static int
fi_ibv_rdm_tagged_find_max_inline_size(struct ibv_pd *pd,
struct ibv_context *context)
{
struct ibv_qp_init_attr qp_attr;
struct ibv_qp *qp = NULL;
struct ibv_cq *cq = ibv_create_cq(context, 1, NULL, NULL, 0);
assert(cq);
int max_inline = 2;
int rst = 0;
memset(&qp_attr, 0, sizeof(qp_attr));
qp_attr.send_cq = cq;
qp_attr.recv_cq = cq;
qp_attr.qp_type = IBV_QPT_RC;
qp_attr.cap.max_send_wr = 1;
qp_attr.cap.max_recv_wr = 1;
qp_attr.cap.max_send_sge = 1;
qp_attr.cap.max_recv_sge = 1;
do {
if (qp)
ibv_destroy_qp(qp);
qp_attr.cap.max_inline_data = max_inline;
qp = ibv_create_qp(pd, &qp_attr);
if (qp)
rst = max_inline;
} while (qp && (max_inline *= 2));
if (rst != 0) {
int pos = rst, neg = max_inline;
do {
max_inline = pos + (neg - pos) / 2;
if (qp)
ibv_destroy_qp(qp);
qp_attr.cap.max_inline_data = max_inline;
qp = ibv_create_qp(pd, &qp_attr);
if (qp)
pos = max_inline;
else
neg = max_inline;
} while (neg - pos > 2);
rst = pos;
}
if (qp) {
ibv_destroy_qp(qp);
}
if (cq) {
ibv_destroy_cq(cq);
}
return rst;
}
static void rping_free_qp(struct rping_cb *cb)
{
ibv_destroy_qp(cb->qp);
ibv_destroy_cq(cb->cq);
ibv_destroy_comp_channel(cb->channel);
ibv_dealloc_pd(cb->pd);
}
int rdma_create_qp(struct rdma_cm_id *id, struct ibv_pd *pd,
struct ibv_qp_init_attr *qp_init_attr)
{
struct cma_id_private *id_priv;
struct ibv_qp *qp;
int ret;
id_priv = container_of(id, struct cma_id_private, id);
if (id->verbs != pd->context)
return ERR(EINVAL);
qp = ibv_create_qp(pd, qp_init_attr);
if (!qp)
return ERR(ENOMEM);
if (ucma_is_ud_ps(id->ps))
ret = ucma_init_ud_qp(id_priv, qp);
else
ret = ucma_init_conn_qp(id_priv, qp);
if (ret)
goto err;
id->qp = qp;
return 0;
err:
ibv_destroy_qp(qp);
return ret;
}
static void
rdma_trans_destroy(void *a)
{
Rdmatrans *rdma;
struct ibv_qp_attr attr;
rdma = a;
if (rdma->connected)
rdma_disconnect(rdma->cm_id);
if (rdma->qp) {
attr.qp_state = IBV_QPS_ERR;
ibv_modify_qp(rdma->qp, &attr, IBV_QP_STATE);
ibv_destroy_qp(rdma->qp);
}
if (rdma->cq)
ibv_destroy_cq(rdma->cq);
if (rdma->ch)
ibv_destroy_comp_channel(rdma->ch);
if (rdma->snd_mr)
ibv_dereg_mr(rdma->snd_mr);
if (rdma->snd_buf)
free(rdma->snd_buf);
if (rdma->rcv_mr)
ibv_dereg_mr(rdma->rcv_mr);
if (rdma->rcv_buf)
free(rdma->rcv_buf);
if (rdma->pd)
ibv_dealloc_pd(rdma->pd);
if (rdma->cm_id)
rdma_destroy_id(rdma->cm_id);
}
static ucs_status_t uct_rc_verbs_ep_tag_qp_create(uct_rc_verbs_iface_t *iface,
uct_rc_verbs_ep_t *ep)
{
#if HAVE_IBV_EX_HW_TM
struct ibv_qp_cap cap;
ucs_status_t status;
int ret;
if (UCT_RC_VERBS_TM_ENABLED(iface)) {
/* Send queue of this QP will be used by FW for HW RNDV. Driver requires
* such a QP to be initialized with zero send queue length. */
status = uct_rc_iface_qp_create(&iface->super, IBV_QPT_RC, &ep->tm_qp,
&cap, iface->tm.xrq.srq, 0);
if (status != UCS_OK) {
return status;
}
status = uct_rc_iface_qp_init(&iface->super, ep->tm_qp);
if (status != UCS_OK) {
ret = ibv_destroy_qp(ep->tm_qp);
if (ret) {
ucs_warn("ibv_destroy_qp() returned %d: %m", ret);
}
return status;
}
uct_rc_iface_add_ep(&iface->super, &ep->super, ep->tm_qp->qp_num);
}
#endif
return UCS_OK;
}
static void free_sa_qp_cache(void)
{
struct mca_btl_openib_sa_qp_cache *cache, *tmp;
cache = sa_qp_cache;
while (NULL != cache) {
/* free cache data */
if (cache->device_name)
free(cache->device_name);
if (NULL != cache->qp)
ibv_destroy_qp(cache->qp);
if (NULL != cache->ah)
ibv_destroy_ah(cache->ah);
if (NULL != cache->cq)
ibv_destroy_cq(cache->cq);
if (NULL != cache->mr)
ibv_dereg_mr(cache->mr);
if (NULL != cache->pd)
ibv_dealloc_pd(cache->pd);
tmp = cache->next;
free(cache);
cache = tmp;
}
sa_qp_cache = NULL;
}
/* destroy ud context */
void mv2_ud_destroy_ctx (mv2_ud_ctx_t *ctx)
{
if (ctx->qp) {
ibv_destroy_qp(ctx->qp);
}
MPIU_Free(ctx);
}
/*
* It seems you can't probe a device / port to see if it supports a
* specific type of QP. You just have to try to make it and see if it
* works. This is a short helper function to try to make a QP of a
* specific type and return whether it worked.
*/
static bool make_qp(struct ibv_pd *pd, struct ibv_cq *cq, enum ibv_qp_type type)
{
struct ibv_qp_init_attr qpia;
struct ibv_qp *qp;
memset(&qpia, 0, sizeof(qpia));
qpia.qp_context = NULL;
qpia.send_cq = cq;
qpia.recv_cq = cq;
qpia.srq = NULL;
qpia.cap.max_send_wr = 1;
qpia.cap.max_recv_wr = 1;
qpia.cap.max_send_sge = 1;
qpia.cap.max_recv_sge = 1;
qpia.cap.max_inline_data = 0;
qpia.qp_type = type;
qpia.sq_sig_all = 0;
qp = ibv_create_qp(pd, &qpia);
if (NULL != qp) {
ibv_destroy_qp(qp);
return true;
}
return false;
}
int
rd_close_qp(CONNECTION *con)
{
if (con->qp)
ibv_destroy_qp(con->qp);
return 0;
}
int __ibv_destroy_qp_1_0(struct ibv_qp_1_0 *qp)
{ fprintf(stderr, "%s:%s:%d \n", __func__, __FILE__, __LINE__);
int ret;
ret = ibv_destroy_qp(qp->real_qp);
if (ret)
return ret;
free(qp);
return 0;
}
static int fio_rdmaio_close_file(struct thread_data *td, struct fio_file *f)
{
struct rdmaio_data *rd = td->io_ops->data;
struct ibv_send_wr *bad_wr;
/* unregister rdma buffer */
/*
* Client sends notification to the server side
*/
/* refer to: http://linux.die.net/man/7/rdma_cm */
if ((rd->is_client == 1) && ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE)
|| (rd->rdma_protocol ==
FIO_RDMA_MEM_READ))) {
if (ibv_post_send(rd->qp, &rd->sq_wr, &bad_wr) != 0) {
log_err("fio: ibv_post_send fail");
return 1;
}
dprint(FD_IO, "fio: close information sent success\n");
rdma_poll_wait(td, IBV_WC_SEND);
}
if (rd->is_client == 1)
rdma_disconnect(rd->cm_id);
else {
rdma_disconnect(rd->child_cm_id);
#if 0
rdma_disconnect(rd->cm_id);
#endif
}
#if 0
if (get_next_channel_event(td, rd->cm_channel, RDMA_CM_EVENT_DISCONNECTED) != 0) {
log_err("fio: wait for RDMA_CM_EVENT_DISCONNECTED\n");
return 1;
}
#endif
ibv_destroy_cq(rd->cq);
ibv_destroy_qp(rd->qp);
if (rd->is_client == 1)
rdma_destroy_id(rd->cm_id);
else {
rdma_destroy_id(rd->child_cm_id);
rdma_destroy_id(rd->cm_id);
}
ibv_destroy_comp_channel(rd->channel);
ibv_dealloc_pd(rd->pd);
return 0;
}
static void uct_ib_md_umr_qp_destroy(uct_ib_md_t *md)
{
#if HAVE_EXP_UMR
if (md->umr_qp != NULL) {
ibv_destroy_qp(md->umr_qp);
}
if (md->umr_cq != NULL) {
ibv_destroy_cq(md->umr_cq);
}
#endif
}
int __ibv_destroy_qp_1_0(struct ibv_qp_1_0 *qp)
{
int ret;
ret = ibv_destroy_qp(qp->real_qp);
if (ret)
return ret;
free(qp);
return 0;
}
int BClientContext::destroy_context () {
if (qp) TEST_NZ(ibv_destroy_qp (qp));
if (recv_memory_mr) TEST_NZ (ibv_dereg_mr (recv_memory_mr));
if (lock_result_mr) TEST_NZ (ibv_dereg_mr (lock_result_mr));
if (cq) TEST_NZ (ibv_destroy_cq (cq));
if (pd) TEST_NZ (ibv_dealloc_pd (pd));
if (ib_ctx) TEST_NZ (ibv_close_device (ib_ctx));
if (sockfd >= 0) TEST_NZ (close (sockfd));
return 0;
}
static void uct_rc_verbs_ep_tag_qp_destroy(uct_rc_verbs_ep_t *ep)
{
#if HAVE_IBV_EX_HW_TM
uct_rc_verbs_iface_t *iface = ucs_derived_of(ep->super.super.super.iface,
uct_rc_verbs_iface_t);
if (UCT_RC_VERBS_TM_ENABLED(iface)) {
if (ibv_destroy_qp(ep->tm_qp)) {
ucs_warn("failed to destroy TM RNDV QP: %m");
}
uct_rc_iface_remove_ep(&iface->super, ep->tm_qp->qp_num);
}
#endif
}
static int resources_destroy(struct resources *res)
{
int rc = 0;
if (res->qp) {
if (ibv_destroy_qp(res->qp)) {
fprintf(stderr, "failed to destroy QP\n");
rc = 1;
}
}
if (res->mr) {
if (ibv_dereg_mr(res->mr)) {
fprintf(stderr, "failed to deregister MR\n");
rc = 1;
}
}
if (res->buf)
free(res->buf);
if (res->cq) {
if (ibv_destroy_cq(res->cq)) {
fprintf(stderr, "failed to destroy CQ\n");
rc = 1;
}
}
if (res->pd) {
if (ibv_dealloc_pd(res->pd)) {
fprintf(stderr, "failed to deallocate PD\n");
rc = 1;
}
}
if (res->ib_ctx) {
if (ibv_close_device(res->ib_ctx)) {
fprintf(stderr, "failed to close device context\n");
rc = 1;
}
}
if (res->sock >= 0) {
if (close(res->sock)) {
fprintf(stderr, "failed to close socket\n");
rc = 1;
}
}
return rc;
}
void fi_ibv_free_xrc_conn_setup(struct fi_ibv_xrc_ep *ep)
{
assert(ep->conn_setup);
if (ep->conn_setup->rsvd_ini_qpn)
ibv_destroy_qp(ep->conn_setup->rsvd_ini_qpn);
if (ep->conn_setup->rsvd_tgt_qpn)
ibv_destroy_qp(ep->conn_setup->rsvd_tgt_qpn);
free(ep->conn_setup);
ep->conn_setup = NULL;
/*Free RDMA CM IDs releasing their associated resources, RDMA CM
* is used for connection setup only with XRC */
if (ep->base_ep.id) {
rdma_destroy_id(ep->base_ep.id);
ep->base_ep.id = NULL;
}
if (ep->tgt_id) {
rdma_destroy_id(ep->tgt_id);
ep->tgt_id = NULL;
}
}
static inline int fi_ibv_get_qp_cap(struct ibv_context *ctx,
struct fi_info *info)
{
struct ibv_pd *pd;
struct ibv_cq *cq;
struct ibv_qp *qp;
struct ibv_qp_init_attr init_attr;
int ret = 0;
pd = ibv_alloc_pd(ctx);
if (!pd) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "ibv_alloc_pd", errno);
return -errno;
}
cq = ibv_create_cq(ctx, 1, NULL, NULL, 0);
if (!cq) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "ibv_create_cq", errno);
ret = -errno;
goto err1;
}
memset(&init_attr, 0, sizeof init_attr);
init_attr.send_cq = cq;
init_attr.recv_cq = cq;
init_attr.cap.max_send_wr = verbs_default_tx_size;
init_attr.cap.max_recv_wr = verbs_default_rx_size;
init_attr.cap.max_send_sge = verbs_default_tx_iov_limit;
init_attr.cap.max_recv_sge = verbs_default_rx_iov_limit;
init_attr.cap.max_inline_data = verbs_default_inline_size;
init_attr.qp_type = IBV_QPT_RC;
qp = ibv_create_qp(pd, &init_attr);
if (!qp) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "ibv_create_qp", errno);
ret = -errno;
goto err2;
}
info->tx_attr->inject_size = init_attr.cap.max_inline_data;
ibv_destroy_qp(qp);
err2:
ibv_destroy_cq(cq);
err1:
ibv_dealloc_pd(pd);
return ret;
}
/** ========================================================================= */
static void destroy_sa_qp(struct oib_port *port)
{
int i;
// if the user just unregistered trap messages those messages may still
// be on this list, wait 5 seconds for the thread to handle the response.
for (i = 0; i < 5000; i++) {
if (!LIST_EMPTY(&port->pending_reg_msg_head)) {
usleep(1000);
}
else {
DBGPRINT("destroy_sa_qp: wait %d ms for LIST_EMPTY\n", i);
break;
}
}
stop_ud_cq_monitor(port);
join_port_thread(port);
/* Free any remaining unregistration messages */
if (!LIST_EMPTY(&port->pending_reg_msg_head)) {
OUTPUT_ERROR("Ignoring Pending Notice un-registation requests\n");
oib_sa_remove_all_pending_reg_msgs(port);
}
if (port->sa_ah)
ibv_destroy_ah(port->sa_ah);
if (port->sa_qp)
ibv_destroy_qp(port->sa_qp);
for (i = 0; i<port->num_userspace_recv_buf; i++)
if (port->recv_bufs)
ibv_dereg_mr(port->recv_bufs[i].mr);
if (port->sa_qp_pd)
ibv_dealloc_pd(port->sa_qp_pd);
if (port->sa_qp_cq)
ibv_destroy_cq(port->sa_qp_cq);
if (port->recv_bufs) {
free(port->recv_bufs);
port->recv_bufs = NULL;
}
if (port->sa_qp_comp_channel)
ibv_destroy_comp_channel(port->sa_qp_comp_channel);
}
static void
rdmasniff_cleanup(pcap_t *handle)
{
struct pcap_rdmasniff *priv = handle->priv;
ibv_dereg_mr(priv->mr);
ibv_destroy_flow(priv->flow);
ibv_destroy_qp(priv->qp);
ibv_destroy_cq(priv->cq);
ibv_dealloc_pd(priv->pd);
ibv_destroy_comp_channel(priv->channel);
ibv_close_device(priv->context);
free(priv->oneshot_buffer);
pcap_cleanup_live_common(handle);
}
static ucs_status_t uct_ib_mlx5_check_dc(uct_ib_device_t *dev)
{
ucs_status_t status = UCS_OK;
struct ibv_context *ctx = dev->ibv_context;
struct ibv_qp_init_attr_ex qp_attr = {};
struct mlx5dv_qp_init_attr dv_attr = {};
struct ibv_pd *pd;
struct ibv_cq *cq;
struct ibv_qp *qp;
pd = ibv_alloc_pd(ctx);
if (pd == NULL) {
ucs_error("ibv_alloc_pd() failed: %m");
return UCS_ERR_IO_ERROR;
}
cq = ibv_create_cq(ctx, 1, NULL, NULL, 0);
if (cq == NULL) {
ucs_error("ibv_create_cq() failed: %m");
status = UCS_ERR_IO_ERROR;
goto err_cq;
}
qp_attr.send_cq = cq;
qp_attr.recv_cq = cq;
qp_attr.cap.max_send_wr = 1;
qp_attr.cap.max_send_sge = 1;
qp_attr.qp_type = IBV_QPT_DRIVER;
qp_attr.comp_mask = IBV_QP_INIT_ATTR_PD;
qp_attr.pd = pd;
dv_attr.comp_mask = MLX5DV_QP_INIT_ATTR_MASK_DC;
dv_attr.dc_init_attr.dc_type = MLX5DV_DCTYPE_DCI;
/* create DCI qp successful means DC is supported */
qp = mlx5dv_create_qp(ctx, &qp_attr, &dv_attr);
if (qp) {
ibv_destroy_qp(qp);
dev->flags |= UCT_IB_DEVICE_FLAG_DC;
}
ibv_destroy_cq(cq);
err_cq:
ibv_dealloc_pd(pd);
return status;
}
int
resource_destroy(resource_t *res)
{
int rc = 0;
// Delete QP
if (res->qp && ibv_destroy_qp(res->qp)){
fprintf(stderr, "failed to destroy QP\n");
rc = 1;
}
// Deregister MR
if(res->mr_list && res->mr_size > 0){
int i;
for(i=0; i<res->mr_size; i++){
ibv_dereg_mr(res->mr_list[i]);
}
free(res->mr_list);
}
if(res->buf){
free(res->buf);
}
// Delete CQ
if (res->scq && ibv_destroy_cq(res->scq)){
fprintf(stderr, "failed to destroy SCQ\n");
rc = 1;
}
if (res->rcq && ibv_destroy_cq(res->rcq)){
fprintf(stderr, "failed to destroy RCQ\n");
rc = 1;
}
if(res->comp_ch && ibv_destroy_comp_channel(res->comp_ch)){
fprintf(stderr, "failed to destroy Complete CH\n");
rc = 1;
}
// Deallocate PD
if (res->pd && ibv_dealloc_pd(res->pd)){
fprintf(stderr, "failed to deallocate PD\n");
rc = 1;
}
if (res->ib_ctx && ibv_close_device(res->ib_ctx)){
fprintf(stderr, "failed to close device context\n");
rc = 1;
}
return rc;
}
void psoib_con_cleanup(psoib_con_info_t *con_info, hca_info_t *hca_info)
{
if (!hca_info) hca_info = &default_hca;
list_del_init(&con_info->next_con_info);
if (con_info->send.bufs.mr) {
psoib_vapi_free(hca_info, &con_info->send.bufs);
con_info->send.bufs.mr = 0;
}
if (con_info->recv.bufs.mr) {
psoib_vapi_free(hca_info, &con_info->recv.bufs);
con_info->recv.bufs.mr = 0;
}
if (con_info->qp) {
ibv_destroy_qp(con_info->qp);
con_info->qp = 0;
}
}
static int destroy_ctx_resources(struct pingpong_context *ctx) {
int test_result = 0;
if (ibv_destroy_qp(ctx->qp)) {
fprintf(stderr, "failed to destroy QP\n");
test_result = 1;
}
if (ibv_destroy_cq(ctx->cq)) {
fprintf(stderr, "failed to destroy CQ\n");
test_result = 1;
}
if (ibv_dereg_mr(ctx->mr)) {
fprintf(stderr, "failed to deregister MR\n");
test_result = 1;
}
if (ctx->channel) {
if (ibv_destroy_comp_channel(ctx->channel)) {
fprintf(stderr, "failed to destroy channel \n");
test_result = 1;
}
}
if (ibv_dealloc_pd(ctx->pd)) {
fprintf(stderr, "failed to deallocate PD\n");
test_result = 1;
}
if (ibv_close_device(ctx->context)) {
fprintf(stderr, "failed to close device context\n");
test_result = 1;
}
free(ctx->buf);
free(ctx);
free(tstamp);
return test_result;
}
static int destroy_ctx_resources(struct pingpong_context *ctx,int num_qps) {
int i;
int test_result = 0;
for (i = 0; i < num_qps; i++) {
if (ibv_destroy_qp(ctx->qp[i])) {
fprintf(stderr, "failed to destroy QP\n");
test_result = 1;
}
}
if (ibv_destroy_cq(ctx->cq)) {
fprintf(stderr, "failed to destroy CQ\n");
test_result = 1;
}
if (ibv_dereg_mr(ctx->mr)) {
fprintf(stderr, "failed to deregister MR\n");
test_result = 1;
}
if (ibv_dealloc_pd(ctx->pd)) {
fprintf(stderr, "failed to deallocate PD\n");
test_result = 1;
}
if (ibv_close_device(ctx->context)) {
fprintf(stderr, "failed to close device context\n");
test_result = 1;
}
free(ctx->buf);
free(ctx->qp);
free(ctx->scnt);
free(ctx->ccnt);
free(ctx);
return test_result;
}
void release_resources(void)
{
int i;
// Destroy the registration cache
reg_cache_destroy(nprocs);
for(i = 0; i < nprocs ; i++) {
if (me != i) {
if(ibv_destroy_qp(conn.qp[i])) {
printf("Exiting\n");
exit(1);
}
}
}
#if 0
if(ibv_destroy_cq(hca.cq)) {
fprintf(stderr,"Couldn't destroy cq %s\n",
ibv_get_device_name(hca.ib_dev));
}
if(ibv_dealloc_pd(hca.pd)) {
fprintf(stderr,"Couldn't free pd %s\n",
ibv_get_device_name(hca.ib_dev));
}
#endif
free(conn.qp);
free(conn.lid);
free(conn.qp_num);
free(rbuf.qp_num);
free(rbuf.lid);
#if 0
free(rbuf.rkey);
free(rbuf.buf);
#endif
}
/* In case if XRC recv qp was closed and sender still don't know about it
* we need close the qp, reset the ib_adrr status to CLOSED and start everything
* from scratch.
*/
static void xoob_restart_connect(mca_btl_base_endpoint_t *endpoint)
{
BTL_VERBOSE(("Restarting the connection for the endpoint"));
OPAL_THREAD_LOCK(&endpoint->ib_addr->addr_lock);
switch (endpoint->ib_addr->status) {
case MCA_BTL_IB_ADDR_CONNECTED:
/* so we have the send qp, we just need the recive site.
* Send request for SRQ numbers */
BTL_VERBOSE(("Restart The IB addr: sid %" PRIx64 " lid %d"
"in MCA_BTL_IB_ADDR_CONNECTED status,"
" Changing to MCA_BTL_IB_ADDR_CLOSED and starting from scratch\n",
endpoint->ib_addr->subnet_id,endpoint->ib_addr->lid));
/* Switching back to closed and starting from scratch */
endpoint->ib_addr->status = MCA_BTL_IB_ADDR_CLOSED;
/* destroy the qp */
/* the reciver site was alredy closed so all pending list must be clean ! */
assert (opal_list_is_empty(&endpoint->qps->no_wqe_pending_frags[0]));
assert (opal_list_is_empty(&endpoint->qps->no_wqe_pending_frags[1]));
if(ibv_destroy_qp(endpoint->qps[0].qp->lcl_qp))
BTL_ERROR(("Failed to destroy QP"));
case MCA_BTL_IB_ADDR_CLOSED:
case MCA_BTL_IB_ADDR_CONNECTING:
BTL_VERBOSE(("Restart The IB addr: sid %" PRIx64 " lid %d"
"in MCA_BTL_IB_ADDR_CONNECTING or MCA_BTL_IB_ADDR_CLOSED status,"
" starting from scratch\n",
endpoint->ib_addr->subnet_id,endpoint->ib_addr->lid));
OPAL_THREAD_UNLOCK(&endpoint->ib_addr->addr_lock);
/* xoob_module_start_connect() should automaticly handle all other cases */
if (OMPI_SUCCESS != xoob_module_start_connect(NULL, endpoint))
BTL_ERROR(("Failed to restart connection from MCA_BTL_IB_ADDR_CONNECTING/CLOSED"));
break;
default :
BTL_ERROR(("Invalid endpoint status %d", endpoint->ib_addr->status));
OPAL_THREAD_UNLOCK(&endpoint->ib_addr->addr_lock);
}
}
static inline int fi_ibv_get_qp_cap(struct ibv_context *ctx,
struct ibv_device_attr *device_attr,
struct fi_info *info)
{
struct ibv_pd *pd;
struct ibv_cq *cq;
struct ibv_qp *qp;
struct ibv_qp_init_attr init_attr;
int ret = 0;
pd = ibv_alloc_pd(ctx);
if (!pd) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "ibv_alloc_pd", errno);
return -errno;
}
cq = ibv_create_cq(ctx, 1, NULL, NULL, 0);
if (!cq) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "ibv_create_cq", errno);
ret = -errno;
goto err1;
}
/* TODO: serialize access to string buffers */
fi_read_file(FI_CONF_DIR, "def_tx_ctx_size",
def_tx_ctx_size, sizeof def_tx_ctx_size);
fi_read_file(FI_CONF_DIR, "def_rx_ctx_size",
def_rx_ctx_size, sizeof def_rx_ctx_size);
fi_read_file(FI_CONF_DIR, "def_tx_iov_limit",
def_tx_iov_limit, sizeof def_tx_iov_limit);
fi_read_file(FI_CONF_DIR, "def_rx_iov_limit",
def_rx_iov_limit, sizeof def_rx_iov_limit);
fi_read_file(FI_CONF_DIR, "def_inject_size",
def_inject_size, sizeof def_inject_size);
memset(&init_attr, 0, sizeof init_attr);
init_attr.send_cq = cq;
init_attr.recv_cq = cq;
init_attr.cap.max_send_wr = MIN(atoi(def_tx_ctx_size), device_attr->max_qp_wr);
init_attr.cap.max_recv_wr = MIN(atoi(def_rx_ctx_size), device_attr->max_qp_wr);
init_attr.cap.max_send_sge = MIN(atoi(def_tx_iov_limit), device_attr->max_sge);
init_attr.cap.max_recv_sge = MIN(atoi(def_rx_iov_limit), device_attr->max_sge);
init_attr.cap.max_inline_data = atoi(def_inject_size);
init_attr.qp_type = IBV_QPT_RC;
qp = ibv_create_qp(pd, &init_attr);
if (!qp) {
VERBS_INFO_ERRNO(FI_LOG_FABRIC, "ibv_create_qp", errno);
ret = -errno;
goto err2;
}
info->tx_attr->inject_size = init_attr.cap.max_inline_data;
info->tx_attr->iov_limit = init_attr.cap.max_send_sge;
info->tx_attr->size = init_attr.cap.max_send_wr;
info->rx_attr->iov_limit = init_attr.cap.max_recv_sge;
/*
* On some HW ibv_create_qp can increase max_recv_wr value more than
* it really supports. So, alignment with device capability is needed.
*/
info->rx_attr->size = MIN(init_attr.cap.max_recv_wr,
device_attr->max_qp_wr);
ibv_destroy_qp(qp);
err2:
ibv_destroy_cq(cq);
err1:
ibv_dealloc_pd(pd);
return ret;
}
RDMAChannel::~RDMAChannel() {
CHECK(!ibv_destroy_qp(qp_)) << "Failed to destroy QP";
}
int rdma_backend_qp_state_rtr(RdmaBackendDev *backend_dev, RdmaBackendQP *qp,
uint8_t qp_type, uint8_t sgid_idx,
union ibv_gid *dgid, uint32_t dqpn,
uint32_t rq_psn, uint32_t qkey, bool use_qkey)
{
struct ibv_qp_attr attr = {};
union ibv_gid ibv_gid = {
.global.interface_id = dgid->global.interface_id,
.global.subnet_prefix = dgid->global.subnet_prefix
};
int rc, attr_mask;
attr.qp_state = IBV_QPS_RTR;
attr_mask = IBV_QP_STATE;
qp->sgid_idx = sgid_idx;
switch (qp_type) {
case IBV_QPT_RC:
attr.path_mtu = IBV_MTU_1024;
attr.dest_qp_num = dqpn;
attr.max_dest_rd_atomic = 1;
attr.min_rnr_timer = 12;
attr.ah_attr.port_num = backend_dev->port_num;
attr.ah_attr.is_global = 1;
attr.ah_attr.grh.hop_limit = 1;
attr.ah_attr.grh.dgid = ibv_gid;
attr.ah_attr.grh.sgid_index = qp->sgid_idx;
attr.rq_psn = rq_psn;
attr_mask |= IBV_QP_AV | IBV_QP_PATH_MTU | IBV_QP_DEST_QPN |
IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC |
IBV_QP_MIN_RNR_TIMER;
trace_rdma_backend_rc_qp_state_rtr(qp->ibqp->qp_num,
be64_to_cpu(ibv_gid.global.
subnet_prefix),
be64_to_cpu(ibv_gid.global.
interface_id),
qp->sgid_idx, dqpn, rq_psn);
break;
case IBV_QPT_UD:
if (use_qkey) {
attr.qkey = qkey;
attr_mask |= IBV_QP_QKEY;
}
trace_rdma_backend_ud_qp_state_rtr(qp->ibqp->qp_num, use_qkey ? qkey :
0);
break;
}
rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask);
if (rc) {
rdma_error_report("ibv_modify_qp fail, rc=%d, errno=%d", rc, errno);
return -EIO;
}
return 0;
}
int rdma_backend_qp_state_rts(RdmaBackendQP *qp, uint8_t qp_type,
uint32_t sq_psn, uint32_t qkey, bool use_qkey)
{
struct ibv_qp_attr attr = {};
int rc, attr_mask;
attr.qp_state = IBV_QPS_RTS;
attr.sq_psn = sq_psn;
attr_mask = IBV_QP_STATE | IBV_QP_SQ_PSN;
switch (qp_type) {
case IBV_QPT_RC:
attr.timeout = 14;
attr.retry_cnt = 7;
attr.rnr_retry = 7;
attr.max_rd_atomic = 1;
attr_mask |= IBV_QP_TIMEOUT | IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY |
IBV_QP_MAX_QP_RD_ATOMIC;
trace_rdma_backend_rc_qp_state_rts(qp->ibqp->qp_num, sq_psn);
break;
case IBV_QPT_UD:
if (use_qkey) {
attr.qkey = qkey;
attr_mask |= IBV_QP_QKEY;
}
trace_rdma_backend_ud_qp_state_rts(qp->ibqp->qp_num, sq_psn,
use_qkey ? qkey : 0);
break;
}
rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask);
if (rc) {
rdma_error_report("ibv_modify_qp fail, rc=%d, errno=%d", rc, errno);
return -EIO;
}
return 0;
}
int rdma_backend_query_qp(RdmaBackendQP *qp, struct ibv_qp_attr *attr,
int attr_mask, struct ibv_qp_init_attr *init_attr)
{
if (!qp->ibqp) {
attr->qp_state = IBV_QPS_RTS;
return 0;
}
return ibv_query_qp(qp->ibqp, attr, attr_mask, init_attr);
}
void rdma_backend_destroy_qp(RdmaBackendQP *qp, RdmaDeviceResources *dev_res)
{
if (qp->ibqp) {
ibv_destroy_qp(qp->ibqp);
}
g_slist_foreach(qp->cqe_ctx_list.list, free_cqe_ctx, dev_res);
rdma_protected_gslist_destroy(&qp->cqe_ctx_list);
}
int rdma_backend_create_srq(RdmaBackendSRQ *srq, RdmaBackendPD *pd,
uint32_t max_wr, uint32_t max_sge,
uint32_t srq_limit)
{
struct ibv_srq_init_attr srq_init_attr = {};
srq_init_attr.attr.max_wr = max_wr;
srq_init_attr.attr.max_sge = max_sge;
srq_init_attr.attr.srq_limit = srq_limit;
srq->ibsrq = ibv_create_srq(pd->ibpd, &srq_init_attr);
if (!srq->ibsrq) {
rdma_error_report("ibv_create_srq failed, errno=%d", errno);
return -EIO;
}
rdma_protected_gslist_init(&srq->cqe_ctx_list);
return 0;
}
int rdma_backend_query_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr)
{
if (!srq->ibsrq) {
return -EINVAL;
}
return ibv_query_srq(srq->ibsrq, srq_attr);
}
int rdma_backend_modify_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr,
int srq_attr_mask)
{
if (!srq->ibsrq) {
return -EINVAL;
}
return ibv_modify_srq(srq->ibsrq, srq_attr, srq_attr_mask);
}
void rdma_backend_destroy_srq(RdmaBackendSRQ *srq, RdmaDeviceResources *dev_res)
{
if (srq->ibsrq) {
ibv_destroy_srq(srq->ibsrq);
}
g_slist_foreach(srq->cqe_ctx_list.list, free_cqe_ctx, dev_res);
rdma_protected_gslist_destroy(&srq->cqe_ctx_list);
}
#define CHK_ATTR(req, dev, member, fmt) ({ \
trace_rdma_check_dev_attr(#member, dev.member, req->member); \
if (req->member > dev.member) { \
rdma_warn_report("%s = "fmt" is higher than host device capability "fmt, \
#member, req->member, dev.member); \
req->member = dev.member; \
} \
})
static int init_device_caps(RdmaBackendDev *backend_dev,
struct ibv_device_attr *dev_attr)
{
struct ibv_device_attr bk_dev_attr;
int rc;
rc = ibv_query_device(backend_dev->context, &bk_dev_attr);
if (rc) {
rdma_error_report("ibv_query_device fail, rc=%d, errno=%d", rc, errno);
return -EIO;
}
dev_attr->max_sge = MAX_SGE;
dev_attr->max_srq_sge = MAX_SGE;
CHK_ATTR(dev_attr, bk_dev_attr, max_mr_size, "%" PRId64);
CHK_ATTR(dev_attr, bk_dev_attr, max_qp, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_sge, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_cq, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_mr, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_pd, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_qp_rd_atom, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_qp_init_rd_atom, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_ah, "%d");
CHK_ATTR(dev_attr, bk_dev_attr, max_srq, "%d");
return 0;
}
static inline void build_mad_hdr(struct ibv_grh *grh, union ibv_gid *sgid,
union ibv_gid *my_gid, int paylen)
{
grh->paylen = htons(paylen);
grh->sgid = *sgid;
grh->dgid = *my_gid;
}
