Quelle  verbs.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
/* Copyright (c) 2015 - 2021 Intel Corporation */
#include "main.h"

/**
 * irdma_query_device - get device attributes
 * @ibdev: device pointer from stack
 * @props: returning device attributes
 * @udata: user data
 */
static int irdma_query_device(struct ib_device *ibdev,
         struct ib_device_attr *props,
         struct ib_udata *udata)
{
 struct irdma_device *iwdev = to_iwdev(ibdev);
 struct irdma_pci_f *rf = iwdev->rf;
 struct pci_dev *pcidev = iwdev->rf->pcidev;
 struct irdma_hw_attrs *hw_attrs = &rf->sc_dev.hw_attrs;

 if (udata->inlen || udata->outlen)
  return -EINVAL;

 memset(props, 0, sizeof(*props));
 addrconf_addr_eui48((u8 *)&props->sys_image_guid,
       iwdev->netdev->dev_addr);
 props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
   irdma_fw_minor_ver(&rf->sc_dev);
 props->device_cap_flags = IB_DEVICE_MEM_WINDOW |
      IB_DEVICE_MEM_MGT_EXTENSIONS;
 props->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
 props->vendor_id = pcidev->vendor;
 props->vendor_part_id = pcidev->device;

 props->hw_ver = rf->pcidev->revision;
 props->page_size_cap = hw_attrs->page_size_cap;
 props->max_mr_size = hw_attrs->max_mr_size;
 props->max_qp = rf->max_qp - rf->used_qps;
 props->max_qp_wr = hw_attrs->max_qp_wr;
 props->max_send_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
 props->max_recv_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
 props->max_cq = rf->max_cq - rf->used_cqs;
 props->max_cqe = rf->max_cqe - 1;
 props->max_mr = rf->max_mr - rf->used_mrs;
 props->max_mw = props->max_mr;
 props->max_pd = rf->max_pd - rf->used_pds;
 props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
 props->max_qp_rd_atom = hw_attrs->max_hw_ird;
 props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
 if (rdma_protocol_roce(ibdev, 1)) {
  props->device_cap_flags |= IB_DEVICE_RC_RNR_NAK_GEN;
  props->max_pkeys = IRDMA_PKEY_TBL_SZ;
 }

 props->max_ah = rf->max_ah;
 props->max_mcast_grp = rf->max_mcg;
 props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
 props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
 props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
#define HCA_CLOCK_TIMESTAMP_MASK 0x1ffff
 if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  props->timestamp_mask = HCA_CLOCK_TIMESTAMP_MASK;

 return 0;
}

/**
 * irdma_query_port - get port attributes
 * @ibdev: device pointer from stack
 * @port: port number for query
 * @props: returning device attributes
 */
static int irdma_query_port(struct ib_device *ibdev, u32 port,
       struct ib_port_attr *props)
{
 struct irdma_device *iwdev = to_iwdev(ibdev);
 struct net_device *netdev = iwdev->netdev;

 /* no need to zero out pros here. done by caller */

 props->max_mtu = IB_MTU_4096;
 props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
 props->lid = 1;
 props->lmc = 0;
 props->sm_lid = 0;
 props->sm_sl = 0;
 if (netif_carrier_ok(netdev) && netif_running(netdev)) {
  props->state = IB_PORT_ACTIVE;
  props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 } else {
  props->state = IB_PORT_DOWN;
  props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
 }

 ib_get_eth_speed(ibdev, port, &props->active_speed,
    &props->active_width);

 if (rdma_protocol_roce(ibdev, 1)) {
  props->gid_tbl_len = 32;
  props->ip_gids = true;
  props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 } else {
  props->gid_tbl_len = 1;
 }
 props->qkey_viol_cntr = 0;
 props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;

 return 0;
}

/**
 * irdma_disassociate_ucontext - Disassociate user context
 * @context: ib user context
 */
static void irdma_disassociate_ucontext(struct ib_ucontext *context)
{
}

static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
        struct vm_area_struct *vma)
{
 u64 pfn;

 if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
  return -EINVAL;

 vma->vm_private_data = ucontext;
 pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
        pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;

 return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
     pgprot_noncached(vma->vm_page_prot), NULL);
}

static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
{
 struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);

 kfree(entry);
}

static struct rdma_user_mmap_entry*
irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
        enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
{
 struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 int ret;

 if (!entry)
  return NULL;

 entry->bar_offset = bar_offset;
 entry->mmap_flag = mmap_flag;

 ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
       &entry->rdma_entry, PAGE_SIZE);
 if (ret) {
  kfree(entry);
  return NULL;
 }
 *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);

 return &entry->rdma_entry;
}

/**
 * irdma_mmap - user memory map
 * @context: context created during alloc
 * @vma: kernel info for user memory map
 */
static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
{
 struct rdma_user_mmap_entry *rdma_entry;
 struct irdma_user_mmap_entry *entry;
 struct irdma_ucontext *ucontext;
 u64 pfn;
 int ret;

 ucontext = to_ucontext(context);

 /* Legacy support for libi40iw with hard-coded mmap key */
 if (ucontext->legacy_mode)
  return irdma_mmap_legacy(ucontext, vma);

 rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 if (!rdma_entry) {
  ibdev_dbg(&ucontext->iwdev->ibdev,
     "VERBS: pgoff[0x%lx] does not have valid entry\n",
     vma->vm_pgoff);
  return -EINVAL;
 }

 entry = to_irdma_mmap_entry(rdma_entry);
 ibdev_dbg(&ucontext->iwdev->ibdev,
    "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
    entry->bar_offset, entry->mmap_flag);

 pfn = (entry->bar_offset +
       pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;

 switch (entry->mmap_flag) {
 case IRDMA_MMAP_IO_NC:
  ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
     pgprot_noncached(vma->vm_page_prot),
     rdma_entry);
  break;
 case IRDMA_MMAP_IO_WC:
  ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
     pgprot_writecombine(vma->vm_page_prot),
     rdma_entry);
  break;
 default:
  ret = -EINVAL;
 }

 if (ret)
  ibdev_dbg(&ucontext->iwdev->ibdev,
     "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
     entry->bar_offset, entry->mmap_flag, ret);
 rdma_user_mmap_entry_put(rdma_entry);

 return ret;
}

/**
 * irdma_alloc_push_page - allocate a push page for qp
 * @iwqp: qp pointer
 */
static void irdma_alloc_push_page(struct irdma_qp *iwqp)
{
 struct irdma_cqp_request *cqp_request;
 struct cqp_cmds_info *cqp_info;
 struct irdma_device *iwdev = iwqp->iwdev;
 struct irdma_sc_qp *qp = &iwqp->sc_qp;
 int status;

 cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 if (!cqp_request)
  return;

 cqp_info = &cqp_request->info;
 cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 cqp_info->post_sq = 1;
 cqp_info->in.u.manage_push_page.info.push_idx = 0;
 cqp_info->in.u.manage_push_page.info.qs_handle =
  qp->vsi->qos[qp->user_pri].qs_handle;
 cqp_info->in.u.manage_push_page.info.free_page = 0;
 cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;

 status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 if (!status && cqp_request->compl_info.op_ret_val <
     iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
  qp->push_idx = cqp_request->compl_info.op_ret_val;
  qp->push_offset = 0;
 }

 irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
}

/**
 * irdma_alloc_ucontext - Allocate the user context data structure
 * @uctx: uverbs context pointer
 * @udata: user data
 *
 * This keeps track of all objects associated with a particular
 * user-mode client.
 */
static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
    struct ib_udata *udata)
{
#define IRDMA_ALLOC_UCTX_MIN_REQ_LEN offsetofend(struct irdma_alloc_ucontext_req, rsvd8)
#define IRDMA_ALLOC_UCTX_MIN_RESP_LEN offsetofend(struct irdma_alloc_ucontext_resp, rsvd)
 struct ib_device *ibdev = uctx->device;
 struct irdma_device *iwdev = to_iwdev(ibdev);
 struct irdma_alloc_ucontext_req req = {};
 struct irdma_alloc_ucontext_resp uresp = {};
 struct irdma_ucontext *ucontext = to_ucontext(uctx);
 struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;

 if (udata->inlen < IRDMA_ALLOC_UCTX_MIN_REQ_LEN ||
     udata->outlen < IRDMA_ALLOC_UCTX_MIN_RESP_LEN)
  return -EINVAL;

 if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
  return -EINVAL;

 if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
  goto ver_error;

 ucontext->iwdev = iwdev;
 ucontext->abi_ver = req.userspace_ver;

 if (req.comp_mask & IRDMA_ALLOC_UCTX_USE_RAW_ATTR)
  ucontext->use_raw_attrs = true;

 /* GEN_1 legacy support with libi40iw */
 if (udata->outlen == IRDMA_ALLOC_UCTX_MIN_RESP_LEN) {
  if (uk_attrs->hw_rev != IRDMA_GEN_1)
   return -EOPNOTSUPP;

  ucontext->legacy_mode = true;
  uresp.max_qps = iwdev->rf->max_qp;
  uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
  uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
  uresp.kernel_ver = req.userspace_ver;
  if (ib_copy_to_udata(udata, &uresp,
         min(sizeof(uresp), udata->outlen)))
   return -EFAULT;
 } else {
  u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];

  ucontext->db_mmap_entry =
   irdma_user_mmap_entry_insert(ucontext, bar_off,
           IRDMA_MMAP_IO_NC,
           &uresp.db_mmap_key);
  if (!ucontext->db_mmap_entry)
   return -ENOMEM;

  uresp.kernel_ver = IRDMA_ABI_VER;
  uresp.feature_flags = uk_attrs->feature_flags;
  uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
  uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
  uresp.max_hw_inline = uk_attrs->max_hw_inline;
  uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
  uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
  uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
  uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
  uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
  uresp.hw_rev = uk_attrs->hw_rev;
  uresp.comp_mask |= IRDMA_ALLOC_UCTX_USE_RAW_ATTR;
  uresp.min_hw_wq_size = uk_attrs->min_hw_wq_size;
  uresp.comp_mask |= IRDMA_ALLOC_UCTX_MIN_HW_WQ_SIZE;
  if (ib_copy_to_udata(udata, &uresp,
         min(sizeof(uresp), udata->outlen))) {
   rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
   return -EFAULT;
  }
 }

 INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 spin_lock_init(&ucontext->qp_reg_mem_list_lock);

 return 0;

ver_error:
 ibdev_err(&iwdev->ibdev,
    "Invalid userspace driver version detected. Detected version %d, should be %d\n",
    req.userspace_ver, IRDMA_ABI_VER);
 return -EINVAL;
}

/**
 * irdma_dealloc_ucontext - deallocate the user context data structure
 * @context: user context created during alloc
 */
static void irdma_dealloc_ucontext(struct ib_ucontext *context)
{
 struct irdma_ucontext *ucontext = to_ucontext(context);

 rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
}

/**
 * irdma_alloc_pd - allocate protection domain
 * @pd: PD pointer
 * @udata: user data
 */
static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
{
#define IRDMA_ALLOC_PD_MIN_RESP_LEN offsetofend(struct irdma_alloc_pd_resp, rsvd)
 struct irdma_pd *iwpd = to_iwpd(pd);
 struct irdma_device *iwdev = to_iwdev(pd->device);
 struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 struct irdma_pci_f *rf = iwdev->rf;
 struct irdma_alloc_pd_resp uresp = {};
 struct irdma_sc_pd *sc_pd;
 u32 pd_id = 0;
 int err;

 if (udata && udata->outlen < IRDMA_ALLOC_PD_MIN_RESP_LEN)
  return -EINVAL;

 err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
          &rf->next_pd);
 if (err)
  return err;

 sc_pd = &iwpd->sc_pd;
 if (udata) {
  struct irdma_ucontext *ucontext =
   rdma_udata_to_drv_context(udata, struct irdma_ucontext,
        ibucontext);
  irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
  uresp.pd_id = pd_id;
  if (ib_copy_to_udata(udata, &uresp,
         min(sizeof(uresp), udata->outlen))) {
   err = -EFAULT;
   goto error;
  }
 } else {
  irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 }

 return 0;
error:
 irdma_free_rsrc(rf, rf->allocated_pds, pd_id);

 return err;
}

/**
 * irdma_dealloc_pd - deallocate pd
 * @ibpd: ptr of pd to be deallocated
 * @udata: user data
 */
static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
{
 struct irdma_pd *iwpd = to_iwpd(ibpd);
 struct irdma_device *iwdev = to_iwdev(ibpd->device);

 irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);

 return 0;
}

/**
 * irdma_get_pbl - Retrieve pbl from a list given a virtual
 * address
 * @va: user virtual address
 * @pbl_list: pbl list to search in (QP's or CQ's)
 */
static struct irdma_pbl *irdma_get_pbl(unsigned long va,
           struct list_head *pbl_list)
{
 struct irdma_pbl *iwpbl;

 list_for_each_entry (iwpbl, pbl_list, list) {
  if (iwpbl->user_base == va) {
   list_del(&iwpbl->list);
   iwpbl->on_list = false;
   return iwpbl;
  }
 }

 return NULL;
}

/**
 * irdma_clean_cqes - clean cq entries for qp
 * @iwqp: qp ptr (user or kernel)
 * @iwcq: cq ptr
 */
static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
{
 struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 unsigned long flags;

 spin_lock_irqsave(&iwcq->lock, flags);
 irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 spin_unlock_irqrestore(&iwcq->lock, flags);
}

static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
{
 if (iwqp->push_db_mmap_entry) {
  rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
  iwqp->push_db_mmap_entry = NULL;
 }
 if (iwqp->push_wqe_mmap_entry) {
  rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
  iwqp->push_wqe_mmap_entry = NULL;
 }
}

static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
      struct irdma_qp *iwqp,
      u64 *push_wqe_mmap_key,
      u64 *push_db_mmap_key)
{
 struct irdma_device *iwdev = ucontext->iwdev;
 u64 rsvd, bar_off;

 rsvd = IRDMA_PF_BAR_RSVD;
 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 /* skip over db page */
 bar_off += IRDMA_HW_PAGE_SIZE;
 /* push wqe page */
 bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
     bar_off, IRDMA_MMAP_IO_WC,
     push_wqe_mmap_key);
 if (!iwqp->push_wqe_mmap_entry)
  return -ENOMEM;

 /* push doorbell page */
 bar_off += IRDMA_HW_PAGE_SIZE;
 iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
     bar_off, IRDMA_MMAP_IO_NC,
     push_db_mmap_key);
 if (!iwqp->push_db_mmap_entry) {
  rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
  return -ENOMEM;
 }

 return 0;
}

/**
 * irdma_destroy_qp - destroy qp
 * @ibqp: qp's ib pointer also to get to device's qp address
 * @udata: user data
 */
static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
{
 struct irdma_qp *iwqp = to_iwqp(ibqp);
 struct irdma_device *iwdev = iwqp->iwdev;

 iwqp->sc_qp.qp_uk.destroy_pending = true;

 if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
  irdma_modify_qp_to_err(&iwqp->sc_qp);

 if (!iwqp->user_mode)
  cancel_delayed_work_sync(&iwqp->dwork_flush);

 if (!iwqp->user_mode) {
  if (iwqp->iwscq) {
   irdma_clean_cqes(iwqp, iwqp->iwscq);
   if (iwqp->iwrcq != iwqp->iwscq)
    irdma_clean_cqes(iwqp, iwqp->iwrcq);
  }
 }

 irdma_qp_rem_ref(&iwqp->ibqp);
 wait_for_completion(&iwqp->free_qp);
 irdma_free_lsmm_rsrc(iwqp);
 irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);

 irdma_remove_push_mmap_entries(iwqp);
 irdma_free_qp_rsrc(iwqp);

 return 0;
}

/**
 * irdma_setup_virt_qp - setup for allocation of virtual qp
 * @iwdev: irdma device
 * @iwqp: qp ptr
 * @init_info: initialize info to return
 */
static void irdma_setup_virt_qp(struct irdma_device *iwdev,
          struct irdma_qp *iwqp,
          struct irdma_qp_init_info *init_info)
{
 struct irdma_pbl *iwpbl = iwqp->iwpbl;
 struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;

 iwqp->page = qpmr->sq_page;
 init_info->shadow_area_pa = qpmr->shadow;
 if (iwpbl->pbl_allocated) {
  init_info->virtual_map = true;
  init_info->sq_pa = qpmr->sq_pbl.idx;
  init_info->rq_pa = qpmr->rq_pbl.idx;
 } else {
  init_info->sq_pa = qpmr->sq_pbl.addr;
  init_info->rq_pa = qpmr->rq_pbl.addr;
 }
}

/**
 * irdma_setup_umode_qp - setup sq and rq size in user mode qp
 * @udata: udata
 * @iwdev: iwarp device
 * @iwqp: qp ptr (user or kernel)
 * @info: initialize info to return
 * @init_attr: Initial QP create attributes
 */
static int irdma_setup_umode_qp(struct ib_udata *udata,
    struct irdma_device *iwdev,
    struct irdma_qp *iwqp,
    struct irdma_qp_init_info *info,
    struct ib_qp_init_attr *init_attr)
{
 struct irdma_ucontext *ucontext = rdma_udata_to_drv_context(udata,
    struct irdma_ucontext, ibucontext);
 struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 struct irdma_create_qp_req req;
 unsigned long flags;
 int ret;

 ret = ib_copy_from_udata(&req, udata,
     min(sizeof(req), udata->inlen));
 if (ret) {
  ibdev_dbg(&iwdev->ibdev, "VERBS: ib_copy_from_data fail\n");
  return ret;
 }

 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 iwqp->user_mode = 1;
 if (req.user_wqe_bufs) {
  info->qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
  spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
  iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
         &ucontext->qp_reg_mem_list);
  spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);

  if (!iwqp->iwpbl) {
   ret = -ENODATA;
   ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
   return ret;
  }
 }

 if (!ucontext->use_raw_attrs) {
  /**
 * Maintain backward compat with older ABI which passes sq and
 * rq depth in quanta in cap.max_send_wr and cap.max_recv_wr.
 * There is no way to compute the correct value of
 * iwqp->max_send_wr/max_recv_wr in the kernel.
 */
  iwqp->max_send_wr = init_attr->cap.max_send_wr;
  iwqp->max_recv_wr = init_attr->cap.max_recv_wr;
  ukinfo->sq_size = init_attr->cap.max_send_wr;
  ukinfo->rq_size = init_attr->cap.max_recv_wr;
  irdma_uk_calc_shift_wq(ukinfo, &ukinfo->sq_shift,
           &ukinfo->rq_shift);
 } else {
  ret = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
         &ukinfo->sq_shift);
  if (ret)
   return ret;

  ret = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
         &ukinfo->rq_shift);
  if (ret)
   return ret;

  iwqp->max_send_wr =
   (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
  iwqp->max_recv_wr =
   (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
  ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
  ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
 }

 irdma_setup_virt_qp(iwdev, iwqp, info);

 return 0;
}

/**
 * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 * @iwdev: iwarp device
 * @iwqp: qp ptr (user or kernel)
 * @info: initialize info to return
 * @init_attr: Initial QP create attributes
 */
static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
    struct irdma_qp *iwqp,
    struct irdma_qp_init_info *info,
    struct ib_qp_init_attr *init_attr)
{
 struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 u32 size;
 int status;
 struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;

 status = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
           &ukinfo->sq_shift);
 if (status)
  return status;

 status = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
           &ukinfo->rq_shift);
 if (status)
  return status;

 iwqp->kqp.sq_wrid_mem =
  kcalloc(ukinfo->sq_depth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 if (!iwqp->kqp.sq_wrid_mem)
  return -ENOMEM;

 iwqp->kqp.rq_wrid_mem =
  kcalloc(ukinfo->rq_depth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);

 if (!iwqp->kqp.rq_wrid_mem) {
  kfree(iwqp->kqp.sq_wrid_mem);
  iwqp->kqp.sq_wrid_mem = NULL;
  return -ENOMEM;
 }

 ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;

 size = (ukinfo->sq_depth + ukinfo->rq_depth) * IRDMA_QP_WQE_MIN_SIZE;
 size += (IRDMA_SHADOW_AREA_SIZE << 3);

 mem->size = ALIGN(size, 256);
 mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
         &mem->pa, GFP_KERNEL);
 if (!mem->va) {
  kfree(iwqp->kqp.sq_wrid_mem);
  iwqp->kqp.sq_wrid_mem = NULL;
  kfree(iwqp->kqp.rq_wrid_mem);
  iwqp->kqp.rq_wrid_mem = NULL;
  return -ENOMEM;
 }

 ukinfo->sq = mem->va;
 info->sq_pa = mem->pa;
 ukinfo->rq = &ukinfo->sq[ukinfo->sq_depth];
 info->rq_pa = info->sq_pa + (ukinfo->sq_depth * IRDMA_QP_WQE_MIN_SIZE);
 ukinfo->shadow_area = ukinfo->rq[ukinfo->rq_depth].elem;
 info->shadow_area_pa =
  info->rq_pa + (ukinfo->rq_depth * IRDMA_QP_WQE_MIN_SIZE);
 ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
 ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;

 iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
 iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
 init_attr->cap.max_send_wr = iwqp->max_send_wr;
 init_attr->cap.max_recv_wr = iwqp->max_recv_wr;

 return 0;
}

static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
{
 struct irdma_pci_f *rf = iwqp->iwdev->rf;
 struct irdma_cqp_request *cqp_request;
 struct cqp_cmds_info *cqp_info;
 struct irdma_create_qp_info *qp_info;
 int status;

 cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 if (!cqp_request)
  return -ENOMEM;

 cqp_info = &cqp_request->info;
 qp_info = &cqp_request->info.in.u.qp_create.info;
 memset(qp_info, 0, sizeof(*qp_info));
 qp_info->mac_valid = true;
 qp_info->cq_num_valid = true;
 qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;

 cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 cqp_info->post_sq = 1;
 cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 status = irdma_handle_cqp_op(rf, cqp_request);
 irdma_put_cqp_request(&rf->cqp, cqp_request);

 return status;
}

static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
            struct irdma_qp_host_ctx_info *ctx_info)
{
 struct irdma_device *iwdev = iwqp->iwdev;
 struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 struct irdma_roce_offload_info *roce_info;
 struct irdma_udp_offload_info *udp_info;

 udp_info = &iwqp->udp_info;
 udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 udp_info->cwnd = iwdev->roce_cwnd;
 udp_info->rexmit_thresh = 2;
 udp_info->rnr_nak_thresh = 2;
 udp_info->src_port = 0xc000;
 udp_info->dst_port = ROCE_V2_UDP_DPORT;
 roce_info = &iwqp->roce_info;
 ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);

 roce_info->rd_en = true;
 roce_info->wr_rdresp_en = true;
 roce_info->bind_en = true;
 roce_info->dcqcn_en = false;
 roce_info->rtomin = 5;

 roce_info->ack_credits = iwdev->roce_ackcreds;
 roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 roce_info->ord_size = dev->hw_attrs.max_hw_ord;

 if (!iwqp->user_mode) {
  roce_info->priv_mode_en = true;
  roce_info->fast_reg_en = true;
  roce_info->udprivcq_en = true;
 }
 roce_info->roce_tver = 0;

 ctx_info->roce_info = &iwqp->roce_info;
 ctx_info->udp_info = &iwqp->udp_info;
 irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
}

static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
          struct irdma_qp_host_ctx_info *ctx_info)
{
 struct irdma_device *iwdev = iwqp->iwdev;
 struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 struct irdma_iwarp_offload_info *iwarp_info;

 iwarp_info = &iwqp->iwarp_info;
 ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 iwarp_info->rd_en = true;
 iwarp_info->wr_rdresp_en = true;
 iwarp_info->bind_en = true;
 iwarp_info->ecn_en = true;
 iwarp_info->rtomin = 5;

 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
  iwarp_info->ib_rd_en = true;
 if (!iwqp->user_mode) {
  iwarp_info->priv_mode_en = true;
  iwarp_info->fast_reg_en = true;
 }
 iwarp_info->ddp_ver = 1;
 iwarp_info->rdmap_ver = 1;

 ctx_info->iwarp_info = &iwqp->iwarp_info;
 ctx_info->iwarp_info_valid = true;
 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 ctx_info->iwarp_info_valid = false;
}

static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
       struct irdma_device *iwdev)
{
 struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;

 if (init_attr->create_flags)
  return -EOPNOTSUPP;

 if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
     init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
     init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags ||
     init_attr->cap.max_send_wr > uk_attrs->max_hw_wq_quanta ||
     init_attr->cap.max_recv_wr > uk_attrs->max_hw_rq_quanta)
  return -EINVAL;

 if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
  if (init_attr->qp_type != IB_QPT_RC &&
      init_attr->qp_type != IB_QPT_UD &&
      init_attr->qp_type != IB_QPT_GSI)
   return -EOPNOTSUPP;
 } else {
  if (init_attr->qp_type != IB_QPT_RC)
   return -EOPNOTSUPP;
 }

 return 0;
}

static void irdma_flush_worker(struct work_struct *work)
{
 struct delayed_work *dwork = to_delayed_work(work);
 struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);

 irdma_generate_flush_completions(iwqp);
}

/**
 * irdma_create_qp - create qp
 * @ibqp: ptr of qp
 * @init_attr: attributes for qp
 * @udata: user data for create qp
 */
static int irdma_create_qp(struct ib_qp *ibqp,
      struct ib_qp_init_attr *init_attr,
      struct ib_udata *udata)
{
#define IRDMA_CREATE_QP_MIN_REQ_LEN offsetofend(struct irdma_create_qp_req, user_compl_ctx)
#define IRDMA_CREATE_QP_MIN_RESP_LEN offsetofend(struct irdma_create_qp_resp, rsvd)
 struct ib_pd *ibpd = ibqp->pd;
 struct irdma_pd *iwpd = to_iwpd(ibpd);
 struct irdma_device *iwdev = to_iwdev(ibpd->device);
 struct irdma_pci_f *rf = iwdev->rf;
 struct irdma_qp *iwqp = to_iwqp(ibqp);
 struct irdma_create_qp_resp uresp = {};
 u32 qp_num = 0;
 int err_code;
 struct irdma_sc_qp *qp;
 struct irdma_sc_dev *dev = &rf->sc_dev;
 struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 struct irdma_qp_init_info init_info = {};
 struct irdma_qp_host_ctx_info *ctx_info;

 err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 if (err_code)
  return err_code;

 if (udata && (udata->inlen < IRDMA_CREATE_QP_MIN_REQ_LEN ||
        udata->outlen < IRDMA_CREATE_QP_MIN_RESP_LEN))
  return -EINVAL;

 init_info.vsi = &iwdev->vsi;
 init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 init_info.qp_uk_init_info.sq_size = init_attr->cap.max_send_wr;
 init_info.qp_uk_init_info.rq_size = init_attr->cap.max_recv_wr;
 init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;

 qp = &iwqp->sc_qp;
 qp->qp_uk.back_qp = iwqp;
 qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;

 iwqp->iwdev = iwdev;
 iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
          256);
 iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
       iwqp->q2_ctx_mem.size,
       &iwqp->q2_ctx_mem.pa,
       GFP_KERNEL);
 if (!iwqp->q2_ctx_mem.va)
  return -ENOMEM;

 init_info.q2 = iwqp->q2_ctx_mem.va;
 init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;

 if (init_attr->qp_type == IB_QPT_GSI)
  qp_num = 1;
 else
  err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
         &qp_num, &rf->next_qp);
 if (err_code)
  goto error;

 iwqp->iwpd = iwpd;
 iwqp->ibqp.qp_num = qp_num;
 qp = &iwqp->sc_qp;
 iwqp->iwscq = to_iwcq(init_attr->send_cq);
 iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 iwqp->host_ctx.va = init_info.host_ctx;
 iwqp->host_ctx.pa = init_info.host_ctx_pa;
 iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;

 init_info.pd = &iwpd->sc_pd;
 init_info.qp_uk_init_info.qp_id = qp_num;
 if (!rdma_protocol_roce(&iwdev->ibdev, 1))
  init_info.qp_uk_init_info.first_sq_wq = 1;
 iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 init_waitqueue_head(&iwqp->waitq);
 init_waitqueue_head(&iwqp->mod_qp_waitq);

 if (udata) {
  init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
  err_code = irdma_setup_umode_qp(udata, iwdev, iwqp, &init_info,
      init_attr);
 } else {
  INIT_DELAYED_WORK(&iwqp->dwork_flush, irdma_flush_worker);
  init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
  err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 }

 if (err_code) {
  ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
  goto error;
 }

 if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
  if (init_attr->qp_type == IB_QPT_RC) {
   init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
   init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
           IRDMA_WRITE_WITH_IMM |
           IRDMA_ROCE;
  } else {
   init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
   init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
           IRDMA_ROCE;
  }
 } else {
  init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
  init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 }

 if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
  init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;

 err_code = irdma_sc_qp_init(qp, &init_info);
 if (err_code) {
  ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
  goto error;
 }

 ctx_info = &iwqp->ctx_info;
 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;

 if (rdma_protocol_roce(&iwdev->ibdev, 1))
  irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 else
  irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);

 err_code = irdma_cqp_create_qp_cmd(iwqp);
 if (err_code)
  goto error;

 refcount_set(&iwqp->refcnt, 1);
 spin_lock_init(&iwqp->lock);
 spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 iwqp->sig_all = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
 rf->qp_table[qp_num] = iwqp;

 if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
  if (dev->ws_add(&iwdev->vsi, 0)) {
   irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
   err_code = -EINVAL;
   goto error;
  }

  irdma_qp_add_qos(&iwqp->sc_qp);
 }

 if (udata) {
  /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
  if (udata->outlen < sizeof(uresp)) {
   uresp.lsmm = 1;
   uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
  } else {
   if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
    uresp.lsmm = 1;
  }
  uresp.actual_sq_size = init_info.qp_uk_init_info.sq_size;
  uresp.actual_rq_size = init_info.qp_uk_init_info.rq_size;
  uresp.qp_id = qp_num;
  uresp.qp_caps = qp->qp_uk.qp_caps;

  err_code = ib_copy_to_udata(udata, &uresp,
         min(sizeof(uresp), udata->outlen));
  if (err_code) {
   ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
   irdma_destroy_qp(&iwqp->ibqp, udata);
   return err_code;
  }
 }

 init_completion(&iwqp->free_qp);
 return 0;

error:
 irdma_free_qp_rsrc(iwqp);
 return err_code;
}

static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
{
 int acc_flags = 0;

 if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
  if (iwqp->roce_info.wr_rdresp_en) {
   acc_flags |= IB_ACCESS_LOCAL_WRITE;
   acc_flags |= IB_ACCESS_REMOTE_WRITE;
  }
  if (iwqp->roce_info.rd_en)
   acc_flags |= IB_ACCESS_REMOTE_READ;
  if (iwqp->roce_info.bind_en)
   acc_flags |= IB_ACCESS_MW_BIND;
 } else {
  if (iwqp->iwarp_info.wr_rdresp_en) {
   acc_flags |= IB_ACCESS_LOCAL_WRITE;
   acc_flags |= IB_ACCESS_REMOTE_WRITE;
  }
  if (iwqp->iwarp_info.rd_en)
   acc_flags |= IB_ACCESS_REMOTE_READ;
  if (iwqp->iwarp_info.bind_en)
   acc_flags |= IB_ACCESS_MW_BIND;
 }
 return acc_flags;
}

/**
 * irdma_query_qp - query qp attributes
 * @ibqp: qp pointer
 * @attr: attributes pointer
 * @attr_mask: Not used
 * @init_attr: qp attributes to return
 */
static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
     int attr_mask, struct ib_qp_init_attr *init_attr)
{
 struct irdma_qp *iwqp = to_iwqp(ibqp);
 struct irdma_sc_qp *qp = &iwqp->sc_qp;

 memset(attr, 0, sizeof(*attr));
 memset(init_attr, 0, sizeof(*init_attr));

 attr->qp_state = iwqp->ibqp_state;
 attr->cur_qp_state = iwqp->ibqp_state;
 attr->cap.max_send_wr = iwqp->max_send_wr;
 attr->cap.max_recv_wr = iwqp->max_recv_wr;
 attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
 attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
 attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
 attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
 attr->port_num = 1;
 if (rdma_protocol_roce(ibqp->device, 1)) {
  attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
  attr->qkey = iwqp->roce_info.qkey;
  attr->rq_psn = iwqp->udp_info.epsn;
  attr->sq_psn = iwqp->udp_info.psn_nxt;
  attr->dest_qp_num = iwqp->roce_info.dest_qp;
  attr->pkey_index = iwqp->roce_info.p_key;
  attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
  attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
  attr->max_rd_atomic = iwqp->roce_info.ord_size;
  attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
 }

 init_attr->event_handler = iwqp->ibqp.event_handler;
 init_attr->qp_context = iwqp->ibqp.qp_context;
 init_attr->send_cq = iwqp->ibqp.send_cq;
 init_attr->recv_cq = iwqp->ibqp.recv_cq;
 init_attr->cap = attr->cap;

 return 0;
}

/**
 * irdma_query_pkey - Query partition key
 * @ibdev: device pointer from stack
 * @port: port number
 * @index: index of pkey
 * @pkey: pointer to store the pkey
 */
static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
       u16 *pkey)
{
 if (index >= IRDMA_PKEY_TBL_SZ)
  return -EINVAL;

 *pkey = IRDMA_DEFAULT_PKEY;
 return 0;
}

static u8 irdma_roce_get_vlan_prio(const struct ib_gid_attr *attr, u8 prio)
{
 struct net_device *ndev;

 rcu_read_lock();
 ndev = rcu_dereference(attr->ndev);
 if (!ndev)
  goto exit;
 if (is_vlan_dev(ndev)) {
  u16 vlan_qos = vlan_dev_get_egress_qos_mask(ndev, prio);

  prio = (vlan_qos & VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
 }
exit:
 rcu_read_unlock();
 return prio;
}

static int irdma_wait_for_suspend(struct irdma_qp *iwqp)
{
 if (!wait_event_timeout(iwqp->iwdev->suspend_wq,
    !iwqp->suspend_pending,
    msecs_to_jiffies(IRDMA_EVENT_TIMEOUT_MS))) {
  iwqp->suspend_pending = false;
  ibdev_warn(&iwqp->iwdev->ibdev,
      "modify_qp timed out waiting for suspend. qp_id = %d, last_ae = 0x%x\n",
      iwqp->ibqp.qp_num, iwqp->last_aeq);
  return -EBUSY;
 }

 return 0;
}

/**
 * irdma_modify_qp_roce - modify qp request
 * @ibqp: qp's pointer for modify
 * @attr: access attributes
 * @attr_mask: state mask
 * @udata: user data
 */
int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
    int attr_mask, struct ib_udata *udata)
{
#define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
#define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
 struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
 struct irdma_qp *iwqp = to_iwqp(ibqp);
 struct irdma_device *iwdev = iwqp->iwdev;
 struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 struct irdma_qp_host_ctx_info *ctx_info;
 struct irdma_roce_offload_info *roce_info;
 struct irdma_udp_offload_info *udp_info;
 struct irdma_modify_qp_info info = {};
 struct irdma_modify_qp_resp uresp = {};
 struct irdma_modify_qp_req ureq = {};
 unsigned long flags;
 u8 issue_modify_qp = 0;
 int ret = 0;

 ctx_info = &iwqp->ctx_info;
 roce_info = &iwqp->roce_info;
 udp_info = &iwqp->udp_info;

 if (udata) {
  /* udata inlen/outlen can be 0 when supporting legacy libi40iw */
  if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
      (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
   return -EINVAL;
 }

 if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
  return -EOPNOTSUPP;

 if (attr_mask & IB_QP_DEST_QPN)
  roce_info->dest_qp = attr->dest_qp_num;

 if (attr_mask & IB_QP_PKEY_INDEX) {
  ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
           &roce_info->p_key);
  if (ret)
   return ret;
 }

 if (attr_mask & IB_QP_QKEY)
  roce_info->qkey = attr->qkey;

 if (attr_mask & IB_QP_PATH_MTU)
  udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);

 if (attr_mask & IB_QP_SQ_PSN) {
  udp_info->psn_nxt = attr->sq_psn;
  udp_info->lsn =  0xffff;
  udp_info->psn_una = attr->sq_psn;
  udp_info->psn_max = attr->sq_psn;
 }

 if (attr_mask & IB_QP_RQ_PSN)
  udp_info->epsn = attr->rq_psn;

 if (attr_mask & IB_QP_RNR_RETRY)
  udp_info->rnr_nak_thresh = attr->rnr_retry;

 if (attr_mask & IB_QP_RETRY_CNT)
  udp_info->rexmit_thresh = attr->retry_cnt;

 ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;

 if (attr_mask & IB_QP_AV) {
  struct irdma_av *av = &iwqp->roce_ah.av;
  const struct ib_gid_attr *sgid_attr =
    attr->ah_attr.grh.sgid_attr;
  u16 vlan_id = VLAN_N_VID;
  u32 local_ip[4];

  memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
  if (attr->ah_attr.ah_flags & IB_AH_GRH) {
   udp_info->ttl = attr->ah_attr.grh.hop_limit;
   udp_info->flow_label = attr->ah_attr.grh.flow_label;
   udp_info->tos = attr->ah_attr.grh.traffic_class;
   udp_info->src_port =
    rdma_get_udp_sport(udp_info->flow_label,
         ibqp->qp_num,
         roce_info->dest_qp);
   irdma_qp_rem_qos(&iwqp->sc_qp);
   dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
   if (iwqp->sc_qp.vsi->dscp_mode)
    ctx_info->user_pri =
    iwqp->sc_qp.vsi->dscp_map[irdma_tos2dscp(udp_info->tos)];
   else
    ctx_info->user_pri = rt_tos2priority(udp_info->tos);
  }
  ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
           ctx_info->roce_info->mac_addr);
  if (ret)
   return ret;
  ctx_info->user_pri = irdma_roce_get_vlan_prio(sgid_attr,
             ctx_info->user_pri);
  if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
   return -ENOMEM;
  iwqp->sc_qp.user_pri = ctx_info->user_pri;
  irdma_qp_add_qos(&iwqp->sc_qp);

  if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
   vlan_id = 0;
  if (vlan_id < VLAN_N_VID) {
   udp_info->insert_vlan_tag = true;
   udp_info->vlan_tag = vlan_id |
    ctx_info->user_pri << VLAN_PRIO_SHIFT;
  } else {
   udp_info->insert_vlan_tag = false;
  }

  av->attrs = attr->ah_attr;
  rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
  rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
  av->net_type = rdma_gid_attr_network_type(sgid_attr);
  if (av->net_type == RDMA_NETWORK_IPV6) {
   __be32 *daddr =
    av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
   __be32 *saddr =
    av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;

   irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
   irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);

   udp_info->ipv4 = false;
   irdma_copy_ip_ntohl(local_ip, daddr);

  } else if (av->net_type == RDMA_NETWORK_IPV4) {
   __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
   __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;

   local_ip[0] = ntohl(daddr);

   udp_info->ipv4 = true;
   udp_info->dest_ip_addr[0] = 0;
   udp_info->dest_ip_addr[1] = 0;
   udp_info->dest_ip_addr[2] = 0;
   udp_info->dest_ip_addr[3] = local_ip[0];

   udp_info->local_ipaddr[0] = 0;
   udp_info->local_ipaddr[1] = 0;
   udp_info->local_ipaddr[2] = 0;
   udp_info->local_ipaddr[3] = ntohl(saddr);
  }
  udp_info->arp_idx =
   irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
          attr->ah_attr.roce.dmac);
 }

 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
  if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
   ibdev_err(&iwdev->ibdev,
      "rd_atomic = %d, above max_hw_ord=%d\n",
      attr->max_rd_atomic,
      dev->hw_attrs.max_hw_ord);
   return -EINVAL;
  }
  if (attr->max_rd_atomic)
   roce_info->ord_size = attr->max_rd_atomic;
  info.ord_valid = true;
 }

 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
  if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
   ibdev_err(&iwdev->ibdev,
      "rd_atomic = %d, above max_hw_ird=%d\n",
       attr->max_dest_rd_atomic,
       dev->hw_attrs.max_hw_ird);
   return -EINVAL;
  }
  if (attr->max_dest_rd_atomic)
   roce_info->ird_size = attr->max_dest_rd_atomic;
 }

 if (attr_mask & IB_QP_ACCESS_FLAGS) {
  if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
   roce_info->wr_rdresp_en = true;
  if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
   roce_info->wr_rdresp_en = true;
  if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
   roce_info->rd_en = true;
 }

 wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));

 ibdev_dbg(&iwdev->ibdev,
    "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
    __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
    iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);

 spin_lock_irqsave(&iwqp->lock, flags);
 if (attr_mask & IB_QP_STATE) {
  if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
     iwqp->ibqp.qp_type, attr_mask)) {
   ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
       iwqp->ibqp.qp_num, iwqp->ibqp_state,
       attr->qp_state);
   ret = -EINVAL;
   goto exit;
  }
  info.curr_iwarp_state = iwqp->iwarp_state;

  switch (attr->qp_state) {
  case IB_QPS_INIT:
   if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
    ret = -EINVAL;
    goto exit;
   }

   if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
    info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
    issue_modify_qp = 1;
   }
   break;
  case IB_QPS_RTR:
   if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
    ret = -EINVAL;
    goto exit;
   }
   info.arp_cache_idx_valid = true;
   info.cq_num_valid = true;
   info.next_iwarp_state = IRDMA_QP_STATE_RTR;
   issue_modify_qp = 1;
   break;
  case IB_QPS_RTS:
   if (iwqp->ibqp_state < IB_QPS_RTR ||
       iwqp->ibqp_state == IB_QPS_ERR) {
    ret = -EINVAL;
    goto exit;
   }

   info.arp_cache_idx_valid = true;
   info.cq_num_valid = true;
   info.ord_valid = true;
   info.next_iwarp_state = IRDMA_QP_STATE_RTS;
   issue_modify_qp = 1;
   if (iwdev->push_mode && udata &&
       iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
       dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
    spin_unlock_irqrestore(&iwqp->lock, flags);
    irdma_alloc_push_page(iwqp);
    spin_lock_irqsave(&iwqp->lock, flags);
   }
   break;
  case IB_QPS_SQD:
   if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
    goto exit;

   if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
    ret = -EINVAL;
    goto exit;
   }

   info.next_iwarp_state = IRDMA_QP_STATE_SQD;
   issue_modify_qp = 1;
   iwqp->suspend_pending = true;
   break;
  case IB_QPS_SQE:
  case IB_QPS_ERR:
  case IB_QPS_RESET:
   if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
    spin_unlock_irqrestore(&iwqp->lock, flags);
    if (udata && udata->inlen) {
     if (ib_copy_from_udata(&ureq, udata,
         min(sizeof(ureq), udata->inlen)))
      return -EINVAL;

     irdma_flush_wqes(iwqp,
         (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
         (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
         IRDMA_REFLUSH);
    }
    return 0;
   }

   info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
   issue_modify_qp = 1;
   break;
  default:
   ret = -EINVAL;
   goto exit;
  }

  iwqp->ibqp_state = attr->qp_state;
 }

 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 spin_unlock_irqrestore(&iwqp->lock, flags);

 if (attr_mask & IB_QP_STATE) {
  if (issue_modify_qp) {
   ctx_info->rem_endpoint_idx = udp_info->arp_idx;
   if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
    return -EINVAL;
   if (info.next_iwarp_state == IRDMA_QP_STATE_SQD) {
    ret = irdma_wait_for_suspend(iwqp);
    if (ret)
     return ret;
   }
   spin_lock_irqsave(&iwqp->lock, flags);
   if (iwqp->iwarp_state == info.curr_iwarp_state) {
    iwqp->iwarp_state = info.next_iwarp_state;
    iwqp->ibqp_state = attr->qp_state;
   }
   if (iwqp->ibqp_state > IB_QPS_RTS &&
       !iwqp->flush_issued) {
    spin_unlock_irqrestore(&iwqp->lock, flags);
    irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
             IRDMA_FLUSH_RQ |
             IRDMA_FLUSH_WAIT);
    iwqp->flush_issued = 1;
   } else {
    spin_unlock_irqrestore(&iwqp->lock, flags);
   }
  } else {
   iwqp->ibqp_state = attr->qp_state;
  }
  if (udata && udata->outlen && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
   struct irdma_ucontext *ucontext;

   ucontext = rdma_udata_to_drv_context(udata,
     struct irdma_ucontext, ibucontext);
   if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
       !iwqp->push_wqe_mmap_entry &&
       !irdma_setup_push_mmap_entries(ucontext, iwqp,
    &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
    uresp.push_valid = 1;
    uresp.push_offset = iwqp->sc_qp.push_offset;
   }
   ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
            udata->outlen));
   if (ret) {
    irdma_remove_push_mmap_entries(iwqp);
    ibdev_dbg(&iwdev->ibdev,
       "VERBS: copy_to_udata failed\n");
    return ret;
   }
  }
 }

 return 0;
exit:
 spin_unlock_irqrestore(&iwqp->lock, flags);

 return ret;
}

/**
 * irdma_modify_qp - modify qp request
 * @ibqp: qp's pointer for modify
 * @attr: access attributes
 * @attr_mask: state mask
 * @udata: user data
 */
int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
      struct ib_udata *udata)
{
#define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
#define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
 struct irdma_qp *iwqp = to_iwqp(ibqp);
 struct irdma_device *iwdev = iwqp->iwdev;
 struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 struct irdma_qp_host_ctx_info *ctx_info;
 struct irdma_tcp_offload_info *tcp_info;
 struct irdma_iwarp_offload_info *offload_info;
 struct irdma_modify_qp_info info = {};
 struct irdma_modify_qp_resp uresp = {};
 struct irdma_modify_qp_req ureq = {};
 u8 issue_modify_qp = 0;
 u8 dont_wait = 0;
 int err;
 unsigned long flags;

 if (udata) {
  /* udata inlen/outlen can be 0 when supporting legacy libi40iw */
  if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
      (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
   return -EINVAL;
 }

 if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
  return -EOPNOTSUPP;

 ctx_info = &iwqp->ctx_info;
 offload_info = &iwqp->iwarp_info;
 tcp_info = &iwqp->tcp_info;
 wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
 ibdev_dbg(&iwdev->ibdev,
    "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d last_aeq=%d hw_tcp_state=%d hw_iwarp_state=%d attr_mask=0x%x\n",
    __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
    iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
    iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);

 spin_lock_irqsave(&iwqp->lock, flags);
 if (attr_mask & IB_QP_STATE) {
  info.curr_iwarp_state = iwqp->iwarp_state;
  switch (attr->qp_state) {
  case IB_QPS_INIT:
  case IB_QPS_RTR:
   if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
    err = -EINVAL;
    goto exit;
   }

   if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
    info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
    issue_modify_qp = 1;
   }
   if (iwdev->push_mode && udata &&
       iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
       dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
    spin_unlock_irqrestore(&iwqp->lock, flags);
    irdma_alloc_push_page(iwqp);
    spin_lock_irqsave(&iwqp->lock, flags);
   }
   break;
  case IB_QPS_RTS:
   if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
       !iwqp->cm_id) {
    err = -EINVAL;
    goto exit;
   }

   issue_modify_qp = 1;
   iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
   iwqp->hte_added = 1;
   info.next_iwarp_state = IRDMA_QP_STATE_RTS;
   info.tcp_ctx_valid = true;
   info.ord_valid = true;
   info.arp_cache_idx_valid = true;
   info.cq_num_valid = true;
   break;
  case IB_QPS_SQD:
   if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
    err = 0;
    goto exit;
   }

   if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
       iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
    err = 0;
    goto exit;
   }

   if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
    err = -EINVAL;
    goto exit;
   }

   info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
   issue_modify_qp = 1;
   break;
  case IB_QPS_SQE:
   if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
    err = -EINVAL;
    goto exit;
   }

   info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
   issue_modify_qp = 1;
   break;
  case IB_QPS_ERR:
  case IB_QPS_RESET:
   if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
    spin_unlock_irqrestore(&iwqp->lock, flags);
    if (udata && udata->inlen) {
     if (ib_copy_from_udata(&ureq, udata,
         min(sizeof(ureq), udata->inlen)))
      return -EINVAL;

     irdma_flush_wqes(iwqp,
         (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
         (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
         IRDMA_REFLUSH);
    }
    return 0;
   }

   if (iwqp->sc_qp.term_flags) {
    spin_unlock_irqrestore(&iwqp->lock, flags);
    irdma_terminate_del_timer(&iwqp->sc_qp);
    spin_lock_irqsave(&iwqp->lock, flags);
   }
   info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
   if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
       iwdev->iw_status &&
       iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
    info.reset_tcp_conn = true;
   else
    dont_wait = 1;

   issue_modify_qp = 1;
   info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
   break;
  default:
   err = -EINVAL;
   goto exit;
  }

  iwqp->ibqp_state = attr->qp_state;
 }
 if (attr_mask & IB_QP_ACCESS_FLAGS) {
  ctx_info->iwarp_info_valid = true;
  if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
   offload_info->wr_rdresp_en = true;
  if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
   offload_info->wr_rdresp_en = true;
  if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
   offload_info->rd_en = true;
 }

 if (ctx_info->iwarp_info_valid) {
  ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
  ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
  irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 }
 spin_unlock_irqrestore(&iwqp->lock, flags);

 if (attr_mask & IB_QP_STATE) {
  if (issue_modify_qp) {
   ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
   if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
    return -EINVAL;
  }

  spin_lock_irqsave(&iwqp->lock, flags);
  if (iwqp->iwarp_state == info.curr_iwarp_state) {
   iwqp->iwarp_state = info.next_iwarp_state;
   iwqp->ibqp_state = attr->qp_state;
  }
  spin_unlock_irqrestore(&iwqp->lock, flags);
 }

 if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
  if (dont_wait) {
   if (iwqp->hw_tcp_state) {
    spin_lock_irqsave(&iwqp->lock, flags);
    iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
    iwqp->last_aeq = IRDMA_AE_RESET_SENT;
    spin_unlock_irqrestore(&iwqp->lock, flags);
   }
   irdma_cm_disconn(iwqp);
  } else {
   int close_timer_started;

   spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);

   if (iwqp->cm_node) {
    refcount_inc(&iwqp->cm_node->refcnt);
    spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
    close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
    if (iwqp->cm_id && close_timer_started == 1)
     irdma_schedule_cm_timer(iwqp->cm_node,
      (struct irdma_puda_buf *)iwqp,
      IRDMA_TIMER_TYPE_CLOSE, 1, 0);

    irdma_rem_ref_cm_node(iwqp->cm_node);
   } else {
    spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
   }
  }
 }
 if (attr_mask & IB_QP_STATE && udata && udata->outlen &&
     dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
  struct irdma_ucontext *ucontext;

  ucontext = rdma_udata_to_drv_context(udata,
     struct irdma_ucontext, ibucontext);
  if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
      !iwqp->push_wqe_mmap_entry &&
      !irdma_setup_push_mmap_entries(ucontext, iwqp,
   &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
   uresp.push_valid = 1;
   uresp.push_offset = iwqp->sc_qp.push_offset;
  }

  err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
           udata->outlen));
  if (err) {
   irdma_remove_push_mmap_entries(iwqp);
   ibdev_dbg(&iwdev->ibdev,
      "VERBS: copy_to_udata failed\n");
   return err;
  }
 }

 return 0;
exit:
 spin_unlock_irqrestore(&iwqp->lock, flags);

 return err;
}

/**
 * irdma_cq_free_rsrc - free up resources for cq
 * @rf: RDMA PCI function
 * @iwcq: cq ptr
 */
static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
{
 struct irdma_sc_cq *cq = &iwcq->sc_cq;

 if (!iwcq->user_mode) {
  dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
      iwcq->kmem.va, iwcq->kmem.pa);
  iwcq->kmem.va = NULL;
  dma_free_coherent(rf->sc_dev.hw->device,
      iwcq->kmem_shadow.size,
      iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
  iwcq->kmem_shadow.va = NULL;
 }

 irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
}

/**
 * irdma_free_cqbuf - worker to free a cq buffer
 * @work: provides access to the cq buffer to free
 */
static void irdma_free_cqbuf(struct work_struct *work)
{
 struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);

 dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
     cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
 cq_buf->kmem_buf.va = NULL;
 kfree(cq_buf);
}

/**
 * irdma_process_resize_list - remove resized cq buffers from the resize_list
 * @iwcq: cq which owns the resize_list
 * @iwdev: irdma device
 * @lcqe_buf: the buffer where the last cqe is received
 */
static int irdma_process_resize_list(struct irdma_cq *iwcq,
         struct irdma_device *iwdev,
         struct irdma_cq_buf *lcqe_buf)
{
 struct list_head *tmp_node, *list_node;
 struct irdma_cq_buf *cq_buf;
 int cnt = 0;

 list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
  cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
  if (cq_buf == lcqe_buf)
   return cnt;

  list_del(&cq_buf->list);
  queue_work(iwdev->cleanup_wq, &cq_buf->work);
  cnt++;
 }

 return cnt;
}

/**
 * irdma_destroy_cq - destroy cq
 * @ib_cq: cq pointer
 * @udata: user data
 */
static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
{
 struct irdma_device *iwdev = to_iwdev(ib_cq->device);
 struct irdma_cq *iwcq = to_iwcq(ib_cq);
 struct irdma_sc_cq *cq = &iwcq->sc_cq;
 struct irdma_sc_dev *dev = cq->dev;
 struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
 struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
 unsigned long flags;

 spin_lock_irqsave(&iwcq->lock, flags);
 if (!list_empty(&iwcq->cmpl_generated))
  irdma_remove_cmpls_list(iwcq);
 if (!list_empty(&iwcq->resize_list))
  irdma_process_resize_list(iwcq, iwdev, NULL);
 spin_unlock_irqrestore(&iwcq->lock, flags);

 irdma_cq_rem_ref(ib_cq);
 wait_for_completion(&iwcq->free_cq);

 irdma_cq_wq_destroy(iwdev->rf, cq);

 spin_lock_irqsave(&iwceq->ce_lock, flags);
 irdma_sc_cleanup_ceqes(cq, ceq);
 spin_unlock_irqrestore(&iwceq->ce_lock, flags);
 irdma_cq_free_rsrc(iwdev->rf, iwcq);

 return 0;
}

/**
 * irdma_resize_cq - resize cq
 * @ibcq: cq to be resized
 * @entries: desired cq size
 * @udata: user data
 */
static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
      struct ib_udata *udata)
{
#define IRDMA_RESIZE_CQ_MIN_REQ_LEN offsetofend(struct irdma_resize_cq_req, user_cq_buffer)
 struct irdma_cq *iwcq = to_iwcq(ibcq);
 struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
 struct irdma_cqp_request *cqp_request;
 struct cqp_cmds_info *cqp_info;
 struct irdma_modify_cq_info *m_info;
 struct irdma_modify_cq_info info = {};
 struct irdma_dma_mem kmem_buf;
 struct irdma_cq_mr *cqmr_buf;
 struct irdma_pbl *iwpbl_buf;
 struct irdma_device *iwdev;
 struct irdma_pci_f *rf;
 struct irdma_cq_buf *cq_buf = NULL;
 unsigned long flags;
 int ret;

 iwdev = to_iwdev(ibcq->device);
 rf = iwdev->rf;

 if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
     IRDMA_FEATURE_CQ_RESIZE))
  return -EOPNOTSUPP;

 if (udata && udata->inlen < IRDMA_RESIZE_CQ_MIN_REQ_LEN)
  return -EINVAL;

 if (entries > rf->max_cqe)
  return -EINVAL;

 if (!iwcq->user_mode) {
  entries++;
  if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
   entries *= 2;
 }

 info.cq_size = max(entries, 4);

 if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
  return 0;

 if (udata) {
  struct irdma_resize_cq_req req = {};
  struct irdma_ucontext *ucontext =
   rdma_udata_to_drv_context(udata, struct irdma_ucontext,
        ibucontext);

  /* CQ resize not supported with legacy GEN_1 libi40iw */
  if (ucontext->legacy_mode)
   return -EOPNOTSUPP;

  if (ib_copy_from_udata(&req, udata,
           min(sizeof(req), udata->inlen)))
   return -EINVAL;

  spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
  iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
       &ucontext->cq_reg_mem_list);
  spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);

  if (!iwpbl_buf)
   return -ENOMEM;

  cqmr_buf = &iwpbl_buf->cq_mr;
  if (iwpbl_buf->pbl_allocated) {
   info.virtual_map = true;
   info.pbl_chunk_size = 1;
   info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
  } else {
   info.cq_pa = cqmr_buf->cq_pbl.addr;
  }
 } else {
  /* Kmode CQ resize */
  int rsize;

  rsize = info.cq_size * sizeof(struct irdma_cqe);
  kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
  kmem_buf.va = dma_alloc_coherent(dev->hw->device,
       kmem_buf.size, &kmem_buf.pa,
       GFP_KERNEL);
  if (!kmem_buf.va)
   return -ENOMEM;

  info.cq_base = kmem_buf.va;
  info.cq_pa = kmem_buf.pa;
  cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
  if (!cq_buf) {
   ret = -ENOMEM;
   goto error;
  }
 }

 cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 if (!cqp_request) {
  ret = -ENOMEM;
  goto error;
 }

 info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
 info.cq_resize = true;

 cqp_info = &cqp_request->info;
 m_info = &cqp_info->in.u.cq_modify.info;
 memcpy(m_info, &info, sizeof(*m_info));

 cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
 cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
 cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
 cqp_info->post_sq = 1;
 ret = irdma_handle_cqp_op(rf, cqp_request);
 irdma_put_cqp_request(&rf->cqp, cqp_request);
 if (ret)
  goto error;

 spin_lock_irqsave(&iwcq->lock, flags);
 if (cq_buf) {
  cq_buf->kmem_buf = iwcq->kmem;
  cq_buf->hw = dev->hw;
  memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
  INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
  list_add_tail(&cq_buf->list, &iwcq->resize_list);
  iwcq->kmem = kmem_buf;
 }

 irdma_sc_cq_resize(&iwcq->sc_cq, &info);
 ibcq->cqe = info.cq_size - 1;
 spin_unlock_irqrestore(&iwcq->lock, flags);

 return 0;
error:
 if (!udata) {
  dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
      kmem_buf.pa);
  kmem_buf.va = NULL;
 }
 kfree(cq_buf);

 return ret;
}

static inline int cq_validate_flags(u32 flags, u8 hw_rev)
{
 /* GEN1 does not support CQ create flags */
 if (hw_rev == IRDMA_GEN_1)
  return flags ? -EOPNOTSUPP : 0;

 return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
}

/**
 * irdma_create_cq - create cq
 * @ibcq: CQ allocated
 * @attr: attributes for cq
 * @attrs: uverbs attribute bundle
 */
static int irdma_create_cq(struct ib_cq *ibcq,
      const struct ib_cq_init_attr *attr,
      struct uverbs_attr_bundle *attrs)
{
#define IRDMA_CREATE_CQ_MIN_REQ_LEN offsetofend(struct irdma_create_cq_req, user_cq_buf)
#define IRDMA_CREATE_CQ_MIN_RESP_LEN offsetofend(struct irdma_create_cq_resp, cq_size)
 struct ib_udata *udata = &attrs->driver_udata;
 struct ib_device *ibdev = ibcq->device;
 struct irdma_device *iwdev = to_iwdev(ibdev);
 struct irdma_pci_f *rf = iwdev->rf;
 struct irdma_cq *iwcq = to_iwcq(ibcq);
 u32 cq_num = 0;
 struct irdma_sc_cq *cq;
 struct irdma_sc_dev *dev = &rf->sc_dev;
 struct irdma_cq_init_info info = {};
 struct irdma_cqp_request *cqp_request;
 struct cqp_cmds_info *cqp_info;
 struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
 unsigned long flags;
 int err_code;
 int entries = attr->cqe;

 err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
 if (err_code)
  return err_code;

 if (udata && (udata->inlen < IRDMA_CREATE_CQ_MIN_REQ_LEN ||
        udata->outlen < IRDMA_CREATE_CQ_MIN_RESP_LEN))
  return -EINVAL;

 err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
        &rf->next_cq);
 if (err_code)
  return err_code;

 cq = &iwcq->sc_cq;
 cq->back_cq = iwcq;
 refcount_set(&iwcq->refcnt, 1);
 spin_lock_init(&iwcq->lock);
 INIT_LIST_HEAD(&iwcq->resize_list);
 INIT_LIST_HEAD(&iwcq->cmpl_generated);
 iwcq->cq_num = cq_num;
 info.dev = dev;
 ukinfo->cq_size = max(entries, 4);
 ukinfo->cq_id = cq_num;
 iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
 if (attr->comp_vector < rf->ceqs_count)
  info.ceq_id = attr->comp_vector;
 info.ceq_id_valid = true;
 info.ceqe_mask = 1;
 info.type = IRDMA_CQ_TYPE_IWARP;
 info.vsi = &iwdev->vsi;

 if (udata) {
  struct irdma_ucontext *ucontext;
  struct irdma_create_cq_req req = {};
  struct irdma_cq_mr *cqmr;
  struct irdma_pbl *iwpbl;
  struct irdma_pbl *iwpbl_shadow;
  struct irdma_cq_mr *cqmr_shadow;

  iwcq->user_mode = true;
  ucontext =
   rdma_udata_to_drv_context(udata, struct irdma_ucontext,
        ibucontext);
  if (ib_copy_from_udata(&req, udata,
           min(sizeof(req), udata->inlen))) {
   err_code = -EFAULT;
   goto cq_free_rsrc;
  }

  spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
  iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
          &ucontext->cq_reg_mem_list);
  spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
  if (!iwpbl) {
   err_code = -EPROTO;
   goto cq_free_rsrc;
  }

  cqmr = &iwpbl->cq_mr;

  if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
      IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
   spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
   iwpbl_shadow = irdma_get_pbl(
     (unsigned long)req.user_shadow_area,
     &ucontext->cq_reg_mem_list);
   spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);

   if (!iwpbl_shadow) {
    err_code = -EPROTO;
    goto cq_free_rsrc;
   }
   cqmr_shadow = &iwpbl_shadow->cq_mr;
   info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
   cqmr->split = true;
  } else {
   info.shadow_area_pa = cqmr->shadow;
  }
  if (iwpbl->pbl_allocated) {
   info.virtual_map = true;
   info.pbl_chunk_size = 1;
   info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
  } else {
   info.cq_base_pa = cqmr->cq_pbl.addr;
  }
 } else {
  /* Kmode allocations */
  int rsize;

  if (entries < 1 || entries > rf->max_cqe) {
   err_code = -EINVAL;
   goto cq_free_rsrc;
  }

  entries++;
  if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
   entries *= 2;
  ukinfo->cq_size = entries;

  rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
  iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
  iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
         iwcq->kmem.size,
         &iwcq->kmem.pa, GFP_KERNEL);
  if (!iwcq->kmem.va) {
   err_code = -ENOMEM;
   goto cq_free_rsrc;
  }

  iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
            64);
  iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
         iwcq->kmem_shadow.size,
         &iwcq->kmem_shadow.pa,
         GFP_KERNEL);
  if (!iwcq->kmem_shadow.va) {
   err_code = -ENOMEM;
   goto cq_free_rsrc;
  }
  info.shadow_area_pa = iwcq->kmem_shadow.pa;
  ukinfo->shadow_area = iwcq->kmem_shadow.va;
  ukinfo->cq_base = iwcq->kmem.va;
  info.cq_base_pa = iwcq->kmem.pa;
 }

 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
      (u32)IRDMA_MAX_CQ_READ_THRESH);

 if (irdma_sc_cq_init(cq, &info)) {
  ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
  err_code = -EPROTO;
  goto cq_free_rsrc;
 }

 cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 if (!cqp_request) {
  err_code = -ENOMEM;
  goto cq_free_rsrc;
 }

 cqp_info = &cqp_request->info;
 cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
 cqp_info->post_sq = 1;
 cqp_info->in.u.cq_create.cq = cq;
 cqp_info->in.u.cq_create.check_overflow = true;
 cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
 err_code = irdma_handle_cqp_op(rf, cqp_request);
 irdma_put_cqp_request(&rf->cqp, cqp_request);
 if (err_code)
  goto cq_free_rsrc;

 if (udata) {
  struct irdma_create_cq_resp resp = {};

  resp.cq_id = info.cq_uk_init_info.cq_id;
  resp.cq_size = info.cq_uk_init_info.cq_size;
  if (ib_copy_to_udata(udata, &resp,
         min(sizeof(resp), udata->outlen))) {
   ibdev_dbg(&iwdev->ibdev,
      "VERBS: copy to user data\n");
   err_code = -EPROTO;
   goto cq_destroy;
  }
 }
 rf->cq_table[cq_num] = iwcq;
 init_completion(&iwcq->free_cq);

 return 0;
cq_destroy:
 irdma_cq_wq_destroy(rf, cq);
cq_free_rsrc:
 irdma_cq_free_rsrc(rf, iwcq);

 return err_code;
}

/**
 * irdma_get_mr_access - get hw MR access permissions from IB access flags
 * @access: IB access flags
 */
static inline u16 irdma_get_mr_access(int access)
{
 u16 hw_access = 0;

 hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
       IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
 hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
       IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
--> --------------------

--> maximum size reached

--> --------------------