void mthca_free_qp(struct mthca_dev *dev,
struct mthca_qp *qp)
{
u8 status;
int size;
int i;
spin_lock_irq(&dev->qp_table.lock);
mthca_array_clear(&dev->qp_table.qp,
qp->qpn & (dev->limits.num_qps - 1));
spin_unlock_irq(&dev->qp_table.lock);
atomic_dec(&qp->refcount);
wait_event(qp->wait, !atomic_read(&qp->refcount));
if (qp->state != IB_QPS_RESET)
mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
mthca_free_mr(dev, &qp->mr);
size = PAGE_ALIGN(qp->send_wqe_offset +
(qp->sq.max << qp->sq.wqe_shift));
if (qp->is_direct) {
pci_free_consistent(dev->pdev, size,
qp->queue.direct.buf,
pci_unmap_addr(&qp->queue.direct, mapping));
} else {
for (i = 0; i < size / PAGE_SIZE; ++i) {
pci_free_consistent(dev->pdev, PAGE_SIZE,
qp->queue.page_list[i].buf,
pci_unmap_addr(&qp->queue.page_list[i],
mapping));
}
}
kfree(qp->wrid);
if (is_sqp(dev, qp)) {
atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
dma_free_coherent(&dev->pdev->dev,
to_msqp(qp)->header_buf_size,
to_msqp(qp)->header_buf,
to_msqp(qp)->header_dma);
}
else
mthca_free(&dev->qp_table.alloc, qp->qpn);
}
int mlx4_destroy_xrc_srq(struct ibv_srq *srq)
{
struct mlx4_context *mctx = to_mctx(srq->context);
struct mlx4_srq *msrq = to_msrq(srq);
struct mlx4_cq *mcq;
int ret;
mcq = to_mcq(msrq->verbs_srq.cq);
mlx4_cq_clean(mcq, 0, msrq);
pthread_spin_lock(&mcq->lock);
mlx4_clear_xsrq(&mctx->xsrq_table, msrq->verbs_srq.srq_num);
pthread_spin_unlock(&mcq->lock);
ret = ibv_cmd_destroy_srq(srq);
if (ret) {
pthread_spin_lock(&mcq->lock);
mlx4_store_xsrq(&mctx->xsrq_table, msrq->verbs_srq.srq_num, msrq);
pthread_spin_unlock(&mcq->lock);
return ret;
}
mlx4_free_db(mctx, MLX4_DB_TYPE_RQ, msrq->db);
mlx4_free_buf(&msrq->buf);
free(msrq->wrid);
free(msrq);
return 0;
}
int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
{
struct mlx4_ib_cq *cq = to_mcq(ibcq);
struct mlx4_ib_qp *cur_qp = NULL;
unsigned long flags;
int npolled;
int err = 0;
spin_lock_irqsave(&cq->lock, flags);
for (npolled = 0; npolled < num_entries; ++npolled) {
err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
if (err)
break;
}
if (npolled)
mlx4_cq_set_ci(&cq->mcq);
spin_unlock_irqrestore(&cq->lock, flags);
if (err == 0 || err == -EAGAIN)
return npolled;
else
return err;
}
int mlx4_arm_cq(struct ibv_cq *ibvcq, int solicited)
{
struct mlx4_cq *cq = to_mcq(ibvcq);
uint32_t doorbell[2];
uint32_t sn;
uint32_t ci;
uint32_t cmd;
sn = cq->arm_sn & 3;
ci = cq->cons_index & 0xffffff;
cmd = solicited ? MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT;
*cq->arm_db = htonl(sn << 28 | cmd | ci);
/*
* Make sure that the doorbell record in host memory is
* written before ringing the doorbell via PCI MMIO.
*/
wmb();
doorbell[0] = htonl(sn << 28 | cmd | cq->cqn);
doorbell[1] = htonl(ci);
mlx4_write64(doorbell, to_mctx(ibvcq->context), MLX4_CQ_DOORBELL);
return 0;
}
int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
{
struct mlx4_ib_cq *mcq = to_mcq(cq);
struct mlx4_ib_dev *dev = to_mdev(cq->device);
return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
}
int mlx4_ib_destroy_srq(struct ib_srq *srq)
{
struct mlx4_ib_dev *dev = to_mdev(srq->device);
struct mlx4_ib_srq *msrq = to_msrq(srq);
struct mlx4_ib_cq *cq;
mlx4_srq_invalidate(dev->dev, &msrq->msrq);
if (srq->xrc_cq && !srq->uobject) {
cq = to_mcq(srq->xrc_cq);
spin_lock_irq(&cq->lock);
__mlx4_ib_cq_clean(cq, -1, msrq);
mlx4_srq_remove(dev->dev, &msrq->msrq);
spin_unlock_irq(&cq->lock);
} else
mlx4_srq_remove(dev->dev, &msrq->msrq);
mlx4_srq_free(dev->dev, &msrq->msrq);
mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
if (srq->uobject) {
mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
ib_umem_release(msrq->umem);
} else {
kfree(msrq->wrid);
mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
&msrq->buf);
mlx4_db_free(dev->dev, &msrq->db);
}
kfree(msrq);
return 0;
}
int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
{
mlx4_cq_arm(&to_mcq(ibcq)->mcq,
(flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
to_mdev(ibcq->device)->uar_map,
MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
return 0;
}
int mlx4_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
{
struct mlx4_ib_dev *dev = to_mdev(cq->device);
struct mlx4_ib_cq *mcq = to_mcq(cq);
mlx4_cq_free(dev->dev, &mcq->mcq);
mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
if (udata) {
mlx4_ib_db_unmap_user(
rdma_udata_to_drv_context(
udata,
struct mlx4_ib_ucontext,
ibucontext),
&mcq->db);
ib_umem_release(mcq->umem);
} else {
int mlx4_ib_destroy_cq(struct ib_cq *cq)
{
struct mlx4_ib_dev *dev = to_mdev(cq->device);
struct mlx4_ib_cq *mcq = to_mcq(cq);
mlx4_cq_free(dev->dev, &mcq->mcq);
mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
if (cq->uobject) {
mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
ib_umem_release(mcq->umem);
} else {
mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
mlx4_db_free(dev->dev, &mcq->db);
}
kfree(mcq);
return 0;
}
int mlx4_poll_cq(struct ibv_cq *ibcq, int ne, struct ibv_wc *wc)
{
struct mlx4_cq *cq = to_mcq(ibcq);
struct mlx4_qp *qp = NULL;
int npolled;
int err = CQ_OK;
pthread_spin_lock(&cq->lock);
for (npolled = 0; npolled < ne; ++npolled) {
err = mlx4_poll_one(cq, &qp, wc + npolled);
if (err != CQ_OK)
break;
}
if (npolled)
update_cons_index(cq);
pthread_spin_unlock(&cq->lock);
return err == CQ_POLL_ERR ? err : npolled;
}
int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
{
struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
struct mlx4_ib_cq *cq = to_mcq(ibcq);
struct mlx4_mtt mtt;
int outst_cqe;
int err;
mutex_lock(&cq->resize_mutex);
if (entries < 1 || entries > dev->dev->caps.max_cqes) {
err = -EINVAL;
goto out;
}
entries = roundup_pow_of_two(entries + 1);
if (entries == ibcq->cqe + 1) {
err = 0;
goto out;
}
if (ibcq->uobject) {
err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
if (err)
goto out;
} else {
/* Can't be smaller than the number of outstanding CQEs */
outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
if (entries < outst_cqe + 1) {
err = 0;
goto out;
}
err = mlx4_alloc_resize_buf(dev, cq, entries);
if (err)
goto out;
}
mtt = cq->buf.mtt;
err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
if (err)
goto err_buf;
mlx4_mtt_cleanup(dev->dev, &mtt);
if (ibcq->uobject) {
cq->buf = cq->resize_buf->buf;
cq->ibcq.cqe = cq->resize_buf->cqe;
ib_umem_release(cq->umem);
cq->umem = cq->resize_umem;
kfree(cq->resize_buf);
cq->resize_buf = NULL;
cq->resize_umem = NULL;
} else {
struct mlx4_ib_cq_buf tmp_buf;
int tmp_cqe = 0;
spin_lock_irq(&cq->lock);
if (cq->resize_buf) {
mlx4_ib_cq_resize_copy_cqes(cq);
tmp_buf = cq->buf;
tmp_cqe = cq->ibcq.cqe;
cq->buf = cq->resize_buf->buf;
cq->ibcq.cqe = cq->resize_buf->cqe;
kfree(cq->resize_buf);
cq->resize_buf = NULL;
}
spin_unlock_irq(&cq->lock);
if (tmp_cqe)
mlx4_ib_free_cq_buf(dev, &tmp_buf, tmp_cqe);
}
goto out;
err_buf:
mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
if (!ibcq->uobject)
mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
cq->resize_buf->cqe);
kfree(cq->resize_buf);
cq->resize_buf = NULL;
if (cq->resize_umem) {
ib_umem_release(cq->resize_umem);
cq->resize_umem = NULL;
}
out:
mutex_unlock(&cq->resize_mutex);
return err;
}
int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
{
struct mthca_dev *dev = to_mdev(ibqp->device);
struct mthca_qp *qp = to_mqp(ibqp);
enum ib_qp_state cur_state, new_state;
void *mailbox = NULL;
struct mthca_qp_param *qp_param;
struct mthca_qp_context *qp_context;
u32 req_param, opt_param;
u8 status;
int err;
if (attr_mask & IB_QP_CUR_STATE) {
if (attr->cur_qp_state != IB_QPS_RTR &&
attr->cur_qp_state != IB_QPS_RTS &&
attr->cur_qp_state != IB_QPS_SQD &&
attr->cur_qp_state != IB_QPS_SQE)
return -EINVAL;
else
cur_state = attr->cur_qp_state;
} else {
spin_lock_irq(&qp->lock);
cur_state = qp->state;
spin_unlock_irq(&qp->lock);
}
if (attr_mask & IB_QP_STATE) {
if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
return -EINVAL;
new_state = attr->qp_state;
} else
new_state = cur_state;
if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
mthca_dbg(dev, "Illegal QP transition "
"%d->%d\n", cur_state, new_state);
return -EINVAL;
}
req_param = state_table[cur_state][new_state].req_param[qp->transport];
opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
if ((req_param & attr_mask) != req_param) {
mthca_dbg(dev, "QP transition "
"%d->%d missing req attr 0x%08x\n",
cur_state, new_state,
req_param & ~attr_mask);
return -EINVAL;
}
if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
mthca_dbg(dev, "QP transition (transport %d) "
"%d->%d has extra attr 0x%08x\n",
qp->transport,
cur_state, new_state,
attr_mask & ~(req_param | opt_param |
IB_QP_STATE));
return -EINVAL;
}
mailbox = kmalloc(sizeof (*qp_param) + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
if (!mailbox)
return -ENOMEM;
qp_param = MAILBOX_ALIGN(mailbox);
qp_context = &qp_param->context;
memset(qp_param, 0, sizeof *qp_param);
qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
(to_mthca_st(qp->transport) << 16));
qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
if (!(attr_mask & IB_QP_PATH_MIG_STATE))
qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
else {
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
switch (attr->path_mig_state) {
case IB_MIG_MIGRATED:
qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
break;
case IB_MIG_REARM:
qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
break;
case IB_MIG_ARMED:
qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
break;
}
}
/* leave sched_queue as 0 */
if (qp->transport == MLX || qp->transport == UD)
qp_context->mtu_msgmax = cpu_to_be32((IB_MTU_2048 << 29) |
(11 << 24));
else if (attr_mask & IB_QP_PATH_MTU) {
qp_context->mtu_msgmax = cpu_to_be32((attr->path_mtu << 29) |
(31 << 24));
}
qp_context->usr_page = cpu_to_be32(MTHCA_KAR_PAGE);
qp_context->local_qpn = cpu_to_be32(qp->qpn);
if (attr_mask & IB_QP_DEST_QPN) {
qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
}
if (qp->transport == MLX)
qp_context->pri_path.port_pkey |=
cpu_to_be32(to_msqp(qp)->port << 24);
else {
if (attr_mask & IB_QP_PORT) {
qp_context->pri_path.port_pkey |=
cpu_to_be32(attr->port_num << 24);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
}
}
if (attr_mask & IB_QP_PKEY_INDEX) {
qp_context->pri_path.port_pkey |=
cpu_to_be32(attr->pkey_index);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
}
if (attr_mask & IB_QP_RNR_RETRY) {
qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
}
if (attr_mask & IB_QP_AV) {
qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
qp_context->pri_path.static_rate = (!!attr->ah_attr.static_rate) << 3;
if (attr->ah_attr.ah_flags & IB_AH_GRH) {
qp_context->pri_path.g_mylmc |= 1 << 7;
qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
qp_context->pri_path.sl_tclass_flowlabel =
cpu_to_be32((attr->ah_attr.sl << 28) |
(attr->ah_attr.grh.traffic_class << 20) |
(attr->ah_attr.grh.flow_label));
memcpy(qp_context->pri_path.rgid,
attr->ah_attr.grh.dgid.raw, 16);
} else {
qp_context->pri_path.sl_tclass_flowlabel =
cpu_to_be32(attr->ah_attr.sl << 28);
}
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
}
if (attr_mask & IB_QP_TIMEOUT) {
qp_context->pri_path.ackto = attr->timeout;
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
}
/* XXX alt_path */
/* leave rdd as 0 */
qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
/* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
(MTHCA_FLIGHT_LIMIT << 24) |
MTHCA_QP_BIT_SRE |
MTHCA_QP_BIT_SWE |
MTHCA_QP_BIT_SAE);
if (qp->sq.policy == IB_SIGNAL_ALL_WR)
qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
if (attr_mask & IB_QP_RETRY_CNT) {
qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
}
if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
qp_context->params1 |= cpu_to_be32(min(attr->max_dest_rd_atomic ?
ffs(attr->max_dest_rd_atomic) - 1 : 0,
7) << 21);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
}
if (attr_mask & IB_QP_SQ_PSN)
qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
if (attr_mask & IB_QP_ACCESS_FLAGS) {
/*
* Only enable RDMA/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 |=
cpu_to_be32(attr->qp_access_flags & 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 |
MTHCA_QP_OPTPAR_RAE);
qp->atomic_rd_en = attr->qp_access_flags;
}
if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
u8 rra_max;
if (qp->resp_depth && !attr->max_rd_atomic) {
/*
* Lowering our responder resources to zero.
* Turn off RDMA/atomics as responder.
* (RWE/RRE/RAE in params2 already zero)
*/
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
MTHCA_QP_OPTPAR_RRE |
MTHCA_QP_OPTPAR_RAE);
}
if (!qp->resp_depth && attr->max_rd_atomic) {
/*
* Increasing our responder resources from
* zero. Turn on RDMA/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 |
MTHCA_QP_OPTPAR_RAE);
}
for (rra_max = 0;
1 << rra_max < attr->max_rd_atomic &&
rra_max < dev->qp_table.rdb_shift;
++rra_max)
; /* nothing */
qp_context->params2 |= cpu_to_be32(rra_max << 21);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
qp->resp_depth = attr->max_rd_atomic;
}
if (qp->rq.policy == IB_SIGNAL_ALL_WR)
qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
if (attr_mask & IB_QP_MIN_RNR_TIMER) {
qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
}
if (attr_mask & IB_QP_RQ_PSN)
qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
qp_context->ra_buff_indx = dev->qp_table.rdb_base +
((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
dev->qp_table.rdb_shift);
qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
if (attr_mask & IB_QP_QKEY) {
qp_context->qkey = cpu_to_be32(attr->qkey);
qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
}
err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
qp->qpn, 0, qp_param, 0, &status);
if (status) {
mthca_warn(dev, "modify QP %d returned status %02x.\n",
state_table[cur_state][new_state].trans, status);
err = -EINVAL;
}
if (!err)
qp->state = new_state;
kfree(mailbox);
if (is_sqp(dev, qp))
store_attrs(to_msqp(qp), attr, attr_mask);
/*
* If we are moving QP0 to RTR, bring the IB link up; if we
* are moving QP0 to RESET or ERROR, bring the link back down.
*/
if (is_qp0(dev, qp)) {
if (cur_state != IB_QPS_RTR &&
new_state == IB_QPS_RTR)
init_port(dev, to_msqp(qp)->port);
if (cur_state != IB_QPS_RESET &&
cur_state != IB_QPS_ERR &&
(new_state == IB_QPS_RESET ||
new_state == IB_QPS_ERR))
mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
}
return err;
}
struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
struct ib_srq_init_attr *init_attr,
struct ib_udata *udata)
{
struct mlx5_ib_dev *dev = to_mdev(pd->device);
struct mlx5_ib_srq *srq;
int desc_size;
int buf_size;
int err;
struct mlx5_create_srq_mbox_in *uninitialized_var(in);
int uninitialized_var(inlen);
int is_xrc;
u32 flgs, xrcdn;
/* Sanity check SRQ size before proceeding */
if (init_attr->attr.max_wr >= dev->mdev->caps.max_srq_wqes) {
mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
init_attr->attr.max_wr,
dev->mdev->caps.max_srq_wqes);
return ERR_PTR(-EINVAL);
}
srq = kmalloc(sizeof(*srq), GFP_KERNEL);
if (!srq)
return ERR_PTR(-ENOMEM);
mutex_init(&srq->mutex);
spin_lock_init(&srq->lock);
srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
srq->msrq.max_gs = init_attr->attr.max_sge;
desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
desc_size = roundup_pow_of_two(desc_size);
desc_size = max_t(int, 32, desc_size);
srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
sizeof(struct mlx5_wqe_data_seg);
srq->msrq.wqe_shift = ilog2(desc_size);
buf_size = srq->msrq.max * desc_size;
mlx5_ib_dbg(dev, "desc_size 0x%x, req wr 0x%x, srq size 0x%x, max_gs 0x%x, max_avail_gather 0x%x\n",
desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs,
srq->msrq.max_avail_gather);
if (pd->uobject)
err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen);
else
err = create_srq_kernel(dev, srq, &in, buf_size, &inlen);
if (err) {
mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
pd->uobject ? "user" : "kernel", err);
goto err_srq;
}
is_xrc = (init_attr->srq_type == IB_SRQT_XRC);
in->ctx.state_log_sz = ilog2(srq->msrq.max);
flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24;
xrcdn = 0;
if (is_xrc) {
xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn);
} else if (init_attr->srq_type == IB_SRQT_BASIC) {
xrcdn = to_mxrcd(dev->devr.x0)->xrcdn;
in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn);
}
in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF));
in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn);
in->ctx.db_record = cpu_to_be64(srq->db.dma);
err = mlx5_core_create_srq(dev->mdev, &srq->msrq, in, inlen);
mlx5_vfree(in);
if (err) {
mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
goto err_usr_kern_srq;
}
mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
srq->msrq.event = mlx5_ib_srq_event;
srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
if (pd->uobject)
if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
mlx5_ib_dbg(dev, "copy to user failed\n");
err = -EFAULT;
goto err_core;
}
init_attr->attr.max_wr = srq->msrq.max - 1;
return &srq->ibsrq;
err_core:
mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
err_usr_kern_srq:
if (pd->uobject)
destroy_srq_user(pd, srq);
else
destroy_srq_kernel(dev, srq);
err_srq:
kfree(srq);
return ERR_PTR(err);
}
struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
struct ib_srq_init_attr *init_attr,
struct ib_udata *udata)
{
struct mlx4_ib_dev *dev = to_mdev(pd->device);
struct mlx4_ib_srq *srq;
struct mlx4_wqe_srq_next_seg *next;
struct mlx4_wqe_data_seg *scatter;
u32 cqn;
u16 xrcdn;
int desc_size;
int buf_size;
int err;
int i;
/* Sanity check SRQ size before proceeding */
if (init_attr->attr.max_wr >= dev->dev->caps.max_srq_wqes ||
init_attr->attr.max_sge > dev->dev->caps.max_srq_sge)
return ERR_PTR(-EINVAL);
srq = kmalloc(sizeof *srq, GFP_KERNEL);
if (!srq)
return ERR_PTR(-ENOMEM);
mutex_init(&srq->mutex);
spin_lock_init(&srq->lock);
srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
srq->msrq.max_gs = init_attr->attr.max_sge;
desc_size = max(32UL,
roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
srq->msrq.max_gs *
sizeof (struct mlx4_wqe_data_seg)));
srq->msrq.wqe_shift = ilog2(desc_size);
buf_size = srq->msrq.max * desc_size;
if (pd->uobject) {
struct mlx4_ib_create_srq ucmd;
if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
err = -EFAULT;
goto err_srq;
}
srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
buf_size, 0, 0);
if (IS_ERR(srq->umem)) {
err = PTR_ERR(srq->umem);
goto err_srq;
}
err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem),
ilog2(srq->umem->page_size), &srq->mtt);
if (err)
goto err_buf;
err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
if (err)
goto err_mtt;
err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
ucmd.db_addr, &srq->db);
if (err)
goto err_mtt;
} else {
err = mlx4_db_alloc(dev->dev, &srq->db, 0, GFP_KERNEL);
if (err)
goto err_srq;
*srq->db.db = 0;
if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf,
GFP_KERNEL)) {
err = -ENOMEM;
goto err_db;
}
srq->head = 0;
srq->tail = srq->msrq.max - 1;
srq->wqe_ctr = 0;
for (i = 0; i < srq->msrq.max; ++i) {
next = get_wqe(srq, i);
next->next_wqe_index =
cpu_to_be16((i + 1) & (srq->msrq.max - 1));
for (scatter = (void *) (next + 1);
(void *) scatter < (void *) next + desc_size;
++scatter)
scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
}
err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
&srq->mtt);
if (err)
goto err_buf;
err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf, GFP_KERNEL);
if (err)
goto err_mtt;
srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL);
if (!srq->wrid) {
err = -ENOMEM;
goto err_mtt;
}
}
cqn = (init_attr->srq_type == IB_SRQT_XRC) ?
to_mcq(init_attr->ext.xrc.cq)->mcq.cqn : 0;
xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
(u16) dev->dev->caps.reserved_xrcds;
err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, cqn, xrcdn, &srq->mtt,
srq->db.dma, &srq->msrq);
if (err)
goto err_wrid;
srq->msrq.event = mlx4_ib_srq_event;
srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
if (pd->uobject)
if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
err = -EFAULT;
goto err_wrid;
}
init_attr->attr.max_wr = srq->msrq.max - 1;
return &srq->ibsrq;
err_wrid:
if (pd->uobject)
mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
else
kfree(srq->wrid);
err_mtt:
mlx4_mtt_cleanup(dev->dev, &srq->mtt);
err_buf:
if (pd->uobject)
ib_umem_release(srq->umem);
else
mlx4_buf_free(dev->dev, buf_size, &srq->buf);
err_db:
if (!pd->uobject)
mlx4_db_free(dev->dev, &srq->db);
err_srq:
kfree(srq);
return ERR_PTR(err);
}
void mlx4_cq_event(struct ibv_cq *cq)
{
to_mcq(cq)->arm_sn++;
}
struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
struct ib_srq_init_attr *init_attr,
struct ib_udata *udata)
{
struct mlx5_ib_dev *dev = to_mdev(pd->device);
struct mlx5_ib_srq *srq;
int desc_size;
int buf_size;
int err;
struct mlx5_srq_attr in = {0};
__u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
/* Sanity check SRQ size before proceeding */
if (init_attr->attr.max_wr >= max_srq_wqes) {
mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
init_attr->attr.max_wr,
max_srq_wqes);
return ERR_PTR(-EINVAL);
}
srq = kmalloc(sizeof(*srq), GFP_KERNEL);
if (!srq)
return ERR_PTR(-ENOMEM);
mutex_init(&srq->mutex);
spin_lock_init(&srq->lock);
srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
srq->msrq.max_gs = init_attr->attr.max_sge;
desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
desc_size = roundup_pow_of_two(desc_size);
desc_size = max_t(int, 32, desc_size);
srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
sizeof(struct mlx5_wqe_data_seg);
srq->msrq.wqe_shift = ilog2(desc_size);
buf_size = srq->msrq.max * desc_size;
mlx5_ib_dbg(dev, "desc_size 0x%x, req wr 0x%x, srq size 0x%x, max_gs 0x%x, max_avail_gather 0x%x\n",
desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs,
srq->msrq.max_avail_gather);
if (pd->uobject)
err = create_srq_user(pd, srq, &in, udata, buf_size);
else
err = create_srq_kernel(dev, srq, &in, buf_size);
if (err) {
mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
pd->uobject ? "user" : "kernel", err);
goto err_srq;
}
in.type = init_attr->srq_type;
in.log_size = ilog2(srq->msrq.max);
in.wqe_shift = srq->msrq.wqe_shift - 4;
if (srq->wq_sig)
in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
if (init_attr->srq_type == IB_SRQT_XRC) {
in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
in.cqn = to_mcq(init_attr->ext.xrc.cq)->mcq.cqn;
} else if (init_attr->srq_type == IB_SRQT_BASIC) {
in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn;
in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
}
in.pd = to_mpd(pd)->pdn;
in.db_record = srq->db.dma;
err = mlx5_core_create_srq(dev->mdev, &srq->msrq, &in);
kvfree(in.pas);
if (err) {
mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
goto err_usr_kern_srq;
}
mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
srq->msrq.event = mlx5_ib_srq_event;
srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
if (pd->uobject)
if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
mlx5_ib_dbg(dev, "copy to user failed\n");
err = -EFAULT;
goto err_core;
}
init_attr->attr.max_wr = srq->msrq.max - 1;
return &srq->ibsrq;
err_core:
mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
err_usr_kern_srq:
if (pd->uobject)
destroy_srq_user(pd, srq);
else
destroy_srq_kernel(dev, srq);
err_srq:
kfree(srq);
return ERR_PTR(err);
}
int mlx4_post_recv(struct ibv_qp *ibqp, struct ibv_recv_wr *wr,
struct ibv_recv_wr **bad_wr)
{
struct mlx4_qp *qp = to_mqp(ibqp);
struct mlx4_wqe_data_seg *scat;
int ret = 0;
int nreq;
int ind;
int i;
pthread_spin_lock(&qp->rq.lock);
/* XXX check that state is OK to post receive */
ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (wq_overflow(&qp->rq, nreq, to_mcq(qp->ibv_qp.recv_cq))) {
ret = -1;
*bad_wr = wr;
goto out;
}
if (wr->num_sge > qp->rq.max_gs) {
ret = -1;
*bad_wr = wr;
goto out;
}
scat = get_recv_wqe(qp, ind);
for (i = 0; i < wr->num_sge; ++i)
__set_data_seg(scat + i, wr->sg_list + i);
if (i < qp->rq.max_gs) {
scat[i].byte_count = 0;
scat[i].lkey = htonl(MLX4_INVALID_LKEY);
scat[i].addr = 0;
}
qp->rq.wrid[ind] = wr->wr_id;
ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
}
out:
if (nreq) {
qp->rq.head += nreq;
/*
* Make sure that descriptors are written before
* doorbell record.
*/
wmb();
*qp->db = htonl(qp->rq.head & 0xffff);
}
pthread_spin_unlock(&qp->rq.lock);
return ret;
}
int mlx4_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
struct ibv_send_wr **bad_wr)
{
struct mlx4_context *ctx;
struct mlx4_qp *qp = to_mqp(ibqp);
void *wqe;
struct mlx4_wqe_ctrl_seg *ctrl;
int ind;
int nreq;
int inl = 0;
int ret = 0;
int size;
int i;
pthread_spin_lock(&qp->sq.lock);
/* XXX check that state is OK to post send */
ind = qp->sq.head;
for (nreq = 0; wr; ++nreq, wr = wr->next) {
if (wq_overflow(&qp->sq, nreq, to_mcq(qp->ibv_qp.send_cq))) {
ret = -1;
*bad_wr = wr;
goto out;
}
if (wr->num_sge > qp->sq.max_gs) {
ret = -1;
*bad_wr = wr;
goto out;
}
if (wr->opcode >= sizeof mlx4_ib_opcode / sizeof mlx4_ib_opcode[0]) {
ret = -1;
*bad_wr = wr;
goto out;
}
ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
qp->sq.wrid[ind & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
ctrl->xrcrb_flags =
(wr->send_flags & IBV_SEND_SIGNALED ?
htonl(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
(wr->send_flags & IBV_SEND_SOLICITED ?
htonl(MLX4_WQE_CTRL_SOLICIT) : 0) |
qp->sq_signal_bits;
if (wr->opcode == IBV_WR_SEND_WITH_IMM ||
wr->opcode == IBV_WR_RDMA_WRITE_WITH_IMM)
ctrl->imm = wr->imm_data;
else
ctrl->imm = 0;
wqe += sizeof *ctrl;
size = sizeof *ctrl / 16;
switch (ibqp->qp_type) {
case IBV_QPT_XRC:
ctrl->xrcrb_flags |= htonl(wr->xrc_remote_srq_num << 8);
/* fall thru */
case IBV_QPT_RC:
case IBV_QPT_UC:
switch (wr->opcode) {
case IBV_WR_ATOMIC_CMP_AND_SWP:
case IBV_WR_ATOMIC_FETCH_AND_ADD:
set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
wr->wr.atomic.rkey);
wqe += sizeof (struct mlx4_wqe_raddr_seg);
set_atomic_seg(wqe, wr);
wqe += sizeof (struct mlx4_wqe_atomic_seg);
size += (sizeof (struct mlx4_wqe_raddr_seg) +
sizeof (struct mlx4_wqe_atomic_seg)) / 16;
break;
case IBV_WR_RDMA_READ:
inl = 1;
/* fall through */
case IBV_WR_RDMA_WRITE:
case IBV_WR_RDMA_WRITE_WITH_IMM:
set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
wr->wr.rdma.rkey);
wqe += sizeof (struct mlx4_wqe_raddr_seg);
size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
break;
default:
/* No extra segments required for sends */
break;
}
break;
case IBV_QPT_UD:
set_datagram_seg(wqe, wr);
wqe += sizeof (struct mlx4_wqe_datagram_seg);
size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
if (to_mah(wr->wr.ud.ah)->tagged) {
ctrl->ins_vlan = 1 << 6;
ctrl->vlan_tag = htons(to_mah(wr->wr.ud.ah)->vlan);
}
break;
default:
break;
}
if (wr->send_flags & IBV_SEND_INLINE && wr->num_sge) {
struct mlx4_wqe_inline_seg *seg;
void *addr;
int len, seg_len;
int num_seg;
int off, to_copy;
inl = 0;
seg = wqe;
wqe += sizeof *seg;
off = ((uintptr_t) wqe) & (MLX4_INLINE_ALIGN - 1);
num_seg = 0;
seg_len = 0;
for (i = 0; i < wr->num_sge; ++i) {
addr = (void *) (uintptr_t) wr->sg_list[i].addr;
len = wr->sg_list[i].length;
inl += len;
if (inl > qp->max_inline_data) {
inl = 0;
ret = -1;
*bad_wr = wr;
goto out;
}
while (len >= MLX4_INLINE_ALIGN - off) {
to_copy = MLX4_INLINE_ALIGN - off;
memcpy(wqe, addr, to_copy);
len -= to_copy;
wqe += to_copy;
addr += to_copy;
seg_len += to_copy;
wmb(); /* see comment below */
seg->byte_count = htonl(MLX4_INLINE_SEG | seg_len);
seg_len = 0;
seg = wqe;
wqe += sizeof *seg;
off = sizeof *seg;
++num_seg;
}
memcpy(wqe, addr, len);
wqe += len;
seg_len += len;
off += len;
}
if (seg_len) {
++num_seg;
/*
* Need a barrier here to make sure
* all the data is visible before the
* byte_count field is set. Otherwise
* the HCA prefetcher could grab the
* 64-byte chunk with this inline
* segment and get a valid (!=
* 0xffffffff) byte count but stale
* data, and end up sending the wrong
* data.
*/
wmb();
seg->byte_count = htonl(MLX4_INLINE_SEG | seg_len);
}
size += (inl + num_seg * sizeof * seg + 15) / 16;
} else {
struct mlx4_wqe_data_seg *seg = wqe;
for (i = wr->num_sge - 1; i >= 0 ; --i)
set_data_seg(seg + i, wr->sg_list + i);
size += wr->num_sge * (sizeof *seg / 16);
}
ctrl->fence_size = (wr->send_flags & IBV_SEND_FENCE ?
MLX4_WQE_CTRL_FENCE : 0) | size;
/*
* Make sure descriptor is fully written before
* setting ownership bit (because HW can start
* executing as soon as we do).
*/
wmb();
ctrl->owner_opcode = htonl(mlx4_ib_opcode[wr->opcode]) |
(ind & qp->sq.wqe_cnt ? htonl(1 << 31) : 0);
/*
* We can improve latency by not stamping the last
* send queue WQE until after ringing the doorbell, so
* only stamp here if there are still more WQEs to post.
*/
if (wr->next)
stamp_send_wqe(qp, (ind + qp->sq_spare_wqes) &
(qp->sq.wqe_cnt - 1));
++ind;
}
out:
ctx = to_mctx(ibqp->context);
if (nreq == 1 && inl && size > 1 && size < ctx->bf_buf_size / 16) {
ctrl->owner_opcode |= htonl((qp->sq.head & 0xffff) << 8);
*(uint32_t *) (&ctrl->vlan_tag) |= qp->doorbell_qpn;
/*
* Make sure that descriptor is written to memory
* before writing to BlueFlame page.
*/
wmb();
++qp->sq.head;
pthread_spin_lock(&ctx->bf_lock);
mlx4_bf_copy(ctx->bf_page + ctx->bf_offset, (unsigned long *) ctrl,
align(size * 16, 64));
wc_wmb();
ctx->bf_offset ^= ctx->bf_buf_size;
pthread_spin_unlock(&ctx->bf_lock);
} else if (nreq) {
qp->sq.head += nreq;
/*
* Make sure that descriptors are written before
* doorbell record.
*/
wmb();
*(uint32_t *) (ctx->uar + MLX4_SEND_DOORBELL) = qp->doorbell_qpn;
}
if (nreq)
stamp_send_wqe(qp, (ind + qp->sq_spare_wqes - 1) &
(qp->sq.wqe_cnt - 1));
pthread_spin_unlock(&qp->sq.lock);
return ret;
}