]> git.baikalelectronics.ru Git - kernel.git/commitdiff
qedr: Add support for QP verbs
authorRam Amrani <Ram.Amrani@cavium.com>
Mon, 10 Oct 2016 10:15:34 +0000 (13:15 +0300)
committerDoug Ledford <dledford@redhat.com>
Fri, 14 Oct 2016 19:00:10 +0000 (15:00 -0400)
Add support for Queue Pair verbs which adds, deletes,
modifies and queries Queue Pairs.

Signed-off-by: Rajesh Borundia <rajesh.borundia@cavium.com>
Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/qedr/main.c
drivers/infiniband/hw/qedr/qedr.h
drivers/infiniband/hw/qedr/qedr_cm.h [new file with mode: 0644]
drivers/infiniband/hw/qedr/qedr_hsi_rdma.h
drivers/infiniband/hw/qedr/verbs.c
drivers/infiniband/hw/qedr/verbs.h
include/uapi/rdma/qedr-abi.h

index 35928abb6b639524fe39da948d8bc06a555e1b33..13ba47b7b99fd6dcfdd31881abb306543d1c2933 100644 (file)
@@ -48,6 +48,8 @@ MODULE_AUTHOR("QLogic Corporation");
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_VERSION(QEDR_MODULE_VERSION);
 
+#define QEDR_WQ_MULTIPLIER_DFT (3)
+
 void qedr_ib_dispatch_event(struct qedr_dev *dev, u8 port_num,
                            enum ib_event_type type)
 {
@@ -94,7 +96,11 @@ static int qedr_register_device(struct qedr_dev *dev)
                                     QEDR_UVERBS(CREATE_CQ) |
                                     QEDR_UVERBS(RESIZE_CQ) |
                                     QEDR_UVERBS(DESTROY_CQ) |
-                                    QEDR_UVERBS(REQ_NOTIFY_CQ);
+                                    QEDR_UVERBS(REQ_NOTIFY_CQ) |
+                                    QEDR_UVERBS(CREATE_QP) |
+                                    QEDR_UVERBS(MODIFY_QP) |
+                                    QEDR_UVERBS(QUERY_QP) |
+                                    QEDR_UVERBS(DESTROY_QP);
 
        dev->ibdev.phys_port_cnt = 1;
        dev->ibdev.num_comp_vectors = dev->num_cnq;
@@ -120,6 +126,11 @@ static int qedr_register_device(struct qedr_dev *dev)
        dev->ibdev.resize_cq = qedr_resize_cq;
        dev->ibdev.req_notify_cq = qedr_arm_cq;
 
+       dev->ibdev.create_qp = qedr_create_qp;
+       dev->ibdev.modify_qp = qedr_modify_qp;
+       dev->ibdev.query_qp = qedr_query_qp;
+       dev->ibdev.destroy_qp = qedr_destroy_qp;
+
        dev->ibdev.query_pkey = qedr_query_pkey;
 
        dev->ibdev.dma_device = &dev->pdev->dev;
@@ -630,6 +641,8 @@ static struct qedr_dev *qedr_add(struct qed_dev *cdev, struct pci_dev *pdev,
                goto init_err;
        }
 
+       dev->wq_multiplier = QEDR_WQ_MULTIPLIER_DFT;
+
        qedr_pci_set_atomic(dev, pdev);
 
        rc = qedr_alloc_resources(dev);
index 9e2846a599cea332f4aca2dbf4457aebf328d11e..e9fe941c48ad743c5057b2358ef69e272ba0bf3c 100644 (file)
@@ -52,6 +52,9 @@
 #define QEDR_MSG_MISC "MISC"
 #define QEDR_MSG_CQ   "  CQ"
 #define QEDR_MSG_MR   "  MR"
+#define QEDR_MSG_RQ   "  RQ"
+#define QEDR_MSG_SQ   "  SQ"
+#define QEDR_MSG_QP   "  QP"
 
 #define QEDR_CQ_MAGIC_NUMBER   (0x11223344)
 
@@ -143,6 +146,8 @@ struct qedr_dev {
        u32                     dp_module;
        u8                      dp_level;
        u8                      num_hwfns;
+       uint                    wq_multiplier;
+
 };
 
 #define QEDR_MAX_SQ_PBL                        (0x8000)
@@ -272,6 +277,122 @@ struct qedr_mm {
        struct list_head entry;
 };
 
+union db_prod32 {
+       struct rdma_pwm_val16_data data;
+       u32 raw;
+};
+
+struct qedr_qp_hwq_info {
+       /* WQE Elements */
+       struct qed_chain pbl;
+       u64 p_phys_addr_tbl;
+       u32 max_sges;
+
+       /* WQE */
+       u16 prod;
+       u16 cons;
+       u16 wqe_cons;
+       u16 max_wr;
+
+       /* DB */
+       void __iomem *db;
+       union db_prod32 db_data;
+};
+
+#define QEDR_INC_SW_IDX(p_info, index)                                 \
+       do {                                                            \
+               p_info->index = (p_info->index + 1) &                   \
+                               qed_chain_get_capacity(p_info->pbl)     \
+       } while (0)
+
+enum qedr_qp_err_bitmap {
+       QEDR_QP_ERR_SQ_FULL = 1,
+       QEDR_QP_ERR_RQ_FULL = 2,
+       QEDR_QP_ERR_BAD_SR = 4,
+       QEDR_QP_ERR_BAD_RR = 8,
+       QEDR_QP_ERR_SQ_PBL_FULL = 16,
+       QEDR_QP_ERR_RQ_PBL_FULL = 32,
+};
+
+struct qedr_qp {
+       struct ib_qp ibqp;      /* must be first */
+       struct qedr_dev *dev;
+
+       struct qedr_qp_hwq_info sq;
+       struct qedr_qp_hwq_info rq;
+
+       u32 max_inline_data;
+
+       /* Lock for QP's */
+       spinlock_t q_lock;
+       struct qedr_cq *sq_cq;
+       struct qedr_cq *rq_cq;
+       struct qedr_srq *srq;
+       enum qed_roce_qp_state state;
+       u32 id;
+       struct qedr_pd *pd;
+       enum ib_qp_type qp_type;
+       struct qed_rdma_qp *qed_qp;
+       u32 qp_id;
+       u16 icid;
+       u16 mtu;
+       int sgid_idx;
+       u32 rq_psn;
+       u32 sq_psn;
+       u32 qkey;
+       u32 dest_qp_num;
+
+       /* Relevant to qps created from kernel space only (ULPs) */
+       u8 prev_wqe_size;
+       u16 wqe_cons;
+       u32 err_bitmap;
+       bool signaled;
+
+       /* SQ shadow */
+       struct {
+               u64 wr_id;
+               enum ib_wc_opcode opcode;
+               u32 bytes_len;
+               u8 wqe_size;
+               bool signaled;
+               dma_addr_t icrc_mapping;
+               u32 *icrc;
+               struct qedr_mr *mr;
+       } *wqe_wr_id;
+
+       /* RQ shadow */
+       struct {
+               u64 wr_id;
+               struct ib_sge sg_list[RDMA_MAX_SGE_PER_RQ_WQE];
+               u8 wqe_size;
+
+               u16 vlan_id;
+               int rc;
+       } *rqe_wr_id;
+
+       /* Relevant to qps created from user space only (applications) */
+       struct qedr_userq usq;
+       struct qedr_userq urq;
+};
+
+static inline int qedr_get_dmac(struct qedr_dev *dev,
+                               struct ib_ah_attr *ah_attr, u8 *mac_addr)
+{
+       union ib_gid zero_sgid = { { 0 } };
+       struct in6_addr in6;
+
+       if (!memcmp(&ah_attr->grh.dgid, &zero_sgid, sizeof(union ib_gid))) {
+               DP_ERR(dev, "Local port GID not supported\n");
+               eth_zero_addr(mac_addr);
+               return -EINVAL;
+       }
+
+       memcpy(&in6, ah_attr->grh.dgid.raw, sizeof(in6));
+       ether_addr_copy(mac_addr, ah_attr->dmac);
+
+       return 0;
+}
+
 static inline
 struct qedr_ucontext *get_qedr_ucontext(struct ib_ucontext *ibucontext)
 {
@@ -293,4 +414,8 @@ static inline struct qedr_cq *get_qedr_cq(struct ib_cq *ibcq)
        return container_of(ibcq, struct qedr_cq, ibcq);
 }
 
+static inline struct qedr_qp *get_qedr_qp(struct ib_qp *ibqp)
+{
+       return container_of(ibqp, struct qedr_qp, ibqp);
+}
 #endif
diff --git a/drivers/infiniband/hw/qedr/qedr_cm.h b/drivers/infiniband/hw/qedr/qedr_cm.h
new file mode 100644 (file)
index 0000000..b8a8b76
--- /dev/null
@@ -0,0 +1,40 @@
+/* QLogic qedr NIC Driver
+ * Copyright (c) 2015-2016  QLogic Corporation
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and /or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef LINUX_QEDR_CM_H_
+#define LINUX_QEDR_CM_H_
+
+static inline u32 qedr_get_ipv4_from_gid(u8 *gid)
+{
+       return *(u32 *)(void *)&gid[12];
+}
+
+#endif
index 84f6520107cc7f48cdebe3e16705fe7edd9c10bb..47705598eec6037357b4f7e3d86b2fc32455e4df 100644 (file)
@@ -158,6 +158,17 @@ struct rdma_srq_sge {
        __le32 l_key;
 };
 
+/* Rdma doorbell data for SQ and RQ */
+struct rdma_pwm_val16_data {
+       __le16 icid;
+       __le16 value;
+};
+
+union rdma_pwm_val16_data_union {
+       struct rdma_pwm_val16_data as_struct;
+       __le32 as_dword;
+};
+
 /* Rdma doorbell data for CQ */
 struct rdma_pwm_val32_data {
        __le16 icid;
index b525c6cf1df05a9760fa5050ddab00cad3d259e8..a0d1c5fffb63200dfb60751ac9b3fb46cf9794fa 100644 (file)
@@ -48,6 +48,7 @@
 #include "qedr.h"
 #include "verbs.h"
 #include <rdma/qedr-abi.h>
+#include "qedr_cm.h"
 
 #define DB_ADDR_SHIFT(addr)            ((addr) << DB_PWM_ADDR_OFFSET_SHIFT)
 
@@ -993,3 +994,1091 @@ int qedr_destroy_cq(struct ib_cq *ibcq)
 
        return 0;
 }
+
+static inline int get_gid_info_from_table(struct ib_qp *ibqp,
+                                         struct ib_qp_attr *attr,
+                                         int attr_mask,
+                                         struct qed_rdma_modify_qp_in_params
+                                         *qp_params)
+{
+       enum rdma_network_type nw_type;
+       struct ib_gid_attr gid_attr;
+       union ib_gid gid;
+       u32 ipv4_addr;
+       int rc = 0;
+       int i;
+
+       rc = ib_get_cached_gid(ibqp->device, attr->ah_attr.port_num,
+                              attr->ah_attr.grh.sgid_index, &gid, &gid_attr);
+       if (rc)
+               return rc;
+
+       if (!memcmp(&gid, &zgid, sizeof(gid)))
+               return -ENOENT;
+
+       if (gid_attr.ndev) {
+               qp_params->vlan_id = rdma_vlan_dev_vlan_id(gid_attr.ndev);
+
+               dev_put(gid_attr.ndev);
+               nw_type = ib_gid_to_network_type(gid_attr.gid_type, &gid);
+               switch (nw_type) {
+               case RDMA_NETWORK_IPV6:
+                       memcpy(&qp_params->sgid.bytes[0], &gid.raw[0],
+                              sizeof(qp_params->sgid));
+                       memcpy(&qp_params->dgid.bytes[0],
+                              &attr->ah_attr.grh.dgid,
+                              sizeof(qp_params->dgid));
+                       qp_params->roce_mode = ROCE_V2_IPV6;
+                       SET_FIELD(qp_params->modify_flags,
+                                 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
+                       break;
+               case RDMA_NETWORK_IB:
+                       memcpy(&qp_params->sgid.bytes[0], &gid.raw[0],
+                              sizeof(qp_params->sgid));
+                       memcpy(&qp_params->dgid.bytes[0],
+                              &attr->ah_attr.grh.dgid,
+                              sizeof(qp_params->dgid));
+                       qp_params->roce_mode = ROCE_V1;
+                       break;
+               case RDMA_NETWORK_IPV4:
+                       memset(&qp_params->sgid, 0, sizeof(qp_params->sgid));
+                       memset(&qp_params->dgid, 0, sizeof(qp_params->dgid));
+                       ipv4_addr = qedr_get_ipv4_from_gid(gid.raw);
+                       qp_params->sgid.ipv4_addr = ipv4_addr;
+                       ipv4_addr =
+                           qedr_get_ipv4_from_gid(attr->ah_attr.grh.dgid.raw);
+                       qp_params->dgid.ipv4_addr = ipv4_addr;
+                       SET_FIELD(qp_params->modify_flags,
+                                 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
+                       qp_params->roce_mode = ROCE_V2_IPV4;
+                       break;
+               }
+       }
+
+       for (i = 0; i < 4; i++) {
+               qp_params->sgid.dwords[i] = ntohl(qp_params->sgid.dwords[i]);
+               qp_params->dgid.dwords[i] = ntohl(qp_params->dgid.dwords[i]);
+       }
+
+       if (qp_params->vlan_id >= VLAN_CFI_MASK)
+               qp_params->vlan_id = 0;
+
+       return 0;
+}
+
+static void qedr_cleanup_user_sq(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       qedr_free_pbl(dev, &qp->usq.pbl_info, qp->usq.pbl_tbl);
+       ib_umem_release(qp->usq.umem);
+}
+
+static void qedr_cleanup_user_rq(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       qedr_free_pbl(dev, &qp->urq.pbl_info, qp->urq.pbl_tbl);
+       ib_umem_release(qp->urq.umem);
+}
+
+static void qedr_cleanup_kernel_sq(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl);
+       kfree(qp->wqe_wr_id);
+}
+
+static void qedr_cleanup_kernel_rq(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl);
+       kfree(qp->rqe_wr_id);
+}
+
+static int qedr_check_qp_attrs(struct ib_pd *ibpd, struct qedr_dev *dev,
+                              struct ib_qp_init_attr *attrs)
+{
+       struct qedr_device_attr *qattr = &dev->attr;
+
+       /* QP0... attrs->qp_type == IB_QPT_GSI */
+       if (attrs->qp_type != IB_QPT_RC && attrs->qp_type != IB_QPT_GSI) {
+               DP_DEBUG(dev, QEDR_MSG_QP,
+                        "create qp: unsupported qp type=0x%x requested\n",
+                        attrs->qp_type);
+               return -EINVAL;
+       }
+
+       if (attrs->cap.max_send_wr > qattr->max_sqe) {
+               DP_ERR(dev,
+                      "create qp: cannot create a SQ with %d elements (max_send_wr=0x%x)\n",
+                      attrs->cap.max_send_wr, qattr->max_sqe);
+               return -EINVAL;
+       }
+
+       if (attrs->cap.max_inline_data > qattr->max_inline) {
+               DP_ERR(dev,
+                      "create qp: unsupported inline data size=0x%x requested (max_inline=0x%x)\n",
+                      attrs->cap.max_inline_data, qattr->max_inline);
+               return -EINVAL;
+       }
+
+       if (attrs->cap.max_send_sge > qattr->max_sge) {
+               DP_ERR(dev,
+                      "create qp: unsupported send_sge=0x%x requested (max_send_sge=0x%x)\n",
+                      attrs->cap.max_send_sge, qattr->max_sge);
+               return -EINVAL;
+       }
+
+       if (attrs->cap.max_recv_sge > qattr->max_sge) {
+               DP_ERR(dev,
+                      "create qp: unsupported recv_sge=0x%x requested (max_recv_sge=0x%x)\n",
+                      attrs->cap.max_recv_sge, qattr->max_sge);
+               return -EINVAL;
+       }
+
+       /* Unprivileged user space cannot create special QP */
+       if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
+               DP_ERR(dev,
+                      "create qp: userspace can't create special QPs of type=0x%x\n",
+                      attrs->qp_type);
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static void qedr_copy_rq_uresp(struct qedr_create_qp_uresp *uresp,
+                              struct qedr_qp *qp)
+{
+       uresp->rq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
+       uresp->rq_icid = qp->icid;
+}
+
+static void qedr_copy_sq_uresp(struct qedr_create_qp_uresp *uresp,
+                              struct qedr_qp *qp)
+{
+       uresp->sq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
+       uresp->sq_icid = qp->icid + 1;
+}
+
+static int qedr_copy_qp_uresp(struct qedr_dev *dev,
+                             struct qedr_qp *qp, struct ib_udata *udata)
+{
+       struct qedr_create_qp_uresp uresp;
+       int rc;
+
+       memset(&uresp, 0, sizeof(uresp));
+       qedr_copy_sq_uresp(&uresp, qp);
+       qedr_copy_rq_uresp(&uresp, qp);
+
+       uresp.atomic_supported = dev->atomic_cap != IB_ATOMIC_NONE;
+       uresp.qp_id = qp->qp_id;
+
+       rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
+       if (rc)
+               DP_ERR(dev,
+                      "create qp: failed a copy to user space with qp icid=0x%x.\n",
+                      qp->icid);
+
+       return rc;
+}
+
+static void qedr_set_qp_init_params(struct qedr_dev *dev,
+                                   struct qedr_qp *qp,
+                                   struct qedr_pd *pd,
+                                   struct ib_qp_init_attr *attrs)
+{
+       qp->pd = pd;
+
+       spin_lock_init(&qp->q_lock);
+
+       qp->qp_type = attrs->qp_type;
+       qp->max_inline_data = attrs->cap.max_inline_data;
+       qp->sq.max_sges = attrs->cap.max_send_sge;
+       qp->state = QED_ROCE_QP_STATE_RESET;
+       qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
+       qp->sq_cq = get_qedr_cq(attrs->send_cq);
+       qp->rq_cq = get_qedr_cq(attrs->recv_cq);
+       qp->dev = dev;
+
+       DP_DEBUG(dev, QEDR_MSG_QP,
+                "QP params:\tpd = %d, qp_type = %d, max_inline_data = %d, state = %d, signaled = %d, use_srq=%d\n",
+                pd->pd_id, qp->qp_type, qp->max_inline_data,
+                qp->state, qp->signaled, (attrs->srq) ? 1 : 0);
+       DP_DEBUG(dev, QEDR_MSG_QP,
+                "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n",
+                qp->sq.max_sges, qp->sq_cq->icid);
+       qp->rq.max_sges = attrs->cap.max_recv_sge;
+       DP_DEBUG(dev, QEDR_MSG_QP,
+                "RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
+                qp->rq.max_sges, qp->rq_cq->icid);
+}
+
+static inline void
+qedr_init_qp_user_params(struct qed_rdma_create_qp_in_params *params,
+                        struct qedr_create_qp_ureq *ureq)
+{
+       /* QP handle to be written in CQE */
+       params->qp_handle_lo = ureq->qp_handle_lo;
+       params->qp_handle_hi = ureq->qp_handle_hi;
+}
+
+static inline void
+qedr_init_qp_kernel_doorbell_sq(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       qp->sq.db = dev->db_addr +
+                   DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
+       qp->sq.db_data.data.icid = qp->icid + 1;
+}
+
+static inline void
+qedr_init_qp_kernel_doorbell_rq(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       qp->rq.db = dev->db_addr +
+                   DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
+       qp->rq.db_data.data.icid = qp->icid;
+}
+
+static inline int
+qedr_init_qp_kernel_params_rq(struct qedr_dev *dev,
+                             struct qedr_qp *qp, struct ib_qp_init_attr *attrs)
+{
+       /* Allocate driver internal RQ array */
+       qp->rqe_wr_id = kcalloc(qp->rq.max_wr, sizeof(*qp->rqe_wr_id),
+                               GFP_KERNEL);
+       if (!qp->rqe_wr_id)
+               return -ENOMEM;
+
+       DP_DEBUG(dev, QEDR_MSG_QP, "RQ max_wr set to %d.\n", qp->rq.max_wr);
+
+       return 0;
+}
+
+static inline int
+qedr_init_qp_kernel_params_sq(struct qedr_dev *dev,
+                             struct qedr_qp *qp,
+                             struct ib_qp_init_attr *attrs,
+                             struct qed_rdma_create_qp_in_params *params)
+{
+       u32 temp_max_wr;
+
+       /* Allocate driver internal SQ array */
+       temp_max_wr = attrs->cap.max_send_wr * dev->wq_multiplier;
+       temp_max_wr = min_t(u32, temp_max_wr, dev->attr.max_sqe);
+
+       /* temp_max_wr < attr->max_sqe < u16 so the casting is safe */
+       qp->sq.max_wr = (u16)temp_max_wr;
+       qp->wqe_wr_id = kcalloc(qp->sq.max_wr, sizeof(*qp->wqe_wr_id),
+                               GFP_KERNEL);
+       if (!qp->wqe_wr_id)
+               return -ENOMEM;
+
+       DP_DEBUG(dev, QEDR_MSG_QP, "SQ max_wr set to %d.\n", qp->sq.max_wr);
+
+       /* QP handle to be written in CQE */
+       params->qp_handle_lo = lower_32_bits((uintptr_t)qp);
+       params->qp_handle_hi = upper_32_bits((uintptr_t)qp);
+
+       return 0;
+}
+
+static inline int qedr_init_qp_kernel_sq(struct qedr_dev *dev,
+                                        struct qedr_qp *qp,
+                                        struct ib_qp_init_attr *attrs)
+{
+       u32 n_sq_elems, n_sq_entries;
+       int rc;
+
+       /* A single work request may take up to QEDR_MAX_SQ_WQE_SIZE elements in
+        * the ring. The ring should allow at least a single WR, even if the
+        * user requested none, due to allocation issues.
+        */
+       n_sq_entries = attrs->cap.max_send_wr;
+       n_sq_entries = min_t(u32, n_sq_entries, dev->attr.max_sqe);
+       n_sq_entries = max_t(u32, n_sq_entries, 1);
+       n_sq_elems = n_sq_entries * QEDR_MAX_SQE_ELEMENTS_PER_SQE;
+       rc = dev->ops->common->chain_alloc(dev->cdev,
+                                          QED_CHAIN_USE_TO_PRODUCE,
+                                          QED_CHAIN_MODE_PBL,
+                                          QED_CHAIN_CNT_TYPE_U32,
+                                          n_sq_elems,
+                                          QEDR_SQE_ELEMENT_SIZE,
+                                          &qp->sq.pbl);
+       if (rc) {
+               DP_ERR(dev, "failed to allocate QP %p SQ\n", qp);
+               return rc;
+       }
+
+       DP_DEBUG(dev, QEDR_MSG_SQ,
+                "SQ Pbl base addr = %llx max_send_wr=%d max_wr=%d capacity=%d, rc=%d\n",
+                qed_chain_get_pbl_phys(&qp->sq.pbl), attrs->cap.max_send_wr,
+                n_sq_entries, qed_chain_get_capacity(&qp->sq.pbl), rc);
+       return 0;
+}
+
+static inline int qedr_init_qp_kernel_rq(struct qedr_dev *dev,
+                                        struct qedr_qp *qp,
+                                        struct ib_qp_init_attr *attrs)
+{
+       u32 n_rq_elems, n_rq_entries;
+       int rc;
+
+       /* A single work request may take up to QEDR_MAX_RQ_WQE_SIZE elements in
+        * the ring. There ring should allow at least a single WR, even if the
+        * user requested none, due to allocation issues.
+        */
+       n_rq_entries = max_t(u32, attrs->cap.max_recv_wr, 1);
+       n_rq_elems = n_rq_entries * QEDR_MAX_RQE_ELEMENTS_PER_RQE;
+       rc = dev->ops->common->chain_alloc(dev->cdev,
+                                          QED_CHAIN_USE_TO_CONSUME_PRODUCE,
+                                          QED_CHAIN_MODE_PBL,
+                                          QED_CHAIN_CNT_TYPE_U32,
+                                          n_rq_elems,
+                                          QEDR_RQE_ELEMENT_SIZE,
+                                          &qp->rq.pbl);
+
+       if (rc) {
+               DP_ERR(dev, "failed to allocate memory for QP %p RQ\n", qp);
+               return -ENOMEM;
+       }
+
+       DP_DEBUG(dev, QEDR_MSG_RQ,
+                "RQ Pbl base addr = %llx max_recv_wr=%d max_wr=%d capacity=%d, rc=%d\n",
+                qed_chain_get_pbl_phys(&qp->rq.pbl), attrs->cap.max_recv_wr,
+                n_rq_entries, qed_chain_get_capacity(&qp->rq.pbl), rc);
+
+       /* n_rq_entries < u16 so the casting is safe */
+       qp->rq.max_wr = (u16)n_rq_entries;
+
+       return 0;
+}
+
+static inline void
+qedr_init_qp_in_params_sq(struct qedr_dev *dev,
+                         struct qedr_pd *pd,
+                         struct qedr_qp *qp,
+                         struct ib_qp_init_attr *attrs,
+                         struct ib_udata *udata,
+                         struct qed_rdma_create_qp_in_params *params)
+{
+       /* QP handle to be written in an async event */
+       params->qp_handle_async_lo = lower_32_bits((uintptr_t)qp);
+       params->qp_handle_async_hi = upper_32_bits((uintptr_t)qp);
+
+       params->signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR);
+       params->fmr_and_reserved_lkey = !udata;
+       params->pd = pd->pd_id;
+       params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
+       params->sq_cq_id = get_qedr_cq(attrs->send_cq)->icid;
+       params->max_sq_sges = 0;
+       params->stats_queue = 0;
+
+       if (udata) {
+               params->sq_num_pages = qp->usq.pbl_info.num_pbes;
+               params->sq_pbl_ptr = qp->usq.pbl_tbl->pa;
+       } else {
+               params->sq_num_pages = qed_chain_get_page_cnt(&qp->sq.pbl);
+               params->sq_pbl_ptr = qed_chain_get_pbl_phys(&qp->sq.pbl);
+       }
+}
+
+static inline void
+qedr_init_qp_in_params_rq(struct qedr_qp *qp,
+                         struct ib_qp_init_attr *attrs,
+                         struct ib_udata *udata,
+                         struct qed_rdma_create_qp_in_params *params)
+{
+       params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
+       params->srq_id = 0;
+       params->use_srq = false;
+
+       if (udata) {
+               params->rq_num_pages = qp->urq.pbl_info.num_pbes;
+               params->rq_pbl_ptr = qp->urq.pbl_tbl->pa;
+       } else {
+               params->rq_num_pages = qed_chain_get_page_cnt(&qp->rq.pbl);
+               params->rq_pbl_ptr = qed_chain_get_pbl_phys(&qp->rq.pbl);
+       }
+}
+
+static inline void qedr_qp_user_print(struct qedr_dev *dev, struct qedr_qp *qp)
+{
+       DP_DEBUG(dev, QEDR_MSG_QP,
+                "create qp: successfully created user QP. qp=%p, sq_addr=0x%llx, sq_len=%zd, rq_addr=0x%llx, rq_len=%zd\n",
+                qp, qp->usq.buf_addr, qp->usq.buf_len, qp->urq.buf_addr,
+                qp->urq.buf_len);
+}
+
+static inline int qedr_init_user_qp(struct ib_ucontext *ib_ctx,
+                                   struct qedr_dev *dev,
+                                   struct qedr_qp *qp,
+                                   struct qedr_create_qp_ureq *ureq)
+{
+       int rc;
+
+       /* SQ - read access only (0), dma sync not required (0) */
+       rc = qedr_init_user_queue(ib_ctx, dev, &qp->usq, ureq->sq_addr,
+                                 ureq->sq_len, 0, 0);
+       if (rc)
+               return rc;
+
+       /* RQ - read access only (0), dma sync not required (0) */
+       rc = qedr_init_user_queue(ib_ctx, dev, &qp->urq, ureq->rq_addr,
+                                 ureq->rq_len, 0, 0);
+
+       if (rc)
+               qedr_cleanup_user_sq(dev, qp);
+       return rc;
+}
+
+static inline int
+qedr_init_kernel_qp(struct qedr_dev *dev,
+                   struct qedr_qp *qp,
+                   struct ib_qp_init_attr *attrs,
+                   struct qed_rdma_create_qp_in_params *params)
+{
+       int rc;
+
+       rc = qedr_init_qp_kernel_sq(dev, qp, attrs);
+       if (rc) {
+               DP_ERR(dev, "failed to init kernel QP %p SQ\n", qp);
+               return rc;
+       }
+
+       rc = qedr_init_qp_kernel_params_sq(dev, qp, attrs, params);
+       if (rc) {
+               dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl);
+               DP_ERR(dev, "failed to init kernel QP %p SQ params\n", qp);
+               return rc;
+       }
+
+       rc = qedr_init_qp_kernel_rq(dev, qp, attrs);
+       if (rc) {
+               qedr_cleanup_kernel_sq(dev, qp);
+               DP_ERR(dev, "failed to init kernel QP %p RQ\n", qp);
+               return rc;
+       }
+
+       rc = qedr_init_qp_kernel_params_rq(dev, qp, attrs);
+       if (rc) {
+               DP_ERR(dev, "failed to init kernel QP %p RQ params\n", qp);
+               qedr_cleanup_kernel_sq(dev, qp);
+               dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl);
+               return rc;
+       }
+
+       return rc;
+}
+
+struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
+                            struct ib_qp_init_attr *attrs,
+                            struct ib_udata *udata)
+{
+       struct qedr_dev *dev = get_qedr_dev(ibpd->device);
+       struct qed_rdma_create_qp_out_params out_params;
+       struct qed_rdma_create_qp_in_params in_params;
+       struct qedr_pd *pd = get_qedr_pd(ibpd);
+       struct ib_ucontext *ib_ctx = NULL;
+       struct qedr_ucontext *ctx = NULL;
+       struct qedr_create_qp_ureq ureq;
+       struct qedr_qp *qp;
+       int rc = 0;
+
+       DP_DEBUG(dev, QEDR_MSG_QP, "create qp: called from %s, pd=%p\n",
+                udata ? "user library" : "kernel", pd);
+
+       rc = qedr_check_qp_attrs(ibpd, dev, attrs);
+       if (rc)
+               return ERR_PTR(rc);
+
+       qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+       if (!qp)
+               return ERR_PTR(-ENOMEM);
+
+       if (attrs->srq)
+               return ERR_PTR(-EINVAL);
+
+       DP_DEBUG(dev, QEDR_MSG_QP,
+                "create qp: sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n",
+                get_qedr_cq(attrs->send_cq),
+                get_qedr_cq(attrs->send_cq)->icid,
+                get_qedr_cq(attrs->recv_cq),
+                get_qedr_cq(attrs->recv_cq)->icid);
+
+       qedr_set_qp_init_params(dev, qp, pd, attrs);
+
+       memset(&in_params, 0, sizeof(in_params));
+
+       if (udata) {
+               if (!(udata && ibpd->uobject && ibpd->uobject->context))
+                       goto err0;
+
+               ib_ctx = ibpd->uobject->context;
+               ctx = get_qedr_ucontext(ib_ctx);
+
+               memset(&ureq, 0, sizeof(ureq));
+               if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) {
+                       DP_ERR(dev,
+                              "create qp: problem copying data from user space\n");
+                       goto err0;
+               }
+
+               rc = qedr_init_user_qp(ib_ctx, dev, qp, &ureq);
+               if (rc)
+                       goto err0;
+
+               qedr_init_qp_user_params(&in_params, &ureq);
+       } else {
+               rc = qedr_init_kernel_qp(dev, qp, attrs, &in_params);
+               if (rc)
+                       goto err0;
+       }
+
+       qedr_init_qp_in_params_sq(dev, pd, qp, attrs, udata, &in_params);
+       qedr_init_qp_in_params_rq(qp, attrs, udata, &in_params);
+
+       qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
+                                             &in_params, &out_params);
+
+       if (!qp->qed_qp)
+               goto err1;
+
+       qp->qp_id = out_params.qp_id;
+       qp->icid = out_params.icid;
+       qp->ibqp.qp_num = qp->qp_id;
+
+       if (udata) {
+               rc = qedr_copy_qp_uresp(dev, qp, udata);
+               if (rc)
+                       goto err2;
+
+               qedr_qp_user_print(dev, qp);
+       } else {
+               qedr_init_qp_kernel_doorbell_sq(dev, qp);
+               qedr_init_qp_kernel_doorbell_rq(dev, qp);
+       }
+
+       DP_DEBUG(dev, QEDR_MSG_QP, "created %s space QP %p\n",
+                udata ? "user" : "kernel", qp);
+
+       return &qp->ibqp;
+
+err2:
+       rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
+       if (rc)
+               DP_ERR(dev, "create qp: fatal fault. rc=%d", rc);
+err1:
+       if (udata) {
+               qedr_cleanup_user_sq(dev, qp);
+               qedr_cleanup_user_rq(dev, qp);
+       } else {
+               qedr_cleanup_kernel_sq(dev, qp);
+               qedr_cleanup_kernel_rq(dev, qp);
+       }
+
+err0:
+       kfree(qp);
+
+       return ERR_PTR(-EFAULT);
+}
+
+enum ib_qp_state qedr_get_ibqp_state(enum qed_roce_qp_state qp_state)
+{
+       switch (qp_state) {
+       case QED_ROCE_QP_STATE_RESET:
+               return IB_QPS_RESET;
+       case QED_ROCE_QP_STATE_INIT:
+               return IB_QPS_INIT;
+       case QED_ROCE_QP_STATE_RTR:
+               return IB_QPS_RTR;
+       case QED_ROCE_QP_STATE_RTS:
+               return IB_QPS_RTS;
+       case QED_ROCE_QP_STATE_SQD:
+               return IB_QPS_SQD;
+       case QED_ROCE_QP_STATE_ERR:
+               return IB_QPS_ERR;
+       case QED_ROCE_QP_STATE_SQE:
+               return IB_QPS_SQE;
+       }
+       return IB_QPS_ERR;
+}
+
+enum qed_roce_qp_state qedr_get_state_from_ibqp(enum ib_qp_state qp_state)
+{
+       switch (qp_state) {
+       case IB_QPS_RESET:
+               return QED_ROCE_QP_STATE_RESET;
+       case IB_QPS_INIT:
+               return QED_ROCE_QP_STATE_INIT;
+       case IB_QPS_RTR:
+               return QED_ROCE_QP_STATE_RTR;
+       case IB_QPS_RTS:
+               return QED_ROCE_QP_STATE_RTS;
+       case IB_QPS_SQD:
+               return QED_ROCE_QP_STATE_SQD;
+       case IB_QPS_ERR:
+               return QED_ROCE_QP_STATE_ERR;
+       default:
+               return QED_ROCE_QP_STATE_ERR;
+       }
+}
+
+static void qedr_reset_qp_hwq_info(struct qedr_qp_hwq_info *qph)
+{
+       qed_chain_reset(&qph->pbl);
+       qph->prod = 0;
+       qph->cons = 0;
+       qph->wqe_cons = 0;
+       qph->db_data.data.value = cpu_to_le16(0);
+}
+
+static int qedr_update_qp_state(struct qedr_dev *dev,
+                               struct qedr_qp *qp,
+                               enum qed_roce_qp_state new_state)
+{
+       int status = 0;
+
+       if (new_state == qp->state)
+               return 1;
+
+       switch (qp->state) {
+       case QED_ROCE_QP_STATE_RESET:
+               switch (new_state) {
+               case QED_ROCE_QP_STATE_INIT:
+                       qp->prev_wqe_size = 0;
+                       qedr_reset_qp_hwq_info(&qp->sq);
+                       qedr_reset_qp_hwq_info(&qp->rq);
+                       break;
+               default:
+                       status = -EINVAL;
+                       break;
+               };
+               break;
+       case QED_ROCE_QP_STATE_INIT:
+               switch (new_state) {
+               case QED_ROCE_QP_STATE_RTR:
+                       /* Update doorbell (in case post_recv was
+                        * done before move to RTR)
+                        */
+                       wmb();
+                       writel(qp->rq.db_data.raw, qp->rq.db);
+                       /* Make sure write takes effect */
+                       mmiowb();
+                       break;
+               case QED_ROCE_QP_STATE_ERR:
+                       break;
+               default:
+                       /* Invalid state change. */
+                       status = -EINVAL;
+                       break;
+               };
+               break;
+       case QED_ROCE_QP_STATE_RTR:
+               /* RTR->XXX */
+               switch (new_state) {
+               case QED_ROCE_QP_STATE_RTS:
+                       break;
+               case QED_ROCE_QP_STATE_ERR:
+                       break;
+               default:
+                       /* Invalid state change. */
+                       status = -EINVAL;
+                       break;
+               };
+               break;
+       case QED_ROCE_QP_STATE_RTS:
+               /* RTS->XXX */
+               switch (new_state) {
+               case QED_ROCE_QP_STATE_SQD:
+                       break;
+               case QED_ROCE_QP_STATE_ERR:
+                       break;
+               default:
+                       /* Invalid state change. */
+                       status = -EINVAL;
+                       break;
+               };
+               break;
+       case QED_ROCE_QP_STATE_SQD:
+               /* SQD->XXX */
+               switch (new_state) {
+               case QED_ROCE_QP_STATE_RTS:
+               case QED_ROCE_QP_STATE_ERR:
+                       break;
+               default:
+                       /* Invalid state change. */
+                       status = -EINVAL;
+                       break;
+               };
+               break;
+       case QED_ROCE_QP_STATE_ERR:
+               /* ERR->XXX */
+               switch (new_state) {
+               case QED_ROCE_QP_STATE_RESET:
+                       break;
+               default:
+                       status = -EINVAL;
+                       break;
+               };
+               break;
+       default:
+               status = -EINVAL;
+               break;
+       };
+
+       return status;
+}
+
+int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
+                  int attr_mask, struct ib_udata *udata)
+{
+       struct qedr_qp *qp = get_qedr_qp(ibqp);
+       struct qed_rdma_modify_qp_in_params qp_params = { 0 };
+       struct qedr_dev *dev = get_qedr_dev(&qp->dev->ibdev);
+       enum ib_qp_state old_qp_state, new_qp_state;
+       int rc = 0;
+
+       DP_DEBUG(dev, QEDR_MSG_QP,
+                "modify qp: qp %p attr_mask=0x%x, state=%d", qp, attr_mask,
+                attr->qp_state);
+
+       old_qp_state = qedr_get_ibqp_state(qp->state);
+       if (attr_mask & IB_QP_STATE)
+               new_qp_state = attr->qp_state;
+       else
+               new_qp_state = old_qp_state;
+
+       if (!ib_modify_qp_is_ok
+           (old_qp_state, new_qp_state, ibqp->qp_type, attr_mask,
+            IB_LINK_LAYER_ETHERNET)) {
+               DP_ERR(dev,
+                      "modify qp: invalid attribute mask=0x%x specified for\n"
+                      "qpn=0x%x of type=0x%x old_qp_state=0x%x, new_qp_state=0x%x\n",
+                      attr_mask, qp->qp_id, ibqp->qp_type, old_qp_state,
+                      new_qp_state);
+               rc = -EINVAL;
+               goto err;
+       }
+
+       /* Translate the masks... */
+       if (attr_mask & IB_QP_STATE) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_RDMA_MODIFY_QP_VALID_NEW_STATE, 1);
+               qp_params.new_state = qedr_get_state_from_ibqp(attr->qp_state);
+       }
+
+       if (attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
+               qp_params.sqd_async = true;
+
+       if (attr_mask & IB_QP_PKEY_INDEX) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_PKEY, 1);
+               if (attr->pkey_index >= QEDR_ROCE_PKEY_TABLE_LEN) {
+                       rc = -EINVAL;
+                       goto err;
+               }
+
+               qp_params.pkey = QEDR_ROCE_PKEY_DEFAULT;
+       }
+
+       if (attr_mask & IB_QP_QKEY)
+               qp->qkey = attr->qkey;
+
+       if (attr_mask & IB_QP_ACCESS_FLAGS) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN, 1);
+               qp_params.incoming_rdma_read_en = attr->qp_access_flags &
+                                                 IB_ACCESS_REMOTE_READ;
+               qp_params.incoming_rdma_write_en = attr->qp_access_flags &
+                                                  IB_ACCESS_REMOTE_WRITE;
+               qp_params.incoming_atomic_en = attr->qp_access_flags &
+                                              IB_ACCESS_REMOTE_ATOMIC;
+       }
+
+       if (attr_mask & (IB_QP_AV | IB_QP_PATH_MTU)) {
+               if (attr_mask & IB_QP_PATH_MTU) {
+                       if (attr->path_mtu < IB_MTU_256 ||
+                           attr->path_mtu > IB_MTU_4096) {
+                               pr_err("error: Only MTU sizes of 256, 512, 1024, 2048 and 4096 are supported by RoCE\n");
+                               rc = -EINVAL;
+                               goto err;
+                       }
+                       qp->mtu = min(ib_mtu_enum_to_int(attr->path_mtu),
+                                     ib_mtu_enum_to_int(iboe_get_mtu
+                                                        (dev->ndev->mtu)));
+               }
+
+               if (!qp->mtu) {
+                       qp->mtu =
+                       ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
+                       pr_err("Fixing zeroed MTU to qp->mtu = %d\n", qp->mtu);
+               }
+
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR, 1);
+
+               qp_params.traffic_class_tos = attr->ah_attr.grh.traffic_class;
+               qp_params.flow_label = attr->ah_attr.grh.flow_label;
+               qp_params.hop_limit_ttl = attr->ah_attr.grh.hop_limit;
+
+               qp->sgid_idx = attr->ah_attr.grh.sgid_index;
+
+               rc = get_gid_info_from_table(ibqp, attr, attr_mask, &qp_params);
+               if (rc) {
+                       DP_ERR(dev,
+                              "modify qp: problems with GID index %d (rc=%d)\n",
+                              attr->ah_attr.grh.sgid_index, rc);
+                       return rc;
+               }
+
+               rc = qedr_get_dmac(dev, &attr->ah_attr,
+                                  qp_params.remote_mac_addr);
+               if (rc)
+                       return rc;
+
+               qp_params.use_local_mac = true;
+               ether_addr_copy(qp_params.local_mac_addr, dev->ndev->dev_addr);
+
+               DP_DEBUG(dev, QEDR_MSG_QP, "dgid=%x:%x:%x:%x\n",
+                        qp_params.dgid.dwords[0], qp_params.dgid.dwords[1],
+                        qp_params.dgid.dwords[2], qp_params.dgid.dwords[3]);
+               DP_DEBUG(dev, QEDR_MSG_QP, "sgid=%x:%x:%x:%x\n",
+                        qp_params.sgid.dwords[0], qp_params.sgid.dwords[1],
+                        qp_params.sgid.dwords[2], qp_params.sgid.dwords[3]);
+               DP_DEBUG(dev, QEDR_MSG_QP, "remote_mac=[%pM]\n",
+                        qp_params.remote_mac_addr);
+;
+
+               qp_params.mtu = qp->mtu;
+               qp_params.lb_indication = false;
+       }
+
+       if (!qp_params.mtu) {
+               /* Stay with current MTU */
+               if (qp->mtu)
+                       qp_params.mtu = qp->mtu;
+               else
+                       qp_params.mtu =
+                           ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
+       }
+
+       if (attr_mask & IB_QP_TIMEOUT) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);
+
+               qp_params.ack_timeout = attr->timeout;
+               if (attr->timeout) {
+                       u32 temp;
+
+                       temp = 4096 * (1UL << attr->timeout) / 1000 / 1000;
+                       /* FW requires [msec] */
+                       qp_params.ack_timeout = temp;
+               } else {
+                       /* Infinite */
+                       qp_params.ack_timeout = 0;
+               }
+       }
+       if (attr_mask & IB_QP_RETRY_CNT) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);
+               qp_params.retry_cnt = attr->retry_cnt;
+       }
+
+       if (attr_mask & IB_QP_RNR_RETRY) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT, 1);
+               qp_params.rnr_retry_cnt = attr->rnr_retry;
+       }
+
+       if (attr_mask & IB_QP_RQ_PSN) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_RQ_PSN, 1);
+               qp_params.rq_psn = attr->rq_psn;
+               qp->rq_psn = attr->rq_psn;
+       }
+
+       if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
+               if (attr->max_rd_atomic > dev->attr.max_qp_req_rd_atomic_resc) {
+                       rc = -EINVAL;
+                       DP_ERR(dev,
+                              "unsupported max_rd_atomic=%d, supported=%d\n",
+                              attr->max_rd_atomic,
+                              dev->attr.max_qp_req_rd_atomic_resc);
+                       goto err;
+               }
+
+               SET_FIELD(qp_params.modify_flags,
+                         QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ, 1);
+               qp_params.max_rd_atomic_req = attr->max_rd_atomic;
+       }
+
+       if (attr_mask & IB_QP_MIN_RNR_TIMER) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER, 1);
+               qp_params.min_rnr_nak_timer = attr->min_rnr_timer;
+       }
+
+       if (attr_mask & IB_QP_SQ_PSN) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_SQ_PSN, 1);
+               qp_params.sq_psn = attr->sq_psn;
+               qp->sq_psn = attr->sq_psn;
+       }
+
+       if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
+               if (attr->max_dest_rd_atomic >
+                   dev->attr.max_qp_resp_rd_atomic_resc) {
+                       DP_ERR(dev,
+                              "unsupported max_dest_rd_atomic=%d, supported=%d\n",
+                              attr->max_dest_rd_atomic,
+                              dev->attr.max_qp_resp_rd_atomic_resc);
+
+                       rc = -EINVAL;
+                       goto err;
+               }
+
+               SET_FIELD(qp_params.modify_flags,
+                         QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP, 1);
+               qp_params.max_rd_atomic_resp = attr->max_dest_rd_atomic;
+       }
+
+       if (attr_mask & IB_QP_DEST_QPN) {
+               SET_FIELD(qp_params.modify_flags,
+                         QED_ROCE_MODIFY_QP_VALID_DEST_QP, 1);
+
+               qp_params.dest_qp = attr->dest_qp_num;
+               qp->dest_qp_num = attr->dest_qp_num;
+       }
+
+       if (qp->qp_type != IB_QPT_GSI)
+               rc = dev->ops->rdma_modify_qp(dev->rdma_ctx,
+                                             qp->qed_qp, &qp_params);
+
+       if (attr_mask & IB_QP_STATE) {
+               if ((qp->qp_type != IB_QPT_GSI) && (!udata))
+                       qedr_update_qp_state(dev, qp, qp_params.new_state);
+               qp->state = qp_params.new_state;
+       }
+
+err:
+       return rc;
+}
+
+static int qedr_to_ib_qp_acc_flags(struct qed_rdma_query_qp_out_params *params)
+{
+       int ib_qp_acc_flags = 0;
+
+       if (params->incoming_rdma_write_en)
+               ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
+       if (params->incoming_rdma_read_en)
+               ib_qp_acc_flags |= IB_ACCESS_REMOTE_READ;
+       if (params->incoming_atomic_en)
+               ib_qp_acc_flags |= IB_ACCESS_REMOTE_ATOMIC;
+       ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
+       return ib_qp_acc_flags;
+}
+
+int qedr_query_qp(struct ib_qp *ibqp,
+                 struct ib_qp_attr *qp_attr,
+                 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
+{
+       struct qed_rdma_query_qp_out_params params;
+       struct qedr_qp *qp = get_qedr_qp(ibqp);
+       struct qedr_dev *dev = qp->dev;
+       int rc = 0;
+
+       memset(&params, 0, sizeof(params));
+
+       rc = dev->ops->rdma_query_qp(dev->rdma_ctx, qp->qed_qp, &params);
+       if (rc)
+               goto err;
+
+       memset(qp_attr, 0, sizeof(*qp_attr));
+       memset(qp_init_attr, 0, sizeof(*qp_init_attr));
+
+       qp_attr->qp_state = qedr_get_ibqp_state(params.state);
+       qp_attr->cur_qp_state = qedr_get_ibqp_state(params.state);
+       qp_attr->path_mtu = iboe_get_mtu(params.mtu);
+       qp_attr->path_mig_state = IB_MIG_MIGRATED;
+       qp_attr->rq_psn = params.rq_psn;
+       qp_attr->sq_psn = params.sq_psn;
+       qp_attr->dest_qp_num = params.dest_qp;
+
+       qp_attr->qp_access_flags = qedr_to_ib_qp_acc_flags(&params);
+
+       qp_attr->cap.max_send_wr = qp->sq.max_wr;
+       qp_attr->cap.max_recv_wr = qp->rq.max_wr;
+       qp_attr->cap.max_send_sge = qp->sq.max_sges;
+       qp_attr->cap.max_recv_sge = qp->rq.max_sges;
+       qp_attr->cap.max_inline_data = qp->max_inline_data;
+       qp_init_attr->cap = qp_attr->cap;
+
+       memcpy(&qp_attr->ah_attr.grh.dgid.raw[0], &params.dgid.bytes[0],
+              sizeof(qp_attr->ah_attr.grh.dgid.raw));
+
+       qp_attr->ah_attr.grh.flow_label = params.flow_label;
+       qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
+       qp_attr->ah_attr.grh.hop_limit = params.hop_limit_ttl;
+       qp_attr->ah_attr.grh.traffic_class = params.traffic_class_tos;
+
+       qp_attr->ah_attr.ah_flags = IB_AH_GRH;
+       qp_attr->ah_attr.port_num = 1;
+       qp_attr->ah_attr.sl = 0;
+       qp_attr->timeout = params.timeout;
+       qp_attr->rnr_retry = params.rnr_retry;
+       qp_attr->retry_cnt = params.retry_cnt;
+       qp_attr->min_rnr_timer = params.min_rnr_nak_timer;
+       qp_attr->pkey_index = params.pkey_index;
+       qp_attr->port_num = 1;
+       qp_attr->ah_attr.src_path_bits = 0;
+       qp_attr->ah_attr.static_rate = 0;
+       qp_attr->alt_pkey_index = 0;
+       qp_attr->alt_port_num = 0;
+       qp_attr->alt_timeout = 0;
+       memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
+
+       qp_attr->sq_draining = (params.state == QED_ROCE_QP_STATE_SQD) ? 1 : 0;
+       qp_attr->max_dest_rd_atomic = params.max_dest_rd_atomic;
+       qp_attr->max_rd_atomic = params.max_rd_atomic;
+       qp_attr->en_sqd_async_notify = (params.sqd_async) ? 1 : 0;
+
+       DP_DEBUG(dev, QEDR_MSG_QP, "QEDR_QUERY_QP: max_inline_data=%d\n",
+                qp_attr->cap.max_inline_data);
+
+err:
+       return rc;
+}
+
+int qedr_destroy_qp(struct ib_qp *ibqp)
+{
+       struct qedr_qp *qp = get_qedr_qp(ibqp);
+       struct qedr_dev *dev = qp->dev;
+       struct ib_qp_attr attr;
+       int attr_mask = 0;
+       int rc = 0;
+
+       DP_DEBUG(dev, QEDR_MSG_QP, "destroy qp: destroying %p, qp type=%d\n",
+                qp, qp->qp_type);
+
+       if (qp->state != (QED_ROCE_QP_STATE_RESET | QED_ROCE_QP_STATE_ERR |
+                         QED_ROCE_QP_STATE_INIT)) {
+               attr.qp_state = IB_QPS_ERR;
+               attr_mask |= IB_QP_STATE;
+
+               /* Change the QP state to ERROR */
+               qedr_modify_qp(ibqp, &attr, attr_mask, NULL);
+       }
+
+       if (qp->qp_type != IB_QPT_GSI) {
+               rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
+               if (rc)
+                       return rc;
+       }
+
+       if (ibqp->uobject && ibqp->uobject->context) {
+               qedr_cleanup_user_sq(dev, qp);
+               qedr_cleanup_user_rq(dev, qp);
+       } else {
+               qedr_cleanup_kernel_sq(dev, qp);
+               qedr_cleanup_kernel_rq(dev, qp);
+       }
+
+       kfree(qp);
+
+       return rc;
+}
index 36c8a692f740aea9d9482ae22ec8ff0767af0eaa..056d6cb31fa2cf491280aa30fc31f53139008513 100644 (file)
@@ -62,5 +62,12 @@ struct ib_cq *qedr_create_cq(struct ib_device *ibdev,
 int qedr_resize_cq(struct ib_cq *, int cqe, struct ib_udata *);
 int qedr_destroy_cq(struct ib_cq *);
 int qedr_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags);
+struct ib_qp *qedr_create_qp(struct ib_pd *, struct ib_qp_init_attr *attrs,
+                            struct ib_udata *);
+int qedr_modify_qp(struct ib_qp *, struct ib_qp_attr *attr,
+                  int attr_mask, struct ib_udata *udata);
+int qedr_query_qp(struct ib_qp *, struct ib_qp_attr *qp_attr,
+                 int qp_attr_mask, struct ib_qp_init_attr *);
+int qedr_destroy_qp(struct ib_qp *ibqp);
 
 #endif
index b0fc5f2125e02eff1c3fcf1b4eab0bd4811c1706..75c270d839c8d5097757baf8644a0426db4fb16f 100644 (file)
@@ -69,4 +69,38 @@ struct qedr_create_cq_uresp {
        __u16 icid;
 };
 
+struct qedr_create_qp_ureq {
+       __u32 qp_handle_hi;
+       __u32 qp_handle_lo;
+
+       /* SQ */
+       /* user space virtual address of SQ buffer */
+       __u64 sq_addr;
+
+       /* length of SQ buffer */
+       __u64 sq_len;
+
+       /* RQ */
+       /* user space virtual address of RQ buffer */
+       __u64 rq_addr;
+
+       /* length of RQ buffer */
+       __u64 rq_len;
+};
+
+struct qedr_create_qp_uresp {
+       __u32 qp_id;
+       __u32 atomic_supported;
+
+       /* SQ */
+       __u32 sq_db_offset;
+       __u16 sq_icid;
+
+       /* RQ */
+       __u32 rq_db_offset;
+       __u16 rq_icid;
+
+       __u32 rq_db2_offset;
+};
+
 #endif /* __QEDR_USER_H__ */