Make `PyDict` iterator compatible with free-threaded build #4439

bschoenmaeckers · 2024-08-14T13:31:14Z

This pulls in the PyCriticalSection_Begin & PyCriticalSection_End functions new in 3.13 and use it to lock the PyDict iterators as d
described here. I'm not sure about the PyCriticalSection struct definition. We cannot use the opaque_struct! macro to define this struct because we have to allocate enough space on the stack so we can pass the uninitialized pointer to PyCriticalSection_Begin. So some help would be appreciated!

depends on #4421
related to #4265

ngoldbaum · 2024-08-14T15:34:13Z

I actually have a branch with these changes (more or less) that I was planning to do separately from that PR. Unfortunately the deadlock I found is caused by something else.

If you're planning to work on this stuff I'd appreciate it if you could comment on the tracking issue so we can coordinate work and avoid duplication.

mejrs

This use of the critical section api seems unwise. This allows users to create several critical sections (and worse) allows them to release them in arbitrary order. I don't think I understand the critical section api well but it seems guaranteed to cause issues.

I can see two obvious solutions:

Replace the implementation with PyObject_GetIter and PyIter_Next (slow?)
Implement some form of internal iteration:

impl PyDict{
    pub fn traverse<B>(&self, f: &mut impl FnMut(Bound<'py, PyAny>, Bound<'py, PyAny>) -> ControlFlow<B>) -> ControlFlow<B> {
        struct Guard { .. };
        impl Drop for Guard { ..release critical section }
        
        let mut cs = std::mem::MaybeUninit::zeroed();
        ffi::PyCriticalSection_Begin(cs.as_mut_ptr(), dict.as_ptr());
        let mut ma_used = ..;
        let mut di_used = ..;
        let key = ...;
        let value = ..;
        
        while PyDict_Next(...) != 0{
           f(key, value)?;
        }
        ControlFlow::Continue(())
    }
}

mejrs · 2024-08-14T20:09:22Z

src/types/dict.rs

+            let cs = unsafe {
+                let mut cs = std::mem::MaybeUninit::zeroed();
+                ffi::PyCriticalSection_Begin(cs.as_mut_ptr(), dict.as_ptr());
+                cs.assume_init()
+            };


This can just assume_init immediately because it is zero-valid. This would only be necessary if you used MaybeUninit::uninit().

Suggested change

let cs = unsafe {

let mut cs = std::mem::MaybeUninit::zeroed();

ffi::PyCriticalSection_Begin(cs.as_mut_ptr(), dict.as_ptr());

cs.assume_init()

};

let cs: ffi::PyCriticalSection = unsafe { std::mem::MaybeUninit::zeroed().assume_init() };

unsafe { ffi::PyCriticalSection_Begin(cs.as_mut_ptr(), dict.as_ptr()) };

mejrs · 2024-08-14T20:12:21Z

src/types/dict.rs

+    #[cfg(Py_GIL_DISABLED)]
+    impl Drop for BorrowedDictIter<'_, '_> {
+        fn drop(&mut self) {
+            unsafe {
+                ffi::PyCriticalSection_End(&mut self.cs);
+            }
+        }
+    }


It should probably implement Drop unconditionally (or not at all)

Suggested change

#[cfg(Py_GIL_DISABLED)]

impl Drop for BorrowedDictIter<'_, '_> {

fn drop(&mut self) {

unsafe {

ffi::PyCriticalSection_End(&mut self.cs);

}

}

}

impl Drop for BorrowedDictIter<'_, '_> {

fn drop(&mut self) {

#[cfg(Py_GIL_DISABLED)]

unsafe {

ffi::PyCriticalSection_End(&mut self.cs);

}

}

}

davidhewitt · 2024-08-16T20:44:27Z

Replace the implementation with PyObject_GetIter and PyIter_Next (slow?)

I think we should seriously consider going this way and benchmark if it's actually a performance concern. We already made the same change for sets a couple of releases back, it wasn't a major performance impact there compared to the wins from the Bound API. Two reasons why we did it for sets:

_PySet_Next (or whatever the API was called) was a private API
It doesn't do the right thing for subclasses of sets with custom __iter__ functions

Similarly our current implementation doesn't respect dict subclasses with custom __iter__ functions? Should it? Probably, in which case we might just want to switch to PyObject_GetIter anyway.

ngoldbaum · 2024-08-23T21:40:36Z

I opened #4477 with a different implementation of the FFI bindings.

bschoenmaeckers · 2024-08-23T22:33:43Z

I opened #4477 with a different implementation of the FFI bindings.

Sorry for the late reply. Your implementation looks good, and the 'opaque_type!' use is exactly like what I was looking for. I will update my MR after the weekend.

bschoenmaeckers · 2024-08-23T22:36:07Z

This use of the critical section api seems unwise. This allows users to create several critical sections (and worse) allows them to release them in arbitrary order. I don't think I understand the critical section api well but it seems guaranteed to cause issues.

I can see two obvious solutions:

Replace the implementation with PyObject_GetIter and PyIter_Next (slow?)
Implement some form of internal iteration:

impl PyDict{

    pub fn traverse<B>(&self, f: &mut impl FnMut(Bound<'py, PyAny>, Bound<'py, PyAny>) -> ControlFlow<B>) -> ControlFlow<B> {

        struct Guard { .. };

        impl Drop for Guard { ..release critical section }

        

        let mut cs = std::mem::MaybeUninit::zeroed();

        ffi::PyCriticalSection_Begin(cs.as_mut_ptr(), dict.as_ptr());

        let mut ma_used = ..;

        let mut di_used = ..;

        let key = ...;

        let value = ..;

        

        while PyDict_Next(...) != 0{

           f(key, value)?;

        }

        ControlFlow::Continue(())

    }

}

Interesting solutions 👀. I will try to implement the first one and test the performance hit after.

codspeed-hq · 2024-08-28T17:15:07Z

CodSpeed Performance Report

Merging #4439 will not alter performance

_{Comparing bschoenmaeckers:PyDict_next_lock (e76c7b3) with main (133eac8)}

Summary

✅ 83 untouched benchmarks

ngoldbaum · 2024-08-28T17:21:43Z

ouch that does seem to be a big perf hit

bschoenmaeckers · 2024-08-28T18:59:17Z

Yea this is really bad, but kind of expected as dict.items() creates a copy of the iterable and saves it into a PyList.

https://github.com/python/cpython/blob/main/Objects/dictobject.c#L3381-L3432

bschoenmaeckers · 2024-08-28T19:09:08Z

I've also looked into iterating a raw dict but this only yields the keys. So it does not protect against modifications of the values before fetching them on the next call.

ngoldbaum · 2024-08-28T19:40:29Z

I wonder if the critical section API is actually problematic in practice. You could try iterating over the same dict in many threads on the free-threaded build as a stress test. I'm not sure if there are other usage patterns that @mejrs might be concerned about.

It would be nice if we could still keep the fast path for dicts and then only degrade to the slow path if we're not handed an instance of PyDict_Type.

davidhewitt · 2024-08-28T21:44:32Z

Yea this is really bad, but kind of expected as dict.items() creates a copy of the iterable and saves it into a PyList.

https://github.com/python/cpython/blob/main/Objects/dictobject.c#L3381-L3432

dict.items() is equivalent to the Python 2 semantics where .items() in Python did create a new list. Is perf any better if you try .dict.call_method0("items") to get an iterable items view?

bschoenmaeckers · 2024-08-28T22:49:38Z

src/types/dict.rs

+            let tuple = pair.downcast::<PyTuple>().unwrap();
+            let key = tuple.get_item(0).unwrap();
+            let value = tuple.get_item(1).unwrap();


Is it wise to use the unchecked variants here instead of unwrap?

Now that I think about this, it is probably not safe, because items() method can return an arbitrary object when overridden in python code.

bschoenmaeckers · 2024-08-28T23:06:23Z

Yea this is really bad, but kind of expected as dict.items() creates a copy of the iterable and saves it into a PyList.
https://github.com/python/cpython/blob/main/Objects/dictobject.c#L3381-L3432

dict.items() is equivalent to the Python 2 semantics where .items() in Python did create a new list. Is perf any better if you try .dict.call_method0("items") to get an iterable items view?

I didn't know that this is different, learning something new every day! It is indeed somewhat faster. We went down from ~87% slowdown to ~63%.

bschoenmaeckers · 2024-08-29T16:27:47Z

It would be nice if we could still keep the fast path for dicts and then only degrade to the slow path if we're not handed an instance of PyDict_Type.

I made the previous fast path available to non-freethreaded builds when the dict is not a subtype of PyDict. This gives us minimal performance regressions on existing code.

davidhewitt · 2024-08-31T13:33:01Z

src/types/dict.rs

+
+                if unsafe { ffi::PyDict_Next(dict.as_ptr(), ppos, &mut key, &mut value) } != 0 {
+                    *remaining -= 1;
+                    let py = dict.py();
+                    // Safety:
+                    // - PyDict_Next returns borrowed values
+                    // - we have already checked that `PyDict_Next` succeeded, so we can assume these to be non-null
+                    Some((
+                        unsafe { key.assume_borrowed_unchecked(py) }.to_owned(),
+                        unsafe { value.assume_borrowed_unchecked(py) }.to_owned(),
+                    ))
+                } else {
+                    None
+                }
+            }


Is there an alternative implementation here where we add a critical section internally here just around the call to PyDict_Next? It means that each iteration has to lock / unlock a mutex, which might also be terrible for performance, but it'd be interesting to try. (If it performs acceptably, we could then also ask freethreaded CPython experts if this is sound. My hunch is that it would be.)

I also thought of this implementation but as of the following issue this is not sufficient.

python/cpython#120858

If I'm reading correctly, isn't the point of that issue precisely that it's permitting us to add locking here around each call to PyDict_Next if we so wanted? The concern about borrowed references is not relevant here because we immediately incref them, and we can do that before releasing the critical section. Cc @colesbury

@davidhewitt is right about the borrowed references issue not being relevant here because PyO3 would be doing it's own locking around PyDict_Next() with incref inside the lock.

That's still not ideal, but it might be a reasonable starting point. It's much better to lock around the entire loop, both because of the performance issue and because you will see a consistent view of the dict. The locking only around PyDict_Next() allows for concurrent modifications in between each call, so you're going to have more cases that panic due to concurrent modifications, which would have been prevented by the GIL or a loop-scoped lock.

Another alternative is to copy the dict inside the iterator and iterate over the copy. It's probably cheaper than locking around each call.

Thanks for clearing this up! Copying the dict sounds like the easiest solution for now. When we finalize a critical section api we can consider moving the responsibility of locking the dict (during the whole iteration) on free-threaded builds to the user and remove the copy() and panic on concurrent modifications.

That said, I think if the loop executes arbitrary python code then it is still possible for the dict to be modified during iteration under a critical section, because it may be suspended by a nested section which then modifies the dict.

I feel like users are more likely to be able to know for their use case if copying or locking per iteration is more acceptable. I wonder if we need to split .iter() into multiple methods?

I opened #4571 to suggest a way forward on this.

ngoldbaum · 2024-10-15T19:40:54Z

src/types/dict.rs

@@ -357,6 +366,24 @@ impl<'py> PyDictMethods<'py> for Bound<'py, PyDict> {
        BoundDictIterator::new(self.clone())
    }

+    #[cfg(all(Py_GIL_DISABLED, not(feature = "nightly")))]


Suggested change

#[cfg(all(Py_GIL_DISABLED, not(feature = "nightly")))]

#[cfg(not(feature = "nightly"))]

The critical section macros are a no-op on the GIL-enabled build, so I think allowing this function to work (and rely on the GIL for locking) makes it easier for users to adopt this function.

Agreed, I also think this should be available on all builds, for simplicity.

ngoldbaum · 2024-10-15T19:44:01Z

src/types/dict.rs

+    {
+        #[cfg(feature = "nightly")]
+        {
+            self.iter().try_for_each(|(key, value)| f(key, value))


Oh wait, I see, it's because try_for_each is implemented in terms of try_fold (sorry for all the noise here!).

Can you add a comment explaining the difference?

You got it. I will add a comment to prevent confusion next time.

davidhewitt

Thanks, this is shaping up nicely. Sorry for the delay on my side, I've been a bit busy / exhausted and it felt like a bigger context switch to get back here.

davidhewitt · 2024-10-19T20:12:15Z

src/types/dict.rs

@@ -357,6 +366,24 @@ impl<'py> PyDictMethods<'py> for Bound<'py, PyDict> {
        BoundDictIterator::new(self.clone())
    }

+    #[cfg(all(Py_GIL_DISABLED, not(feature = "nightly")))]


Agreed, I also think this should be available on all builds, for simplicity.

davidhewitt · 2024-10-19T20:13:49Z

src/types/dict.rs

+    /// Iterates over the contents of this dictionary while holding a critical section on the dict.
+    /// This is useful when the GIL is disabled and the dictionary is shared between threads.
+    /// It is not guaranteed that the dictionary will not be modified during iteration when the
+    /// closure calls arbitrary Python code that releases the current critical section.


New doc looks great 👍

I think it's potentially worth mentioning that this is roughly a performance optimization for stable where we can't optimise .iter().try_for_each() at the moment. (And that for infallible iteration, .locked_for_each() has no advantage over .iter().for_each().)

And that for infallible iteration, .locked_for_each() has no advantage over .iter().for_each().

I do not understand the second statement, could you elaborate why this is the case?

I was thinking that we can optimize .for_each already on stable so if they don't need to deal with possible errors, then that's probably more optimal? Maybe I missed something.

Ok I understand now, I will add it to the docs.

davidhewitt · 2024-10-19T20:17:39Z

src/types/dict.rs

+        if dict.is_exact_instance_of::<PyDict>() {
+            return BoundDictIterator::DictIter {
+                dict,
+                ppos: 0,
+                di_used: remaining,
+                remaining,
+            };
+        };
+
+        let items = dict.call_method0(intern!(dict.py(), "items")).unwrap();
+        let iter = PyIterator::from_object(&items).unwrap();
+        BoundDictIterator::ItemIter { iter, remaining }


I think I would still like it split out, yes please. I think it's probably correct in the long run down that road, but I also fear that (a) no other dict methods currently respect subclasses and (b) we probably want to synchronise such a change with lists and tuple iterators too, as well as consider what other methods to adjust at the same time. (Probably in combination with a decision on #4490.)

davidhewitt · 2024-10-19T20:20:12Z

src/types/dict.rs

+            #[cfg(not(Py_GIL_DISABLED))]
+            BoundDictIterator::DictIter { .. } => f(self),
+            #[cfg(Py_GIL_DISABLED)]
+            BoundDictIterator::DictIter { ref dict, .. } => {
+                crate::sync::with_critical_section(dict.clone().as_ref(), || f(self))
+            }


Think this can be simplified given that with_critical_section should optimize away trivially on the gil-enabled build, also can remove a clone:

Suggested change

#[cfg(not(Py_GIL_DISABLED))]

BoundDictIterator::DictIter { .. } => f(self),

#[cfg(Py_GIL_DISABLED)]

BoundDictIterator::DictIter { ref dict, .. } => {

crate::sync::with_critical_section(dict.clone().as_ref(), || f(self))

}

BoundDictIterator::DictIter { dict } => {

crate::sync::with_critical_section(dict.as_any(), || f(self))

}

I don't think the borrow checker will allow this because we need a mut reference to self while iterating. So it will not let is have a second reference without Copy to get a strong ref to Bound.

Ah, good point. I think if we split the BoundDictIterator into the main struct and a private inner, we might be able to make it work? That's potentially a good idea anyway so that we don't have to make the enum members public, and it would be desirable to avoid the redundant .clone() as a performance optimization :)

(So probably gets simpler with #4439 (comment))

davidhewitt · 2024-10-19T20:43:22Z

src/types/dict.rs

+
+    #[inline]
+    #[cfg(all(Py_GIL_DISABLED, not(feature = "nightly")))]
+    fn find_map<B, F>(&mut self, f: F) -> Option<B>


Given the large-ish volume of new code, could we please add some tests for these? Broadly these all seem reasonable, but I would love to have some coverage of them to give me more certainty we don't break them later.

Added tests.

davidhewitt

This looks great to me, thanks very much and sorry for the many rounds of slow reviews!

bschoenmaeckers force-pushed the PyDict_next_lock branch 2 times, most recently from d90dc42 to c0136f7 Compare August 14, 2024 13:57

mejrs requested changes Aug 14, 2024

View reviewed changes

ngoldbaum mentioned this pull request Aug 16, 2024

Tracking issue for no-gil/freethreaded work #4265

Open

14 tasks

ngoldbaum mentioned this pull request Aug 23, 2024

add bindings for critical section API #4477

Merged

bschoenmaeckers force-pushed the PyDict_next_lock branch from c0136f7 to 562fb4e Compare August 28, 2024 16:50

bschoenmaeckers commented Aug 28, 2024

View reviewed changes

bschoenmaeckers changed the title ~~Add PyCriticalSection lock to Dict iterator~~ Make PyDict iterator compatible with free-threaded build Aug 29, 2024

bschoenmaeckers force-pushed the PyDict_next_lock branch 3 times, most recently from 0b8f4c6 to ae0ee72 Compare August 29, 2024 16:53

davidhewitt reviewed Aug 31, 2024

View reviewed changes

bschoenmaeckers force-pushed the PyDict_next_lock branch from d5a7330 to afab9b2 Compare September 3, 2024 18:08

bschoenmaeckers marked this pull request as ready for review September 3, 2024 18:09

bschoenmaeckers force-pushed the PyDict_next_lock branch from afab9b2 to 3ace91a Compare September 3, 2024 18:44

davidhewitt mentioned this pull request Sep 13, 2024

Don't use PyList.get_item_unchecked() on free-threaded build #4539

Merged

ngoldbaum mentioned this pull request Sep 20, 2024

Add locked iterations APIs for dicts and lists #4571

Open

bschoenmaeckers force-pushed the PyDict_next_lock branch 2 times, most recently from ac9fcf9 to f79dc05 Compare October 8, 2024 07:40

ngoldbaum reviewed Oct 15, 2024

View reviewed changes

davidhewitt reviewed Oct 19, 2024

View reviewed changes

davidhewitt added this to the 0.23 milestone Oct 19, 2024

bschoenmaeckers added 13 commits October 21, 2024 16:58

Iterate over dict items in DictIterator

01615c9

Use python instead of C API to get dict items

addfbc3

Use plain dict iterator when not on free-threaded & not a subclass

ce3a916

Copy dict on free-threaded builds to prevent concurrent modifications

8d9b37b

Add test for dict subclass iters

6c05268

Implement PyDict::locked_for_each

5224d5f

Lock BoundDictIterator::next on each iteration

c7ca632

Implement locked fold & try_fold

59615ca

Implement all,any,find,find_map,position when not on nightly

bb4392b

Add changelog

2996df7

Use critical section wrapper

5bdbfb2

Make dict::locked_for_each available on all builds

a656c8b

Remove item() iter

d6d029c

bschoenmaeckers force-pushed the PyDict_next_lock branch from f79dc05 to a5ba737 Compare October 22, 2024 15:59

bschoenmaeckers added 3 commits October 22, 2024 18:18

Add tests for PyDict::iter() reducers

0804996

Add more docs to locked_for_each

3d1ec80

Move iter implementation into inner struct

e76c7b3

bschoenmaeckers force-pushed the PyDict_next_lock branch from a5ba737 to e76c7b3 Compare October 22, 2024 16:18

davidhewitt approved these changes Oct 24, 2024

View reviewed changes

davidhewitt added this pull request to the merge queue Oct 24, 2024

Merged via the queue into PyO3:main with commit 602fbba Oct 24, 2024
77 of 78 checks passed

bschoenmaeckers deleted the PyDict_next_lock branch October 24, 2024 08:01

ngoldbaum mentioned this pull request Dec 10, 2024

Implement locked iteration for PyList #4789

Open

	#[cfg(all(Py_GIL_DISABLED, not(feature = "nightly")))]
	#[cfg(not(feature = "nightly"))]

Make PyDict iterator compatible with free-threaded build #4439

Make PyDict iterator compatible with free-threaded build #4439

Conversation

bschoenmaeckers commented Aug 14, 2024 • edited Loading

ngoldbaum commented Aug 14, 2024

mejrs left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

davidhewitt commented Aug 16, 2024

ngoldbaum commented Aug 23, 2024

bschoenmaeckers commented Aug 23, 2024

bschoenmaeckers commented Aug 23, 2024

codspeed-hq bot commented Aug 28, 2024 • edited Loading

Merging #4439 will not alter performance

Summary

ngoldbaum commented Aug 28, 2024

bschoenmaeckers commented Aug 28, 2024 • edited Loading

bschoenmaeckers commented Aug 28, 2024

ngoldbaum commented Aug 28, 2024

davidhewitt commented Aug 28, 2024

Choose a reason for hiding this comment

bschoenmaeckers Aug 28, 2024 • edited Loading

Choose a reason for hiding this comment

bschoenmaeckers commented Aug 28, 2024 • edited Loading

bschoenmaeckers commented Aug 29, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

This comment was marked as resolved.

ngoldbaum Oct 15, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

davidhewitt left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

davidhewitt left a comment

Choose a reason for hiding this comment

Make `PyDict` iterator compatible with free-threaded build #4439

Make `PyDict` iterator compatible with free-threaded build #4439

bschoenmaeckers commented Aug 14, 2024 •

edited

Loading

mejrs left a comment •

edited

Loading

codspeed-hq bot commented Aug 28, 2024 •

edited

Loading

bschoenmaeckers commented Aug 28, 2024 •

edited

Loading

bschoenmaeckers Aug 28, 2024 •

edited

Loading

bschoenmaeckers commented Aug 28, 2024 •

edited

Loading

ngoldbaum Oct 15, 2024 •

edited

Loading