/*
* dapl_hca_alloc
*
* alloc and initialize an HCA struct
*
* Input:
* name
* port
*
* Output:
* hca_ptr
*
* Returns:
* none
*
*/
DAPL_HCA *dapl_hca_alloc(char *name, char *port)
{
DAPL_HCA *hca_ptr;
hca_ptr = dapl_os_alloc(sizeof(DAPL_HCA));
if (NULL == hca_ptr) {
goto bail;
}
dapl_os_memzero(hca_ptr, sizeof(DAPL_HCA));
dapl_os_lock_init(&hca_ptr->lock);
dapl_llist_init_head(&hca_ptr->ia_list_head);
hca_ptr->name = dapl_os_strdup(name);
if (NULL == hca_ptr->name) {
goto bail;
}
hca_ptr->ib_hca_handle = IB_INVALID_HANDLE;
hca_ptr->port_num = dapl_os_strtol(port, NULL, 0);
return (hca_ptr);
bail:
if (NULL != hca_ptr)
dapl_os_free(hca_ptr, sizeof(DAPL_HCA));
return NULL;
}
DAPL_LMR *dapl_lmr_alloc(IN DAPL_IA * ia,
IN DAT_MEM_TYPE mem_type,
IN DAT_REGION_DESCRIPTION region_desc,
IN DAT_VLEN length,
IN DAT_PZ_HANDLE pz_handle,
IN DAT_MEM_PRIV_FLAGS mem_priv)
{
DAPL_LMR *lmr;
/* Allocate LMR */
lmr = (DAPL_LMR *) dapl_os_alloc(sizeof(DAPL_LMR));
if (NULL == lmr) {
return (NULL);
}
/* zero the structure */
dapl_os_memzero(lmr, sizeof(DAPL_LMR));
/*
* initialize the header
*/
lmr->header.provider = ia->header.provider;
lmr->header.magic = DAPL_MAGIC_LMR;
lmr->header.handle_type = DAT_HANDLE_TYPE_LMR;
lmr->header.owner_ia = ia;
lmr->header.user_context.as_64 = 0;
lmr->header.user_context.as_ptr = NULL;
dapl_llist_init_entry(&lmr->header.ia_list_entry);
dapl_ia_link_lmr(ia, lmr);
dapl_os_lock_init(&lmr->header.lock);
/*
* initialize the body
*/
lmr->param.ia_handle = (DAT_IA_HANDLE) ia;
lmr->param.mem_type = mem_type;
lmr->param.region_desc = region_desc;
lmr->param.length = length;
lmr->param.pz_handle = pz_handle;
lmr->param.mem_priv = mem_priv;
dapl_os_atomic_set(&lmr->lmr_ref_count, 0);
return (lmr);
}
/*
* dapl_cno_alloc
*
* alloc and initialize an EVD struct
*
* Input:
* ia
*
* Returns:
* cno_ptr, or null on failure.
*/
DAPL_CNO *
dapl_cno_alloc(
IN DAPL_IA *ia_ptr,
IN DAT_OS_WAIT_PROXY_AGENT wait_agent)
{
DAPL_CNO *cno_ptr;
cno_ptr = (DAPL_CNO *) dapl_os_alloc(sizeof (DAPL_CNO));
if (!cno_ptr) {
return (NULL);
}
/* zero the structure */
(void) dapl_os_memzero(cno_ptr, sizeof (DAPL_CNO));
/*
* Initialize the header.
*/
cno_ptr->header.provider = ia_ptr->header.provider;
cno_ptr->header.magic = DAPL_MAGIC_CNO;
cno_ptr->header.handle_type = DAT_HANDLE_TYPE_CNO;
cno_ptr->header.owner_ia = ia_ptr;
cno_ptr->header.user_context.as_64 = 0;
cno_ptr->header.user_context.as_ptr = NULL;
dapl_llist_init_entry(&cno_ptr->header.ia_list_entry);
dapl_llist_init_head(&cno_ptr->evd_list_head);
dapl_os_lock_init(&cno_ptr->header.lock);
/*
* Initialize the body
*/
cno_ptr->cno_waiters = 0;
cno_ptr->cno_ref_count = 0;
cno_ptr->cno_state = DAPL_CNO_STATE_UNTRIGGERED;
cno_ptr->cno_evd_triggered = NULL;
cno_ptr->cno_wait_agent = wait_agent;
(void) dapl_os_wait_object_init(&cno_ptr->cno_wait_object);
return (cno_ptr);
}
/*
* dapl_sp_alloc
*
* alloc and initialize a PSP INFO struct
*
* Input:
* IA INFO struct ptr
*
* Output:
* sp_ptr
*
* Returns:
* NULL
* pointer to sp info struct
*
*/
DAPL_SP *
dapls_sp_alloc(
IN DAPL_IA *ia_ptr,
IN DAT_BOOLEAN is_psp)
{
DAPL_SP *sp_ptr;
/* Allocate EP */
sp_ptr = (DAPL_SP *)dapl_os_alloc(sizeof (DAPL_SP));
if (sp_ptr == NULL) {
return (NULL);
}
/* zero the structure */
(void) dapl_os_memzero(sp_ptr, sizeof (DAPL_SP));
/*
* initialize the header
*/
sp_ptr->header.provider = ia_ptr->header.provider;
if (is_psp) {
sp_ptr->header.magic = DAPL_MAGIC_PSP;
sp_ptr->header.handle_type = DAT_HANDLE_TYPE_PSP;
} else {
sp_ptr->header.magic = DAPL_MAGIC_RSP;
sp_ptr->header.handle_type = DAT_HANDLE_TYPE_RSP;
}
sp_ptr->header.owner_ia = ia_ptr;
sp_ptr->header.user_context.as_64 = 0;
sp_ptr->header.user_context.as_ptr = NULL;
dapl_llist_init_entry(&sp_ptr->header.ia_list_entry);
dapl_os_lock_init(&sp_ptr->header.lock);
/*
* Initialize the Body (set to NULL above)
*/
dapl_llist_init_head(&sp_ptr->cr_list_head);
return (sp_ptr);
}
/*
* dapl_ep_connect
*
* DAPL Requirements Version xxx, 6.5.7
*
* Request a connection be established between the local Endpoint
* and a remote Endpoint. This operation is used by the active/client
* side of a connection
*
* Input:
* ep_handle
* remote_ia_address
* remote_conn_qual
* timeout
* private_data_size
* privaet_data
* qos
* connect_flags
*
* Output:
* None
*
* Returns:
* DAT_SUCCESS
* DAT_INSUFFICIENT_RESOUCRES
* DAT_INVALID_PARAMETER
* DAT_MODLE_NOT_SUPPORTED
*/
DAT_RETURN
dapl_ep_connect(
IN DAT_EP_HANDLE ep_handle,
IN DAT_IA_ADDRESS_PTR remote_ia_address,
IN DAT_CONN_QUAL remote_conn_qual,
IN DAT_TIMEOUT timeout,
IN DAT_COUNT private_data_size,
IN const DAT_PVOID private_data,
IN DAT_QOS qos,
IN DAT_CONNECT_FLAGS connect_flags)
{
DAPL_EP *ep_ptr;
DAPL_PRIVATE prd;
DAPL_EP alloc_ep;
DAT_RETURN dat_status;
dapl_dbg_log(DAPL_DBG_TYPE_API | DAPL_DBG_TYPE_CM,
"dapl_ep_connect (%p, {%u.%u.%u.%u}, %X, %d, %d, %p, %x, %x)\n",
ep_handle,
remote_ia_address->sa_data[2],
remote_ia_address->sa_data[3],
remote_ia_address->sa_data[4],
remote_ia_address->sa_data[5],
remote_conn_qual,
timeout,
private_data_size,
private_data,
qos,
connect_flags);
dat_status = DAT_SUCCESS;
ep_ptr = (DAPL_EP *) ep_handle;
/*
* Verify parameter & state. The connection handle must be good
* at this point.
*/
if (DAPL_BAD_HANDLE(ep_ptr, DAPL_MAGIC_EP)) {
dat_status = DAT_ERROR(DAT_INVALID_HANDLE,
DAT_INVALID_HANDLE_EP);
goto bail;
}
if (DAPL_BAD_HANDLE(ep_ptr->param.connect_evd_handle, DAPL_MAGIC_EVD)) {
dat_status = DAT_ERROR(DAT_INVALID_HANDLE,
DAT_INVALID_HANDLE_EVD_CONN);
goto bail;
}
/*
* If the endpoint needs a QP, associated the QP with it.
* This needs to be done carefully, in order to:
* * Avoid allocating under a lock.
* * Not step on data structures being altered by
* routines with which we are racing.
* So we:
* * Confirm that a new QP is needed and is not forbidden by the
* current state.
* * Allocate it into a separate EP.
* * Take the EP lock.
* * Reconfirm that the EP is in a state where it needs a QP.
* * Assign the QP and release the lock.
*/
if (ep_ptr->qp_state == DAPL_QP_STATE_UNATTACHED) {
if (ep_ptr->param.pz_handle == NULL ||
DAPL_BAD_HANDLE(ep_ptr->param.pz_handle, DAPL_MAGIC_PZ)) {
dat_status = DAT_ERROR(DAT_INVALID_STATE,
DAT_INVALID_STATE_EP_NOTREADY);
goto bail;
}
alloc_ep = *ep_ptr;
dat_status = dapls_ib_qp_alloc(ep_ptr->header.owner_ia,
&alloc_ep, ep_ptr);
if (dat_status != DAT_SUCCESS) {
dat_status = DAT_ERROR(DAT_INSUFFICIENT_RESOURCES,
DAT_RESOURCE_MEMORY);
goto bail;
}
dapl_os_lock(&ep_ptr->header.lock);
/*
* PZ shouldn't have changed since we're only racing with
* dapl_cr_accept()
*/
if (ep_ptr->qp_state != DAPL_QP_STATE_UNATTACHED) {
/* Bail, cleaning up. */
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = dapls_ib_qp_free(ep_ptr->header.owner_ia,
&alloc_ep);
if (dat_status != DAT_SUCCESS) {
dapl_dbg_log(DAPL_DBG_TYPE_WARN,
"ep_connect: ib_qp_free failed with %x\n",
dat_status);
}
dat_status = DAT_ERROR(DAT_INVALID_STATE,
dapls_ep_state_subtype(ep_ptr));
goto bail;
}
ep_ptr->qp_handle = alloc_ep.qp_handle;
ep_ptr->qpn = alloc_ep.qpn;
ep_ptr->qp_state = alloc_ep.qp_state;
dapl_os_unlock(&ep_ptr->header.lock);
}
/*
* We do state checks and transitions under lock.
* The only code we're racing against is dapl_cr_accept.
*/
dapl_os_lock(&ep_ptr->header.lock);
/*
* Verify the attributes of the EP handle before we connect it. Test
* all of the handles to make sure they are currently valid.
* Specifically:
* pz_handle required
* recv_evd_handle optional, but must be valid
* request_evd_handle optional, but must be valid
* connect_evd_handle required
*/
if (ep_ptr->param.pz_handle == NULL ||
DAPL_BAD_HANDLE(ep_ptr->param.pz_handle, DAPL_MAGIC_PZ) ||
ep_ptr->param.connect_evd_handle == NULL ||
DAPL_BAD_HANDLE(ep_ptr->param.connect_evd_handle,
DAPL_MAGIC_EVD) ||
!(((DAPL_EVD *)ep_ptr->param.connect_evd_handle)->evd_flags &
DAT_EVD_CONNECTION_FLAG) ||
(ep_ptr->param.recv_evd_handle != DAT_HANDLE_NULL &&
(DAPL_BAD_HANDLE(ep_ptr->param.recv_evd_handle,
DAPL_MAGIC_EVD))) ||
(ep_ptr->param.request_evd_handle != DAT_HANDLE_NULL &&
(DAPL_BAD_HANDLE(ep_ptr->param.request_evd_handle,
DAPL_MAGIC_EVD)))) {
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_INVALID_STATE,
DAT_INVALID_STATE_EP_NOTREADY);
goto bail;
}
/*
* Check both the EP state and the QP state: if we don't have a QP
* we need to attach one now.
*/
if (ep_ptr->qp_state == DAPL_QP_STATE_UNATTACHED) {
dat_status = dapls_ib_qp_alloc(ep_ptr->header.owner_ia,
ep_ptr, ep_ptr);
if (dat_status != DAT_SUCCESS) {
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_INSUFFICIENT_RESOURCES,
DAT_RESOURCE_TEP);
goto bail;
}
}
if (ep_ptr->param.ep_state != DAT_EP_STATE_UNCONNECTED) {
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_INVALID_STATE,
dapls_ep_state_subtype(ep_ptr));
goto bail;
}
if (qos != DAT_QOS_BEST_EFFORT ||
connect_flags != DAT_CONNECT_DEFAULT_FLAG) {
/*
* At this point we only support one QOS level
*/
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_MODEL_NOT_SUPPORTED, 0);
goto bail;
}
/*
* Verify the private data size doesn't exceed the max
*/
if (private_data_size > DAPL_CONSUMER_MAX_PRIVATE_DATA_SIZE) {
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_INVALID_PARAMETER, DAT_INVALID_ARG5);
goto bail;
}
/*
* transition the state before requesting a connection to avoid
* race conditions
*/
ep_ptr->param.ep_state = DAT_EP_STATE_ACTIVE_CONNECTION_PENDING;
/*
* At this point we're committed, and done with the endpoint
* except for the connect, so we can drop the lock.
*/
dapl_os_unlock(&ep_ptr->header.lock);
/*
* fill in the private data
*/
(void) dapl_os_memzero(&prd, sizeof (DAPL_PRIVATE));
if (private_data_size > 0)
(void) dapl_os_memcpy(prd.private_data, private_data,
private_data_size);
/* Copy the connection qualifiers */
(void) dapl_os_memcpy(ep_ptr->param.remote_ia_address_ptr,
remote_ia_address, sizeof (DAT_SOCK_ADDR6));
ep_ptr->param.remote_port_qual = remote_conn_qual;
dat_status = dapls_ib_connect(ep_handle,
remote_ia_address, remote_conn_qual,
private_data_size, &prd, timeout);
if (dat_status != DAT_SUCCESS) {
DAPL_EVD *evd_ptr;
if (dat_status == DAT_ERROR(DAT_INVALID_ADDRESS,
DAT_INVALID_ADDRESS_UNREACHABLE)) {
/* Unreachable IP address */
evd_ptr = (DAPL_EVD *)ep_ptr->param.connect_evd_handle;
if (evd_ptr != NULL) {
(void) dapls_evd_post_connection_event(evd_ptr,
DAT_CONNECTION_EVENT_UNREACHABLE,
(DAT_HANDLE) ep_ptr, 0, 0);
}
ep_ptr->param.ep_state = DAT_EP_STATE_DISCONNECTED;
dat_status = DAT_SUCCESS;
} else if (dat_status == DAT_ERROR(DAT_INVALID_PARAMETER,
DAT_INVALID_ADDRESS_UNREACHABLE)) {
/* Non-existant connection qualifier */
evd_ptr = (DAPL_EVD *)ep_ptr->param.connect_evd_handle;
if (evd_ptr != NULL) {
(void) dapls_evd_post_connection_event(evd_ptr,
DAT_CONNECTION_EVENT_NON_PEER_REJECTED,
(DAT_HANDLE) ep_ptr, 0, 0);
}
ep_ptr->param.ep_state = DAT_EP_STATE_DISCONNECTED;
dat_status = DAT_SUCCESS;
} else {
ep_ptr->param.ep_state = DAT_EP_STATE_UNCONNECTED;
}
}
bail:
dapl_dbg_log(DAPL_DBG_TYPE_RTN | DAPL_DBG_TYPE_CM,
"dapl_ep_connect () returns 0x%x\n", dat_status);
return (dat_status);
}
/*
* dapl_cr_accept
*
* DAPL Requirements Version xxx, 6.4.2.1
*
* Establish a connection between active remote side requesting Endpoint
* and passic side local Endpoint.
*
* Input:
* cr_handle
* ep_handle
* private_data_size
* private_data
*
* Output:
* none
*
* Returns:
* DAT_SUCCESS
* DAT_INVALID_PARAMETER
* DAT_INVALID_ATTRIBUTE
*/
DAT_RETURN
dapl_cr_accept(
IN DAT_CR_HANDLE cr_handle,
IN DAT_EP_HANDLE ep_handle,
IN DAT_COUNT private_data_size,
IN const DAT_PVOID private_data)
{
DAPL_EP *ep_ptr;
DAT_RETURN dat_status;
DAPL_PRIVATE prd;
DAPL_CR *cr_ptr;
DAT_EP_STATE entry_ep_state;
DAT_EP_HANDLE entry_ep_handle;
dapl_dbg_log(DAPL_DBG_TYPE_API,
"dapl_cr_accept(%p, %p, %d, %p)\n",
cr_handle, ep_handle, private_data_size, private_data);
if (DAPL_BAD_HANDLE(cr_handle, DAPL_MAGIC_CR)) {
dat_status = DAT_ERROR(DAT_INVALID_HANDLE,
DAT_INVALID_HANDLE_CR);
goto bail;
}
cr_ptr = (DAPL_CR *) cr_handle;
/*
* Return an error if we have an ep_handle and the CR already has an
* EP, indicating this is an RSP connection or PSP_PROVIDER_FLAG was
* specified.
*/
if (ep_handle != NULL &&
(DAPL_BAD_HANDLE(ep_handle, DAPL_MAGIC_EP) ||
cr_ptr->param.local_ep_handle != NULL)) {
dat_status = DAT_ERROR(DAT_INVALID_HANDLE,
DAT_INVALID_HANDLE_EP);
goto bail;
}
if ((0 != private_data_size) && (NULL == private_data)) {
dat_status = DAT_ERROR(DAT_INVALID_PARAMETER, DAT_INVALID_ARG4);
goto bail;
}
/*
* Verify the private data size doesn't exceed the max
*/
if (private_data_size > DAPL_CONSUMER_MAX_PRIVATE_DATA_SIZE) {
dat_status = DAT_ERROR(DAT_INVALID_PARAMETER, DAT_INVALID_ARG3);
goto bail;
}
/*
* ep_handle is NULL if the user specified DAT_PSP_PROVIDER_FLAG
* OR this is an RSP connection; retrieve it from the cr.
*/
if (ep_handle == NULL) {
ep_handle = cr_ptr->param.local_ep_handle;
if ((((DAPL_EP *) ep_handle)->param.ep_state !=
DAT_EP_STATE_TENTATIVE_CONNECTION_PENDING) &&
(((DAPL_EP *)ep_handle)->param.ep_state !=
DAT_EP_STATE_PASSIVE_CONNECTION_PENDING)) {
return (DAT_INVALID_STATE);
}
} else {
/* ensure this EP isn't connected or in use */
if (((DAPL_EP *)ep_handle)->param.ep_state !=
DAT_EP_STATE_UNCONNECTED) {
return (DAT_INVALID_STATE);
}
}
ep_ptr = (DAPL_EP *) ep_handle;
/*
* Verify the attributes of the EP handle before we connect it. Test
* all of the handles to make sure they are currently valid.
* Specifically:
* pz_handle required
* recv_evd_handle optional, but must be valid
* request_evd_handle optional, but must be valid
* connect_evd_handle required
* We do all verification and state change under lock, at which
* point the EP state should protect us from most races.
*/
dapl_os_lock(&ep_ptr->header.lock);
if ((ep_ptr->param.pz_handle == NULL) ||
DAPL_BAD_HANDLE(ep_ptr->param.pz_handle, DAPL_MAGIC_PZ) ||
(ep_ptr->param.connect_evd_handle == NULL) ||
DAPL_BAD_HANDLE(ep_ptr->param.connect_evd_handle, DAPL_MAGIC_EVD) ||
!(((DAPL_EVD *)ep_ptr->param.connect_evd_handle)->evd_flags &
DAT_EVD_CONNECTION_FLAG) ||
(ep_ptr->param.recv_evd_handle != DAT_HANDLE_NULL &&
(DAPL_BAD_HANDLE(ep_ptr->param.recv_evd_handle, DAPL_MAGIC_EVD))) ||
(ep_ptr->param.request_evd_handle != DAT_HANDLE_NULL &&
(DAPL_BAD_HANDLE(ep_ptr->param.request_evd_handle,
DAPL_MAGIC_EVD)))) {
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_INVALID_HANDLE,
DAT_INVALID_HANDLE_EP);
goto bail;
}
if (ep_ptr->qp_state == DAPL_QP_STATE_UNATTACHED) {
/*
* If we are lazy attaching the QP then we may need to
* hook it up here. Typically, we run this code only for
* DAT_PSP_PROVIDER_FLAG
*/
dat_status = dapls_ib_qp_alloc(cr_ptr->header.owner_ia, ep_ptr,
ep_ptr);
if (dat_status != DAT_SUCCESS) {
/* This is not a great error code, but spec allows */
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = DAT_ERROR(DAT_INVALID_HANDLE,
DAT_INVALID_HANDLE_EP);
goto bail;
}
}
entry_ep_state = ep_ptr->param.ep_state;
entry_ep_handle = cr_ptr->param.local_ep_handle;
ep_ptr->param.ep_state = DAT_EP_STATE_COMPLETION_PENDING;
ep_ptr->cm_handle = cr_ptr->ib_cm_handle;
ep_ptr->cr_ptr = cr_ptr;
ep_ptr->param.remote_ia_address_ptr = cr_ptr->param.
remote_ia_address_ptr;
cr_ptr->param.local_ep_handle = ep_handle;
/*
* private data
*/
(void) dapl_os_memcpy(prd.private_data, private_data,
private_data_size);
(void) dapl_os_memzero(prd.private_data + private_data_size,
sizeof (DAPL_PRIVATE) - private_data_size);
dapl_os_unlock(&ep_ptr->header.lock);
dat_status = dapls_ib_accept_connection(cr_handle, ep_handle, &prd);
/*
* If the provider failed, unwind the damage so we are back at
* the initial state.
*/
if (dat_status != DAT_SUCCESS) {
ep_ptr->param.ep_state = entry_ep_state;
cr_ptr->param.local_ep_handle = entry_ep_handle;
} else {
/*
* Make this CR invalid. We need to hang on to it until
* the connection terminates, but it's destroyed from
* the app point of view.
*/
cr_ptr->header.magic = DAPL_MAGIC_CR_DESTROYED;
}
bail:
return (dat_status);
}
DAT_RETURN
dapls_ib_connect(IN DAT_EP_HANDLE ep_handle,
IN DAT_IA_ADDRESS_PTR remote_ia_address, IN DAT_CONN_QUAL remote_conn_qual,
IN DAT_COUNT prd_size, IN DAPL_PRIVATE *prd_ptr, IN DAT_TIMEOUT timeout)
{
dapl_ep_connect_t args;
DAPL_EP *ep_p = (DAPL_EP *)ep_handle;
struct sockaddr *s;
char addr_buf[64];
ib_gid_t dgid;
int retval;
struct sockaddr_in6 *v6addr;
struct sockaddr_in *v4addr;
dapl_ia_addr_t *sap;
s = (struct sockaddr *)remote_ia_address;
dapl_dbg_log(DAPL_DBG_TYPE_CM,
"dapls_ib_connect: ep 0x%p\n"
" addr %s, conn_qual %016llu, ep_hkey %016llx\n"
" prd_size %d, timeout 0x%x\n",
ep_p, dapls_inet_ntop(s, addr_buf, 64), remote_conn_qual,
ep_p->qp_handle->ep_hkey, prd_size, timeout);
if (ep_p->qp_handle == NULL) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"dapls_ib_connect: ep 0x%p, addr %s, conn_qual %016llu, "
"qp_handle == NULL\n", ep_p, dapls_inet_ntop(s,
addr_buf, 64), remote_conn_qual);
return (DAT_INVALID_PARAMETER);
}
if (timeout == DAT_TIMEOUT_INFINITE) {
args.epc_timeout = 0;
} else {
args.epc_timeout = timeout;
}
/* resolve remote address to dgid */
retval = dapls_ns_lookup_address(ep_p->header.owner_ia,
remote_ia_address, timeout, &dgid);
if (retval == DAT_SUCCESS) {
args.epc_dgid = dgid;
} else if ((retval & DAT_SUBTYPE_MASK) ==
DAT_INVALID_ADDRESS_UNREACHABLE) {
/* let the kernel driver look up the dgid from ATS */
args.epc_dgid.gid_guid = 0ULL;
args.epc_dgid.gid_prefix = 0ULL;
} else {
return (retval);
}
args.epc_sid = remote_conn_qual;
args.epc_hkey = ep_p->qp_handle->ep_hkey;
sap = (dapl_ia_addr_t *)&args.epc_raddr_sadata;
/*
* filled in the remote_ia_address for consistent though
* not necessary when dapls_ns_lookup_address has resolved the dgid
*/
switch (s->sa_family) {
case AF_INET:
/* LINTED: E_BAD_PTR_CAST_ALIGN */
v4addr = (struct sockaddr_in *)s;
sap->iad_v4pad[0] = 0;
sap->iad_v4pad[1] = 0;
sap->iad_v4pad[2] = 0;
sap->iad_v4 = v4addr->sin_addr;
break;
case AF_INET6:
/* LINTED: E_BAD_PTR_CAST_ALIGN */
v6addr = (struct sockaddr_in6 *)s;
sap->iad_v6 = v6addr->sin6_addr;
break;
}
/* establish the hello message */
(void) dapl_os_memzero((void *)&prd_ptr->hello_msg,
sizeof (DAPL_HELLO_MSG));
/* on ATS leave the msg blank to avoid confusion to 3rd parties */
if ((args.epc_dgid.gid_guid | args.epc_dgid.gid_prefix)) {
prd_ptr->hello_msg.hi_checksum = DAPL_CHECKSUM;
prd_ptr->hello_msg.hi_clen = prd_size;
prd_ptr->hello_msg.hi_mid = 0;
prd_ptr->hello_msg.hi_vers = DAPL_HELLO_MSG_VERS;
/* fill in local address */
s = (struct sockaddr *)
&ep_p->header.owner_ia->hca_ptr->hca_address;
prd_ptr->hello_msg.hi_ipv = (uint8_t)s->sa_family;
switch (s->sa_family) {
case AF_INET:
/* LINTED: E_BAD_PTR_CAST_ALIGN */
v4addr = (struct sockaddr_in *)s;
prd_ptr->hello_msg.hi_port = v4addr->sin_port;
prd_ptr->hello_msg.hi_v4ipaddr = v4addr->sin_addr;
break;
case AF_INET6:
/* LINTED: E_BAD_PTR_CAST_ALIGN */
v6addr = (struct sockaddr_in6 *)s;
prd_ptr->hello_msg.hi_port = v6addr->sin6_port;
prd_ptr->hello_msg.hi_v6ipaddr = v6addr->sin6_addr;
break;
default:
break; /* fall through */
}
}
if (prd_size > 0) {
(void) dapl_os_memcpy((void *)&args.epc_priv[0],
(void *)prd_ptr, sizeof (DAPL_PRIVATE));
} else {
(void) dapl_os_memcpy((void *)
&args.epc_priv[DAPL_CONSUMER_MAX_PRIVATE_DATA_SIZE],
(void *)&prd_ptr->hello_msg, sizeof (DAPL_HELLO_MSG));
}
args.epc_priv_sz = sizeof (DAPL_PRIVATE);
retval = ioctl(ep_p->header.owner_ia->hca_ptr->ib_hca_handle->ia_fd,
DAPL_EP_CONNECT, &args);
if (retval != 0) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"dapls_ib_connect: connect failed %s, retval %d\n\n",
strerror(errno), retval);
return (dapls_convert_error(errno, retval));
}
dapl_dbg_log(DAPL_DBG_TYPE_CM,
"dapls_ib_connect: connected to %s\n\n",
dapls_inet_ntop(s, addr_buf, 64));
return (DAT_SUCCESS);
}
static int
dapls_ns_resolve_addr(int af, struct sockaddr *addr, DAT_TIMEOUT timeout)
{
struct sockaddr_storage sock;
struct sockaddr_in *v4dest;
struct sockaddr_in6 *v6dest;
struct pollfd pollfd;
int fd, retval;
int tmo;
int ip_version;
if (af == AF_INET) {
ip_version = 4;
} else if (af == AF_INET6) {
ip_version = 6;
} else {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_resolve_addr: invalid af %d\n", af);
return (-1);
}
fd = socket(af, SOCK_STREAM, 0);
if (fd < 0) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_resolve_addr: ipv%d, cannot create socket %s\n",
ip_version, strerror(errno));
return (-1);
}
/*
* set socket to non-blocking mode
*/
retval = fcntl(fd, F_SETFL, O_NONBLOCK);
if (retval < 0) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_resolve_addr: ipv%d, fcntl failed: %s\n",
ip_version, strerror(errno));
(void) close(fd);
return (-1);
}
/*
* connect to the discard port (9) at the dest IP
*/
(void) dapl_os_memzero(&sock, sizeof (sock));
if (af == AF_INET) {
v4dest = (struct sockaddr_in *)&sock;
v4dest->sin_family = AF_INET;
v4dest->sin_addr.s_addr =
/* LINTED: E_BAD_PTR_CAST_ALIGN */
((struct sockaddr_in *)addr)->sin_addr.s_addr;
v4dest->sin_port = htons(9);
retval = connect(fd, (struct sockaddr *)v4dest,
sizeof (struct sockaddr_in));
} else {
v6dest = (struct sockaddr_in6 *)&sock;
v6dest->sin6_family = AF_INET6;
/* LINTED: E_BAD_PTR_CAST_ALIGN */
(void) dapl_os_memcpy(&v6dest->sin6_addr,
&((struct sockaddr_in6 *)addr)->sin6_addr,
sizeof (struct sockaddr_in6));
v6dest->sin6_port = htons(9);
retval = connect(fd, (struct sockaddr *)v6dest,
sizeof (struct sockaddr_in6));
}
/*
* we can return immediately if connect succeeds
*/
if (retval == 0) {
(void) close(fd);
return (0);
}
/*
* receiving a RST means that the arp/nd entry should
* already be resolved
*/
if (retval < 0 && errno == ECONNREFUSED) {
errno = 0;
(void) close(fd);
return (0);
}
/*
* for all other cases, we poll on the fd
*/
pollfd.fd = fd;
pollfd.events = POLLIN | POLLOUT;
pollfd.revents = 0;
if (timeout == DAT_TIMEOUT_INFINITE ||
timeout == 0) {
/*
* -1 means infinite
*/
tmo = -1;
} else {
/*
* convert timeout from usecs to msecs
*/
tmo = timeout/1000;
}
retval = poll(&pollfd, 1, tmo);
if (retval > 0) {
int so_error = 0, len = sizeof (so_error);
retval = getsockopt(fd, SOL_SOCKET, SO_ERROR,
&so_error, &len);
if (retval == 0) {
/*
* we only return 0 if so_error == 0 or
* so_error == ECONNREFUSED. for all other
* cases retval is non-zero.
*/
if (so_error != 0 && so_error != ECONNREFUSED) {
retval = -1;
errno = so_error;
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_resolve_addr: ipv%d, so_error: %s\n",
ip_version, strerror(errno));
}
} else {
/*
* if retval != 0, it must be -1. and errno must
* have been set by getsockopt.
*/
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_resolve_addr: ipv%d, getsockopt: %s\n",
ip_version, strerror(errno));
}
} else {
if (retval == 0) {
errno = ETIMEDOUT;
}
retval = -1;
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_resolve_addr: ipv%d, poll: %s\n",
ip_version, strerror(errno));
}
(void) close(fd);
return (retval);
}
DAT_RETURN
dapls_ns_lookup_v6(
IN DAPL_IA *ia_ptr,
IN struct sockaddr_in6 *addr,
IN DAT_TIMEOUT timeout,
OUT ib_gid_t *gid)
{
struct lifreq lifr;
uchar_t *mac;
int s, retries = 0;
s = socket(AF_INET6, SOCK_DGRAM, 0);
if (s < 0) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_lookup_v6: socket: %s\n", strerror(errno));
return (DAT_INTERNAL_ERROR);
}
if (dapls_ns_subnet_match_v6(s, ia_ptr, addr) != 0) {
(void) close(s);
return (DAT_INVALID_ADDRESS);
}
(void) dapl_os_memzero(&lifr, sizeof (lifr));
(void) dapl_os_memcpy(&lifr.lifr_nd.lnr_addr, addr, sizeof (*addr));
(void) dapl_os_strcpy(lifr.lifr_name, ia_ptr->hca_ptr->name);
again:;
if (ioctl(s, SIOCLIFGETND, (caddr_t)&lifr) < 0) {
/*
* if SIOCLIFGETND failed, we force the ND
* cache to be filled by connecting to the
* destination IP address.
*/
if (retries < NS_MAX_RETRIES &&
dapls_ns_send_packet_v6(s, addr) == 0 &&
dapls_ns_resolve_addr(AF_INET6, (struct sockaddr *)addr,
timeout) == 0) {
retries++;
goto again;
}
dapl_dbg_log(DAPL_DBG_TYPE_ERR, "ns_lookup_v6: giving up\n");
(void) close(s);
return (DAT_ERROR(DAT_INVALID_ADDRESS,
DAT_INVALID_ADDRESS_UNREACHABLE));
}
if (lifr.lifr_nd.lnr_hdw_len == 0 && retries <= NS_MAX_RETRIES) {
/*
* lnr_hdw_len == 0 means that the ND entry
* is still incomplete. we need to retry the ioctl.
*/
retries++;
(void) sleep(1);
goto again;
}
(void) close(s);
mac = (uchar_t *)lifr.lifr_nd.lnr_hdw_addr;
if (lifr.lifr_nd.lnr_hdw_len >= sizeof (ipoib_mac_t)) {
ib_gid_t tmp_gid;
/* LINTED: E_BAD_PTR_CAST_ALIGN */
(void) dapl_os_memcpy(&tmp_gid,
&((ipoib_mac_t *)mac)->ipoib_gidpref, sizeof (ib_gid_t));
/*
* gids from the ND table are in network order, convert
* the gids from network order to host byte order
*/
gid->gid_prefix = BETOH_64(tmp_gid.gid_prefix);
gid->gid_guid = BETOH_64(tmp_gid.gid_guid);
} else {
int i, len;
len = lifr.lifr_nd.lnr_hdw_len;
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_lookup_v6: failed, non IB address: "
"len = %d, addr = 0x", len);
if (len > 0) {
for (i = 0; i < len; i++) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"%02x", (int)mac[i] & 0xff);
}
} else {
dapl_dbg_log(DAPL_DBG_TYPE_ERR, "0");
}
dapl_dbg_log(DAPL_DBG_TYPE_ERR, "\n");
return (DAT_INVALID_ADDRESS);
}
return (DAT_SUCCESS);
}
DAT_RETURN
dapls_ns_lookup_v4(
IN DAPL_IA *ia_ptr,
IN struct sockaddr_in *addr,
IN DAT_TIMEOUT timeout,
OUT ib_gid_t *gid)
{
struct xarpreq ar;
struct sockaddr_in *sin;
uchar_t *mac;
int s, retries = 0;
(void) dapl_os_memzero(&ar, sizeof (ar));
sin = (struct sockaddr_in *)&ar.xarp_pa;
sin->sin_family = AF_INET;
sin->sin_addr.s_addr = addr->sin_addr.s_addr;
ar.xarp_ha.sdl_family = AF_LINK;
s = socket(AF_INET, SOCK_DGRAM, 0);
if (s < 0) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_lookup_v4: socket: %s\n", strerror(errno));
return (DAT_INTERNAL_ERROR);
}
if (dapls_ns_subnet_match_v4(s, ia_ptr, addr) != 0) {
(void) close(s);
return (DAT_INVALID_ADDRESS);
}
again:;
if (ioctl(s, SIOCGXARP, (caddr_t)&ar) < 0) {
/*
* if SIOCGXARP failed, we force the ARP
* cache to be filled by connecting to the
* destination IP address.
*/
if (retries <= NS_MAX_RETRIES &&
dapls_ns_resolve_addr(AF_INET, (struct sockaddr *)addr,
timeout) == 0) {
retries++;
goto again;
}
dapl_dbg_log(DAPL_DBG_TYPE_ERR, "ns_lookup_v4: giving up\n");
(void) close(s);
return (DAT_ERROR(DAT_INVALID_ADDRESS,
DAT_INVALID_ADDRESS_UNREACHABLE));
}
if ((ar.xarp_flags & ATF_COM) == 0 &&
ar.xarp_ha.sdl_type == IFT_IB && retries <= NS_MAX_RETRIES) {
/*
* we get here if arp resolution is still incomplete
*/
retries++;
(void) sleep(1);
goto again;
}
(void) close(s);
mac = (uchar_t *)LLADDR(&ar.xarp_ha);
if (ar.xarp_flags & ATF_COM &&
ar.xarp_ha.sdl_type == IFT_IB &&
ar.xarp_ha.sdl_alen >= sizeof (ipoib_mac_t)) {
ib_gid_t tmp_gid;
/* LINTED: E_BAD_PTR_CAST_ALIGN */
(void) dapl_os_memcpy(&tmp_gid,
&((ipoib_mac_t *)mac)->ipoib_gidpref, sizeof (ib_gid_t));
/*
* gids from the ARP table are in network order, convert
* the gids from network order to host byte order
*/
gid->gid_prefix = BETOH_64(tmp_gid.gid_prefix);
gid->gid_guid = BETOH_64(tmp_gid.gid_guid);
} else {
int i, len;
len = ar.xarp_ha.sdl_alen;
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"ns_lookup_v4: failed, non IB address: "
"len = %d, addr = 0x", len);
if (len > 0) {
for (i = 0; i < len; i++) {
dapl_dbg_log(DAPL_DBG_TYPE_ERR,
"%02x", (int)mac[i] & 0xff);
}
} else {
dapl_dbg_log(DAPL_DBG_TYPE_ERR, "0");
}
dapl_dbg_log(DAPL_DBG_TYPE_ERR, "\n");
return (DAT_INVALID_ADDRESS);
}
return (DAT_SUCCESS);
}