Patch _tkinter.c to look in base prefix
#421
Conversation
charliermarsh
force-pushed
the
charlie/tkinter
branch
6 times, most recently
from
bf37889 to
b9e7b1d
Compare
|
(To clarify, I'm looking for sign-off on the approach and input on whether we want to apply a separate patch for each Python minor, or cutoff support at some later version, like 3.11, or 3.12, or 3.13.) |
There was a problem hiding this comment.
Historically I've tried to ensure that all Python X.Y versions behave the same. So nominally we should carry patch variants for all versions, as necessary.
I agree that this is the type of patch that upstream may want to take. However, I could see aspects of it getting stalled up. e.g. I wouldn't be surprised if some distros didn't install things at /usr/lib/tcl and the hardcoding of this path is probably wrong. (This is the type of thing that should be controllable via a configure argument and the actual path should also be sniffed via configure.)
I glanced at the patch and looking for the tcl files in <prefix>/lib/tcl should be reasonable for PBS. So the new functionality passes the smell test.
cpython-unix/build-cpython.sh
Outdated
| @@ -179,6 +179,15 @@ else | |||
| patch -p1 -i ${ROOT}/patch-ctypes-callproc-legacy.patch | |||
| fi | |||
|
|
|||
| # On Windows, CPython looks for the Tcl/Tk libraries relative to the base prefix, | |||
| # with we want. But on Unix, it doesn't. This patch applies similar behavior on Unix, | |||
|
Sounds good. I'll go ahead and add patches for the remaining minors. Yeah, I briefly chatted with @zooba about this in DMs, just to ask for any context on why the Tcl/Tk behavior differs, and he suggested that it could make sense for us to add a build flag in CPython to enable standalone behavior, since it tends to be easier to get patches merged if they're not enabled by default and therefore don't impact redistributors. |
charliermarsh
force-pushed
the
charlie/tkinter
branch
from
da1bd7e to
20d138d
Compare
|
Ok, I somewhat painstakingly added and manually tested a patch for each minor. |
charliermarsh
force-pushed
the
charlie/tkinter
branch
from
20d138d to
cc78c11
Compare
There was a problem hiding this comment.
I don't know anything about Tcl/Tk, so I don't fully understand what this code needs to do, but commented on a couple things I noticed. Couple other things:
- It seems like the return value of
_get_tcl_lib_pathis never decrefed and is just leaked? But if this is only called in one-time init code paths, probably nobody cares, one-time leaks are normal. And it looks like that was already the case in the Windows-only codepaths. - It seems like some code duplication between Windows and non-Windows could probably be reduced here, but for now this keeps the patches simpler. Might be worth doing that if this is upstreamed in future.
| + return NULL; | ||
| + } | ||
| + if (str_path != NULL) { | ||
| + setenv("TCL_LIBRARY", path, 1); |
There was a problem hiding this comment.
It looks like path is used un-initialized here?
(Same in other versions.)
| Tcl_FindExecutable(PyBytes_AS_STRING(cexe)); | ||
| + | ||
| + if (set_var) { | ||
| + unsetenv("TCL_LIBRARY"); |
There was a problem hiding this comment.
Should path be freed here? (If it were ever allocated/initialized above, which it doesn't seem to be.) The equivalent in the Windows code path (wcs_path) is freed here.
(Same in other versions.)
|
Thank you @carljm! |
## Summary In #421, @carljm pointed out that `path` was uninitialized in `setenv("TCL_LIBRARY", path, 1);`. It turns out we need to set it via `PyUnicode_AsUTF8`. I confirmed that the existing code _was_ running; it's just that`setenv` returns an error (rather than crashing), and we weren't checking the result (though the same is true in the Windows path). If you, e.g., try to `printf` `path` just before, you get a segmentation fault. This code isn't necessary to fix the motivating Tkinter limitation, so that's another reason it wasn't caught.
Summary
This PR attempts to adresss one of the known quirks around Tcl/tk files. On Windows, CPython will look for the Tcl/tk files at
{base_prefix}/lib. That works well for python-build-standalone, since we ship the files in that location. However, on Unix, CPython does not check that location -- instead, for python-build-standalone, it just looks relative to the executable. This works for non-virtual environments, since CPython will look in{prefix}/lib. But not for virtual environments.The patch applied here ports the Windows discovery logic to Unix, so that we always hook up the right paths.
The advantage of this approach (vis-a-vis others that were considered) is that it doesn't rely on injecting an environment variable into the user's environment.
For now, I've only patched Python 3.13. Unfortunately, I'll need a slightly different patch for each Python minor. I've already authored them locally, but I want to hold off on wiring them all up until we're aligned on this general approach.
In theory, I think this change could be upstreamed.
Test Plan
./python/install/bin/python -c "from tkinter import Tk; Tk()"./python/install/bin/python -m venv .venv.venv/bin/python -c "from tkinter import Tk; Tk()"