/*
 * Copyright 2024 Vsevolod Stakhov
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *    http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
#include "config.h"
#include "mem_pool.h"
#include "fstring.h"
#include "logger.h"
#include "ottery.h"
#include "unix-std.h"
#include "khash.h"
#include "cryptobox.h"
#include "contrib/uthash/utlist.h"
#include "mem_pool_internal.h"

#ifdef WITH_JEMALLOC
#include <jemalloc/jemalloc.h>
#if (JEMALLOC_VERSION_MAJOR == 3 && JEMALLOC_VERSION_MINOR >= 6) || (JEMALLOC_VERSION_MAJOR > 3)
#define HAVE_MALLOC_SIZE 1
#define sys_alloc_size(sz) nallocx(sz, 0)
#endif
#elif defined(__APPLE__)
#include <malloc/malloc.h>
#define HAVE_MALLOC_SIZE 1
#define sys_alloc_size(sz) malloc_good_size(sz)
#endif

#ifdef HAVE_SCHED_YIELD
#include <sched.h>
#endif

/* Sleep time for spin lock in nanoseconds */
#define MUTEX_SLEEP_TIME 10000000L
#define MUTEX_SPIN_COUNT 100

#define POOL_MTX_LOCK() \
	do {                \
	} while (0)
#define POOL_MTX_UNLOCK() \
	do {                  \
	} while (0)

/*
 * This define specify whether we should check all pools for free space for new object
 * or just begin scan from current (recently attached) pool
 * If MEMORY_GREEDY is defined, then we scan all pools to find free space (more CPU usage, slower
 * but requires less memory). If it is not defined check only current pool and if object is too large
 * to place in it allocate new one (this may cause huge CPU usage in some cases too, but generally faster than
 * greedy method)
 */
#undef MEMORY_GREEDY


static inline uint32_t
rspamd_entry_hash(const char *str)
{
	return (unsigned int) rspamd_cryptobox_fast_hash(str, strlen(str), rspamd_hash_seed());
}

static inline int
rspamd_entry_equal(const char *k1, const char *k2)
{
	return strcmp(k1, k2) == 0;
}


KHASH_INIT(mempool_entry, const char *, struct rspamd_mempool_entry_point *,
		   1, rspamd_entry_hash, rspamd_entry_equal)

static khash_t(mempool_entry) *mempool_entries = NULL;


/* Internal statistic */
static rspamd_mempool_stat_t *mem_pool_stat = NULL;
/* Environment variable */
static gboolean env_checked = FALSE;
static gboolean always_malloc = FALSE;

/**
 * Function that return free space in pool page
 * @param x pool page struct
 */
static gsize
pool_chain_free(struct _pool_chain *chain)
{
	int64_t occupied = chain->pos - chain->begin + MIN_MEM_ALIGNMENT;

	return (occupied < (int64_t) chain->slice_size ? chain->slice_size - occupied : 0);
}

/* By default allocate 4Kb chunks of memory */
#define FIXED_POOL_SIZE 4096

static inline struct rspamd_mempool_entry_point *
rspamd_mempool_entry_new(const char *loc)
{
	struct rspamd_mempool_entry_point **pentry, *entry;
	int r;
	khiter_t k;

	k = kh_put(mempool_entry, mempool_entries, loc, &r);

	if (r >= 0) {
		pentry = &kh_value(mempool_entries, k);
		entry = g_malloc0(sizeof(*entry));
		*pentry = entry;
		memset(entry, 0, sizeof(*entry));
		rspamd_strlcpy(entry->src, loc, sizeof(entry->src));
#ifdef HAVE_GETPAGESIZE
		entry->cur_suggestion = MAX(getpagesize(), FIXED_POOL_SIZE);
#else
		entry->cur_suggestion = MAX(sysconf(_SC_PAGESIZE), FIXED_POOL_SIZE);
#endif
	}
	else {
		g_assert_not_reached();
	}

	return entry;
}

RSPAMD_CONSTRUCTOR(rspamd_mempool_entries_ctor)
{
	if (mempool_entries == NULL) {
		mempool_entries = kh_init(mempool_entry);
	}
}

RSPAMD_DESTRUCTOR(rspamd_mempool_entries_dtor)
{
	struct rspamd_mempool_entry_point *elt;

	kh_foreach_value(mempool_entries, elt, {
		g_free(elt);
	});

	kh_destroy(mempool_entry, mempool_entries);
	mempool_entries = NULL;
}

static inline struct rspamd_mempool_entry_point *
rspamd_mempool_get_entry(const char *loc)
{
	khiter_t k;
	struct rspamd_mempool_entry_point *elt;

	if (G_UNLIKELY(!mempool_entries)) {
		rspamd_mempool_entries_ctor();
	}

	k = kh_get(mempool_entry, mempool_entries, loc);

	if (k != kh_end(mempool_entries)) {
		elt = kh_value(mempool_entries, k);

		return elt;
	}

	return rspamd_mempool_entry_new(loc);
}

static struct _pool_chain *
rspamd_mempool_chain_new(gsize size, gsize alignment, enum rspamd_mempool_chain_type pool_type)
{
	struct _pool_chain *chain;
	gsize total_size = size + sizeof(struct _pool_chain) + alignment,
		  optimal_size = 0;
	gpointer map;

	g_assert(size > 0);

	if (pool_type == RSPAMD_MEMPOOL_SHARED) {
#if defined(HAVE_MMAP_ANON)
		map = mmap(NULL,
				   total_size,
				   PROT_READ | PROT_WRITE,
				   MAP_ANON | MAP_SHARED,
				   -1,
				   0);
		if (map == MAP_FAILED) {
			g_error("%s: failed to allocate %" G_GSIZE_FORMAT " bytes",
					G_STRLOC, total_size);
			abort();
		}
		chain = map;
		chain->begin = ((uint8_t *) chain) + sizeof(struct _pool_chain);
#elif defined(HAVE_MMAP_ZERO)
		int fd;

		fd = open("/dev/zero", O_RDWR);
		if (fd == -1) {
			return NULL;
		}
		map = mmap(NULL,
				   size + sizeof(struct _pool_chain),
				   PROT_READ | PROT_WRITE,
				   MAP_SHARED,
				   fd,
				   0);
		if (map == MAP_FAILED) {
			msg_err("cannot allocate %z bytes, aborting", size +
															  sizeof(struct _pool_chain));
			abort();
		}
		chain = map;
		chain->begin = ((uint8_t *) chain) + sizeof(struct _pool_chain);
#else
#error No mmap methods are defined
#endif
		g_atomic_int_inc(&mem_pool_stat->shared_chunks_allocated);
		g_atomic_int_add(&mem_pool_stat->bytes_allocated, total_size);
	}
	else {
#ifdef HAVE_MALLOC_SIZE
		optimal_size = sys_alloc_size(total_size);
#endif
		total_size = MAX(total_size, optimal_size);
		int ret = posix_memalign(&map, alignment, total_size);

		if (ret != 0 || map == NULL) {
			g_error("%s: failed to allocate %" G_GSIZE_FORMAT " bytes: %d - %s",
					G_STRLOC, total_size, ret, strerror(errno));
			abort();
		}

		chain = map;
		chain->begin = ((uint8_t *) chain) + sizeof(struct _pool_chain);
		g_atomic_int_add(&mem_pool_stat->bytes_allocated, total_size);
		g_atomic_int_inc(&mem_pool_stat->chunks_allocated);
	}

	chain->pos = align_ptr(chain->begin, alignment);
	chain->slice_size = total_size - sizeof(struct _pool_chain);

	return chain;
}


/**
 * Get the current pool of the specified type, creating the corresponding
 * array if it's absent
 * @param pool
 * @param pool_type
 * @return
 */
static struct _pool_chain *
rspamd_mempool_get_chain(rspamd_mempool_t *pool,
						 enum rspamd_mempool_chain_type pool_type)
{
	g_assert(pool_type >= 0 && pool_type < RSPAMD_MEMPOOL_MAX);

	return pool->priv->pools[pool_type];
}

static void
rspamd_mempool_append_chain(rspamd_mempool_t *pool,
							struct _pool_chain *chain,
							enum rspamd_mempool_chain_type pool_type)
{
	g_assert(pool_type >= 0 && pool_type < RSPAMD_MEMPOOL_MAX);
	g_assert(chain != NULL);

	LL_PREPEND(pool->priv->pools[pool_type], chain);
}

/**
 * Allocate new memory poll
 * @param size size of pool's page
 * @return new memory pool object
 */
rspamd_mempool_t *
rspamd_mempool_new_(gsize size, const char *tag, int flags, const char *loc)
{
	rspamd_mempool_t *new_pool;
	gpointer map;

	/* Allocate statistic structure if it is not allocated before */
	if (mem_pool_stat == NULL) {
#if defined(HAVE_MMAP_ANON)
		map = mmap(NULL,
				   sizeof(rspamd_mempool_stat_t),
				   PROT_READ | PROT_WRITE,
				   MAP_ANON | MAP_SHARED,
				   -1,
				   0);
		if (map == MAP_FAILED) {
			msg_err("cannot allocate %z bytes, aborting",
					sizeof(rspamd_mempool_stat_t));
			abort();
		}
		mem_pool_stat = (rspamd_mempool_stat_t *) map;
#elif defined(HAVE_MMAP_ZERO)
		int fd;

		fd = open("/dev/zero", O_RDWR);
		g_assert(fd != -1);
		map = mmap(NULL,
				   sizeof(rspamd_mempool_stat_t),
				   PROT_READ | PROT_WRITE,
				   MAP_SHARED,
				   fd,
				   0);
		if (map == MAP_FAILED) {
			msg_err("cannot allocate %z bytes, aborting",
					sizeof(rspamd_mempool_stat_t));
			abort();
		}
		mem_pool_stat = (rspamd_mempool_stat_t *) map;
#else
#error No mmap methods are defined
#endif
		memset(map, 0, sizeof(rspamd_mempool_stat_t));
	}

	if (!env_checked) {
		/* Check G_SLICE=always-malloc to allow memory pool debug */
		const char *g_slice;

		g_slice = getenv("VALGRIND");
		if (g_slice != NULL) {
			always_malloc = TRUE;
		}
		env_checked = TRUE;
	}

	struct rspamd_mempool_entry_point *entry = rspamd_mempool_get_entry(loc);
	gsize total_size;

	if (size == 0 && entry) {
		size = entry->cur_suggestion;
	}

	total_size = sizeof(rspamd_mempool_t) +
				 sizeof(struct rspamd_mempool_specific) +
				 MIN_MEM_ALIGNMENT +
				 sizeof(struct _pool_chain) +
				 size;

	if (G_UNLIKELY(flags & RSPAMD_MEMPOOL_DEBUG)) {
		total_size += sizeof(GHashTable *);
	}
	/*
	 * Memory layout:
	 * struct rspamd_mempool_t
	 * <optional debug hash table>
	 * struct rspamd_mempool_specific
	 * struct _pool_chain
	 * alignment (if needed)
	 * memory chunk
	 */
	unsigned char *mem_chunk;
	int ret = posix_memalign((void **) &mem_chunk, MIN_MEM_ALIGNMENT,
							 total_size);
	gsize priv_offset;

	if (ret != 0 || mem_chunk == NULL) {
		g_error("%s: failed to allocate %" G_GSIZE_FORMAT " bytes: %d - %s",
				G_STRLOC, total_size, ret, strerror(errno));
		abort();
	}

	/* Set memory layout */
	new_pool = (rspamd_mempool_t *) mem_chunk;
	if (G_UNLIKELY(flags & RSPAMD_MEMPOOL_DEBUG)) {
		/* Allocate debug table */
		GHashTable *debug_tbl;

		debug_tbl = g_hash_table_new(rspamd_str_hash, rspamd_str_equal);
		memcpy(mem_chunk + sizeof(rspamd_mempool_t), &debug_tbl,
			   sizeof(GHashTable *));
		priv_offset = sizeof(rspamd_mempool_t) + sizeof(GHashTable *);
	}
	else {
		priv_offset = sizeof(rspamd_mempool_t);
	}

	new_pool->priv = (struct rspamd_mempool_specific *) (mem_chunk +
														 priv_offset);
	/* Zero memory for specific and for the first chain */
	memset(new_pool->priv, 0, sizeof(struct rspamd_mempool_specific) + sizeof(struct _pool_chain));

	new_pool->priv->entry = entry;
	new_pool->priv->elt_len = size;
	new_pool->priv->flags = flags;

	if (tag) {
		rspamd_strlcpy(new_pool->tag.tagname, tag, sizeof(new_pool->tag.tagname));
	}
	else {
		new_pool->tag.tagname[0] = '\0';
	}

	/* Generate new uid */
	uint64_t uid = rspamd_random_uint64_fast();
	rspamd_encode_hex_buf((unsigned char *) &uid, sizeof(uid),
						  new_pool->tag.uid, sizeof(new_pool->tag.uid) - 1);
	new_pool->tag.uid[sizeof(new_pool->tag.uid) - 1] = '\0';

	mem_pool_stat->pools_allocated++;

	/* Now we can attach one chunk to speed up simple allocations */
	struct _pool_chain *nchain;

	nchain = (struct _pool_chain *) (mem_chunk +
									 priv_offset +
									 sizeof(struct rspamd_mempool_specific));

	unsigned char *unaligned = mem_chunk +
							   priv_offset +
							   sizeof(struct rspamd_mempool_specific) +
							   sizeof(struct _pool_chain);

	nchain->slice_size = size;
	nchain->begin = unaligned;
	nchain->slice_size = size;
	nchain->pos = align_ptr(unaligned, MIN_MEM_ALIGNMENT);
	new_pool->priv->pools[RSPAMD_MEMPOOL_NORMAL] = nchain;
	new_pool->priv->used_memory = size;

	/* Adjust stats */
	g_atomic_int_add(&mem_pool_stat->bytes_allocated,
					 (int) size);
	g_atomic_int_add(&mem_pool_stat->chunks_allocated, 1);

	return new_pool;
}

static void *
memory_pool_alloc_common(rspamd_mempool_t *pool, gsize size, gsize alignment,
						 enum rspamd_mempool_chain_type pool_type,
						 const char *loc)
	RSPAMD_ATTR_ALLOC_SIZE(2) RSPAMD_ATTR_ALLOC_ALIGN(MIN_MEM_ALIGNMENT) RSPAMD_ATTR_RETURNS_NONNUL;


void rspamd_mempool_notify_alloc_(rspamd_mempool_t *pool, gsize size, const char *loc)
{
	if (pool && G_UNLIKELY(pool->priv->flags & RSPAMD_MEMPOOL_DEBUG)) {
		GHashTable *debug_tbl = *(GHashTable **) (((unsigned char *) pool + sizeof(*pool)));
		gpointer ptr;

		ptr = g_hash_table_lookup(debug_tbl, loc);

		if (ptr) {
			ptr = GSIZE_TO_POINTER(GPOINTER_TO_SIZE(ptr) + size);
		}
		else {
			ptr = GSIZE_TO_POINTER(size);
		}

		g_hash_table_insert(debug_tbl, (gpointer) loc, ptr);
	}
}

static void *
memory_pool_alloc_common(rspamd_mempool_t *pool, gsize size, gsize alignment,
						 enum rspamd_mempool_chain_type pool_type, const char *loc)
{
	uint8_t *tmp;
	struct _pool_chain *new, *cur;
	gsize free = 0;

	if (pool) {
		POOL_MTX_LOCK();
		pool->priv->used_memory += size;

		if (G_UNLIKELY(pool->priv->flags & RSPAMD_MEMPOOL_DEBUG)) {
			rspamd_mempool_notify_alloc_(pool, size, loc);
		}

		if (always_malloc && pool_type != RSPAMD_MEMPOOL_SHARED) {
			void *ptr;

			if (alignment <= G_MEM_ALIGN) {
				ptr = g_malloc(size);
			}
			else {
				ptr = g_malloc(size + alignment);
				ptr = align_ptr(ptr, alignment);
			}
			POOL_MTX_UNLOCK();

			if (pool->priv->trash_stack == NULL) {
				pool->priv->trash_stack = g_ptr_array_sized_new(128);
			}

			g_ptr_array_add(pool->priv->trash_stack, ptr);

			return ptr;
		}

		cur = rspamd_mempool_get_chain(pool, pool_type);

		/* Find free space in pool chain */
		if (cur) {
			free = pool_chain_free(cur);
		}

		if (cur == NULL || free < size + alignment) {
			if (free < size) {
				pool->priv->wasted_memory += free;
			}

			/* Allocate new chain element */
			if (pool->priv->elt_len >= size + alignment) {
				pool->priv->entry->elts[pool->priv->entry->cur_elts].fragmentation += size;
				new = rspamd_mempool_chain_new(pool->priv->elt_len, alignment,
											   pool_type);
			}
			else {
				mem_pool_stat->oversized_chunks++;
				g_atomic_int_add(&mem_pool_stat->fragmented_size,
								 free);
				pool->priv->entry->elts[pool->priv->entry->cur_elts].fragmentation += free;
				new = rspamd_mempool_chain_new(size + pool->priv->elt_len, alignment,
											   pool_type);
			}

			/* Connect to pool subsystem */
			rspamd_mempool_append_chain(pool, new, pool_type);
			/* No need to align again, aligned by rspamd_mempool_chain_new */
			tmp = new->pos;
			new->pos = tmp + size;
			POOL_MTX_UNLOCK();

			return tmp;
		}

		/* No need to allocate page */
		tmp = align_ptr(cur->pos, alignment);
		cur->pos = tmp + size;
		POOL_MTX_UNLOCK();

		return tmp;
	}

	abort();
}


void *
rspamd_mempool_alloc_(rspamd_mempool_t *pool, gsize size, gsize alignment, const char *loc)
{
	return memory_pool_alloc_common(pool, size, alignment, RSPAMD_MEMPOOL_NORMAL, loc);
}

/*
 * This is sqrt(SIZE_MAX+1), as s1*s2 <= SIZE_MAX
 * if both s1 < MUL_NO_OVERFLOW and s2 < MUL_NO_OVERFLOW
 */
#define MUL_NO_OVERFLOW (1UL << (sizeof(gsize) * 4))

void *
rspamd_mempool_alloc_array_(rspamd_mempool_t *pool, gsize nmemb, gsize size, gsize alignment, const char *loc)
{
	if ((nmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
		nmemb > 0 && G_MAXSIZE / nmemb < size) {

		g_error("alloc_array: overflow %" G_GSIZE_FORMAT " * %" G_GSIZE_FORMAT "",
				nmemb, size);
		g_abort();
	}
	return memory_pool_alloc_common(pool, size * nmemb, alignment, RSPAMD_MEMPOOL_NORMAL, loc);
}

void *
rspamd_mempool_alloc0_(rspamd_mempool_t *pool, gsize size, gsize alignment, const char *loc)
{
	void *pointer = rspamd_mempool_alloc_(pool, size, alignment, loc);
	memset(pointer, 0, size);

	return pointer;
}
void *
rspamd_mempool_alloc0_shared_(rspamd_mempool_t *pool, gsize size, gsize alignment, const char *loc)
{
	void *pointer = rspamd_mempool_alloc_shared_(pool, size, alignment, loc);

	memset(pointer, 0, size);
	return pointer;
}

void *
rspamd_mempool_alloc_shared_(rspamd_mempool_t *pool, gsize size, gsize alignment, const char *loc)
{
	return memory_pool_alloc_common(pool, size, alignment, RSPAMD_MEMPOOL_SHARED, loc);
}


char *
rspamd_mempool_strdup_(rspamd_mempool_t *pool, const char *src, const char *loc)
{
	if (src == NULL) {
		return NULL;
	}
	return rspamd_mempool_strdup_len_(pool, src, strlen(src), loc);
}

char *
rspamd_mempool_strdup_len_(rspamd_mempool_t *pool, const char *src, gsize len, const char *loc)
{
	char *newstr;

	if (src == NULL) {
		return NULL;
	}

	newstr = rspamd_mempool_alloc_(pool, len + 1, MIN_MEM_ALIGNMENT, loc);
	memcpy(newstr, src, len);
	newstr[len] = '\0';

	return newstr;
}

char *
rspamd_mempool_ftokdup_(rspamd_mempool_t *pool, const rspamd_ftok_t *src,
						const char *loc)
{
	char *newstr;

	if (src == NULL) {
		return NULL;
	}

	newstr = rspamd_mempool_alloc_(pool, src->len + 1, MIN_MEM_ALIGNMENT, loc);
	memcpy(newstr, src->begin, src->len);
	newstr[src->len] = '\0';

	return newstr;
}

void rspamd_mempool_add_destructor_full(rspamd_mempool_t *pool,
										rspamd_mempool_destruct_t func,
										void *data,
										const char *function,
										const char *line)
{
	struct _pool_destructors *cur;

	POOL_MTX_LOCK();
	cur = rspamd_mempool_alloc_(pool, sizeof(*cur),
								RSPAMD_ALIGNOF(struct _pool_destructors), line);
	cur->func = func;
	cur->data = data;
	cur->function = function;
	cur->loc = line;
	cur->next = NULL;

	if (pool->priv->dtors_tail) {
		pool->priv->dtors_tail->next = cur;
		pool->priv->dtors_tail = cur;
	}
	else {
		pool->priv->dtors_head = cur;
		pool->priv->dtors_tail = cur;
	}

	POOL_MTX_UNLOCK();
}

void rspamd_mempool_replace_destructor(rspamd_mempool_t *pool,
									   rspamd_mempool_destruct_t func,
									   void *old_data,
									   void *new_data)
{
	struct _pool_destructors *tmp;

	LL_FOREACH(pool->priv->dtors_head, tmp)
	{
		if (tmp->func == func && tmp->data == old_data) {
			tmp->func = func;
			tmp->data = new_data;
			break;
		}
	}
}

static int
cmp_int(gconstpointer a, gconstpointer b)
{
	int i1 = *(const int *) a, i2 = *(const int *) b;

	return i1 - i2;
}

static void
rspamd_mempool_adjust_entry(struct rspamd_mempool_entry_point *e)
{
	int sz[G_N_ELEMENTS(e->elts)], sel_pos, sel_neg;
	unsigned int i, jitter;

	for (i = 0; i < G_N_ELEMENTS(sz); i++) {
		sz[i] = e->elts[i].fragmentation - (int) e->elts[i].leftover;
	}

	qsort(sz, G_N_ELEMENTS(sz), sizeof(int), cmp_int);
	jitter = rspamd_random_uint64_fast() % 10;
	/*
	 * Take stochastic quantiles
	 */
	sel_pos = sz[50 + jitter];
	sel_neg = sz[4 + jitter];

	if (-sel_neg > sel_pos) {
		/* We need to reduce current suggestion */
		e->cur_suggestion /= (1 + (((double) -sel_neg) / e->cur_suggestion)) * 1.5;
	}
	else {
		/* We still want to grow */
		e->cur_suggestion *= (1 + (((double) sel_pos) / e->cur_suggestion)) * 1.5;
	}

	/* Some sane limits counting mempool architecture */
	if (e->cur_suggestion < 1024) {
		e->cur_suggestion = 1024;
	}
	else if (e->cur_suggestion > 1024 * 1024 * 10) {
		e->cur_suggestion = 1024 * 1024 * 10;
	}

	memset(e->elts, 0, sizeof(e->elts));
}

static void
rspamd_mempool_variables_cleanup(rspamd_mempool_t *pool)
{
	if (pool->priv->variables) {
		struct rspamd_mempool_variable *var;
		kh_foreach_value_ptr(pool->priv->variables, var, {
			if (var->dtor) {
				var->dtor(var->data);
			}
		});

		if (pool->priv->entry && pool->priv->entry->cur_vars <
									 kh_size(pool->priv->variables)) {
			/*
			 * Increase preallocated size in two cases:
			 * 1) Our previous guess was zero
			 * 2) Our new variables count is not more than twice larger than
			 * previous count
			 * 3) Our variables count is less than some hard limit
			 */
			static const unsigned int max_preallocated_vars = 512;

			unsigned int cur_size = kh_size(pool->priv->variables);
			unsigned int old_guess = pool->priv->entry->cur_vars;
			unsigned int new_guess;

			if (old_guess == 0) {
				new_guess = MIN(cur_size, max_preallocated_vars);
			}
			else {
				if (old_guess * 2 < cur_size) {
					new_guess = MIN(cur_size, max_preallocated_vars);
				}
				else {
					/* Too large step */
					new_guess = MIN(old_guess * 2, max_preallocated_vars);
				}
			}

			pool->priv->entry->cur_vars = new_guess;
		}

		kh_destroy(rspamd_mempool_vars_hash, pool->priv->variables);
		pool->priv->variables = NULL;
	}
}

void rspamd_mempool_destructors_enforce(rspamd_mempool_t *pool)
{
	struct _pool_destructors *destructor;

	POOL_MTX_LOCK();

	LL_FOREACH(pool->priv->dtors_head, destructor)
	{
		/* Avoid calling destructors for NULL pointers */
		if (destructor->data != NULL) {
			destructor->func(destructor->data);
		}
	}

	pool->priv->dtors_head = pool->priv->dtors_tail = NULL;

	rspamd_mempool_variables_cleanup(pool);

	POOL_MTX_UNLOCK();
}

struct mempool_debug_elt {
	gsize sz;
	const char *loc;
};

static int
rspamd_mempool_debug_elt_cmp(const void *a, const void *b)
{
	const struct mempool_debug_elt *e1 = a, *e2 = b;

	/* Inverse order */
	return (int) ((gssize) e2->sz) - ((gssize) e1->sz);
}

void rspamd_mempool_delete(rspamd_mempool_t *pool)
{
	struct _pool_chain *cur, *tmp;
	struct _pool_destructors *destructor;
	gpointer ptr;
	unsigned int i;
	gsize len;

	POOL_MTX_LOCK();

	cur = pool->priv->pools[RSPAMD_MEMPOOL_NORMAL];

	if (G_UNLIKELY(pool->priv->flags & RSPAMD_MEMPOOL_DEBUG)) {
		GHashTable *debug_tbl = *(GHashTable **) (((unsigned char *) pool) + sizeof(*pool));
		/* Show debug info */
		gsize ndtor = 0;
		LL_COUNT(pool->priv->dtors_head, destructor, ndtor);
		msg_info_pool("destructing of the memory pool %p; elt size = %z; "
					  "used memory = %Hz; wasted memory = %Hd; "
					  "vars = %z; destructors = %z",
					  pool,
					  pool->priv->elt_len,
					  pool->priv->used_memory,
					  pool->priv->wasted_memory,
					  pool->priv->variables ? (gsize) kh_size(pool->priv->variables) : (gsize) 0,
					  ndtor);

		GHashTableIter it;
		gpointer k, v;
		GArray *sorted_debug_size = g_array_sized_new(FALSE, FALSE,
													  sizeof(struct mempool_debug_elt),
													  g_hash_table_size(debug_tbl));

		g_hash_table_iter_init(&it, debug_tbl);

		while (g_hash_table_iter_next(&it, &k, &v)) {
			struct mempool_debug_elt e;
			e.loc = (const char *) k;
			e.sz = GPOINTER_TO_SIZE(v);
			g_array_append_val(sorted_debug_size, e);
		}

		g_array_sort(sorted_debug_size, rspamd_mempool_debug_elt_cmp);

		for (unsigned int _i = 0; _i < sorted_debug_size->len; _i++) {
			struct mempool_debug_elt *e;

			e = &g_array_index(sorted_debug_size, struct mempool_debug_elt, _i);
			msg_info_pool("allocated %Hz from %s", e->sz, e->loc);
		}

		g_array_free(sorted_debug_size, TRUE);
		g_hash_table_unref(debug_tbl);
	}

	if (cur && mempool_entries) {
		pool->priv->entry->elts[pool->priv->entry->cur_elts].leftover =
			pool_chain_free(cur);

		pool->priv->entry->cur_elts = (pool->priv->entry->cur_elts + 1) %
									  G_N_ELEMENTS(pool->priv->entry->elts);

		if (pool->priv->entry->cur_elts == 0) {
			rspamd_mempool_adjust_entry(pool->priv->entry);
		}
	}

	/* Call all pool destructors */
	LL_FOREACH(pool->priv->dtors_head, destructor)
	{
		/* Avoid calling destructors for NULL pointers */
		if (destructor->data != NULL) {
			destructor->func(destructor->data);
		}
	}

	rspamd_mempool_variables_cleanup(pool);

	if (pool->priv->trash_stack) {
		for (i = 0; i < pool->priv->trash_stack->len; i++) {
			ptr = g_ptr_array_index(pool->priv->trash_stack, i);
			g_free(ptr);
		}

		g_ptr_array_free(pool->priv->trash_stack, TRUE);
	}

	for (i = 0; i < G_N_ELEMENTS(pool->priv->pools); i++) {
		if (pool->priv->pools[i]) {
			LL_FOREACH_SAFE(pool->priv->pools[i], cur, tmp)
			{
				g_atomic_int_add(&mem_pool_stat->bytes_allocated,
								 -((int) cur->slice_size));
				g_atomic_int_add(&mem_pool_stat->chunks_allocated, -1);

				len = cur->slice_size + sizeof(struct _pool_chain);

				if (i == RSPAMD_MEMPOOL_SHARED) {
					munmap((void *) cur, len);
				}
				else {
					/* The last pool is special, it is a part of the initial chunk */
					if (cur->next != NULL) {
						free(cur); /* Not g_free as we use system allocator */
					}
				}
			}
		}
	}

	g_atomic_int_inc(&mem_pool_stat->pools_freed);
	POOL_MTX_UNLOCK();
	free(pool); /* allocated by posix_memalign */
}

void rspamd_mempool_stat(rspamd_mempool_stat_t *st)
{
	if (mem_pool_stat != NULL) {
		st->pools_allocated = mem_pool_stat->pools_allocated;
		st->pools_freed = mem_pool_stat->pools_freed;
		st->shared_chunks_allocated = mem_pool_stat->shared_chunks_allocated;
		st->bytes_allocated = mem_pool_stat->bytes_allocated;
		st->chunks_allocated = mem_pool_stat->chunks_allocated;
		st->chunks_freed = mem_pool_stat->chunks_freed;
		st->oversized_chunks = mem_pool_stat->oversized_chunks;
	}
}

void rspamd_mempool_stat_reset(void)
{
	if (mem_pool_stat != NULL) {
		memset(mem_pool_stat, 0, sizeof(rspamd_mempool_stat_t));
	}
}

gsize rspamd_mempool_suggest_size_(const char *loc)
{
	return 0;
}

#if !defined(HAVE_PTHREAD_PROCESS_SHARED) || defined(DISABLE_PTHREAD_MUTEX)
/*
 * Own emulation
 */
static inline int
__mutex_spin(rspamd_mempool_mutex_t *mutex)
{
	/* check spin count */
	if (g_atomic_int_dec_and_test(&mutex->spin)) {
		/* This may be deadlock, so check owner of this lock */
		if (mutex->owner == getpid()) {
			/* This mutex was locked by calling process, so it is just double lock and we can easily unlock it */
			g_atomic_int_set(&mutex->spin, MUTEX_SPIN_COUNT);
			return 0;
		}
		else if (kill(mutex->owner, 0) == -1) {
			/* Owner process was not found, so release lock */
			g_atomic_int_set(&mutex->spin, MUTEX_SPIN_COUNT);
			return 0;
		}
		/* Spin again */
		g_atomic_int_set(&mutex->spin, MUTEX_SPIN_COUNT);
	}

#ifdef HAVE_SCHED_YIELD
	(void) sched_yield();
#elif defined(HAVE_NANOSLEEP)
	struct timespec ts;
	ts.tv_sec = 0;
	ts.tv_nsec = MUTEX_SLEEP_TIME;
	/* Spin */
	while (nanosleep(&ts, &ts) == -1 && errno == EINTR)
		;
#else
#error No methods to spin are defined
#endif
	return 1;
}

static void
memory_pool_mutex_spin(rspamd_mempool_mutex_t *mutex)
{
	while (!g_atomic_int_compare_and_exchange(&mutex->lock, 0, 1)) {
		if (!__mutex_spin(mutex)) {
			return;
		}
	}
}

rspamd_mempool_mutex_t *
rspamd_mempool_get_mutex(rspamd_mempool_t *pool)
{
	rspamd_mempool_mutex_t *res;
	if (pool != NULL) {
		res =
			rspamd_mempool_alloc_shared(pool, sizeof(rspamd_mempool_mutex_t));
		res->lock = 0;
		res->owner = 0;
		res->spin = MUTEX_SPIN_COUNT;
		return res;
	}
	return NULL;
}

void rspamd_mempool_lock_mutex(rspamd_mempool_mutex_t *mutex)
{
	memory_pool_mutex_spin(mutex);
	mutex->owner = getpid();
}

void rspamd_mempool_unlock_mutex(rspamd_mempool_mutex_t *mutex)
{
	mutex->owner = 0;
	(void) g_atomic_int_compare_and_exchange(&mutex->lock, 1, 0);
}

rspamd_mempool_rwlock_t *
rspamd_mempool_get_rwlock(rspamd_mempool_t *pool)
{
	rspamd_mempool_rwlock_t *lock;

	lock = rspamd_mempool_alloc_shared(pool, sizeof(rspamd_mempool_rwlock_t));
	lock->__r_lock = rspamd_mempool_get_mutex(pool);
	lock->__w_lock = rspamd_mempool_get_mutex(pool);

	return lock;
}

void rspamd_mempool_rlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	/* Spin on write lock */
	while (g_atomic_int_get(&lock->__w_lock->lock)) {
		if (!__mutex_spin(lock->__w_lock)) {
			break;
		}
	}

	g_atomic_int_inc(&lock->__r_lock->lock);
	lock->__r_lock->owner = getpid();
}

void rspamd_mempool_wlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	/* Spin on write lock first */
	rspamd_mempool_lock_mutex(lock->__w_lock);
	/* Now we have write lock set up */
	/* Wait all readers */
	while (g_atomic_int_get(&lock->__r_lock->lock)) {
		__mutex_spin(lock->__r_lock);
	}
}

void rspamd_mempool_runlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	if (g_atomic_int_get(&lock->__r_lock->lock)) {
		(void) g_atomic_int_dec_and_test(&lock->__r_lock->lock);
	}
}

void rspamd_mempool_wunlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	rspamd_mempool_unlock_mutex(lock->__w_lock);
}
#else

/*
 * Pthread bases shared mutexes
 */
rspamd_mempool_mutex_t *
rspamd_mempool_get_mutex(rspamd_mempool_t *pool)
{
	rspamd_mempool_mutex_t *res;
	pthread_mutexattr_t mattr;

	if (pool != NULL) {
		res =
			rspamd_mempool_alloc_shared(pool, sizeof(rspamd_mempool_mutex_t));

		pthread_mutexattr_init(&mattr);
		pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
		pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST);
		pthread_mutex_init(res, &mattr);
		rspamd_mempool_add_destructor(pool,
									  (rspamd_mempool_destruct_t) pthread_mutex_destroy, res);
		pthread_mutexattr_destroy(&mattr);

		return res;
	}
	return NULL;
}

void rspamd_mempool_lock_mutex(rspamd_mempool_mutex_t *mutex)
{
	pthread_mutex_lock(mutex);
}

void rspamd_mempool_unlock_mutex(rspamd_mempool_mutex_t *mutex)
{
	pthread_mutex_unlock(mutex);
}

rspamd_mempool_rwlock_t *
rspamd_mempool_get_rwlock(rspamd_mempool_t *pool)
{
	rspamd_mempool_rwlock_t *res;
	pthread_rwlockattr_t mattr;

	if (pool != NULL) {
		res =
			rspamd_mempool_alloc_shared(pool, sizeof(rspamd_mempool_rwlock_t));

		pthread_rwlockattr_init(&mattr);
		pthread_rwlockattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
		pthread_rwlock_init(res, &mattr);
		rspamd_mempool_add_destructor(pool,
									  (rspamd_mempool_destruct_t) pthread_rwlock_destroy, res);
		pthread_rwlockattr_destroy(&mattr);

		return res;
	}
	return NULL;
}

void rspamd_mempool_rlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	pthread_rwlock_rdlock(lock);
}

void rspamd_mempool_wlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	pthread_rwlock_wrlock(lock);
}

void rspamd_mempool_runlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	pthread_rwlock_unlock(lock);
}

void rspamd_mempool_wunlock_rwlock(rspamd_mempool_rwlock_t *lock)
{
	pthread_rwlock_unlock(lock);
}
#endif

#define RSPAMD_MEMPOOL_VARS_HASH_SEED 0xb32ad7c55eb2e647ULL
void rspamd_mempool_set_variable(rspamd_mempool_t *pool,
								 const char *name,
								 gpointer value,
								 rspamd_mempool_destruct_t destructor)
{
	if (pool->priv->variables == NULL) {

		pool->priv->variables = kh_init(rspamd_mempool_vars_hash);

		if (pool->priv->entry->cur_vars > 0) {
			/* Preallocate */
			kh_resize(rspamd_mempool_vars_hash,
					  pool->priv->variables,
					  pool->priv->entry->cur_vars);
		}
	}

	int hv = rspamd_cryptobox_fast_hash(name, strlen(name),
										RSPAMD_MEMPOOL_VARS_HASH_SEED);
	khiter_t it;
	int r;

	it = kh_put(rspamd_mempool_vars_hash, pool->priv->variables, hv, &r);

	if (it == kh_end(pool->priv->variables)) {
		g_assert_not_reached();
	}
	else {
		struct rspamd_mempool_variable *pvar;

		if (r == 0) {
			/* Existing entry, maybe need cleanup */
			pvar = &kh_val(pool->priv->variables, it);

			if (pvar->dtor) {
				pvar->dtor(pvar->data);
			}
		}

		pvar = &kh_val(pool->priv->variables, it);
		pvar->data = value;
		pvar->dtor = destructor;
	}
}

gpointer
rspamd_mempool_get_variable(rspamd_mempool_t *pool, const char *name)
{
	if (pool->priv->variables == NULL) {
		return NULL;
	}

	khiter_t it;
	int hv = rspamd_cryptobox_fast_hash(name, strlen(name),
										RSPAMD_MEMPOOL_VARS_HASH_SEED);

	it = kh_get(rspamd_mempool_vars_hash, pool->priv->variables, hv);

	if (it != kh_end(pool->priv->variables)) {
		struct rspamd_mempool_variable *pvar;

		pvar = &kh_val(pool->priv->variables, it);
		return pvar->data;
	}

	return NULL;
}

gpointer
rspamd_mempool_steal_variable(rspamd_mempool_t *pool, const char *name)
{
	if (pool->priv->variables == NULL) {
		return NULL;
	}

	khiter_t it;
	int hv = rspamd_cryptobox_fast_hash(name, strlen(name),
										RSPAMD_MEMPOOL_VARS_HASH_SEED);

	it = kh_get(rspamd_mempool_vars_hash, pool->priv->variables, hv);

	if (it != kh_end(pool->priv->variables)) {
		struct rspamd_mempool_variable *pvar;

		pvar = &kh_val(pool->priv->variables, it);
		kh_del(rspamd_mempool_vars_hash, pool->priv->variables, it);

		return pvar->data;
	}

	return NULL;
}

void rspamd_mempool_remove_variable(rspamd_mempool_t *pool, const char *name)
{
	if (pool->priv->variables != NULL) {
		khiter_t it;
		int hv = rspamd_cryptobox_fast_hash(name, strlen(name),
											RSPAMD_MEMPOOL_VARS_HASH_SEED);

		it = kh_get(rspamd_mempool_vars_hash, pool->priv->variables, hv);

		if (it != kh_end(pool->priv->variables)) {
			struct rspamd_mempool_variable *pvar;

			pvar = &kh_val(pool->priv->variables, it);

			if (pvar->dtor) {
				pvar->dtor(pvar->data);
			}

			kh_del(rspamd_mempool_vars_hash, pool->priv->variables, it);
		}
	}
}

GList *
rspamd_mempool_glist_prepend(rspamd_mempool_t *pool, GList *l, gpointer p)
{
	GList *cell;

	cell = rspamd_mempool_alloc(pool, sizeof(*cell));
	cell->prev = NULL;
	cell->data = p;

	if (l == NULL) {
		cell->next = NULL;
	}
	else {
		cell->next = l;
		l->prev = cell;
	}

	return cell;
}

GList *
rspamd_mempool_glist_append(rspamd_mempool_t *pool, GList *l, gpointer p)
{
	GList *cell, *cur;

	cell = rspamd_mempool_alloc(pool, sizeof(*cell));
	cell->next = NULL;
	cell->data = p;

	if (l) {
		for (cur = l; cur->next != NULL; cur = cur->next) {}
		cur->next = cell;
		cell->prev = cur;
	}
	else {
		l = cell;
		l->prev = NULL;
	}

	return l;
}

gsize rspamd_mempool_get_used_size(rspamd_mempool_t *pool)
{
	return pool->priv->used_memory;
}

gsize rspamd_mempool_get_wasted_size(rspamd_mempool_t *pool)
{
	return pool->priv->wasted_memory;
}