/**
* Destroy a slab.
*
* @param[in] handle Handle to the allocator being used.
* @param[in] slab Slab to be destroyed.
*
* @return FI_SUCCESS On successful slab destruction.
*
* @return -FI_EINVAL On invalid handle or slab being given as parameters.
*/
static int __destroy_slab(struct gnix_mbox_alloc_handle *handle,
struct gnix_slab *slab)
{
size_t total_size;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
if (!handle || !slab) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Invalid argument handle or slab.\n");
return -FI_EINVAL;
}
total_size = handle->page_size * __page_count(handle);
_gnix_free_bitmap(slab->used);
free(slab->used);
COND_ACQUIRE(handle->nic_handle->requires_lock, &handle->nic_handle->lock);
GNI_MemDeregister(handle->nic_handle->gni_nic_hndl,
&slab->memory_handle);
COND_RELEASE(handle->nic_handle->requires_lock, &handle->nic_handle->lock);
munmap(slab->base, total_size);
free(slab);
return FI_SUCCESS;
}
static int __gnix_deregister_region(
void *handle,
void *context)
{
struct gnix_fid_mem_desc *mr = (struct gnix_fid_mem_desc *) handle;
gni_return_t ret;
struct gnix_fid_domain *domain;
struct gnix_nic *nic;
domain = mr->domain;
nic = mr->nic;
COND_ACQUIRE(nic->requires_lock, &nic->lock);
ret = GNI_MemDeregister(nic->gni_nic_hndl, &mr->mem_hndl);
COND_RELEASE(nic->requires_lock, &nic->lock);
if (ret == GNI_RC_SUCCESS) {
/* release reference to domain */
_gnix_ref_put(domain);
/* release reference to nic */
_gnix_ref_put(nic);
} else {
GNIX_INFO(FI_LOG_MR, "failed to deregister memory"
" region, entry=%p ret=%i\n", handle, ret);
}
return ret;
}
DIRECT_FN STATIC int gnix_shutdown(struct fid_ep *ep, uint64_t flags)
{
int ret;
struct gnix_fid_ep *ep_priv;
struct fi_eq_cm_entry eq_entry = {0};
if (!ep)
return -FI_EINVAL;
ep_priv = container_of(ep, struct gnix_fid_ep, ep_fid.fid);
COND_ACQUIRE(ep_priv->requires_lock, &ep_priv->vc_lock);
eq_entry.fid = &ep_priv->ep_fid.fid;
ret = fi_eq_write(&ep_priv->eq->eq_fid, FI_SHUTDOWN, &eq_entry,
sizeof(eq_entry), 0);
if (ret != sizeof(eq_entry)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"fi_eq_write failed, err: %d\n", ret);
} else {
ret = FI_SUCCESS;
}
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
return ret;
}
static int __gnix_amo_send_cntr_req(void *arg)
{
struct gnix_fab_req *req = (struct gnix_fab_req *)arg;
struct gnix_fid_ep *ep = req->gnix_ep;
struct gnix_nic *nic = ep->nic;
struct gnix_tx_descriptor *txd;
gni_return_t status;
int rc;
int inject_err = _gnix_req_inject_err(req);
rc = _gnix_nic_tx_alloc(nic, &txd);
if (rc) {
GNIX_INFO(FI_LOG_EP_DATA,
"_gnix_nic_tx_alloc() failed: %d\n",
rc);
return -FI_ENOSPC;
}
txd->req = req;
txd->completer_fn = __gnix_amo_txd_cntr_complete;
if (req->type == GNIX_FAB_RQ_AMO) {
txd->amo_cntr_hdr.flags = FI_REMOTE_WRITE;
} else {
txd->amo_cntr_hdr.flags = FI_REMOTE_READ;
}
COND_ACQUIRE(nic->requires_lock, &nic->lock);
if (inject_err) {
_gnix_nic_txd_err_inject(nic, txd);
status = GNI_RC_SUCCESS;
} else {
status = GNI_SmsgSendWTag(req->vc->gni_ep,
&txd->amo_cntr_hdr,
sizeof(txd->amo_cntr_hdr),
NULL, 0, txd->id,
GNIX_SMSG_T_AMO_CNTR);
}
COND_RELEASE(nic->requires_lock, &nic->lock);
if (status == GNI_RC_NOT_DONE) {
_gnix_nic_tx_free(nic, txd);
GNIX_INFO(FI_LOG_EP_DATA,
"GNI_SmsgSendWTag returned %s\n",
gni_err_str[status]);
} else if (status != GNI_RC_SUCCESS) {
_gnix_nic_tx_free(nic, txd);
GNIX_WARN(FI_LOG_EP_DATA,
"GNI_SmsgSendWTag returned %s\n",
gni_err_str[status]);
} else {
GNIX_INFO(FI_LOG_EP_DATA, "Sent RMA CQ data, req: %p\n", req);
}
return gnixu_to_fi_errno(status);
}
/* Check for a connection response on an FI_EP_MSG. */
int _gnix_ep_progress(struct gnix_fid_ep *ep)
{
int ret, bytes_read;
struct gnix_pep_sock_connresp resp;
/* No lock, fast exit. */
if (ep->conn_state != GNIX_EP_CONNECTING) {
return FI_SUCCESS;
}
COND_ACQUIRE(ep->requires_lock, &ep->vc_lock);
if (ep->conn_state != GNIX_EP_CONNECTING) {
COND_RELEASE(ep->requires_lock, &ep->vc_lock);
return FI_SUCCESS;
}
/* Check for a connection response. */
bytes_read = read(ep->conn_fd, &resp, sizeof(resp));
if (bytes_read >= 0) {
if (bytes_read == sizeof(resp)) {
/* Received response. */
ret = __gnix_ep_connresp(ep, &resp);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"__gnix_pep_connreq failed, %d\n",
ret);
}
} else {
GNIX_FATAL(FI_LOG_EP_CTRL,
"Unexpected read size: %d\n",
bytes_read);
}
} else if (errno != EAGAIN) {
GNIX_WARN(FI_LOG_EP_CTRL, "Read error: %s\n", strerror(errno));
}
COND_RELEASE(ep->requires_lock, &ep->vc_lock);
return FI_SUCCESS;
}
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;
}
static int __nic_rx_overrun(struct gnix_nic *nic)
{
int i, max_id, ret;
struct gnix_vc *vc;
gni_return_t status;
gni_cq_entry_t cqe;
GNIX_WARN(FI_LOG_EP_DATA, "\n");
/* clear out the CQ */
/*
* TODO: really need to process CQEs better for error reporting,
* etc.
*/
while ((status = GNI_CqGetEvent(nic->rx_cq, &cqe)) == GNI_RC_SUCCESS);
assert(status == GNI_RC_NOT_DONE);
COND_ACQUIRE(nic->requires_lock, &nic->vc_id_lock);
max_id = nic->vc_id_table_count;
COND_RELEASE(nic->requires_lock, &nic->vc_id_lock);
/*
* TODO: optimization would
* be to keep track of last time
* this happened and where smsg msgs.
* were found.
*/
for (i = 0; i < max_id; i++) {
ret = _gnix_test_bit(&nic->vc_id_bitmap, i);
if (ret) {
vc = __gnix_nic_elem_by_rem_id(nic, i);
ret = _gnix_vc_dequeue_smsg(vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_vc_dqueue_smsg returned %d\n",
ret);
}
}
}
return FI_SUCCESS;
}
/**
* Create a slab from a handle and append to the slab list.
*
* @param[in] handle Handle to the allocator being used.
*
* @return FI_SUCCESS On successful slab creation.
*
* @return -FI_ENOMEM if failure to allocate memory for slab or bitmap.
* @return [Unspec] if failure in alloc_bitmap. Will return error code from
* alloc_bitmap.
* @return [Unspec] if failure in GNI_MemRegister. Converts gni_return_t
* status code to FI_ERRNO value.
*/
static int __create_slab(struct gnix_mbox_alloc_handle *handle)
{
struct gnix_slab *slab;
gni_return_t status;
char error_buf[256];
char *error;
size_t total_size;
int ret;
int vmdh_index = -1;
int flags = GNI_MEM_READWRITE;
struct gnix_auth_key *info;
GNIX_TRACE(FI_LOG_EP_CTRL, "\n");
slab = calloc(1, sizeof(*slab));
if (!slab) {
error = strerror_r(errno, error_buf, sizeof(error_buf));
GNIX_WARN(FI_LOG_EP_CTRL,
"Error allocating slab: %s\n",
error);
ret = -FI_ENOMEM;
goto err_slab_calloc;
}
total_size = handle->page_size * __page_count(handle);
GNIX_DEBUG(FI_LOG_EP_CTRL, "total_size requested for mmap: %zu.\n",
total_size);
slab->used = calloc(1, sizeof(*(slab->used)));
if (!slab->used) {
error = strerror_r(errno, error_buf, sizeof(error_buf));
GNIX_WARN(FI_LOG_EP_CTRL,
"Error allocating bitmap: %s\n",
error);
ret = -FI_ENOMEM;
goto err_bitmap_calloc;
}
slab->base = mmap(0, total_size, (PROT_READ | PROT_WRITE), MAP_SHARED,
handle->fd, handle->last_offset);
if (slab->base == MAP_FAILED) {
error = strerror_r(errno, error_buf, sizeof(error_buf));
GNIX_WARN(FI_LOG_EP_CTRL, "%s\n", error);
ret = -FI_ENOMEM;
goto err_mmap;
}
ret = _gnix_alloc_bitmap(slab->used, __mbox_count(handle), NULL);
if (ret) {
GNIX_WARN(FI_LOG_EP_CTRL, "Error allocating bitmap.\n");
goto err_alloc_bitmap;
}
COND_ACQUIRE(handle->nic_handle->requires_lock, &handle->nic_handle->lock);
if (handle->nic_handle->using_vmdh) {
info = _gnix_auth_key_lookup(GNIX_PROV_DEFAULT_AUTH_KEY,
GNIX_PROV_DEFAULT_AUTH_KEYLEN);
assert(info);
if (!handle->nic_handle->mdd_resources_set) {
/* check to see if the ptag registration limit was set
* yet or not -- becomes read-only after success */
_gnix_auth_key_enable(info);
status = GNI_SetMddResources(
handle->nic_handle->gni_nic_hndl,
(info->attr.prov_key_limit +
info->attr.user_key_limit));
assert(status == GNI_RC_SUCCESS);
handle->nic_handle->mdd_resources_set = 1;
}
vmdh_index = _gnix_get_next_reserved_key(info);
if (vmdh_index <= 0) {
GNIX_FATAL(FI_LOG_DOMAIN,
"failed to get reserved key for mbox "
"registration, rc=%d\n",
vmdh_index);
}
flags |= GNI_MEM_USE_VMDH;
}
status = GNI_MemRegister(handle->nic_handle->gni_nic_hndl,
(uint64_t) slab->base, total_size,
handle->cq_handle,
flags, vmdh_index,
&slab->memory_handle);
COND_RELEASE(handle->nic_handle->requires_lock, &handle->nic_handle->lock);
if (status != GNI_RC_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL, "GNI_MemRegister failed: %s\n",
gni_err_str[status]);
ret = gnixu_to_fi_errno(status);
goto err_memregister;
}
slab->allocator = handle;
gnix_slist_insert_tail(&slab->list_entry, &handle->slab_list);
handle->last_offset += total_size;
return ret;
err_memregister:
_gnix_free_bitmap(slab->used);
err_alloc_bitmap:
munmap(slab->base, total_size);
err_mmap:
free(slab->used);
err_bitmap_calloc:
free(slab);
err_slab_calloc:
return ret;
}
static int gnix_cq_set_wait(struct gnix_fid_cq *cq)
{
int ret = FI_SUCCESS;
GNIX_TRACE(FI_LOG_CQ, "\n");
struct fi_wait_attr requested = {
.wait_obj = cq->attr.wait_obj,
.flags = 0
};
switch (cq->attr.wait_obj) {
case FI_WAIT_UNSPEC:
case FI_WAIT_FD:
case FI_WAIT_MUTEX_COND:
ret = gnix_wait_open(&cq->domain->fabric->fab_fid,
&requested, &cq->wait);
break;
case FI_WAIT_SET:
ret = _gnix_wait_set_add(cq->attr.wait_set, &cq->cq_fid.fid);
if (!ret)
cq->wait = cq->attr.wait_set;
break;
default:
break;
}
return ret;
}
static void free_cq_entry(struct slist_entry *item)
{
struct gnix_cq_entry *entry;
entry = container_of(item, struct gnix_cq_entry, item);
free(entry->the_entry);
free(entry);
}
static struct slist_entry *alloc_cq_entry(size_t size)
{
struct gnix_cq_entry *entry = malloc(sizeof(*entry));
if (!entry) {
GNIX_DEBUG(FI_LOG_CQ, "out of memory\n");
goto err;
}
entry->the_entry = malloc(size);
if (!entry->the_entry) {
GNIX_DEBUG(FI_LOG_CQ, "out of memory\n");
goto cleanup;
}
return &entry->item;
cleanup:
free(entry);
err:
return NULL;
}
static int __gnix_cq_progress(struct gnix_fid_cq *cq)
{
return _gnix_prog_progress(&cq->pset);
}
/*******************************************************************************
* Exposed helper functions
******************************************************************************/
ssize_t _gnix_cq_add_event(struct gnix_fid_cq *cq, struct gnix_fid_ep *ep,
void *op_context, uint64_t flags, size_t len,
void *buf, uint64_t data, uint64_t tag,
fi_addr_t src_addr)
{
struct gnix_cq_entry *event;
struct slist_entry *item;
uint64_t mask;
ssize_t ret = FI_SUCCESS;
if (ep) {
if (ep->info && ep->info->mode & FI_NOTIFY_FLAGS_ONLY) {
mask = (FI_REMOTE_CQ_DATA | FI_MULTI_RECV);
if (flags & FI_RMA_EVENT) {
mask |= (FI_REMOTE_READ | FI_REMOTE_WRITE |
FI_RMA);
}
flags &= mask;
}
}
COND_ACQUIRE(cq->requires_lock, &cq->lock);
item = _gnix_queue_get_free(cq->events);
if (!item) {
GNIX_DEBUG(FI_LOG_CQ, "error creating cq_entry\n");
ret = -FI_ENOMEM;
goto err;
}
event = container_of(item, struct gnix_cq_entry, item);
assert(event->the_entry);
fill_function[cq->attr.format](event->the_entry, op_context, flags,
len, buf, data, tag);
event->src_addr = src_addr;
_gnix_queue_enqueue(cq->events, &event->item);
GNIX_DEBUG(FI_LOG_CQ, "Added event: %lx\n", op_context);
if (cq->wait)
_gnix_signal_wait_obj(cq->wait);
err:
COND_RELEASE(cq->requires_lock, &cq->lock);
return ret;
}
ssize_t _gnix_cq_add_error(struct gnix_fid_cq *cq, void *op_context,
uint64_t flags, size_t len, void *buf,
uint64_t data, uint64_t tag, size_t olen,
int err, int prov_errno, void *err_data,
size_t err_data_size)
{
struct fi_cq_err_entry *error;
struct gnix_cq_entry *event;
struct slist_entry *item;
ssize_t ret = FI_SUCCESS;
GNIX_INFO(FI_LOG_CQ, "creating error event entry\n");
COND_ACQUIRE(cq->requires_lock, &cq->lock);
item = _gnix_queue_get_free(cq->errors);
if (!item) {
GNIX_WARN(FI_LOG_CQ, "error creating error entry\n");
ret = -FI_ENOMEM;
goto err;
}
event = container_of(item, struct gnix_cq_entry, item);
error = event->the_entry;
error->op_context = op_context;
error->flags = flags;
error->len = len;
error->buf = buf;
error->data = data;
error->tag = tag;
error->olen = olen;
error->err = err;
error->prov_errno = prov_errno;
error->err_data = err_data;
error->err_data_size = err_data_size;
_gnix_queue_enqueue(cq->errors, &event->item);
if (cq->wait)
_gnix_signal_wait_obj(cq->wait);
err:
COND_RELEASE(cq->requires_lock, &cq->lock);
return ret;
}
int _gnix_cq_poll_obj_add(struct gnix_fid_cq *cq, void *obj,
int (*prog_fn)(void *data))
{
return _gnix_prog_obj_add(&cq->pset, obj, prog_fn);
}
int _gnix_cq_poll_obj_rem(struct gnix_fid_cq *cq, void *obj,
int (*prog_fn)(void *data))
{
return _gnix_prog_obj_rem(&cq->pset, obj, prog_fn);
}
static void __cq_destruct(void *obj)
{
struct gnix_fid_cq *cq = (struct gnix_fid_cq *) obj;
_gnix_ref_put(cq->domain);
switch (cq->attr.wait_obj) {
case FI_WAIT_NONE:
break;
case FI_WAIT_SET:
_gnix_wait_set_remove(cq->wait, &cq->cq_fid.fid);
break;
case FI_WAIT_UNSPEC:
case FI_WAIT_FD:
case FI_WAIT_MUTEX_COND:
assert(cq->wait);
gnix_wait_close(&cq->wait->fid);
break;
default:
GNIX_WARN(FI_LOG_CQ, "format: %d unsupported.\n",
cq->attr.wait_obj);
break;
}
_gnix_prog_fini(&cq->pset);
_gnix_queue_destroy(cq->events);
_gnix_queue_destroy(cq->errors);
fastlock_destroy(&cq->lock);
free(cq->cq_fid.ops);
free(cq->cq_fid.fid.ops);
free(cq);
}
/*******************************************************************************
* API functions.
******************************************************************************/
static int gnix_cq_close(fid_t fid)
{
struct gnix_fid_cq *cq;
int references_held;
GNIX_TRACE(FI_LOG_CQ, "\n");
cq = container_of(fid, struct gnix_fid_cq, cq_fid);
references_held = _gnix_ref_put(cq);
if (references_held) {
GNIX_INFO(FI_LOG_CQ, "failed to fully close cq due to lingering "
"references. references=%i cq=%p\n",
references_held, cq);
}
return FI_SUCCESS;
}
static ssize_t __gnix_cq_readfrom(struct fid_cq *cq, void *buf,
size_t count, fi_addr_t *src_addr)
{
struct gnix_fid_cq *cq_priv;
struct gnix_cq_entry *event;
struct slist_entry *temp;
ssize_t read_count = 0;
if (!cq || !buf || !count)
return -FI_EINVAL;
cq_priv = container_of(cq, struct gnix_fid_cq, cq_fid);
__gnix_cq_progress(cq_priv);
if (_gnix_queue_peek(cq_priv->errors))
return -FI_EAVAIL;
COND_ACQUIRE(cq_priv->requires_lock, &cq_priv->lock);
while (_gnix_queue_peek(cq_priv->events) && count--) {
temp = _gnix_queue_dequeue(cq_priv->events);
event = container_of(temp, struct gnix_cq_entry, item);
assert(event->the_entry);
memcpy(buf, event->the_entry, cq_priv->entry_size);
if (src_addr)
memcpy(&src_addr[read_count], &event->src_addr, sizeof(fi_addr_t));
_gnix_queue_enqueue_free(cq_priv->events, &event->item);
buf = (void *) ((uint8_t *) buf + cq_priv->entry_size);
read_count++;
}
COND_RELEASE(cq_priv->requires_lock, &cq_priv->lock);
return read_count ?: -FI_EAGAIN;
}
static ssize_t __gnix_cq_sreadfrom(int blocking, struct fid_cq *cq, void *buf,
size_t count, fi_addr_t *src_addr,
const void *cond, int timeout)
{
struct gnix_fid_cq *cq_priv;
cq_priv = container_of(cq, struct gnix_fid_cq, cq_fid);
if ((blocking && !cq_priv->wait) ||
(blocking && cq_priv->attr.wait_obj == FI_WAIT_SET))
return -FI_EINVAL;
if (_gnix_queue_peek(cq_priv->errors))
return -FI_EAVAIL;
if (cq_priv->wait)
gnix_wait_wait((struct fid_wait *)cq_priv->wait, timeout);
return __gnix_cq_readfrom(cq, buf, count, src_addr);
}
DIRECT_FN STATIC ssize_t gnix_cq_sreadfrom(struct fid_cq *cq, void *buf,
size_t count, fi_addr_t *src_addr,
const void *cond, int timeout)
{
return __gnix_cq_sreadfrom(1, cq, buf, count, src_addr, cond, timeout);
}
DIRECT_FN STATIC ssize_t gnix_cq_read(struct fid_cq *cq,
void *buf,
size_t count)
{
return __gnix_cq_sreadfrom(0, cq, buf, count, NULL, NULL, 0);
}
DIRECT_FN STATIC ssize_t gnix_cq_sread(struct fid_cq *cq, void *buf,
size_t count, const void *cond,
int timeout)
{
return __gnix_cq_sreadfrom(1, cq, buf, count, NULL, cond, timeout);
}
DIRECT_FN STATIC ssize_t gnix_cq_readfrom(struct fid_cq *cq, void *buf,
size_t count, fi_addr_t *src_addr)
{
return __gnix_cq_sreadfrom(0, cq, buf, count, src_addr, NULL, 0);
}
DIRECT_FN STATIC ssize_t gnix_cq_readerr(struct fid_cq *cq,
struct fi_cq_err_entry *buf,
uint64_t flags)
{
struct gnix_fid_cq *cq_priv;
struct gnix_cq_entry *event;
struct slist_entry *entry;
size_t err_data_cpylen;
struct fi_cq_err_entry *gnix_cq_err;
ssize_t read_count = 0;
if (!cq || !buf)
return -FI_EINVAL;
cq_priv = container_of(cq, struct gnix_fid_cq, cq_fid);
/*
* we need to progress cq. some apps may be only using
* cq to check for errors.
*/
_gnix_prog_progress(&cq_priv->pset);
COND_ACQUIRE(cq_priv->requires_lock, &cq_priv->lock);
entry = _gnix_queue_dequeue(cq_priv->errors);
if (!entry) {
read_count = -FI_EAGAIN;
goto err;
}
event = container_of(entry, struct gnix_cq_entry, item);
gnix_cq_err = event->the_entry;
buf->op_context = gnix_cq_err->op_context;
buf->flags = gnix_cq_err->flags;
buf->len = gnix_cq_err->len;
buf->buf = gnix_cq_err->buf;
buf->data = gnix_cq_err->data;
buf->tag = gnix_cq_err->tag;
buf->olen = gnix_cq_err->olen;
buf->err = gnix_cq_err->err;
buf->prov_errno = gnix_cq_err->prov_errno;
if (gnix_cq_err->err_data != NULL) {
/*
* Note: If the api version is >= 1.5 then copy err_data into
* buf->err_data and copy at most buf->err_data_size.
* If buf->err_data_size is zero or the api version is < 1.5,
* use the old method of allocating space in provider.
*/
if (FI_VERSION_LT(cq_priv->domain->fabric->fab_fid.api_version,
FI_VERSION(1, 5)) || buf->err_data_size == 0) {
err_data_cpylen = sizeof(cq_priv->err_data);
memcpy(cq_priv->err_data, gnix_cq_err->err_data,
err_data_cpylen);
buf->err_data = cq_priv->err_data;
} else {
if (buf->err_data == NULL)
return -FI_EINVAL;
err_data_cpylen = MIN(buf->err_data_size,
gnix_cq_err->err_data_size);
memcpy(buf->err_data, gnix_cq_err->err_data, err_data_cpylen);
buf->err_data_size = err_data_cpylen;
}
free(gnix_cq_err->err_data);
gnix_cq_err->err_data = NULL;
} else {
if (FI_VERSION_LT(cq_priv->domain->fabric->fab_fid.api_version,
FI_VERSION(1, 5))) {
buf->err_data = NULL;
} else {
buf->err_data_size = 0;
}
}
_gnix_queue_enqueue_free(cq_priv->errors, &event->item);
read_count++;
err:
COND_RELEASE(cq_priv->requires_lock, &cq_priv->lock);
return read_count;
}
DIRECT_FN STATIC const char *gnix_cq_strerror(struct fid_cq *cq, int prov_errno,
const void *prov_data, char *buf,
size_t len)
{
return NULL;
}
DIRECT_FN STATIC int gnix_cq_signal(struct fid_cq *cq)
{
struct gnix_fid_cq *cq_priv;
cq_priv = container_of(cq, struct gnix_fid_cq, cq_fid);
if (cq_priv->wait)
_gnix_signal_wait_obj(cq_priv->wait);
return FI_SUCCESS;
}
static int gnix_cq_control(struct fid *cq, int command, void *arg)
{
switch (command) {
case FI_GETWAIT:
return -FI_ENOSYS;
default:
return -FI_EINVAL;
}
}
DIRECT_FN int gnix_cq_open(struct fid_domain *domain, struct fi_cq_attr *attr,
struct fid_cq **cq, void *context)
{
struct gnix_fid_domain *domain_priv;
struct gnix_fid_cq *cq_priv;
struct fi_ops_cq *cq_ops;
struct fi_ops *fi_cq_ops;
int ret = FI_SUCCESS;
GNIX_TRACE(FI_LOG_CQ, "\n");
cq_ops = calloc(1, sizeof(*cq_ops));
if (!cq_ops) {
return -FI_ENOMEM;
}
fi_cq_ops = calloc(1, sizeof(*fi_cq_ops));
if (!fi_cq_ops) {
ret = -FI_ENOMEM;
goto free_cq_ops;
}
*cq_ops = gnix_cq_ops;
*fi_cq_ops = gnix_cq_fi_ops;
ret = verify_cq_attr(attr, cq_ops, fi_cq_ops);
if (ret)
goto free_fi_cq_ops;
domain_priv = container_of(domain, struct gnix_fid_domain, domain_fid);
if (!domain_priv) {
ret = -FI_EINVAL;
goto free_fi_cq_ops;
}
cq_priv = calloc(1, sizeof(*cq_priv));
if (!cq_priv) {
ret = -FI_ENOMEM;
goto free_fi_cq_ops;
}
cq_priv->requires_lock = (domain_priv->thread_model !=
FI_THREAD_COMPLETION);
cq_priv->domain = domain_priv;
cq_priv->attr = *attr;
_gnix_ref_init(&cq_priv->ref_cnt, 1, __cq_destruct);
_gnix_ref_get(cq_priv->domain);
_gnix_prog_init(&cq_priv->pset);
cq_priv->cq_fid.fid.fclass = FI_CLASS_CQ;
cq_priv->cq_fid.fid.context = context;
cq_priv->cq_fid.fid.ops = fi_cq_ops;
cq_priv->cq_fid.ops = cq_ops;
/*
* Although we don't need to store entry_size since we're already
* storing the format, this might provide a performance benefit
* when allocating storage.
*/
cq_priv->entry_size = format_sizes[cq_priv->attr.format];
fastlock_init(&cq_priv->lock);
ret = gnix_cq_set_wait(cq_priv);
if (ret)
goto free_cq_priv;
ret = _gnix_queue_create(&cq_priv->events, alloc_cq_entry,
free_cq_entry, cq_priv->entry_size,
cq_priv->attr.size);
if (ret)
goto free_cq_priv;
ret = _gnix_queue_create(&cq_priv->errors, alloc_cq_entry,
free_cq_entry, sizeof(struct fi_cq_err_entry),
0);
if (ret)
goto free_gnix_queue;
*cq = &cq_priv->cq_fid;
return ret;
free_gnix_queue:
_gnix_queue_destroy(cq_priv->events);
free_cq_priv:
_gnix_ref_put(cq_priv->domain);
fastlock_destroy(&cq_priv->lock);
free(cq_priv);
free_fi_cq_ops:
free(fi_cq_ops);
free_cq_ops:
free(cq_ops);
return ret;
}
/*******************************************************************************
* FI_OPS_* data structures.
******************************************************************************/
static const struct fi_ops gnix_cq_fi_ops = {
.size = sizeof(struct fi_ops),
.close = gnix_cq_close,
.bind = fi_no_bind,
.control = gnix_cq_control,
.ops_open = fi_no_ops_open
};
static const struct fi_ops_cq gnix_cq_ops = {
.size = sizeof(struct fi_ops_cq),
.read = gnix_cq_read,
.readfrom = gnix_cq_readfrom,
.readerr = gnix_cq_readerr,
.sread = gnix_cq_sread,
.sreadfrom = gnix_cq_sreadfrom,
.signal = gnix_cq_signal,
.strerror = gnix_cq_strerror
};
DIRECT_FN STATIC int gnix_accept(struct fid_ep *ep, const void *param,
size_t paramlen)
{
int ret;
struct gnix_vc *vc;
struct gnix_fid_ep *ep_priv;
struct gnix_pep_sock_conn *conn;
struct gnix_pep_sock_connresp resp;
struct fi_eq_cm_entry eq_entry, *eqe_ptr;
struct gnix_mbox *mbox = NULL;
struct gnix_av_addr_entry av_entry;
if (!ep || (paramlen && !param) || paramlen > GNIX_CM_DATA_MAX_SIZE)
return -FI_EINVAL;
ep_priv = container_of(ep, struct gnix_fid_ep, ep_fid.fid);
COND_ACQUIRE(ep_priv->requires_lock, &ep_priv->vc_lock);
/* Look up and unpack the connection request used to create this EP. */
conn = (struct gnix_pep_sock_conn *)ep_priv->info->handle;
if (!conn || conn->fid.fclass != FI_CLASS_CONNREQ) {
ret = -FI_EINVAL;
goto err_unlock;
}
/* Create new VC without CM data. */
av_entry.gnix_addr = ep_priv->dest_addr.gnix_addr;
av_entry.cm_nic_cdm_id = ep_priv->dest_addr.cm_nic_cdm_id;
ret = _gnix_vc_alloc(ep_priv, &av_entry, &vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL, "Failed to create VC: %d\n", ret);
goto err_unlock;
}
ep_priv->vc = vc;
ep_priv->vc->peer_caps = conn->req.peer_caps;
ep_priv->vc->peer_id = conn->req.vc_id;
ret = _gnix_mbox_alloc(vc->ep->nic->mbox_hndl, &mbox);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA, "_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto err_mbox_alloc;
}
vc->smsg_mbox = mbox;
/* Initialize the GNI connection. */
ret = _gnix_vc_smsg_init(vc, conn->req.vc_id,
&conn->req.vc_mbox_attr,
&conn->req.cq_irq_mdh);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"_gnix_vc_smsg_init returned %s\n",
fi_strerror(-ret));
goto err_smsg_init;
}
vc->conn_state = GNIX_VC_CONNECTED;
/* Send ACK with VC attrs to allow peer to initialize GNI connection. */
resp.cmd = GNIX_PEP_SOCK_RESP_ACCEPT;
resp.vc_id = vc->vc_id;
resp.vc_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
resp.vc_mbox_attr.msg_buffer = mbox->base;
resp.vc_mbox_attr.buff_size = vc->ep->nic->mem_per_mbox;
resp.vc_mbox_attr.mem_hndl = *mbox->memory_handle;
resp.vc_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
resp.vc_mbox_attr.mbox_maxcredit =
ep_priv->domain->params.mbox_maxcredit;
resp.vc_mbox_attr.msg_maxsize =
ep_priv->domain->params.mbox_msg_maxsize;
resp.cq_irq_mdh = ep_priv->nic->irq_mem_hndl;
resp.peer_caps = ep_priv->caps;
resp.cm_data_len = paramlen;
if (paramlen) {
eqe_ptr = (struct fi_eq_cm_entry *)resp.eqe_buf;
memcpy(eqe_ptr->data, param, paramlen);
}
ret = write(conn->sock_fd, &resp, sizeof(resp));
if (ret != sizeof(resp)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to send resp, errno: %d\n",
errno);
ret = -FI_EIO;
goto err_write;
}
/* Notify user that this side is connected. */
eq_entry.fid = &ep_priv->ep_fid.fid;
ret = fi_eq_write(&ep_priv->eq->eq_fid, FI_CONNECTED, &eq_entry,
sizeof(eq_entry), 0);
if (ret != sizeof(eq_entry)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"fi_eq_write failed, err: %d\n", ret);
goto err_eq_write;
}
/* Free the connection request. */
free(conn);
ep_priv->conn_state = GNIX_EP_CONNECTED;
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
GNIX_DEBUG(FI_LOG_EP_CTRL, "Sent conn accept: %p\n", ep_priv);
return FI_SUCCESS;
err_eq_write:
err_write:
err_smsg_init:
_gnix_mbox_free(ep_priv->vc->smsg_mbox);
ep_priv->vc->smsg_mbox = NULL;
err_mbox_alloc:
_gnix_vc_destroy(ep_priv->vc);
ep_priv->vc = NULL;
err_unlock:
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
return ret;
}
DIRECT_FN STATIC int gnix_connect(struct fid_ep *ep, const void *addr,
const void *param, size_t paramlen)
{
int ret;
struct gnix_fid_ep *ep_priv;
struct sockaddr_in saddr;
struct gnix_pep_sock_connreq req;
struct fi_eq_cm_entry *eqe_ptr;
struct gnix_vc *vc;
struct gnix_mbox *mbox = NULL;
struct gnix_av_addr_entry av_entry;
if (!ep || !addr || (paramlen && !param) ||
paramlen > GNIX_CM_DATA_MAX_SIZE)
return -FI_EINVAL;
ep_priv = container_of(ep, struct gnix_fid_ep, ep_fid.fid);
COND_ACQUIRE(ep_priv->requires_lock, &ep_priv->vc_lock);
if (ep_priv->conn_state != GNIX_EP_UNCONNECTED) {
ret = -FI_EINVAL;
goto err_unlock;
}
ret = _gnix_pe_to_ip(addr, &saddr);
if (ret != FI_SUCCESS) {
GNIX_INFO(FI_LOG_EP_CTRL,
"Failed to translate gnix_ep_name to IP\n");
goto err_unlock;
}
/* Create new VC without CM data. */
av_entry.gnix_addr = ep_priv->dest_addr.gnix_addr;
av_entry.cm_nic_cdm_id = ep_priv->dest_addr.cm_nic_cdm_id;
ret = _gnix_vc_alloc(ep_priv, &av_entry, &vc);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to create VC:: %d\n",
ret);
goto err_unlock;
}
ep_priv->vc = vc;
ret = _gnix_mbox_alloc(vc->ep->nic->mbox_hndl, &mbox);
if (ret != FI_SUCCESS) {
GNIX_WARN(FI_LOG_EP_DATA,
"_gnix_mbox_alloc returned %s\n",
fi_strerror(-ret));
goto err_mbox_alloc;
}
vc->smsg_mbox = mbox;
ep_priv->conn_fd = socket(AF_INET, SOCK_STREAM, 0);
if (ep_priv->conn_fd < 0) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to create connect socket, errno: %d\n",
errno);
ret = -FI_ENOSPC;
goto err_socket;
}
/* Currently blocks until connected. */
ret = connect(ep_priv->conn_fd, (struct sockaddr *)&saddr,
sizeof(saddr));
if (ret) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to connect, errno: %d\n",
errno);
ret = -FI_EIO;
goto err_connect;
}
req.info = *ep_priv->info;
/* Note addrs are swapped. */
memcpy(&req.dest_addr, (void *)&ep_priv->src_addr,
sizeof(req.dest_addr));
memcpy(&ep_priv->dest_addr, addr, sizeof(ep_priv->dest_addr));
memcpy(&req.src_addr, addr, sizeof(req.src_addr));
if (ep_priv->info->tx_attr)
req.tx_attr = *ep_priv->info->tx_attr;
if (ep_priv->info->rx_attr)
req.rx_attr = *ep_priv->info->rx_attr;
if (ep_priv->info->ep_attr)
req.ep_attr = *ep_priv->info->ep_attr;
if (ep_priv->info->domain_attr)
req.domain_attr = *ep_priv->info->domain_attr;
if (ep_priv->info->fabric_attr)
req.fabric_attr = *ep_priv->info->fabric_attr;
req.fabric_attr.fabric = NULL;
req.domain_attr.domain = NULL;
req.vc_id = vc->vc_id;
req.vc_mbox_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
req.vc_mbox_attr.msg_buffer = mbox->base;
req.vc_mbox_attr.buff_size = vc->ep->nic->mem_per_mbox;
req.vc_mbox_attr.mem_hndl = *mbox->memory_handle;
req.vc_mbox_attr.mbox_offset = (uint64_t)mbox->offset;
req.vc_mbox_attr.mbox_maxcredit =
ep_priv->domain->params.mbox_maxcredit;
req.vc_mbox_attr.msg_maxsize = ep_priv->domain->params.mbox_msg_maxsize;
req.cq_irq_mdh = ep_priv->nic->irq_mem_hndl;
req.peer_caps = ep_priv->caps;
req.cm_data_len = paramlen;
if (paramlen) {
eqe_ptr = (struct fi_eq_cm_entry *)req.eqe_buf;
memcpy(eqe_ptr->data, param, paramlen);
}
ret = write(ep_priv->conn_fd, &req, sizeof(req));
if (ret != sizeof(req)) {
GNIX_WARN(FI_LOG_EP_CTRL,
"Failed to send req, errno: %d\n",
errno);
ret = -FI_EIO;
goto err_write;
}
/* set fd to non-blocking now since we can't block within the eq
* progress system
*/
fi_fd_nonblock(ep_priv->conn_fd);
ep_priv->conn_state = GNIX_EP_CONNECTING;
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
GNIX_DEBUG(FI_LOG_EP_CTRL, "Sent conn req: %p, %s\n",
ep_priv, inet_ntoa(saddr.sin_addr));
return FI_SUCCESS;
err_write:
err_connect:
close(ep_priv->conn_fd);
ep_priv->conn_fd = -1;
err_socket:
_gnix_mbox_free(ep_priv->vc->smsg_mbox);
ep_priv->vc->smsg_mbox = NULL;
err_mbox_alloc:
_gnix_vc_destroy(ep_priv->vc);
ep_priv->vc = NULL;
err_unlock:
COND_RELEASE(ep_priv->requires_lock, &ep_priv->vc_lock);
return ret;
}
ssize_t _gnix_atomic(struct gnix_fid_ep *ep,
enum gnix_fab_req_type fr_type,
const struct fi_msg_atomic *msg,
const struct fi_ioc *comparev,
void **compare_desc,
size_t compare_count,
struct fi_ioc *resultv,
void **result_desc,
size_t result_count,
uint64_t flags)
{
struct gnix_vc *vc;
struct gnix_fab_req *req;
struct gnix_fid_mem_desc *md = NULL;
int rc, len;
struct fid_mr *auto_mr = NULL;
void *mdesc = NULL;
uint64_t compare_operand = 0;
void *loc_addr = NULL;
int dt_len, dt_align;
int connected;
if (!(flags & FI_INJECT) && !ep->send_cq &&
(((fr_type == GNIX_FAB_RQ_AMO ||
fr_type == GNIX_FAB_RQ_NAMO_AX ||
fr_type == GNIX_FAB_RQ_NAMO_AX_S) &&
!ep->write_cntr) ||
((fr_type == GNIX_FAB_RQ_FAMO ||
fr_type == GNIX_FAB_RQ_CAMO ||
fr_type == GNIX_FAB_RQ_NAMO_FAX ||
fr_type == GNIX_FAB_RQ_NAMO_FAX_S) &&
!ep->read_cntr))) {
return -FI_ENOCQ;
}
if (!ep || !msg || !msg->msg_iov ||
msg->msg_iov[0].count != 1 ||
msg->iov_count != GNIX_MAX_ATOMIC_IOV_LIMIT ||
!msg->rma_iov)
return -FI_EINVAL;
/*
* see fi_atomic man page
*/
if ((msg->op != FI_ATOMIC_READ) &&
!msg->msg_iov[0].addr)
return -FI_EINVAL;
if (flags & FI_TRIGGER) {
struct fi_triggered_context *trigger_context =
(struct fi_triggered_context *)msg->context;
if ((trigger_context->event_type != FI_TRIGGER_THRESHOLD) ||
(flags & FI_INJECT)) {
return -FI_EINVAL;
}
}
if (fr_type == GNIX_FAB_RQ_CAMO) {
if (!comparev || !comparev[0].addr || compare_count != 1)
return -FI_EINVAL;
compare_operand = *(uint64_t *)comparev[0].addr;
}
dt_len = ofi_datatype_size(msg->datatype);
dt_align = dt_len - 1;
len = dt_len * msg->msg_iov->count;
if (msg->rma_iov->addr & dt_align) {
GNIX_INFO(FI_LOG_EP_DATA,
"Invalid target alignment: %d (mask 0x%x)\n",
msg->rma_iov->addr, dt_align);
return -FI_EINVAL;
}
/* need a memory descriptor for all fetching and comparison AMOs */
if (fr_type == GNIX_FAB_RQ_FAMO ||
fr_type == GNIX_FAB_RQ_CAMO ||
fr_type == GNIX_FAB_RQ_NAMO_FAX ||
fr_type == GNIX_FAB_RQ_NAMO_FAX_S) {
if (!resultv || !resultv[0].addr || result_count != 1)
return -FI_EINVAL;
loc_addr = resultv[0].addr;
if ((uint64_t)loc_addr & dt_align) {
GNIX_INFO(FI_LOG_EP_DATA,
"Invalid source alignment: %d (mask 0x%x)\n",
loc_addr, dt_align);
return -FI_EINVAL;
}
if (!result_desc || !result_desc[0]) {
rc = _gnix_mr_reg(&ep->domain->domain_fid.fid,
loc_addr, len, FI_READ | FI_WRITE,
0, 0, 0, &auto_mr,
NULL, ep->auth_key, GNIX_PROV_REG);
if (rc != FI_SUCCESS) {
GNIX_INFO(FI_LOG_EP_DATA,
"Failed to auto-register local buffer: %d\n",
rc);
return rc;
}
flags |= FI_LOCAL_MR;
mdesc = (void *)auto_mr;
GNIX_INFO(FI_LOG_EP_DATA, "auto-reg MR: %p\n",
auto_mr);
} else {
mdesc = result_desc[0];
}
}
/* setup fabric request */
req = _gnix_fr_alloc(ep);
if (!req) {
GNIX_INFO(FI_LOG_EP_DATA, "_gnix_fr_alloc() failed\n");
rc = -FI_ENOSPC;
goto err_fr_alloc;
}
req->type = fr_type;
req->gnix_ep = ep;
req->user_context = msg->context;
req->work_fn = _gnix_amo_post_req;
if (mdesc) {
md = container_of(mdesc, struct gnix_fid_mem_desc, mr_fid);
}
req->amo.loc_md = (void *)md;
req->amo.loc_addr = (uint64_t)loc_addr;
if ((fr_type == GNIX_FAB_RQ_NAMO_AX) ||
(fr_type == GNIX_FAB_RQ_NAMO_FAX) ||
(fr_type == GNIX_FAB_RQ_NAMO_AX_S) ||
(fr_type == GNIX_FAB_RQ_NAMO_FAX_S)) {
req->amo.first_operand =
*(uint64_t *)msg->msg_iov[0].addr;
req->amo.second_operand =
*((uint64_t *)(msg->msg_iov[0].addr) + 1);
} else if (msg->op == FI_ATOMIC_READ) {
req->amo.first_operand = 0xFFFFFFFFFFFFFFFF; /* operand to FAND */
} else if (msg->op == FI_CSWAP) {
req->amo.first_operand = compare_operand;
req->amo.second_operand = *(uint64_t *)msg->msg_iov[0].addr;
} else if (msg->op == FI_MSWAP) {
req->amo.first_operand = ~compare_operand;
req->amo.second_operand = *(uint64_t *)msg->msg_iov[0].addr;
req->amo.second_operand &= compare_operand;
} else {
req->amo.first_operand = *(uint64_t *)msg->msg_iov[0].addr;
}
req->amo.rem_addr = msg->rma_iov->addr;
req->amo.rem_mr_key = msg->rma_iov->key;
req->amo.len = len;
req->amo.imm = msg->data;
req->amo.datatype = msg->datatype;
req->amo.op = msg->op;
req->flags = flags;
/* Inject interfaces always suppress completions. If
* SELECTIVE_COMPLETION is set, honor any setting. Otherwise, always
* deliver a completion. */
if ((flags & GNIX_SUPPRESS_COMPLETION) ||
(ep->send_selective_completion && !(flags & FI_COMPLETION))) {
req->flags &= ~FI_COMPLETION;
} else {
req->flags |= FI_COMPLETION;
}
COND_ACQUIRE(ep->requires_lock, &ep->vc_lock);
/* find VC for target */
rc = _gnix_vc_ep_get_vc(ep, msg->addr, &vc);
if (rc) {
GNIX_INFO(FI_LOG_EP_DATA,
"_gnix_vc_ep_get_vc() failed, addr: %lx, rc:\n",
msg->addr, rc);
goto err_get_vc;
}
req->vc = vc;
rc = _gnix_vc_queue_tx_req(req);
connected = (vc->conn_state == GNIX_VC_CONNECTED);
COND_RELEASE(ep->requires_lock, &ep->vc_lock);
/*
*If a new VC was allocated, progress CM before returning.
* If the VC is connected and there's a backlog, poke
* the nic progress engine befure returning.
*/
if (!connected) {
_gnix_cm_nic_progress(ep->cm_nic);
} else if (!dlist_empty(&vc->tx_queue)) {
_gnix_nic_progress(vc->ep->nic);
}
return rc;
err_get_vc:
COND_RELEASE(ep->requires_lock, &ep->vc_lock);
err_fr_alloc:
if (auto_mr) {
fi_close(&auto_mr->fid);
}
return rc;
}
int _gnix_amo_post_req(void *data)
{
struct gnix_fab_req *fab_req = (struct gnix_fab_req *)data;
struct gnix_fid_ep *ep = fab_req->gnix_ep;
struct gnix_nic *nic = ep->nic;
struct gnix_fid_mem_desc *loc_md;
struct gnix_tx_descriptor *txd;
gni_mem_handle_t mdh;
gni_return_t status;
int rc;
int inject_err = _gnix_req_inject_err(fab_req);
if (!gnix_ops_allowed(ep, fab_req->vc->peer_caps, fab_req->flags)) {
GNIX_DEBUG(FI_LOG_EP_DATA, "flags:0x%llx, %s\n", fab_req->flags,
fi_tostr(&fab_req->flags, FI_TYPE_OP_FLAGS));
GNIX_DEBUG(FI_LOG_EP_DATA, "caps:0x%llx, %s\n",
ep->caps, fi_tostr(&ep->caps, FI_TYPE_CAPS));
GNIX_DEBUG(FI_LOG_EP_DATA, "peer_caps:0x%llx, %s\n",
fab_req->vc->peer_caps,
fi_tostr(&fab_req->vc->peer_caps, FI_TYPE_OP_FLAGS));
rc = __gnix_amo_post_err(fab_req, FI_EOPNOTSUPP);
if (rc != FI_SUCCESS)
GNIX_WARN(FI_LOG_EP_DATA,
"__gnix_amo_post_err() failed: %d\n", rc);
return -FI_ECANCELED;
}
rc = _gnix_nic_tx_alloc(nic, &txd);
if (rc) {
GNIX_INFO(FI_LOG_EP_DATA, "_gnix_nic_tx_alloc() failed: %d\n",
rc);
return -FI_ENOSPC;
}
txd->completer_fn = __gnix_amo_txd_complete;
txd->req = fab_req;
/* Mem handle CRC is not validated during FMA operations. Skip this
* costly calculation. */
_gnix_convert_key_to_mhdl_no_crc(
(gnix_mr_key_t *)&fab_req->amo.rem_mr_key,
&mdh);
loc_md = (struct gnix_fid_mem_desc *)fab_req->amo.loc_md;
txd->gni_desc.type = GNI_POST_AMO;
txd->gni_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT; /* check flags */
txd->gni_desc.dlvr_mode = GNI_DLVMODE_PERFORMANCE; /* check flags */
txd->gni_desc.local_addr = (uint64_t)fab_req->amo.loc_addr;
if (loc_md) {
txd->gni_desc.local_mem_hndl = loc_md->mem_hndl;
}
txd->gni_desc.remote_addr = (uint64_t)fab_req->amo.rem_addr;
txd->gni_desc.remote_mem_hndl = mdh;
txd->gni_desc.length = fab_req->amo.len;
txd->gni_desc.rdma_mode = 0; /* check flags */
txd->gni_desc.src_cq_hndl = nic->tx_cq; /* check flags */
txd->gni_desc.amo_cmd = _gnix_atomic_cmd(fab_req->amo.datatype,
fab_req->amo.op,
fab_req->type);
txd->gni_desc.first_operand = fab_req->amo.first_operand;
txd->gni_desc.second_operand = fab_req->amo.second_operand;
GNIX_DEBUG(FI_LOG_EP_DATA, "fo:%016lx so:%016lx\n",
txd->gni_desc.first_operand, txd->gni_desc.second_operand);
GNIX_DEBUG(FI_LOG_EP_DATA, "amo_cmd:%x\n",
txd->gni_desc.amo_cmd);
GNIX_LOG_DUMP_TXD(txd);
COND_ACQUIRE(nic->requires_lock, &nic->lock);
if (OFI_UNLIKELY(inject_err)) {
_gnix_nic_txd_err_inject(nic, txd);
status = GNI_RC_SUCCESS;
} else {
status = GNI_PostFma(fab_req->vc->gni_ep, &txd->gni_desc);
}
COND_RELEASE(nic->requires_lock, &nic->lock);
if (status != GNI_RC_SUCCESS) {
_gnix_nic_tx_free(nic, txd);
GNIX_INFO(FI_LOG_EP_DATA, "GNI_Post*() failed: %s\n",
gni_err_str[status]);
}
return gnixu_to_fi_errno(status);
}
static inline void *__gnix_generic_register(
struct gnix_fid_domain *domain,
struct gnix_fid_mem_desc *md,
void *address,
size_t length,
gni_cq_handle_t dst_cq_hndl,
int flags,
int vmdh_index)
{
struct gnix_nic *nic;
gni_return_t grc = GNI_RC_SUCCESS;
int rc;
pthread_mutex_lock(&gnix_nic_list_lock);
/* If the nic list is empty, create a nic */
if (unlikely((dlist_empty(&gnix_nic_list_ptag[domain->ptag])))) {
/* release the lock because we are not checking the list after
this point. Additionally, gnix_nic_alloc takes the
lock to add the nic. */
pthread_mutex_unlock(&gnix_nic_list_lock);
rc = gnix_nic_alloc(domain, NULL, &nic);
if (rc) {
GNIX_INFO(FI_LOG_MR,
"could not allocate nic to do mr_reg,"
" ret=%i\n", rc);
return NULL;
}
} else {
nic = dlist_first_entry(&gnix_nic_list_ptag[domain->ptag],
struct gnix_nic, ptag_nic_list);
if (unlikely(nic == NULL)) {
GNIX_ERR(FI_LOG_MR, "Failed to find nic on "
"ptag list\n");
pthread_mutex_unlock(&gnix_nic_list_lock);
return NULL;
}
_gnix_ref_get(nic);
pthread_mutex_unlock(&gnix_nic_list_lock);
}
COND_ACQUIRE(nic->requires_lock, &nic->lock);
grc = GNI_MemRegister(nic->gni_nic_hndl, (uint64_t) address,
length, dst_cq_hndl, flags,
vmdh_index, &md->mem_hndl);
COND_RELEASE(nic->requires_lock, &nic->lock);
if (unlikely(grc != GNI_RC_SUCCESS)) {
GNIX_INFO(FI_LOG_MR, "failed to register memory with uGNI, "
"ret=%s\n", gni_err_str[grc]);
_gnix_ref_put(nic);
return NULL;
}
/* set up the mem desc */
md->nic = nic;
md->domain = domain;
/* take references on domain */
_gnix_ref_get(md->domain);
return md;
}
