int fi_ibv_process_xrc_connreq(struct fi_ibv_ep *ep,
struct fi_ibv_connreq *connreq)
{
struct fi_ibv_xrc_ep *xrc_ep = container_of(ep, struct fi_ibv_xrc_ep,
base_ep);
int ret;
assert(ep->info->src_addr);
assert(ep->info->dest_addr);
xrc_ep->conn_setup = calloc(1, sizeof(*xrc_ep->conn_setup));
if (!xrc_ep->conn_setup)
return -FI_ENOMEM;
/* This endpoint was created on the passive side of a connection
* request. The reciprocal connection request will go back to the
* passive port indicated by the active side */
ofi_addr_set_port(ep->info->src_addr, 0);
ofi_addr_set_port(ep->info->dest_addr, connreq->xrc.port);
ret = fi_ibv_create_ep(NULL, NULL, 0, ep->info, NULL, &ep->id);
if (ret) {
VERBS_WARN(FI_LOG_EP_CTRL,
"Creation of INI cm_id failed %d\n", ret);
goto create_err;
}
xrc_ep->tgt_id = connreq->id;
xrc_ep->tgt_id->context = &ep->util_ep.ep_fid.fid;
return FI_SUCCESS;
create_err:
free(xrc_ep->conn_setup);
return ret;
}
static int fi_ibv_get_param_int(char *param_name, char *param_str,
size_t *param_default)
{
char *param_help;
size_t len, ret;
int param;
len = strlen(param_str) + 50;
param_help = malloc(len);
ret = snprintf(param_help, len, "%s (default: %zu)", param_str,
*param_default);
if (ret >= len) {
VERBS_WARN(FI_LOG_EP_DATA,
"param_help string size insufficient!\n");
free(param_help);
assert(0);
return -FI_ETOOSMALL;
}
fi_param_define(&fi_ibv_prov, param_name, FI_PARAM_INT, param_help);
if (!fi_param_get_int(&fi_ibv_prov, param_name, ¶m))
*param_default = param;
free(param_help);
return 0;
}
static int fi_ibv_rai_to_fi(struct rdma_addrinfo *rai, struct fi_info *fi)
{
fi->addr_format = ofi_translate_addr_format(rai->ai_family);
if (fi->addr_format == FI_FORMAT_UNSPEC) {
VERBS_WARN(FI_LOG_FABRIC, "Unknown address format\n");
return -FI_EINVAL;
}
if (rai->ai_src_len) {
free(fi->src_addr);
if (!(fi->src_addr = malloc(rai->ai_src_len)))
return -FI_ENOMEM;
memcpy(fi->src_addr, rai->ai_src_addr, rai->ai_src_len);
fi->src_addrlen = rai->ai_src_len;
}
if (rai->ai_dst_len) {
free(fi->dest_addr);
if (!(fi->dest_addr = malloc(rai->ai_dst_len)))
return -FI_ENOMEM;
memcpy(fi->dest_addr, rai->ai_dst_addr, rai->ai_dst_len);
fi->dest_addrlen = rai->ai_dst_len;
}
return 0;
}
int fi_ibv_verify_xrc_cm_data(struct fi_ibv_xrc_cm_data *remote,
int private_data_len)
{
if (sizeof(*remote) > private_data_len) {
VERBS_WARN(FI_LOG_EP_CTRL,
"XRC MSG EP CM data length mismatch\n");
return -FI_EINVAL;
}
if (remote->version != FI_IBV_XRC_VERSION) {
VERBS_WARN(FI_LOG_EP_CTRL,
"XRC MSG EP connection protocol mismatch "
"(local %"PRIu8", remote %"PRIu8")\n",
FI_IBV_XRC_VERSION, remote->version);
return -FI_EINVAL;
}
return FI_SUCCESS;
}
static int fi_ibv_get_srcaddr_devs(struct fi_info **info)
{
struct fi_info *fi, *add_info;
struct fi_info *fi_unconf = NULL, *fi_prev = NULL;
struct verbs_dev_info *dev;
struct verbs_addr *addr;
int ret = 0;
DEFINE_LIST(verbs_devs);
ret = fi_ibv_getifaddrs(&verbs_devs);
if (ret)
return ret;
if (dlist_empty(&verbs_devs)) {
VERBS_WARN(FI_LOG_CORE, "No interface address found\n");
return 0;
}
for (fi = *info; fi; fi = fi->next) {
dlist_foreach_container(&verbs_devs, struct verbs_dev_info, dev, entry)
if (!strncmp(fi->domain_attr->name, dev->name, strlen(dev->name))) {
dlist_foreach_container(&dev->addrs, struct verbs_addr, addr, entry) {
/* When a device has multiple interfaces/addresses configured
* duplicate fi_info and add the address info. fi->src_addr
* would have been set in the previous iteration */
if (fi->src_addr) {
if (!(add_info = fi_dupinfo(fi))) {
ret = -FI_ENOMEM;
goto out;
}
add_info->next = fi->next;
fi->next = add_info;
fi = add_info;
}
ret = fi_ibv_rai_to_fi(addr->rai, fi);
if (ret)
goto out;
}
break;
}
}
int fi_ibv_connect_xrc(struct fi_ibv_xrc_ep *ep, struct sockaddr *addr,
int reciprocal, void *param, size_t paramlen)
{
struct fi_ibv_domain *domain = fi_ibv_ep_to_domain(&ep->base_ep);
struct sockaddr *peer_addr;
int ret;
assert(ep->base_ep.id && !ep->base_ep.ibv_qp && !ep->ini_conn);
peer_addr = rdma_get_local_addr(ep->base_ep.id);
if (peer_addr)
ofi_straddr_dbg(&fi_ibv_prov, FI_LOG_FABRIC,
"XRC connect src_addr", peer_addr);
peer_addr = rdma_get_peer_addr(ep->base_ep.id);
if (peer_addr)
ofi_straddr_dbg(&fi_ibv_prov, FI_LOG_FABRIC,
"XRC connect dest_addr", peer_addr);
if (!reciprocal) {
ep->conn_setup = calloc(1, sizeof(*ep->conn_setup));
if (!ep->conn_setup)
return -FI_ENOMEM;
}
fastlock_acquire(&domain->xrc.ini_mgmt_lock);
ret = fi_ibv_get_shared_ini_conn(ep, &ep->ini_conn);
if (ret) {
VERBS_WARN(FI_LOG_FABRIC,
"Get of shared XRC INI connection failed %d\n", ret);
fastlock_release(&domain->xrc.ini_mgmt_lock);
if (!reciprocal) {
free(ep->conn_setup);
ep->conn_setup = NULL;
}
return ret;
}
fi_ibv_add_pending_ini_conn(ep, reciprocal, param, paramlen);
fi_ibv_sched_ini_conn(ep->ini_conn);
fastlock_release(&domain->xrc.ini_mgmt_lock);
return FI_SUCCESS;
}
static inline int
fi_ibv_poll_events(struct fi_ibv_cq *_cq, int timeout)
{
int ret, rc;
void *context;
struct pollfd fds[2];
char data;
fds[0].fd = _cq->channel->fd;
fds[1].fd = _cq->signal_fd[0];
fds[0].events = fds[1].events = POLLIN;
rc = poll(fds, 2, timeout);
if (rc == 0)
return -FI_EAGAIN;
else if (rc < 0)
return -errno;
if (fds[0].revents & POLLIN) {
ret = ibv_get_cq_event(_cq->channel, &_cq->cq, &context);
if (ret)
return ret;
ofi_atomic_inc32(&_cq->nevents);
rc--;
}
if (fds[1].revents & POLLIN) {
do {
ret = read(fds[1].fd, &data, 1);
} while (ret > 0);
ret = -FI_EAGAIN;
rc--;
}
if (rc) {
VERBS_WARN(FI_LOG_CQ, "Unknown poll error: check revents\n");
return -FI_EOTHER;
}
return ret;
}
static int fi_ibv_reap_comp(struct fi_ibv_msg_ep *ep)
{
struct fi_ibv_wce *wce = NULL;
int got_wc = 0;
int ret = 0;
fastlock_acquire(&ep->scq->lock);
while (ofi_atomic_get32(&ep->comp_pending) > 0) {
if (!wce) {
wce = util_buf_alloc(ep->scq->wce_pool);
if (!wce) {
fastlock_release(&ep->scq->lock);
return -FI_ENOMEM;
}
memset(wce, 0, sizeof(*wce));
}
ret = fi_ibv_poll_cq(ep->scq, &wce->wc);
if (ret < 0) {
VERBS_WARN(FI_LOG_EP_DATA,
"Failed to read completion for signaled send\n");
util_buf_release(ep->scq->wce_pool, wce);
fastlock_release(&ep->scq->lock);
return ret;
} else if (ret > 0) {
slist_insert_tail(&wce->entry, &ep->scq->wcq);
got_wc = 1;
wce = NULL;
}
}
if (wce)
util_buf_release(ep->scq->wce_pool, wce);
if (got_wc && ep->scq->channel)
ret = fi_ibv_cq_signal(&ep->scq->cq_fid);
fastlock_release(&ep->scq->lock);
return ret;
}
void fi_ibv_prev_xrc_conn_state(struct fi_ibv_xrc_ep *ep)
{
switch (ep->conn_state) {
case FI_IBV_XRC_UNCONNECTED:
break;
case FI_IBV_XRC_ORIG_CONNECTING:
ep->conn_state = FI_IBV_XRC_UNCONNECTED;
break;
case FI_IBV_XRC_ORIG_CONNECTED:
ep->conn_state = FI_IBV_XRC_ORIG_CONNECTING;
break;
case FI_IBV_XRC_RECIP_CONNECTING:
ep->conn_state = FI_IBV_XRC_ORIG_CONNECTED;
break;
case FI_IBV_XRC_CONNECTED:
ep->conn_state = FI_IBV_XRC_RECIP_CONNECTING;
break;
default:
assert(0);
VERBS_WARN(FI_LOG_FABRIC, "Unkown XRC connection state %d\n",
ep->conn_state);
}
}
static int fi_ibv_msg_ep_enable(struct fid_ep *ep)
{
struct ibv_qp_init_attr attr;
struct fi_ibv_msg_ep *_ep;
struct ibv_pd *pd;
_ep = container_of(ep, struct fi_ibv_msg_ep, ep_fid);
if (!_ep->eq) {
VERBS_WARN(FI_LOG_EP_CTRL,
"Endpoint is not bound to an event queue\n");
return -FI_ENOEQ;
}
if (!_ep->scq && !_ep->rcq) {
VERBS_WARN(FI_LOG_EP_CTRL, "Endpoint is not bound to "
"a send or receive completion queue\n");
return -FI_ENOCQ;
}
if (!_ep->scq && (ofi_send_allowed(_ep->info->caps) ||
ofi_rma_initiate_allowed(_ep->info->caps))) {
VERBS_WARN(FI_LOG_EP_CTRL, "Endpoint is not bound to "
"a send completion queue when it has transmit "
"capabilities enabled (FI_SEND | FI_RMA).\n");
return -FI_ENOCQ;
}
if (!_ep->rcq && ofi_recv_allowed(_ep->info->caps)) {
VERBS_WARN(FI_LOG_EP_CTRL, "Endpoint is not bound to "
"a receive completion queue when it has receive "
"capabilities enabled. (FI_RECV)\n");
return -FI_ENOCQ;
}
memset(&attr, 0, sizeof attr);
if (_ep->scq) {
attr.cap.max_send_wr = _ep->info->tx_attr->size;
attr.cap.max_send_sge = _ep->info->tx_attr->iov_limit;
attr.send_cq = _ep->scq->cq;
pd = _ep->scq->domain->pd;
} else {
attr.send_cq = _ep->rcq->cq;
pd = _ep->rcq->domain->pd;
}
if (_ep->rcq) {
attr.cap.max_recv_wr = _ep->info->rx_attr->size;
attr.cap.max_recv_sge = _ep->info->rx_attr->iov_limit;
attr.recv_cq = _ep->rcq->cq;
} else {
attr.recv_cq = _ep->scq->cq;
}
attr.cap.max_inline_data = _ep->info->tx_attr->inject_size;
if (_ep->srq_ep) {
attr.srq =_ep->srq_ep->srq;
/* Use of SRQ, no need to allocate recv_wr entries in the QP */
attr.cap.max_recv_wr = 0;
/* Override the default ops to prevent the user from posting WRs to a
* QP where a SRQ is attached to */
_ep->ep_fid.msg = fi_ibv_msg_srq_ep_ops_msg(_ep);
}
attr.qp_type = IBV_QPT_RC;
attr.sq_sig_all = 0;
attr.qp_context = _ep;
return rdma_create_qp(_ep->id, pd, &attr) ?
-errno : 0;
}
static int
fi_ibv_mr_reg(struct fid *fid, const void *buf, size_t len,
uint64_t access, uint64_t offset, uint64_t requested_key,
uint64_t flags, struct fid_mr **mr, void *context)
{
struct fi_ibv_mem_desc *md;
int fi_ibv_access = 0;
struct fid_domain *domain;
if (flags)
return -FI_EBADFLAGS;
if (fid->fclass != FI_CLASS_DOMAIN) {
return -FI_EINVAL;
}
domain = container_of(fid, struct fid_domain, fid);
md = calloc(1, sizeof *md);
if (!md)
return -FI_ENOMEM;
md->domain = container_of(domain, struct fi_ibv_domain, domain_fid);
md->mr_fid.fid.fclass = FI_CLASS_MR;
md->mr_fid.fid.context = context;
md->mr_fid.fid.ops = &fi_ibv_mr_ops;
/* Enable local write access by default for FI_EP_RDM which hides local
* registration requirements. This allows to avoid buffering or double
* registration */
if (!(md->domain->info->caps & FI_LOCAL_MR))
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
/* Local read access to an MR is enabled by default in verbs */
if (access & FI_RECV)
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
/* iWARP spec requires Remote Write access for an MR that is used
* as a data sink for a Remote Read */
if (access & FI_READ) {
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
if (md->domain->verbs->device->transport_type == IBV_TRANSPORT_IWARP)
fi_ibv_access |= IBV_ACCESS_REMOTE_WRITE;
}
if (access & FI_WRITE)
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE;
if (access & FI_REMOTE_READ)
fi_ibv_access |= IBV_ACCESS_REMOTE_READ;
/* Verbs requires Local Write access too for Remote Write access */
if (access & FI_REMOTE_WRITE)
fi_ibv_access |= IBV_ACCESS_LOCAL_WRITE |
IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_ATOMIC;
md->mr = ibv_reg_mr(md->domain->pd, (void *) buf, len, fi_ibv_access);
if (!md->mr)
goto err;
md->mr_fid.mem_desc = (void *) (uintptr_t) md->mr->lkey;
md->mr_fid.key = md->mr->rkey;
*mr = &md->mr_fid;
if(md->domain->eq && (md->domain->eq_flags & FI_REG_MR)) {
struct fi_eq_entry entry = {
.fid = &md->mr_fid.fid,
.context = context
};
fi_ibv_eq_write_event(md->domain->eq, FI_MR_COMPLETE,
&entry, sizeof(entry));
}
return 0;
err:
free(md);
return -errno;
}
static int fi_ibv_mr_regv(struct fid *fid, const struct iovec * iov,
size_t count, uint64_t access, uint64_t offset, uint64_t requested_key,
uint64_t flags, struct fid_mr **mr, void *context)
{
if (count > VERBS_MR_IOV_LIMIT) {
VERBS_WARN(FI_LOG_FABRIC,
"iov count > %d not supported\n",
VERBS_MR_IOV_LIMIT);
return -FI_EINVAL;
}
return fi_ibv_mr_reg(fid, iov->iov_base, iov->iov_len, access, offset,
requested_key, flags, mr, context);
}
static int fi_ibv_mr_regattr(struct fid *fid, const struct fi_mr_attr *attr,
uint64_t flags, struct fid_mr **mr)
{
return fi_ibv_mr_regv(fid, attr->mr_iov, attr->iov_count, attr->access,
0, attr->requested_key, flags, mr, attr->context);
}
static int fi_ibv_domain_bind(struct fid *fid, struct fid *bfid, uint64_t flags)
{
struct fi_ibv_domain *domain;
struct fi_ibv_eq *eq;
domain = container_of(fid, struct fi_ibv_domain, domain_fid.fid);
switch (bfid->fclass) {
case FI_CLASS_EQ:
eq = container_of(bfid, struct fi_ibv_eq, eq_fid);
domain->eq = eq;
domain->eq_flags = flags;
break;
default:
return -EINVAL;
}
return 0;
}
static int fi_ibv_domain_close(fid_t fid)
{
struct fi_ibv_domain *domain;
int ret;
domain = container_of(fid, struct fi_ibv_domain, domain_fid.fid);
if (domain->rdm) {
rdma_destroy_ep(domain->rdm_cm->listener);
free(domain->rdm_cm);
}
if (domain->pd) {
ret = ibv_dealloc_pd(domain->pd);
if (ret)
return -ret;
domain->pd = NULL;
}
fi_freeinfo(domain->info);
free(domain);
return 0;
}
static int fi_ibv_open_device_by_name(struct fi_ibv_domain *domain, const char *name)
{
struct ibv_context **dev_list;
int i, ret = -FI_ENODEV;
if (!name)
return -FI_EINVAL;
dev_list = rdma_get_devices(NULL);
if (!dev_list)
return -errno;
for (i = 0; dev_list[i] && ret; i++) {
if (domain->rdm) {
ret = strncmp(name, ibv_get_device_name(dev_list[i]->device),
strlen(name) - strlen(verbs_rdm_domain.suffix));
} else {
ret = strcmp(name, ibv_get_device_name(dev_list[i]->device));
}
if (!ret)
domain->verbs = dev_list[i];
}
rdma_free_devices(dev_list);
return ret;
}
static struct fi_ops fi_ibv_fid_ops = {
.size = sizeof(struct fi_ops),
.close = fi_ibv_domain_close,
.bind = fi_ibv_domain_bind,
.control = fi_no_control,
.ops_open = fi_no_ops_open,
};
static struct fi_ops_mr fi_ibv_domain_mr_ops = {
.size = sizeof(struct fi_ops_mr),
.reg = fi_ibv_mr_reg,
.regv = fi_ibv_mr_regv,
.regattr = fi_ibv_mr_regattr,
};
static struct fi_ops_domain fi_ibv_domain_ops = {
.size = sizeof(struct fi_ops_domain),
.av_open = fi_no_av_open,
.cq_open = fi_ibv_cq_open,
.endpoint = fi_ibv_open_ep,
.scalable_ep = fi_no_scalable_ep,
.cntr_open = fi_no_cntr_open,
.poll_open = fi_no_poll_open,
.stx_ctx = fi_no_stx_context,
.srx_ctx = fi_ibv_srq_context,
};
static struct fi_ops_domain fi_ibv_rdm_domain_ops = {
.size = sizeof(struct fi_ops_domain),
.av_open = fi_ibv_rdm_av_open,
.cq_open = fi_ibv_rdm_cq_open,
.endpoint = fi_ibv_rdm_open_ep,
.scalable_ep = fi_no_scalable_ep,
.cntr_open = fi_rbv_rdm_cntr_open,
.poll_open = fi_no_poll_open,
.stx_ctx = fi_no_stx_context,
.srx_ctx = fi_no_srx_context,
};
static int
fi_ibv_domain(struct fid_fabric *fabric, struct fi_info *info,
struct fid_domain **domain, void *context)
{
struct fi_ibv_domain *_domain;
struct fi_ibv_fabric *fab;
struct fi_info *fi;
int ret;
fi = fi_ibv_get_verbs_info(info->domain_attr->name);
if (!fi)
return -FI_EINVAL;
fab = container_of(fabric, struct fi_ibv_fabric, util_fabric.fabric_fid);
ret = ofi_check_domain_attr(&fi_ibv_prov, fabric->api_version,
fi->domain_attr, info->domain_attr);
if (ret)
return ret;
_domain = calloc(1, sizeof *_domain);
if (!_domain)
return -FI_ENOMEM;
_domain->info = fi_dupinfo(info);
if (!_domain->info)
goto err1;
_domain->rdm = FI_IBV_EP_TYPE_IS_RDM(info);
if (_domain->rdm) {
_domain->rdm_cm = calloc(1, sizeof(*_domain->rdm_cm));
if (!_domain->rdm_cm) {
ret = -FI_ENOMEM;
goto err2;
}
}
ret = fi_ibv_open_device_by_name(_domain, info->domain_attr->name);
if (ret)
goto err2;
_domain->pd = ibv_alloc_pd(_domain->verbs);
if (!_domain->pd) {
ret = -errno;
goto err2;
}
_domain->domain_fid.fid.fclass = FI_CLASS_DOMAIN;
_domain->domain_fid.fid.context = context;
_domain->domain_fid.fid.ops = &fi_ibv_fid_ops;
_domain->domain_fid.mr = &fi_ibv_domain_mr_ops;
if (_domain->rdm) {
_domain->domain_fid.ops = &fi_ibv_rdm_domain_ops;
_domain->rdm_cm->ec = rdma_create_event_channel();
if (!_domain->rdm_cm->ec) {
VERBS_INFO(FI_LOG_EP_CTRL,
"Failed to create listener event channel: %s\n",
strerror(errno));
ret = -FI_EOTHER;
goto err2;
}
if (fi_fd_nonblock(_domain->rdm_cm->ec->fd) != 0) {
VERBS_INFO_ERRNO(FI_LOG_EP_CTRL, "fcntl", errno);
ret = -FI_EOTHER;
goto err3;
}
if (rdma_create_id(_domain->rdm_cm->ec,
&_domain->rdm_cm->listener, NULL, RDMA_PS_TCP))
{
VERBS_INFO(FI_LOG_EP_CTRL, "Failed to create cm listener: %s\n",
strerror(errno));
ret = -FI_EOTHER;
goto err3;
}
_domain->rdm_cm->is_bound = 0;
} else {
_domain->domain_fid.ops = &fi_ibv_domain_ops;
}
_domain->fab = fab;
*domain = &_domain->domain_fid;
return 0;
err3:
if (_domain->rdm)
rdma_destroy_event_channel(_domain->rdm_cm->ec);
err2:
if (_domain->rdm)
free(_domain->rdm_cm);
fi_freeinfo(_domain->info);
err1:
free(_domain);
return ret;
}
static int fi_ibv_trywait(struct fid_fabric *fabric, struct fid **fids, int count)
{
struct fi_ibv_cq *cq;
int ret, i;
for (i = 0; i < count; i++) {
switch (fids[i]->fclass) {
case FI_CLASS_CQ:
cq = container_of(fids[i], struct fi_ibv_cq, cq_fid.fid);
ret = cq->trywait(fids[i]);
if (ret)
return ret;
break;
case FI_CLASS_EQ:
/* We are always ready to wait on an EQ since
* rdmacm EQ is based on an fd */
continue;
case FI_CLASS_CNTR:
case FI_CLASS_WAIT:
return -FI_ENOSYS;
default:
return -FI_EINVAL;
}
}
return FI_SUCCESS;
}
static int fi_ibv_fabric_close(fid_t fid)
{
struct fi_ibv_fabric *fab;
int ret;
fab = container_of(fid, struct fi_ibv_fabric, util_fabric.fabric_fid.fid);
ret = ofi_fabric_close(&fab->util_fabric);
if (ret)
return ret;
free(fab);
return 0;
}
static struct fi_ops fi_ibv_fi_ops = {
.size = sizeof(struct fi_ops),
.close = fi_ibv_fabric_close,
.bind = fi_no_bind,
.control = fi_no_control,
.ops_open = fi_no_ops_open,
};
static struct fi_ops_fabric fi_ibv_ops_fabric = {
.size = sizeof(struct fi_ops_fabric),
.domain = fi_ibv_domain,
.passive_ep = fi_ibv_passive_ep,
.eq_open = fi_ibv_eq_open,
.wait_open = fi_no_wait_open,
.trywait = fi_ibv_trywait
};
int fi_ibv_fabric(struct fi_fabric_attr *attr, struct fid_fabric **fabric,
void *context)
{
struct fi_ibv_fabric *fab;
struct fi_info *info;
int ret;
ret = fi_ibv_init_info();
if (ret)
return ret;
fab = calloc(1, sizeof(*fab));
if (!fab)
return -FI_ENOMEM;
for (info = verbs_info; info; info = info->next) {
ret = ofi_fabric_init(&fi_ibv_prov, info->fabric_attr, attr,
&fab->util_fabric, context);
if (ret != -FI_ENODATA)
break;
}
if (ret) {
free(fab);
return ret;
}
*fabric = &fab->util_fabric.fabric_fid;
(*fabric)->fid.ops = &fi_ibv_fi_ops;
(*fabric)->ops = &fi_ibv_ops_fabric;
return 0;
}
/* Builds a list of interfaces that correspond to active verbs devices */
static int fi_ibv_getifaddrs(struct dlist_entry *verbs_devs)
{
struct ifaddrs *ifaddr, *ifa;
char name[INET6_ADDRSTRLEN];
struct rdma_addrinfo *rai;
struct rdma_cm_id *id;
const char *ret_ptr;
int ret, num_verbs_ifs = 0;
char *iface = NULL;
size_t iface_len = 0;
int exact_match = 0;
ret = getifaddrs(&ifaddr);
if (ret) {
VERBS_WARN(FI_LOG_FABRIC,
"Unable to get interface addresses\n");
return ret;
}
/* select best iface name based on user's input */
if (fi_param_get_str(&fi_ibv_prov, "iface", &iface) == FI_SUCCESS) {
iface_len = strlen(iface);
if (iface_len > IFNAMSIZ) {
VERBS_INFO(FI_LOG_EP_CTRL,
"Too long iface name: %s, max: %d\n",
iface, IFNAMSIZ);
return -FI_EINVAL;
}
for (ifa = ifaddr; ifa && !exact_match; ifa = ifa->ifa_next)
exact_match = !strcmp(ifa->ifa_name, iface);
}
for (ifa = ifaddr; ifa; ifa = ifa->ifa_next) {
if (!ifa->ifa_addr || !(ifa->ifa_flags & IFF_UP) ||
!strcmp(ifa->ifa_name, "lo"))
continue;
if(iface) {
if(exact_match) {
if(strcmp(ifa->ifa_name, iface))
continue;
} else {
if(strncmp(ifa->ifa_name, iface, iface_len))
continue;
}
}
switch (ifa->ifa_addr->sa_family) {
case AF_INET:
ret_ptr = inet_ntop(AF_INET, &ofi_sin_addr(ifa->ifa_addr),
name, INET6_ADDRSTRLEN);
break;
case AF_INET6:
ret_ptr = inet_ntop(AF_INET6, &ofi_sin6_addr(ifa->ifa_addr),
name, INET6_ADDRSTRLEN);
break;
default:
continue;
}
if (!ret_ptr) {
VERBS_WARN(FI_LOG_FABRIC,
"inet_ntop failed: %s(%d)\n",
strerror(errno), errno);
goto err1;
}
ret = fi_ibv_create_ep(name, NULL, FI_NUMERICHOST | FI_SOURCE,
NULL, &rai, &id);
if (ret)
continue;
ret = fi_ibv_add_rai(verbs_devs, id, rai);
if (ret)
goto err2;
VERBS_DBG(FI_LOG_FABRIC, "Found active interface for verbs device: "
"%s with address: %s\n",
ibv_get_device_name(id->verbs->device), name);
rdma_destroy_ep(id);
num_verbs_ifs++;
}
freeifaddrs(ifaddr);
return num_verbs_ifs ? 0 : -FI_ENODATA;
err2:
rdma_destroy_ep(id);
err1:
fi_ibv_verbs_devs_free(verbs_devs);
freeifaddrs(ifaddr);
return ret;
}