/*
- * Copyright (C) 2006-2011 Andre Noll <maan@systemlinux.org>
+ * Copyright (C) 2006 Andre Noll <maan@tuebingen.mpg.de>
*
* Licensed under the GPL v2. For licencing details see COPYING.
*/
/** \file sched.c Paraslash's scheduling functions. */
#include <regex.h>
-#include <dirent.h> /* readdir() */
-#include <assert.h>
-#include <sys/time.h>
#include "para.h"
#include "ipc.h"
#include "time.h"
#include "error.h"
-static struct list_head pre_select_list, post_select_list;
-static int initialized;
-
-static struct timeval now_struct;
-struct timeval *now = &now_struct;
-
-/*
- * Remove a task from the scheduler.
- *
- * \param t The task to remove.
+/**
+ * The possible states of a task.
*
- * If the pre_select pointer of \a t is not \p NULL, it is removed from
- * the pre_select list of the scheduler. Same goes for \a post_select.
+ * In addition to the states listed here, a task may also enter zombie state.
+ * This happens when its ->post_select function returns negative, the ->status
+ * field is then set to this return value. Such tasks are not scheduled any
+ * more (i.e. ->pre_select() and ->post_select() are no longer called), but
+ * they stay on the scheduler task list until \ref task_reap() or
+ * \ref sched_shutdown() is called.
*/
-static void unregister_task(struct task *t)
-{
- if (!initialized)
- return;
- PARA_INFO_LOG("unregistering %s (%s)\n", t->status,
- t->error <0? para_strerror(-t->error) : "shutdown");
- if (t->pre_select)
- list_del(&t->pre_select_node);
- if (t->post_select)
- list_del(&t->post_select_node);
- t->error = -E_TASK_UNREGISTERED;
-}
+enum task_status {
+ /** Task has been reaped and may be removed from the task list. */
+ TS_DEAD,
+ /** Task is active. */
+ TS_RUNNING,
+};
+struct task {
+ /** A copy of the task name supplied when the task was registered. */
+ char *name;
+ /** Copied during task_register(). */
+ struct task_info info;
+ /* TS_RUNNING, TS_DEAD, or zombie (negative value). */
+ int status;
+ /** Position of the task in the task list of the scheduler. */
+ struct list_head node;
+ /** If less than zero, the task was notified by another task. */
+ int notification;
+};
+
+static struct timeval now_struct;
+const struct timeval *now = &now_struct;
static void sched_preselect(struct sched *s)
{
struct task *t, *tmp;
- list_for_each_entry_safe(t, tmp, &pre_select_list, pre_select_node) {
- if (t->error >= 0 && t->pre_select)
- t->pre_select(s, t);
-// PARA_INFO_LOG("%s \n", t->status);
- if (t->error >= 0)
+
+ list_for_each_entry_safe(t, tmp, &s->task_list, node) {
+ if (t->status < 0)
continue;
- /*
- * We have to check whether the list is empty because the call
- * to ->pre_select() might have called sched_shutdown(). In
- * this case t has been unregistered already, so we must not
- * unregister it again.
- */
- if (list_empty(&pre_select_list))
- return;
- unregister_task(t);
+ if (t->notification != 0)
+ sched_min_delay(s);
+ if (t->info.pre_select)
+ t->info.pre_select(s, t->info.context);
}
}
+static void unlink_and_free_task(struct task *t)
+{
+ PARA_INFO_LOG("freeing task %s\n", t->name);
+ list_del(&t->node);
+ free(t->name);
+ free(t);
+}
+
//#define SCHED_DEBUG 1
static inline void call_post_select(struct sched *s, struct task *t)
{
+ int ret;
+
#ifndef SCHED_DEBUG
- t->post_select(s, t);
+ ret = t->info.post_select(s, t->info.context);
#else
struct timeval t1, t2, diff;
unsigned long pst;
- gettimeofday(&t1, NULL);
- t->post_select(s, t);
- gettimeofday(&t2, NULL);
+ clock_get_realtime(&t1);
+ ret = t->info.post_select(s, t->info.context);
+ clock_get_realtime(&t2);
tv_diff(&t1, &t2, &diff);
pst = tv2ms(&diff);
if (pst > 50)
PARA_WARNING_LOG("%s: post_select time: %lums\n",
- t->status, pst);
+ t->name, pst);
#endif
+ t->status = ret < 0? ret : TS_RUNNING;
}
-static void sched_post_select(struct sched *s)
+static unsigned sched_post_select(struct sched *s)
{
struct task *t, *tmp;
+ unsigned num_running_tasks = 0;
- list_for_each_entry_safe(t, tmp, &post_select_list, post_select_node) {
- if (t->error >= 0)
- call_post_select(s, t);
-// PARA_INFO_LOG("%s: %d\n", t->status, t->ret);
- if (t->error >= 0)
- continue;
- /* nec., see sched_preselect() */
- if (list_empty(&post_select_list))
- return;
- unregister_task(t);
+ list_for_each_entry_safe(t, tmp, &s->task_list, node) {
+ if (t->status == TS_DEAD) /* task has been reaped */
+ unlink_and_free_task(t);
+ else if (t->status == TS_RUNNING) {
+ call_post_select(s, t); /* sets t->status */
+ t->notification = 0;
+ if (t->status == TS_RUNNING)
+ num_running_tasks++;
+ }
}
+ return num_running_tasks;
}
/**
- * The core function for all paraslash programs.
+ * The core function of all paraslash programs.
*
* \param s Pointer to the scheduler struct.
*
* the fd sets of \a s. Next, it calls para_select() and makes the result available
* to the registered tasks by calling their post_select hook.
*
- * \return Zero if no more tasks are left in either of the two lists, negative
- * if para_select returned an error.
+ * \return Zero if no more tasks are left in the task list, negative if the
+ * select function returned an error.
*
- * \sa task, now.
+ * \sa \ref now.
*/
int schedule(struct sched *s)
{
int ret;
+ unsigned num_running_tasks;
- if (!initialized)
- return -E_NOT_INITIALIZED;
if (!s->select_function)
s->select_function = para_select;
again:
FD_ZERO(&s->wfds);
s->select_timeout = s->default_timeout;
s->max_fileno = -1;
- gettimeofday(now, NULL);
+ clock_get_realtime(&now_struct);
sched_preselect(s);
- if (list_empty(&pre_select_list) && list_empty(&post_select_list))
- return 0;
ret = s->select_function(s->max_fileno + 1, &s->rfds, &s->wfds,
&s->select_timeout);
if (ret < 0)
FD_ZERO(&s->rfds);
FD_ZERO(&s->wfds);
}
- gettimeofday(now, NULL);
- sched_post_select(s);
- if (list_empty(&pre_select_list) && list_empty(&post_select_list))
+ clock_get_realtime(&now_struct);
+ num_running_tasks = sched_post_select(s);
+ if (num_running_tasks == 0)
return 0;
goto again;
}
-/*
- * Initialize the paraslash scheduler.
+/**
+ * Obtain the error status of a task and deallocate its resources.
+ *
+ * \param tptr Identifies the task to reap.
+ *
+ * This function is similar to wait(2) in that it returns information about a
+ * terminated task and allows to release the resources associated with the
+ * task. Until this function is called, the terminated task remains in a zombie
+ * state.
+ *
+ * \return If \a tptr is \p NULL, or \a *tptr is \p NULL, the function does
+ * nothing and returns zero. Otherwise, it is checked whether the task
+ * identified by \a tptr is still running. If it is, the function returns zero
+ * and again, no action is taken. Otherwise the (negative) error code of the
+ * terminated task is returned and \a *tptr is set to \p NULL. The task will
+ * then be removed removed from the scheduler task list.
+ *
+ * \sa \ref sched_shutdown(), wait(2).
*/
-static void init_sched(void)
+int task_reap(struct task **tptr)
{
- PARA_INFO_LOG("initializing scheduler\n");
- INIT_LIST_HEAD(&pre_select_list);
- INIT_LIST_HEAD(&post_select_list);
- initialized = 1;
+ struct task *t;
+ int ret;
+
+ if (!tptr)
+ return 0;
+ t = *tptr;
+ if (!t)
+ return 0;
+ if (t->status >= 0)
+ return 0;
+ ret = t->status;
+ /*
+ * With list_for_each_entry_safe() it is only safe to remove the
+ * _current_ list item. Since we are being called from the loop in
+ * schedule() via some task's ->post_select() function, freeing the
+ * given task here would result in use-after-free bugs in schedule().
+ * So we only set the task status to TS_DEAD which tells schedule() to
+ * free the task in the next iteration of its loop.
+ */
+ t->status = TS_DEAD;
+
+ *tptr = NULL;
+ return ret;
}
/**
- * Add a task to the scheduler.
+ * Deallocate all resources of all tasks of a scheduler instance.
*
- * \param t the task to add
+ * \param s The scheduler instance.
*
- * If the pre_select pointer of \a t is not \p NULL, it is added to
- * the pre_select list of the scheduler. Same goes for post_select.
- *
- * \sa task::pre_select, task::post_select
+ * This should only be called after \ref schedule() has returned.
*/
-void register_task(struct task *t)
+void sched_shutdown(struct sched *s)
{
- if (!initialized)
- init_sched();
- PARA_INFO_LOG("registering %s (%p)\n", t->status, t);
- if (t->pre_select) {
- PARA_DEBUG_LOG("pre_select: %p\n", &t->pre_select);
- list_add_tail(&t->pre_select_node, &pre_select_list);
- }
- if (t->post_select) {
- PARA_DEBUG_LOG("post_select: %p\n", &t->post_select);
- list_add_tail(&t->post_select_node, &post_select_list);
+ struct task *t, *tmp;
+
+ list_for_each_entry_safe(t, tmp, &s->task_list, node) {
+ if (t->status == TS_RUNNING)
+ /* The task list should contain only terminated tasks. */
+ PARA_WARNING_LOG("shutting down running task %s\n",
+ t->name);
+ unlink_and_free_task(t);
}
}
/**
- * Unregister all tasks.
+ * Add a task to the scheduler task list.
+ *
+ * \param info Task information supplied by the caller.
+ * \param s The scheduler instance.
*
- * This will cause \a schedule() to return immediately because both the
- * \a pre_select_list and the \a post_select_list are empty.
+ * \return A pointer to a newly allocated task structure. It will be
+ * freed by sched_shutdown().
*/
-void sched_shutdown(void)
+struct task *task_register(struct task_info *info, struct sched *s)
{
- struct task *t, *tmp;
+ struct task *t = para_malloc(sizeof(*t));
- if (!initialized)
- return;
- list_for_each_entry_safe(t, tmp, &pre_select_list, pre_select_node)
- unregister_task(t);
- list_for_each_entry_safe(t, tmp, &post_select_list, post_select_node)
- unregister_task(t);
- initialized = 0;
+ assert(info->post_select);
+
+ if (!s->task_list.next)
+ INIT_LIST_HEAD(&s->task_list);
+
+ t->info = *info;
+ t->name = para_strdup(info->name);
+ t->notification = 0;
+ t->status = TS_RUNNING;
+ list_add_tail(&t->node, &s->task_list);
+ return t;
}
/**
* Get the list of all registered tasks.
*
+ * \param s The scheduler instance to get the task list from.
+ *
* \return The task list.
*
- * Each entry of the list contains an identifier which is simply a hex number
- * that may be used in \a kill_task() to terminate the task.
+ * Each entry of the list contains an identifier which is simply a hex number.
* The result is dynamically allocated and must be freed by the caller.
*/
-char *get_task_list(void)
+char *get_task_list(struct sched *s)
{
struct task *t, *tmp;
char *msg = NULL;
- if (!initialized)
- return NULL;
- list_for_each_entry_safe(t, tmp, &pre_select_list, pre_select_node) {
+ list_for_each_entry_safe(t, tmp, &s->task_list, node) {
char *tmp_msg;
- tmp_msg = make_message("%s%p\tpre\t%s\n", msg? msg : "", t, t->status);
+ tmp_msg = make_message("%s%p\t%s\t%s\n", msg? msg : "", t,
+ t->status == TS_DEAD? "dead" :
+ (t->status == TS_RUNNING? "running" : "zombie"),
+ t->name);
free(msg);
msg = tmp_msg;
}
- list_for_each_entry_safe(t, tmp, &post_select_list, post_select_node) {
- char *tmp_msg;
-// if (t->pre_select)
-// continue;
- tmp_msg = make_message("%s%p\tpost\t%s\n", msg? msg : "", t, t->status);
- free(msg);
- msg = tmp_msg;
- }
- //PARA_DEBUG_LOG("task list:\n%s", msg);
return msg;
}
/**
- * Simulate an error for the given task.
+ * Set the notification value of a task.
+ *
+ * \param t The task to notify.
+ * \param err A positive error code.
+ *
+ * Tasks which honor notifications are supposed to call \ref
+ * task_get_notification() in their post_select function and act on the
+ * returned notification value.
*
- * \param id The task identifier.
+ * If the scheduler detects during its pre_select loop that at least one task
+ * has been notified, the loop terminates, and the post_select methods of all
+ * taks are immediately called again.
*
- * Find the task identified by \a id, set the tasks' error value to
- * \p -E_TASK_KILLED and unregister the task.
+ * The notification for a task is reset after the call to its post_select
+ * method.
*
- * \return Positive on success, negative on errors (e.g. if \a id does not
- * correspond to a registered task).
+ * \sa \ref task_get_notification().
*/
-int kill_task(char *id)
+void task_notify(struct task *t, int err)
{
- struct task *t, *tmp;
- char buf[20];
+ assert(err > 0);
+ if (t->notification == -err) /* ignore subsequent notifications */
+ return;
+ PARA_INFO_LOG("notifying task %s: %s\n", t->name, para_strerror(err));
+ t->notification = -err;
+}
- if (!initialized)
- return -E_NOT_INITIALIZED;
- list_for_each_entry_safe(t, tmp, &pre_select_list, pre_select_node) {
- sprintf(buf, "%p", t);
- if (strcmp(id, buf))
- continue;
- t->error = -E_TASK_KILLED;
- return 1;
- }
- list_for_each_entry_safe(t, tmp, &post_select_list, post_select_node) {
- sprintf(buf, "%p", t);
- if (strcmp(id, buf))
- continue;
- t->error = -E_TASK_KILLED;
+/**
+ * Return the notification value of a task.
+ *
+ * \param t The task to get the notification value from.
+ *
+ * \return The notification value. If this is negative, the task has been
+ * notified by another task. Tasks are supposed to check for notifications by
+ * calling this function from their post_select method.
+ *
+ * \sa \ref task_notify().
+ */
+int task_get_notification(const struct task *t)
+{
+ return t->notification;
+}
+
+/**
+ * Return the status value of a task.
+ *
+ * \param t The task to get the status value from.
+ *
+ * \return Zero if task does not exist, one if task is running, negative error
+ * code if task has terminated.
+ */
+int task_status(const struct task *t)
+{
+ if (!t)
+ return 0;
+ if (t->status == TS_DEAD) /* pretend dead tasks don't exist */
+ return 0;
+ if (t->status == TS_RUNNING)
return 1;
- }
- return -E_NO_SUCH_TASK;
+ return t->status;
+}
+
+/**
+ * Set the notification value of all tasks of a scheduler instance.
+ *
+ * \param s The scheduler instance whose tasks should be notified.
+ * \param err A positive error code.
+ *
+ * This simply iterates over all existing tasks of \a s and sets each
+ * task's notification value to \p -err.
+ */
+void task_notify_all(struct sched *s, int err)
+{
+ struct task *t;
+
+ list_for_each_entry(t, &s->task_list, node)
+ task_notify(t, err);
}
/**
*/
void sched_min_delay(struct sched *s)
{
- s->select_timeout.tv_sec = 0;
- s->select_timeout.tv_usec = 1;
+ s->select_timeout.tv_sec = s->select_timeout.tv_usec = 0;
}
/**
* \param barrier Absolute time before select() should return.
* \param s Pointer to the scheduler struct.
*
- * If \a barrier is in the past, this function does nothing.
+ * \return If \a barrier is in the past, this function does nothing and returns
+ * zero. Otherwise it returns one.
*
* \sa sched_request_barrier_or_min_delay().
*/
-void sched_request_barrier(struct timeval *barrier, struct sched *s)
+int sched_request_barrier(struct timeval *barrier, struct sched *s)
{
struct timeval diff;
if (tv_diff(now, barrier, &diff) > 0)
- return;
+ return 0;
sched_request_timeout(&diff, s);
+ return 1;
}
/**
* \param barrier Absolute time before select() should return.
* \param s Pointer to the scheduler struct.
*
- * If \a barrier is in the past, this function requests a minimal timeout.
+ * \return If \a barrier is in the past, this function requests a minimal
+ * timeout and returns zero. Otherwise it returns one.
*
* \sa sched_min_delay(), sched_request_barrier().
*/
-void sched_request_barrier_or_min_delay(struct timeval *barrier, struct sched *s)
+int sched_request_barrier_or_min_delay(struct timeval *barrier, struct sched *s)
{
struct timeval diff;
- if (tv_diff(now, barrier, &diff) > 0)
- return sched_min_delay(s);
+ if (tv_diff(now, barrier, &diff) > 0) {
+ sched_min_delay(s);
+ return 0;
+ }
sched_request_timeout(&diff, s);
+ return 1;
}