sched.c

   1 /*
   2  * Copyright (C) 2006-2014 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 #include <assert.h>
  11
  12 #include "para.h"
  13 #include "ipc.h"
  14 #include "fd.h"
  15 #include "list.h"
  16 #include "sched.h"
  17 #include "string.h"
  18 #include "time.h"
  19 #include "error.h"
  20
  21 /**
  22  * The possible states of a task.
  23  *
  24  * In addition to the states listed here, a task may also enter zombie state.
  25  * This happens when its ->post_select function returns negative, the ->status
  26  * field is then set to this return value. Such tasks are not scheduled any
  27  * more (i.e. ->pre_select() and ->post_select() are no longer called), but
  28  * they stay on the scheduler task list until \ref task_reap() or
  29  * \ref sched_shutdown() is called.
  30  */
  31 enum task_status {
  32         /** Task has been reaped and may be removed from the task list. */
  33         TS_DEAD,
  34         /** Task is active. */
  35         TS_RUNNING,
  36 };
  37
  38 struct task {
  39         /** A copy of the task name supplied when the task was registered. */
  40         char *name;
  41         /** Copied during task_register(). */
  42         struct task_info info;
  43         /* TS_RUNNING, TS_DEAD, or zombie (negative value). */
  44         int status;
  45         /** Position of the task in the task list of the scheduler. */
  46         struct list_head node;
  47         /** If less than zero, the task was notified by another task. */
  48         int notification;
  49 };
  50
  51 static struct timeval now_struct;
  52 const struct timeval *now = &now_struct;
  53
  54 static inline bool timeout_is_zero(struct sched *s)
  55 {
  56         struct timeval *tv = &s->select_timeout;
  57         return tv->tv_sec == 0 && tv->tv_usec == 0;
  58 }
  59
  60 static void sched_preselect(struct sched *s)
  61 {
  62         struct task *t, *tmp;
  63
  64         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
  65                 if (t->status < 0)
  66                         continue;
  67                 if (t->notification != 0)
  68                         sched_min_delay(s);
  69                 if (t->info.pre_select)
  70                         t->info.pre_select(s, t->info.context);
  71         }
  72 }
  73
  74 static void unlink_and_free_task(struct task *t)
  75 {
  76         PARA_INFO_LOG("freeing task %s\n", t->name);
  77         list_del(&t->node);
  78         free(t->name);
  79         free(t);
  80 }
  81
  82 //#define SCHED_DEBUG 1
  83 static inline void call_post_select(struct sched *s, struct task *t)
  84 {
  85         int ret;
  86
  87 #ifndef SCHED_DEBUG
  88         ret = t->info.post_select(s, t->info.context);
  89 #else
  90         struct timeval t1, t2, diff;
  91         unsigned long pst;
  92
  93         clock_get_realtime(&t1);
  94         ret = t->info.post_select(s, t->info.context);
  95         clock_get_realtime(&t2);
  96         tv_diff(&t1, &t2, &diff);
  97         pst = tv2ms(&diff);
  98         if (pst > 50)
  99                 PARA_WARNING_LOG("%s: post_select time: %lums\n",
 100                         t->name, pst);
 101 #endif
 102         t->status = ret < 0? ret : TS_RUNNING;
 103 }
 104
 105 static unsigned sched_post_select(struct sched *s)
 106 {
 107         struct task *t, *tmp;
 108         unsigned num_running_tasks = 0;
 109
 110         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 111                 if (t->status == TS_DEAD) /* task has been reaped */
 112                         unlink_and_free_task(t);
 113                 else if (t->status == TS_RUNNING) {
 114                         call_post_select(s, t); /* sets t->status */
 115                         t->notification = 0;
 116                         if (t->status == TS_RUNNING)
 117                                 num_running_tasks++;
 118                 }
 119         }
 120         return num_running_tasks;
 121 }
 122
 123 /**
 124  * The core function of all paraslash programs.
 125  *
 126  * \param s Pointer to the scheduler struct.
 127  *
 128  * This function updates the global \a now pointer, calls all registered
 129  * pre_select hooks which may set the timeout and add any file descriptors to
 130  * the fd sets of \a s.  Next, it calls para_select() and makes the result available
 131  * to the registered tasks by calling their post_select hook.
 132  *
 133  * \return Zero if no more tasks are left in the task list, negative if the
 134  * select function returned an error.
 135  *
 136  * \sa \ref now.
 137  */
 138 int schedule(struct sched *s)
 139 {
 140         int ret;
 141         unsigned num_running_tasks;
 142
 143         if (!s->select_function)
 144                 s->select_function = para_select;
 145 again:
 146         FD_ZERO(&s->rfds);
 147         FD_ZERO(&s->wfds);
 148         s->select_timeout = s->default_timeout;
 149         s->max_fileno = -1;
 150         clock_get_realtime(&now_struct);
 151         sched_preselect(s);
 152         ret = s->select_function(s->max_fileno + 1, &s->rfds, &s->wfds,
 153                 &s->select_timeout);
 154         if (ret < 0)
 155                 return ret;
 156         if (ret == 0) {
 157                 /*
 158                  * APUE: Be careful not to check the descriptor sets on return
 159                  * unless the return value is greater than zero. The return
 160                  * state of the descriptor sets is implementation dependent if
 161                  * either a signal is caught or the timer expires.
 162                  */
 163                 FD_ZERO(&s->rfds);
 164                 FD_ZERO(&s->wfds);
 165         }
 166         clock_get_realtime(&now_struct);
 167         num_running_tasks = sched_post_select(s);
 168         if (num_running_tasks == 0)
 169                 return 0;
 170         goto again;
 171 }
 172
 173 /**
 174  * Obtain the error status of a task and deallocate its resources.
 175  *
 176  * \param tptr Identifies the task to reap.
 177  *
 178  * This function is similar to wait(2) in that it returns information about a
 179  * terminated task and allows to release the resources associated with the
 180  * task. Until this function is called, the terminated task remains in a zombie
 181  * state.
 182  *
 183  * \return If \a tptr is \p NULL, or \a *tptr is \p NULL, the function does
 184  * nothing and returns zero. Otherwise, it is checked whether the task
 185  * identified by \a tptr is still running. If it is, the function returns zero
 186  * and again, no action is taken. Otherwise the (negative) error code of the
 187  * terminated task is returned and \a *tptr is set to \p NULL. The task will
 188  * then be removed removed from the scheduler task list.
 189  *
 190  * \sa \ref sched_shutdown(), wait(2).
 191  */
 192 int task_reap(struct task **tptr)
 193 {
 194         struct task *t;
 195         int ret;
 196
 197         if (!tptr)
 198                 return 0;
 199         t = *tptr;
 200         if (!t)
 201                 return 0;
 202         if (t->status >= 0)
 203                 return 0;
 204         ret = t->status;
 205         /*
 206          * With list_for_each_entry_safe() it is only safe to remove the
 207          * _current_ list item. Since we are being called from the loop in
 208          * schedule() via some task's ->post_select() function, freeing the
 209          * given task here would result in use-after-free bugs in schedule().
 210          * So we only set the task status to TS_DEAD which tells schedule() to
 211          * free the task in the next iteration of its loop.
 212          */
 213         t->status = TS_DEAD;
 214
 215         *tptr = NULL;
 216         return ret;
 217 }
 218
 219 /**
 220  * Deallocate all resources of all tasks of a scheduler instance.
 221  *
 222  * \param s The scheduler instance.
 223  *
 224  * This should only be called after \ref schedule() has returned.
 225  */
 226 void sched_shutdown(struct sched *s)
 227 {
 228         struct task *t, *tmp;
 229
 230         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 231                 if (t->status == TS_RUNNING)
 232                         /* The task list should contain only terminated tasks. */
 233                         PARA_WARNING_LOG("shutting down running task %s\n",
 234                                 t->name);
 235                 unlink_and_free_task(t);
 236         }
 237 }
 238
 239 /**
 240  * Add a task to the scheduler task list.
 241  *
 242  * \param info Task information supplied by the caller.
 243  * \param s The scheduler instance.
 244  *
 245  * \return A pointer to a newly allocated task structure. It will be
 246  * freed by sched_shutdown().
 247  */
 248 struct task *task_register(struct task_info *info, struct sched *s)
 249 {
 250         struct task *t = para_malloc(sizeof(*t));
 251
 252         assert(info->post_select);
 253
 254         if (!s->task_list.next)
 255                 INIT_LIST_HEAD(&s->task_list);
 256
 257         t->info = *info;
 258         t->name = para_strdup(info->name);
 259         t->notification = 0;
 260         t->status = TS_RUNNING;
 261         list_add_tail(&t->node, &s->task_list);
 262         return t;
 263 }
 264
 265 /**
 266  * Get the list of all registered tasks.
 267  *
 268  * \param s The scheduler instance to get the task list from.
 269  *
 270  * \return The task list.
 271  *
 272  * Each entry of the list contains an identifier which is simply a hex number.
 273  * The result is dynamically allocated and must be freed by the caller.
 274  */
 275 char *get_task_list(struct sched *s)
 276 {
 277         struct task *t, *tmp;
 278         char *msg = NULL;
 279
 280         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 281                 char *tmp_msg;
 282                 tmp_msg = make_message("%s%p\t%s\t%s\n", msg? msg : "", t,
 283                         t->status == TS_DEAD? "dead" :
 284                                 (t->status == TS_RUNNING? "running" : "zombie"),
 285                         t->name);
 286                 free(msg);
 287                 msg = tmp_msg;
 288         }
 289         return msg;
 290 }
 291
 292 /**
 293  * Set the notification value of a task.
 294  *
 295  * \param t The task to notify.
 296  * \param err A positive error code.
 297  *
 298  * Tasks which honor notifications are supposed to call \ref
 299  * task_get_notification() in their post_select function and act on the
 300  * returned notification value.
 301  *
 302  * If the scheduler detects during its pre_select loop that at least one task
 303  * has been notified, the loop terminates, and the post_select methods of all
 304  * taks are immediately called again.
 305  *
 306  * The notification for a task is reset after the call to its post_select
 307  * method.
 308  *
 309  * \sa \ref task_get_notification().
 310  */
 311 void task_notify(struct task *t, int err)
 312 {
 313         assert(err > 0);
 314         if (t->notification == -err) /* ignore subsequent notifications */
 315                 return;
 316         PARA_INFO_LOG("notifying task %s: %s\n", t->name, para_strerror(err));
 317         t->notification = -err;
 318 }
 319
 320 /**
 321  * Return the notification value of a task.
 322  *
 323  * \param t The task to get the notification value from.
 324  *
 325  * \return The notification value. If this is negative, the task has been
 326  * notified by another task. Tasks are supposed to check for notifications by
 327  * calling this function from their post_select method.
 328  *
 329  * \sa \ref task_notify().
 330  */
 331 int task_get_notification(const struct task *t)
 332 {
 333         return t->notification;
 334 }
 335
 336 /**
 337  * Return the status value of a task.
 338  *
 339  * \param t The task to get the status value from.
 340  *
 341  * \return Zero if task does not exist, one if task is running, negative error
 342  * code if task has terminated.
 343  */
 344 int task_status(const struct task *t)
 345 {
 346         if (!t)
 347                 return 0;
 348         if (t->status == TS_DEAD) /* pretend dead tasks don't exist */
 349                 return 0;
 350         if (t->status == TS_RUNNING)
 351                 return 1;
 352         return t->status;
 353 }
 354
 355 /**
 356  * Set the notification value of all tasks of a scheduler instance.
 357  *
 358  * \param s The scheduler instance whose tasks should be notified.
 359  * \param err A positive error code.
 360  *
 361  * This simply iterates over all existing tasks of \a s and sets each
 362  * task's notification value to \p -err.
 363  */
 364 void task_notify_all(struct sched *s, int err)
 365 {
 366         struct task *t;
 367
 368         list_for_each_entry(t, &s->task_list, node)
 369                 task_notify(t, err);
 370 }
 371
 372 /**
 373  * Set the select timeout to the minimal possible value.
 374  *
 375  * \param s Pointer to the scheduler struct.
 376  *
 377  * This causes the next select() call to return immediately.
 378  */
 379 void sched_min_delay(struct sched *s)
 380 {
 381         s->select_timeout.tv_sec = s->select_timeout.tv_usec = 0;
 382 }
 383
 384 /**
 385  * Impose an upper bound for the timeout of the next select() call.
 386  *
 387  * \param to Maximal allowed timeout.
 388  * \param s Pointer to the scheduler struct.
 389  *
 390  * If the current scheduler timeout is already smaller than \a to, this
 391  * function does nothing. Otherwise the timeout for the next select() call is
 392  * set to the given value.
 393  *
 394  * \sa sched_request_timeout_ms().
 395  */
 396 void sched_request_timeout(struct timeval *to, struct sched *s)
 397 {
 398         if (tv_diff(&s->select_timeout, to, NULL) > 0)
 399                 s->select_timeout = *to;
 400 }
 401
 402 /**
 403  * Force the next select() call to return before the given amount of milliseconds.
 404  *
 405  * \param ms The maximal allowed timeout in milliseconds.
 406  * \param s Pointer to the scheduler struct.
 407  *
 408  * Like sched_request_timeout() this imposes an upper bound on the timeout
 409  * value for the next select() call.
 410  */
 411 void sched_request_timeout_ms(long unsigned ms, struct sched *s)
 412 {
 413         struct timeval tv;
 414         ms2tv(ms, &tv);
 415         sched_request_timeout(&tv, s);
 416 }
 417
 418 /**
 419  * Force the next select() call to return before the given future time.
 420  *
 421  * \param barrier Absolute time before select() should return.
 422  * \param s Pointer to the scheduler struct.
 423  *
 424  * \return If \a barrier is in the past, this function does nothing and returns
 425  * zero. Otherwise it returns one.
 426  *
 427  * \sa sched_request_barrier_or_min_delay().
 428  */
 429 int sched_request_barrier(struct timeval *barrier, struct sched *s)
 430 {
 431         struct timeval diff;
 432
 433         if (tv_diff(now, barrier, &diff) > 0)
 434                 return 0;
 435         sched_request_timeout(&diff, s);
 436         return 1;
 437 }
 438
 439 /**
 440  * Force the next select() call to return before the given time.
 441  *
 442  * \param barrier Absolute time before select() should return.
 443  * \param s Pointer to the scheduler struct.
 444  *
 445  * \return If \a barrier is in the past, this function requests a minimal
 446  * timeout and returns zero. Otherwise it returns one.
 447  *
 448  * \sa sched_min_delay(), sched_request_barrier().
 449  */
 450 int sched_request_barrier_or_min_delay(struct timeval *barrier, struct sched *s)
 451 {
 452         struct timeval diff;
 453
 454         if (tv_diff(now, barrier, &diff) > 0) {
 455                 sched_min_delay(s);
 456                 return 0;
 457         }
 458         sched_request_timeout(&diff, s);
 459         return 1;
 460 }