main-afb-daemon: Export API after initialization
[src/app-framework-binder.git] / src / jobs.c
index f7acebf..9768f26 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2016, 2017 "IoT.bzh"
+ * Copyright (C) 2016, 2017, 2018 "IoT.bzh"
  * Author José Bollo <jose.bollo@iot.bzh>
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
 
 #define _GNU_SOURCE
 
+#if defined(NO_JOBS_WATCHDOG)
+#   define HAS_WATCHDOG 0
+#else
+#   define HAS_WATCHDOG 1
+#endif
+
 #include <stdlib.h>
+#include <stdint.h>
 #include <unistd.h>
 #include <signal.h>
+#include <string.h>
 #include <time.h>
 #include <sys/syscall.h>
 #include <pthread.h>
 #include <errno.h>
 #include <assert.h>
+#include <sys/eventfd.h>
+
+#include <systemd/sd-event.h>
+#include "fdev.h"
+#if HAS_WATCHDOG
+#include <systemd/sd-daemon.h>
+#endif
 
 #include "jobs.h"
 #include "sig-monitor.h"
 #include "verbose.h"
 
-/* describes pending job */
+#if defined(REMOVE_SYSTEMD_EVENT)
+#include "fdev-epoll.h"
+#endif
+
+#define EVENT_TIMEOUT_TOP      ((uint64_t)-1)
+#define EVENT_TIMEOUT_CHILD    ((uint64_t)10000)
+
+/** Internal shortcut for callback */
+typedef void (*job_cb_t)(int, void*);
+
+/** Description of a pending job */
 struct job
 {
-       struct job *next;   /* link to the next job enqueued */
-       void *group;        /* group of the request */
-       void (*callback)(int,void*,void*,void*);     /* processing callback */
-       void *arg1;         /* first arg */
-       void *arg2;         /* second arg */
-       void *arg3;         /* second arg */
-       int timeout;        /* timeout in second for processing the request */
-       int blocked;        /* is an other request blocking this one ? */
+       struct job *next;    /**< link to the next job enqueued */
+       const void *group;   /**< group of the request */
+       job_cb_t callback;   /**< processing callback */
+       void *arg;           /**< argument */
+       int timeout;         /**< timeout in second for processing the request */
+       unsigned blocked: 1; /**< is an other request blocking this one ? */
+       unsigned dropped: 1; /**< is removed ? */
 };
 
-/** control of threads */
+/** Description of handled event loops */
+struct evloop
+{
+       unsigned state;        /**< encoded state */
+       int efd;               /**< event notification */
+       struct sd_event *sdev; /**< the systemd event loop */
+       pthread_cond_t  cond;  /**< condition */
+       struct fdev *fdev;     /**< handling of events */
+};
+
+#define EVLOOP_STATE_WAIT           1U
+#define EVLOOP_STATE_RUN            2U
+#define EVLOOP_STATE_LOCK           4U
+
+/** Description of threads */
 struct thread
 {
-       struct thread *next;  /**< next thread of the list */
-       struct thread *upper; /**< upper same thread */
-       struct job *job;      /**< currently processed job */
-       pthread_t tid;        /**< the thread id */
-       unsigned stop: 1;     /**< stop requested */
-       unsigned lowered: 1;  /**< has a lower same thread */
+       struct thread *next;   /**< next thread of the list */
+       struct thread *upper;  /**< upper same thread */
+       struct job *job;       /**< currently processed job */
+       pthread_t tid;         /**< the thread id */
+       volatile unsigned stop: 1;      /**< stop requested */
+       volatile unsigned waits: 1;     /**< is waiting? */
 };
 
+/**
+ * Description of synchonous callback
+ */
+struct sync
+{
+       struct thread thread;   /**< thread loop data */
+       union {
+               void (*callback)(int, void*);   /**< the synchronous callback */
+               void (*enter)(int signum, void *closure, struct jobloop *jobloop);
+                               /**< the entering synchronous routine */
+       };
+       void *arg;              /**< the argument of the callback */
+};
+
+
 /* synchronisation of threads */
 static pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
 static pthread_cond_t  cond = PTHREAD_COND_INITIALIZER;
@@ -62,34 +115,38 @@ static pthread_cond_t  cond = PTHREAD_COND_INITIALIZER;
 static int allowed = 0; /** allowed count of threads */
 static int started = 0; /** started count of threads */
 static int running = 0; /** running count of threads */
-static int remains = 0; /** remaining count of jobs that can be created */
+static int remains = 0; /** allowed count of waiting jobs */
 
 /* list of threads */
 static struct thread *threads;
-static _Thread_local struct thread *current;
+static _Thread_local struct thread *current_thread;
+static _Thread_local struct evloop *current_evloop;
 
 /* queue of pending jobs */
 static struct job *first_job;
-static struct job *first_evloop;
 static struct job *free_jobs;
 
+/* event loop */
+static struct evloop evloop[1];
+
+#if defined(REMOVE_SYSTEMD_EVENT)
+static struct fdev_epoll *fdevepoll;
+static int waitevt;
+#endif
+
 /**
  * Create a new job with the given parameters
- * @param group the group of the job
- * @param timeout the timeout of the job (0 if none)
+ * @param group    the group of the job
+ * @param timeout  the timeout of the job (0 if none)
  * @param callback the function that achieves the job
- * @param arg1 the first argument of the callback
- * @param arg2 the second argument of the callback
- * @param arg3 the third argument of the callback
+ * @param arg      the argument of the callback
  * @return the created job unblock or NULL when no more memory
  */
 static struct job *job_create(
-               void *group,
+               const void *group,
                int timeout,
-               void (*callback)(int, void*, void *, void*),
-               void *arg1,
-               void *arg2,
-               void *arg3)
+               job_cb_t callback,
+               void *arg)
 {
        struct job *job;
 
@@ -98,12 +155,12 @@ static struct job *job_create(
        if (job)
                free_jobs = job->next;
        else {
-               /* allocation  without blocking */
+               /* allocation without blocking */
                pthread_mutex_unlock(&mutex);
                job = malloc(sizeof *job);
                pthread_mutex_lock(&mutex);
                if (!job) {
-                       errno = -ENOMEM;
+                       errno = ENOMEM;
                        goto end;
                }
        }
@@ -111,80 +168,63 @@ static struct job *job_create(
        job->group = group;
        job->timeout = timeout;
        job->callback = callback;
-       job->arg1 = arg1;
-       job->arg2 = arg2;
-       job->arg3 = arg3;
+       job->arg = arg;
        job->blocked = 0;
+       job->dropped = 0;
 end:
        return job;
 }
 
 /**
- * Adds 'job1' and 'job2' at the end of the list of jobs, marking it
+ * Adds 'job' at the end of the list of jobs, marking it
  * as blocked if an other job with the same group is pending.
- * @param job1 the first job to add
- * @param job2 the second job to add or NULL
+ * @param job the job to add
  */
-static void job_add2(struct job *job1, struct job *job2)
+static void job_add(struct job *job)
 {
-       void *group1, *group2, *group;
+       const void *group;
        struct job *ijob, **pjob;
 
        /* prepare to add */
-       group1 = job1->group;
-       job1->next = job2;
-       if (!job2)
-               group2 = NULL;
-       else {
-               job2->next = NULL;
-               group2 = job2->group;
-               if (group2 && group2 == group1)
-                       job2->blocked = 1;
-       }
+       group = job->group;
+       job->next = NULL;
 
-       /* search end and blackers */
+       /* search end and blockers */
        pjob = &first_job;
        ijob = first_job;
        while (ijob) {
-               group = ijob->group;
-               if (group) {
-                       if (group == group1)
-                               job1->blocked = 1;
-                       if (group == group2)
-                               job2->blocked = 1;
-               }
+               if (group && ijob->group == group)
+                       job->blocked = 1;
                pjob = &ijob->next;
                ijob = ijob->next;
        }
 
        /* queue the jobs */
-       *pjob = job1;
+       *pjob = job;
 }
 
 /**
  * Get the next job to process or NULL if none.
- * The returned job if any isn't removed from
- * the list of jobs.
- * @return the job to process
+ * @return the first job that isn't blocked or NULL
  */
 static inline struct job *job_get()
 {
-       struct job *job;
-
-       job = first_job;
+       struct job *job = first_job;
        while (job && job->blocked)
                job = job->next;
        return job;
 }
 
 /**
- * Releases the processed 'job'
+ * Releases the processed 'job': removes it
+ * from the list of jobs and unblock the first
+ * pending job of the same group if any.
  * @param job the job to release
  */
 static inline void job_release(struct job *job)
 {
        struct job *ijob, **pjob;
-       void *group;
+       const void *group;
 
        /* first unqueue the job */
        pjob = &first_job;
@@ -210,170 +250,272 @@ static inline void job_release(struct job *job)
        free_jobs = job;
 }
 
-/** monitored call to the job */
-static void job_call(int signum, void *arg)
+/**
+ * Monitored cancel callback for a job.
+ * This function is called by the monitor
+ * to cancel the job when the safe environment
+ * is set.
+ * @param signum 0 on normal flow or the number
+ *               of the signal that interrupted the normal
+ *               flow, isn't used
+ * @param arg    the job to run
+ */
+static void job_cancel(int signum, void *arg)
 {
        struct job *job = arg;
-       job->callback(signum, job->arg1, job->arg2, job->arg3);
+       job->callback(SIGABRT, job->arg);
 }
 
-/** monitored cancel of the job */
-static void job_cancel(int signum, void *arg)
+#if defined(REMOVE_SYSTEMD_EVENT)
+/**
+ * Gets a fdev_epoll item.
+ * @return a fdev_epoll or NULL in case of error
+ */
+static struct fdev_epoll *get_fdevepoll()
+{
+       struct fdev_epoll *result;
+
+       result = fdevepoll;
+       if (!result)
+               result = fdevepoll = fdev_epoll_create();
+
+       return result;
+}
+#endif
+
+/**
+ * Monitored normal callback for events.
+ * This function is called by the monitor
+ * to run the event loop when the safe environment
+ * is set.
+ * @param signum 0 on normal flow or the number
+ *               of the signal that interrupted the normal
+ *               flow
+ * @param arg     the events to run
+ */
+static void evloop_run(int signum, void *arg)
 {
-       job_call(SIGABRT, arg);
+       int rc;
+       struct sd_event *se;
+       struct evloop *el = arg;
+
+       if (!signum) {
+               current_evloop = el;
+               __atomic_store_n(&el->state, EVLOOP_STATE_LOCK|EVLOOP_STATE_RUN|EVLOOP_STATE_WAIT, __ATOMIC_RELAXED);
+               se = el->sdev;
+               rc = sd_event_prepare(se);
+               if (rc < 0) {
+                       errno = -rc;
+                       CRITICAL("sd_event_prepare returned an error (state: %d): %m", sd_event_get_state(se));
+                       abort();
+               } else {
+                       if (rc == 0) {
+                               rc = sd_event_wait(se, (uint64_t)(int64_t)-1);
+                               if (rc < 0) {
+                                       errno = -rc;
+                                       ERROR("sd_event_wait returned an error (state: %d): %m", sd_event_get_state(se));
+                               }
+                       }
+                       __atomic_and_fetch(&el->state, ~(EVLOOP_STATE_WAIT), __ATOMIC_RELAXED);
+
+                       if (rc > 0) {
+                               rc = sd_event_dispatch(se);
+                               if (rc < 0) {
+                                       errno = -rc;
+                                       ERROR("sd_event_dispatch returned an error (state: %d): %m", sd_event_get_state(se));
+                               }
+                       }
+               }
+       }
+       __atomic_and_fetch(&el->state, ~(EVLOOP_STATE_WAIT|EVLOOP_STATE_RUN), __ATOMIC_RELAXED);
+}
+
+
+#if defined(REMOVE_SYSTEMD_EVENT)
+/**
+ * Monitored normal loop for waiting events.
+ * @param signum 0 on normal flow or the number
+ *               of the signal that interrupted the normal
+ *               flow
+ * @param arg     the events to run
+ */
+static void monitored_wait_and_dispatch(int signum, void *arg)
+{
+       struct fdev_epoll *fdev_epoll = arg;
+       if (!signum) {
+               fdev_epoll_wait_and_dispatch(fdev_epoll, -1);
+       }
 }
+#endif
 
-/* main loop of processing threads */
-static void thread_run(struct thread *me)
+/**
+ * Main processing loop of threads processing jobs.
+ * The loop must be called with the mutex locked
+ * and it returns with the mutex locked.
+ * @param me the description of the thread to use
+ * TODO: how are timeout handled when reentering?
+ */
+static void thread_run(volatile struct thread *me)
 {
        struct thread **prv;
        struct job *job;
+#if !defined(REMOVE_SYSTEMD_EVENT)
+       struct evloop *el;
+#endif
 
-       /* init */
+       /* initialize description of itself and link it in the list */
        me->tid = pthread_self();
        me->stop = 0;
-       me->lowered = 0;
-       me->upper = current;
-       if (current)
-               current->lowered = 1;
-       else
+       me->waits = 0;
+       me->upper = current_thread;
+       if (!current_thread) {
+               started++;
                sig_monitor_init_timeouts();
-       current = me;
+       }
        me->next = threads;
-       threads = me;
+       threads = (struct thread*)me;
+       current_thread = (struct thread*)me;
 
        /* loop until stopped */
-       running++;
        while (!me->stop) {
+               /* release the event loop */
+               if (current_evloop) {
+                       __atomic_and_fetch(&current_evloop->state, ~EVLOOP_STATE_LOCK, __ATOMIC_RELAXED);
+                       current_evloop = NULL;
+               }
+
                /* get a job */
                job = job_get();
-               if (!job && first_job && running == 0) {
-                       /* sad situation!! should not happen */
-                       ERROR("threads are blocked!");
-                       job = first_job;
-                       first_job = job->next;
-               }
                if (job) {
+                       /* prepare running the job */
+                       remains++; /* increases count of job that can wait */
+                       job->blocked = 1; /* mark job as blocked */
+                       me->job = job; /* record the job (only for terminate) */
+
                        /* run the job */
-                       remains++;
-                       job->blocked = 1;
-                       me->job = job;
                        pthread_mutex_unlock(&mutex);
-                       sig_monitor(job->timeout, job_call, job);
+                       sig_monitor(job->timeout, job->callback, job->arg);
                        pthread_mutex_lock(&mutex);
+
+                       /* release the run job */
                        job_release(job);
+#if !defined(REMOVE_SYSTEMD_EVENT)
                } else {
-                       /* no job, check evloop */
-                       job = first_evloop;
-                       if (job) {
-                               /* evloop */
-                               first_evloop = job->next;
+                       /* no job, check events */
+                       el = &evloop[0];
+                       if (el->sdev && !__atomic_load_n(&el->state, __ATOMIC_RELAXED)) {
+                               /* run the events */
+                               __atomic_store_n(&el->state, EVLOOP_STATE_LOCK|EVLOOP_STATE_RUN|EVLOOP_STATE_WAIT, __ATOMIC_RELAXED);
+                               current_evloop = el;
                                pthread_mutex_unlock(&mutex);
-                               sig_monitor(job->timeout, job_call, job);
+                               sig_monitor(0, evloop_run, el);
                                pthread_mutex_lock(&mutex);
-                               job->next = first_evloop;
-                               first_evloop = job;
                        } else {
-                               /* no job and not evloop */
+                               /* no job and not events */
                                running--;
+                               if (!running)
+                                       ERROR("Entering job deep sleep! Check your bindings.");
+                               me->waits = 1;
                                pthread_cond_wait(&cond, &mutex);
+                               me->waits = 0;
                                running++;
                        }
+#else
+               } else if (waitevt) {
+                       /* no job and not events */
+                       running--;
+                       if (!running)
+                               ERROR("Entering job deep sleep! Check your bindings.");
+                       me->waits = 1;
+                       pthread_cond_wait(&cond, &mutex);
+                       me->waits = 0;
+                       running++;
+               } else {
+                       /* wait for events */
+                       waitevt = 1;
+                       pthread_mutex_unlock(&mutex);
+                       sig_monitor(0, monitored_wait_and_dispatch, get_fdevepoll());
+                       pthread_mutex_lock(&mutex);
+                       waitevt = 0;
+#endif
                }
        }
-       running--;
 
-       /* uninit */
+       /* release the event loop */
+       if (current_evloop) {
+               __atomic_and_fetch(&current_evloop->state, ~EVLOOP_STATE_LOCK, __ATOMIC_RELAXED);
+               current_evloop = NULL;
+       }
+
+       /* unlink the current thread and cleanup */
        prv = &threads;
        while (*prv != me)
                prv = &(*prv)->next;
        *prv = me->next;
-       current = me->upper;
-       if (current)
-               current->lowered = 0;
-       else
+       current_thread = me->upper;
+       if (!current_thread) {
                sig_monitor_clean_timeouts();
-       pthread_mutex_unlock(&mutex);
+               started--;
+       }
 }
 
-/* main loop of processing threads */
-static void *thread_create(void *data)
+/**
+ * Entry point for created threads.
+ * @param data not used
+ * @return NULL
+ */
+static void *thread_main(void *data)
 {
        struct thread me;
 
        pthread_mutex_lock(&mutex);
+       running++;
        thread_run(&me);
+       running--;
        pthread_mutex_unlock(&mutex);
        return NULL;
 }
 
-/* start a new thread */
+/**
+ * Starts a new thread
+ * @return 0 in case of success or -1 in case of error
+ */
 static int start_one_thread()
 {
        pthread_t tid;
        int rc;
 
-       assert(started < allowed);
-
-       started++;
-       rc = pthread_create(&tid, NULL, thread_create, NULL);
+       rc = pthread_create(&tid, NULL, thread_main, NULL);
        if (rc != 0) {
-               started--;
-               errno = rc;
+               /* errno = rc; */
                WARNING("not able to start thread: %m");
                rc = -1;
        }
        return rc;
 }
 
-static int start_one_thread_if_needed()
-{
-       int rc;
-
-       if (started == running && started < allowed) {
-               /* all threads are busy and a new can be started */
-               rc = start_one_thread();
-               if (rc < 0 && started == 0)
-                       return rc; /* no thread available */
-       }
-       return 0;
-}
-
-int jobs_queue0(
-               void *group,
-               int timeout,
-               void (*callback)(int signum))
-{
-       return jobs_queue3(group, timeout, (void(*)(int,void*,void*,void*))callback, NULL, NULL, NULL);
-}
-
+/**
+ * Queues a new asynchronous job represented by 'callback' and 'arg'
+ * for the 'group' and the 'timeout'.
+ * Jobs are queued FIFO and are possibly executed in parallel
+ * concurrently except for job of the same group that are
+ * executed sequentially in FIFO order.
+ * @param group    The group of the job or NULL when no group.
+ * @param timeout  The maximum execution time in seconds of the job
+ *                 or 0 for unlimited time.
+ * @param callback The function to execute for achieving the job.
+ *                 Its first parameter is either 0 on normal flow
+ *                 or the signal number that broke the normal flow.
+ *                 The remaining parameter is the parameter 'arg1'
+ *                 given here.
+ * @param arg      The second argument for 'callback'
+ * @return 0 in case of success or -1 in case of error
+ */
 int jobs_queue(
-               void *group,
+               const void *group,
                int timeout,
                void (*callback)(int, void*),
                void *arg)
-{
-       return jobs_queue3(group, timeout, (void(*)(int,void*,void*,void*))callback, arg, NULL, NULL);
-}
-
-int jobs_queue2(
-               void *group,
-               int timeout,
-               void (*callback)(int, void*, void*),
-               void *arg1,
-               void *arg2)
-{
-       return jobs_queue3(group, timeout, (void(*)(int,void*,void*,void*))callback, arg1, arg2, NULL);
-}
-
-/* queue the job to the 'callback' using a separate thread if available */
-int jobs_queue3(
-               void *group,
-               int timeout,
-               void (*callback)(int, void*, void *, void*),
-               void *arg1,
-               void *arg2,
-               void *arg3)
 {
        const char *info;
        struct job *job;
@@ -382,7 +524,7 @@ int jobs_queue3(
        pthread_mutex_lock(&mutex);
 
        /* allocates the job */
-       job = job_create(group, timeout, callback, arg1, arg2, arg3);
+       job = job_create(group, timeout, callback, arg);
        if (!job) {
                errno = ENOMEM;
                info = "out of memory";
@@ -397,20 +539,22 @@ int jobs_queue3(
        }
 
        /* start a thread if needed */
-       rc = start_one_thread_if_needed();
-       if (rc < 0) {
-               /* failed to start threading */
-               info = "can't start first thread";
-               goto error2;
+       if (running == started && started < allowed) {
+               /* all threads are busy and a new can be started */
+               rc = start_one_thread();
+               if (rc < 0 && started == 0) {
+                       info = "can't start first thread";
+                       goto error2;
+               }
        }
 
        /* queues the job */
        remains--;
-       job_add2(job, NULL);
-       pthread_mutex_unlock(&mutex);
+       job_add(job);
 
        /* signal an existing job */
        pthread_cond_signal(&cond);
+       pthread_mutex_unlock(&mutex);
        return 0;
 
 error2:
@@ -422,116 +566,357 @@ error:
        return -1;
 }
 
-/* initialise the threads */
-int jobs_init(int allowed_count, int start_count, int waiter_count)
+/**
+ * Internal helper function for 'jobs_enter'.
+ * @see jobs_enter, jobs_leave
+ */
+static void enter_cb(int signum, void *closure)
 {
-       /* records the allowed count */
-       allowed = allowed_count;
-       started = 0;
-       running = 0;
-       remains = waiter_count;
+       struct sync *sync = closure;
+       sync->enter(signum, sync->arg, (void*)&sync->thread);
+}
+
+/**
+ * Internal helper function for 'jobs_call'.
+ * @see jobs_call
+ */
+static void call_cb(int signum, void *closure)
+{
+       struct sync *sync = closure;
+       sync->callback(signum, sync->arg);
+       jobs_leave((void*)&sync->thread);
+}
+
+/**
+ * Internal helper for synchronous jobs. It enters
+ * a new thread loop for evaluating the given job
+ * as recorded by the couple 'sync_cb' and 'sync'.
+ * @see jobs_call, jobs_enter, jobs_leave
+ */
+static int do_sync(
+               const void *group,
+               int timeout,
+               void (*sync_cb)(int signum, void *closure),
+               struct sync *sync
+)
+{
+       struct job *job;
 
-       /* start at least one thread */
        pthread_mutex_lock(&mutex);
-       while (started < start_count && start_one_thread() == 0);
-       pthread_mutex_unlock(&mutex);
 
-       /* end */
-       return -(started != start_count);
+       /* allocates the job */
+       job = job_create(group, timeout, sync_cb, sync);
+       if (!job) {
+               ERROR("out of memory");
+               errno = ENOMEM;
+               pthread_mutex_unlock(&mutex);
+               return -1;
+       }
+
+       /* queues the job */
+       job_add(job);
+
+       /* run until stopped */
+       thread_run(&sync->thread);
+       pthread_mutex_unlock(&mutex);
+       return 0;
 }
 
-int jobs_invoke0(
+/**
+ * Enter a synchronisation point: activates the job given by 'callback'
+ * and 'closure' using 'group' and 'timeout' to control sequencing and
+ * execution time.
+ * @param group the group for sequencing jobs
+ * @param timeout the time in seconds allocated to the job
+ * @param callback the callback that will handle the job.
+ *                 it receives 3 parameters: 'signum' that will be 0
+ *                 on normal flow or the catched signal number in case
+ *                 of interrupted flow, the context 'closure' as given and
+ *                 a 'jobloop' reference that must be used when the job is
+ *                 terminated to unlock the current execution flow.
+ * @param closure the argument to the callback
+ * @return 0 on success or -1 in case of error
+ */
+int jobs_enter(
+               const void *group,
                int timeout,
-               void (*callback)(int signum))
+               void (*callback)(int signum, void *closure, struct jobloop *jobloop),
+               void *closure
+)
 {
-       return jobs_invoke3(timeout, (void(*)(int,void*,void*,void*))callback, NULL, NULL, NULL);
+       struct sync sync;
+
+       sync.enter = callback;
+       sync.arg = closure;
+       return do_sync(group, timeout, enter_cb, &sync);
+}
+
+/**
+ * Unlocks the execution flow designed by 'jobloop'.
+ * @param jobloop indication of the flow to unlock
+ * @return 0 in case of success of -1 on error
+ */
+int jobs_leave(struct jobloop *jobloop)
+{
+       struct thread *t;
+
+       pthread_mutex_lock(&mutex);
+       t = threads;
+       while (t && t != (struct thread*)jobloop)
+               t = t->next;
+       if (!t) {
+               errno = EINVAL;
+       } else {
+               t->stop = 1;
+               if (t->waits)
+                       pthread_cond_broadcast(&cond);
+       }
+       pthread_mutex_unlock(&mutex);
+       return -!t;
 }
 
-int jobs_invoke(
+/**
+ * Calls synchronously the job represented by 'callback' and 'arg1'
+ * for the 'group' and the 'timeout' and waits for its completion.
+ * @param group    The group of the job or NULL when no group.
+ * @param timeout  The maximum execution time in seconds of the job
+ *                 or 0 for unlimited time.
+ * @param callback The function to execute for achieving the job.
+ *                 Its first parameter is either 0 on normal flow
+ *                 or the signal number that broke the normal flow.
+ *                 The remaining parameter is the parameter 'arg1'
+ *                 given here.
+ * @param arg      The second argument for 'callback'
+ * @return 0 in case of success or -1 in case of error
+ */
+int jobs_call(
+               const void *group,
                int timeout,
                void (*callback)(int, void*),
                void *arg)
 {
-       return jobs_invoke3(timeout, (void(*)(int,void*,void*,void*))callback, arg, NULL, NULL);
+       struct sync sync;
+
+       sync.callback = callback;
+       sync.arg = arg;
+
+       return do_sync(group, timeout, call_cb, &sync);
 }
 
-int jobs_invoke2(
-               int timeout,
-               void (*callback)(int, void*, void*),
-               void *arg1,
-               void *arg2)
+/**
+ * Internal callback for evloop management.
+ * The effect of this function is hidden: it exits
+ * the waiting poll if any. Then it wakes up a thread
+ * awaiting the evloop using signal.
+ */
+static int on_evloop_efd(sd_event_source *s, int fd, uint32_t revents, void *userdata)
+{
+       uint64_t x;
+       struct evloop *evloop = userdata;
+       read(evloop->efd, &x, sizeof x);
+       pthread_mutex_lock(&mutex);
+       pthread_cond_broadcast(&evloop->cond);
+       pthread_mutex_unlock(&mutex);
+       return 1;
+}
+
+/* temporary hack */
+#if !defined(REMOVE_SYSTEMD_EVENT)
+__attribute__((unused))
+#endif
+static void evloop_callback(void *arg, uint32_t event, struct fdev *fdev)
+{
+       sig_monitor(0, evloop_run, arg);
+}
+
+/**
+ * Gets a sd_event item for the current thread.
+ * @return a sd_event or NULL in case of error
+ */
+static struct sd_event *get_sd_event_locked()
+{
+       struct evloop *el;
+       uint64_t x;
+       int rc;
+
+       /* creates the evloop on need */
+       el = &evloop[0];
+       if (!el->sdev) {
+               /* start the creation */
+               el->state = 0;
+               /* creates the eventfd for waking up polls */
+               el->efd = eventfd(0, EFD_CLOEXEC);
+               if (el->efd < 0) {
+                       ERROR("can't make eventfd for events");
+                       goto error1;
+               }
+               /* create the systemd event loop */
+               rc = sd_event_new(&el->sdev);
+               if (rc < 0) {
+                       ERROR("can't make new event loop");
+                       goto error2;
+               }
+               /* put the eventfd in the event loop */
+               rc = sd_event_add_io(el->sdev, NULL, el->efd, EPOLLIN, on_evloop_efd, el);
+               if (rc < 0) {
+                       ERROR("can't register eventfd");
+#if !defined(REMOVE_SYSTEMD_EVENT)
+                       sd_event_unref(el->sdev);
+                       el->sdev = NULL;
+error2:
+                       close(el->efd);
+error1:
+                       return NULL;
+               }
+#else
+                       goto error3;
+               }
+               /* handle the event loop */
+               el->fdev = fdev_epoll_add(get_fdevepoll(), sd_event_get_fd(el->sdev));
+               if (!el->fdev) {
+                       ERROR("can't create fdev");
+error3:
+                       sd_event_unref(el->sdev);
+error2:
+                       close(el->efd);
+error1:
+                       memset(el, 0, sizeof *el);
+                       return NULL;
+               }
+               fdev_set_autoclose(el->fdev, 0);
+               fdev_set_events(el->fdev, EPOLLIN);
+               fdev_set_callback(el->fdev, evloop_callback, el);
+#endif
+       }
+
+       /* attach the event loop to the current thread */
+       if (current_evloop != el) {
+               if (current_evloop)
+                       __atomic_and_fetch(&current_evloop->state, ~EVLOOP_STATE_LOCK, __ATOMIC_RELAXED);
+               current_evloop = el;
+               __atomic_or_fetch(&el->state, EVLOOP_STATE_LOCK, __ATOMIC_RELAXED);
+       }
+
+       /* wait for a modifiable event loop */
+       while (__atomic_load_n(&el->state, __ATOMIC_RELAXED) & EVLOOP_STATE_WAIT) {
+               x = 1;
+               write(el->efd, &x, sizeof x);
+               pthread_cond_wait(&el->cond, &mutex);
+       }
+
+       return el->sdev;
+}
+
+/**
+ * Gets a sd_event item for the current thread.
+ * @return a sd_event or NULL in case of error
+ */
+struct sd_event *jobs_get_sd_event()
 {
-       return jobs_invoke3(timeout, (void(*)(int,void*,void*,void*))callback, arg1, arg2, NULL);
+       struct sd_event *result;
+
+       pthread_mutex_lock(&mutex);
+       result = get_sd_event_locked();
+       pthread_mutex_unlock(&mutex);
+
+       return result;
 }
 
-static void unlock_invoker(int signum, void *arg1, void *arg2, void *arg3)
+#if defined(REMOVE_SYSTEMD_EVENT)
+/**
+ * Gets the fdev_epoll item.
+ * @return a fdev_epoll or NULL in case of error
+ */
+struct fdev_epoll *jobs_get_fdev_epoll()
 {
-       struct thread *t = arg1;
+       struct fdev_epoll *result;
+
        pthread_mutex_lock(&mutex);
-       t->stop = 1;
+       result = get_fdevepoll();
        pthread_mutex_unlock(&mutex);
+
+       return result;
 }
+#endif
 
-/* invoke the job to the 'callback' using a separate thread if available */
-int jobs_invoke3(
-               int timeout,
-               void (*callback)(int, void*, void *, void*),
-               void *arg1,
-               void *arg2,
-               void *arg3)
+/**
+ * Enter the jobs processing loop.
+ * @param allowed_count Maximum count of thread for jobs including this one
+ * @param start_count   Count of thread to start now, must be lower.
+ * @param waiter_count  Maximum count of jobs that can be waiting.
+ * @param start         The start routine to activate (can't be NULL)
+ * @return 0 in case of success or -1 in case of error.
+ */
+int jobs_start(int allowed_count, int start_count, int waiter_count, void (*start)(int signum, void* arg), void *arg)
 {
-       const char *info;
-       struct job *job1, *job2;
-       int rc;
+       int rc, launched;
        struct thread me;
-       
+       struct job *job;
+
+       assert(allowed_count >= 1);
+       assert(start_count >= 0);
+       assert(waiter_count > 0);
+       assert(start_count <= allowed_count);
+
+       rc = -1;
        pthread_mutex_lock(&mutex);
 
-       /* allocates the job */
-       job1 = job_create(&me, timeout, callback, arg1, arg2, arg3);
-       job2 = job_create(&me, 0, unlock_invoker, &me, NULL, NULL);
-       if (!job1 || !job2) {
-               errno = ENOMEM;
-               info = "out of memory";
+       /* check whether already running */
+       if (current_thread || allowed) {
+               ERROR("thread already started");
+               errno = EINVAL;
                goto error;
        }
 
-       /* start a thread if needed */
-       rc = start_one_thread_if_needed();
-       if (rc < 0) {
-               /* failed to start threading */
-               info = "can't start first thread";
-               goto error;
+       /* records the allowed count */
+       allowed = allowed_count;
+       started = 0;
+       running = 0;
+       remains = waiter_count;
+
+#if HAS_WATCHDOG
+       /* set the watchdog */
+       if (sd_watchdog_enabled(0, NULL))
+               sd_event_set_watchdog(get_sd_event_locked(), 1);
+#endif
+
+       /* start at least one thread */
+       launched = 0;
+       while ((launched + 1) < start_count) {
+               if (start_one_thread() != 0) {
+                       ERROR("Not all threads can be started");
+                       goto error;
+               }
+               launched++;
        }
 
-       /* queues the job */
-       job_add2(job1, job2);
+       /* queue the start job */
+       job = job_create(NULL, 0, start, arg);
+       if (!job) {
+               ERROR("out of memory");
+               errno = ENOMEM;
+               goto error;
+       }
+       job_add(job);
+       remains--;
 
-       /* run untill stopped */
+       /* run until end */
        thread_run(&me);
-       pthread_mutex_unlock(&mutex);
-       return 0;
-
+       rc = 0;
 error:
-       if (job1) {
-               job1->next = free_jobs;
-               free_jobs = job1;
-       }
-       if (job2) {
-               job2->next = free_jobs;
-               free_jobs = job2;
-       }
-       ERROR("can't process job with threads: %s, %m", info);
        pthread_mutex_unlock(&mutex);
-       return -1;
+       return rc;
 }
 
-/* terminate all the threads and all pending requests */
+/**
+ * Terminate all the threads and cancel all pending jobs.
+ */
 void jobs_terminate()
 {
        struct job *job, *head, *tail;
-       pthread_t me, other;
+       pthread_t me, *others;
        struct thread *t;
+       int count;
 
        /* how am i? */
        me = pthread_self();
@@ -539,23 +924,42 @@ void jobs_terminate()
        /* request all threads to stop */
        pthread_mutex_lock(&mutex);
        allowed = 0;
-       for(;;) {
-               /* search the next thread to stop */
-               t = threads;
-               while (t && pthread_equal(t->tid, me))
-                       t = t->next;
-               if (!t)
-                       break;
-               /* stop it */
-               other = t->tid;
+
+       /* count the number of threads */
+       count = 0;
+       t = threads;
+       while (t) {
+               if (!t->upper && !pthread_equal(t->tid, me))
+                       count++;
+               t = t->next;
+       }
+
+       /* fill the array of threads */
+       others = alloca(count * sizeof *others);
+       count = 0;
+       t = threads;
+       while (t) {
+               if (!t->upper && !pthread_equal(t->tid, me))
+                       others[count++] = t->tid;
+               t = t->next;
+       }
+
+       /* stops the threads */
+       t = threads;
+       while (t) {
                t->stop = 1;
-               pthread_mutex_unlock(&mutex);
-               pthread_cond_broadcast(&cond);
-               pthread_join(other, NULL);
-               pthread_mutex_lock(&mutex);
+               t = t->next;
        }
 
+       /* wait the threads */
+       pthread_cond_broadcast(&cond);
+       pthread_mutex_unlock(&mutex);
+       while (count)
+               pthread_join(others[--count], NULL);
+       pthread_mutex_lock(&mutex);
+
        /* cancel pending jobs of other threads */
+       remains = 0;
        head = first_job;
        first_job = NULL;
        tail = NULL;
@@ -565,7 +969,7 @@ void jobs_terminate()
                head = job->next;
 
                /* search if job is stacked for current */
-               t = current;
+               t = current_thread;
                while (t && t->job != job)
                        t = t->upper;
                if (t) {
@@ -587,42 +991,3 @@ void jobs_terminate()
        pthread_mutex_unlock(&mutex);
 }
 
-int jobs_add_event_loop(void *key, int timeout, void (*evloop)(int signum, void*), void *closure)
-{
-       struct job *job;
-
-       pthread_mutex_lock(&mutex);
-       job = job_create(key, timeout, (void (*)(int,  void *, void *, void *))evloop, closure, NULL, NULL);
-       if (job) {
-               /* adds the loop */
-               job->next = first_evloop;
-               first_evloop = job;
-
-               /* signal the loop */
-               pthread_cond_signal(&cond);
-       }
-       pthread_mutex_unlock(&mutex);
-       return -!job;
-}
-
-int jobs_add_me()
-{
-       struct thread me;
-
-       /* check whether already running */
-       if (current) {
-               ERROR("thread already running");
-               errno = EINVAL;
-               return -1;
-       }
-
-       /* allowed... */
-       pthread_mutex_lock(&mutex);
-       allowed++;
-       thread_run(&me);
-       allowed--;
-       pthread_mutex_unlock(&mutex);
-       return 0;
-}
-
-