rspamd/src/libmime/message.c

/*
 * 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 "util.h"
#include "rspamd.h"
#include "message.h"
#include "libserver/html/html.h"
#include "images.h"
#include "archives.h"
#include "tokenizers/tokenizers.h"
#include "smtp_parsers.h"
#include "mime_parser.h"
#include "mime_encoding.h"
#include "lang_detection.h"
#include "libutil/multipattern.h"
#include "libserver/mempool_vars_internal.h"

#ifdef WITH_SNOWBALL
#include "libstemmer.h"
#endif

#include <math.h>
#include <unicode/uchar.h>
#include "sodium.h"
#include "libserver/cfg_file_private.h"
#include "lua/lua_common.h"
#include "contrib/uthash/utlist.h"
#include "contrib/t1ha/t1ha.h"
#include "received.h"

#define GTUBE_SYMBOL "GTUBE"

#define SET_PART_RAW(part) ((part)->flags &= ~RSPAMD_MIME_TEXT_PART_FLAG_UTF)
#define SET_PART_UTF(part) ((part)->flags |= RSPAMD_MIME_TEXT_PART_FLAG_UTF)

static const char gtube_pattern_reject[] = "XJS*C4JDBQADN1.NSBN3*2IDNEN*"
										   "GTUBE-STANDARD-ANTI-UBE-TEST-EMAIL*C.34X";
static const char gtube_pattern_add_header[] = "YJS*C4JDBQADN1.NSBN3*2IDNEN*"
											   "GTUBE-STANDARD-ANTI-UBE-TEST-EMAIL*C.34X";
static const char gtube_pattern_rewrite_subject[] = "ZJS*C4JDBQADN1.NSBN3*2IDNEN*"
													"GTUBE-STANDARD-ANTI-UBE-TEST-EMAIL*C.34X";
static const char gtube_pattern_no_action[] = "AJS*C4JDBQADN1.NSBN3*2IDNEN*"
											  "GTUBE-STANDARD-ANTI-UBE-TEST-EMAIL*C.34X";
struct rspamd_multipattern *gtube_matcher = NULL;
static const uint64_t words_hash_seed = 0xdeadbabe;

static void
free_byte_array_callback(void *pointer)
{
	GByteArray *arr = (GByteArray *) pointer;
	g_byte_array_free(arr, TRUE);
}

static void
rspamd_mime_part_extract_words(struct rspamd_task *task,
							   struct rspamd_mime_text_part *part)
{
	rspamd_stat_token_t *w;
	unsigned int i, total_len = 0, short_len = 0;

	if (part->utf_words) {
		rspamd_stem_words(part->utf_words, task->task_pool, part->language,
						  task->lang_det);

		for (i = 0; i < part->utf_words->len; i++) {
			uint64_t h;

			w = &g_array_index(part->utf_words, rspamd_stat_token_t, i);

			if (w->stemmed.len > 0) {
				/*
				 * We use static hash seed if we would want to use that in shingles
				 * computation in future
				 */
				h = rspamd_cryptobox_fast_hash_specific(
					RSPAMD_CRYPTOBOX_HASHFAST_INDEPENDENT,
					w->stemmed.begin, w->stemmed.len, words_hash_seed);
				g_array_append_val(part->normalized_hashes, h);
				total_len += w->stemmed.len;

				if (w->stemmed.len <= 3) {
					short_len++;
				}

				if (w->flags & RSPAMD_STAT_TOKEN_FLAG_TEXT &&
					!(w->flags & RSPAMD_STAT_TOKEN_FLAG_SKIPPED)) {
					part->nwords++;
				}
			}

			if (w->flags & (RSPAMD_STAT_TOKEN_FLAG_BROKEN_UNICODE |
							RSPAMD_STAT_TOKEN_FLAG_NORMALISED |
							RSPAMD_STAT_TOKEN_FLAG_INVISIBLE_SPACES)) {
				task->flags |= RSPAMD_TASK_FLAG_BAD_UNICODE;
			}
		}

		if (part->utf_words->len) {
			double *avg_len_p, *short_len_p;

			avg_len_p = rspamd_mempool_get_variable(task->task_pool,
													RSPAMD_MEMPOOL_AVG_WORDS_LEN);

			if (avg_len_p == NULL) {
				avg_len_p = rspamd_mempool_alloc(task->task_pool,
												 sizeof(double));
				*avg_len_p = total_len;
				rspamd_mempool_set_variable(task->task_pool,
											RSPAMD_MEMPOOL_AVG_WORDS_LEN, avg_len_p, NULL);
			}
			else {
				*avg_len_p += total_len;
			}

			short_len_p = rspamd_mempool_get_variable(task->task_pool,
													  RSPAMD_MEMPOOL_SHORT_WORDS_CNT);

			if (short_len_p == NULL) {
				short_len_p = rspamd_mempool_alloc(task->task_pool,
												   sizeof(double));
				*short_len_p = short_len;
				rspamd_mempool_set_variable(task->task_pool,
											RSPAMD_MEMPOOL_SHORT_WORDS_CNT, avg_len_p, NULL);
			}
			else {
				*short_len_p += short_len;
			}
		}
	}
}

static void
rspamd_mime_part_create_words(struct rspamd_task *task,
							  struct rspamd_mime_text_part *part)
{
	enum rspamd_tokenize_type tok_type;

	if (IS_TEXT_PART_UTF(part)) {

#if U_ICU_VERSION_MAJOR_NUM < 50
		/* Hack to prevent hang with Thai in old libicu */
		const char *p = part->utf_stripped_content->data, *end;
		unsigned int i = 0;
		end = p + part->utf_stripped_content->len;
		int32_t uc, sc;

		tok_type = RSPAMD_TOKENIZE_UTF;

		while (p + i < end) {
			U8_NEXT(p, i, part->utf_stripped_content->len, uc);

			if (((int32_t) uc) < 0) {
				tok_type = RSPAMD_TOKENIZE_RAW;
				break;
			}

			if (u_isalpha(uc)) {
				sc = ublock_getCode(uc);

				if (sc == UBLOCK_THAI) {
					msg_info_task("enable workaround for Thai characters for old libicu");
					tok_type = RSPAMD_TOKENIZE_RAW;
					break;
				}
			}
		}
#else
		tok_type = RSPAMD_TOKENIZE_UTF;
#endif
	}
	else {
		tok_type = RSPAMD_TOKENIZE_RAW;
	}

	part->utf_words = rspamd_tokenize_text(
		part->utf_stripped_content->data,
		part->utf_stripped_content->len,
		&part->utf_stripped_text,
		tok_type, task->cfg,
		part->exceptions,
		NULL,
		NULL,
		task->task_pool);


	if (part->utf_words) {
		part->normalized_hashes = g_array_sized_new(FALSE, FALSE,
													sizeof(uint64_t), part->utf_words->len);
		rspamd_normalize_words(part->utf_words, task->task_pool);
	}
}

static void
rspamd_mime_part_detect_language(struct rspamd_task *task,
								 struct rspamd_mime_text_part *part)
{
	struct rspamd_lang_detector_res *lang;

	if (!IS_TEXT_PART_EMPTY(part) && part->utf_words && part->utf_words->len > 0 &&
		task->lang_det) {
		if (rspamd_language_detector_detect(task, task->lang_det, part)) {
			lang = g_ptr_array_index(part->languages, 0);
			part->language = lang->lang;

			msg_info_task("detected part language: %s", part->language);
		}
		else {
			part->language = "en"; /* Safe fallback */
		}
	}
}

static void
rspamd_strip_newlines_parse(struct rspamd_task *task,
							const char *begin, const char *pe,
							struct rspamd_mime_text_part *part)
{
	const char *p = begin, *c = begin;
	gboolean crlf_added = FALSE, is_utf = IS_TEXT_PART_UTF(part);
	gboolean url_open_bracket = FALSE;
	UChar32 uc;

	enum {
		normal_char,
		seen_cr,
		seen_lf,
	} state = normal_char;

	while (p < pe) {
		if (U8_IS_LEAD(*p) && is_utf) {
			int32_t off = p - begin;
			U8_NEXT(begin, off, pe - begin, uc);

			if (uc != -1) {
				while (p < pe && off < (pe - begin)) {
					if (IS_ZERO_WIDTH_SPACE(uc)) {
						/* Invisible space ! */
						task->flags |= RSPAMD_TASK_FLAG_BAD_UNICODE;
						part->spaces++;

						if (p > c) {
							g_byte_array_append(part->utf_stripped_content,
												(const uint8_t *) c, p - c);
							c = begin + off;
							p = c;
						}

						U8_NEXT(begin, off, pe - begin, uc);

						if (!IS_ZERO_WIDTH_SPACE(uc)) {
							break;
						}

						part->double_spaces++;
						p = begin + off;
						c = p;
					}
					else {
						break;
					}
				}
			}
		}

		if (G_UNLIKELY(p >= pe)) {
			/*
			 * This is reached when there is a utf8 part and we
			 * have zero width spaces at the end of the text
			 * So we just check overflow and refuse to access *p if it is
			 * after our real content.
			 */
			break;
		}
		else if (*p == '\r') {
			switch (state) {
			case normal_char:
				state = seen_cr;
				if (p > c) {
					g_byte_array_append(part->utf_stripped_content,
										(const uint8_t *) c, p - c);
				}

				crlf_added = FALSE;
				c = p + 1;
				break;
			case seen_cr:
				/* Double \r\r */
				if (!crlf_added) {
					g_byte_array_append(part->utf_stripped_content,
										(const uint8_t *) " ", 1);
					crlf_added = TRUE;
					g_ptr_array_add(part->newlines,
									(((gpointer) (goffset) (part->utf_stripped_content->len))));
				}

				part->nlines++;
				part->empty_lines++;
				c = p + 1;
				break;
			case seen_lf:
				/* Likely \r\n\r...*/
				state = seen_cr;
				c = p + 1;
				break;
			}

			url_open_bracket = FALSE;

			p++;
		}
		else if (*p == '\n') {
			switch (state) {
			case normal_char:
				state = seen_lf;

				if (p > c) {
					g_byte_array_append(part->utf_stripped_content,
										(const uint8_t *) c, p - c);
				}

				c = p + 1;

				if (IS_TEXT_PART_HTML(part) || !url_open_bracket) {
					g_byte_array_append(part->utf_stripped_content,
										(const uint8_t *) " ", 1);
					g_ptr_array_add(part->newlines,
									(((gpointer) (goffset) (part->utf_stripped_content->len))));
					crlf_added = TRUE;
				}
				else {
					crlf_added = FALSE;
				}

				break;
			case seen_cr:
				/* \r\n */
				if (!crlf_added) {
					if (IS_TEXT_PART_HTML(part) || !url_open_bracket) {
						g_byte_array_append(part->utf_stripped_content,
											(const uint8_t *) " ", 1);
						crlf_added = TRUE;
					}

					g_ptr_array_add(part->newlines,
									(((gpointer) (goffset) (part->utf_stripped_content->len))));
				}

				c = p + 1;
				state = seen_lf;

				break;
			case seen_lf:
				/* Double \n\n */
				if (!crlf_added) {
					g_byte_array_append(part->utf_stripped_content,
										(const uint8_t *) " ", 1);
					crlf_added = TRUE;
					g_ptr_array_add(part->newlines,
									(((gpointer) (goffset) (part->utf_stripped_content->len))));
				}

				part->nlines++;
				part->empty_lines++;

				c = p + 1;
				break;
			}
			url_open_bracket = FALSE;

			p++;
		}
		else {
			if ((*p) == '<') {
				url_open_bracket = TRUE;
			}
			else if ((*p) == '>') {
				url_open_bracket = FALSE;
			}

			switch (state) {
			case normal_char:
				if (*p == ' ') {
					part->spaces++;

					if (p > begin && *(p - 1) == ' ') {
						part->double_spaces++;
					}
				}
				else {
					part->non_spaces++;

					if ((*p) & 0x80) {
						part->non_ascii_chars++;
					}
					else {
						if (g_ascii_isupper(*p)) {
							part->capital_letters++;
						}
						else if (g_ascii_isdigit(*p)) {
							part->numeric_characters++;
						}

						part->ascii_chars++;
					}
				}
				break;
			case seen_cr:
			case seen_lf:
				part->nlines++;

				if (!crlf_added) {
					g_ptr_array_add(part->newlines,
									(((gpointer) (goffset) (part->utf_stripped_content->len))));
				}

				/* Skip initial spaces */
				if (*p == ' ') {
					if (!crlf_added) {
						g_byte_array_append(part->utf_stripped_content,
											(const uint8_t *) " ", 1);
					}

					while (p < pe && *p == ' ') {
						p++;
						c++;
						part->spaces++;
					}

					if (p < pe && (*p == '\r' || *p == '\n')) {
						part->empty_lines++;
					}
				}

				state = normal_char;
				continue;
			}

			p++;
		}
	}

	/* Leftover */
	if (p > c) {
		if (p > pe) {
			p = pe;
		}

		switch (state) {
		case normal_char:
			g_byte_array_append(part->utf_stripped_content,
								(const uint8_t *) c, p - c);

			while (c < p) {
				if (*c == ' ') {
					part->spaces++;

					if (c > begin && *(c - 1) == ' ') {
						part->double_spaces++;
					}
				}
				else {
					part->non_spaces++;

					if ((*c) & 0x80) {
						part->non_ascii_chars++;
					}
					else {
						part->ascii_chars++;
					}
				}

				c++;
			}
			break;
		default:

			if (!crlf_added) {
				g_byte_array_append(part->utf_stripped_content,
									(const uint8_t *) " ", 1);
				g_ptr_array_add(part->newlines,
								(((gpointer) (goffset) (part->utf_stripped_content->len))));
			}

			part->nlines++;
			break;
		}
	}
}

static void
rspamd_u_text_dtor(void *p)
{
	utext_close((UText *) p);
}

static void
rspamd_normalize_text_part(struct rspamd_task *task,
						   struct rspamd_mime_text_part *part)
{
	const char *p, *end;
	unsigned int i;
	goffset off;
	struct rspamd_process_exception *ex;
	UErrorCode uc_err = U_ZERO_ERROR;

	part->newlines = g_ptr_array_sized_new(128);

	if (IS_TEXT_PART_EMPTY(part)) {
		part->utf_stripped_content = g_byte_array_new();
	}
	else {
		part->utf_stripped_content = g_byte_array_sized_new(part->utf_content.len);

		p = (const char *) part->utf_content.begin;
		end = p + part->utf_content.len;

		rspamd_strip_newlines_parse(task, p, end, part);

		for (i = 0; i < part->newlines->len; i++) {
			ex = rspamd_mempool_alloc(task->task_pool, sizeof(*ex));
			off = (goffset) g_ptr_array_index(part->newlines, i);
			g_ptr_array_index(part->newlines, i) = (gpointer) (goffset) (part->utf_stripped_content->data + off);
			ex->pos = off;
			ex->len = 0;
			ex->type = RSPAMD_EXCEPTION_NEWLINE;
			part->exceptions = g_list_prepend(part->exceptions, ex);
		}
	}

	if (IS_TEXT_PART_UTF(part)) {
		utext_openUTF8(&part->utf_stripped_text,
					   part->utf_stripped_content->data,
					   part->utf_stripped_content->len,
					   &uc_err);

		if (!U_SUCCESS(uc_err)) {
			msg_warn_task("cannot open text from utf content");
			/* Probably, should be an assertion */
		}
		else {
			rspamd_mempool_add_destructor(task->task_pool,
										  rspamd_u_text_dtor,
										  &part->utf_stripped_text);
		}
	}

	rspamd_mempool_add_destructor(task->task_pool,
								  (rspamd_mempool_destruct_t) free_byte_array_callback,
								  part->utf_stripped_content);
	rspamd_mempool_notify_alloc(task->task_pool,
								part->utf_stripped_content->len);
	rspamd_mempool_add_destructor(task->task_pool,
								  (rspamd_mempool_destruct_t) rspamd_ptr_array_free_hard,
								  part->newlines);
}

#define MIN3(a, b, c) ((a) < (b) ? ((a) < (c) ? (a) : (c)) : ((b) < (c) ? (b) : (c)))

static unsigned int
rspamd_words_levenshtein_distance(struct rspamd_task *task,
								  GArray *w1, GArray *w2)
{
	unsigned int s1len, s2len, x, y, lastdiag, olddiag;
	unsigned int *column, ret;
	uint64_t h1, h2;
	int eq;
	static const unsigned int max_words = 8192;

	s1len = w1->len;
	s2len = w2->len;

	if (s1len + s2len > max_words) {
		msg_info_task("cannot direct compare multipart/alternative parts with more than %ud words in total: "
					  "(%ud words in one part and %ud in another)",
					  max_words, s1len, s2len);

		/* Use approximate comparison of number of words */
		if (s1len > s2len) {
			return s1len - s2len;
		}
		else {
			return s2len - s1len;
		}
	}

	column = g_malloc0((s1len + 1) * sizeof(unsigned int));

	for (y = 1; y <= s1len; y++) {
		column[y] = y;
	}

	for (x = 1; x <= s2len; x++) {
		column[0] = x;

		for (y = 1, lastdiag = x - 1; y <= s1len; y++) {
			olddiag = column[y];
			h1 = g_array_index(w1, uint64_t, y - 1);
			h2 = g_array_index(w2, uint64_t, x - 1);
			eq = (h1 == h2) ? 1 : 0;
			/*
			 * Cost of replacement is twice higher than cost of add/delete
			 * to calculate percentage properly
			 */
			column[y] = MIN3(column[y] + 1, column[y - 1] + 1,
							 lastdiag + (eq * 2));
			lastdiag = olddiag;
		}
	}

	ret = column[s1len];
	g_free(column);

	return ret;
}

static int
rspamd_multipattern_gtube_cb(struct rspamd_multipattern *mp,
							 unsigned int strnum,
							 int match_start,
							 int match_pos,
							 const char *text,
							 gsize len,
							 void *context)
{
	struct rspamd_task *task = (struct rspamd_task *) context;

	if (strnum > 0) {
		if (task->cfg->gtube_patterns_policy == RSPAMD_GTUBE_ALL) {
			return strnum + 1;
		}

		return 0;
	}

	return strnum + 1; /* To distinguish from zero */
}

static enum rspamd_action_type
rspamd_check_gtube(struct rspamd_task *task, struct rspamd_mime_text_part *part)
{
	static const gsize max_check_size = 8 * 1024;
	int ret;
	enum rspamd_action_type act = METRIC_ACTION_NOACTION;
	enum rspamd_gtube_patterns_policy policy = task->cfg ? task->cfg->gtube_patterns_policy : RSPAMD_GTUBE_REJECT;
	g_assert(part != NULL);

	if (gtube_matcher == NULL && policy != RSPAMD_GTUBE_DISABLED) {
		gtube_matcher = rspamd_multipattern_create(RSPAMD_MULTIPATTERN_DEFAULT);

		rspamd_multipattern_add_pattern(gtube_matcher,
										gtube_pattern_reject,
										RSPAMD_MULTIPATTERN_DEFAULT);
		rspamd_multipattern_add_pattern(gtube_matcher,
										gtube_pattern_add_header,
										RSPAMD_MULTIPATTERN_DEFAULT);
		rspamd_multipattern_add_pattern(gtube_matcher,
										gtube_pattern_rewrite_subject,
										RSPAMD_MULTIPATTERN_DEFAULT);
		rspamd_multipattern_add_pattern(gtube_matcher,
										gtube_pattern_no_action,
										RSPAMD_MULTIPATTERN_DEFAULT);

		GError *err = NULL;
		rspamd_multipattern_compile(gtube_matcher, RSPAMD_MULTIPATTERN_COMPILE_NO_FS, &err);

		if (err != NULL) {
			/* It will be expensive, but I don't care, still better than to abort */
			msg_err("cannot compile gtube matcher: %s", err->message);
			g_error_free(err);
		}
	}

	if (part->utf_content.len >= sizeof(gtube_pattern_reject) &&
		part->utf_content.len <= max_check_size &&
		policy != RSPAMD_GTUBE_DISABLED) {
		if ((ret = rspamd_multipattern_lookup(gtube_matcher, part->utf_content.begin,
											  part->utf_content.len,
											  rspamd_multipattern_gtube_cb, task, NULL)) > 0) {

			switch (ret) {
			case 1:
				act = METRIC_ACTION_REJECT;
				break;
			case 2:
				act = METRIC_ACTION_ADD_HEADER;
				break;
			case 3:
				act = METRIC_ACTION_REWRITE_SUBJECT;
				break;
			case 4:
				act = METRIC_ACTION_NOACTION;
				break;
			}

			if (ret != 0) {
				task->flags |= RSPAMD_TASK_FLAG_SKIP;
				task->flags |= RSPAMD_TASK_FLAG_GTUBE;
				msg_info_task(
					"gtube %s pattern has been found in part of length %uz",
					rspamd_action_to_str(act),
					part->utf_content.len);
			}
		}
	}

	return act;
}

static int
exceptions_compare_func(gconstpointer a, gconstpointer b)
{
	const struct rspamd_process_exception *ea = a, *eb = b;

	return ea->pos - eb->pos;
}

static gboolean
rspamd_message_process_plain_text_part(struct rspamd_task *task,
									   struct rspamd_mime_text_part *text_part)
{
	if (text_part->parsed.len == 0) {
		text_part->flags |= RSPAMD_MIME_TEXT_PART_FLAG_EMPTY;

		return TRUE;
	}

	rspamd_mime_text_part_maybe_convert(task, text_part);

	if (text_part->utf_raw_content != NULL) {
		/* Just have the same content */
		text_part->utf_content.begin = (const char *) text_part->utf_raw_content->data;
		text_part->utf_content.len = text_part->utf_raw_content->len;
	}
	else {
		/*
		 * We ignore unconverted parts from now as it is dangerous
		 * to treat them as text parts
		 */
		text_part->utf_content.begin = NULL;
		text_part->utf_content.len = 0;

		return FALSE;
	}

	return TRUE;
}

static gboolean
rspamd_message_process_html_text_part(struct rspamd_task *task,
									  struct rspamd_mime_text_part *text_part,
									  uint16_t *cur_url_order)
{
	text_part->flags |= RSPAMD_MIME_TEXT_PART_FLAG_HTML;

	if (text_part->parsed.len == 0) {
		text_part->flags |= RSPAMD_MIME_TEXT_PART_FLAG_EMPTY;

		return TRUE;
	}

	rspamd_mime_text_part_maybe_convert(task, text_part);

	if (text_part->utf_raw_content == NULL) {
		return FALSE;
	}


	text_part->html = rspamd_html_process_part_full(
		task,
		text_part->utf_raw_content,
		&text_part->exceptions,
		MESSAGE_FIELD(task, urls),
		text_part->mime_part->urls,
		task->cfg ? task->cfg->enable_css_parser : true,
		cur_url_order);
	rspamd_html_get_parsed_content(text_part->html, &text_part->utf_content);

	if (text_part->utf_content.len == 0) {
		text_part->flags |= RSPAMD_MIME_TEXT_PART_FLAG_EMPTY;
	}

	return TRUE;
}

enum rspamd_message_part_is_text_result {
	RSPAMD_MESSAGE_PART_IS_TEXT_PLAIN = 0,
	RSPAMD_MESSAGE_PART_IS_TEXT_HTML,
	RSPAMD_MESSAGE_PART_IS_NOT_TEXT
};

static enum rspamd_message_part_is_text_result
rspamd_message_part_can_be_parsed_as_text(struct rspamd_task *task,
										  struct rspamd_mime_part *mime_part)
{
	enum rspamd_message_part_is_text_result res = RSPAMD_MESSAGE_PART_IS_NOT_TEXT;

	if ((mime_part->ct && (mime_part->ct->flags & RSPAMD_CONTENT_TYPE_TEXT)) ||
		(mime_part->detected_type && strcmp(mime_part->detected_type, "text") == 0)) {

		res = RSPAMD_MESSAGE_PART_IS_TEXT_PLAIN;
		rspamd_ftok_t html_tok, xhtml_tok;

		html_tok.begin = "html";
		html_tok.len = 4;
		xhtml_tok.begin = "xhtml";
		xhtml_tok.len = 5;

		if (rspamd_ftok_casecmp(&mime_part->ct->subtype, &html_tok) == 0 ||
			rspamd_ftok_casecmp(&mime_part->ct->subtype, &xhtml_tok) == 0 ||
			(mime_part->detected_ext &&
			 strcmp(mime_part->detected_ext, "html") == 0)) {
			res = RSPAMD_MESSAGE_PART_IS_TEXT_HTML;
		}
	}

	/* Skip attachments */
	if (res != RSPAMD_MESSAGE_PART_IS_NOT_TEXT &&
		(mime_part->cd && mime_part->cd->type == RSPAMD_CT_ATTACHMENT)) {
		if (!task->cfg->check_text_attachements) {
			debug_task("skip attachments for checking as text parts");
			return RSPAMD_MESSAGE_PART_IS_NOT_TEXT;
		}
	}

	return res;
}

static gboolean
rspamd_message_process_text_part_maybe(struct rspamd_task *task,
									   struct rspamd_mime_part *mime_part,
									   enum rspamd_message_part_is_text_result is_text,
									   uint16_t *cur_url_order)
{
	struct rspamd_mime_text_part *text_part;
	unsigned int flags = 0;
	enum rspamd_action_type act;

	/* Skip attachments */
	if ((mime_part->cd && mime_part->cd->type == RSPAMD_CT_ATTACHMENT)) {
		flags |= RSPAMD_MIME_TEXT_PART_ATTACHMENT;
	}

	text_part = rspamd_mempool_alloc0(task->task_pool,
									  sizeof(struct rspamd_mime_text_part));
	text_part->mime_part = mime_part;
	text_part->raw.begin = mime_part->raw_data.begin;
	text_part->raw.len = mime_part->raw_data.len;
	text_part->parsed.begin = mime_part->parsed_data.begin;
	text_part->parsed.len = mime_part->parsed_data.len;
	text_part->utf_stripped_text = (UText) UTEXT_INITIALIZER;
	text_part->flags |= flags;

	if (is_text == RSPAMD_MESSAGE_PART_IS_TEXT_HTML) {
		if (!rspamd_message_process_html_text_part(task, text_part, cur_url_order)) {
			return FALSE;
		}
	}
	else {
		if (!rspamd_message_process_plain_text_part(task, text_part)) {
			return FALSE;
		}
	}

	g_ptr_array_add(MESSAGE_FIELD(task, text_parts), text_part);
	mime_part->part_type = RSPAMD_MIME_PART_TEXT;
	mime_part->specific.txt = text_part;

	act = rspamd_check_gtube(task, text_part);
	if (act != METRIC_ACTION_NOACTION) {
		struct rspamd_action *action;
		double score = NAN;

		action = rspamd_config_get_action_by_type(task->cfg, act);

		if (action) {
			score = action->threshold;

			rspamd_add_passthrough_result(task, action,
										  RSPAMD_PASSTHROUGH_CRITICAL,
										  score, "Gtube pattern",
										  "GTUBE", 0, NULL);
		}

		rspamd_task_insert_result(task, GTUBE_SYMBOL, 0, NULL);

		return TRUE;
	}

	/* Post process part */
	rspamd_normalize_text_part(task, text_part);

	if (!IS_TEXT_PART_HTML(text_part)) {
		if (mime_part->parent_part) {
			struct rspamd_mime_part *parent = mime_part->parent_part;

			if (IS_PART_MULTIPART(parent) && parent->specific.mp->children->len == 2) {
				/*
				 * Use strict extraction mode: we will extract missing urls from
				 * an html part if needed
				 */
				rspamd_url_text_extract(task->task_pool, task, text_part, cur_url_order,
										RSPAMD_URL_FIND_STRICT);
			}
			else {
				/*
				 * Fall back to full text extraction using TLD patterns
				 */
				rspamd_url_text_extract(task->task_pool, task, text_part, cur_url_order,
										RSPAMD_URL_FIND_ALL);
			}
		}
		else {
			/*
			 * Fall back to full text extraction using TLD patterns
			*/
			rspamd_url_text_extract(task->task_pool, task, text_part, cur_url_order,
									RSPAMD_URL_FIND_ALL);
		}
	}
	else {
		rspamd_url_text_extract(task->task_pool, task, text_part, cur_url_order,
								RSPAMD_URL_FIND_STRICT);
	}

	if (text_part->exceptions) {
		text_part->exceptions = g_list_sort(text_part->exceptions,
											exceptions_compare_func);
		rspamd_mempool_add_destructor(task->task_pool,
									  (rspamd_mempool_destruct_t) g_list_free,
									  text_part->exceptions);
	}

	rspamd_mime_part_create_words(task, text_part);

	return TRUE;
}

/* Creates message from various data using libmagic to detect type */
static void
rspamd_message_from_data(struct rspamd_task *task, const unsigned char *start,
						 gsize len)
{
	struct rspamd_content_type *ct = NULL;
	struct rspamd_mime_part *part;
	const char *mb = "application/octet-stream";
	char *mid;
	rspamd_ftok_t srch, *tok;
	char cdbuf[1024];

	g_assert(start != NULL);

	part = rspamd_mempool_alloc0(task->task_pool, sizeof(*part));

	part->raw_data.begin = start;
	part->raw_data.len = len;
	part->parsed_data.begin = start;
	part->parsed_data.len = len;
	part->part_number = MESSAGE_FIELD(task, parts)->len;
	part->urls = g_ptr_array_new();
	part->raw_headers = rspamd_message_headers_new();
	part->headers_order = NULL;

	tok = rspamd_task_get_request_header(task, "Content-Type");

	if (tok) {
		/* We have Content-Type defined */
		ct = rspamd_content_type_parse(tok->begin, tok->len,
									   task->task_pool);
		part->ct = ct;
	}
	else if (task->cfg && task->cfg->libs_ctx) {
		lua_State *L = task->cfg->lua_state;

		if (rspamd_lua_require_function(L,
										"lua_magic", "detect_mime_part")) {

			struct rspamd_mime_part **pmime;
			struct rspamd_task **ptask;

			pmime = lua_newuserdata(L, sizeof(struct rspamd_mime_part *));
			rspamd_lua_setclass(L, rspamd_mimepart_classname, -1);
			*pmime = part;
			ptask = lua_newuserdata(L, sizeof(struct rspamd_task *));
			rspamd_lua_setclass(L, rspamd_task_classname, -1);
			*ptask = task;

			if (lua_pcall(L, 2, 2, 0) != 0) {
				msg_err_task("cannot detect type: %s", lua_tostring(L, -1));
			}
			else {
				if (lua_istable(L, -1)) {
					lua_pushstring(L, "ct");
					lua_gettable(L, -2);

					if (lua_isstring(L, -1)) {
						mb = rspamd_mempool_strdup(task->task_pool,
												   lua_tostring(L, -1));
					}
				}
			}

			lua_settop(L, 0);
		}
		else {
			msg_err_task("cannot require lua_magic.detect_mime_part");
		}

		if (mb) {
			srch.begin = mb;
			srch.len = strlen(mb);
			ct = rspamd_content_type_parse(srch.begin, srch.len,
										   task->task_pool);

			if (!part->ct) {
				msg_info_task("construct fake mime of type: %s", mb);
				part->ct = ct;
			}
			else {
				/* Check sanity */
				if (part->ct && (part->ct->flags & RSPAMD_CONTENT_TYPE_TEXT)) {
					RSPAMD_FTOK_FROM_STR(&srch, "application");

					if (rspamd_ftok_cmp(&ct->type, &srch) == 0) {
						msg_info_task("construct fake mime of type: %s", mb);
						part->ct = ct;
					}
				}
				else {
					msg_info_task("construct fake mime of type: %T/%T, detected %s",
								  &part->ct->type, &part->ct->subtype, mb);
				}
			}

			part->detected_ct = ct;
		}
	}


	tok = rspamd_task_get_request_header(task, "Filename");

	if (tok) {
		rspamd_snprintf(cdbuf, sizeof(cdbuf), "inline; filename=\"%T\"", tok);
	}
	else {
		rspamd_snprintf(cdbuf, sizeof(cdbuf), "inline");
	}

	part->cd = rspamd_content_disposition_parse(cdbuf, strlen(cdbuf),
												task->task_pool);

	g_ptr_array_add(MESSAGE_FIELD(task, parts), part);
	rspamd_mime_parser_calc_digest(part);

	/* Generate message ID */
	mid = rspamd_mime_message_id_generate("localhost.localdomain");
	rspamd_mempool_add_destructor(task->task_pool,
								  (rspamd_mempool_destruct_t) g_free, mid);
	MESSAGE_FIELD(task, message_id) = mid;
	task->queue_id = mid;
}

static void
rspamd_message_dtor(struct rspamd_message *msg)
{
	unsigned int i;
	struct rspamd_mime_part *p;
	struct rspamd_mime_text_part *tp;


	PTR_ARRAY_FOREACH(msg->parts, i, p)
	{
		if (p->raw_headers) {
			rspamd_message_headers_unref(p->raw_headers);
		}

		if (IS_PART_MULTIPART(p)) {
			if (p->specific.mp->children) {
				g_ptr_array_free(p->specific.mp->children, TRUE);
			}
		}

		if (p->part_type == RSPAMD_MIME_PART_CUSTOM_LUA &&
			p->specific.lua_specific.cbref != -1) {
			luaL_unref(msg->task->cfg->lua_state,
					   LUA_REGISTRYINDEX,
					   p->specific.lua_specific.cbref);
		}

		if (p->urls) {
			g_ptr_array_unref(p->urls);
		}
	}

	PTR_ARRAY_FOREACH(msg->text_parts, i, tp)
	{
		if (tp->utf_words) {
			g_array_free(tp->utf_words, TRUE);
		}
		if (tp->normalized_hashes) {
			g_array_free(tp->normalized_hashes, TRUE);
		}
		if (tp->languages) {
			g_ptr_array_unref(tp->languages);
		}
	}

	rspamd_message_headers_unref(msg->raw_headers);

	g_ptr_array_unref(msg->text_parts);
	g_ptr_array_unref(msg->parts);

	kh_destroy(rspamd_url_hash, msg->urls);
}

struct rspamd_message *
rspamd_message_new(struct rspamd_task *task)
{
	struct rspamd_message *msg;

	msg = rspamd_mempool_alloc0(task->task_pool, sizeof(*msg));

	msg->raw_headers = rspamd_message_headers_new();
	msg->urls = kh_init(rspamd_url_hash);
	msg->parts = g_ptr_array_sized_new(4);
	msg->text_parts = g_ptr_array_sized_new(2);
	msg->task = task;

	REF_INIT_RETAIN(msg, rspamd_message_dtor);

	return msg;
}

gboolean
rspamd_message_parse(struct rspamd_task *task)
{
	const char *p;
	gsize len;
	unsigned int i;
	GError *err = NULL;
	uint64_t n[2], seed;

	if (RSPAMD_TASK_IS_EMPTY(task)) {
		/* Don't do anything with empty task */
		task->flags |= RSPAMD_TASK_FLAG_SKIP_PROCESS;
		return TRUE;
	}

	p = task->msg.begin;
	len = task->msg.len;

	/* Skip any space characters to avoid some bad messages to be unparsed */
	while (len > 0 && g_ascii_isspace(*p)) {
		p++;
		len--;
	}

	/*
	 * Exim somehow uses mailbox format for messages being scanned:
	 * From xxx@xxx.com Fri May 13 19:08:48 2016
	 *
	 * So we check if a task has this line to avoid possible issues
	 */
	if (len > sizeof("From ") - 1) {
		if (memcmp(p, "From ", sizeof("From ") - 1) == 0) {
			/* Skip to CRLF */
			msg_info_task("mailbox input detected, enable workaround");
			p += sizeof("From ") - 1;
			len -= sizeof("From ") - 1;

			while (len > 0 && *p != '\n') {
				p++;
				len--;
			}
			while (len > 0 && g_ascii_isspace(*p)) {
				p++;
				len--;
			}
		}
	}

	task->msg.begin = p;
	task->msg.len = len;

	/* Cleanup old message */
	if (task->message) {
		rspamd_message_unref(task->message);
	}

	task->message = rspamd_message_new(task);

	if (task->flags & RSPAMD_TASK_FLAG_MIME) {
		enum rspamd_mime_parse_error ret;

		debug_task("construct mime parser from string length %d",
				   (int) task->msg.len);
		ret = rspamd_mime_parse_task(task, &err);

		switch (ret) {
		case RSPAMD_MIME_PARSE_FATAL:
			msg_err_task("cannot construct mime from stream: %e", err);

			if (task->cfg && (!task->cfg->allow_raw_input)) {
				msg_err_task("cannot construct mime from stream");
				if (err) {
					task->err = err;
				}

				return FALSE;
			}
			else {
				task->flags &= ~RSPAMD_TASK_FLAG_MIME;
				rspamd_message_from_data(task, p, len);
			}
			break;
		case RSPAMD_MIME_PARSE_NESTING:
			msg_warn_task("cannot construct full mime from stream: %e", err);
			task->flags |= RSPAMD_TASK_FLAG_BROKEN_HEADERS;
			break;
		case RSPAMD_MIME_PARSE_OK:
		default:
			break;
		}

		if (err) {
			g_error_free(err);
		}
	}
	else {
		rspamd_message_from_data(task, p, len);
	}


	if (MESSAGE_FIELD(task, message_id) == NULL) {
		MESSAGE_FIELD(task, message_id) = "undef";
	}

	debug_task("found %ud parts in message", MESSAGE_FIELD(task, parts)->len);
	if (task->queue_id == NULL) {
		task->queue_id = "undef";
	}

	rspamd_received_maybe_fix_task(task);

	struct rspamd_mime_part *part;

	/* Blake2b applied to string 'rspamd' */
	static const unsigned char RSPAMD_ALIGNED(32) hash_key[] = {
		0xef,
		0x43,
		0xae,
		0x80,
		0xcc,
		0x8d,
		0xc3,
		0x4c,
		0x6f,
		0x1b,
		0xd6,
		0x18,
		0x1b,
		0xae,
		0x87,
		0x74,
		0x0c,
		0xca,
		0xf7,
		0x8e,
		0x5f,
		0x2e,
		0x54,
		0x32,
		0xf6,
		0x79,
		0xb9,
		0x27,
		0x26,
		0x96,
		0x20,
		0x92,
		0x70,
		0x07,
		0x85,
		0xeb,
		0x83,
		0xf7,
		0x89,
		0xe0,
		0xd7,
		0x32,
		0x2a,
		0xd2,
		0x1a,
		0x64,
		0x41,
		0xef,
		0x49,
		0xff,
		0xc3,
		0x8c,
		0x54,
		0xf9,
		0x67,
		0x74,
		0x30,
		0x1e,
		0x70,
		0x2e,
		0xb7,
		0x12,
		0x09,
		0xfe,
	};

	memcpy(&seed, hash_key, sizeof(seed));

	PTR_ARRAY_FOREACH(MESSAGE_FIELD(task, parts), i, part)
	{
		n[0] = t1ha2_atonce128(&n[1],
							   part->digest, sizeof(part->digest),
							   seed);

		seed = n[0] ^ n[1];
	}

	memcpy(MESSAGE_FIELD(task, digest), n, sizeof(n));

	if (MESSAGE_FIELD(task, subject)) {
		p = MESSAGE_FIELD(task, subject);
		len = strlen(p);
		n[0] = t1ha2_atonce128(&n[1],
							   p, len,
							   seed);
		memcpy(MESSAGE_FIELD(task, digest), n, sizeof(n));
	}

	if (task->queue_id) {
		msg_info_task("loaded message; id: <%s>; queue-id: <%s>; size: %z; "
					  "checksum: <%*xs>",
					  MESSAGE_FIELD(task, message_id), task->queue_id, task->msg.len,
					  (int) sizeof(MESSAGE_FIELD(task, digest)), MESSAGE_FIELD(task, digest));
	}
	else {
		msg_info_task("loaded message; id: <%s>; size: %z; "
					  "checksum: <%*xs>",
					  MESSAGE_FIELD(task, message_id), task->msg.len,
					  (int) sizeof(MESSAGE_FIELD(task, digest)), MESSAGE_FIELD(task, digest));
	}

	return TRUE;
}


/*
 * A helper structure to store text parts positions, if it was C++, I could just use std::pair,
 * but here I have to make it all manually, sigh...
 */
struct rspamd_mime_part_text_position {
	unsigned pos;
	enum rspamd_message_part_is_text_result res;
};

/* Place html parts first during analysis */
static int
rspamd_mime_text_part_position_compare_func(const void *v1, const void *v2)
{
	const struct rspamd_mime_part_text_position *p1 = (const struct rspamd_mime_part_text_position *) v1;
	const struct rspamd_mime_part_text_position *p2 = (const struct rspamd_mime_part_text_position *) v2;

	if (p1->res == p2->res) {
		return (int) p2->pos - (int) p1->pos;
	}
	else {
		if (p1->res == RSPAMD_MESSAGE_PART_IS_TEXT_HTML) {
			return -1;
		}
		else {
			return 1;
		}
	}
}

void rspamd_message_process(struct rspamd_task *task)
{
	unsigned int i;
	struct rspamd_mime_text_part *p1, *p2;
	double diff, *pdiff;
	unsigned int tw, *ptw, dw;
	struct rspamd_mime_part *part;
	lua_State *L = NULL;
	int magic_func_pos = -1, content_func_pos = -1, old_top = -1, funcs_top = -1;

	if (task->cfg) {
		L = task->cfg->lua_state;
	}

	rspamd_archives_process(task);

	if (L) {
		old_top = lua_gettop(L);
	}

	if (L && rspamd_lua_require_function(L,
										 "lua_magic", "detect_mime_part")) {
		magic_func_pos = lua_gettop(L);
	}
	else {
		msg_err_task("cannot require lua_magic.detect_mime_part");
	}

	if (L && rspamd_lua_require_function(L,
										 "lua_content", "maybe_process_mime_part")) {
		content_func_pos = lua_gettop(L);
	}
	else {
		msg_err_task("cannot require lua_content.maybe_process_mime_part");
	}

	if (L) {
		funcs_top = lua_gettop(L);
	}

	GArray *detected_text_parts = g_array_sized_new(FALSE, FALSE, sizeof(struct rspamd_mime_part_text_position), 2);

	PTR_ARRAY_FOREACH(MESSAGE_FIELD(task, parts), i, part)
	{
		if (magic_func_pos != -1 && part->parsed_data.len > 0) {
			struct rspamd_mime_part **pmime;
			struct rspamd_task **ptask;

			lua_pushcfunction(L, &rspamd_lua_traceback);
			int err_idx = lua_gettop(L);
			lua_pushvalue(L, magic_func_pos);
			pmime = lua_newuserdata(L, sizeof(struct rspamd_mime_part *));
			rspamd_lua_setclass(L, rspamd_mimepart_classname, -1);
			*pmime = part;
			ptask = lua_newuserdata(L, sizeof(struct rspamd_task *));
			rspamd_lua_setclass(L, rspamd_task_classname, -1);
			*ptask = task;

			if (lua_pcall(L, 2, 2, err_idx) != 0) {
				msg_err_task("cannot detect type: %s", lua_tostring(L, -1));
			}
			else {
				if (lua_istable(L, -1)) {
					const char *mb;

					/* First returned value */
					part->detected_ext = rspamd_mempool_strdup(task->task_pool,
															   lua_tostring(L, -2));

					lua_pushstring(L, "ct");
					lua_gettable(L, -2);

					if (lua_isstring(L, -1)) {
						mb = lua_tostring(L, -1);

						if (mb) {
							rspamd_ftok_t srch;

							srch.begin = mb;
							srch.len = strlen(mb);
							part->detected_ct = rspamd_content_type_parse(srch.begin,
																		  srch.len,
																		  task->task_pool);
						}
					}

					lua_pop(L, 1);

					lua_pushstring(L, "type");
					lua_gettable(L, -2);

					if (lua_isstring(L, -1)) {
						part->detected_type = rspamd_mempool_strdup(task->task_pool,
																	lua_tostring(L, -1));
					}

					lua_pop(L, 1);

					lua_pushstring(L, "no_text");
					lua_gettable(L, -2);

					if (lua_isboolean(L, -1)) {
						if (!!lua_toboolean(L, -1)) {
							part->flags |= RSPAMD_MIME_PART_NO_TEXT_EXTRACTION;
						}
					}

					lua_pop(L, 1);
				}
			}

			lua_settop(L, funcs_top);
		}

		/* Now detect content */
		if (content_func_pos != -1 && part->parsed_data.len > 0 &&
			part->part_type == RSPAMD_MIME_PART_UNDEFINED) {
			struct rspamd_mime_part **pmime;
			struct rspamd_task **ptask;

			lua_pushcfunction(L, &rspamd_lua_traceback);
			int err_idx = lua_gettop(L);
			lua_pushvalue(L, content_func_pos);
			pmime = lua_newuserdata(L, sizeof(struct rspamd_mime_part *));
			rspamd_lua_setclass(L, rspamd_mimepart_classname, -1);
			*pmime = part;
			ptask = lua_newuserdata(L, sizeof(struct rspamd_task *));
			rspamd_lua_setclass(L, rspamd_task_classname, -1);
			*ptask = task;

			if (lua_pcall(L, 2, 0, err_idx) != 0) {
				msg_err_task("cannot detect content: %s", lua_tostring(L, -1));
			}

			lua_settop(L, funcs_top);
		}

		/* Try to detect image before checking for text */
		rspamd_images_process_mime_part_maybe(task, part);

		if (part->part_type == RSPAMD_MIME_PART_UNDEFINED &&
			!(part->flags & RSPAMD_MIME_PART_NO_TEXT_EXTRACTION)) {
			enum rspamd_message_part_is_text_result res = rspamd_message_part_can_be_parsed_as_text(task, part);

			if (res != RSPAMD_MESSAGE_PART_IS_NOT_TEXT) {
				struct rspamd_mime_part_text_position p = {
					.pos = i,
					.res = res};
				g_array_append_val(detected_text_parts, p);
			}
		}
	}

	uint16_t cur_url_order = 0;
	g_array_sort(detected_text_parts, rspamd_mime_text_part_position_compare_func);
	/* One more iteration to process text parts in a more specific order */
	for (i = 0; i < detected_text_parts->len; i++) {
		part = g_ptr_array_index(MESSAGE_FIELD(task, parts),
								 g_array_index(detected_text_parts, struct rspamd_mime_part_text_position, i).pos);
		rspamd_message_process_text_part_maybe(task, part,
											   g_array_index(detected_text_parts, struct rspamd_mime_part_text_position, i).res, &cur_url_order);
	}

	g_array_free(detected_text_parts, TRUE);

	if (old_top != -1) {
		lua_settop(L, old_top);
	}

	/* Parse urls inside Subject header */
	if (MESSAGE_FIELD(task, subject)) {
		rspamd_url_find_multiple(task->task_pool, MESSAGE_FIELD(task, subject),
								 strlen(MESSAGE_FIELD(task, subject)),
								 RSPAMD_URL_FIND_STRICT, NULL,
								 rspamd_url_task_subject_callback,
								 task);
	}

	/* Calculate average words length and number of short words */
	struct rspamd_mime_text_part *text_part;
	double *var;
	unsigned int total_words = 0;

	PTR_ARRAY_FOREACH(MESSAGE_FIELD(task, text_parts), i, text_part)
	{
		if (!text_part->language) {
			rspamd_mime_part_detect_language(task, text_part);
		}

		rspamd_mime_part_extract_words(task, text_part);

		if (text_part->utf_words) {
			total_words += text_part->nwords;
		}
	}

	/* Calculate distance for 2-parts messages */
	if (i == 2) {
		p1 = g_ptr_array_index(MESSAGE_FIELD(task, text_parts), 0);
		p2 = g_ptr_array_index(MESSAGE_FIELD(task, text_parts), 1);

		/* First of all check parent object */
		if (p1->mime_part->parent_part) {
			rspamd_ftok_t srch;

			srch.begin = "alternative";
			srch.len = 11;

			if (rspamd_ftok_cmp(&p1->mime_part->parent_part->ct->subtype, &srch) == 0) {
				if (!IS_TEXT_PART_EMPTY(p1) && !IS_TEXT_PART_EMPTY(p2) &&
					p1->normalized_hashes && p2->normalized_hashes) {
					/*
					 * We also detect language on one part and propagate it to
					 * another one
					 */
					struct rspamd_mime_text_part *sel;

					/* Prefer HTML as text part is not displayed normally */
					if (IS_TEXT_PART_HTML(p1)) {
						sel = p1;
					}
					else if (IS_TEXT_PART_HTML(p2)) {
						sel = p2;
					}
					else {
						if (p1->utf_content.len > p2->utf_content.len) {
							sel = p1;
						}
						else {
							sel = p2;
						}
					}

					if (sel->language && sel->language[0]) {
						/* Propagate language */
						if (sel == p1) {
							if (p2->languages) {
								g_ptr_array_unref(p2->languages);
							}

							p2->language = sel->language;
							p2->languages = g_ptr_array_ref(sel->languages);
						}
						else {
							if (p1->languages) {
								g_ptr_array_unref(p1->languages);
							}

							p1->language = sel->language;
							p1->languages = g_ptr_array_ref(sel->languages);
						}
					}

					tw = p1->normalized_hashes->len + p2->normalized_hashes->len;

					if (tw > 0) {
						dw = rspamd_words_levenshtein_distance(task,
															   p1->normalized_hashes,
															   p2->normalized_hashes);
						diff = dw / (double) tw;

						msg_debug_task(
							"different words: %d, total words: %d, "
							"got diff between parts of %.2f",
							dw, tw,
							diff);

						pdiff = rspamd_mempool_alloc(task->task_pool,
													 sizeof(double));
						*pdiff = diff;
						rspamd_mempool_set_variable(task->task_pool,
													"parts_distance",
													pdiff,
													NULL);
						ptw = rspamd_mempool_alloc(task->task_pool,
												   sizeof(int));
						*ptw = tw;
						rspamd_mempool_set_variable(task->task_pool,
													"total_words",
													ptw,
													NULL);
					}
				}
			}
		}
		else {
			debug_task(
				"message contains two parts but they are in different multi-parts");
		}
	}

	if (total_words > 0) {
		var = rspamd_mempool_get_variable(task->task_pool,
										  RSPAMD_MEMPOOL_AVG_WORDS_LEN);

		if (var) {
			*var /= (double) total_words;
		}

		var = rspamd_mempool_get_variable(task->task_pool,
										  RSPAMD_MEMPOOL_SHORT_WORDS_CNT);

		if (var) {
			*var /= (double) total_words;
		}
	}

	rspamd_images_link(task);
	rspamd_tokenize_meta_words(task);
}


struct rspamd_message *
rspamd_message_ref(struct rspamd_message *msg)
{
	REF_RETAIN(msg);

	return msg;
}

void rspamd_message_unref(struct rspamd_message *msg)
{
	if (msg) {
		REF_RELEASE(msg);
	}
}

void rspamd_message_update_digest(struct rspamd_message *msg,
								  const void *input, gsize len)
{
	uint64_t n[2];
	/* Sanity */
	G_STATIC_ASSERT(sizeof(n) == sizeof(msg->digest));

	memcpy(n, msg->digest, sizeof(msg->digest));
	n[0] = t1ha2_atonce128(&n[1], input, len, n[0]);
	memcpy(msg->digest, n, sizeof(msg->digest));
}