords << 2);
+ tx->txreq.sg_count = ndesc;
+ tx->map_len = (hdrwords + 2) << 2;
+ tx->txreq.map_addr = &tx->hdr;
+ ret = ipath_sdma_verbs_send(dd, ss, dwords, tx);
+ if (ret) {
+ /* save ss and length in dwords */
+ tx->ss = ss;
+ tx->len = dwords;
+ qp->s_tx = tx;
+ }
+ goto bail;
+ }
+
+ /* Allocate a buffer and copy the header and payload to it. */
+ tx->map_len = (plen + 1) << 2;
+ piobuf = kmalloc(tx->map_len, GFP_ATOMIC);
+ if (unlikely(piobuf == NULL)) {
+ ret = -EBUSY;
+ goto err_tx;
+ }
+ tx->txreq.map_addr = piobuf;
+ tx->txreq.flags |= IPATH_SDMA_TXREQ_F_FREEBUF;
+ tx->txreq.sg_count = 1;
+
+ *piobuf++ = (__force u32) cpu_to_le32(plen);
+ *piobuf++ = (__force u32) cpu_to_le32(control);
+ memcpy(piobuf, hdr, hdrwords << 2);
+ ipath_copy_from_sge(piobuf + hdrwords, ss, len);
+
+ ret = ipath_sdma_verbs_send(dd, NULL, 0, tx);
+ /*
+ * If we couldn't queue the DMA request, save the info
+ * and try again later rather than destroying the
+ * buffer and undoing the side effects of the copy.
+ */
+ if (ret) {
+ tx->ss = NULL;
+ tx->len = 0;
+ qp->s_tx = tx;
+ }
+ dev->n_unaligned++;
+ goto bail;
+
+err_tx:
+ if (atomic_dec_and_test(&qp->refcount))
+ wake_up(&qp->wait);
+ put_txreq(dev, tx);
+bail:
+ return ret;
+}
+
+static int ipath_verbs_send_pio(struct ipath_qp *qp,
+ struct ipath_ib_header *ibhdr, u32 hdrwords,