[linux.git] / net / core / flow.c

/* flow.c: Generic flow cache.
 *
 * Copyright (C) 2003 Alexey N. Kuznetsov ([email protected])
 * Copyright (C) 2003 David S. Miller ([email protected])
 */

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/list.h>
#include <linux/jhash.h>
#include <linux/interrupt.h>
#include <linux/mm.h>
#include <linux/random.h>
#include <linux/init.h>
#include <linux/slab.h>
#include <linux/smp.h>
#include <linux/completion.h>
#include <linux/percpu.h>
#include <linux/bitops.h>
#include <linux/notifier.h>
#include <linux/cpu.h>
#include <linux/cpumask.h>
#include <linux/mutex.h>
#include <net/flow.h>
#include <asm/atomic.h>
#include <asm/semaphore.h>
#include <linux/security.h>

struct flow_cache_entry {
	struct flow_cache_entry	*next;
	u16			family;
	u8			dir;
	struct flowi		key;
	u32			genid;
	u32			sk_sid;
	void			*object;
	atomic_t		*object_ref;
};

atomic_t flow_cache_genid = ATOMIC_INIT(0);

static u32 flow_hash_shift;
#define flow_hash_size	(1 << flow_hash_shift)
static DEFINE_PER_CPU(struct flow_cache_entry **, flow_tables) = { NULL };

#define flow_table(cpu) (per_cpu(flow_tables, cpu))

static kmem_cache_t *flow_cachep __read_mostly;

static int flow_lwm, flow_hwm;

struct flow_percpu_info {
	int hash_rnd_recalc;
	u32 hash_rnd;
	int count;
} ____cacheline_aligned;
static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };

#define flow_hash_rnd_recalc(cpu) \
	(per_cpu(flow_hash_info, cpu).hash_rnd_recalc)
#define flow_hash_rnd(cpu) \
	(per_cpu(flow_hash_info, cpu).hash_rnd)
#define flow_count(cpu) \
	(per_cpu(flow_hash_info, cpu).count)

static struct timer_list flow_hash_rnd_timer;

#define FLOW_HASH_RND_PERIOD	(10 * 60 * HZ)

struct flow_flush_info {
	atomic_t cpuleft;
	struct completion completion;
};
static DEFINE_PER_CPU(struct tasklet_struct, flow_flush_tasklets) = { NULL };

#define flow_flush_tasklet(cpu) (&per_cpu(flow_flush_tasklets, cpu))

static void flow_cache_new_hashrnd(unsigned long arg)
{
	int i;

	for_each_cpu(i)
		flow_hash_rnd_recalc(i) = 1;

	flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
	add_timer(&flow_hash_rnd_timer);
}

static void __flow_cache_shrink(int cpu, int shrink_to)
{
	struct flow_cache_entry *fle, **flp;
	int i;

	for (i = 0; i < flow_hash_size; i++) {
		int k = 0;

		flp = &flow_table(cpu)[i];
		while ((fle = *flp) != NULL && k < shrink_to) {
			k++;
			flp = &fle->next;
		}
		while ((fle = *flp) != NULL) {
			*flp = fle->next;
			if (fle->object)
				atomic_dec(fle->object_ref);
			kmem_cache_free(flow_cachep, fle);
			flow_count(cpu)--;
		}
	}
}

static void flow_cache_shrink(int cpu)
{
	int shrink_to = flow_lwm / flow_hash_size;

	__flow_cache_shrink(cpu, shrink_to);
}

static void flow_new_hash_rnd(int cpu)
{
	get_random_bytes(&flow_hash_rnd(cpu), sizeof(u32));
	flow_hash_rnd_recalc(cpu) = 0;

	__flow_cache_shrink(cpu, 0);
}

static u32 flow_hash_code(struct flowi *key, int cpu)
{
	u32 *k = (u32 *) key;

	return (jhash2(k, (sizeof(*key) / sizeof(u32)), flow_hash_rnd(cpu)) &
		(flow_hash_size - 1));
}

#if (BITS_PER_LONG == 64)
typedef u64 flow_compare_t;
#else
typedef u32 flow_compare_t;
#endif

extern void flowi_is_missized(void);

/* I hear what you're saying, use memcmp.  But memcmp cannot make
 * important assumptions that we can here, such as alignment and
 * constant size.
 */
static int flow_key_compare(struct flowi *key1, struct flowi *key2)
{
	flow_compare_t *k1, *k1_lim, *k2;
	const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);

	if (sizeof(struct flowi) % sizeof(flow_compare_t))
		flowi_is_missized();

	k1 = (flow_compare_t *) key1;
	k1_lim = k1 + n_elem;

	k2 = (flow_compare_t *) key2;

	do {
		if (*k1++ != *k2++)
			return 1;
	} while (k1 < k1_lim);

	return 0;
}

void *flow_cache_lookup(struct flowi *key, u32 sk_sid, u16 family, u8 dir,
			flow_resolve_t resolver)
{
	struct flow_cache_entry *fle, **head;
	unsigned int hash;
	int cpu;

	local_bh_disable();
	cpu = smp_processor_id();

	fle = NULL;
	/* Packet really early in init?  Making flow_cache_init a
	 * pre-smp initcall would solve this.  --RR */
	if (!flow_table(cpu))
		goto nocache;

	if (flow_hash_rnd_recalc(cpu))
		flow_new_hash_rnd(cpu);
	hash = flow_hash_code(key, cpu);

	head = &flow_table(cpu)[hash];
	for (fle = *head; fle; fle = fle->next) {
		if (fle->family == family &&
		    fle->dir == dir &&
		    fle->sk_sid == sk_sid &&
		    flow_key_compare(key, &fle->key) == 0) {
			if (fle->genid == atomic_read(&flow_cache_genid)) {
				void *ret = fle->object;

				if (ret)
					atomic_inc(fle->object_ref);
				local_bh_enable();

				return ret;
			}
			break;
		}
	}

	if (!fle) {
		if (flow_count(cpu) > flow_hwm)
			flow_cache_shrink(cpu);

		fle = kmem_cache_alloc(flow_cachep, SLAB_ATOMIC);
		if (fle) {
			fle->next = *head;
			*head = fle;
			fle->family = family;
			fle->dir = dir;
			fle->sk_sid = sk_sid;
			memcpy(&fle->key, key, sizeof(*key));
			fle->object = NULL;
			flow_count(cpu)++;
		}
	}

nocache:
	{
		void *obj;
		atomic_t *obj_ref;

		resolver(key, sk_sid, family, dir, &obj, &obj_ref);

		if (fle) {
			fle->genid = atomic_read(&flow_cache_genid);

			if (fle->object)
				atomic_dec(fle->object_ref);

			fle->object = obj;
			fle->object_ref = obj_ref;
			if (obj)
				atomic_inc(fle->object_ref);
		}
		local_bh_enable();

		return obj;
	}
}

static void flow_cache_flush_tasklet(unsigned long data)
{
	struct flow_flush_info *info = (void *)data;
	int i;
	int cpu;

	cpu = smp_processor_id();
	for (i = 0; i < flow_hash_size; i++) {
		struct flow_cache_entry *fle;

		fle = flow_table(cpu)[i];
		for (; fle; fle = fle->next) {
			unsigned genid = atomic_read(&flow_cache_genid);

			if (!fle->object || fle->genid == genid)
				continue;

			fle->object = NULL;
			atomic_dec(fle->object_ref);
		}
	}

	if (atomic_dec_and_test(&info->cpuleft))
		complete(&info->completion);
}

static void flow_cache_flush_per_cpu(void *) __attribute__((__unused__));
static void flow_cache_flush_per_cpu(void *data)
{
	struct flow_flush_info *info = data;
	int cpu;
	struct tasklet_struct *tasklet;

	cpu = smp_processor_id();

	tasklet = flow_flush_tasklet(cpu);
	tasklet->data = (unsigned long)info;
	tasklet_schedule(tasklet);
}

void flow_cache_flush(void)
{
	struct flow_flush_info info;
	static DEFINE_MUTEX(flow_flush_sem);

	/* Don't want cpus going down or up during this. */
	lock_cpu_hotplug();
	mutex_lock(&flow_flush_sem);
	atomic_set(&info.cpuleft, num_online_cpus());
	init_completion(&info.completion);

	local_bh_disable();
	smp_call_function(flow_cache_flush_per_cpu, &info, 1, 0);
	flow_cache_flush_tasklet((unsigned long)&info);
	local_bh_enable();

	wait_for_completion(&info.completion);
	mutex_unlock(&flow_flush_sem);
	unlock_cpu_hotplug();
}

static void __devinit flow_cache_cpu_prepare(int cpu)
{
	struct tasklet_struct *tasklet;
	unsigned long order;

	for (order = 0;
	     (PAGE_SIZE << order) <
		     (sizeof(struct flow_cache_entry *)*flow_hash_size);
	     order++)
		/* NOTHING */;

	flow_table(cpu) = (struct flow_cache_entry **)
		__get_free_pages(GFP_KERNEL, order);
	if (!flow_table(cpu))
		panic("NET: failed to allocate flow cache order %lu\n", order);

	memset(flow_table(cpu), 0, PAGE_SIZE << order);

	flow_hash_rnd_recalc(cpu) = 1;
	flow_count(cpu) = 0;

	tasklet = flow_flush_tasklet(cpu);
	tasklet_init(tasklet, flow_cache_flush_tasklet, 0);
}

#ifdef CONFIG_HOTPLUG_CPU
static int flow_cache_cpu(struct notifier_block *nfb,
			  unsigned long action,
			  void *hcpu)
{
	if (action == CPU_DEAD)
		__flow_cache_shrink((unsigned long)hcpu, 0);
	return NOTIFY_OK;
}
#endif /* CONFIG_HOTPLUG_CPU */

static int __init flow_cache_init(void)
{
	int i;

	flow_cachep = kmem_cache_create("flow_cache",
					sizeof(struct flow_cache_entry),
					0, SLAB_HWCACHE_ALIGN,
					NULL, NULL);

	if (!flow_cachep)
		panic("NET: failed to allocate flow cache slab\n");

	flow_hash_shift = 10;
	flow_lwm = 2 * flow_hash_size;
	flow_hwm = 4 * flow_hash_size;

	init_timer(&flow_hash_rnd_timer);
	flow_hash_rnd_timer.function = flow_cache_new_hashrnd;
	flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
	add_timer(&flow_hash_rnd_timer);

	for_each_cpu(i)
		flow_cache_cpu_prepare(i);

	hotcpu_notifier(flow_cache_cpu, 0);
	return 0;
}

module_init(flow_cache_init);

EXPORT_SYMBOL(flow_cache_genid);
EXPORT_SYMBOL(flow_cache_lookup);
Commit	Line	Data
1da177e4 LT	1	/* flow.c: Generic flow cache.
	2	*
	3	* Copyright (C) 2003 Alexey N. Kuznetsov ([email protected])
	4	* Copyright (C) 2003 David S. Miller ([email protected])
	5	*/
	6
	7	#include <linux/kernel.h>
	8	#include <linux/module.h>
	9	#include <linux/list.h>
	10	#include <linux/jhash.h>
	11	#include <linux/interrupt.h>
	12	#include <linux/mm.h>
	13	#include <linux/random.h>
	14	#include <linux/init.h>
	15	#include <linux/slab.h>
	16	#include <linux/smp.h>
	17	#include <linux/completion.h>
	18	#include <linux/percpu.h>
	19	#include <linux/bitops.h>
	20	#include <linux/notifier.h>
	21	#include <linux/cpu.h>
	22	#include <linux/cpumask.h>
4a3e2f71	23	#include <linux/mutex.h>
1da177e4 LT	24	#include <net/flow.h>
	25	#include <asm/atomic.h>
	26	#include <asm/semaphore.h>
df71837d	27	#include <linux/security.h>
1da177e4 LT	28
	29	struct flow_cache_entry {
	30	struct flow_cache_entry *next;
	31	u16 family;
	32	u8 dir;
	33	struct flowi key;
	34	u32 genid;
df71837d	35	u32 sk_sid;
1da177e4 LT	36	void *object;
	37	atomic_t *object_ref;
	38	};
	39
	40	atomic_t flow_cache_genid = ATOMIC_INIT(0);
	41
	42	static u32 flow_hash_shift;
	43	#define flow_hash_size (1 << flow_hash_shift)
	44	static DEFINE_PER_CPU(struct flow_cache_entry **, flow_tables) = { NULL };
	45
	46	#define flow_table(cpu) (per_cpu(flow_tables, cpu))
	47
ba89966c	48	static kmem_cache_t *flow_cachep __read_mostly;
1da177e4 LT	49
	50	static int flow_lwm, flow_hwm;
	51
	52	struct flow_percpu_info {
	53	int hash_rnd_recalc;
	54	u32 hash_rnd;
	55	int count;
	56	} ____cacheline_aligned;
	57	static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };
	58
	59	#define flow_hash_rnd_recalc(cpu) \
	60	(per_cpu(flow_hash_info, cpu).hash_rnd_recalc)
	61	#define flow_hash_rnd(cpu) \
	62	(per_cpu(flow_hash_info, cpu).hash_rnd)
	63	#define flow_count(cpu) \
	64	(per_cpu(flow_hash_info, cpu).count)
	65
	66	static struct timer_list flow_hash_rnd_timer;
	67
	68	#define FLOW_HASH_RND_PERIOD (10 * 60 * HZ)
	69
	70	struct flow_flush_info {
	71	atomic_t cpuleft;
	72	struct completion completion;
	73	};
	74	static DEFINE_PER_CPU(struct tasklet_struct, flow_flush_tasklets) = { NULL };
	75
	76	#define flow_flush_tasklet(cpu) (&per_cpu(flow_flush_tasklets, cpu))
	77
	78	static void flow_cache_new_hashrnd(unsigned long arg)
	79	{
	80	int i;
	81
	82	for_each_cpu(i)
	83	flow_hash_rnd_recalc(i) = 1;
	84
	85	flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
	86	add_timer(&flow_hash_rnd_timer);
	87	}
	88
	89	static void __flow_cache_shrink(int cpu, int shrink_to)
	90	{
	91	struct flow_cache_entry fle, *flp;
	92	int i;
	93
	94	for (i = 0; i < flow_hash_size; i++) {
	95	int k = 0;
	96
	97	flp = &flow_table(cpu)[i];
	98	while ((fle = *flp) != NULL && k < shrink_to) {
	99	k++;
	100	flp = &fle->next;
	101	}
	102	while ((fle = *flp) != NULL) {
	103	*flp = fle->next;
	104	if (fle->object)
	105	atomic_dec(fle->object_ref);
	106	kmem_cache_free(flow_cachep, fle);
	107	flow_count(cpu)--;
	108	}
	109	}
	110	}
	111
	112	static void flow_cache_shrink(int cpu)
113	{
114	int shrink_to = flow_lwm / flow_hash_size;
115
116	__flow_cache_shrink(cpu, shrink_to);
117	}
118
119	static void flow_new_hash_rnd(int cpu)
120	{
121	get_random_bytes(&flow_hash_rnd(cpu), sizeof(u32));
122	flow_hash_rnd_recalc(cpu) = 0;
123
124	__flow_cache_shrink(cpu, 0);
125	}
126
127	static u32 flow_hash_code(struct flowi *key, int cpu)
128	{
129	u32 k = (u32 ) key;
130
131	return (jhash2(k, (sizeof(*key) / sizeof(u32)), flow_hash_rnd(cpu)) &
132	(flow_hash_size - 1));
133	}
134
135	#if (BITS_PER_LONG == 64)
136	typedef u64 flow_compare_t;
137	#else
138	typedef u32 flow_compare_t;
139	#endif
140
141	extern void flowi_is_missized(void);
142
143	/* I hear what you're saying, use memcmp. But memcmp cannot make
144	* important assumptions that we can here, such as alignment and
145	* constant size.
146	*/
147	static int flow_key_compare(struct flowi key1, struct flowi key2)
148	{
149	flow_compare_t k1, k1_lim, *k2;
150	const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
151
152	if (sizeof(struct flowi) % sizeof(flow_compare_t))
153	flowi_is_missized();
154
155	k1 = (flow_compare_t *) key1;
156	k1_lim = k1 + n_elem;
157
158	k2 = (flow_compare_t *) key2;
159
160	do {
161	if (k1++ != k2++)
162	return 1;
163	} while (k1 < k1_lim);
164
165	return 0;
166	}
167
df71837d	168	void flow_cache_lookup(struct flowi key, u32 sk_sid, u16 family, u8 dir,
1da177e4 LT	169	flow_resolve_t resolver)
	170	{
	171	struct flow_cache_entry fle, *head;
	172	unsigned int hash;
	173	int cpu;
	174
	175	local_bh_disable();
	176	cpu = smp_processor_id();
	177
	178	fle = NULL;
	179	/* Packet really early in init? Making flow_cache_init a
	180	* pre-smp initcall would solve this. --RR */
	181	if (!flow_table(cpu))
	182	goto nocache;
	183
	184	if (flow_hash_rnd_recalc(cpu))
	185	flow_new_hash_rnd(cpu);
	186	hash = flow_hash_code(key, cpu);
	187
	188	head = &flow_table(cpu)[hash];
	189	for (fle = *head; fle; fle = fle->next) {
	190	if (fle->family == family &&
	191	fle->dir == dir &&
df71837d	192	fle->sk_sid == sk_sid &&
1da177e4 LT	193	flow_key_compare(key, &fle->key) == 0) {
	194	if (fle->genid == atomic_read(&flow_cache_genid)) {
	195	void *ret = fle->object;
	196
	197	if (ret)
	198	atomic_inc(fle->object_ref);
	199	local_bh_enable();
	200
	201	return ret;
	202	}
	203	break;
	204	}
	205	}
	206
	207	if (!fle) {
	208	if (flow_count(cpu) > flow_hwm)
	209	flow_cache_shrink(cpu);
	210
	211	fle = kmem_cache_alloc(flow_cachep, SLAB_ATOMIC);
	212	if (fle) {
	213	fle->next = *head;
	214	*head = fle;
	215	fle->family = family;
	216	fle->dir = dir;
df71837d	217	fle->sk_sid = sk_sid;
1da177e4 LT	218	memcpy(&fle->key, key, sizeof(*key));
	219	fle->object = NULL;
	220	flow_count(cpu)++;
	221	}
	222	}
	223
	224	nocache:
	225	{
	226	void *obj;
	227	atomic_t *obj_ref;
	228
df71837d	229	resolver(key, sk_sid, family, dir, &obj, &obj_ref);
1da177e4 LT	230
	231	if (fle) {
	232	fle->genid = atomic_read(&flow_cache_genid);
	233
	234	if (fle->object)
	235	atomic_dec(fle->object_ref);
	236
	237	fle->object = obj;
	238	fle->object_ref = obj_ref;
	239	if (obj)
	240	atomic_inc(fle->object_ref);
	241	}
	242	local_bh_enable();
	243
	244	return obj;
	245	}
	246	}
	247
	248	static void flow_cache_flush_tasklet(unsigned long data)
	249	{
	250	struct flow_flush_info info = (void )data;
	251	int i;
	252	int cpu;
	253
	254	cpu = smp_processor_id();
	255	for (i = 0; i < flow_hash_size; i++) {
	256	struct flow_cache_entry *fle;
	257
	258	fle = flow_table(cpu)[i];
	259	for (; fle; fle = fle->next) {
	260	unsigned genid = atomic_read(&flow_cache_genid);
	261
	262	if (!fle->object \|\| fle->genid == genid)
	263	continue;
	264
	265	fle->object = NULL;
	266	atomic_dec(fle->object_ref);
	267	}
	268	}
	269
	270	if (atomic_dec_and_test(&info->cpuleft))
	271	complete(&info->completion);
	272	}
	273
	274	static void flow_cache_flush_per_cpu(void *) __attribute__((__unused__));
	275	static void flow_cache_flush_per_cpu(void *data)
	276	{
	277	struct flow_flush_info *info = data;
	278	int cpu;
	279	struct tasklet_struct *tasklet;
	280
	281	cpu = smp_processor_id();
	282
	283	tasklet = flow_flush_tasklet(cpu);
	284	tasklet->data = (unsigned long)info;
	285	tasklet_schedule(tasklet);
	286	}
	287
	288	void flow_cache_flush(void)
	289	{
	290	struct flow_flush_info info;
4a3e2f71	291	static DEFINE_MUTEX(flow_flush_sem);
1da177e4 LT	292
	293	/* Don't want cpus going down or up during this. */
	294	lock_cpu_hotplug();
4a3e2f71	295	mutex_lock(&flow_flush_sem);
1da177e4 LT	296	atomic_set(&info.cpuleft, num_online_cpus());
	297	init_completion(&info.completion);
	298
	299	local_bh_disable();
	300	smp_call_function(flow_cache_flush_per_cpu, &info, 1, 0);
	301	flow_cache_flush_tasklet((unsigned long)&info);
	302	local_bh_enable();
	303
	304	wait_for_completion(&info.completion);
4a3e2f71	305	mutex_unlock(&flow_flush_sem);
1da177e4 LT	306	unlock_cpu_hotplug();
	307	}
	308
	309	static void __devinit flow_cache_cpu_prepare(int cpu)
	310	{
	311	struct tasklet_struct *tasklet;
	312	unsigned long order;
	313
	314	for (order = 0;
	315	(PAGE_SIZE << order) <
	316	(sizeof(struct flow_cache_entry )flow_hash_size);
	317	order++)
	318	/* NOTHING */;
	319
	320	flow_table(cpu) = (struct flow_cache_entry **)
	321	__get_free_pages(GFP_KERNEL, order);
	322	if (!flow_table(cpu))
	323	panic("NET: failed to allocate flow cache order %lu\n", order);
	324
	325	memset(flow_table(cpu), 0, PAGE_SIZE << order);
	326
	327	flow_hash_rnd_recalc(cpu) = 1;
	328	flow_count(cpu) = 0;
	329
	330	tasklet = flow_flush_tasklet(cpu);
	331	tasklet_init(tasklet, flow_cache_flush_tasklet, 0);
	332	}
	333
	334	#ifdef CONFIG_HOTPLUG_CPU
	335	static int flow_cache_cpu(struct notifier_block *nfb,
	336	unsigned long action,
	337	void *hcpu)
	338	{
	339	if (action == CPU_DEAD)
	340	__flow_cache_shrink((unsigned long)hcpu, 0);
	341	return NOTIFY_OK;
	342	}
	343	#endif /* CONFIG_HOTPLUG_CPU */
	344
	345	static int __init flow_cache_init(void)
	346	{
	347	int i;
	348
	349	flow_cachep = kmem_cache_create("flow_cache",
	350	sizeof(struct flow_cache_entry),
	351	0, SLAB_HWCACHE_ALIGN,
	352	NULL, NULL);
	353
	354	if (!flow_cachep)
	355	panic("NET: failed to allocate flow cache slab\n");
	356
	357	flow_hash_shift = 10;
	358	flow_lwm = 2 * flow_hash_size;
	359	flow_hwm = 4 * flow_hash_size;
	360
	361	init_timer(&flow_hash_rnd_timer);
	362	flow_hash_rnd_timer.function = flow_cache_new_hashrnd;
	363	flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
	364	add_timer(&flow_hash_rnd_timer);
	365
	366	for_each_cpu(i)
	367	flow_cache_cpu_prepare(i);
	368
	369	hotcpu_notifier(flow_cache_cpu, 0);
370	return 0;
371	}
372
373	module_init(flow_cache_init);
374
375	EXPORT_SYMBOL(flow_cache_genid);
376	EXPORT_SYMBOL(flow_cache_lookup);