The reason we have a manifest file for python is that our goal is to
keep python-core as small as posible and add other python packages only
when the user needs them, hence why we split upstream python into several
packages.
Although our manifest file has several issues:
- Its unorganized and hard to read and understand it for an average
human being.
- When a new package needs to be added, the user actually has to modify
the script that creates the manifest, then call the script to create
a new manifest, and then submit a patch for both the script and the
manifest, so its a little convoluted.
- Git complains every single time a patch is submitted to the manifest,
since it violates some of its guidelines.
- It changes or may change with every release of python, its impossible
to know if the required files for a certain package have changed
(it could have more or less dependencies), the only way of doing so
would be to install and test them all one by one on separate individual
images, and even then we wouldnt know if they require less dependencies,
we would just know if an extra dependency is required since it would
complain, lets face it, this isnt feasible.
- The same thing happens for new packages, if someone wants to add a
new package, its dependencies need to be checked manually one by one.
This patch fixes those issues, while adding some additional features.
Features/Fixes:
- A new manifest format is used (JSON), easy to read and understand.
This file is parsed by the python recipe and python packages read
from here are passed directly to bitbake during parsing time.
- It provides an automatic manifest creation task (explained below),
which automagically checks for every package dependencies and adds
them to the new manifest, hence we will have on each package exactly
what that package needs to be run, providing finer granularity.
- Dependencies are also checked automagically for new packages (explained below).
- Fixes the manifest in the following ways:
* python-core should be base and all packages should depend on it,
fixes lang, string, codecs, etc.
* Fixes packages with repeated files (e.g. bssdb and db, or
netclient and mime, and many others).
- Sitecustomize was fixed since encoding was deprecated.
- The JSON manifest file invalidates bitbake's cache, so if it changes
the python package will be rebuilt.
- It creates a solution for users that want precompiled bytecode files
(*.pyc) INCLUDE_PYCS = "1" can be set by the user on their local.conf to
include such files, some argument they get faster boot time, even when the
files would be created on their first run?, but they also sometimes give a
magic number error and take up space, so we leave it to the user to
decide if they want them or not.
- Fixes python-core dependencies, e.g.
When python is run on an image, it TRIES to import everything it needs,
but it doesnt necessarily fails when it doesnt find something, so even if
we didnt know, we had errors like (trimmed on purpose):
# trying /usr/lib/python2.7/_locale.so
# trying /usr/lib/python2.7/lib-dynload/_locale.so
# trying /usr/lib/python2.7/_sysconfigdata.so
while it didnt complain about _locale it should have imported it,
after creating a new manifest with the automated script we get:
# trying /usr/lib/python2.7/lib-dynload/_locale.so
dlopen("/usr/lib/python2.7/lib-dynload/_locale.so", 2);
import _locale # dynamically loaded from /usr/lib/python2.7/lib-dynload/_locale.so
How to use (after a new release of python, or maybe before every OE
release):
- A new task called create_manifest was added to the python package,
which may be invoked via:
$ bitbake python -c create_manifest
This task runs a script on native python on our HOST system, and since
the python and python-native packages come from the same source, we can
use it to know the dependencies of each module as if we were doing it
on an image, this script is called create_manifest.py and in a very
simplistic way it does the following:
1. Reads the JSON manifest file and creates a dictionary data structure
with all of our python packages, their FILES, RDEPENDS and SUMMARY.
2. Loops through all of them and runs every module listed on them
asynchronously, determining every dependency that they have.
3. These module dependencies are then handled, to be able to know which
packages contain those files and which should RDEPEND on one another.
4. The data structure that comes out of this, is then used to create a
new manifest file which is automatically copied onto the user's python
directory replacing the old one.
Create_manifest script features:
- Handles modules which dont exist anymore (new release for example).
- Handles modules that are builtin.
- Deals with modules which were not compiled (e.g. bsddb or ossaudiodev)
- Deals with packages which include folders.
- Deals with packages which include FILES with a wildcard.
- The manifest can be constructed on a multilib environment as well.
- This method works for both python modules and shared libraries used
by python.
How to add a new package:
- If a user wants to add a new package all that has to be done is
modify the python2-manifest.json file, and add the required file(s)
to the FILES list, the script should handle all the rest.
Real example:
We want to add a web browser package, including the file webbrowser.py
which at the moment is on python-misc.
"webbrowser": {
"files": ["${libdir}/python2.7/lib-dynload/webbrowser.py"],
"rdepends": [],
"summary": "Python Web Browser support"}
Run bitbake python -c create_manifest and the resulting manifest
should be completed after a few seconds, showing something like:
"webbrowser": {
"files": ["${libdir}/python2.7/webbrowser.py"],
"rdepends": ["core","fcntl","io","pickle","shell","subprocess"],
"summary": "Python Web Browser support"}
Known errors/issues:
- Some special packages are handled differently: core, misc,
modules,dev, staticdev.
All these should be handled manually, because they either include
binaries, static libraries, include files, etc. (something that we
cant import).
Specifically static libraries are not not supported by this method
and have to be handled by the user.
- The change should be transparent to the user, other than the fact
that now we CANT build python-foo (it was pretty dumb anyway, since
what building python-foo actually did was building the whole python
package anyway), but doing IMAGE_INSTALL_append = " python-foo"
would create an image with the requested package with no issues.
[YOCTO #11510] [YOCTO #11694] [YOCTO #11695]
(From OE-Core rev: 6959e2e4dba5bbfa6ffd49c44e738cc1c38bc280)
Signed-off-by: Alejandro Hernandez <alejandro.hernandez@linux.intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The bsddb module is deprecated and requires an old version of Berkeley DB that
some may be unhappy with even shipping, so expose a way to disable the module.
(From OE-Core rev: 701250dae6c5d3f464bf6d7c46c19d59d1c00bec)
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
python-ptest needs python-tests package installed in order to
run any tests.
This patch adds python-tests as a runtime dependency, so the test suite
will be present in the image.
While in there, also removed several build host references.
[YOCTO #12144]
(From OE-Core rev: ff83e15289e4b47cd3926220a0039bf97ec35120)
Signed-off-by: Juro Bystricky <juro.bystricky@intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The compiled .pyc files contain time stamp corresponding to the compile time.
This prevents binary reproducibility. This patch allows to achieve binary
reproducibility by overriding the build time stamp by the value
exported via SOURCE_DATE_EPOCH.
Patch by Bernhard M. Wiedemann, backported from https://github.com/python/cpython/pull/296
[YOCTO#11241]
(From OE-Core rev: 2a044f1e4f5c63e11e631b31f741c7aabfa6f601)
Signed-off-by: Juro Bystricky <juro.bystricky@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
In upstream, the following commit:
e711cafab1
...
commit e711cafab13efc9c1fe6c5cd75826401445eb585
Author: Benjamin Peterson <benjamin@python.org>
Date: Wed Jun 11 16:44:04 2008 +0000
Merged revisions 64104,64117 via svnmerge from
svn+ssh://pythondev@svn.python.org/python/trunk
...
(see diff in setup.py)
It assigned libraries for multiprocessing module according
the host_platform, but not pass it to Extension.
In glibc, the following commit caused two definition of
sem_getvalue are different.
https://sourceware.org/git/?p=glibc.git;a=commit;h=042e1521c794a945edc43b5bfa7e69ad70420524
(see diff in nptl/sem_getvalue.c for detail)
`__new_sem_getvalue' is the latest sem_getvalue@@GLIBC_2.1
and `__old_sem_getvalue' is to compat the old version
sem_getvalue@GLIBC_2.0.
If not explicitly link to library pthread (-lpthread), it will
load glibc's sem_getvalue randomly at runtime.
Such as build python on linux x86_64 host and run the python
on linux x86_32 target. If not link library pthread, it caused
multiprocessing bounded semaphore could not work correctly.
...
>>> import multiprocessing
>>> pool_sema = multiprocessing.BoundedSemaphore(value=1)
>>> pool_sema.acquire()
True
>>> pool_sema.release()
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: semaphore or lock released too many times
...
And the semaphore issue also caused multiprocessing.Queue().put() hung.
(From OE-Core rev: ca1542cdf6b6437a2f3dcdb33ac5216bf841c04a)
Signed-off-by: Hongxu Jia <hongxu.jia@windriver.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
