xref: /freebsd/contrib/xz/src/liblzma/lz/lz_encoder.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 ///////////////////////////////////////////////////////////////////////////////
2 //
3 /// \file       lz_encoder.c
4 /// \brief      LZ in window
5 ///
6 //  Authors:    Igor Pavlov
7 //              Lasse Collin
8 //
9 //  This file has been put into the public domain.
10 //  You can do whatever you want with this file.
11 //
12 ///////////////////////////////////////////////////////////////////////////////
13 
14 #include "lz_encoder.h"
15 #include "lz_encoder_hash.h"
16 
17 // See lz_encoder_hash.h. This is a bit hackish but avoids making
18 // endianness a conditional in makefiles.
19 #if defined(WORDS_BIGENDIAN) && !defined(HAVE_SMALL)
20 #	include "lz_encoder_hash_table.h"
21 #endif
22 
23 #include "memcmplen.h"
24 
25 
26 struct lzma_coder_s {
27 	/// LZ-based encoder e.g. LZMA
28 	lzma_lz_encoder lz;
29 
30 	/// History buffer and match finder
31 	lzma_mf mf;
32 
33 	/// Next coder in the chain
34 	lzma_next_coder next;
35 };
36 
37 
38 /// \brief      Moves the data in the input window to free space for new data
39 ///
40 /// mf->buffer is a sliding input window, which keeps mf->keep_size_before
41 /// bytes of input history available all the time. Now and then we need to
42 /// "slide" the buffer to make space for the new data to the end of the
43 /// buffer. At the same time, data older than keep_size_before is dropped.
44 ///
45 static void
46 move_window(lzma_mf *mf)
47 {
48 	// Align the move to a multiple of 16 bytes. Some LZ-based encoders
49 	// like LZMA use the lowest bits of mf->read_pos to know the
50 	// alignment of the uncompressed data. We also get better speed
51 	// for memmove() with aligned buffers.
52 	assert(mf->read_pos > mf->keep_size_before);
53 	const uint32_t move_offset
54 		= (mf->read_pos - mf->keep_size_before) & ~UINT32_C(15);
55 
56 	assert(mf->write_pos > move_offset);
57 	const size_t move_size = mf->write_pos - move_offset;
58 
59 	assert(move_offset + move_size <= mf->size);
60 
61 	memmove(mf->buffer, mf->buffer + move_offset, move_size);
62 
63 	mf->offset += move_offset;
64 	mf->read_pos -= move_offset;
65 	mf->read_limit -= move_offset;
66 	mf->write_pos -= move_offset;
67 
68 	return;
69 }
70 
71 
72 /// \brief      Tries to fill the input window (mf->buffer)
73 ///
74 /// If we are the last encoder in the chain, our input data is in in[].
75 /// Otherwise we call the next filter in the chain to process in[] and
76 /// write its output to mf->buffer.
77 ///
78 /// This function must not be called once it has returned LZMA_STREAM_END.
79 ///
80 static lzma_ret
81 fill_window(lzma_coder *coder, const lzma_allocator *allocator,
82 		const uint8_t *in, size_t *in_pos, size_t in_size,
83 		lzma_action action)
84 {
85 	assert(coder->mf.read_pos <= coder->mf.write_pos);
86 
87 	// Move the sliding window if needed.
88 	if (coder->mf.read_pos >= coder->mf.size - coder->mf.keep_size_after)
89 		move_window(&coder->mf);
90 
91 	// Maybe this is ugly, but lzma_mf uses uint32_t for most things
92 	// (which I find cleanest), but we need size_t here when filling
93 	// the history window.
94 	size_t write_pos = coder->mf.write_pos;
95 	lzma_ret ret;
96 	if (coder->next.code == NULL) {
97 		// Not using a filter, simply memcpy() as much as possible.
98 		lzma_bufcpy(in, in_pos, in_size, coder->mf.buffer,
99 				&write_pos, coder->mf.size);
100 
101 		ret = action != LZMA_RUN && *in_pos == in_size
102 				? LZMA_STREAM_END : LZMA_OK;
103 
104 	} else {
105 		ret = coder->next.code(coder->next.coder, allocator,
106 				in, in_pos, in_size,
107 				coder->mf.buffer, &write_pos,
108 				coder->mf.size, action);
109 	}
110 
111 	coder->mf.write_pos = write_pos;
112 
113 	// Silence Valgrind. lzma_memcmplen() can read extra bytes
114 	// and Valgrind will give warnings if those bytes are uninitialized
115 	// because Valgrind cannot see that the values of the uninitialized
116 	// bytes are eventually ignored.
117 	memzero(coder->mf.buffer + write_pos, LZMA_MEMCMPLEN_EXTRA);
118 
119 	// If end of stream has been reached or flushing completed, we allow
120 	// the encoder to process all the input (that is, read_pos is allowed
121 	// to reach write_pos). Otherwise we keep keep_size_after bytes
122 	// available as prebuffer.
123 	if (ret == LZMA_STREAM_END) {
124 		assert(*in_pos == in_size);
125 		ret = LZMA_OK;
126 		coder->mf.action = action;
127 		coder->mf.read_limit = coder->mf.write_pos;
128 
129 	} else if (coder->mf.write_pos > coder->mf.keep_size_after) {
130 		// This needs to be done conditionally, because if we got
131 		// only little new input, there may be too little input
132 		// to do any encoding yet.
133 		coder->mf.read_limit = coder->mf.write_pos
134 				- coder->mf.keep_size_after;
135 	}
136 
137 	// Restart the match finder after finished LZMA_SYNC_FLUSH.
138 	if (coder->mf.pending > 0
139 			&& coder->mf.read_pos < coder->mf.read_limit) {
140 		// Match finder may update coder->pending and expects it to
141 		// start from zero, so use a temporary variable.
142 		const uint32_t pending = coder->mf.pending;
143 		coder->mf.pending = 0;
144 
145 		// Rewind read_pos so that the match finder can hash
146 		// the pending bytes.
147 		assert(coder->mf.read_pos >= pending);
148 		coder->mf.read_pos -= pending;
149 
150 		// Call the skip function directly instead of using
151 		// mf_skip(), since we don't want to touch mf->read_ahead.
152 		coder->mf.skip(&coder->mf, pending);
153 	}
154 
155 	return ret;
156 }
157 
158 
159 static lzma_ret
160 lz_encode(lzma_coder *coder, const lzma_allocator *allocator,
161 		const uint8_t *restrict in, size_t *restrict in_pos,
162 		size_t in_size,
163 		uint8_t *restrict out, size_t *restrict out_pos,
164 		size_t out_size, lzma_action action)
165 {
166 	while (*out_pos < out_size
167 			&& (*in_pos < in_size || action != LZMA_RUN)) {
168 		// Read more data to coder->mf.buffer if needed.
169 		if (coder->mf.action == LZMA_RUN && coder->mf.read_pos
170 				>= coder->mf.read_limit)
171 			return_if_error(fill_window(coder, allocator,
172 					in, in_pos, in_size, action));
173 
174 		// Encode
175 		const lzma_ret ret = coder->lz.code(coder->lz.coder,
176 				&coder->mf, out, out_pos, out_size);
177 		if (ret != LZMA_OK) {
178 			// Setting this to LZMA_RUN for cases when we are
179 			// flushing. It doesn't matter when finishing or if
180 			// an error occurred.
181 			coder->mf.action = LZMA_RUN;
182 			return ret;
183 		}
184 	}
185 
186 	return LZMA_OK;
187 }
188 
189 
190 static bool
191 lz_encoder_prepare(lzma_mf *mf, const lzma_allocator *allocator,
192 		const lzma_lz_options *lz_options)
193 {
194 	// For now, the dictionary size is limited to 1.5 GiB. This may grow
195 	// in the future if needed, but it needs a little more work than just
196 	// changing this check.
197 	if (lz_options->dict_size < LZMA_DICT_SIZE_MIN
198 			|| lz_options->dict_size
199 				> (UINT32_C(1) << 30) + (UINT32_C(1) << 29)
200 			|| lz_options->nice_len > lz_options->match_len_max)
201 		return true;
202 
203 	mf->keep_size_before = lz_options->before_size + lz_options->dict_size;
204 
205 	mf->keep_size_after = lz_options->after_size
206 			+ lz_options->match_len_max;
207 
208 	// To avoid constant memmove()s, allocate some extra space. Since
209 	// memmove()s become more expensive when the size of the buffer
210 	// increases, we reserve more space when a large dictionary is
211 	// used to make the memmove() calls rarer.
212 	//
213 	// This works with dictionaries up to about 3 GiB. If bigger
214 	// dictionary is wanted, some extra work is needed:
215 	//   - Several variables in lzma_mf have to be changed from uint32_t
216 	//     to size_t.
217 	//   - Memory usage calculation needs something too, e.g. use uint64_t
218 	//     for mf->size.
219 	uint32_t reserve = lz_options->dict_size / 2;
220 	if (reserve > (UINT32_C(1) << 30))
221 		reserve /= 2;
222 
223 	reserve += (lz_options->before_size + lz_options->match_len_max
224 			+ lz_options->after_size) / 2 + (UINT32_C(1) << 19);
225 
226 	const uint32_t old_size = mf->size;
227 	mf->size = mf->keep_size_before + reserve + mf->keep_size_after;
228 
229 	// Deallocate the old history buffer if it exists but has different
230 	// size than what is needed now.
231 	if (mf->buffer != NULL && old_size != mf->size) {
232 		lzma_free(mf->buffer, allocator);
233 		mf->buffer = NULL;
234 	}
235 
236 	// Match finder options
237 	mf->match_len_max = lz_options->match_len_max;
238 	mf->nice_len = lz_options->nice_len;
239 
240 	// cyclic_size has to stay smaller than 2 Gi. Note that this doesn't
241 	// mean limiting dictionary size to less than 2 GiB. With a match
242 	// finder that uses multibyte resolution (hashes start at e.g. every
243 	// fourth byte), cyclic_size would stay below 2 Gi even when
244 	// dictionary size is greater than 2 GiB.
245 	//
246 	// It would be possible to allow cyclic_size >= 2 Gi, but then we
247 	// would need to be careful to use 64-bit types in various places
248 	// (size_t could do since we would need bigger than 32-bit address
249 	// space anyway). It would also require either zeroing a multigigabyte
250 	// buffer at initialization (waste of time and RAM) or allow
251 	// normalization in lz_encoder_mf.c to access uninitialized
252 	// memory to keep the code simpler. The current way is simple and
253 	// still allows pretty big dictionaries, so I don't expect these
254 	// limits to change.
255 	mf->cyclic_size = lz_options->dict_size + 1;
256 
257 	// Validate the match finder ID and setup the function pointers.
258 	switch (lz_options->match_finder) {
259 #ifdef HAVE_MF_HC3
260 	case LZMA_MF_HC3:
261 		mf->find = &lzma_mf_hc3_find;
262 		mf->skip = &lzma_mf_hc3_skip;
263 		break;
264 #endif
265 #ifdef HAVE_MF_HC4
266 	case LZMA_MF_HC4:
267 		mf->find = &lzma_mf_hc4_find;
268 		mf->skip = &lzma_mf_hc4_skip;
269 		break;
270 #endif
271 #ifdef HAVE_MF_BT2
272 	case LZMA_MF_BT2:
273 		mf->find = &lzma_mf_bt2_find;
274 		mf->skip = &lzma_mf_bt2_skip;
275 		break;
276 #endif
277 #ifdef HAVE_MF_BT3
278 	case LZMA_MF_BT3:
279 		mf->find = &lzma_mf_bt3_find;
280 		mf->skip = &lzma_mf_bt3_skip;
281 		break;
282 #endif
283 #ifdef HAVE_MF_BT4
284 	case LZMA_MF_BT4:
285 		mf->find = &lzma_mf_bt4_find;
286 		mf->skip = &lzma_mf_bt4_skip;
287 		break;
288 #endif
289 
290 	default:
291 		return true;
292 	}
293 
294 	// Calculate the sizes of mf->hash and mf->son and check that
295 	// nice_len is big enough for the selected match finder.
296 	const uint32_t hash_bytes = lz_options->match_finder & 0x0F;
297 	if (hash_bytes > mf->nice_len)
298 		return true;
299 
300 	const bool is_bt = (lz_options->match_finder & 0x10) != 0;
301 	uint32_t hs;
302 
303 	if (hash_bytes == 2) {
304 		hs = 0xFFFF;
305 	} else {
306 		// Round dictionary size up to the next 2^n - 1 so it can
307 		// be used as a hash mask.
308 		hs = lz_options->dict_size - 1;
309 		hs |= hs >> 1;
310 		hs |= hs >> 2;
311 		hs |= hs >> 4;
312 		hs |= hs >> 8;
313 		hs >>= 1;
314 		hs |= 0xFFFF;
315 
316 		if (hs > (UINT32_C(1) << 24)) {
317 			if (hash_bytes == 3)
318 				hs = (UINT32_C(1) << 24) - 1;
319 			else
320 				hs >>= 1;
321 		}
322 	}
323 
324 	mf->hash_mask = hs;
325 
326 	++hs;
327 	if (hash_bytes > 2)
328 		hs += HASH_2_SIZE;
329 	if (hash_bytes > 3)
330 		hs += HASH_3_SIZE;
331 /*
332 	No match finder uses this at the moment.
333 	if (mf->hash_bytes > 4)
334 		hs += HASH_4_SIZE;
335 */
336 
337 	const uint32_t old_hash_count = mf->hash_count;
338 	const uint32_t old_sons_count = mf->sons_count;
339 	mf->hash_count = hs;
340 	mf->sons_count = mf->cyclic_size;
341 	if (is_bt)
342 		mf->sons_count *= 2;
343 
344 	// Deallocate the old hash array if it exists and has different size
345 	// than what is needed now.
346 	if (old_hash_count != mf->hash_count
347 			|| old_sons_count != mf->sons_count) {
348 		lzma_free(mf->hash, allocator);
349 		mf->hash = NULL;
350 
351 		lzma_free(mf->son, allocator);
352 		mf->son = NULL;
353 	}
354 
355 	// Maximum number of match finder cycles
356 	mf->depth = lz_options->depth;
357 	if (mf->depth == 0) {
358 		if (is_bt)
359 			mf->depth = 16 + mf->nice_len / 2;
360 		else
361 			mf->depth = 4 + mf->nice_len / 4;
362 	}
363 
364 	return false;
365 }
366 
367 
368 static bool
369 lz_encoder_init(lzma_mf *mf, const lzma_allocator *allocator,
370 		const lzma_lz_options *lz_options)
371 {
372 	// Allocate the history buffer.
373 	if (mf->buffer == NULL) {
374 		// lzma_memcmplen() is used for the dictionary buffer
375 		// so we need to allocate a few extra bytes to prevent
376 		// it from reading past the end of the buffer.
377 		mf->buffer = lzma_alloc(mf->size + LZMA_MEMCMPLEN_EXTRA,
378 				allocator);
379 		if (mf->buffer == NULL)
380 			return true;
381 
382 		// Keep Valgrind happy with lzma_memcmplen() and initialize
383 		// the extra bytes whose value may get read but which will
384 		// effectively get ignored.
385 		memzero(mf->buffer + mf->size, LZMA_MEMCMPLEN_EXTRA);
386 	}
387 
388 	// Use cyclic_size as initial mf->offset. This allows
389 	// avoiding a few branches in the match finders. The downside is
390 	// that match finder needs to be normalized more often, which may
391 	// hurt performance with huge dictionaries.
392 	mf->offset = mf->cyclic_size;
393 	mf->read_pos = 0;
394 	mf->read_ahead = 0;
395 	mf->read_limit = 0;
396 	mf->write_pos = 0;
397 	mf->pending = 0;
398 
399 #if UINT32_MAX >= SIZE_MAX / 4
400 	// Check for integer overflow. (Huge dictionaries are not
401 	// possible on 32-bit CPU.)
402 	if (mf->hash_count > SIZE_MAX / sizeof(uint32_t)
403 			|| mf->sons_count > SIZE_MAX / sizeof(uint32_t))
404 		return true;
405 #endif
406 
407 	// Allocate and initialize the hash table. Since EMPTY_HASH_VALUE
408 	// is zero, we can use lzma_alloc_zero() or memzero() for mf->hash.
409 	//
410 	// We don't need to initialize mf->son, but not doing that may
411 	// make Valgrind complain in normalization (see normalize() in
412 	// lz_encoder_mf.c). Skipping the initialization is *very* good
413 	// when big dictionary is used but only small amount of data gets
414 	// actually compressed: most of the mf->son won't get actually
415 	// allocated by the kernel, so we avoid wasting RAM and improve
416 	// initialization speed a lot.
417 	if (mf->hash == NULL) {
418 		mf->hash = lzma_alloc_zero(mf->hash_count * sizeof(uint32_t),
419 				allocator);
420 		mf->son = lzma_alloc(mf->sons_count * sizeof(uint32_t),
421 				allocator);
422 
423 		if (mf->hash == NULL || mf->son == NULL) {
424 			lzma_free(mf->hash, allocator);
425 			mf->hash = NULL;
426 
427 			lzma_free(mf->son, allocator);
428 			mf->son = NULL;
429 
430 			return true;
431 		}
432 	} else {
433 /*
434 		for (uint32_t i = 0; i < mf->hash_count; ++i)
435 			mf->hash[i] = EMPTY_HASH_VALUE;
436 */
437 		memzero(mf->hash, mf->hash_count * sizeof(uint32_t));
438 	}
439 
440 	mf->cyclic_pos = 0;
441 
442 	// Handle preset dictionary.
443 	if (lz_options->preset_dict != NULL
444 			&& lz_options->preset_dict_size > 0) {
445 		// If the preset dictionary is bigger than the actual
446 		// dictionary, use only the tail.
447 		mf->write_pos = my_min(lz_options->preset_dict_size, mf->size);
448 		memcpy(mf->buffer, lz_options->preset_dict
449 				+ lz_options->preset_dict_size - mf->write_pos,
450 				mf->write_pos);
451 		mf->action = LZMA_SYNC_FLUSH;
452 		mf->skip(mf, mf->write_pos);
453 	}
454 
455 	mf->action = LZMA_RUN;
456 
457 	return false;
458 }
459 
460 
461 extern uint64_t
462 lzma_lz_encoder_memusage(const lzma_lz_options *lz_options)
463 {
464 	// Old buffers must not exist when calling lz_encoder_prepare().
465 	lzma_mf mf = {
466 		.buffer = NULL,
467 		.hash = NULL,
468 		.son = NULL,
469 		.hash_count = 0,
470 		.sons_count = 0,
471 	};
472 
473 	// Setup the size information into mf.
474 	if (lz_encoder_prepare(&mf, NULL, lz_options))
475 		return UINT64_MAX;
476 
477 	// Calculate the memory usage.
478 	return ((uint64_t)(mf.hash_count) + mf.sons_count) * sizeof(uint32_t)
479 			+ mf.size + sizeof(lzma_coder);
480 }
481 
482 
483 static void
484 lz_encoder_end(lzma_coder *coder, const lzma_allocator *allocator)
485 {
486 	lzma_next_end(&coder->next, allocator);
487 
488 	lzma_free(coder->mf.son, allocator);
489 	lzma_free(coder->mf.hash, allocator);
490 	lzma_free(coder->mf.buffer, allocator);
491 
492 	if (coder->lz.end != NULL)
493 		coder->lz.end(coder->lz.coder, allocator);
494 	else
495 		lzma_free(coder->lz.coder, allocator);
496 
497 	lzma_free(coder, allocator);
498 	return;
499 }
500 
501 
502 static lzma_ret
503 lz_encoder_update(lzma_coder *coder, const lzma_allocator *allocator,
504 		const lzma_filter *filters_null lzma_attribute((__unused__)),
505 		const lzma_filter *reversed_filters)
506 {
507 	if (coder->lz.options_update == NULL)
508 		return LZMA_PROG_ERROR;
509 
510 	return_if_error(coder->lz.options_update(
511 			coder->lz.coder, reversed_filters));
512 
513 	return lzma_next_filter_update(
514 			&coder->next, allocator, reversed_filters + 1);
515 }
516 
517 
518 extern lzma_ret
519 lzma_lz_encoder_init(lzma_next_coder *next, const lzma_allocator *allocator,
520 		const lzma_filter_info *filters,
521 		lzma_ret (*lz_init)(lzma_lz_encoder *lz,
522 			const lzma_allocator *allocator, const void *options,
523 			lzma_lz_options *lz_options))
524 {
525 #ifdef HAVE_SMALL
526 	// We need that the CRC32 table has been initialized.
527 	lzma_crc32_init();
528 #endif
529 
530 	// Allocate and initialize the base data structure.
531 	if (next->coder == NULL) {
532 		next->coder = lzma_alloc(sizeof(lzma_coder), allocator);
533 		if (next->coder == NULL)
534 			return LZMA_MEM_ERROR;
535 
536 		next->code = &lz_encode;
537 		next->end = &lz_encoder_end;
538 		next->update = &lz_encoder_update;
539 
540 		next->coder->lz.coder = NULL;
541 		next->coder->lz.code = NULL;
542 		next->coder->lz.end = NULL;
543 
544 		next->coder->mf.buffer = NULL;
545 		next->coder->mf.hash = NULL;
546 		next->coder->mf.son = NULL;
547 		next->coder->mf.hash_count = 0;
548 		next->coder->mf.sons_count = 0;
549 
550 		next->coder->next = LZMA_NEXT_CODER_INIT;
551 	}
552 
553 	// Initialize the LZ-based encoder.
554 	lzma_lz_options lz_options;
555 	return_if_error(lz_init(&next->coder->lz, allocator,
556 			filters[0].options, &lz_options));
557 
558 	// Setup the size information into next->coder->mf and deallocate
559 	// old buffers if they have wrong size.
560 	if (lz_encoder_prepare(&next->coder->mf, allocator, &lz_options))
561 		return LZMA_OPTIONS_ERROR;
562 
563 	// Allocate new buffers if needed, and do the rest of
564 	// the initialization.
565 	if (lz_encoder_init(&next->coder->mf, allocator, &lz_options))
566 		return LZMA_MEM_ERROR;
567 
568 	// Initialize the next filter in the chain, if any.
569 	return lzma_next_filter_init(&next->coder->next, allocator,
570 			filters + 1);
571 }
572 
573 
574 extern LZMA_API(lzma_bool)
575 lzma_mf_is_supported(lzma_match_finder mf)
576 {
577 	bool ret = false;
578 
579 #ifdef HAVE_MF_HC3
580 	if (mf == LZMA_MF_HC3)
581 		ret = true;
582 #endif
583 
584 #ifdef HAVE_MF_HC4
585 	if (mf == LZMA_MF_HC4)
586 		ret = true;
587 #endif
588 
589 #ifdef HAVE_MF_BT2
590 	if (mf == LZMA_MF_BT2)
591 		ret = true;
592 #endif
593 
594 #ifdef HAVE_MF_BT3
595 	if (mf == LZMA_MF_BT3)
596 		ret = true;
597 #endif
598 
599 #ifdef HAVE_MF_BT4
600 	if (mf == LZMA_MF_BT4)
601 		ret = true;
602 #endif
603 
604 	return ret;
605 }
606