The _PyRuntime.ceval structure should be made "per-interpreter". I don't want to make the GIL per-interpreter: that's out of the scope of this issue. So I propose to only move a few fields to make more ceval fields "per interpreter".

New changeset dab8423d220243efabbbcafafc12d90145539b50 by Victor Stinner in branch 'master':
bpo-39984: Add PyInterpreterState.ceval (GH-19047)
https://github.com/python/cpython/commit/dab8423d220243efabbbcafafc12d90145539b50

While trying to move pending to PyInterpreterState, I hit in issue in _PyEval_AddPendingCall(): signal.raise_signal() can call it with tstate=NULL.

See https://bugs.python.org/issue37127#msg364489

New changeset d7fabc116269e4650a684eb04f9ecd84421aa247 by Victor Stinner in branch 'master':
bpo-39984: Pass tstate to handle_signals() (GH-19050)
https://github.com/python/cpython/commit/d7fabc116269e4650a684eb04f9ecd84421aa247

New changeset 23ef89db7ae46d160650263cc80479c2ed6693fb by Victor Stinner in branch 'master':
bpo-39984: _PyThreadState_DeleteCurrent() takes tstate (GH-19051)
https://github.com/python/cpython/commit/23ef89db7ae46d160650263cc80479c2ed6693fb

New changeset 29356e03d4f8800b04f799efe7a10e3ce8b16f61 by Victor Stinner in branch 'master':
bpo-39877: Fix take_gil() for daemon threads (GH-19054)
https://github.com/python/cpython/commit/29356e03d4f8800b04f799efe7a10e3ce8b16f61

New changeset 56bfdebfb17ea9d3245b1f222e92b8e3b1ed6118 by Victor Stinner in branch 'master':
bpo-39984: Pass tstate to _PyEval_SignalAsyncExc() (GH-19049)
https://github.com/python/cpython/commit/56bfdebfb17ea9d3245b1f222e92b8e3b1ed6118

New changeset 8849e5962ba481d5d414b3467a256aba2134b4da by Victor Stinner in branch 'master':
bpo-39984: trip_signal() uses PyGILState_GetThisThreadState() (GH-19061)
https://github.com/python/cpython/commit/8849e5962ba481d5d414b3467a256aba2134b4da

New changeset 50e6e991781db761c496561a995541ca8d83ff87 by Victor Stinner in branch 'master':
bpo-39984: Move pending calls to PyInterpreterState (GH-19066)
https://github.com/python/cpython/commit/50e6e991781db761c496561a995541ca8d83ff87

I moved pending calls to PyInterpreterState.

There are remaining issues with using pending calls in subinterpreters, but I propose to continue the discussion in bpo-37127.

Awesome!  Thanks for doing this, Victor.  I'll take a look when I can and adjust the changes for bpo-33608.  If you'll recall, I made a similar change as part of the solution for that issue, which we later reverted due to problems we discovered with daemon threads during runtime finalization.

> Awesome!  Thanks for doing this, Victor.  I'll take a look when I can and adjust the changes for bpo-33608.  If you'll recall, I made a similar change as part of the solution for that issue, which we later reverted due to problems we discovered with daemon threads during runtime finalization.

I recall the revert dance, yeah. It took one year to fix the daemon threads and related issues! So far, buildbots look happy. I also ran many manual tests which makes me confident that the code is now reliable even with all recent changes done for isolating thread states.

I reopen the issue because of bpo-40082 "trip_signal() gets NULL tstate on Windows on CTRL+C" regression.

> I reopen the issue because of bpo-40082 "trip_signal() gets NULL tstate on Windows on CTRL+C" regression.

bpo-40082 is fixed, so I close the issue again.