diff options
Diffstat (limited to 'drivers/infiniband/hw/mthca/mthca_qp.c')
-rw-r--r-- | drivers/infiniband/hw/mthca/mthca_qp.c | 201 |
1 files changed, 161 insertions, 40 deletions
diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c index 5fa00669f9b..7450550db73 100644 --- a/drivers/infiniband/hw/mthca/mthca_qp.c +++ b/drivers/infiniband/hw/mthca/mthca_qp.c @@ -36,6 +36,8 @@ */ #include <linux/init.h> +#include <linux/string.h> +#include <linux/slab.h> #include <rdma/ib_verbs.h> #include <rdma/ib_cache.h> @@ -338,8 +340,7 @@ static const struct { [UC] = (IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN | - IB_QP_RQ_PSN | - IB_QP_MAX_DEST_RD_ATOMIC), + IB_QP_RQ_PSN), [RC] = (IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN | @@ -368,8 +369,7 @@ static const struct { .trans = MTHCA_TRANS_RTR2RTS, .req_param = { [UD] = IB_QP_SQ_PSN, - [UC] = (IB_QP_SQ_PSN | - IB_QP_MAX_QP_RD_ATOMIC), + [UC] = IB_QP_SQ_PSN, [RC] = (IB_QP_TIMEOUT | IB_QP_RETRY_CNT | IB_QP_RNR_RETRY | @@ -446,8 +446,6 @@ static const struct { [UD] = (IB_QP_PKEY_INDEX | IB_QP_QKEY), [UC] = (IB_QP_AV | - IB_QP_MAX_QP_RD_ATOMIC | - IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_CUR_STATE | IB_QP_ALT_PATH | IB_QP_ACCESS_FLAGS | @@ -478,7 +476,7 @@ static const struct { .opt_param = { [UD] = (IB_QP_CUR_STATE | IB_QP_QKEY), - [UC] = (IB_QP_CUR_STATE), + [UC] = IB_QP_CUR_STATE, [RC] = (IB_QP_CUR_STATE | IB_QP_MIN_RNR_TIMER), [MLX] = (IB_QP_CUR_STATE | @@ -586,6 +584,13 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) return -EINVAL; } + if ((attr_mask & IB_QP_PKEY_INDEX) && + attr->pkey_index >= dev->limits.pkey_table_len) { + mthca_dbg(dev, "PKey index (%u) too large. max is %d\n", + attr->pkey_index,dev->limits.pkey_table_len-1); + return -EINVAL; + } + mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL); if (IS_ERR(mailbox)) return PTR_ERR(mailbox); @@ -725,15 +730,16 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) } if (attr_mask & IB_QP_ACCESS_FLAGS) { + qp_context->params2 |= + cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ? + MTHCA_QP_BIT_RWE : 0); + /* - * Only enable RDMA/atomics if we have responder - * resources set to a non-zero value. + * Only enable RDMA reads and atomics if we have + * responder resources set to a non-zero value. */ if (qp->resp_depth) { qp_context->params2 |= - cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ? - MTHCA_QP_BIT_RWE : 0); - qp_context->params2 |= cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ? MTHCA_QP_BIT_RRE : 0); qp_context->params2 |= @@ -754,31 +760,27 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) if (qp->resp_depth && !attr->max_dest_rd_atomic) { /* * Lowering our responder resources to zero. - * Turn off RDMA/atomics as responder. - * (RWE/RRE/RAE in params2 already zero) + * Turn off reads RDMA and atomics as responder. + * (RRE/RAE in params2 already zero) */ - qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE | - MTHCA_QP_OPTPAR_RRE | + qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE | MTHCA_QP_OPTPAR_RAE); } if (!qp->resp_depth && attr->max_dest_rd_atomic) { /* * Increasing our responder resources from - * zero. Turn on RDMA/atomics as appropriate. + * zero. Turn on RDMA reads and atomics as + * appropriate. */ qp_context->params2 |= - cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_WRITE ? - MTHCA_QP_BIT_RWE : 0); - qp_context->params2 |= cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ? MTHCA_QP_BIT_RRE : 0); qp_context->params2 |= cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_QP_BIT_RAE : 0); - qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE | - MTHCA_QP_OPTPAR_RRE | + qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE | MTHCA_QP_OPTPAR_RAE); } @@ -869,7 +871,10 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL); mthca_wq_init(&qp->sq); + qp->sq.last = get_send_wqe(qp, qp->sq.max - 1); + mthca_wq_init(&qp->rq); + qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1); if (mthca_is_memfree(dev)) { *qp->sq.db = 0; @@ -880,6 +885,50 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask) return err; } +static void mthca_adjust_qp_caps(struct mthca_dev *dev, + struct mthca_pd *pd, + struct mthca_qp *qp) +{ + int max_data_size; + + /* + * Calculate the maximum size of WQE s/g segments, excluding + * the next segment and other non-data segments. + */ + max_data_size = min(dev->limits.max_desc_sz, 1 << qp->sq.wqe_shift) - + sizeof (struct mthca_next_seg); + + switch (qp->transport) { + case MLX: + max_data_size -= 2 * sizeof (struct mthca_data_seg); + break; + + case UD: + if (mthca_is_memfree(dev)) + max_data_size -= sizeof (struct mthca_arbel_ud_seg); + else + max_data_size -= sizeof (struct mthca_tavor_ud_seg); + break; + + default: + max_data_size -= sizeof (struct mthca_raddr_seg); + break; + } + + /* We don't support inline data for kernel QPs (yet). */ + if (!pd->ibpd.uobject) + qp->max_inline_data = 0; + else + qp->max_inline_data = max_data_size - MTHCA_INLINE_HEADER_SIZE; + + qp->sq.max_gs = min_t(int, dev->limits.max_sg, + max_data_size / sizeof (struct mthca_data_seg)); + qp->rq.max_gs = min_t(int, dev->limits.max_sg, + (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) - + sizeof (struct mthca_next_seg)) / + sizeof (struct mthca_data_seg)); +} + /* * Allocate and register buffer for WQEs. qp->rq.max, sq.max, * rq.max_gs and sq.max_gs must all be assigned. @@ -897,27 +946,53 @@ static int mthca_alloc_wqe_buf(struct mthca_dev *dev, size = sizeof (struct mthca_next_seg) + qp->rq.max_gs * sizeof (struct mthca_data_seg); + if (size > dev->limits.max_desc_sz) + return -EINVAL; + for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size; qp->rq.wqe_shift++) ; /* nothing */ - size = sizeof (struct mthca_next_seg) + - qp->sq.max_gs * sizeof (struct mthca_data_seg); + size = qp->sq.max_gs * sizeof (struct mthca_data_seg); switch (qp->transport) { case MLX: size += 2 * sizeof (struct mthca_data_seg); break; + case UD: - if (mthca_is_memfree(dev)) - size += sizeof (struct mthca_arbel_ud_seg); - else - size += sizeof (struct mthca_tavor_ud_seg); + size += mthca_is_memfree(dev) ? + sizeof (struct mthca_arbel_ud_seg) : + sizeof (struct mthca_tavor_ud_seg); + break; + + case UC: + size += sizeof (struct mthca_raddr_seg); + break; + + case RC: + size += sizeof (struct mthca_raddr_seg); + /* + * An atomic op will require an atomic segment, a + * remote address segment and one scatter entry. + */ + size = max_t(int, size, + sizeof (struct mthca_atomic_seg) + + sizeof (struct mthca_raddr_seg) + + sizeof (struct mthca_data_seg)); break; + default: - /* bind seg is as big as atomic + raddr segs */ - size += sizeof (struct mthca_bind_seg); + break; } + /* Make sure that we have enough space for a bind request */ + size = max_t(int, size, sizeof (struct mthca_bind_seg)); + + size += sizeof (struct mthca_next_seg); + + if (size > dev->limits.max_desc_sz) + return -EINVAL; + for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size; qp->sq.wqe_shift++) ; /* nothing */ @@ -1061,6 +1136,8 @@ static int mthca_alloc_qp_common(struct mthca_dev *dev, return ret; } + mthca_adjust_qp_caps(dev, pd, qp); + /* * If this is a userspace QP, we're done now. The doorbells * will be allocated and buffers will be initialized in @@ -1112,8 +1189,10 @@ static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap, struct mthca_qp *qp) { /* Sanity check QP size before proceeding */ - if (cap->max_send_wr > 65536 || cap->max_recv_wr > 65536 || - cap->max_send_sge > 64 || cap->max_recv_sge > 64) + if (cap->max_send_wr > dev->limits.max_wqes || + cap->max_recv_wr > dev->limits.max_wqes || + cap->max_send_sge > dev->limits.max_sg || + cap->max_recv_sge > dev->limits.max_sg) return -EINVAL; if (mthca_is_memfree(dev)) { @@ -1479,8 +1558,8 @@ int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, } wqe += sizeof (struct mthca_atomic_seg); - size += sizeof (struct mthca_raddr_seg) / 16 + - sizeof (struct mthca_atomic_seg); + size += (sizeof (struct mthca_raddr_seg) + + sizeof (struct mthca_atomic_seg)) / 16; break; case IB_WR_RDMA_WRITE: @@ -1630,6 +1709,7 @@ int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, { struct mthca_dev *dev = to_mdev(ibqp->device); struct mthca_qp *qp = to_mqp(ibqp); + __be32 doorbell[2]; unsigned long flags; int err = 0; int nreq; @@ -1647,6 +1727,22 @@ int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, ind = qp->rq.next_ind; for (nreq = 0; wr; ++nreq, wr = wr->next) { + if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) { + nreq = 0; + + doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0); + doorbell[1] = cpu_to_be32(qp->qpn << 8); + + wmb(); + + mthca_write64(doorbell, + dev->kar + MTHCA_RECEIVE_DOORBELL, + MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); + + qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB; + size0 = 0; + } + if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { mthca_err(dev, "RQ %06x full (%u head, %u tail," " %d max, %d nreq)\n", qp->qpn, @@ -1704,8 +1800,6 @@ int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, out: if (likely(nreq)) { - __be32 doorbell[2]; - doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0); doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq); @@ -1728,6 +1822,7 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, { struct mthca_dev *dev = to_mdev(ibqp->device); struct mthca_qp *qp = to_mqp(ibqp); + __be32 doorbell[2]; void *wqe; void *prev_wqe; unsigned long flags; @@ -1747,6 +1842,34 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, ind = qp->sq.head & (qp->sq.max - 1); for (nreq = 0; wr; ++nreq, wr = wr->next) { + if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) { + nreq = 0; + + doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) | + ((qp->sq.head & 0xffff) << 8) | + f0 | op0); + doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0); + + qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB; + size0 = 0; + + /* + * Make sure that descriptors are written before + * doorbell record. + */ + wmb(); + *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff); + + /* + * Make sure doorbell record is written before we + * write MMIO send doorbell. + */ + wmb(); + mthca_write64(doorbell, + dev->kar + MTHCA_SEND_DOORBELL, + MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); + } + if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { mthca_err(dev, "SQ %06x full (%u head, %u tail," " %d max, %d nreq)\n", qp->qpn, @@ -1799,8 +1922,8 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, } wqe += sizeof (struct mthca_atomic_seg); - size += sizeof (struct mthca_raddr_seg) / 16 + - sizeof (struct mthca_atomic_seg); + size += (sizeof (struct mthca_raddr_seg) + + sizeof (struct mthca_atomic_seg)) / 16; break; case IB_WR_RDMA_READ: @@ -1923,8 +2046,6 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, out: if (likely(nreq)) { - __be32 doorbell[2]; - doorbell[0] = cpu_to_be32((nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0); |