TSAN race in _waiting_release

[ccotter]

Bug report

Bug description:

A TSAN (gcc-12) build of the latest version of cpython reports a race in the test_send_recv_different_interpreters_and_threads test:

  Read of size 4 at 0x7ffe8d0c9338 by main thread:
    #0 _waiting_finish_releasing Modules/_interpchannelsmodule.c:504 (_interpchannels.cpython-314-x86_64-linux-gnu.so+0x88f4) (BuildId: 233ea3cec8e2d6b4a5322cdab2aa3fcc24ab85f8)
    #1 channel_send_wait Modules/_interpchannelsmodule.c:1852 (_interpchannels.cpython-314-x86_64-linux-gnu.so+0x88f4)
    #2 channelsmod_send Modules/_interpchannelsmodule.c:3096 (_interpchannels.cpython-314-x86_64-linux-gnu.so+0x8dd0) (BuildId: 233ea3cec8e2d6b4a5322cdab2aa3fcc24ab85f8)
    #3 cfunction_call Objects/methodobject.c:540 (python+0x549e55) (BuildId: f667f6a4a823ca3db627f9e7bf3836f7af978c6d)

    #0 _waiting_release Modules/_interpchannelsmodule.c:498 (_interpchannels.cpython-314-x86_64-linux-gnu.so+0x37c7) (BuildId: 233ea3cec8e2d6b4a5322cdab2aa3fcc24ab85f8)
    #1 channel_recv Modules/_interpchannelsmodule.c:1937 (_interpchannels.cpython-314-x86_64-linux-gnu.so+0x863a) (BuildId: 233ea3cec8e2d6b4a5322cdab2aa3fcc24ab85f8)
    #2 channelsmod_recv Modules/_interpchannelsmodule.c:3189 (_interpchannels.cpython-314-x86_64-linux-gnu.so+0x863a)

Indeed,

cpython/Modules/_interpchannelsmodule.c

Line 504 in 8b209fd

while (waiting->status == WAITING_RELEASING) {

spin waits for waiting->status to change state, with

cpython/Modules/_interpchannelsmodule.c

Line 498 in 8b209fd

waiting->status = WAITING_RELEASED;

triggering the spinning thread to exit the spin. Neither happens under a mutex, or with any release-acquire sequencing.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-123828: fix data raceing for _waiting_release method in _interpchannelsmodule.c #124107

[ZeroIntensity]

cc @ericsnowcurrently

[n-bes]

Can confirm, reproduced on Python-3.13.0b4

• Build 182

  • build script
  • logs

• Build 183

  • build script
  • logs

• Build 186

  • build script
  • logs

p.s. summary table

If I remember correctly, this OOM related with this issue