Inefficient signal handling in multithreaded applications

[vstinner]

BPO	40010
Nosy	@vstinner, @ericsnowcurrently
PRs	bpo-40010: Optimize signal handling in multithreaded applications #19067 bpo-40010: Pass tstate to ceval GIL functions #19077 bpo-39877: 4th take_gil() fix for daemon threads #19080 bpo-40010: COMPUTE_EVAL_BREAKER() checks for subinterpreter #19087 bpo-40010: Optimize pending calls in multithreaded applications #19091

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2020-03-25.17:33:17.981>
created_at = <Date 2020-03-19.00:57:19.125>
labels = ['interpreter-core', '3.9', 'performance']
title = 'Inefficient signal handling in multithreaded applications'
updated_at = <Date 2020-05-05.15:27:40.795>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2020-05-05.15:27:40.795>
actor = 'eric.snow'
assignee = 'none'
closed = True
closed_date = <Date 2020-03-25.17:33:17.981>
closer = 'vstinner'
components = ['Interpreter Core']
creation = <Date 2020-03-19.00:57:19.125>
creator = 'vstinner'
dependencies = []
files = []
hgrepos = []
issue_num = 40010
keywords = ['patch']
message_count = 8.0
messages = ['364580', '364583', '364611', '364628', '364659', '364666', '364667', '368171']
nosy_count = 2.0
nosy_names = ['vstinner', 'eric.snow']
pr_nums = ['19067', '19077', '19080', '19087', '19091']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue40010'
versions = ['Python 3.9']

[vstinner]

When a thread gets a signal, SIGNAL_PENDING_SIGNALS() sets signals_pending to 1 and eval_breaker to 1. In this case, _PyEval_EvalFrameDefault() calls handle_signals(), but since it's not the main thread, it does nothing and signals_pending value remains 1. Moreover, eval_breaker value remains 1 which means that the following code will be called before executing *each* bytecode instruction.

    if (_Py_atomic_load_relaxed(eval_breaker)) {
        (...)
        opcode = _Py_OPCODE(*next_instr);
        if (opcode == SETUP_FINALLY || ...) {
            ...
        }
        if (_Py_atomic_load_relaxed(&ceval->signals_pending)) {
            if (handle_signals(tstate) != 0) {
                goto error;
            }
        }
        if (_Py_atomic_load_relaxed(&ceval->pending.calls_to_do)) {
            ...
        }

        if (_Py_atomic_load_relaxed(&ceval->gil_drop_request)) {
            ...
        }
        if (tstate->async_exc != NULL) {
            ...
        }
    }

This is inefficient.

I'm working on a PR modifying SIGNAL_PENDING_SIGNALS() to not set eval_breaker to 1 if the current thread is not the main thread.

[vstinner]

I found this issue while working on PR 19066 of bpo-39984.

[vstinner]

New changeset 5a3a71d by Victor Stinner in branch 'master':
bpo-40010: Optimize signal handling in multithreaded applications (GH-19067)
5a3a71d

[vstinner]

New changeset a36adfa by Victor Stinner in branch 'master':
bpo-39877: 4th take_gil() fix for daemon threads (GH-19080)
a36adfa

[vstinner]

New changeset da2914d by Victor Stinner in branch 'master':
bpo-40010: Pass tstate to ceval GIL functions (GH-19077)
da2914d

[vstinner]

New changeset d2a8e5b by Victor Stinner in branch 'master':
bpo-40010: COMPUTE_EVAL_BREAKER() checks for subinterpreter (GH-19087)
d2a8e5b

[vstinner]

New changeset d831688 by Victor Stinner in branch 'master':
bpo-40010: Optimize pending calls in multithreaded applications (GH-19091)
d831688

[ericsnowcurrently]

Good catch on this, Victor. Thanks for doing it.