eal: add IPC asynchronous request
authorAnatoly Burakov <anatoly.burakov@intel.com>
Sat, 31 Mar 2018 17:06:16 +0000 (18:06 +0100)
committerThomas Monjalon <thomas@monjalon.net>
Wed, 4 Apr 2018 21:47:59 +0000 (23:47 +0200)
This API is similar to the blocking API that is already present,
but reply will be received in a separate callback by the caller
(callback specified at the time of request, rather than registering
for it in advance).

Under the hood, we create a separate thread to deal with replies to
asynchronous requests, that will just wait to be notified by the
main thread, or woken up on a timer.

Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Jianfeng Tan <jianfeng.tan@intel.com>
lib/librte_eal/common/eal_common_proc.c
lib/librte_eal/common/include/rte_eal.h
lib/librte_eal/rte_eal_version.map

index b704f5a..f98622f 100644 (file)
@@ -26,6 +26,7 @@
 #include <rte_errno.h>
 #include <rte_lcore.h>
 #include <rte_log.h>
+#include <rte_tailq.h>
 
 #include "eal_private.h"
 #include "eal_filesystem.h"
@@ -60,13 +61,32 @@ struct mp_msg_internal {
        struct rte_mp_msg msg;
 };
 
+struct async_request_param {
+       rte_mp_async_reply_t clb;
+       struct rte_mp_reply user_reply;
+       struct timespec end;
+       int n_responses_processed;
+};
+
 struct pending_request {
        TAILQ_ENTRY(pending_request) next;
-       int reply_received;
+       enum {
+               REQUEST_TYPE_SYNC,
+               REQUEST_TYPE_ASYNC
+       } type;
        char dst[PATH_MAX];
        struct rte_mp_msg *request;
        struct rte_mp_msg *reply;
-       pthread_cond_t cond;
+       int reply_received;
+       RTE_STD_C11
+       union {
+               struct {
+                       struct async_request_param *param;
+               } async;
+               struct {
+                       pthread_cond_t cond;
+               } sync;
+       };
 };
 
 TAILQ_HEAD(pending_request_list, pending_request);
@@ -74,9 +94,12 @@ TAILQ_HEAD(pending_request_list, pending_request);
 static struct {
        struct pending_request_list requests;
        pthread_mutex_t lock;
+       pthread_cond_t async_cond;
 } pending_requests = {
        .requests = TAILQ_HEAD_INITIALIZER(pending_requests.requests),
-       .lock = PTHREAD_MUTEX_INITIALIZER
+       .lock = PTHREAD_MUTEX_INITIALIZER,
+       .async_cond = PTHREAD_COND_INITIALIZER
+       /**< used in async requests only */
 };
 
 /* forward declarations */
@@ -273,7 +296,12 @@ process_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
                        memcpy(sync_req->reply, msg, sizeof(*msg));
                        /* -1 indicates that we've been asked to ignore */
                        sync_req->reply_received = m->type == MP_REP ? 1 : -1;
-                       pthread_cond_signal(&sync_req->cond);
+
+                       if (sync_req->type == REQUEST_TYPE_SYNC)
+                               pthread_cond_signal(&sync_req->sync.cond);
+                       else if (sync_req->type == REQUEST_TYPE_ASYNC)
+                               pthread_cond_signal(
+                                       &pending_requests.async_cond);
                } else
                        RTE_LOG(ERR, EAL, "Drop mp reply: %s\n", msg->name);
                pthread_mutex_unlock(&pending_requests.lock);
@@ -319,6 +347,189 @@ mp_handle(void *arg __rte_unused)
        return NULL;
 }
 
+static int
+timespec_cmp(const struct timespec *a, const struct timespec *b)
+{
+       if (a->tv_sec < b->tv_sec)
+               return -1;
+       if (a->tv_sec > b->tv_sec)
+               return 1;
+       if (a->tv_nsec < b->tv_nsec)
+               return -1;
+       if (a->tv_nsec > b->tv_nsec)
+               return 1;
+       return 0;
+}
+
+enum async_action {
+       ACTION_NONE, /**< don't do anything */
+       ACTION_FREE, /**< free the action entry, but don't trigger callback */
+       ACTION_TRIGGER /**< trigger callback, then free action entry */
+};
+
+static enum async_action
+process_async_request(struct pending_request *sr, const struct timespec *now)
+{
+       struct async_request_param *param;
+       struct rte_mp_reply *reply;
+       bool timeout, received, last_msg;
+
+       param = sr->async.param;
+       reply = &param->user_reply;
+
+       /* did we timeout? */
+       timeout = timespec_cmp(&param->end, now) <= 0;
+
+       /* did we receive a response? */
+       received = sr->reply_received != 0;
+
+       /* if we didn't time out, and we didn't receive a response, ignore */
+       if (!timeout && !received)
+               return ACTION_NONE;
+
+       /* if we received a response, adjust relevant data and copy mesasge. */
+       if (sr->reply_received == 1 && sr->reply) {
+               struct rte_mp_msg *msg, *user_msgs, *tmp;
+
+               msg = sr->reply;
+               user_msgs = reply->msgs;
+
+               tmp = realloc(user_msgs, sizeof(*msg) *
+                               (reply->nb_received + 1));
+               if (!tmp) {
+                       RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
+                               sr->dst, sr->request->name);
+                       /* this entry is going to be removed and its message
+                        * dropped, but we don't want to leak memory, so
+                        * continue.
+                        */
+               } else {
+                       user_msgs = tmp;
+                       reply->msgs = user_msgs;
+                       memcpy(&user_msgs[reply->nb_received],
+                                       msg, sizeof(*msg));
+                       reply->nb_received++;
+               }
+
+               /* mark this request as processed */
+               param->n_responses_processed++;
+       } else if (sr->reply_received == -1) {
+               /* we were asked to ignore this process */
+               reply->nb_sent--;
+       }
+       free(sr->reply);
+
+       last_msg = param->n_responses_processed == reply->nb_sent;
+
+       return last_msg ? ACTION_TRIGGER : ACTION_FREE;
+}
+
+static void
+trigger_async_action(struct pending_request *sr)
+{
+       struct async_request_param *param;
+       struct rte_mp_reply *reply;
+
+       param = sr->async.param;
+       reply = &param->user_reply;
+
+       param->clb(sr->request, reply);
+
+       /* clean up */
+       free(sr->async.param->user_reply.msgs);
+       free(sr->async.param);
+       free(sr->request);
+}
+
+static void *
+async_reply_handle(void *arg __rte_unused)
+{
+       struct pending_request *sr;
+       struct timeval now;
+       struct timespec timeout, ts_now;
+       while (1) {
+               struct pending_request *trigger = NULL;
+               int ret;
+               bool nowait = false;
+               bool timedwait = false;
+
+               pthread_mutex_lock(&pending_requests.lock);
+
+               /* scan through the list and see if there are any timeouts that
+                * are earlier than our current timeout.
+                */
+               TAILQ_FOREACH(sr, &pending_requests.requests, next) {
+                       if (sr->type != REQUEST_TYPE_ASYNC)
+                               continue;
+                       if (!timedwait || timespec_cmp(&sr->async.param->end,
+                                       &timeout) < 0) {
+                               memcpy(&timeout, &sr->async.param->end,
+                                       sizeof(timeout));
+                               timedwait = true;
+                       }
+
+                       /* sometimes, we don't even wait */
+                       if (sr->reply_received) {
+                               nowait = true;
+                               break;
+                       }
+               }
+
+               if (nowait)
+                       ret = 0;
+               else if (timedwait)
+                       ret = pthread_cond_timedwait(
+                                       &pending_requests.async_cond,
+                                       &pending_requests.lock, &timeout);
+               else
+                       ret = pthread_cond_wait(&pending_requests.async_cond,
+                                       &pending_requests.lock);
+
+               if (gettimeofday(&now, NULL) < 0) {
+                       RTE_LOG(ERR, EAL, "Cannot get current time\n");
+                       break;
+               }
+               ts_now.tv_nsec = now.tv_usec * 1000;
+               ts_now.tv_sec = now.tv_sec;
+
+               if (ret == 0 || ret == ETIMEDOUT) {
+                       struct pending_request *next;
+                       /* we've either been woken up, or we timed out */
+
+                       /* we have still the lock, check if anything needs
+                        * processing.
+                        */
+                       TAILQ_FOREACH_SAFE(sr, &pending_requests.requests, next,
+                                       next) {
+                               enum async_action action;
+                               if (sr->type != REQUEST_TYPE_ASYNC)
+                                       continue;
+
+                               action = process_async_request(sr, &ts_now);
+                               if (action == ACTION_FREE) {
+                                       TAILQ_REMOVE(&pending_requests.requests,
+                                                       sr, next);
+                                       free(sr);
+                               } else if (action == ACTION_TRIGGER &&
+                                               trigger == NULL) {
+                                       TAILQ_REMOVE(&pending_requests.requests,
+                                                       sr, next);
+                                       trigger = sr;
+                               }
+                       }
+               }
+               pthread_mutex_unlock(&pending_requests.lock);
+               if (trigger) {
+                       trigger_async_action(trigger);
+                       free(trigger);
+               }
+       };
+
+       RTE_LOG(ERR, EAL, "ERROR: asynchronous requests disabled\n");
+
+       return NULL;
+}
+
 static int
 open_socket_fd(void)
 {
@@ -382,7 +593,7 @@ rte_mp_channel_init(void)
        char thread_name[RTE_MAX_THREAD_NAME_LEN];
        char path[PATH_MAX];
        int dir_fd;
-       pthread_t tid;
+       pthread_t mp_handle_tid, async_reply_handle_tid;
 
        /* create filter path */
        create_socket_path("*", path, sizeof(path));
@@ -419,7 +630,16 @@ rte_mp_channel_init(void)
                return -1;
        }
 
-       if (pthread_create(&tid, NULL, mp_handle, NULL) < 0) {
+       if (pthread_create(&mp_handle_tid, NULL, mp_handle, NULL) < 0) {
+               RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
+                       strerror(errno));
+               close(mp_fd);
+               mp_fd = -1;
+               return -1;
+       }
+
+       if (pthread_create(&async_reply_handle_tid, NULL,
+                       async_reply_handle, NULL) < 0) {
                RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
                        strerror(errno));
                close(mp_fd);
@@ -430,7 +650,11 @@ rte_mp_channel_init(void)
 
        /* try best to set thread name */
        snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "rte_mp_handle");
-       rte_thread_setname(tid, thread_name);
+       rte_thread_setname(mp_handle_tid, thread_name);
+
+       /* try best to set thread name */
+       snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "rte_mp_async_handle");
+       rte_thread_setname(async_reply_handle_tid, thread_name);
 
        /* unlock the directory */
        flock(dir_fd, LOCK_UN);
@@ -602,18 +826,77 @@ rte_mp_sendmsg(struct rte_mp_msg *msg)
 }
 
 static int
-mp_request_one(const char *dst, struct rte_mp_msg *req,
+mp_request_async(const char *dst, struct rte_mp_msg *req,
+               struct async_request_param *param)
+{
+       struct rte_mp_msg *reply_msg;
+       struct pending_request *sync_req, *exist;
+       int ret;
+
+       sync_req = malloc(sizeof(*sync_req));
+       reply_msg = malloc(sizeof(*reply_msg));
+       if (sync_req == NULL || reply_msg == NULL) {
+               RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
+               rte_errno = ENOMEM;
+               ret = -1;
+               goto fail;
+       }
+
+       memset(sync_req, 0, sizeof(*sync_req));
+       memset(reply_msg, 0, sizeof(*reply_msg));
+
+       sync_req->type = REQUEST_TYPE_ASYNC;
+       strcpy(sync_req->dst, dst);
+       sync_req->request = req;
+       sync_req->reply = reply_msg;
+       sync_req->async.param = param;
+
+       /* queue already locked by caller */
+
+       exist = find_sync_request(dst, req->name);
+       if (!exist) {
+               TAILQ_INSERT_TAIL(&pending_requests.requests, sync_req, next);
+       } else {
+               RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
+               rte_errno = EEXIST;
+               ret = -1;
+               goto fail;
+       }
+
+       ret = send_msg(dst, req, MP_REQ);
+       if (ret < 0) {
+               RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
+                       dst, req->name);
+               ret = -1;
+               goto fail;
+       } else if (ret == 0) {
+               ret = 0;
+               goto fail;
+       }
+
+       param->user_reply.nb_sent++;
+
+       return 0;
+fail:
+       free(sync_req);
+       free(reply_msg);
+       return ret;
+}
+
+static int
+mp_request_sync(const char *dst, struct rte_mp_msg *req,
               struct rte_mp_reply *reply, const struct timespec *ts)
 {
        int ret;
        struct rte_mp_msg msg, *tmp;
        struct pending_request sync_req, *exist;
 
+       sync_req.type = REQUEST_TYPE_SYNC;
        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_cond_init(&sync_req.sync.cond, NULL);
 
        pthread_mutex_lock(&pending_requests.lock);
        exist = find_sync_request(dst, req->name);
@@ -637,7 +920,7 @@ mp_request_one(const char *dst, struct rte_mp_msg *req,
        reply->nb_sent++;
 
        do {
-               ret = pthread_cond_timedwait(&sync_req.cond,
+               ret = pthread_cond_timedwait(&sync_req.sync.cond,
                                &pending_requests.lock, ts);
        } while (ret != 0 && ret != ETIMEDOUT);
 
@@ -703,7 +986,7 @@ rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 
        /* 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);
+               return mp_request_sync(eal_mp_socket_path(), req, reply, &end);
 
        /* for primary process, broadcast request, and collect reply 1 by 1 */
        mp_dir = opendir(mp_dir_path);
@@ -732,7 +1015,7 @@ rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
                snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
                         ent->d_name);
 
-               if (mp_request_one(path, req, reply, &end))
+               if (mp_request_sync(path, req, reply, &end))
                        ret = -1;
        }
        /* unlock the directory */
@@ -744,9 +1027,158 @@ rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 }
 
 int __rte_experimental
-rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
+rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
+               rte_mp_async_reply_t clb)
 {
+       struct rte_mp_msg *copy;
+       struct pending_request *dummy;
+       struct async_request_param *param;
+       struct rte_mp_reply *reply;
+       int dir_fd, ret = 0;
+       DIR *mp_dir;
+       struct dirent *ent;
+       struct timeval now;
+       struct timespec *end;
+       bool dummy_used = false;
+
+       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;
+       }
+       copy = malloc(sizeof(*copy));
+       dummy = malloc(sizeof(*dummy));
+       param = malloc(sizeof(*param));
+       if (copy == NULL || dummy == NULL || param == NULL) {
+               RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
+               rte_errno = ENOMEM;
+               goto fail;
+       }
+
+       memset(copy, 0, sizeof(*copy));
+       memset(dummy, 0, sizeof(*dummy));
+       memset(param, 0, sizeof(*param));
+
+       /* copy message */
+       memcpy(copy, req, sizeof(*copy));
+
+       param->n_responses_processed = 0;
+       param->clb = clb;
+       end = &param->end;
+       reply = &param->user_reply;
+
+       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;
+
+       /* we have to lock the request queue here, as we will be adding a bunch
+        * of requests to the queue at once, and some of the replies may arrive
+        * before we add all of the requests to the queue.
+        */
+       pthread_mutex_lock(&pending_requests.lock);
 
+       /* we have to ensure that callback gets triggered even if we don't send
+        * anything, therefore earlier we have allocated a dummy request. fill
+        * it, and put it on the queue if we don't send any requests.
+        */
+       dummy->type = REQUEST_TYPE_ASYNC;
+       dummy->request = copy;
+       dummy->reply = NULL;
+       dummy->async.param = param;
+       dummy->reply_received = 1; /* short-circuit the timeout */
+
+       /* for secondary process, send request to the primary process only */
+       if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
+               ret = mp_request_async(eal_mp_socket_path(), copy, param);
+
+               /* if we didn't send anything, put dummy request on the queue */
+               if (ret == 0 && reply->nb_sent == 0) {
+                       TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
+                                       next);
+                       dummy_used = true;
+               }
+
+               pthread_mutex_unlock(&pending_requests.lock);
+
+               /* if we couldn't send anything, clean up */
+               if (ret != 0)
+                       goto fail;
+               return 0;
+       }
+
+       /* for primary process, broadcast request */
+       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;
+               goto unlock_fail;
+       }
+       dir_fd = dirfd(mp_dir);
+
+       /* lock the directory to prevent processes spinning up while we send */
+       if (flock(dir_fd, LOCK_EX)) {
+               RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
+                       mp_dir_path);
+               rte_errno = errno;
+               goto closedir_fail;
+       }
+
+       while ((ent = readdir(mp_dir))) {
+               char path[PATH_MAX];
+
+               if (fnmatch(mp_filter, ent->d_name, 0) != 0)
+                       continue;
+
+               snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
+                        ent->d_name);
+
+               if (mp_request_async(path, copy, param))
+                       ret = -1;
+       }
+       /* if we didn't send anything, put dummy request on the queue */
+       if (ret == 0 && reply->nb_sent == 0) {
+               TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
+               dummy_used = true;
+       }
+
+       /* trigger async request thread wake up */
+       pthread_cond_signal(&pending_requests.async_cond);
+
+       /* finally, unlock the queue */
+       pthread_mutex_unlock(&pending_requests.lock);
+
+       /* unlock the directory */
+       flock(dir_fd, LOCK_UN);
+
+       /* dir_fd automatically closed on closedir */
+       closedir(mp_dir);
+
+       /* if dummy was unused, free it */
+       if (!dummy_used)
+               free(dummy);
+
+       return ret;
+closedir_fail:
+       closedir(mp_dir);
+unlock_fail:
+       pthread_mutex_unlock(&pending_requests.lock);
+fail:
+       free(dummy);
+       free(param);
+       free(copy);
+       return -1;
+}
+
+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)
index d1cc89e..b804ff5 100644 (file)
@@ -229,6 +229,16 @@ struct rte_mp_reply {
  */
 typedef int (*rte_mp_t)(const struct rte_mp_msg *msg, const void *peer);
 
+/**
+ * Asynchronous reply function typedef used by other components.
+ *
+ * As we create socket channel for primary/secondary communication, use
+ * this function typedef to register action for coming responses to asynchronous
+ * requests.
+ */
+typedef int (*rte_mp_async_reply_t)(const struct rte_mp_msg *request,
+               const struct rte_mp_reply *reply);
+
 /**
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice
@@ -317,6 +327,32 @@ int __rte_experimental
 rte_mp_request_sync(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 request to the peer process and expect a reply in a separate callback.
+ *
+ * This function sends a request message to the peer process, and will not
+ * block. Instead, reply will be received in a separate callback.
+ *
+ * @param req
+ *   The req argument contains the customized request message.
+ *
+ * @param ts
+ *   The ts argument specifies how long we can wait for the peer(s) to reply.
+ *
+ * @param clb
+ *   The callback to trigger when all responses for this request have arrived.
+ *
+ * @return
+ *  - On success, return 0.
+ *  - On failure, return -1, and the reason will be stored in rte_errno.
+ */
+int __rte_experimental
+rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
+               rte_mp_async_reply_t clb);
+
 /**
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice
index 1715468..8299946 100644 (file)
@@ -226,6 +226,7 @@ EXPERIMENTAL {
        rte_mp_action_unregister;
        rte_mp_reply;
        rte_mp_request_sync;
+       rte_mp_request_async;
        rte_mp_sendmsg;
        rte_service_attr_get;
        rte_service_attr_reset_all;