int reserved_qps;
int num_srqs;
int max_srq_wqes;
+ int max_srq_sge;
int reserved_srqs;
int num_eecs;
int reserved_eecs;
int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
enum ib_srq_attr_mask attr_mask);
int mthca_query_srq(struct ib_srq *srq, struct ib_srq_attr *srq_attr);
+int mthca_max_srq_sge(struct mthca_dev *dev);
void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
enum ib_event_type event_type);
void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr);
mdev->limits.reserved_srqs = dev_lim->reserved_srqs;
mdev->limits.reserved_eecs = dev_lim->reserved_eecs;
mdev->limits.max_desc_sz = dev_lim->max_desc_sz;
+ mdev->limits.max_srq_sge = mthca_max_srq_sge(mdev);
/*
* Subtract 1 from the limit because we need to allocate a
* spare CQE so the HCA HW can tell the difference between an
props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
props->max_srq = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
props->max_srq_wr = mdev->limits.max_srq_wqes;
- props->max_srq_sge = mdev->limits.max_sg;
+ props->max_srq_sge = mdev->limits.max_srq_sge;
props->local_ca_ack_delay = mdev->limits.local_ca_ack_delay;
props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
IB_ATOMIC_HCA : IB_ATOMIC_NONE;
/* Sanity check SRQ size before proceeding */
if (attr->max_wr > dev->limits.max_srq_wqes ||
- attr->max_sge > dev->limits.max_sg)
+ attr->max_sge > dev->limits.max_srq_sge)
return -EINVAL;
srq->max = attr->max_wr;
return err;
}
+int mthca_max_srq_sge(struct mthca_dev *dev)
+{
+ if (mthca_is_memfree(dev))
+ return dev->limits.max_sg;
+
+ /*
+ * SRQ allocations are based on powers of 2 for Tavor,
+ * (although they only need to be multiples of 16 bytes).
+ *
+ * Therefore, we need to base the max number of sg entries on
+ * the largest power of 2 descriptor size that is <= to the
+ * actual max WQE descriptor size, rather than return the
+ * max_sg value given by the firmware (which is based on WQE
+ * sizes as multiples of 16, not powers of 2).
+ *
+ * If SRQ implementation is changed for Tavor to be based on
+ * multiples of 16, the calculation below can be deleted and
+ * the FW max_sg value returned.
+ */
+ return min_t(int, dev->limits.max_sg,
+ ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
+ sizeof (struct mthca_next_seg)) /
+ sizeof (struct mthca_data_seg));
+}
+
int __devinit mthca_init_srq_table(struct mthca_dev *dev)
{
int err;