static int fi_ibv_fabric_close(fid_t fid)
{
struct fi_ibv_fabric *fab;
fab = container_of(fid, struct fi_ibv_fabric, fabric_fid.fid);
util_buf_pool_destroy(fab->wce_pool);
util_buf_pool_destroy(fab->epe_pool);
free(fid);
return 0;
}
static int rxm_ep_txrx_res_open(struct rxm_ep *rxm_ep)
{
struct rxm_domain *rxm_domain;
uint8_t local_mr;
int ret;
rxm_domain = container_of(rxm_ep->util_ep.domain, struct rxm_domain, util_domain);
local_mr = rxm_ep->msg_info->mode & FI_LOCAL_MR ? 1 : 0;
FI_DBG(&rxm_prov, FI_LOG_EP_CTRL, "MSG provider mode & FI_LOCAL_MR: %d\n",
local_mr);
ret = rxm_buf_pool_create(local_mr, rxm_ep->msg_info->tx_attr->size,
sizeof(struct rxm_pkt), &rxm_ep->tx_pool, rxm_domain->msg_domain);
if (ret)
return ret;
ret = rxm_buf_pool_create(local_mr, rxm_ep->msg_info->rx_attr->size,
sizeof(struct rxm_rx_buf), &rxm_ep->rx_pool, rxm_domain->msg_domain);
if (ret)
goto err1;
rxm_ep->txe_fs = rxm_txe_fs_create(rxm_ep->rxm_info->tx_attr->size);
if (!rxm_ep->txe_fs) {
ret = -FI_ENOMEM;
goto err2;
}
ofi_key_idx_init(&rxm_ep->tx_key_idx, fi_size_bits(rxm_ep->rxm_info->tx_attr->size));
ret = rxm_recv_queue_init(&rxm_ep->recv_queue, rxm_ep->rxm_info->rx_attr->size);
if (ret)
goto err3;
ret = rxm_recv_queue_init(&rxm_ep->trecv_queue, rxm_ep->rxm_info->rx_attr->size);
if (ret)
goto err4;
return 0;
err4:
rxm_recv_queue_close(&rxm_ep->recv_queue);
err3:
rxm_txe_fs_free(rxm_ep->txe_fs);
err2:
util_buf_pool_destroy(rxm_ep->tx_pool);
err1:
util_buf_pool_destroy(rxm_ep->rx_pool);
return ret;
}
void ofi_mr_cache_cleanup(struct ofi_mr_cache *cache)
{
struct ofi_mr_entry *entry;
struct dlist_entry *tmp;
FI_INFO(cache->domain->prov, FI_LOG_MR, "MR cache stats: "
"searches %zu, deletes %zu, hits %zu\n",
cache->search_cnt, cache->delete_cnt, cache->hit_cnt);
util_mr_cache_process_events(cache);
dlist_foreach_container_safe(&cache->lru_list, struct ofi_mr_entry,
entry, lru_entry, tmp) {
assert(entry->use_cnt == 0);
util_mr_uncache_entry(cache, entry);
dlist_remove_init(&entry->lru_entry);
util_mr_free_entry(cache, entry);
}
cache->mr_storage.destroy(&cache->mr_storage);
ofi_monitor_del_queue(&cache->nq);
ofi_atomic_dec32(&cache->domain->ref);
util_buf_pool_destroy(cache->entry_pool);
assert(cache->cached_cnt == 0);
assert(cache->cached_size == 0);
}
static void rxm_ep_txrx_res_close(struct rxm_ep *rxm_ep)
{
struct slist_entry *entry;
struct rxm_rx_buf *rx_buf;
rxm_recv_queue_close(&rxm_ep->trecv_queue);
rxm_recv_queue_close(&rxm_ep->recv_queue);
if (rxm_ep->txe_fs)
rxm_txe_fs_free(rxm_ep->txe_fs);
while(!slist_empty(&rxm_ep->rx_buf_list)) {
entry = slist_remove_head(&rxm_ep->rx_buf_list);
rx_buf = container_of(entry, struct rxm_rx_buf, entry);
util_buf_release(rxm_ep->rx_pool, rx_buf);
}
util_buf_pool_destroy(rxm_ep->rx_pool);
util_buf_pool_destroy(rxm_ep->tx_pool);
}
int fi_ibv_fabric(struct fi_fabric_attr *attr, struct fid_fabric **fabric,
void *context)
{
struct fi_ibv_fabric *fab;
int ret;
ret = fi_ibv_init_info();
if (ret)
return ret;
ret = fi_ibv_find_fabric(attr);
if (ret)
return ret;
fab = calloc(1, sizeof(*fab));
if (!fab)
return -FI_ENOMEM;
fab->fabric_fid.fid.fclass = FI_CLASS_FABRIC;
fab->fabric_fid.fid.context = context;
fab->fabric_fid.fid.ops = &fi_ibv_fi_ops;
fab->fabric_fid.ops = &fi_ibv_ops_fabric;
fab->wce_pool = util_buf_pool_create(sizeof(struct fi_ibv_wce), 16, 0, VERBS_WCE_CNT);
if (!fab->wce_pool) {
FI_WARN(&fi_ibv_prov, FI_LOG_FABRIC, "Failed to create wce_pool\n");
ret = -FI_ENOMEM;
goto err1;
}
fab->epe_pool = util_buf_pool_create(sizeof(struct fi_ibv_msg_epe), 16, 0, VERBS_EPE_CNT);
if (!fab->epe_pool) {
FI_WARN(&fi_ibv_prov, FI_LOG_FABRIC, "Failed to create epe_pool\n");
ret = -FI_ENOMEM;
goto err2;
}
*fabric = &fab->fabric_fid;
return 0;
err2:
util_buf_pool_destroy(fab->wce_pool);
err1:
free(fab);
return ret;
}
static int rxd_cq_close(struct fid *fid)
{
int ret;
struct rxd_cq *cq;
cq = container_of(fid, struct rxd_cq, util_cq.cq_fid.fid);
fastlock_acquire(&cq->domain->lock);
dlist_remove(&cq->dom_entry);
fastlock_release(&cq->domain->lock);
fastlock_destroy(&cq->lock);
ret = fi_close(&cq->dg_cq->fid);
if (ret)
return ret;
ret = ofi_cq_cleanup(&cq->util_cq);
if (ret)
return ret;
util_buf_pool_destroy(cq->unexp_pool);
free(cq);
return 0;
}
int tcpx_progress_close(struct tcpx_progress *progress)
{
util_buf_pool_destroy(progress->pe_entry_pool);
return FI_SUCCESS;
}
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;
}