ucs_status_t progress_remote_cq(uct_ugni_smsg_iface_t *iface)
{
gni_return_t ugni_rc;
gni_cq_entry_t event_data;
uct_ugni_ep_t tl_ep;
uct_ugni_ep_t *ugni_ep;
uct_ugni_smsg_ep_t *ep;
ugni_rc = GNI_CqGetEvent(iface->remote_cq, &event_data);
if(GNI_RC_NOT_DONE == ugni_rc){
return UCS_OK;
}
if (GNI_RC_SUCCESS != ugni_rc || !GNI_CQ_STATUS_OK(event_data) || GNI_CQ_OVERRUN(event_data)) {
if(GNI_RC_ERROR_RESOURCE == ugni_rc || (GNI_RC_SUCCESS == ugni_rc && GNI_CQ_OVERRUN(event_data))){
ucs_debug("Detected remote CQ overrun. ungi_rc = %d [%s]", ugni_rc, gni_err_str[ugni_rc]);
uct_ugni_smsg_handle_remote_overflow(iface);
return UCS_OK;
}
ucs_error("GNI_CqGetEvent falied with unhandled error. Error status %s %d ",
gni_err_str[ugni_rc], ugni_rc);
return UCS_ERR_IO_ERROR;
}
tl_ep.hash_key = GNI_CQ_GET_INST_ID(event_data);
ugni_ep = sglib_hashed_uct_ugni_ep_t_find_member(iface->super.eps, &tl_ep);
ep = ucs_derived_of(ugni_ep, uct_ugni_smsg_ep_t);
process_mbox(iface, ep);
return UCS_INPROGRESS;
}
static ucs_status_t progress_local_cq(uct_ugni_smsg_iface_t *iface){
gni_return_t ugni_rc;
gni_cq_entry_t event_data;
uct_ugni_smsg_desc_t message_data;
uct_ugni_smsg_desc_t *message_pointer;
ugni_rc = GNI_CqGetEvent(iface->super.local_cq, &event_data);
if(GNI_RC_NOT_DONE == ugni_rc){
return UCS_OK;
}
if((GNI_RC_SUCCESS != ugni_rc && !event_data) || GNI_CQ_OVERRUN(event_data)){
/* TODO: handle overruns */
ucs_error("Error posting data. CQ overrun = %d", (int)GNI_CQ_OVERRUN(event_data));
return UCS_ERR_NO_RESOURCE;
}
message_data.msg_id = GNI_CQ_GET_MSG_ID(event_data);
message_pointer = sglib_hashed_uct_ugni_smsg_desc_t_find_member(iface->smsg_list,&message_data);
ucs_assert(NULL != message_pointer);
message_pointer->ep->outstanding--;
iface->super.outstanding--;
uct_ugni_ep_check_flush(message_pointer->ep);
sglib_hashed_uct_ugni_smsg_desc_t_delete(iface->smsg_list,message_pointer);
ucs_mpool_put(message_pointer);
return UCS_INPROGRESS;
}
static void progress_local_cq(uct_ugni_smsg_iface_t *iface){
gni_return_t ugni_rc;
gni_cq_entry_t event_data;
uct_ugni_smsg_desc_t message_data;
uct_ugni_smsg_desc_t *message_pointer;
if(0 == iface->super.outstanding){
return;
}
ugni_rc = GNI_CqGetEvent(iface->super.local_cq, &event_data);
if(GNI_RC_NOT_DONE == ugni_rc){
return;
}
if((GNI_RC_SUCCESS != ugni_rc && !event_data) || GNI_CQ_OVERRUN(event_data)){
/* TODO: handle overruns */
ucs_error("Error posting data. CQ overrun = %d", (int)GNI_CQ_OVERRUN(event_data));
return;
}
message_data.msg_id = GNI_CQ_GET_MSG_ID(event_data);
message_pointer = sglib_hashed_uct_ugni_smsg_desc_t_find_member(iface->smsg_list,&message_data);
message_pointer->ep->outstanding--;
iface->super.outstanding--;
sglib_hashed_uct_ugni_smsg_desc_t_delete(iface->smsg_list,message_pointer);
ucs_mpool_put(message_pointer);
}
int mca_btl_ugni_progress_remote_smsg (mca_btl_ugni_module_t *btl)
{
mca_btl_base_endpoint_t *ep;
gni_cq_entry_t event_data;
gni_return_t grc;
uint64_t inst_id;
grc = mca_btl_ugni_gni_cq_get_event (btl->devices, btl->smsg_remote_cq, &event_data);
if (GNI_RC_NOT_DONE == grc) {
return 0;
}
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc || !GNI_CQ_STATUS_OK(event_data) ||
GNI_CQ_OVERRUN(event_data))) {
if (GNI_RC_ERROR_RESOURCE == grc ||
(GNI_RC_SUCCESS == grc && GNI_CQ_OVERRUN(event_data))) {
/* recover from smsg cq overrun */
return mca_btl_ugni_handle_remote_smsg_overrun (btl);
}
BTL_ERROR(("unhandled error in GNI_CqGetEvent"));
/* unhandled error: crash */
assert (0);
return mca_btl_rc_ugni_to_opal (grc);
}
BTL_VERBOSE(("REMOTE CQ: Got event 0x%" PRIx64 ". msg id = %" PRIu64
". ok = %d, type = %" PRIu64, (uint64_t) event_data,
GNI_CQ_GET_INST_ID(event_data), GNI_CQ_STATUS_OK(event_data),
GNI_CQ_GET_TYPE(event_data)));
inst_id = GNI_CQ_GET_INST_ID(event_data);
ep = (mca_btl_base_endpoint_t *) opal_pointer_array_get_item (&btl->endpoints, inst_id);
if (OPAL_UNLIKELY(MCA_BTL_UGNI_EP_STATE_CONNECTED != ep->state)) {
/* due to the nature of datagrams we may get a smsg completion before
we get mailbox info from the peer */
BTL_VERBOSE(("event occurred on an unconnected endpoint! ep state = %d", ep->state));
return 0;
}
return mca_btl_ugni_smsg_process (ep);
}
static inline int
mca_btl_ugni_progress_rdma (mca_btl_ugni_module_t *ugni_module)
{
ompi_common_ugni_post_desc_t *desc;
mca_btl_ugni_base_frag_t *frag;
gni_cq_entry_t event_data = 0;
uint32_t recoverable = 1;
gni_return_t rc;
rc = GNI_CqGetEvent (ugni_module->rdma_local_cq, &event_data);
if (GNI_RC_NOT_DONE == rc) {
return 0;
}
if (OPAL_UNLIKELY((GNI_RC_SUCCESS != rc && !event_data) || GNI_CQ_OVERRUN(event_data))) {
/* TODO -- need to handle overrun -- how do we do this without an event?
will the event eventually come back? Ask Cray */
BTL_ERROR(("unhandled post error! ugni rc = %d", rc));
assert (0);
return ompi_common_rc_ugni_to_ompi (rc);
}
rc = GNI_GetCompleted (ugni_module->rdma_local_cq, event_data, (gni_post_descriptor_t **) &desc);
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != rc && GNI_RC_TRANSACTION_ERROR != rc)) {
BTL_ERROR(("Error in GNI_GetComplete %s", gni_err_str[rc]));
return ompi_common_rc_ugni_to_ompi (rc);
}
frag = MCA_BTL_UGNI_DESC_TO_FRAG(desc);
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != rc || !GNI_CQ_STATUS_OK(event_data))) {
(void) GNI_CqErrorRecoverable (event_data, &recoverable);
if (OPAL_UNLIKELY(++frag->post_desc.tries >= mca_btl_ugni_component.rdma_max_retries ||
!recoverable)) {
/* give up */
BTL_ERROR(("giving up on frag %p", (void *) frag));
frag->cbfunc (frag, OMPI_ERROR);
return OMPI_ERROR;
}
/* repost transaction */
mca_btl_ugni_repost (frag, OMPI_SUCCESS);
return 0;
}
BTL_VERBOSE(("RDMA/FMA complete for frag %p", (void *) frag));
frag->cbfunc (frag, ompi_common_rc_ugni_to_ompi (rc));
return 1;
}
static int __nic_rx_progress(struct gnix_nic *nic)
{
int ret = FI_SUCCESS;
gni_return_t status = GNI_RC_NOT_DONE;
gni_cq_entry_t cqe;
status = GNI_CqTestEvent(nic->rx_cq);
if (status == GNI_RC_NOT_DONE)
return FI_SUCCESS;
COND_ACQUIRE(nic->requires_lock, &nic->lock);
do {
status = GNI_CqGetEvent(nic->rx_cq, &cqe);
if (unlikely(status == GNI_RC_NOT_DONE)) {
ret = FI_SUCCESS;
break;
}
if (likely(status == GNI_RC_SUCCESS)) {
/* Find and schedule the associated VC. */
ret = __process_rx_cqe(nic, cqe);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"process_rx_cqe() failed: %d\n",
ret);
}
} else if (status == GNI_RC_ERROR_RESOURCE) {
/* The remote CQ was overrun. Events related to any VC
* could have been missed. Schedule each VC to be sure
* all messages are processed. */
assert(GNI_CQ_OVERRUN(cqe));
__nic_rx_overrun(nic);
} else {
GNIX_WARN(FI_LOG_EP_DATA,
"GNI_CqGetEvent returned %s\n",
gni_err_str[status]);
ret = gnixu_to_fi_errno(status);
break;
}
} while (1);
COND_RELEASE(nic->requires_lock, &nic->lock);
return ret;
}
void uct_ugni_progress(void *arg)
{
gni_cq_entry_t event_data = 0;
gni_post_descriptor_t *event_post_desc_ptr;
uct_ugni_base_desc_t *desc;
uct_ugni_iface_t * iface = (uct_ugni_iface_t *)arg;
gni_return_t ugni_rc;
ugni_rc = GNI_CqGetEvent(iface->local_cq, &event_data);
if (GNI_RC_NOT_DONE == ugni_rc) {
goto out;
}
if ((GNI_RC_SUCCESS != ugni_rc && !event_data) || GNI_CQ_OVERRUN(event_data)) {
ucs_error("GNI_CqGetEvent falied. Error status %s %d ",
gni_err_str[ugni_rc], ugni_rc);
return;
}
ugni_rc = GNI_GetCompleted(iface->local_cq, event_data, &event_post_desc_ptr);
if (GNI_RC_SUCCESS != ugni_rc && GNI_RC_TRANSACTION_ERROR != ugni_rc) {
ucs_error("GNI_GetCompleted falied. Error status %s %d %d",
gni_err_str[ugni_rc], ugni_rc, GNI_RC_TRANSACTION_ERROR);
return;
}
desc = (uct_ugni_base_desc_t *)event_post_desc_ptr;
ucs_trace_async("Completion received on %p", desc);
if (NULL != desc->comp_cb) {
uct_invoke_completion(desc->comp_cb, UCS_OK);
}
--iface->outstanding;
--desc->ep->outstanding;
if (ucs_likely(0 == desc->not_ready_to_free)) {
ucs_mpool_put(desc);
}
uct_ugni_ep_check_flush(desc->ep);
out:
/* have a go a processing the pending queue */
ucs_arbiter_dispatch(&iface->arbiter, 1, uct_ugni_ep_process_pending, NULL);
return;
}
void uct_ugni_progress(void *arg)
{
gni_cq_entry_t event_data = 0;
gni_post_descriptor_t *event_post_desc_ptr;
uct_ugni_base_desc_t *desc;
uct_ugni_iface_t * iface = (uct_ugni_iface_t *)arg;
gni_return_t ugni_rc;
ugni_rc = GNI_CqGetEvent(iface->local_cq, &event_data);
if (GNI_RC_NOT_DONE == ugni_rc) {
return;
}
if ((GNI_RC_SUCCESS != ugni_rc && !event_data) || GNI_CQ_OVERRUN(event_data)) {
ucs_error("GNI_CqGetEvent falied. Error status %s %d ",
gni_err_str[ugni_rc], ugni_rc);
return;
}
ugni_rc = GNI_GetCompleted(iface->local_cq, event_data, &event_post_desc_ptr);
if (GNI_RC_SUCCESS != ugni_rc && GNI_RC_TRANSACTION_ERROR != ugni_rc) {
ucs_error("GNI_GetCompleted falied. Error status %s %d %d",
gni_err_str[ugni_rc], ugni_rc, GNI_RC_TRANSACTION_ERROR);
return;
}
desc = (uct_ugni_base_desc_t *)event_post_desc_ptr;
ucs_trace_async("Completion received on %p", desc);
if (NULL != desc->comp_cb) {
uct_invoke_completion(desc->comp_cb);
}
--iface->outstanding;
--desc->ep->outstanding;
if (ucs_likely(desc->not_ready_to_free == 0)) {
ucs_mpool_put(desc);
}
return;
}
static inline int mca_btl_ugni_progress_rdma (mca_btl_ugni_module_t *ugni_module, int which_cq)
{
mca_btl_ugni_post_descriptor_t *post_desc = NULL;
gni_cq_entry_t event_data = 0;
gni_post_descriptor_t *desc;
uint32_t recoverable = 1;
gni_return_t grc;
gni_cq_handle_t the_cq;
the_cq = (which_cq == 0) ? ugni_module->rdma_local_cq : ugni_module->rdma_local_irq_cq;
OPAL_THREAD_LOCK(&ugni_module->device->dev_lock);
grc = GNI_CqGetEvent (the_cq, &event_data);
if (GNI_RC_NOT_DONE == grc) {
OPAL_THREAD_UNLOCK(&ugni_module->device->dev_lock);
return 0;
}
if (OPAL_UNLIKELY((GNI_RC_SUCCESS != grc && !event_data) || GNI_CQ_OVERRUN(event_data))) {
/* TODO -- need to handle overrun -- how do we do this without an event?
will the event eventually come back? Ask Cray */
BTL_ERROR(("unhandled post error! ugni rc = %d %s", grc, gni_err_str[grc]));
OPAL_THREAD_UNLOCK(&ugni_module->device->dev_lock);
return opal_common_rc_ugni_to_opal (grc);
}
grc = GNI_GetCompleted (the_cq, event_data, &desc);
OPAL_THREAD_UNLOCK(&ugni_module->device->dev_lock);
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc && GNI_RC_TRANSACTION_ERROR != grc)) {
BTL_ERROR(("Error in GNI_GetComplete %s", gni_err_str[grc]));
return opal_common_rc_ugni_to_opal (grc);
}
post_desc = MCA_BTL_UGNI_DESC_TO_PDESC(desc);
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc || !GNI_CQ_STATUS_OK(event_data))) {
(void) GNI_CqErrorRecoverable (event_data, &recoverable);
if (OPAL_UNLIKELY(++post_desc->desc.tries >= mca_btl_ugni_component.rdma_max_retries ||
!recoverable)) {
char char_buffer[1024];
GNI_CqErrorStr (event_data, char_buffer, 1024);
/* give up */
BTL_ERROR(("giving up on desciptor %p, recoverable %d: %s", (void *) post_desc,
recoverable, char_buffer));
#if OPAL_ENABLE_DEBUG
btl_ugni_dump_post_desc (post_desc);
#endif
mca_btl_ugni_post_desc_complete (ugni_module, post_desc, OPAL_ERROR);
return OPAL_ERROR;
}
mca_btl_ugni_repost (ugni_module, post_desc);
return 0;
}
mca_btl_ugni_post_desc_complete (ugni_module, post_desc, opal_common_rc_ugni_to_opal (grc));
return 1;
}
