/* Must call with cq->lock held */
ssize_t fi_ibv_poll_cq(struct fi_ibv_cq *cq, struct ibv_wc *wc)
{
struct fi_ibv_msg_epe *epe;
struct slist_entry *entry;
ssize_t ret;
ret = ibv_poll_cq(cq->cq, 1, wc);
if (ret <= 0)
return ret;
if (wc->opcode == IBV_WC_RECV || wc->opcode == IBV_WC_RECV_RDMA_WITH_IMM)
return ret;
/* TODO Handle the case when app posts a send with same wr_id */
if ((wc->wr_id & cq->wr_id_mask) != cq->send_signal_wr_id)
return ret;
entry = slist_remove_first_match(&cq->ep_list, fi_ibv_match_ep_id, (void *)wc->wr_id);
if (!entry) {
FI_WARN(&fi_ibv_prov, FI_LOG_CQ, "No matching EP for :"
"given signaled send completion\n");
return -FI_EOTHER;
}
epe = container_of(entry, struct fi_ibv_msg_epe, entry);
atomic_sub(&epe->ep->unsignaled_send_cnt,
VERBS_SEND_SIGNAL_THRESH(epe->ep));
atomic_dec(&epe->ep->comp_pending);
util_buf_release(cq->domain->fab->epe_pool, epe);
return 0;
}
static struct psmx_unexp *psmx_am_search_and_dequeue_unexp(
struct psmx_fid_domain *domain,
const void *src_addr)
{
struct slist_entry *item;
fastlock_acquire(&domain->unexp_queue.lock);
item = slist_remove_first_match(&domain->unexp_queue.list,
match_unexp, src_addr);
fastlock_release(&domain->unexp_queue.lock);
if (!item)
return NULL;
return container_of(item, struct psmx_unexp, list_entry);
}
static struct psmx_am_request *psmx_am_search_and_dequeue_recv(
struct psmx_fid_domain *domain,
const void *src_addr)
{
struct slist_entry *item;
pthread_mutex_lock(&domain->recv_queue.lock);
item = slist_remove_first_match(&domain->recv_queue.list,
match_recv, src_addr);
pthread_mutex_unlock(&domain->recv_queue.lock);
if (!item)
return NULL;
return container_of(item, struct psmx_am_request, list_entry);
}
int _gnix_wait_set_remove(struct fid_wait *wait, struct fid *wait_obj)
{
struct gnix_fid_wait *wait_priv;
struct gnix_wait_entry *wait_entry;
struct slist_entry *found;
GNIX_TRACE(WAIT_SUB, "\n");
wait_priv = container_of(wait, struct gnix_fid_wait, wait.fid);
found = slist_remove_first_match(&wait_priv->set, gnix_match_fid,
wait_obj);
if (found) {
wait_entry = container_of(found, struct gnix_wait_entry,
entry);
free(wait_entry);
return FI_SUCCESS;
}
return -FI_EINVAL;
}
int gnix_ep_open(struct fid_domain *domain, struct fi_info *info,
struct fid_ep **ep, void *context)
{
int ret = FI_SUCCESS;
int tsret = FI_SUCCESS;
uint32_t cdm_id;
struct gnix_fid_domain *domain_priv;
struct gnix_fid_ep *ep_priv;
gnix_hashtable_attr_t gnix_ht_attr;
gnix_ht_key_t *key_ptr;
struct gnix_tag_storage_attr untagged_attr = {
.type = GNIX_TAG_LIST,
.use_src_addr_matching = 1,
};
bool free_list_inited = false;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
if ((domain == NULL) || (info == NULL) || (ep == NULL) ||
(info->ep_attr == NULL))
return -FI_EINVAL;
/*
* TODO: need to implement other endpoint types
*/
if (info->ep_attr->type != FI_EP_RDM)
return -FI_ENOSYS;
domain_priv = container_of(domain, struct gnix_fid_domain, domain_fid);
ep_priv = calloc(1, sizeof *ep_priv);
if (!ep_priv)
return -FI_ENOMEM;
/* init untagged storages */
tsret = _gnix_posted_tag_storage_init(
&ep_priv->posted_recv_queue, &untagged_attr);
if (tsret)
return tsret;
tsret = _gnix_unexpected_tag_storage_init(
&ep_priv->unexp_recv_queue, &untagged_attr);
if (tsret)
return tsret;
/* init tagged storages */
tsret = _gnix_posted_tag_storage_init(
&ep_priv->tagged_posted_recv_queue, NULL);
if (tsret)
return tsret;
tsret = _gnix_unexpected_tag_storage_init(
&ep_priv->tagged_unexp_recv_queue, NULL);
if (tsret)
return tsret;
ep_priv->ep_fid.fid.fclass = FI_CLASS_EP;
ep_priv->ep_fid.fid.context = context;
ep_priv->ep_fid.fid.ops = &gnix_ep_fi_ops;
ep_priv->ep_fid.ops = &gnix_ep_ops;
ep_priv->domain = domain_priv;
ep_priv->type = info->ep_attr->type;
_gnix_ref_init(&ep_priv->ref_cnt, 1, __ep_destruct);
fastlock_init(&ep_priv->recv_comp_lock);
fastlock_init(&ep_priv->recv_queue_lock);
fastlock_init(&ep_priv->tagged_queue_lock);
slist_init(&ep_priv->pending_recv_comp_queue);
ep_priv->caps = info->caps & GNIX_EP_RDM_CAPS;
if (info->tx_attr)
ep_priv->op_flags = info->tx_attr->op_flags;
if (info->rx_attr)
ep_priv->op_flags |= info->rx_attr->op_flags;
ep_priv->op_flags &= GNIX_EP_OP_FLAGS;
ep_priv->min_multi_recv = GNIX_OPT_MIN_MULTI_RECV_DEFAULT;
ret = __fr_freelist_init(ep_priv);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Error allocating gnix_fab_req freelist (%s)",
fi_strerror(-ret));
goto err;
} else
free_list_inited = true;
ep_priv->ep_fid.msg = &gnix_ep_msg_ops;
ep_priv->ep_fid.rma = &gnix_ep_rma_ops;
ep_priv->ep_fid.tagged = &gnix_ep_tagged_ops;
ep_priv->ep_fid.atomic = &gnix_ep_atomic_ops;
ep_priv->ep_fid.cm = &gnix_cm_ops;
if (ep_priv->type == FI_EP_RDM) {
if (info->src_addr != NULL) {
ret = __gnix_ep_bound_prep(domain_priv,
info,
ep_priv);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"__gnix_ep_bound_prep returned error (%s)",
fi_strerror(-ret));
goto err;
}
} else {
fastlock_acquire(&domain_priv->cm_nic_lock);
/*
* if a cm_nic has not yet been allocated for this
* domain, do it now. Reuse the embedded gnix_nic
* in the cm_nic as the nic for this endpoint
* to reduce demand on Aries hw resources.
*/
if (domain_priv->cm_nic == NULL) {
ret = _gnix_cm_nic_alloc(domain_priv,
info,
&domain_priv->cm_nic);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_cm_nic_alloc returned %s\n",
fi_strerror(-ret));
fastlock_release(
&domain_priv->cm_nic_lock);
goto err;
}
ep_priv->cm_nic = domain_priv->cm_nic;
ep_priv->nic = ep_priv->cm_nic->nic;
_gnix_ref_get(ep_priv->nic);
} else {
ep_priv->cm_nic = domain_priv->cm_nic;
_gnix_ref_get(ep_priv->cm_nic);
}
fastlock_release(&domain_priv->cm_nic_lock);
ep_priv->my_name.gnix_addr.device_addr =
ep_priv->cm_nic->my_name.gnix_addr.device_addr;
ep_priv->my_name.cm_nic_cdm_id =
ep_priv->cm_nic->my_name.gnix_addr.cdm_id;
ret = _gnix_get_new_cdm_id(domain_priv, &cdm_id);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"gnix_get_new_cdm_id call returned %s\n",
fi_strerror(-ret));
goto err;
}
ep_priv->my_name.gnix_addr.cdm_id = cdm_id;
}
key_ptr = (gnix_ht_key_t *)&ep_priv->my_name.gnix_addr;
ret = _gnix_ht_insert(ep_priv->cm_nic->addr_to_ep_ht,
*key_ptr,
ep_priv);
if ((ret != FI_SUCCESS) && (ret != -FI_ENOSPC)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"__gnix_ht_insert returned %d\n",
ret);
goto err;
}
gnix_ht_attr.ht_initial_size = domain_priv->params.ct_init_size;
gnix_ht_attr.ht_maximum_size = domain_priv->params.ct_max_size;
gnix_ht_attr.ht_increase_step = domain_priv->params.ct_step;
gnix_ht_attr.ht_increase_type = GNIX_HT_INCREASE_MULT;
gnix_ht_attr.ht_collision_thresh = 500;
gnix_ht_attr.ht_hash_seed = 0xdeadbeefbeefdead;
gnix_ht_attr.ht_internal_locking = 0;
gnix_ht_attr.destructor = __gnix_vc_destroy_ht_entry;
ep_priv->vc_ht = calloc(1, sizeof(struct gnix_hashtable));
if (ep_priv->vc_ht == NULL)
goto err;
ret = _gnix_ht_init(ep_priv->vc_ht, &gnix_ht_attr);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"gnix_ht_init call returned %d\n",
ret);
goto err;
}
fastlock_init(&ep_priv->vc_ht_lock);
} else {
ep_priv->cm_nic = NULL;
ep_priv->vc = NULL;
}
ep_priv->progress_fn = NULL;
ep_priv->rx_progress_fn = NULL;
if (ep_priv->nic == NULL) {
ret = gnix_nic_alloc(domain_priv, NULL, &ep_priv->nic);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_nic_alloc call returned %d\n",
ret);
goto err;
}
}
/*
* if smsg callbacks not present hook them up now
*/
if (ep_priv->nic->smsg_callbacks == NULL)
ep_priv->nic->smsg_callbacks = gnix_ep_smsg_callbacks;
_gnix_ref_get(ep_priv->domain);
*ep = &ep_priv->ep_fid;
return ret;
err:
if (free_list_inited == true)
__fr_freelist_destroy(ep_priv);
if (ep_priv->vc_ht != NULL) {
_gnix_ht_destroy(ep_priv->vc_ht); /* may not be initialized but
okay */
free(ep_priv->vc_ht);
ep_priv->vc_ht = NULL;
}
if (ep_priv->cm_nic != NULL)
ret = _gnix_cm_nic_free(ep_priv->cm_nic);
if (ep_priv->nic != NULL)
ret = _gnix_nic_free(ep_priv->nic);
free(ep_priv);
return ret;
}
static int __match_context(struct slist_entry *item, const void *arg)
{
struct gnix_fab_req *req;
req = container_of(item, struct gnix_fab_req, slist);
return req->user_context == arg;
}
static inline struct gnix_fab_req *__find_tx_req(
struct gnix_fid_ep *ep,
void *context)
{
struct gnix_fab_req *req = NULL;
struct slist_entry *entry;
struct gnix_vc *vc;
GNIX_HASHTABLE_ITERATOR(ep->vc_ht, iter);
GNIX_DEBUG(FI_LOG_EP_CTRL, "searching VCs for the correct context to"
" cancel, context=%p", context);
fastlock_acquire(&ep->vc_ht_lock);
while ((vc = _gnix_ht_iterator_next(&iter))) {
fastlock_acquire(&vc->tx_queue_lock);
entry = slist_remove_first_match(&vc->tx_queue,
__match_context, context);
fastlock_release(&vc->tx_queue_lock);
if (entry) {
req = container_of(entry, struct gnix_fab_req, slist);
break;
}
}
fastlock_release(&ep->vc_ht_lock);
return req;
}
static inline struct gnix_fab_req *__find_rx_req(
struct gnix_fid_ep *ep,
void *context)
{
struct gnix_fab_req *req = NULL;
fastlock_acquire(&ep->recv_queue_lock);
req = _gnix_remove_req_by_context(&ep->posted_recv_queue, context);
fastlock_release(&ep->recv_queue_lock);
if (req)
return req;
fastlock_acquire(&ep->tagged_queue_lock);
req = _gnix_remove_req_by_context(&ep->tagged_posted_recv_queue,
context);
fastlock_release(&ep->tagged_queue_lock);
return req;
}
static ssize_t gnix_ep_cancel(fid_t fid, void *context)
{
int ret = FI_SUCCESS;
struct gnix_fid_ep *ep;
struct gnix_fab_req *req;
struct gnix_fid_cq *err_cq = NULL;
struct gnix_fid_cntr *err_cntr = NULL;
void *addr;
uint64_t tag, flags;
size_t len;
int is_send = 0;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
ep = container_of(fid, struct gnix_fid_ep, ep_fid.fid);
if (!ep->domain)
return -FI_EDOMAIN;
/* without context, we will have to find a request that matches
* a recv or send request. Try the send requests first.
*/
GNIX_INFO(FI_LOG_EP_CTRL, "looking for event to cancel\n");
req = __find_tx_req(ep, context);
if (!req) {
req = __find_rx_req(ep, context);
if (req) {
err_cq = ep->recv_cq;
err_cntr = ep->recv_cntr;
}
} else {
is_send = 1;
err_cq = ep->send_cq;
err_cntr = ep->send_cntr;
}
GNIX_INFO(FI_LOG_EP_CTRL, "finished searching\n");
if (!req)
return -FI_ENOENT;
if (err_cq) {
/* add canceled event */
if (!(req->type == GNIX_FAB_RQ_RDMA_READ ||
req->type == GNIX_FAB_RQ_RDMA_WRITE)) {
if (!is_send) {
addr = (void *) req->msg.recv_addr;
len = req->msg.recv_len;
} else {
addr = (void *) req->msg.send_addr;
len = req->msg.send_len;
}
tag = req->msg.tag;
} else {
/* rma information */
addr = (void *) req->rma.loc_addr;
len = req->rma.len;
tag = 0;
}
flags = req->flags;
_gnix_cq_add_error(err_cq, context, flags, len, addr, 0 /* data */,
tag, len, FI_ECANCELED, FI_ECANCELED, 0);
}
if (err_cntr) {
/* signal increase in cntr errs */
_gnix_cntr_inc_err(err_cntr);
}
_gnix_fr_free(ep, req);
return ret;
}
