This manual is for Tarlz (version 0.14, 12 March 2019).
Copyright © 2013-2019 Antonio Diaz Diaz.
This manual is free documentation: you have unlimited permission to copy, distribute and modify it.
Tarlz is a massively parallel (multi-threaded) combined implementation of the tar archiver and the lzip compressor. Tarlz creates, lists and extracts archives in a simplified posix pax format compressed with lzip, keeping the alignment between tar members and lzip members. This method adds an indexed lzip layer on top of the tar archive, making it possible to decode the archive safely in parallel. The resulting multimember tar.lz archive is fully backward compatible with standard tar tools like GNU tar, which treat it like any other tar.lz archive. Tarlz can append files to the end of such compressed archives.
Tarlz can create tar archives with five levels of compression granularity; per file (--no-solid), per block (--bsolid, default), per directory (--dsolid), appendable solid (--asolid), and solid (--solid).
Of course, compressing each file (or each directory) individually can't achieve a compression ratio as high as compressing solidly the whole tar archive, but it has the following advantages:
Tarlz protects the extended records with a CRC in a way compatible with standard tar tools. See crc32.
Tarlz does not understand other tar formats like 'gnu', 'oldgnu', 'star' or 'v7'.
The format for running tarlz is:
tarlz [options] [files]
On archive creation or appending tarlz archives the files specified, but removes from member names any leading and trailing slashes and any filename prefixes containing a '..' component. On extraction, leading and trailing slashes are also removed from member names, and archive members containing a '..' component in the filename are skipped. Tarlz detects when the archive being created or enlarged is among the files to be dumped, appended or concatenated, and skips it.
On extraction and listing, tarlz removes leading './' strings from member names in the archive or given in the command line, so that 'tarlz -xf foo ./bar baz' extracts members 'bar' and './baz' from archive 'foo'.
If several compression levels or '--*solid' options are given, the last setting is used. For example '-9 --solid --uncompressed -1' is equivalent to '-1 --solid'
tarlz supports the following options:
Note that a process can only have one current working directory (CWD).
Therefore multi-threading can't be used to create an archive if a '-C'
option appears after a relative filename in the command line.
As tarlz removes leading slashes from member names, the '-C' option may
be used in combination with '--diff' when absolute filenames were used
on archive creation: 'tarlz -C / -d'. Alternatively, tarlz may be
run from the root directory to perform the comparison.
Note that the number of usable threads is limited during compression to
ceil( uncompressed_size / data_size ) (see Minimum archive sizes),
and during decompression to the number of lzip members in the tar.lz
archive, which you can find by running 'lzip -lv archive.tar.lz'.
-0 .. -9
|Level||Dictionary size||Match length limit
|-0||64 KiB||16 bytes
|-1||1 MiB||5 bytes
|-2||1.5 MiB||6 bytes
|-3||2 MiB||8 bytes
|-4||3 MiB||12 bytes
|-5||4 MiB||20 bytes
|-6||8 MiB||36 bytes
|-7||16 MiB||68 bytes
|-8||24 MiB||132 bytes
|-9||32 MiB||273 bytes
tar='tarlz --solid'. See --data-size, to set the target block size.
Exit status: 0 for a normal exit, 1 for environmental problems (file not found, invalid flags, I/O errors, etc), 2 to indicate a corrupt or invalid input file, 3 for an internal consistency error (eg, bug) which caused tarlz to panic.
In the diagram below, a box like this:
+---+ | | <-- the vertical bars might be missing +---+
represents one byte; a box like this:
+==============+ | | +==============+
represents a variable number of bytes or a fixed but large number of bytes (for example 512).
A tar.lz file consists of a series of lzip members (compressed data sets). The members simply appear one after another in the file, with no additional information before, between, or after them.
Each lzip member contains one or more tar members in a simplified posix pax interchange format. The only pax typeflag value supported by tarlz (in addition to the typeflag values defined by the ustar format) is 'x'. The pax format is an extension on top of the ustar format that removes the size limitations of the ustar format.
Each tar member contains one file archived, and is represented by the following sequence:
Each tar member must be contiguously stored in a lzip member for the parallel decoding operations like '--list' to work. If any tar member is split over two or more lzip members, the archive must be decoded sequentially. See Multi-threaded tar.
At the end of the archive file there are two 512-byte blocks filled with binary zeros, interpreted as an end-of-archive indicator. These EOF blocks are either compressed in a separate lzip member or compressed along with the tar members contained in the last lzip member.
The diagram below shows the correspondence between each tar member (formed by one or two headers plus optional data) in the tar archive and each lzip member in the resulting multimember tar.lz archive, when per file compression is used:
tar +========+======+=================+===============+========+======+========+ | header | data | extended header | extended data | header | data | EOF | +========+======+=================+===============+========+======+========+ tar.lz +===============+=================================================+========+ | member | member | member | +===============+=================================================+========+
The pax header block is identical to the ustar header block described below except that the typeflag has the value 'x' (extended). The size field is the size of the extended header data in bytes. Most other fields in the pax header block are zeroed on archive creation to prevent trouble if the archive is read by an ustar tool, and are ignored by tarlz on archive extraction. See flawed-compat.
The pax extended header data consists of one or more records, each of
them constructed as follows:
"%d %s=%s\n", <length>, <keyword>, <value>
The <length>, <blank>, <keyword>, <equals-sign>, and <newline> in the record must be limited to the portable character set. The <length> field contains the decimal length of the record in bytes, including the trailing <newline>. The <value> field is stored as-is, without conversion to UTF-8 nor any other transformation.
These are the <keyword> fields currently supported by tarlz:
The ustar header block has a length of 512 bytes and is structured as shown in the following table. All lengths and offsets are in decimal.
|Field Name||Offset||Length (in bytes)
All characters in the header block are coded using the ISO/IEC 646:1991 (ASCII) standard, except in fields storing names for files, users, and groups. For maximum portability between implementations, names should only contain characters from the portable filename character set. But if an implementation supports the use of characters outside of '/' and the portable filename character set in names for files, users, and groups, tarlz will use the byte values in these names unmodified.
The fields name, linkname, and prefix are null-terminated character strings except when all characters in the array contain non-null characters including the last character.
The name and the prefix fields produce the pathname of the file. A new pathname is formed, if prefix is not an empty string (its first character is not null), by concatenating prefix (up to the first null character), a <slash> character, and name; otherwise, name is used alone. In either case, name is terminated at the first null character. If prefix begins with a null character, it is ignored. In this manner, pathnames of at most 256 characters can be supported. If a pathname does not fit in the space provided, an extended record is used to store the pathname.
The linkname field does not use the prefix to produce a pathname. If the linkname does not fit in the 100 characters provided, an extended record is used to store the linkname.
The mode field provides 12 access permission bits. The following table shows the symbolic name of each bit and its octal value:
|Bit Name||Value||Bit Name||Value||Bit Name||Value
The uid and gid fields are the user and group ID of the owner and group of the file, respectively.
The size field contains the octal representation of the size of the file in bytes. If the typeflag field specifies a file of type '0' (regular file) or '7' (high performance regular file), the number of logical records following the header is (size / 512) rounded to the next integer. For all other values of typeflag, tarlz either sets the size field to 0 or ignores it, and does not store or expect any logical records following the header. If the file size is larger than 8_589_934_591 bytes (octal 77777777777), an extended record is used to store the file size.
The mtime field contains the octal representation of the modification time of the file at the time it was archived, obtained from the stat() function.
The chksum field contains the octal representation of the value of the simple sum of all bytes in the header logical record. Each byte in the header is treated as an unsigned value. When calculating the checksum, the chksum field is treated as if it were all <space> characters.
The typeflag field contains a single character specifying the type of file archived:
The magic field contains the ASCII null-terminated string "ustar". The version field contains the characters "00" (0x30,0x30). The fields uname, and gname are null-terminated character strings except when all characters in the array contain non-null characters including the last character. Each numeric field contains a leading space- or zero-filled, optionally null-terminated octal number using digits from the ISO/IEC 646:1991 (ASCII) standard. Tarlz is able to decode numeric fields 1 byte larger than standard ustar by not requiring a terminating null character.
Tarlz is meant to reliably detect invalid or corrupt metadata during extraction and to not create safety risks in the archives it creates. In order to achieve these goals, tarlz makes some changes to the variant of the pax format that it uses. This chapter describes these changes and the concrete reasons to implement them.
The posix pax format has a serious flaw. The metadata stored in pax extended records are not protected by any kind of check sequence. Corruption in a long filename may cause the extraction of the file in the wrong place without warning. Corruption in a large file size may cause the truncation of the file or the appending of garbage to the file, both followed by a spurious warning about a corrupt header far from the place of the undetected corruption.
Metadata like filename and file size must be always protected in an archive format because of the adverse effects of undetected corruption in them, potentially much worse that undetected corruption in the data. Even more so in the case of pax because the amount of metadata it stores is potentially large, making undetected corruption more probable.
Because of the above, tarlz protects the extended records with a CRC in a way compatible with standard tar tools. See key_crc32.
In order to allow the extraction of pax archives by a tar utility conforming to the POSIX-2:1993 standard, POSIX.1-2008 recommends selecting extended header field values that allow such tar to create a regular file containing the extended header records as data. This approach is broken because if the extended header is needed because of a long filename, the name and prefix fields will be unable to contain the full pathname of the file. Therefore the files corresponding to both the extended header and the overridden ustar header will be extracted using truncated filenames, perhaps overwriting existing files or directories. It may be a security risk to extract a file with a truncated filename.
To avoid this problem, tarlz writes extended headers with all fields zeroed except size, chksum, typeflag, magic and version. This prevents old tar programs from extracting the extended records as a file in the wrong place. Tarlz also sets to zero those fields of the ustar header overridden by extended records.
If an extended header is required for any reason (for example a file size larger than 8 GiB or a link name longer than 100 bytes), tarlz moves the filename also to the extended header to prevent an ustar tool from trying to extract the file or link. This also makes easier during parallel extraction or listing the detection of a tar member split between two lzip members at the boundary between the extended header and the ustar header.
The tarlz format is mainly ustar. Extended pax headers are used only when needed because the length of a filename or link name, or the size of a file exceed the limits of the ustar format. Adding extended headers to each member just to record subsecond timestamps seems wasteful for a backup format.
There is no portable way to tell what charset a text string is coded into. Therefore, tarlz stores all fields representing text strings unmodified, without conversion to UTF-8 nor any other transformation. This prevents accidental double UTF-8 conversions. If the need arises this behavior will be adjusted with a command line option in the future.
Safely decoding an arbitrary tar archive in parallel is impossible. For example, if a tar archive containing another tar archive is decoded starting from some position other than the beginning, there is no way to know if the first header found there belongs to the outer tar archive or to the inner tar archive. Tar is a format inherently serial; it was designed for tapes.
In the case of compressed tar archives, the start of each compressed block determines one point through which the tar archive can be decoded in parallel. Therefore, in tar.lz archives the decoding operations can't be parallelized if the tar members are not aligned with the lzip members. Tar archives compressed with plzip can't be decoded in parallel because tar and plzip do not have a way to align both sets of members. Certainly one can decompress one such archive with a multi-threaded tool like plzip, but the increase in speed is not as large as it could be because plzip must serialize the decompressed data and pass them to tar, which decodes them sequentially, one tar member at a time.
On the other hand, if the tar.lz archive is created with a tool like tarlz, which can guarantee the alignment between tar members and lzip members because it controls both archiving and compression, then the lzip format becomes an indexed layer on top of the tar archive which makes possible decoding it safely in parallel.
Tarlz is able to automatically decode aligned and unaligned multimember tar.lz archives, keeping backwards compatibility. If tarlz finds a member misalignment during multi-threaded decoding, it switches to single-threaded mode and continues decoding the archive. Currently only the '--list' option is able to do multi-threaded decoding.
If the files in the archive are large, multi-threaded '--list' on a regular (seekable) tar.lz archive can be hundreds of times faster than sequential '--list' because, in addition to using several processors, it only needs to decompress part of each lzip member. See the following example listing the Silesia corpus on a dual core machine:
tarlz -9 --no-solid -cf silesia.tar.lz silesia time lzip -cd silesia.tar.lz | tar -tf - (5.032s) time plzip -cd silesia.tar.lz | tar -tf - (3.256s) time tarlz -tf silesia.tar.lz (0.020s)
When creating or appending to a compressed archive using multi-threaded block compression, tarlz puts tar members together in blocks and compresses as many blocks simultaneously as worker threads are chosen, creating a multimember compressed archive.
For this to work as expected (and roughly multiply the compression speed by the number of available processors), the uncompressed archive must be at least as large as the number of worker threads times the block size (see --data-size). Else some processors will not get any data to compress, and compression will be proportionally slower. The maximum speed increase achievable on a given archive is limited by the ratio (uncompressed_size / data_size). For example, a tarball the size of gcc or linux will scale up to 10 or 12 processors at level -9.
The following table shows the minimum uncompressed archive size needed for full use of N processors at a given compression level, using the default data size for each level:
|-0||2 MiB||4 MiB||8 MiB||16 MiB||64 MiB||256 MiB
|-1||4 MiB||8 MiB||16 MiB||32 MiB||128 MiB||512 MiB
|-2||6 MiB||12 MiB||24 MiB||48 MiB||192 MiB||768 MiB
|-3||8 MiB||16 MiB||32 MiB||64 MiB||256 MiB||1 GiB
|-4||12 MiB||24 MiB||48 MiB||96 MiB||384 MiB||1.5 GiB
|-5||16 MiB||32 MiB||64 MiB||128 MiB||512 MiB||2 GiB
|-6||32 MiB||64 MiB||128 MiB||256 MiB||1 GiB||4 GiB
|-7||64 MiB||128 MiB||256 MiB||512 MiB||2 GiB||8 GiB
|-8||96 MiB||192 MiB||384 MiB||768 MiB||3 GiB||12 GiB
|-9||128 MiB||256 MiB||512 MiB||1 GiB||4 GiB||16 GiB
Example 1: Create a multimember compressed archive 'archive.tar.lz' containing files 'a', 'b' and 'c'.
tarlz -cf archive.tar.lz a b c
Example 2: Append files 'd' and 'e' to the multimember compressed archive 'archive.tar.lz'.
tarlz -rf archive.tar.lz d e
Example 3: Create a solidly compressed appendable archive 'archive.tar.lz' containing files 'a', 'b' and 'c'. Then append files 'd' and 'e' to the archive.
tarlz --asolid -cf archive.tar.lz a b c tarlz --asolid -rf archive.tar.lz d e
Example 4: Create a compressed appendable archive containing directories 'dir1', 'dir2' and 'dir3' with a separate lzip member per directory. Then append files 'a', 'b', 'c', 'd' and 'e' to the archive, all of them contained in a single lzip member. The resulting archive 'archive.tar.lz' contains 5 lzip members (including the EOF member).
tarlz --dsolid -cf archive.tar.lz dir1 dir2 dir3 tarlz --asolid -rf archive.tar.lz a b c d e
Example 5: Create a solidly compressed archive 'archive.tar.lz' containing files 'a', 'b' and 'c'. Note that no more files can be later appended to the archive.
tarlz --solid -cf archive.tar.lz a b c
Example 6: Extract all files from archive 'archive.tar.lz'.
tarlz -xf archive.tar.lz
Example 7: Extract files 'a' and 'c' from archive 'archive.tar.lz'.
tarlz -xf archive.tar.lz a c
Example 8: Copy the contents of directory 'sourcedir' to the directory 'destdir'.
tarlz -C sourcedir -c . | tarlz -C destdir -x
There are probably bugs in tarlz. There are certainly errors and omissions in this manual. If you report them, they will get fixed. If you don't, no one will ever know about them and they will remain unfixed for all eternity, if not longer.
If you find a bug in tarlz, please send electronic mail to firstname.lastname@example.org. Include the version number, which you can find by running 'tarlz --version'.