sched.c

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