Skip to content

gh-133171: Make the executor_list thread-safe - #140038

Closed
aisk wants to merge 7 commits into
python:mainfrom
aisk:executor-list-lock
Closed

gh-133171: Make the executor_list thread-safe#140038
aisk wants to merge 7 commits into
python:mainfrom
aisk:executor-list-lock

Conversation

@aisk

@aiskaisk commented Oct 13, 2025•
edited by bedevere-app Bot
Loading

Copy link
Copy Markdown
Member

The idea is simple and naive: just add a PyMutex to the PyInterpreterState for the executor_list under the nogil build, and lock or unlock the mutex in all functions that operate on the executor_list.

Since I'm not very familiar with the related code, there might be a better way to do this, so any feedback would be appreciated!

@bedevere-appbedevere-appBot mentioned this pull request Oct 13, 2025
Open
7 tasks
@aiskaisk changed the title gh-133171: Add lock on operations to executor_listgh-133171: Make the executor_list thread-safeOct 13, 2025
ZeroIntensity
ZeroIntensity reviewed Oct 13, 2025
View reviewed changes
Comment threadInclude/internal/pycore_interp_structs.h
ZeroIntensity
ZeroIntensity reviewed Oct 14, 2025
View reviewed changes

@ZeroIntensityZeroIntensity left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not entirely sure that a mutex is safe here, because we're locking code paths that call Py_DECREF. @Fidget-Spinner, can these paths escape?

Comment threadInclude/internal/pycore_interp_structs.h
@Fidget-Spinner

Fidget-Spinner commented Oct 14, 2025

Copy link
Copy Markdown
Member

In theory, Py_DECREF(executor) could clear the (future) constants pool that we plan to add, which would hold arbitrary constant objects, which can escape.

@corona10

corona10 commented Oct 15, 2025•
edited
Loading

Copy link
Copy Markdown
Member

I am not sure if making executor_list operations thread-safe at this point is good idea.
There is no way to verify that we make the thread-safe until we actually integrate the whole pipeline. I do believe that we have to do more things before we start to work on this task.

@Fidget-Spinner

Fidget-Spinner commented Nov 7, 2025

Copy link
Copy Markdown
Member

@corona10 we need the executor list to be thread safe for it to work with the plan here #133171 (comment).

@aisk could you resume the PR please?

Fidget-Spinner
Fidget-Spinner reviewed Nov 7, 2025
View reviewed changes
Comment threadPython/optimizer.c
void
_Py_Executors_InvalidateAll(PyInterpreterState *interp, int is_invalidation)
{
EXECUTOR_LIST_LOCK(interp);

@Fidget-SpinnerFidget-SpinnerNov 7, 2025

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should be a stop the world event.

Comment threadPython/optimizer.c
}
/* Clearing an executor can deallocate others, so we need to make a list of
* executors to invalidate first */
EXECUTOR_LIST_LOCK(interp);

@Fidget-SpinnerFidget-SpinnerNov 7, 2025

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is fine. nice.!

Comment threadPython/optimizer.c
@corona10

corona10 commented Nov 7, 2025

Copy link
Copy Markdown
Member

I will try to re review this PR in this weekend.

@corona10

corona10 commented Nov 7, 2025

Copy link
Copy Markdown
Member

Share the DM context between @Fidget-Spinner and I.

Before merging this PR, we should investigate the current blocker #133171 (comment) to verify this PR in practice.

@aisk

aisk commented Nov 9, 2025

Copy link
Copy Markdown
MemberAuthor

Hi @ZeroIntensity@corona10@Fidget-Spinner , thanks for the review and feedback. I'll bring this PR up to date while waiting for #133171 (comment) to be resolved.

Fidget-Spinner
Fidget-Spinner reviewed Nov 9, 2025
View reviewed changes
Comment threadPython/optimizer.c
Comment on lines +303 to +307
/* Release the lock before calling _Py_ClearExecutorDeletionList
* to avoid deadlock, since it also tries to acquire the same lock */
EXECUTOR_LIST_UNLOCK(interp);
_Py_ClearExecutorDeletionList(interp);
EXECUTOR_LIST_LOCK(interp);

@Fidget-SpinnerFidget-SpinnerNov 9, 2025

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do we still need this?

@Fidget-Spinner

Fidget-Spinner commented Nov 10, 2025

Copy link
Copy Markdown
Member

Sorry, now that I think about it, locking isn't the right choice here. We should instead move the executor list to tstate.

Really sorry for taking up your time!

Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@Fidget-SpinnerFidget-SpinnerFidget-Spinner left review comments
@ZeroIntensityZeroIntensityZeroIntensity left review comments
@ericsnowcurrentlyericsnowcurrentlyAwaiting requested review from ericsnowcurrentlyericsnowcurrently is a code owner
@markshannonmarkshannonAwaiting requested review from markshannonmarkshannon is a code owner

Assignees

No one assigned

Labels

awaiting reviewskip news

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@aisk@Fidget-Spinner@corona10@ZeroIntensity