Quelle  rdma.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * NVMe over Fabrics RDMA target.
 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
 */
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/atomic.h>
#include <linux/blk-integrity.h>
#include <linux/ctype.h>
#include <linux/delay.h>
#include <linux/err.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/nvme.h>
#include <linux/slab.h>
#include <linux/string.h>
#include <linux/wait.h>
#include <linux/inet.h>
#include <linux/unaligned.h>

#include <rdma/ib_verbs.h>
#include <rdma/rdma_cm.h>
#include <rdma/rw.h>
#include <rdma/ib_cm.h>

#include <linux/nvme-rdma.h>
#include "nvmet.h"

/*
 * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
 */
#define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE PAGE_SIZE
#define NVMET_RDMA_MAX_INLINE_SGE  4
#define NVMET_RDMA_MAX_INLINE_DATA_SIZE  max_t(int, SZ_16K, PAGE_SIZE)

/* Assume mpsmin == device_page_size == 4KB */
#define NVMET_RDMA_MAX_MDTS   8
#define NVMET_RDMA_MAX_METADATA_MDTS  5

#define NVMET_RDMA_BACKLOG 128

#define NVMET_RDMA_DISCRETE_RSP_TAG  -1

struct nvmet_rdma_srq;

struct nvmet_rdma_cmd {
 struct ib_sge  sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
 struct ib_cqe  cqe;
 struct ib_recv_wr wr;
 struct scatterlist inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
 struct nvme_command     *nvme_cmd;
 struct nvmet_rdma_queue *queue;
 struct nvmet_rdma_srq   *nsrq;
};

enum {
 NVMET_RDMA_REQ_INLINE_DATA = (1 << 0),
};

struct nvmet_rdma_rsp {
 struct ib_sge  send_sge;
 struct ib_cqe  send_cqe;
 struct ib_send_wr send_wr;

 struct nvmet_rdma_cmd *cmd;
 struct nvmet_rdma_queue *queue;

 struct ib_cqe  read_cqe;
 struct ib_cqe  write_cqe;
 struct rdma_rw_ctx rw;

 struct nvmet_req req;

 bool   allocated;
 u8   n_rdma;
 u32   flags;
 u32   invalidate_rkey;

 struct list_head wait_list;
 int   tag;
};

enum nvmet_rdma_queue_state {
 NVMET_RDMA_Q_CONNECTING,
 NVMET_RDMA_Q_LIVE,
 NVMET_RDMA_Q_DISCONNECTING,
};

struct nvmet_rdma_queue {
 struct rdma_cm_id *cm_id;
 struct ib_qp  *qp;
 struct nvmet_port *port;
 struct ib_cq  *cq;
 atomic_t  sq_wr_avail;
 struct nvmet_rdma_device *dev;
 struct nvmet_rdma_srq   *nsrq;
 spinlock_t  state_lock;
 enum nvmet_rdma_queue_state state;
 struct nvmet_cq  nvme_cq;
 struct nvmet_sq  nvme_sq;

 struct nvmet_rdma_rsp *rsps;
 struct sbitmap  rsp_tags;
 struct nvmet_rdma_cmd *cmds;

 struct work_struct release_work;
 struct list_head rsp_wait_list;
 struct list_head rsp_wr_wait_list;
 spinlock_t  rsp_wr_wait_lock;

 int   idx;
 int   host_qid;
 int   comp_vector;
 int   recv_queue_size;
 int   send_queue_size;

 struct list_head queue_list;
};

struct nvmet_rdma_port {
 struct nvmet_port *nport;
 struct sockaddr_storage addr;
 struct rdma_cm_id *cm_id;
 struct delayed_work repair_work;
};

struct nvmet_rdma_srq {
 struct ib_srq            *srq;
 struct nvmet_rdma_cmd    *cmds;
 struct nvmet_rdma_device *ndev;
};

struct nvmet_rdma_device {
 struct ib_device *device;
 struct ib_pd  *pd;
 struct nvmet_rdma_srq **srqs;
 int   srq_count;
 size_t   srq_size;
 struct kref  ref;
 struct list_head entry;
 int   inline_data_size;
 int   inline_page_count;
};

static bool nvmet_rdma_use_srq;
module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
MODULE_PARM_DESC(use_srq, "Use shared receive queue.");

static int srq_size_set(const char *val, const struct kernel_param *kp);
static const struct kernel_param_ops srq_size_ops = {
 .set = srq_size_set,
 .get = param_get_int,
};

static int nvmet_rdma_srq_size = 1024;
module_param_cb(srq_size, &srq_size_ops, &nvmet_rdma_srq_size, 0644);
MODULE_PARM_DESC(srq_size, "set Shared Receive Queue (SRQ) size, should >= 256 (default: 1024)");

static DEFINE_IDA(nvmet_rdma_queue_ida);
static LIST_HEAD(nvmet_rdma_queue_list);
static DEFINE_MUTEX(nvmet_rdma_queue_mutex);

static LIST_HEAD(device_list);
static DEFINE_MUTEX(device_list_mutex);

static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc);
static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
    struct nvmet_rdma_rsp *r);
static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
    struct nvmet_rdma_rsp *r,
    int tag);

static const struct nvmet_fabrics_ops nvmet_rdma_ops;

static int srq_size_set(const char *val, const struct kernel_param *kp)
{
 int n = 0, ret;

 ret = kstrtoint(val, 10, &n);
 if (ret != 0 || n < 256)
  return -EINVAL;

 return param_set_int(val, kp);
}

static int num_pages(int len)
{
 return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
}

static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
{
 return nvme_is_write(rsp->req.cmd) &&
  rsp->req.transfer_len &&
  !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
}

static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
{
 return !nvme_is_write(rsp->req.cmd) &&
  rsp->req.transfer_len &&
  !rsp->req.cqe->status &&
  !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
}

static inline struct nvmet_rdma_rsp *
nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
{
 struct nvmet_rdma_rsp *rsp = NULL;
 int tag;

 tag = sbitmap_get(&queue->rsp_tags);
 if (tag >= 0)
  rsp = &queue->rsps[tag];

 if (unlikely(!rsp)) {
  int ret;

  rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
  if (unlikely(!rsp))
   return NULL;
  ret = nvmet_rdma_alloc_rsp(queue->dev, rsp,
    NVMET_RDMA_DISCRETE_RSP_TAG);
  if (unlikely(ret)) {
   kfree(rsp);
   return NULL;
  }
 }

 return rsp;
}

static inline void
nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
{
 if (unlikely(rsp->tag == NVMET_RDMA_DISCRETE_RSP_TAG)) {
  nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
  kfree(rsp);
  return;
 }

 sbitmap_clear_bit(&rsp->queue->rsp_tags, rsp->tag);
}

static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
    struct nvmet_rdma_cmd *c)
{
 struct scatterlist *sg;
 struct ib_sge *sge;
 int i;

 if (!ndev->inline_data_size)
  return;

 sg = c->inline_sg;
 sge = &c->sge[1];

 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
  if (sge->length)
   ib_dma_unmap_page(ndev->device, sge->addr,
     sge->length, DMA_FROM_DEVICE);
  if (sg_page(sg))
   __free_page(sg_page(sg));
 }
}

static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
    struct nvmet_rdma_cmd *c)
{
 struct scatterlist *sg;
 struct ib_sge *sge;
 struct page *pg;
 int len;
 int i;

 if (!ndev->inline_data_size)
  return 0;

 sg = c->inline_sg;
 sg_init_table(sg, ndev->inline_page_count);
 sge = &c->sge[1];
 len = ndev->inline_data_size;

 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
  pg = alloc_page(GFP_KERNEL);
  if (!pg)
   goto out_err;
  sg_assign_page(sg, pg);
  sge->addr = ib_dma_map_page(ndev->device,
   pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
  if (ib_dma_mapping_error(ndev->device, sge->addr))
   goto out_err;
  sge->length = min_t(int, len, PAGE_SIZE);
  sge->lkey = ndev->pd->local_dma_lkey;
  len -= sge->length;
 }

 return 0;
out_err:
 for (; i >= 0; i--, sg--, sge--) {
  if (sge->length)
   ib_dma_unmap_page(ndev->device, sge->addr,
     sge->length, DMA_FROM_DEVICE);
  if (sg_page(sg))
   __free_page(sg_page(sg));
 }
 return -ENOMEM;
}

static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
   struct nvmet_rdma_cmd *c, bool admin)
{
 /* NVMe command / RDMA RECV */
 c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 if (!c->nvme_cmd)
  goto out;

 c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
   sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
  goto out_free_cmd;

 c->sge[0].length = sizeof(*c->nvme_cmd);
 c->sge[0].lkey = ndev->pd->local_dma_lkey;

 if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
  goto out_unmap_cmd;

 c->cqe.done = nvmet_rdma_recv_done;

 c->wr.wr_cqe = &c->cqe;
 c->wr.sg_list = c->sge;
 c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;

 return 0;

out_unmap_cmd:
 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
   sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
out_free_cmd:
 kfree(c->nvme_cmd);

out:
 return -ENOMEM;
}

static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
  struct nvmet_rdma_cmd *c, bool admin)
{
 if (!admin)
  nvmet_rdma_free_inline_pages(ndev, c);
 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
    sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 kfree(c->nvme_cmd);
}

static struct nvmet_rdma_cmd *
nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
  int nr_cmds, bool admin)
{
 struct nvmet_rdma_cmd *cmds;
 int ret = -EINVAL, i;

 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 if (!cmds)
  goto out;

 for (i = 0; i < nr_cmds; i++) {
  ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
  if (ret)
   goto out_free;
 }

 return cmds;

out_free:
 while (--i >= 0)
  nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 kfree(cmds);
out:
 return ERR_PTR(ret);
}

static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
  struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
{
 int i;

 for (i = 0; i < nr_cmds; i++)
  nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 kfree(cmds);
}

static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
  struct nvmet_rdma_rsp *r, int tag)
{
 /* NVMe CQE / RDMA SEND */
 r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
 if (!r->req.cqe)
  goto out;

 r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
   sizeof(*r->req.cqe), DMA_TO_DEVICE);
 if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
  goto out_free_rsp;

 if (ib_dma_pci_p2p_dma_supported(ndev->device))
  r->req.p2p_client = &ndev->device->dev;
 r->send_sge.length = sizeof(*r->req.cqe);
 r->send_sge.lkey = ndev->pd->local_dma_lkey;

 r->send_cqe.done = nvmet_rdma_send_done;

 r->send_wr.wr_cqe = &r->send_cqe;
 r->send_wr.sg_list = &r->send_sge;
 r->send_wr.num_sge = 1;
 r->send_wr.send_flags = IB_SEND_SIGNALED;

 /* Data In / RDMA READ */
 r->read_cqe.done = nvmet_rdma_read_data_done;
 /* Data Out / RDMA WRITE */
 r->write_cqe.done = nvmet_rdma_write_data_done;
 r->tag = tag;

 return 0;

out_free_rsp:
 kfree(r->req.cqe);
out:
 return -ENOMEM;
}

static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
  struct nvmet_rdma_rsp *r)
{
 ib_dma_unmap_single(ndev->device, r->send_sge.addr,
    sizeof(*r->req.cqe), DMA_TO_DEVICE);
 kfree(r->req.cqe);
}

static int
nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
{
 struct nvmet_rdma_device *ndev = queue->dev;
 int nr_rsps = queue->recv_queue_size * 2;
 int ret = -ENOMEM, i;

 if (sbitmap_init_node(&queue->rsp_tags, nr_rsps, -1, GFP_KERNEL,
   NUMA_NO_NODE, false, true))
  goto out;

 queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
   GFP_KERNEL);
 if (!queue->rsps)
  goto out_free_sbitmap;

 for (i = 0; i < nr_rsps; i++) {
  struct nvmet_rdma_rsp *rsp = &queue->rsps[i];

  ret = nvmet_rdma_alloc_rsp(ndev, rsp, i);
  if (ret)
   goto out_free;
 }

 return 0;

out_free:
 while (--i >= 0)
  nvmet_rdma_free_rsp(ndev, &queue->rsps[i]);
 kfree(queue->rsps);
out_free_sbitmap:
 sbitmap_free(&queue->rsp_tags);
out:
 return ret;
}

static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
{
 struct nvmet_rdma_device *ndev = queue->dev;
 int i, nr_rsps = queue->recv_queue_size * 2;

 for (i = 0; i < nr_rsps; i++)
  nvmet_rdma_free_rsp(ndev, &queue->rsps[i]);
 kfree(queue->rsps);
 sbitmap_free(&queue->rsp_tags);
}

static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
  struct nvmet_rdma_cmd *cmd)
{
 int ret;

 ib_dma_sync_single_for_device(ndev->device,
  cmd->sge[0].addr, cmd->sge[0].length,
  DMA_FROM_DEVICE);

 if (cmd->nsrq)
  ret = ib_post_srq_recv(cmd->nsrq->srq, &cmd->wr, NULL);
 else
  ret = ib_post_recv(cmd->queue->qp, &cmd->wr, NULL);

 if (unlikely(ret))
  pr_err("post_recv cmd failed\n");

 return ret;
}

static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
{
 spin_lock(&queue->rsp_wr_wait_lock);
 while (!list_empty(&queue->rsp_wr_wait_list)) {
  struct nvmet_rdma_rsp *rsp;
  bool ret;

  rsp = list_entry(queue->rsp_wr_wait_list.next,
    struct nvmet_rdma_rsp, wait_list);
  list_del(&rsp->wait_list);

  spin_unlock(&queue->rsp_wr_wait_lock);
  ret = nvmet_rdma_execute_command(rsp);
  spin_lock(&queue->rsp_wr_wait_lock);

  if (!ret) {
   list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
   break;
  }
 }
 spin_unlock(&queue->rsp_wr_wait_lock);
}

static u16 nvmet_rdma_check_pi_status(struct ib_mr *sig_mr)
{
 struct ib_mr_status mr_status;
 int ret;
 u16 status = 0;

 ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
 if (ret) {
  pr_err("ib_check_mr_status failed, ret %d\n", ret);
  return NVME_SC_INVALID_PI;
 }

 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
  switch (mr_status.sig_err.err_type) {
  case IB_SIG_BAD_GUARD:
   status = NVME_SC_GUARD_CHECK;
   break;
  case IB_SIG_BAD_REFTAG:
   status = NVME_SC_REFTAG_CHECK;
   break;
  case IB_SIG_BAD_APPTAG:
   status = NVME_SC_APPTAG_CHECK;
   break;
  }
  pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
         mr_status.sig_err.err_type,
         mr_status.sig_err.expected,
         mr_status.sig_err.actual);
 }

 return status;
}

static void nvmet_rdma_set_sig_domain(struct blk_integrity *bi,
  struct nvme_command *cmd, struct ib_sig_domain *domain,
  u16 control, u8 pi_type)
{
 domain->sig_type = IB_SIG_TYPE_T10_DIF;
 domain->sig.dif.bg_type = IB_T10DIF_CRC;
 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
 if (control & NVME_RW_PRINFO_PRCHK_REF)
  domain->sig.dif.ref_remap = true;

 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.lbat);
 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.lbatm);
 domain->sig.dif.app_escape = true;
 if (pi_type == NVME_NS_DPS_PI_TYPE3)
  domain->sig.dif.ref_escape = true;
}

static void nvmet_rdma_set_sig_attrs(struct nvmet_req *req,
         struct ib_sig_attrs *sig_attrs)
{
 struct nvme_command *cmd = req->cmd;
 u16 control = le16_to_cpu(cmd->rw.control);
 u8 pi_type = req->ns->pi_type;
 struct blk_integrity *bi;

 bi = bdev_get_integrity(req->ns->bdev);

 memset(sig_attrs, 0, sizeof(*sig_attrs));

 if (control & NVME_RW_PRINFO_PRACT) {
  /* for WRITE_INSERT/READ_STRIP no wire domain */
  sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
  nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
       pi_type);
  /* Clear the PRACT bit since HCA will generate/verify the PI */
  control &= ~NVME_RW_PRINFO_PRACT;
  cmd->rw.control = cpu_to_le16(control);
  /* PI is added by the HW */
  req->transfer_len += req->metadata_len;
 } else {
  /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
  nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
       pi_type);
  nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
       pi_type);
 }

 if (control & NVME_RW_PRINFO_PRCHK_REF)
  sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
 if (control & NVME_RW_PRINFO_PRCHK_GUARD)
  sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
 if (control & NVME_RW_PRINFO_PRCHK_APP)
  sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
}

static int nvmet_rdma_rw_ctx_init(struct nvmet_rdma_rsp *rsp, u64 addr, u32 key,
      struct ib_sig_attrs *sig_attrs)
{
 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 struct nvmet_req *req = &rsp->req;
 int ret;

 if (req->metadata_len)
  ret = rdma_rw_ctx_signature_init(&rsp->rw, cm_id->qp,
   cm_id->port_num, req->sg, req->sg_cnt,
   req->metadata_sg, req->metadata_sg_cnt, sig_attrs,
   addr, key, nvmet_data_dir(req));
 else
  ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
           req->sg, req->sg_cnt, 0, addr, key,
           nvmet_data_dir(req));

 return ret;
}

static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
{
 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 struct nvmet_req *req = &rsp->req;

 if (req->metadata_len)
  rdma_rw_ctx_destroy_signature(&rsp->rw, cm_id->qp,
   cm_id->port_num, req->sg, req->sg_cnt,
   req->metadata_sg, req->metadata_sg_cnt,
   nvmet_data_dir(req));
 else
  rdma_rw_ctx_destroy(&rsp->rw, cm_id->qp, cm_id->port_num,
        req->sg, req->sg_cnt, nvmet_data_dir(req));
}

static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
{
 struct nvmet_rdma_queue *queue = rsp->queue;

 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);

 if (rsp->n_rdma)
  nvmet_rdma_rw_ctx_destroy(rsp);

 if (rsp->req.sg != rsp->cmd->inline_sg)
  nvmet_req_free_sgls(&rsp->req);

 if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
  nvmet_rdma_process_wr_wait_list(queue);

 nvmet_rdma_put_rsp(rsp);
}

static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
{
 if (queue->nvme_sq.ctrl) {
  nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 } else {
  /*
 * we didn't setup the controller yet in case
 * of admin connect error, just disconnect and
 * cleanup the queue
 */
  nvmet_rdma_queue_disconnect(queue);
 }
}

static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
{
 struct nvmet_rdma_rsp *rsp =
  container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 struct nvmet_rdma_queue *queue = wc->qp->qp_context;

 nvmet_rdma_release_rsp(rsp);

 if (unlikely(wc->status != IB_WC_SUCCESS &&
       wc->status != IB_WC_WR_FLUSH_ERR)) {
  pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
   wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
  nvmet_rdma_error_comp(queue);
 }
}

static void nvmet_rdma_queue_response(struct nvmet_req *req)
{
 struct nvmet_rdma_rsp *rsp =
  container_of(req, struct nvmet_rdma_rsp, req);
 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 struct ib_send_wr *first_wr;

 if (rsp->invalidate_rkey) {
  rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
  rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 } else {
  rsp->send_wr.opcode = IB_WR_SEND;
 }

 if (nvmet_rdma_need_data_out(rsp)) {
  if (rsp->req.metadata_len)
   first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
     cm_id->port_num, &rsp->write_cqe, NULL);
  else
   first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
     cm_id->port_num, NULL, &rsp->send_wr);
 } else {
  first_wr = &rsp->send_wr;
 }

 nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);

 ib_dma_sync_single_for_device(rsp->queue->dev->device,
  rsp->send_sge.addr, rsp->send_sge.length,
  DMA_TO_DEVICE);

 if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
  pr_err("sending cmd response failed\n");
  nvmet_rdma_release_rsp(rsp);
 }
}

static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
{
 struct nvmet_rdma_rsp *rsp =
  container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
 u16 status = 0;

 WARN_ON(rsp->n_rdma <= 0);
 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 rsp->n_rdma = 0;

 if (unlikely(wc->status != IB_WC_SUCCESS)) {
  nvmet_rdma_rw_ctx_destroy(rsp);
  nvmet_req_uninit(&rsp->req);
  nvmet_rdma_release_rsp(rsp);
  if (wc->status != IB_WC_WR_FLUSH_ERR) {
   pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
    wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
   nvmet_rdma_error_comp(queue);
  }
  return;
 }

 if (rsp->req.metadata_len)
  status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
 nvmet_rdma_rw_ctx_destroy(rsp);

 if (unlikely(status))
  nvmet_req_complete(&rsp->req, status);
 else
  rsp->req.execute(&rsp->req);
}

static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc)
{
 struct nvmet_rdma_rsp *rsp =
  container_of(wc->wr_cqe, struct nvmet_rdma_rsp, write_cqe);
 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 u16 status;

 if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
  return;

 WARN_ON(rsp->n_rdma <= 0);
 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 rsp->n_rdma = 0;

 if (unlikely(wc->status != IB_WC_SUCCESS)) {
  nvmet_rdma_rw_ctx_destroy(rsp);
  nvmet_req_uninit(&rsp->req);
  nvmet_rdma_release_rsp(rsp);
  if (wc->status != IB_WC_WR_FLUSH_ERR) {
   pr_info("RDMA WRITE for CQE failed with status %s (%d).\n",
    ib_wc_status_msg(wc->status), wc->status);
   nvmet_rdma_error_comp(queue);
  }
  return;
 }

 /*
 * Upon RDMA completion check the signature status
 * - if succeeded send good NVMe response
 * - if failed send bad NVMe response with appropriate error
 */
 status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
 if (unlikely(status))
  rsp->req.cqe->status = cpu_to_le16(status << 1);
 nvmet_rdma_rw_ctx_destroy(rsp);

 if (unlikely(ib_post_send(cm_id->qp, &rsp->send_wr, NULL))) {
  pr_err("sending cmd response failed\n");
  nvmet_rdma_release_rsp(rsp);
 }
}

static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
  u64 off)
{
 int sg_count = num_pages(len);
 struct scatterlist *sg;
 int i;

 sg = rsp->cmd->inline_sg;
 for (i = 0; i < sg_count; i++, sg++) {
  if (i < sg_count - 1)
   sg_unmark_end(sg);
  else
   sg_mark_end(sg);
  sg->offset = off;
  sg->length = min_t(int, len, PAGE_SIZE - off);
  len -= sg->length;
  if (!i)
   off = 0;
 }

 rsp->req.sg = rsp->cmd->inline_sg;
 rsp->req.sg_cnt = sg_count;
}

static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
{
 struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 u64 off = le64_to_cpu(sgl->addr);
 u32 len = le32_to_cpu(sgl->length);

 if (!nvme_is_write(rsp->req.cmd)) {
  rsp->req.error_loc =
   offsetof(struct nvme_common_command, opcode);
  return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
 }

 if (off + len > rsp->queue->dev->inline_data_size) {
  pr_err("invalid inline data offset!\n");
  return NVME_SC_SGL_INVALID_OFFSET | NVME_STATUS_DNR;
 }

 /* no data command? */
 if (!len)
  return 0;

 nvmet_rdma_use_inline_sg(rsp, len, off);
 rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 rsp->req.transfer_len += len;
 return 0;
}

static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
  struct nvme_keyed_sgl_desc *sgl, bool invalidate)
{
 u64 addr = le64_to_cpu(sgl->addr);
 u32 key = get_unaligned_le32(sgl->key);
 struct ib_sig_attrs sig_attrs;
 int ret;

 rsp->req.transfer_len = get_unaligned_le24(sgl->length);

 /* no data command? */
 if (!rsp->req.transfer_len)
  return 0;

 if (rsp->req.metadata_len)
  nvmet_rdma_set_sig_attrs(&rsp->req, &sig_attrs);

 ret = nvmet_req_alloc_sgls(&rsp->req);
 if (unlikely(ret < 0))
  goto error_out;

 ret = nvmet_rdma_rw_ctx_init(rsp, addr, key, &sig_attrs);
 if (unlikely(ret < 0))
  goto error_out;
 rsp->n_rdma += ret;

 if (invalidate)
  rsp->invalidate_rkey = key;

 return 0;

error_out:
 rsp->req.transfer_len = 0;
 return NVME_SC_INTERNAL;
}

static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
{
 struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;

 switch (sgl->type >> 4) {
 case NVME_SGL_FMT_DATA_DESC:
  switch (sgl->type & 0xf) {
  case NVME_SGL_FMT_OFFSET:
   return nvmet_rdma_map_sgl_inline(rsp);
  default:
   pr_err("invalid SGL subtype: %#x\n", sgl->type);
   rsp->req.error_loc =
    offsetof(struct nvme_common_command, dptr);
   return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
  }
 case NVME_KEY_SGL_FMT_DATA_DESC:
  switch (sgl->type & 0xf) {
  case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
   return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
  case NVME_SGL_FMT_ADDRESS:
   return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
  default:
   pr_err("invalid SGL subtype: %#x\n", sgl->type);
   rsp->req.error_loc =
    offsetof(struct nvme_common_command, dptr);
   return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
  }
 default:
  pr_err("invalid SGL type: %#x\n", sgl->type);
  rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
  return NVME_SC_SGL_INVALID_TYPE | NVME_STATUS_DNR;
 }
}

static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
{
 struct nvmet_rdma_queue *queue = rsp->queue;

 if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
   &queue->sq_wr_avail) < 0)) {
  pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
    1 + rsp->n_rdma, queue->idx,
    queue->nvme_sq.ctrl->cntlid);
  atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
  return false;
 }

 if (nvmet_rdma_need_data_in(rsp)) {
  if (rdma_rw_ctx_post(&rsp->rw, queue->qp,
    queue->cm_id->port_num, &rsp->read_cqe, NULL))
   nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 } else {
  rsp->req.execute(&rsp->req);
 }

 return true;
}

static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
  struct nvmet_rdma_rsp *cmd)
{
 u16 status;

 ib_dma_sync_single_for_cpu(queue->dev->device,
  cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
  DMA_FROM_DEVICE);
 ib_dma_sync_single_for_cpu(queue->dev->device,
  cmd->send_sge.addr, cmd->send_sge.length,
  DMA_TO_DEVICE);

 if (!nvmet_req_init(&cmd->req, &queue->nvme_sq, &nvmet_rdma_ops))
  return;

 status = nvmet_rdma_map_sgl(cmd);
 if (status)
  goto out_err;

 if (unlikely(!nvmet_rdma_execute_command(cmd))) {
  spin_lock(&queue->rsp_wr_wait_lock);
  list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
  spin_unlock(&queue->rsp_wr_wait_lock);
 }

 return;

out_err:
 nvmet_req_complete(&cmd->req, status);
}

static bool nvmet_rdma_recv_not_live(struct nvmet_rdma_queue *queue,
  struct nvmet_rdma_rsp *rsp)
{
 unsigned long flags;
 bool ret = true;

 spin_lock_irqsave(&queue->state_lock, flags);
 /*
 * recheck queue state is not live to prevent a race condition
 * with RDMA_CM_EVENT_ESTABLISHED handler.
 */
 if (queue->state == NVMET_RDMA_Q_LIVE)
  ret = false;
 else if (queue->state == NVMET_RDMA_Q_CONNECTING)
  list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 else
  nvmet_rdma_put_rsp(rsp);
 spin_unlock_irqrestore(&queue->state_lock, flags);
 return ret;
}

static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
{
 struct nvmet_rdma_cmd *cmd =
  container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
 struct nvmet_rdma_rsp *rsp;

 if (unlikely(wc->status != IB_WC_SUCCESS)) {
  if (wc->status != IB_WC_WR_FLUSH_ERR) {
   pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
    wc->wr_cqe, ib_wc_status_msg(wc->status),
    wc->status);
   nvmet_rdma_error_comp(queue);
  }
  return;
 }

 if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
  pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
  nvmet_rdma_error_comp(queue);
  return;
 }

 cmd->queue = queue;
 rsp = nvmet_rdma_get_rsp(queue);
 if (unlikely(!rsp)) {
  /*
 * we get here only under memory pressure,
 * silently drop and have the host retry
 * as we can't even fail it.
 */
  nvmet_rdma_post_recv(queue->dev, cmd);
  return;
 }
 rsp->queue = queue;
 rsp->cmd = cmd;
 rsp->flags = 0;
 rsp->req.cmd = cmd->nvme_cmd;
 rsp->req.port = queue->port;
 rsp->n_rdma = 0;
 rsp->invalidate_rkey = 0;

 if (unlikely(queue->state != NVMET_RDMA_Q_LIVE) &&
     nvmet_rdma_recv_not_live(queue, rsp))
  return;

 nvmet_rdma_handle_command(queue, rsp);
}

static void nvmet_rdma_destroy_srq(struct nvmet_rdma_srq *nsrq)
{
 nvmet_rdma_free_cmds(nsrq->ndev, nsrq->cmds, nsrq->ndev->srq_size,
        false);
 ib_destroy_srq(nsrq->srq);

 kfree(nsrq);
}

static void nvmet_rdma_destroy_srqs(struct nvmet_rdma_device *ndev)
{
 int i;

 if (!ndev->srqs)
  return;

 for (i = 0; i < ndev->srq_count; i++)
  nvmet_rdma_destroy_srq(ndev->srqs[i]);

 kfree(ndev->srqs);
}

static struct nvmet_rdma_srq *
nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
{
 struct ib_srq_init_attr srq_attr = { NULL, };
 size_t srq_size = ndev->srq_size;
 struct nvmet_rdma_srq *nsrq;
 struct ib_srq *srq;
 int ret, i;

 nsrq = kzalloc(sizeof(*nsrq), GFP_KERNEL);
 if (!nsrq)
  return ERR_PTR(-ENOMEM);

 srq_attr.attr.max_wr = srq_size;
 srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
 srq_attr.attr.srq_limit = 0;
 srq_attr.srq_type = IB_SRQT_BASIC;
 srq = ib_create_srq(ndev->pd, &srq_attr);
 if (IS_ERR(srq)) {
  ret = PTR_ERR(srq);
  goto out_free;
 }

 nsrq->cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 if (IS_ERR(nsrq->cmds)) {
  ret = PTR_ERR(nsrq->cmds);
  goto out_destroy_srq;
 }

 nsrq->srq = srq;
 nsrq->ndev = ndev;

 for (i = 0; i < srq_size; i++) {
  nsrq->cmds[i].nsrq = nsrq;
  ret = nvmet_rdma_post_recv(ndev, &nsrq->cmds[i]);
  if (ret)
   goto out_free_cmds;
 }

 return nsrq;

out_free_cmds:
 nvmet_rdma_free_cmds(ndev, nsrq->cmds, srq_size, false);
out_destroy_srq:
 ib_destroy_srq(srq);
out_free:
 kfree(nsrq);
 return ERR_PTR(ret);
}

static int nvmet_rdma_init_srqs(struct nvmet_rdma_device *ndev)
{
 int i, ret;

 if (!ndev->device->attrs.max_srq_wr || !ndev->device->attrs.max_srq) {
  /*
 * If SRQs aren't supported we just go ahead and use normal
 * non-shared receive queues.
 */
  pr_info("SRQ requested but not supported.\n");
  return 0;
 }

 ndev->srq_size = min(ndev->device->attrs.max_srq_wr,
        nvmet_rdma_srq_size);
 ndev->srq_count = min(ndev->device->num_comp_vectors,
         ndev->device->attrs.max_srq);

 ndev->srqs = kcalloc(ndev->srq_count, sizeof(*ndev->srqs), GFP_KERNEL);
 if (!ndev->srqs)
  return -ENOMEM;

 for (i = 0; i < ndev->srq_count; i++) {
  ndev->srqs[i] = nvmet_rdma_init_srq(ndev);
  if (IS_ERR(ndev->srqs[i])) {
   ret = PTR_ERR(ndev->srqs[i]);
   goto err_srq;
  }
 }

 return 0;

err_srq:
 while (--i >= 0)
  nvmet_rdma_destroy_srq(ndev->srqs[i]);
 kfree(ndev->srqs);
 return ret;
}

static void nvmet_rdma_free_dev(struct kref *ref)
{
 struct nvmet_rdma_device *ndev =
  container_of(ref, struct nvmet_rdma_device, ref);

 mutex_lock(&device_list_mutex);
 list_del(&ndev->entry);
 mutex_unlock(&device_list_mutex);

 nvmet_rdma_destroy_srqs(ndev);
 ib_dealloc_pd(ndev->pd);

 kfree(ndev);
}

static struct nvmet_rdma_device *
nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
{
 struct nvmet_rdma_port *port = cm_id->context;
 struct nvmet_port *nport = port->nport;
 struct nvmet_rdma_device *ndev;
 int inline_page_count;
 int inline_sge_count;
 int ret;

 mutex_lock(&device_list_mutex);
 list_for_each_entry(ndev, &device_list, entry) {
  if (ndev->device->node_guid == cm_id->device->node_guid &&
      kref_get_unless_zero(&ndev->ref))
   goto out_unlock;
 }

 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 if (!ndev)
  goto out_err;

 inline_page_count = num_pages(nport->inline_data_size);
 inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
    cm_id->device->attrs.max_recv_sge) - 1;
 if (inline_page_count > inline_sge_count) {
  pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
   nport->inline_data_size, cm_id->device->name,
   inline_sge_count * PAGE_SIZE);
  nport->inline_data_size = inline_sge_count * PAGE_SIZE;
  inline_page_count = inline_sge_count;
 }
 ndev->inline_data_size = nport->inline_data_size;
 ndev->inline_page_count = inline_page_count;

 if (nport->pi_enable && !(cm_id->device->attrs.kernel_cap_flags &
      IBK_INTEGRITY_HANDOVER)) {
  pr_warn("T10-PI is not supported by device %s. Disabling it\n",
   cm_id->device->name);
  nport->pi_enable = false;
 }

 ndev->device = cm_id->device;
 kref_init(&ndev->ref);

 ndev->pd = ib_alloc_pd(ndev->device, 0);
 if (IS_ERR(ndev->pd))
  goto out_free_dev;

 if (nvmet_rdma_use_srq) {
  ret = nvmet_rdma_init_srqs(ndev);
  if (ret)
   goto out_free_pd;
 }

 list_add(&ndev->entry, &device_list);
out_unlock:
 mutex_unlock(&device_list_mutex);
 pr_debug("added %s.\n", ndev->device->name);
 return ndev;

out_free_pd:
 ib_dealloc_pd(ndev->pd);
out_free_dev:
 kfree(ndev);
out_err:
 mutex_unlock(&device_list_mutex);
 return NULL;
}

static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
{
 struct ib_qp_init_attr qp_attr = { };
 struct nvmet_rdma_device *ndev = queue->dev;
 int nr_cqe, ret, i, factor;

 /*
 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 */
 nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;

 queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
       queue->comp_vector, IB_POLL_WORKQUEUE);
 if (IS_ERR(queue->cq)) {
  ret = PTR_ERR(queue->cq);
  pr_err("failed to create CQ cqe= %d ret= %d\n",
         nr_cqe + 1, ret);
  goto out;
 }

 qp_attr.qp_context = queue;
 qp_attr.event_handler = nvmet_rdma_qp_event;
 qp_attr.send_cq = queue->cq;
 qp_attr.recv_cq = queue->cq;
 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 qp_attr.qp_type = IB_QPT_RC;
 /* +1 for drain */
 qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 factor = rdma_rw_mr_factor(ndev->device, queue->cm_id->port_num,
       1 << NVMET_RDMA_MAX_MDTS);
 qp_attr.cap.max_rdma_ctxs = queue->send_queue_size * factor;
 qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
     ndev->device->attrs.max_send_sge);

 if (queue->nsrq) {
  qp_attr.srq = queue->nsrq->srq;
 } else {
  /* +1 for drain */
  qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
  qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
 }

 if (queue->port->pi_enable && queue->host_qid)
  qp_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;

 ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
 if (ret) {
  pr_err("failed to create_qp ret= %d\n", ret);
  goto err_destroy_cq;
 }
 queue->qp = queue->cm_id->qp;

 atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);

 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
   __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
   qp_attr.cap.max_send_wr, queue->cm_id);

 if (!queue->nsrq) {
  for (i = 0; i < queue->recv_queue_size; i++) {
   queue->cmds[i].queue = queue;
   ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
   if (ret)
    goto err_destroy_qp;
  }
 }

out:
 return ret;

err_destroy_qp:
 rdma_destroy_qp(queue->cm_id);
err_destroy_cq:
 ib_cq_pool_put(queue->cq, nr_cqe + 1);
 goto out;
}

static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
{
 ib_drain_qp(queue->qp);
 if (queue->cm_id)
  rdma_destroy_id(queue->cm_id);
 ib_destroy_qp(queue->qp);
 ib_cq_pool_put(queue->cq, queue->recv_queue_size + 2 *
         queue->send_queue_size + 1);
}

static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
{
 pr_debug("freeing queue %d\n", queue->idx);

 nvmet_sq_destroy(&queue->nvme_sq);
 nvmet_cq_put(&queue->nvme_cq);

 nvmet_rdma_destroy_queue_ib(queue);
 if (!queue->nsrq) {
  nvmet_rdma_free_cmds(queue->dev, queue->cmds,
    queue->recv_queue_size,
    !queue->host_qid);
 }
 nvmet_rdma_free_rsps(queue);
 ida_free(&nvmet_rdma_queue_ida, queue->idx);
 kfree(queue);
}

static void nvmet_rdma_release_queue_work(struct work_struct *w)
{
 struct nvmet_rdma_queue *queue =
  container_of(w, struct nvmet_rdma_queue, release_work);
 struct nvmet_rdma_device *dev = queue->dev;

 nvmet_rdma_free_queue(queue);

 kref_put(&dev->ref, nvmet_rdma_free_dev);
}

static int
nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
    struct nvmet_rdma_queue *queue)
{
 struct nvme_rdma_cm_req *req;

 req = (struct nvme_rdma_cm_req *)conn->private_data;
 if (!req || conn->private_data_len == 0)
  return NVME_RDMA_CM_INVALID_LEN;

 if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
  return NVME_RDMA_CM_INVALID_RECFMT;

 queue->host_qid = le16_to_cpu(req->qid);

 /*
 * req->hsqsize corresponds to our recv queue size plus 1
 * req->hrqsize corresponds to our send queue size
 */
 queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
 queue->send_queue_size = le16_to_cpu(req->hrqsize);

 if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
  return NVME_RDMA_CM_INVALID_HSQSIZE;

 /* XXX: Should we enforce some kind of max for IO queues? */

 return 0;
}

static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
    enum nvme_rdma_cm_status status)
{
 struct nvme_rdma_cm_rej rej;

 pr_debug("rejecting connect request: status %d (%s)\n",
   status, nvme_rdma_cm_msg(status));

 rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
 rej.sts = cpu_to_le16(status);

 return rdma_reject(cm_id, (void *)&rej, sizeof(rej),
      IB_CM_REJ_CONSUMER_DEFINED);
}

static struct nvmet_rdma_queue *
nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
  struct rdma_cm_id *cm_id,
  struct rdma_cm_event *event)
{
 struct nvmet_rdma_port *port = cm_id->context;
 struct nvmet_rdma_queue *queue;
 int ret;

 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
 if (!queue) {
  ret = NVME_RDMA_CM_NO_RSC;
  goto out_reject;
 }

 nvmet_cq_init(&queue->nvme_cq);
 ret = nvmet_sq_init(&queue->nvme_sq, &queue->nvme_cq);
 if (ret) {
  ret = NVME_RDMA_CM_NO_RSC;
  goto out_free_queue;
 }

 ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
 if (ret)
  goto out_destroy_sq;

 /*
 * Schedules the actual release because calling rdma_destroy_id from
 * inside a CM callback would trigger a deadlock. (great API design..)
 */
 INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
 queue->dev = ndev;
 queue->cm_id = cm_id;
 queue->port = port->nport;

 spin_lock_init(&queue->state_lock);
 queue->state = NVMET_RDMA_Q_CONNECTING;
 INIT_LIST_HEAD(&queue->rsp_wait_list);
 INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
 spin_lock_init(&queue->rsp_wr_wait_lock);
 INIT_LIST_HEAD(&queue->queue_list);

 queue->idx = ida_alloc(&nvmet_rdma_queue_ida, GFP_KERNEL);
 if (queue->idx < 0) {
  ret = NVME_RDMA_CM_NO_RSC;
  goto out_destroy_sq;
 }

 /*
 * Spread the io queues across completion vectors,
 * but still keep all admin queues on vector 0.
 */
 queue->comp_vector = !queue->host_qid ? 0 :
  queue->idx % ndev->device->num_comp_vectors;


 ret = nvmet_rdma_alloc_rsps(queue);
 if (ret) {
  ret = NVME_RDMA_CM_NO_RSC;
  goto out_ida_remove;
 }

 if (ndev->srqs) {
  queue->nsrq = ndev->srqs[queue->comp_vector % ndev->srq_count];
 } else {
  queue->cmds = nvmet_rdma_alloc_cmds(ndev,
    queue->recv_queue_size,
    !queue->host_qid);
  if (IS_ERR(queue->cmds)) {
   ret = NVME_RDMA_CM_NO_RSC;
   goto out_free_responses;
  }
 }

 ret = nvmet_rdma_create_queue_ib(queue);
 if (ret) {
  pr_err("%s: creating RDMA queue failed (%d).\n",
   __func__, ret);
  ret = NVME_RDMA_CM_NO_RSC;
  goto out_free_cmds;
 }

 return queue;

out_free_cmds:
 if (!queue->nsrq) {
  nvmet_rdma_free_cmds(queue->dev, queue->cmds,
    queue->recv_queue_size,
    !queue->host_qid);
 }
out_free_responses:
 nvmet_rdma_free_rsps(queue);
out_ida_remove:
 ida_free(&nvmet_rdma_queue_ida, queue->idx);
out_destroy_sq:
 nvmet_sq_destroy(&queue->nvme_sq);
out_free_queue:
 nvmet_cq_put(&queue->nvme_cq);
 kfree(queue);
out_reject:
 nvmet_rdma_cm_reject(cm_id, ret);
 return NULL;
}

static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
{
 struct nvmet_rdma_queue *queue = priv;

 switch (event->event) {
 case IB_EVENT_COMM_EST:
  rdma_notify(queue->cm_id, event->event);
  break;
 case IB_EVENT_QP_LAST_WQE_REACHED:
  pr_debug("received last WQE reached event for queue=0x%p\n",
    queue);
  break;
 default:
  pr_err("received IB QP event: %s (%d)\n",
         ib_event_msg(event->event), event->event);
  break;
 }
}

static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
  struct nvmet_rdma_queue *queue,
  struct rdma_conn_param *p)
{
 struct rdma_conn_param  param = { };
 struct nvme_rdma_cm_rep priv = { };
 int ret = -ENOMEM;

 param.rnr_retry_count = 7;
 param.flow_control = 1;
 param.initiator_depth = min_t(u8, p->initiator_depth,
  queue->dev->device->attrs.max_qp_init_rd_atom);
 param.private_data = &priv;
 param.private_data_len = sizeof(priv);
 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
 priv.crqsize = cpu_to_le16(queue->recv_queue_size);

 ret = rdma_accept(cm_id, ¶m);
 if (ret)
  pr_err("rdma_accept failed (error code = %d)\n", ret);

 return ret;
}

static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
  struct rdma_cm_event *event)
{
 struct nvmet_rdma_device *ndev;
 struct nvmet_rdma_queue *queue;
 int ret = -EINVAL;

 ndev = nvmet_rdma_find_get_device(cm_id);
 if (!ndev) {
  nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
  return -ECONNREFUSED;
 }

 queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
 if (!queue) {
  ret = -ENOMEM;
  goto put_device;
 }

 if (queue->host_qid == 0) {
  struct nvmet_rdma_queue *q;
  int pending = 0;

  /* Check for pending controller teardown */
  mutex_lock(&nvmet_rdma_queue_mutex);
  list_for_each_entry(q, &nvmet_rdma_queue_list, queue_list) {
   if (q->nvme_sq.ctrl == queue->nvme_sq.ctrl &&
       q->state == NVMET_RDMA_Q_DISCONNECTING)
    pending++;
  }
  mutex_unlock(&nvmet_rdma_queue_mutex);
  if (pending > NVMET_RDMA_BACKLOG)
   return NVME_SC_CONNECT_CTRL_BUSY;
 }

 ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
 if (ret) {
  /*
 * Don't destroy the cm_id in free path, as we implicitly
 * destroy the cm_id here with non-zero ret code.
 */
  queue->cm_id = NULL;
  goto free_queue;
 }

 mutex_lock(&nvmet_rdma_queue_mutex);
 list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
 mutex_unlock(&nvmet_rdma_queue_mutex);

 return 0;

free_queue:
 nvmet_rdma_free_queue(queue);
put_device:
 kref_put(&ndev->ref, nvmet_rdma_free_dev);

 return ret;
}

static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
{
 unsigned long flags;

 spin_lock_irqsave(&queue->state_lock, flags);
 if (queue->state != NVMET_RDMA_Q_CONNECTING) {
  pr_warn("trying to establish a connected queue\n");
  goto out_unlock;
 }
 queue->state = NVMET_RDMA_Q_LIVE;

 while (!list_empty(&queue->rsp_wait_list)) {
  struct nvmet_rdma_rsp *cmd;

  cmd = list_first_entry(&queue->rsp_wait_list,
     struct nvmet_rdma_rsp, wait_list);
  list_del(&cmd->wait_list);

  spin_unlock_irqrestore(&queue->state_lock, flags);
  nvmet_rdma_handle_command(queue, cmd);
  spin_lock_irqsave(&queue->state_lock, flags);
 }

out_unlock:
 spin_unlock_irqrestore(&queue->state_lock, flags);
}

static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
{
 bool disconnect = false;
 unsigned long flags;

 pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);

 spin_lock_irqsave(&queue->state_lock, flags);
 switch (queue->state) {
 case NVMET_RDMA_Q_CONNECTING:
  while (!list_empty(&queue->rsp_wait_list)) {
   struct nvmet_rdma_rsp *rsp;

   rsp = list_first_entry(&queue->rsp_wait_list,
            struct nvmet_rdma_rsp,
            wait_list);
   list_del(&rsp->wait_list);
   nvmet_rdma_put_rsp(rsp);
  }
  fallthrough;
 case NVMET_RDMA_Q_LIVE:
  queue->state = NVMET_RDMA_Q_DISCONNECTING;
  disconnect = true;
  break;
 case NVMET_RDMA_Q_DISCONNECTING:
  break;
 }
 spin_unlock_irqrestore(&queue->state_lock, flags);

 if (disconnect) {
  rdma_disconnect(queue->cm_id);
  queue_work(nvmet_wq, &queue->release_work);
 }
}

static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
{
 bool disconnect = false;

 mutex_lock(&nvmet_rdma_queue_mutex);
 if (!list_empty(&queue->queue_list)) {
  list_del_init(&queue->queue_list);
  disconnect = true;
 }
 mutex_unlock(&nvmet_rdma_queue_mutex);

 if (disconnect)
  __nvmet_rdma_queue_disconnect(queue);
}

static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
  struct nvmet_rdma_queue *queue)
{
 WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);

 mutex_lock(&nvmet_rdma_queue_mutex);
 if (!list_empty(&queue->queue_list))
  list_del_init(&queue->queue_list);
 mutex_unlock(&nvmet_rdma_queue_mutex);

 pr_err("failed to connect queue %d\n", queue->idx);
 queue_work(nvmet_wq, &queue->release_work);
}

/**
 * nvmet_rdma_device_removal() - Handle RDMA device removal
 * @cm_id: rdma_cm id, used for nvmet port
 * @queue:      nvmet rdma queue (cm id qp_context)
 *
 * DEVICE_REMOVAL event notifies us that the RDMA device is about
 * to unplug. Note that this event can be generated on a normal
 * queue cm_id and/or a device bound listener cm_id (where in this
 * case queue will be null).
 *
 * We registered an ib_client to handle device removal for queues,
 * so we only need to handle the listening port cm_ids. In this case
 * we nullify the priv to prevent double cm_id destruction and destroying
 * the cm_id implicitly by returning a non-zero rc to the callout.
 */
static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
  struct nvmet_rdma_queue *queue)
{
 struct nvmet_rdma_port *port;

 if (queue) {
  /*
 * This is a queue cm_id. we have registered
 * an ib_client to handle queues removal
 * so don't interfere and just return.
 */
  return 0;
 }

 port = cm_id->context;

 /*
 * This is a listener cm_id. Make sure that
 * future remove_port won't invoke a double
 * cm_id destroy. use atomic xchg to make sure
 * we don't compete with remove_port.
 */
 if (xchg(&port->cm_id, NULL) != cm_id)
  return 0;

 /*
 * We need to return 1 so that the core will destroy
 * its own ID.  What a great API design..
 */
 return 1;
}

static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
  struct rdma_cm_event *event)
{
 struct nvmet_rdma_queue *queue = NULL;
 int ret = 0;

 if (cm_id->qp)
  queue = cm_id->qp->qp_context;

 pr_debug("%s (%d): status %d id %p\n",
  rdma_event_msg(event->event), event->event,
  event->status, cm_id);

 switch (event->event) {
 case RDMA_CM_EVENT_CONNECT_REQUEST:
  ret = nvmet_rdma_queue_connect(cm_id, event);
  break;
 case RDMA_CM_EVENT_ESTABLISHED:
  nvmet_rdma_queue_established(queue);
  break;
 case RDMA_CM_EVENT_ADDR_CHANGE:
  if (!queue) {
   struct nvmet_rdma_port *port = cm_id->context;

   queue_delayed_work(nvmet_wq, &port->repair_work, 0);
   break;
  }
  fallthrough;
 case RDMA_CM_EVENT_DISCONNECTED:
 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
  nvmet_rdma_queue_disconnect(queue);
  break;
 case RDMA_CM_EVENT_DEVICE_REMOVAL:
  ret = nvmet_rdma_device_removal(cm_id, queue);
  break;
 case RDMA_CM_EVENT_REJECTED:
  pr_debug("Connection rejected: %s\n",
    rdma_reject_msg(cm_id, event->status));
  fallthrough;
 case RDMA_CM_EVENT_UNREACHABLE:
 case RDMA_CM_EVENT_CONNECT_ERROR:
  nvmet_rdma_queue_connect_fail(cm_id, queue);
  break;
 default:
  pr_err("received unrecognized RDMA CM event %d\n",
   event->event);
  break;
 }

 return ret;
}

static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
{
 struct nvmet_rdma_queue *queue, *n;

 mutex_lock(&nvmet_rdma_queue_mutex);
 list_for_each_entry_safe(queue, n, &nvmet_rdma_queue_list, queue_list) {
  if (queue->nvme_sq.ctrl != ctrl)
   continue;
  list_del_init(&queue->queue_list);
  __nvmet_rdma_queue_disconnect(queue);
 }
 mutex_unlock(&nvmet_rdma_queue_mutex);
}

static void nvmet_rdma_destroy_port_queues(struct nvmet_rdma_port *port)
{
 struct nvmet_rdma_queue *queue, *tmp;
 struct nvmet_port *nport = port->nport;

 mutex_lock(&nvmet_rdma_queue_mutex);
 list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
     queue_list) {
  if (queue->port != nport)
   continue;

  list_del_init(&queue->queue_list);
  __nvmet_rdma_queue_disconnect(queue);
 }
 mutex_unlock(&nvmet_rdma_queue_mutex);
}

static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
{
 struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);

 if (cm_id)
  rdma_destroy_id(cm_id);

 /*
 * Destroy the remaining queues, which are not belong to any
 * controller yet. Do it here after the RDMA-CM was destroyed
 * guarantees that no new queue will be created.
 */
 nvmet_rdma_destroy_port_queues(port);
}

static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
{
 struct sockaddr *addr = (struct sockaddr *)&port->addr;
 struct rdma_cm_id *cm_id;
 int ret;

 cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
   RDMA_PS_TCP, IB_QPT_RC);
 if (IS_ERR(cm_id)) {
  pr_err("CM ID creation failed\n");
  return PTR_ERR(cm_id);
 }

 /*
 * Allow both IPv4 and IPv6 sockets to bind a single port
 * at the same time.
 */
 ret = rdma_set_afonly(cm_id, 1);
 if (ret) {
  pr_err("rdma_set_afonly failed (%d)\n", ret);
  goto out_destroy_id;
 }

 ret = rdma_bind_addr(cm_id, addr);
 if (ret) {
  pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
  goto out_destroy_id;
 }

 ret = rdma_listen(cm_id, NVMET_RDMA_BACKLOG);
 if (ret) {
  pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
  goto out_destroy_id;
 }

 port->cm_id = cm_id;
 return 0;

out_destroy_id:
 rdma_destroy_id(cm_id);
 return ret;
}

static void nvmet_rdma_repair_port_work(struct work_struct *w)
{
 struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
   struct nvmet_rdma_port, repair_work);
 int ret;

 nvmet_rdma_disable_port(port);
 ret = nvmet_rdma_enable_port(port);
 if (ret)
  queue_delayed_work(nvmet_wq, &port->repair_work, 5 * HZ);
}

static int nvmet_rdma_add_port(struct nvmet_port *nport)
{
 struct nvmet_rdma_port *port;
 __kernel_sa_family_t af;
 int ret;

 port = kzalloc(sizeof(*port), GFP_KERNEL);
 if (!port)
  return -ENOMEM;

 nport->priv = port;
 port->nport = nport;
 INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);

 switch (nport->disc_addr.adrfam) {
 case NVMF_ADDR_FAMILY_IP4:
  af = AF_INET;
  break;
 case NVMF_ADDR_FAMILY_IP6:
  af = AF_INET6;
  break;
 default:
  pr_err("address family %d not supported\n",
   nport->disc_addr.adrfam);
  ret = -EINVAL;
  goto out_free_port;
 }

 if (nport->inline_data_size < 0) {
  nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
 } else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
  pr_warn("inline_data_size %u is too large, reducing to %u\n",
   nport->inline_data_size,
   NVMET_RDMA_MAX_INLINE_DATA_SIZE);
  nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
 }

 if (nport->max_queue_size < 0) {
  nport->max_queue_size = NVME_RDMA_DEFAULT_QUEUE_SIZE;
 } else if (nport->max_queue_size > NVME_RDMA_MAX_QUEUE_SIZE) {
  pr_warn("max_queue_size %u is too large, reducing to %u\n",
   nport->max_queue_size, NVME_RDMA_MAX_QUEUE_SIZE);
  nport->max_queue_size = NVME_RDMA_MAX_QUEUE_SIZE;
 }

 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
   nport->disc_addr.trsvcid, &port->addr);
 if (ret) {
  pr_err("malformed ip/port passed: %s:%s\n",
   nport->disc_addr.traddr, nport->disc_addr.trsvcid);
  goto out_free_port;
 }

 ret = nvmet_rdma_enable_port(port);
 if (ret)
  goto out_free_port;

 pr_info("enabling port %d (%pISpcs)\n",
  le16_to_cpu(nport->disc_addr.portid),
  (struct sockaddr *)&port->addr);

 return 0;

out_free_port:
 kfree(port);
 return ret;
}

static void nvmet_rdma_remove_port(struct nvmet_port *nport)
{
 struct nvmet_rdma_port *port = nport->priv;

 cancel_delayed_work_sync(&port->repair_work);
 nvmet_rdma_disable_port(port);
 kfree(port);
}

static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
  struct nvmet_port *nport, char *traddr)
{
 struct nvmet_rdma_port *port = nport->priv;
 struct rdma_cm_id *cm_id = port->cm_id;

 if (inet_addr_is_any(&cm_id->route.addr.src_addr)) {
  struct nvmet_rdma_rsp *rsp =
   container_of(req, struct nvmet_rdma_rsp, req);
  struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
  struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;

  sprintf(traddr, "%pISc", addr);
 } else {
  memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
 }
}

static ssize_t nvmet_rdma_host_port_addr(struct nvmet_ctrl *ctrl,
  char *traddr, size_t traddr_len)
{
 struct nvmet_sq *nvme_sq = ctrl->sqs[0];
 struct nvmet_rdma_queue *queue =
  container_of(nvme_sq, struct nvmet_rdma_queue, nvme_sq);

 return snprintf(traddr, traddr_len, "%pISc",
   (struct sockaddr *)&queue->cm_id->route.addr.dst_addr);
}

static u8 nvmet_rdma_get_mdts(const struct nvmet_ctrl *ctrl)
{
 if (ctrl->pi_support)
  return NVMET_RDMA_MAX_METADATA_MDTS;
 return NVMET_RDMA_MAX_MDTS;
}

static u16 nvmet_rdma_get_max_queue_size(const struct nvmet_ctrl *ctrl)
{
 if (ctrl->pi_support)
  return NVME_RDMA_MAX_METADATA_QUEUE_SIZE;
 return NVME_RDMA_MAX_QUEUE_SIZE;
}

static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
 .owner   = THIS_MODULE,
 .type   = NVMF_TRTYPE_RDMA,
 .msdbd   = 1,
 .flags   = NVMF_KEYED_SGLS | NVMF_METADATA_SUPPORTED,
 .add_port  = nvmet_rdma_add_port,
 .remove_port  = nvmet_rdma_remove_port,
 .queue_response  = nvmet_rdma_queue_response,
 .delete_ctrl  = nvmet_rdma_delete_ctrl,
 .disc_traddr  = nvmet_rdma_disc_port_addr,
 .host_traddr  = nvmet_rdma_host_port_addr,
 .get_mdts  = nvmet_rdma_get_mdts,
 .get_max_queue_size = nvmet_rdma_get_max_queue_size,
};

static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
{
 struct nvmet_rdma_queue *queue, *tmp;
 struct nvmet_rdma_device *ndev;
 bool found = false;

 mutex_lock(&device_list_mutex);
 list_for_each_entry(ndev, &device_list, entry) {
  if (ndev->device == ib_device) {
   found = true;
   break;
  }
 }
 mutex_unlock(&device_list_mutex);

 if (!found)
  return;

 /*
 * IB Device that is used by nvmet controllers is being removed,
 * delete all queues using this device.
 */
 mutex_lock(&nvmet_rdma_queue_mutex);
 list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
     queue_list) {
  if (queue->dev->device != ib_device)
   continue;

  pr_info("Removing queue %d\n", queue->idx);
  list_del_init(&queue->queue_list);
  __nvmet_rdma_queue_disconnect(queue);
 }
 mutex_unlock(&nvmet_rdma_queue_mutex);

 flush_workqueue(nvmet_wq);
}

static struct ib_client nvmet_rdma_ib_client = {
 .name   = "nvmet_rdma",
 .remove = nvmet_rdma_remove_one
};

static int __init nvmet_rdma_init(void)
{
 int ret;

 ret = ib_register_client(&nvmet_rdma_ib_client);
 if (ret)
  return ret;

 ret = nvmet_register_transport(&nvmet_rdma_ops);
 if (ret)
  goto err_ib_client;

 return 0;

err_ib_client:
 ib_unregister_client(&nvmet_rdma_ib_client);
 return ret;
}

static void __exit nvmet_rdma_exit(void)
{
 nvmet_unregister_transport(&nvmet_rdma_ops);
 ib_unregister_client(&nvmet_rdma_ib_client);
 WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
 ida_destroy(&nvmet_rdma_queue_ida);
}

module_init(nvmet_rdma_init);
module_exit(nvmet_rdma_exit);

MODULE_DESCRIPTION("NVMe target RDMA transport driver");
MODULE_LICENSE("GPL v2");
MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */