Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

gh-115103: Delay reuse of mimalloc pages that store PyObjects #115435

Merged
merged 3 commits into from
Mar 6, 2024

Conversation

colesbury
Copy link
Contributor

@colesbury colesbury commented Feb 13, 2024

This implements the delayed reuse of mimalloc pages that contain Python objects in the free-threaded build.

Allocations of the same size class are grouped in data structures called pages. These are different from operating system pages. For thread-safety, we want to ensure that memory used to store PyObjects remains valid as long as there may be concurrent lock-free readers; we want to delay using it for other size classes, in other heaps, or returning it to the operating system.

When a mimalloc page becomes empty, instead of immediately freeing it, we tag it with a QSBR goal and insert it into a per-thread state linked list of pages to be freed. When mimalloc needs a fresh page, we process the queue and free any still empty pages that are now deemed safe to be freed. Pages waiting to be freed are still available for allocations of the same size class and allocating from a page prevent it from being freed. There is additional logic to handle abandoned pages when threads exit.

See https://peps.python.org/pep-0703/#mimalloc-page-reuse

@colesbury
Copy link
Contributor Author

!buildbot nogil

@bedevere-bot
Copy link

🤖 New build scheduled with the buildbot fleet by @colesbury for commit 99bd660 🤖

The command will test the builders whose names match following regular expression: nogil

The builders matched are:

  • AMD64 Ubuntu NoGIL Refleaks PR
  • x86-64 MacOS Intel ASAN NoGIL PR
  • ARM64 MacOS M1 Refleaks NoGIL PR
  • x86-64 MacOS Intel NoGIL PR
  • AMD64 Windows Server 2022 NoGIL PR
  • AMD64 Ubuntu NoGIL PR
  • ARM64 MacOS M1 NoGIL PR

Copy link
Contributor

@DinoV DinoV left a comment

Choose a reason for hiding this comment

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

Looks good to me, just a few nits really...

mi_page_t *page = llist_data(node, mi_page_t, qsbr_node);
if (!mi_page_all_free(page)) {
// We allocated from this page some point after the delayed free
page->qsbr_goal = 0;
Copy link
Contributor

Choose a reason for hiding this comment

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

nit: Seems like we could consistently use _PyMem_mi_page_clear_qsbr?

return;
}

page->qsbr_goal = 0;
Copy link
Contributor

Choose a reason for hiding this comment

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

_PyMem_mi_page_clear_qsbr?

if (!_Py_qbsr_goal_reached(tstate->qsbr, page->qsbr_goal)) {
return false;
}
_PyMem_mi_page_clear_qsbr(page);
Copy link
Contributor

Choose a reason for hiding this comment

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

This side effect seems surprising given the name of the function... It seems like it'd be better to call _PyMem_mi_page_clear_qsbr explicitly in mi_segment_page_clear and _PyMem_mi_page_maybe_free

Copy link
Contributor Author

@colesbury colesbury left a comment

Choose a reason for hiding this comment

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

I've rebased now that #116343 landed.

  • I fixed a bug with reclaiming segments that led to the buildbot failures (see mi_segment_reclaim). Large pages only contain a single block. Prior to this PR, they were either empty and freed immediately on reclamation or full and so didn't set right_page_reclaimed. The right_page_reclaimed logic doesn't handle large pages properly because large pages use a different allocation code path than small and medium pages. This means we have slightly worse re-use than we'd like when reclaiming large pages, but at least it's correct now.
  • I changed _PyMem_mi_page_is_safe_to_free to only handle reclaimed pages, instead of also _PyMem_mi_page_maybe_free case. The qsbr_goal == 0 has different implications in the two cases. When reclaiming an abandoned page, qsbr_goal == 0 means that we can free it immediately, while in mi_heap_page_collect it means that we haven't started the QSBR process.
  • _mi_page_retire should clear the previous QSBR goal, if one exists (since we just freed a block).

@colesbury
Copy link
Contributor Author

!buildbot nogil

@bedevere-bot
Copy link

🤖 New build scheduled with the buildbot fleet by @colesbury for commit 1645d3a 🤖

The command will test the builders whose names match following regular expression: nogil

The builders matched are:

  • x86-64 MacOS Intel ASAN NoGIL PR
  • ARM64 MacOS M1 NoGIL PR
  • AMD64 Ubuntu NoGIL PR
  • x86-64 MacOS Intel NoGIL PR
  • ARM64 MacOS M1 Refleaks NoGIL PR
  • AMD64 Ubuntu NoGIL Refleaks PR
  • AMD64 Windows Server 2022 NoGIL PR

@colesbury
Copy link
Contributor Author

!buildbot nogil

@bedevere-bot
Copy link

🤖 New build scheduled with the buildbot fleet by @colesbury for commit 3453eb9 🤖

The command will test the builders whose names match following regular expression: nogil

The builders matched are:

  • x86-64 MacOS Intel ASAN NoGIL PR
  • ARM64 MacOS M1 NoGIL PR
  • AMD64 Ubuntu NoGIL PR
  • x86-64 MacOS Intel NoGIL PR
  • ARM64 MacOS M1 Refleaks NoGIL PR
  • AMD64 Ubuntu NoGIL Refleaks PR
  • AMD64 Windows Server 2022 NoGIL PR

@colesbury
Copy link
Contributor Author

The failed buildbot looks unrelated to this PR.

@colesbury colesbury merged commit c012c8a into python:main Mar 6, 2024
38 of 39 checks passed
@colesbury colesbury deleted the gh-115103-mimalloc-qsbr branch March 6, 2024 14:42
@ericsnowcurrently
Copy link
Member

CC @daanx

adorilson pushed a commit to adorilson/cpython that referenced this pull request Mar 25, 2024
…ython#115435)

This implements the delayed reuse of mimalloc pages that contain Python
objects in the free-threaded build.

Allocations of the same size class are grouped in data structures called
pages. These are different from operating system pages. For thread-safety, we
want to ensure that memory used to store PyObjects remains valid as long as
there may be concurrent lock-free readers; we want to delay using it for
other size classes, in other heaps, or returning it to the operating system.

When a mimalloc page becomes empty, instead of immediately freeing it, we tag
it with a QSBR goal and insert it into a per-thread state linked list of
pages to be freed. When mimalloc needs a fresh page, we process the queue and
free any still empty pages that are now deemed safe to be freed. Pages
waiting to be freed are still available for allocations of the same size
class and allocating from a page prevent it from being freed. There is
additional logic to handle abandoned pages when threads exit.
diegorusso pushed a commit to diegorusso/cpython that referenced this pull request Apr 17, 2024
…ython#115435)

This implements the delayed reuse of mimalloc pages that contain Python
objects in the free-threaded build.

Allocations of the same size class are grouped in data structures called
pages. These are different from operating system pages. For thread-safety, we
want to ensure that memory used to store PyObjects remains valid as long as
there may be concurrent lock-free readers; we want to delay using it for
other size classes, in other heaps, or returning it to the operating system.

When a mimalloc page becomes empty, instead of immediately freeing it, we tag
it with a QSBR goal and insert it into a per-thread state linked list of
pages to be freed. When mimalloc needs a fresh page, we process the queue and
free any still empty pages that are now deemed safe to be freed. Pages
waiting to be freed are still available for allocations of the same size
class and allocating from a page prevent it from being freed. There is
additional logic to handle abandoned pages when threads exit.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants