Allow choice to trap or not program faults
[src/app-framework-binder.git] / src / jobs.c
index 93e864f..a6357c8 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"
 
-#if 0
-#define _alert_ "do you really want to remove monitoring?"
-#define sig_monitor_init_timeouts()  ((void)0)
-#define sig_monitor_clean_timeouts() ((void)0)
-#define sig_monitor(to,cb,arg)       (cb(0,arg))
+#if defined(REMOVE_SYSTEMD_EVENT)
+#include "fdev-epoll.h"
 #endif
 
 #define EVENT_TIMEOUT_TOP      ((uint64_t)-1)
@@ -59,26 +68,28 @@ struct job
 };
 
 /** Description of handled event loops */
-struct events
+struct evloop
 {
-       struct events *next;
-       struct sd_event *event;
-       uint64_t timeout;
-       unsigned used: 1;
-       unsigned runs: 1;
+       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 */
-       struct events *events; /**< currently processed job */
        pthread_t tid;         /**< the thread id */
-       unsigned stop: 1;      /**< stop requested */
-       unsigned lowered: 1;   /**< has a lower same thread */
-       unsigned waits: 1;     /**< is waiting? */
+       volatile unsigned stop: 1;      /**< stop requested */
+       volatile unsigned waits: 1;     /**< is waiting? */
 };
 
 /**
@@ -100,22 +111,29 @@ struct sync
 static pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
 static pthread_cond_t  cond = PTHREAD_COND_INITIALIZER;
 
-/* count allowed, started and waiting threads */
+/* count allowed, started and running threads */
 static int allowed = 0; /** allowed count of threads */
 static int started = 0; /** started count of threads */
-static int waiting = 0; /** waiting count of threads */
+static int running = 0; /** running count of threads */
 static int remains = 0; /** allowed count of waiting jobs */
-static int nevents = 0; /** count of events */
 
 /* 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 events *first_events;
 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
@@ -137,7 +155,7 @@ 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);
@@ -197,18 +215,6 @@ static inline struct job *job_get()
        return job;
 }
 
-/**
- * Get the next events to process or NULL if none.
- * @return the first events that isn't running or NULL
- */
-static inline struct events *events_get()
-{
-       struct events *events = first_events;
-       while (events && events->used)
-               events = events->next;
-       return events;
-}
-
 /**
  * Releases the processed 'job': removes it
  * from the list of jobs and unblock the first
@@ -260,6 +266,23 @@ static void job_cancel(int signum, void *arg)
        job->callback(SIGABRT, job->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
@@ -270,13 +293,61 @@ static void job_cancel(int signum, void *arg)
  *               flow
  * @param arg     the events to run
  */
-static void events_call(int signum, void *arg)
+static void evloop_run(int signum, void *arg)
 {
-       struct events *events = arg;
-       if (!signum)
-               sd_event_run(events->event, events->timeout);
+       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 processing loop of threads processing jobs.
  * The loop must be called with the mutex locked
@@ -286,35 +357,35 @@ static void events_call(int signum, void *arg)
  */
 static void thread_run(volatile struct thread *me)
 {
-       struct thread **prv, *thr;
+       struct thread **prv;
        struct job *job;
-       struct events *events;
-       uint64_t evto;
+#if !defined(REMOVE_SYSTEMD_EVENT)
+       struct evloop *el;
+#endif
 
        /* initialize description of itself and link it in the list */
        me->tid = pthread_self();
        me->stop = 0;
-       me->lowered = 0;
        me->waits = 0;
-       me->upper = current;
-       if (current) {
-               current->lowered = 1;
-               evto = EVENT_TIMEOUT_CHILD;
-               me->events = current->events;
-       } else {
+       me->upper = current_thread;
+       if (!current_thread) {
                started++;
                sig_monitor_init_timeouts();
-               evto = EVENT_TIMEOUT_TOP;
-               me->events = NULL;
        }
        me->next = threads;
        threads = (struct thread*)me;
-       current = (struct thread*)me;
+       current_thread = (struct thread*)me;
 
        /* loop until stopped */
        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(first_job);
+               job = job_get();
                if (job) {
                        /* prepare running the job */
                        remains++; /* increases count of job that can wait */
@@ -328,55 +399,61 @@ static void thread_run(volatile struct thread *me)
 
                        /* release the run job */
                        job_release(job);
-
-                       /* release event if any */
-                       events = me->events;
-                       if (events) {
-                               events->used = 0;
-                               me->events = NULL;
-                       }
+#if !defined(REMOVE_SYSTEMD_EVENT)
                } else {
                        /* no job, check events */
-                       events = me->events;
-                       if (!events || events->runs)
-                               events = events_get();
-                       if (events) {
+                       el = &evloop[0];
+                       if (el->sdev && !__atomic_load_n(&el->state, __ATOMIC_RELAXED)) {
                                /* run the events */
-                               events->used = 1;
-                               events->runs = 1;
-                               events->timeout = evto;
-                               me->events = 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(0, events_call, events);
+                               sig_monitor(0, evloop_run, el);
                                pthread_mutex_lock(&mutex);
-                               events->used = 0;
-                               events->runs = 0;
-                               me->events = NULL;
-                               thr = me->upper;
-                               while (thr && thr->events == events) {
-                                       thr->events = NULL;
-                                       thr = thr->upper;
-                               }
                        } else {
                                /* no job and not events */
-                               waiting++;
+                               running--;
+                               if (!running)
+                                       ERROR("Entering job deep sleep! Check your bindings.");
                                me->waits = 1;
                                pthread_cond_wait(&cond, &mutex);
                                me->waits = 0;
-                               waiting--;
+                               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
                }
        }
 
+       /* 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();
                started--;
        }
@@ -392,7 +469,9 @@ static void *thread_main(void *data)
        struct thread me;
 
        pthread_mutex_lock(&mutex);
+       running++;
        thread_run(&me);
+       running--;
        pthread_mutex_unlock(&mutex);
        return NULL;
 }
@@ -460,7 +539,7 @@ int jobs_queue(
        }
 
        /* start a thread if needed */
-       if (waiting == 0 && started < allowed) {
+       if (running == started && started < allowed) {
                /* all threads are busy and a new can be started */
                rc = start_one_thread();
                if (rc < 0 && started == 0) {
@@ -555,7 +634,7 @@ static int do_sync(
  *                 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 arg the argument to the callback
+ * @param closure the argument to the callback
  * @return 0 on success or -1 in case of error
  */
 int jobs_enter(
@@ -624,73 +703,141 @@ int jobs_call(
        return do_sync(group, timeout, call_cb, &sync);
 }
 
+/**
+ * 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
  */
-struct sd_event *jobs_get_sd_event()
+static struct sd_event *get_sd_event_locked()
 {
-       struct events *events;
-       struct thread *me;
+       struct evloop *el;
+       uint64_t x;
        int rc;
 
-       pthread_mutex_lock(&mutex);
-
-       /* search events on stack */
-       me = current;
-       while (me && !me->events)
-               me = me->upper;
-       if (me)
-               /* return the stacked events */
-               events = me->events;
-       else {
-               /* search an available events */
-               events = events_get();
-               if (!events) {
-                       /* not found, check if creation possible */
-                       if (nevents >= allowed) {
-                               ERROR("not possible to add a new event");
-                               events = NULL;
-                       } else {
-                               events = malloc(sizeof *events);
-                               if (events && (rc = sd_event_new(&events->event)) >= 0) {
-                                       if (nevents < started || start_one_thread() >= 0) {
-                                               events->used = 0;
-                                               events->runs = 0;
-                                               events->next = first_events;
-                                               first_events = events;
-                                       } else {
-                                               ERROR("can't start thread for events");
-                                               sd_event_unref(events->event);
-                                               free(events);
-                                               events = NULL;
-                                       }
-                               } else {
-                                       if (!events) {
-                                               ERROR("out of memory");
-                                               errno = ENOMEM;
-                                       } else {
-                                               free(events);
-                                               ERROR("creation of sd_event failed: %m");
-                                               events = NULL;
-                                               errno = -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;
                }
-               if (events) {
-                       me = current;
-                       if (me) {
-                               events->used = 1;
-                               me->events = events;
-                       } else {
-                               WARNING("event returned for unknown thread!");
-                       }
+               /* 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()
+{
+       struct sd_event *result;
+
+       pthread_mutex_lock(&mutex);
+       result = get_sd_event_locked();
+       pthread_mutex_unlock(&mutex);
+
+       return result;
+}
+
+#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 fdev_epoll *result;
+
+       pthread_mutex_lock(&mutex);
+       result = get_fdevepoll();
        pthread_mutex_unlock(&mutex);
-       return events ? events->event : NULL;
+
+       return result;
 }
+#endif
 
 /**
  * Enter the jobs processing loop.
@@ -700,7 +847,7 @@ struct sd_event *jobs_get_sd_event()
  * @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))
+int jobs_start(int allowed_count, int start_count, int waiter_count, void (*start)(int signum, void* arg), void *arg)
 {
        int rc, launched;
        struct thread me;
@@ -715,24 +862,24 @@ int jobs_start(int allowed_count, int start_count, int waiter_count, void (*star
        pthread_mutex_lock(&mutex);
 
        /* check whether already running */
-       if (current || allowed) {
+       if (current_thread || allowed) {
                ERROR("thread already started");
                errno = EINVAL;
                goto error;
        }
 
-       /* start */
-       if (sig_monitor_init() < 0) {
-               ERROR("failed to initialise signal handlers");
-               goto error;
-       }
-
        /* records the allowed count */
        allowed = allowed_count;
        started = 0;
-       waiting = 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) {
@@ -744,7 +891,7 @@ int jobs_start(int allowed_count, int start_count, int waiter_count, void (*star
        }
 
        /* queue the start job */
-       job = job_create(NULL, 0, (job_cb_t)start, NULL);
+       job = job_create(NULL, 0, start, arg);
        if (!job) {
                ERROR("out of memory");
                errno = ENOMEM;
@@ -822,7 +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) {