[API-NEXT,PATCHv3,06/13] linux-gen: fdserver: new fdserver added

Message ID 1471679163-17240-7-git-send-email-christophe.milard@linaro.org
State New
Headers show

Commit Message

Christophe Milard Aug. 20, 2016, 7:45 a.m.
A fdserver is added and started at init time.
The role of the fdserver (file descriptor server) is to enable sharing of
file descriptors between unrelated processes: processes which wish to
share a file descriptor may register their files descriptor to the server
and processes wishing to use the shared file descriptors can do a lookup.
When registration occurs, a triple {context, key, fd} is provided to
the server. The context identifies the client and scope (i.e. shmem).
The key is implemented as a long, and can be whatever.
The server won't care as long as keys are unique.
The file descriptor can be retrieved by another process providing the same
context and key.
Files descriptors passed this way are converted on the fly during the unix
domain socket communication that occurs between the server and its clients.
This is done by using the ancillary control data part of the msg.

Signed-off-by: Christophe Milard <christophe.milard@linaro.org>

---
 platform/linux-generic/Makefile.am                 |   2 +
 platform/linux-generic/_fdserver.c                 | 655 +++++++++++++++++++++
 .../linux-generic/include/_fdserver_internal.h     |  38 ++
 platform/linux-generic/include/odp_internal.h      |   4 +
 platform/linux-generic/odp_init.c                  |  13 +
 5 files changed, 712 insertions(+)
 create mode 100644 platform/linux-generic/_fdserver.c
 create mode 100644 platform/linux-generic/include/_fdserver_internal.h

-- 
2.7.4

Comments

Brian Brooks Aug. 29, 2016, 3:02 a.m. | #1
On 08/20 09:45:56, Christophe Milard wrote:
> A fdserver is added and started at init time.

> The role of the fdserver (file descriptor server) is to enable sharing of

> file descriptors between unrelated processes: processes which wish to

> share a file descriptor may register their files descriptor to the server

> and processes wishing to use the shared file descriptors can do a lookup.

> When registration occurs, a triple {context, key, fd} is provided to

> the server. The context identifies the client and scope (i.e. shmem).

> The key is implemented as a long, and can be whatever.

> The server won't care as long as keys are unique.

> The file descriptor can be retrieved by another process providing the same

> context and key.

> Files descriptors passed this way are converted on the fly during the unix

> domain socket communication that occurs between the server and its clients.

> This is done by using the ancillary control data part of the msg.

> 

> Signed-off-by: Christophe Milard <christophe.milard@linaro.org>

> ---

>  platform/linux-generic/Makefile.am                 |   2 +

>  platform/linux-generic/_fdserver.c                 | 655 +++++++++++++++++++++

>  .../linux-generic/include/_fdserver_internal.h     |  38 ++

>  platform/linux-generic/include/odp_internal.h      |   4 +

>  platform/linux-generic/odp_init.c                  |  13 +

>  5 files changed, 712 insertions(+)

>  create mode 100644 platform/linux-generic/_fdserver.c

>  create mode 100644 platform/linux-generic/include/_fdserver_internal.h

> 

> diff --git a/platform/linux-generic/Makefile.am b/platform/linux-generic/Makefile.am

> index 271f5b5..cb7c8e9 100644

> --- a/platform/linux-generic/Makefile.am

> +++ b/platform/linux-generic/Makefile.am

> @@ -112,6 +112,7 @@ odpdrvplatinclude_HEADERS = \

>  		  $(srcdir)/include/odp/drv/plat/strong_types.h

>  

>  noinst_HEADERS = \

> +		  ${srcdir}/include/_fdserver_internal.h \

>  		  ${srcdir}/include/odp_align_internal.h \

>  		  ${srcdir}/include/odp_atomic_internal.h \

>  		  ${srcdir}/include/odp_buffer_inlines.h \

> @@ -154,6 +155,7 @@ noinst_HEADERS = \

>  		  ${srcdir}/Makefile.inc

>  

>  __LIB__libodp_linux_la_SOURCES = \

> +			   _fdserver.c \

>  			   odp_atomic.c \

>  			   odp_barrier.c \

>  			   odp_buffer.c \

> diff --git a/platform/linux-generic/_fdserver.c b/platform/linux-generic/_fdserver.c

> new file mode 100644

> index 0000000..bf36eb2

> --- /dev/null

> +++ b/platform/linux-generic/_fdserver.c

> @@ -0,0 +1,655 @@

> +/* Copyright (c) 2016, Linaro Limited

> + * All rights reserved.

> + *

> + * SPDX-License-Identifier:     BSD-3-Clause

> + */

> +

> +/*

> + * This file implements a file descriptor sharing server enabling

> + * sharing of file descriptors between processes, regardless of fork time.

> + *

> + * File descriptors are process scoped, but they can be "sent and converted

> + * on the fly" between processes using special unix domain socket ancillary

> + * data.

> + * The receiving process gets a file descriptor "pointing" to the same thing

> + * as the one sent (but the value of the file descriptor itself may be different

> + * from the one sent).

> + * Because ODP applications are responsible for creating ODP threads (i.e.

> + * pthreads or linux processes), ODP has no control on the order things happen:

> + * Nothing prevent a thread A to fork B and C, and then C creating a pktio

> + * which will be used by A and B to send/receive packets.

> + * Assuming this pktio uses a file descriptor, the latter will need to be

> + * shared between the processes, despite the "non convenient" fork time.

> + * The shared memory allocator is likely to use this as well to be able to

> + * share memory regardless of fork() time.

> + * This server handles a table of {(context,key)<-> fd} pair, and is

> + * interfaced by the following functions:

> + *

> + * _odp_fdserver_register_fd(context, key, fd_to_send);

> + * _odp_fdserver_deregister_fd(context, key);

> + * _odp_fdserver_lookup_fd(context, key);

> + *

> + * which are used to register/deregister or querry for file descriptor based

> + * on a context and key value couple, which has to be unique.

> + *

> + * Note again that the file descriptors stored here are local to this server

> + * process and get converted both when registered or looked up.

> + */

> +

> +#include <odp_posix_extensions.h>

> +#include <odp/api/spinlock.h>

> +#include <odp_internal.h>

> +#include <odp_debug_internal.h>

> +#include <_fdserver_internal.h>

> +

> +#include <stdio.h>

> +#include <stdlib.h>

> +#include <errno.h>

> +#include <string.h>

> +#include <sys/types.h>

> +#include <signal.h>

> +#include <sys/socket.h>

> +#include <sys/un.h>

> +#include <unistd.h>

> +#include <inttypes.h>

> +#include <sys/mman.h>

> +#include <sys/wait.h>

> +

> +#define FDSERVER_SOCKPATH_MAXLEN 32

> +#define FDSERVER_SOCKPATH_FORMAT "/tmp/odp-%d-fdserver"

> +

> +#define FDSERVER_BACKLOG 5

> +

> +#ifndef MAP_ANONYMOUS

> +#define MAP_ANONYMOUS MAP_ANON

> +#endif

> +

> +/* when accessing the client functions, clients should be mutexed: */

> +odp_spinlock_t *client_lock;

> +

> +/* define the tables of file descriptors handled by this server: */

> +#define FDSERVER_MAX_ENTRIES 256

> +typedef struct fdentry_s {

> +	fd_server_context_e context;

> +	uint64_t key;

> +	int  fd;

> +} fdentry_t;

> +static fdentry_t fd_table[FDSERVER_MAX_ENTRIES];

> +static int fd_table_nb_entries;


This data will be COW on fork, which means it needs to be owned by some other
process instead of this shared library. Is this in your mind?

> +/*

> + * define the message struct used for communication between client and server

> + * (this single message is used in both direction)

> + * The file descriptors are sent out of band as ancillary data for conversion.

> + */

> +typedef struct fd_server_msg {

> +	int command;

> +	fd_server_context_e context;

> +	uint64_t key;

> +} fdserver_msg_t;

> +/* possible commands are: */

> +#define FD_REGISTER_REQ		1  /* client -> server */

> +#define FD_REGISTER_ACK		2  /* server -> client */

> +#define FD_REGISTER_NACK	3  /* server -> client */

> +#define FD_LOOKUP_REQ		4  /* client -> server */

> +#define FD_LOOKUP_ACK		5  /* server -> client */

> +#define FD_LOOKUP_NACK		6  /* server -> client */

> +#define FD_DEREGISTER_REQ	7  /* client -> server */

> +#define FD_DEREGISTER_ACK	8  /* server -> client */

> +#define FD_DEREGISTER_NACK	9  /* server -> client */

> +#define FD_SERVERSTOP_REQ	10 /* client -> server (stops) */

> +

> +/*

> + * Client and server function:

> + * Send a fdserver_msg, possibly including a file descriptor, on the socket

> + * This function is used both by:

> + * -the client (sending a FD_REGISTER_REQ with a file descriptor to be shared,

> + *  or FD_LOOKUP_REQ/FD_DEREGISTER_REQ without a file descirptor)

> + * -the server (sending FD_REGISTER_ACK/NACK, FD_LOOKUP_NACK,

> + *  FD_DEREGISTER_ACK/NACK... without a fd or a

> + *  FD_LOOKUP_ACK with a fd)

> + * This function make use of the ancillary data (control data) to pass and

> + * convert file descriptors over UNIX sockets

> + * Return -1 on error, 0 on success.

> + */

> +static int send_fdserver_msg(int sock, int command,

> +			     fd_server_context_e context, uint64_t key,

> +			     int fd_to_send)

> +{

> +	struct msghdr socket_message;

> +	struct iovec io_vector[1]; /* one msg frgmt only */

> +	struct cmsghdr *control_message = NULL;

> +	int *fd_location;

> +	fdserver_msg_t msg;

> +	int res;

> +

> +	char ancillary_data[CMSG_SPACE(sizeof(int))];

> +

> +	/* prepare the register request body (single framgent): */

> +	msg.command = command;

> +	msg.context = context;

> +	msg.key = key;

> +	io_vector[0].iov_base = &msg;

> +	io_vector[0].iov_len = sizeof(fdserver_msg_t);

> +

> +	/* initialize socket message */

> +	memset(&socket_message, 0, sizeof(struct msghdr));

> +	socket_message.msg_iov = io_vector;

> +	socket_message.msg_iovlen = 1;

> +

> +	if (fd_to_send >= 0) {

> +		/* provide space for the ancillary data */

> +		memset(ancillary_data, 0, CMSG_SPACE(sizeof(int)));

> +		socket_message.msg_control = ancillary_data;

> +		socket_message.msg_controllen = CMSG_SPACE(sizeof(int));

> +

> +		/* initialize a single ancillary data element for fd passing */

> +		control_message = CMSG_FIRSTHDR(&socket_message);

> +		control_message->cmsg_level = SOL_SOCKET;

> +		control_message->cmsg_type = SCM_RIGHTS;

> +		control_message->cmsg_len = CMSG_LEN(sizeof(int));

> +		fd_location = (int *)(void *)CMSG_DATA(control_message);

> +		*fd_location = fd_to_send;

> +	}

> +	res = sendmsg(sock, &socket_message, 0);

> +	if (res < 0) {

> +		ODP_ERR("send_fdserver_msg: %s\n", strerror(errno));

> +		return(-1);

> +	}

> +

> +	return 0;

> +}

> +

> +/*

> + * Client and server function

> + * Receive a fdserver_msg, possibly including a file descriptor, on the

> + * given socket.

> + * This function is used both by:

> + * -the server (receiving a FD_REGISTER_REQ with a file descriptor to be shared,

> + *  or FD_LOOKUP_REQ, FD_DEREGISTER_REQ without a file descirptor)

> + * -the client (receiving FD_REGISTER_ACK...without a fd or a FD_LOOKUP_ACK with

> + * a fd)

> + * This function make use of the ancillary data (control data) to pass and

> + * convert file descriptors over UNIX sockets.

> + * Return -1 on error, 0 on success.

> + */

> +static int recv_fdserver_msg(int sock, int *command,

> +			     fd_server_context_e *context, uint64_t *key,

> +			     int *recvd_fd)

> +{

> +	struct msghdr socket_message;

> +	struct iovec io_vector[1]; /* one msg frgmt only */

> +	struct cmsghdr *control_message = NULL;

> +	int *fd_location;

> +	fdserver_msg_t msg;

> +	char ancillary_data[CMSG_SPACE(sizeof(int))];

> +

> +	memset(&socket_message, 0, sizeof(struct msghdr));

> +	memset(ancillary_data, 0, CMSG_SPACE(sizeof(int)));

> +

> +	/* setup a place to fill in message contents */

> +	io_vector[0].iov_base = &msg;

> +	io_vector[0].iov_len = sizeof(fdserver_msg_t);

> +	socket_message.msg_iov = io_vector;

> +	socket_message.msg_iovlen = 1;

> +

> +	/* provide space for the ancillary data */

> +	socket_message.msg_control = ancillary_data;

> +	socket_message.msg_controllen = CMSG_SPACE(sizeof(int));

> +

> +	/* receive the message */

> +	if (recvmsg(sock, &socket_message, MSG_CMSG_CLOEXEC) < 0) {

> +		ODP_ERR("recv_fdserver_msg: %s\n", strerror(errno));

> +		return(-1);

> +	}

> +

> +	*command = msg.command;

> +	*context = msg.context;

> +	*key = msg.key;

> +

> +	/* grab the converted file descriptor (if any) */

> +	*recvd_fd = -1;

> +

> +	if ((socket_message.msg_flags & MSG_CTRUNC) == MSG_CTRUNC)

> +		return 0;

> +

> +	/* iterate ancillary elements to find the file descriptor: */

> +	for (control_message = CMSG_FIRSTHDR(&socket_message);

> +	     control_message != NULL;

> +	     control_message = CMSG_NXTHDR(&socket_message, control_message)) {

> +		if ((control_message->cmsg_level == SOL_SOCKET) &&

> +		    (control_message->cmsg_type == SCM_RIGHTS)) {

> +			fd_location = (int *)(void *)CMSG_DATA(control_message);

> +			*recvd_fd = *fd_location;

> +			break;

> +		}

> +	}

> +

> +	return 0;

> +}

> +

> +/* opens and returns a connected socket to the server */

> +static int get_socket(void)

> +{

> +	char sockpath[FDSERVER_SOCKPATH_MAXLEN];

> +	int s_sock; /* server socket */

> +	struct sockaddr_un remote;

> +	int len;

> +

> +	/* construct the named socket path: */

> +	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,

> +		 odp_global_data.main_pid);

> +

> +	s_sock = socket(AF_UNIX, SOCK_STREAM, 0);

> +	if (s_sock == -1) {

> +		ODP_ERR("cannot connect to server: %s\n", strerror(errno));

> +		return(-1);

> +	}

> +

> +	remote.sun_family = AF_UNIX;

> +	strcpy(remote.sun_path, sockpath);

> +	len = strlen(remote.sun_path) + sizeof(remote.sun_family);

> +	if (connect(s_sock, (struct sockaddr *)&remote, len) == -1) {

> +		ODP_ERR("cannot connect to server: %s\n", strerror(errno));

> +		close(s_sock);

> +		return(-1);

> +	}

> +

> +	return s_sock;

> +}

> +

> +/*

> + * Client function:

> + * Register a file descriptor to the server. Return -1 on error.

> + */

> +int _odp_fdserver_register_fd(fd_server_context_e context, uint64_t key,

> +			      int fd_to_send)

> +{

> +	int s_sock; /* server socket */

> +	int res;

> +	int command;

> +	int fd;

> +

> +	odp_spinlock_lock(client_lock);

> +

> +	ODP_DBG("FD client register: pid=%d key=%" PRIu64 ", fd=%d\n",

> +		getpid(), key, fd_to_send);

> +

> +	s_sock = get_socket();

> +	if (s_sock < 0) {

> +		odp_spinlock_unlock(client_lock);

> +		return(-1);

> +	}

> +

> +	res =  send_fdserver_msg(s_sock, FD_REGISTER_REQ, context, key,

> +				 fd_to_send);

> +	if (res < 0) {

> +		ODP_ERR("fd registration failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);

> +

> +	if ((res < 0) || (command != FD_REGISTER_ACK)) {

> +		ODP_ERR("fd registration failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	close(s_sock);

> +

> +	odp_spinlock_unlock(client_lock);

> +	return 0;

> +}

> +

> +/*

> + * Client function:

> + * Deregister a file descriptor from the server. Return -1 on error.

> + */

> +int _odp_fdserver_deregister_fd(fd_server_context_e context, uint64_t key)

> +{

> +	int s_sock; /* server socket */

> +	int res;

> +	int command;

> +	int fd;

> +

> +	odp_spinlock_lock(client_lock);

> +

> +	ODP_DBG("FD client deregister: pid=%d key=%" PRIu64 "\n",

> +		getpid(), key);

> +

> +	s_sock = get_socket();

> +	if (s_sock < 0) {

> +		odp_spinlock_unlock(client_lock);

> +		return(-1);

> +	}

> +

> +	res =  send_fdserver_msg(s_sock, FD_DEREGISTER_REQ, context, key, -1);

> +	if (res < 0) {

> +		ODP_ERR("fd de-registration failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);

> +

> +	if ((res < 0) || (command != FD_DEREGISTER_ACK)) {

> +		ODP_ERR("fd de-registration failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	close(s_sock);

> +

> +	odp_spinlock_unlock(client_lock);

> +	return 0;

> +}

> +

> +/*

> + * client function:

> + * lookup a file descriptor from the server. return -1 on error,

> + * or the file descriptor on success (>=0).

> + */

> +int _odp_fdserver_lookup_fd(fd_server_context_e context, uint64_t key)

> +{

> +	int s_sock; /* server socket */

> +	int res;

> +	int command;

> +	int fd;

> +

> +	odp_spinlock_lock(client_lock);

> +

> +	s_sock = get_socket();

> +	if (s_sock < 0) {

> +		odp_spinlock_unlock(client_lock);

> +		return(-1);

> +	}

> +

> +	res =  send_fdserver_msg(s_sock, FD_LOOKUP_REQ, context, key, -1);

> +	if (res < 0) {

> +		ODP_ERR("fd lookup failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);

> +

> +	if ((res < 0) || (command != FD_LOOKUP_ACK)) {

> +		ODP_ERR("fd lookup failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	close(s_sock);

> +	ODP_DBG("FD client lookup: pid=%d, key=%" PRIu64 ", fd=%d\n",

> +		getpid(), key, fd);

> +

> +	odp_spinlock_unlock(client_lock);

> +	return fd;

> +}

> +

> +/*

> + * request server terminaison:

> + */

> +static int stop_server(void)

> +{

> +	int s_sock; /* server socket */

> +	int res;

> +

> +	odp_spinlock_lock(client_lock);

> +

> +	ODP_DBG("FD sending server stop request\n");

> +

> +	s_sock = get_socket();

> +	if (s_sock < 0) {

> +		odp_spinlock_unlock(client_lock);

> +		return(-1);

> +	}

> +

> +	res =  send_fdserver_msg(s_sock, FD_SERVERSTOP_REQ, 0, 0, -1);

> +	if (res < 0) {

> +		ODP_ERR("fd stop request failure\n");

> +		close(s_sock);

> +		odp_spinlock_unlock(client_lock);

> +		return -1;

> +	}

> +

> +	close(s_sock);

> +

> +	odp_spinlock_unlock(client_lock);

> +	return 0;

> +}

> +

> +/*

> + * server function

> + * receive a client request and handle it.

> + * Always returns 0 unless a stop request is received.

> + */

> +static int handle_request(int client_sock)

> +{

> +	int command;

> +	fd_server_context_e context;

> +	uint64_t key;

> +	int fd;

> +	int i;

> +

> +	/* get a client request: */

> +	recv_fdserver_msg(client_sock, &command, &context, &key, &fd);


Might be good to handle return value on this one.

> +	switch (command) {

> +	case FD_REGISTER_REQ:

> +		if ((fd < 0) || (context >= FD_SRV_CTX_END)) {

> +			ODP_ERR("Invalid register fd or context\n");

> +			send_fdserver_msg(client_sock, FD_REGISTER_NACK,

> +					  FD_SRV_CTX_NA, 0, -1);

> +			return 0;

> +		}

> +

> +		/* store the file descriptor in table: */

> +		if (fd_table_nb_entries < FDSERVER_MAX_ENTRIES) {

> +			fd_table[fd_table_nb_entries].context = context;

> +			fd_table[fd_table_nb_entries].key     = key;

> +			fd_table[fd_table_nb_entries++].fd    = fd;

> +			ODP_DBG("storing {ctx=%d, key=%" PRIu64 "}->fd=%d\n",

> +				context, key, fd);

> +		} else {

> +			ODP_ERR("FD table full\n");

> +			send_fdserver_msg(client_sock, FD_REGISTER_NACK,

> +					  FD_SRV_CTX_NA, 0, -1);

> +			return 0;

> +		}

> +

> +		send_fdserver_msg(client_sock, FD_REGISTER_ACK,

> +				  FD_SRV_CTX_NA, 0, -1);

> +		break;

> +

> +	case FD_LOOKUP_REQ:

> +		if (context >= FD_SRV_CTX_END) {

> +			ODP_ERR("invalid lookup context\n");

> +			send_fdserver_msg(client_sock, FD_LOOKUP_NACK,

> +					  FD_SRV_CTX_NA, 0, -1);

> +			return 0;

> +		}

> +

> +		/* search key in table and sent reply: */

> +		for (i = 0; i < fd_table_nb_entries; i++) {

> +			if ((fd_table[i].context == context) &&

> +			    (fd_table[i].key == key)) {

> +				fd = fd_table[i].fd;

> +				ODP_DBG("lookup {ctx=%d,"

> +					" key=%" PRIu64 "}->fd=%d\n",

> +					context, key, fd);

> +				send_fdserver_msg(client_sock,

> +						  FD_LOOKUP_ACK, context, key,

> +						  fd);

> +				return 0;

> +			}

> +		}

> +

> +		/* context+key not found... send nack */

> +		send_fdserver_msg(client_sock, FD_LOOKUP_NACK, context, key,

> +				  -1);

> +		break;

> +

> +	case FD_DEREGISTER_REQ:

> +		if (context >= FD_SRV_CTX_END) {

> +			ODP_ERR("invalid deregister context\n");

> +			send_fdserver_msg(client_sock, FD_DEREGISTER_NACK,

> +					  FD_SRV_CTX_NA, 0, -1);

> +			return 0;

> +		}

> +

> +		/* search key in table and remove it if found, and reply: */

> +		for (i = 0; i < fd_table_nb_entries; i++) {

> +			if ((fd_table[i].context == context) &&

> +			    (fd_table[i].key == key)) {

> +				ODP_DBG("drop {ctx=%d,"

> +					" key=%" PRIu64 "}->fd=%d\n",

> +					context, key, fd_table[i].fd);

> +				close(fd_table[i].fd);

> +				fd_table[i] = fd_table[--fd_table_nb_entries];

> +				send_fdserver_msg(client_sock,

> +						  FD_DEREGISTER_ACK,

> +						  context, key, -1);

> +				return 0;

> +			}

> +		}

> +

> +		/* key not found... send nack */

> +		send_fdserver_msg(client_sock, FD_DEREGISTER_NACK,

> +				  context, key, -1);

> +		break;

> +

> +	case FD_SERVERSTOP_REQ:

> +		ODP_DBG("Stoping FD server\n");

> +		return 1;

> +

> +	default:

> +		ODP_ERR("Unexpected request\n");

> +		break;

> +	}

> +	return 0;

> +}

> +

> +/*

> + * server function

> + * loop forever, handling client requests one by one

> + */

> +static void wait_requests(int sock)

> +{

> +	int c_socket; /* client connection */

> +	unsigned int addr_sz;

> +	struct sockaddr_un remote;

> +

> +	for (;;) {

> +		addr_sz = sizeof(remote);

> +		c_socket = accept(sock, (struct sockaddr *)&remote, &addr_sz);

> +		if (c_socket == -1) {

> +				ODP_ERR("wait_requests: %s\n", strerror(errno));

> +				return;

> +		}

> +

> +		if (handle_request(c_socket))

> +			break;

> +		close(c_socket);

> +	}

> +	close(c_socket);

> +}

> +

> +/*

> + * Create a unix domain socket and fork a process to listen to incoming

> + * requests.

> + */

> +int _odp_fdserver_init_global(void)

> +{

> +	char sockpath[FDSERVER_SOCKPATH_MAXLEN];

> +	int sock;

> +	struct sockaddr_un local;

> +	pid_t server_pid;

> +	int res;

> +

> +	/* create the client spinlock that any client can see: */

> +	client_lock = mmap(NULL, sizeof(odp_spinlock_t), PROT_READ | PROT_WRITE,

> +			   MAP_SHARED | MAP_ANONYMOUS, -1, 0);

> +

> +	odp_spinlock_init(client_lock);

> +

> +	/* construct the server named socket path: */

> +	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,

> +		 odp_global_data.main_pid);

> +

> +	/* create UNIX domain socket: */

> +	sock = socket(AF_UNIX, SOCK_STREAM, 0);

> +	if (sock == -1) {

> +		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));

> +		return(-1);

> +	}

> +

> +	/* remove previous named socket if it already exists: */

> +	unlink(sockpath);

> +

> +	/* bind to new named socket: */

> +	local.sun_family = AF_UNIX;

> +	strncpy(local.sun_path, sockpath, sizeof(local.sun_path));

> +	res = bind(sock, (struct sockaddr *)&local, sizeof(struct sockaddr_un));

> +	if (res == -1) {

> +		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));

> +		close(sock);

> +		return(-1);

> +	}

> +

> +	/* listen for incoming conections: */

> +	if (listen(sock, FDSERVER_BACKLOG) == -1) {

> +		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));

> +		close(sock);

> +		return(-1);

> +	}

> +

> +	/* fork a server process: */

> +	server_pid = fork();

> +	if (server_pid == -1) {

> +		ODP_ERR("Could not fork!\n");

> +		close(sock);

> +		return(-1);

> +	}

> +

> +	if (server_pid == 0) { /*child */

> +		/* TODO: pin the server on appropriate service cpu mask */

> +		/* when (if) we can agree on the usage of service mask  */

> +

> +		wait_requests(sock); /* Returns when server is stopped  */

> +		close(sock);

> +		exit(0);

> +	}

> +

> +	/* parent */

> +	close(sock);

> +	return 0;

> +}

> +

> +/*

> + * Terminate the server

> + */

> +int _odp_fdserver_term_global(void)

> +{

> +	int status;

> +	char sockpath[FDSERVER_SOCKPATH_MAXLEN];

> +

> +	/* close the server and wait for child terminaison*/

> +	stop_server();

> +	wait(&status);

> +

> +	/* construct the server named socket path: */

> +	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,

> +		 odp_global_data.main_pid);

> +

> +	/* delete the UNIX domain socket: */

> +	unlink(sockpath);

> +

> +	return 0;

> +}

> diff --git a/platform/linux-generic/include/_fdserver_internal.h b/platform/linux-generic/include/_fdserver_internal.h

> new file mode 100644

> index 0000000..480ac02

> --- /dev/null

> +++ b/platform/linux-generic/include/_fdserver_internal.h

> @@ -0,0 +1,38 @@

> +/* Copyright (c) 2016, Linaro Limited

> + * All rights reserved.

> + *

> + * SPDX-License-Identifier:     BSD-3-Clause

> + */

> +

> +#ifndef _FD_SERVER_INTERNAL_H

> +#define _FD_SERVER_INTERNAL_H

> +

> +#ifdef __cplusplus

> +extern "C" {

> +#endif

> +

> +/*

> + * the following enum defines the different contextes by which the

> + * FD server may be used: In the FD server, the keys used to store/retrieve

> + * a file descriptor are actually context based:

> + * Both the context and the key are stored at fd registration time,

> + * and both the context and the key are used to retrieve a fd.

> + * In other words a context identifies a FD server usage, so that different

> + * unrelated fd server users do not have to guarantee key unicity between

> + * them.

> + */

> +typedef enum fd_server_context {

> +	FD_SRV_CTX_NA,  /* Not Applicable   */

> +	FD_SRV_CTX_END, /* upper enum limit */

> +} fd_server_context_e;

> +

> +int _odp_fdserver_register_fd(fd_server_context_e context, uint64_t key,

> +			      int fd);

> +int _odp_fdserver_deregister_fd(fd_server_context_e context, uint64_t key);

> +int _odp_fdserver_lookup_fd(fd_server_context_e context, uint64_t key);

> +

> +#ifdef __cplusplus

> +}

> +#endif

> +

> +#endif

> diff --git a/platform/linux-generic/include/odp_internal.h b/platform/linux-generic/include/odp_internal.h

> index 8bad450..fd770b5 100644

> --- a/platform/linux-generic/include/odp_internal.h

> +++ b/platform/linux-generic/include/odp_internal.h

> @@ -54,6 +54,7 @@ enum init_stage {

>  	CPUMASK_INIT,

>  	TIME_INIT,

>  	SYSINFO_INIT,

> +	FDSERVER_INIT,

>  	SHM_INIT,

>  	THREAD_INIT,

>  	POOL_INIT,

> @@ -119,6 +120,9 @@ int odp_tm_term_global(void);

>  int _odp_int_name_tbl_init_global(void);

>  int _odp_int_name_tbl_term_global(void);

>  

> +int _odp_fdserver_init_global(void);

> +int _odp_fdserver_term_global(void);

> +

>  int cpuinfo_parser(FILE *file, system_info_t *sysinfo);

>  uint64_t odp_cpu_hz_current(int id);

>  

> diff --git a/platform/linux-generic/odp_init.c b/platform/linux-generic/odp_init.c

> index f534759..ec84fd0 100644

> --- a/platform/linux-generic/odp_init.c

> +++ b/platform/linux-generic/odp_init.c

> @@ -51,6 +51,12 @@ int odp_init_global(odp_instance_t *instance,

>  	}

>  	stage = SYSINFO_INIT;

>  

> +	if (_odp_fdserver_init_global()) {

> +		ODP_ERR("ODP fdserver init failed.\n");

> +		goto init_failed;

> +	}

> +	stage = FDSERVER_INIT;

> +

>  	if (odp_shm_init_global()) {

>  		ODP_ERR("ODP shm init failed.\n");

>  		goto init_failed;

> @@ -218,6 +224,13 @@ int _odp_term_global(enum init_stage stage)

>  		}

>  		/* Fall through */

>  

> +	case FDSERVER_INIT:

> +		if (_odp_fdserver_term_global()) {

> +			ODP_ERR("ODP fdserver term failed.\n");

> +			rc = -1;

> +		}

> +		/* Fall through */

> +

>  	case SYSINFO_INIT:

>  		if (odp_system_info_term()) {

>  			ODP_ERR("ODP system info term failed.\n");

> -- 

> 2.7.4

>
Maxim Uvarov Aug. 31, 2016, 6:56 a.m. | #2
Hello Brian,

it looks like fixes for your comments can be in separate patch. If 
everything is ok with patch series I would merge
it now and when Christophe will be back from vacation he can send 
updated patch. Does that work for you?

Thank you,
Maxim.

On 08/29/16 06:02, Brian Brooks wrote:
> On 08/20 09:45:56, Christophe Milard wrote:

>> A fdserver is added and started at init time.

>> The role of the fdserver (file descriptor server) is to enable sharing of

>> file descriptors between unrelated processes: processes which wish to

>> share a file descriptor may register their files descriptor to the server

>> and processes wishing to use the shared file descriptors can do a lookup.

>> When registration occurs, a triple {context, key, fd} is provided to

>> the server. The context identifies the client and scope (i.e. shmem).

>> The key is implemented as a long, and can be whatever.

>> The server won't care as long as keys are unique.

>> The file descriptor can be retrieved by another process providing the same

>> context and key.

>> Files descriptors passed this way are converted on the fly during the unix

>> domain socket communication that occurs between the server and its clients.

>> This is done by using the ancillary control data part of the msg.

>>

>> Signed-off-by: Christophe Milard <christophe.milard@linaro.org>

>> ---

>>   platform/linux-generic/Makefile.am                 |   2 +

>>   platform/linux-generic/_fdserver.c                 | 655 +++++++++++++++++++++

>>   .../linux-generic/include/_fdserver_internal.h     |  38 ++

>>   platform/linux-generic/include/odp_internal.h      |   4 +

>>   platform/linux-generic/odp_init.c                  |  13 +

>>   5 files changed, 712 insertions(+)

>>   create mode 100644 platform/linux-generic/_fdserver.c

>>   create mode 100644 platform/linux-generic/include/_fdserver_internal.h

>>

>> diff --git a/platform/linux-generic/Makefile.am b/platform/linux-generic/Makefile.am

>> index 271f5b5..cb7c8e9 100644

>> --- a/platform/linux-generic/Makefile.am

>> +++ b/platform/linux-generic/Makefile.am

>> @@ -112,6 +112,7 @@ odpdrvplatinclude_HEADERS = \

>>   		  $(srcdir)/include/odp/drv/plat/strong_types.h

>>   

>>   noinst_HEADERS = \

>> +		  ${srcdir}/include/_fdserver_internal.h \

>>   		  ${srcdir}/include/odp_align_internal.h \

>>   		  ${srcdir}/include/odp_atomic_internal.h \

>>   		  ${srcdir}/include/odp_buffer_inlines.h \

>> @@ -154,6 +155,7 @@ noinst_HEADERS = \

>>   		  ${srcdir}/Makefile.inc

>>   

>>   __LIB__libodp_linux_la_SOURCES = \

>> +			   _fdserver.c \

>>   			   odp_atomic.c \

>>   			   odp_barrier.c \

>>   			   odp_buffer.c \

>> diff --git a/platform/linux-generic/_fdserver.c b/platform/linux-generic/_fdserver.c

>> new file mode 100644

>> index 0000000..bf36eb2

>> --- /dev/null

>> +++ b/platform/linux-generic/_fdserver.c

>> @@ -0,0 +1,655 @@

>> +/* Copyright (c) 2016, Linaro Limited

>> + * All rights reserved.

>> + *

>> + * SPDX-License-Identifier:     BSD-3-Clause

>> + */

>> +

>> +/*

>> + * This file implements a file descriptor sharing server enabling

>> + * sharing of file descriptors between processes, regardless of fork time.

>> + *

>> + * File descriptors are process scoped, but they can be "sent and converted

>> + * on the fly" between processes using special unix domain socket ancillary

>> + * data.

>> + * The receiving process gets a file descriptor "pointing" to the same thing

>> + * as the one sent (but the value of the file descriptor itself may be different

>> + * from the one sent).

>> + * Because ODP applications are responsible for creating ODP threads (i.e.

>> + * pthreads or linux processes), ODP has no control on the order things happen:

>> + * Nothing prevent a thread A to fork B and C, and then C creating a pktio

>> + * which will be used by A and B to send/receive packets.

>> + * Assuming this pktio uses a file descriptor, the latter will need to be

>> + * shared between the processes, despite the "non convenient" fork time.

>> + * The shared memory allocator is likely to use this as well to be able to

>> + * share memory regardless of fork() time.

>> + * This server handles a table of {(context,key)<-> fd} pair, and is

>> + * interfaced by the following functions:

>> + *

>> + * _odp_fdserver_register_fd(context, key, fd_to_send);

>> + * _odp_fdserver_deregister_fd(context, key);

>> + * _odp_fdserver_lookup_fd(context, key);

>> + *

>> + * which are used to register/deregister or querry for file descriptor based

>> + * on a context and key value couple, which has to be unique.

>> + *

>> + * Note again that the file descriptors stored here are local to this server

>> + * process and get converted both when registered or looked up.

>> + */

>> +

>> +#include <odp_posix_extensions.h>

>> +#include <odp/api/spinlock.h>

>> +#include <odp_internal.h>

>> +#include <odp_debug_internal.h>

>> +#include <_fdserver_internal.h>

>> +

>> +#include <stdio.h>

>> +#include <stdlib.h>

>> +#include <errno.h>

>> +#include <string.h>

>> +#include <sys/types.h>

>> +#include <signal.h>

>> +#include <sys/socket.h>

>> +#include <sys/un.h>

>> +#include <unistd.h>

>> +#include <inttypes.h>

>> +#include <sys/mman.h>

>> +#include <sys/wait.h>

>> +

>> +#define FDSERVER_SOCKPATH_MAXLEN 32

>> +#define FDSERVER_SOCKPATH_FORMAT "/tmp/odp-%d-fdserver"

>> +

>> +#define FDSERVER_BACKLOG 5

>> +

>> +#ifndef MAP_ANONYMOUS

>> +#define MAP_ANONYMOUS MAP_ANON

>> +#endif

>> +

>> +/* when accessing the client functions, clients should be mutexed: */

>> +odp_spinlock_t *client_lock;

>> +

>> +/* define the tables of file descriptors handled by this server: */

>> +#define FDSERVER_MAX_ENTRIES 256

>> +typedef struct fdentry_s {

>> +	fd_server_context_e context;

>> +	uint64_t key;

>> +	int  fd;

>> +} fdentry_t;

>> +static fdentry_t fd_table[FDSERVER_MAX_ENTRIES];

>> +static int fd_table_nb_entries;

> This data will be COW on fork, which means it needs to be owned by some other

> process instead of this shared library. Is this in your mind?

>

>> +/*

>> + * define the message struct used for communication between client and server

>> + * (this single message is used in both direction)

>> + * The file descriptors are sent out of band as ancillary data for conversion.

>> + */

>> +typedef struct fd_server_msg {

>> +	int command;

>> +	fd_server_context_e context;

>> +	uint64_t key;

>> +} fdserver_msg_t;

>> +/* possible commands are: */

>> +#define FD_REGISTER_REQ		1  /* client -> server */

>> +#define FD_REGISTER_ACK		2  /* server -> client */

>> +#define FD_REGISTER_NACK	3  /* server -> client */

>> +#define FD_LOOKUP_REQ		4  /* client -> server */

>> +#define FD_LOOKUP_ACK		5  /* server -> client */

>> +#define FD_LOOKUP_NACK		6  /* server -> client */

>> +#define FD_DEREGISTER_REQ	7  /* client -> server */

>> +#define FD_DEREGISTER_ACK	8  /* server -> client */

>> +#define FD_DEREGISTER_NACK	9  /* server -> client */

>> +#define FD_SERVERSTOP_REQ	10 /* client -> server (stops) */

>> +

>> +/*

>> + * Client and server function:

>> + * Send a fdserver_msg, possibly including a file descriptor, on the socket

>> + * This function is used both by:

>> + * -the client (sending a FD_REGISTER_REQ with a file descriptor to be shared,

>> + *  or FD_LOOKUP_REQ/FD_DEREGISTER_REQ without a file descirptor)

>> + * -the server (sending FD_REGISTER_ACK/NACK, FD_LOOKUP_NACK,

>> + *  FD_DEREGISTER_ACK/NACK... without a fd or a

>> + *  FD_LOOKUP_ACK with a fd)

>> + * This function make use of the ancillary data (control data) to pass and

>> + * convert file descriptors over UNIX sockets

>> + * Return -1 on error, 0 on success.

>> + */

>> +static int send_fdserver_msg(int sock, int command,

>> +			     fd_server_context_e context, uint64_t key,

>> +			     int fd_to_send)

>> +{

>> +	struct msghdr socket_message;

>> +	struct iovec io_vector[1]; /* one msg frgmt only */

>> +	struct cmsghdr *control_message = NULL;

>> +	int *fd_location;

>> +	fdserver_msg_t msg;

>> +	int res;

>> +

>> +	char ancillary_data[CMSG_SPACE(sizeof(int))];

>> +

>> +	/* prepare the register request body (single framgent): */

>> +	msg.command = command;

>> +	msg.context = context;

>> +	msg.key = key;

>> +	io_vector[0].iov_base = &msg;

>> +	io_vector[0].iov_len = sizeof(fdserver_msg_t);

>> +

>> +	/* initialize socket message */

>> +	memset(&socket_message, 0, sizeof(struct msghdr));

>> +	socket_message.msg_iov = io_vector;

>> +	socket_message.msg_iovlen = 1;

>> +

>> +	if (fd_to_send >= 0) {

>> +		/* provide space for the ancillary data */

>> +		memset(ancillary_data, 0, CMSG_SPACE(sizeof(int)));

>> +		socket_message.msg_control = ancillary_data;

>> +		socket_message.msg_controllen = CMSG_SPACE(sizeof(int));

>> +

>> +		/* initialize a single ancillary data element for fd passing */

>> +		control_message = CMSG_FIRSTHDR(&socket_message);

>> +		control_message->cmsg_level = SOL_SOCKET;

>> +		control_message->cmsg_type = SCM_RIGHTS;

>> +		control_message->cmsg_len = CMSG_LEN(sizeof(int));

>> +		fd_location = (int *)(void *)CMSG_DATA(control_message);

>> +		*fd_location = fd_to_send;

>> +	}

>> +	res = sendmsg(sock, &socket_message, 0);

>> +	if (res < 0) {

>> +		ODP_ERR("send_fdserver_msg: %s\n", strerror(errno));

>> +		return(-1);

>> +	}

>> +

>> +	return 0;

>> +}

>> +

>> +/*

>> + * Client and server function

>> + * Receive a fdserver_msg, possibly including a file descriptor, on the

>> + * given socket.

>> + * This function is used both by:

>> + * -the server (receiving a FD_REGISTER_REQ with a file descriptor to be shared,

>> + *  or FD_LOOKUP_REQ, FD_DEREGISTER_REQ without a file descirptor)

>> + * -the client (receiving FD_REGISTER_ACK...without a fd or a FD_LOOKUP_ACK with

>> + * a fd)

>> + * This function make use of the ancillary data (control data) to pass and

>> + * convert file descriptors over UNIX sockets.

>> + * Return -1 on error, 0 on success.

>> + */

>> +static int recv_fdserver_msg(int sock, int *command,

>> +			     fd_server_context_e *context, uint64_t *key,

>> +			     int *recvd_fd)

>> +{

>> +	struct msghdr socket_message;

>> +	struct iovec io_vector[1]; /* one msg frgmt only */

>> +	struct cmsghdr *control_message = NULL;

>> +	int *fd_location;

>> +	fdserver_msg_t msg;

>> +	char ancillary_data[CMSG_SPACE(sizeof(int))];

>> +

>> +	memset(&socket_message, 0, sizeof(struct msghdr));

>> +	memset(ancillary_data, 0, CMSG_SPACE(sizeof(int)));

>> +

>> +	/* setup a place to fill in message contents */

>> +	io_vector[0].iov_base = &msg;

>> +	io_vector[0].iov_len = sizeof(fdserver_msg_t);

>> +	socket_message.msg_iov = io_vector;

>> +	socket_message.msg_iovlen = 1;

>> +

>> +	/* provide space for the ancillary data */

>> +	socket_message.msg_control = ancillary_data;

>> +	socket_message.msg_controllen = CMSG_SPACE(sizeof(int));

>> +

>> +	/* receive the message */

>> +	if (recvmsg(sock, &socket_message, MSG_CMSG_CLOEXEC) < 0) {

>> +		ODP_ERR("recv_fdserver_msg: %s\n", strerror(errno));

>> +		return(-1);

>> +	}

>> +

>> +	*command = msg.command;

>> +	*context = msg.context;

>> +	*key = msg.key;

>> +

>> +	/* grab the converted file descriptor (if any) */

>> +	*recvd_fd = -1;

>> +

>> +	if ((socket_message.msg_flags & MSG_CTRUNC) == MSG_CTRUNC)

>> +		return 0;

>> +

>> +	/* iterate ancillary elements to find the file descriptor: */

>> +	for (control_message = CMSG_FIRSTHDR(&socket_message);

>> +	     control_message != NULL;

>> +	     control_message = CMSG_NXTHDR(&socket_message, control_message)) {

>> +		if ((control_message->cmsg_level == SOL_SOCKET) &&

>> +		    (control_message->cmsg_type == SCM_RIGHTS)) {

>> +			fd_location = (int *)(void *)CMSG_DATA(control_message);

>> +			*recvd_fd = *fd_location;

>> +			break;

>> +		}

>> +	}

>> +

>> +	return 0;

>> +}

>> +

>> +/* opens and returns a connected socket to the server */

>> +static int get_socket(void)

>> +{

>> +	char sockpath[FDSERVER_SOCKPATH_MAXLEN];

>> +	int s_sock; /* server socket */

>> +	struct sockaddr_un remote;

>> +	int len;

>> +

>> +	/* construct the named socket path: */

>> +	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,

>> +		 odp_global_data.main_pid);

>> +

>> +	s_sock = socket(AF_UNIX, SOCK_STREAM, 0);

>> +	if (s_sock == -1) {

>> +		ODP_ERR("cannot connect to server: %s\n", strerror(errno));

>> +		return(-1);

>> +	}

>> +

>> +	remote.sun_family = AF_UNIX;

>> +	strcpy(remote.sun_path, sockpath);

>> +	len = strlen(remote.sun_path) + sizeof(remote.sun_family);

>> +	if (connect(s_sock, (struct sockaddr *)&remote, len) == -1) {

>> +		ODP_ERR("cannot connect to server: %s\n", strerror(errno));

>> +		close(s_sock);

>> +		return(-1);

>> +	}

>> +

>> +	return s_sock;

>> +}

>> +

>> +/*

>> + * Client function:

>> + * Register a file descriptor to the server. Return -1 on error.

>> + */

>> +int _odp_fdserver_register_fd(fd_server_context_e context, uint64_t key,

>> +			      int fd_to_send)

>> +{

>> +	int s_sock; /* server socket */

>> +	int res;

>> +	int command;

>> +	int fd;

>> +

>> +	odp_spinlock_lock(client_lock);

>> +

>> +	ODP_DBG("FD client register: pid=%d key=%" PRIu64 ", fd=%d\n",

>> +		getpid(), key, fd_to_send);

>> +

>> +	s_sock = get_socket();

>> +	if (s_sock < 0) {

>> +		odp_spinlock_unlock(client_lock);

>> +		return(-1);

>> +	}

>> +

>> +	res =  send_fdserver_msg(s_sock, FD_REGISTER_REQ, context, key,

>> +				 fd_to_send);

>> +	if (res < 0) {

>> +		ODP_ERR("fd registration failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);

>> +

>> +	if ((res < 0) || (command != FD_REGISTER_ACK)) {

>> +		ODP_ERR("fd registration failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	close(s_sock);

>> +

>> +	odp_spinlock_unlock(client_lock);

>> +	return 0;

>> +}

>> +

>> +/*

>> + * Client function:

>> + * Deregister a file descriptor from the server. Return -1 on error.

>> + */

>> +int _odp_fdserver_deregister_fd(fd_server_context_e context, uint64_t key)

>> +{

>> +	int s_sock; /* server socket */

>> +	int res;

>> +	int command;

>> +	int fd;

>> +

>> +	odp_spinlock_lock(client_lock);

>> +

>> +	ODP_DBG("FD client deregister: pid=%d key=%" PRIu64 "\n",

>> +		getpid(), key);

>> +

>> +	s_sock = get_socket();

>> +	if (s_sock < 0) {

>> +		odp_spinlock_unlock(client_lock);

>> +		return(-1);

>> +	}

>> +

>> +	res =  send_fdserver_msg(s_sock, FD_DEREGISTER_REQ, context, key, -1);

>> +	if (res < 0) {

>> +		ODP_ERR("fd de-registration failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);

>> +

>> +	if ((res < 0) || (command != FD_DEREGISTER_ACK)) {

>> +		ODP_ERR("fd de-registration failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	close(s_sock);

>> +

>> +	odp_spinlock_unlock(client_lock);

>> +	return 0;

>> +}

>> +

>> +/*

>> + * client function:

>> + * lookup a file descriptor from the server. return -1 on error,

>> + * or the file descriptor on success (>=0).

>> + */

>> +int _odp_fdserver_lookup_fd(fd_server_context_e context, uint64_t key)

>> +{

>> +	int s_sock; /* server socket */

>> +	int res;

>> +	int command;

>> +	int fd;

>> +

>> +	odp_spinlock_lock(client_lock);

>> +

>> +	s_sock = get_socket();

>> +	if (s_sock < 0) {

>> +		odp_spinlock_unlock(client_lock);

>> +		return(-1);

>> +	}

>> +

>> +	res =  send_fdserver_msg(s_sock, FD_LOOKUP_REQ, context, key, -1);

>> +	if (res < 0) {

>> +		ODP_ERR("fd lookup failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);

>> +

>> +	if ((res < 0) || (command != FD_LOOKUP_ACK)) {

>> +		ODP_ERR("fd lookup failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	close(s_sock);

>> +	ODP_DBG("FD client lookup: pid=%d, key=%" PRIu64 ", fd=%d\n",

>> +		getpid(), key, fd);

>> +

>> +	odp_spinlock_unlock(client_lock);

>> +	return fd;

>> +}

>> +

>> +/*

>> + * request server terminaison:

>> + */

>> +static int stop_server(void)

>> +{

>> +	int s_sock; /* server socket */

>> +	int res;

>> +

>> +	odp_spinlock_lock(client_lock);

>> +

>> +	ODP_DBG("FD sending server stop request\n");

>> +

>> +	s_sock = get_socket();

>> +	if (s_sock < 0) {

>> +		odp_spinlock_unlock(client_lock);

>> +		return(-1);

>> +	}

>> +

>> +	res =  send_fdserver_msg(s_sock, FD_SERVERSTOP_REQ, 0, 0, -1);

>> +	if (res < 0) {

>> +		ODP_ERR("fd stop request failure\n");

>> +		close(s_sock);

>> +		odp_spinlock_unlock(client_lock);

>> +		return -1;

>> +	}

>> +

>> +	close(s_sock);

>> +

>> +	odp_spinlock_unlock(client_lock);

>> +	return 0;

>> +}

>> +

>> +/*

>> + * server function

>> + * receive a client request and handle it.

>> + * Always returns 0 unless a stop request is received.

>> + */

>> +static int handle_request(int client_sock)

>> +{

>> +	int command;

>> +	fd_server_context_e context;

>> +	uint64_t key;

>> +	int fd;

>> +	int i;

>> +

>> +	/* get a client request: */

>> +	recv_fdserver_msg(client_sock, &command, &context, &key, &fd);

> Might be good to handle return value on this one.

>

>> +	switch (command) {

>> +	case FD_REGISTER_REQ:

>> +		if ((fd < 0) || (context >= FD_SRV_CTX_END)) {

>> +			ODP_ERR("Invalid register fd or context\n");

>> +			send_fdserver_msg(client_sock, FD_REGISTER_NACK,

>> +					  FD_SRV_CTX_NA, 0, -1);

>> +			return 0;

>> +		}

>> +

>> +		/* store the file descriptor in table: */

>> +		if (fd_table_nb_entries < FDSERVER_MAX_ENTRIES) {

>> +			fd_table[fd_table_nb_entries].context = context;

>> +			fd_table[fd_table_nb_entries].key     = key;

>> +			fd_table[fd_table_nb_entries++].fd    = fd;

>> +			ODP_DBG("storing {ctx=%d, key=%" PRIu64 "}->fd=%d\n",

>> +				context, key, fd);

>> +		} else {

>> +			ODP_ERR("FD table full\n");

>> +			send_fdserver_msg(client_sock, FD_REGISTER_NACK,

>> +					  FD_SRV_CTX_NA, 0, -1);

>> +			return 0;

>> +		}

>> +

>> +		send_fdserver_msg(client_sock, FD_REGISTER_ACK,

>> +				  FD_SRV_CTX_NA, 0, -1);

>> +		break;

>> +

>> +	case FD_LOOKUP_REQ:

>> +		if (context >= FD_SRV_CTX_END) {

>> +			ODP_ERR("invalid lookup context\n");

>> +			send_fdserver_msg(client_sock, FD_LOOKUP_NACK,

>> +					  FD_SRV_CTX_NA, 0, -1);

>> +			return 0;

>> +		}

>> +

>> +		/* search key in table and sent reply: */

>> +		for (i = 0; i < fd_table_nb_entries; i++) {

>> +			if ((fd_table[i].context == context) &&

>> +			    (fd_table[i].key == key)) {

>> +				fd = fd_table[i].fd;

>> +				ODP_DBG("lookup {ctx=%d,"

>> +					" key=%" PRIu64 "}->fd=%d\n",

>> +					context, key, fd);

>> +				send_fdserver_msg(client_sock,

>> +						  FD_LOOKUP_ACK, context, key,

>> +						  fd);

>> +				return 0;

>> +			}

>> +		}

>> +

>> +		/* context+key not found... send nack */

>> +		send_fdserver_msg(client_sock, FD_LOOKUP_NACK, context, key,

>> +				  -1);

>> +		break;

>> +

>> +	case FD_DEREGISTER_REQ:

>> +		if (context >= FD_SRV_CTX_END) {

>> +			ODP_ERR("invalid deregister context\n");

>> +			send_fdserver_msg(client_sock, FD_DEREGISTER_NACK,

>> +					  FD_SRV_CTX_NA, 0, -1);

>> +			return 0;

>> +		}

>> +

>> +		/* search key in table and remove it if found, and reply: */

>> +		for (i = 0; i < fd_table_nb_entries; i++) {

>> +			if ((fd_table[i].context == context) &&

>> +			    (fd_table[i].key == key)) {

>> +				ODP_DBG("drop {ctx=%d,"

>> +					" key=%" PRIu64 "}->fd=%d\n",

>> +					context, key, fd_table[i].fd);

>> +				close(fd_table[i].fd);

>> +				fd_table[i] = fd_table[--fd_table_nb_entries];

>> +				send_fdserver_msg(client_sock,

>> +						  FD_DEREGISTER_ACK,

>> +						  context, key, -1);

>> +				return 0;

>> +			}

>> +		}

>> +

>> +		/* key not found... send nack */

>> +		send_fdserver_msg(client_sock, FD_DEREGISTER_NACK,

>> +				  context, key, -1);

>> +		break;

>> +

>> +	case FD_SERVERSTOP_REQ:

>> +		ODP_DBG("Stoping FD server\n");

>> +		return 1;

>> +

>> +	default:

>> +		ODP_ERR("Unexpected request\n");

>> +		break;

>> +	}

>> +	return 0;

>> +}

>> +

>> +/*

>> + * server function

>> + * loop forever, handling client requests one by one

>> + */

>> +static void wait_requests(int sock)

>> +{

>> +	int c_socket; /* client connection */

>> +	unsigned int addr_sz;

>> +	struct sockaddr_un remote;

>> +

>> +	for (;;) {

>> +		addr_sz = sizeof(remote);

>> +		c_socket = accept(sock, (struct sockaddr *)&remote, &addr_sz);

>> +		if (c_socket == -1) {

>> +				ODP_ERR("wait_requests: %s\n", strerror(errno));

>> +				return;

>> +		}

>> +

>> +		if (handle_request(c_socket))

>> +			break;

>> +		close(c_socket);

>> +	}

>> +	close(c_socket);

>> +}

>> +

>> +/*

>> + * Create a unix domain socket and fork a process to listen to incoming

>> + * requests.

>> + */

>> +int _odp_fdserver_init_global(void)

>> +{

>> +	char sockpath[FDSERVER_SOCKPATH_MAXLEN];

>> +	int sock;

>> +	struct sockaddr_un local;

>> +	pid_t server_pid;

>> +	int res;

>> +

>> +	/* create the client spinlock that any client can see: */

>> +	client_lock = mmap(NULL, sizeof(odp_spinlock_t), PROT_READ | PROT_WRITE,

>> +			   MAP_SHARED | MAP_ANONYMOUS, -1, 0);

>> +

>> +	odp_spinlock_init(client_lock);

>> +

>> +	/* construct the server named socket path: */

>> +	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,

>> +		 odp_global_data.main_pid);

>> +

>> +	/* create UNIX domain socket: */

>> +	sock = socket(AF_UNIX, SOCK_STREAM, 0);

>> +	if (sock == -1) {

>> +		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));

>> +		return(-1);

>> +	}

>> +

>> +	/* remove previous named socket if it already exists: */

>> +	unlink(sockpath);

>> +

>> +	/* bind to new named socket: */

>> +	local.sun_family = AF_UNIX;

>> +	strncpy(local.sun_path, sockpath, sizeof(local.sun_path));

>> +	res = bind(sock, (struct sockaddr *)&local, sizeof(struct sockaddr_un));

>> +	if (res == -1) {

>> +		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));

>> +		close(sock);

>> +		return(-1);

>> +	}

>> +

>> +	/* listen for incoming conections: */

>> +	if (listen(sock, FDSERVER_BACKLOG) == -1) {

>> +		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));

>> +		close(sock);

>> +		return(-1);

>> +	}

>> +

>> +	/* fork a server process: */

>> +	server_pid = fork();

>> +	if (server_pid == -1) {

>> +		ODP_ERR("Could not fork!\n");

>> +		close(sock);

>> +		return(-1);

>> +	}

>> +

>> +	if (server_pid == 0) { /*child */

>> +		/* TODO: pin the server on appropriate service cpu mask */

>> +		/* when (if) we can agree on the usage of service mask  */

>> +

>> +		wait_requests(sock); /* Returns when server is stopped  */

>> +		close(sock);

>> +		exit(0);

>> +	}

>> +

>> +	/* parent */

>> +	close(sock);

>> +	return 0;

>> +}

>> +

>> +/*

>> + * Terminate the server

>> + */

>> +int _odp_fdserver_term_global(void)

>> +{

>> +	int status;

>> +	char sockpath[FDSERVER_SOCKPATH_MAXLEN];

>> +

>> +	/* close the server and wait for child terminaison*/

>> +	stop_server();

>> +	wait(&status);

>> +

>> +	/* construct the server named socket path: */

>> +	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,

>> +		 odp_global_data.main_pid);

>> +

>> +	/* delete the UNIX domain socket: */

>> +	unlink(sockpath);

>> +

>> +	return 0;

>> +}

>> diff --git a/platform/linux-generic/include/_fdserver_internal.h b/platform/linux-generic/include/_fdserver_internal.h

>> new file mode 100644

>> index 0000000..480ac02

>> --- /dev/null

>> +++ b/platform/linux-generic/include/_fdserver_internal.h

>> @@ -0,0 +1,38 @@

>> +/* Copyright (c) 2016, Linaro Limited

>> + * All rights reserved.

>> + *

>> + * SPDX-License-Identifier:     BSD-3-Clause

>> + */

>> +

>> +#ifndef _FD_SERVER_INTERNAL_H

>> +#define _FD_SERVER_INTERNAL_H

>> +

>> +#ifdef __cplusplus

>> +extern "C" {

>> +#endif

>> +

>> +/*

>> + * the following enum defines the different contextes by which the

>> + * FD server may be used: In the FD server, the keys used to store/retrieve

>> + * a file descriptor are actually context based:

>> + * Both the context and the key are stored at fd registration time,

>> + * and both the context and the key are used to retrieve a fd.

>> + * In other words a context identifies a FD server usage, so that different

>> + * unrelated fd server users do not have to guarantee key unicity between

>> + * them.

>> + */

>> +typedef enum fd_server_context {

>> +	FD_SRV_CTX_NA,  /* Not Applicable   */

>> +	FD_SRV_CTX_END, /* upper enum limit */

>> +} fd_server_context_e;

>> +

>> +int _odp_fdserver_register_fd(fd_server_context_e context, uint64_t key,

>> +			      int fd);

>> +int _odp_fdserver_deregister_fd(fd_server_context_e context, uint64_t key);

>> +int _odp_fdserver_lookup_fd(fd_server_context_e context, uint64_t key);

>> +

>> +#ifdef __cplusplus

>> +}

>> +#endif

>> +

>> +#endif

>> diff --git a/platform/linux-generic/include/odp_internal.h b/platform/linux-generic/include/odp_internal.h

>> index 8bad450..fd770b5 100644

>> --- a/platform/linux-generic/include/odp_internal.h

>> +++ b/platform/linux-generic/include/odp_internal.h

>> @@ -54,6 +54,7 @@ enum init_stage {

>>   	CPUMASK_INIT,

>>   	TIME_INIT,

>>   	SYSINFO_INIT,

>> +	FDSERVER_INIT,

>>   	SHM_INIT,

>>   	THREAD_INIT,

>>   	POOL_INIT,

>> @@ -119,6 +120,9 @@ int odp_tm_term_global(void);

>>   int _odp_int_name_tbl_init_global(void);

>>   int _odp_int_name_tbl_term_global(void);

>>   

>> +int _odp_fdserver_init_global(void);

>> +int _odp_fdserver_term_global(void);

>> +

>>   int cpuinfo_parser(FILE *file, system_info_t *sysinfo);

>>   uint64_t odp_cpu_hz_current(int id);

>>   

>> diff --git a/platform/linux-generic/odp_init.c b/platform/linux-generic/odp_init.c

>> index f534759..ec84fd0 100644

>> --- a/platform/linux-generic/odp_init.c

>> +++ b/platform/linux-generic/odp_init.c

>> @@ -51,6 +51,12 @@ int odp_init_global(odp_instance_t *instance,

>>   	}

>>   	stage = SYSINFO_INIT;

>>   

>> +	if (_odp_fdserver_init_global()) {

>> +		ODP_ERR("ODP fdserver init failed.\n");

>> +		goto init_failed;

>> +	}

>> +	stage = FDSERVER_INIT;

>> +

>>   	if (odp_shm_init_global()) {

>>   		ODP_ERR("ODP shm init failed.\n");

>>   		goto init_failed;

>> @@ -218,6 +224,13 @@ int _odp_term_global(enum init_stage stage)

>>   		}

>>   		/* Fall through */

>>   

>> +	case FDSERVER_INIT:

>> +		if (_odp_fdserver_term_global()) {

>> +			ODP_ERR("ODP fdserver term failed.\n");

>> +			rc = -1;

>> +		}

>> +		/* Fall through */

>> +

>>   	case SYSINFO_INIT:

>>   		if (odp_system_info_term()) {

>>   			ODP_ERR("ODP system info term failed.\n");

>> -- 

>> 2.7.4

>>

Patch

diff --git a/platform/linux-generic/Makefile.am b/platform/linux-generic/Makefile.am
index 271f5b5..cb7c8e9 100644
--- a/platform/linux-generic/Makefile.am
+++ b/platform/linux-generic/Makefile.am
@@ -112,6 +112,7 @@  odpdrvplatinclude_HEADERS = \
 		  $(srcdir)/include/odp/drv/plat/strong_types.h
 
 noinst_HEADERS = \
+		  ${srcdir}/include/_fdserver_internal.h \
 		  ${srcdir}/include/odp_align_internal.h \
 		  ${srcdir}/include/odp_atomic_internal.h \
 		  ${srcdir}/include/odp_buffer_inlines.h \
@@ -154,6 +155,7 @@  noinst_HEADERS = \
 		  ${srcdir}/Makefile.inc
 
 __LIB__libodp_linux_la_SOURCES = \
+			   _fdserver.c \
 			   odp_atomic.c \
 			   odp_barrier.c \
 			   odp_buffer.c \
diff --git a/platform/linux-generic/_fdserver.c b/platform/linux-generic/_fdserver.c
new file mode 100644
index 0000000..bf36eb2
--- /dev/null
+++ b/platform/linux-generic/_fdserver.c
@@ -0,0 +1,655 @@ 
+/* Copyright (c) 2016, Linaro Limited
+ * All rights reserved.
+ *
+ * SPDX-License-Identifier:     BSD-3-Clause
+ */
+
+/*
+ * This file implements a file descriptor sharing server enabling
+ * sharing of file descriptors between processes, regardless of fork time.
+ *
+ * File descriptors are process scoped, but they can be "sent and converted
+ * on the fly" between processes using special unix domain socket ancillary
+ * data.
+ * The receiving process gets a file descriptor "pointing" to the same thing
+ * as the one sent (but the value of the file descriptor itself may be different
+ * from the one sent).
+ * Because ODP applications are responsible for creating ODP threads (i.e.
+ * pthreads or linux processes), ODP has no control on the order things happen:
+ * Nothing prevent a thread A to fork B and C, and then C creating a pktio
+ * which will be used by A and B to send/receive packets.
+ * Assuming this pktio uses a file descriptor, the latter will need to be
+ * shared between the processes, despite the "non convenient" fork time.
+ * The shared memory allocator is likely to use this as well to be able to
+ * share memory regardless of fork() time.
+ * This server handles a table of {(context,key)<-> fd} pair, and is
+ * interfaced by the following functions:
+ *
+ * _odp_fdserver_register_fd(context, key, fd_to_send);
+ * _odp_fdserver_deregister_fd(context, key);
+ * _odp_fdserver_lookup_fd(context, key);
+ *
+ * which are used to register/deregister or querry for file descriptor based
+ * on a context and key value couple, which has to be unique.
+ *
+ * Note again that the file descriptors stored here are local to this server
+ * process and get converted both when registered or looked up.
+ */
+
+#include <odp_posix_extensions.h>
+#include <odp/api/spinlock.h>
+#include <odp_internal.h>
+#include <odp_debug_internal.h>
+#include <_fdserver_internal.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <string.h>
+#include <sys/types.h>
+#include <signal.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <unistd.h>
+#include <inttypes.h>
+#include <sys/mman.h>
+#include <sys/wait.h>
+
+#define FDSERVER_SOCKPATH_MAXLEN 32
+#define FDSERVER_SOCKPATH_FORMAT "/tmp/odp-%d-fdserver"
+#define FDSERVER_BACKLOG 5
+
+#ifndef MAP_ANONYMOUS
+#define MAP_ANONYMOUS MAP_ANON
+#endif
+
+/* when accessing the client functions, clients should be mutexed: */
+odp_spinlock_t *client_lock;
+
+/* define the tables of file descriptors handled by this server: */
+#define FDSERVER_MAX_ENTRIES 256
+typedef struct fdentry_s {
+	fd_server_context_e context;
+	uint64_t key;
+	int  fd;
+} fdentry_t;
+static fdentry_t fd_table[FDSERVER_MAX_ENTRIES];
+static int fd_table_nb_entries;
+
+/*
+ * define the message struct used for communication between client and server
+ * (this single message is used in both direction)
+ * The file descriptors are sent out of band as ancillary data for conversion.
+ */
+typedef struct fd_server_msg {
+	int command;
+	fd_server_context_e context;
+	uint64_t key;
+} fdserver_msg_t;
+/* possible commands are: */
+#define FD_REGISTER_REQ		1  /* client -> server */
+#define FD_REGISTER_ACK		2  /* server -> client */
+#define FD_REGISTER_NACK	3  /* server -> client */
+#define FD_LOOKUP_REQ		4  /* client -> server */
+#define FD_LOOKUP_ACK		5  /* server -> client */
+#define FD_LOOKUP_NACK		6  /* server -> client */
+#define FD_DEREGISTER_REQ	7  /* client -> server */
+#define FD_DEREGISTER_ACK	8  /* server -> client */
+#define FD_DEREGISTER_NACK	9  /* server -> client */
+#define FD_SERVERSTOP_REQ	10 /* client -> server (stops) */
+
+/*
+ * Client and server function:
+ * Send a fdserver_msg, possibly including a file descriptor, on the socket
+ * This function is used both by:
+ * -the client (sending a FD_REGISTER_REQ with a file descriptor to be shared,
+ *  or FD_LOOKUP_REQ/FD_DEREGISTER_REQ without a file descirptor)
+ * -the server (sending FD_REGISTER_ACK/NACK, FD_LOOKUP_NACK,
+ *  FD_DEREGISTER_ACK/NACK... without a fd or a
+ *  FD_LOOKUP_ACK with a fd)
+ * This function make use of the ancillary data (control data) to pass and
+ * convert file descriptors over UNIX sockets
+ * Return -1 on error, 0 on success.
+ */
+static int send_fdserver_msg(int sock, int command,
+			     fd_server_context_e context, uint64_t key,
+			     int fd_to_send)
+{
+	struct msghdr socket_message;
+	struct iovec io_vector[1]; /* one msg frgmt only */
+	struct cmsghdr *control_message = NULL;
+	int *fd_location;
+	fdserver_msg_t msg;
+	int res;
+
+	char ancillary_data[CMSG_SPACE(sizeof(int))];
+
+	/* prepare the register request body (single framgent): */
+	msg.command = command;
+	msg.context = context;
+	msg.key = key;
+	io_vector[0].iov_base = &msg;
+	io_vector[0].iov_len = sizeof(fdserver_msg_t);
+
+	/* initialize socket message */
+	memset(&socket_message, 0, sizeof(struct msghdr));
+	socket_message.msg_iov = io_vector;
+	socket_message.msg_iovlen = 1;
+
+	if (fd_to_send >= 0) {
+		/* provide space for the ancillary data */
+		memset(ancillary_data, 0, CMSG_SPACE(sizeof(int)));
+		socket_message.msg_control = ancillary_data;
+		socket_message.msg_controllen = CMSG_SPACE(sizeof(int));
+
+		/* initialize a single ancillary data element for fd passing */
+		control_message = CMSG_FIRSTHDR(&socket_message);
+		control_message->cmsg_level = SOL_SOCKET;
+		control_message->cmsg_type = SCM_RIGHTS;
+		control_message->cmsg_len = CMSG_LEN(sizeof(int));
+		fd_location = (int *)(void *)CMSG_DATA(control_message);
+		*fd_location = fd_to_send;
+	}
+	res = sendmsg(sock, &socket_message, 0);
+	if (res < 0) {
+		ODP_ERR("send_fdserver_msg: %s\n", strerror(errno));
+		return(-1);
+	}
+
+	return 0;
+}
+
+/*
+ * Client and server function
+ * Receive a fdserver_msg, possibly including a file descriptor, on the
+ * given socket.
+ * This function is used both by:
+ * -the server (receiving a FD_REGISTER_REQ with a file descriptor to be shared,
+ *  or FD_LOOKUP_REQ, FD_DEREGISTER_REQ without a file descirptor)
+ * -the client (receiving FD_REGISTER_ACK...without a fd or a FD_LOOKUP_ACK with
+ * a fd)
+ * This function make use of the ancillary data (control data) to pass and
+ * convert file descriptors over UNIX sockets.
+ * Return -1 on error, 0 on success.
+ */
+static int recv_fdserver_msg(int sock, int *command,
+			     fd_server_context_e *context, uint64_t *key,
+			     int *recvd_fd)
+{
+	struct msghdr socket_message;
+	struct iovec io_vector[1]; /* one msg frgmt only */
+	struct cmsghdr *control_message = NULL;
+	int *fd_location;
+	fdserver_msg_t msg;
+	char ancillary_data[CMSG_SPACE(sizeof(int))];
+
+	memset(&socket_message, 0, sizeof(struct msghdr));
+	memset(ancillary_data, 0, CMSG_SPACE(sizeof(int)));
+
+	/* setup a place to fill in message contents */
+	io_vector[0].iov_base = &msg;
+	io_vector[0].iov_len = sizeof(fdserver_msg_t);
+	socket_message.msg_iov = io_vector;
+	socket_message.msg_iovlen = 1;
+
+	/* provide space for the ancillary data */
+	socket_message.msg_control = ancillary_data;
+	socket_message.msg_controllen = CMSG_SPACE(sizeof(int));
+
+	/* receive the message */
+	if (recvmsg(sock, &socket_message, MSG_CMSG_CLOEXEC) < 0) {
+		ODP_ERR("recv_fdserver_msg: %s\n", strerror(errno));
+		return(-1);
+	}
+
+	*command = msg.command;
+	*context = msg.context;
+	*key = msg.key;
+
+	/* grab the converted file descriptor (if any) */
+	*recvd_fd = -1;
+
+	if ((socket_message.msg_flags & MSG_CTRUNC) == MSG_CTRUNC)
+		return 0;
+
+	/* iterate ancillary elements to find the file descriptor: */
+	for (control_message = CMSG_FIRSTHDR(&socket_message);
+	     control_message != NULL;
+	     control_message = CMSG_NXTHDR(&socket_message, control_message)) {
+		if ((control_message->cmsg_level == SOL_SOCKET) &&
+		    (control_message->cmsg_type == SCM_RIGHTS)) {
+			fd_location = (int *)(void *)CMSG_DATA(control_message);
+			*recvd_fd = *fd_location;
+			break;
+		}
+	}
+
+	return 0;
+}
+
+/* opens and returns a connected socket to the server */
+static int get_socket(void)
+{
+	char sockpath[FDSERVER_SOCKPATH_MAXLEN];
+	int s_sock; /* server socket */
+	struct sockaddr_un remote;
+	int len;
+
+	/* construct the named socket path: */
+	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,
+		 odp_global_data.main_pid);
+
+	s_sock = socket(AF_UNIX, SOCK_STREAM, 0);
+	if (s_sock == -1) {
+		ODP_ERR("cannot connect to server: %s\n", strerror(errno));
+		return(-1);
+	}
+
+	remote.sun_family = AF_UNIX;
+	strcpy(remote.sun_path, sockpath);
+	len = strlen(remote.sun_path) + sizeof(remote.sun_family);
+	if (connect(s_sock, (struct sockaddr *)&remote, len) == -1) {
+		ODP_ERR("cannot connect to server: %s\n", strerror(errno));
+		close(s_sock);
+		return(-1);
+	}
+
+	return s_sock;
+}
+
+/*
+ * Client function:
+ * Register a file descriptor to the server. Return -1 on error.
+ */
+int _odp_fdserver_register_fd(fd_server_context_e context, uint64_t key,
+			      int fd_to_send)
+{
+	int s_sock; /* server socket */
+	int res;
+	int command;
+	int fd;
+
+	odp_spinlock_lock(client_lock);
+
+	ODP_DBG("FD client register: pid=%d key=%" PRIu64 ", fd=%d\n",
+		getpid(), key, fd_to_send);
+
+	s_sock = get_socket();
+	if (s_sock < 0) {
+		odp_spinlock_unlock(client_lock);
+		return(-1);
+	}
+
+	res =  send_fdserver_msg(s_sock, FD_REGISTER_REQ, context, key,
+				 fd_to_send);
+	if (res < 0) {
+		ODP_ERR("fd registration failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);
+
+	if ((res < 0) || (command != FD_REGISTER_ACK)) {
+		ODP_ERR("fd registration failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	close(s_sock);
+
+	odp_spinlock_unlock(client_lock);
+	return 0;
+}
+
+/*
+ * Client function:
+ * Deregister a file descriptor from the server. Return -1 on error.
+ */
+int _odp_fdserver_deregister_fd(fd_server_context_e context, uint64_t key)
+{
+	int s_sock; /* server socket */
+	int res;
+	int command;
+	int fd;
+
+	odp_spinlock_lock(client_lock);
+
+	ODP_DBG("FD client deregister: pid=%d key=%" PRIu64 "\n",
+		getpid(), key);
+
+	s_sock = get_socket();
+	if (s_sock < 0) {
+		odp_spinlock_unlock(client_lock);
+		return(-1);
+	}
+
+	res =  send_fdserver_msg(s_sock, FD_DEREGISTER_REQ, context, key, -1);
+	if (res < 0) {
+		ODP_ERR("fd de-registration failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);
+
+	if ((res < 0) || (command != FD_DEREGISTER_ACK)) {
+		ODP_ERR("fd de-registration failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	close(s_sock);
+
+	odp_spinlock_unlock(client_lock);
+	return 0;
+}
+
+/*
+ * client function:
+ * lookup a file descriptor from the server. return -1 on error,
+ * or the file descriptor on success (>=0).
+ */
+int _odp_fdserver_lookup_fd(fd_server_context_e context, uint64_t key)
+{
+	int s_sock; /* server socket */
+	int res;
+	int command;
+	int fd;
+
+	odp_spinlock_lock(client_lock);
+
+	s_sock = get_socket();
+	if (s_sock < 0) {
+		odp_spinlock_unlock(client_lock);
+		return(-1);
+	}
+
+	res =  send_fdserver_msg(s_sock, FD_LOOKUP_REQ, context, key, -1);
+	if (res < 0) {
+		ODP_ERR("fd lookup failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	res = recv_fdserver_msg(s_sock, &command, &context, &key, &fd);
+
+	if ((res < 0) || (command != FD_LOOKUP_ACK)) {
+		ODP_ERR("fd lookup failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	close(s_sock);
+	ODP_DBG("FD client lookup: pid=%d, key=%" PRIu64 ", fd=%d\n",
+		getpid(), key, fd);
+
+	odp_spinlock_unlock(client_lock);
+	return fd;
+}
+
+/*
+ * request server terminaison:
+ */
+static int stop_server(void)
+{
+	int s_sock; /* server socket */
+	int res;
+
+	odp_spinlock_lock(client_lock);
+
+	ODP_DBG("FD sending server stop request\n");
+
+	s_sock = get_socket();
+	if (s_sock < 0) {
+		odp_spinlock_unlock(client_lock);
+		return(-1);
+	}
+
+	res =  send_fdserver_msg(s_sock, FD_SERVERSTOP_REQ, 0, 0, -1);
+	if (res < 0) {
+		ODP_ERR("fd stop request failure\n");
+		close(s_sock);
+		odp_spinlock_unlock(client_lock);
+		return -1;
+	}
+
+	close(s_sock);
+
+	odp_spinlock_unlock(client_lock);
+	return 0;
+}
+
+/*
+ * server function
+ * receive a client request and handle it.
+ * Always returns 0 unless a stop request is received.
+ */
+static int handle_request(int client_sock)
+{
+	int command;
+	fd_server_context_e context;
+	uint64_t key;
+	int fd;
+	int i;
+
+	/* get a client request: */
+	recv_fdserver_msg(client_sock, &command, &context, &key, &fd);
+	switch (command) {
+	case FD_REGISTER_REQ:
+		if ((fd < 0) || (context >= FD_SRV_CTX_END)) {
+			ODP_ERR("Invalid register fd or context\n");
+			send_fdserver_msg(client_sock, FD_REGISTER_NACK,
+					  FD_SRV_CTX_NA, 0, -1);
+			return 0;
+		}
+
+		/* store the file descriptor in table: */
+		if (fd_table_nb_entries < FDSERVER_MAX_ENTRIES) {
+			fd_table[fd_table_nb_entries].context = context;
+			fd_table[fd_table_nb_entries].key     = key;
+			fd_table[fd_table_nb_entries++].fd    = fd;
+			ODP_DBG("storing {ctx=%d, key=%" PRIu64 "}->fd=%d\n",
+				context, key, fd);
+		} else {
+			ODP_ERR("FD table full\n");
+			send_fdserver_msg(client_sock, FD_REGISTER_NACK,
+					  FD_SRV_CTX_NA, 0, -1);
+			return 0;
+		}
+
+		send_fdserver_msg(client_sock, FD_REGISTER_ACK,
+				  FD_SRV_CTX_NA, 0, -1);
+		break;
+
+	case FD_LOOKUP_REQ:
+		if (context >= FD_SRV_CTX_END) {
+			ODP_ERR("invalid lookup context\n");
+			send_fdserver_msg(client_sock, FD_LOOKUP_NACK,
+					  FD_SRV_CTX_NA, 0, -1);
+			return 0;
+		}
+
+		/* search key in table and sent reply: */
+		for (i = 0; i < fd_table_nb_entries; i++) {
+			if ((fd_table[i].context == context) &&
+			    (fd_table[i].key == key)) {
+				fd = fd_table[i].fd;
+				ODP_DBG("lookup {ctx=%d,"
+					" key=%" PRIu64 "}->fd=%d\n",
+					context, key, fd);
+				send_fdserver_msg(client_sock,
+						  FD_LOOKUP_ACK, context, key,
+						  fd);
+				return 0;
+			}
+		}
+
+		/* context+key not found... send nack */
+		send_fdserver_msg(client_sock, FD_LOOKUP_NACK, context, key,
+				  -1);
+		break;
+
+	case FD_DEREGISTER_REQ:
+		if (context >= FD_SRV_CTX_END) {
+			ODP_ERR("invalid deregister context\n");
+			send_fdserver_msg(client_sock, FD_DEREGISTER_NACK,
+					  FD_SRV_CTX_NA, 0, -1);
+			return 0;
+		}
+
+		/* search key in table and remove it if found, and reply: */
+		for (i = 0; i < fd_table_nb_entries; i++) {
+			if ((fd_table[i].context == context) &&
+			    (fd_table[i].key == key)) {
+				ODP_DBG("drop {ctx=%d,"
+					" key=%" PRIu64 "}->fd=%d\n",
+					context, key, fd_table[i].fd);
+				close(fd_table[i].fd);
+				fd_table[i] = fd_table[--fd_table_nb_entries];
+				send_fdserver_msg(client_sock,
+						  FD_DEREGISTER_ACK,
+						  context, key, -1);
+				return 0;
+			}
+		}
+
+		/* key not found... send nack */
+		send_fdserver_msg(client_sock, FD_DEREGISTER_NACK,
+				  context, key, -1);
+		break;
+
+	case FD_SERVERSTOP_REQ:
+		ODP_DBG("Stoping FD server\n");
+		return 1;
+
+	default:
+		ODP_ERR("Unexpected request\n");
+		break;
+	}
+	return 0;
+}
+
+/*
+ * server function
+ * loop forever, handling client requests one by one
+ */
+static void wait_requests(int sock)
+{
+	int c_socket; /* client connection */
+	unsigned int addr_sz;
+	struct sockaddr_un remote;
+
+	for (;;) {
+		addr_sz = sizeof(remote);
+		c_socket = accept(sock, (struct sockaddr *)&remote, &addr_sz);
+		if (c_socket == -1) {
+				ODP_ERR("wait_requests: %s\n", strerror(errno));
+				return;
+		}
+
+		if (handle_request(c_socket))
+			break;
+		close(c_socket);
+	}
+	close(c_socket);
+}
+
+/*
+ * Create a unix domain socket and fork a process to listen to incoming
+ * requests.
+ */
+int _odp_fdserver_init_global(void)
+{
+	char sockpath[FDSERVER_SOCKPATH_MAXLEN];
+	int sock;
+	struct sockaddr_un local;
+	pid_t server_pid;
+	int res;
+
+	/* create the client spinlock that any client can see: */
+	client_lock = mmap(NULL, sizeof(odp_spinlock_t), PROT_READ | PROT_WRITE,
+			   MAP_SHARED | MAP_ANONYMOUS, -1, 0);
+
+	odp_spinlock_init(client_lock);
+
+	/* construct the server named socket path: */
+	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,
+		 odp_global_data.main_pid);
+
+	/* create UNIX domain socket: */
+	sock = socket(AF_UNIX, SOCK_STREAM, 0);
+	if (sock == -1) {
+		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));
+		return(-1);
+	}
+
+	/* remove previous named socket if it already exists: */
+	unlink(sockpath);
+
+	/* bind to new named socket: */
+	local.sun_family = AF_UNIX;
+	strncpy(local.sun_path, sockpath, sizeof(local.sun_path));
+	res = bind(sock, (struct sockaddr *)&local, sizeof(struct sockaddr_un));
+	if (res == -1) {
+		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));
+		close(sock);
+		return(-1);
+	}
+
+	/* listen for incoming conections: */
+	if (listen(sock, FDSERVER_BACKLOG) == -1) {
+		ODP_ERR("_odp_fdserver_init_global: %s\n", strerror(errno));
+		close(sock);
+		return(-1);
+	}
+
+	/* fork a server process: */
+	server_pid = fork();
+	if (server_pid == -1) {
+		ODP_ERR("Could not fork!\n");
+		close(sock);
+		return(-1);
+	}
+
+	if (server_pid == 0) { /*child */
+		/* TODO: pin the server on appropriate service cpu mask */
+		/* when (if) we can agree on the usage of service mask  */
+
+		wait_requests(sock); /* Returns when server is stopped  */
+		close(sock);
+		exit(0);
+	}
+
+	/* parent */
+	close(sock);
+	return 0;
+}
+
+/*
+ * Terminate the server
+ */
+int _odp_fdserver_term_global(void)
+{
+	int status;
+	char sockpath[FDSERVER_SOCKPATH_MAXLEN];
+
+	/* close the server and wait for child terminaison*/
+	stop_server();
+	wait(&status);
+
+	/* construct the server named socket path: */
+	snprintf(sockpath, FDSERVER_SOCKPATH_MAXLEN, FDSERVER_SOCKPATH_FORMAT,
+		 odp_global_data.main_pid);
+
+	/* delete the UNIX domain socket: */
+	unlink(sockpath);
+
+	return 0;
+}
diff --git a/platform/linux-generic/include/_fdserver_internal.h b/platform/linux-generic/include/_fdserver_internal.h
new file mode 100644
index 0000000..480ac02
--- /dev/null
+++ b/platform/linux-generic/include/_fdserver_internal.h
@@ -0,0 +1,38 @@ 
+/* Copyright (c) 2016, Linaro Limited
+ * All rights reserved.
+ *
+ * SPDX-License-Identifier:     BSD-3-Clause
+ */
+
+#ifndef _FD_SERVER_INTERNAL_H
+#define _FD_SERVER_INTERNAL_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * the following enum defines the different contextes by which the
+ * FD server may be used: In the FD server, the keys used to store/retrieve
+ * a file descriptor are actually context based:
+ * Both the context and the key are stored at fd registration time,
+ * and both the context and the key are used to retrieve a fd.
+ * In other words a context identifies a FD server usage, so that different
+ * unrelated fd server users do not have to guarantee key unicity between
+ * them.
+ */
+typedef enum fd_server_context {
+	FD_SRV_CTX_NA,  /* Not Applicable   */
+	FD_SRV_CTX_END, /* upper enum limit */
+} fd_server_context_e;
+
+int _odp_fdserver_register_fd(fd_server_context_e context, uint64_t key,
+			      int fd);
+int _odp_fdserver_deregister_fd(fd_server_context_e context, uint64_t key);
+int _odp_fdserver_lookup_fd(fd_server_context_e context, uint64_t key);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/platform/linux-generic/include/odp_internal.h b/platform/linux-generic/include/odp_internal.h
index 8bad450..fd770b5 100644
--- a/platform/linux-generic/include/odp_internal.h
+++ b/platform/linux-generic/include/odp_internal.h
@@ -54,6 +54,7 @@  enum init_stage {
 	CPUMASK_INIT,
 	TIME_INIT,
 	SYSINFO_INIT,
+	FDSERVER_INIT,
 	SHM_INIT,
 	THREAD_INIT,
 	POOL_INIT,
@@ -119,6 +120,9 @@  int odp_tm_term_global(void);
 int _odp_int_name_tbl_init_global(void);
 int _odp_int_name_tbl_term_global(void);
 
+int _odp_fdserver_init_global(void);
+int _odp_fdserver_term_global(void);
+
 int cpuinfo_parser(FILE *file, system_info_t *sysinfo);
 uint64_t odp_cpu_hz_current(int id);
 
diff --git a/platform/linux-generic/odp_init.c b/platform/linux-generic/odp_init.c
index f534759..ec84fd0 100644
--- a/platform/linux-generic/odp_init.c
+++ b/platform/linux-generic/odp_init.c
@@ -51,6 +51,12 @@  int odp_init_global(odp_instance_t *instance,
 	}
 	stage = SYSINFO_INIT;
 
+	if (_odp_fdserver_init_global()) {
+		ODP_ERR("ODP fdserver init failed.\n");
+		goto init_failed;
+	}
+	stage = FDSERVER_INIT;
+
 	if (odp_shm_init_global()) {
 		ODP_ERR("ODP shm init failed.\n");
 		goto init_failed;
@@ -218,6 +224,13 @@  int _odp_term_global(enum init_stage stage)
 		}
 		/* Fall through */
 
+	case FDSERVER_INIT:
+		if (_odp_fdserver_term_global()) {
+			ODP_ERR("ODP fdserver term failed.\n");
+			rc = -1;
+		}
+		/* Fall through */
+
 	case SYSINFO_INIT:
 		if (odp_system_info_term()) {
 			ODP_ERR("ODP system info term failed.\n");