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;
}
int c2_init_cq(struct c2_dev *c2dev, int entries,
struct c2_ucontext *ctx, struct c2_cq *cq)
{
struct c2wr_cq_create_req wr;
struct c2wr_cq_create_rep *reply;
unsigned long peer_pa;
struct c2_vq_req *vq_req;
int err;
might_sleep();
cq->ibcq.cqe = entries - 1;
cq->is_kernel = !ctx;
/* Allocate a shared pointer */
cq->mq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&cq->mq.shared_dma, GFP_KERNEL);
if (!cq->mq.shared)
return -ENOMEM;
/* Allocate pages for the message pool */
err = c2_alloc_cq_buf(c2dev, &cq->mq, entries + 1, C2_CQ_MSG_SIZE);
if (err)
goto bail0;
vq_req = vq_req_alloc(c2dev);
if (!vq_req) {
err = -ENOMEM;
goto bail1;
}
memset(&wr, 0, sizeof(wr));
c2_wr_set_id(&wr, CCWR_CQ_CREATE);
wr.hdr.context = (unsigned long) vq_req;
wr.rnic_handle = c2dev->adapter_handle;
wr.msg_size = cpu_to_be32(cq->mq.msg_size);
wr.depth = cpu_to_be32(cq->mq.q_size);
wr.shared_ht = cpu_to_be64(cq->mq.shared_dma);
wr.msg_pool = cpu_to_be64(cq->mq.host_dma);
wr.user_context = (u64) (unsigned long) (cq);
vq_req_get(c2dev, vq_req);
err = vq_send_wr(c2dev, (union c2wr *) & wr);
if (err) {
vq_req_put(c2dev, vq_req);
goto bail2;
}
err = vq_wait_for_reply(c2dev, vq_req);
if (err)
goto bail2;
reply = (struct c2wr_cq_create_rep *) (unsigned long) (vq_req->reply_msg);
if (!reply) {
err = -ENOMEM;
goto bail2;
}
if ((err = c2_errno(reply)) != 0)
goto bail3;
cq->adapter_handle = reply->cq_handle;
cq->mq.index = be32_to_cpu(reply->mq_index);
peer_pa = c2dev->pa + be32_to_cpu(reply->adapter_shared);
cq->mq.peer = ioremap_nocache(peer_pa, PAGE_SIZE);
if (!cq->mq.peer) {
err = -ENOMEM;
goto bail3;
}
vq_repbuf_free(c2dev, reply);
vq_req_free(c2dev, vq_req);
spin_lock_init(&cq->lock);
atomic_set(&cq->refcount, 1);
init_waitqueue_head(&cq->wait);
/*
* Use the MQ index allocated by the adapter to
* store the CQ in the qptr_array
*/
cq->cqn = cq->mq.index;
c2dev->qptr_array[cq->cqn] = cq;
return 0;
bail3:
vq_repbuf_free(c2dev, reply);
bail2:
vq_req_free(c2dev, vq_req);
bail1:
c2_free_cq_buf(c2dev, &cq->mq);
bail0:
c2_free_mqsp(cq->mq.shared);
return err;
}
/*
* Called by c2_probe to initialize the RNIC. This principally
* involves initalizing the various limits and resouce pools that
* comprise the RNIC instance.
*/
int __devinit c2_rnic_init(struct c2_dev *c2dev)
{
int err;
u32 qsize, msgsize;
void *q1_pages;
void *q2_pages;
void __iomem *mmio_regs;
/* Device capabilities */
c2dev->device_cap_flags =
(IB_DEVICE_RESIZE_MAX_WR |
IB_DEVICE_CURR_QP_STATE_MOD |
IB_DEVICE_SYS_IMAGE_GUID |
IB_DEVICE_ZERO_STAG |
IB_DEVICE_MEM_WINDOW);
/* Allocate the qptr_array */
c2dev->qptr_array = vmalloc(C2_MAX_CQS * sizeof(void *));
if (!c2dev->qptr_array) {
return -ENOMEM;
}
/* Inialize the qptr_array */
memset(c2dev->qptr_array, 0, C2_MAX_CQS * sizeof(void *));
c2dev->qptr_array[0] = (void *) &c2dev->req_vq;
c2dev->qptr_array[1] = (void *) &c2dev->rep_vq;
c2dev->qptr_array[2] = (void *) &c2dev->aeq;
/* Initialize data structures */
init_waitqueue_head(&c2dev->req_vq_wo);
spin_lock_init(&c2dev->vqlock);
spin_lock_init(&c2dev->lock);
/* Allocate MQ shared pointer pool for kernel clients. User
* mode client pools are hung off the user context
*/
err = c2_init_mqsp_pool(c2dev, GFP_KERNEL, &c2dev->kern_mqsp_pool);
if (err) {
goto bail0;
}
/* Allocate shared pointers for Q0, Q1, and Q2 from
* the shared pointer pool.
*/
c2dev->hint_count = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&c2dev->hint_count_dma,
GFP_KERNEL);
c2dev->req_vq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&c2dev->req_vq.shared_dma,
GFP_KERNEL);
c2dev->rep_vq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&c2dev->rep_vq.shared_dma,
GFP_KERNEL);
c2dev->aeq.shared = c2_alloc_mqsp(c2dev, c2dev->kern_mqsp_pool,
&c2dev->aeq.shared_dma, GFP_KERNEL);
if (!c2dev->hint_count || !c2dev->req_vq.shared ||
!c2dev->rep_vq.shared || !c2dev->aeq.shared) {
err = -ENOMEM;
goto bail1;
}
mmio_regs = c2dev->kva;
/* Initialize the Verbs Request Queue */
c2_mq_req_init(&c2dev->req_vq, 0,
be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q0_QSIZE)),
be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q0_MSGSIZE)),
mmio_regs +
be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q0_POOLSTART)),
mmio_regs +
be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q0_SHARED)),
C2_MQ_ADAPTER_TARGET);
/* Initialize the Verbs Reply Queue */
qsize = be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q1_QSIZE));
msgsize = be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q1_MSGSIZE));
q1_pages = dma_alloc_coherent(&c2dev->pcidev->dev, qsize * msgsize,
&c2dev->rep_vq.host_dma, GFP_KERNEL);
if (!q1_pages) {
err = -ENOMEM;
goto bail1;
}
pci_unmap_addr_set(&c2dev->rep_vq, mapping, c2dev->rep_vq.host_dma);
pr_debug("%s rep_vq va %p dma %llx\n", __func__, q1_pages,
(unsigned long long) c2dev->rep_vq.host_dma);
c2_mq_rep_init(&c2dev->rep_vq,
1,
qsize,
msgsize,
q1_pages,
mmio_regs +
be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q1_SHARED)),
C2_MQ_HOST_TARGET);
/* Initialize the Asynchronus Event Queue */
qsize = be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q2_QSIZE));
msgsize = be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q2_MSGSIZE));
q2_pages = dma_alloc_coherent(&c2dev->pcidev->dev, qsize * msgsize,
&c2dev->aeq.host_dma, GFP_KERNEL);
if (!q2_pages) {
err = -ENOMEM;
goto bail2;
}
pci_unmap_addr_set(&c2dev->aeq, mapping, c2dev->aeq.host_dma);
pr_debug("%s aeq va %p dma %llx\n", __func__, q2_pages,
(unsigned long long) c2dev->aeq.host_dma);
c2_mq_rep_init(&c2dev->aeq,
2,
qsize,
msgsize,
q2_pages,
mmio_regs +
be32_to_cpu((__force __be32) readl(mmio_regs + C2_REGS_Q2_SHARED)),
C2_MQ_HOST_TARGET);
/* Initialize the verbs request allocator */
err = vq_init(c2dev);
if (err)
goto bail3;
/* Enable interrupts on the adapter */
writel(0, c2dev->regs + C2_IDIS);
/* create the WR init message */
err = c2_adapter_init(c2dev);
if (err)
goto bail4;
c2dev->init++;
/* open an adapter instance */
err = c2_rnic_open(c2dev);
if (err)
goto bail4;
/* Initialize cached the adapter limits */
if (c2_rnic_query(c2dev, &c2dev->props))
goto bail5;
/* Initialize the PD pool */
err = c2_init_pd_table(c2dev);
if (err)
goto bail5;
/* Initialize the QP pool */
c2_init_qp_table(c2dev);
return 0;
bail5:
c2_rnic_close(c2dev);
bail4:
vq_term(c2dev);
bail3:
dma_free_coherent(&c2dev->pcidev->dev,
c2dev->aeq.q_size * c2dev->aeq.msg_size,
q2_pages, pci_unmap_addr(&c2dev->aeq, mapping));
bail2:
dma_free_coherent(&c2dev->pcidev->dev,
c2dev->rep_vq.q_size * c2dev->rep_vq.msg_size,
q1_pages, pci_unmap_addr(&c2dev->rep_vq, mapping));
bail1:
c2_free_mqsp_pool(c2dev, c2dev->kern_mqsp_pool);
bail0:
vfree(c2dev->qptr_array);
return err;
}