bufferevent_ratelim.c

/*
 * Copyright (c) 2007-2009 Niels Provos and Nick Mathewson
 * Copyright (c) 2002-2006 Niels Provos <provos@citi.umich.edu>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. The name of the author may not be used to endorse or promote products
 *    derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include <sys/types.h>
#include <limits.h>
#include <string.h>
#include <stdlib.h>

#include "event2/event.h"
#include "event2/event_struct.h"
#include "event2/util.h"
#include "event2/bufferevent.h"
#include "event2/bufferevent_struct.h"
#include "event2/buffer.h"

#include "ratelim-internal.h"

#include "bufferevent-internal.h"
#include "mm-internal.h"
#include "util-internal.h"

int
ev_token_bucket_init(struct ev_token_bucket *bucket,
    const struct ev_token_bucket_cfg *cfg,
    ev_uint32_t current_tick,
    int reinitialize)
{
	if (reinitialize) {
		/* on reinitialization, we only clip downwards, since we've
		   already used who-knows-how-much bandwidth this tick.  We
		   leave "last_updated" as it is; the next update will add the
		   appropriate amount of bandwidth to the bucket.
		*/
		if (bucket->read_limit > cfg->read_maximum)
			bucket->read_limit = cfg->read_maximum;
		if (bucket->write_limit > cfg->write_maximum)
			bucket->write_limit = cfg->write_maximum;
	} else {
		bucket->read_limit = cfg->read_rate;
		bucket->write_limit = cfg->write_rate;
		bucket->last_updated = current_tick;
	}
	return 0;
}

int
ev_token_bucket_update(struct ev_token_bucket *bucket,
    const struct ev_token_bucket_cfg *cfg,
    ev_uint32_t current_tick)
{
	/* It's okay if the tick number overflows, since we'll just
	 * wrap around when we do the unsigned substraction. */
	unsigned n_ticks = current_tick - bucket->last_updated;

	/* Make sure some ticks actually happened, and that time didn't
	 * roll back. */
	if (n_ticks == 0 || n_ticks > INT_MAX)
		return 0;

	/* Naively, we would say
	       bucket->limit += n_ticks * cfg->rate;

	       if (bucket->limit > cfg->maximum)
	           bucket->limit = cfg->maximum;

	   But we're worried about overflow, so we do it like this:
	*/

	if ((cfg->read_maximum - bucket->read_limit) / n_ticks < cfg->read_rate)
		bucket->read_limit = cfg->read_maximum;
	else
		bucket->read_limit += n_ticks * cfg->read_rate;


	if ((cfg->write_maximum - bucket->write_limit) / n_ticks < cfg->write_rate)
		bucket->write_limit = cfg->write_maximum;
	else
		bucket->write_limit += n_ticks * cfg->write_rate;


	bucket->last_updated = current_tick;

	return 1;
}

ev_uint32_t
ev_token_bucket_get_tick(const struct timeval *tv,
    const struct ev_token_bucket_cfg *cfg)
{
	/* This computation uses two multiplies and a divide.  We could do
	 * fewer if we knew that the tick length was an integer number of
	 * seconds, or if we knew it divided evenly into a second.  We should
	 * investigate that more.
	 */

	/* We cast to an ev_uint64_t first, since we don't want to overflow
	 * before we do the final divide. */
	ev_uint64_t msec = (ev_uint64_t)tv->tv_sec * 1000 + tv->tv_usec / 1000;
	return (unsigned)(msec / cfg->msec_per_tick);
}

struct ev_token_bucket_cfg *
ev_token_bucket_cfg_new(ev_uint32_t read_rate, ev_uint32_t read_burst,
    ev_uint32_t write_rate, ev_uint32_t write_burst,
    const struct timeval *tick_len)
{
	struct ev_token_bucket_cfg *r;
	struct timeval g;
	if (! tick_len) {
		g.tv_sec = 1;
		g.tv_usec = 0;
		tick_len = &g;
	}
	if (read_rate > read_burst || write_rate > write_burst ||
	    read_rate < 1 || write_rate < 1)
		return NULL;
	r = mm_calloc(1, sizeof(struct ev_token_bucket_cfg));
	if (!r)
		return NULL;
	r->read_rate = read_rate;
	r->write_rate = write_rate;
	r->read_maximum = read_burst;
	r->write_maximum = write_burst;
	memcpy(&r->tick_timeout, tick_len, sizeof(struct timeval));
	r->msec_per_tick = (tick_len->tv_sec * 1000) + tick_len->tv_usec/1000;
	return r;
}

void
ev_token_bucket_cfg_free(struct ev_token_bucket_cfg *cfg)
{
	mm_free(cfg);
}

/* No matter how big our bucket gets, don't try to read more than this
 * much in a single read operation. */
#define MAX_TO_READ_EVER 16384
/* No matter how big our bucket gets, don't try to write more than this
 * much in a single write operation. */
#define MAX_TO_WRITE_EVER 16384

#define LOCK_GROUP(g) EVLOCK_LOCK((g)->lock, 0)
#define UNLOCK_GROUP(g) EVLOCK_UNLOCK((g)->lock, 0)

static int _bev_group_suspend_reading(struct bufferevent_rate_limit_group *g);
static int _bev_group_suspend_writing(struct bufferevent_rate_limit_group *g);

/** Helper: figure out the maximum amount we should write if is_write, or
    the maximum amount we should read if is_read.  Return that maximum, or
    0 if our bucket is wholly exhausted.
 */
static inline int
_bufferevent_get_rlim_max(struct bufferevent_private *bev, int is_write)
{
	/* needs lock on bev. */
	int max_so_far = is_write?MAX_TO_WRITE_EVER:MAX_TO_READ_EVER;
	struct timeval now;

#define LIM(x)						\
	(is_write ? (x).write_limit : (x).read_limit)

#define GROUP_SUSPENDED(g)			\
	(is_write ? (g)->write_suspended : (g)->read_suspended)

	/* Sets max_so_far to MIN(x, max_so_far) */
#define CLAMPTO(x)				\
	do {					\
		if (max_so_far > (x))		\
			max_so_far = (x);	\
	} while (0);

	if (!bev->rate_limiting)
		return max_so_far;

	/* If rate-limiting is enabled at all, update the appropriate
	   bucket, and take the smaller of our rate limit and the group
	   rate limit.
	 */

	if (bev->rate_limiting->cfg) {
		unsigned tick;

		event_base_gettimeofday_cached(bev->bev.ev_base, &now);
		tick = ev_token_bucket_get_tick(&now, bev->rate_limiting->cfg);
		ev_token_bucket_update(&bev->rate_limiting->limit,
		    bev->rate_limiting->cfg, tick);
		max_so_far = LIM(bev->rate_limiting->limit);
	}
	if (bev->rate_limiting->group) {
		struct bufferevent_rate_limit_group *g =
		    bev->rate_limiting->group;
		ev_uint32_t share;
		LOCK_GROUP(g);
		if (GROUP_SUSPENDED(g)) {
			/* We can get here if we failed to lock this
			 * particular bufferevent while suspending the whole
			 * group. */
			if (is_write)
				bufferevent_suspend_write(&bev->bev,
				    BEV_SUSPEND_BW_GROUP);
			else
				bufferevent_suspend_read(&bev->bev,
				    BEV_SUSPEND_BW_GROUP);
			share = 0;
		} else {
			/* XXXX probably we should divide among the active
			 * members, not the total members. */
			share = LIM(g->rate_limit) / g->n_members;
			if (share < g->min_share)
				share = g->min_share;
		}
		UNLOCK_GROUP(g);
		CLAMPTO(share);
	}

	return max_so_far;
}

int
_bufferevent_get_read_max(struct bufferevent_private *bev)
{
	return _bufferevent_get_rlim_max(bev, 0);
}

int
_bufferevent_get_write_max(struct bufferevent_private *bev)
{
	return _bufferevent_get_rlim_max(bev, 1);
}

int
_bufferevent_decrement_read_buckets(struct bufferevent_private *bev, int bytes)
{
	/* XXXXX Make sure all users of this function check its return value */
	int r = 0;
	/* need to hold lock on bev */
	if (!bev->rate_limiting)
		return 0;

	if (bev->rate_limiting->cfg) {
		bev->rate_limiting->limit.read_limit -= bytes;
		if (bev->rate_limiting->limit.read_limit <= 0) {
			bufferevent_suspend_read(&bev->bev, BEV_SUSPEND_BW);
			if (event_add(&bev->rate_limiting->refill_bucket_event,
				&bev->rate_limiting->cfg->tick_timeout) < 0)
				r = -1;
		}
	}

	if (bev->rate_limiting->group) {
		LOCK_GROUP(bev->rate_limiting->group);
		bev->rate_limiting->group->rate_limit.read_limit -= bytes;
		if (bev->rate_limiting->group->rate_limit.read_limit <= 0) {
			_bev_group_suspend_reading(bev->rate_limiting->group);
		}
		UNLOCK_GROUP(bev->rate_limiting->group);
	}

	return r;
}

int
_bufferevent_decrement_write_buckets(struct bufferevent_private *bev, int bytes)
{
	/* XXXXX Make sure all users of this function check its return value */
	int r = 0;
	/* need to hold lock */
	if (!bev->rate_limiting)
		return 0;

	if (bev->rate_limiting->cfg) {
		bev->rate_limiting->limit.write_limit -= bytes;
		if (bev->rate_limiting->limit.write_limit <= 0) {
			bufferevent_suspend_write(&bev->bev, BEV_SUSPEND_BW);
			if (event_add(&bev->rate_limiting->refill_bucket_event,
				&bev->rate_limiting->cfg->tick_timeout) < 0)
				r = -1;
		}
	}

	if (bev->rate_limiting->group) {
		LOCK_GROUP(bev->rate_limiting->group);
		bev->rate_limiting->group->rate_limit.write_limit -= bytes;
		if (bev->rate_limiting->group->rate_limit.write_limit <= 0) {
			_bev_group_suspend_writing(bev->rate_limiting->group);
		}
		UNLOCK_GROUP(bev->rate_limiting->group);
	}

	return r;
}

/** Stop reading on every bufferevent in <b>g</b> */
static int
_bev_group_suspend_reading(struct bufferevent_rate_limit_group *g)
{
	/* Needs group lock */
	struct bufferevent_private *bev;
	g->read_suspended = 1;
	g->pending_unsuspend_read = 0;

	/* Note that in this loop we call EVLOCK_TRY_LOCK instead of BEV_LOCK,
	   to prevent a deadlock.  (Ordinarily, the group lock nests inside
	   the bufferevent locks.  If we are unable to lock any individual
	   bufferevent, it will find out later when it looks at its limit
	   and sees that its group is suspended.
	*/
	TAILQ_FOREACH(bev, &g->members, rate_limiting->next_in_group) {
		if (EVLOCK_TRY_LOCK(bev->lock)) {
			bufferevent_suspend_read(&bev->bev,
			    BEV_SUSPEND_BW_GROUP);
			EVLOCK_UNLOCK(bev->lock, 0);
		}
	}
	return 0;
}

/** Stop writing on every bufferevent in <b>g</b> */
static int
_bev_group_suspend_writing(struct bufferevent_rate_limit_group *g)
{
	/* Needs group lock */
	struct bufferevent_private *bev;
	g->write_suspended = 1;
	g->pending_unsuspend_write = 0;
	TAILQ_FOREACH(bev, &g->members, rate_limiting->next_in_group) {
		if (EVLOCK_TRY_LOCK(bev->lock)) {
			bufferevent_suspend_write(&bev->bev,
			    BEV_SUSPEND_BW_GROUP);
			EVLOCK_UNLOCK(bev->lock, 0);
		}
	}
	return 0;
}

/** Timer callback invoked on a single bufferevent with one or more exhausted
    buckets when they are ready to refill. */
static void
_bev_refill_callback(evutil_socket_t fd, short what, void *arg)
{
	unsigned tick;
	struct timeval now;
	struct bufferevent_private *bev = arg;
	int again = 0;
	BEV_LOCK(&bev->bev);
	if (!bev->rate_limiting || !bev->rate_limiting->cfg) {
		BEV_UNLOCK(&bev->bev);
		return;
	}

	/* First, update the bucket */
	event_base_gettimeofday_cached(bev->bev.ev_base, &now);
	tick = ev_token_bucket_get_tick(&now,
	    bev->rate_limiting->cfg);
	ev_token_bucket_update(&bev->rate_limiting->limit,
	    bev->rate_limiting->cfg,
	    tick);

	/* Now unsuspend any read/write operations as appropriate. */
	if ((bev->read_suspended & BEV_SUSPEND_BW)) {
		if (bev->rate_limiting->limit.read_limit > 0)
			bufferevent_unsuspend_read(&bev->bev, BEV_SUSPEND_BW);
		else
			again = 1;
	}
	if ((bev->write_suspended & BEV_SUSPEND_BW)) {
		if (bev->rate_limiting->limit.write_limit > 0)
			bufferevent_unsuspend_write(&bev->bev, BEV_SUSPEND_BW);
		else
			again = 1;
	}
	if (again) {
		/* One or more of the buckets may need another refill if they
		   started negative.

		   XXXX if we need to be quiet for more ticks, we should
		   maybe figure out what timeout we really want.
		*/
		/* XXXX Handle event_add failure somehow */
		event_add(&bev->rate_limiting->refill_bucket_event,
		    &bev->rate_limiting->cfg->tick_timeout);
	}
	BEV_UNLOCK(&bev->bev);
}

/** Helper: grab a random element from a bufferevent group. */
static struct bufferevent_private *
_bev_group_random_element(struct bufferevent_rate_limit_group *group)
{
	int which;
	struct bufferevent_private *bev;

	/* requires group lock */

	if (!group->n_members)
		return NULL;

	EVUTIL_ASSERT(! TAILQ_EMPTY(&group->members));

	which = _evutil_weakrand() % group->n_members;

	bev = TAILQ_FIRST(&group->members);
	while (which--)
		bev = TAILQ_NEXT(bev, rate_limiting->next_in_group);

	return bev;
}

/** Iterate over the elements of a rate-limiting group 'g' with a random
    starting point, assigning each to the variable 'bev', and executing the
    block 'block'.

    We do this in a half-baked effort to get fairness among group members.
    XXX Round-robin or some kind of priority queue would be even more fair.
 */
#define FOREACH_RANDOM_ORDER(block)			\
	do {						\
		first = _bev_group_random_element(g);	\
		for (bev = first; bev != TAILQ_END(&g->members); \
		    bev = TAILQ_NEXT(bev, rate_limiting->next_in_group)) { \
			block ;					 \
		}						 \
		for (bev = TAILQ_FIRST(&g->members); bev && bev != first; \
		    bev = TAILQ_NEXT(bev, rate_limiting->next_in_group)) { \
			block ;						\
		}							\
	} while (0)

/** Callback invoked every tick to add more elements to the group bucket
    and unsuspend group members as needed.
 */
static void
_bev_group_refill_callback(evutil_socket_t fd, short what, void *arg)
{
	struct bufferevent_rate_limit_group *g = arg;
	unsigned tick;
	struct timeval now;
	int again = 0;
	struct bufferevent_private *bev, *first;

	event_base_gettimeofday_cached(event_get_base(&g->master_refill_event), &now);

	LOCK_GROUP(g);
	tick = ev_token_bucket_get_tick(&now, &g->rate_limit_cfg);
	ev_token_bucket_update(&g->rate_limit, &g->rate_limit_cfg, tick);

	if (g->pending_unsuspend_read ||
	    (g->read_suspended && (g->rate_limit.read_limit >= g->min_share))) {
		g->read_suspended = 0;
		again = 0;
		FOREACH_RANDOM_ORDER({
			if (EVLOCK_TRY_LOCK(bev->lock)) {
				bufferevent_unsuspend_read(&bev->bev,
				    BEV_SUSPEND_BW_GROUP);
				EVLOCK_UNLOCK(bev->lock, 0);
			} else {
				again = 1;
			}
		});
		g->pending_unsuspend_read = again;
	}
	if (g->pending_unsuspend_write ||
	    (g->write_suspended && (g->rate_limit.write_limit >= g->min_share))){
		g->write_suspended = 0;
		again = 0;
		FOREACH_RANDOM_ORDER({
			if (EVLOCK_TRY_LOCK(bev->lock)) {
				bufferevent_unsuspend_write(&bev->bev,
				    BEV_SUSPEND_BW_GROUP);
				EVLOCK_UNLOCK(bev->lock, 0);
			} else {
				again = 1;
			}
		});
		g->pending_unsuspend_write = again;
	}

	/* XXXX Rather than waiting to the next tick to unsuspend stuff
	 * with pending_unsuspend_write/read, we should do it on the
	 * next iteration of the mainloop.
	 */

	UNLOCK_GROUP(g);
}

int
bufferevent_set_rate_limit(struct bufferevent *bev,
    struct ev_token_bucket_cfg *cfg)
{
	struct bufferevent_private *bevp =
	    EVUTIL_UPCAST(bev, struct bufferevent_private, bev);
	int r = -1;
	struct bufferevent_rate_limit *rlim;
	struct timeval now;
	ev_uint32_t tick;
	/* XXX reference-count cfg */

	BEV_LOCK(bev);

	if (cfg == NULL) {
		if (bevp->rate_limiting) {
			bevp->rate_limiting->cfg = NULL;
			bufferevent_unsuspend_read(bev, BEV_SUSPEND_BW);
			bufferevent_unsuspend_write(bev, BEV_SUSPEND_BW);
		}
		r = 0;
		goto done;
	}

	event_base_gettimeofday_cached(bev->ev_base, &now);
	tick = ev_token_bucket_get_tick(&now, cfg);

	if (bevp->rate_limiting && bevp->rate_limiting->cfg == cfg) {
		;
	} else if (bevp->rate_limiting) {
		bevp->rate_limiting->cfg = cfg;
		ev_token_bucket_init(&bevp->rate_limiting->limit, cfg, tick, 1);
		if (bevp->rate_limiting->limit.read_limit > 0)
			bufferevent_unsuspend_read(bev, BEV_SUSPEND_BW);
		else
			bufferevent_suspend_read(bev, BEV_SUSPEND_BW);
		if (bevp->rate_limiting->limit.write_limit > 0)
			bufferevent_unsuspend_write(bev, BEV_SUSPEND_BW);
		else
			bufferevent_suspend_write(bev, BEV_SUSPEND_BW);
	} else {
		rlim = mm_calloc(1, sizeof(struct bufferevent_rate_limit));
		if (!rlim)
			goto done;
		rlim->cfg = cfg;
		ev_token_bucket_init(&rlim->limit, cfg, tick, 0);
		evtimer_assign(&rlim->refill_bucket_event, bev->ev_base,
		    _bev_refill_callback, bevp);
		bevp->rate_limiting = rlim;
	}
	r = 0;
done:
	BEV_UNLOCK(bev);
	return r;
}

struct bufferevent_rate_limit_group *
bufferevent_rate_limit_group_new(struct event_base *base,
    const struct ev_token_bucket_cfg *cfg)
{
	struct bufferevent_rate_limit_group *g;
	struct timeval now;
	ev_uint32_t tick;

	event_base_gettimeofday_cached(base, &now);
	tick = ev_token_bucket_get_tick(&now, cfg);

	g = mm_calloc(1, sizeof(struct bufferevent_rate_limit_group));
	if (!g)
		return NULL;
	memcpy(&g->rate_limit_cfg, cfg, sizeof(g->rate_limit_cfg));
	TAILQ_INIT(&g->members);

	ev_token_bucket_init(&g->rate_limit, cfg, tick, 0);

	g->min_share = 64;
	event_assign(&g->master_refill_event, base, -1, EV_PERSIST,
	    _bev_group_refill_callback, g);
	/*XXXX handle event_add failure */
	event_add(&g->master_refill_event, &cfg->tick_timeout);

	EVTHREAD_ALLOC_LOCK(g->lock, EVTHREAD_LOCKTYPE_RECURSIVE);
	return g;
}

int
bufferevent_add_to_rate_limit_group(struct bufferevent *bev,
    struct bufferevent_rate_limit_group *g)
{
	int wsuspend, rsuspend;
	struct bufferevent_private *bevp =
	    EVUTIL_UPCAST(bev, struct bufferevent_private, bev);
	BEV_LOCK(bev);

	if (!bevp->rate_limiting) {
		struct bufferevent_rate_limit *rlim;
		rlim = mm_calloc(1, sizeof(struct bufferevent_rate_limit));
		if (!rlim) {
			BEV_UNLOCK(bev);
			return -1;
		}
		evtimer_assign(&rlim->refill_bucket_event, bev->ev_base,
		    _bev_refill_callback, bevp);
		bevp->rate_limiting = rlim;
	}

	if (bevp->rate_limiting->group == g) {
		BEV_UNLOCK(bev);
		return 0;
	}
	if (bevp->rate_limiting->group)
		bufferevent_remove_from_rate_limit_group(bev);

	LOCK_GROUP(g);
	bevp->rate_limiting->group = g;
	++g->n_members;
	TAILQ_INSERT_TAIL(&g->members, bevp, rate_limiting->next_in_group);

	rsuspend = g->read_suspended;
	wsuspend = g->write_suspended;

	UNLOCK_GROUP(g);

	if (rsuspend)
		bufferevent_suspend_read(bev, BEV_SUSPEND_BW_GROUP);
	if (wsuspend)
		bufferevent_suspend_write(bev, BEV_SUSPEND_BW_GROUP);

	BEV_UNLOCK(bev);
	return 0;
}

int
bufferevent_remove_from_rate_limit_group(struct bufferevent *bev)
{
	struct bufferevent_private *bevp =
	    EVUTIL_UPCAST(bev, struct bufferevent_private, bev);
	BEV_LOCK(bev);
	if (bevp->rate_limiting && bevp->rate_limiting->group) {
		struct bufferevent_rate_limit_group *g =
		    bevp->rate_limiting->group;
		LOCK_GROUP(g);
		bevp->rate_limiting->group = NULL;
		--g->n_members;
		TAILQ_REMOVE(&g->members, bevp, rate_limiting->next_in_group);
		UNLOCK_GROUP(g);
	}
	bufferevent_unsuspend_read(bev, BEV_SUSPEND_BW_GROUP);
	bufferevent_unsuspend_write(bev, BEV_SUSPEND_BW_GROUP);
	BEV_UNLOCK(bev);
	return 0;
}