| blob | 12d2530cda4831e133ffaebad7d0f3f22ae302f2 |
2 The .xz File Format
3 ===================
5 Version 1.2.1 (2024-04-08)
8 0. Preface
9 0.1. Notices and Acknowledgements
10 0.2. Getting the Latest Version
11 0.3. Version History
12 1. Conventions
13 1.1. Byte and Its Representation
14 1.2. Multibyte Integers
15 2. Overall Structure of .xz File
16 2.1. Stream
17 2.1.1. Stream Header
18 2.1.1.1. Header Magic Bytes
19 2.1.1.2. Stream Flags
20 2.1.1.3. CRC32
21 2.1.2. Stream Footer
22 2.1.2.1. CRC32
23 2.1.2.2. Backward Size
24 2.1.2.3. Stream Flags
25 2.1.2.4. Footer Magic Bytes
26 2.2. Stream Padding
27 3. Block
28 3.1. Block Header
29 3.1.1. Block Header Size
30 3.1.2. Block Flags
31 3.1.3. Compressed Size
32 3.1.4. Uncompressed Size
33 3.1.5. List of Filter Flags
34 3.1.6. Header Padding
35 3.1.7. CRC32
36 3.2. Compressed Data
37 3.3. Block Padding
38 3.4. Check
39 4. Index
40 4.1. Index Indicator
41 4.2. Number of Records
42 4.3. List of Records
43 4.3.1. Unpadded Size
44 4.3.2. Uncompressed Size
45 4.4. Index Padding
46 4.5. CRC32
47 5. Filter Chains
48 5.1. Alignment
49 5.2. Security
50 5.3. Filters
51 5.3.1. LZMA2
52 5.3.2. Branch/Call/Jump Filters for Executables
53 5.3.3. Delta
54 5.3.3.1. Format of the Encoded Output
55 5.4. Custom Filter IDs
56 5.4.1. Reserved Custom Filter ID Ranges
57 6. Cyclic Redundancy Checks
58 7. References
61 0. Preface
63 This document describes the .xz file format (filename suffix
64 ".xz", MIME type "application/x-xz"). It is intended that this
65 this format replace the old .lzma format used by LZMA SDK and
66 LZMA Utils.
69 0.1. Notices and Acknowledgements
71 This file format was designed by Lasse Collin
72 <lasse.collin@tukaani.org> and Igor Pavlov.
74 Special thanks for helping with this document goes to
75 Ville Koskinen. Thanks for helping with this document goes to
76 Mark Adler, H. Peter Anvin, Mikko Pouru, and Lars Wirzenius.
78 This document has been put into the public domain.
81 0.2. Getting the Latest Version
83 The latest official version of this document can be downloaded
84 from <https://tukaani.org/xz/xz-file-format.txt>.
86 Specific versions of this document have a filename
87 xz-file-format-X.Y.Z.txt where X.Y.Z is the version number.
88 For example, the version 1.0.0 of this document is available
89 at <https://tukaani.org/xz/xz-file-format-1.0.0.txt>.
92 0.3. Version History
94 Version Date Description
96 1.2.1 2024-04-08 The URLs of this specification and
97 XZ Utils were changed back to the
98 original ones in Sections 0.2 and 7.
100 1.2.0 2024-01-19 Added RISC-V filter and updated URLs in
101 Sections 0.2 and 7. The URL of this
102 specification was changed.
104 1.1.0 2022-12-11 Added ARM64 filter and clarified 32-bit
105 ARM endianness in Section 5.3.2,
106 language improvements in Section 5.4
108 1.0.4 2009-08-27 Language improvements in Sections 1.2,
109 2.1.1.2, 3.1.1, 3.1.2, and 5.3.1
111 1.0.3 2009-06-05 Spelling fixes in Sections 5.1 and 5.4
113 1.0.2 2009-06-04 Typo fixes in Sections 4 and 5.3.1
115 1.0.1 2009-06-01 Typo fix in Section 0.3 and minor
116 clarifications to Sections 2, 2.2,
117 3.3, 4.4, and 5.3.2
119 1.0.0 2009-01-14 The first official version
122 1. Conventions
124 The key words "MUST", "MUST NOT", "REQUIRED", "SHOULD",
125 "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
126 document are to be interpreted as described in [RFC-2119].
128 Indicating a warning means displaying a message, returning
129 appropriate exit status, or doing something else to let the
130 user know that something worth warning occurred. The operation
131 SHOULD still finish if a warning is indicated.
133 Indicating an error means displaying a message, returning
134 appropriate exit status, or doing something else to let the
135 user know that something prevented successfully finishing the
136 operation. The operation MUST be aborted once an error has
137 been indicated.
140 1.1. Byte and Its Representation
142 In this document, byte is always 8 bits.
144 A "null byte" has all bits unset. That is, the value of a null
145 byte is 0x00.
147 To represent byte blocks, this document uses notation that
148 is similar to the notation used in [RFC-1952]:
150 +-------+
151 | Foo | One byte.
152 +-------+
154 +---+---+
155 | Foo | Two bytes; that is, some of the vertical bars
156 +---+---+ can be missing.
158 +=======+
159 | Foo | Zero or more bytes.
160 +=======+
162 In this document, a boxed byte or a byte sequence declared
163 using this notation is called "a field". The example field
164 above would be called "the Foo field" or plain "Foo".
166 If there are many fields, they may be split to multiple lines.
167 This is indicated with an arrow ("--->"):
169 +=====+
170 | Foo |
171 +=====+
173 +=====+
174 ---> | Bar |
175 +=====+
177 The above is equivalent to this:
179 +=====+=====+
180 | Foo | Bar |
181 +=====+=====+
184 1.2. Multibyte Integers
186 Multibyte integers of static length, such as CRC values,
187 are stored in little endian byte order (least significant
188 byte first).
190 When smaller values are more likely than bigger values (for
191 example file sizes), multibyte integers are encoded in a
192 variable-length representation:
193 - Numbers in the range [0, 127] are copied as is, and take
194 one byte of space.
195 - Bigger numbers will occupy two or more bytes. All but the
196 last byte of the multibyte representation have the highest
197 (eighth) bit set.
199 For now, the value of the variable-length integers is limited
200 to 63 bits, which limits the encoded size of the integer to
201 nine bytes. These limits may be increased in the future if
202 needed.
204 The following C code illustrates encoding and decoding of
205 variable-length integers. The functions return the number of
206 bytes occupied by the integer (1-9), or zero on error.
208 #include <stddef.h>
209 #include <inttypes.h>
211 size_t
212 encode(uint8_t buf[static 9], uint64_t num)
213 {
214 if (num > UINT64_MAX / 2)
215 return 0;
217 size_t i = 0;
219 while (num >= 0x80) {
220 buf[i++] = (uint8_t)(num) | 0x80;
221 num >>= 7;
222 }
224 buf[i++] = (uint8_t)(num);
226 return i;
227 }
229 size_t
230 decode(const uint8_t buf[], size_t size_max, uint64_t *num)
231 {
232 if (size_max == 0)
233 return 0;
235 if (size_max > 9)
236 size_max = 9;
238 *num = buf[0] & 0x7F;
239 size_t i = 0;
241 while (buf[i++] & 0x80) {
242 if (i >= size_max || buf[i] == 0x00)
243 return 0;
245 *num |= (uint64_t)(buf[i] & 0x7F) << (i * 7);
246 }
248 return i;
249 }
252 2. Overall Structure of .xz File
254 A standalone .xz files consist of one or more Streams which may
255 have Stream Padding between or after them:
257 +========+================+========+================+
258 | Stream | Stream Padding | Stream | Stream Padding | ...
259 +========+================+========+================+
261 The sizes of Stream and Stream Padding are always multiples
262 of four bytes, thus the size of every valid .xz file MUST be
263 a multiple of four bytes.
265 While a typical file contains only one Stream and no Stream
266 Padding, a decoder handling standalone .xz files SHOULD support
267 files that have more than one Stream or Stream Padding.
269 In contrast to standalone .xz files, when the .xz file format
270 is used as an internal part of some other file format or
271 communication protocol, it usually is expected that the decoder
272 stops after the first Stream, and doesn't look for Stream
273 Padding or possibly other Streams.
276 2.1. Stream
278 +-+-+-+-+-+-+-+-+-+-+-+-+=======+=======+ +=======+
279 | Stream Header | Block | Block | ... | Block |
280 +-+-+-+-+-+-+-+-+-+-+-+-+=======+=======+ +=======+
282 +=======+-+-+-+-+-+-+-+-+-+-+-+-+
283 ---> | Index | Stream Footer |
284 +=======+-+-+-+-+-+-+-+-+-+-+-+-+
286 All the above fields have a size that is a multiple of four. If
287 Stream is used as an internal part of another file format, it
288 is RECOMMENDED to make the Stream start at an offset that is
289 a multiple of four bytes.
291 Stream Header, Index, and Stream Footer are always present in
292 a Stream. The maximum size of the Index field is 16 GiB (2^34).
294 There are zero or more Blocks. The maximum number of Blocks is
295 limited only by the maximum size of the Index field.
297 Total size of a Stream MUST be less than 8 EiB (2^63 bytes).
298 The same limit applies to the total amount of uncompressed
299 data stored in a Stream.
301 If an implementation supports handling .xz files with multiple
302 concatenated Streams, it MAY apply the above limits to the file
303 as a whole instead of limiting per Stream basis.
306 2.1.1. Stream Header
308 +---+---+---+---+---+---+-------+------+--+--+--+--+
309 | Header Magic Bytes | Stream Flags | CRC32 |
310 +---+---+---+---+---+---+-------+------+--+--+--+--+
313 2.1.1.1. Header Magic Bytes
315 The first six (6) bytes of the Stream are so called Header
316 Magic Bytes. They can be used to identify the file type.
318 Using a C array and ASCII:
319 const uint8_t HEADER_MAGIC[6]
320 = { 0xFD, '7', 'z', 'X', 'Z', 0x00 };
322 In plain hexadecimal:
323 FD 37 7A 58 5A 00
325 Notes:
326 - The first byte (0xFD) was chosen so that the files cannot
327 be erroneously detected as being in .lzma format, in which
328 the first byte is in the range [0x00, 0xE0].
329 - The sixth byte (0x00) was chosen to prevent applications
330 from misdetecting the file as a text file.
332 If the Header Magic Bytes don't match, the decoder MUST
333 indicate an error.
336 2.1.1.2. Stream Flags
338 The first byte of Stream Flags is always a null byte. In the
339 future, this byte may be used to indicate a new Stream version
340 or other Stream properties.
342 The second byte of Stream Flags is a bit field:
344 Bit(s) Mask Description
345 0-3 0x0F Type of Check (see Section 3.4):
346 ID Size Check name
347 0x00 0 bytes None
348 0x01 4 bytes CRC32
349 0x02 4 bytes (Reserved)
350 0x03 4 bytes (Reserved)
351 0x04 8 bytes CRC64
352 0x05 8 bytes (Reserved)
353 0x06 8 bytes (Reserved)
354 0x07 16 bytes (Reserved)
355 0x08 16 bytes (Reserved)
356 0x09 16 bytes (Reserved)
357 0x0A 32 bytes SHA-256
358 0x0B 32 bytes (Reserved)
359 0x0C 32 bytes (Reserved)
360 0x0D 64 bytes (Reserved)
361 0x0E 64 bytes (Reserved)
362 0x0F 64 bytes (Reserved)
363 4-7 0xF0 Reserved for future use; MUST be zero for now.
365 Implementations SHOULD support at least the Check IDs 0x00
366 (None) and 0x01 (CRC32). Supporting other Check IDs is
367 OPTIONAL. If an unsupported Check is used, the decoder SHOULD
368 indicate a warning or error.
370 If any reserved bit is set, the decoder MUST indicate an error.
371 It is possible that there is a new field present which the
372 decoder is not aware of, and can thus parse the Stream Header
373 incorrectly.
376 2.1.1.3. CRC32
378 The CRC32 is calculated from the Stream Flags field. It is
379 stored as an unsigned 32-bit little endian integer. If the
380 calculated value does not match the stored one, the decoder
381 MUST indicate an error.
383 The idea is that Stream Flags would always be two bytes, even
384 if new features are needed. This way old decoders will be able
385 to verify the CRC32 calculated from Stream Flags, and thus
386 distinguish between corrupt files (CRC32 doesn't match) and
387 files that the decoder doesn't support (CRC32 matches but
388 Stream Flags has reserved bits set).
391 2.1.2. Stream Footer
393 +-+-+-+-+---+---+---+---+-------+------+----------+---------+
394 | CRC32 | Backward Size | Stream Flags | Footer Magic Bytes |
395 +-+-+-+-+---+---+---+---+-------+------+----------+---------+
398 2.1.2.1. CRC32
400 The CRC32 is calculated from the Backward Size and Stream Flags
401 fields. It is stored as an unsigned 32-bit little endian
402 integer. If the calculated value does not match the stored one,
403 the decoder MUST indicate an error.
405 The reason to have the CRC32 field before the Backward Size and
406 Stream Flags fields is to keep the four-byte fields aligned to
407 a multiple of four bytes.
410 2.1.2.2. Backward Size
412 Backward Size is stored as a 32-bit little endian integer,
413 which indicates the size of the Index field as multiple of
414 four bytes, minimum value being four bytes:
416 real_backward_size = (stored_backward_size + 1) * 4;
418 If the stored value does not match the real size of the Index
419 field, the decoder MUST indicate an error.
421 Using a fixed-size integer to store Backward Size makes
422 it slightly simpler to parse the Stream Footer when the
423 application needs to parse the Stream backwards.
426 2.1.2.3. Stream Flags
428 This is a copy of the Stream Flags field from the Stream
429 Header. The information stored to Stream Flags is needed
430 when parsing the Stream backwards. The decoder MUST compare
431 the Stream Flags fields in both Stream Header and Stream
432 Footer, and indicate an error if they are not identical.
435 2.1.2.4. Footer Magic Bytes
437 As the last step of the decoding process, the decoder MUST
438 verify the existence of Footer Magic Bytes. If they don't
439 match, an error MUST be indicated.
441 Using a C array and ASCII:
442 const uint8_t FOOTER_MAGIC[2] = { 'Y', 'Z' };
444 In hexadecimal:
445 59 5A
447 The primary reason to have Footer Magic Bytes is to make
448 it easier to detect incomplete files quickly, without
449 uncompressing. If the file does not end with Footer Magic Bytes
450 (excluding Stream Padding described in Section 2.2), it cannot
451 be undamaged, unless someone has intentionally appended garbage
452 after the end of the Stream.
455 2.2. Stream Padding
457 Only the decoders that support decoding of concatenated Streams
458 MUST support Stream Padding.
460 Stream Padding MUST contain only null bytes. To preserve the
461 four-byte alignment of consecutive Streams, the size of Stream
462 Padding MUST be a multiple of four bytes. Empty Stream Padding
463 is allowed. If these requirements are not met, the decoder MUST
464 indicate an error.
466 Note that non-empty Stream Padding is allowed at the end of the
467 file; there doesn't need to be a new Stream after non-empty
468 Stream Padding. This can be convenient in certain situations
469 [GNU-tar].
471 The possibility of Stream Padding MUST be taken into account
472 when designing an application that parses Streams backwards,
473 and the application supports concatenated Streams.
476 3. Block
478 +==============+=================+===============+=======+
479 | Block Header | Compressed Data | Block Padding | Check |
480 +==============+=================+===============+=======+
483 3.1. Block Header
485 +-------------------+-------------+=================+
486 | Block Header Size | Block Flags | Compressed Size |
487 +-------------------+-------------+=================+
489 +===================+======================+
490 ---> | Uncompressed Size | List of Filter Flags |
491 +===================+======================+
493 +================+--+--+--+--+
494 ---> | Header Padding | CRC32 |
495 +================+--+--+--+--+
498 3.1.1. Block Header Size
500 This field overlaps with the Index Indicator field (see
501 Section 4.1).
503 This field contains the size of the Block Header field,
504 including the Block Header Size field itself. Valid values are
505 in the range [0x01, 0xFF], which indicate the size of the Block
506 Header as multiples of four bytes, minimum size being eight
507 bytes:
509 real_header_size = (encoded_header_size + 1) * 4;
511 If a Block Header bigger than 1024 bytes is needed in the
512 future, a new field can be added between the Block Header and
513 Compressed Data fields. The presence of this new field would
514 be indicated in the Block Header field.
517 3.1.2. Block Flags
519 The Block Flags field is a bit field:
521 Bit(s) Mask Description
522 0-1 0x03 Number of filters (1-4)
523 2-5 0x3C Reserved for future use; MUST be zero for now.
524 6 0x40 The Compressed Size field is present.
525 7 0x80 The Uncompressed Size field is present.
527 If any reserved bit is set, the decoder MUST indicate an error.
528 It is possible that there is a new field present which the
529 decoder is not aware of, and can thus parse the Block Header
530 incorrectly.
533 3.1.3. Compressed Size
535 This field is present only if the appropriate bit is set in
536 the Block Flags field (see Section 3.1.2).
538 The Compressed Size field contains the size of the Compressed
539 Data field, which MUST be non-zero. Compressed Size is stored
540 using the encoding described in Section 1.2. If the Compressed
541 Size doesn't match the size of the Compressed Data field, the
542 decoder MUST indicate an error.
545 3.1.4. Uncompressed Size
547 This field is present only if the appropriate bit is set in
548 the Block Flags field (see Section 3.1.2).
550 The Uncompressed Size field contains the size of the Block
551 after uncompressing. Uncompressed Size is stored using the
552 encoding described in Section 1.2. If the Uncompressed Size
553 does not match the real uncompressed size, the decoder MUST
554 indicate an error.
556 Storing the Compressed Size and Uncompressed Size fields serves
557 several purposes:
558 - The decoder knows how much memory it needs to allocate
559 for a temporary buffer in multithreaded mode.
560 - Simple error detection: wrong size indicates a broken file.
561 - Seeking forwards to a specific location in streamed mode.
563 It should be noted that the only reliable way to determine
564 the real uncompressed size is to uncompress the Block,
565 because the Block Header and Index fields may contain
566 (intentionally or unintentionally) invalid information.
569 3.1.5. List of Filter Flags
571 +================+================+ +================+
572 | Filter 0 Flags | Filter 1 Flags | ... | Filter n Flags |
573 +================+================+ +================+
575 The number of Filter Flags fields is stored in the Block Flags
576 field (see Section 3.1.2).
578 The format of each Filter Flags field is as follows:
580 +===========+====================+===================+
581 | Filter ID | Size of Properties | Filter Properties |
582 +===========+====================+===================+
584 Both Filter ID and Size of Properties are stored using the
585 encoding described in Section 1.2. Size of Properties indicates
586 the size of the Filter Properties field as bytes. The list of
587 officially defined Filter IDs and the formats of their Filter
588 Properties are described in Section 5.3.
590 Filter IDs greater than or equal to 0x4000_0000_0000_0000
591 (2^62) are reserved for implementation-specific internal use.
592 These Filter IDs MUST never be used in List of Filter Flags.
595 3.1.6. Header Padding
597 This field contains as many null byte as it is needed to make
598 the Block Header have the size specified in Block Header Size.
599 If any of the bytes are not null bytes, the decoder MUST
600 indicate an error. It is possible that there is a new field
601 present which the decoder is not aware of, and can thus parse
602 the Block Header incorrectly.
605 3.1.7. CRC32
607 The CRC32 is calculated over everything in the Block Header
608 field except the CRC32 field itself. It is stored as an
609 unsigned 32-bit little endian integer. If the calculated
610 value does not match the stored one, the decoder MUST indicate
611 an error.
613 By verifying the CRC32 of the Block Header before parsing the
614 actual contents allows the decoder to distinguish between
615 corrupt and unsupported files.
618 3.2. Compressed Data
620 The format of Compressed Data depends on Block Flags and List
621 of Filter Flags. Excluding the descriptions of the simplest
622 filters in Section 5.3, the format of the filter-specific
623 encoded data is out of scope of this document.
626 3.3. Block Padding
628 Block Padding MUST contain 0-3 null bytes to make the size of
629 the Block a multiple of four bytes. This can be needed when
630 the size of Compressed Data is not a multiple of four. If any
631 of the bytes in Block Padding are not null bytes, the decoder
632 MUST indicate an error.
635 3.4. Check
637 The type and size of the Check field depends on which bits
638 are set in the Stream Flags field (see Section 2.1.1.2).
640 The Check, when used, is calculated from the original
641 uncompressed data. If the calculated Check does not match the
642 stored one, the decoder MUST indicate an error. If the selected
643 type of Check is not supported by the decoder, it SHOULD
644 indicate a warning or error.
647 4. Index
649 +-----------------+===================+
650 | Index Indicator | Number of Records |
651 +-----------------+===================+
653 +=================+===============+-+-+-+-+
654 ---> | List of Records | Index Padding | CRC32 |
655 +=================+===============+-+-+-+-+
657 Index serves several purposes. Using it, one can
658 - verify that all Blocks in a Stream have been processed;
659 - find out the uncompressed size of a Stream; and
660 - quickly access the beginning of any Block (random access).
663 4.1. Index Indicator
665 This field overlaps with the Block Header Size field (see
666 Section 3.1.1). The value of Index Indicator is always 0x00.
669 4.2. Number of Records
671 This field indicates how many Records there are in the List
672 of Records field, and thus how many Blocks there are in the
673 Stream. The value is stored using the encoding described in
674 Section 1.2. If the decoder has decoded all the Blocks of the
675 Stream, and then notices that the Number of Records doesn't
676 match the real number of Blocks, the decoder MUST indicate an
677 error.
680 4.3. List of Records
682 List of Records consists of as many Records as indicated by the
683 Number of Records field:
685 +========+========+
686 | Record | Record | ...
687 +========+========+
689 Each Record contains information about one Block:
691 +===============+===================+
692 | Unpadded Size | Uncompressed Size |
693 +===============+===================+
695 If the decoder has decoded all the Blocks of the Stream, it
696 MUST verify that the contents of the Records match the real
697 Unpadded Size and Uncompressed Size of the respective Blocks.
699 Implementation hint: It is possible to verify the Index with
700 constant memory usage by calculating for example SHA-256 of
701 both the real size values and the List of Records, then
702 comparing the hash values. Implementing this using
703 non-cryptographic hash like CRC32 SHOULD be avoided unless
704 small code size is important.
706 If the decoder supports random-access reading, it MUST verify
707 that Unpadded Size and Uncompressed Size of every completely
708 decoded Block match the sizes stored in the Index. If only
709 partial Block is decoded, the decoder MUST verify that the
710 processed sizes don't exceed the sizes stored in the Index.
713 4.3.1. Unpadded Size
715 This field indicates the size of the Block excluding the Block
716 Padding field. That is, Unpadded Size is the size of the Block
717 Header, Compressed Data, and Check fields. Unpadded Size is
718 stored using the encoding described in Section 1.2. The value
719 MUST never be zero; with the current structure of Blocks, the
720 actual minimum value for Unpadded Size is five.
722 Implementation note: Because the size of the Block Padding
723 field is not included in Unpadded Size, calculating the total
724 size of a Stream or doing random-access reading requires
725 calculating the actual size of the Blocks by rounding Unpadded
726 Sizes up to the next multiple of four.
728 The reason to exclude Block Padding from Unpadded Size is to
729 ease making a raw copy of Compressed Data without Block
730 Padding. This can be useful, for example, if someone wants
731 to convert Streams to some other file format quickly.
734 4.3.2. Uncompressed Size
736 This field indicates the Uncompressed Size of the respective
737 Block as bytes. The value is stored using the encoding
738 described in Section 1.2.
741 4.4. Index Padding
743 This field MUST contain 0-3 null bytes to pad the Index to
744 a multiple of four bytes. If any of the bytes are not null
745 bytes, the decoder MUST indicate an error.
748 4.5. CRC32
750 The CRC32 is calculated over everything in the Index field
751 except the CRC32 field itself. The CRC32 is stored as an
752 unsigned 32-bit little endian integer. If the calculated
753 value does not match the stored one, the decoder MUST indicate
754 an error.
757 5. Filter Chains
759 The Block Flags field defines how many filters are used. When
760 more than one filter is used, the filters are chained; that is,
761 the output of one filter is the input of another filter. The
762 following figure illustrates the direction of data flow.
764 v Uncompressed Data ^
765 | Filter 0 |
766 Encoder | Filter 1 | Decoder
767 | Filter n |
768 v Compressed Data ^
771 5.1. Alignment
773 Alignment of uncompressed input data is usually the job of
774 the application producing the data. For example, to get the
775 best results, an archiver tool should make sure that all
776 PowerPC executable files in the archive stream start at
777 offsets that are multiples of four bytes.
779 Some filters, for example LZMA2, can be configured to take
780 advantage of specified alignment of input data. Note that
781 taking advantage of aligned input can be beneficial also when
782 a filter is not the first filter in the chain. For example,
783 if you compress PowerPC executables, you may want to use the
784 PowerPC filter and chain that with the LZMA2 filter. Because
785 not only the input but also the output alignment of the PowerPC
786 filter is four bytes, it is now beneficial to set LZMA2
787 settings so that the LZMA2 encoder can take advantage of its
788 four-byte-aligned input data.
790 The output of the last filter in the chain is stored to the
791 Compressed Data field, which is is guaranteed to be aligned
792 to a multiple of four bytes relative to the beginning of the
793 Stream. This can increase
794 - speed, if the filtered data is handled multiple bytes at
795 a time by the filter-specific encoder and decoder,
796 because accessing aligned data in computer memory is
797 usually faster; and
798 - compression ratio, if the output data is later compressed
799 with an external compression tool.
802 5.2. Security
804 If filters would be allowed to be chained freely, it would be
805 possible to create malicious files, that would be very slow to
806 decode. Such files could be used to create denial of service
807 attacks.
809 Slow files could occur when multiple filters are chained:
811 v Compressed input data
812 | Filter 1 decoder (last filter)
813 | Filter 0 decoder (non-last filter)
814 v Uncompressed output data
816 The decoder of the last filter in the chain produces a lot of
817 output from little input. Another filter in the chain takes the
818 output of the last filter, and produces very little output
819 while consuming a lot of input. As a result, a lot of data is
820 moved inside the filter chain, but the filter chain as a whole
821 gets very little work done.
823 To prevent this kind of slow files, there are restrictions on
824 how the filters can be chained. These restrictions MUST be
825 taken into account when designing new filters.
827 The maximum number of filters in the chain has been limited to
828 four, thus there can be at maximum of three non-last filters.
829 Of these three non-last filters, only two are allowed to change
830 the size of the data.
832 The non-last filters, that change the size of the data, MUST
833 have a limit how much the decoder can compress the data: the
834 decoder SHOULD produce at least n bytes of output when the
835 filter is given 2n bytes of input. This limit is not
836 absolute, but significant deviations MUST be avoided.
838 The above limitations guarantee that if the last filter in the
839 chain produces 4n bytes of output, the chain as a whole will
840 produce at least n bytes of output.
843 5.3. Filters
845 5.3.1. LZMA2
847 LZMA (Lempel-Ziv-Markov chain-Algorithm) is a general-purpose
848 compression algorithm with high compression ratio and fast
849 decompression. LZMA is based on LZ77 and range coding
850 algorithms.
852 LZMA2 is an extension on top of the original LZMA. LZMA2 uses
853 LZMA internally, but adds support for flushing the encoder,
854 uncompressed chunks, eases stateful decoder implementations,
855 and improves support for multithreading. Thus, the plain LZMA
856 will not be supported in this file format.
858 Filter ID: 0x21
859 Size of Filter Properties: 1 byte
860 Changes size of data: Yes
861 Allow as a non-last filter: No
862 Allow as the last filter: Yes
864 Preferred alignment:
865 Input data: Adjustable to 1/2/4/8/16 byte(s)
866 Output data: 1 byte
868 The format of the one-byte Filter Properties field is as
869 follows:
871 Bits Mask Description
872 0-5 0x3F Dictionary Size
873 6-7 0xC0 Reserved for future use; MUST be zero for now.
875 Dictionary Size is encoded with one-bit mantissa and five-bit
876 exponent. The smallest dictionary size is 4 KiB and the biggest
877 is 4 GiB.
879 Raw value Mantissa Exponent Dictionary size
880 0 2 11 4 KiB
881 1 3 11 6 KiB
882 2 2 12 8 KiB
883 3 3 12 12 KiB
884 4 2 13 16 KiB
885 5 3 13 24 KiB
886 6 2 14 32 KiB
887 ... ... ... ...
888 35 3 27 768 MiB
889 36 2 28 1024 MiB
890 37 3 29 1536 MiB
891 38 2 30 2048 MiB
892 39 3 30 3072 MiB
893 40 2 31 4096 MiB - 1 B
895 Instead of having a table in the decoder, the dictionary size
896 can be decoded using the following C code:
898 const uint8_t bits = get_dictionary_flags() & 0x3F;
899 if (bits > 40)
900 return DICTIONARY_TOO_BIG; // Bigger than 4 GiB
902 uint32_t dictionary_size;
903 if (bits == 40) {
904 dictionary_size = UINT32_MAX;
905 } else {
906 dictionary_size = 2 | (bits & 1);
907 dictionary_size <<= bits / 2 + 11;
908 }
911 5.3.2. Branch/Call/Jump Filters for Executables
913 These filters convert relative branch, call, and jump
914 instructions to their absolute counterparts in executable
915 files. This conversion increases redundancy and thus
916 compression ratio.
918 Size of Filter Properties: 0 or 4 bytes
919 Changes size of data: No
920 Allow as a non-last filter: Yes
921 Allow as the last filter: No
923 Below is the list of filters in this category. The alignment
924 is the same for both input and output data.
926 Filter ID Alignment Description
927 0x04 1 byte x86 filter (BCJ)
928 0x05 4 bytes PowerPC (big endian) filter
929 0x06 16 bytes IA64 filter
930 0x07 4 bytes ARM filter [1]
931 0x08 2 bytes ARM Thumb filter [1]
932 0x09 4 bytes SPARC filter
933 0x0A 4 bytes ARM64 filter [2]
934 0x0B 2 bytes RISC-V filter
936 [1] These are for little endian instruction encoding.
937 This must not be confused with data endianness.
938 A processor configured for big endian data access
939 may still use little endian instruction encoding.
940 The filters don't care about the data endianness.
942 [2] 4096-byte alignment gives the best results
943 because the address in the ADRP instruction
944 is a multiple of 4096 bytes.
946 If the size of Filter Properties is four bytes, the Filter
947 Properties field contains the start offset used for address
948 conversions. It is stored as an unsigned 32-bit little endian
949 integer. The start offset MUST be a multiple of the alignment
950 of the filter as listed in the table above; if it isn't, the
951 decoder MUST indicate an error. If the size of Filter
952 Properties is zero, the start offset is zero.
954 Setting the start offset may be useful if an executable has
955 multiple sections, and there are many cross-section calls.
956 Taking advantage of this feature usually requires usage of
957 the Subblock filter, whose design is not complete yet.
960 5.3.3. Delta
962 The Delta filter may increase compression ratio when the value
963 of the next byte correlates with the value of an earlier byte
964 at specified distance.
966 Filter ID: 0x03
967 Size of Filter Properties: 1 byte
968 Changes size of data: No
969 Allow as a non-last filter: Yes
970 Allow as the last filter: No
972 Preferred alignment:
973 Input data: 1 byte
974 Output data: Same as the original input data
976 The Properties byte indicates the delta distance, which can be
977 1-256 bytes backwards from the current byte: 0x00 indicates
978 distance of 1 byte and 0xFF distance of 256 bytes.
981 5.3.3.1. Format of the Encoded Output
983 The code below illustrates both encoding and decoding with
984 the Delta filter.
986 // Distance is in the range [1, 256].
987 const unsigned int distance = get_properties_byte() + 1;
988 uint8_t pos = 0;
989 uint8_t delta[256];
991 memset(delta, 0, sizeof(delta));
993 while (1) {
994 const int byte = read_byte();
995 if (byte == EOF)
996 break;
998 uint8_t tmp = delta[(uint8_t)(distance + pos)];
999 if (is_encoder) {
1000 tmp = (uint8_t)(byte) - tmp;
1001 delta[pos] = (uint8_t)(byte);
1002 } else {
1003 tmp = (uint8_t)(byte) + tmp;
1004 delta[pos] = tmp;
1005 }
1007 write_byte(tmp);
1008 --pos;
1009 }
1012 5.4. Custom Filter IDs
1014 If a developer wants to use custom Filter IDs, there are two
1015 choices. The first choice is to contact Lasse Collin and ask
1016 him to allocate a range of IDs for the developer.
1018 The second choice is to generate a 40-bit random integer
1019 which the developer can use as a personal Developer ID.
1020 To minimize the risk of collisions, Developer ID has to be
1021 a randomly generated integer, not manually selected "hex word".
1022 The following command, which works on many free operating
1023 systems, can be used to generate Developer ID:
1025 dd if=/dev/urandom bs=5 count=1 | hexdump
1027 The developer can then use the Developer ID to create unique
1028 (well, hopefully unique) Filter IDs.
1030 Bits Mask Description
1031 0-15 0x0000_0000_0000_FFFF Filter ID
1032 16-55 0x00FF_FFFF_FFFF_0000 Developer ID
1033 56-62 0x3F00_0000_0000_0000 Static prefix: 0x3F
1035 The resulting 63-bit integer will use 9 bytes of space when
1036 stored using the encoding described in Section 1.2. To get
1037 a shorter ID, see the beginning of this Section how to
1038 request a custom ID range.
1041 5.4.1. Reserved Custom Filter ID Ranges
1043 Range Description
1044 0x0000_0300 - 0x0000_04FF Reserved to ease .7z compatibility
1045 0x0002_0000 - 0x0007_FFFF Reserved to ease .7z compatibility
1046 0x0200_0000 - 0x07FF_FFFF Reserved to ease .7z compatibility
1049 6. Cyclic Redundancy Checks
1051 There are several incompatible variations to calculate CRC32
1052 and CRC64. For simplicity and clarity, complete examples are
1053 provided to calculate the checks as they are used in this file
1054 format. Implementations MAY use different code as long as it
1055 gives identical results.
1057 The program below reads data from standard input, calculates
1058 the CRC32 and CRC64 values, and prints the calculated values
1059 as big endian hexadecimal strings to standard output.
1061 #include <stddef.h>
1062 #include <inttypes.h>
1063 #include <stdio.h>
1065 uint32_t crc32_table[256];
1066 uint64_t crc64_table[256];
1068 void
1069 init(void)
1070 {
1071 static const uint32_t poly32 = UINT32_C(0xEDB88320);
1072 static const uint64_t poly64
1073 = UINT64_C(0xC96C5795D7870F42);
1075 for (size_t i = 0; i < 256; ++i) {
1076 uint32_t crc32 = i;
1077 uint64_t crc64 = i;
1079 for (size_t j = 0; j < 8; ++j) {
1080 if (crc32 & 1)
1081 crc32 = (crc32 >> 1) ^ poly32;
1082 else
1083 crc32 >>= 1;
1085 if (crc64 & 1)
1086 crc64 = (crc64 >> 1) ^ poly64;
1087 else
1088 crc64 >>= 1;
1089 }
1091 crc32_table[i] = crc32;
1092 crc64_table[i] = crc64;
1093 }
1094 }
1096 uint32_t
1097 crc32(const uint8_t *buf, size_t size, uint32_t crc)
1098 {
1099 crc = ~crc;
1100 for (size_t i = 0; i < size; ++i)
1101 crc = crc32_table[buf[i] ^ (crc & 0xFF)]
1102 ^ (crc >> 8);
1103 return ~crc;
1104 }
1106 uint64_t
1107 crc64(const uint8_t *buf, size_t size, uint64_t crc)
1108 {
1109 crc = ~crc;
1110 for (size_t i = 0; i < size; ++i)
1111 crc = crc64_table[buf[i] ^ (crc & 0xFF)]
1112 ^ (crc >> 8);
1113 return ~crc;
1114 }
1116 int
1117 main()
1118 {
1119 init();
1121 uint32_t value32 = 0;
1122 uint64_t value64 = 0;
1123 uint64_t total_size = 0;
1124 uint8_t buf[8192];
1126 while (1) {
1127 const size_t buf_size
1128 = fread(buf, 1, sizeof(buf), stdin);
1129 if (buf_size == 0)
1130 break;
1132 total_size += buf_size;
1133 value32 = crc32(buf, buf_size, value32);
1134 value64 = crc64(buf, buf_size, value64);
1135 }
1137 printf("Bytes: %" PRIu64 "\n", total_size);
1138 printf("CRC-32: 0x%08" PRIX32 "\n", value32);
1139 printf("CRC-64: 0x%016" PRIX64 "\n", value64);
1141 return 0;
1142 }
1145 7. References
1147 LZMA SDK - The original LZMA implementation
1148 https://7-zip.org/sdk.html
1150 LZMA Utils - LZMA adapted to POSIX-like systems
1151 https://tukaani.org/lzma/
1153 XZ Utils - The next generation of LZMA Utils
1154 https://tukaani.org/xz/
1156 [RFC-1952]
1157 GZIP file format specification version 4.3
1158 https://www.ietf.org/rfc/rfc1952.txt
1159 - Notation of byte boxes in section "2.1. Overall conventions"
1161 [RFC-2119]
1162 Key words for use in RFCs to Indicate Requirement Levels
1163 https://www.ietf.org/rfc/rfc2119.txt
1165 [GNU-tar]
1166 GNU tar 1.35 manual
1167 https://www.gnu.org/software/tar/manual/html_node/Blocking-Factor.html
1168 - Node 9.4.2 "Blocking Factor", paragraph that begins
1169 "gzip will complain about trailing garbage"
1170 - Note that this URL points to the latest version of the
1171 manual, and may some day not contain the note which is in
1172 1.35. For the exact version of the manual, download GNU
1173 tar 1.35: ftp://ftp.gnu.org/pub/gnu/tar/tar-1.35.tar.gz