gh-94242: Update calculation of _MAX_WINDOWS_WORKERS #94283
Conversation
The calculation of _MAX_WINDOWS_WORKERS contained two off-by-one errors. They cancelled each other out so this does not change the value of _MAX_WINDOWS_WORKERS, just how it is determined. This could avoid future confusion about where the 61 value comes from. Refs python#94242
|
All commit authors signed the Contributor License Agreement. |
|
Most changes to Python require a NEWS entry. Please add it using the blurb_it web app or the blurb command-line tool. |
| # It can wait on, at most, 64 objects. There is an overhead of two to three | ||
| # objects: | ||
| # - the result queue reader | ||
| # - the thread wakeup reader | ||
| _MAX_WINDOWS_WORKERS = 63 - 2 | ||
| # - sometimes _PyOS_SigintEvent | ||
| _MAX_WINDOWS_WORKERS = 64 - 3 |
There was a problem hiding this comment.
I'd prefer for the previous line to state that on Windows waiting for processes to finish uses _winapi.WaitForMultipleObjects. Ideally, _winapi.MAXIMUM_WAIT_OBJECTS would be defined, which you'd have to add in Modules/_winapi.c. Then we'd have:
# It can wait on, at most, MAXIMUM_WAIT_OBJECTS - 1 (63) objects. An index of
# the wait list is reserved for the interpreter's SIGINT event. There is an
# additional overhead of two objects:
# - the result queue reader
# - the thread wakeup reader
_MAX_WINDOWS_WORKERS = _winapi.MAXIMUM_WAIT_OBJECTS - 3There was a problem hiding this comment.
I understand the desire to use _winapi.MAXIMUM_WAIT_OBJECTS but as a casual contributor to cpython I don't want to start altering _winapi.c - it's not practical for me to test such changes.
There was a problem hiding this comment.
You could still use MAXIMUM_WAIT_OBJECTS in the comment, and make it clear that the SIGINT overhead is due to the _winapi implementation, not overhead added by this module. For example:
# In Windows, waiting for processes uses _winapi.WaitForMultipleObjects.
# It can wait on, at most, MAXIMUM_WAIT_OBJECTS - 1 (64 - 1) objects because
# it reserves an index of the wait list for the interpreter's SIGINT event.
# There is an additional overhead of two objects:
# - the result queue reader
# - the thread wakeup reader
_MAX_WINDOWS_WORKERS = 64 - 3
The calculation of _MAX_WINDOWS_WORKERS contained two off-by-one errors.
They cancelled each other out so this does not change the value of
_MAX_WINDOWS_WORKERS, just how it is determined. This could avoid future
confusion about where the 61 value comes from.