eal: add synchronous multi-process communication
authorJianfeng Tan <jianfeng.tan@intel.com>
Tue, 30 Jan 2018 06:58:09 +0000 (06:58 +0000)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 30 Jan 2018 14:17:23 +0000 (15:17 +0100)
We need the synchronous way for multi-process communication,
i.e., blockingly waiting for reply message when we send a request
to the peer process.

We add two APIs rte_eal_mp_request() and rte_eal_mp_reply() for
such use case. By invoking rte_eal_mp_request(), a request message
is sent out, and then it waits there for a reply message. The caller
can specify the timeout. And the response messages will be collected
and returned so that the caller can decide how to translate them.

The API rte_eal_mp_reply() is always called by an mp action handler.
Here we add another parameter for rte_eal_mp_t so that the action
handler knows which peer address to reply.

       sender-process                receiver-process
   ----------------------            ----------------

    thread-n
     |_rte_eal_mp_request() ----------> mp-thread
        |_timedwait()                    |_process_msg()
                                           |_action()
                                               |_rte_eal_mp_reply()
        mp_thread  <---------------------|
                  |_process_msg()
                     |_signal(send_thread)
    thread-m <----------|
     |_collect-reply

 * A secondary process is only allowed to talk to the primary process.
 * If there are multiple secondary processes for the primary process,
   it will send request to peer1, collect response from peer1; then
   send request to peer2, collect response from peer2, and so on.
 * When thread-n is sending request, thread-m of that process can send
   request at the same time.
 * For pair <action_name, peer>, we guarantee that only one such request
   is on the fly.

Suggested-by: Anatoly Burakov <anatoly.burakov@intel.com>
Suggested-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Signed-off-by: Jianfeng Tan <jianfeng.tan@intel.com>
Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
doc/guides/rel_notes/release_18_02.rst
lib/librte_eal/common/eal_common_proc.c
lib/librte_eal/common/include/rte_eal.h
lib/librte_eal/rte_eal_version.map

index 4126c74..b7ee203 100644 (file)
@@ -175,6 +175,8 @@ New Features
 
   * ``rte_mp_register`` and ``rte_mp_unregister`` are for action (un)registration.
   * ``rte_mp_sendmsg`` is for sending a message without blocking for a response.
+  * ``rte_mp_request`` is for sending a request message and will block until
+    it gets a reply message which is sent from the peer by ``rte_mp_reply``.
 
 API Changes
 -----------
index f63c9c2..b974837 100644 (file)
@@ -13,6 +13,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/time.h>
 #include <sys/types.h>
 #include <sys/socket.h>
 #include <sys/un.h>
@@ -46,6 +47,50 @@ TAILQ_HEAD(action_entry_list, action_entry);
 static struct action_entry_list action_entry_list =
        TAILQ_HEAD_INITIALIZER(action_entry_list);
 
+enum mp_type {
+       MP_MSG, /* Share message with peers, will not block */
+       MP_REQ, /* Request for information, Will block for a reply */
+       MP_REP, /* Response to previously-received request */
+};
+
+struct mp_msg_internal {
+       int type;
+       struct rte_mp_msg msg;
+};
+
+struct sync_request {
+       TAILQ_ENTRY(sync_request) next;
+       int reply_received;
+       char dst[PATH_MAX];
+       struct rte_mp_msg *request;
+       struct rte_mp_msg *reply;
+       pthread_cond_t cond;
+};
+
+TAILQ_HEAD(sync_request_list, sync_request);
+
+static struct {
+       struct sync_request_list requests;
+       pthread_mutex_t lock;
+} sync_requests = {
+       .requests = TAILQ_HEAD_INITIALIZER(sync_requests.requests),
+       .lock = PTHREAD_MUTEX_INITIALIZER
+};
+
+static struct sync_request *
+find_sync_request(const char *dst, const char *act_name)
+{
+       struct sync_request *r;
+
+       TAILQ_FOREACH(r, &sync_requests.requests, next) {
+               if (!strcmp(r->dst, dst) &&
+                   !strcmp(r->request->name, act_name))
+                       break;
+       }
+
+       return r;
+}
+
 int
 rte_eal_primary_proc_alive(const char *config_file_path)
 {
@@ -149,19 +194,21 @@ rte_mp_action_unregister(const char *name)
 }
 
 static int
-read_msg(struct rte_mp_msg *msg)
+read_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 {
        int msglen;
        struct iovec iov;
        struct msghdr msgh;
-       char control[CMSG_SPACE(sizeof(msg->fds))];
+       char control[CMSG_SPACE(sizeof(m->msg.fds))];
        struct cmsghdr *cmsg;
-       int buflen = sizeof(*msg) - sizeof(msg->fds);
+       int buflen = sizeof(*m) - sizeof(m->msg.fds);
 
        memset(&msgh, 0, sizeof(msgh));
-       iov.iov_base = msg;
+       iov.iov_base = m;
        iov.iov_len  = buflen;
 
+       msgh.msg_name = s;
+       msgh.msg_namelen = sizeof(*s);
        msgh.msg_iov = &iov;
        msgh.msg_iovlen = 1;
        msgh.msg_control = control;
@@ -183,7 +230,7 @@ read_msg(struct rte_mp_msg *msg)
                cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
                if ((cmsg->cmsg_level == SOL_SOCKET) &&
                        (cmsg->cmsg_type == SCM_RIGHTS)) {
-                       memcpy(msg->fds, CMSG_DATA(cmsg), sizeof(msg->fds));
+                       memcpy(m->msg.fds, CMSG_DATA(cmsg), sizeof(m->msg.fds));
                        break;
                }
        }
@@ -192,12 +239,28 @@ read_msg(struct rte_mp_msg *msg)
 }
 
 static void
-process_msg(struct rte_mp_msg *msg)
+process_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 {
+       struct sync_request *sync_req;
        struct action_entry *entry;
+       struct rte_mp_msg *msg = &m->msg;
        rte_mp_t action = NULL;
 
        RTE_LOG(DEBUG, EAL, "msg: %s\n", msg->name);
+
+       if (m->type == MP_REP) {
+               pthread_mutex_lock(&sync_requests.lock);
+               sync_req = find_sync_request(s->sun_path, msg->name);
+               if (sync_req) {
+                       memcpy(sync_req->reply, msg, sizeof(*msg));
+                       sync_req->reply_received = 1;
+                       pthread_cond_signal(&sync_req->cond);
+               } else
+                       RTE_LOG(ERR, EAL, "Drop mp reply: %s\n", msg->name);
+               pthread_mutex_unlock(&sync_requests.lock);
+               return;
+       }
+
        pthread_mutex_lock(&mp_mutex_action);
        entry = find_action_entry_by_name(msg->name);
        if (entry != NULL)
@@ -206,18 +269,19 @@ process_msg(struct rte_mp_msg *msg)
 
        if (!action)
                RTE_LOG(ERR, EAL, "Cannot find action: %s\n", msg->name);
-       else if (action(msg) < 0)
+       else if (action(msg, s->sun_path) < 0)
                RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
 }
 
 static void *
 mp_handle(void *arg __rte_unused)
 {
-       struct rte_mp_msg msg;
+       struct mp_msg_internal msg;
+       struct sockaddr_un sa;
 
        while (1) {
-               if (read_msg(&msg) == 0)
-                       process_msg(&msg);
+               if (read_msg(&msg, &sa) == 0)
+                       process_msg(&msg, &sa);
        }
 
        return NULL;
@@ -336,16 +400,20 @@ rte_mp_channel_init(void)
  *
  */
 static int
-send_msg(const char *dst_path, struct rte_mp_msg *msg)
+send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 {
        int snd;
        struct iovec iov;
        struct msghdr msgh;
        struct cmsghdr *cmsg;
        struct sockaddr_un dst;
+       struct mp_msg_internal m;
        int fd_size = msg->num_fds * sizeof(int);
        char control[CMSG_SPACE(fd_size)];
 
+       m.type = type;
+       memcpy(&m.msg, msg, sizeof(*msg));
+
        memset(&dst, 0, sizeof(dst));
        dst.sun_family = AF_UNIX;
        snprintf(dst.sun_path, sizeof(dst.sun_path), "%s", dst_path);
@@ -353,8 +421,8 @@ send_msg(const char *dst_path, struct rte_mp_msg *msg)
        memset(&msgh, 0, sizeof(msgh));
        memset(control, 0, sizeof(control));
 
-       iov.iov_base = msg;
-       iov.iov_len = sizeof(*msg) - sizeof(msg->fds);
+       iov.iov_base = &m;
+       iov.iov_len = sizeof(m) - sizeof(msg->fds);
 
        msgh.msg_name = &dst;
        msgh.msg_namelen = sizeof(dst);
@@ -396,14 +464,17 @@ send_msg(const char *dst_path, struct rte_mp_msg *msg)
 }
 
 static int
-mp_send(struct rte_mp_msg *msg)
+mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 {
        int ret = 0;
        DIR *mp_dir;
        struct dirent *ent;
 
-       if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
-               if (send_msg(eal_mp_socket_path(), msg) < 0)
+       if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
+               peer = eal_mp_socket_path();
+
+       if (peer) {
+               if (send_msg(peer, msg, type) < 0)
                        return -1;
                else
                        return 0;
@@ -421,11 +492,11 @@ mp_send(struct rte_mp_msg *msg)
                if (fnmatch(mp_filter, ent->d_name, 0) != 0)
                        continue;
 
-               if (send_msg(ent->d_name, msg) < 0)
+               if (send_msg(ent->d_name, msg, type) < 0)
                        ret = -1;
        }
-       closedir(mp_dir);
 
+       closedir(mp_dir);
        return ret;
 }
 
@@ -464,5 +535,150 @@ rte_mp_sendmsg(struct rte_mp_msg *msg)
                return -1;
 
        RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
-       return mp_send(msg);
+       return mp_send(msg, NULL, MP_MSG);
+}
+
+static int
+mp_request_one(const char *dst, struct rte_mp_msg *req,
+              struct rte_mp_reply *reply, const struct timespec *ts)
+{
+       int ret;
+       struct timeval now;
+       struct rte_mp_msg msg, *tmp;
+       struct sync_request sync_req, *exist;
+
+       sync_req.reply_received = 0;
+       strcpy(sync_req.dst, dst);
+       sync_req.request = req;
+       sync_req.reply = &msg;
+       pthread_cond_init(&sync_req.cond, NULL);
+
+       pthread_mutex_lock(&sync_requests.lock);
+       exist = find_sync_request(dst, req->name);
+       if (!exist)
+               TAILQ_INSERT_TAIL(&sync_requests.requests, &sync_req, next);
+       pthread_mutex_unlock(&sync_requests.lock);
+       if (exist) {
+               RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
+               rte_errno = -EEXIST;
+               return -1;
+       }
+
+       ret = send_msg(dst, req, MP_REQ);
+       if (ret < 0) {
+               RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
+                       dst, req->name);
+               return -1;
+       } else if (ret == 0)
+               return 0;
+
+       reply->nb_sent++;
+
+       pthread_mutex_lock(&sync_requests.lock);
+       do {
+               pthread_cond_timedwait(&sync_req.cond, &sync_requests.lock, ts);
+               /* Check spurious wakeups */
+               if (sync_req.reply_received == 1)
+                       break;
+               /* Check if time is out */
+               if (gettimeofday(&now, NULL) < 0)
+                       break;
+               if (now.tv_sec < ts->tv_sec)
+                       break;
+               else if (now.tv_sec == ts->tv_sec &&
+                        now.tv_usec * 1000 < ts->tv_nsec)
+                       break;
+       } while (1);
+       /* We got the lock now */
+       TAILQ_REMOVE(&sync_requests.requests, &sync_req, next);
+       pthread_mutex_unlock(&sync_requests.lock);
+
+       if (sync_req.reply_received == 0) {
+               RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
+                       dst, req->name);
+               rte_errno = -ETIMEDOUT;
+               return -1;
+       }
+
+       tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
+       if (!tmp) {
+               RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
+                       dst, req->name);
+               rte_errno = -ENOMEM;
+               return -1;
+       }
+       memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
+       reply->msgs = tmp;
+       reply->nb_received++;
+       return 0;
+}
+
+int __rte_experimental
+rte_mp_request(struct rte_mp_msg *req, struct rte_mp_reply *reply,
+               const struct timespec *ts)
+{
+       int ret = 0;
+       DIR *mp_dir;
+       struct dirent *ent;
+       struct timeval now;
+       struct timespec end;
+
+       RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
+
+       if (check_input(req) == false)
+               return -1;
+       if (gettimeofday(&now, NULL) < 0) {
+               RTE_LOG(ERR, EAL, "Faile to get current time\n");
+               rte_errno = errno;
+               return -1;
+       }
+
+       end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
+       end.tv_sec = now.tv_sec + ts->tv_sec +
+                       (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
+
+       reply->nb_sent = 0;
+       reply->nb_received = 0;
+       reply->msgs = NULL;
+
+       /* for secondary process, send request to the primary process only */
+       if (rte_eal_process_type() == RTE_PROC_SECONDARY)
+               return mp_request_one(eal_mp_socket_path(), req, reply, &end);
+
+       /* for primary process, broadcast request, and collect reply 1 by 1 */
+       mp_dir = opendir(mp_dir_path);
+       if (!mp_dir) {
+               RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
+               rte_errno = errno;
+               return -1;
+       }
+
+       while ((ent = readdir(mp_dir))) {
+               if (fnmatch(mp_filter, ent->d_name, 0) != 0)
+                       continue;
+
+               if (mp_request_one(ent->d_name, req, reply, &end))
+                       ret = -1;
+       }
+
+       closedir(mp_dir);
+       return ret;
+}
+
+int __rte_experimental
+rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
+{
+
+       RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
+
+       if (check_input(msg) == false)
+               return -1;
+
+       if (peer == NULL) {
+               RTE_LOG(ERR, EAL, "peer is not specified\n");
+               rte_errno = -EINVAL;
+               return -1;
+       }
+
+       return mp_send(msg, peer, MP_REP);
 }
index 2d022c0..08c6637 100644 (file)
@@ -13,6 +13,7 @@
 
 #include <stdint.h>
 #include <sched.h>
+#include <time.h>
 
 #include <rte_config.h>
 #include <rte_compat.h>
@@ -214,13 +215,19 @@ struct rte_mp_msg {
        int fds[RTE_MP_MAX_FD_NUM];
 };
 
+struct rte_mp_reply {
+       int nb_sent;
+       int nb_received;
+       struct rte_mp_msg *msgs; /* caller to free */
+};
+
 /**
  * Action function typedef used by other components.
  *
  * As we create  socket channel for primary/secondary communication, use
  * this function typedef to register action for coming messages.
  */
-typedef int (*rte_mp_t)(const struct rte_mp_msg *msg);
+typedef int (*rte_mp_t)(const struct rte_mp_msg *msg, const void *peer);
 
 /**
  * @warning
@@ -281,6 +288,57 @@ rte_mp_action_unregister(const char *name);
 int __rte_experimental
 rte_mp_sendmsg(struct rte_mp_msg *msg);
 
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice
+ *
+ * Send a request to the peer process and expect a reply.
+ *
+ * This function sends a request message to the peer process, and will
+ * block until receiving reply message from the peer process.
+ *
+ * @note The caller is responsible to free reply->replies.
+ *
+ * @param req
+ *   The req argument contains the customized request message.
+ *
+ * @param reply
+ *   The reply argument will be for storing all the replied messages;
+ *   the caller is responsible for free reply->replies.
+ *
+ * @param ts
+ *   The ts argument specifies how long we can wait for the peer(s) to reply.
+ *
+ * @return
+ *  - On success, return 0.
+ *  - On failure, return -1, and the reason will be stored in rte_errno.
+ */
+int __rte_experimental
+rte_mp_request(struct rte_mp_msg *req, struct rte_mp_reply *reply,
+              const struct timespec *ts);
+
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice
+ *
+ * Send a reply to the peer process.
+ *
+ * This function will send a reply message in response to a request message
+ * received previously.
+ *
+ * @param msg
+ *   The msg argument contains the customized message.
+ *
+ * @param peer
+ *   The peer argument is the pointer to the peer socket path.
+ *
+ * @return
+ *  - On success, return 0.
+ *  - On failure, return -1, and the reason will be stored in rte_errno.
+ */
+int __rte_experimental
+rte_mp_reply(struct rte_mp_msg *msg, const char *peer);
+
 /**
  * Usage function typedef used by the application usage function.
  *
index 24deaef..4146907 100644 (file)
@@ -223,6 +223,8 @@ EXPERIMENTAL {
        rte_mp_action_register;
        rte_mp_action_unregister;
        rte_mp_sendmsg;
+       rte_mp_request;
+       rte_mp_reply;
        rte_service_attr_get;
        rte_service_attr_reset_all;
        rte_service_component_register;