-
Notifications
You must be signed in to change notification settings - Fork 760
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
Added PyRefMap
and PyRefMapMut
for borrowing data nested within a PyRef
/PyRefMut
#4203
base: main
Are you sure you want to change the base?
Conversation
…ts are now opaque, w.r.t. the owner, including both the `PyRef` flavor and the underlying pyclass.
…p` struct is now opaque w.r.t. only the `PyRef` flavor and the `Frozen`-ness of the pyclass so that the struct can use either `PyRef` or `PyRefMut` as the owner. `PyRefMapMut` is no longer opaque since it only needs to support `PyRefMut` where `T` is `Frozen`.
…are now aliases for a single struct `PyRefMapBase` that has a `Mut: Boolean` generic parameter indicating whether it supports `DerefMut`. The underlying `PyRef`/`PyRefMut` flavor is opaque, as well as the pyclass `Frozen`-ness, but the pyclass type is visible.
…ar to the 4th draft, with a single base struct being reused for both `PyRefMap` and `PyRefMapMut`, but the pyclass type is now opaque as well.
PyRefMap
and PyRefMapMut
for borrowing data nested PyRef
/PyRefMut
PyRefMap
and PyRefMapMut
for borrowing data nested within a PyRef
/PyRefMut
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks! I think this could be a bit more simpified ;)
pub struct PyRefMapBase<'py, U: 'py, Mut: Boolean> { | ||
// Either `PyRef` or `PyRefMut` guarding the opaque pyclass from which | ||
// the target is borrowed. | ||
owner: Box<dyn OpaquePyRef<'py>>, | ||
// Pointer to the `Deref` target which is (probably) borrowed from `owner`. | ||
// This pointer is derived from either `&U` or `&mut U` as indicated by | ||
// the `Mut` parameter. | ||
target: NonNull<U>, | ||
// Marks whether mutable methods are supported. If `Mut` is `True` then | ||
// `owner` is a `PyRefMut` and `target` was derived from `&mut U`, so | ||
// the pointer may be mutably dereferenced safely; if `False`, then | ||
// `owner` may be either `PyRef` or `PyRefMut` and `target` was derived | ||
// from `&U` so mutable dereferencing is forbidden. | ||
_mut: PhantomData<Mut> | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I feel like this could be a lot simpler, something like:
pub struct PyRefMap<'py, U: 'py> {
value: NonNull<U>,
borrow: &'py dyn PyClassBorrowChecker,
}
similar for a mutable variant. I haven't tried if that works.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the suggestion, I'll play around with it a bit to see if it can be simplified.
However, I'm not sure that your method would work unless I am missing something. My motivation for storing the original PyRef
/PyRefMut
within the PyRefMap
struct was to hold the borrow on the T: PyClass
that the NonNull<U>
is pointing into so that the pointer is guaranteed to be safely dereferenceable. If I allow the PyRef
to be dropped, its Drop
impl will call release_borrow
and stop protecting T
from being mutably aliased. The inner Bound<T>
would also get dropped, which would decrement the reference count and potentially let T
get garbage collected. The next time PyRefMap
get's dereferenced would then be UB.
So to prevent all this I would have to do something messy, perhaps ManuallyDrop
the PyRef
and then add a Drop
impl for PyRefMap
that releases the borrow; and then I'd still have to find a way to decrement the T
's reference count, which wouldn't be possible if I only had access to &dyn PyClassBorrowChecker
. I'm also not sure how I would even get a reference to the borrow checker with the 'py
lifetime; it seems like the only way to get it is by borrowing it from a &'a Bound<'py, T>
which could only give you &'a dyn PyClassBorrowChecker
. But even if there is a way to get around all this, it seems a lot more complicated/unsafe to me than just storing the original PyRef
and letting it handle all the protection and cleanup with its existing mechanisms.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If I allow the PyRef to be dropped, its Drop impl will call release_borrow and stop protecting...
Of course. I was just sketching out an approach. I would really prefer to not allocate in PyRefMap::new
, this is another approach to erase the pyclass type without boxing things.
A solution would be to refactor that into its own struct:
pub struct BorrowRef<'b>{
borrow: &'b ???,
}
impl Drop for BorrowRef {
...
}
pub struct PyRef<'p, T: PyClass> {
value: NonNull<T>
borrow: BorrowRef<'p>
}
pub struct PyRefMap<'b, U>{
value: NonNull<U>
borrow: BorrowRef<'p>
}
Another solution is your simplified solution that stores the PyRef inline but does not erase the type.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I would really prefer to not allocate in
PyRefMap::new
Yes I agree allocating wouldn't be ideal, and I had pretty mixed feeling about taking that approach.
Also, the current implementation of pyclass and pyref/pyrefmut is fairly complicated. We'd be open to changes to it if that would make this simpler or more efficient.
I took you up on this and was able to simplify PyRef
/PyRefMut
quite a bit without breaking anything. I will submit a PR for it soon to see what you guys think.
Another solution is your simplified solution that stores the PyRef inline but does not erase the type.
I personally do prefer storing the PyRef
inline, and keeping the type un-erased is probably no big deal. Users can just implement their own wrapper if they want to erase the type, so I agree that it isn't worth allocating a box internally.
I have a new implementation you may like based on the reworks to PyRef
/PyRefMut
that I'll be proposing soon, so let's circle back after that.
Also, the current implementation of pyclass and pyref/pyrefmut is fairly complicated. We'd be open to changes to it if that would make this simpler or more efficient. |
Thank you for working on this! I am 👍 on this feature although just wanted to give a heads up it might take me a few days to contribute to any review given it's a tricky one from soundness angle and it's not something I'd actively had in my head until you pushed the PR! 😂 |
No problem at all David! I know I've been spamming you with PR's this week 😉 However, I'd actually like to put this PR on a brief hold until the following PR's get resolved:
|
This draft PR proposes a possible solution for issue #2300 and PR #4195, which involves adding a
Ref::map
-like mechanism for borrowing data nested within aPyRef
orPyRefMut
.Unlike some of the approaches proposed in #2300 which attempt to create a
PyRef<U>
to some typeU
nested within aPyRef<T: PyClass>
, this implementation instead defines a wrapper type that encapsulates the originalPyRef
alongside a raw pointer toU
which it then dereferences to. This has the benefit of not relying on any of thepyo3
internals that may be in flux, and makes the safety relative easy to reason about since there is nothing magic going on.Simplified version
In the simplest case of supporting only immutable borrows from immutable
PyRef
's, the approach boils down to this:This can then be used like so:
Full version
The above method should extend seamlessly to mutable borrows through
PyRefMut
, however this requires adding another wrapper type (e.g.PyRefMutMap
) and all the associatedimpl
's. Then another wrapper may be needed if you want to support shared borrows from aPyRefMut
... this could be somewhat be avoided by making a single wrapper with some extra generic parameters, but then naming the types gets tedious. So, I ultimately decided to use a single base wrapper typePyRefMapBase<'py, U, Mut>
where the containedPyRef<'py, T>
orPyRefMut<'py, T>
is stored as an opaque trait objectBox<dyn OpaquePyRef<'py>>
and the pointer is stored asNonNull<U>
(which can be derived from either&U
or&mut U
). The mutability is then handled with the generic parameterMut: Boolean
which is used in type bounds to statically enforce that mutable dereferences only occur when derived fromPyRefMut<T>
and&mut U
. Finally, two type aliasesPyRefMap<'py, U>
andPyRefMapMut<'py, U>
are defined for the immutable and mutable cases, which are the actual names that users will reference (instead ofPyRefMapBase
directly).This approach of
Box
ing thePyRef
/PyRefMut
has the added benefit of allowing the pyclass typeT
to be erased from the type name so that only the lifetime and the target type need to be included in signatures (i.e. instead ofPyRefMap<'py, MyClass, [i32; 100]>
you simply havePyRefMap<'py, [i32; 100]>
. This comes at the cost of a pointer-sizedBox
allocation when thePyRef
/PyRefMut
is converted to aPyRefMap
/PyRefMapMut
, but I would imagine that is negligible in the context of interfacing with Python.ToDo
pycell
module.PyRef
->PyRefMap
conversions should be associated methods; I know they probably should due to deref conflicts, but I am in denial because standard methods are just so much nicer.compile_fail
tests to make sure it can't be abused.AsRef
/AsMut
impls