int c2_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
{
struct c2_mq_shared __iomem *shared;
struct c2_cq *cq;
unsigned long flags;
int ret = 0;
cq = to_c2cq(ibcq);
shared = cq->mq.peer;
if ((notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_NEXT_COMP)
writeb(C2_CQ_NOTIFICATION_TYPE_NEXT, &shared->notification_type);
else if ((notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
writeb(C2_CQ_NOTIFICATION_TYPE_NEXT_SE, &shared->notification_type);
else
return -EINVAL;
writeb(CQ_WAIT_FOR_DMA | CQ_ARMED, &shared->armed);
/*
* Now read back shared->armed to make the PCI
* write synchronous. This is necessary for
* correct cq notification semantics.
*/
readb(&shared->armed);
if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
spin_lock_irqsave(&cq->lock, flags);
ret = !c2_mq_empty(&cq->mq);
spin_unlock_irqrestore(&cq->lock, flags);
}
return ret;
}
void c2_free_qp(struct c2_dev *c2dev, struct c2_qp *qp)
{
struct c2_cq *send_cq;
struct c2_cq *recv_cq;
send_cq = to_c2cq(qp->ibqp.send_cq);
recv_cq = to_c2cq(qp->ibqp.recv_cq);
/*
* Lock CQs here, so that CQ polling code can do QP lookup
* without taking a lock.
*/
c2_lock_cqs(send_cq, recv_cq);
c2_free_qpn(c2dev, qp->qpn);
c2_unlock_cqs(send_cq, recv_cq);
/*
* Destory qp in the rnic...
*/
destroy_qp(c2dev, qp);
/*
* Mark any unreaped CQEs as null and void.
*/
c2_cq_clean(c2dev, qp, send_cq->cqn);
if (send_cq != recv_cq)
c2_cq_clean(c2dev, qp, recv_cq->cqn);
/*
* Unmap the MQs and return the shared pointers
* to the message pool.
*/
iounmap(qp->sq_mq.peer);
iounmap(qp->rq_mq.peer);
c2_free_mqsp(qp->sq_mq.shared);
c2_free_mqsp(qp->rq_mq.shared);
atomic_dec(&qp->refcount);
wait_event(qp->wait, !atomic_read(&qp->refcount));
}
int c2_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
{
struct c2_dev *c2dev = to_c2dev(ibcq->device);
struct c2_cq *cq = to_c2cq(ibcq);
unsigned long flags;
int npolled, err;
spin_lock_irqsave(&cq->lock, flags);
for (npolled = 0; npolled < num_entries; ++npolled) {
err = c2_poll_one(c2dev, cq, entry + npolled);
if (err)
break;
}
spin_unlock_irqrestore(&cq->lock, flags);
return npolled;
}
int c2_alloc_qp(struct c2_dev *c2dev,
struct c2_pd *pd,
struct ib_qp_init_attr *qp_attrs, struct c2_qp *qp)
{
struct c2wr_qp_create_req wr;
struct c2wr_qp_create_rep *reply;
struct c2_vq_req *vq_req;
struct c2_cq *send_cq = to_c2cq(qp_attrs->send_cq);
struct c2_cq *recv_cq = to_c2cq(qp_attrs->recv_cq);
unsigned long peer_pa;
u32 q_size, msg_size, mmap_size;
void __iomem *mmap;
int err;
err = c2_alloc_qpn(c2dev, qp);
if (err)
return err;
qp->ibqp.qp_num = qp->qpn;
qp->ibqp.qp_type = IB_QPT_RC;
/* Allocate the SQ and RQ shared pointers */
qp->sq_mq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&qp->sq_mq.shared_dma, GFP_KERNEL);
if (!qp->sq_mq.shared) {
err = -ENOMEM;
goto bail0;
}
qp->rq_mq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&qp->rq_mq.shared_dma, GFP_KERNEL);
if (!qp->rq_mq.shared) {
err = -ENOMEM;
goto bail1;
}
/* Allocate the verbs request */
vq_req = vq_req_alloc(c2dev);
if (vq_req == NULL) {
err = -ENOMEM;
goto bail2;
}
/* Initialize the work request */
memset(&wr, 0, sizeof(wr));
c2_wr_set_id(&wr, CCWR_QP_CREATE);
wr.hdr.context = (unsigned long) vq_req;
wr.rnic_handle = c2dev->adapter_handle;
wr.sq_cq_handle = send_cq->adapter_handle;
wr.rq_cq_handle = recv_cq->adapter_handle;
wr.sq_depth = cpu_to_be32(qp_attrs->cap.max_send_wr + 1);
wr.rq_depth = cpu_to_be32(qp_attrs->cap.max_recv_wr + 1);
wr.srq_handle = 0;
wr.flags = cpu_to_be32(QP_RDMA_READ | QP_RDMA_WRITE | QP_MW_BIND |
QP_ZERO_STAG | QP_RDMA_READ_RESPONSE);
wr.send_sgl_depth = cpu_to_be32(qp_attrs->cap.max_send_sge);
wr.recv_sgl_depth = cpu_to_be32(qp_attrs->cap.max_recv_sge);
wr.rdma_write_sgl_depth = cpu_to_be32(qp_attrs->cap.max_send_sge);
wr.shared_sq_ht = cpu_to_be64(qp->sq_mq.shared_dma);
wr.shared_rq_ht = cpu_to_be64(qp->rq_mq.shared_dma);
wr.ord = cpu_to_be32(C2_MAX_ORD_PER_QP);
wr.ird = cpu_to_be32(C2_MAX_IRD_PER_QP);
wr.pd_id = pd->pd_id;
wr.user_context = (unsigned long) qp;
vq_req_get(c2dev, vq_req);
/* Send the WR to the adapter */
err = vq_send_wr(c2dev, (union c2wr *) & wr);
if (err) {
vq_req_put(c2dev, vq_req);
goto bail3;
}
/* Wait for the verb reply */
err = vq_wait_for_reply(c2dev, vq_req);
if (err) {
goto bail3;
}
/* Process the reply */
reply = (struct c2wr_qp_create_rep *) (unsigned long) (vq_req->reply_msg);
if (!reply) {
err = -ENOMEM;
goto bail3;
}
if ((err = c2_wr_get_result(reply)) != 0) {
goto bail4;
}
/* Fill in the kernel QP struct */
atomic_set(&qp->refcount, 1);
qp->adapter_handle = reply->qp_handle;
qp->state = IB_QPS_RESET;
qp->send_sgl_depth = qp_attrs->cap.max_send_sge;
qp->rdma_write_sgl_depth = qp_attrs->cap.max_send_sge;
qp->recv_sgl_depth = qp_attrs->cap.max_recv_sge;
init_waitqueue_head(&qp->wait);
/* Initialize the SQ MQ */
q_size = be32_to_cpu(reply->sq_depth);
msg_size = be32_to_cpu(reply->sq_msg_size);
peer_pa = c2dev->pa + be32_to_cpu(reply->sq_mq_start);
mmap_size = PAGE_ALIGN(sizeof(struct c2_mq_shared) + msg_size * q_size);
mmap = ioremap_nocache(peer_pa, mmap_size);
if (!mmap) {
err = -ENOMEM;
goto bail5;
}
c2_mq_req_init(&qp->sq_mq,
be32_to_cpu(reply->sq_mq_index),
q_size,
msg_size,
mmap + sizeof(struct c2_mq_shared), /* pool start */
mmap, /* peer */
C2_MQ_ADAPTER_TARGET);
/* Initialize the RQ mq */
q_size = be32_to_cpu(reply->rq_depth);
msg_size = be32_to_cpu(reply->rq_msg_size);
peer_pa = c2dev->pa + be32_to_cpu(reply->rq_mq_start);
mmap_size = PAGE_ALIGN(sizeof(struct c2_mq_shared) + msg_size * q_size);
mmap = ioremap_nocache(peer_pa, mmap_size);
if (!mmap) {
err = -ENOMEM;
goto bail6;
}
c2_mq_req_init(&qp->rq_mq,
be32_to_cpu(reply->rq_mq_index),
q_size,
msg_size,
mmap + sizeof(struct c2_mq_shared), /* pool start */
mmap, /* peer */
C2_MQ_ADAPTER_TARGET);
vq_repbuf_free(c2dev, reply);
vq_req_free(c2dev, vq_req);
return 0;
bail6:
iounmap(qp->sq_mq.peer);
bail5:
destroy_qp(c2dev, qp);
bail4:
vq_repbuf_free(c2dev, reply);
bail3:
vq_req_free(c2dev, vq_req);
bail2:
c2_free_mqsp(qp->rq_mq.shared);
bail1:
c2_free_mqsp(qp->sq_mq.shared);
bail0:
c2_free_qpn(c2dev, qp->qpn);
return err;
}
