sched.c

   1 /*
   2  * Copyright (C) 2006-2014 Andre Noll <maan@systemlinux.org>
   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 struct task {
  22         /** The task name supplied when the task was registered(). */
  23         char name[255];
  24         /** Copied from the task_info struct during task_register(). */
  25         void (*pre_select)(struct sched *s, struct task *t);
  26         /** Copied from the task_info struct during task_register(). */
  27         int (*post_select)(struct sched *s, struct task *t);
  28         /** Whether this task is active (>=0) or in error state (<0). */
  29         int status;
  30         /** Position of the task in the task list of the scheduler. */
  31         struct list_head node;
  32         /** If less than zero, the task was notified by another task. */
  33         int notification;
  34         /** True if task is in error state and exit status has been queried. */
  35         bool dead;
  36         /** Usually a pointer to the struct containing this task. */
  37         void *context;
  38 };
  39
  40 static struct timeval now_struct;
  41 struct timeval *now = &now_struct;
  42
  43 static inline bool timeout_is_zero(struct sched *s)
  44 {
  45         struct timeval *tv = &s->select_timeout;
  46         return tv->tv_sec == 0 && tv->tv_usec == 0;
  47 }
  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->pre_select)
  59                         t->pre_select(s, t);
  60         }
  61 }
  62
  63 static void unlink_and_free_task(struct task *t)
  64 {
  65         PARA_INFO_LOG("freeing task %s\n", t->name);
  66         list_del(&t->node);
  67         free(t);
  68 }
  69
  70 //#define SCHED_DEBUG 1
  71 static inline void call_post_select(struct sched *s, struct task *t)
  72 {
  73 #ifndef SCHED_DEBUG
  74         t->status = t->post_select(s, t);
  75 #else
  76         struct timeval t1, t2, diff;
  77         unsigned long pst;
  78
  79         clock_get_realtime(&t1);
  80         t->status = t->post_select(s, t);
  81         clock_get_realtime(&t2);
  82         tv_diff(&t1, &t2, &diff);
  83         pst = tv2ms(&diff);
  84         if (pst > 50)
  85                 PARA_WARNING_LOG("%s: post_select time: %lums\n",
  86                         t->name, pst);
  87 #endif
  88 }
  89
  90 static unsigned sched_post_select(struct sched *s)
  91 {
  92         struct task *t, *tmp;
  93         unsigned num_running_tasks = 0;
  94
  95         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
  96                 if (t->status < 0) {
  97                         if (t->dead) /* task has been reaped */
  98                                 unlink_and_free_task(t);
  99                         continue;
 100                 }
 101                 call_post_select(s, t);
 102                 t->notification = 0;
 103                 if (t->status >= 0)
 104                         num_running_tasks++;
 105         }
 106         return num_running_tasks;
 107 }
 108
 109 /**
 110  * The core function of all paraslash programs.
 111  *
 112  * \param s Pointer to the scheduler struct.
 113  *
 114  * This function updates the global \a now pointer, calls all registered
 115  * pre_select hooks which may set the timeout and add any file descriptors to
 116  * the fd sets of \a s.  Next, it calls para_select() and makes the result available
 117  * to the registered tasks by calling their post_select hook.
 118  *
 119  * \return Zero if no more tasks are left in the task list, negative if the
 120  * select function returned an error.
 121  *
 122  * \sa \ref task, \ref now.
 123  */
 124 int schedule(struct sched *s)
 125 {
 126         int ret;
 127         unsigned num_running_tasks;
 128
 129         if (!s->select_function)
 130                 s->select_function = para_select;
 131 again:
 132         FD_ZERO(&s->rfds);
 133         FD_ZERO(&s->wfds);
 134         s->select_timeout = s->default_timeout;
 135         s->max_fileno = -1;
 136         clock_get_realtime(now);
 137         sched_preselect(s);
 138         ret = s->select_function(s->max_fileno + 1, &s->rfds, &s->wfds,
 139                 &s->select_timeout);
 140         if (ret < 0)
 141                 return ret;
 142         if (ret == 0) {
 143                 /*
 144                  * APUE: Be careful not to check the descriptor sets on return
 145                  * unless the return value is greater than zero. The return
 146                  * state of the descriptor sets is implementation dependent if
 147                  * either a signal is caught or the timer expires.
 148                  */
 149                 FD_ZERO(&s->rfds);
 150                 FD_ZERO(&s->wfds);
 151         }
 152         clock_get_realtime(now);
 153         num_running_tasks = sched_post_select(s);
 154         if (num_running_tasks == 0)
 155                 return 0;
 156         goto again;
 157 }
 158
 159 /**
 160  * Obtain the error status of a task and deallocate its resources.
 161  *
 162  * \param tptr Identifies the task to reap.
 163  *
 164  * This function is similar to wait(2) in that it returns information about a
 165  * terminated task and allows to release the resources associated with the
 166  * task. Until this function is called, the terminated task remains in a zombie
 167  * state.
 168  *
 169  * \return If \a tptr is \p NULL, or \a *tptr is \p NULL, the function does
 170  * nothing and returns zero. Otherwise, it is checked whether the task
 171  * identified by \a tptr is still running. If it is, the function returns zero
 172  * and again, no action is taken. Otherwise the (negative) error code of the
 173  * terminated task is returned and \a *tptr is set to \p NULL. The task will
 174  * then be removed removed from the scheduler task list.
 175  *
 176  * \sa \ref sched_shutdown(), wait(2).
 177  */
 178 int task_reap(struct task **tptr)
 179 {
 180         struct task *t;
 181
 182         if (!tptr)
 183                 return 0;
 184         t = *tptr;
 185         if (!t)
 186                 return 0;
 187         if (t->status >= 0)
 188                 return 0;
 189         if (t->dead) /* will be freed in sched_post_select() */
 190                 return 0;
 191         /*
 192          * With list_for_each_entry_safe() it is only safe to remove the
 193          * _current_ list item. Since we are being called from the loop in
 194          * schedule() via some task's ->post_select() function, freeing the
 195          * given task here would result in use-after-free bugs in schedule().
 196          * So we only set t->dead which tells schedule() to free the task in
 197          * the next iteration of its loop.
 198          */
 199         t->dead = true;
 200         *tptr = NULL;
 201         return t->status;
 202 }
 203
 204 /**
 205  * Deallocate all resources of all tasks of a scheduler instance.
 206  *
 207  * \param s The scheduler instance.
 208  *
 209  * This should only be called after \ref schedule() has returned.
 210  */
 211 void sched_shutdown(struct sched *s)
 212 {
 213         struct task *t, *tmp;
 214
 215         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 216                 if (t->status >= 0)
 217                         /* The task list should contain only terminated tasks. */
 218                         PARA_WARNING_LOG("shutting down running task %s\n",
 219                                 t->name);
 220                 unlink_and_free_task(t);
 221         }
 222 }
 223
 224 /**
 225  * Add a task to the scheduler task list.
 226  *
 227  * \param info Task information supplied by the caller.
 228  * \param s The scheduler instance.
 229  *
 230  * \return A pointer to a newly allocated task structure. It will be
 231  * freed by sched_shutdown().
 232  */
 233 struct task *task_register(struct task_info *info, struct sched *s)
 234 {
 235         struct task *t = para_malloc(sizeof(*t));
 236
 237         assert(info->post_select);
 238
 239         if (!s->task_list.next)
 240                 INIT_LIST_HEAD(&s->task_list);
 241
 242         snprintf(t->name, sizeof(t->name) - 1, "%s", info->name);
 243         t->name[sizeof(t->name) - 1] = '\0';
 244         t->notification = 0;
 245         t->status = 0;
 246         t->dead = false;
 247         t->pre_select = info->pre_select;
 248         t->post_select = info->post_select;
 249         t->context = info->context;
 250         list_add_tail(&t->node, &s->task_list);
 251         return t;
 252 }
 253
 254 /**
 255  * Obtain the context pointer of a task.
 256  *
 257  * \param t Return this task's context pointer.
 258  *
 259  * \return A pointer to the memory location specified previously as \a
 260  * task_info->context when the task was registered with \ref task_register().
 261  */
 262 void *task_context(struct task *t)
 263 {
 264         return t->context;
 265 }
 266
 267 /**
 268  * Get the list of all registered tasks.
 269  *
 270  * \param s The scheduler instance to get the task list from.
 271  *
 272  * \return The task list.
 273  *
 274  * Each entry of the list contains an identifier which is simply a hex number.
 275  * The result is dynamically allocated and must be freed by the caller.
 276  */
 277 char *get_task_list(struct sched *s)
 278 {
 279         struct task *t, *tmp;
 280         char *msg = NULL;
 281
 282         list_for_each_entry_safe(t, tmp, &s->task_list, node) {
 283                 char *tmp_msg;
 284                 tmp_msg = make_message("%s%p\t%s\t%s\n", msg? msg : "", t,
 285                         t->status < 0? (t->dead? "dead" : "zombie") : "running",
 286                         t->name);
 287                 free(msg);
 288                 msg = tmp_msg;
 289         }
 290         return msg;
 291 }
 292
 293 /**
 294  * Set the notification value of a task.
 295  *
 296  * \param t The task to notify.
 297  * \param err A positive error code.
 298  *
 299  * Tasks which honor notifications are supposed to call \ref
 300  * task_get_notification() in their post_select function and act on the
 301  * returned notification value.
 302  *
 303  * If the scheduler detects during its pre_select loop that at least one task
 304  * has been notified, the loop terminates, and the post_select methods of all
 305  * taks are immediately called again.
 306  *
 307  * The notification for a task is reset after the call to its post_select
 308  * method.
 309  *
 310  * \sa \ref task_get_notification().
 311  */
 312 void task_notify(struct task *t, int err)
 313 {
 314         assert(err > 0);
 315         if (t->notification == -err) /* ignore subsequent notifications */
 316                 return;
 317         PARA_INFO_LOG("notifying task %s: %s\n", t->name, para_strerror(err));
 318         t->notification = -err;
 319 }
 320
 321 /**
 322  * Return the notification value of a task.
 323  *
 324  * \param t The task to get the notification value from.
 325  *
 326  * \return The notification value. If this is negative, the task has been
 327  * notified by another task. Tasks are supposed to check for notifications by
 328  * calling this function from their post_select method.
 329  *
 330  * \sa \ref task_notify().
 331  */
 332 int task_get_notification(const struct task *t)
 333 {
 334         return t->notification;
 335 }
 336
 337 /**
 338  * Return the status value of a task.
 339  *
 340  * \param t The task to get the status value from.
 341  *
 342  * \return Zero if task does not exist, one if task is running, negative error
 343  * code if task has terminated.
 344  */
 345 int task_status(const struct task *t)
 346 {
 347         if (!t)
 348                 return 0;
 349         if (t->dead)
 350                 return 0;
 351         if (t->status >= 0)
 352                 return 1;
 353         return t->status;
 354 }
 355
 356 /**
 357  * Set the notification value of all tasks of a scheduler instance.
 358  *
 359  * \param s The scheduler instance whose tasks should be notified.
 360  * \param err A positive error code.
 361  *
 362  * This simply iterates over all existing tasks of \a s and sets each
 363  * task's notification value to \p -err.
 364  */
 365 void task_notify_all(struct sched *s, int err)
 366 {
 367         struct task *t;
 368
 369         list_for_each_entry(t, &s->task_list, node)
 370                 task_notify(t, err);
 371 }
 372
 373 /**
 374  * Set the select timeout to the minimal possible value.
 375  *
 376  * \param s Pointer to the scheduler struct.
 377  *
 378  * This causes the next select() call to return immediately.
 379  */
 380 void sched_min_delay(struct sched *s)
 381 {
 382         s->select_timeout.tv_sec = s->select_timeout.tv_usec = 0;
 383 }
 384
 385 /**
 386  * Impose an upper bound for the timeout of the next select() call.
 387  *
 388  * \param to Maximal allowed timeout.
 389  * \param s Pointer to the scheduler struct.
 390  *
 391  * If the current scheduler timeout is already smaller than \a to, this
 392  * function does nothing. Otherwise the timeout for the next select() call is
 393  * set to the given value.
 394  *
 395  * \sa sched_request_timeout_ms().
 396  */
 397 void sched_request_timeout(struct timeval *to, struct sched *s)
 398 {
 399         if (tv_diff(&s->select_timeout, to, NULL) > 0)
 400                 s->select_timeout = *to;
 401 }
 402
 403 /**
 404  * Force the next select() call to return before the given amount of milliseconds.
 405  *
 406  * \param ms The maximal allowed timeout in milliseconds.
 407  * \param s Pointer to the scheduler struct.
 408  *
 409  * Like sched_request_timeout() this imposes an upper bound on the timeout
 410  * value for the next select() call.
 411  */
 412 void sched_request_timeout_ms(long unsigned ms, struct sched *s)
 413 {
 414         struct timeval tv;
 415         ms2tv(ms, &tv);
 416         sched_request_timeout(&tv, s);
 417 }
 418
 419 /**
 420  * Force the next select() call to return before the given future time.
 421  *
 422  * \param barrier Absolute time before select() should return.
 423  * \param s Pointer to the scheduler struct.
 424  *
 425  * \return If \a barrier is in the past, this function does nothing and returns
 426  * zero. Otherwise it returns one.
 427  *
 428  * \sa sched_request_barrier_or_min_delay().
 429  */
 430 int sched_request_barrier(struct timeval *barrier, struct sched *s)
 431 {
 432         struct timeval diff;
 433
 434         if (tv_diff(now, barrier, &diff) > 0)
 435                 return 0;
 436         sched_request_timeout(&diff, s);
 437         return 1;
 438 }
 439
 440 /**
 441  * Force the next select() call to return before the given time.
 442  *
 443  * \param barrier Absolute time before select() should return.
 444  * \param s Pointer to the scheduler struct.
 445  *
 446  * \return If \a barrier is in the past, this function requests a minimal
 447  * timeout and returns zero. Otherwise it returns one.
 448  *
 449  * \sa sched_min_delay(), sched_request_barrier().
 450  */
 451 int sched_request_barrier_or_min_delay(struct timeval *barrier, struct sched *s)
 452 {
 453         struct timeval diff;
 454
 455         if (tv_diff(now, barrier, &diff) > 0) {
 456                 sched_min_delay(s);
 457                 return 0;
 458         }
 459         sched_request_timeout(&diff, s);
 460         return 1;
 461 }