sched.c

   1 /*
   2  * Copyright (C) 2006 Andre Noll <maan@tuebingen.mpg.de>
   3  *
   4  * Licensed under the GPL v2. For licencing details see COPYING.
   5  */
   6
   7 /** \file sched.c Paraslash's scheduling functions. */
   8
   9 #include <regex.h>
  10
  11 #include "para.h"
  12 #include "ipc.h"
  13 #include "fd.h"
  14 #include "list.h"
  15 #include "sched.h"
  16 #include "string.h"
  17 #include "time.h"
  18 #include "error.h"
  19
  20 /**
  21  * The possible states of a task.
  22  *
  23  * In addition to the states listed here, a task may also enter zombie state.
  24  * This happens when its ->post_select function returns negative, the ->status
  25  * field is then set to this return value. Such tasks are not scheduled any
  26  * more (i.e. ->pre_select() and ->post_select() are no longer called), but
  27  * they stay on the scheduler task list until \ref task_reap() or
  28  * \ref sched_shutdown() is called.
  29  */
  30 enum task_status {
  31         /** Task has been reaped and may be removed from the task list. */
  32         TS_DEAD,
  33         /** Task is active. */
  34         TS_RUNNING,
  35 };
  36
  37 struct task {
  38         /** A copy of the task name supplied when the task was registered. */
  39         char *name;
  40         /** Copied during task_register(). */
  41         struct task_info info;
  42         /* TS_RUNNING, TS_DEAD, or zombie (negative value). */
  43         int status;
  44         /** Position of the task in the task list of the scheduler. */
  45         struct list_head node;
  46         /** If less than zero, the task was notified by another task. */
  47         int notification;
  48 };
  49
  50 static struct timeval now_struct;
  51 const struct timeval *now = &now_struct;
  52
  53 static void sched_preselect(struct sched *s)
  54 {
  55         struct task *t, *tmp;
  56
  57         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
  58                 if (t->status < 0)
  59                         continue;
  60                 if (t->notification != 0)
  61                         sched_min_delay(s);
  62                 if (t->info.pre_select)
  63                         t->info.pre_select(s, t->info.context);
  64         }
  65 }
  66
  67 static void unlink_and_free_task(struct task *t)
  68 {
  69         PARA_INFO_LOG("freeing task %s\n", t->name);
  70         list_del(&t->node);
  71         free(t->name);
  72         free(t);
  73 }
  74
  75 //#define SCHED_DEBUG 1
  76 static inline void call_post_select(struct sched *s, struct task *t)
  77 {
  78         int ret;
  79
  80 #ifndef SCHED_DEBUG
  81         ret = t->info.post_select(s, t->info.context);
  82 #else
  83         struct timeval t1, t2, diff;
  84         unsigned long pst;
  85
  86         clock_get_realtime(&t1);
  87         ret = t->info.post_select(s, t->info.context);
  88         clock_get_realtime(&t2);
  89         tv_diff(&t1, &t2, &diff);
  90         pst = tv2ms(&diff);
  91         if (pst > 50)
  92                 PARA_WARNING_LOG("%s: post_select time: %lums\n",
  93                         t->name, pst);
  94 #endif
  95         t->status = ret < 0? ret : TS_RUNNING;
  96 }
  97
  98 static unsigned sched_post_select(struct sched *s)
  99 {
 100         struct task *t, *tmp;
 101         unsigned num_running_tasks = 0;
 102
 103         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 104                 if (t->status == TS_DEAD) /* task has been reaped */
 105                         unlink_and_free_task(t);
 106                 else if (t->status == TS_RUNNING) {
 107                         call_post_select(s, t); /* sets t->status */
 108                         t->notification = 0;
 109                         if (t->status == TS_RUNNING)
 110                                 num_running_tasks++;
 111                 }
 112         }
 113         return num_running_tasks;
 114 }
 115
 116 /**
 117  * The core function of all paraslash programs.
 118  *
 119  * \param s Pointer to the scheduler struct.
 120  *
 121  * This function updates the global \a now pointer, calls all registered
 122  * pre_select hooks which may set the timeout and add any file descriptors to
 123  * the fd sets of \a s.  Next, it calls para_select() and makes the result available
 124  * to the registered tasks by calling their post_select hook.
 125  *
 126  * \return Zero if no more tasks are left in the task list, negative if the
 127  * select function returned an error.
 128  *
 129  * \sa \ref now.
 130  */
 131 int schedule(struct sched *s)
 132 {
 133         int ret;
 134         unsigned num_running_tasks;
 135
 136         if (!s->select_function)
 137                 s->select_function = para_select;
 138 again:
 139         FD_ZERO(&s->rfds);
 140         FD_ZERO(&s->wfds);
 141         s->select_timeout = s->default_timeout;
 142         s->max_fileno = -1;
 143         clock_get_realtime(&now_struct);
 144         sched_preselect(s);
 145         ret = s->select_function(s->max_fileno + 1, &s->rfds, &s->wfds,
 146                 &s->select_timeout);
 147         if (ret < 0)
 148                 return ret;
 149         if (ret == 0) {
 150                 /*
 151                  * APUE: Be careful not to check the descriptor sets on return
 152                  * unless the return value is greater than zero. The return
 153                  * state of the descriptor sets is implementation dependent if
 154                  * either a signal is caught or the timer expires.
 155                  */
 156                 FD_ZERO(&s->rfds);
 157                 FD_ZERO(&s->wfds);
 158         }
 159         clock_get_realtime(&now_struct);
 160         num_running_tasks = sched_post_select(s);
 161         if (num_running_tasks == 0)
 162                 return 0;
 163         goto again;
 164 }
 165
 166 /**
 167  * Obtain the error status of a task and deallocate its resources.
 168  *
 169  * \param tptr Identifies the task to reap.
 170  *
 171  * This function is similar to wait(2) in that it returns information about a
 172  * terminated task and allows to release the resources associated with the
 173  * task. Until this function is called, the terminated task remains in a zombie
 174  * state.
 175  *
 176  * \return If \a tptr is \p NULL, or \a *tptr is \p NULL, the function does
 177  * nothing and returns zero. Otherwise, it is checked whether the task
 178  * identified by \a tptr is still running. If it is, the function returns zero
 179  * and again, no action is taken. Otherwise the (negative) error code of the
 180  * terminated task is returned and \a *tptr is set to \p NULL. The task will
 181  * then be removed removed from the scheduler task list.
 182  *
 183  * \sa \ref sched_shutdown(), wait(2).
 184  */
 185 int task_reap(struct task **tptr)
 186 {
 187         struct task *t;
 188         int ret;
 189
 190         if (!tptr)
 191                 return 0;
 192         t = *tptr;
 193         if (!t)
 194                 return 0;
 195         if (t->status >= 0)
 196                 return 0;
 197         ret = t->status;
 198         /*
 199          * With list_for_each_entry_safe() it is only safe to remove the
 200          * _current_ list item. Since we are being called from the loop in
 201          * schedule() via some task's ->post_select() function, freeing the
 202          * given task here would result in use-after-free bugs in schedule().
 203          * So we only set the task status to TS_DEAD which tells schedule() to
 204          * free the task in the next iteration of its loop.
 205          */
 206         t->status = TS_DEAD;
 207
 208         *tptr = NULL;
 209         return ret;
 210 }
 211
 212 /**
 213  * Deallocate all resources of all tasks of a scheduler instance.
 214  *
 215  * \param s The scheduler instance.
 216  *
 217  * This should only be called after \ref schedule() has returned.
 218  */
 219 void sched_shutdown(struct sched *s)
 220 {
 221         struct task *t, *tmp;
 222
 223         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 224                 if (t->status == TS_RUNNING)
 225                         /* The task list should contain only terminated tasks. */
 226                         PARA_WARNING_LOG("shutting down running task %s\n",
 227                                 t->name);
 228                 unlink_and_free_task(t);
 229         }
 230 }
 231
 232 /**
 233  * Add a task to the scheduler task list.
 234  *
 235  * \param info Task information supplied by the caller.
 236  * \param s The scheduler instance.
 237  *
 238  * \return A pointer to a newly allocated task structure. It will be
 239  * freed by sched_shutdown().
 240  */
 241 struct task *task_register(struct task_info *info, struct sched *s)
 242 {
 243         struct task *t = para_malloc(sizeof(*t));
 244
 245         assert(info->post_select);
 246
 247         if (!s->task_list.next)
 248                 INIT_LIST_HEAD(&s->task_list);
 249
 250         t->info = *info;
 251         t->name = para_strdup(info->name);
 252         t->notification = 0;
 253         t->status = TS_RUNNING;
 254         list_add_tail(&t->node, &s->task_list);
 255         return t;
 256 }
 257
 258 /**
 259  * Get the list of all registered tasks.
 260  *
 261  * \param s The scheduler instance to get the task list from.
 262  *
 263  * \return The task list.
 264  *
 265  * Each entry of the list contains an identifier which is simply a hex number.
 266  * The result is dynamically allocated and must be freed by the caller.
 267  */
 268 char *get_task_list(struct sched *s)
 269 {
 270         struct task *t, *tmp;
 271         char *msg = NULL;
 272
 273         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 274                 char *tmp_msg;
 275                 tmp_msg = make_message("%s%p\t%s\t%s\n", msg? msg : "", t,
 276                         t->status == TS_DEAD? "dead" :
 277                                 (t->status == TS_RUNNING? "running" : "zombie"),
 278                         t->name);
 279                 free(msg);
 280                 msg = tmp_msg;
 281         }
 282         return msg;
 283 }
 284
 285 /**
 286  * Set the notification value of a task.
 287  *
 288  * \param t The task to notify.
 289  * \param err A positive error code.
 290  *
 291  * Tasks which honor notifications are supposed to call \ref
 292  * task_get_notification() in their post_select function and act on the
 293  * returned notification value.
 294  *
 295  * If the scheduler detects during its pre_select loop that at least one task
 296  * has been notified, the loop terminates, and the post_select methods of all
 297  * taks are immediately called again.
 298  *
 299  * The notification for a task is reset after the call to its post_select
 300  * method.
 301  *
 302  * \sa \ref task_get_notification().
 303  */
 304 void task_notify(struct task *t, int err)
 305 {
 306         assert(err > 0);
 307         if (t->notification == -err) /* ignore subsequent notifications */
 308                 return;
 309         PARA_INFO_LOG("notifying task %s: %s\n", t->name, para_strerror(err));
 310         t->notification = -err;
 311 }
 312
 313 /**
 314  * Return the notification value of a task.
 315  *
 316  * \param t The task to get the notification value from.
 317  *
 318  * \return The notification value. If this is negative, the task has been
 319  * notified by another task. Tasks are supposed to check for notifications by
 320  * calling this function from their post_select method.
 321  *
 322  * \sa \ref task_notify().
 323  */
 324 int task_get_notification(const struct task *t)
 325 {
 326         return t->notification;
 327 }
 328
 329 /**
 330  * Return the status value of a task.
 331  *
 332  * \param t The task to get the status value from.
 333  *
 334  * \return Zero if task does not exist, one if task is running, negative error
 335  * code if task has terminated.
 336  */
 337 int task_status(const struct task *t)
 338 {
 339         if (!t)
 340                 return 0;
 341         if (t->status == TS_DEAD) /* pretend dead tasks don't exist */
 342                 return 0;
 343         if (t->status == TS_RUNNING)
 344                 return 1;
 345         return t->status;
 346 }
 347
 348 /**
 349  * Set the notification value of all tasks of a scheduler instance.
 350  *
 351  * \param s The scheduler instance whose tasks should be notified.
 352  * \param err A positive error code.
 353  *
 354  * This simply iterates over all existing tasks of \a s and sets each
 355  * task's notification value to \p -err.
 356  */
 357 void task_notify_all(struct sched *s, int err)
 358 {
 359         struct task *t;
 360
 361         list_for_each_entry(t, &s->task_list, node)
 362                 task_notify(t, err);
 363 }
 364
 365 /**
 366  * Set the select timeout to the minimal possible value.
 367  *
 368  * \param s Pointer to the scheduler struct.
 369  *
 370  * This causes the next select() call to return immediately.
 371  */
 372 void sched_min_delay(struct sched *s)
 373 {
 374         s->select_timeout.tv_sec = s->select_timeout.tv_usec = 0;
 375 }
 376
 377 /**
 378  * Impose an upper bound for the timeout of the next select() call.
 379  *
 380  * \param to Maximal allowed timeout.
 381  * \param s Pointer to the scheduler struct.
 382  *
 383  * If the current scheduler timeout is already smaller than \a to, this
 384  * function does nothing. Otherwise the timeout for the next select() call is
 385  * set to the given value.
 386  *
 387  * \sa sched_request_timeout_ms().
 388  */
 389 void sched_request_timeout(struct timeval *to, struct sched *s)
 390 {
 391         if (tv_diff(&s->select_timeout, to, NULL) > 0)
 392                 s->select_timeout = *to;
 393 }
 394
 395 /**
 396  * Force the next select() call to return before the given amount of milliseconds.
 397  *
 398  * \param ms The maximal allowed timeout in milliseconds.
 399  * \param s Pointer to the scheduler struct.
 400  *
 401  * Like sched_request_timeout() this imposes an upper bound on the timeout
 402  * value for the next select() call.
 403  */
 404 void sched_request_timeout_ms(long unsigned ms, struct sched *s)
 405 {
 406         struct timeval tv;
 407         ms2tv(ms, &tv);
 408         sched_request_timeout(&tv, s);
 409 }
 410
 411 /**
 412  * Force the next select() call to return before the given future time.
 413  *
 414  * \param barrier Absolute time before select() should return.
 415  * \param s Pointer to the scheduler struct.
 416  *
 417  * \return If \a barrier is in the past, this function does nothing and returns
 418  * zero. Otherwise it returns one.
 419  *
 420  * \sa sched_request_barrier_or_min_delay().
 421  */
 422 int sched_request_barrier(struct timeval *barrier, struct sched *s)
 423 {
 424         struct timeval diff;
 425
 426         if (tv_diff(now, barrier, &diff) > 0)
 427                 return 0;
 428         sched_request_timeout(&diff, s);
 429         return 1;
 430 }
 431
 432 /**
 433  * Force the next select() call to return before the given time.
 434  *
 435  * \param barrier Absolute time before select() should return.
 436  * \param s Pointer to the scheduler struct.
 437  *
 438  * \return If \a barrier is in the past, this function requests a minimal
 439  * timeout and returns zero. Otherwise it returns one.
 440  *
 441  * \sa sched_min_delay(), sched_request_barrier().
 442  */
 443 int sched_request_barrier_or_min_delay(struct timeval *barrier, struct sched *s)
 444 {
 445         struct timeval diff;
 446
 447         if (tv_diff(now, barrier, &diff) > 0) {
 448                 sched_min_delay(s);
 449                 return 0;
 450         }
 451         sched_request_timeout(&diff, s);
 452         return 1;
 453 }