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. #include <event2/event.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 event *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(evutil_socket_t fd, short events, void *arg)
  195. {
  196. (void) fd;
  197. (void) events;
  198. (void) arg;
  199. log_debug(LD_SCHED, "Scheduler event callback called");
  200. /* Run the scheduler. This is a mandatory function. */
  201. /* We might as well assert on this. If this function doesn't exist, no cells
  202. * are getting scheduled. Things are very broken. scheduler_t says the run()
  203. * function is mandatory. */
  204. tor_assert(the_scheduler->run);
  205. the_scheduler->run();
  206. /* Schedule itself back in if it has more work. */
  207. /* Again, might as well assert on this mandatory scheduler_t function. If it
  208. * doesn't exist, there's no way to tell libevent to run the scheduler again
  209. * in the future. */
  210. tor_assert(the_scheduler->schedule);
  211. the_scheduler->schedule();
  212. }
  213. /** Using the global options, select the scheduler we should be using. */
  214. static void
  215. select_scheduler(void)
  216. {
  217. scheduler_t *new_scheduler = NULL;
  218. #ifdef TOR_UNIT_TESTS
  219. /* This is hella annoying to set in the options for every test that passes
  220. * through the scheduler and there are many so if we don't explicitly have
  221. * a list of types set, just put the vanilla one. */
  222. if (get_options()->SchedulerTypes_ == NULL) {
  223. the_scheduler = get_vanilla_scheduler();
  224. return;
  225. }
  226. #endif /* defined(TOR_UNIT_TESTS) */
  227. /* This list is ordered that is first entry has the first priority. Thus, as
  228. * soon as we find a scheduler type that we can use, we use it and stop. */
  229. SMARTLIST_FOREACH_BEGIN(get_options()->SchedulerTypes_, int *, type) {
  230. switch (*type) {
  231. case SCHEDULER_VANILLA:
  232. new_scheduler = get_vanilla_scheduler();
  233. goto end;
  234. case SCHEDULER_KIST:
  235. if (!scheduler_can_use_kist()) {
  236. #ifdef HAVE_KIST_SUPPORT
  237. if (!have_logged_kist_suddenly_disabled) {
  238. /* We should only log this once in most cases. If it was the kernel
  239. * losing support for kist that caused scheduler_can_use_kist() to
  240. * return false, then this flag makes sure we only log this message
  241. * once. If it was the consensus that switched from "yes use kist"
  242. * to "no don't use kist", then we still set the flag so we log
  243. * once, but we unset the flag elsewhere if we ever can_use_kist()
  244. * again.
  245. */
  246. have_logged_kist_suddenly_disabled = 1;
  247. log_notice(LD_SCHED, "Scheduler type KIST has been disabled by "
  248. "the consensus or no kernel support.");
  249. }
  250. #else /* !(defined(HAVE_KIST_SUPPORT)) */
  251. log_info(LD_SCHED, "Scheduler type KIST not built in");
  252. #endif /* defined(HAVE_KIST_SUPPORT) */
  253. continue;
  254. }
  255. /* This flag will only get set in one of two cases:
  256. * 1 - the kernel lost support for kist. In that case, we don't expect to
  257. * ever end up here
  258. * 2 - the consensus went from "yes use kist" to "no don't use kist".
  259. * We might end up here if the consensus changes back to "yes", in which
  260. * case we might want to warn the user again if it goes back to "no"
  261. * yet again. Thus we unset the flag */
  262. have_logged_kist_suddenly_disabled = 0;
  263. new_scheduler = get_kist_scheduler();
  264. scheduler_kist_set_full_mode();
  265. goto end;
  266. case SCHEDULER_KIST_LITE:
  267. new_scheduler = get_kist_scheduler();
  268. scheduler_kist_set_lite_mode();
  269. goto end;
  270. case SCHEDULER_NONE:
  271. /* fallthrough */
  272. default:
  273. /* Our option validation should have caught this. */
  274. tor_assert_unreached();
  275. }
  276. } SMARTLIST_FOREACH_END(type);
  277. end:
  278. if (new_scheduler == NULL) {
  279. log_err(LD_SCHED, "Tor was unable to select a scheduler type. Please "
  280. "make sure Schedulers is correctly configured with "
  281. "what Tor does support.");
  282. /* We weren't able to choose a scheduler which means that none of the ones
  283. * set in Schedulers are supported or usable. We will respect the user
  284. * wishes of using what it has been configured and don't do a sneaky
  285. * fallback. Because this can be changed at runtime, we have to stop tor
  286. * right now. */
  287. exit(1); // XXXX bad exit
  288. }
  289. /* Set the chosen scheduler. */
  290. the_scheduler = new_scheduler;
  291. }
  292. /**
  293. * Helper function called from a few different places. It changes the
  294. * scheduler implementation, if necessary. And if it did, it then tells the
  295. * old one to free its state and the new one to initialize.
  296. */
  297. static void
  298. set_scheduler(void)
  299. {
  300. const scheduler_t *old_scheduler = the_scheduler;
  301. scheduler_types_t old_scheduler_type = SCHEDULER_NONE;
  302. /* We keep track of the type in order to log only if the type switched. We
  303. * can't just use the scheduler pointers because KIST and KISTLite share the
  304. * same object. */
  305. if (the_scheduler) {
  306. old_scheduler_type = the_scheduler->type;
  307. }
  308. /* From the options, select the scheduler type to set. */
  309. select_scheduler();
  310. tor_assert(the_scheduler);
  311. /* We look at the pointer difference in case the old sched and new sched
  312. * share the same scheduler object, as is the case with KIST and KISTLite. */
  313. if (old_scheduler != the_scheduler) {
  314. /* Allow the old scheduler to clean up, if needed. */
  315. if (old_scheduler && old_scheduler->free_all) {
  316. old_scheduler->free_all();
  317. }
  318. /* Initialize the new scheduler. */
  319. if (the_scheduler->init) {
  320. the_scheduler->init();
  321. }
  322. }
  323. /* Finally we notice log if we switched schedulers. We use the type in case
  324. * two schedulers share a scheduler object. */
  325. if (old_scheduler_type != the_scheduler->type) {
  326. log_notice(LD_CONFIG, "Scheduler type %s has been enabled.",
  327. get_scheduler_type_string(the_scheduler->type));
  328. }
  329. }
  330. /*****************************************************************************
  331. * Scheduling system private function definitions
  332. *
  333. * Functions that can only be accessed from scheduler*.c
  334. *****************************************************************************/
  335. /** Helper that logs channel scheduler_state changes. Use this instead of
  336. * setting scheduler_state directly. */
  337. void scheduler_set_channel_state(channel_t *chan, int new_state){
  338. log_debug(LD_SCHED, "chan %s changed from scheduler state %d to %d",
  339. chan->global_identifier, chan->scheduler_state, new_state);
  340. chan->scheduler_state = new_state;
  341. }
  342. /** Return the pending channel list. */
  343. smartlist_t *
  344. get_channels_pending(void)
  345. {
  346. return channels_pending;
  347. }
  348. /** Comparison function to use when sorting pending channels. */
  349. MOCK_IMPL(int,
  350. scheduler_compare_channels, (const void *c1_v, const void *c2_v))
  351. {
  352. const channel_t *c1 = NULL, *c2 = NULL;
  353. /* These are a workaround for -Wbad-function-cast throwing a fit */
  354. const circuitmux_policy_t *p1, *p2;
  355. uintptr_t p1_i, p2_i;
  356. tor_assert(c1_v);
  357. tor_assert(c2_v);
  358. c1 = (const channel_t *)(c1_v);
  359. c2 = (const channel_t *)(c2_v);
  360. if (c1 != c2) {
  361. if (circuitmux_get_policy(c1->cmux) ==
  362. circuitmux_get_policy(c2->cmux)) {
  363. /* Same cmux policy, so use the mux comparison */
  364. return circuitmux_compare_muxes(c1->cmux, c2->cmux);
  365. } else {
  366. /*
  367. * Different policies; not important to get this edge case perfect
  368. * because the current code never actually gives different channels
  369. * different cmux policies anyway. Just use this arbitrary but
  370. * definite choice.
  371. */
  372. p1 = circuitmux_get_policy(c1->cmux);
  373. p2 = circuitmux_get_policy(c2->cmux);
  374. p1_i = (uintptr_t)p1;
  375. p2_i = (uintptr_t)p2;
  376. return (p1_i < p2_i) ? -1 : 1;
  377. }
  378. } else {
  379. /* c1 == c2, so always equal */
  380. return 0;
  381. }
  382. }
  383. /*****************************************************************************
  384. * Scheduling system global functions
  385. *
  386. * Functions that can be accessed from anywhere in Tor.
  387. *****************************************************************************/
  388. /**
  389. * This is how the scheduling system is notified of Tor's configuration
  390. * changing. For example: a SIGHUP was issued.
  391. */
  392. void
  393. scheduler_conf_changed(void)
  394. {
  395. /* Let the scheduler decide what it should do. */
  396. set_scheduler();
  397. /* Then tell the (possibly new) scheduler that we have new options. */
  398. if (the_scheduler->on_new_options) {
  399. the_scheduler->on_new_options();
  400. }
  401. }
  402. /**
  403. * Whenever we get a new consensus, this function is called.
  404. */
  405. void
  406. scheduler_notify_networkstatus_changed(const networkstatus_t *old_c,
  407. const networkstatus_t *new_c)
  408. {
  409. /* Maybe the consensus param made us change the scheduler. */
  410. set_scheduler();
  411. /* Then tell the (possibly new) scheduler that we have a new consensus */
  412. if (the_scheduler->on_new_consensus) {
  413. the_scheduler->on_new_consensus(old_c, new_c);
  414. }
  415. }
  416. /**
  417. * Free everything scheduling-related from main.c. Note this is only called
  418. * when Tor is shutting down, while scheduler_t->free_all() is called both when
  419. * Tor is shutting down and when we are switching schedulers.
  420. */
  421. void
  422. scheduler_free_all(void)
  423. {
  424. log_debug(LD_SCHED, "Shutting down scheduler");
  425. if (run_sched_ev) {
  426. if (event_del(run_sched_ev) < 0) {
  427. log_warn(LD_BUG, "Problem deleting run_sched_ev");
  428. }
  429. tor_event_free(run_sched_ev);
  430. run_sched_ev = NULL;
  431. }
  432. if (channels_pending) {
  433. /* We don't have ownership of the objects in this list. */
  434. smartlist_free(channels_pending);
  435. channels_pending = NULL;
  436. }
  437. if (the_scheduler && the_scheduler->free_all) {
  438. the_scheduler->free_all();
  439. }
  440. the_scheduler = NULL;
  441. }
  442. /** Mark a channel as no longer ready to accept writes. */
  443. MOCK_IMPL(void,
  444. scheduler_channel_doesnt_want_writes,(channel_t *chan))
  445. {
  446. IF_BUG_ONCE(!chan) {
  447. return;
  448. }
  449. IF_BUG_ONCE(!channels_pending) {
  450. return;
  451. }
  452. /* If it's already in pending, we can put it in waiting_to_write */
  453. if (chan->scheduler_state == SCHED_CHAN_PENDING) {
  454. /*
  455. * It's in channels_pending, so it shouldn't be in any of
  456. * the other lists. It can't write any more, so it goes to
  457. * channels_waiting_to_write.
  458. */
  459. smartlist_pqueue_remove(channels_pending,
  460. scheduler_compare_channels,
  461. offsetof(channel_t, sched_heap_idx),
  462. chan);
  463. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_TO_WRITE);
  464. log_debug(LD_SCHED,
  465. "Channel " U64_FORMAT " at %p went from pending "
  466. "to waiting_to_write",
  467. U64_PRINTF_ARG(chan->global_identifier), chan);
  468. } else {
  469. /*
  470. * It's not in pending, so it can't become waiting_to_write; it's
  471. * either not in any of the lists (nothing to do) or it's already in
  472. * waiting_for_cells (remove it, can't write any more).
  473. */
  474. if (chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS) {
  475. scheduler_set_channel_state(chan, SCHED_CHAN_IDLE);
  476. log_debug(LD_SCHED,
  477. "Channel " U64_FORMAT " at %p left waiting_for_cells",
  478. U64_PRINTF_ARG(chan->global_identifier), chan);
  479. }
  480. }
  481. }
  482. /** Mark a channel as having waiting cells. */
  483. MOCK_IMPL(void,
  484. scheduler_channel_has_waiting_cells,(channel_t *chan))
  485. {
  486. IF_BUG_ONCE(!chan) {
  487. return;
  488. }
  489. IF_BUG_ONCE(!channels_pending) {
  490. return;
  491. }
  492. /* First, check if it's also writeable */
  493. if (chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS) {
  494. /*
  495. * It's in channels_waiting_for_cells, so it shouldn't be in any of
  496. * the other lists. It has waiting cells now, so it goes to
  497. * channels_pending.
  498. */
  499. scheduler_set_channel_state(chan, SCHED_CHAN_PENDING);
  500. smartlist_pqueue_add(channels_pending,
  501. scheduler_compare_channels,
  502. offsetof(channel_t, sched_heap_idx),
  503. chan);
  504. log_debug(LD_SCHED,
  505. "Channel " U64_FORMAT " at %p went from waiting_for_cells "
  506. "to pending",
  507. U64_PRINTF_ARG(chan->global_identifier), chan);
  508. /* If we made a channel pending, we potentially have scheduling work to
  509. * do. */
  510. the_scheduler->schedule();
  511. } else {
  512. /*
  513. * It's not in waiting_for_cells, so it can't become pending; it's
  514. * either not in any of the lists (we add it to waiting_to_write)
  515. * or it's already in waiting_to_write or pending (we do nothing)
  516. */
  517. if (!(chan->scheduler_state == SCHED_CHAN_WAITING_TO_WRITE ||
  518. chan->scheduler_state == SCHED_CHAN_PENDING)) {
  519. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_TO_WRITE);
  520. log_debug(LD_SCHED,
  521. "Channel " U64_FORMAT " at %p entered waiting_to_write",
  522. U64_PRINTF_ARG(chan->global_identifier), chan);
  523. }
  524. }
  525. }
  526. /** Add the scheduler event to the set of pending events with next_run being
  527. * the longest time libevent should wait before triggering the event. */
  528. void
  529. scheduler_ev_add(const struct timeval *next_run)
  530. {
  531. tor_assert(run_sched_ev);
  532. tor_assert(next_run);
  533. if (BUG(event_add(run_sched_ev, next_run) < 0)) {
  534. log_warn(LD_SCHED, "Adding to libevent failed. Next run time was set to: "
  535. "%ld.%06ld", next_run->tv_sec, (long)next_run->tv_usec);
  536. return;
  537. }
  538. }
  539. /** Make the scheduler event active with the given flags. */
  540. void
  541. scheduler_ev_active(int flags)
  542. {
  543. tor_assert(run_sched_ev);
  544. event_active(run_sched_ev, flags, 1);
  545. }
  546. /*
  547. * Initialize everything scheduling-related from config.c. Note this is only
  548. * called when Tor is starting up, while scheduler_t->init() is called both
  549. * when Tor is starting up and when we are switching schedulers.
  550. */
  551. void
  552. scheduler_init(void)
  553. {
  554. log_debug(LD_SCHED, "Initting scheduler");
  555. // Two '!' because we really do want to check if the pointer is non-NULL
  556. IF_BUG_ONCE(!!run_sched_ev) {
  557. log_warn(LD_SCHED, "We should not already have a libevent scheduler event."
  558. "I'll clean the old one up, but this is odd.");
  559. tor_event_free(run_sched_ev);
  560. run_sched_ev = NULL;
  561. }
  562. run_sched_ev = tor_event_new(tor_libevent_get_base(), -1,
  563. 0, scheduler_evt_callback, NULL);
  564. channels_pending = smartlist_new();
  565. set_scheduler();
  566. }
  567. /*
  568. * If a channel is going away, this is how the scheduling system is informed
  569. * so it can do any freeing necessary. This ultimately calls
  570. * scheduler_t->on_channel_free() so the current scheduler can release any
  571. * state specific to this channel.
  572. */
  573. MOCK_IMPL(void,
  574. scheduler_release_channel,(channel_t *chan))
  575. {
  576. IF_BUG_ONCE(!chan) {
  577. return;
  578. }
  579. IF_BUG_ONCE(!channels_pending) {
  580. return;
  581. }
  582. if (chan->scheduler_state == SCHED_CHAN_PENDING) {
  583. if (SCHED_BUG(smartlist_pos(channels_pending, chan) == -1, chan)) {
  584. log_warn(LD_SCHED, "Scheduler asked to release channel %" PRIu64 " "
  585. "but it wasn't in channels_pending",
  586. chan->global_identifier);
  587. } else {
  588. smartlist_pqueue_remove(channels_pending,
  589. scheduler_compare_channels,
  590. offsetof(channel_t, sched_heap_idx),
  591. chan);
  592. }
  593. }
  594. if (the_scheduler->on_channel_free) {
  595. the_scheduler->on_channel_free(chan);
  596. }
  597. scheduler_set_channel_state(chan, SCHED_CHAN_IDLE);
  598. }
  599. /** Mark a channel as ready to accept writes */
  600. void
  601. scheduler_channel_wants_writes(channel_t *chan)
  602. {
  603. IF_BUG_ONCE(!chan) {
  604. return;
  605. }
  606. IF_BUG_ONCE(!channels_pending) {
  607. return;
  608. }
  609. /* If it's already in waiting_to_write, we can put it in pending */
  610. if (chan->scheduler_state == SCHED_CHAN_WAITING_TO_WRITE) {
  611. /*
  612. * It can write now, so it goes to channels_pending.
  613. */
  614. log_debug(LD_SCHED, "chan=%" PRIu64 " became pending",
  615. chan->global_identifier);
  616. smartlist_pqueue_add(channels_pending,
  617. scheduler_compare_channels,
  618. offsetof(channel_t, sched_heap_idx),
  619. chan);
  620. scheduler_set_channel_state(chan, SCHED_CHAN_PENDING);
  621. log_debug(LD_SCHED,
  622. "Channel " U64_FORMAT " at %p went from waiting_to_write "
  623. "to pending",
  624. U64_PRINTF_ARG(chan->global_identifier), chan);
  625. /* We just made a channel pending, we have scheduling work to do. */
  626. the_scheduler->schedule();
  627. } else {
  628. /*
  629. * It's not in SCHED_CHAN_WAITING_TO_WRITE, so it can't become pending;
  630. * it's either idle and goes to WAITING_FOR_CELLS, or it's a no-op.
  631. */
  632. if (!(chan->scheduler_state == SCHED_CHAN_WAITING_FOR_CELLS ||
  633. chan->scheduler_state == SCHED_CHAN_PENDING)) {
  634. scheduler_set_channel_state(chan, SCHED_CHAN_WAITING_FOR_CELLS);
  635. log_debug(LD_SCHED,
  636. "Channel " U64_FORMAT " at %p entered waiting_for_cells",
  637. U64_PRINTF_ARG(chan->global_identifier), chan);
  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 %d."
  653. " Num cells on cmux: %d. Connection outbuf len: %lu.",
  654. chan->global_identifier,
  655. channel_state_to_string(chan->state),
  656. chan->scheduler_state, 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) */