int fi_wait_init(struct util_fabric *fabric, struct fi_wait_attr *attr,
struct util_wait *wait)
{
struct fid_poll *poll_fid;
struct fi_poll_attr poll_attr;
int ret;
wait->prov = fabric->prov;
ofi_atomic_initialize32(&wait->ref, 0);
wait->wait_fid.fid.fclass = FI_CLASS_WAIT;
switch (attr->wait_obj) {
case FI_WAIT_UNSPEC:
case FI_WAIT_FD:
wait->wait_obj = FI_WAIT_FD;
break;
case FI_WAIT_MUTEX_COND:
wait->wait_obj = FI_WAIT_MUTEX_COND;
break;
default:
assert(0);
return -FI_EINVAL;
}
memset(&poll_attr, 0, sizeof poll_attr);
ret = fi_poll_create_(fabric->prov, NULL, &poll_attr, &poll_fid);
if (ret)
return ret;
wait->pollset = container_of(poll_fid, struct util_poll, poll_fid);
wait->fabric = fabric;
ofi_atomic_inc32(&fabric->ref);
return 0;
}
static int util_domain_init(struct util_domain *domain,
const struct fi_info *info)
{
ofi_atomic_initialize32(&domain->ref, 0);
fastlock_init(&domain->lock);
domain->info_domain_caps = info->caps | info->domain_attr->caps;
domain->info_domain_mode = info->mode | info->domain_attr->mode;
domain->mr_mode = info->domain_attr->mr_mode;
domain->addr_format = info->addr_format;
domain->av_type = info->domain_attr->av_type;
domain->name = strdup(info->domain_attr->name);
return domain->name ? 0 : -FI_ENOMEM;
}
int psmx2_stx_ctx(struct fid_domain *domain, struct fi_tx_attr *attr,
struct fid_stx **stx, void *context)
{
struct psmx2_fid_domain *domain_priv;
struct psmx2_trx_ctxt *trx_ctxt;
struct psmx2_fid_stx *stx_priv;
int err = -FI_EINVAL;
domain_priv = container_of(domain, struct psmx2_fid_domain,
util_domain.domain_fid.fid);
if (!domain_priv)
goto errout;
stx_priv = (struct psmx2_fid_stx *) calloc(1, sizeof *stx_priv);
if (!stx_priv) {
err = -FI_ENOMEM;
goto errout;
}
trx_ctxt = psmx2_trx_ctxt_alloc(domain_priv, NULL/*src_addr*/, 0);
if (!trx_ctxt) {
err = -FI_ENOMEM;
goto errout_free_stx;
}
psmx2_domain_acquire(domain_priv);
stx_priv->stx.fid.fclass = FI_CLASS_STX_CTX;
stx_priv->stx.fid.context = context;
stx_priv->stx.fid.ops = &psmx2_fi_ops_stx;
stx_priv->stx.ops = &psmx2_stx_ops;
stx_priv->domain = domain_priv;
stx_priv->tx = trx_ctxt;
ofi_atomic_initialize32(&stx_priv->ref, 0);
psmx2_lock(&domain_priv->trx_ctxt_lock, 1);
dlist_insert_before(&trx_ctxt->entry,
&domain_priv->trx_ctxt_list);
psmx2_unlock(&domain_priv->trx_ctxt_lock, 1);
*stx = &stx_priv->stx;
return 0;
errout_free_stx:
free(stx_priv);
errout:
return err;
}
int ofi_wait_fd_add(struct util_wait *wait, int fd, uint32_t events,
ofi_wait_fd_try_func wait_try, void *arg, void *context)
{
struct ofi_wait_fd_entry *fd_entry;
struct dlist_entry *entry;
struct util_wait_fd *wait_fd = container_of(wait, struct util_wait_fd,
util_wait);
int ret = 0;
fastlock_acquire(&wait_fd->lock);
entry = dlist_find_first_match(&wait_fd->fd_list, ofi_wait_fd_match, &fd);
if (entry) {
FI_DBG(wait->prov, FI_LOG_EP_CTRL,
"Given fd (%d) already added to wait list - %p \n",
fd, wait_fd);
fd_entry = container_of(entry, struct ofi_wait_fd_entry, entry);
ofi_atomic_inc32(&fd_entry->ref);
goto out;
}
ret = fi_epoll_add(wait_fd->epoll_fd, fd, events, context);
if (ret) {
FI_WARN(wait->prov, FI_LOG_FABRIC, "Unable to add fd to epoll\n");
goto out;
}
fd_entry = calloc(1, sizeof *fd_entry);
if (!fd_entry) {
ret = -FI_ENOMEM;
fi_epoll_del(wait_fd->epoll_fd, fd);
goto out;
}
fd_entry->fd = fd;
fd_entry->wait_try = wait_try;
fd_entry->arg = arg;
ofi_atomic_initialize32(&fd_entry->ref, 1);
dlist_insert_tail(&fd_entry->entry, &wait_fd->fd_list);
out:
fastlock_release(&wait_fd->lock);
return ret;
}
static int psmx2_domain_init(struct psmx2_fid_domain *domain,
struct psmx2_ep_name *src_addr)
{
int err;
err = fastlock_init(&domain->mr_lock);
if (err) {
FI_WARN(&psmx2_prov, FI_LOG_CORE,
"fastlock_init(mr_lock) returns %d\n", err);
goto err_out;
}
domain->mr_map = rbtNew(&psmx2_key_compare);
if (!domain->mr_map) {
FI_WARN(&psmx2_prov, FI_LOG_CORE,
"rbtNew failed\n");
goto err_out_destroy_mr_lock;
}
domain->mr_reserved_key = 1;
domain->max_atomic_size = INT_MAX;
ofi_atomic_initialize32(&domain->sep_cnt, 0);
fastlock_init(&domain->sep_lock);
dlist_init(&domain->sep_list);
dlist_init(&domain->trx_ctxt_list);
fastlock_init(&domain->trx_ctxt_lock);
if (domain->progress_thread_enabled)
psmx2_domain_start_progress(domain);
return 0;
err_out_destroy_mr_lock:
fastlock_destroy(&domain->mr_lock);
err_out:
return err;
}
static int
usdf_dom_rdc_alloc_data(struct usdf_domain *udp)
{
struct usdf_rdm_connection *rdc;
int ret;
int i;
udp->dom_rdc_hashtab = calloc(USDF_RDM_HASH_SIZE,
sizeof(*udp->dom_rdc_hashtab));
if (udp->dom_rdc_hashtab == NULL) {
return -FI_ENOMEM;
}
SLIST_INIT(&udp->dom_rdc_free);
ofi_atomic_initialize32(&udp->dom_rdc_free_cnt, 0);
for (i = 0; i < USDF_RDM_FREE_BLOCK; ++i) {
rdc = calloc(1, sizeof(*rdc));
if (rdc == NULL) {
return -FI_ENOMEM;
}
ret = usdf_timer_alloc(usdf_rdm_rdc_timeout, rdc,
&rdc->dc_timer);
if (ret != 0) {
free(rdc);
return ret;
}
rdc->dc_flags = USDF_DCS_UNCONNECTED | USDF_DCF_NEW_RX;
rdc->dc_next_rx_seq = 0;
rdc->dc_next_tx_seq = 0;
rdc->dc_last_rx_ack = rdc->dc_next_tx_seq - 1;
TAILQ_INIT(&rdc->dc_wqe_posted);
TAILQ_INIT(&rdc->dc_wqe_sent);
SLIST_INSERT_HEAD(&udp->dom_rdc_free, rdc, dc_addr_link);
ofi_atomic_inc32(&udp->dom_rdc_free_cnt);
}
udp->dom_rdc_total = USDF_RDM_FREE_BLOCK;
return 0;
}
static int util_eq_init(struct fid_fabric *fabric, struct util_eq *eq,
const struct fi_eq_attr *attr)
{
struct fi_wait_attr wait_attr;
struct fid_wait *wait;
int ret;
ofi_atomic_initialize32(&eq->ref, 0);
slist_init(&eq->list);
fastlock_init(&eq->lock);
switch (attr->wait_obj) {
case FI_WAIT_NONE:
break;
case FI_WAIT_UNSPEC:
case FI_WAIT_FD:
case FI_WAIT_MUTEX_COND:
memset(&wait_attr, 0, sizeof wait_attr);
wait_attr.wait_obj = attr->wait_obj;
eq->internal_wait = 1;
ret = fi_wait_open(fabric, &wait_attr, &wait);
if (ret)
return ret;
eq->wait = container_of(wait, struct util_wait, wait_fid);
break;
case FI_WAIT_SET:
eq->wait = container_of(attr->wait_set, struct util_wait,
wait_fid);
break;
default:
assert(0);
return -FI_EINVAL;
}
return 0;
}
int sock_cq_open(struct fid_domain *domain, struct fi_cq_attr *attr,
struct fid_cq **cq, void *context)
{
struct sock_domain *sock_dom;
struct sock_cq *sock_cq;
struct fi_wait_attr wait_attr;
struct sock_fid_list *list_entry;
struct sock_wait *wait;
int ret;
sock_dom = container_of(domain, struct sock_domain, dom_fid);
ret = sock_cq_verify_attr(attr);
if (ret)
return ret;
sock_cq = calloc(1, sizeof(*sock_cq));
if (!sock_cq)
return -FI_ENOMEM;
ofi_atomic_initialize32(&sock_cq->ref, 0);
sock_cq->cq_fid.fid.fclass = FI_CLASS_CQ;
sock_cq->cq_fid.fid.context = context;
sock_cq->cq_fid.fid.ops = &sock_cq_fi_ops;
sock_cq->cq_fid.ops = &sock_cq_ops;
if (attr == NULL) {
sock_cq->attr = _sock_cq_def_attr;
} else {
sock_cq->attr = *attr;
if (attr->size == 0)
sock_cq->attr.size = _sock_cq_def_attr.size;
}
sock_cq->domain = sock_dom;
sock_cq->cq_entry_size = sock_cq_entry_size(sock_cq);
sock_cq_set_report_fn(sock_cq);
dlist_init(&sock_cq->tx_list);
dlist_init(&sock_cq->rx_list);
dlist_init(&sock_cq->ep_list);
dlist_init(&sock_cq->overflow_list);
ret = ofi_rbfdinit(&sock_cq->cq_rbfd, sock_cq->attr.size *
sock_cq->cq_entry_size);
if (ret)
goto err1;
ret = ofi_rbinit(&sock_cq->addr_rb,
sock_cq->attr.size * sizeof(fi_addr_t));
if (ret)
goto err2;
ret = ofi_rbinit(&sock_cq->cqerr_rb, sock_cq->attr.size *
sizeof(struct fi_cq_err_entry));
if (ret)
goto err3;
fastlock_init(&sock_cq->lock);
switch (sock_cq->attr.wait_obj) {
case FI_WAIT_NONE:
case FI_WAIT_UNSPEC:
case FI_WAIT_FD:
break;
case FI_WAIT_MUTEX_COND:
wait_attr.flags = 0;
wait_attr.wait_obj = FI_WAIT_MUTEX_COND;
ret = sock_wait_open(&sock_dom->fab->fab_fid, &wait_attr,
&sock_cq->waitset);
if (ret) {
ret = -FI_EINVAL;
goto err4;
}
sock_cq->signal = 1;
break;
case FI_WAIT_SET:
if (!attr) {
ret = -FI_EINVAL;
goto err4;
}
sock_cq->waitset = attr->wait_set;
sock_cq->signal = 1;
wait = container_of(attr->wait_set, struct sock_wait, wait_fid);
list_entry = calloc(1, sizeof(*list_entry));
if (!list_entry) {
ret = -FI_ENOMEM;
goto err4;
}
dlist_init(&list_entry->entry);
list_entry->fid = &sock_cq->cq_fid.fid;
dlist_insert_after(&list_entry->entry, &wait->fid_list);
break;
default:
break;
}
*cq = &sock_cq->cq_fid;
ofi_atomic_inc32(&sock_dom->ref);
fastlock_init(&sock_cq->list_lock);
return 0;
err4:
ofi_rbfree(&sock_cq->cqerr_rb);
err3:
ofi_rbfree(&sock_cq->addr_rb);
err2:
ofi_rbfdfree(&sock_cq->cq_rbfd);
err1:
free(sock_cq);
return ret;
}
int psmx2_sep_open(struct fid_domain *domain, struct fi_info *info,
struct fid_ep **sep, void *context)
{
struct psmx2_fid_domain *domain_priv;
struct psmx2_fid_ep *ep_priv;
struct psmx2_fid_sep *sep_priv;
struct psmx2_ep_name ep_name;
struct psmx2_ep_name *src_addr;
struct psmx2_trx_ctxt *trx_ctxt;
size_t ctxt_cnt = 1;
size_t ctxt_size;
int err = -FI_EINVAL;
int i;
domain_priv = container_of(domain, struct psmx2_fid_domain,
util_domain.domain_fid.fid);
if (!domain_priv)
goto errout;
if (info && info->ep_attr) {
if (info->ep_attr->tx_ctx_cnt > psmx2_env.sep_trx_ctxt) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL,
"tx_ctx_cnt %"PRIu64" exceed limit %d.\n",
info->ep_attr->tx_ctx_cnt,
psmx2_env.sep_trx_ctxt);
goto errout;
}
if (info->ep_attr->rx_ctx_cnt > psmx2_env.sep_trx_ctxt) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL,
"rx_ctx_cnt %"PRIu64" exceed limit %d.\n",
info->ep_attr->rx_ctx_cnt,
psmx2_env.sep_trx_ctxt);
goto errout;
}
ctxt_cnt = info->ep_attr->tx_ctx_cnt;
if (ctxt_cnt < info->ep_attr->rx_ctx_cnt)
ctxt_cnt = info->ep_attr->rx_ctx_cnt;
if (ctxt_cnt == 0) {
FI_INFO(&psmx2_prov, FI_LOG_EP_CTRL,
"tx_ctx_cnt and rx_ctx_cnt are 0, use 1.\n");
ctxt_cnt = 1;
}
}
ctxt_size = ctxt_cnt * sizeof(struct psmx2_sep_ctxt);
sep_priv = (struct psmx2_fid_sep *) calloc(1, sizeof(*sep_priv) + ctxt_size);
if (!sep_priv) {
err = -FI_ENOMEM;
goto errout;
}
sep_priv->ep.fid.fclass = FI_CLASS_SEP;
sep_priv->ep.fid.context = context;
sep_priv->ep.fid.ops = &psmx2_fi_ops_sep;
sep_priv->ep.ops = &psmx2_sep_ops;
sep_priv->ep.cm = &psmx2_cm_ops;
sep_priv->domain = domain_priv;
sep_priv->ctxt_cnt = ctxt_cnt;
ofi_atomic_initialize32(&sep_priv->ref, 0);
src_addr = NULL;
if (info && info->src_addr) {
if (info->addr_format == FI_ADDR_STR)
src_addr = psmx2_string_to_ep_name(info->src_addr);
else
src_addr = info->src_addr;
}
for (i = 0; i < ctxt_cnt; i++) {
trx_ctxt = psmx2_trx_ctxt_alloc(domain_priv, src_addr, i);
if (!trx_ctxt) {
err = -FI_ENOMEM;
goto errout_free_ctxt;
}
sep_priv->ctxts[i].trx_ctxt = trx_ctxt;
err = psmx2_ep_open_internal(domain_priv, info, &ep_priv, context,
trx_ctxt);
if (err)
goto errout_free_ctxt;
/* override the ops so the fid can't be closed individually */
ep_priv->ep.fid.ops = &psmx2_fi_ops_sep_ctxt;
trx_ctxt->ep = ep_priv;
sep_priv->ctxts[i].ep = ep_priv;
}
sep_priv->type = PSMX2_EP_SCALABLE;
sep_priv->service = PSMX2_ANY_SERVICE;
if (src_addr) {
sep_priv->service = src_addr->service;
if (info->addr_format == FI_ADDR_STR)
free(src_addr);
}
if (sep_priv->service == PSMX2_ANY_SERVICE)
sep_priv->service = ((getpid() & 0x7FFF) << 16) +
((uintptr_t)sep_priv & 0xFFFF);
sep_priv->id = ofi_atomic_inc32(&domain_priv->sep_cnt);
psmx2_lock(&domain_priv->sep_lock, 1);
dlist_insert_before(&sep_priv->entry, &domain_priv->sep_list);
psmx2_unlock(&domain_priv->sep_lock, 1);
psmx2_lock(&domain_priv->trx_ctxt_lock, 1);
for (i = 0; i< ctxt_cnt; i++) {
dlist_insert_before(&sep_priv->ctxts[i].trx_ctxt->entry,
&domain_priv->trx_ctxt_list);
}
psmx2_unlock(&domain_priv->trx_ctxt_lock, 1);
ep_name.epid = sep_priv->ctxts[0].trx_ctxt->psm2_epid;
ep_name.sep_id = sep_priv->id;
ep_name.type = sep_priv->type;
ofi_ns_add_local_name(&domain_priv->fabric->name_server,
&sep_priv->service, &ep_name);
psmx2_domain_acquire(domain_priv);
*sep = &sep_priv->ep;
/* Make sure the AM handler is installed to answer SEP query */
psmx2_am_init(sep_priv->ctxts[0].trx_ctxt);
return 0;
errout_free_ctxt:
while (i) {
if (sep_priv->ctxts[i].ep)
psmx2_ep_close_internal(sep_priv->ctxts[i].ep);
if (sep_priv->ctxts[i].trx_ctxt)
psmx2_trx_ctxt_free(sep_priv->ctxts[i].trx_ctxt);
i--;
}
free(sep_priv);
errout:
return err;
}
int psmx2_ep_open_internal(struct psmx2_fid_domain *domain_priv,
struct fi_info *info,
struct psmx2_fid_ep **ep_out, void *context,
struct psmx2_trx_ctxt *trx_ctxt)
{
struct psmx2_fid_ep *ep_priv;
uint64_t ep_cap;
int err = -FI_EINVAL;
if (info)
ep_cap = info->caps;
else
ep_cap = FI_TAGGED;
if (info && info->ep_attr && info->ep_attr->auth_key) {
if (info->ep_attr->auth_key_size != sizeof(psm2_uuid_t)) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL,
"Invalid auth_key_len %"PRIu64
", should be %"PRIu64".\n",
info->ep_attr->auth_key_size,
sizeof(psm2_uuid_t));
goto errout;
}
if (memcmp(domain_priv->fabric->uuid, info->ep_attr->auth_key,
sizeof(psm2_uuid_t))) {
FI_WARN(&psmx2_prov, FI_LOG_EP_CTRL,
"Invalid auth_key: %s\n",
psmx2_uuid_to_string((void *)info->ep_attr->auth_key));
goto errout;
}
}
ep_priv = (struct psmx2_fid_ep *) calloc(1, sizeof *ep_priv);
if (!ep_priv) {
err = -FI_ENOMEM;
goto errout;
}
ep_priv->ep.fid.fclass = FI_CLASS_EP;
ep_priv->ep.fid.context = context;
ep_priv->ep.fid.ops = &psmx2_fi_ops;
ep_priv->ep.ops = &psmx2_ep_ops;
ep_priv->ep.cm = &psmx2_cm_ops;
ep_priv->domain = domain_priv;
ep_priv->rx = trx_ctxt;
if (!(info && info->ep_attr && info->ep_attr->tx_ctx_cnt == FI_SHARED_CONTEXT))
ep_priv->tx = trx_ctxt;
ofi_atomic_initialize32(&ep_priv->ref, 0);
PSMX2_CTXT_TYPE(&ep_priv->nocomp_send_context) = PSMX2_NOCOMP_SEND_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_send_context) = ep_priv;
PSMX2_CTXT_TYPE(&ep_priv->nocomp_tsend_context) = PSMX2_NOCOMP_TSEND_CONTEXT;
PSMX2_CTXT_EP(&ep_priv->nocomp_tsend_context) = ep_priv;
if (ep_cap & FI_TAGGED)
ep_priv->ep.tagged = &psmx2_tagged_ops;
if (ep_cap & FI_MSG)
ep_priv->ep.msg = &psmx2_msg_ops;
if (ep_cap & FI_RMA)
ep_priv->ep.rma = &psmx2_rma_ops;
if (ep_cap & FI_ATOMICS)
ep_priv->ep.atomic = &psmx2_atomic_ops;
ep_priv->caps = ep_cap;
err = psmx2_domain_enable_ep(domain_priv, ep_priv);
if (err)
goto errout_free_ep;
psmx2_domain_acquire(domain_priv);
if (info) {
if (info->tx_attr)
ep_priv->tx_flags = info->tx_attr->op_flags;
if (info->rx_attr)
ep_priv->rx_flags = info->rx_attr->op_flags;
}
psmx2_ep_optimize_ops(ep_priv);
PSMX2_EP_INIT_OP_CONTEXT(ep_priv);
*ep_out = ep_priv;
return 0;
errout_free_ep:
free(ep_priv);
errout:
return err;
}
static int psmx2_domain_init(struct psmx2_fid_domain *domain,
struct psmx2_src_name *src_addr)
{
int err;
psmx2_am_global_init();
psmx2_atomic_global_init();
domain->base_trx_ctxt = psmx2_trx_ctxt_alloc(domain, src_addr, -1);
if (!domain->base_trx_ctxt)
return -FI_ENODEV;
err = fastlock_init(&domain->mr_lock);
if (err) {
FI_WARN(&psmx2_prov, FI_LOG_CORE,
"fastlock_init(mr_lock) returns %d\n", err);
goto err_out_free_trx_ctxt;
}
domain->mr_map = rbtNew(&psmx2_key_compare);
if (!domain->mr_map) {
FI_WARN(&psmx2_prov, FI_LOG_CORE,
"rbtNew failed\n");
goto err_out_destroy_mr_lock;
}
domain->mr_reserved_key = 1;
err = fastlock_init(&domain->vl_lock);
if (err) {
FI_WARN(&psmx2_prov, FI_LOG_CORE,
"fastlock_init(vl_lock) returns %d\n", err);
goto err_out_delete_mr_map;
}
memset(domain->vl_map, 0, sizeof(domain->vl_map));
domain->vl_alloc = 0;
ofi_atomic_initialize32(&domain->sep_cnt, 0);
fastlock_init(&domain->sep_lock);
dlist_init(&domain->sep_list);
dlist_init(&domain->trx_ctxt_list);
fastlock_init(&domain->trx_ctxt_lock);
dlist_insert_before(&domain->base_trx_ctxt->entry, &domain->trx_ctxt_list);
/* Set active domain before psmx2_domain_enable_ep() installs the
* AM handlers to ensure that psmx2_active_fabric->active_domain
* is always non-NULL inside the handlers. Notice that the vlaue
* active_domain becomes NULL again only when the domain is closed.
* At that time the AM handlers are gone with the PSM endpoint.
*/
domain->fabric->active_domain = domain;
if (psmx2_domain_enable_ep(domain, NULL) < 0)
goto err_out_reset_active_domain;
if (domain->progress_thread_enabled)
psmx2_domain_start_progress(domain);
psmx2_am_init(domain->base_trx_ctxt);
return 0;
err_out_reset_active_domain:
domain->fabric->active_domain = NULL;
fastlock_destroy(&domain->vl_lock);
err_out_delete_mr_map:
rbtDelete(domain->mr_map);
err_out_destroy_mr_lock:
fastlock_destroy(&domain->mr_lock);
err_out_free_trx_ctxt:
psmx2_trx_ctxt_free(domain->base_trx_ctxt);
return err;
}
static int
usdf_fabric_open(struct fi_fabric_attr *fattrp, struct fid_fabric **fabric,
void *context)
{
struct fid_fabric *ff;
struct usdf_fabric *fp;
struct usdf_usnic_info *dp;
struct usdf_dev_entry *dep;
struct epoll_event ev;
struct sockaddr_in sin;
int ret;
int d;
USDF_TRACE("\n");
/* Make sure this fabric exists */
dp = __usdf_devinfo;
for (d = 0; d < dp->uu_num_devs; ++d) {
dep = &dp->uu_info[d];
if (dep->ue_dev_ok &&
usdf_fabric_checkname(0, &(dep->ue_dattr), fattrp->name)) {
break;
}
}
if (d >= dp->uu_num_devs) {
USDF_INFO("device \"%s\" does not exit, returning -FI_ENODEV\n",
fattrp->name);
return -FI_ENODEV;
}
fp = calloc(1, sizeof(*fp));
if (fp == NULL) {
USDF_INFO("unable to allocate memory for fabric\n");
return -FI_ENOMEM;
}
fp->fab_epollfd = -1;
fp->fab_arp_sockfd = -1;
LIST_INIT(&fp->fab_domain_list);
fp->fab_attr.fabric = fab_utof(fp);
fp->fab_attr.name = strdup(fattrp->name);
fp->fab_attr.prov_name = strdup(USDF_PROV_NAME);
fp->fab_attr.prov_version = USDF_PROV_VERSION;
if (fp->fab_attr.name == NULL ||
fp->fab_attr.prov_name == NULL) {
ret = -FI_ENOMEM;
goto fail;
}
fp->fab_fid.fid.fclass = FI_CLASS_FABRIC;
fp->fab_fid.fid.context = context;
fp->fab_fid.fid.ops = &usdf_fi_ops;
fp->fab_fid.ops = &usdf_ops_fabric;
fp->fab_dev_attrs = &dep->ue_dattr;
fp->fab_epollfd = epoll_create(1024);
if (fp->fab_epollfd == -1) {
ret = -errno;
USDF_INFO("unable to allocate epoll fd\n");
goto fail;
}
fp->fab_eventfd = eventfd(0, EFD_NONBLOCK | EFD_SEMAPHORE);
if (fp->fab_eventfd == -1) {
ret = -errno;
USDF_INFO("unable to allocate event fd\n");
goto fail;
}
fp->fab_poll_item.pi_rtn = usdf_fabric_progression_cb;
fp->fab_poll_item.pi_context = fp;
ev.events = EPOLLIN;
ev.data.ptr = &fp->fab_poll_item;
ret = epoll_ctl(fp->fab_epollfd, EPOLL_CTL_ADD, fp->fab_eventfd, &ev);
if (ret == -1) {
ret = -errno;
USDF_INFO("unable to EPOLL_CTL_ADD\n");
goto fail;
}
/* initialize timer subsystem */
ret = usdf_timer_init(fp);
if (ret != 0) {
USDF_INFO("unable to initialize timer\n");
goto fail;
}
/* create and bind socket for ARP resolution */
memset(&sin, 0, sizeof(sin));
sin.sin_family = AF_INET;
sin.sin_addr.s_addr = fp->fab_dev_attrs->uda_ipaddr_be;
fp->fab_arp_sockfd = socket(AF_INET, SOCK_DGRAM, 0);
if (fp->fab_arp_sockfd == -1) {
USDF_INFO("unable to create socket\n");
goto fail;
}
ret = bind(fp->fab_arp_sockfd, (struct sockaddr *) &sin, sizeof(sin));
if (ret == -1) {
ret = -errno;
goto fail;
}
ofi_atomic_initialize32(&fp->fab_refcnt, 0);
ofi_atomic_initialize32(&fp->num_blocked_waiting, 0);
ret = pthread_create(&fp->fab_thread, NULL,
usdf_fabric_progression_thread, fp);
if (ret != 0) {
ret = -ret;
USDF_INFO("unable to create progress thread\n");
goto fail;
}
fattrp->fabric = fab_utof(fp);
fattrp->prov_version = USDF_PROV_VERSION;
*fabric = fab_utof(fp);
USDF_INFO("successfully opened %s/%s\n", fattrp->name,
fp->fab_dev_attrs->uda_ifname);
return 0;
fail:
free(fp->fab_attr.name);
free(fp->fab_attr.prov_name);
ff = fab_utof(fp);
usdf_fabric_close(&ff->fid);
USDF_DBG("returning %d (%s)\n", ret, fi_strerror(-ret));
return ret;
}
int
usdf_domain_open(struct fid_fabric *fabric, struct fi_info *info,
struct fid_domain **domain, void *context)
{
struct usdf_fabric *fp;
struct usdf_domain *udp;
struct sockaddr_in *sin;
size_t addrlen;
int ret;
#if ENABLE_DEBUG
char requested[INET_ADDRSTRLEN], actual[INET_ADDRSTRLEN];
#endif
USDF_TRACE_SYS(DOMAIN, "\n");
sin = NULL;
fp = fab_fidtou(fabric);
if (info->domain_attr != NULL) {
/* No versioning information available here. */
if (!usdf_domain_checkname(0, fp->fab_dev_attrs,
info->domain_attr->name)) {
USDF_WARN_SYS(DOMAIN, "domain name mismatch\n");
return -FI_ENODATA;
}
if (ofi_check_mr_mode(fabric->api_version,
OFI_MR_BASIC_MAP | FI_MR_LOCAL,
info->domain_attr->mr_mode)) {
/* the caller ignored our fi_getinfo results */
USDF_WARN_SYS(DOMAIN, "MR mode (%d) not supported\n",
info->domain_attr->mr_mode);
return -FI_ENODATA;
}
}
udp = calloc(1, sizeof *udp);
if (udp == NULL) {
USDF_DBG("unable to alloc mem for domain\n");
ret = -FI_ENOMEM;
goto fail;
}
USDF_DBG("uda_devname=%s\n", fp->fab_dev_attrs->uda_devname);
/*
* Make sure address format is good and matches this fabric
*/
switch (info->addr_format) {
case FI_SOCKADDR:
addrlen = sizeof(struct sockaddr);
sin = info->src_addr;
break;
case FI_SOCKADDR_IN:
addrlen = sizeof(struct sockaddr_in);
sin = info->src_addr;
break;
case FI_ADDR_STR:
sin = usdf_format_to_sin(info, info->src_addr);
goto skip_size_check;
default:
ret = -FI_EINVAL;
goto fail;
}
if (info->src_addrlen != addrlen) {
ret = -FI_EINVAL;
goto fail;
}
skip_size_check:
if (sin->sin_family != AF_INET ||
sin->sin_addr.s_addr != fp->fab_dev_attrs->uda_ipaddr_be) {
USDF_DBG_SYS(DOMAIN, "requested src_addr (%s) != fabric addr (%s)\n",
inet_ntop(AF_INET, &sin->sin_addr.s_addr,
requested, sizeof(requested)),
inet_ntop(AF_INET, &fp->fab_dev_attrs->uda_ipaddr_be,
actual, sizeof(actual)));
ret = -FI_EINVAL;
usdf_free_sin_if_needed(info, sin);
goto fail;
}
usdf_free_sin_if_needed(info, sin);
ret = usd_open(fp->fab_dev_attrs->uda_devname, &udp->dom_dev);
if (ret != 0) {
goto fail;
}
udp->dom_fid.fid.fclass = FI_CLASS_DOMAIN;
udp->dom_fid.fid.context = context;
udp->dom_fid.fid.ops = &usdf_fid_ops;
udp->dom_fid.ops = &usdf_domain_ops;
udp->dom_fid.mr = &usdf_domain_mr_ops;
ret = pthread_spin_init(&udp->dom_progress_lock,
PTHREAD_PROCESS_PRIVATE);
if (ret != 0) {
ret = -ret;
goto fail;
}
TAILQ_INIT(&udp->dom_tx_ready);
TAILQ_INIT(&udp->dom_hcq_list);
udp->dom_info = fi_dupinfo(info);
if (udp->dom_info == NULL) {
ret = -FI_ENOMEM;
goto fail;
}
if (udp->dom_info->dest_addr != NULL) {
free(udp->dom_info->dest_addr);
udp->dom_info->dest_addr = NULL;
}
ret = usdf_dom_rdc_alloc_data(udp);
if (ret != 0) {
goto fail;
}
udp->dom_fabric = fp;
LIST_INSERT_HEAD(&fp->fab_domain_list, udp, dom_link);
ofi_atomic_initialize32(&udp->dom_refcnt, 0);
ofi_atomic_inc32(&fp->fab_refcnt);
*domain = &udp->dom_fid;
return 0;
fail:
if (udp != NULL) {
if (udp->dom_info != NULL) {
fi_freeinfo(udp->dom_info);
}
if (udp->dom_dev != NULL) {
usd_close(udp->dom_dev);
}
usdf_dom_rdc_free_data(udp);
free(udp);
}
return ret;
}
static ssize_t mrail_send(struct fid_ep *ep_fid, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr, void *context)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_send_common(ep_fid, &iov, &desc, 1, len, dest_addr, 0,
context, mrail_comp_flag(ep_fid));
}
static ssize_t mrail_inject(struct fid_ep *ep_fid, const void *buf, size_t len,
fi_addr_t dest_addr)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_send_common(ep_fid, &iov, NULL, 1, len, dest_addr, 0,
NULL, mrail_inject_flags(ep_fid));
}
static ssize_t mrail_injectdata(struct fid_ep *ep_fid, const void *buf,
size_t len, uint64_t data, fi_addr_t dest_addr)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_send_common(ep_fid, &iov, NULL, 1, len, dest_addr, data,
NULL, (mrail_inject_flags(ep_fid) |
FI_REMOTE_CQ_DATA));
}
static ssize_t
mrail_tsendmsg(struct fid_ep *ep_fid, const struct fi_msg_tagged *msg,
uint64_t flags)
{
return mrail_tsend_common(ep_fid, msg->msg_iov, msg->desc, msg->iov_count,
ofi_total_iov_len(msg->msg_iov, msg->iov_count),
msg->addr, msg->tag, msg->data, msg->context,
flags | mrail_comp_flag(ep_fid));
}
static ssize_t mrail_tsend(struct fid_ep *ep_fid, const void *buf, size_t len,
void *desc, fi_addr_t dest_addr, uint64_t tag,
void *context)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_tsend_common(ep_fid, &iov, &desc, 1, len, dest_addr, tag,
0, context, mrail_comp_flag(ep_fid));
}
static ssize_t mrail_tsenddata(struct fid_ep *ep_fid, const void *buf, size_t len,
void *desc, uint64_t data, fi_addr_t dest_addr,
uint64_t tag, void *context)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_tsend_common(ep_fid, &iov, &desc, 1, len, dest_addr, tag,
data, context, (mrail_comp_flag(ep_fid) |
FI_REMOTE_CQ_DATA));
}
static ssize_t mrail_tinject(struct fid_ep *ep_fid, const void *buf, size_t len,
fi_addr_t dest_addr, uint64_t tag)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_tsend_common(ep_fid, &iov, NULL, 1, len, dest_addr, tag,
0, NULL, mrail_inject_flags(ep_fid));
}
static ssize_t mrail_tinjectdata(struct fid_ep *ep_fid, const void *buf,
size_t len, uint64_t data, fi_addr_t dest_addr,
uint64_t tag)
{
struct iovec iov = {
.iov_base = (void *)buf,
.iov_len = len,
};
return mrail_tsend_common(ep_fid, &iov, NULL, 1, len, dest_addr, tag,
data, NULL, (mrail_inject_flags(ep_fid) |
FI_REMOTE_CQ_DATA));
}
static struct mrail_unexp_msg_entry *
mrail_get_unexp_msg_entry(struct mrail_recv_queue *recv_queue, void *context)
{
// TODO use buf pool
// context would be mrail_ep from which u can get the buf pool
struct mrail_unexp_msg_entry *unexp_msg_entry =
malloc(sizeof(*unexp_msg_entry) + sizeof(struct fi_cq_tagged_entry));
return unexp_msg_entry;
}
static int mrail_getname(fid_t fid, void *addr, size_t *addrlen)
{
struct mrail_ep *mrail_ep =
container_of(fid, struct mrail_ep, util_ep.ep_fid.fid);
struct mrail_domain *mrail_domain =
container_of(mrail_ep->util_ep.domain, struct mrail_domain,
util_domain);
size_t i, offset = 0, rail_addrlen;
int ret;
if (*addrlen < mrail_domain->addrlen)
return -FI_ETOOSMALL;
for (i = 0; i < mrail_ep->num_eps; i++) {
rail_addrlen = *addrlen - offset;
ret = fi_getname(&mrail_ep->rails[i].ep->fid,
(char *)addr + offset, &rail_addrlen);
if (ret) {
FI_WARN(&mrail_prov, FI_LOG_EP_CTRL,
"Unable to get name for rail: %zd\n", i);
return ret;
}
offset += rail_addrlen;
}
return 0;
}
static void mrail_tx_buf_init(void *pool_ctx, void *buf)
{
struct mrail_ep *mrail_ep = pool_ctx;
struct mrail_tx_buf *tx_buf = buf;
tx_buf->ep = mrail_ep;
tx_buf->hdr.version = MRAIL_HDR_VERSION;
}
static void mrail_ep_free_bufs(struct mrail_ep *mrail_ep)
{
if (mrail_ep->req_pool)
util_buf_pool_destroy(mrail_ep->req_pool);
if (mrail_ep->ooo_recv_pool)
util_buf_pool_destroy(mrail_ep->ooo_recv_pool);
if (mrail_ep->tx_buf_pool)
util_buf_pool_destroy(mrail_ep->tx_buf_pool);
if (mrail_ep->recv_fs)
mrail_recv_fs_free(mrail_ep->recv_fs);
}
static int mrail_ep_alloc_bufs(struct mrail_ep *mrail_ep)
{
struct util_buf_attr attr = {
.size = sizeof(struct mrail_tx_buf),
.alignment = sizeof(void *),
.max_cnt = 0,
.chunk_cnt = 64,
.alloc_hndlr = NULL,
.free_hndlr = NULL,
.init = mrail_tx_buf_init,
.ctx = mrail_ep,
};
size_t buf_size, rxq_total_size = 0;
struct fi_info *fi;
int ret;
for (fi = mrail_ep->info->next; fi; fi = fi->next)
rxq_total_size += fi->rx_attr->size;
mrail_ep->recv_fs = mrail_recv_fs_create(rxq_total_size, mrail_init_recv,
mrail_ep);
if (!mrail_ep->recv_fs)
return -FI_ENOMEM;
ret = util_buf_pool_create(&mrail_ep->ooo_recv_pool,
sizeof(struct mrail_ooo_recv),
sizeof(void *), 0, 64);
if (!mrail_ep->ooo_recv_pool)
goto err;
ret = util_buf_pool_create_attr(&attr, &mrail_ep->tx_buf_pool);
if (!mrail_ep->tx_buf_pool)
goto err;
buf_size = (sizeof(struct mrail_req) +
(mrail_ep->num_eps * sizeof(struct mrail_subreq)));
ret = util_buf_pool_create(&mrail_ep->req_pool, buf_size,
sizeof(void *), 0, 64);
if (ret)
goto err;
return 0;
err:
mrail_ep_free_bufs(mrail_ep);
return ret;
}
static int mrail_ep_close(fid_t fid)
{
struct mrail_ep *mrail_ep =
container_of(fid, struct mrail_ep, util_ep.ep_fid.fid);
int ret, retv = 0;
size_t i;
mrail_ep_free_bufs(mrail_ep);
for (i = 0; i < mrail_ep->num_eps; i++) {
ret = fi_close(&mrail_ep->rails[i].ep->fid);
if (ret)
retv = ret;
}
free(mrail_ep->rails);
ret = ofi_endpoint_close(&mrail_ep->util_ep);
if (ret)
retv = ret;
free(mrail_ep);
return retv;
}
static int mrail_ep_bind(struct fid *ep_fid, struct fid *bfid, uint64_t flags)
{
struct mrail_ep *mrail_ep =
container_of(ep_fid, struct mrail_ep, util_ep.ep_fid.fid);
struct mrail_cq *mrail_cq;
struct mrail_av *mrail_av;
struct util_cntr *cntr;
int ret = 0;
size_t i;
switch (bfid->fclass) {
case FI_CLASS_AV:
mrail_av = container_of(bfid, struct mrail_av,
util_av.av_fid.fid);
ret = ofi_ep_bind_av(&mrail_ep->util_ep, &mrail_av->util_av);
if (ret)
return ret;
for (i = 0; i < mrail_ep->num_eps; i++) {
ret = fi_ep_bind(mrail_ep->rails[i].ep,
&mrail_av->avs[i]->fid, flags);
if (ret)
return ret;
}
break;
case FI_CLASS_CQ:
mrail_cq = container_of(bfid, struct mrail_cq,
util_cq.cq_fid.fid);
ret = ofi_ep_bind_cq(&mrail_ep->util_ep, &mrail_cq->util_cq,
flags);
if (ret)
return ret;
for (i = 0; i < mrail_ep->num_eps; i++) {
ret = fi_ep_bind(mrail_ep->rails[i].ep,
&mrail_cq->cqs[i]->fid, flags);
if (ret)
return ret;
}
break;
case FI_CLASS_CNTR:
cntr = container_of(bfid, struct util_cntr, cntr_fid.fid);
ret = ofi_ep_bind_cntr(&mrail_ep->util_ep, cntr, flags);
if (ret)
return ret;
break;
case FI_CLASS_EQ:
ret = -FI_ENOSYS;
break;
default:
FI_WARN(&mrail_prov, FI_LOG_EP_CTRL, "invalid fid class\n");
ret = -FI_EINVAL;
break;
}
return ret;
}
static int mrail_ep_ctrl(struct fid *fid, int command, void *arg)
{
struct mrail_ep *mrail_ep;
size_t i, buf_recv_min = sizeof(struct mrail_hdr);
int ret;
mrail_ep = container_of(fid, struct mrail_ep, util_ep.ep_fid.fid);
switch (command) {
case FI_ENABLE:
if (!mrail_ep->util_ep.rx_cq || !mrail_ep->util_ep.tx_cq)
return -FI_ENOCQ;
if (!mrail_ep->util_ep.av)
return -FI_ENOAV;
for (i = 0; i < mrail_ep->num_eps; i++) {
ret = fi_setopt(&mrail_ep->rails[i].ep->fid,
FI_OPT_ENDPOINT, FI_OPT_BUFFERED_MIN,
&buf_recv_min, sizeof(buf_recv_min));
if (ret)
return ret;
ret = fi_enable(mrail_ep->rails[i].ep);
if (ret)
return ret;
}
break;
default:
return -FI_ENOSYS;
}
return 0;
}
static struct fi_ops mrail_ep_fi_ops = {
.size = sizeof(struct fi_ops),
.close = mrail_ep_close,
.bind = mrail_ep_bind,
.control = mrail_ep_ctrl,
.ops_open = fi_no_ops_open,
};
static int mrail_ep_setopt(fid_t fid, int level, int optname,
const void *optval, size_t optlen)
{
struct mrail_ep *mrail_ep;
size_t i;
int ret = 0;
mrail_ep = container_of(fid, struct mrail_ep, util_ep.ep_fid.fid);
for (i = 0; i < mrail_ep->num_eps; i++) {
ret = fi_setopt(&mrail_ep->rails[i].ep->fid, level, optname,
optval, optlen);
if (ret)
return ret;
}
return ret;
}
static struct fi_ops_ep mrail_ops_ep = {
.size = sizeof(struct fi_ops_ep),
.cancel = fi_no_cancel,
.getopt = fi_no_getopt,
.setopt = mrail_ep_setopt,
.tx_ctx = fi_no_tx_ctx,
.rx_ctx = fi_no_rx_ctx,
.rx_size_left = fi_no_rx_size_left,
.tx_size_left = fi_no_tx_size_left,
};
static struct fi_ops_cm mrail_ops_cm = {
.size = sizeof(struct fi_ops_cm),
.setname = fi_no_setname,
.getname = mrail_getname,
.getpeer = fi_no_getpeer,
.connect = fi_no_connect,
.listen = fi_no_listen,
.accept = fi_no_accept,
.reject = fi_no_reject,
.shutdown = fi_no_shutdown,
.join = fi_no_join,
};
static struct fi_ops_msg mrail_ops_msg = {
.size = sizeof(struct fi_ops_msg),
.recv = mrail_recv,
.recvv = fi_no_msg_recvv,
.recvmsg = mrail_recvmsg,
.send = mrail_send,
.sendv = fi_no_msg_sendv,
.sendmsg = mrail_sendmsg,
.inject = mrail_inject,
.senddata = fi_no_msg_senddata,
.injectdata = mrail_injectdata,
};
struct fi_ops_tagged mrail_ops_tagged = {
.size = sizeof(struct fi_ops_tagged),
.recv = mrail_trecv,
.recvv = fi_no_tagged_recvv,
.recvmsg = mrail_trecvmsg,
.send = mrail_tsend,
.sendv = fi_no_tagged_sendv,
.sendmsg = mrail_tsendmsg,
.inject = mrail_tinject,
.senddata = mrail_tsenddata,
.injectdata = mrail_tinjectdata,
};
void mrail_ep_progress(struct util_ep *ep)
{
struct mrail_ep *mrail_ep;
mrail_ep = container_of(ep, struct mrail_ep, util_ep);
mrail_progress_deferred_reqs(mrail_ep);
}
int mrail_ep_open(struct fid_domain *domain_fid, struct fi_info *info,
struct fid_ep **ep_fid, void *context)
{
struct mrail_domain *mrail_domain =
container_of(domain_fid, struct mrail_domain,
util_domain.domain_fid);
struct mrail_ep *mrail_ep;
struct fi_info *fi;
size_t i;
int ret;
if (strcmp(mrail_domain->info->domain_attr->name,
info->domain_attr->name)) {
FI_WARN(&mrail_prov, FI_LOG_EP_CTRL, "info domain name: %s "
"doesn't match fid_domain name: %s!\n",
info->domain_attr->name,
mrail_domain->info->domain_attr->name);
return -FI_EINVAL;
}
mrail_ep = calloc(1, sizeof(*mrail_ep));
if (!mrail_ep)
return -FI_ENOMEM;
// TODO detect changes b/w mrail_domain->info and info arg
// this may be difficult and we may not support such changes
mrail_ep->info = mrail_domain->info;
mrail_ep->num_eps = mrail_domain->num_domains;
ret = ofi_endpoint_init(domain_fid, &mrail_util_prov, info, &mrail_ep->util_ep,
context, &mrail_ep_progress);
if (ret) {
goto free_ep;
}
mrail_ep->rails = calloc(mrail_ep->num_eps, sizeof(*mrail_ep->rails));
if (!mrail_ep->rails) {
ret = -FI_ENOMEM;
goto err;
}
for (i = 0, fi = mrail_ep->info->next; fi; fi = fi->next, i++) {
fi->tx_attr->op_flags &= ~FI_COMPLETION;
ret = fi_endpoint(mrail_domain->domains[i], fi,
&mrail_ep->rails[i].ep, mrail_ep);
if (ret) {
FI_WARN(&mrail_prov, FI_LOG_EP_CTRL,
"Unable to open EP\n");
goto err;
}
mrail_ep->rails[i].info = fi;
}
ret = mrail_ep_alloc_bufs(mrail_ep);
if (ret)
goto err;
slist_init(&mrail_ep->deferred_reqs);
if (mrail_ep->info->caps & FI_DIRECTED_RECV) {
mrail_recv_queue_init(&mrail_prov, &mrail_ep->recv_queue,
mrail_match_recv_addr,
mrail_match_unexp_addr,
mrail_get_unexp_msg_entry);
mrail_recv_queue_init(&mrail_prov, &mrail_ep->trecv_queue,
mrail_match_recv_addr_tag,
mrail_match_unexp_addr_tag,
mrail_get_unexp_msg_entry);
} else {
mrail_recv_queue_init(&mrail_prov, &mrail_ep->recv_queue,
mrail_match_recv_any,
mrail_match_unexp_any,
mrail_get_unexp_msg_entry);
mrail_recv_queue_init(&mrail_prov, &mrail_ep->trecv_queue,
mrail_match_recv_tag,
mrail_match_unexp_tag,
mrail_get_unexp_msg_entry);
}
ofi_atomic_initialize32(&mrail_ep->tx_rail, 0);
ofi_atomic_initialize32(&mrail_ep->rx_rail, 0);
*ep_fid = &mrail_ep->util_ep.ep_fid;
(*ep_fid)->fid.ops = &mrail_ep_fi_ops;
(*ep_fid)->ops = &mrail_ops_ep;
(*ep_fid)->cm = &mrail_ops_cm;
(*ep_fid)->msg = &mrail_ops_msg;
(*ep_fid)->tagged = &mrail_ops_tagged;
(*ep_fid)->rma = &mrail_ops_rma;
return 0;
err:
mrail_ep_close(&mrail_ep->util_ep.ep_fid.fid);
free_ep:
free(mrail_ep);
return ret;
}
int psmx_ep_open(struct fid_domain *domain, struct fi_info *info,
struct fid_ep **ep, void *context)
{
struct psmx_fid_domain *domain_priv;
struct psmx_fid_ep *ep_priv;
int err;
uint64_t ep_cap;
if (info)
ep_cap = info->caps;
else
ep_cap = FI_TAGGED;
domain_priv = container_of(domain, struct psmx_fid_domain,
util_domain.domain_fid.fid);
if (!domain_priv)
return -FI_EINVAL;
if (info && info->ep_attr && info->ep_attr->auth_key) {
if (info->ep_attr->auth_key_size != sizeof(psm_uuid_t)) {
FI_WARN(&psmx_prov, FI_LOG_EP_CTRL,
"Invalid auth_key_len %d, should be %d.\n",
info->ep_attr->auth_key_size,
sizeof(psm_uuid_t));
return -FI_EINVAL;
}
if (memcmp(domain_priv->fabric->uuid, info->ep_attr->auth_key,
sizeof(psm_uuid_t))) {
FI_WARN(&psmx_prov, FI_LOG_EP_CTRL,
"Invalid auth_key: %s\n",
psmx_uuid_to_string((void *)info->ep_attr->auth_key));
return -FI_EINVAL;
}
}
err = psmx_domain_check_features(domain_priv, ep_cap);
if (err)
return err;
ep_priv = (struct psmx_fid_ep *) calloc(1, sizeof *ep_priv);
if (!ep_priv)
return -FI_ENOMEM;
ep_priv->ep.fid.fclass = FI_CLASS_EP;
ep_priv->ep.fid.context = context;
ep_priv->ep.fid.ops = &psmx_fi_ops;
ep_priv->ep.ops = &psmx_ep_ops;
ep_priv->ep.cm = &psmx_cm_ops;
ep_priv->domain = domain_priv;
ofi_atomic_initialize32(&ep_priv->ref, 0);
PSMX_CTXT_TYPE(&ep_priv->nocomp_send_context) = PSMX_NOCOMP_SEND_CONTEXT;
PSMX_CTXT_EP(&ep_priv->nocomp_send_context) = ep_priv;
PSMX_CTXT_TYPE(&ep_priv->nocomp_recv_context) = PSMX_NOCOMP_RECV_CONTEXT;
PSMX_CTXT_EP(&ep_priv->nocomp_recv_context) = ep_priv;
if (ep_cap & FI_TAGGED)
ep_priv->ep.tagged = &psmx_tagged_ops;
if (ep_cap & FI_MSG)
ep_priv->ep.msg = &psmx_msg_ops;
if ((ep_cap & FI_MSG) && psmx_env.am_msg)
ep_priv->ep.msg = &psmx_msg2_ops;
if (ep_cap & FI_RMA)
ep_priv->ep.rma = &psmx_rma_ops;
if (ep_cap & FI_ATOMICS)
ep_priv->ep.atomic = &psmx_atomic_ops;
ep_priv->caps = ep_cap;
err = psmx_domain_enable_ep(domain_priv, ep_priv);
if (err) {
free(ep_priv);
return err;
}
psmx_domain_acquire(domain_priv);
if (info) {
if (info->tx_attr)
ep_priv->tx_flags = info->tx_attr->op_flags;
if (info->rx_attr)
ep_priv->rx_flags = info->rx_attr->op_flags;
}
psmx_ep_optimize_ops(ep_priv);
ep_priv->service = PSMX_ANY_SERVICE;
if (info && info->src_addr)
ep_priv->service = ((struct psmx_src_name *)info->src_addr)->service;
if (ep_priv->service == PSMX_ANY_SERVICE)
ep_priv->service = ((getpid() & 0x7FFF) << 16) +
((uintptr_t)ep_priv & 0xFFFF);
ofi_ns_add_local_name(&ep_priv->domain->fabric->name_server,
&ep_priv->service, &ep_priv->domain->psm_epid);
*ep = &ep_priv->ep;
return 0;
}
void ofi_monitor_init(struct ofi_mem_monitor *monitor)
{
ofi_atomic_initialize32(&monitor->refcnt, 0);
}