rxr_init.c

/*
 * Copyright (c) 2019 Amazon.com, Inc. or its affiliates.
 * All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include <rdma/fi_errno.h>

#include <ofi_prov.h>
#include "rxr.h"
#include "efa.h"

struct fi_info *shm_info;

struct fi_provider *lower_efa_prov;
struct efa_ep_addr *local_efa_addr;


struct rxr_env rxr_env = {
	.rx_window_size	= RXR_DEF_MAX_RX_WINDOW,
	.tx_max_credits = RXR_DEF_MAX_TX_CREDITS,
	.tx_min_credits = RXR_DEF_MIN_TX_CREDITS,
	.tx_queue_size = 0,
	.enable_sas_ordering = 1,
	.enable_atomic_ordering = 1,
	.enable_shm_transfer = 1,
	.shm_av_size = 128,
	.shm_max_medium_size = 4096,
	.recvwin_size = RXR_RECVWIN_SIZE,
	.cq_size = RXR_DEF_CQ_SIZE,
	.max_memcpy_size = 4096,
	.mtu_size = 0,
	.tx_size = 0,
	.rx_size = 0,
	.tx_iov_limit = 0,
	.rx_iov_limit = 0,
	.rx_copy_unexp = 1,
	.rx_copy_ooo = 1,
	.max_timeout = RXR_DEF_RNR_MAX_TIMEOUT,
	.timeout_interval = 0, /* 0 is random timeout */
	.efa_cq_read_size = 50,
	.shm_cq_read_size = 50,
	.efa_max_medium_msg_size = 65536,
	.efa_max_long_msg_size = 131072,
	.efa_max_emulated_read_size = 0,
	.efa_max_long_write_size = 65536,
	.efa_read_segment_size = 1073741824,
};

static void rxr_init_env(void)
{
	fi_param_get_int(&rxr_prov, "rx_window_size", &rxr_env.rx_window_size);
	fi_param_get_int(&rxr_prov, "tx_max_credits", &rxr_env.tx_max_credits);
	fi_param_get_int(&rxr_prov, "tx_min_credits", &rxr_env.tx_min_credits);
	fi_param_get_int(&rxr_prov, "tx_queue_size", &rxr_env.tx_queue_size);
	fi_param_get_int(&rxr_prov, "enable_sas_ordering", &rxr_env.enable_sas_ordering);
	fi_param_get_int(&rxr_prov, "enable_shm_transfer", &rxr_env.enable_shm_transfer);
	fi_param_get_int(&rxr_prov, "shm_av_size", &rxr_env.shm_av_size);
	fi_param_get_int(&rxr_prov, "shm_max_medium_size", &rxr_env.shm_max_medium_size);
	fi_param_get_int(&rxr_prov, "recvwin_size", &rxr_env.recvwin_size);
	fi_param_get_int(&rxr_prov, "cq_size", &rxr_env.cq_size);
	fi_param_get_size_t(&rxr_prov, "max_memcpy_size",
			    &rxr_env.max_memcpy_size);
	fi_param_get_bool(&rxr_prov, "mr_cache_enable",
			  &efa_mr_cache_enable);
	fi_param_get_size_t(&rxr_prov, "mr_max_cached_count",
			    &efa_mr_max_cached_count);
	fi_param_get_size_t(&rxr_prov, "mr_max_cached_size",
			    &efa_mr_max_cached_size);
	fi_param_get_size_t(&rxr_prov, "mtu_size",
			    &rxr_env.mtu_size);
	fi_param_get_size_t(&rxr_prov, "tx_size", &rxr_env.tx_size);
	fi_param_get_size_t(&rxr_prov, "rx_size", &rxr_env.rx_size);
	fi_param_get_size_t(&rxr_prov, "tx_iov_limit", &rxr_env.tx_iov_limit);
	fi_param_get_size_t(&rxr_prov, "rx_iov_limit", &rxr_env.rx_iov_limit);
	fi_param_get_bool(&rxr_prov, "rx_copy_unexp",
			  &rxr_env.rx_copy_unexp);
	fi_param_get_bool(&rxr_prov, "rx_copy_ooo",
			  &rxr_env.rx_copy_ooo);
	fi_param_get_int(&rxr_prov, "max_timeout", &rxr_env.max_timeout);
	fi_param_get_int(&rxr_prov, "timeout_interval",
			 &rxr_env.timeout_interval);
	fi_param_get_size_t(&rxr_prov, "efa_cq_read_size",
			 &rxr_env.efa_cq_read_size);
	fi_param_get_size_t(&rxr_prov, "shm_cq_read_size",
			 &rxr_env.shm_cq_read_size);
	fi_param_get_size_t(&rxr_prov, "inter_max_medium_message_size",
			    &rxr_env.efa_max_medium_msg_size);
	fi_param_get_size_t(&rxr_prov, "inter_max_long_message_size",
			    &rxr_env.efa_max_long_msg_size);
	fi_param_get_size_t(&rxr_prov, "inter_max_emulated_read_size",
			    &rxr_env.efa_max_emulated_read_size);
	fi_param_get_size_t(&rxr_prov, "inter_max_long_write_size",
			    &rxr_env.efa_max_long_write_size);
	fi_param_get_size_t(&rxr_prov, "read_segment_size",
			    &rxr_env.efa_read_segment_size);
}

/*
 * Stringify the void *addr to a string smr_name formatted as `gid_qpn`, which
 * will be used to insert into shm provider's AV. Then shm uses smr_name as
 * ep_name to create the shared memory region.
 *
 * The IPv6 address length is 46, but the max supported name length for shm is 32.
 * The string `gid_qpn` could be truncated during snprintf.
 * The current way works because the IPv6 addresses starting with FE in hexadecimals represent
 * link local IPv6 addresses, which has reserved first 64 bits (FE80::/64).
 * e.g., fe80:0000:0000:0000:0436:29ff:fe8e:ceaa -> fe80::436:29ff:fe8e:ceaa
 * And the length of string `gid_qpn` (fe80::436:29ff:fe8e:ceaa_***) will not exceed 32.
 * If the address is NOT link local, we need to think another reasonable way to
 * generate the string.
 */
int rxr_ep_efa_addr_to_str(const void *addr, char *smr_name)
{
	char gid[INET6_ADDRSTRLEN] = { 0 };
	uint16_t qpn;
	int ret;

	if (!inet_ntop(AF_INET6, ((struct efa_ep_addr *)addr)->raw, gid, INET6_ADDRSTRLEN)) {
		printf("Failed to get current EFA's GID, errno: %d\n", errno);
		return 0;
	}
	qpn = ((struct efa_ep_addr *)addr)->qpn;

	ret = snprintf(smr_name, NAME_MAX, "%ld_%s_%d", (size_t) getuid(), gid, qpn);

	return (ret <= 0) ? ret : FI_SUCCESS;
}

void rxr_info_to_core_mr_modes(uint32_t version,
			       const struct fi_info *hints,
			       struct fi_info *core_info)
{
	if (hints && hints->domain_attr &&
	    (hints->domain_attr->mr_mode & (FI_MR_SCALABLE | FI_MR_BASIC))) {
		core_info->mode = FI_LOCAL_MR | FI_MR_ALLOCATED;
		core_info->domain_attr->mr_mode = hints->domain_attr->mr_mode;
	} else if (FI_VERSION_LT(version, FI_VERSION(1, 5))) {
		core_info->mode |= FI_LOCAL_MR | FI_MR_ALLOCATED;
		core_info->domain_attr->mr_mode = FI_MR_UNSPEC;
	} else {
		core_info->domain_attr->mr_mode |=
			FI_MR_LOCAL | FI_MR_ALLOCATED;
		if (!hints)
			core_info->domain_attr->mr_mode |= OFI_MR_BASIC_MAP;
		else {
			if (hints->domain_attr)
				core_info->domain_attr->mr_mode |=
					hints->domain_attr->mr_mode & OFI_MR_BASIC_MAP;
			core_info->addr_format = hints->addr_format;
		}
#ifdef HAVE_LIBCUDA
		core_info->domain_attr->mr_mode |= FI_MR_HMEM;
#endif
	}
}

static int rxr_copy_attr(const struct fi_info *info, struct fi_info *dup)
{
	if (info->src_addr) {
		dup->src_addrlen = info->src_addrlen;
		dup->src_addr = mem_dup(info->src_addr,
					info->src_addrlen);
		if (!dup->src_addr)
			return -FI_ENOMEM;
	}
	if (info->dest_addr) {
		dup->dest_addrlen = info->dest_addrlen;
		dup->dest_addr = mem_dup(info->dest_addr,
					 info->dest_addrlen);
		if (!dup->dest_addr)
			return -FI_ENOMEM;
	}
	if (info->fabric_attr) {
		if (info->fabric_attr->name) {
			dup->fabric_attr->name =
				strdup(info->fabric_attr->name);
			if (!dup->fabric_attr->name)
				return -FI_ENOMEM;
		}
	}
	if (info->domain_attr) {
		if (info->domain_attr->name) {
			dup->domain_attr->name =
				strdup(info->domain_attr->name);
			if (!dup->domain_attr->name)
				return -FI_ENOMEM;
		}
	}
	if (info->nic) {
		dup->nic = ofi_nic_dup(info->nic);
		if (!dup->nic)
			return -FI_ENOMEM;
	}
	return 0;
}

static int rxr_info_to_core(uint32_t version, const struct fi_info *rxr_info,
			    struct fi_info **core_info)
{
	int ret = 0;
	*core_info = fi_allocinfo();
	if (!*core_info)
		return -FI_ENOMEM;

	rxr_info_to_core_mr_modes(version, rxr_info, *core_info);
	(*core_info)->caps = FI_MSG;
	(*core_info)->ep_attr->type = FI_EP_RDM;
	(*core_info)->tx_attr->op_flags = FI_TRANSMIT_COMPLETE;

	/*
	 * Skip copying address, domain, fabric info.
	 */
	if (!rxr_info)
		return 0;

	ret = rxr_copy_attr(rxr_info, *core_info);
	if (ret)
		fi_freeinfo(*core_info);
	return ret;
}

/* Explicitly set all necessary bits before calling shm provider's getinfo function */
void rxr_set_shm_hints(struct fi_info *shm_hints)
{
	shm_hints->caps = FI_MSG | FI_TAGGED | FI_RECV | FI_SEND | FI_READ
			   | FI_WRITE | FI_REMOTE_READ | FI_REMOTE_WRITE
			   | FI_MULTI_RECV | FI_RMA;
	shm_hints->domain_attr->av_type = FI_AV_TABLE;
	shm_hints->domain_attr->mr_mode = FI_MR_VIRT_ADDR;
	shm_hints->domain_attr->caps |= FI_LOCAL_COMM;
	shm_hints->tx_attr->msg_order = FI_ORDER_SAS;
	shm_hints->rx_attr->msg_order = FI_ORDER_SAS;
	shm_hints->fabric_attr->name = strdup("shm");
	shm_hints->fabric_attr->prov_name = strdup("shm");
	shm_hints->ep_attr->type = FI_EP_RDM;
}

/* Pass tx/rx attr that user specifies down to core provider */
void rxr_reset_rx_tx_to_core(const struct fi_info *user_info,
			     struct fi_info *core_info)
{
	/* rx attr */
	core_info->rx_attr->total_buffered_recv =
		user_info->rx_attr->total_buffered_recv < core_info->rx_attr->total_buffered_recv ?
		user_info->rx_attr->total_buffered_recv : core_info->rx_attr->total_buffered_recv;
	core_info->rx_attr->size =
		user_info->rx_attr->size < core_info->rx_attr->size ?
		user_info->rx_attr->size : core_info->rx_attr->size;
	core_info->rx_attr->iov_limit =
		user_info->rx_attr->iov_limit < core_info->rx_attr->iov_limit ?
		user_info->rx_attr->iov_limit : core_info->rx_attr->iov_limit;
	/* tx attr */
	core_info->tx_attr->inject_size =
		user_info->tx_attr->inject_size < core_info->tx_attr->inject_size ?
		user_info->tx_attr->inject_size : core_info->tx_attr->inject_size;
	core_info->tx_attr->size =
		user_info->tx_attr->size < core_info->tx_attr->size ?
		user_info->tx_attr->size : core_info->tx_attr->size;
	core_info->tx_attr->iov_limit =
		user_info->tx_attr->iov_limit < core_info->tx_attr->iov_limit ?
		user_info->tx_attr->iov_limit : core_info->tx_attr->iov_limit;
}

void rxr_set_rx_tx_size(struct fi_info *info,
			const struct fi_info *core_info)
{
	if (rxr_env.tx_size > 0)
		info->tx_attr->size = rxr_env.tx_size;
	else
		info->tx_attr->size = core_info->tx_attr->size;

	if (rxr_env.rx_size > 0)
		info->rx_attr->size = rxr_env.rx_size;
	else
		info->rx_attr->size = core_info->rx_attr->size;

	if (rxr_env.tx_iov_limit > 0)
		info->tx_attr->iov_limit = rxr_env.tx_iov_limit;

	if (rxr_env.rx_iov_limit > 0)
		info->rx_attr->iov_limit = rxr_env.rx_iov_limit;
}

static int rxr_dgram_info_to_rxr(uint32_t version,
				 const struct fi_info *core_info,
				 struct fi_info *info) {
	rxr_set_rx_tx_size(info, core_info);
	return 0;
}

static int rxr_info_to_rxr(uint32_t version, const struct fi_info *core_info,
			   struct fi_info *info, const struct fi_info *hints)
{
	uint64_t atomic_ordering;

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

	*info->tx_attr = *rxr_info.tx_attr;
	*info->rx_attr = *rxr_info.rx_attr;
	*info->ep_attr = *rxr_info.ep_attr;
	*info->domain_attr = *rxr_info.domain_attr;

	/* TODO: update inject_size when we implement inject */
	info->tx_attr->inject_size = 0;
	rxr_info.tx_attr->inject_size = info->tx_attr->inject_size;

	info->addr_format = core_info->addr_format;
	info->domain_attr->ep_cnt = core_info->domain_attr->ep_cnt;
	info->domain_attr->cq_cnt = core_info->domain_attr->cq_cnt;
	info->domain_attr->mr_key_size = core_info->domain_attr->mr_key_size;

	/*
	 * Handle user-provided hints and adapt the info object passed back up
	 * based on EFA-specific constraints.
	 */
	if (hints) {
		if (hints->tx_attr) {

			/* Disable packet reordering if the app doesn't need it */
			if (!(hints->tx_attr->msg_order & FI_ORDER_SAS))
				rxr_env.enable_sas_ordering = 0;

			/* Disable atomic ordering if the app doesn't need it */

			atomic_ordering = FI_ORDER_ATOMIC_RAR | FI_ORDER_ATOMIC_RAW |
					  FI_ORDER_ATOMIC_WAR | FI_ORDER_ATOMIC_WAW;

			if (!(hints->tx_attr->msg_order & atomic_ordering))
				rxr_env.enable_atomic_ordering = 0;
		}

		/* We only support manual progress for RMA operations */
		if (hints->caps & FI_RMA) {
			info->domain_attr->control_progress = FI_PROGRESS_MANUAL;
			info->domain_attr->data_progress = FI_PROGRESS_MANUAL;
		}

		/* Use a table for AV if the app has no strong requirement */
		if (!hints->domain_attr || hints->domain_attr->av_type == FI_AV_UNSPEC)
			info->domain_attr->av_type = FI_AV_TABLE;

#ifdef HAVE_LIBCUDA
		/* If the application requires HMEM support, we will add FI_MR_HMEM
		 * to mr_mode, because we need application to provide descriptor
		 * for cuda buffer.
		 * Note we did not add FI_MR_LOCAL here because according
		 * to FI_MR man page:
		 *
		 *     "If FI_MR_HMEM is set, but FI_MR_LOCAL is unset,
		 *      only device buffers must be registered when used locally.
		 *      "
		 * which means FI_MR_HMEM implies FI_MR_LOCAL for cuda buffer
		 */
		if (hints->caps & FI_HMEM) {
			if (hints->domain_attr &&
			    !(hints->domain_attr->mr_mode & FI_MR_HMEM)) {
				FI_INFO(&rxr_prov, FI_LOG_CORE,
				        "FI_HMEM capability requires device registrations (FI_MR_HMEM)\n");
				return -FI_ENODATA;
			}

			info->domain_attr->mr_mode |= FI_MR_HMEM;

			/*
			 * If in this case application add FI_MR_LOCAL to hints,
			 * it would mean that application want provide descriptor
			 * for system memory too, which we are able to use, so
			 * we add FI_MR_LOCAL to mr_mode.
			 *
			 * TODO: add FI_MR_LOCAL to mr_mode for any applcations
			 * the requested it, not just CUDA application.
			 */
			if (hints->domain_attr->mr_mode & FI_MR_LOCAL)
				info->domain_attr->mr_mode |= FI_MR_LOCAL;
		}
#endif
	}

	rxr_set_rx_tx_size(info, core_info);
	return 0;
}

/*
 * For the RDM endpoint, translate user hints to hints for the lower layer and
 * call getinfo on the lower layer.
 */
int rxr_get_lower_rdm_info(uint32_t version, const char *node,
			   const char *service, uint64_t flags,
			   const struct util_prov *util_prov,
			   const struct fi_info *util_hints,
			   struct fi_info **core_info)
{
	struct fi_info *core_hints = NULL;
	int ret;

	ret = ofi_prov_check_info(&rxr_util_prov, version, util_hints);
	if (ret)
		return ret;

	ret = rxr_info_to_core(version, util_hints, &core_hints);
	if (ret)
		return ret;

	ret = lower_efa_prov->getinfo(version, node, service, flags,
				      core_hints, core_info);
	fi_freeinfo(core_hints);
	return ret;
}

/*
 * Call getinfo on lower efa provider to get all locally qualified fi_info
 * structure, then store the corresponding efa nic GIDs
 */
int rxr_get_local_gids(struct fi_provider *lower_efa_prov)
{
	struct fi_info *core_info, *cur;
	struct efa_ep_addr *cur_efa_addr;
	int ret;

	cur_efa_addr = local_efa_addr = NULL;
	core_info = cur = NULL;

	ret = lower_efa_prov->getinfo(rxr_prov.fi_version, NULL, NULL, 0, NULL, &core_info);
	if (ret)
		return ret;

	local_efa_addr = (struct efa_ep_addr *)malloc(sizeof(struct efa_ep_addr));
	if (!local_efa_addr) {
		ret = -FI_ENOMEM;
		goto out;
	}
	local_efa_addr->next = NULL;

	cur_efa_addr = local_efa_addr;
	for (cur = core_info; cur; cur = cur->next) {
		memcpy(cur_efa_addr->raw, ((struct efa_ep_addr *)cur->src_addr)->raw, 16);
		if (cur->next) {
			cur_efa_addr->next = (struct efa_ep_addr *)malloc(sizeof(struct efa_ep_addr));
			if (!cur_efa_addr->next) {
				ret = -FI_ENOMEM;
				goto out;
			}
			cur_efa_addr = cur_efa_addr->next;
			cur_efa_addr->next = NULL;
		}
	}

out:
	fi_freeinfo(core_info);
	return ret;
}

static int rxr_dgram_getinfo(uint32_t version, const char *node,
			     const char *service, uint64_t flags,
			     const struct fi_info *hints, struct fi_info **info,
			     struct fi_info **tail)
{
	struct fi_info *core_info, *util_info, *cur;
	int ret;

	core_info = NULL;

	ret = lower_efa_prov->getinfo(version, node, service,
				      flags, hints, &core_info);

	if (ret)
		return ret;

	ret = -FI_ENODATA;

	for (cur = core_info; cur; cur = cur->next) {
		/* Skip non DGRAM info structs */
		if (cur->ep_attr->type != FI_EP_DGRAM)
			continue;

		ret = 0;

		util_info = fi_dupinfo(cur);
		if (!util_info) {
			ret = -FI_ENOMEM;
			fi_freeinfo(*info);
			goto out;
		}

		rxr_dgram_info_to_rxr(version, cur, util_info);

		if (!*info)
			*info = util_info;
		else
			(*tail)->next = util_info;
		*tail = util_info;
	}

out:
	fi_freeinfo(core_info);
	return ret;
}

static int rxr_getinfo(uint32_t version, const char *node,
		       const char *service, uint64_t flags,
		       const struct fi_info *hints, struct fi_info **info)
{
	struct fi_info *core_info, *util_info, *cur, *tail;
	struct fi_info *shm_hints;
	int ret;

	*info = tail = core_info = NULL;

	if (hints && hints->ep_attr && hints->ep_attr->type == FI_EP_DGRAM)
		goto dgram_info;

	/*
	 * Using the shm provider comes with some overheads, particularly in the
	 * progress engine when polling an empty completion queue, so avoid
	 * initializing the provider if the app provides a hint that it does not
	 * require node-local communication. We can still loopback over the EFA
	 * device in cases where the app violates the hint and continues
	 * communicating with node-local peers.
	 */
	if (hints
	    /* If the app requires explicitly remote communication */
	    && (hints->caps & FI_REMOTE_COMM)
	    /* but not local communication */
	    && !(hints->caps & FI_LOCAL_COMM))
		rxr_env.enable_shm_transfer = 0;

	ret = rxr_get_lower_rdm_info(version, node, service, flags,
				     &rxr_util_prov, hints, &core_info);

	if (ret == -FI_ENODATA)
		goto dgram_info;

	if (ret)
		return ret;

	for (cur = core_info; cur; cur = cur->next) {
		util_info = fi_allocinfo();
		if (!util_info) {
			ret = -FI_ENOMEM;
			goto free_info;
		}

		ret = rxr_info_to_rxr(version, cur, util_info, hints);
		if (ret)
			goto free_info;

		ret = rxr_copy_attr(cur, util_info);
		if (ret)
			goto free_info;

		ofi_alter_info(util_info, hints, version);
		if (!*info)
			*info = util_info;
		else
			tail->next = util_info;
		tail = util_info;
	}

dgram_info:
	ret = rxr_dgram_getinfo(version, node, service, flags, hints, info,
				&tail);
	/*
	 * Ignore dgram getinfo return code if rdm getinfo was successful.
	 */
	if (ret == -FI_ENODATA && *info)
		ret = 0;

	if (!ret && rxr_env.enable_shm_transfer && !shm_info) {
		shm_info = fi_allocinfo();
		shm_hints = fi_allocinfo();
		rxr_set_shm_hints(shm_hints);
		ret = fi_getinfo(FI_VERSION(1, 8), NULL, NULL,
		                 OFI_GETINFO_HIDDEN, shm_hints, &shm_info);
		fi_freeinfo(shm_hints);
		if (ret) {
			FI_WARN(&rxr_prov, FI_LOG_CORE, "Disabling EFA shared memory support; failed to get shm provider's info: %s\n",
				fi_strerror(-ret));
			rxr_env.enable_shm_transfer = 0;
			ret = 0;
		} else {
			assert(!strcmp(shm_info->fabric_attr->name, "shm"));
		}
	}

	fi_freeinfo(core_info);
	return ret;
free_info:
	fi_freeinfo(core_info);
	fi_freeinfo(util_info);
	fi_freeinfo(*info);
	*info = NULL;
	return ret;
}

static void rxr_fini(void)
{
	struct efa_ep_addr *cur;

	if (lower_efa_prov)
		lower_efa_prov->cleanup();

	if (rxr_env.enable_shm_transfer) {
		/* Cleanup all local efa nic GIDs */
		while (local_efa_addr) {
			cur = local_efa_addr;
			local_efa_addr = local_efa_addr->next;
			free(cur);
		}
		if (shm_info)
			fi_freeinfo(shm_info);
	}

#if HAVE_EFA_DL
	ofi_monitor_cleanup();
	ofi_mem_fini();
#endif
}

struct fi_provider rxr_prov = {
	.name = "efa",
	.version = FI_VERSION(RXR_MAJOR_VERSION, RXR_MINOR_VERSION),
	.fi_version = RXR_FI_VERSION,
	.getinfo = rxr_getinfo,
	.fabric = rxr_fabric,
	.cleanup = rxr_fini
};

EFA_INI
{
	fi_param_define(&rxr_prov, "rx_window_size", FI_PARAM_INT,
			"Defines the maximum window size that a receiver will return for matched large messages. (Default: 128).");
	fi_param_define(&rxr_prov, "tx_max_credits", FI_PARAM_INT,
			"Defines the maximum number of credits a sender requests from a receiver (Default: 64).");
	fi_param_define(&rxr_prov, "tx_min_credits", FI_PARAM_INT,
			"Defines the minimum number of credits a sender requests from a receiver (Default: 32).");
	fi_param_define(&rxr_prov, "tx_queue_size", FI_PARAM_INT,
			"Defines the maximum number of unacknowledged sends with the NIC.");
	fi_param_define(&rxr_prov, "enable_sas_ordering", FI_PARAM_INT,
			"Enable packet reordering for the RDM endpoint. This is always enabled when FI_ORDER_SAS is requested by the application. (Default: 1)");
	fi_param_define(&rxr_prov, "enable_atomic_ordering", FI_PARAM_INT,
			"Enable atomic reordering for the RDM endpoint. This is always enabled when FI_ORDER_ATOMIC_RAW or FI_ORDER_ATOMIC_RAR or FI_ORDER_ATOMIC_WAR or FI_ORDER_ATOMIC_WAW is requested by the application. (Default: 1)");
	fi_param_define(&rxr_prov, "enable_shm_transfer", FI_PARAM_INT,
			"Enable using SHM provider to provide the communication between processes on the same system. (Default: 1)");
	fi_param_define(&rxr_prov, "shm_av_size", FI_PARAM_INT,
			"Defines the maximum number of entries in SHM provider's address vector (Default 128).");
	fi_param_define(&rxr_prov, "shm_max_medium_size", FI_PARAM_INT,
			"Defines the switch point between small/medium message and large message. The message larger than this switch point will be transferred with large message protocol (Default 4096).");
	fi_param_define(&rxr_prov, "recvwin_size", FI_PARAM_INT,
			"Defines the size of sliding receive window. (Default: 16384)");
	fi_param_define(&rxr_prov, "cq_size", FI_PARAM_INT,
			"Define the size of completion queue. (Default: 8192)");
	fi_param_define(&rxr_prov, "mr_cache_enable", FI_PARAM_BOOL,
			"Enables using the mr cache and in-line registration instead of a bounce buffer for iov's larger than max_memcpy_size. Defaults to true. When disabled, only uses a bounce buffer.");
	fi_param_define(&rxr_prov, "mr_max_cached_count", FI_PARAM_SIZE_T,
			"Sets the maximum number of memory registrations that can be cached at any time.");
	fi_param_define(&rxr_prov, "mr_max_cached_size", FI_PARAM_SIZE_T,
			"Sets the maximum amount of memory that cached memory registrations can hold onto at any time.");
	fi_param_define(&rxr_prov, "max_memcpy_size", FI_PARAM_SIZE_T,
			"Threshold size switch between using memory copy into a pre-registered bounce buffer and memory registration on the user buffer. (Default: 4096)");
	fi_param_define(&rxr_prov, "mtu_size", FI_PARAM_SIZE_T,
			"Override the MTU size of the device.");
	fi_param_define(&rxr_prov, "tx_size", FI_PARAM_SIZE_T,
			"Set the maximum number of transmit operations before the provider returns -FI_EAGAIN. For only the RDM endpoint, this parameter will cause transmit operations to be queued when this value is set higher than the default and the transmit queue is full.");
	fi_param_define(&rxr_prov, "rx_size", FI_PARAM_SIZE_T,
			"Set the maximum number of receive operations before the provider returns -FI_EAGAIN.");
	fi_param_define(&rxr_prov, "tx_iov_limit", FI_PARAM_SIZE_T,
			"Maximum transmit iov_limit.");
	fi_param_define(&rxr_prov, "rx_iov_limit", FI_PARAM_SIZE_T,
			"Maximum receive iov_limit.");
	fi_param_define(&rxr_prov, "rx_copy_unexp", FI_PARAM_BOOL,
			"Enables the use of a separate pool of bounce-buffers to copy unexpected messages out of the pre-posted receive buffers. (Default: 1)");
	fi_param_define(&rxr_prov, "rx_copy_ooo", FI_PARAM_BOOL,
			"Enables the use of a separate pool of bounce-buffers to copy out-of-order RTM packets out of the pre-posted receive buffers. (Default: 1)");
	fi_param_define(&rxr_prov, "max_timeout", FI_PARAM_INT,
			"Set the maximum timeout (us) for backoff to a peer after a receiver not ready error. (Default: 1000000)");
	fi_param_define(&rxr_prov, "timeout_interval", FI_PARAM_INT,
			"Set the time interval (us) for the base timeout to use for exponential backoff to a peer after a receiver not ready error. (Default: 0 [random])");
	fi_param_define(&rxr_prov, "efa_cq_read_size", FI_PARAM_SIZE_T,
			"Set the number of EFA completion entries to read for one loop for one iteration of the progress engine. (Default: 50)");
	fi_param_define(&rxr_prov, "shm_cq_read_size", FI_PARAM_SIZE_T,
			"Set the number of SHM completion entries to read for one loop for one iteration of the progress engine. (Default: 50)");
	fi_param_define(&rxr_prov, "inter_max_medium_message_size", FI_PARAM_INT,
			"The maximum message size for inter EFA medium message protocol, messages whose size is larger than this value will be sent either by read message protocol (depend on firmware support), or long message protocol (Default 65536).");
	fi_param_define(&rxr_prov, "inter_max_long_message_size", FI_PARAM_INT,
			"The maximum message size for inter EFA long message protocol, messages whose size is larger than this value will be sent either by read message protocol (if firmware support) (Default 131072).");
	fi_param_define(&rxr_prov, "inter_max_emulated_read_size", FI_PARAM_INT,
			"The maximum message size for inter EFA emulated read protocol. Read requests whose size is larger than this value will be implemented via RDMA read (if firmware support), (Default 0 [RDMA read is always used]).");
	fi_param_define(&rxr_prov, "inter_max_long_write_size", FI_PARAM_INT,
			"The maximum message size for inter EFA emulated write protocol. Write requests whose size is larger than this value will be implemented via read write protocol (write by read).");
	fi_param_define(&rxr_prov, "efa_read_segment_size", FI_PARAM_INT,
			"Calls to RDMA read is segmented using this value.");
	rxr_init_env();

#if HAVE_EFA_DL
	ofi_mem_init();
	ofi_monitor_init();
#endif

	lower_efa_prov = init_lower_efa_prov();
	if (!lower_efa_prov)
		return NULL;

	if (rxr_env.enable_shm_transfer && rxr_get_local_gids(lower_efa_prov))
		return NULL;

	return &rxr_prov;
}