Hi,

Plist files have actually three flavors : XML ones, binary ones, and now (starting from Mac OS X 10.7 Lion) json one. The plistlib.readPlist function can only read XML plist files and thus cannot read binary and json ones.

The binary format is open and described by Apple (http://opensource.apple.com/source/CF/CF-550/CFBinaryPList.c).

Here is the diff (from Python 2.7 implementation of plistlib) to transparently read both binary and json formats.

API of plistlib remains unchanged, since format detection is done by plistlib.readPlist. 
An InvalidFileException is raised in case of malformed binary file.


57,58c57
<     "Plist", "Data", "Dict",
<     "InvalidFileException",
---
>     "Plist", "Data", "Dict"
64d62
< import json
66d63
< import os
68d64
< import struct
81,89c77,78
<     header = pathOrFile.read(8)
<     pathOrFile.seek(0)
<     if header == '<?xml ve' or header[2:] == '<?xml ': #XML plist file, without or with BOM 
<         p = PlistParser()
<         rootObject = p.parse(pathOrFile)
<     elif header == 'bplist00': #binary plist file
<         rootObject =  readBinaryPlistFile(pathOrFile)
<     else: #json plist file
<         rootObject = json.load(pathOrFile)
---
>     p = PlistParser()
>     rootObject = p.parse(pathOrFile)
195,285d183
< 
< # timestamp 0 of binary plists corresponds to 1/1/2001 (year of Mac OS X 10.0), instead of 1/1/1970.
< MAC_OS_X_TIME_OFFSET = (31 * 365 + 8) * 86400
< 
< class InvalidFileException(ValueError):
<     def __str__(self):
<         return "Invalid file"
<     def __unicode__(self):
<         return "Invalid file"
< 
< def readBinaryPlistFile(in_file):
<     """
<     Read a binary plist file, following the description of the binary format: http://opensource.apple.com/source/CF/CF-550/CFBinaryPList.c
<     Raise InvalidFileException in case of error, otherwise return the root object, as usual
<     """
<     in_file.seek(-32, os.SEEK_END)
<     trailer = in_file.read(32)
<     if len(trailer) != 32:
<         return InvalidFileException()
<     offset_size, ref_size, num_objects, top_object, offset_table_offset = struct.unpack('>6xBB4xL4xL4xL', trailer)
<     in_file.seek(offset_table_offset)
<     object_offsets = []
<     offset_format = '>' + {1: 'B', 2: 'H', 4: 'L', 8: 'Q', }[offset_size] * num_objects
<     ref_format = {1: 'B', 2: 'H', 4: 'L', 8: 'Q', }[ref_size]
<     int_format = {0: (1, '>B'), 1: (2, '>H'), 2: (4, '>L'), 3: (8, '>Q'), }
<     object_offsets = struct.unpack(offset_format, in_file.read(offset_size * num_objects))
<     def getSize(token_l):
<         """ return the size of the next object."""
<         if token_l == 0xF:
<             m = ord(in_file.read(1)) & 0x3
<             s, f = int_format[m]
<             return struct.unpack(f, in_file.read(s))[0]
<         return token_l
<     def readNextObject(offset):
<         """ read the object at offset. May recursively read sub-objects (content of an array/dict/set) """
<         in_file.seek(offset)
<         token = in_file.read(1)
<         token_h, token_l = ord(token) & 0xF0, ord(token) & 0x0F #high and low parts 
<         if token == '\x00':
<             return None
<         elif token == '\x08':
<             return False
<         elif token == '\x09':
<             return True
<         elif token == '\x0f':
<             return ''
<         elif token_h == 0x10: #int
<             result = 0
<             for k in xrange((2 << token_l) - 1):
<                 result = (result << 8) + ord(in_file.read(1))
<             return result
<         elif token_h == 0x20: #real
<             if token_l == 2:
<                 return struct.unpack('>f', in_file.read(4))[0]
<             elif token_l == 3:
<                 return struct.unpack('>d', in_file.read(8))[0]
<         elif token_h == 0x30: #date
<             f = struct.unpack('>d', in_file.read(8))[0]
<             return datetime.datetime.utcfromtimestamp(f + MAC_OS_X_TIME_OFFSET)
<         elif token_h == 0x80: #data
<             s = getSize(token_l)
<             return in_file.read(s)
<         elif token_h == 0x50: #ascii string
<             s = getSize(token_l)
<             return in_file.read(s)
<         elif token_h == 0x60: #unicode string
<             s = getSize(token_l)
<             return in_file.read(s * 2).decode('utf-16be')
<         elif token_h == 0x80: #uid
<             return in_file.read(token_l + 1)
<         elif token_h == 0xA0: #array
<             s = getSize(token_l)
<             obj_refs = struct.unpack('>' + ref_format * s, in_file.read(s * ref_size))
<             return map(lambda x: readNextObject(object_offsets[x]), obj_refs)
<         elif token_h == 0xC0: #set
<             s = getSize(token_l)
<             obj_refs = struct.unpack('>' + ref_format * s, in_file.read(s * ref_size))
<             return set(map(lambda x: readNextObject(object_offsets[x]), obj_refs))
<         elif token_h == 0xD0: #dict
<             result = {}
<             s = getSize(token_l)
<             key_refs = struct.unpack('>' + ref_format * s, in_file.read(s * ref_size))
<             obj_refs = struct.unpack('>' + ref_format * s, in_file.read(s * ref_size))
<             for k, o in zip(key_refs, obj_refs):
<                 key = readNextObject(object_offsets[k])
<                 obj = readNextObject(object_offsets[o])
<                 result[key] = obj
<             return result
<         raise InvalidFileException()
<     return readNextObject(object_offsets[top_object])
<

Thanks for the patch.  Could you upload it as a context diff?

Here is the new patch. I assumed that you meant to use diff -c instead of the raw diff command.

Hmm.  Apparently what I meant was -u instead of -c (unified diff).  I just use the 'hg diff' command myself, which does the right thing :)  Of course, to do that you need to have a checkout.  (We can probably use the context diff.)

This patch is for Python 2. New features are accepted only for Python 3.3+. I ported the patch, but since I have no Mac, I can't check.

To date code was specified incorrectly.

The length of integers was calculated incorrectly. To convert integers, you can use int.from_bytes.

Objects identity was not preserved.

I'm not sure that the recognition of XML done enough. Should consider UTF-16 and UTF-32 with the BOM and without.

Need tests.

Also I'm a bit cleaned up and modernizing the code. I believe that it should be rewritten in a more object-oriented style. It is also worth to implement writer.

storchaka > I'm trying to take care of your remarks.
So, I'm working on a more object-oriented code, with both write and read functions. I just need to write some test cases.
IMHO, we should add a new parameter to the writePlist function, to allow the use of the binary or the json format of plist files instead of the default XML one.

Keep it simple: if a few functions work, there is no need at all to add classes.  Before doing more work though I suggest you wait for the feedback of the Mac maintainers.

I (as one of the Mac maintainers) like the new functionality, but would like to see some changes:

1) as others have noted it is odd that binary and json plists can be read but not written

2) there need to be tests, and I'd add two or even three set of tests:

   a. tests that read pre-generated files in the various formats
      (tests that we're compatible with the format generated by Apple)

   b. tests that use Apple tools to generated plists in various formats,
      and check that the library can read them
      (these tests would be skipped on platforms other than OSX)

   c. if there are read and write functions: check that the writer
      generates files that can be read back in.

3) there is a new public function for reading binary plist files, 
   I'd keep that private and add a "format" argument to readPlist
   when there is a need for forcing the usage of a specific format
   (and to mirror the (currently hypothetical) format argument for
   writePlist).

Don't worry about rearchitecturing plistlib, it might need work in that regard but that need not be part of this issue and makes it harder to review the changes. I'm also far from convinced that a redesign of the code is needed.

I'm working on a class, BinaryPlistParser, which allow to both read and write binary files.

I've also added a parameter fmt to writePlist and readPlist, to specify the format ('json', 'xml1' or 'binary1', using XML by default). These constants are used by Apple for its plutil program.

I'm now working on integrating these three formats to the test_plistlib.py. However, the json is less expressive than the other two, since it cannot handle dates.

Here is the new patch, allowing read and write binary, json and xml plist files.

It includes both the plistlib.py and test/test_plistlib.py patches.
JSON format does not allow dates and data, so XML is used by default to write files.
I use the json library to write JSON plist files, but its output is slightly different from the Apple default output: keys of dictionaries are in different order. Thus, I removed the test_appleformattingfromliteral test for JSON files.

Similarly, my binary writer does not write the same binary files as the Apple library: my library writes the content of compound objects (dicts, lists and sets) before the object itself, while Apple writes the object before its content. Copying the Apple behavior results in some additional weird lines of code, for little benefit. Thus, I also removed the  test_appleformattingfromliteral test for binary files.

Other tests are made for all the three formats.

Hi,

I noticed in the latest message that d9pounces posted that "JSON format does not allow dates and data, so XML is used by default to write files.". Rthe XML version of plists also do not really 'support' those types, and they are converted as follows:

NSData -> Base64 encoded data
NSDate -> ISO 8601 formatted string

(from http://en.wikipedia.org/wiki/Property_list#Mac_OS_X)

So really it should be the same thing when converting to json no?

The plutil (Apple's command-line tool to convert plist files from a format to another) returns an error if you try to convert a XML plist with dates to JSON.

Where are you even seeing these json property lists? I just checked the most recent documentation for NSPropertyListSerialization, and they have not updated the enum for NSPropertyListFormat. It seems that if even Apple doesn't support writing json property lists with their own apis then we shouldn't worry about supporting it?

see: https://developer.apple.com/library/ios/#documentation/Cocoa/Reference/Foundation/Classes/NSPropertyListSerialization_Class/Reference/Reference.html

enum {
   NSPropertyListOpenStepFormat = kCFPropertyListOpenStepFormat,
   NSPropertyListXMLFormat_v1_0 = kCFPropertyListXMLFormat_v1_0,
   NSPropertyListBinaryFormat_v1_0 = kCFPropertyListBinaryFormat_v1_0
}; NSPropertyListFormat;
typedef NSUInteger NSPropertyListFormat;

plutil(1) supports writing json format. 

That written, the opensource parts of CoreFoundation on opensource.apple.com don't support reading or writing json files. 

I'm therefore -1 w.r.t. adding support for json formatted plist files, support for json can be added when Apple actually supports that it the system libraries and hence the format is stable.

are any more changes needed to the code that is already posted as a patch in this bug report? or are the changes you wanted to see happen in msg157669 not happen yet?

d9pouces: are you willing to sign a contributor agreement? The agreement is needed before we can add these changes to the stdlib, and I'd like to that for the 3.4 release.

More information on the contributor agreement: http://www.python.org/psf/contrib/contrib-form/

I just signed this agreement. Thanks for accepting this patch!

I've started work on integrating the latest patch.

Some notes (primarily for my own use):

* I'll drop support for the JSON format, that format is not used
  by Apple's libraries (although the plutil tool can convert plists
  to JSON)

* The patch (plistlib_with_tests.diff) no longer applies cleanly,
  there are two rejected hunks in test_plist.py. 

* There is no documentation update in the patch, I'm working on that

* There is also an NextStep format (see <http://code.google.com/p/networkpx/wiki/PlistSpec>. 
  That format is not supported by plutil, but is supported by 
  Apple's libraries. Support for this can be added later.

* plistlib needs futher work as well, it would be nice to move closer
  to the PEP8 coding style (although this should be done carefully to
  maintain compatibility with existing code.

* plistlib.Data should be deprecated. The class was needed in python 2.7,
  but python 3 already has a unambiguous representation for binary data:
  the bytes type.

* It might be nice to add some options from the json library to 
  the serialization function:

  * skipkeys=True/False: skip keys that aren't strings and cannot
    be represented in a plist file, instead of raising TypeError
  * check_circular=True/False: explicitly check for circular recursion
    and raise an exception instead of hoping that some other
    exception occurs
  * default=function(obj): a function that is called on values that
    cannot be represented, should either return a value that can
    be encoded or raise TypeError
  
* Issue #11101 mentions that it would be nice to have an option
  to ignore keys whose value is None. I'm not sure if this is really
  useful, and if it is it might be better to add a "skip values" option
  that ignores items where the value cannot be encoded.

I've attached issue14455-v2.txt with an updated patch. The patch is still very much a work in progress, I haven't had as much time to work on this as I'd like.

This version:
* Should apply cleanly to the tip of the default branch

* Move around some code.

* Doesn't pass unit tests (most likely because I've botched the manual
  merge). The unittests also don't cover the functionality I've added.

* Adds documentation

* Adds 'skipkeys' and 'sort_keys' to the write functions (with the 
  same semantics as these keywords have with json.dump)

* Adds 'data_as_bytes' to the read functions. Then this option is true
  binary data is returned as an instance of bytes instead of plistlib.Data
  The latter is still the default.

v3 is still a work in progress, and still fails some tests

* Replaced test data by data generated by a helper script (to make it
  easier to update)

* Use 'subtest' feature of unittest library in 3.4

* Small tweaks to plist library (the dump/load/dumps/loads function at
  the end probably won't survive)

* Updated link to the CFBinaryPlist.c source code (this should be a 
  newer version of the file)

* Added option to readPlist to pass in the dictionary type (defaults
  to plistlib._InternalDict). This is primarily useful for testing
  and debugging, but also mirrors the 'sortkeys' keyword argument for
  the writer (when the order can be important for writing the user should
  have some way to detect the order when reading).

The data generated by the binary plist generator does't match the data generated by Cocoa in OSX 10.8 (and generated by the helper script), I haven't fully debugged that problem yet.  The generated binary plist
and the Cocoa version can both be parsed by plistlib, and result in the same data structure

See also:

#18168: request for the sort_keys option
#11101: request for an option to ignore 'None' values when writing
#9256: datetime.datetime objects created by plistlib don't include timezone information
(and looking at the code I'd say that timezones are ignored when *writing* plist files as well)
#10733: Apple's plist can create malformed XML (control characters) than cannot be read by plistlib

The test failure I'm getting is caused by a difference in the order in which items are written to the archive. I'm working on a fix.

v4 passes the included tests. 

The testsuite isn't finished yet.

The fifth version of the patch should be much cleaner.

Changes:

* Coding style cleanup, the new code uses PEP8 conformant names for
  methods and variables.

* Explicitly make private classes private by prefixing their name with
  an underscore (including the old XML parser/generator classes)

* Remove support in the binary plist code for sets and uuids, neither can
  be written to plist files by Apple's code.

* More tests

* There is no support for JSON because JSON is not supported by Apple's
  propertylist APIs either. The command-line tool plutil does support
  JSON output, but the C and Objective-C APIs do not.

* There is no support for the old OpenStep format. That format is badly
  documented, has been deprecated for a long time, and writing it is 
  not supported by Apple's plist libraries.  The OpenStep format also 
  seems to be much more limited than the two modern ones.

Open issues:

* The patch contains plistlib.{dump,dumps,load,loads} which mirror the
  functions with same name from the pickle and json modules. 

  Is is usefull to add those functions to plistlib, and deprecate the 
  older functions?

  Advantages:
  + Cleaner API
  + Provides a clean path to remove plistlib.Data and 
    plistlib._InternalDict (the latter is already deprecated)

  Disadvantages:
  - Very (too?) close to needless code churn

* Is renaming PlistParser and PlistWriter ok? Both are private and a 
  quick search on google seems to indicate that nobody directory uses
  these classes.

  If renaming is ok, should methods/variables be renamed to PEP8 style?

* Should a 'default' keyword be added to the serialization functions
  (simular to the keyword in json.dump)

  I don't have a usecase for this, the only reason to add is is 
  consistency with the json module

* Should a 'skipvalues' keyword be added to the serialization functions
  (to ignore values that cannot be serialized, see #11101)

  I'm not convinced that this would be a good idea.

* Should a 'check_circular' keyword be added to the 
  serialization functions (again similar to the same keyword for
  json.dump)? 

  This would avoid relying on the recursion limit to break infinite loops
  when serializing circular datastructures.

  Would need to check if binary plist can contain circular data structures
  when they are written using Apple's libraries.

I intend to commit my latest version of the patch during the europython sprints, with a minor change: don't include dump(s) and load(s), that change (and the other items on "open issues" in my last post) can be addressed later.

Let me review your patch.

Any review would be greatly appreciated. One thing I'm not too happy about is the use of magic numbers in the binary plist support code, but I think that using constants or a dispatch table would not make the code any clearer.

I have added comments on Rietveld.

I have to apologize for unwitting misleading of d9pouces. Functional version of the patch is enough Pythonic and it looks more clear to me than object-oriented one.

I've attached a slightly updated version of the patch. The documentation now lists dump and load as the primary API, with the old API as a deprecated alternative.

The code hasn't been changed to relect this yet, but does contain a number of tweaks and bugfixes (and a new testcase that ensures that decoding UTF-16 and UTF-32 files actually works, after testing that those encodings are supported by Apple's tools).

NOTE: no review of this version is needed, I'm mostly posting as backup.

This version should be better:

* There should be no lines longer than 80 characters

* Changed coding style of the (private) XML plist writer classes
  to PEP8

* Public API is now dump/dumps and load/loads, the old API is still
  available and deprecated

  -> As mentioned before I'm not entirely sure if doing this is the
     right solution, but it does give us a clean path to remove
     deprecated functionality later on (in particular the Dict and Data
     classes). 
  -> What I haven't done yet, and probably should, is to write a 2to3
     fixer that converts code to the new API.

* Grouped source in more logical segments (deprecated, xml support, 
  binary support, generic code).

* The previous item means that I've moved even more code around, I looked
  into minimizing the patch but the original module could not be 
  easily extended without moving code around.

* Added some more tests
  -> I might add more tests when I manage to run coverage.py, for some
     reason the instructions in the dev-guide don't work for me with
     a framework install :-(

I might add a documentation comment to the binary plist support code that
gives an overview of the file format with pointers to more information, but other than that and possible test coverage improvements the patch should be done.

Updated test-data generator: it now encodes the data using base64, to
make it easier to generate a file with limited line lengths.

v8 of the patch contains 1 change from v7: the test data is encoded
in base64. This was primarily done to ensure that the file has 
usable line lengths.  A nice side effect is that it is now harder than
ever to manually change the test data, as the comment mentions there is
a script for generating that data.

As always I'd appreciate feedback on this patch, especially on deprecating the current public API and introducing a new (PEP8 compliant) one.

Ronald, I think v8 of the patch is missing (and plistlib_generate_testdata.py was uploaded twice).

Actually attach the latest version of the patch.

I'd really like to include this patch in 3.4, but haven't managed to do any opensource work in the previous period and don't know when I'll be able to actually commit this (and more importantly, be available when issues crop up) :-(

I added a lot of comments on Rietveld.

I for the most part agree with the comments and will provide an updated patch on thursday. Would you mind if I committed that without further review (due to cutting it awfully close to the deadline for beta 1)?

Some comments I want to reply to specifically:

* "Can the code be simpler, with only one pass?"

   Maybe, but not right now. 

* "This is inconsistent with _flatten()."

  I'll add a comment that explains why this is: _flatten (and this code) 
  can deal with arbitrary keys, but that is not supported by Apple's code.

  The type check in _write_object ensures that it is not possible to write
  archives that cannot be read back by Apple's Cocoa frameworks.

* "unusual indentation" (several times)

  I'll have to look at other stdlib code to find suitable indentation, this
  is indentation I've used in my code for a long time (I've also used 
  camelCase instead of pep8_style names for methods for a long time, which 
  is probably why I never noticed that I forgot to convert some method 
  name when cleaning up the naming conventions used in this module).

It's too large and complicated patch. I would like to have a chance to quick review it before committing. You will have time to commit.

I've attached an updated version of the patch that should fix most of the issues found during review.

I've also changed the two FMT_ constants to an enum.Enum (but still expose the constants themselves as module global names because that's IMHO more convenient).

FYI I'm completely away from the computer during the weekend and will have very limited time to work from later today (18:00 CET).

I have added a few comments on Rietveld. Besides formatting nitpicks your have forgot third argument in new warns and missed some details in tests. As for the rest the patch LGTM. If you have no time I will fixed this minor issues and will commited the patch.

I'm not sure about docstrings text ("return" vs "returns", I don't remember what is better), but we can bikeshed it after beta 1.

Updated patch after next round of reviews.

New changeset 673ca119dbd0 by Ronald Oussoren in branch 'default':
Issue #14455: plistlib now supports binary plists and has an updated API.
http://hg.python.org/cpython/rev/673ca119dbd0

New changeset 602e0a0ec67e by Ned Deily in branch 'default':
Issue #14455: Fix maybe_open typo in Plist.fromFile().
http://hg.python.org/cpython/rev/602e0a0ec67e

These changes are worth to mention in What's News.

"versionchanged" below writePlistToBytes() is wrong. Perhaps below dump() too. "versionadded" is needed for new functions: dump(), dumps(), load(), loads().

Currently negative integers are not supported in binary format. Here is a patch which adds support for negative integers and large integers up to 128 bit.

oops... thanks for the patch. I'll review later this week, in particular the 128 bit integer support because I don't know if Apple's libraries support those.

According to [1] Apple's libraries write any signed 128-bit integers, but read only integers from -2**63 to 2**64-1 (e.g. signed and unsigned 64-bit integers).

[1] http://opensource.apple.com/source/CF/CF-550/CFBinaryPList.c

Yet one nitpick. Perhaps _write_object() should raise TypeError instead of InvalidFileException.

I'm working on an update for your patch that addresses these comments:

* I don't like supporting 128 bit integers because Apple's public APIs
  don't support those values. That is, the value 'kCFNumberSInt128Type'
  is not in a public header for the OSX 10.9 SDK.

* The test_int method that you introduced tests conversions to and from XML,
  and doesn't test the problem with negative values in binary plists.

* I don't understand why test_int converts a value to a binary plist twice,
  that is the 'data2 = plistlib.dumps(pl2)' bit.

* I'm adding negative integers to the _create method as well, with the 
  corresponding changes to the binary TESTDATA dictionary.

* I don't understand your comment about the writePlistToBytes documentation,
  there was no versionchanged in the documentation.  The version changed 
  for dump was wrong, that should be versionadded (and the other new functions
  should have a versionadded as well)

* I agree that this change should be mentioned in What's New.

* I agree that _write_object should raise TypeError

BTW. What about out-of-range integer values? Those currently raise struct.error, I'd prefer to raise TypeError instead because the use of the
struct module should be an implementation detail.

And a final question: integers with '2 ** 63 <= value < 2 ** 64' (e.g. values that are in the range of uint64_t but not in the range of int64_t) can be written to a binary plist, but will be read back as a negative value (which is the same behavior as in Apple's code). Should we warn about this in the documentation?

I'll post an updated patch later today.

The attached patch should fix the open issues:

* Negative integers are supported (based on Serhiy's patch), but without
  support for 128-bit integer (as per my previous comment)

* Test updates for this

* Updated version tags in the documentation

* Documented the odd behavior for 64-bit unsigned values larger than the
  largest 64-bit signed value.

* Raise TypeError when trying to write an object that isn't supported
  (with test)

* Raise OverflowError when trying to write an integer that cannot be 
  represented in a binary plist

* Add entry to "What's New"

> * I don't like supporting 128 bit integers because Apple's public APIs
>   don't support those values. That is, the value 'kCFNumberSInt128Type'
>   is not in a public header for the OSX 10.9 SDK.

At least we should support integers from -2**63 to 2**64-1 (signed and 
unsigned 64-bit).

> * The test_int method that you introduced tests conversions to and from XML,
> and doesn't test the problem with negative values in binary plists.

Indeed.

> * I don't understand why test_int converts a value to a binary plist twice,
>   that is the 'data2 = plistlib.dumps(pl2)' bit.

I have copied this from test_bytes(). I suppose that pl2 can be int subclass. 
Agree, for now this check is redundant.

> * I don't understand your comment about the writePlistToBytes documentation,
> there was no versionchanged in the documentation.  The version changed for
> dump was wrong, that should be versionadded (and the other new functions
> should have a versionadded as well)

http://hg.python.org/cpython/file/673ca119dbd0/Doc/library/plistlib.rst#l165

> BTW. What about out-of-range integer values? Those currently raise
> struct.error, I'd prefer to raise TypeError instead because the use of the
> struct module should be an implementation detail.

Agree. Especially if OSX SDK doesn't support deserialization of integers 
larger than 64-bit. Perhaps we should add this check for XML format too. And 
document this limitation.

> And a final question: integers with '2 ** 63 <= value < 2 ** 64' (e.g.
> values that are in the range of uint64_t but not in the range of int64_t)
> can be written to a binary plist, but will be read back as a negative value
> (which is the same behavior as in Apple's code). Should we warn about this
> in the documentation?

These values should be written as 128-bit integers (token b'\x14').

Attached a script (using PyObjC) that demonstrates the behavior of Apple's Foundation framework with large integers. The same behavior should occur when the script is rewritten in Objective-C.

Updated patch.

I can't test on OSX, but I see that Apple's code can write any 128-bit integers and read signed and unsigned 64-bit integers.

Can Apple's utilities read this file? What is a result?

Conversion to XML results in:

$ plutil -convert xml1 -o - 18446744073709551615.plist 
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
	<key>a</key>
	<integer>18446744073709551615</integer>
</dict>
</plist>

This is the same as what I get with my latest patch:

>>> import plistlib
>>> plistlib.load(open('18446744073709551615.plist', 'rb'))
__main__:1: ResourceWarning: unclosed file <_io.BufferedReader name='18446744073709551615.plist'>
{'a': 18446744073709551615}

(and I have check that I can create a binary plist with a negative integer in this shell session)

New changeset 1a8149ba3000 by Ronald Oussoren in branch 'default':
Issue #14455: Fix some issues with plistlib
http://hg.python.org/cpython/rev/1a8149ba3000

I see that plistlib incorrectly writes large ints from 2**63 to 2**64-1 as negative values.

>>> d = plistlib.dumps({'a': 18446744073709551615}, fmt=plistlib.FMT_BINARY)
>>> plistlib.loads(d)
{'a': -1}

My patch did this correct (as 128-bit integer), and as you can see the produced file is accepted by Apple's plutil.

However, I have no idea how to write that file using Apple's APIs. 

I'd prefer to either be compatible with Apple's API (current behavior), or just outright reject values that cannot be represented as a 64-bit signed integer.  

The file you generated happens to work, but as there is no way to create such as file using a public API there is little reason to expect that this will keep functioning in the future.  

The CFBinaryPlist code appears to be shared between support for binary plists and keyed archiving (more or less Cocoa's equivalent for pickle) and supports other values that cannot be put in plist files, such as sets.  The original patch supported sets in the binary plist reader and writer, I ripped that out because such objects cannot be serialised using Apple's plist APIs.

Keep in mind that this module is intended for interop with Apple's data format.

> However, I have no idea how to write that file using Apple's APIs.

Look in CFBinaryPList.c. It have a code for creating 128-bit integers:

	    CFSInt128Struct val;
	    val.high = 0;
	    val.low = bigint;
	    *plist = CFNumberCreate(allocator, kCFNumberSInt128Type, &val);

And I suppose that you have at least one way to create such file -- just 
convert plist file in XML format to binary format.

> Keep in mind that this module is intended for interop with Apple's data
> format.

Apple's tool can read and write integers from 2**63 to 2**64-1.

Here is a patch against current sources.

kCFNumberSInt128Type is not public API, see the list of number types in <https://developer.apple.com/library/mac/documentation/corefoundation/Reference/CFNumberRef/Reference/reference.html>.

I agree that CFBinaryPlist.c contains support for those, and for writing binary plists that contain sets, but you cannot create a 128 bit CFNumber object using a public API, and the public API for writing plists won't accept data structures containing sets.

You have at least one way to create a 128 bit CFNumber. Read plist file (and 
you can create plist in XML format with big integers in any text editor).

In any case it is not good to produce incorrect plist for big integers. If you 
don't want to support integers over 2**63, just reject them.

Reopening because Cocoa behaves differently that I had noticed earlier...

The (Objective-C) code below serialises an NSDictionary with an unsigned long of value ULLONG_MAX and then reads it back. I had expected that restored value contained a negative number, but it actually reads back the correct value.

I'm going to do some more spelunking to find out what's going on here, and will adjust the plistlib code to fully represent all values of unsigned 64-bit integers (likely based on your code for supporting 128-bit integers)



Output (on a 64-bit system running OSX 10.9):

$ ./demo 
2014-01-15 15:34:18.196 demo[77580:507] input dictionary: {
    key = 18446744073709551615;
}   value 18446744073709551615
2014-01-15 15:34:18.198 demo[77580:507] as binary plist: <62706c69 73743030 d1010253 6b657914 00000000 00000000 ffffffff ffffffff 080b0f00 00000000 00010100 00000000 00000300 00000000 00000000 00000000 000020>
2014-01-15 15:34:18.198 demo[77580:507] Restored as {
    key = 18446744073709551615;
}


Code:

/*
 * To use:
 *  $ cc -o demo demo.c -framework Cocoa
 *  $ ./demo
 */
#import <Cocoa/Cocoa.h>

int main(void)
{
	NSAutoreleasePool* pool = [[NSAutoreleasePool alloc] init];
	NSNumber* value = [NSNumber numberWithUnsignedLongLong:ULLONG_MAX];

	NSDictionary* dict = [NSDictionary dictionaryWithObjectsAndKeys:value, @"key", nil];
	NSLog(@"input dictionary: %@   value %llu", dict, ULLONG_MAX);

        NSData* serialized = [NSPropertyListSerialization
            dataWithPropertyList:dict
                          format: NSPropertyListBinaryFormat_v1_0
                         options: 0
                           error: nil];
        NSLog(@"as binary plist: %@", serialized);

        NSDictionary* restored = [NSPropertyListSerialization
            propertyListWithData:serialized
                         options:0
                          format:nil
                           error:nil];
        NSLog(@"Restored as %@", restored);
	return 0;
}

> I'm going to do some more spelunking to find out what's going on here, and
> will adjust the plistlib code to fully represent all values of unsigned
> 64-bit integers (likely based on your code for supporting 128-bit integers)

My last patch supports only values up to 2**64-1.

Perhaps you will want to add new test case in 
Mac/Tools/plistlib_generate_testdata.py.

New changeset 0121c2b7dcce by Ronald Oussoren in branch 'default':
Issue #14455: fix handling of unsigned long long values for binary plist files
http://hg.python.org/cpython/rev/0121c2b7dcce

Serhiy: the issue should now be fixed.

I finally understand why I was so sure that Apple's code serialised large positive numbers as negative numbers: due to a bug in PyObjC large positive numbers end up as NSNumber values that are interpreted as negative values.

The patch tweaks the test generator to do the right thing by explicitly creating the NSNumber value instead of relying on PyObjC's automatic conversion. 

Now I just have to hunt down this bug in PyObjC :-)

New changeset 728f626ee337 by R David Murray in branch 'default':
whatsnew: plistlib new api and deprecations (#14455)
http://hg.python.org/cpython/rev/728f626ee337