scheduler.c 26 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766
  1. /* Copyright (c) 2013-2017, The Tor Project, Inc. */
  2. /* See LICENSE for licensing information */
  3. #include "or.h"
  4. #include "config.h"
  5. #include "compat_libevent.h"
  6. #define SCHEDULER_PRIVATE_
  7. #define SCHEDULER_KIST_PRIVATE
  8. #include "scheduler.h"
  9. #include "main.h"
  10. #include "buffers.h"
  11. #define TOR_CHANNEL_INTERNAL_
  12. #include "channeltls.h"
  13. /**
  14. * \file scheduler.c
  15. * \brief Channel scheduling system: decides which channels should send and
  16. * receive when.
  17. *
  18. * This module is the global/common parts of the scheduling system. This system
  19. * is what decides what channels get to send cells on their circuits and when.
  20. *
  21. * Terms:
  22. * - "Scheduling system": the collection of scheduler*.{h,c} files and their
  23. * aggregate behavior.
  24. * - "Scheduler implementation": a scheduler_t. The scheduling system has one
  25. * active scheduling implementation at a time.
  26. *
  27. * In this file you will find state that any scheduler implementation can have
  28. * access to as well as the functions the rest of Tor uses to interact with the
  29. * scheduling system.
  30. *
  31. * The earliest versions of Tor approximated a kind of round-robin system
  32. * among active connections, but only approximated it. It would only consider
  33. * one connection (roughly equal to a channel in today's terms) at a time, and
  34. * thus could only prioritize circuits against others on the same connection.
  35. *
  36. * Then in response to the KIST paper[0], Tor implemented a global
  37. * circuit scheduler. It was supposed to prioritize circuits across many
  38. * channels, but wasn't effective. It is preserved in scheduler_vanilla.c.
  39. *
  40. * [0]: http://www.robgjansen.com/publications/kist-sec2014.pdf
  41. *
  42. * Then we actually got around to implementing KIST for real. We decided to
  43. * modularize the scheduler so new ones can be implemented. You can find KIST
  44. * in scheduler_kist.c.
  45. *
  46. * Channels have one of four scheduling states based on whether or not they
  47. * have cells to send and whether or not they are able to send.
  48. *
  49. * <ol>
  50. * <li>
  51. * Not open for writes, no cells to send.
  52. * <ul><li> Not much to do here, and the channel will have scheduler_state
  53. * == SCHED_CHAN_IDLE
  54. * <li> Transitions from:
  55. * <ul>
  56. * <li>Open for writes/has cells by simultaneously draining all circuit
  57. * queues and filling the output buffer.
  58. * </ul>
  59. * <li> Transitions to:
  60. * <ul>
  61. * <li> Not open for writes/has cells by arrival of cells on an attached
  62. * circuit (this would be driven from append_cell_to_circuit_queue())
  63. * <li> Open for writes/no cells by a channel type specific path;
  64. * driven from connection_or_flushed_some() for channel_tls_t.
  65. * </ul>
  66. * </ul>
  67. *
  68. * <li> Open for writes, no cells to send
  69. * <ul>
  70. * <li>Not much here either; this will be the state an idle but open
  71. * channel can be expected to settle in. It will have scheduler_state
  72. * == SCHED_CHAN_WAITING_FOR_CELLS
  73. * <li> Transitions from:
  74. * <ul>
  75. * <li>Not open for writes/no cells by flushing some of the output
  76. * buffer.
  77. * <li>Open for writes/has cells by the scheduler moving cells from
  78. * circuit queues to channel output queue, but not having enough
  79. * to fill the output queue.
  80. * </ul>
  81. * <li> Transitions to:
  82. * <ul>
  83. * <li>Open for writes/has cells by arrival of new cells on an attached
  84. * circuit, in append_cell_to_circuit_queue()
  85. * </ul>
  86. * </ul>
  87. *
  88. * <li>Not open for writes, cells to send
  89. * <ul>
  90. * <li>This is the state of a busy circuit limited by output bandwidth;
  91. * cells have piled up in the circuit queues waiting to be relayed.
  92. * The channel will have scheduler_state == SCHED_CHAN_WAITING_TO_WRITE.
  93. * <li> Transitions from:
  94. * <ul>
  95. * <li>Not open for writes/no cells by arrival of cells on an attached
  96. * circuit
  97. * <li>Open for writes/has cells by filling an output buffer without
  98. * draining all cells from attached circuits
  99. * </ul>
  100. * <li> Transitions to:
  101. * <ul>
  102. * <li>Opens for writes/has cells by draining some of the output buffer
  103. * via the connection_or_flushed_some() path (for channel_tls_t).
  104. * </ul>
  105. * </ul>
  106. *
  107. * <li>Open for writes, cells to send
  108. * <ul>
  109. * <li>This connection is ready to relay some cells and waiting for
  110. * the scheduler to choose it. The channel will have scheduler_state ==
  111. * SCHED_CHAN_PENDING.
  112. * <li>Transitions from:
  113. * <ul>
  114. * <li>Not open for writes/has cells by the connection_or_flushed_some()
  115. * path
  116. * <li>Open for writes/no cells by the append_cell_to_circuit_queue()
  117. * path
  118. * </ul>
  119. * <li> Transitions to:
  120. * <ul>
  121. * <li>Not open for writes/no cells by draining all circuit queues and
  122. * simultaneously filling the output buffer.
  123. * <li>Not open for writes/has cells by writing enough cells to fill the
  124. * output buffer
  125. * <li>Open for writes/no cells by draining all attached circuit queues
  126. * without also filling the output buffer
  127. * </ul>
  128. * </ul>
  129. * </ol>
  130. *
  131. * Other event-driven parts of the code move channels between these scheduling
  132. * states by calling scheduler functions. The scheduling system builds up a
  133. * list of channels in the SCHED_CHAN_PENDING state that the scheduler
  134. * implementation should then use when it runs. Scheduling implementations need
  135. * to properly update channel states during their scheduler_t->run() function
  136. * as that is the only opportunity for channels to move from SCHED_CHAN_PENDING
  137. * to any other state.
  138. *
  139. * The remainder of this file is a small amount of state that any scheduler
  140. * implementation should have access to, and the functions the rest of Tor uses
  141. * to interact with the scheduling system.
  142. */
  143. /*****************************************************************************
  144. * Scheduling system state
  145. *
  146. * State that can be accessed from any scheduler implementation (but not
  147. * outside the scheduling system)
  148. *****************************************************************************/
  149. /** DOCDOC */
  150. STATIC const scheduler_t *the_scheduler;
  151. /**
  152. * We keep a list of channels that are pending - i.e, have cells to write
  153. * and can accept them to send. The enum scheduler_state in channel_t
  154. * is reserved for our use.
  155. *
  156. * Priority queue of channels that can write and have cells (pending work)
  157. */
  158. STATIC smartlist_t *channels_pending = NULL;
  159. /**
  160. * This event runs the scheduler from its callback, and is manually
  161. * activated whenever a channel enters open for writes/cells to send.
  162. */
  163. STATIC struct mainloop_event_t *run_sched_ev = NULL;
  164. static int have_logged_kist_suddenly_disabled = 0;
  165. /*****************************************************************************
  166. * Scheduling system static function definitions
  167. *
  168. * Functions that can only be accessed from this file.
  169. *****************************************************************************/
  170. /** Return a human readable string for the given scheduler type. */
  171. static const char *
  172. get_scheduler_type_string(scheduler_types_t type)
  173. {
  174. switch (type) {
  175. case SCHEDULER_VANILLA:
  176. return "Vanilla";
  177. case SCHEDULER_KIST:
  178. return "KIST";
  179. case SCHEDULER_KIST_LITE:
  180. return "KISTLite";
  181. case SCHEDULER_NONE:
  182. /* fallthrough */
  183. default:
  184. tor_assert_unreached();
  185. return "(N/A)";
  186. }
  187. }
  188. /**
  189. * Scheduler event callback; this should get triggered once per event loop
  190. * if any scheduling work was created during the event loop.
  191. */
  192. static void
  193. scheduler_evt_callback(mainloop_event_t *event, void *arg)
  194. {
  195. (void) event;
  196. (void) arg;
  197. log_debug(LD_SCHED, "Scheduler event callback called");
  198. /* Run the scheduler. This is a mandatory function. */
  199. /* We might as well assert on this. If this function doesn't exist, no cells
  200. * are getting scheduled. Things are very broken. scheduler_t says the run()
  201. * function is mandatory. */
  202. tor_assert(the_scheduler->run);
  203. the_scheduler->run();
  204. /* Schedule itself back in if it has more work. */
  205. /* Again, might as well assert on this mandatory scheduler_t function. If it
  206. * doesn't exist, there's no way to tell libevent to run the scheduler again
  207. * in the future. */
  208. tor_assert(the_scheduler->schedule);
  209. the_scheduler->schedule();
  210. }
  211. /** Using the global options, select the scheduler we should be using. */
  212. static void
  213. select_scheduler(void)
  214. {
  215. scheduler_t *new_scheduler = NULL;
  216. #ifdef TOR_UNIT_TESTS
  217. /* This is hella annoying to set in the options for every test that passes
  218. * through the scheduler and there are many so if we don't explicitly have
  219. * a list of types set, just put the vanilla one. */
  220. if (get_options()->SchedulerTypes_ == NULL) {
  221. the_scheduler = get_vanilla_scheduler();
  222. return;
  223. }
  224. #endif /* defined(TOR_UNIT_TESTS) */
  225. /* This list is ordered that is first entry has the first priority. Thus, as
  226. * soon as we find a scheduler type that we can use, we use it and stop. */
  227. SMARTLIST_FOREACH_BEGIN(get_options()->SchedulerTypes_, int *, type) {
  228. switch (*type) {
  229. case SCHEDULER_VANILLA:
  230. new_scheduler = get_vanilla_scheduler();
  231. goto end;
  232. case SCHEDULER_KIST:
  233. if (!scheduler_can_use_kist()) {
  234. #ifdef HAVE_KIST_SUPPORT
  235. if (!have_logged_kist_suddenly_disabled) {
  236. /* We should only log this once in most cases. If it was the kernel
  237. * losing support for kist that caused scheduler_can_use_kist() to
  238. * return false, then this flag makes sure we only log this message
  239. * once. If it was the consensus that switched from "yes use kist"
  240. * to "no don't use kist", then we still set the flag so we log
  241. * once, but we unset the flag elsewhere if we ever can_use_kist()
  242. * again.
  243. */
  244. have_logged_kist_suddenly_disabled = 1;
  245. log_notice(LD_SCHED, "Scheduler type KIST has been disabled by "
  246. "the consensus or no kernel support.");
  247. }
  248. #else /* !(defined(HAVE_KIST_SUPPORT)) */
  249. log_info(LD_SCHED, "Scheduler type KIST not built in");
  250. #endif /* defined(HAVE_KIST_SUPPORT) */
  251. continue;
  252. }
  253. /* This flag will only get set in one of two cases:
  254. * 1 - the kernel lost support for kist. In that case, we don't expect to
  255. * ever end up here
  256. * 2 - the consensus went from "yes use kist" to "no don't use kist".
  257. * We might end up here if the consensus changes back to "yes", in which
  258. * case we might want to warn the user again if it goes back to "no"
  259. * yet again. Thus we unset the flag */
  260. have_logged_kist_suddenly_disabled = 0;
  261. new_scheduler = get_kist_scheduler();
  262. scheduler_kist_set_full_mode();
  263. goto end;
  264. case SCHEDULER_KIST_LITE:
  265. new_scheduler = get_kist_scheduler();
  266. scheduler_kist_set_lite_mode();
  267. goto end;
  268. case SCHEDULER_NONE:
  269. /* fallthrough */
  270. default:
  271. /* Our option validation should have caught this. */
  272. tor_assert_unreached();
  273. }
  274. } SMARTLIST_FOREACH_END(type);
  275. end:
  276. if (new_scheduler == NULL) {
  277. log_err(LD_SCHED, "Tor was unable to select a scheduler type. Please "
  278. "make sure Schedulers is correctly configured with "
  279. "what Tor does support.");
  280. /* We weren't able to choose a scheduler which means that none of the ones
  281. * set in Schedulers are supported or usable. We will respect the user
  282. * wishes of using what it has been configured and don't do a sneaky
  283. * fallback. Because this can be changed at runtime, we have to stop tor
  284. * right now. */
  285. exit(1); // XXXX bad exit
  286. }
  287. /* Set the chosen scheduler. */
  288. the_scheduler = new_scheduler;
  289. }
  290. /**
  291. * Helper function called from a few different places. It changes the
  292. * scheduler implementation, if necessary. And if it did, it then tells the
  293. * old one to free its state and the new one to initialize.
  294. */
  295. static void
  296. set_scheduler(void)
  297. {
  298. const scheduler_t *old_scheduler = the_scheduler;
  299. scheduler_types_t old_scheduler_type = SCHEDULER_NONE;
  300. /* We keep track of the type in order to log only if the type switched. We
  301. * can't just use the scheduler pointers because KIST and KISTLite share the
  302. * same object. */
  303. if (the_scheduler) {
  304. old_scheduler_type = the_scheduler->type;
  305. }
  306. /* From the options, select the scheduler type to set. */
  307. select_scheduler();
  308. tor_assert(the_scheduler);
  309. /* We look at the pointer difference in case the old sched and new sched
  310. * share the same scheduler object, as is the case with KIST and KISTLite. */
  311. if (old_scheduler != the_scheduler) {
  312. /* Allow the old scheduler to clean up, if needed. */
  313. if (old_scheduler && old_scheduler->free_all) {
  314. old_scheduler->free_all();
  315. }
  316. /* Initialize the new scheduler. */
  317. if (the_scheduler->init) {
  318. the_scheduler->init();
  319. }
  320. }
  321. /* Finally we notice log if we switched schedulers. We use the type in case
  322. * two schedulers share a scheduler object. */
  323. if (old_scheduler_type != the_scheduler->type) {
  324. log_notice(LD_CONFIG, "Scheduler type %s has been enabled.",
  325. get_scheduler_type_string(the_scheduler->type));
  326. }
  327. }
  328. /*****************************************************************************
  329. * Scheduling system private function definitions
  330. *
  331. * Functions that can only be accessed from scheduler*.c
  332. *****************************************************************************/
  333. /** Returns human readable string for the given channel scheduler state. */
  334. const char *
  335. get_scheduler_state_string(int scheduler_state)
  336. {
  337. switch (scheduler_state) {
  338. case SCHED_CHAN_IDLE:
  339. return "IDLE";
  340. case SCHED_CHAN_WAITING_FOR_CELLS:
  341. return "WAITING_FOR_CELLS";
  342. case SCHED_CHAN_WAITING_TO_WRITE:
  343. return "WAITING_TO_WRITE";
  344. case SCHED_CHAN_PENDING:
  345. return "PENDING";
  346. default:
  347. return "(invalid)";
  348. }
  349. }
  350. /** Helper that logs channel scheduler_state changes. Use this instead of
  351. * setting scheduler_state directly. */
  352. void
  353. scheduler_set_channel_state(channel_t *chan, int new_state)
  354. {
  355. log_debug(LD_SCHED, "chan %" PRIu64 " changed from scheduler state %s to %s",
  356. chan->global_identifier,
  357. get_scheduler_state_string(chan->scheduler_state),
  358. get_scheduler_state_string(new_state));
  359. chan->scheduler_state = new_state;
  360. }
  361. /** Return the pending channel list. */
  362. smartlist_t *
  363. get_channels_pending(void)
  364. {
  365. return channels_pending;
  366. }
  367. /** Comparison function to use when sorting pending channels. */
  368. MOCK_IMPL(int,
  369. scheduler_compare_channels, (const void *c1_v, const void *c2_v))
  370. {
  371. const channel_t *c1 = NULL, *c2 = NULL;
  372. /* These are a workaround for -Wbad-function-cast throwing a fit */
  373. const circuitmux_policy_t *p1, *p2;
  374. uintptr_t p1_i, p2_i;
  375. tor_assert(c1_v);
  376. tor_assert(c2_v);
  377. c1 = (const channel_t *)(c1_v);
  378. c2 = (const channel_t *)(c2_v);
  379. if (c1 != c2) {
  380. if (circuitmux_get_policy(c1->cmux) ==
  381. circuitmux_get_policy(c2->cmux)) {
  382. /* Same cmux policy, so use the mux comparison */
  383. return circuitmux_compare_muxes(c1->cmux, c2->cmux);
  384. } else {
  385. /*
  386. * Different policies; not important to get this edge case perfect
  387. * because the current code never actually gives different channels
  388. * different cmux policies anyway. Just use this arbitrary but
  389. * definite choice.
  390. */
  391. p1 = circuitmux_get_policy(c1->cmux);
  392. p2 = circuitmux_get_policy(c2->cmux);
  393. p1_i = (uintptr_t)p1;
  394. p2_i = (uintptr_t)p2;
  395. return (p1_i < p2_i) ? -1 : 1;
  396. }
  397. } else {
  398. /* c1 == c2, so always equal */
  399. return 0;
  400. }
  401. }
  402. /*****************************************************************************
  403. * Scheduling system global functions
  404. *
  405. * Functions that can be accessed from anywhere in Tor.
  406. *****************************************************************************/
  407. /**
  408. * This is how the scheduling system is notified of Tor's configuration
  409. * changing. For example: a SIGHUP was issued.
  410. */
  411. void
  412. scheduler_conf_changed(void)
  413. {
  414. /* Let the scheduler decide what it should do. */
  415. set_scheduler();
  416. /* Then tell the (possibly new) scheduler that we have new options. */
  417. if (the_scheduler->on_new_options) {
  418. the_scheduler->on_new_options();
  419. }
  420. }
  421. /**
  422. * Whenever we get a new consensus, this function is called.
  423. */
  424. void
  425. scheduler_notify_networkstatus_changed(void)
  426. {
  427. /* Maybe the consensus param made us change the scheduler. */
  428. set_scheduler();
  429. /* Then tell the (possibly new) scheduler that we have a new consensus */
  430. if (the_scheduler->on_new_consensus) {
  431. the_scheduler->on_new_consensus();
  432. }
  433. }
  434. /**
  435. * Free everything scheduling-related from main.c. Note this is only called
  436. * when Tor is shutting down, while scheduler_t->free_all() is called both when
  437. * Tor is shutting down and when we are switching schedulers.
  438. */
  439. void
  440. scheduler_free_all(void)
  441. {
  442. log_debug(LD_SCHED, "Shutting down scheduler");
  443. if (run_sched_ev) {
  444. mainloop_event_free(run_sched_ev);
  445. run_sched_ev = NULL;
  446. }
  447. if (channels_pending) {
  448. /* We don't have ownership of the objects in this list. */
  449. smartlist_free(channels_pending);
  450. channels_pending = NULL;
  451. }
  452. if (the_scheduler && the_scheduler->free_all) {
  453. the_scheduler->free_all();
  454. }
  455. the_scheduler = NULL;
  456. }
  457. /** Mark a channel as no longer ready to accept writes. */
  458. MOCK_IMPL(void,
  459. scheduler_channel_doesnt_want_writes,(channel_t *chan))
  460. {
  461. IF_BUG_ONCE(!chan) {
  462. return;
  463. }
  464. IF_BUG_ONCE(!channels_pending) {
  465. return;
  466. }
  467. /* If it's already in pending, we can put it in waiting_to_write */
  468. if (chan->scheduler_state == SCHED_CHAN_PENDING) {
  469. /*
  470. * It's in channels_pending, so it shouldn't be in any of
  471. * the other lists. It can't write any more, so it goes to
  472. * channels_waiting_to_write.
  473. */
  474. smartlist_pqueue_remove(channels_pending,
  475. scheduler_compare_channels,
  476. offsetof(channel_t, sched_heap_idx),
  477. chan);
  478. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_TO_WRITE);
  479. } else {
  480. /*
  481. * It's not in pending, so it can't become waiting_to_write; it's
  482. * either not in any of the lists (nothing to do) or it's already in
  483. * waiting_for_cells (remove it, can't write any more).
  484. */
  485. if (chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS) {
  486. scheduler_set_channel_state(chan, SCHED_CHAN_IDLE);
  487. }
  488. }
  489. }
  490. /** Mark a channel as having waiting cells. */
  491. MOCK_IMPL(void,
  492. scheduler_channel_has_waiting_cells,(channel_t *chan))
  493. {
  494. IF_BUG_ONCE(!chan) {
  495. return;
  496. }
  497. IF_BUG_ONCE(!channels_pending) {
  498. return;
  499. }
  500. /* First, check if it's also writeable */
  501. if (chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS) {
  502. /*
  503. * It's in channels_waiting_for_cells, so it shouldn't be in any of
  504. * the other lists. It has waiting cells now, so it goes to
  505. * channels_pending.
  506. */
  507. scheduler_set_channel_state(chan, SCHED_CHAN_PENDING);
  508. if (!SCHED_BUG(chan->sched_heap_idx != -1, chan)) {
  509. smartlist_pqueue_add(channels_pending,
  510. scheduler_compare_channels,
  511. offsetof(channel_t, sched_heap_idx),
  512. chan);
  513. }
  514. /* If we made a channel pending, we potentially have scheduling work to
  515. * do. */
  516. the_scheduler->schedule();
  517. } else {
  518. /*
  519. * It's not in waiting_for_cells, so it can't become pending; it's
  520. * either not in any of the lists (we add it to waiting_to_write)
  521. * or it's already in waiting_to_write or pending (we do nothing)
  522. */
  523. if (!(chan->scheduler_state == SCHED_CHAN_WAITING_TO_WRITE ||
  524. chan->scheduler_state == SCHED_CHAN_PENDING)) {
  525. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_TO_WRITE);
  526. }
  527. }
  528. }
  529. /** Add the scheduler event to the set of pending events with next_run being
  530. * the longest time libevent should wait before triggering the event. */
  531. void
  532. scheduler_ev_add(const struct timeval *next_run)
  533. {
  534. tor_assert(run_sched_ev);
  535. tor_assert(next_run);
  536. if (BUG(mainloop_event_schedule(run_sched_ev, next_run) < 0)) {
  537. log_warn(LD_SCHED, "Adding to libevent failed. Next run time was set to: "
  538. "%ld.%06ld", next_run->tv_sec, (long)next_run->tv_usec);
  539. return;
  540. }
  541. }
  542. /** Make the scheduler event active with the given flags. */
  543. void
  544. scheduler_ev_active(void)
  545. {
  546. tor_assert(run_sched_ev);
  547. mainloop_event_activate(run_sched_ev);
  548. }
  549. /*
  550. * Initialize everything scheduling-related from config.c. Note this is only
  551. * called when Tor is starting up, while scheduler_t->init() is called both
  552. * when Tor is starting up and when we are switching schedulers.
  553. */
  554. void
  555. scheduler_init(void)
  556. {
  557. log_debug(LD_SCHED, "Initting scheduler");
  558. // Two '!' because we really do want to check if the pointer is non-NULL
  559. IF_BUG_ONCE(!!run_sched_ev) {
  560. log_warn(LD_SCHED, "We should not already have a libevent scheduler event."
  561. "I'll clean the old one up, but this is odd.");
  562. mainloop_event_free(run_sched_ev);
  563. run_sched_ev = NULL;
  564. }
  565. run_sched_ev = mainloop_event_new(scheduler_evt_callback, NULL);
  566. channels_pending = smartlist_new();
  567. set_scheduler();
  568. }
  569. /*
  570. * If a channel is going away, this is how the scheduling system is informed
  571. * so it can do any freeing necessary. This ultimately calls
  572. * scheduler_t->on_channel_free() so the current scheduler can release any
  573. * state specific to this channel.
  574. */
  575. MOCK_IMPL(void,
  576. scheduler_release_channel,(channel_t *chan))
  577. {
  578. IF_BUG_ONCE(!chan) {
  579. return;
  580. }
  581. IF_BUG_ONCE(!channels_pending) {
  582. return;
  583. }
  584. /* Try to remove the channel from the pending list regardless of its
  585. * scheduler state. We can release a channel in many places in the tor code
  586. * so we can't rely on the channel state (PENDING) to remove it from the
  587. * list.
  588. *
  589. * For instance, the channel can change state from OPEN to CLOSING while
  590. * being handled in the scheduler loop leading to the channel being in
  591. * PENDING state but not in the pending list. Furthermore, we release the
  592. * channel when it changes state to close and a second time when we free it.
  593. * Not ideal at all but for now that is the way it is. */
  594. if (chan->sched_heap_idx != -1) {
  595. smartlist_pqueue_remove(channels_pending,
  596. scheduler_compare_channels,
  597. offsetof(channel_t, sched_heap_idx),
  598. chan);
  599. }
  600. if (the_scheduler->on_channel_free) {
  601. the_scheduler->on_channel_free(chan);
  602. }
  603. scheduler_set_channel_state(chan, SCHED_CHAN_IDLE);
  604. }
  605. /** Mark a channel as ready to accept writes */
  606. void
  607. scheduler_channel_wants_writes(channel_t *chan)
  608. {
  609. IF_BUG_ONCE(!chan) {
  610. return;
  611. }
  612. IF_BUG_ONCE(!channels_pending) {
  613. return;
  614. }
  615. /* If it's already in waiting_to_write, we can put it in pending */
  616. if (chan->scheduler_state == SCHED_CHAN_WAITING_TO_WRITE) {
  617. /*
  618. * It can write now, so it goes to channels_pending.
  619. */
  620. scheduler_set_channel_state(chan, SCHED_CHAN_PENDING);
  621. if (!SCHED_BUG(chan->sched_heap_idx != -1, chan)) {
  622. smartlist_pqueue_add(channels_pending,
  623. scheduler_compare_channels,
  624. offsetof(channel_t, sched_heap_idx),
  625. chan);
  626. }
  627. /* We just made a channel pending, we have scheduling work to do. */
  628. the_scheduler->schedule();
  629. } else {
  630. /*
  631. * It's not in SCHED_CHAN_WAITING_TO_WRITE, so it can't become pending;
  632. * it's either idle and goes to WAITING_FOR_CELLS, or it's a no-op.
  633. */
  634. if (!(chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS ||
  635. chan->scheduler_state == SCHED_CHAN_PENDING)) {
  636. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_FOR_CELLS);
  637. }
  638. }
  639. }
  640. /* Log warn the given channel and extra scheduler context as well. This is
  641. * used by SCHED_BUG() in order to be able to extract as much information as
  642. * we can when we hit a bug. Channel chan can be NULL. */
  643. void
  644. scheduler_bug_occurred(const channel_t *chan)
  645. {
  646. char buf[128];
  647. if (chan != NULL) {
  648. const size_t outbuf_len =
  649. buf_datalen(TO_CONN(BASE_CHAN_TO_TLS((channel_t *) chan)->conn)->outbuf);
  650. tor_snprintf(buf, sizeof(buf),
  651. "Channel %" PRIu64 " in state %s and scheduler state %s."
  652. " Num cells on cmux: %d. Connection outbuf len: %lu.",
  653. chan->global_identifier,
  654. channel_state_to_string(chan->state),
  655. get_scheduler_state_string(chan->scheduler_state),
  656. circuitmux_num_cells(chan->cmux),
  657. (unsigned long)outbuf_len);
  658. }
  659. {
  660. char *msg;
  661. /* Rate limit every 60 seconds. If we start seeing this every 60 sec, we
  662. * know something is stuck/wrong. It *should* be loud but not too much. */
  663. static ratelim_t rlimit = RATELIM_INIT(60);
  664. if ((msg = rate_limit_log(&rlimit, approx_time()))) {
  665. log_warn(LD_BUG, "%s Num pending channels: %d. "
  666. "Channel in pending list: %s.%s",
  667. (chan != NULL) ? buf : "No channel in bug context.",
  668. smartlist_len(channels_pending),
  669. (smartlist_pos(channels_pending, chan) == -1) ? "no" : "yes",
  670. msg);
  671. tor_free(msg);
  672. }
  673. }
  674. }
  675. #ifdef TOR_UNIT_TESTS
  676. /*
  677. * Notify scheduler that a channel's queue position may have changed.
  678. */
  679. void
  680. scheduler_touch_channel(channel_t *chan)
  681. {
  682. IF_BUG_ONCE(!chan) {
  683. return;
  684. }
  685. if (chan->scheduler_state == SCHED_CHAN_PENDING) {
  686. /* Remove and re-add it */
  687. smartlist_pqueue_remove(channels_pending,
  688. scheduler_compare_channels,
  689. offsetof(channel_t, sched_heap_idx),
  690. chan);
  691. smartlist_pqueue_add(channels_pending,
  692. scheduler_compare_channels,
  693. offsetof(channel_t, sched_heap_idx),
  694. chan);
  695. }
  696. /* else no-op, since it isn't in the queue */
  697. }
  698. #endif /* defined(TOR_UNIT_TESTS) */