GeezerVison is showing github.com — 22 free-threading race conditions · Issue #149816 · python/cpython
22 free-threading race conditions #149816Labels interpreter-core(Objects, Python, Grammar, and Parser dirs)(Objects, Python, Grammar, and Parser dirs)topic-free-threadingtype-bugAn unexpected behavior, bug, or errorAn unexpected behavior, bug, or error DescriptionBug report
Changes
2026-05-13: Filed issue.
2026-06-08: Added new findings (16) and (86).
Bug description
I found 22 free-threading race conditions during a partial Xint Code scan of cpython. These only apply to the free-threaded build, and were all found and tested on commit 0534774 built with ./configure --disable-gil --enable-asan on an M3 Mac.
I'm attaching a zip file with detailed writeups of each, as well as reproduction cases for 15 of the 22 issues. I don't have scripts reproducing the findings numbered 21, 36, 82, 94, 106, 115, or 124. Race conditions can be difficult to trigger, and even the scripts I provided are trying to win tight races and may not always work.
I know this is a lot. @colesbury suggested I create a combined issue to make the triage discussion easier.
Let me know if you have any questions or concerns. The issues were found and written up by an automated system, but I have put additional work into the validation and reporting. I do want to be respectful of everyone's time while helping to make Python better.
Writeups
2026-05-13: Initial writeups and test scripts: cpython-ft.zip
2026-06-08: New writeups and test scripts for (16) and (86): cpython-ft-2026-06-08.zip
Many of the scripts are expected to crash, however some of them only produce exceptions or corrupted output. You should be able to run any of the scripts with PYTHON_GIL=1 set in your environment if you want to see a baseline without the race condition.
Finding List
- (16) Cross-interpreter syslog race
- (17) Unlocked
__init__ races with PRNG state access in Modules/_randommodule.c
- (21) Racy EVP_MD cache overwrite leaks references in
Modules/_hashopenssl.c
- (36) Borrowed type lookup races to use-after-free in
Objects/typeobject.c
- (49) SNI callback callable race use-after-free in
Modules/_ssl.c
- (61) Racy weakref head load before incref in
Objects/typeobject.c
- (62) Concurrent kwargs growth causes heap overwrite in
Objects/call.c
- (69) Unsynchronized extra pointer dereference in len in
Modules/_elementtree.c
- (82) Non-atomic list slot memmove in shared list delete in
Objects/listobject.c
- (84) Iterator path bypasses buffered object lock in
Modules/_io/bufferedio.c
- (86) Cross-interpreter XID registry race
- (87) Unsynchronized Element.text borrowed-pointer race in
Modules/_elementtree.c
- (89) Unsynchronized dict iteration causes borrowed-ref UAF in
Modules/_pickle.c
- (91) Racy list item borrow causes UAF in
Modules/_pickle.c
- (94) Async-exception setter races thread-state free in
Python/pystate.c
- (96) Racy GC callback list iteration in
Python/gc_free_threading.c
- (106) Immediate decref races lock-free reader in
Modules/_ctypes/_ctypes.c
- (108) Borrowed dict used after lock release in
Objects/dictobject.c
- (115) Split clear frees keys without QSBR in
Objects/dictobject.c
- (124) Stale keys race in attribute hint fastpath in
Python/bytecodes.c
- (125 wontfix) Struct reinit races with pack/unpack in
Modules/_struct.c
- (128) Borrowed list item raced before incref in
Objects/bytesobject.c
- (129) Reentrant
__index__ causes released-buffer dereference in Objects/memoryobject.c
- (132) Non-atomic exports race in memoryview.hex in
Objects/memoryobject.c
CPython versions tested on:
3.14
Operating systems tested on:
macOS
Linked PRs
Reacted by bpak-theori, Yongtao Huang, Ivy Xu, Thomas Kowalski, Abhishek Tiwari and Edward Xu Reacted by Tim Becker, Duprat and Maurycy Pawłowski-Wieroński Activityskirpichev commented on May 14, 2026
(125) Struct reinit races with pack/unpack in Modules/_struct.c
Is this is same as #143379? lunixbochs commented on May 14, 2026 ContributorAuthor More actions
(125) Struct reinit races with pack/unpack in Modules/_struct.c
Is this is same as #143379?
Looks pretty similar, just as a threading race instead of reentrancy. skirpichev commented on May 14, 2026 Then I think it's not something worth fixing, on same ground. Reacted by Ryan Hileman and Maurycy Pawłowski-Wieroński kumaraditya303 commented on May 14, 2026 I don't think it is worth fixing issues of calling __init__ in multiple threads. lunixbochs commented on May 14, 2026 ContributorAuthor More actions Neither of the mentioned init bugs (struct / random) are init racing with itself, they're init racing with something else. The struct bug is far more contrived because it requires you to call init a second time while using the struct object. The random init/sample race seems fine to fix because if someone ever does manage to hit that they won't even get a crash, just incorrect numbers out of their seeded rng.
There's a separate question on the cpython philosophy of "should a script be able to intentionally crash or memory corrupt the interpreter using the stdlib and no ctypes?", which should inform whether more contrived reentrancy issues / race conditions should be fixed picnixz commented on May 15, 2026 I will have a look at all issues related to SSL and crypto myself in 10 days unless someone beats me to it. added 4 commits that reference this issue on May 15, 2026gh-149816: Fix race condition in `memoryview` with free-threading (#1… gh-149816: Fix a race condition in `_PyBytes_FromList` with free-thre… sobolevn commented on May 16, 2026 (36) does not seem correct, because _PyType_Lookup is supposed to return the borrowed reference by design.
picnixz commented on Oct 12, 2026
MetadataMetadatainterpreter-core(Objects, Python, Grammar, and Parser dirs)(Objects, Python, Grammar, and Parser dirs)topic-free-threadingtype-bugAn unexpected behavior, bug, or errorAn unexpected behavior, bug, or error DevelopmentNo branches or pull requestsIssue actions
|