scheduler.c 26 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768
  1. /* Copyright (c) 2013-2018, The Tor Project, Inc. */
  2. /* See LICENSE for licensing information */
  3. #include "or/or.h"
  4. #include "or/config.h"
  5. #include "common/compat_libevent.h"
  6. #define SCHEDULER_PRIVATE_
  7. #define SCHEDULER_KIST_PRIVATE
  8. #include "or/scheduler.h"
  9. #include "or/main.h"
  10. #include "lib/container/buffers.h"
  11. #define TOR_CHANNEL_INTERNAL_
  12. #include "or/channeltls.h"
  13. #include "or/or_connection_st.h"
  14. /**
  15. * \file scheduler.c
  16. * \brief Channel scheduling system: decides which channels should send and
  17. * receive when.
  18. *
  19. * This module is the global/common parts of the scheduling system. This system
  20. * is what decides what channels get to send cells on their circuits and when.
  21. *
  22. * Terms:
  23. * - "Scheduling system": the collection of scheduler*.{h,c} files and their
  24. * aggregate behavior.
  25. * - "Scheduler implementation": a scheduler_t. The scheduling system has one
  26. * active scheduling implementation at a time.
  27. *
  28. * In this file you will find state that any scheduler implementation can have
  29. * access to as well as the functions the rest of Tor uses to interact with the
  30. * scheduling system.
  31. *
  32. * The earliest versions of Tor approximated a kind of round-robin system
  33. * among active connections, but only approximated it. It would only consider
  34. * one connection (roughly equal to a channel in today's terms) at a time, and
  35. * thus could only prioritize circuits against others on the same connection.
  36. *
  37. * Then in response to the KIST paper[0], Tor implemented a global
  38. * circuit scheduler. It was supposed to prioritize circuits across many
  39. * channels, but wasn't effective. It is preserved in scheduler_vanilla.c.
  40. *
  41. * [0]: http://www.robgjansen.com/publications/kist-sec2014.pdf
  42. *
  43. * Then we actually got around to implementing KIST for real. We decided to
  44. * modularize the scheduler so new ones can be implemented. You can find KIST
  45. * in scheduler_kist.c.
  46. *
  47. * Channels have one of four scheduling states based on whether or not they
  48. * have cells to send and whether or not they are able to send.
  49. *
  50. * <ol>
  51. * <li>
  52. * Not open for writes, no cells to send.
  53. * <ul><li> Not much to do here, and the channel will have scheduler_state
  54. * == SCHED_CHAN_IDLE
  55. * <li> Transitions from:
  56. * <ul>
  57. * <li>Open for writes/has cells by simultaneously draining all circuit
  58. * queues and filling the output buffer.
  59. * </ul>
  60. * <li> Transitions to:
  61. * <ul>
  62. * <li> Not open for writes/has cells by arrival of cells on an attached
  63. * circuit (this would be driven from append_cell_to_circuit_queue())
  64. * <li> Open for writes/no cells by a channel type specific path;
  65. * driven from connection_or_flushed_some() for channel_tls_t.
  66. * </ul>
  67. * </ul>
  68. *
  69. * <li> Open for writes, no cells to send
  70. * <ul>
  71. * <li>Not much here either; this will be the state an idle but open
  72. * channel can be expected to settle in. It will have scheduler_state
  73. * == SCHED_CHAN_WAITING_FOR_CELLS
  74. * <li> Transitions from:
  75. * <ul>
  76. * <li>Not open for writes/no cells by flushing some of the output
  77. * buffer.
  78. * <li>Open for writes/has cells by the scheduler moving cells from
  79. * circuit queues to channel output queue, but not having enough
  80. * to fill the output queue.
  81. * </ul>
  82. * <li> Transitions to:
  83. * <ul>
  84. * <li>Open for writes/has cells by arrival of new cells on an attached
  85. * circuit, in append_cell_to_circuit_queue()
  86. * </ul>
  87. * </ul>
  88. *
  89. * <li>Not open for writes, cells to send
  90. * <ul>
  91. * <li>This is the state of a busy circuit limited by output bandwidth;
  92. * cells have piled up in the circuit queues waiting to be relayed.
  93. * The channel will have scheduler_state == SCHED_CHAN_WAITING_TO_WRITE.
  94. * <li> Transitions from:
  95. * <ul>
  96. * <li>Not open for writes/no cells by arrival of cells on an attached
  97. * circuit
  98. * <li>Open for writes/has cells by filling an output buffer without
  99. * draining all cells from attached circuits
  100. * </ul>
  101. * <li> Transitions to:
  102. * <ul>
  103. * <li>Opens for writes/has cells by draining some of the output buffer
  104. * via the connection_or_flushed_some() path (for channel_tls_t).
  105. * </ul>
  106. * </ul>
  107. *
  108. * <li>Open for writes, cells to send
  109. * <ul>
  110. * <li>This connection is ready to relay some cells and waiting for
  111. * the scheduler to choose it. The channel will have scheduler_state ==
  112. * SCHED_CHAN_PENDING.
  113. * <li>Transitions from:
  114. * <ul>
  115. * <li>Not open for writes/has cells by the connection_or_flushed_some()
  116. * path
  117. * <li>Open for writes/no cells by the append_cell_to_circuit_queue()
  118. * path
  119. * </ul>
  120. * <li> Transitions to:
  121. * <ul>
  122. * <li>Not open for writes/no cells by draining all circuit queues and
  123. * simultaneously filling the output buffer.
  124. * <li>Not open for writes/has cells by writing enough cells to fill the
  125. * output buffer
  126. * <li>Open for writes/no cells by draining all attached circuit queues
  127. * without also filling the output buffer
  128. * </ul>
  129. * </ul>
  130. * </ol>
  131. *
  132. * Other event-driven parts of the code move channels between these scheduling
  133. * states by calling scheduler functions. The scheduling system builds up a
  134. * list of channels in the SCHED_CHAN_PENDING state that the scheduler
  135. * implementation should then use when it runs. Scheduling implementations need
  136. * to properly update channel states during their scheduler_t->run() function
  137. * as that is the only opportunity for channels to move from SCHED_CHAN_PENDING
  138. * to any other state.
  139. *
  140. * The remainder of this file is a small amount of state that any scheduler
  141. * implementation should have access to, and the functions the rest of Tor uses
  142. * to interact with the scheduling system.
  143. */
  144. /*****************************************************************************
  145. * Scheduling system state
  146. *
  147. * State that can be accessed from any scheduler implementation (but not
  148. * outside the scheduling system)
  149. *****************************************************************************/
  150. /** DOCDOC */
  151. STATIC const scheduler_t *the_scheduler;
  152. /**
  153. * We keep a list of channels that are pending - i.e, have cells to write
  154. * and can accept them to send. The enum scheduler_state in channel_t
  155. * is reserved for our use.
  156. *
  157. * Priority queue of channels that can write and have cells (pending work)
  158. */
  159. STATIC smartlist_t *channels_pending = NULL;
  160. /**
  161. * This event runs the scheduler from its callback, and is manually
  162. * activated whenever a channel enters open for writes/cells to send.
  163. */
  164. STATIC struct mainloop_event_t *run_sched_ev = NULL;
  165. static int have_logged_kist_suddenly_disabled = 0;
  166. /*****************************************************************************
  167. * Scheduling system static function definitions
  168. *
  169. * Functions that can only be accessed from this file.
  170. *****************************************************************************/
  171. /** Return a human readable string for the given scheduler type. */
  172. static const char *
  173. get_scheduler_type_string(scheduler_types_t type)
  174. {
  175. switch (type) {
  176. case SCHEDULER_VANILLA:
  177. return "Vanilla";
  178. case SCHEDULER_KIST:
  179. return "KIST";
  180. case SCHEDULER_KIST_LITE:
  181. return "KISTLite";
  182. case SCHEDULER_NONE:
  183. /* fallthrough */
  184. default:
  185. tor_assert_unreached();
  186. return "(N/A)";
  187. }
  188. }
  189. /**
  190. * Scheduler event callback; this should get triggered once per event loop
  191. * if any scheduling work was created during the event loop.
  192. */
  193. static void
  194. scheduler_evt_callback(mainloop_event_t *event, void *arg)
  195. {
  196. (void) event;
  197. (void) arg;
  198. log_debug(LD_SCHED, "Scheduler event callback called");
  199. /* Run the scheduler. This is a mandatory function. */
  200. /* We might as well assert on this. If this function doesn't exist, no cells
  201. * are getting scheduled. Things are very broken. scheduler_t says the run()
  202. * function is mandatory. */
  203. tor_assert(the_scheduler->run);
  204. the_scheduler->run();
  205. /* Schedule itself back in if it has more work. */
  206. /* Again, might as well assert on this mandatory scheduler_t function. If it
  207. * doesn't exist, there's no way to tell libevent to run the scheduler again
  208. * in the future. */
  209. tor_assert(the_scheduler->schedule);
  210. the_scheduler->schedule();
  211. }
  212. /** Using the global options, select the scheduler we should be using. */
  213. static void
  214. select_scheduler(void)
  215. {
  216. scheduler_t *new_scheduler = NULL;
  217. #ifdef TOR_UNIT_TESTS
  218. /* This is hella annoying to set in the options for every test that passes
  219. * through the scheduler and there are many so if we don't explicitly have
  220. * a list of types set, just put the vanilla one. */
  221. if (get_options()->SchedulerTypes_ == NULL) {
  222. the_scheduler = get_vanilla_scheduler();
  223. return;
  224. }
  225. #endif /* defined(TOR_UNIT_TESTS) */
  226. /* This list is ordered that is first entry has the first priority. Thus, as
  227. * soon as we find a scheduler type that we can use, we use it and stop. */
  228. SMARTLIST_FOREACH_BEGIN(get_options()->SchedulerTypes_, int *, type) {
  229. switch (*type) {
  230. case SCHEDULER_VANILLA:
  231. new_scheduler = get_vanilla_scheduler();
  232. goto end;
  233. case SCHEDULER_KIST:
  234. if (!scheduler_can_use_kist()) {
  235. #ifdef HAVE_KIST_SUPPORT
  236. if (!have_logged_kist_suddenly_disabled) {
  237. /* We should only log this once in most cases. If it was the kernel
  238. * losing support for kist that caused scheduler_can_use_kist() to
  239. * return false, then this flag makes sure we only log this message
  240. * once. If it was the consensus that switched from "yes use kist"
  241. * to "no don't use kist", then we still set the flag so we log
  242. * once, but we unset the flag elsewhere if we ever can_use_kist()
  243. * again.
  244. */
  245. have_logged_kist_suddenly_disabled = 1;
  246. log_notice(LD_SCHED, "Scheduler type KIST has been disabled by "
  247. "the consensus or no kernel support.");
  248. }
  249. #else /* !(defined(HAVE_KIST_SUPPORT)) */
  250. log_info(LD_SCHED, "Scheduler type KIST not built in");
  251. #endif /* defined(HAVE_KIST_SUPPORT) */
  252. continue;
  253. }
  254. /* This flag will only get set in one of two cases:
  255. * 1 - the kernel lost support for kist. In that case, we don't expect to
  256. * ever end up here
  257. * 2 - the consensus went from "yes use kist" to "no don't use kist".
  258. * We might end up here if the consensus changes back to "yes", in which
  259. * case we might want to warn the user again if it goes back to "no"
  260. * yet again. Thus we unset the flag */
  261. have_logged_kist_suddenly_disabled = 0;
  262. new_scheduler = get_kist_scheduler();
  263. scheduler_kist_set_full_mode();
  264. goto end;
  265. case SCHEDULER_KIST_LITE:
  266. new_scheduler = get_kist_scheduler();
  267. scheduler_kist_set_lite_mode();
  268. goto end;
  269. case SCHEDULER_NONE:
  270. /* fallthrough */
  271. default:
  272. /* Our option validation should have caught this. */
  273. tor_assert_unreached();
  274. }
  275. } SMARTLIST_FOREACH_END(type);
  276. end:
  277. if (new_scheduler == NULL) {
  278. log_err(LD_SCHED, "Tor was unable to select a scheduler type. Please "
  279. "make sure Schedulers is correctly configured with "
  280. "what Tor does support.");
  281. /* We weren't able to choose a scheduler which means that none of the ones
  282. * set in Schedulers are supported or usable. We will respect the user
  283. * wishes of using what it has been configured and don't do a sneaky
  284. * fallback. Because this can be changed at runtime, we have to stop tor
  285. * right now. */
  286. exit(1); // XXXX bad exit
  287. }
  288. /* Set the chosen scheduler. */
  289. the_scheduler = new_scheduler;
  290. }
  291. /**
  292. * Helper function called from a few different places. It changes the
  293. * scheduler implementation, if necessary. And if it did, it then tells the
  294. * old one to free its state and the new one to initialize.
  295. */
  296. static void
  297. set_scheduler(void)
  298. {
  299. const scheduler_t *old_scheduler = the_scheduler;
  300. scheduler_types_t old_scheduler_type = SCHEDULER_NONE;
  301. /* We keep track of the type in order to log only if the type switched. We
  302. * can't just use the scheduler pointers because KIST and KISTLite share the
  303. * same object. */
  304. if (the_scheduler) {
  305. old_scheduler_type = the_scheduler->type;
  306. }
  307. /* From the options, select the scheduler type to set. */
  308. select_scheduler();
  309. tor_assert(the_scheduler);
  310. /* We look at the pointer difference in case the old sched and new sched
  311. * share the same scheduler object, as is the case with KIST and KISTLite. */
  312. if (old_scheduler != the_scheduler) {
  313. /* Allow the old scheduler to clean up, if needed. */
  314. if (old_scheduler && old_scheduler->free_all) {
  315. old_scheduler->free_all();
  316. }
  317. /* Initialize the new scheduler. */
  318. if (the_scheduler->init) {
  319. the_scheduler->init();
  320. }
  321. }
  322. /* Finally we notice log if we switched schedulers. We use the type in case
  323. * two schedulers share a scheduler object. */
  324. if (old_scheduler_type != the_scheduler->type) {
  325. log_notice(LD_CONFIG, "Scheduler type %s has been enabled.",
  326. get_scheduler_type_string(the_scheduler->type));
  327. }
  328. }
  329. /*****************************************************************************
  330. * Scheduling system private function definitions
  331. *
  332. * Functions that can only be accessed from scheduler*.c
  333. *****************************************************************************/
  334. /** Returns human readable string for the given channel scheduler state. */
  335. const char *
  336. get_scheduler_state_string(int scheduler_state)
  337. {
  338. switch (scheduler_state) {
  339. case SCHED_CHAN_IDLE:
  340. return "IDLE";
  341. case SCHED_CHAN_WAITING_FOR_CELLS:
  342. return "WAITING_FOR_CELLS";
  343. case SCHED_CHAN_WAITING_TO_WRITE:
  344. return "WAITING_TO_WRITE";
  345. case SCHED_CHAN_PENDING:
  346. return "PENDING";
  347. default:
  348. return "(invalid)";
  349. }
  350. }
  351. /** Helper that logs channel scheduler_state changes. Use this instead of
  352. * setting scheduler_state directly. */
  353. void
  354. scheduler_set_channel_state(channel_t *chan, int new_state)
  355. {
  356. log_debug(LD_SCHED, "chan %" PRIu64 " changed from scheduler state %s to %s",
  357. chan->global_identifier,
  358. get_scheduler_state_string(chan->scheduler_state),
  359. get_scheduler_state_string(new_state));
  360. chan->scheduler_state = new_state;
  361. }
  362. /** Return the pending channel list. */
  363. smartlist_t *
  364. get_channels_pending(void)
  365. {
  366. return channels_pending;
  367. }
  368. /** Comparison function to use when sorting pending channels. */
  369. MOCK_IMPL(int,
  370. scheduler_compare_channels, (const void *c1_v, const void *c2_v))
  371. {
  372. const channel_t *c1 = NULL, *c2 = NULL;
  373. /* These are a workaround for -Wbad-function-cast throwing a fit */
  374. const circuitmux_policy_t *p1, *p2;
  375. uintptr_t p1_i, p2_i;
  376. tor_assert(c1_v);
  377. tor_assert(c2_v);
  378. c1 = (const channel_t *)(c1_v);
  379. c2 = (const channel_t *)(c2_v);
  380. if (c1 != c2) {
  381. if (circuitmux_get_policy(c1->cmux) ==
  382. circuitmux_get_policy(c2->cmux)) {
  383. /* Same cmux policy, so use the mux comparison */
  384. return circuitmux_compare_muxes(c1->cmux, c2->cmux);
  385. } else {
  386. /*
  387. * Different policies; not important to get this edge case perfect
  388. * because the current code never actually gives different channels
  389. * different cmux policies anyway. Just use this arbitrary but
  390. * definite choice.
  391. */
  392. p1 = circuitmux_get_policy(c1->cmux);
  393. p2 = circuitmux_get_policy(c2->cmux);
  394. p1_i = (uintptr_t)p1;
  395. p2_i = (uintptr_t)p2;
  396. return (p1_i < p2_i) ? -1 : 1;
  397. }
  398. } else {
  399. /* c1 == c2, so always equal */
  400. return 0;
  401. }
  402. }
  403. /*****************************************************************************
  404. * Scheduling system global functions
  405. *
  406. * Functions that can be accessed from anywhere in Tor.
  407. *****************************************************************************/
  408. /**
  409. * This is how the scheduling system is notified of Tor's configuration
  410. * changing. For example: a SIGHUP was issued.
  411. */
  412. void
  413. scheduler_conf_changed(void)
  414. {
  415. /* Let the scheduler decide what it should do. */
  416. set_scheduler();
  417. /* Then tell the (possibly new) scheduler that we have new options. */
  418. if (the_scheduler->on_new_options) {
  419. the_scheduler->on_new_options();
  420. }
  421. }
  422. /**
  423. * Whenever we get a new consensus, this function is called.
  424. */
  425. void
  426. scheduler_notify_networkstatus_changed(void)
  427. {
  428. /* Maybe the consensus param made us change the scheduler. */
  429. set_scheduler();
  430. /* Then tell the (possibly new) scheduler that we have a new consensus */
  431. if (the_scheduler->on_new_consensus) {
  432. the_scheduler->on_new_consensus();
  433. }
  434. }
  435. /**
  436. * Free everything scheduling-related from main.c. Note this is only called
  437. * when Tor is shutting down, while scheduler_t->free_all() is called both when
  438. * Tor is shutting down and when we are switching schedulers.
  439. */
  440. void
  441. scheduler_free_all(void)
  442. {
  443. log_debug(LD_SCHED, "Shutting down scheduler");
  444. if (run_sched_ev) {
  445. mainloop_event_free(run_sched_ev);
  446. run_sched_ev = NULL;
  447. }
  448. if (channels_pending) {
  449. /* We don't have ownership of the objects in this list. */
  450. smartlist_free(channels_pending);
  451. channels_pending = NULL;
  452. }
  453. if (the_scheduler && the_scheduler->free_all) {
  454. the_scheduler->free_all();
  455. }
  456. the_scheduler = NULL;
  457. }
  458. /** Mark a channel as no longer ready to accept writes. */
  459. MOCK_IMPL(void,
  460. scheduler_channel_doesnt_want_writes,(channel_t *chan))
  461. {
  462. IF_BUG_ONCE(!chan) {
  463. return;
  464. }
  465. IF_BUG_ONCE(!channels_pending) {
  466. return;
  467. }
  468. /* If it's already in pending, we can put it in waiting_to_write */
  469. if (chan->scheduler_state == SCHED_CHAN_PENDING) {
  470. /*
  471. * It's in channels_pending, so it shouldn't be in any of
  472. * the other lists. It can't write any more, so it goes to
  473. * channels_waiting_to_write.
  474. */
  475. smartlist_pqueue_remove(channels_pending,
  476. scheduler_compare_channels,
  477. offsetof(channel_t, sched_heap_idx),
  478. chan);
  479. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_TO_WRITE);
  480. } else {
  481. /*
  482. * It's not in pending, so it can't become waiting_to_write; it's
  483. * either not in any of the lists (nothing to do) or it's already in
  484. * waiting_for_cells (remove it, can't write any more).
  485. */
  486. if (chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS) {
  487. scheduler_set_channel_state(chan, SCHED_CHAN_IDLE);
  488. }
  489. }
  490. }
  491. /** Mark a channel as having waiting cells. */
  492. MOCK_IMPL(void,
  493. scheduler_channel_has_waiting_cells,(channel_t *chan))
  494. {
  495. IF_BUG_ONCE(!chan) {
  496. return;
  497. }
  498. IF_BUG_ONCE(!channels_pending) {
  499. return;
  500. }
  501. /* First, check if it's also writeable */
  502. if (chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS) {
  503. /*
  504. * It's in channels_waiting_for_cells, so it shouldn't be in any of
  505. * the other lists. It has waiting cells now, so it goes to
  506. * channels_pending.
  507. */
  508. scheduler_set_channel_state(chan, SCHED_CHAN_PENDING);
  509. if (!SCHED_BUG(chan->sched_heap_idx != -1, chan)) {
  510. smartlist_pqueue_add(channels_pending,
  511. scheduler_compare_channels,
  512. offsetof(channel_t, sched_heap_idx),
  513. chan);
  514. }
  515. /* If we made a channel pending, we potentially have scheduling work to
  516. * do. */
  517. the_scheduler->schedule();
  518. } else {
  519. /*
  520. * It's not in waiting_for_cells, so it can't become pending; it's
  521. * either not in any of the lists (we add it to waiting_to_write)
  522. * or it's already in waiting_to_write or pending (we do nothing)
  523. */
  524. if (!(chan->scheduler_state == SCHED_CHAN_WAITING_TO_WRITE ||
  525. chan->scheduler_state == SCHED_CHAN_PENDING)) {
  526. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_TO_WRITE);
  527. }
  528. }
  529. }
  530. /** Add the scheduler event to the set of pending events with next_run being
  531. * the longest time libevent should wait before triggering the event. */
  532. void
  533. scheduler_ev_add(const struct timeval *next_run)
  534. {
  535. tor_assert(run_sched_ev);
  536. tor_assert(next_run);
  537. if (BUG(mainloop_event_schedule(run_sched_ev, next_run) < 0)) {
  538. log_warn(LD_SCHED, "Adding to libevent failed. Next run time was set to: "
  539. "%ld.%06ld", next_run->tv_sec, (long)next_run->tv_usec);
  540. return;
  541. }
  542. }
  543. /** Make the scheduler event active with the given flags. */
  544. void
  545. scheduler_ev_active(void)
  546. {
  547. tor_assert(run_sched_ev);
  548. mainloop_event_activate(run_sched_ev);
  549. }
  550. /*
  551. * Initialize everything scheduling-related from config.c. Note this is only
  552. * called when Tor is starting up, while scheduler_t->init() is called both
  553. * when Tor is starting up and when we are switching schedulers.
  554. */
  555. void
  556. scheduler_init(void)
  557. {
  558. log_debug(LD_SCHED, "Initting scheduler");
  559. // Two '!' because we really do want to check if the pointer is non-NULL
  560. IF_BUG_ONCE(!!run_sched_ev) {
  561. log_warn(LD_SCHED, "We should not already have a libevent scheduler event."
  562. "I'll clean the old one up, but this is odd.");
  563. mainloop_event_free(run_sched_ev);
  564. run_sched_ev = NULL;
  565. }
  566. run_sched_ev = mainloop_event_new(scheduler_evt_callback, NULL);
  567. channels_pending = smartlist_new();
  568. set_scheduler();
  569. }
  570. /*
  571. * If a channel is going away, this is how the scheduling system is informed
  572. * so it can do any freeing necessary. This ultimately calls
  573. * scheduler_t->on_channel_free() so the current scheduler can release any
  574. * state specific to this channel.
  575. */
  576. MOCK_IMPL(void,
  577. scheduler_release_channel,(channel_t *chan))
  578. {
  579. IF_BUG_ONCE(!chan) {
  580. return;
  581. }
  582. IF_BUG_ONCE(!channels_pending) {
  583. return;
  584. }
  585. /* Try to remove the channel from the pending list regardless of its
  586. * scheduler state. We can release a channel in many places in the tor code
  587. * so we can't rely on the channel state (PENDING) to remove it from the
  588. * list.
  589. *
  590. * For instance, the channel can change state from OPEN to CLOSING while
  591. * being handled in the scheduler loop leading to the channel being in
  592. * PENDING state but not in the pending list. Furthermore, we release the
  593. * channel when it changes state to close and a second time when we free it.
  594. * Not ideal at all but for now that is the way it is. */
  595. if (chan->sched_heap_idx != -1) {
  596. smartlist_pqueue_remove(channels_pending,
  597. scheduler_compare_channels,
  598. offsetof(channel_t, sched_heap_idx),
  599. chan);
  600. }
  601. if (the_scheduler->on_channel_free) {
  602. the_scheduler->on_channel_free(chan);
  603. }
  604. scheduler_set_channel_state(chan, SCHED_CHAN_IDLE);
  605. }
  606. /** Mark a channel as ready to accept writes */
  607. void
  608. scheduler_channel_wants_writes(channel_t *chan)
  609. {
  610. IF_BUG_ONCE(!chan) {
  611. return;
  612. }
  613. IF_BUG_ONCE(!channels_pending) {
  614. return;
  615. }
  616. /* If it's already in waiting_to_write, we can put it in pending */
  617. if (chan->scheduler_state == SCHED_CHAN_WAITING_TO_WRITE) {
  618. /*
  619. * It can write now, so it goes to channels_pending.
  620. */
  621. scheduler_set_channel_state(chan, SCHED_CHAN_PENDING);
  622. if (!SCHED_BUG(chan->sched_heap_idx != -1, chan)) {
  623. smartlist_pqueue_add(channels_pending,
  624. scheduler_compare_channels,
  625. offsetof(channel_t, sched_heap_idx),
  626. chan);
  627. }
  628. /* We just made a channel pending, we have scheduling work to do. */
  629. the_scheduler->schedule();
  630. } else {
  631. /*
  632. * It's not in SCHED_CHAN_WAITING_TO_WRITE, so it can't become pending;
  633. * it's either idle and goes to WAITING_FOR_CELLS, or it's a no-op.
  634. */
  635. if (!(chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS ||
  636. chan->scheduler_state == SCHED_CHAN_PENDING)) {
  637. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_FOR_CELLS);
  638. }
  639. }
  640. }
  641. /* Log warn the given channel and extra scheduler context as well. This is
  642. * used by SCHED_BUG() in order to be able to extract as much information as
  643. * we can when we hit a bug. Channel chan can be NULL. */
  644. void
  645. scheduler_bug_occurred(const channel_t *chan)
  646. {
  647. char buf[128];
  648. if (chan != NULL) {
  649. const size_t outbuf_len =
  650. buf_datalen(TO_CONN(BASE_CHAN_TO_TLS((channel_t *) chan)->conn)->outbuf);
  651. tor_snprintf(buf, sizeof(buf),
  652. "Channel %" PRIu64 " in state %s and scheduler state %s."
  653. " Num cells on cmux: %d. Connection outbuf len: %lu.",
  654. chan->global_identifier,
  655. channel_state_to_string(chan->state),
  656. get_scheduler_state_string(chan->scheduler_state),
  657. circuitmux_num_cells(chan->cmux),
  658. (unsigned long)outbuf_len);
  659. }
  660. {
  661. char *msg;
  662. /* Rate limit every 60 seconds. If we start seeing this every 60 sec, we
  663. * know something is stuck/wrong. It *should* be loud but not too much. */
  664. static ratelim_t rlimit = RATELIM_INIT(60);
  665. if ((msg = rate_limit_log(&rlimit, approx_time()))) {
  666. log_warn(LD_BUG, "%s Num pending channels: %d. "
  667. "Channel in pending list: %s.%s",
  668. (chan != NULL) ? buf : "No channel in bug context.",
  669. smartlist_len(channels_pending),
  670. (smartlist_pos(channels_pending, chan) == -1) ? "no" : "yes",
  671. msg);
  672. tor_free(msg);
  673. }
  674. }
  675. }
  676. #ifdef TOR_UNIT_TESTS
  677. /*
  678. * Notify scheduler that a channel's queue position may have changed.
  679. */
  680. void
  681. scheduler_touch_channel(channel_t *chan)
  682. {
  683. IF_BUG_ONCE(!chan) {
  684. return;
  685. }
  686. if (chan->scheduler_state == SCHED_CHAN_PENDING) {
  687. /* Remove and re-add it */
  688. smartlist_pqueue_remove(channels_pending,
  689. scheduler_compare_channels,
  690. offsetof(channel_t, sched_heap_idx),
  691. chan);
  692. smartlist_pqueue_add(channels_pending,
  693. scheduler_compare_channels,
  694. offsetof(channel_t, sched_heap_idx),
  695. chan);
  696. }
  697. /* else no-op, since it isn't in the queue */
  698. }
  699. #endif /* defined(TOR_UNIT_TESTS) */