eliminating phar's interception of zend_compile in favor of a new to-be-added hook in PHP 5.3+,
allow "include 'phar:///path/to/my.phar';" to work as equivalent to
php /path/to/my.phar
this slight change in scripting still allows inclusion and execution of phar stub, but removes the need to
check and modify path in zend_compile, which allows us to play much nicer with external tools like
debuggers/opcode caches
Maybe we don't need to call a non-existent dtor if we're going to physically apply zend_hash_graceful_reverse_destroy()?
- This works on my box, please test under *nix/OSX
- Make cached manifest test platform agnostic
- Comment out zend_(init|destroy)_rsrc_list() and associated references
@Greg: cached manifest test (now) passes here regardless, make of that what you will
copy virtual_dirs to avoid segfault on multi-process
fix metadata reading for phar.cache_list
initialize manifest to exact size needed (performance increase)
fix freeing of signature on error to use the correct persist value (fixes segfault on error in cache.list)
reset EG(regular_list) so it is identical to how we found it
at startup. This caches the manifest, so that on first access to a phar archive, no file manifest parsing occurs.
This could use further tweaking. For instance, the full copy of the manifest into the current process may be unnecessary if refcounting could be external
to the manifest. This would be another significant gain. With APC, I measure a slight perf increase to 19 req/sec up from 16 req/sec, without it approaches
regular PHP at 3.8 req/sec (regular is 4 req/sec). This is benching phpMyAdmin
for the contents of the exported private key to Phar->setSignatureAlgorithm, and expects the public key to be in
a file named blah.phar.pubkey in the same directory as the phar blah.phar. This works with openssl static or
shared and fails gracefully if openssl is not present without adding a dependency. config.w32 needs updating to match config.m4 [DOC]
to open not-yet-loaded phar and fails on compressed files
# By Gregory's request
# Sorry, can't find how to write test case for that - it reproduces
# for me only under bytecode-caching. Suggestions welcome.
to open not-yet-loaded phar and fails on compressed files
# By Gregory's request
# Sorry, can't find how to write test case for that - it reproduces
# for me only under bytecode-caching. Suggestions welcome.
* found felipe's segfault in util.c and fixed the segfault (3 tests fail due to odd behavior of . and .. on this machine)
* fixed serious flaws in the setting/resetting of is_data - now it works properly. Assume
all new PharData are tar-based, and allow passing Phar::ZIP to PharData constructor to override this
* fix broken earlier commit, introduced segfault that broke 20 tests here
* found felipe's segfault in util.c and fixed the segfault (3 tests fail due to odd behavior of . and .. on this machine)
* fixed serious flaws in the setting/resetting of is_data - now it works properly. Assume
all new PharData are tar-based, and allow passing Phar::ZIP to PharData constructor to override this
* fix broken earlier commit, introduced segfault that broke 20 tests here
this is done by removing zlib/bz2 explicit dependencies because they are unnecessary
we only ever use the stream filter, and the check for existence has
been moved to runtime where it is after startup
this is done by removing zlib/bz2 explicit dependencies because they are unnecessary
we only ever use the stream filter, and the check for existence has
been moved to runtime where it is after startup
phar_find_in_include_path. We were allowing data-based phars to be executed, and actually marking phar-based phar archives
without '.phar' in the name as data-based phars, which would allow modifying them even if phar.readonly=0. Add test for this sinister case
- This would've fixed that test... removing clean section
@Greg: I commented out the call that breaks the Windows build, pending a decision about its future.
this also resulted in a major fix for mounted directories, which were recycling the 'link' field which
could cause stupid conflicts with actual links, so move that to new 'tmp' field.
1 - executable phars must contain '.phar' in the filename
2 - non-executable phars must not contain '.phar' and must have an extension of at least 1 character
In addition, phar filenames must exist if opened for read, and the directory containing the phar must exist if opened for creation
if opened for creation, the file must not already exist
[DOC]
5.3 code expects the proposed patch for stream wrapper in include_path to be committed
5.2 code only supports phar stream wrapper in include_path.
this is a 2-step process. After this, more magic, particularly in funcinterceptors.c will be
converted to use phar_resolve_path, which is far safer than the current implementation.
this needs windows and 5.2 testing unix/windows
- If readonly=0, why not $phardata->convertToPhar()?
- Known issue with directories creating 'as-file' copies within the archive (all formats)
@Greg/Marcus/Tony: This passes all tests on my box, 5.2/5.3/release_ts/debug_ts, and I can't find any more memleaks. Obviously this is too good to be true, so if conversion is still messy elsewhere please feel free to fix, or bug and assign to me.
Note: two tests currently fail. IMHO we should be throwing E_ERROR on encountering a corrupted archive, not trying to throw a trail of exceptions...
New tests still to be written, not all functionality is in place yet.
@Greg: this breaks a handful of tests due to the change in stub.h - will fix in a bit. The only one that's interesting is you can't do strlen(Phar::createDefaultStub()) and expect anything other than an exception now.
# TODO: implement directory mounting. Involves scanning paths that are not exact matches for a partial match.
# this also means maintaining a per-phar hash of mounted directories for quick matching
intercepted file functions now fall through if the file is not found in the phar, this allows access to external libraries
actually use include_path for locating files for inclusion and in file_get_contents/fopen when include_path is requested.
This allows applications like Zend Framework MVC implementation to function properly
1) rename is_explicit_alias to is_temporary_alias for clarity and flip the value
2) fix setAlias so that it sets a permanent to-be-saved alias, and restores the old one on error
3) fix Phar constructor to work with sub-directories in RecursiveDirectoryIterator
re-organize, create util.c, move entry_info/archive_data/entry_data access methods to this file
refactor entry->fp, now this is abstracted with phar_get_efp() and phar_seek_efp(), fixes all weird dependency issues
permanently solve the "millions of file pointers" issue for read access. All compressed files are read into a single
temporary stream, and their constraints are controlled by the entry->fp abstraction
Improvements in this zip implementation over ext/zip:
* full read/write support for bzip2 compressed files
* much more efficient access for accessing only a few files within large zip files, as crc/header validation is
done just-in-time
* full stream support for opendir/rename/rmdir/mkdir as well as all of the other stream funcs
* full support for setting file perms via Phar::chmod(), stored as zip-standard extra field
* no problem with large zips and many open file pointers
# TODO: add big-endian system support for tar/zip file format headers, otherwise the implementation is complete
# TODO: test on windows and fix any windows-specific issues
# TODO: verify zips created work with unzip/winzip/windows explorer and so on
out of the box regardless of server configuration with phar file format
split up stub.h strings into 2046 byte chunks because MS VC 6 is friggin stupid
the new default stub allows creation of phars that run identically
1) with Phar extension
2) without Phar extension
3) extracted to disk from the phar
this makes the default phar format quite interesting as it eliminates the only drawback of the extension
