[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <ZvluU69LMB6nuD6r@pollux>
Date: Sun, 29 Sep 2024 17:12:19 +0200
From: Danilo Krummrich <dakr@...nel.org>
To: Gary Guo <gary@...yguo.net>
Cc: ojeda@...nel.org, alex.gaynor@...il.com, wedsonaf@...il.com,
boqun.feng@...il.com, bjorn3_gh@...tonmail.com,
benno.lossin@...ton.me, a.hindborg@...sung.com,
aliceryhl@...gle.com, akpm@...ux-foundation.org,
daniel.almeida@...labora.com, faith.ekstrand@...labora.com,
boris.brezillon@...labora.com, lina@...hilina.net,
mcanal@...lia.com, zhiw@...dia.com, cjia@...dia.com,
jhubbard@...dia.com, airlied@...hat.com, ajanulgu@...hat.com,
lyude@...hat.com, linux-kernel@...r.kernel.org,
rust-for-linux@...r.kernel.org, linux-mm@...ck.org
Subject: Re: [PATCH v7 15/26] rust: alloc: implement `collect` for `IntoIter`
On Sat, Sep 28, 2024 at 08:27:34PM +0100, Gary Guo wrote:
> On Thu, 12 Sep 2024 00:52:51 +0200
> Danilo Krummrich <dakr@...nel.org> wrote:
>
> > Currently, we can't implement `FromIterator`. There are a couple of
> > issues with this trait in the kernel, namely:
> >
> > - Rust's specialization feature is unstable. This prevents us to
> > optimze for the special case where `I::IntoIter` equals `Vec`'s
> > `IntoIter` type.
> > - We also can't use `I::IntoIter`'s type ID either to work around this,
> > since `FromIterator` doesn't require this type to be `'static`.
> > - `FromIterator::from_iter` does return `Self` instead of
> > `Result<Self, AllocError>`, hence we can't properly handle allocation
> > failures.
> > - Neither `Iterator::collect` nor `FromIterator::from_iter` can handle
> > additional allocation flags.
> >
> > Instead, provide `IntoIter::collect`, such that we can at least convert
> > `IntoIter` into a `Vec` again.
> >
> > Reviewed-by: Alice Ryhl <aliceryhl@...gle.com>
> > Signed-off-by: Danilo Krummrich <dakr@...nel.org>
>
> A question is how useful is this? The way this can be used seems
> fairly limited: you `into_iter`, consume a few elements, and the
> `collect()`?
Well, it allows us to convert back to a `Vec`, which we otherwise can't until we
implement our own `FromIterator` trait.
Also note that we want to have this specialization of `collect` for performance
reasons anyways. The Rust stdlib uses specialization traits (which aren't yet
stable) for this optimization.
>
> It feels whatever user this serves, it would make more sense for them
> to use a `VecDeque` and just `pop_front()`.
We don't have `VecDeque` (yet).
>
> Also, inline comments below.
>
> > ---
> > rust/kernel/alloc/kvec.rs | 86 +++++++++++++++++++++++++++++++++++++++
> > 1 file changed, 86 insertions(+)
> >
> > diff --git a/rust/kernel/alloc/kvec.rs b/rust/kernel/alloc/kvec.rs
> > index e91761c5c52d..686e969463f8 100644
> > --- a/rust/kernel/alloc/kvec.rs
> > +++ b/rust/kernel/alloc/kvec.rs
> > @@ -690,6 +690,92 @@ impl<T, A> IntoIter<T, A>
> > fn as_raw_mut_slice(&mut self) -> *mut [T] {
> > ptr::slice_from_raw_parts_mut(self.ptr, self.len)
> > }
> > +
> > + fn into_raw_parts(self) -> (*mut T, NonNull<T>, usize, usize) {
> > + let me = ManuallyDrop::new(self);
> > + let ptr = me.ptr;
> > + let buf = me.buf;
> > + let len = me.len;
> > + let cap = me.cap;
> > + (ptr, buf, len, cap)
> > + }
> > +
> > + /// Same as `Iterator::collect` but specialized for `Vec`'s `IntoIter`.
> > + ///
> > + /// # Examples
> > + ///
> > + /// ```
> > + /// let v = kernel::kvec![1, 2, 3]?;
> > + /// let mut it = v.into_iter();
> > + ///
> > + /// assert_eq!(it.next(), Some(1));
> > + ///
> > + /// let v = it.collect(GFP_KERNEL);
> > + /// assert_eq!(v, [2, 3]);
> > + ///
> > + /// # Ok::<(), Error>(())
> > + /// ```
> > + /// # Implementation Details
> > + ///
> > + /// Currently, we can't implement `FromIterator`. There are a couple of issues with this trait
> > + /// in the kernel, namely:
> > + ///
> > + /// - Rust's specialization feature is unstable. This prevents us to optimze for the special
> > + /// case where `I::IntoIter` equals `Vec`'s `IntoIter` type.
> > + /// - We also can't use `I::IntoIter`'s type ID either to work around this, since `FromIterator`
> > + /// doesn't require this type to be `'static`.
> > + /// - `FromIterator::from_iter` does return `Self` instead of `Result<Self, AllocError>`, hence
> > + /// we can't properly handle allocation failures.
> > + /// - Neither `Iterator::collect` nor `FromIterator::from_iter` can handle additional allocation
> > + /// flags.
> > + ///
> > + /// Instead, provide `IntoIter::collect`, such that we can at least convert a `IntoIter` into a
> > + /// `Vec` again.
> > + ///
> > + /// Note that `IntoIter::collect` doesn't require `Flags`, since it re-uses the existing backing
> > + /// buffer. However, this backing buffer may be shrunk to the actual count of elements.
> > + pub fn collect(self, flags: Flags) -> Vec<T, A> {
> > + let (mut ptr, buf, len, mut cap) = self.into_raw_parts();
> > + let has_advanced = ptr != buf.as_ptr();
> > +
> > + if has_advanced {
> > + // Copy the contents we have advanced to at the beginning of the buffer.
> > + //
> > + // SAFETY:
> > + // - `ptr` is valid for reads of `len * size_of::<T>()` bytes,
> > + // - `buf.as_ptr()` is valid for writes of `len * size_of::<T>()` bytes,
> > + // - `ptr` and `buf.as_ptr()` are not be subject to aliasing restrictions relative to
> > + // each other,
> > + // - both `ptr` and `buf.ptr()` are properly aligned.
> > + unsafe { ptr::copy(ptr, buf.as_ptr(), len) };
> > + ptr = buf.as_ptr();
> > + }
> > +
> > + // This can never fail, `len` is guaranteed to be smaller than `cap`.
> > + let layout = core::alloc::Layout::array::<T>(len).unwrap();
>
> nit: could be `unwrap_unchecked()`. Although feel free to leave it to
> avoid an additional unsafe.
This will be replaced with `ArrayLayout` anyways.
>
> > +
> > + // SAFETY: `buf` points to the start of the backing buffer and `len` is guaranteed to be
> > + // smaller than `cap`. Depending on `alloc` this operation may shrink the buffer or leaves
> > + // it as it is.
> > + ptr = match unsafe { A::realloc(Some(buf.cast()), layout, flags) } {
> > + // If we fail to shrink, which likely can't even happen, continue with the existing
> > + // buffer.
> > + Err(_) => ptr,
> > + Ok(ptr) => {
> > + cap = len;
> > + ptr.as_ptr().cast()
> > + }
> > + };
>
> This should be moved to `Vec::shrink_to_fit`. And then this function
> can just `Vec::from_raw_parts(...)` and then `vec.shrink_to_fit`.
I'll put it on my list for a follow up patch, I really think we should focus on
landing the series now.
>
> But my question would be why this function needs to shrink in the first
> place.
It's meant as an optimization for `Iterator::collect` in the mid / long term. A
user expects that the resulting allocation isn't larger than actually needed for
`Iterator::collect`.
>
> > +
> > + // SAFETY: If the iterator has been advanced, the advanced elements have been copied to
> > + // the beginning of the buffer and `len` has been adjusted accordingly.
> > + //
> > + // - `ptr` is guaranteed to point to the start of the backing buffer.
> > + // - `cap` is either the original capacity or, after shrinking the buffer, equal to `len`.
> > + // - `alloc` is guaranteed to be unchanged since `into_iter` has been called on the original
> > + // `Vec`.
> > + unsafe { Vec::from_raw_parts(ptr, len, cap) }
> > + }
> > }
> >
> > impl<T, A> Iterator for IntoIter<T, A>
>
Powered by blists - more mailing lists