qed: Duplicate qed_roce.[ch] to qed_rdma.[ch]
authorKalderon, Michal <Michal.Kalderon@cavium.com>
Wed, 21 Jun 2017 13:22:44 +0000 (16:22 +0300)
committerDavid S. Miller <davem@davemloft.net>
Wed, 21 Jun 2017 19:32:59 +0000 (15:32 -0400)
This patch adds files that will contain common code for RoCE/iWARP.
The files are currently identical to qed_roce.c / qed_roce.h and
intentionally not added to the makefile. The next patch in the series
will modify the files so that roce specific code is left in qed_roce
and common roce/iwarp code will be placed in qed_rdma

This patch is the result of a simple
cp qed_rdma.c qed_roce.c
cp qed_rdma.h qed_roce.h

Signed-off-by: Michal Kalderon <Michal.Kalderon@cavium.com>
Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/qlogic/qed/qed_rdma.c [new file with mode: 0644]
drivers/net/ethernet/qlogic/qed/qed_rdma.h [new file with mode: 0644]

diff --git a/drivers/net/ethernet/qlogic/qed/qed_rdma.c b/drivers/net/ethernet/qlogic/qed/qed_rdma.c
new file mode 100644 (file)
index 0000000..9dc41a8
--- /dev/null
@@ -0,0 +1,2805 @@
+/* QLogic qed NIC Driver
+ * Copyright (c) 2015-2017  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.
+ */
+#include <linux/types.h>
+#include <asm/byteorder.h>
+#include <linux/bitops.h>
+#include <linux/delay.h>
+#include <linux/dma-mapping.h>
+#include <linux/errno.h>
+#include <linux/io.h>
+#include <linux/kernel.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/string.h>
+#include "qed.h"
+#include "qed_cxt.h"
+#include "qed_hsi.h"
+#include "qed_hw.h"
+#include "qed_init_ops.h"
+#include "qed_int.h"
+#include "qed_ll2.h"
+#include "qed_mcp.h"
+#include "qed_reg_addr.h"
+#include "qed_roce.h"
+#include <linux/qed/qed_roce_if.h>
+#include "qed_sp.h"
+
+static void qed_roce_free_real_icid(struct qed_hwfn *p_hwfn, u16 icid);
+
+static int
+qed_roce_async_event(struct qed_hwfn *p_hwfn,
+                    u8 fw_event_code,
+                    u16 echo, union event_ring_data *data, u8 fw_return_code)
+{
+       if (fw_event_code == ROCE_ASYNC_EVENT_DESTROY_QP_DONE) {
+               u16 icid =
+                   (u16)le32_to_cpu(data->rdma_data.rdma_destroy_qp_data.cid);
+
+               /* icid release in this async event can occur only if the icid
+                * was offloaded to the FW. In case it wasn't offloaded this is
+                * handled in qed_roce_sp_destroy_qp.
+                */
+               qed_roce_free_real_icid(p_hwfn, icid);
+       } else {
+               struct qed_rdma_events *events = &p_hwfn->p_rdma_info->events;
+
+               events->affiliated_event(p_hwfn->p_rdma_info->events.context,
+                                        fw_event_code,
+                                    (void *)&data->rdma_data.async_handle);
+       }
+
+       return 0;
+}
+
+static int qed_rdma_bmap_alloc(struct qed_hwfn *p_hwfn,
+                              struct qed_bmap *bmap, u32 max_count, char *name)
+{
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "max_count = %08x\n", max_count);
+
+       bmap->max_count = max_count;
+
+       bmap->bitmap = kcalloc(BITS_TO_LONGS(max_count), sizeof(long),
+                              GFP_KERNEL);
+       if (!bmap->bitmap)
+               return -ENOMEM;
+
+       snprintf(bmap->name, QED_RDMA_MAX_BMAP_NAME, "%s", name);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "0\n");
+       return 0;
+}
+
+static int qed_rdma_bmap_alloc_id(struct qed_hwfn *p_hwfn,
+                                 struct qed_bmap *bmap, u32 *id_num)
+{
+       *id_num = find_first_zero_bit(bmap->bitmap, bmap->max_count);
+       if (*id_num >= bmap->max_count)
+               return -EINVAL;
+
+       __set_bit(*id_num, bmap->bitmap);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "%s bitmap: allocated id %d\n",
+                  bmap->name, *id_num);
+
+       return 0;
+}
+
+static void qed_bmap_set_id(struct qed_hwfn *p_hwfn,
+                           struct qed_bmap *bmap, u32 id_num)
+{
+       if (id_num >= bmap->max_count)
+               return;
+
+       __set_bit(id_num, bmap->bitmap);
+}
+
+static void qed_bmap_release_id(struct qed_hwfn *p_hwfn,
+                               struct qed_bmap *bmap, u32 id_num)
+{
+       bool b_acquired;
+
+       if (id_num >= bmap->max_count)
+               return;
+
+       b_acquired = test_and_clear_bit(id_num, bmap->bitmap);
+       if (!b_acquired) {
+               DP_NOTICE(p_hwfn, "%s bitmap: id %d already released\n",
+                         bmap->name, id_num);
+               return;
+       }
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "%s bitmap: released id %d\n",
+                  bmap->name, id_num);
+}
+
+static int qed_bmap_test_id(struct qed_hwfn *p_hwfn,
+                           struct qed_bmap *bmap, u32 id_num)
+{
+       if (id_num >= bmap->max_count)
+               return -1;
+
+       return test_bit(id_num, bmap->bitmap);
+}
+
+static bool qed_bmap_is_empty(struct qed_bmap *bmap)
+{
+       return bmap->max_count == find_first_bit(bmap->bitmap, bmap->max_count);
+}
+
+static u32 qed_rdma_get_sb_id(void *p_hwfn, u32 rel_sb_id)
+{
+       /* First sb id for RoCE is after all the l2 sb */
+       return FEAT_NUM((struct qed_hwfn *)p_hwfn, QED_PF_L2_QUE) + rel_sb_id;
+}
+
+static int qed_rdma_alloc(struct qed_hwfn *p_hwfn,
+                         struct qed_ptt *p_ptt,
+                         struct qed_rdma_start_in_params *params)
+{
+       struct qed_rdma_info *p_rdma_info;
+       u32 num_cons, num_tasks;
+       int rc = -ENOMEM;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Allocating RDMA\n");
+
+       /* Allocate a struct with current pf rdma info */
+       p_rdma_info = kzalloc(sizeof(*p_rdma_info), GFP_KERNEL);
+       if (!p_rdma_info)
+               return rc;
+
+       p_hwfn->p_rdma_info = p_rdma_info;
+       p_rdma_info->proto = PROTOCOLID_ROCE;
+
+       num_cons = qed_cxt_get_proto_cid_count(p_hwfn, p_rdma_info->proto,
+                                              NULL);
+
+       p_rdma_info->num_qps = num_cons / 2;
+
+       num_tasks = qed_cxt_get_proto_tid_count(p_hwfn, PROTOCOLID_ROCE);
+
+       /* Each MR uses a single task */
+       p_rdma_info->num_mrs = num_tasks;
+
+       /* Queue zone lines are shared between RoCE and L2 in such a way that
+        * they can be used by each without obstructing the other.
+        */
+       p_rdma_info->queue_zone_base = (u16)RESC_START(p_hwfn, QED_L2_QUEUE);
+       p_rdma_info->max_queue_zones = (u16)RESC_NUM(p_hwfn, QED_L2_QUEUE);
+
+       /* Allocate a struct with device params and fill it */
+       p_rdma_info->dev = kzalloc(sizeof(*p_rdma_info->dev), GFP_KERNEL);
+       if (!p_rdma_info->dev)
+               goto free_rdma_info;
+
+       /* Allocate a struct with port params and fill it */
+       p_rdma_info->port = kzalloc(sizeof(*p_rdma_info->port), GFP_KERNEL);
+       if (!p_rdma_info->port)
+               goto free_rdma_dev;
+
+       /* Allocate bit map for pd's */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->pd_map, RDMA_MAX_PDS,
+                                "PD");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate pd_map, rc = %d\n",
+                          rc);
+               goto free_rdma_port;
+       }
+
+       /* Allocate DPI bitmap */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->dpi_map,
+                                p_hwfn->dpi_count, "DPI");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate DPI bitmap, rc = %d\n", rc);
+               goto free_pd_map;
+       }
+
+       /* Allocate bitmap for cq's. The maximum number of CQs is bounded to
+        * twice the number of QPs.
+        */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->cq_map,
+                                p_rdma_info->num_qps * 2, "CQ");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate cq bitmap, rc = %d\n", rc);
+               goto free_dpi_map;
+       }
+
+       /* Allocate bitmap for toggle bit for cq icids
+        * We toggle the bit every time we create or resize cq for a given icid.
+        * The maximum number of CQs is bounded to  twice the number of QPs.
+        */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->toggle_bits,
+                                p_rdma_info->num_qps * 2, "Toggle");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate toogle bits, rc = %d\n", rc);
+               goto free_cq_map;
+       }
+
+       /* Allocate bitmap for itids */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->tid_map,
+                                p_rdma_info->num_mrs, "MR");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate itids bitmaps, rc = %d\n", rc);
+               goto free_toggle_map;
+       }
+
+       /* Allocate bitmap for cids used for qps. */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->cid_map, num_cons,
+                                "CID");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate cid bitmap, rc = %d\n", rc);
+               goto free_tid_map;
+       }
+
+       /* Allocate bitmap for cids used for responders/requesters. */
+       rc = qed_rdma_bmap_alloc(p_hwfn, &p_rdma_info->real_cid_map, num_cons,
+                                "REAL_CID");
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                          "Failed to allocate real cid bitmap, rc = %d\n", rc);
+               goto free_cid_map;
+       }
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Allocation successful\n");
+       return 0;
+
+free_cid_map:
+       kfree(p_rdma_info->cid_map.bitmap);
+free_tid_map:
+       kfree(p_rdma_info->tid_map.bitmap);
+free_toggle_map:
+       kfree(p_rdma_info->toggle_bits.bitmap);
+free_cq_map:
+       kfree(p_rdma_info->cq_map.bitmap);
+free_dpi_map:
+       kfree(p_rdma_info->dpi_map.bitmap);
+free_pd_map:
+       kfree(p_rdma_info->pd_map.bitmap);
+free_rdma_port:
+       kfree(p_rdma_info->port);
+free_rdma_dev:
+       kfree(p_rdma_info->dev);
+free_rdma_info:
+       kfree(p_rdma_info);
+
+       return rc;
+}
+
+static void qed_rdma_bmap_free(struct qed_hwfn *p_hwfn,
+                              struct qed_bmap *bmap, bool check)
+{
+       int weight = bitmap_weight(bmap->bitmap, bmap->max_count);
+       int last_line = bmap->max_count / (64 * 8);
+       int last_item = last_line * 8 +
+           DIV_ROUND_UP(bmap->max_count % (64 * 8), 64);
+       u64 *pmap = (u64 *)bmap->bitmap;
+       int line, item, offset;
+       u8 str_last_line[200] = { 0 };
+
+       if (!weight || !check)
+               goto end;
+
+       DP_NOTICE(p_hwfn,
+                 "%s bitmap not free - size=%d, weight=%d, 512 bits per line\n",
+                 bmap->name, bmap->max_count, weight);
+
+       /* print aligned non-zero lines, if any */
+       for (item = 0, line = 0; line < last_line; line++, item += 8)
+               if (bitmap_weight((unsigned long *)&pmap[item], 64 * 8))
+                       DP_NOTICE(p_hwfn,
+                                 "line 0x%04x: 0x%016llx 0x%016llx 0x%016llx 0x%016llx 0x%016llx 0x%016llx 0x%016llx 0x%016llx\n",
+                                 line,
+                                 pmap[item],
+                                 pmap[item + 1],
+                                 pmap[item + 2],
+                                 pmap[item + 3],
+                                 pmap[item + 4],
+                                 pmap[item + 5],
+                                 pmap[item + 6], pmap[item + 7]);
+
+       /* print last unaligned non-zero line, if any */
+       if ((bmap->max_count % (64 * 8)) &&
+           (bitmap_weight((unsigned long *)&pmap[item],
+                          bmap->max_count - item * 64))) {
+               offset = sprintf(str_last_line, "line 0x%04x: ", line);
+               for (; item < last_item; item++)
+                       offset += sprintf(str_last_line + offset,
+                                         "0x%016llx ", pmap[item]);
+               DP_NOTICE(p_hwfn, "%s\n", str_last_line);
+       }
+
+end:
+       kfree(bmap->bitmap);
+       bmap->bitmap = NULL;
+}
+
+static void qed_rdma_resc_free(struct qed_hwfn *p_hwfn)
+{
+       struct qed_rdma_info *p_rdma_info = p_hwfn->p_rdma_info;
+
+       qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->cid_map, 1);
+       qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->pd_map, 1);
+       qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->dpi_map, 1);
+       qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->cq_map, 1);
+       qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->toggle_bits, 0);
+       qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->tid_map, 1);
+
+       kfree(p_rdma_info->port);
+       kfree(p_rdma_info->dev);
+
+       kfree(p_rdma_info);
+}
+
+static void qed_rdma_free(struct qed_hwfn *p_hwfn)
+{
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Freeing RDMA\n");
+
+       qed_rdma_resc_free(p_hwfn);
+}
+
+static void qed_rdma_get_guid(struct qed_hwfn *p_hwfn, u8 *guid)
+{
+       guid[0] = p_hwfn->hw_info.hw_mac_addr[0] ^ 2;
+       guid[1] = p_hwfn->hw_info.hw_mac_addr[1];
+       guid[2] = p_hwfn->hw_info.hw_mac_addr[2];
+       guid[3] = 0xff;
+       guid[4] = 0xfe;
+       guid[5] = p_hwfn->hw_info.hw_mac_addr[3];
+       guid[6] = p_hwfn->hw_info.hw_mac_addr[4];
+       guid[7] = p_hwfn->hw_info.hw_mac_addr[5];
+}
+
+static void qed_rdma_init_events(struct qed_hwfn *p_hwfn,
+                                struct qed_rdma_start_in_params *params)
+{
+       struct qed_rdma_events *events;
+
+       events = &p_hwfn->p_rdma_info->events;
+
+       events->unaffiliated_event = params->events->unaffiliated_event;
+       events->affiliated_event = params->events->affiliated_event;
+       events->context = params->events->context;
+}
+
+static void qed_rdma_init_devinfo(struct qed_hwfn *p_hwfn,
+                                 struct qed_rdma_start_in_params *params)
+{
+       struct qed_rdma_device *dev = p_hwfn->p_rdma_info->dev;
+       struct qed_dev *cdev = p_hwfn->cdev;
+       u32 pci_status_control;
+       u32 num_qps;
+
+       /* Vendor specific information */
+       dev->vendor_id = cdev->vendor_id;
+       dev->vendor_part_id = cdev->device_id;
+       dev->hw_ver = 0;
+       dev->fw_ver = (FW_MAJOR_VERSION << 24) | (FW_MINOR_VERSION << 16) |
+                     (FW_REVISION_VERSION << 8) | (FW_ENGINEERING_VERSION);
+
+       qed_rdma_get_guid(p_hwfn, (u8 *)&dev->sys_image_guid);
+       dev->node_guid = dev->sys_image_guid;
+
+       dev->max_sge = min_t(u32, RDMA_MAX_SGE_PER_SQ_WQE,
+                            RDMA_MAX_SGE_PER_RQ_WQE);
+
+       if (cdev->rdma_max_sge)
+               dev->max_sge = min_t(u32, cdev->rdma_max_sge, dev->max_sge);
+
+       dev->max_inline = ROCE_REQ_MAX_INLINE_DATA_SIZE;
+
+       dev->max_inline = (cdev->rdma_max_inline) ?
+                         min_t(u32, cdev->rdma_max_inline, dev->max_inline) :
+                         dev->max_inline;
+
+       dev->max_wqe = QED_RDMA_MAX_WQE;
+       dev->max_cnq = (u8)FEAT_NUM(p_hwfn, QED_RDMA_CNQ);
+
+       /* The number of QPs may be higher than QED_ROCE_MAX_QPS, because
+        * it is up-aligned to 16 and then to ILT page size within qed cxt.
+        * This is OK in terms of ILT but we don't want to configure the FW
+        * above its abilities
+        */
+       num_qps = ROCE_MAX_QPS;
+       num_qps = min_t(u64, num_qps, p_hwfn->p_rdma_info->num_qps);
+       dev->max_qp = num_qps;
+
+       /* CQs uses the same icids that QPs use hence they are limited by the
+        * number of icids. There are two icids per QP.
+        */
+       dev->max_cq = num_qps * 2;
+
+       /* The number of mrs is smaller by 1 since the first is reserved */
+       dev->max_mr = p_hwfn->p_rdma_info->num_mrs - 1;
+       dev->max_mr_size = QED_RDMA_MAX_MR_SIZE;
+
+       /* The maximum CQE capacity per CQ supported.
+        * max number of cqes will be in two layer pbl,
+        * 8 is the pointer size in bytes
+        * 32 is the size of cq element in bytes
+        */
+       if (params->cq_mode == QED_RDMA_CQ_MODE_32_BITS)
+               dev->max_cqe = QED_RDMA_MAX_CQE_32_BIT;
+       else
+               dev->max_cqe = QED_RDMA_MAX_CQE_16_BIT;
+
+       dev->max_mw = 0;
+       dev->max_fmr = QED_RDMA_MAX_FMR;
+       dev->max_mr_mw_fmr_pbl = (PAGE_SIZE / 8) * (PAGE_SIZE / 8);
+       dev->max_mr_mw_fmr_size = dev->max_mr_mw_fmr_pbl * PAGE_SIZE;
+       dev->max_pkey = QED_RDMA_MAX_P_KEY;
+
+       dev->max_qp_resp_rd_atomic_resc = RDMA_RING_PAGE_SIZE /
+                                         (RDMA_RESP_RD_ATOMIC_ELM_SIZE * 2);
+       dev->max_qp_req_rd_atomic_resc = RDMA_RING_PAGE_SIZE /
+                                        RDMA_REQ_RD_ATOMIC_ELM_SIZE;
+       dev->max_dev_resp_rd_atomic_resc = dev->max_qp_resp_rd_atomic_resc *
+                                          p_hwfn->p_rdma_info->num_qps;
+       dev->page_size_caps = QED_RDMA_PAGE_SIZE_CAPS;
+       dev->dev_ack_delay = QED_RDMA_ACK_DELAY;
+       dev->max_pd = RDMA_MAX_PDS;
+       dev->max_ah = p_hwfn->p_rdma_info->num_qps;
+       dev->max_stats_queues = (u8)RESC_NUM(p_hwfn, QED_RDMA_STATS_QUEUE);
+
+       /* Set capablities */
+       dev->dev_caps = 0;
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_RNR_NAK, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_PORT_ACTIVE_EVENT, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_PORT_CHANGE_EVENT, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_RESIZE_CQ, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_BASE_MEMORY_EXT, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_BASE_QUEUE_EXT, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_ZBVA, 1);
+       SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_LOCAL_INV_FENCE, 1);
+
+       /* Check atomic operations support in PCI configuration space. */
+       pci_read_config_dword(cdev->pdev,
+                             cdev->pdev->pcie_cap + PCI_EXP_DEVCTL2,
+                             &pci_status_control);
+
+       if (pci_status_control & PCI_EXP_DEVCTL2_LTR_EN)
+               SET_FIELD(dev->dev_caps, QED_RDMA_DEV_CAP_ATOMIC_OP, 1);
+}
+
+static void qed_rdma_init_port(struct qed_hwfn *p_hwfn)
+{
+       struct qed_rdma_port *port = p_hwfn->p_rdma_info->port;
+       struct qed_rdma_device *dev = p_hwfn->p_rdma_info->dev;
+
+       port->port_state = p_hwfn->mcp_info->link_output.link_up ?
+                          QED_RDMA_PORT_UP : QED_RDMA_PORT_DOWN;
+
+       port->max_msg_size = min_t(u64,
+                                  (dev->max_mr_mw_fmr_size *
+                                   p_hwfn->cdev->rdma_max_sge),
+                                  BIT(31));
+
+       port->pkey_bad_counter = 0;
+}
+
+static int qed_rdma_init_hw(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt)
+{
+       u32 ll2_ethertype_en;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Initializing HW\n");
+       p_hwfn->b_rdma_enabled_in_prs = false;
+
+       qed_wr(p_hwfn, p_ptt, PRS_REG_ROCE_DEST_QP_MAX_PF, 0);
+
+       p_hwfn->rdma_prs_search_reg = PRS_REG_SEARCH_ROCE;
+
+       /* We delay writing to this reg until first cid is allocated. See
+        * qed_cxt_dynamic_ilt_alloc function for more details
+        */
+       ll2_ethertype_en = qed_rd(p_hwfn, p_ptt, PRS_REG_LIGHT_L2_ETHERTYPE_EN);
+       qed_wr(p_hwfn, p_ptt, PRS_REG_LIGHT_L2_ETHERTYPE_EN,
+              (ll2_ethertype_en | 0x01));
+
+       if (qed_cxt_get_proto_cid_start(p_hwfn, PROTOCOLID_ROCE) % 2) {
+               DP_NOTICE(p_hwfn, "The first RoCE's cid should be even\n");
+               return -EINVAL;
+       }
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Initializing HW - Done\n");
+       return 0;
+}
+
+static int qed_rdma_start_fw(struct qed_hwfn *p_hwfn,
+                            struct qed_rdma_start_in_params *params,
+                            struct qed_ptt *p_ptt)
+{
+       struct rdma_init_func_ramrod_data *p_ramrod;
+       struct qed_rdma_cnq_params *p_cnq_pbl_list;
+       struct rdma_init_func_hdr *p_params_header;
+       struct rdma_cnq_params *p_cnq_params;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       u32 cnq_id, sb_id;
+       u16 igu_sb_id;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Starting FW\n");
+
+       /* Save the number of cnqs for the function close ramrod */
+       p_hwfn->p_rdma_info->num_cnqs = params->desired_cnq;
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent, RDMA_RAMROD_FUNC_INIT,
+                                p_hwfn->p_rdma_info->proto, &init_data);
+       if (rc)
+               return rc;
+
+       p_ramrod = &p_ent->ramrod.roce_init_func.rdma;
+
+       p_params_header = &p_ramrod->params_header;
+       p_params_header->cnq_start_offset = (u8)RESC_START(p_hwfn,
+                                                          QED_RDMA_CNQ_RAM);
+       p_params_header->num_cnqs = params->desired_cnq;
+
+       if (params->cq_mode == QED_RDMA_CQ_MODE_16_BITS)
+               p_params_header->cq_ring_mode = 1;
+       else
+               p_params_header->cq_ring_mode = 0;
+
+       for (cnq_id = 0; cnq_id < params->desired_cnq; cnq_id++) {
+               sb_id = qed_rdma_get_sb_id(p_hwfn, cnq_id);
+               igu_sb_id = qed_get_igu_sb_id(p_hwfn, sb_id);
+               p_ramrod->cnq_params[cnq_id].sb_num = cpu_to_le16(igu_sb_id);
+               p_cnq_params = &p_ramrod->cnq_params[cnq_id];
+               p_cnq_pbl_list = &params->cnq_pbl_list[cnq_id];
+
+               p_cnq_params->sb_index = p_hwfn->pf_params.rdma_pf_params.gl_pi;
+               p_cnq_params->num_pbl_pages = p_cnq_pbl_list->num_pbl_pages;
+
+               DMA_REGPAIR_LE(p_cnq_params->pbl_base_addr,
+                              p_cnq_pbl_list->pbl_ptr);
+
+               /* we assume here that cnq_id and qz_offset are the same */
+               p_cnq_params->queue_zone_num =
+                       cpu_to_le16(p_hwfn->p_rdma_info->queue_zone_base +
+                                   cnq_id);
+       }
+
+       return qed_spq_post(p_hwfn, p_ent, NULL);
+}
+
+static int qed_rdma_alloc_tid(void *rdma_cxt, u32 *itid)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Allocate TID\n");
+
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       rc = qed_rdma_bmap_alloc_id(p_hwfn,
+                                   &p_hwfn->p_rdma_info->tid_map, itid);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+       if (rc)
+               goto out;
+
+       rc = qed_cxt_dynamic_ilt_alloc(p_hwfn, QED_ELEM_TASK, *itid);
+out:
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Allocate TID - done, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_rdma_reserve_lkey(struct qed_hwfn *p_hwfn)
+{
+       struct qed_rdma_device *dev = p_hwfn->p_rdma_info->dev;
+
+       /* The first DPI is reserved for the Kernel */
+       __set_bit(0, p_hwfn->p_rdma_info->dpi_map.bitmap);
+
+       /* Tid 0 will be used as the key for "reserved MR".
+        * The driver should allocate memory for it so it can be loaded but no
+        * ramrod should be passed on it.
+        */
+       qed_rdma_alloc_tid(p_hwfn, &dev->reserved_lkey);
+       if (dev->reserved_lkey != RDMA_RESERVED_LKEY) {
+               DP_NOTICE(p_hwfn,
+                         "Reserved lkey should be equal to RDMA_RESERVED_LKEY\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int qed_rdma_setup(struct qed_hwfn *p_hwfn,
+                         struct qed_ptt *p_ptt,
+                         struct qed_rdma_start_in_params *params)
+{
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "RDMA setup\n");
+
+       spin_lock_init(&p_hwfn->p_rdma_info->lock);
+
+       qed_rdma_init_devinfo(p_hwfn, params);
+       qed_rdma_init_port(p_hwfn);
+       qed_rdma_init_events(p_hwfn, params);
+
+       rc = qed_rdma_reserve_lkey(p_hwfn);
+       if (rc)
+               return rc;
+
+       rc = qed_rdma_init_hw(p_hwfn, p_ptt);
+       if (rc)
+               return rc;
+
+       qed_spq_register_async_cb(p_hwfn, PROTOCOLID_ROCE,
+                                 qed_roce_async_event);
+
+       return qed_rdma_start_fw(p_hwfn, params, p_ptt);
+}
+
+void qed_roce_stop(struct qed_hwfn *p_hwfn)
+{
+       struct qed_bmap *rcid_map = &p_hwfn->p_rdma_info->real_cid_map;
+       int wait_count = 0;
+
+       /* when destroying a_RoCE QP the control is returned to the user after
+        * the synchronous part. The asynchronous part may take a little longer.
+        * We delay for a short while if an async destroy QP is still expected.
+        * Beyond the added delay we clear the bitmap anyway.
+        */
+       while (bitmap_weight(rcid_map->bitmap, rcid_map->max_count)) {
+               msleep(100);
+               if (wait_count++ > 20) {
+                       DP_NOTICE(p_hwfn, "cid bitmap wait timed out\n");
+                       break;
+               }
+       }
+       qed_spq_unregister_async_cb(p_hwfn, PROTOCOLID_ROCE);
+}
+
+static int qed_rdma_stop(void *rdma_cxt)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct rdma_close_func_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       struct qed_ptt *p_ptt;
+       u32 ll2_ethertype_en;
+       int rc = -EBUSY;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "RDMA stop\n");
+
+       p_ptt = qed_ptt_acquire(p_hwfn);
+       if (!p_ptt) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Failed to acquire PTT\n");
+               return rc;
+       }
+
+       /* Disable RoCE search */
+       qed_wr(p_hwfn, p_ptt, p_hwfn->rdma_prs_search_reg, 0);
+       p_hwfn->b_rdma_enabled_in_prs = false;
+
+       qed_wr(p_hwfn, p_ptt, PRS_REG_ROCE_DEST_QP_MAX_PF, 0);
+
+       ll2_ethertype_en = qed_rd(p_hwfn, p_ptt, PRS_REG_LIGHT_L2_ETHERTYPE_EN);
+
+       qed_wr(p_hwfn, p_ptt, PRS_REG_LIGHT_L2_ETHERTYPE_EN,
+              (ll2_ethertype_en & 0xFFFE));
+
+       qed_roce_stop(p_hwfn);
+       qed_ptt_release(p_hwfn, p_ptt);
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       /* Stop RoCE */
+       rc = qed_sp_init_request(p_hwfn, &p_ent, RDMA_RAMROD_FUNC_CLOSE,
+                                p_hwfn->p_rdma_info->proto, &init_data);
+       if (rc)
+               goto out;
+
+       p_ramrod = &p_ent->ramrod.rdma_close_func;
+
+       p_ramrod->num_cnqs = p_hwfn->p_rdma_info->num_cnqs;
+       p_ramrod->cnq_start_offset = (u8)RESC_START(p_hwfn, QED_RDMA_CNQ_RAM);
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+
+out:
+       qed_rdma_free(p_hwfn);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "RDMA stop done, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_rdma_add_user(void *rdma_cxt,
+                            struct qed_rdma_add_user_out_params *out_params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       u32 dpi_start_offset;
+       u32 returned_id = 0;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Adding User\n");
+
+       /* Allocate DPI */
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       rc = qed_rdma_bmap_alloc_id(p_hwfn, &p_hwfn->p_rdma_info->dpi_map,
+                                   &returned_id);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+
+       out_params->dpi = (u16)returned_id;
+
+       /* Calculate the corresponding DPI address */
+       dpi_start_offset = p_hwfn->dpi_start_offset;
+
+       out_params->dpi_addr = (u64)((u8 __iomem *)p_hwfn->doorbells +
+                                    dpi_start_offset +
+                                    ((out_params->dpi) * p_hwfn->dpi_size));
+
+       out_params->dpi_phys_addr = p_hwfn->cdev->db_phys_addr +
+                                   dpi_start_offset +
+                                   ((out_params->dpi) * p_hwfn->dpi_size);
+
+       out_params->dpi_size = p_hwfn->dpi_size;
+       out_params->wid_count = p_hwfn->wid_count;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Adding user - done, rc = %d\n", rc);
+       return rc;
+}
+
+static struct qed_rdma_port *qed_rdma_query_port(void *rdma_cxt)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct qed_rdma_port *p_port = p_hwfn->p_rdma_info->port;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "RDMA Query port\n");
+
+       /* Link may have changed */
+       p_port->port_state = p_hwfn->mcp_info->link_output.link_up ?
+                            QED_RDMA_PORT_UP : QED_RDMA_PORT_DOWN;
+
+       p_port->link_speed = p_hwfn->mcp_info->link_output.speed;
+
+       p_port->max_msg_size = RDMA_MAX_DATA_SIZE_IN_WQE;
+
+       return p_port;
+}
+
+static struct qed_rdma_device *qed_rdma_query_device(void *rdma_cxt)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Query device\n");
+
+       /* Return struct with device parameters */
+       return p_hwfn->p_rdma_info->dev;
+}
+
+static void qed_rdma_free_tid(void *rdma_cxt, u32 itid)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "itid = %08x\n", itid);
+
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       qed_bmap_release_id(p_hwfn, &p_hwfn->p_rdma_info->tid_map, itid);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
+static void qed_rdma_cnq_prod_update(void *rdma_cxt, u8 qz_offset, u16 prod)
+{
+       struct qed_hwfn *p_hwfn;
+       u16 qz_num;
+       u32 addr;
+
+       p_hwfn = (struct qed_hwfn *)rdma_cxt;
+
+       if (qz_offset > p_hwfn->p_rdma_info->max_queue_zones) {
+               DP_NOTICE(p_hwfn,
+                         "queue zone offset %d is too large (max is %d)\n",
+                         qz_offset, p_hwfn->p_rdma_info->max_queue_zones);
+               return;
+       }
+
+       qz_num = p_hwfn->p_rdma_info->queue_zone_base + qz_offset;
+       addr = GTT_BAR0_MAP_REG_USDM_RAM +
+              USTORM_COMMON_QUEUE_CONS_OFFSET(qz_num);
+
+       REG_WR16(p_hwfn, addr, prod);
+
+       /* keep prod updates ordered */
+       wmb();
+}
+
+static int qed_fill_rdma_dev_info(struct qed_dev *cdev,
+                                 struct qed_dev_rdma_info *info)
+{
+       struct qed_hwfn *p_hwfn = QED_LEADING_HWFN(cdev);
+
+       memset(info, 0, sizeof(*info));
+
+       info->rdma_type = QED_RDMA_TYPE_ROCE;
+       info->user_dpm_enabled = (p_hwfn->db_bar_no_edpm == 0);
+
+       qed_fill_dev_info(cdev, &info->common);
+
+       return 0;
+}
+
+static int qed_rdma_get_sb_start(struct qed_dev *cdev)
+{
+       int feat_num;
+
+       if (cdev->num_hwfns > 1)
+               feat_num = FEAT_NUM(QED_LEADING_HWFN(cdev), QED_PF_L2_QUE);
+       else
+               feat_num = FEAT_NUM(QED_LEADING_HWFN(cdev), QED_PF_L2_QUE) *
+                          cdev->num_hwfns;
+
+       return feat_num;
+}
+
+static int qed_rdma_get_min_cnq_msix(struct qed_dev *cdev)
+{
+       int n_cnq = FEAT_NUM(QED_LEADING_HWFN(cdev), QED_RDMA_CNQ);
+       int n_msix = cdev->int_params.rdma_msix_cnt;
+
+       return min_t(int, n_cnq, n_msix);
+}
+
+static int qed_rdma_set_int(struct qed_dev *cdev, u16 cnt)
+{
+       int limit = 0;
+
+       /* Mark the fastpath as free/used */
+       cdev->int_params.fp_initialized = cnt ? true : false;
+
+       if (cdev->int_params.out.int_mode != QED_INT_MODE_MSIX) {
+               DP_ERR(cdev,
+                      "qed roce supports only MSI-X interrupts (detected %d).\n",
+                      cdev->int_params.out.int_mode);
+               return -EINVAL;
+       } else if (cdev->int_params.fp_msix_cnt) {
+               limit = cdev->int_params.rdma_msix_cnt;
+       }
+
+       if (!limit)
+               return -ENOMEM;
+
+       return min_t(int, cnt, limit);
+}
+
+static int qed_rdma_get_int(struct qed_dev *cdev, struct qed_int_info *info)
+{
+       memset(info, 0, sizeof(*info));
+
+       if (!cdev->int_params.fp_initialized) {
+               DP_INFO(cdev,
+                       "Protocol driver requested interrupt information, but its support is not yet configured\n");
+               return -EINVAL;
+       }
+
+       if (cdev->int_params.out.int_mode == QED_INT_MODE_MSIX) {
+               int msix_base = cdev->int_params.rdma_msix_base;
+
+               info->msix_cnt = cdev->int_params.rdma_msix_cnt;
+               info->msix = &cdev->int_params.msix_table[msix_base];
+
+               DP_VERBOSE(cdev, QED_MSG_RDMA, "msix_cnt = %d msix_base=%d\n",
+                          info->msix_cnt, msix_base);
+       }
+
+       return 0;
+}
+
+static int qed_rdma_alloc_pd(void *rdma_cxt, u16 *pd)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       u32 returned_id;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Alloc PD\n");
+
+       /* Allocates an unused protection domain */
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       rc = qed_rdma_bmap_alloc_id(p_hwfn,
+                                   &p_hwfn->p_rdma_info->pd_map, &returned_id);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+
+       *pd = (u16)returned_id;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Alloc PD - done, rc = %d\n", rc);
+       return rc;
+}
+
+static void qed_rdma_free_pd(void *rdma_cxt, u16 pd)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "pd = %08x\n", pd);
+
+       /* Returns a previously allocated protection domain for reuse */
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       qed_bmap_release_id(p_hwfn, &p_hwfn->p_rdma_info->pd_map, pd);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
+static enum qed_rdma_toggle_bit
+qed_rdma_toggle_bit_create_resize_cq(struct qed_hwfn *p_hwfn, u16 icid)
+{
+       struct qed_rdma_info *p_info = p_hwfn->p_rdma_info;
+       enum qed_rdma_toggle_bit toggle_bit;
+       u32 bmap_id;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", icid);
+
+       /* the function toggle the bit that is related to a given icid
+        * and returns the new toggle bit's value
+        */
+       bmap_id = icid - qed_cxt_get_proto_cid_start(p_hwfn, p_info->proto);
+
+       spin_lock_bh(&p_info->lock);
+       toggle_bit = !test_and_change_bit(bmap_id,
+                                         p_info->toggle_bits.bitmap);
+       spin_unlock_bh(&p_info->lock);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "QED_RDMA_TOGGLE_BIT_= %d\n",
+                  toggle_bit);
+
+       return toggle_bit;
+}
+
+static int qed_rdma_create_cq(void *rdma_cxt,
+                             struct qed_rdma_create_cq_in_params *params,
+                             u16 *icid)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct qed_rdma_info *p_info = p_hwfn->p_rdma_info;
+       struct rdma_create_cq_ramrod_data *p_ramrod;
+       enum qed_rdma_toggle_bit toggle_bit;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       u32 returned_id, start_cid;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "cq_handle = %08x%08x\n",
+                  params->cq_handle_hi, params->cq_handle_lo);
+
+       /* Allocate icid */
+       spin_lock_bh(&p_info->lock);
+       rc = qed_rdma_bmap_alloc_id(p_hwfn, &p_info->cq_map, &returned_id);
+       spin_unlock_bh(&p_info->lock);
+
+       if (rc) {
+               DP_NOTICE(p_hwfn, "Can't create CQ, rc = %d\n", rc);
+               return rc;
+       }
+
+       start_cid = qed_cxt_get_proto_cid_start(p_hwfn,
+                                               p_info->proto);
+       *icid = returned_id + start_cid;
+
+       /* Check if icid requires a page allocation */
+       rc = qed_cxt_dynamic_ilt_alloc(p_hwfn, QED_ELEM_CXT, *icid);
+       if (rc)
+               goto err;
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = *icid;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       /* Send create CQ ramrod */
+       rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                RDMA_RAMROD_CREATE_CQ,
+                                p_info->proto, &init_data);
+       if (rc)
+               goto err;
+
+       p_ramrod = &p_ent->ramrod.rdma_create_cq;
+
+       p_ramrod->cq_handle.hi = cpu_to_le32(params->cq_handle_hi);
+       p_ramrod->cq_handle.lo = cpu_to_le32(params->cq_handle_lo);
+       p_ramrod->dpi = cpu_to_le16(params->dpi);
+       p_ramrod->is_two_level_pbl = params->pbl_two_level;
+       p_ramrod->max_cqes = cpu_to_le32(params->cq_size);
+       DMA_REGPAIR_LE(p_ramrod->pbl_addr, params->pbl_ptr);
+       p_ramrod->pbl_num_pages = cpu_to_le16(params->pbl_num_pages);
+       p_ramrod->cnq_id = (u8)RESC_START(p_hwfn, QED_RDMA_CNQ_RAM) +
+                          params->cnq_id;
+       p_ramrod->int_timeout = params->int_timeout;
+
+       /* toggle the bit for every resize or create cq for a given icid */
+       toggle_bit = qed_rdma_toggle_bit_create_resize_cq(p_hwfn, *icid);
+
+       p_ramrod->toggle_bit = toggle_bit;
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+       if (rc) {
+               /* restore toggle bit */
+               qed_rdma_toggle_bit_create_resize_cq(p_hwfn, *icid);
+               goto err;
+       }
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Created CQ, rc = %d\n", rc);
+       return rc;
+
+err:
+       /* release allocated icid */
+       spin_lock_bh(&p_info->lock);
+       qed_bmap_release_id(p_hwfn, &p_info->cq_map, returned_id);
+       spin_unlock_bh(&p_info->lock);
+       DP_NOTICE(p_hwfn, "Create CQ failed, rc = %d\n", rc);
+
+       return rc;
+}
+
+static int
+qed_rdma_destroy_cq(void *rdma_cxt,
+                   struct qed_rdma_destroy_cq_in_params *in_params,
+                   struct qed_rdma_destroy_cq_out_params *out_params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct rdma_destroy_cq_output_params *p_ramrod_res;
+       struct rdma_destroy_cq_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       dma_addr_t ramrod_res_phys;
+       enum protocol_type proto;
+       int rc = -ENOMEM;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", in_params->icid);
+
+       p_ramrod_res =
+           (struct rdma_destroy_cq_output_params *)
+           dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
+                              sizeof(struct rdma_destroy_cq_output_params),
+                              &ramrod_res_phys, GFP_KERNEL);
+       if (!p_ramrod_res) {
+               DP_NOTICE(p_hwfn,
+                         "qed destroy cq failed: cannot allocate memory (ramrod)\n");
+               return rc;
+       }
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = in_params->icid;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+       proto = p_hwfn->p_rdma_info->proto;
+       /* Send destroy CQ ramrod */
+       rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                RDMA_RAMROD_DESTROY_CQ,
+                                proto, &init_data);
+       if (rc)
+               goto err;
+
+       p_ramrod = &p_ent->ramrod.rdma_destroy_cq;
+       DMA_REGPAIR_LE(p_ramrod->output_params_addr, ramrod_res_phys);
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+       if (rc)
+               goto err;
+
+       out_params->num_cq_notif = le16_to_cpu(p_ramrod_res->cnq_num);
+
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         sizeof(struct rdma_destroy_cq_output_params),
+                         p_ramrod_res, ramrod_res_phys);
+
+       /* Free icid */
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+
+       qed_bmap_release_id(p_hwfn,
+                           &p_hwfn->p_rdma_info->cq_map,
+                           (in_params->icid -
+                            qed_cxt_get_proto_cid_start(p_hwfn, proto)));
+
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Destroyed CQ, rc = %d\n", rc);
+       return rc;
+
+err:   dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         sizeof(struct rdma_destroy_cq_output_params),
+                         p_ramrod_res, ramrod_res_phys);
+
+       return rc;
+}
+
+static void qed_rdma_set_fw_mac(u16 *p_fw_mac, u8 *p_qed_mac)
+{
+       p_fw_mac[0] = cpu_to_le16((p_qed_mac[0] << 8) + p_qed_mac[1]);
+       p_fw_mac[1] = cpu_to_le16((p_qed_mac[2] << 8) + p_qed_mac[3]);
+       p_fw_mac[2] = cpu_to_le16((p_qed_mac[4] << 8) + p_qed_mac[5]);
+}
+
+static void qed_rdma_copy_gids(struct qed_rdma_qp *qp, __le32 *src_gid,
+                              __le32 *dst_gid)
+{
+       u32 i;
+
+       if (qp->roce_mode == ROCE_V2_IPV4) {
+               /* The IPv4 addresses shall be aligned to the highest word.
+                * The lower words must be zero.
+                */
+               memset(src_gid, 0, sizeof(union qed_gid));
+               memset(dst_gid, 0, sizeof(union qed_gid));
+               src_gid[3] = cpu_to_le32(qp->sgid.ipv4_addr);
+               dst_gid[3] = cpu_to_le32(qp->dgid.ipv4_addr);
+       } else {
+               /* GIDs and IPv6 addresses coincide in location and size */
+               for (i = 0; i < ARRAY_SIZE(qp->sgid.dwords); i++) {
+                       src_gid[i] = cpu_to_le32(qp->sgid.dwords[i]);
+                       dst_gid[i] = cpu_to_le32(qp->dgid.dwords[i]);
+               }
+       }
+}
+
+static enum roce_flavor qed_roce_mode_to_flavor(enum roce_mode roce_mode)
+{
+       enum roce_flavor flavor;
+
+       switch (roce_mode) {
+       case ROCE_V1:
+               flavor = PLAIN_ROCE;
+               break;
+       case ROCE_V2_IPV4:
+               flavor = RROCE_IPV4;
+               break;
+       case ROCE_V2_IPV6:
+               flavor = ROCE_V2_IPV6;
+               break;
+       default:
+               flavor = MAX_ROCE_MODE;
+               break;
+       }
+       return flavor;
+}
+
+void qed_roce_free_cid_pair(struct qed_hwfn *p_hwfn, u16 cid)
+{
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       qed_bmap_release_id(p_hwfn, &p_hwfn->p_rdma_info->cid_map, cid);
+       qed_bmap_release_id(p_hwfn, &p_hwfn->p_rdma_info->cid_map, cid + 1);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
+static int qed_roce_alloc_cid(struct qed_hwfn *p_hwfn, u16 *cid)
+{
+       struct qed_rdma_info *p_rdma_info = p_hwfn->p_rdma_info;
+       u32 responder_icid;
+       u32 requester_icid;
+       int rc;
+
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       rc = qed_rdma_bmap_alloc_id(p_hwfn, &p_rdma_info->cid_map,
+                                   &responder_icid);
+       if (rc) {
+               spin_unlock_bh(&p_rdma_info->lock);
+               return rc;
+       }
+
+       rc = qed_rdma_bmap_alloc_id(p_hwfn, &p_rdma_info->cid_map,
+                                   &requester_icid);
+
+       spin_unlock_bh(&p_rdma_info->lock);
+       if (rc)
+               goto err;
+
+       /* the two icid's should be adjacent */
+       if ((requester_icid - responder_icid) != 1) {
+               DP_NOTICE(p_hwfn, "Failed to allocate two adjacent qp's'\n");
+               rc = -EINVAL;
+               goto err;
+       }
+
+       responder_icid += qed_cxt_get_proto_cid_start(p_hwfn,
+                                                     p_rdma_info->proto);
+       requester_icid += qed_cxt_get_proto_cid_start(p_hwfn,
+                                                     p_rdma_info->proto);
+
+       /* If these icids require a new ILT line allocate DMA-able context for
+        * an ILT page
+        */
+       rc = qed_cxt_dynamic_ilt_alloc(p_hwfn, QED_ELEM_CXT, responder_icid);
+       if (rc)
+               goto err;
+
+       rc = qed_cxt_dynamic_ilt_alloc(p_hwfn, QED_ELEM_CXT, requester_icid);
+       if (rc)
+               goto err;
+
+       *cid = (u16)responder_icid;
+       return rc;
+
+err:
+       spin_lock_bh(&p_rdma_info->lock);
+       qed_bmap_release_id(p_hwfn, &p_rdma_info->cid_map, responder_icid);
+       qed_bmap_release_id(p_hwfn, &p_rdma_info->cid_map, requester_icid);
+
+       spin_unlock_bh(&p_rdma_info->lock);
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                  "Allocate CID - failed, rc = %d\n", rc);
+       return rc;
+}
+
+static void qed_roce_set_real_cid(struct qed_hwfn *p_hwfn, u32 cid)
+{
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       qed_bmap_set_id(p_hwfn, &p_hwfn->p_rdma_info->real_cid_map, cid);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
+static int qed_roce_sp_create_responder(struct qed_hwfn *p_hwfn,
+                                       struct qed_rdma_qp *qp)
+{
+       struct roce_create_qp_resp_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       enum roce_flavor roce_flavor;
+       struct qed_spq_entry *p_ent;
+       u16 regular_latency_queue;
+       enum protocol_type proto;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       /* Allocate DMA-able memory for IRQ */
+       qp->irq_num_pages = 1;
+       qp->irq = dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
+                                    RDMA_RING_PAGE_SIZE,
+                                    &qp->irq_phys_addr, GFP_KERNEL);
+       if (!qp->irq) {
+               rc = -ENOMEM;
+               DP_NOTICE(p_hwfn,
+                         "qed create responder failed: cannot allocate memory (irq). rc = %d\n",
+                         rc);
+               return rc;
+       }
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent, ROCE_RAMROD_CREATE_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc)
+               goto err;
+
+       p_ramrod = &p_ent->ramrod.roce_create_qp_resp;
+
+       p_ramrod->flags = 0;
+
+       roce_flavor = qed_roce_mode_to_flavor(qp->roce_mode);
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_ROCE_FLAVOR, roce_flavor);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_RDMA_RD_EN,
+                 qp->incoming_rdma_read_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_RDMA_WR_EN,
+                 qp->incoming_rdma_write_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_ATOMIC_EN,
+                 qp->incoming_atomic_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_E2E_FLOW_CONTROL_EN,
+                 qp->e2e_flow_control_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_SRQ_FLG, qp->use_srq);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_RESERVED_KEY_EN,
+                 qp->fmr_and_reserved_lkey);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_MIN_RNR_NAK_TIMER,
+                 qp->min_rnr_nak_timer);
+
+       p_ramrod->max_ird = qp->max_rd_atomic_resp;
+       p_ramrod->traffic_class = qp->traffic_class_tos;
+       p_ramrod->hop_limit = qp->hop_limit_ttl;
+       p_ramrod->irq_num_pages = qp->irq_num_pages;
+       p_ramrod->p_key = cpu_to_le16(qp->pkey);
+       p_ramrod->flow_label = cpu_to_le32(qp->flow_label);
+       p_ramrod->dst_qp_id = cpu_to_le32(qp->dest_qp);
+       p_ramrod->mtu = cpu_to_le16(qp->mtu);
+       p_ramrod->initial_psn = cpu_to_le32(qp->rq_psn);
+       p_ramrod->pd = cpu_to_le16(qp->pd);
+       p_ramrod->rq_num_pages = cpu_to_le16(qp->rq_num_pages);
+       DMA_REGPAIR_LE(p_ramrod->rq_pbl_addr, qp->rq_pbl_ptr);
+       DMA_REGPAIR_LE(p_ramrod->irq_pbl_addr, qp->irq_phys_addr);
+       qed_rdma_copy_gids(qp, p_ramrod->src_gid, p_ramrod->dst_gid);
+       p_ramrod->qp_handle_for_async.hi = cpu_to_le32(qp->qp_handle_async.hi);
+       p_ramrod->qp_handle_for_async.lo = cpu_to_le32(qp->qp_handle_async.lo);
+       p_ramrod->qp_handle_for_cqe.hi = cpu_to_le32(qp->qp_handle.hi);
+       p_ramrod->qp_handle_for_cqe.lo = cpu_to_le32(qp->qp_handle.lo);
+       p_ramrod->cq_cid = cpu_to_le32((p_hwfn->hw_info.opaque_fid << 16) |
+                                      qp->rq_cq_id);
+
+       regular_latency_queue = qed_get_cm_pq_idx(p_hwfn, PQ_FLAGS_OFLD);
+
+       p_ramrod->regular_latency_phy_queue =
+           cpu_to_le16(regular_latency_queue);
+       p_ramrod->low_latency_phy_queue =
+           cpu_to_le16(regular_latency_queue);
+
+       p_ramrod->dpi = cpu_to_le16(qp->dpi);
+
+       qed_rdma_set_fw_mac(p_ramrod->remote_mac_addr, qp->remote_mac_addr);
+       qed_rdma_set_fw_mac(p_ramrod->local_mac_addr, qp->local_mac_addr);
+
+       p_ramrod->udp_src_port = qp->udp_src_port;
+       p_ramrod->vlan_id = cpu_to_le16(qp->vlan_id);
+       p_ramrod->srq_id.srq_idx = cpu_to_le16(qp->srq_id);
+       p_ramrod->srq_id.opaque_fid = cpu_to_le16(p_hwfn->hw_info.opaque_fid);
+
+       p_ramrod->stats_counter_id = RESC_START(p_hwfn, QED_RDMA_STATS_QUEUE) +
+                                    qp->stats_queue;
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                  "rc = %d regular physical queue = 0x%x\n", rc,
+                  regular_latency_queue);
+
+       if (rc)
+               goto err;
+
+       qp->resp_offloaded = true;
+       qp->cq_prod = 0;
+
+       proto = p_hwfn->p_rdma_info->proto;
+       qed_roce_set_real_cid(p_hwfn, qp->icid -
+                             qed_cxt_get_proto_cid_start(p_hwfn, proto));
+
+       return rc;
+
+err:
+       DP_NOTICE(p_hwfn, "create responder - failed, rc = %d\n", rc);
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         qp->irq_num_pages * RDMA_RING_PAGE_SIZE,
+                         qp->irq, qp->irq_phys_addr);
+
+       return rc;
+}
+
+static int qed_roce_sp_create_requester(struct qed_hwfn *p_hwfn,
+                                       struct qed_rdma_qp *qp)
+{
+       struct roce_create_qp_req_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       enum roce_flavor roce_flavor;
+       struct qed_spq_entry *p_ent;
+       u16 regular_latency_queue;
+       enum protocol_type proto;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       /* Allocate DMA-able memory for ORQ */
+       qp->orq_num_pages = 1;
+       qp->orq = dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
+                                    RDMA_RING_PAGE_SIZE,
+                                    &qp->orq_phys_addr, GFP_KERNEL);
+       if (!qp->orq) {
+               rc = -ENOMEM;
+               DP_NOTICE(p_hwfn,
+                         "qed create requester failed: cannot allocate memory (orq). rc = %d\n",
+                         rc);
+               return rc;
+       }
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid + 1;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                ROCE_RAMROD_CREATE_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc)
+               goto err;
+
+       p_ramrod = &p_ent->ramrod.roce_create_qp_req;
+
+       p_ramrod->flags = 0;
+
+       roce_flavor = qed_roce_mode_to_flavor(qp->roce_mode);
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_REQ_RAMROD_DATA_ROCE_FLAVOR, roce_flavor);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_REQ_RAMROD_DATA_FMR_AND_RESERVED_EN,
+                 qp->fmr_and_reserved_lkey);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_REQ_RAMROD_DATA_SIGNALED_COMP, qp->signal_all);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_REQ_RAMROD_DATA_ERR_RETRY_CNT, qp->retry_cnt);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_REQ_RAMROD_DATA_RNR_NAK_CNT,
+                 qp->rnr_retry_cnt);
+
+       p_ramrod->max_ord = qp->max_rd_atomic_req;
+       p_ramrod->traffic_class = qp->traffic_class_tos;
+       p_ramrod->hop_limit = qp->hop_limit_ttl;
+       p_ramrod->orq_num_pages = qp->orq_num_pages;
+       p_ramrod->p_key = cpu_to_le16(qp->pkey);
+       p_ramrod->flow_label = cpu_to_le32(qp->flow_label);
+       p_ramrod->dst_qp_id = cpu_to_le32(qp->dest_qp);
+       p_ramrod->ack_timeout_val = cpu_to_le32(qp->ack_timeout);
+       p_ramrod->mtu = cpu_to_le16(qp->mtu);
+       p_ramrod->initial_psn = cpu_to_le32(qp->sq_psn);
+       p_ramrod->pd = cpu_to_le16(qp->pd);
+       p_ramrod->sq_num_pages = cpu_to_le16(qp->sq_num_pages);
+       DMA_REGPAIR_LE(p_ramrod->sq_pbl_addr, qp->sq_pbl_ptr);
+       DMA_REGPAIR_LE(p_ramrod->orq_pbl_addr, qp->orq_phys_addr);
+       qed_rdma_copy_gids(qp, p_ramrod->src_gid, p_ramrod->dst_gid);
+       p_ramrod->qp_handle_for_async.hi = cpu_to_le32(qp->qp_handle_async.hi);
+       p_ramrod->qp_handle_for_async.lo = cpu_to_le32(qp->qp_handle_async.lo);
+       p_ramrod->qp_handle_for_cqe.hi = cpu_to_le32(qp->qp_handle.hi);
+       p_ramrod->qp_handle_for_cqe.lo = cpu_to_le32(qp->qp_handle.lo);
+       p_ramrod->cq_cid =
+           cpu_to_le32((p_hwfn->hw_info.opaque_fid << 16) | qp->sq_cq_id);
+
+       regular_latency_queue = qed_get_cm_pq_idx(p_hwfn, PQ_FLAGS_OFLD);
+
+       p_ramrod->regular_latency_phy_queue =
+           cpu_to_le16(regular_latency_queue);
+       p_ramrod->low_latency_phy_queue =
+           cpu_to_le16(regular_latency_queue);
+
+       p_ramrod->dpi = cpu_to_le16(qp->dpi);
+
+       qed_rdma_set_fw_mac(p_ramrod->remote_mac_addr, qp->remote_mac_addr);
+       qed_rdma_set_fw_mac(p_ramrod->local_mac_addr, qp->local_mac_addr);
+
+       p_ramrod->udp_src_port = qp->udp_src_port;
+       p_ramrod->vlan_id = cpu_to_le16(qp->vlan_id);
+       p_ramrod->stats_counter_id = RESC_START(p_hwfn, QED_RDMA_STATS_QUEUE) +
+                                    qp->stats_queue;
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+
+       if (rc)
+               goto err;
+
+       qp->req_offloaded = true;
+       proto = p_hwfn->p_rdma_info->proto;
+       qed_roce_set_real_cid(p_hwfn,
+                             qp->icid + 1 -
+                             qed_cxt_get_proto_cid_start(p_hwfn, proto));
+
+       return rc;
+
+err:
+       DP_NOTICE(p_hwfn, "Create requested - failed, rc = %d\n", rc);
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         qp->orq_num_pages * RDMA_RING_PAGE_SIZE,
+                         qp->orq, qp->orq_phys_addr);
+       return rc;
+}
+
+static int qed_roce_sp_modify_responder(struct qed_hwfn *p_hwfn,
+                                       struct qed_rdma_qp *qp,
+                                       bool move_to_err, u32 modify_flags)
+{
+       struct roce_modify_qp_resp_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       if (move_to_err && !qp->resp_offloaded)
+               return 0;
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                ROCE_EVENT_MODIFY_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc) {
+               DP_NOTICE(p_hwfn, "rc = %d\n", rc);
+               return rc;
+       }
+
+       p_ramrod = &p_ent->ramrod.roce_modify_qp_resp;
+
+       p_ramrod->flags = 0;
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_MOVE_TO_ERR_FLG, move_to_err);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_RDMA_RD_EN,
+                 qp->incoming_rdma_read_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_RDMA_WR_EN,
+                 qp->incoming_rdma_write_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_ATOMIC_EN,
+                 qp->incoming_atomic_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_CREATE_QP_RESP_RAMROD_DATA_E2E_FLOW_CONTROL_EN,
+                 qp->e2e_flow_control_en);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_RDMA_OPS_EN_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_P_KEY_FLG,
+                 GET_FIELD(modify_flags, QED_ROCE_MODIFY_QP_VALID_PKEY));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_ADDRESS_VECTOR_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_MAX_IRD_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_MIN_RNR_NAK_TIMER_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER));
+
+       p_ramrod->fields = 0;
+       SET_FIELD(p_ramrod->fields,
+                 ROCE_MODIFY_QP_RESP_RAMROD_DATA_MIN_RNR_NAK_TIMER,
+                 qp->min_rnr_nak_timer);
+
+       p_ramrod->max_ird = qp->max_rd_atomic_resp;
+       p_ramrod->traffic_class = qp->traffic_class_tos;
+       p_ramrod->hop_limit = qp->hop_limit_ttl;
+       p_ramrod->p_key = cpu_to_le16(qp->pkey);
+       p_ramrod->flow_label = cpu_to_le32(qp->flow_label);
+       p_ramrod->mtu = cpu_to_le16(qp->mtu);
+       qed_rdma_copy_gids(qp, p_ramrod->src_gid, p_ramrod->dst_gid);
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Modify responder, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_roce_sp_modify_requester(struct qed_hwfn *p_hwfn,
+                                       struct qed_rdma_qp *qp,
+                                       bool move_to_sqd,
+                                       bool move_to_err, u32 modify_flags)
+{
+       struct roce_modify_qp_req_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       if (move_to_err && !(qp->req_offloaded))
+               return 0;
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid + 1;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                ROCE_EVENT_MODIFY_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc) {
+               DP_NOTICE(p_hwfn, "rc = %d\n", rc);
+               return rc;
+       }
+
+       p_ramrod = &p_ent->ramrod.roce_modify_qp_req;
+
+       p_ramrod->flags = 0;
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_MOVE_TO_ERR_FLG, move_to_err);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_MOVE_TO_SQD_FLG, move_to_sqd);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_EN_SQD_ASYNC_NOTIFY,
+                 qp->sqd_async);
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_P_KEY_FLG,
+                 GET_FIELD(modify_flags, QED_ROCE_MODIFY_QP_VALID_PKEY));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_ADDRESS_VECTOR_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_MAX_ORD_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_RNR_NAK_CNT_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_ERR_RETRY_CNT_FLG,
+                 GET_FIELD(modify_flags, QED_ROCE_MODIFY_QP_VALID_RETRY_CNT));
+
+       SET_FIELD(p_ramrod->flags,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_ACK_TIMEOUT_FLG,
+                 GET_FIELD(modify_flags,
+                           QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT));
+
+       p_ramrod->fields = 0;
+       SET_FIELD(p_ramrod->fields,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_ERR_RETRY_CNT, qp->retry_cnt);
+
+       SET_FIELD(p_ramrod->fields,
+                 ROCE_MODIFY_QP_REQ_RAMROD_DATA_RNR_NAK_CNT,
+                 qp->rnr_retry_cnt);
+
+       p_ramrod->max_ord = qp->max_rd_atomic_req;
+       p_ramrod->traffic_class = qp->traffic_class_tos;
+       p_ramrod->hop_limit = qp->hop_limit_ttl;
+       p_ramrod->p_key = cpu_to_le16(qp->pkey);
+       p_ramrod->flow_label = cpu_to_le32(qp->flow_label);
+       p_ramrod->ack_timeout_val = cpu_to_le32(qp->ack_timeout);
+       p_ramrod->mtu = cpu_to_le16(qp->mtu);
+       qed_rdma_copy_gids(qp, p_ramrod->src_gid, p_ramrod->dst_gid);
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Modify requester, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_roce_sp_destroy_qp_responder(struct qed_hwfn *p_hwfn,
+                                           struct qed_rdma_qp *qp,
+                                           u32 *num_invalidated_mw,
+                                           u32 *cq_prod)
+{
+       struct roce_destroy_qp_resp_output_params *p_ramrod_res;
+       struct roce_destroy_qp_resp_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       dma_addr_t ramrod_res_phys;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       *num_invalidated_mw = 0;
+       *cq_prod = qp->cq_prod;
+
+       if (!qp->resp_offloaded) {
+               /* If a responder was never offload, we need to free the cids
+                * allocated in create_qp as a FW async event will never arrive
+                */
+               u32 cid;
+
+               cid = qp->icid -
+                     qed_cxt_get_proto_cid_start(p_hwfn,
+                                                 p_hwfn->p_rdma_info->proto);
+               qed_roce_free_cid_pair(p_hwfn, (u16)cid);
+
+               return 0;
+       }
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                ROCE_RAMROD_DESTROY_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc)
+               return rc;
+
+       p_ramrod = &p_ent->ramrod.roce_destroy_qp_resp;
+
+       p_ramrod_res = (struct roce_destroy_qp_resp_output_params *)
+           dma_alloc_coherent(&p_hwfn->cdev->pdev->dev, sizeof(*p_ramrod_res),
+                              &ramrod_res_phys, GFP_KERNEL);
+
+       if (!p_ramrod_res) {
+               rc = -ENOMEM;
+               DP_NOTICE(p_hwfn,
+                         "qed destroy responder failed: cannot allocate memory (ramrod). rc = %d\n",
+                         rc);
+               return rc;
+       }
+
+       DMA_REGPAIR_LE(p_ramrod->output_params_addr, ramrod_res_phys);
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+       if (rc)
+               goto err;
+
+       *num_invalidated_mw = le32_to_cpu(p_ramrod_res->num_invalidated_mw);
+       *cq_prod = le32_to_cpu(p_ramrod_res->cq_prod);
+       qp->cq_prod = *cq_prod;
+
+       /* Free IRQ - only if ramrod succeeded, in case FW is still using it */
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         qp->irq_num_pages * RDMA_RING_PAGE_SIZE,
+                         qp->irq, qp->irq_phys_addr);
+
+       qp->resp_offloaded = false;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Destroy responder, rc = %d\n", rc);
+
+err:
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         sizeof(struct roce_destroy_qp_resp_output_params),
+                         p_ramrod_res, ramrod_res_phys);
+
+       return rc;
+}
+
+static int qed_roce_sp_destroy_qp_requester(struct qed_hwfn *p_hwfn,
+                                           struct qed_rdma_qp *qp,
+                                           u32 *num_bound_mw)
+{
+       struct roce_destroy_qp_req_output_params *p_ramrod_res;
+       struct roce_destroy_qp_req_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       dma_addr_t ramrod_res_phys;
+       int rc = -ENOMEM;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       if (!qp->req_offloaded)
+               return 0;
+
+       p_ramrod_res = (struct roce_destroy_qp_req_output_params *)
+                      dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
+                                         sizeof(*p_ramrod_res),
+                                         &ramrod_res_phys, GFP_KERNEL);
+       if (!p_ramrod_res) {
+               DP_NOTICE(p_hwfn,
+                         "qed destroy requester failed: cannot allocate memory (ramrod)\n");
+               return rc;
+       }
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid + 1;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent, ROCE_RAMROD_DESTROY_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc)
+               goto err;
+
+       p_ramrod = &p_ent->ramrod.roce_destroy_qp_req;
+       DMA_REGPAIR_LE(p_ramrod->output_params_addr, ramrod_res_phys);
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+       if (rc)
+               goto err;
+
+       *num_bound_mw = le32_to_cpu(p_ramrod_res->num_bound_mw);
+
+       /* Free ORQ - only if ramrod succeeded, in case FW is still using it */
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev,
+                         qp->orq_num_pages * RDMA_RING_PAGE_SIZE,
+                         qp->orq, qp->orq_phys_addr);
+
+       qp->req_offloaded = false;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Destroy requester, rc = %d\n", rc);
+
+err:
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev, sizeof(*p_ramrod_res),
+                         p_ramrod_res, ramrod_res_phys);
+
+       return rc;
+}
+
+static int qed_roce_query_qp(struct qed_hwfn *p_hwfn,
+                            struct qed_rdma_qp *qp,
+                            struct qed_rdma_query_qp_out_params *out_params)
+{
+       struct roce_query_qp_resp_output_params *p_resp_ramrod_res;
+       struct roce_query_qp_req_output_params *p_req_ramrod_res;
+       struct roce_query_qp_resp_ramrod_data *p_resp_ramrod;
+       struct roce_query_qp_req_ramrod_data *p_req_ramrod;
+       struct qed_sp_init_data init_data;
+       dma_addr_t resp_ramrod_res_phys;
+       dma_addr_t req_ramrod_res_phys;
+       struct qed_spq_entry *p_ent;
+       bool rq_err_state;
+       bool sq_err_state;
+       bool sq_draining;
+       int rc = -ENOMEM;
+
+       if ((!(qp->resp_offloaded)) && (!(qp->req_offloaded))) {
+               /* We can't send ramrod to the fw since this qp wasn't offloaded
+                * to the fw yet
+                */
+               out_params->draining = false;
+               out_params->rq_psn = qp->rq_psn;
+               out_params->sq_psn = qp->sq_psn;
+               out_params->state = qp->cur_state;
+
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "No QPs as no offload\n");
+               return 0;
+       }
+
+       if (!(qp->resp_offloaded)) {
+               DP_NOTICE(p_hwfn,
+                         "The responder's qp should be offloded before requester's\n");
+               return -EINVAL;
+       }
+
+       /* Send a query responder ramrod to FW to get RQ-PSN and state */
+       p_resp_ramrod_res = (struct roce_query_qp_resp_output_params *)
+           dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
+                              sizeof(*p_resp_ramrod_res),
+                              &resp_ramrod_res_phys, GFP_KERNEL);
+       if (!p_resp_ramrod_res) {
+               DP_NOTICE(p_hwfn,
+                         "qed query qp failed: cannot allocate memory (ramrod)\n");
+               return rc;
+       }
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.cid = qp->icid;
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+       rc = qed_sp_init_request(p_hwfn, &p_ent, ROCE_RAMROD_QUERY_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc)
+               goto err_resp;
+
+       p_resp_ramrod = &p_ent->ramrod.roce_query_qp_resp;
+       DMA_REGPAIR_LE(p_resp_ramrod->output_params_addr, resp_ramrod_res_phys);
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+       if (rc)
+               goto err_resp;
+
+       out_params->rq_psn = le32_to_cpu(p_resp_ramrod_res->psn);
+       rq_err_state = GET_FIELD(le32_to_cpu(p_resp_ramrod_res->err_flag),
+                                ROCE_QUERY_QP_RESP_OUTPUT_PARAMS_ERROR_FLG);
+
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev, sizeof(*p_resp_ramrod_res),
+                         p_resp_ramrod_res, resp_ramrod_res_phys);
+
+       if (!(qp->req_offloaded)) {
+               /* Don't send query qp for the requester */
+               out_params->sq_psn = qp->sq_psn;
+               out_params->draining = false;
+
+               if (rq_err_state)
+                       qp->cur_state = QED_ROCE_QP_STATE_ERR;
+
+               out_params->state = qp->cur_state;
+
+               return 0;
+       }
+
+       /* Send a query requester ramrod to FW to get SQ-PSN and state */
+       p_req_ramrod_res = (struct roce_query_qp_req_output_params *)
+                          dma_alloc_coherent(&p_hwfn->cdev->pdev->dev,
+                                             sizeof(*p_req_ramrod_res),
+                                             &req_ramrod_res_phys,
+                                             GFP_KERNEL);
+       if (!p_req_ramrod_res) {
+               rc = -ENOMEM;
+               DP_NOTICE(p_hwfn,
+                         "qed query qp failed: cannot allocate memory (ramrod)\n");
+               return rc;
+       }
+
+       /* Get SPQ entry */
+       init_data.cid = qp->icid + 1;
+       rc = qed_sp_init_request(p_hwfn, &p_ent, ROCE_RAMROD_QUERY_QP,
+                                PROTOCOLID_ROCE, &init_data);
+       if (rc)
+               goto err_req;
+
+       p_req_ramrod = &p_ent->ramrod.roce_query_qp_req;
+       DMA_REGPAIR_LE(p_req_ramrod->output_params_addr, req_ramrod_res_phys);
+
+       rc = qed_spq_post(p_hwfn, p_ent, NULL);
+       if (rc)
+               goto err_req;
+
+       out_params->sq_psn = le32_to_cpu(p_req_ramrod_res->psn);
+       sq_err_state = GET_FIELD(le32_to_cpu(p_req_ramrod_res->flags),
+                                ROCE_QUERY_QP_REQ_OUTPUT_PARAMS_ERR_FLG);
+       sq_draining =
+               GET_FIELD(le32_to_cpu(p_req_ramrod_res->flags),
+                         ROCE_QUERY_QP_REQ_OUTPUT_PARAMS_SQ_DRAINING_FLG);
+
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev, sizeof(*p_req_ramrod_res),
+                         p_req_ramrod_res, req_ramrod_res_phys);
+
+       out_params->draining = false;
+
+       if (rq_err_state || sq_err_state)
+               qp->cur_state = QED_ROCE_QP_STATE_ERR;
+       else if (sq_draining)
+               out_params->draining = true;
+       out_params->state = qp->cur_state;
+
+       return 0;
+
+err_req:
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev, sizeof(*p_req_ramrod_res),
+                         p_req_ramrod_res, req_ramrod_res_phys);
+       return rc;
+err_resp:
+       dma_free_coherent(&p_hwfn->cdev->pdev->dev, sizeof(*p_resp_ramrod_res),
+                         p_resp_ramrod_res, resp_ramrod_res_phys);
+       return rc;
+}
+
+static int qed_roce_destroy_qp(struct qed_hwfn *p_hwfn, struct qed_rdma_qp *qp)
+{
+       u32 num_invalidated_mw = 0;
+       u32 num_bound_mw = 0;
+       u32 cq_prod;
+       int rc;
+
+       /* Destroys the specified QP */
+       if ((qp->cur_state != QED_ROCE_QP_STATE_RESET) &&
+           (qp->cur_state != QED_ROCE_QP_STATE_ERR) &&
+           (qp->cur_state != QED_ROCE_QP_STATE_INIT)) {
+               DP_NOTICE(p_hwfn,
+                         "QP must be in error, reset or init state before destroying it\n");
+               return -EINVAL;
+       }
+
+       if (qp->cur_state != QED_ROCE_QP_STATE_RESET) {
+               rc = qed_roce_sp_destroy_qp_responder(p_hwfn, qp,
+                                                     &num_invalidated_mw,
+                                                     &cq_prod);
+               if (rc)
+                       return rc;
+
+               /* Send destroy requester ramrod */
+               rc = qed_roce_sp_destroy_qp_requester(p_hwfn, qp,
+                                                     &num_bound_mw);
+               if (rc)
+                       return rc;
+
+               if (num_invalidated_mw != num_bound_mw) {
+                       DP_NOTICE(p_hwfn,
+                                 "number of invalidate memory windows is different from bounded ones\n");
+                       return -EINVAL;
+               }
+       }
+
+       return 0;
+}
+
+static int qed_rdma_query_qp(void *rdma_cxt,
+                            struct qed_rdma_qp *qp,
+                            struct qed_rdma_query_qp_out_params *out_params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       /* The following fields are filled in from qp and not FW as they can't
+        * be modified by FW
+        */
+       out_params->mtu = qp->mtu;
+       out_params->dest_qp = qp->dest_qp;
+       out_params->incoming_atomic_en = qp->incoming_atomic_en;
+       out_params->e2e_flow_control_en = qp->e2e_flow_control_en;
+       out_params->incoming_rdma_read_en = qp->incoming_rdma_read_en;
+       out_params->incoming_rdma_write_en = qp->incoming_rdma_write_en;
+       out_params->dgid = qp->dgid;
+       out_params->flow_label = qp->flow_label;
+       out_params->hop_limit_ttl = qp->hop_limit_ttl;
+       out_params->traffic_class_tos = qp->traffic_class_tos;
+       out_params->timeout = qp->ack_timeout;
+       out_params->rnr_retry = qp->rnr_retry_cnt;
+       out_params->retry_cnt = qp->retry_cnt;
+       out_params->min_rnr_nak_timer = qp->min_rnr_nak_timer;
+       out_params->pkey_index = 0;
+       out_params->max_rd_atomic = qp->max_rd_atomic_req;
+       out_params->max_dest_rd_atomic = qp->max_rd_atomic_resp;
+       out_params->sqd_async = qp->sqd_async;
+
+       rc = qed_roce_query_qp(p_hwfn, qp, out_params);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Query QP, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_rdma_destroy_qp(void *rdma_cxt, struct qed_rdma_qp *qp)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       int rc = 0;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x\n", qp->icid);
+
+       rc = qed_roce_destroy_qp(p_hwfn, qp);
+
+       /* free qp params struct */
+       kfree(qp);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "QP destroyed\n");
+       return rc;
+}
+
+static struct qed_rdma_qp *
+qed_rdma_create_qp(void *rdma_cxt,
+                  struct qed_rdma_create_qp_in_params *in_params,
+                  struct qed_rdma_create_qp_out_params *out_params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct qed_rdma_qp *qp;
+       u8 max_stats_queues;
+       int rc;
+
+       if (!rdma_cxt || !in_params || !out_params || !p_hwfn->p_rdma_info) {
+               DP_ERR(p_hwfn->cdev,
+                      "qed roce create qp failed due to NULL entry (rdma_cxt=%p, in=%p, out=%p, roce_info=?\n",
+                      rdma_cxt, in_params, out_params);
+               return NULL;
+       }
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                  "qed rdma create qp called with qp_handle = %08x%08x\n",
+                  in_params->qp_handle_hi, in_params->qp_handle_lo);
+
+       /* Some sanity checks... */
+       max_stats_queues = p_hwfn->p_rdma_info->dev->max_stats_queues;
+       if (in_params->stats_queue >= max_stats_queues) {
+               DP_ERR(p_hwfn->cdev,
+                      "qed rdma create qp failed due to invalid statistics queue %d. maximum is %d\n",
+                      in_params->stats_queue, max_stats_queues);
+               return NULL;
+       }
+
+       qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+       if (!qp)
+               return NULL;
+
+       rc = qed_roce_alloc_cid(p_hwfn, &qp->icid);
+       qp->qpid = ((0xFF << 16) | qp->icid);
+
+       DP_INFO(p_hwfn, "ROCE qpid=%x\n", qp->qpid);
+
+       if (rc) {
+               kfree(qp);
+               return NULL;
+       }
+
+       qp->cur_state = QED_ROCE_QP_STATE_RESET;
+       qp->qp_handle.hi = cpu_to_le32(in_params->qp_handle_hi);
+       qp->qp_handle.lo = cpu_to_le32(in_params->qp_handle_lo);
+       qp->qp_handle_async.hi = cpu_to_le32(in_params->qp_handle_async_hi);
+       qp->qp_handle_async.lo = cpu_to_le32(in_params->qp_handle_async_lo);
+       qp->use_srq = in_params->use_srq;
+       qp->signal_all = in_params->signal_all;
+       qp->fmr_and_reserved_lkey = in_params->fmr_and_reserved_lkey;
+       qp->pd = in_params->pd;
+       qp->dpi = in_params->dpi;
+       qp->sq_cq_id = in_params->sq_cq_id;
+       qp->sq_num_pages = in_params->sq_num_pages;
+       qp->sq_pbl_ptr = in_params->sq_pbl_ptr;
+       qp->rq_cq_id = in_params->rq_cq_id;
+       qp->rq_num_pages = in_params->rq_num_pages;
+       qp->rq_pbl_ptr = in_params->rq_pbl_ptr;
+       qp->srq_id = in_params->srq_id;
+       qp->req_offloaded = false;
+       qp->resp_offloaded = false;
+       qp->e2e_flow_control_en = qp->use_srq ? false : true;
+       qp->stats_queue = in_params->stats_queue;
+
+       out_params->icid = qp->icid;
+       out_params->qp_id = qp->qpid;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Create QP, rc = %d\n", rc);
+       return qp;
+}
+
+static int qed_roce_modify_qp(struct qed_hwfn *p_hwfn,
+                             struct qed_rdma_qp *qp,
+                             enum qed_roce_qp_state prev_state,
+                             struct qed_rdma_modify_qp_in_params *params)
+{
+       u32 num_invalidated_mw = 0, num_bound_mw = 0;
+       int rc = 0;
+
+       /* Perform additional operations according to the current state and the
+        * next state
+        */
+       if (((prev_state == QED_ROCE_QP_STATE_INIT) ||
+            (prev_state == QED_ROCE_QP_STATE_RESET)) &&
+           (qp->cur_state == QED_ROCE_QP_STATE_RTR)) {
+               /* Init->RTR or Reset->RTR */
+               rc = qed_roce_sp_create_responder(p_hwfn, qp);
+               return rc;
+       } else if ((prev_state == QED_ROCE_QP_STATE_RTR) &&
+                  (qp->cur_state == QED_ROCE_QP_STATE_RTS)) {
+               /* RTR-> RTS */
+               rc = qed_roce_sp_create_requester(p_hwfn, qp);
+               if (rc)
+                       return rc;
+
+               /* Send modify responder ramrod */
+               rc = qed_roce_sp_modify_responder(p_hwfn, qp, false,
+                                                 params->modify_flags);
+               return rc;
+       } else if ((prev_state == QED_ROCE_QP_STATE_RTS) &&
+                  (qp->cur_state == QED_ROCE_QP_STATE_RTS)) {
+               /* RTS->RTS */
+               rc = qed_roce_sp_modify_responder(p_hwfn, qp, false,
+                                                 params->modify_flags);
+               if (rc)
+                       return rc;
+
+               rc = qed_roce_sp_modify_requester(p_hwfn, qp, false, false,
+                                                 params->modify_flags);
+               return rc;
+       } else if ((prev_state == QED_ROCE_QP_STATE_RTS) &&
+                  (qp->cur_state == QED_ROCE_QP_STATE_SQD)) {
+               /* RTS->SQD */
+               rc = qed_roce_sp_modify_requester(p_hwfn, qp, true, false,
+                                                 params->modify_flags);
+               return rc;
+       } else if ((prev_state == QED_ROCE_QP_STATE_SQD) &&
+                  (qp->cur_state == QED_ROCE_QP_STATE_SQD)) {
+               /* SQD->SQD */
+               rc = qed_roce_sp_modify_responder(p_hwfn, qp, false,
+                                                 params->modify_flags);
+               if (rc)
+                       return rc;
+
+               rc = qed_roce_sp_modify_requester(p_hwfn, qp, false, false,
+                                                 params->modify_flags);
+               return rc;
+       } else if ((prev_state == QED_ROCE_QP_STATE_SQD) &&
+                  (qp->cur_state == QED_ROCE_QP_STATE_RTS)) {
+               /* SQD->RTS */
+               rc = qed_roce_sp_modify_responder(p_hwfn, qp, false,
+                                                 params->modify_flags);
+               if (rc)
+                       return rc;
+
+               rc = qed_roce_sp_modify_requester(p_hwfn, qp, false, false,
+                                                 params->modify_flags);
+
+               return rc;
+       } else if (qp->cur_state == QED_ROCE_QP_STATE_ERR) {
+               /* ->ERR */
+               rc = qed_roce_sp_modify_responder(p_hwfn, qp, true,
+                                                 params->modify_flags);
+               if (rc)
+                       return rc;
+
+               rc = qed_roce_sp_modify_requester(p_hwfn, qp, false, true,
+                                                 params->modify_flags);
+               return rc;
+       } else if (qp->cur_state == QED_ROCE_QP_STATE_RESET) {
+               /* Any state -> RESET */
+               u32 cq_prod;
+
+               /* Send destroy responder ramrod */
+               rc = qed_roce_sp_destroy_qp_responder(p_hwfn,
+                                                     qp,
+                                                     &num_invalidated_mw,
+                                                     &cq_prod);
+
+               if (rc)
+                       return rc;
+
+               qp->cq_prod = cq_prod;
+
+               rc = qed_roce_sp_destroy_qp_requester(p_hwfn, qp,
+                                                     &num_bound_mw);
+
+               if (num_invalidated_mw != num_bound_mw) {
+                       DP_NOTICE(p_hwfn,
+                                 "number of invalidate memory windows is different from bounded ones\n");
+                       return -EINVAL;
+               }
+       } else {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "0\n");
+       }
+
+       return rc;
+}
+
+static int qed_rdma_modify_qp(void *rdma_cxt,
+                             struct qed_rdma_qp *qp,
+                             struct qed_rdma_modify_qp_in_params *params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       enum qed_roce_qp_state prev_state;
+       int rc = 0;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "icid = %08x params->new_state=%d\n",
+                  qp->icid, params->new_state);
+
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+               return rc;
+       }
+
+       if (GET_FIELD(params->modify_flags,
+                     QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN)) {
+               qp->incoming_rdma_read_en = params->incoming_rdma_read_en;
+               qp->incoming_rdma_write_en = params->incoming_rdma_write_en;
+               qp->incoming_atomic_en = params->incoming_atomic_en;
+       }
+
+       /* Update QP structure with the updated values */
+       if (GET_FIELD(params->modify_flags, QED_ROCE_MODIFY_QP_VALID_ROCE_MODE))
+               qp->roce_mode = params->roce_mode;
+       if (GET_FIELD(params->modify_flags, QED_ROCE_MODIFY_QP_VALID_PKEY))
+               qp->pkey = params->pkey;
+       if (GET_FIELD(params->modify_flags,
+                     QED_ROCE_MODIFY_QP_VALID_E2E_FLOW_CONTROL_EN))
+               qp->e2e_flow_control_en = params->e2e_flow_control_en;
+       if (GET_FIELD(params->modify_flags, QED_ROCE_MODIFY_QP_VALID_DEST_QP))
+               qp->dest_qp = params->dest_qp;
+       if (GET_FIELD(params->modify_flags,
+                     QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR)) {
+               /* Indicates that the following parameters have changed:
+                * Traffic class, flow label, hop limit, source GID,
+                * destination GID, loopback indicator
+                */
+               qp->traffic_class_tos = params->traffic_class_tos;
+               qp->flow_label = params->flow_label;
+               qp->hop_limit_ttl = params->hop_limit_ttl;
+
+               qp->sgid = params->sgid;
+               qp->dgid = params->dgid;
+               qp->udp_src_port = 0;
+               qp->vlan_id = params->vlan_id;
+               qp->mtu = params->mtu;
+               qp->lb_indication = params->lb_indication;
+               memcpy((u8 *)&qp->remote_mac_addr[0],
+                      (u8 *)&params->remote_mac_addr[0], ETH_ALEN);
+               if (params->use_local_mac) {
+                       memcpy((u8 *)&qp->local_mac_addr[0],
+                              (u8 *)&params->local_mac_addr[0], ETH_ALEN);
+               } else {
+                       memcpy((u8 *)&qp->local_mac_addr[0],
+                              (u8 *)&p_hwfn->hw_info.hw_mac_addr, ETH_ALEN);
+               }
+       }
+       if (GET_FIELD(params->modify_flags, QED_ROCE_MODIFY_QP_VALID_RQ_PSN))
+               qp->rq_psn = params->rq_psn;
+       if (GET_FIELD(params->modify_flags, QED_ROCE_MODIFY_QP_VALID_SQ_PSN))
+               qp->sq_psn = params->sq_psn;
+       if (GET_FIELD(params->modify_flags,
+                     QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ))
+               qp->max_rd_atomic_req = params->max_rd_atomic_req;
+       if (GET_FIELD(params->modify_flags,
+                     QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP))
+               qp->max_rd_atomic_resp = params->max_rd_atomic_resp;
+       if (GET_FIELD(params->modify_flags,
+                     QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT))
+               qp->ack_timeout = params->ack_timeout;
+       if (GET_FIELD(params->modify_flags, QED_ROCE_MODIFY_QP_VALID_RETRY_CNT))
+               qp->retry_cnt = params->retry_cnt;
+       if (GET_FIELD(params->modify_flags,
+                     QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT))
+               qp->rnr_retry_cnt = params->rnr_retry_cnt;
+       if (GET_FIELD(params->modify_flags,
+                     QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER))
+               qp->min_rnr_nak_timer = params->min_rnr_nak_timer;
+
+       qp->sqd_async = params->sqd_async;
+
+       prev_state = qp->cur_state;
+       if (GET_FIELD(params->modify_flags,
+                     QED_RDMA_MODIFY_QP_VALID_NEW_STATE)) {
+               qp->cur_state = params->new_state;
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "qp->cur_state=%d\n",
+                          qp->cur_state);
+       }
+
+       rc = qed_roce_modify_qp(p_hwfn, qp, prev_state, params);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Modify QP, rc = %d\n", rc);
+       return rc;
+}
+
+static int
+qed_rdma_register_tid(void *rdma_cxt,
+                     struct qed_rdma_register_tid_in_params *params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct rdma_register_tid_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       enum rdma_tid_type tid_type;
+       u8 fw_return_code;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "itid = %08x\n", params->itid);
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent, RDMA_RAMROD_REGISTER_MR,
+                                p_hwfn->p_rdma_info->proto, &init_data);
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+               return rc;
+       }
+
+       if (p_hwfn->p_rdma_info->last_tid < params->itid)
+               p_hwfn->p_rdma_info->last_tid = params->itid;
+
+       p_ramrod = &p_ent->ramrod.rdma_register_tid;
+
+       p_ramrod->flags = 0;
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_TWO_LEVEL_PBL,
+                 params->pbl_two_level);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_ZERO_BASED, params->zbva);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_PHY_MR, params->phy_mr);
+
+       /* Don't initialize D/C field, as it may override other bits. */
+       if (!(params->tid_type == QED_RDMA_TID_FMR) && !(params->dma_mr))
+               SET_FIELD(p_ramrod->flags,
+                         RDMA_REGISTER_TID_RAMROD_DATA_PAGE_SIZE_LOG,
+                         params->page_size_log - 12);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_REMOTE_READ,
+                 params->remote_read);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_REMOTE_WRITE,
+                 params->remote_write);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_REMOTE_ATOMIC,
+                 params->remote_atomic);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_LOCAL_WRITE,
+                 params->local_write);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_LOCAL_READ, params->local_read);
+
+       SET_FIELD(p_ramrod->flags,
+                 RDMA_REGISTER_TID_RAMROD_DATA_ENABLE_MW_BIND,
+                 params->mw_bind);
+
+       SET_FIELD(p_ramrod->flags1,
+                 RDMA_REGISTER_TID_RAMROD_DATA_PBL_PAGE_SIZE_LOG,
+                 params->pbl_page_size_log - 12);
+
+       SET_FIELD(p_ramrod->flags2,
+                 RDMA_REGISTER_TID_RAMROD_DATA_DMA_MR, params->dma_mr);
+
+       switch (params->tid_type) {
+       case QED_RDMA_TID_REGISTERED_MR:
+               tid_type = RDMA_TID_REGISTERED_MR;
+               break;
+       case QED_RDMA_TID_FMR:
+               tid_type = RDMA_TID_FMR;
+               break;
+       case QED_RDMA_TID_MW_TYPE1:
+               tid_type = RDMA_TID_MW_TYPE1;
+               break;
+       case QED_RDMA_TID_MW_TYPE2A:
+               tid_type = RDMA_TID_MW_TYPE2A;
+               break;
+       default:
+               rc = -EINVAL;
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+               return rc;
+       }
+       SET_FIELD(p_ramrod->flags1,
+                 RDMA_REGISTER_TID_RAMROD_DATA_TID_TYPE, tid_type);
+
+       p_ramrod->itid = cpu_to_le32(params->itid);
+       p_ramrod->key = params->key;
+       p_ramrod->pd = cpu_to_le16(params->pd);
+       p_ramrod->length_hi = (u8)(params->length >> 32);
+       p_ramrod->length_lo = DMA_LO_LE(params->length);
+       if (params->zbva) {
+               /* Lower 32 bits of the registered MR address.
+                * In case of zero based MR, will hold FBO
+                */
+               p_ramrod->va.hi = 0;
+               p_ramrod->va.lo = cpu_to_le32(params->fbo);
+       } else {
+               DMA_REGPAIR_LE(p_ramrod->va, params->vaddr);
+       }
+       DMA_REGPAIR_LE(p_ramrod->pbl_base, params->pbl_ptr);
+
+       /* DIF */
+       if (params->dif_enabled) {
+               SET_FIELD(p_ramrod->flags2,
+                         RDMA_REGISTER_TID_RAMROD_DATA_DIF_ON_HOST_FLG, 1);
+               DMA_REGPAIR_LE(p_ramrod->dif_error_addr,
+                              params->dif_error_addr);
+               DMA_REGPAIR_LE(p_ramrod->dif_runt_addr, params->dif_runt_addr);
+       }
+
+       rc = qed_spq_post(p_hwfn, p_ent, &fw_return_code);
+       if (rc)
+               return rc;
+
+       if (fw_return_code != RDMA_RETURN_OK) {
+               DP_NOTICE(p_hwfn, "fw_return_code = %d\n", fw_return_code);
+               return -EINVAL;
+       }
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "Register TID, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_rdma_deregister_tid(void *rdma_cxt, u32 itid)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct rdma_deregister_tid_ramrod_data *p_ramrod;
+       struct qed_sp_init_data init_data;
+       struct qed_spq_entry *p_ent;
+       struct qed_ptt *p_ptt;
+       u8 fw_return_code;
+       int rc;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "itid = %08x\n", itid);
+
+       /* Get SPQ entry */
+       memset(&init_data, 0, sizeof(init_data));
+       init_data.opaque_fid = p_hwfn->hw_info.opaque_fid;
+       init_data.comp_mode = QED_SPQ_MODE_EBLOCK;
+
+       rc = qed_sp_init_request(p_hwfn, &p_ent, RDMA_RAMROD_DEREGISTER_MR,
+                                p_hwfn->p_rdma_info->proto, &init_data);
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+               return rc;
+       }
+
+       p_ramrod = &p_ent->ramrod.rdma_deregister_tid;
+       p_ramrod->itid = cpu_to_le32(itid);
+
+       rc = qed_spq_post(p_hwfn, p_ent, &fw_return_code);
+       if (rc) {
+               DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "rc = %d\n", rc);
+               return rc;
+       }
+
+       if (fw_return_code == RDMA_RETURN_DEREGISTER_MR_BAD_STATE_ERR) {
+               DP_NOTICE(p_hwfn, "fw_return_code = %d\n", fw_return_code);
+               return -EINVAL;
+       } else if (fw_return_code == RDMA_RETURN_NIG_DRAIN_REQ) {
+               /* Bit indicating that the TID is in use and a nig drain is
+                * required before sending the ramrod again
+                */
+               p_ptt = qed_ptt_acquire(p_hwfn);
+               if (!p_ptt) {
+                       rc = -EBUSY;
+                       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                                  "Failed to acquire PTT\n");
+                       return rc;
+               }
+
+               rc = qed_mcp_drain(p_hwfn, p_ptt);
+               if (rc) {
+                       qed_ptt_release(p_hwfn, p_ptt);
+                       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                                  "Drain failed\n");
+                       return rc;
+               }
+
+               qed_ptt_release(p_hwfn, p_ptt);
+
+               /* Resend the ramrod */
+               rc = qed_sp_init_request(p_hwfn, &p_ent,
+                                        RDMA_RAMROD_DEREGISTER_MR,
+                                        p_hwfn->p_rdma_info->proto,
+                                        &init_data);
+               if (rc) {
+                       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                                  "Failed to init sp-element\n");
+                       return rc;
+               }
+
+               rc = qed_spq_post(p_hwfn, p_ent, &fw_return_code);
+               if (rc) {
+                       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                                  "Ramrod failed\n");
+                       return rc;
+               }
+
+               if (fw_return_code != RDMA_RETURN_OK) {
+                       DP_NOTICE(p_hwfn, "fw_return_code = %d\n",
+                                 fw_return_code);
+                       return rc;
+               }
+       }
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "De-registered TID, rc = %d\n", rc);
+       return rc;
+}
+
+static void qed_roce_free_real_icid(struct qed_hwfn *p_hwfn, u16 icid)
+{
+       struct qed_rdma_info *p_rdma_info = p_hwfn->p_rdma_info;
+       u32 start_cid, cid, xcid;
+
+       /* an even icid belongs to a responder while an odd icid belongs to a
+        * requester. The 'cid' received as an input can be either. We calculate
+        * the "partner" icid and call it xcid. Only if both are free then the
+        * "cid" map can be cleared.
+        */
+       start_cid = qed_cxt_get_proto_cid_start(p_hwfn, p_rdma_info->proto);
+       cid = icid - start_cid;
+       xcid = cid ^ 1;
+
+       spin_lock_bh(&p_rdma_info->lock);
+
+       qed_bmap_release_id(p_hwfn, &p_rdma_info->real_cid_map, cid);
+       if (qed_bmap_test_id(p_hwfn, &p_rdma_info->real_cid_map, xcid) == 0) {
+               qed_bmap_release_id(p_hwfn, &p_rdma_info->cid_map, cid);
+               qed_bmap_release_id(p_hwfn, &p_rdma_info->cid_map, xcid);
+       }
+
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
+static void *qed_rdma_get_rdma_ctx(struct qed_dev *cdev)
+{
+       return QED_LEADING_HWFN(cdev);
+}
+
+static bool qed_rdma_allocated_qps(struct qed_hwfn *p_hwfn)
+{
+       bool result;
+
+       /* if rdma info has not been allocated, naturally there are no qps */
+       if (!p_hwfn->p_rdma_info)
+               return false;
+
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       if (!p_hwfn->p_rdma_info->cid_map.bitmap)
+               result = false;
+       else
+               result = !qed_bmap_is_empty(&p_hwfn->p_rdma_info->cid_map);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+       return result;
+}
+
+static void qed_rdma_dpm_conf(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt)
+{
+       u32 val;
+
+       val = (p_hwfn->dcbx_no_edpm || p_hwfn->db_bar_no_edpm) ? 0 : 1;
+
+       qed_wr(p_hwfn, p_ptt, DORQ_REG_PF_DPM_ENABLE, val);
+       DP_VERBOSE(p_hwfn, (QED_MSG_DCB | QED_MSG_RDMA),
+                  "Changing DPM_EN state to %d (DCBX=%d, DB_BAR=%d)\n",
+                  val, p_hwfn->dcbx_no_edpm, p_hwfn->db_bar_no_edpm);
+}
+
+void qed_roce_dpm_dcbx(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt)
+{
+       u8 val;
+
+       /* if any QPs are already active, we want to disable DPM, since their
+        * context information contains information from before the latest DCBx
+        * update. Otherwise enable it.
+        */
+       val = qed_rdma_allocated_qps(p_hwfn) ? true : false;
+       p_hwfn->dcbx_no_edpm = (u8)val;
+
+       qed_rdma_dpm_conf(p_hwfn, p_ptt);
+}
+
+void qed_rdma_dpm_bar(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt)
+{
+       p_hwfn->db_bar_no_edpm = true;
+
+       qed_rdma_dpm_conf(p_hwfn, p_ptt);
+}
+
+static int qed_rdma_start(void *rdma_cxt,
+                         struct qed_rdma_start_in_params *params)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+       struct qed_ptt *p_ptt;
+       int rc = -EBUSY;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                  "desired_cnq = %08x\n", params->desired_cnq);
+
+       p_ptt = qed_ptt_acquire(p_hwfn);
+       if (!p_ptt)
+               goto err;
+
+       rc = qed_rdma_alloc(p_hwfn, p_ptt, params);
+       if (rc)
+               goto err1;
+
+       rc = qed_rdma_setup(p_hwfn, p_ptt, params);
+       if (rc)
+               goto err2;
+
+       qed_ptt_release(p_hwfn, p_ptt);
+
+       return rc;
+
+err2:
+       qed_rdma_free(p_hwfn);
+err1:
+       qed_ptt_release(p_hwfn, p_ptt);
+err:
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "RDMA start - error, rc = %d\n", rc);
+       return rc;
+}
+
+static int qed_rdma_init(struct qed_dev *cdev,
+                        struct qed_rdma_start_in_params *params)
+{
+       return qed_rdma_start(QED_LEADING_HWFN(cdev), params);
+}
+
+static void qed_rdma_remove_user(void *rdma_cxt, u16 dpi)
+{
+       struct qed_hwfn *p_hwfn = (struct qed_hwfn *)rdma_cxt;
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "dpi = %08x\n", dpi);
+
+       spin_lock_bh(&p_hwfn->p_rdma_info->lock);
+       qed_bmap_release_id(p_hwfn, &p_hwfn->p_rdma_info->dpi_map, dpi);
+       spin_unlock_bh(&p_hwfn->p_rdma_info->lock);
+}
+
+static int qed_roce_ll2_set_mac_filter(struct qed_dev *cdev,
+                                      u8 *old_mac_address,
+                                      u8 *new_mac_address)
+{
+       struct qed_hwfn *p_hwfn = QED_LEADING_HWFN(cdev);
+       struct qed_ptt *p_ptt;
+       int rc = 0;
+
+       p_ptt = qed_ptt_acquire(p_hwfn);
+       if (!p_ptt) {
+               DP_ERR(cdev,
+                      "qed roce ll2 mac filter set: failed to acquire PTT\n");
+               return -EINVAL;
+       }
+
+       if (old_mac_address)
+               qed_llh_remove_mac_filter(p_hwfn, p_ptt, old_mac_address);
+       if (new_mac_address)
+               rc = qed_llh_add_mac_filter(p_hwfn, p_ptt, new_mac_address);
+
+       qed_ptt_release(p_hwfn, p_ptt);
+
+       if (rc)
+               DP_ERR(cdev,
+                      "qed roce ll2 mac filter set: failed to add MAC filter\n");
+
+       return rc;
+}
+
+static const struct qed_rdma_ops qed_rdma_ops_pass = {
+       .common = &qed_common_ops_pass,
+       .fill_dev_info = &qed_fill_rdma_dev_info,
+       .rdma_get_rdma_ctx = &qed_rdma_get_rdma_ctx,
+       .rdma_init = &qed_rdma_init,
+       .rdma_add_user = &qed_rdma_add_user,
+       .rdma_remove_user = &qed_rdma_remove_user,
+       .rdma_stop = &qed_rdma_stop,
+       .rdma_query_port = &qed_rdma_query_port,
+       .rdma_query_device = &qed_rdma_query_device,
+       .rdma_get_start_sb = &qed_rdma_get_sb_start,
+       .rdma_get_rdma_int = &qed_rdma_get_int,
+       .rdma_set_rdma_int = &qed_rdma_set_int,
+       .rdma_get_min_cnq_msix = &qed_rdma_get_min_cnq_msix,
+       .rdma_cnq_prod_update = &qed_rdma_cnq_prod_update,
+       .rdma_alloc_pd = &qed_rdma_alloc_pd,
+       .rdma_dealloc_pd = &qed_rdma_free_pd,
+       .rdma_create_cq = &qed_rdma_create_cq,
+       .rdma_destroy_cq = &qed_rdma_destroy_cq,
+       .rdma_create_qp = &qed_rdma_create_qp,
+       .rdma_modify_qp = &qed_rdma_modify_qp,
+       .rdma_query_qp = &qed_rdma_query_qp,
+       .rdma_destroy_qp = &qed_rdma_destroy_qp,
+       .rdma_alloc_tid = &qed_rdma_alloc_tid,
+       .rdma_free_tid = &qed_rdma_free_tid,
+       .rdma_register_tid = &qed_rdma_register_tid,
+       .rdma_deregister_tid = &qed_rdma_deregister_tid,
+       .ll2_acquire_connection = &qed_ll2_acquire_connection,
+       .ll2_establish_connection = &qed_ll2_establish_connection,
+       .ll2_terminate_connection = &qed_ll2_terminate_connection,
+       .ll2_release_connection = &qed_ll2_release_connection,
+       .ll2_post_rx_buffer = &qed_ll2_post_rx_buffer,
+       .ll2_prepare_tx_packet = &qed_ll2_prepare_tx_packet,
+       .ll2_set_fragment_of_tx_packet = &qed_ll2_set_fragment_of_tx_packet,
+       .ll2_set_mac_filter = &qed_roce_ll2_set_mac_filter,
+       .ll2_get_stats = &qed_ll2_get_stats,
+};
+
+const struct qed_rdma_ops *qed_get_rdma_ops(void)
+{
+       return &qed_rdma_ops_pass;
+}
+EXPORT_SYMBOL(qed_get_rdma_ops);
diff --git a/drivers/net/ethernet/qlogic/qed/qed_rdma.h b/drivers/net/ethernet/qlogic/qed/qed_rdma.h
new file mode 100644 (file)
index 0000000..b178d99
--- /dev/null
@@ -0,0 +1,178 @@
+/* QLogic qed NIC Driver
+ * Copyright (c) 2015-2017  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 _QED_ROCE_H
+#define _QED_ROCE_H
+#include <linux/types.h>
+#include <linux/bitops.h>
+#include <linux/kernel.h>
+#include <linux/list.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/qed/qed_if.h>
+#include <linux/qed/qed_roce_if.h>
+#include "qed.h"
+#include "qed_dev_api.h"
+#include "qed_hsi.h"
+#include "qed_ll2.h"
+
+#define QED_RDMA_MAX_FMR                    (RDMA_MAX_TIDS)
+#define QED_RDMA_MAX_P_KEY                  (1)
+#define QED_RDMA_MAX_WQE                    (0x7FFF)
+#define QED_RDMA_MAX_SRQ_WQE_ELEM           (0x7FFF)
+#define QED_RDMA_PAGE_SIZE_CAPS             (0xFFFFF000)
+#define QED_RDMA_ACK_DELAY                  (15)
+#define QED_RDMA_MAX_MR_SIZE                (0x10000000000ULL)
+#define QED_RDMA_MAX_CQS                    (RDMA_MAX_CQS)
+#define QED_RDMA_MAX_MRS                    (RDMA_MAX_TIDS)
+/* Add 1 for header element */
+#define QED_RDMA_MAX_SRQ_ELEM_PER_WQE      (RDMA_MAX_SGE_PER_RQ_WQE + 1)
+#define QED_RDMA_MAX_SGE_PER_SRQ_WQE        (RDMA_MAX_SGE_PER_RQ_WQE)
+#define QED_RDMA_SRQ_WQE_ELEM_SIZE          (16)
+#define QED_RDMA_MAX_SRQS                   (32 * 1024)
+
+#define QED_RDMA_MAX_CQE_32_BIT             (0x7FFFFFFF - 1)
+#define QED_RDMA_MAX_CQE_16_BIT             (0x7FFF - 1)
+
+enum qed_rdma_toggle_bit {
+       QED_RDMA_TOGGLE_BIT_CLEAR = 0,
+       QED_RDMA_TOGGLE_BIT_SET = 1
+};
+
+#define QED_RDMA_MAX_BMAP_NAME (10)
+struct qed_bmap {
+       unsigned long *bitmap;
+       u32 max_count;
+       char name[QED_RDMA_MAX_BMAP_NAME];
+};
+
+struct qed_rdma_info {
+       /* spin lock to protect bitmaps */
+       spinlock_t lock;
+
+       struct qed_bmap cq_map;
+       struct qed_bmap pd_map;
+       struct qed_bmap tid_map;
+       struct qed_bmap qp_map;
+       struct qed_bmap srq_map;
+       struct qed_bmap cid_map;
+       struct qed_bmap real_cid_map;
+       struct qed_bmap dpi_map;
+       struct qed_bmap toggle_bits;
+       struct qed_rdma_events events;
+       struct qed_rdma_device *dev;
+       struct qed_rdma_port *port;
+       u32 last_tid;
+       u8 num_cnqs;
+       u32 num_qps;
+       u32 num_mrs;
+       u16 queue_zone_base;
+       u16 max_queue_zones;
+       enum protocol_type proto;
+};
+
+struct qed_rdma_qp {
+       struct regpair qp_handle;
+       struct regpair qp_handle_async;
+       u32 qpid;
+       u16 icid;
+       enum qed_roce_qp_state cur_state;
+       bool use_srq;
+       bool signal_all;
+       bool fmr_and_reserved_lkey;
+
+       bool incoming_rdma_read_en;
+       bool incoming_rdma_write_en;
+       bool incoming_atomic_en;
+       bool e2e_flow_control_en;
+
+       u16 pd;
+       u16 pkey;
+       u32 dest_qp;
+       u16 mtu;
+       u16 srq_id;
+       u8 traffic_class_tos;
+       u8 hop_limit_ttl;
+       u16 dpi;
+       u32 flow_label;
+       bool lb_indication;
+       u16 vlan_id;
+       u32 ack_timeout;
+       u8 retry_cnt;
+       u8 rnr_retry_cnt;
+       u8 min_rnr_nak_timer;
+       bool sqd_async;
+       union qed_gid sgid;
+       union qed_gid dgid;
+       enum roce_mode roce_mode;
+       u16 udp_src_port;
+       u8 stats_queue;
+
+       /* requeseter */
+       u8 max_rd_atomic_req;
+       u32 sq_psn;
+       u16 sq_cq_id;
+       u16 sq_num_pages;
+       dma_addr_t sq_pbl_ptr;
+       void *orq;
+       dma_addr_t orq_phys_addr;
+       u8 orq_num_pages;
+       bool req_offloaded;
+
+       /* responder */
+       u8 max_rd_atomic_resp;
+       u32 rq_psn;
+       u16 rq_cq_id;
+       u16 rq_num_pages;
+       dma_addr_t rq_pbl_ptr;
+       void *irq;
+       dma_addr_t irq_phys_addr;
+       u8 irq_num_pages;
+       bool resp_offloaded;
+       u32 cq_prod;
+
+       u8 remote_mac_addr[6];
+       u8 local_mac_addr[6];
+
+       void *shared_queue;
+       dma_addr_t shared_queue_phys_addr;
+};
+
+#if IS_ENABLED(CONFIG_QED_RDMA)
+void qed_rdma_dpm_bar(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt);
+void qed_roce_dpm_dcbx(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt);
+#else
+static inline void qed_rdma_dpm_bar(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt) {}
+
+static inline void qed_roce_dpm_dcbx(struct qed_hwfn *p_hwfn,
+                                    struct qed_ptt *p_ptt) {}
+#endif
+#endif