diff options
Diffstat (limited to 'src/ext/timeouts/timeout.c')
-rw-r--r-- | src/ext/timeouts/timeout.c | 744 |
1 files changed, 744 insertions, 0 deletions
diff --git a/src/ext/timeouts/timeout.c b/src/ext/timeouts/timeout.c new file mode 100644 index 0000000000..e78f57d25b --- /dev/null +++ b/src/ext/timeouts/timeout.c @@ -0,0 +1,744 @@ +/* ========================================================================== + * timeout.c - Tickless hierarchical timing wheel. + * -------------------------------------------------------------------------- + * Copyright (c) 2013, 2014 William Ahern + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sublicense, and/or sell copies of the Software, and to permit + * persons to whom the Software is furnished to do so, subject to the + * following conditions: + * + * The above copyright notice and this permission notice shall be included + * in all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN + * NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, + * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR + * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE + * USE OR OTHER DEALINGS IN THE SOFTWARE. + * ========================================================================== + */ +#include <limits.h> /* CHAR_BIT */ + +#include <stddef.h> /* NULL */ +#include <stdlib.h> /* malloc(3) free(3) */ +#include <stdio.h> /* FILE fprintf(3) */ + +#include <inttypes.h> /* UINT64_C uint64_t */ + +#include <string.h> /* memset(3) */ + +#include <errno.h> /* errno */ + +#include <sys/queue.h> /* TAILQ(3) */ + +#include "timeout.h" + +#if TIMEOUT_DEBUG - 0 +#include "timeout-debug.h" +#endif + +#ifdef TIMEOUT_DISABLE_RELATIVE_ACCESS +#define TO_SET_TIMEOUTS(to, T) ((void)0) +#else +#define TO_SET_TIMEOUTS(to, T) ((to)->timeouts = (T)) +#endif + +/* + * A N C I L L A R Y R O U T I N E S + * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +#define abstime_t timeout_t /* for documentation purposes */ +#define reltime_t timeout_t /* "" */ + +#if !defined countof +#define countof(a) (sizeof (a) / sizeof *(a)) +#endif + +#if !defined endof +#define endof(a) (&(a)[countof(a)]) +#endif + +#if !defined MIN +#define MIN(a, b) (((a) < (b))? (a) : (b)) +#endif + +#if !defined MAX +#define MAX(a, b) (((a) > (b))? (a) : (b)) +#endif + +#if !defined TAILQ_CONCAT +#define TAILQ_CONCAT(head1, head2, field) do { \ + if (!TAILQ_EMPTY(head2)) { \ + *(head1)->tqh_last = (head2)->tqh_first; \ + (head2)->tqh_first->field.tqe_prev = (head1)->tqh_last; \ + (head1)->tqh_last = (head2)->tqh_last; \ + TAILQ_INIT((head2)); \ + } \ +} while (0) +#endif + +#if !defined TAILQ_FOREACH_SAFE +#define TAILQ_FOREACH_SAFE(var, head, field, tvar) \ + for ((var) = TAILQ_FIRST(head); \ + (var) && ((tvar) = TAILQ_NEXT(var, field), 1); \ + (var) = (tvar)) +#endif + + +/* + * B I T M A N I P U L A T I O N R O U T I N E S + * + * The macros and routines below implement wheel parameterization. The + * inputs are: + * + * WHEEL_BIT - The number of value bits mapped in each wheel. The + * lowest-order WHEEL_BIT bits index the lowest-order (highest + * resolution) wheel, the next group of WHEEL_BIT bits the + * higher wheel, etc. + * + * WHEEL_NUM - The number of wheels. WHEEL_BIT * WHEEL_NUM = the number of + * value bits used by all the wheels. For the default of 6 and + * 4, only the low 24 bits are processed. Any timeout value + * larger than this will cycle through again. + * + * The implementation uses bit fields to remember which slot in each wheel + * is populated, and to generate masks of expiring slots according to the + * current update interval (i.e. the "tickless" aspect). The slots to + * process in a wheel are (populated-set & interval-mask). + * + * WHEEL_BIT cannot be larger than 6 bits because 2^6 -> 64 is the largest + * number of slots which can be tracked in a uint64_t integer bit field. + * WHEEL_BIT cannot be smaller than 3 bits because of our rotr and rotl + * routines, which only operate on all the value bits in an integer, and + * there's no integer smaller than uint8_t. + * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +#if !defined WHEEL_BIT +#define WHEEL_BIT 6 +#endif + +#if !defined WHEEL_NUM +#define WHEEL_NUM 4 +#endif + +#define WHEEL_LEN (1U << WHEEL_BIT) +#define WHEEL_MAX (WHEEL_LEN - 1) +#define WHEEL_MASK (WHEEL_LEN - 1) +#define TIMEOUT_MAX ((TIMEOUT_C(1) << (WHEEL_BIT * WHEEL_NUM)) - 1) + +#include "timeout-bitops.c" + +#if WHEEL_BIT == 6 +#define ctz(n) ctz64(n) +#define clz(n) clz64(n) +#define fls(n) ((int)(64 - clz64(n))) +#else +#define ctz(n) ctz32(n) +#define clz(n) clz32(n) +#define fls(n) ((int)(32 - clz32(n))) +#endif + +#if WHEEL_BIT == 6 +#define WHEEL_C(n) UINT64_C(n) +#define WHEEL_PRIu PRIu64 +#define WHEEL_PRIx PRIx64 + +typedef uint64_t wheel_t; + +#elif WHEEL_BIT == 5 + +#define WHEEL_C(n) UINT32_C(n) +#define WHEEL_PRIu PRIu32 +#define WHEEL_PRIx PRIx32 + +typedef uint32_t wheel_t; + +#elif WHEEL_BIT == 4 + +#define WHEEL_C(n) UINT16_C(n) +#define WHEEL_PRIu PRIu16 +#define WHEEL_PRIx PRIx16 + +typedef uint16_t wheel_t; + +#elif WHEEL_BIT == 3 + +#define WHEEL_C(n) UINT8_C(n) +#define WHEEL_PRIu PRIu8 +#define WHEEL_PRIx PRIx8 + +typedef uint8_t wheel_t; + +#else +#error invalid WHEEL_BIT value +#endif + + +static inline wheel_t rotl(const wheel_t v, int c) { + if (!(c &= (sizeof v * CHAR_BIT - 1))) + return v; + + return (v << c) | (v >> (sizeof v * CHAR_BIT - c)); +} /* rotl() */ + + +static inline wheel_t rotr(const wheel_t v, int c) { + if (!(c &= (sizeof v * CHAR_BIT - 1))) + return v; + + return (v >> c) | (v << (sizeof v * CHAR_BIT - c)); +} /* rotr() */ + + +/* + * T I M E R R O U T I N E S + * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +TAILQ_HEAD(timeout_list, timeout); + +struct timeouts { + struct timeout_list wheel[WHEEL_NUM][WHEEL_LEN], expired; + + wheel_t pending[WHEEL_NUM]; + + timeout_t curtime; + timeout_t hertz; +}; /* struct timeouts */ + + +static struct timeouts *timeouts_init(struct timeouts *T, timeout_t hz) { + unsigned i, j; + + for (i = 0; i < countof(T->wheel); i++) { + for (j = 0; j < countof(T->wheel[i]); j++) { + TAILQ_INIT(&T->wheel[i][j]); + } + } + + TAILQ_INIT(&T->expired); + + for (i = 0; i < countof(T->pending); i++) { + T->pending[i] = 0; + } + + T->curtime = 0; + T->hertz = (hz)? hz : TIMEOUT_mHZ; + + return T; +} /* timeouts_init() */ + + +TIMEOUT_PUBLIC struct timeouts *timeouts_open(timeout_t hz, int *error) { + struct timeouts *T; + + if ((T = malloc(sizeof *T))) + return timeouts_init(T, hz); + + *error = errno; + + return NULL; +} /* timeouts_open() */ + + +static void timeouts_reset(struct timeouts *T) { + struct timeout_list reset; + struct timeout *to; + unsigned i, j; + + TAILQ_INIT(&reset); + + for (i = 0; i < countof(T->wheel); i++) { + for (j = 0; j < countof(T->wheel[i]); j++) { + TAILQ_CONCAT(&reset, &T->wheel[i][j], tqe); + } + } + + TAILQ_CONCAT(&reset, &T->expired, tqe); + + TAILQ_FOREACH(to, &reset, tqe) { + to->pending = NULL; + TO_SET_TIMEOUTS(to, NULL); + } +} /* timeouts_reset() */ + + +TIMEOUT_PUBLIC void timeouts_close(struct timeouts *T) { + /* + * NOTE: Delete installed timeouts so timeout_pending() and + * timeout_expired() worked as expected. + */ + timeouts_reset(T); + + free(T); +} /* timeouts_close() */ + + +TIMEOUT_PUBLIC timeout_t timeouts_hz(struct timeouts *T) { + return T->hertz; +} /* timeouts_hz() */ + + +TIMEOUT_PUBLIC void timeouts_del(struct timeouts *T, struct timeout *to) { + if (to->pending) { + TAILQ_REMOVE(to->pending, to, tqe); + + if (to->pending != &T->expired && TAILQ_EMPTY(to->pending)) { + ptrdiff_t index = to->pending - &T->wheel[0][0]; + int wheel = index / WHEEL_LEN; + int slot = index % WHEEL_LEN; + + T->pending[wheel] &= ~(WHEEL_C(1) << slot); + } + + to->pending = NULL; + TO_SET_TIMEOUTS(to, NULL); + } +} /* timeouts_del() */ + + +static inline reltime_t timeout_rem(struct timeouts *T, struct timeout *to) { + return to->expires - T->curtime; +} /* timeout_rem() */ + + +static inline int timeout_wheel(timeout_t timeout) { + /* must be called with timeout != 0, so fls input is nonzero */ + return (fls(MIN(timeout, TIMEOUT_MAX)) - 1) / WHEEL_BIT; +} /* timeout_wheel() */ + + +static inline int timeout_slot(int wheel, timeout_t expires) { + return WHEEL_MASK & ((expires >> (wheel * WHEEL_BIT)) - !!wheel); +} /* timeout_slot() */ + + +static void timeouts_sched(struct timeouts *T, struct timeout *to, timeout_t expires) { + timeout_t rem; + int wheel, slot; + + timeouts_del(T, to); + + to->expires = expires; + + TO_SET_TIMEOUTS(to, T); + + if (expires > T->curtime) { + rem = timeout_rem(T, to); + + /* rem is nonzero since: + * rem == timeout_rem(T,to), + * == to->expires - T->curtime + * and above we have expires > T->curtime. + */ + wheel = timeout_wheel(rem); + slot = timeout_slot(wheel, to->expires); + + to->pending = &T->wheel[wheel][slot]; + TAILQ_INSERT_TAIL(to->pending, to, tqe); + + T->pending[wheel] |= WHEEL_C(1) << slot; + } else { + to->pending = &T->expired; + TAILQ_INSERT_TAIL(to->pending, to, tqe); + } +} /* timeouts_sched() */ + + +#ifndef TIMEOUT_DISABLE_INTERVALS +static void timeouts_readd(struct timeouts *T, struct timeout *to) { + to->expires += to->interval; + + if (to->expires <= T->curtime) { + /* If we've missed the next firing of this timeout, reschedule + * it to occur at the next multiple of its interval after + * the last time that it fired. + */ + timeout_t n = T->curtime - to->expires; + timeout_t r = n % to->interval; + to->expires = T->curtime + (to->interval - r); + } + + timeouts_sched(T, to, to->expires); +} /* timeouts_readd() */ +#endif + + +TIMEOUT_PUBLIC void timeouts_add(struct timeouts *T, struct timeout *to, timeout_t timeout) { +#ifndef TIMEOUT_DISABLE_INTERVALS + if (to->flags & TIMEOUT_INT) + to->interval = MAX(1, timeout); +#endif + + if (to->flags & TIMEOUT_ABS) + timeouts_sched(T, to, timeout); + else + timeouts_sched(T, to, T->curtime + timeout); +} /* timeouts_add() */ + + +TIMEOUT_PUBLIC void timeouts_update(struct timeouts *T, abstime_t curtime) { + timeout_t elapsed = curtime - T->curtime; + struct timeout_list todo; + int wheel; + + TAILQ_INIT(&todo); + + /* + * There's no avoiding looping over every wheel. It's best to keep + * WHEEL_NUM smallish. + */ + for (wheel = 0; wheel < WHEEL_NUM; wheel++) { + wheel_t pending; + + /* + * Calculate the slots expiring in this wheel + * + * If the elapsed time is greater than the maximum period of + * the wheel, mark every position as expiring. + * + * Otherwise, to determine the expired slots fill in all the + * bits between the last slot processed and the current + * slot, inclusive of the last slot. We'll bitwise-AND this + * with our pending set below. + * + * If a wheel rolls over, force a tick of the next higher + * wheel. + */ + if ((elapsed >> (wheel * WHEEL_BIT)) > WHEEL_MAX) { + pending = (wheel_t)~WHEEL_C(0); + } else { + wheel_t _elapsed = WHEEL_MASK & (elapsed >> (wheel * WHEEL_BIT)); + int oslot, nslot; + + /* + * TODO: It's likely that at least one of the + * following three bit fill operations is redundant + * or can be replaced with a simpler operation. + */ + oslot = WHEEL_MASK & (T->curtime >> (wheel * WHEEL_BIT)); + pending = rotl(((UINT64_C(1) << _elapsed) - 1), oslot); + + nslot = WHEEL_MASK & (curtime >> (wheel * WHEEL_BIT)); + pending |= rotr(rotl(((WHEEL_C(1) << _elapsed) - 1), nslot), _elapsed); + pending |= WHEEL_C(1) << nslot; + } + + while (pending & T->pending[wheel]) { + /* ctz input cannot be zero: loop condition. */ + int slot = ctz(pending & T->pending[wheel]); + TAILQ_CONCAT(&todo, &T->wheel[wheel][slot], tqe); + T->pending[wheel] &= ~(UINT64_C(1) << slot); + } + + if (!(0x1 & pending)) + break; /* break if we didn't wrap around end of wheel */ + + /* if we're continuing, the next wheel must tick at least once */ + elapsed = MAX(elapsed, (WHEEL_LEN << (wheel * WHEEL_BIT))); + } + + T->curtime = curtime; + + while (!TAILQ_EMPTY(&todo)) { + struct timeout *to = TAILQ_FIRST(&todo); + + TAILQ_REMOVE(&todo, to, tqe); + to->pending = NULL; + + timeouts_sched(T, to, to->expires); + } + + return; +} /* timeouts_update() */ + + +TIMEOUT_PUBLIC void timeouts_step(struct timeouts *T, reltime_t elapsed) { + timeouts_update(T, T->curtime + elapsed); +} /* timeouts_step() */ + + +TIMEOUT_PUBLIC bool timeouts_pending(struct timeouts *T) { + wheel_t pending = 0; + int wheel; + + for (wheel = 0; wheel < WHEEL_NUM; wheel++) { + pending |= T->pending[wheel]; + } + + return !!pending; +} /* timeouts_pending() */ + + +TIMEOUT_PUBLIC bool timeouts_expired(struct timeouts *T) { + return !TAILQ_EMPTY(&T->expired); +} /* timeouts_expired() */ + + +/* + * Calculate the interval before needing to process any timeouts pending on + * any wheel. + * + * (This is separated from the public API routine so we can evaluate our + * wheel invariant assertions irrespective of the expired queue.) + * + * This might return a timeout value sooner than any installed timeout if + * only higher-order wheels have timeouts pending. We can only know when to + * process a wheel, not precisely when a timeout is scheduled. Our timeout + * accuracy could be off by 2^(N*M)-1 units where N is the wheel number and + * M is WHEEL_BIT. Only timeouts which have fallen through to wheel 0 can be + * known exactly. + * + * We should never return a timeout larger than the lowest actual timeout. + */ +static timeout_t timeouts_int(struct timeouts *T) { + timeout_t timeout = ~TIMEOUT_C(0), _timeout; + timeout_t relmask; + int wheel, slot; + + relmask = 0; + + for (wheel = 0; wheel < WHEEL_NUM; wheel++) { + if (T->pending[wheel]) { + slot = WHEEL_MASK & (T->curtime >> (wheel * WHEEL_BIT)); + + /* ctz input cannot be zero: T->pending[wheel] is + * nonzero, so rotr() is nonzero. */ + _timeout = (ctz(rotr(T->pending[wheel], slot)) + !!wheel) << (wheel * WHEEL_BIT); + /* +1 to higher order wheels as those timeouts are one rotation in the future (otherwise they'd be on a lower wheel or expired) */ + + _timeout -= relmask & T->curtime; + /* reduce by how much lower wheels have progressed */ + + timeout = MIN(_timeout, timeout); + } + + relmask <<= WHEEL_BIT; + relmask |= WHEEL_MASK; + } + + return timeout; +} /* timeouts_int() */ + + +/* + * Calculate the interval our caller can wait before needing to process + * events. + */ +TIMEOUT_PUBLIC timeout_t timeouts_timeout(struct timeouts *T) { + if (!TAILQ_EMPTY(&T->expired)) + return 0; + + return timeouts_int(T); +} /* timeouts_timeout() */ + + +TIMEOUT_PUBLIC struct timeout *timeouts_get(struct timeouts *T) { + if (!TAILQ_EMPTY(&T->expired)) { + struct timeout *to = TAILQ_FIRST(&T->expired); + + TAILQ_REMOVE(&T->expired, to, tqe); + to->pending = NULL; + TO_SET_TIMEOUTS(to, NULL); + +#ifndef TIMEOUT_DISABLE_INTERVALS + if ((to->flags & TIMEOUT_INT) && to->interval > 0) + timeouts_readd(T, to); +#endif + + return to; + } else { + return 0; + } +} /* timeouts_get() */ + + +/* + * Use dumb looping to locate the earliest timeout pending on the wheel so + * our invariant assertions can check the result of our optimized code. + */ +static struct timeout *timeouts_min(struct timeouts *T) { + struct timeout *to, *min = NULL; + unsigned i, j; + + for (i = 0; i < countof(T->wheel); i++) { + for (j = 0; j < countof(T->wheel[i]); j++) { + TAILQ_FOREACH(to, &T->wheel[i][j], tqe) { + if (!min || to->expires < min->expires) + min = to; + } + } + } + + return min; +} /* timeouts_min() */ + + +/* + * Check some basic algorithm invariants. If these invariants fail then + * something is definitely broken. + */ +#define report(...) do { \ + if ((fp)) \ + fprintf(fp, __VA_ARGS__); \ +} while (0) + +#define check(expr, ...) do { \ + if (!(expr)) { \ + report(__VA_ARGS__); \ + return 0; \ + } \ +} while (0) + +TIMEOUT_PUBLIC bool timeouts_check(struct timeouts *T, FILE *fp) { + timeout_t timeout; + struct timeout *to; + + if ((to = timeouts_min(T))) { + check(to->expires > T->curtime, "missed timeout (expires:%" TIMEOUT_PRIu " <= curtime:%" TIMEOUT_PRIu ")\n", to->expires, T->curtime); + + timeout = timeouts_int(T); + check(timeout <= to->expires - T->curtime, "wrong soft timeout (soft:%" TIMEOUT_PRIu " > hard:%" TIMEOUT_PRIu ") (expires:%" TIMEOUT_PRIu "; curtime:%" TIMEOUT_PRIu ")\n", timeout, (to->expires - T->curtime), to->expires, T->curtime); + + timeout = timeouts_timeout(T); + check(timeout <= to->expires - T->curtime, "wrong soft timeout (soft:%" TIMEOUT_PRIu " > hard:%" TIMEOUT_PRIu ") (expires:%" TIMEOUT_PRIu "; curtime:%" TIMEOUT_PRIu ")\n", timeout, (to->expires - T->curtime), to->expires, T->curtime); + } else { + timeout = timeouts_timeout(T); + + if (!TAILQ_EMPTY(&T->expired)) + check(timeout == 0, "wrong soft timeout (soft:%" TIMEOUT_PRIu " != hard:%" TIMEOUT_PRIu ")\n", timeout, TIMEOUT_C(0)); + else + check(timeout == ~TIMEOUT_C(0), "wrong soft timeout (soft:%" TIMEOUT_PRIu " != hard:%" TIMEOUT_PRIu ")\n", timeout, ~TIMEOUT_C(0)); + } + + return 1; +} /* timeouts_check() */ + + +#define ENTER \ + do { \ + static const int pc0 = __LINE__; \ + switch (pc0 + it->pc) { \ + case __LINE__: (void)0 + +#define SAVE_AND_DO(do_statement) \ + do { \ + it->pc = __LINE__ - pc0; \ + do_statement; \ + case __LINE__: (void)0; \ + } while (0) + +#define YIELD(rv) \ + SAVE_AND_DO(return (rv)) + +#define LEAVE \ + SAVE_AND_DO(break); \ + } \ + } while (0) + +TIMEOUT_PUBLIC struct timeout *timeouts_next(struct timeouts *T, struct timeouts_it *it) { + struct timeout *to; + + ENTER; + + if (it->flags & TIMEOUTS_EXPIRED) { + if (it->flags & TIMEOUTS_CLEAR) { + while ((to = timeouts_get(T))) { + YIELD(to); + } + } else { + TAILQ_FOREACH_SAFE(to, &T->expired, tqe, it->to) { + YIELD(to); + } + } + } + + if (it->flags & TIMEOUTS_PENDING) { + for (it->i = 0; it->i < countof(T->wheel); it->i++) { + for (it->j = 0; it->j < countof(T->wheel[it->i]); it->j++) { + TAILQ_FOREACH_SAFE(to, &T->wheel[it->i][it->j], tqe, it->to) { + YIELD(to); + } + } + } + } + + LEAVE; + + return NULL; +} /* timeouts_next */ + +#undef LEAVE +#undef YIELD +#undef SAVE_AND_DO +#undef ENTER + + +/* + * T I M E O U T R O U T I N E S + * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +TIMEOUT_PUBLIC struct timeout *timeout_init(struct timeout *to, int flags) { + memset(to, 0, sizeof *to); + + to->flags = flags; + + return to; +} /* timeout_init() */ + + +#ifndef TIMEOUT_DISABLE_RELATIVE_ACCESS +TIMEOUT_PUBLIC bool timeout_pending(struct timeout *to) { + return to->pending && to->pending != &to->timeouts->expired; +} /* timeout_pending() */ + + +TIMEOUT_PUBLIC bool timeout_expired(struct timeout *to) { + return to->pending && to->pending == &to->timeouts->expired; +} /* timeout_expired() */ + + +TIMEOUT_PUBLIC void timeout_del(struct timeout *to) { + timeouts_del(to->timeouts, to); +} /* timeout_del() */ +#endif + + +/* + * V E R S I O N I N T E R F A C E S + * + * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ + +TIMEOUT_PUBLIC int timeout_version(void) { + return TIMEOUT_VERSION; +} /* timeout_version() */ + + +TIMEOUT_PUBLIC const char *timeout_vendor(void) { + return TIMEOUT_VENDOR; +} /* timeout_version() */ + + +TIMEOUT_PUBLIC int timeout_v_rel(void) { + return TIMEOUT_V_REL; +} /* timeout_version() */ + + +TIMEOUT_PUBLIC int timeout_v_abi(void) { + return TIMEOUT_V_ABI; +} /* timeout_version() */ + + +TIMEOUT_PUBLIC int timeout_v_api(void) { + return TIMEOUT_V_API; +} /* timeout_version() */ + |