gh-116738: Make lzma module thread-safe

[yoney]

Make the attributes in lzma module thread-safe on the free-threading build. Attributes (check, eof, needs_input, unused_data) are now stored atomically or accessed via mutex-protected getters.

Even though the operations are protected by locks, the attributes exposed via PyMemberDef (check, eof, needs_input, unused_data) should still be updated atomically while holding the lock, or accessed via mutex‑protected getters, because they can be read without acquiring the lock.

This issue was not addressed in #140711. A small addition that repeatedly reads these attributes shows the issue under tsan build.

WARNING: ThreadSanitizer: data race (pid=2367860)
  Atomic read of size 1 at 0x7fdbc9506ee0 by thread T19:
    #0 _Py_atomic_load_char_relaxed /workspace/ft_lzma_tsan/./Include/cpython/pyatomic_gcc.h:311 (python+0x99e392) (BuildId: ce943d12d16f92ec125f1db1856a1617d3d75854)
    #1 PyMember_GetOne /workspace/ft_lzma_tsan/Python/structmember.c:37 (python+0x99e392)
    #2 member_get /workspace/ft_lzma_tsan/Objects/descrobject.c:180 (python+0x60f4d5) (BuildId: ce943d12d16f92ec125f1db1856a1617d3d75854)
...
...
...
Previous write of size 1 at 0x7fdbc9506ee0 by thread T11:
    #0 decompress /workspace/ft_lzma_tsan/./Modules/_lzmamodule.c:1067 (_lzma.cpython-315td-x86_64-linux-gnu.so+0x7256) (BuildId: 978a710f12e938f2fbc3233c01c40cf693105ef1)
    #1 _lzma_LZMADecompressor_decompress_impl /workspace/ft_lzma_tsan/./Modules/_lzmamodule.c:1139 (_lzma.cpython-315td-x86_64-linux-gnu.so+0x7256)
    #2 _lzma_LZMADecompressor_decompress /workspace/ft_lzma_tsan/./Modules/clinic/_lzmamodule.c.h:157 (_lzma.cpython-315td-x86_64-linux-gnu.so+0x7256)
...
...
...
SUMMARY: ThreadSanitizer: data race /workspace/ft_lzma_tsan/./Modules/_lzmamodule.c:1043 in decompress

cc: @mpage @colesbury

• Issue: Audit all built-in modules for thread safety #116738

[emmatyping]

unused_data should be b"" if not set, so I don't think this check is needed. It's initialized in __new__ https://github.com/python/cpython/blob/6a4f10325d58deb1906b39d68dc8e84f4c2bf5a4/Modules/_lzmamodule.c#L1239C5-L1239C76

I think we should probably initialize unused_data in bz2 in __new__ as well.

[emmatyping]

One other possibility would be to standardize on the implementation in compression.zstd https://github.com/python/cpython/blob/main/Modules/_zstd/decompressor.c#L583-L611

[yoney]

Thank you @emmatyping for pointing this out.

I actually wanted to assert(self->unused_data != NULL), but I was a little bit worried about the code block below. I think PyBytes_FromStringAndSize() failure is very unlikely in L1044, and I believe it could happen only with allocation errors, but this was the main reason. With Py_T_OBJECT_EX in PyMemberDef, reading a NULL
raises AttributeError, so I was trying to keep the current behaviour.

cpython/Modules/_lzmamodule.c

Lines 1040 to 1049 in 08bc03f

	if (d->eof) {
	d->needs_input = 0;
	if (lzs->avail_in > 0) {
	Py_XSETREF(d->unused_data,
	PyBytes_FromStringAndSize((char *)lzs->next_in, lzs->avail_in));
	if (d->unused_data == NULL) {
	goto error;
	}
	}
	}

So, how do you think we should change it?