예제 #1
0
static int fi_bgq_bind_sep(struct fid *fid, struct fid *bfid,
		uint64_t flags)
{
	int ret = 0;
	struct fi_bgq_sep *bgq_sep = container_of(fid, struct fi_bgq_sep, ep_fid);
	struct fi_bgq_av *bgq_av;

	if (!fid || !bfid) {
		errno = FI_EINVAL;
		return -errno;
	}

	switch (bfid->fclass) {
	case FI_CLASS_AV:
		bgq_av = container_of(bfid, struct fi_bgq_av, av_fid);
		fi_bgq_ref_inc(&bgq_av->ref_cnt, "address vector");
		bgq_sep->av = bgq_av;
		break;
	default:
		errno = FI_ENOSYS;
		return -errno;
	}

	return ret;
}
예제 #2
0
int fi_bgq_progress_init (struct fi_bgq_domain *bgq_domain, const uint64_t max_threads) {

	assert(max_threads < (64/Kernel_ProcessCount()));

	bgq_domain->progress.max_threads = max_threads;
	bgq_domain->progress.num_threads_active = 0;
	bgq_domain->progress.memptr = NULL;

	if (0 == max_threads) return 0;

	size_t i, j;

	const size_t bytes = sizeof(union fi_bgq_progress_data) * max_threads;
	const size_t alignment = 128;

	void * memptr = malloc(bytes+alignment);
	uint32_t cnk_rc = 0;
	cnk_rc = Kernel_L2AtomicsAllocate(memptr, bytes+alignment);
	assert(0==cnk_rc);
	if (cnk_rc != 0) {
		/* Error allocating l2atomic memory */
		free(memptr);
		bgq_domain->progress.memptr = NULL;
		return -1;
	}

	union fi_bgq_progress_data *data = (union fi_bgq_progress_data *)(((uintptr_t)memptr + alignment) & (~(alignment-1)));
	const uint64_t npackets = sizeof(data[0].data) / sizeof(uint64_t);
	const size_t nep = sizeof(bgq_domain->progress.thread[0].tag_ep) / sizeof(struct fi_bgq_ep *);
	for (i=0; i<max_threads; ++i) {
		l2atomic_fifo_initialize(&bgq_domain->progress.thread[i].consumer,
			&bgq_domain->progress.thread[i].producer,
			&data[i].l2atomic, npackets);

		bgq_domain->progress.thread[i].tag_ep_count = 0;
		bgq_domain->progress.thread[i].msg_ep_count = 0;
		bgq_domain->progress.thread[i].all_ep_count = 0;
		bgq_domain->progress.thread[i].pthread = 0;
		bgq_domain->progress.thread[i].enabled = 0;
		bgq_domain->progress.thread[i].active = 0;

		fi_bgq_ref_inc(&bgq_domain->ref_cnt, "domain");
		bgq_domain->progress.thread[i].bgq_domain = bgq_domain;

		for (j=0; j<nep; ++j) {
			bgq_domain->progress.thread[i].tag_ep[j] = NULL;
			bgq_domain->progress.thread[i].msg_ep[j] = NULL;
			bgq_domain->progress.thread[i].all_ep[j] = NULL;
		}
	}

	bgq_domain->progress.memptr = memptr;

	return 0;
}
예제 #3
0
static int fi_bgq_tx_ctx(struct fid_ep *sep, int index,
			struct fi_tx_attr *attr, struct fid_ep **tx_ep,
			void *context)
{
	int ret;
	struct fi_info info = {0};
	struct fi_tx_attr tx_attr = {0};
	struct fi_ep_attr ep_attr = {0};
	struct fi_domain_attr dom_attr = {0};
	struct fi_fabric_attr fab_attr = {0};
	struct fi_bgq_sep *bgq_sep;
	struct fi_bgq_ep  *bgq_tx_ep;

	if (!sep || !attr || !tx_ep) {
		errno = FI_EINVAL;
		return -errno;
	}

	bgq_sep = container_of(sep, struct fi_bgq_sep, ep_fid);

	uint64_t caps = attr->caps;	/* TODO - "By default, a transmit context inherits the properties of its associated endpoint. However, applications may request context specific attributes through the attr parameter." */

	if ((caps & FI_MSG || caps & FI_TAGGED) && (caps & FI_RECV)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"FI_MSG|FI_TAGGED with FI_RECV capability specified for a TX context\n");
		caps &= ~FI_RECV;
	}

	if ((caps & FI_RMA || caps & FI_ATOMIC) && (caps & FI_REMOTE_READ || caps & FI_REMOTE_WRITE)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"FI_RMA|FI_ATOMIC with FI_REMOTE_READ|FI_REMOTE_WRITE capability specified for a TX context\n");
		caps &= ~FI_REMOTE_READ;
		caps &= ~FI_REMOTE_WRITE;
	}

	if (caps & FI_MSG || caps & FI_TAGGED) {
		caps |= FI_SEND;
	}

	if (caps & FI_RMA || caps & FI_ATOMIC) {
		caps |= FI_READ;
		caps |= FI_WRITE;
	}

	if (ofi_recv_allowed(caps) || ofi_rma_target_allowed(caps)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"RX capabilities specified for TX context\n");
		errno = FI_EINVAL;
		return -errno;
	}

	if (!ofi_send_allowed(caps) && !ofi_rma_initiate_allowed(caps)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"TX capabilities not specified for TX context\n");
		errno = FI_EINVAL;
		return -errno;
	}

	if (bgq_sep->domain->tx.count >= fi_bgq_domain_get_tx_max(bgq_sep->domain)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"TX ctx count exceeded (max %lu, created %lu)\n",
				fi_bgq_domain_get_tx_max(bgq_sep->domain), bgq_sep->domain->tx.count);
		errno = FI_EINVAL;
		return -errno;
	}

	info.caps = caps;
	info.mode = attr->mode;

	info.tx_attr = &tx_attr;
	memcpy(info.tx_attr, attr, sizeof(*info.tx_attr));

	info.ep_attr = &ep_attr;
	memcpy(info.ep_attr, bgq_sep->info->ep_attr, sizeof(*info.ep_attr));

	info.domain_attr = &dom_attr;
	memcpy(info.domain_attr, bgq_sep->info->domain_attr, sizeof(*info.domain_attr));

	info.fabric_attr = &fab_attr;
	memcpy(info.fabric_attr, bgq_sep->info->fabric_attr, sizeof(*info.fabric_attr));

	ret = fi_bgq_endpoint_rx_tx((struct fid_domain *)bgq_sep->domain,
		&info, tx_ep, context, -1, index);
	if (ret) {
		goto err;
	}

	bgq_tx_ep = container_of(*tx_ep, struct fi_bgq_ep, ep_fid);
	bgq_tx_ep->ep_fid.fid.fclass = FI_CLASS_TX_CTX;

	bgq_tx_ep->av = bgq_sep->av;
	fi_bgq_ref_inc(&bgq_tx_ep->av->ref_cnt, "address vector");

	bgq_tx_ep->sep = container_of(sep, struct fi_bgq_sep, ep_fid);

	++ bgq_sep->domain->tx.count;

	fi_bgq_ref_inc(&bgq_sep->ref_cnt, "scalable endpoint");

	attr->caps = caps;

	return 0;

err:
	return -errno;
}
예제 #4
0
int fi_bgq_scalable_ep (struct fid_domain *domain,
	struct fi_info *info,
	struct fid_ep **sep,
	void *context)
{
	struct fi_bgq_sep *bgq_sep = NULL;

	if (!info || !domain) {
		errno = FI_EINVAL;
		goto err;
	}

	void * memptr = NULL;
	memptr = malloc(sizeof(struct fi_bgq_sep)+L2_CACHE_LINE_SIZE);
	if (!memptr) {
		errno = FI_ENOMEM;
		goto err;
	}
	memset(memptr, 0, sizeof(struct fi_bgq_sep)+L2_CACHE_LINE_SIZE);
	bgq_sep = (struct fi_bgq_sep *)(((uintptr_t)memptr+L2_CACHE_LINE_SIZE) & ~(L2_CACHE_LINE_SIZE-1));
	bgq_sep->memptr = memptr;
	memptr = NULL;

	bgq_sep->domain = (struct fi_bgq_domain *) domain;

	bgq_sep->ep_fid.fid.fclass	= FI_CLASS_SEP;
	bgq_sep->ep_fid.fid.context	= context;
	bgq_sep->ep_fid.fid.ops		= &fi_bgq_fi_ops;
	bgq_sep->ep_fid.ops		= &fi_bgq_sep_ops;

	bgq_sep->info = calloc(1, sizeof (struct fi_info));
	if (!bgq_sep->info) {
		errno = FI_ENOMEM;
		goto err;
	}
	memcpy(bgq_sep->info, info, sizeof (struct fi_info));
	bgq_sep->info->next = NULL;
	bgq_sep->info->ep_attr = calloc(1, sizeof(struct fi_ep_attr));
	if (!bgq_sep->info->ep_attr) {
		errno = FI_ENOMEM;
		goto err;
	}
	memcpy(bgq_sep->info->ep_attr, info->ep_attr, sizeof(struct fi_ep_attr));

	/*
	 * fi_endpoint.3
	 *
	 * "tx_ctx_cnt - Transmit Context Count
	 * 	Number of transmit contexts to associate with the endpoint. If
	 * 	not specified (0), 1 context will be assigned if the endpoint
	 * 	supports outbound transfers."
	 */
	if (0 == bgq_sep->info->ep_attr->tx_ctx_cnt) {
		bgq_sep->info->ep_attr->tx_ctx_cnt = 1;
	}

	/*
	 * fi_endpoint.3
	 *
	 * "rx_ctx_cnt - Receive Context Count
	 * 	Number of receive contexts to associate with the endpoint. If
	 * 	not specified, 1 context will be assigned if the endpoint
	 * 	supports inbound transfers."
	 */
	if (0 == bgq_sep->info->ep_attr->rx_ctx_cnt) {
		bgq_sep->info->ep_attr->rx_ctx_cnt = 1;
	}

	fi_bgq_ref_init(&bgq_sep->domain->fabric->node, &bgq_sep->ref_cnt, "scalable endpoint");
	fi_bgq_ref_inc(&bgq_sep->domain->ref_cnt, "domain");

	*sep = &bgq_sep->ep_fid;

	return 0;
err:
	if (bgq_sep) {
		if (bgq_sep->info) {
			if (bgq_sep->info->ep_attr)
				free(bgq_sep->info->ep_attr);
			free(bgq_sep->info);
		}
		memptr = bgq_sep->memptr;
		free(memptr);
	}
	return -errno;
}
예제 #5
0
static int fi_bgq_rx_ctx(struct fid_ep *sep, int index,
			struct fi_rx_attr *attr, struct fid_ep **rx_ep,
			void *context)
{
	int ret;
	struct fi_info info = {0};
	struct fi_bgq_sep *bgq_sep;
	struct fi_bgq_ep  *bgq_rx_ep;

	if (!sep || !attr || !rx_ep) {
		errno = FI_EINVAL;
		return -errno;
	}

	bgq_sep = container_of(sep, struct fi_bgq_sep, ep_fid);

	uint64_t caps = attr->caps;	/* TODO - "By default, a receive context inherits the properties of its associated endpoint. However, applications may request context specific attributes through the attr parameter." */

	if ((caps & FI_MSG || caps & FI_TAGGED) && (caps & FI_SEND)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"FI_MSG|FI_TAGGED with FI_SEND capability specified for a RX context\n");
		caps &= ~FI_SEND;
	}

	if ((caps & FI_RMA || caps & FI_ATOMIC) && (caps & FI_READ || caps & FI_WRITE)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"FI_RMA|FI_ATOMIC with FI_READ|FI_WRITE capability specified for a RX context\n");
		caps &= ~FI_READ;
		caps &= ~FI_WRITE;
	}

	if (caps & FI_MSG || caps & FI_TAGGED) {
		caps |= FI_RECV;
	}

	if (caps & FI_RMA || caps & FI_ATOMIC) {
		caps |= FI_REMOTE_READ;
		caps |= FI_REMOTE_WRITE;
	}

	if (ofi_send_allowed(caps) || ofi_rma_initiate_allowed(caps)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"TX capabilities specified for RX context\n");
		errno = FI_EINVAL;
		return -errno;
	}

	if (!ofi_recv_allowed(caps) && !ofi_rma_target_allowed(caps)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"RX capabilities not specified for RX context\n");
		errno = FI_EINVAL;
		return -errno;
	}

	if (bgq_sep->domain->rx.count >= fi_bgq_domain_get_rx_max(bgq_sep->domain)) {
		FI_LOG(fi_bgq_global.prov, FI_LOG_DEBUG, FI_LOG_DOMAIN,
				"RX ctx count exceeded (max %lu, created %lu)\n",
				fi_bgq_domain_get_rx_max(bgq_sep->domain), bgq_sep->domain->rx.count);
		errno = FI_EINVAL;
		return -errno;
	}

	info.caps = caps;
	info.mode = attr->mode;

	info.rx_attr = calloc(1, sizeof(*info.rx_attr));
	if (!info.rx_attr) {
		errno = FI_ENOMEM;
		goto err;
	}

	info.rx_attr->caps     = caps;
	info.rx_attr->mode     = attr->mode;
	info.rx_attr->op_flags = attr->op_flags;
	info.rx_attr->msg_order = attr->msg_order;
	info.rx_attr->total_buffered_recv = attr->total_buffered_recv;
	info.rx_attr->iov_limit = attr->iov_limit;

	info.ep_attr = calloc(1, sizeof(*info.ep_attr));
	if (!info.ep_attr) {
		errno = FI_ENOMEM;
		goto err;
	}
	memcpy(info.ep_attr, bgq_sep->info->ep_attr,
			sizeof(*info.ep_attr));

	info.domain_attr = calloc(1, sizeof(*info.domain_attr));
	if (!info.domain_attr) {
		errno = FI_ENOMEM;
		goto err;
	}
	memcpy(info.domain_attr, bgq_sep->info->domain_attr,
			sizeof(*info.domain_attr));

	info.fabric_attr = calloc(1, sizeof(*info.fabric_attr));
	if (!info.fabric_attr) {
		errno = FI_ENOMEM;
		goto err;
	}
	memcpy(info.fabric_attr, bgq_sep->info->fabric_attr,
			sizeof(*info.fabric_attr));

	ret = fi_bgq_endpoint_rx_tx(&bgq_sep->domain->domain_fid, &info,
			rx_ep, context, index, -1);
	if (ret) {
		goto err;
	}

	bgq_rx_ep = container_of(*rx_ep, struct fi_bgq_ep, ep_fid);
	bgq_rx_ep->ep_fid.fid.fclass = FI_CLASS_RX_CTX;

	bgq_rx_ep->sep = container_of(sep, struct fi_bgq_sep, ep_fid);

	bgq_rx_ep->av = bgq_sep->av;
	fi_bgq_ref_inc(&bgq_rx_ep->av->ref_cnt, "address vector");

	++ bgq_sep->domain->rx.count;

	fi_bgq_ref_inc(&bgq_sep->ref_cnt, "scalable endpoint");

	return 0;

err:
	if (info.fabric_attr)
		free(info.fabric_attr);
	if (info.domain_attr)
		free(info.domain_attr);
	if (info.ep_attr)
		free(info.ep_attr);
	if (info.tx_attr)
		free(info.tx_attr);
	return -errno;
}