ziplimit.txt
A1) Hard limits of the Zip archive format (without Zip64 extensions):
Number of entries in Zip archive: 64 Ki (2^16 - 1 entries)
Compressed size of archive entry: 4 GiByte (2^32 - 1 Bytes)
Uncompressed size of entry: 4 GiByte (2^32 - 1 Bytes)
Size of single-volume Zip archive: 4 GiByte (2^32 - 1 Bytes)
Per-volume size of multi-volume archives: 4 GiByte (2^32 - 1 Bytes)
Number of parts for multi-volume archives: 64 Ki (2^16 - 1 parts)
Total size of multi-volume archive: 256 TiByte (4G * 64k)
The number of archive entries and of multivolume parts are limited by
the structure of the "end-of-central-directory" record, where the these
numbers are stored in 2-Byte fields.
Some Zip and/or UnZip implementations (for example Info-ZIP's) allow
handling of archives with more than 64k entries. (The information
from "number of entries" field in the "end-of-central-directory" record
is not really neccessary to retrieve the contents of a Zip archive;
it should rather be used for consistency checks.)
Length of an archive entry name: 64 KiByte (2^16 - 1)
Length of archive member comment: 64 KiByte (2^16 - 1)
Total length of "extra field": 64 KiByte (2^16 - 1)
Length of a single e.f. block: 64 KiByte (2^16 - 1)
Length of archive comment: 64 KiByte (2^16 - 1)
Additional limitation claimed by PKWARE:
Size of local-header structure (fixed fields of 30 Bytes + filename
local extra field): < 64 KiByte
Size of central-directory structure (46 Bytes + filename +
central extra field + member comment): < 64 KiByte
A2) Hard limits of the Zip archive format with Zip64 extensions:
In 2001, PKWARE has published version 4.5 of the Zip format specification
(together with the release of PKZIP for Windows 4.5). This specification
defines new extra field blocks that allow to break the size limits of the
standard zipfile structures. This extended "Zip64" format enlarges the
theoretical limits to the following values:
Number of entries in Zip archive: 16 Ei (2^64 - 1 entries)
Compressed size of archive entry: 16 EiByte (2^64 - 1 Bytes)
Uncompressed size of entry: 16 EiByte (2^64 - 1 Bytes)
Size of single-volume Zip archive: 16 EiByte (2^64 - 1 Bytes)
Per-volume size of multi-volume archives: 16 EiByte (2^64 - 1 Bytes)
Number of parts for multi-volume archives: 4 Gi (2^32 - 1 parts)
Total size of multi-volume archive: 2^96 Byte (16 Ei * 4Gi)
The Info-ZIP software releases (beginning with Zip 3.0 and UnZip 6.0)
support Zip64 archives on selected environments (where the underlying
operating system capabilities are sufficient, e.g. Unix, VMS and Win32).
B) Implementation limits of UnZip:
1. Size limits caused by file I/O and decompression handling:
a) Without "Zip64" and "LargeFile" extensions:
Size of Zip archive: 2 GiByte (2^31 - 1 Bytes)
Compressed size of archive entry: 2 GiByte (2^31 - 1 Bytes)
b) With "Zip64" enabled and "LargeFile" supported:
Size of Zip archive: 8 EiByte (2^63 - 1 Bytes)
Compressed size of archive entry: 8 EiByte (2^63 - 1 Bytes)
Uncompressed size of entry: 8 EiByte (2^63 - 1 Bytes)
Note: On some systems, even UnZip without "LargeFile" extensions enabled
may support archive sizes up to 4 GiByte. To get this support, the
target environment has to meet the following requirements:
a) The compiler's intrinsic "long" data types must be able to hold
integer numbers of 2^32. In other words - the standard intrinsic
integer types "long" and "unsigned long" have to be wider than
32 bit.
b) The system has to supply a C runtime library that is compatible
with the more-than-32-bit-wide "long int" type of condition a)
c) The standard file positioning functions fseek(), ftell() (and/or
the Unix style lseek() and tell() functions) have to be capable
to move to absolute file offsets of up to 4 GiByte from the file
start.
On 32-bit CPU hardware, you generally cannot expect that a C compiler
provides a "long int" type that is wider than 32-bit. So, many of the
most popular systems (i386, PowerPC, 680x0, et. al) are out of luck.
You may find environment that provide all requirements on systems
with 64-bit CPU hardware. Examples might be Cray number crunchers,
Compaq (former DEC) Alpha AXP machines, or Intel/AMD x64 computers.
The number of Zip archive entries is unlimited. The "number-of-entries"
field of the "end-of-central-dir" record is checked against the "number
of entries found in the central directory" modulus 64k (2^16) (without
Zip64 extension) or modulus 2^64 (with Zip64 extensions enabled for
Zip64 archives).
Multi-volume archive extraction is not (yet) supported.
Memory requirements are mostly independent of the archive size
and archive contents.
In general, UnZip needs a fixed amount of internal buffer space
plus the size to hold the complete information of the currently
processed entry's local header. Here, a large extra field
(could be up to 64 kByte) may exceed the available memory
for MSDOS 16-bit executables (when they were compiled in small
or medium memory model, with a fixed 64 KiByte limit on data space).
The other exception where memory requirements scale with "larger"
archives is the "restore directory attributes" feature. Here, the
directory attributes info for each restored directory has to be held
in memory until the whole archive has been processed. So, the amount
of memory needed to keep this info scales with the number of restored
directories and may cause memory problems when a lot of directories
are restored in a single run.
C) Implementation limits of the Zip executables:
1. Size limits caused by file I/O and compression handling:
a) Without "Zip64" and "LargeFile" extensions:
Size of Zip archive: 2 GiByte (2^31 - 1 Bytes)
Compressed size of archive entry: 2 GiByte (2^31 - 1 Bytes)
Uncompressed size of entry: 2 GiByte (2^31 - 1 Bytes),
(could/should be 4 GiBytes...)
b) With "Zip64" enabled and "LargeFile" supported:
Size of Zip archive: 8 EiByte (2^63 - 1 Bytes)
Compressed size of archive entry: 8 EiByte (2^63 - 1 Bytes)
Uncompressed size of entry: 8 EiByte (2^63 - 1 Bytes)
Multi-volume archive creation now supported in the form of split
archives. Currently up to 99,999 splits are supported.
2. Limits caused by handling of archive contents lists
2.1. Number of archive entries (freshen, update, delete)
a) 16-bit executable: 64k (2^16 -1) or 32k (2^15 - 1),
(unsigned vs. signed type of size_t)
a1) 16-bit executable: <16k ((2^16)/4)
(The smaller limit a1) results from the array size limit of
the "qsort()" function.)
32-bit executable: <1G ((2^32)/4)
(usual system limit of the "qsort()" function on 32-bit systems)
64-bit executable: <2Ei ((2^64)/8)
(theoretical limit of 64-bit flat memory model, the actual limit of
currently available OS implementations is several orders of magnitude
lower)
b) stack space needed by qsort to sort list of archive entries
NOTE: In the current executables, overflows of limits a) and b) are NOT
checked!
c) amount of free memory to hold "central directory information" of
all archive entries; one entry needs:
128 bytes (Zip64), 96 bytes (32-bit) resp. 80 bytes (16-bit)
+ 3 * length of entry name
+ length of zip entry comment (when present)
+ length of extra field(s) (when present, e.g.: UT needs 9 bytes)
+ some bytes for book-keeping of memory allocation
Conclusion:
For systems with limited memory space (MSDOS, small AMIGAs, other
environments without virtual memory), the number of archive entries
is most often limited by condition c).
For example, with approx. 100 kBytes of free memory after loading and
initializing the program, a 16-bit DOS Zip cannot process more than 600
to 1000 (+) archive entries. (For the 16-bit Windows DLL or the 16-bit
OS/2 port, limit c) is less important because Windows or OS/2 executables
are not restricted to the 1024k area of real mode memory. These 16-bit
ports are limited by conditions a1) and b), say: at maximum approx.
16000 entries!)
2.2. Number of "new" entries (add operation)
In addition to the restrictions above (2.1.), the following limits
caused by the handling of the "new files" list apply:
a) 16-bit executable: <16k ((2^64)/4)
b) stack size required for "qsort" operation on "new entries" list.
NOTE: In the current executables, the overflow checks for these limits
are missing!
c) amount of free memory to hold the directory info list for new entries;
one entry needs:
32 bytes (Zip64), 24 bytes (32-bit) resp. 22 bytes (16-bit)
+ 3 * length of filename
NOTE: For larger systems, the actual usability limits may be more
performance issues (how long you want to wait) rather than available
memory and other resources.
D) Some technical remarks:
1. For executables without support for "Zip64" archives and "LargeFile"
I/O extensions, the 2GiByte size limit on archive files is a consequence
of the portable C implementation used for the Info-ZIP programs.
Zip archive processing requires random access to the archive file for
jumping between different parts of the archive's structure.
In standard C, this is done via stdio functions fseek()/ftell() resp.
unix-io functions lseek()/tell(). In many (most?) C implementations,
these functions use "signed long" variables to hold offset pointers
into sequential files. In most cases, this is a signed 32-bit number,
which is limited to ca. 2E+09. There may be specific C runtime library
implementations that interpret the offset numbers as unsigned, but for
us, this is not reliable in the context of portable programming.
2. Similarly, for executables without "Zip64" and "LargeFile" support,
the 2GiByte limit on the size of a single compressed archive member
is again a consequence of the implementation in C.
The variables used internally to count the size of the compressed
data stream are of type "long", which is guaranted to be at least
32-bit wide on all supported environments.
But, why do we use "signed" long and not "unsigned long"?
Throughout the I/O handling of the compressed data stream, the sign bit
of the "long" numbers is (mis-)used as a kind of overflow detection.
In the end, this is caused by the fact that standard C lacks any
overflow checking on integer arithmetics and does not support access
to the underlying hardware's overflow detection (the status bits,
especially "carry" and "overflow" of the CPU's flags-register) in a
system-independent manner.
So, we "misuse" the most-significant bit of the compressed data size
counters as carry bit for efficient overflow/underflow detection. We
could change the code to a different method of overflow detection, by
using a bunch of "sanity" comparisons (kind of "is the calculated result
plausible when compared with the operands"). But, this would "blow up"
the code of the "inner loop", with remarkable loss of processing speed.
Or, we could reduce the amount of consistency checks of the compressed
data (e.g. detection of premature end of stream) to an absolute minimum,
at the cost of the programs' stability when processing corrupted data.
3. The argumentation above is somewhat out-dated. Beginning with the
releases of Zip 3 and UnZip 6, Info-ZIP programs support archive
sizes larger than 4GiB on systems where the required underlying
support for 64-bit file offsets and file sizes is available from
the OS (and the C runtime environment).
For executables with support for "Zip64" archive format and "LargeFile"
extension, the I/O limits are lifted by applying extended 64-bit off_t
file offsets. All limits discussed above are then based on integer
sizes of 64 bits instead of 32, this should allow to handle file and
archive sizes up to the limits of manufacturable hardware for the
foreseeable future. The reduction of the theoretical limits from
(2^64 - 1) to (2^63 - 1) because of the throughout use of signed
numbers can be neglected with the currently imaginable hardware.
However, this new support partially breaks compatibility with older
"legacy" systems. And it should be noted that the portability and
readability of the UnZip and Zip code has suffered somehow caused
by the extensive use of non-standard language extension needed for
64-bit support on the major target systems.
Please report any problems to: Zip-Bugs at www.info-zip.org
Last updated: 25 May 2008, Ed Gordon
02 January 2009, Christian Spieler