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
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) {
FI_WARN(&fi_ibv_prov, 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_info *fi;
int ret;
fi = fi_ibv_get_verbs_info(info->domain_attr->name);
if (!fi)
return -FI_EINVAL;
ret = ofi_check_domain_attr(&fi_ibv_prov, fi->domain_attr,
info->domain_attr, FI_MATCH_EXACT);
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 = container_of(fabric, struct fi_ibv_fabric, fabric_fid);
*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)
{
free(fid);
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;
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;
*fabric = &fab->fabric_fid;
return 0;
}