* [PATCH v5 01/12] rust: xarray: minor formatting fixes
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
@ 2026-09-02 13:25 ` Andreas Hindborg
2026-09-02 13:25 ` [PATCH v5 02/12] rust: xarray: add debug format for `StoreError` Andreas Hindborg
` (10 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:25 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm, Daniel Gomez, Mukesh Kumar Chaurasiya (IBM),
Tamir Duberstein, Liam R. Howlett
Fix formatting in xarray module to comply with kernel coding
guidelines:
- Update use clauses to use vertical layout with each import on its
own line.
- Add trailing empty comments to preserve formatting and prevent
rustfmt from collapsing imports.
- Break long assert_eq! statement in documentation across multiple
lines for better readability.
Reviewed-by: Gary Guo <gary@garyguo.net>
Reviewed-by: Tamir Duberstein <tamird@gmail.com>
Acked-by: Tamir Duberstein <tamird@gmail.com>
Reviewed-by: Daniel Gomez <da.gomez@samsung.com>
Acked-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Mukesh Kumar Chaurasiya (IBM) <mkchauras@gmail.com>
Reviewed-by: Daniel Almeida <daniel.almeida@collabora.com>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/kernel/xarray.rs | 30 +++++++++++++++++++++++-------
1 file changed, 23 insertions(+), 7 deletions(-)
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index 987c9c0c21989..02f93ae1f92de 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -4,20 +4,33 @@
//!
//! C header: [`include/linux/xarray.h`](srctree/include/linux/xarray.h)
-use crate::{
+use core::{
+ iter,
+ marker::PhantomData,
+ pin::Pin,
+ ptr::NonNull, //
+};
+use kernel::{
alloc,
bindings,
- build_assert::build_assert,
- error::{Error, Result},
+ build_assert::build_assert, //
+ error::{
+ Error,
+ Result, //
+ },
ffi::c_void,
types::{
ForeignOwnable,
NotThreadSafe,
Opaque, //
- }, //
+ },
+};
+use pin_init::{
+ pin_data,
+ pin_init,
+ pinned_drop,
+ PinInit, //
};
-use core::{iter, marker::PhantomData, pin::Pin, ptr::NonNull};
-use pin_init::{pin_data, pin_init, pinned_drop, PinInit};
/// An array which efficiently maps sparse integer indices to owned objects.
///
@@ -50,7 +63,10 @@
/// *guard.get_mut(0).unwrap() = 0xffff;
/// assert_eq!(guard.get(0).copied(), Some(0xffff));
///
-/// assert_eq!(guard.store(0, beef, GFP_KERNEL)?.as_deref().copied(), Some(0xffff));
+/// assert_eq!(
+/// guard.store(0, beef, GFP_KERNEL)?.as_deref().copied(),
+/// Some(0xffff)
+/// );
/// assert_eq!(guard.get(0).copied(), Some(0xbeef));
///
/// guard.remove(0);
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 02/12] rust: xarray: add debug format for `StoreError`
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
2026-09-02 13:25 ` [PATCH v5 01/12] rust: xarray: minor formatting fixes Andreas Hindborg
@ 2026-09-02 13:25 ` Andreas Hindborg
2026-09-02 13:25 ` [PATCH v5 03/12] xarray: move xas_result() and xa_zero_to_null() to the header Andreas Hindborg
` (9 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:25 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm, Daniel Gomez, Tamir Duberstein, Liam R. Howlett
Add a `Debug` implementation for `StoreError<T>` to enable better error
reporting and debugging. The implementation only displays the `error`
field and omits the `value` field, as `T` may not implement `Debug`.
Reviewed-by: Gary Guo <gary@garyguo.net>
Reviewed-by: Daniel Gomez <da.gomez@samsung.com>
Acked-by: Tamir Duberstein <tamird@gmail.com>
Acked-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Reviewed-by: Daniel Almeida <daniel.almeida@collabora.com>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/kernel/xarray.rs | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index 02f93ae1f92d..4335caab8cca 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -19,6 +19,7 @@
Result, //
},
ffi::c_void,
+ fmt,
types::{
ForeignOwnable,
NotThreadSafe,
@@ -193,6 +194,14 @@ pub struct StoreError<T> {
pub value: T,
}
+impl<T> fmt::Debug for StoreError<T> {
+ fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+ f.debug_struct("StoreError")
+ .field("error", &self.error)
+ .finish()
+ }
+}
+
impl<T> From<StoreError<T>> for Error {
#[inline]
fn from(value: StoreError<T>) -> Self {
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 03/12] xarray: move xas_result() and xa_zero_to_null() to the header
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
2026-09-02 13:25 ` [PATCH v5 01/12] rust: xarray: minor formatting fixes Andreas Hindborg
2026-09-02 13:25 ` [PATCH v5 02/12] rust: xarray: add debug format for `StoreError` Andreas Hindborg
@ 2026-09-02 13:25 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 04/12] rust: xarray: add `XArrayState` Andreas Hindborg
` (8 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:25 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm
The functions xas_result() and xa_zero_to_null() are static in
lib/xarray.c. Upcoming Rust XArray changes need to call them from
Rust helper functions. Move them to include/linux/xarray.h as static
inline functions, following the approach of commit 79ada2ae6615
("xarray: extract helper from __xa_{insert,cmpxchg}").
No functional change.
Suggested-by: Tamir Duberstein <tamird@kernel.org>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
include/linux/xarray.h | 26 ++++++++++++++++++++++++++
lib/xarray.c | 12 ------------
2 files changed, 26 insertions(+), 12 deletions(-)
diff --git a/include/linux/xarray.h b/include/linux/xarray.h
index be850174e802e..db2520864aaa7 100644
--- a/include/linux/xarray.h
+++ b/include/linux/xarray.h
@@ -191,6 +191,17 @@ static inline bool xa_is_zero(const void *entry)
return unlikely(entry == XA_ZERO_ENTRY);
}
+/**
+ * xa_zero_to_null() - Convert an internal zero entry into a NULL pointer.
+ * @entry: XArray entry.
+ *
+ * Return: %NULL if @entry is a zero entry, @entry otherwise.
+ */
+static inline void *xa_zero_to_null(void *entry)
+{
+ return xa_is_zero(entry) ? NULL : entry;
+}
+
/**
* xa_is_err() - Report whether an XArray operation returned an error
* @entry: Result from calling an XArray function
@@ -1437,6 +1448,21 @@ static inline int xas_error(const struct xa_state *xas)
return xa_err(xas->xa_node);
}
+/**
+ * xas_result() - Extract the result of an XArray operation.
+ * @xas: XArray operation state.
+ * @curr: Entry returned by the operation.
+ *
+ * Return: @curr if the operation succeeded, the error encoded in @xas
+ * otherwise.
+ */
+static inline void *xas_result(struct xa_state *xas, void *curr)
+{
+ if (xas_error(xas))
+ curr = xas->xa_node;
+ return curr;
+}
+
/**
* xas_set_err() - Note an error in the xa_state.
* @xas: XArray operation state.
diff --git a/lib/xarray.c b/lib/xarray.c
index 9a8b4916540cf..cddb44f6b4ab2 100644
--- a/lib/xarray.c
+++ b/lib/xarray.c
@@ -437,11 +437,6 @@ static unsigned long max_index(void *entry)
return (XA_CHUNK_SIZE << xa_to_node(entry)->shift) - 1;
}
-static inline void *xa_zero_to_null(void *entry)
-{
- return xa_is_zero(entry) ? NULL : entry;
-}
-
static void xas_shrink(struct xa_state *xas)
{
struct xarray *xa = xas->xa;
@@ -1624,13 +1619,6 @@ void *xa_load(struct xarray *xa, unsigned long index)
}
EXPORT_SYMBOL(xa_load);
-static void *xas_result(struct xa_state *xas, void *curr)
-{
- if (xas_error(xas))
- curr = xas->xa_node;
- return curr;
-}
-
/**
* __xa_erase() - Erase this entry from the XArray while locked.
* @xa: XArray.
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 04/12] rust: xarray: add `XArrayState`
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (2 preceding siblings ...)
2026-09-02 13:25 ` [PATCH v5 03/12] xarray: move xas_result() and xa_zero_to_null() to the header Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 05/12] rust: xarray: simplify `Guard::load` Andreas Hindborg
` (7 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm
Add `XArrayState` as internal state for XArray iteration and entry
operations. This struct wraps the C `xa_state` structure and holds a
reference to a `Guard` to ensure exclusive access to the XArray for the
lifetime of the state object. `XArrayState` is generic over the guard
borrow through a `Deref` bound, so it can hold either a shared or a
mutable reference to the guard.
Use the new state in `Guard::load` by replacing the call to `xa_load`
with `xas_load`. The `xa_load` function takes the RCU lock internally,
which we do not need, since the `Guard` already holds an exclusive lock
on the `XArray`. The `xas_load` function operates on `xa_state` and
assumes the required locks are already held.
Unlike `xa_load`, `xas_load` does not filter out internal entries.
Arrays created with `AllocKind::Alloc1` store `XA_ZERO_ENTRY` at index
0 when they are expanded from empty, so convert zero entries to `NULL`
like the C normal API does, exposing the `xa_zero_to_null` helper for
this purpose.
The `XAS_RESTART` constant is also exposed through the bindings helper
to properly initialize the `xa_node` field.
The `guard` field of `XArrayState` is not read until a later patch
adds `into_guard`, so it is annotated with `#[expect(dead_code)]`
until then.
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/bindings/bindings_helper.h | 7 ++++
rust/helpers/xarray.c | 5 +++
rust/kernel/xarray.rs | 74 ++++++++++++++++++++++++++++++++++++++---
3 files changed, 81 insertions(+), 5 deletions(-)
diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h
index 1124785e210b3..419e6b74fedc3 100644
--- a/rust/bindings/bindings_helper.h
+++ b/rust/bindings/bindings_helper.h
@@ -124,6 +124,13 @@ const xa_mark_t RUST_CONST_HELPER_XA_PRESENT = XA_PRESENT;
const gfp_t RUST_CONST_HELPER_XA_FLAGS_ALLOC = XA_FLAGS_ALLOC;
const gfp_t RUST_CONST_HELPER_XA_FLAGS_ALLOC1 = XA_FLAGS_ALLOC1;
+/*
+ * `XAS_RESTART` is `((struct xa_node *)3UL)` -- a sentinel pointer value, not
+ * an address. Cast to `size_t` so bindgen emits a plain `usize` constant; for
+ * pointer-typed macro values bindgen otherwise generates a `pub static mut`,
+ * see https://github.com/rust-lang/rust-bindgen/issues/3347.
+ */
+const size_t RUST_CONST_HELPER_XAS_RESTART = (size_t)XAS_RESTART;
const vm_flags_t RUST_CONST_HELPER_VM_MERGEABLE = VM_MERGEABLE;
const vm_flags_t RUST_CONST_HELPER_VM_READ = VM_READ;
diff --git a/rust/helpers/xarray.c b/rust/helpers/xarray.c
index 08979b3043410..79799c55c3d73 100644
--- a/rust/helpers/xarray.c
+++ b/rust/helpers/xarray.c
@@ -26,3 +26,8 @@ __rust_helper void rust_helper_xa_unlock(struct xarray *xa)
{
return xa_unlock(xa);
}
+
+__rust_helper void *rust_helper_xa_zero_to_null(void *entry)
+{
+ return xa_zero_to_null(entry);
+}
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index 4335caab8ccae..e8082df2b4797 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -8,7 +8,10 @@
iter,
marker::PhantomData,
pin::Pin,
- ptr::NonNull, //
+ ptr::{
+ null_mut,
+ NonNull, //
+ },
};
use kernel::{
alloc,
@@ -214,10 +217,8 @@ fn load<F, U>(&self, index: usize, f: F) -> Option<U>
where
F: FnOnce(NonNull<c_void>) -> U,
{
- // SAFETY: `self.xa.xa` is always valid by the type invariant.
- let ptr = unsafe { bindings::xa_load(self.xa.xa.get(), index) };
- let ptr = NonNull::new(ptr.cast())?;
- Some(f(ptr))
+ let mut state = XArrayState::new(self, index);
+ Some(f(state.load()?))
}
/// Provides a reference to the element at the given index.
@@ -300,6 +301,69 @@ pub fn store(
}
}
+/// Internal state for XArray iteration and entry operations.
+///
+/// `R` is the borrow held on the guard: either `&Guard` for read-only callers
+/// or `&mut Guard` for entry-style APIs that need to surrender the borrow back
+/// via [`XArrayState::into_guard`].
+///
+/// # Invariants
+///
+/// - `state` is always a valid `bindings::xa_state`.
+/// - `state.xa` aliases the xarray reachable through `guard`.
+pub(crate) struct XArrayState<R> {
+ // The borrow is held to guarantee exclusive access to the array. It is
+ // not read until a later patch adds `into_guard`, so silence the dead
+ // code warning until then.
+ #[expect(dead_code)]
+ guard: R,
+ state: bindings::xa_state,
+}
+
+impl<'a, R, T> XArrayState<R>
+where
+ T: ForeignOwnable + 'a,
+ R: core::ops::Deref<Target = Guard<'a, T>>,
+{
+ #[inline]
+ fn new(guard: R, index: usize) -> Self {
+ let xa_ptr = guard.xa.xa.get();
+ // INVARIANT: `state` is initialized to a valid `xa_state` whose `xa` field aliases the
+ // xarray reachable through `guard`.
+ Self {
+ guard,
+ state: bindings::xa_state {
+ xa: xa_ptr,
+ xa_index: index,
+ xa_shift: 0,
+ xa_sibs: 0,
+ xa_offset: 0,
+ xa_pad: 0,
+ xa_node: bindings::XAS_RESTART as *mut bindings::xa_node,
+ xa_alloc: null_mut(),
+ xa_update: None,
+ xa_lru: null_mut(),
+ },
+ }
+ }
+
+ fn load(&mut self) -> Option<NonNull<c_void>> {
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant. By the same
+ // invariant, `self.state.xa` aliases the xarray reachable through `self.guard`, whose
+ // lock we hold.
+ let ptr = unsafe { bindings::xas_load(&raw mut self.state) };
+
+ // Unlike the normal API, `xas_load` does not filter out internal entries. Arrays
+ // created with [`AllocKind::Alloc1`] store `XA_ZERO_ENTRY` at index 0 when they are
+ // expanded from empty, so convert zero entries to `NULL` like `xa_load` does. Retry
+ // entries cannot be observed here because they require concurrent modification of the
+ // array, and we hold the lock.
+ //
+ // SAFETY: `xa_zero_to_null` only inspects the value of `ptr`.
+ NonNull::new(unsafe { bindings::xa_zero_to_null(ptr) }.cast())
+ }
+}
+
// SAFETY: `XArray<T>` has no shared mutable state so it is `Send` iff `T` is `Send`.
unsafe impl<T: ForeignOwnable + Send> Send for XArray<T> {}
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 05/12] rust: xarray: simplify `Guard::load`
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (3 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 04/12] rust: xarray: add `XArrayState` Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 06/12] rust: xarray: add `find_next` and `find_next_mut` Andreas Hindborg
` (6 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm
Simplify the implementation by removing the closure-based API from
`Guard::load` in favor of returning `Option<NonNull<c_void>>` directly.
The closure-based API existed to avoid passing around untyped
pointers. The following patches add find and entry operations that
need to store the returned pointer in entry objects and pass it
between internal functions, where the closure style does not scale.
Change `load` to return the pointer directly, establishing the style
used for the rest of the series. Pointers are still only converted to
references at the public API boundary.
Reviewed-by: Daniel Almeida <daniel.almeida@collabora.com>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/kernel/xarray.rs | 25 +++++++++++--------------
1 file changed, 11 insertions(+), 14 deletions(-)
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index e8082df2b479..a14f874ad630 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -213,28 +213,25 @@ fn from(value: StoreError<T>) -> Self {
}
impl<'a, T: ForeignOwnable> Guard<'a, T> {
- fn load<F, U>(&self, index: usize, f: F) -> Option<U>
- where
- F: FnOnce(NonNull<c_void>) -> U,
- {
- let mut state = XArrayState::new(self, index);
- Some(f(state.load()?))
+ #[inline]
+ fn load(&self, index: usize) -> Option<NonNull<c_void>> {
+ XArrayState::new(self, index).load()
}
/// Provides a reference to the element at the given index.
+ #[inline]
pub fn get(&self, index: usize) -> Option<T::Borrowed<'_>> {
- self.load(index, |ptr| {
- // SAFETY: `ptr` came from `T::into_foreign`.
- unsafe { T::borrow(ptr.as_ptr()) }
- })
+ let ptr = self.load(index)?;
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ Some(unsafe { T::borrow(ptr.as_ptr()) })
}
/// Provides a mutable reference to the element at the given index.
+ #[inline]
pub fn get_mut(&mut self, index: usize) -> Option<T::BorrowedMut<'_>> {
- self.load(index, |ptr| {
- // SAFETY: `ptr` came from `T::into_foreign`.
- unsafe { T::borrow_mut(ptr.as_ptr()) }
- })
+ let ptr = self.load(index)?;
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ Some(unsafe { T::borrow_mut(ptr.as_ptr()) })
}
/// Removes and returns the element at the given index.
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 06/12] rust: xarray: add `find_next` and `find_next_mut`
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (4 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 05/12] rust: xarray: simplify `Guard::load` Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 07/12] rust: xarray: add entry API Andreas Hindborg
` (5 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm
Add methods to find the next element in an XArray starting from a
given index. The methods return a tuple containing the index where the
element was found and a reference to the element.
The implementation uses the XArray state API via `xas_find` to avoid taking
the rcu lock as an exclusive lock is already held by `Guard`.
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/kernel/xarray.rs | 101 ++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 101 insertions(+)
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index a14f874ad630..9993783fc854 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -219,6 +219,22 @@ fn load(&self, index: usize) -> Option<NonNull<c_void>> {
}
/// Provides a reference to the element at the given index.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray}};
+ /// let xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc1), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// // Expanding an empty `Alloc1` array stores an internal zero entry at
+ /// // index 0. It must not be visible through the API.
+ /// guard.store(5, KBox::new(0xcafeu32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// assert_eq!(guard.get(0), None);
+ /// assert_eq!(guard.find_next(0).map(|(i, v)| (i, *v)), Some((5, 0xcafe)));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
#[inline]
pub fn get(&self, index: usize) -> Option<T::Borrowed<'_>> {
let ptr = self.load(index)?;
@@ -234,6 +250,67 @@ pub fn get_mut(&mut self, index: usize) -> Option<T::BorrowedMut<'_>> {
Some(unsafe { T::borrow_mut(ptr.as_ptr()) })
}
+ fn load_next(&self, index: usize) -> Option<(usize, NonNull<c_void>)> {
+ XArrayState::new(self, index).load_next(usize::MAX)
+ }
+
+ /// Finds the next element starting from the given index.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(10, KBox::new(10u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// guard.store(20, KBox::new(20u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Some((found_index, value)) = guard.find_next(11) {
+ /// assert_eq!(found_index, 20);
+ /// assert_eq!(*value, 20);
+ /// }
+ ///
+ /// if let Some((found_index, value)) = guard.find_next(5) {
+ /// assert_eq!(found_index, 10);
+ /// assert_eq!(*value, 10);
+ /// }
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn find_next(&self, index: usize) -> Option<(usize, T::Borrowed<'_>)> {
+ self.load_next(index)
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ .map(|(index, ptr)| (index, unsafe { T::borrow(ptr.as_ptr()) }))
+ }
+
+ /// Finds the next element starting from the given index, returning a mutable reference.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(10, KBox::new(10u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// guard.store(20, KBox::new(20u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Some((found_index, mut_value)) = guard.find_next_mut(5) {
+ /// assert_eq!(found_index, 10);
+ /// *mut_value = 0x99;
+ /// }
+ ///
+ /// assert_eq!(guard.get(10).copied(), Some(0x99));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn find_next_mut(&mut self, index: usize) -> Option<(usize, T::BorrowedMut<'_>)> {
+ self.load_next(index)
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ .map(move |(index, ptr)| (index, unsafe { T::borrow_mut(ptr.as_ptr()) }))
+ }
+
/// Removes and returns the element at the given index.
pub fn remove(&mut self, index: usize) -> Option<T> {
// SAFETY:
@@ -359,6 +436,30 @@ fn load(&mut self) -> Option<NonNull<c_void>> {
// SAFETY: `xa_zero_to_null` only inspects the value of `ptr`.
NonNull::new(unsafe { bindings::xa_zero_to_null(ptr) }.cast())
}
+
+ fn load_next(&mut self, max: usize) -> Option<(usize, NonNull<c_void>)> {
+ loop {
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant. By the same
+ // invariant, `self.state.xa` aliases the xarray reachable through `self.guard`,
+ // whose lock we hold.
+ let ptr = unsafe { bindings::xas_find(&raw mut self.state, max) };
+ if ptr.is_null() {
+ break None;
+ }
+
+ // Unlike the normal API, `xas_find` does not filter out internal entries. Arrays
+ // created with [`AllocKind::Alloc1`] store `XA_ZERO_ENTRY` at index 0 when they
+ // are expanded from empty. Skip zero entries and continue the search, like the
+ // `xas_retry` loop in `xa_find` does. Retry entries cannot be observed here
+ // because they require concurrent modification of the array, and we hold the
+ // lock.
+ //
+ // SAFETY: `xa_zero_to_null` only inspects the value of `ptr`.
+ if let Some(ptr) = NonNull::new(unsafe { bindings::xa_zero_to_null(ptr) }) {
+ break Some((self.state.xa_index, ptr));
+ }
+ }
+ }
}
// SAFETY: `XArray<T>` has no shared mutable state so it is `Send` iff `T` is `Send`.
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 07/12] rust: xarray: add entry API
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (5 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 06/12] rust: xarray: add `find_next` and `find_next_mut` Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 08/12] rust: mm: add abstractions for allocating from a `sheaf` Andreas Hindborg
` (4 subsequent siblings)
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm
Add an Entry API for XArray that provides ergonomic access to array
slots that may be vacant or occupied. The API follows the pattern of
Rust's standard library HashMap entry API, allowing efficient
conditional insertion and modification of entries.
`Guard::entry` returns an `Entry` that is either `Vacant` or
`Occupied`. The functionality provided by the two entry types is
motivated by the needs of the Rust null block driver memory backing
implementation:
- `VacantEntry::insert` inserts a value and returns a borrow of it.
`VacantEntry::insert_entry` inserts a value and returns an
`OccupiedEntry` instead, for callers that need to perform further
operations on the slot after insertion.
- `OccupiedEntry` provides access to the value through `Deref` and
`DerefMut`, replaces the value through `insert` or `swap`, and
removes it from the array through `remove`.
- `Entry::is_occupied` and the `index` methods support conditional
logic and bookkeeping in callers.
- `into_guard` releases the borrow of the slot and returns the
underlying guard, so a caller can continue operating on the array
without dropping the lock.
- `Guard::find_next_entry` and `Guard::find_next_entry_circular`
return an `OccupiedEntry` for the next occupied slot, the latter
wrapping around at the end of the array. `Guard::insert_entry`
combines lookup and insertion into one operation.
The implementation uses the XArray state API (`xas_*` functions) for
efficient operations without requiring multiple lookups. Helper
functions are added to rust/helpers/xarray.c to wrap static inline C
functions that are not directly accessible from Rust.
Also update MAINTAINERS to cover the new rust files.
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
MAINTAINERS | 1 +
rust/bindings/bindings_helper.h | 1 +
rust/helpers/xarray.c | 10 ++
rust/kernel/xarray.rs | 174 ++++++++++++++++++-
rust/kernel/xarray/entry.rs | 362 ++++++++++++++++++++++++++++++++++++++++
5 files changed, 544 insertions(+), 4 deletions(-)
diff --git a/MAINTAINERS b/MAINTAINERS
index 8014b9f8253ed..b2c9911b88d1d 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -29347,6 +29347,7 @@ B: https://github.com/Rust-for-Linux/linux/issues
C: https://rust-for-linux.zulipchat.com
T: git https://github.com/Rust-for-Linux/linux.git xarray-next
F: rust/kernel/xarray.rs
+F: rust/kernel/xarray/
XBOX DVD IR REMOTE
M: Benjamin Valentin <benpicco@googlemail.com>
diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h
index 419e6b74fedc3..5dda2bb36e3c2 100644
--- a/rust/bindings/bindings_helper.h
+++ b/rust/bindings/bindings_helper.h
@@ -121,6 +121,7 @@ const blk_features_t RUST_CONST_HELPER_BLK_FEAT_ROTATIONAL = BLK_FEAT_ROTATIONAL
const fop_flags_t RUST_CONST_HELPER_FOP_UNSIGNED_OFFSET = FOP_UNSIGNED_OFFSET;
const xa_mark_t RUST_CONST_HELPER_XA_PRESENT = XA_PRESENT;
+const xa_mark_t RUST_CONST_HELPER_XA_FREE_MARK = XA_FREE_MARK;
const gfp_t RUST_CONST_HELPER_XA_FLAGS_ALLOC = XA_FLAGS_ALLOC;
const gfp_t RUST_CONST_HELPER_XA_FLAGS_ALLOC1 = XA_FLAGS_ALLOC1;
diff --git a/rust/helpers/xarray.c b/rust/helpers/xarray.c
index 79799c55c3d73..d7548a17ae0d0 100644
--- a/rust/helpers/xarray.c
+++ b/rust/helpers/xarray.c
@@ -27,7 +27,17 @@ __rust_helper void rust_helper_xa_unlock(struct xarray *xa)
return xa_unlock(xa);
}
+__rust_helper void *rust_helper_xas_result(struct xa_state *xas, void *curr)
+{
+ return xas_result(xas, curr);
+}
+
__rust_helper void *rust_helper_xa_zero_to_null(void *entry)
{
return xa_zero_to_null(entry);
}
+
+__rust_helper int rust_helper_xas_error(const struct xa_state *xas)
+{
+ return xas_error(xas);
+}
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index 9993783fc854a..cf7248bddb8b9 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -13,11 +13,17 @@
NonNull, //
},
};
+pub use entry::{
+ Entry,
+ OccupiedEntry,
+ VacantEntry, //
+};
use kernel::{
alloc,
bindings,
build_assert::build_assert, //
error::{
+ to_result,
Error,
Result, //
},
@@ -250,6 +256,35 @@ pub fn get_mut(&mut self, index: usize) -> Option<T::BorrowedMut<'_>> {
Some(unsafe { T::borrow_mut(ptr.as_ptr()) })
}
+ /// Gets an entry for the specified index, which can be vacant or occupied.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// assert!(guard.get(42).is_none());
+ ///
+ /// match guard.entry(42) {
+ /// Entry::Vacant(entry) => {
+ /// entry.insert(KBox::new(0x1337u32, GFP_KERNEL)?)?;
+ /// }
+ /// Entry::Occupied(_) => unreachable!("We did not insert an entry yet"),
+ /// }
+ ///
+ /// assert_eq!(guard.get(42), Some(&0x1337));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn entry<'b>(&'b mut self, index: usize) -> Entry<'a, 'b, T> {
+ match self.load(index) {
+ None => Entry::Vacant(VacantEntry::new(self, index)),
+ Some(ptr) => Entry::Occupied(OccupiedEntry::new(self, index, ptr)),
+ }
+ }
+
fn load_next(&self, index: usize) -> Option<(usize, NonNull<c_void>)> {
XArrayState::new(self, index).load_next(usize::MAX)
}
@@ -311,6 +346,66 @@ pub fn find_next_mut(&mut self, index: usize) -> Option<(usize, T::BorrowedMut<'
.map(move |(index, ptr)| (index, unsafe { T::borrow_mut(ptr.as_ptr()) }))
}
+ /// Finds the next occupied entry starting from the given index.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(10, KBox::new(10u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// guard.store(20, KBox::new(20u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Some(entry) = guard.find_next_entry(5) {
+ /// assert_eq!(entry.index(), 10);
+ /// let value = entry.remove();
+ /// assert_eq!(*value, 10);
+ /// }
+ ///
+ /// assert_eq!(guard.get(10), None);
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn find_next_entry<'b>(&'b mut self, index: usize) -> Option<OccupiedEntry<'a, 'b, T>> {
+ let mut state = XArrayState::new(self, index);
+ let (_, ptr) = state.load_next(usize::MAX)?;
+ Some(OccupiedEntry { state, ptr })
+ }
+
+ /// Finds the next occupied entry starting at the given index, wrapping around.
+ ///
+ /// Searches for an entry starting at `index` up to the maximum index. If no entry
+ /// is found, wraps around and searches from index 0 up to `index`.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(100, KBox::new(42u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// let entry = guard.find_next_entry_circular(101);
+ /// assert_eq!(entry.map(|e| e.index()), Some(100));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn find_next_entry_circular<'b>(
+ &'b mut self,
+ index: usize,
+ ) -> Option<OccupiedEntry<'a, 'b, T>> {
+ let mut state = XArrayState::new(self, index);
+
+ let (_, ptr) = state.load_next(usize::MAX).or_else(|| {
+ state.restart_at(0);
+ state.load_next(index)
+ })?;
+
+ Some(OccupiedEntry { state, ptr })
+ }
+
/// Removes and returns the element at the given index.
pub fn remove(&mut self, index: usize) -> Option<T> {
// SAFETY:
@@ -386,10 +481,6 @@ pub fn store(
/// - `state` is always a valid `bindings::xa_state`.
/// - `state.xa` aliases the xarray reachable through `guard`.
pub(crate) struct XArrayState<R> {
- // The borrow is held to guarantee exclusive access to the array. It is
- // not read until a later patch adds `into_guard`, so silence the dead
- // code warning until then.
- #[expect(dead_code)]
guard: R,
state: bindings::xa_state,
}
@@ -460,8 +551,83 @@ fn load_next(&mut self, max: usize) -> Option<(usize, NonNull<c_void>)> {
}
}
}
+
+ fn status(&self) -> Result {
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant.
+ to_result(unsafe { bindings::xas_error(&self.state) })
+ }
+
+ /// Resets the state so the next operation walks the tree from the root,
+ /// starting at `index`.
+ fn restart_at(&mut self, index: usize) {
+ self.state.xa_index = index;
+ self.state.xa_node = bindings::XAS_RESTART as *mut bindings::xa_node;
+ }
}
+// Operations that modify the array require exclusive access to the guard, so
+// they are only implemented for `XArrayState<&mut Guard>`.
+impl<'a, 'b, T: ForeignOwnable> XArrayState<&'b mut Guard<'a, T>> {
+ /// Stores `new` at the index of this state, returning the previous entry.
+ ///
+ /// The slot at the index of this state must be occupied. Storing to an
+ /// occupied slot is a simple pointer swap that cannot fail, by design of
+ /// the xarray data structure.
+ fn replace(&mut self, new: *mut c_void) -> *mut c_void {
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant. By the same
+ // invariant, `self.state.xa` aliases the xarray reachable through `self.guard`, whose
+ // lock we hold.
+ let old = unsafe {
+ bindings::xas_result(
+ &raw mut self.state,
+ bindings::xa_zero_to_null(bindings::xas_store(&raw mut self.state, new)),
+ )
+ };
+
+ // SAFETY: `old` is a valid return value from `xas_result`.
+ let errno = unsafe { bindings::xa_err(old) };
+
+ // NOTE: Storing to an occupied slot never fails. This is by design of
+ // the xarray data structure. If a slot is occupied, a store is a
+ // simple pointer swap.
+ debug_assert!(errno == 0);
+
+ old
+ }
+
+ fn insert(&mut self, value: T) -> Result<*mut c_void, StoreError<T>> {
+ let new = T::into_foreign(value).cast();
+
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant. By the same
+ // invariant, `self.state.xa` aliases the xarray reachable through `self.guard`,
+ // whose lock we hold. `new` came from `T::into_foreign`.
+ unsafe { bindings::xas_store(&mut self.state, new) };
+
+ // All arrays created by this abstraction have `XA_FLAGS_TRACK_FREE` set, so the
+ // free mark must be cleared for a newly occupied index, as `__xa_store` does.
+ // This is a no-op if the store above failed.
+ //
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant, and we hold
+ // the lock on the xarray it refers to.
+ unsafe { bindings::xas_clear_mark(&self.state, bindings::XA_FREE_MARK) };
+
+ self.status().map(|()| new).map_err(|error| {
+ // SAFETY: `new` came from `T::into_foreign` and `xas_store` does not take
+ // ownership of the value on error.
+ let value = unsafe { T::from_foreign(new) };
+ StoreError { value, error }
+ })
+ }
+
+ /// Consumes `self` and returns the inner `&mut Guard`.
+ #[inline]
+ pub(crate) fn into_guard(self) -> &'b mut Guard<'a, T> {
+ self.guard
+ }
+}
+
+mod entry;
+
// SAFETY: `XArray<T>` has no shared mutable state so it is `Send` iff `T` is `Send`.
unsafe impl<T: ForeignOwnable + Send> Send for XArray<T> {}
diff --git a/rust/kernel/xarray/entry.rs b/rust/kernel/xarray/entry.rs
new file mode 100644
index 0000000000000..c6c5385e6b4f9
--- /dev/null
+++ b/rust/kernel/xarray/entry.rs
@@ -0,0 +1,362 @@
+// SPDX-License-Identifier: GPL-2.0
+
+use super::{
+ Guard,
+ StoreError,
+ XArrayState, //
+};
+use core::ptr::NonNull;
+use kernel::{
+ prelude::*,
+ types::ForeignOwnable, //
+};
+
+/// Represents either a vacant or occupied entry in an XArray.
+pub enum Entry<'a, 'b, T: ForeignOwnable> {
+ /// A vacant entry that can have a value inserted.
+ Vacant(VacantEntry<'a, 'b, T>),
+ /// An occupied entry containing a value.
+ Occupied(OccupiedEntry<'a, 'b, T>),
+}
+
+impl<T: ForeignOwnable> Entry<'_, '_, T> {
+ /// Returns true if this entry is occupied.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// let entry = guard.entry(42);
+ /// assert_eq!(entry.is_occupied(), false);
+ /// drop(entry);
+ ///
+ /// guard.store(42, KBox::new(0x1337u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// let entry = guard.entry(42);
+ /// assert_eq!(entry.is_occupied(), true);
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ #[inline]
+ pub fn is_occupied(&self) -> bool {
+ matches!(self, Entry::Occupied(_))
+ }
+}
+
+/// A view into a vacant entry in an XArray.
+pub struct VacantEntry<'a, 'b, T: ForeignOwnable> {
+ state: XArrayState<&'b mut Guard<'a, T>>,
+}
+
+impl<'a, 'b, T> VacantEntry<'a, 'b, T>
+where
+ T: ForeignOwnable,
+{
+ pub(crate) fn new(guard: &'b mut Guard<'a, T>, index: usize) -> Self {
+ Self {
+ state: XArrayState::new(guard, index),
+ }
+ }
+
+ /// Consumes the entry and returns a mutable reference to the underlying
+ /// guard.
+ ///
+ /// This releases the slot reservation but retains the lock guard so the
+ /// caller can perform further operations on the array.
+ #[inline]
+ pub fn into_guard(self) -> &'b mut Guard<'a, T> {
+ self.state.into_guard()
+ }
+
+ /// Inserts a value into this vacant entry.
+ ///
+ /// Returns a reference to the newly inserted value.
+ ///
+ /// - This method will fail if the nodes on the path to the index
+ /// represented by this entry are not present in the XArray.
+ /// - This method will not drop the XArray lock.
+ ///
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// assert_eq!(guard.get(42), None);
+ ///
+ /// if let Entry::Vacant(entry) = guard.entry(42) {
+ /// let value = KBox::new(0x1337u32, GFP_ATOMIC)?;
+ /// let borrowed = entry.insert(value)?;
+ /// assert_eq!(*borrowed, 0x1337);
+ /// }
+ ///
+ /// assert_eq!(guard.get(42).copied(), Some(0x1337));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn insert(mut self, value: T) -> Result<T::BorrowedMut<'b>, StoreError<T>> {
+ let new = self.state.insert(value)?;
+
+ // SAFETY: `new` came from `T::into_foreign`. The entry has exclusive
+ // ownership of `new` as it holds a mutable reference to `Guard`.
+ Ok(unsafe { T::borrow_mut(new) })
+ }
+
+ /// Inserts a value and returns an occupied entry representing the newly inserted value.
+ ///
+ /// - This method will fail if the nodes on the path to the index
+ /// represented by this entry are not present in the XArray.
+ /// - This method will not drop the XArray lock.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// assert_eq!(guard.get(42), None);
+ ///
+ /// if let Entry::Vacant(entry) = guard.entry(42) {
+ /// let value = KBox::new(0x1337u32, GFP_ATOMIC)?;
+ /// let occupied = entry.insert_entry(value)?;
+ /// assert_eq!(occupied.index(), 42);
+ /// }
+ ///
+ /// assert_eq!(guard.get(42).copied(), Some(0x1337));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn insert_entry(mut self, value: T) -> Result<OccupiedEntry<'a, 'b, T>, StoreError<T>> {
+ let new = self.state.insert(value)?;
+
+ Ok(OccupiedEntry::<'a, 'b, T> {
+ state: self.state,
+ // SAFETY: `new` came from `T::into_foreign` and is guaranteed non-null.
+ ptr: unsafe { core::ptr::NonNull::new_unchecked(new) },
+ })
+ }
+
+ /// Returns the index of this vacant entry.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// assert_eq!(guard.get(42), None);
+ ///
+ /// if let Entry::Vacant(entry) = guard.entry(42) {
+ /// assert_eq!(entry.index(), 42);
+ /// }
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ #[inline]
+ pub fn index(&self) -> usize {
+ self.state.state.xa_index
+ }
+}
+
+/// A view into an occupied entry in an XArray.
+pub struct OccupiedEntry<'a, 'b, T: ForeignOwnable> {
+ pub(crate) state: XArrayState<&'b mut Guard<'a, T>>,
+ pub(crate) ptr: NonNull<c_void>,
+}
+
+impl<'a, 'b, T> OccupiedEntry<'a, 'b, T>
+where
+ T: ForeignOwnable,
+{
+ pub(crate) fn new(guard: &'b mut Guard<'a, T>, index: usize, ptr: NonNull<c_void>) -> Self {
+ Self {
+ state: XArrayState::new(guard, index),
+ ptr,
+ }
+ }
+
+ /// Consumes the entry and returns a mutable reference to the underlying
+ /// guard.
+ ///
+ /// This releases the borrow on the entry's slot but retains the lock
+ /// guard so the caller can perform further operations on the array.
+ #[inline]
+ pub fn into_guard(self) -> &'b mut Guard<'a, T> {
+ self.state.into_guard()
+ }
+
+ /// Removes the value from this occupied entry and returns it, consuming the entry.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(42, KBox::new(0x1337u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ /// assert_eq!(guard.get(42).copied(), Some(0x1337));
+ ///
+ /// if let Entry::Occupied(entry) = guard.entry(42) {
+ /// let value = entry.remove();
+ /// assert_eq!(*value, 0x1337);
+ /// }
+ ///
+ /// assert_eq!(guard.get(42), None);
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn remove(mut self) -> T {
+ let ptr = self.state.replace(core::ptr::null_mut());
+
+ // SAFETY:
+ // - `ptr` came from `T::into_foreign`.
+ // - As this method takes self by value, the lifetimes of any [`T::Borrowed`] and
+ // [`T::BorrowedMut`] we have created must have ended.
+ unsafe { T::from_foreign(ptr.cast()) }
+ }
+
+ /// Returns the index of this occupied entry.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(42, KBox::new(0x1337u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Entry::Occupied(entry) = guard.entry(42) {
+ /// assert_eq!(entry.index(), 42);
+ /// }
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ #[inline]
+ pub fn index(&self) -> usize {
+ self.state.state.xa_index
+ }
+
+ /// Replaces the value in this occupied entry and returns the old value.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(42, KBox::new(0x1337u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Entry::Occupied(mut entry) = guard.entry(42) {
+ /// let new_value = KBox::new(0x9999u32, GFP_ATOMIC)?;
+ /// let old_value = entry.insert(new_value);
+ /// assert_eq!(*old_value, 0x1337);
+ /// }
+ ///
+ /// assert_eq!(guard.get(42).copied(), Some(0x9999));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn insert(&mut self, value: T) -> T {
+ let new = T::into_foreign(value).cast();
+ // SAFETY: `new` came from `T::into_foreign` and is guaranteed non-null.
+ self.ptr = unsafe { NonNull::new_unchecked(new) };
+
+ let old = self.state.replace(new);
+
+ // SAFETY:
+ // - `old` came from `T::into_foreign`.
+ // - As this method takes `self` by mutable reference, the lifetimes of any
+ // [`T::Borrowed`] and [`T::BorrowedMut`] we have created must have ended.
+ unsafe { T::from_foreign(old) }
+ }
+
+ /// Converts this occupied entry into a mutable reference to the value in the slot represented
+ /// by the entry.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(42, KBox::new(0x1337u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Entry::Occupied(entry) = guard.entry(42) {
+ /// let value_ref = entry.into_mut();
+ /// *value_ref = 0x9999;
+ /// }
+ ///
+ /// assert_eq!(guard.get(42).copied(), Some(0x9999));
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn into_mut(self) -> T::BorrowedMut<'b> {
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ unsafe { T::borrow_mut(self.ptr.as_ptr()) }
+ }
+
+ /// Swaps the value in this entry with the provided value.
+ ///
+ /// Returns the old value that was in the entry.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray, Entry}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// guard.store(42, KBox::new(100u32, GFP_ATOMIC)?, GFP_ATOMIC)?;
+ ///
+ /// if let Entry::Occupied(mut entry) = guard.entry(42) {
+ /// let mut other = 200u32;
+ /// entry.swap(&mut other);
+ /// assert_eq!(other, 100);
+ /// assert_eq!(*entry, 200);
+ /// }
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn swap<U>(&mut self, other: &mut U)
+ where
+ T: ForeignOwnable<Borrowed<'b> = &'b U, BorrowedMut<'b> = &'b mut U> + 'b,
+ U: 'b,
+ {
+ use core::ops::DerefMut;
+ core::mem::swap(self.deref_mut(), other);
+ }
+}
+
+impl<'a, 'b, T, U> core::ops::Deref for OccupiedEntry<'a, 'b, T>
+where
+ T: ForeignOwnable<Borrowed<'b> = &'b U, BorrowedMut<'b> = &'b mut U> + 'b,
+ U: 'b,
+{
+ type Target = U;
+
+ fn deref(&self) -> &Self::Target {
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ unsafe { T::borrow(self.ptr.as_ptr()) }
+ }
+}
+
+impl<'a, 'b, T, U> core::ops::DerefMut for OccupiedEntry<'a, 'b, T>
+where
+ T: ForeignOwnable<Borrowed<'b> = &'b U, BorrowedMut<'b> = &'b mut U> + 'b,
+ U: 'b,
+{
+ fn deref_mut(&mut self) -> &mut Self::Target {
+ // SAFETY: `ptr` came from `T::into_foreign`.
+ unsafe { T::borrow_mut(self.ptr.as_ptr()) }
+ }
+}
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 08/12] rust: mm: add abstractions for allocating from a `sheaf`
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (6 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 07/12] rust: xarray: add entry API Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-03 10:09 ` Vlastimil Babka (SUSE)
2026-09-02 13:26 ` [PATCH v5 09/12] rust: mm: sheaf: allow use of C initialized static caches Andreas Hindborg
` (3 subsequent siblings)
11 siblings, 1 reply; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm, Vlastimil Babka, Liam R. Howlett, Lorenzo Stoakes
Add Rust APIs for allocating objects from a `sheaf`.
Introduce a reduced abstraction `KMemCacheInit` for `struct kmem_cache` to
support management of the `Sheaf`s.
Initialize objects using in-place initialization when objects are allocated
from a `Sheaf`. This is different from C which tends to do some
initialization when the cache is filled. This approach is chosen because
there is no destructor/drop capability in `struct kmem_cache` that can be
invoked when the cache is dropped.
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: linux-mm@kvack.org
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/kernel/mm.rs | 1 +
rust/kernel/mm/sheaf.rs | 466 ++++++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 467 insertions(+)
diff --git a/rust/kernel/mm.rs b/rust/kernel/mm.rs
index 4764d7b68f2a7..1aa44424b0d53 100644
--- a/rust/kernel/mm.rs
+++ b/rust/kernel/mm.rs
@@ -18,6 +18,7 @@
};
use core::{ops::Deref, ptr::NonNull};
+pub mod sheaf;
pub mod virt;
use virt::VmaRef;
diff --git a/rust/kernel/mm/sheaf.rs b/rust/kernel/mm/sheaf.rs
new file mode 100644
index 0000000000000..8f310a06d8404
--- /dev/null
+++ b/rust/kernel/mm/sheaf.rs
@@ -0,0 +1,466 @@
+// SPDX-License-Identifier: GPL-2.0
+
+//! Slub allocator sheaf abstraction.
+//!
+//! Sheaves are percpu array-based caching layers for the slub allocator.
+//! They provide a mechanism for pre-allocating objects that can later
+//! be retrieved without risking allocation failure, making them useful in
+//! contexts where memory allocation must be guaranteed to succeed.
+//!
+//! The term "sheaf" is the english word for a bundle of straw. In this context
+//! it means a bundle of pre-allocated objects. A per-NUMA-node cache of sheaves
+//! is called a "barn". Because you store your sheafs in barns.
+//!
+//! # Use cases
+//!
+//! Sheaves are particularly useful when:
+//!
+//! - Allocations must be guaranteed to succeed in a restricted context (e.g.,
+//! while holding locks or in atomic context).
+//! - Multiple allocations need to be performed as a batch operation.
+//! - Fast-path allocation performance is critical, as sheaf allocations avoid
+//! atomic operations by using local locks with preemption disabled.
+//!
+//! # Architecture
+//!
+//! The sheaf system consists of three main components:
+//!
+//! - [`KMemCache`]: A slab cache configured with sheaf support.
+//! - [`Sheaf`]: A pre-filled container of objects from a specific cache.
+//! - [`SBox`]: An owned allocation from a sheaf, similar to a `Box`.
+//!
+//! # Example
+//!
+//! ```
+//! use kernel::c_str;
+//! use kernel::mm::sheaf::{KMemCache, KMemCacheInit, Sheaf, SBox};
+//! use kernel::prelude::*;
+//!
+//! struct MyObject {
+//! value: u32,
+//! }
+//!
+//! impl KMemCacheInit<MyObject> for MyObject {
+//! fn init() -> impl Init<MyObject> {
+//! init!(MyObject { value: 0 })
+//! }
+//! }
+//!
+//! // Create a cache with sheaf capacity of 16 objects.
+//! let cache = KMemCache::<MyObject>::new(c_str!("my_cache"), 16)?;
+//!
+//! // Pre-fill a sheaf with 8 objects.
+//! let mut sheaf = cache.as_arc_borrow().sheaf(8, GFP_KERNEL)?;
+//!
+//! // Allocations from the sheaf are guaranteed to succeed until empty.
+//! let obj = sheaf.alloc().unwrap();
+//!
+//! // Return the sheaf when done, attempting to refill it.
+//! sheaf.return_refill(GFP_KERNEL);
+//! # Ok::<(), Error>(())
+//! ```
+//!
+//! # Constraints
+//!
+//! - Sheaves are slower when `CONFIG_SLUB_TINY` or `CONFIG_SLUB_DEBUG` is
+//! enabled due to cpu sheaves being disabled. All prefilled sheaves become
+//! "oversize" and go through a slower allocation path.
+//! - The sheaf capacity is fixed at cache creation time.
+
+use core::{
+ convert::Infallible,
+ marker::PhantomData,
+ ops::{Deref, DerefMut},
+ ptr::NonNull,
+};
+
+use kernel::prelude::*;
+
+use crate::sync::{Arc, ArcBorrow};
+
+/// A slab cache with sheaf support.
+///
+/// This type wraps a kernel `kmem_cache` configured with a sheaf capacity,
+/// enabling pre-allocation of objects via [`Sheaf`].
+///
+/// For now, this type only exists for sheaf management.
+///
+/// # Type parameter
+///
+/// - `T`: The type of objects managed by this cache. Must implement
+/// [`KMemCacheInit`] to provide initialization logic for new allocations.
+///
+/// # Context
+///
+/// Dropping the last reference to a `KMemCache` destroys the cache via
+/// `kmem_cache_destroy`, which may sleep. [`Sheaf`] and [`SBox`] instances
+/// created from the cache each hold a reference, so the last reference may
+/// be dropped when one of those is dropped. The last reference must not be
+/// dropped from a context where sleeping is not allowed.
+///
+/// # Invariants
+///
+/// - `cache` is a valid pointer to a `kmem_cache` created with
+/// `__kmem_cache_create_args`.
+/// - The cache is valid for the lifetime of this struct.
+pub struct KMemCache<T: KMemCacheInit<T>> {
+ cache: NonNull<bindings::kmem_cache>,
+ _p: PhantomData<T>,
+}
+
+// SAFETY: `KMemCache<T>` owns a `kmem_cache`, which is internally
+// synchronized and has no thread affinity. The cache may be destroyed from a
+// thread other than the one that created it.
+unsafe impl<T: KMemCacheInit<T> + Send> Send for KMemCache<T> {}
+
+// SAFETY: All operations available through `&KMemCache<T>` are
+// internally synchronized by the C side, so the cache may be used from
+// multiple threads concurrently if the objects it manages can be sent between
+// threads.
+unsafe impl<T: KMemCacheInit<T> + Send> Sync for KMemCache<T> {}
+
+impl<T: KMemCacheInit<T>> KMemCache<T> {
+ /// Creates a new slab cache with sheaf support.
+ ///
+ /// Creates a kernel slab cache for objects of type `T` with the specified
+ /// sheaf capacity. The cache uses the provided `name` for identification
+ /// in `/sys/kernel/slab/` and debugging output.
+ ///
+ /// # Arguments
+ ///
+ /// - `name`: A string identifying the cache. This name appears in sysfs and
+ /// debugging output.
+ /// - `sheaf_capacity`: The maximum number of objects a sheaf from this
+ /// cache can hold. A capacity of zero disables sheaf support.
+ ///
+ /// # Errors
+ ///
+ /// Returns an error if:
+ ///
+ /// - The cache could not be created due to memory pressure.
+ /// - The size of `T` cannot be represented as a `c_uint`.
+ pub fn new(name: &CStr, sheaf_capacity: u32) -> Result<Arc<Self>>
+ where
+ T: KMemCacheInit<T>,
+ {
+ let flags = 0;
+ let mut args: bindings::kmem_cache_args = pin_init::zeroed();
+ args.sheaf_capacity = sheaf_capacity;
+
+ // With an alignment of zero, the slab allocator only guarantees
+ // `ARCH_SLAB_MINALIGN`, which may be smaller than the alignment of `T`.
+ args.align = core::mem::align_of::<T>().try_into()?;
+
+ // NOTE: We are not initializing at object allocation time, because
+ // there is no matching teardown function on the C side machinery.
+ args.ctor = None;
+
+ // SAFETY: `name` is a valid C string, `args` is properly initialized,
+ // and the size of `T` has been validated to fit in a `c_uint`.
+ let ptr = unsafe {
+ bindings::__kmem_cache_create_args(
+ name.as_char_ptr(),
+ core::mem::size_of::<T>().try_into()?,
+ &mut args,
+ flags,
+ )
+ };
+
+ // INVARIANT: `ptr` was returned by `__kmem_cache_create_args` and is
+ // non-null (checked below). The cache is valid until
+ // `kmem_cache_destroy` is called in `Drop`.
+ Ok(Arc::new(
+ Self {
+ cache: NonNull::new(ptr).ok_or(ENOMEM)?,
+ _p: PhantomData,
+ },
+ GFP_KERNEL,
+ )?)
+ }
+
+ /// Creates a pre-filled sheaf from this cache.
+ ///
+ /// Allocates a sheaf and pre-fills it with `size` objects. Once created,
+ /// allocations from the sheaf via [`Sheaf::alloc`] are guaranteed to
+ /// succeed until the sheaf is depleted.
+ ///
+ /// # Arguments
+ ///
+ /// - `size`: The number of objects to pre-allocate. Must not exceed the
+ /// cache's `sheaf_capacity`.
+ /// - `gfp`: Allocation flags controlling how memory is obtained. Use
+ /// [`GFP_KERNEL`] for normal allocations that may sleep, or
+ /// [`GFP_NOWAIT`] for non-blocking allocations.
+ ///
+ /// # Errors
+ ///
+ /// Returns [`ENOMEM`] if the sheaf or its objects could not be allocated.
+ ///
+ /// # Warnings
+ ///
+ /// The kernel will warn if `size` exceeds `sheaf_capacity`.
+ pub fn sheaf(
+ self: ArcBorrow<'_, Self>,
+ size: usize,
+ gfp: kernel::alloc::Flags,
+ ) -> Result<Sheaf<T>> {
+ // SAFETY: `self.as_raw()` returns a valid cache pointer, and `size`
+ // has been validated to fit in a `c_uint`.
+ let ptr = unsafe {
+ bindings::kmem_cache_prefill_sheaf(self.as_raw(), gfp.as_raw(), size.try_into()?)
+ };
+
+ // INVARIANT: `ptr` was returned by `kmem_cache_prefill_sheaf` and is
+ // non-null (checked below). `cache` is the cache from which this sheaf
+ // was created. `dropped` is false since the sheaf has not been returned.
+ Ok(Sheaf {
+ sheaf: NonNull::new(ptr).ok_or(ENOMEM)?,
+ cache: self.into(),
+ dropped: false,
+ })
+ }
+
+ #[inline]
+ fn as_raw(&self) -> *mut bindings::kmem_cache {
+ self.cache.as_ptr()
+ }
+}
+
+impl<T: KMemCacheInit<T>> Drop for KMemCache<T> {
+ fn drop(&mut self) {
+ // SAFETY: `self.as_raw()` returns a valid cache pointer that was
+ // created by `__kmem_cache_create_args`. As all objects allocated from
+ // this hold a reference on `self`, they must have been dropped for this
+ // `drop` method to execute.
+ unsafe { bindings::kmem_cache_destroy(self.as_raw()) };
+ }
+}
+
+/// Trait for types that can be initialized in a slab cache.
+///
+/// This trait provides the initialization logic for objects allocated from a
+/// [`KMemCache`]. When the slab allocator creates new objects, it invokes the
+/// constructor to ensure objects are in a valid initial state.
+///
+/// # Implementation
+///
+/// Implementors must provide [`init`](KMemCacheInit::init), which returns
+/// a in-place initializer for the type.
+///
+/// # Example
+///
+/// ```
+/// use kernel::mm::sheaf::KMemCacheInit;
+/// use kernel::prelude::*;
+///
+/// struct MyData {
+/// counter: u32,
+/// name: [u8; 16],
+/// }
+///
+/// impl KMemCacheInit<MyData> for MyData {
+/// fn init() -> impl Init<MyData> {
+/// init!(MyData {
+/// counter: 0,
+/// name: [0; 16],
+/// })
+/// }
+/// }
+/// ```
+pub trait KMemCacheInit<T> {
+ /// Returns an initializer for creating new objects of type `T`.
+ ///
+ /// The initializer is applied to newly allocated objects when they are
+ /// allocated from a sheaf via [`Sheaf::alloc`]. The cache itself has no
+ /// constructor. The initializer should set all fields to their default or
+ /// initial values.
+ fn init() -> impl Init<T, Infallible>;
+}
+
+/// A pre-filled container of slab objects.
+///
+/// A sheaf holds a set of pre-allocated objects from a [`KMemCache`].
+/// Allocations from a sheaf are guaranteed to succeed until the sheaf is
+/// depleted, making sheaves useful in contexts where allocation failure is
+/// not acceptable.
+///
+/// Sheaves provide faster allocation than direct allocation because they use
+/// local locks with preemption disabled rather than atomic operations.
+///
+/// # Lifecycle
+///
+/// Sheaves are created via [`KMemCache::sheaf`] and should be returned to the
+/// allocator when no longer needed via [`Sheaf::return_refill`]. If a sheaf is
+/// simply dropped, it is returned with `GFP_NOWAIT` flags, which may result in
+/// the sheaf being flushed and freed rather than being cached for reuse.
+///
+/// A sheaf holds a reference to the [`KMemCache`] it was created from.
+/// Dropping the sheaf may thus drop the last reference to the cache and
+/// destroy the cache, which may sleep. See the `# Context` section of
+/// [`KMemCache`].
+///
+/// # Invariants
+///
+/// - `sheaf` is a valid pointer to a `slab_sheaf` obtained from
+/// `kmem_cache_prefill_sheaf`.
+/// - `cache` is the cache from which this sheaf was created.
+/// - `dropped` tracks whether the sheaf has been explicitly returned.
+pub struct Sheaf<T: KMemCacheInit<T>> {
+ sheaf: NonNull<bindings::slab_sheaf>,
+ cache: Arc<KMemCache<T>>,
+ dropped: bool,
+}
+
+// SAFETY: A prefilled sheaf is exclusively owned by the caller and has no
+// affinity to the CPU or thread that created it: `kmem_cache_alloc_from_sheaf`
+// does not touch percpu state, and `kmem_cache_return_sheaf` reattaches the
+// sheaf to the CPU that is current at return time. Thus the sheaf may be sent
+// to another thread if the objects it manages can.
+unsafe impl<T: KMemCacheInit<T> + Send> Send for Sheaf<T> {}
+
+// NOTE: `Sheaf` is deliberately not `Sync`. The C side mutates sheaf state
+// without synchronization, relying on the caller's exclusive ownership. The
+// mutable receivers of the methods on `Sheaf` enforce this exclusivity.
+
+impl<T: KMemCacheInit<T>> Sheaf<T> {
+ #[inline]
+ fn as_raw(&self) -> *mut bindings::slab_sheaf {
+ self.sheaf.as_ptr()
+ }
+
+ /// Return the sheaf and try to refill using `flags`.
+ ///
+ /// If the sheaf cannot simply become the percpu spare sheaf, but there's
+ /// space for a full sheaf in the barn, we try to refill the sheaf back to
+ /// the cache's sheaf_capacity to avoid handling partially full sheaves.
+ ///
+ /// If the refill fails because gfp is e.g. GFP_NOWAIT, or the barn is full,
+ /// the sheaf is instead flushed and freed.
+ pub fn return_refill(mut self, flags: kernel::alloc::Flags) {
+ self.dropped = true;
+ // SAFETY: `self.cache.as_raw()` and `self.as_raw()` return valid
+ // pointers to the cache and sheaf respectively.
+ unsafe {
+ bindings::kmem_cache_return_sheaf(self.cache.as_raw(), flags.as_raw(), self.as_raw())
+ };
+ drop(self);
+ }
+
+ /// Allocates an object from the sheaf.
+ ///
+ /// Returns a new [`SBox`] containing an initialized object, or [`None`]
+ /// if the sheaf is depleted. Allocations are guaranteed to succeed as
+ /// long as the sheaf contains pre-allocated objects.
+ ///
+ /// The `gfp` flags passed to `kmem_cache_alloc_from_sheaf` are set to zero,
+ /// meaning no additional flags like `__GFP_ZERO` or `__GFP_ACCOUNT` are
+ /// applied.
+ ///
+ /// The returned `T` is initialized as part of this function.
+ pub fn alloc(&mut self) -> Option<SBox<T>> {
+ // SAFETY: `self.cache.as_raw()` and `self.as_raw()` return valid
+ // pointers. The function returns NULL when the sheaf is empty.
+ let ptr = unsafe {
+ bindings::kmem_cache_alloc_from_sheaf_noprof(self.cache.as_raw(), 0, self.as_raw())
+ };
+
+ let ptr = NonNull::new(ptr.cast::<T>())?;
+
+ // SAFETY:
+ // - `ptr` is a valid, non-null pointer as it was just returned by the
+ // cache.
+ // - The initializer is infallible, so an error is never returned.
+ unsafe { T::init().__init(ptr.as_ptr()) }.expect("Initializer is infallible");
+
+ // INVARIANT: `ptr` was returned by `kmem_cache_alloc_from_sheaf_noprof`
+ // and initialized above. `cache` is the cache from which this object
+ // was allocated. The object remains valid until freed in `Drop`.
+ Some(SBox {
+ ptr,
+ cache: self.cache.clone(),
+ })
+ }
+}
+
+impl<T: KMemCacheInit<T>> Drop for Sheaf<T> {
+ fn drop(&mut self) {
+ if !self.dropped {
+ // SAFETY: `self.cache.as_raw()` and `self.as_raw()` return valid
+ // pointers. Using `GFP_NOWAIT` because the drop may occur in a
+ // context where sleeping is not permitted.
+ unsafe {
+ bindings::kmem_cache_return_sheaf(
+ self.cache.as_raw(),
+ GFP_NOWAIT.as_raw(),
+ self.as_raw(),
+ )
+ };
+ }
+ }
+}
+
+/// An owned allocation from a cache sheaf.
+///
+/// `SBox` is similar to `Box` but is backed by a slab cache allocation obtained
+/// through a [`Sheaf`]. It provides owned access to an initialized object and
+/// ensures the object is properly freed back to the cache when dropped.
+///
+/// The contained `T` is initialized when the `SBox` is returned from alloc and
+/// dropped when the `SBox` is dropped.
+///
+/// An `SBox` holds a reference to the [`KMemCache`] it was allocated from.
+/// Dropping the `SBox` may thus drop the last reference to the cache and
+/// destroy the cache, which may sleep. See the `# Context` section of
+/// [`KMemCache`].
+///
+/// # Invariants
+///
+/// - `ptr` points to a valid, initialized object of type `T`.
+/// - `cache` is the cache from which this object was allocated.
+/// - The object remains valid for the lifetime of the `SBox`.
+pub struct SBox<T: KMemCacheInit<T>> {
+ ptr: NonNull<T>,
+ cache: Arc<KMemCache<T>>,
+}
+
+// SAFETY: `SBox<T>` owns a `T`. Sheaf allocated objects are ordinary slab
+// objects that may be freed from any thread, so an `SBox<T>` may be sent to
+// another thread if `T` can.
+unsafe impl<T: KMemCacheInit<T> + Send> Send for SBox<T> {}
+
+// SAFETY: `SBox<T>` has no interior mutability, so sharing `&SBox<T>` between
+// threads only shares `&T`.
+unsafe impl<T: KMemCacheInit<T> + Sync> Sync for SBox<T> {}
+
+impl<T: KMemCacheInit<T>> Deref for SBox<T> {
+ type Target = T;
+
+ #[inline]
+ fn deref(&self) -> &Self::Target {
+ // SAFETY: `ptr` is valid and properly aligned per the type invariants.
+ unsafe { self.ptr.as_ref() }
+ }
+}
+
+impl<T: KMemCacheInit<T>> DerefMut for SBox<T> {
+ #[inline]
+ fn deref_mut(&mut self) -> &mut Self::Target {
+ // SAFETY: `ptr` is valid and properly aligned per the type invariants,
+ // and we have exclusive access via `&mut self`.
+ unsafe { self.ptr.as_mut() }
+ }
+}
+
+impl<T: KMemCacheInit<T>> Drop for SBox<T> {
+ fn drop(&mut self) {
+ // SAFETY: By type invariant, `ptr` points to a valid and initialized
+ // object. We do not touch `ptr` after returning it to the cache.
+ unsafe { core::ptr::drop_in_place(self.ptr.as_ptr()) };
+
+ // SAFETY: `self.ptr` was allocated from `self.cache` via
+ // `kmem_cache_alloc_from_sheaf_noprof` and is valid.
+ unsafe {
+ bindings::kmem_cache_free(self.cache.as_raw(), self.ptr.as_ptr().cast());
+ }
+ }
+}
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* Re: [PATCH v5 08/12] rust: mm: add abstractions for allocating from a `sheaf`
2026-09-02 13:26 ` [PATCH v5 08/12] rust: mm: add abstractions for allocating from a `sheaf` Andreas Hindborg
@ 2026-09-03 10:09 ` Vlastimil Babka (SUSE)
0 siblings, 0 replies; 15+ messages in thread
From: Vlastimil Babka (SUSE) @ 2026-09-03 10:09 UTC (permalink / raw)
To: Andreas Hindborg, Tamir Duberstein, Miguel Ojeda, Boqun Feng,
Gary Guo, Björn Roy Baron, Benno Lossin, Alice Ryhl,
Trevor Gross, Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Harry Yoo, Hao Li, Christoph Lameter,
David Rientjes, Roman Gushchin
Cc: rust-for-linux, linux-kernel, linux-fsdevel, linux-mm
On 9/2/26 15:26, Andreas Hindborg wrote:
> Add Rust APIs for allocating objects from a `sheaf`.
>
> Introduce a reduced abstraction `KMemCacheInit` for `struct kmem_cache` to
> support management of the `Sheaf`s.
>
> Initialize objects using in-place initialization when objects are allocated
> from a `Sheaf`. This is different from C which tends to do some
> initialization when the cache is filled. This approach is chosen because
> there is no destructor/drop capability in `struct kmem_cache` that can be
> invoked when the cache is dropped.
>
> Cc: Vlastimil Babka <vbabka@suse.cz>
> Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
> Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
> Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> Cc: linux-mm@kvack.org
> Assisted-by: LLM
> Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
Acked-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
Note that currently all caches have sheaves (except CONFIG_SLUB_TINY or
slab_debug enabled), and sheaf_capacity on creation is treated as a minimum,
which the automated sizing calculation can make larger - in case you want to
reflect this detail in some of the description comments.
^ permalink raw reply [flat|nested] 15+ messages in thread
* [PATCH v5 09/12] rust: mm: sheaf: allow use of C initialized static caches
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (7 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 08/12] rust: mm: add abstractions for allocating from a `sheaf` Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-03 10:11 ` Vlastimil Babka (SUSE)
2026-09-02 13:26 ` [PATCH v5 10/12] xarray, radix-tree: enable sheaf support for kmem_cache Andreas Hindborg
` (2 subsequent siblings)
11 siblings, 1 reply; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm, Vlastimil Babka, Liam R. Howlett, Lorenzo Stoakes
Extend the sheaf abstraction to support caches initialized by C at kernel
boot time, in addition to dynamically created Rust caches.
Introduce `KMemCache<T>` as a transparent wrapper around `kmem_cache` for
static caches with `'static` lifetime. Rename the previous `KMemCache<T>`
to `KMemCacheHandle<T>` to represent dynamically created, reference-counted
caches.
Add `Static` and `Dynamic` marker types along with `StaticSheaf` and
`DynamicSheaf` type aliases to distinguish sheaves from each cache type.
The `Sheaf` type now carries lifetime and allocation mode type parameters.
Add `SBox::into_ptr()` and `SBox::static_from_ptr()` methods for passing
allocations through C code via raw pointers.
Add `KMemCache::from_raw()` for wrapping C-initialized static caches and
`Sheaf::refill()` for replenishing a sheaf to a minimum size.
Export `kmem_cache_prefill_sheaf`, `kmem_cache_return_sheaf`,
`kmem_cache_refill_sheaf`, and `kmem_cache_alloc_from_sheaf_noprof` to
allow Rust module code to use the sheaf API.
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: linux-mm@kvack.org
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
mm/slub.c | 4 +
rust/kernel/mm/sheaf.rs | 401 ++++++++++++++++++++++++++++++++++++++++++------
2 files changed, 356 insertions(+), 49 deletions(-)
diff --git a/mm/slub.c b/mm/slub.c
index 0337e60db5ac..d81dbf2abb86 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -5101,6 +5101,7 @@ kmem_cache_prefill_sheaf(struct kmem_cache *s, gfp_t gfp, unsigned int size)
return sheaf;
}
+EXPORT_SYMBOL(kmem_cache_prefill_sheaf);
/*
* Use this to return a sheaf obtained by kmem_cache_prefill_sheaf()
@@ -5156,6 +5157,7 @@ void kmem_cache_return_sheaf(struct kmem_cache *s, gfp_t gfp,
barn_put_full_sheaf(barn, sheaf);
stat(s, BARN_PUT);
}
+EXPORT_SYMBOL(kmem_cache_return_sheaf);
/*
* Refill a sheaf previously returned by kmem_cache_prefill_sheaf to at least
@@ -5211,6 +5213,7 @@ int kmem_cache_refill_sheaf(struct kmem_cache *s, gfp_t gfp,
*sheafp = sheaf;
return 0;
}
+EXPORT_SYMBOL(kmem_cache_refill_sheaf);
/*
* Allocate from a sheaf obtained by kmem_cache_prefill_sheaf()
@@ -5249,6 +5252,7 @@ kmem_cache_alloc_from_sheaf_noprof(struct kmem_cache *s, gfp_t gfp,
return ret;
}
+EXPORT_SYMBOL(kmem_cache_alloc_from_sheaf_noprof);
unsigned int kmem_cache_sheaf_size(struct slab_sheaf *sheaf)
{
diff --git a/rust/kernel/mm/sheaf.rs b/rust/kernel/mm/sheaf.rs
index 8f310a06d840..f6df856c7f4b 100644
--- a/rust/kernel/mm/sheaf.rs
+++ b/rust/kernel/mm/sheaf.rs
@@ -23,17 +23,26 @@
//!
//! # Architecture
//!
-//! The sheaf system consists of three main components:
+//! The sheaf system supports two modes of operation:
+//!
+//! - **Static caches**: [`KMemCache`] represents a cache initialized by C code at
+//! kernel boot time. These have `'static` lifetime and produce [`StaticSheaf`]
+//! instances.
+//! - **Dynamic caches**: [`KMemCacheHandle`] wraps a cache created at runtime by
+//! Rust code. These are reference-counted and produce [`DynamicSheaf`] instances.
+//!
+//! Both modes use the same core types:
//!
-//! - [`KMemCache`]: A slab cache configured with sheaf support.
//! - [`Sheaf`]: A pre-filled container of objects from a specific cache.
//! - [`SBox`]: An owned allocation from a sheaf, similar to a `Box`.
//!
//! # Example
//!
+//! Using a dynamically created cache:
+//!
//! ```
//! use kernel::c_str;
-//! use kernel::mm::sheaf::{KMemCache, KMemCacheInit, Sheaf, SBox};
+//! use kernel::mm::sheaf::{KMemCacheHandle, KMemCacheInit, Sheaf, SBox};
//! use kernel::prelude::*;
//!
//! struct MyObject {
@@ -47,7 +56,7 @@
//! }
//!
//! // Create a cache with sheaf capacity of 16 objects.
-//! let cache = KMemCache::<MyObject>::new(c_str!("my_cache"), 16)?;
+//! let cache = KMemCacheHandle::<MyObject>::new(c_str!("my_cache"), 16)?;
//!
//! // Pre-fill a sheaf with 8 objects.
//! let mut sheaf = cache.as_arc_borrow().sheaf(8, GFP_KERNEL)?;
@@ -76,7 +85,114 @@
use kernel::prelude::*;
-use crate::sync::{Arc, ArcBorrow};
+use crate::{
+ sync::{Arc, ArcBorrow},
+ types::Opaque,
+};
+
+/// A slab cache with sheaf support.
+///
+/// This type is a transparent wrapper around a kernel `kmem_cache`. It can be
+/// used with caches created either by C code or via [`KMemCacheHandle`].
+///
+/// When a reference to this type has `'static` lifetime (i.e., `&'static
+/// KMemCache<T>`), it typically represents a cache initialized by C at boot
+/// time. Such references produce [`StaticSheaf`] instances via [`sheaf`].
+///
+/// [`sheaf`]: KMemCache::sheaf
+///
+/// # Type parameter
+///
+/// - `T`: The type of objects managed by this cache. Must implement
+/// [`KMemCacheInit`] to provide initialization logic for allocations.
+#[repr(transparent)]
+pub struct KMemCache<T: KMemCacheInit<T>> {
+ inner: Opaque<bindings::kmem_cache>,
+ _p: PhantomData<T>,
+}
+
+// SAFETY: The C `kmem_cache` is internally synchronized and has no thread
+// affinity, so a `KMemCache<T>` may be sent to another thread if the objects
+// it manages can.
+unsafe impl<T: KMemCacheInit<T> + Send> Send for KMemCache<T> {}
+
+// SAFETY: All operations available through `&KMemCache<T>` (creating sheaves
+// and allocating objects) are internally synchronized by the C side, so the
+// cache may be used from multiple threads concurrently if the objects it
+// manages can be sent between threads.
+unsafe impl<T: KMemCacheInit<T> + Send> Sync for KMemCache<T> {}
+
+impl<T: KMemCacheInit<T>> KMemCache<T> {
+ /// Creates a pre-filled sheaf from this cache.
+ ///
+ /// Allocates a sheaf and pre-fills it with `size` objects. Once created,
+ /// allocations from the sheaf via [`Sheaf::alloc`] are guaranteed to
+ /// succeed until the sheaf is depleted.
+ ///
+ /// # Arguments
+ ///
+ /// - `size`: The number of objects to pre-allocate. Must not exceed the
+ /// cache's `sheaf_capacity`.
+ /// - `gfp`: Allocation flags controlling how memory is obtained. Use
+ /// [`GFP_KERNEL`] for normal allocations that may sleep, or
+ /// [`GFP_NOWAIT`] for non-blocking allocations.
+ ///
+ /// # Errors
+ ///
+ /// Returns [`ENOMEM`] if the sheaf or its objects could not be allocated.
+ ///
+ /// # Warnings
+ ///
+ /// The kernel will warn if `size` exceeds `sheaf_capacity`.
+ pub fn sheaf(
+ &'static self,
+ size: usize,
+ gfp: kernel::alloc::Flags,
+ ) -> Result<Sheaf<'static, T, Static>> {
+ // SAFETY: `self.as_raw()` returns a valid cache pointer, and `size`
+ // has been validated to fit in a `c_uint`.
+ let ptr = unsafe {
+ bindings::kmem_cache_prefill_sheaf(self.inner.get(), gfp.as_raw(), size.try_into()?)
+ };
+
+ // INVARIANT: `ptr` was returned by `kmem_cache_prefill_sheaf` and is
+ // non-null (checked below). `cache` is the cache from which this sheaf
+ // was created. `dropped` is false since the sheaf has not been returned.
+ Ok(Sheaf {
+ sheaf: NonNull::new(ptr).ok_or(ENOMEM)?,
+ // SAFETY: `self` is a valid reference, so the pointer is non-null.
+ cache: CacheRef::Static(unsafe {
+ NonNull::new_unchecked((&raw const *self).cast_mut())
+ }),
+ dropped: false,
+ _p: PhantomData,
+ })
+ }
+
+ #[inline]
+ fn as_raw(&self) -> *mut bindings::kmem_cache {
+ self.inner.get()
+ }
+
+ /// Creates a reference to a [`KMemCache`] from a raw pointer.
+ ///
+ /// This is useful for wrapping a C-initialized static `kmem_cache`, such as
+ /// the global `radix_tree_node_cachep` used by XArrays.
+ ///
+ /// # Safety
+ ///
+ /// - `ptr` must be a valid pointer to a `kmem_cache` that was created for
+ /// objects of type `T`.
+ /// - The cache must remain valid for the lifetime `'a`.
+ /// - The caller must ensure that the cache was configured appropriately for
+ /// the type `T`, including proper size and alignment.
+ pub unsafe fn from_raw<'a>(ptr: *mut bindings::kmem_cache) -> &'a Self {
+ // SAFETY: The caller guarantees that `ptr` is a valid pointer to a
+ // `kmem_cache` created for objects of type `T`, that it remains valid
+ // for lifetime `'a`, and that the cache is properly configured for `T`.
+ unsafe { &*ptr.cast::<Self>() }
+ }
+}
/// A slab cache with sheaf support.
///
@@ -92,9 +208,9 @@
///
/// # Context
///
-/// Dropping the last reference to a `KMemCache` destroys the cache via
+/// Dropping the last reference to a `KMemCacheHandle` destroys the cache via
/// `kmem_cache_destroy`, which may sleep. [`Sheaf`] and [`SBox`] instances
-/// created from the cache each hold a reference, so the last reference may
+/// created from the handle each hold a reference, so the last reference may
/// be dropped when one of those is dropped. The last reference must not be
/// dropped from a context where sleeping is not allowed.
///
@@ -103,23 +219,23 @@
/// - `cache` is a valid pointer to a `kmem_cache` created with
/// `__kmem_cache_create_args`.
/// - The cache is valid for the lifetime of this struct.
-pub struct KMemCache<T: KMemCacheInit<T>> {
- cache: NonNull<bindings::kmem_cache>,
- _p: PhantomData<T>,
+#[repr(transparent)]
+pub struct KMemCacheHandle<T: KMemCacheInit<T>> {
+ cache: NonNull<KMemCache<T>>,
}
-// SAFETY: `KMemCache<T>` owns a `kmem_cache`, which is internally
+// SAFETY: `KMemCacheHandle<T>` owns a `kmem_cache`, which is internally
// synchronized and has no thread affinity. The cache may be destroyed from a
// thread other than the one that created it.
-unsafe impl<T: KMemCacheInit<T> + Send> Send for KMemCache<T> {}
+unsafe impl<T: KMemCacheInit<T> + Send> Send for KMemCacheHandle<T> {}
-// SAFETY: All operations available through `&KMemCache<T>` are
+// SAFETY: All operations available through `&KMemCacheHandle<T>` are
// internally synchronized by the C side, so the cache may be used from
// multiple threads concurrently if the objects it manages can be sent between
// threads.
-unsafe impl<T: KMemCacheInit<T> + Send> Sync for KMemCache<T> {}
+unsafe impl<T: KMemCacheInit<T> + Send> Sync for KMemCacheHandle<T> {}
-impl<T: KMemCacheInit<T>> KMemCache<T> {
+impl<T: KMemCacheInit<T>> KMemCacheHandle<T> {
/// Creates a new slab cache with sheaf support.
///
/// Creates a kernel slab cache for objects of type `T` with the specified
@@ -171,8 +287,7 @@ pub fn new(name: &CStr, sheaf_capacity: u32) -> Result<Arc<Self>>
// `kmem_cache_destroy` is called in `Drop`.
Ok(Arc::new(
Self {
- cache: NonNull::new(ptr).ok_or(ENOMEM)?,
- _p: PhantomData,
+ cache: NonNull::new(ptr.cast()).ok_or(ENOMEM)?,
},
GFP_KERNEL,
)?)
@@ -199,11 +314,11 @@ pub fn new(name: &CStr, sheaf_capacity: u32) -> Result<Arc<Self>>
/// # Warnings
///
/// The kernel will warn if `size` exceeds `sheaf_capacity`.
- pub fn sheaf(
- self: ArcBorrow<'_, Self>,
+ pub fn sheaf<'a>(
+ self: ArcBorrow<'a, Self>,
size: usize,
gfp: kernel::alloc::Flags,
- ) -> Result<Sheaf<T>> {
+ ) -> Result<Sheaf<'a, T, Dynamic>> {
// SAFETY: `self.as_raw()` returns a valid cache pointer, and `size`
// has been validated to fit in a `c_uint`.
let ptr = unsafe {
@@ -215,18 +330,19 @@ pub fn sheaf(
// was created. `dropped` is false since the sheaf has not been returned.
Ok(Sheaf {
sheaf: NonNull::new(ptr).ok_or(ENOMEM)?,
- cache: self.into(),
+ cache: CacheRef::Arc(self.into()),
dropped: false,
+ _p: PhantomData,
})
}
#[inline]
fn as_raw(&self) -> *mut bindings::kmem_cache {
- self.cache.as_ptr()
+ self.cache.as_ptr().cast()
}
}
-impl<T: KMemCacheInit<T>> Drop for KMemCache<T> {
+impl<T: KMemCacheInit<T>> Drop for KMemCacheHandle<T> {
fn drop(&mut self) {
// SAFETY: `self.as_raw()` returns a valid cache pointer that was
// created by `__kmem_cache_create_args`. As all objects allocated from
@@ -239,13 +355,13 @@ fn drop(&mut self) {
/// Trait for types that can be initialized in a slab cache.
///
/// This trait provides the initialization logic for objects allocated from a
-/// [`KMemCache`]. When the slab allocator creates new objects, it invokes the
-/// constructor to ensure objects are in a valid initial state.
+/// [`KMemCache`]. The initializer is called when objects are allocated from a
+/// sheaf via [`Sheaf::alloc`].
///
/// # Implementation
///
-/// Implementors must provide [`init`](KMemCacheInit::init), which returns
-/// a in-place initializer for the type.
+/// Implementors must provide [`init`](KMemCacheInit::init), which returns an
+/// infallible initializer for the type.
///
/// # Example
///
@@ -270,13 +386,34 @@ fn drop(&mut self) {
pub trait KMemCacheInit<T> {
/// Returns an initializer for creating new objects of type `T`.
///
- /// The initializer is applied to newly allocated objects when they are
- /// allocated from a sheaf via [`Sheaf::alloc`]. The cache itself has no
- /// constructor. The initializer should set all fields to their default or
- /// initial values.
+ /// The initializer is applied to newly allocated objects when they are allocated from a sheaf
+ /// via [`Sheaf::alloc`]. The initializer should set all fields to their default or initial
+ /// values.
fn init() -> impl Init<T, Infallible>;
}
+/// Marker type for sheaves from static caches.
+///
+/// Used as a type parameter for [`Sheaf`] to indicate the sheaf was created
+/// from a `&'static KMemCache<T>`.
+pub enum Static {}
+
+/// Marker type for sheaves from dynamic caches.
+///
+/// Used as a type parameter for [`Sheaf`] to indicate the sheaf was created
+/// from a [`KMemCacheHandle`] via [`ArcBorrow`].
+pub enum Dynamic {}
+
+/// A sheaf from a static cache.
+///
+/// This is a [`Sheaf`] backed by a `&'static KMemCache<T>`.
+pub type StaticSheaf<'a, T> = Sheaf<'a, T, Static>;
+
+/// A sheaf from a dynamic cache.
+///
+/// This is a [`Sheaf`] backed by a reference-counted [`KMemCacheHandle`].
+pub type DynamicSheaf<'a, T> = Sheaf<'a, T, Dynamic>;
+
/// A pre-filled container of slab objects.
///
/// A sheaf holds a set of pre-allocated objects from a [`KMemCache`].
@@ -287,17 +424,28 @@ pub trait KMemCacheInit<T> {
/// Sheaves provide faster allocation than direct allocation because they use
/// local locks with preemption disabled rather than atomic operations.
///
+/// # Type parameters
+///
+/// - `'a`: The lifetime of the cache reference.
+/// - `T`: The type of objects in this sheaf.
+/// - `A`: Either [`Static`] or [`Dynamic`], indicating whether the backing
+/// cache is a static reference or a reference-counted handle.
+///
+/// For convenience, [`StaticSheaf`] and [`DynamicSheaf`] type aliases are
+/// provided.
+///
/// # Lifecycle
///
-/// Sheaves are created via [`KMemCache::sheaf`] and should be returned to the
-/// allocator when no longer needed via [`Sheaf::return_refill`]. If a sheaf is
-/// simply dropped, it is returned with `GFP_NOWAIT` flags, which may result in
-/// the sheaf being flushed and freed rather than being cached for reuse.
+/// Sheaves are created via [`KMemCache::sheaf`] or [`KMemCacheHandle::sheaf`]
+/// and should be returned to the allocator when no longer needed via
+/// [`Sheaf::return_refill`]. If a sheaf is simply dropped, it is returned with
+/// `GFP_NOWAIT` flags, which may result in the sheaf being flushed and freed
+/// rather than being cached for reuse.
///
-/// A sheaf holds a reference to the [`KMemCache`] it was created from.
-/// Dropping the sheaf may thus drop the last reference to the cache and
-/// destroy the cache, which may sleep. See the `# Context` section of
-/// [`KMemCache`].
+/// A sheaf created from a [`KMemCacheHandle`] holds a reference to the
+/// handle. Dropping the sheaf may thus drop the last reference to the handle
+/// and destroy the cache, which may sleep. See the `# Context` section of
+/// [`KMemCacheHandle`].
///
/// # Invariants
///
@@ -305,10 +453,11 @@ pub trait KMemCacheInit<T> {
/// `kmem_cache_prefill_sheaf`.
/// - `cache` is the cache from which this sheaf was created.
/// - `dropped` tracks whether the sheaf has been explicitly returned.
-pub struct Sheaf<T: KMemCacheInit<T>> {
+pub struct Sheaf<'a, T: KMemCacheInit<T>, A> {
sheaf: NonNull<bindings::slab_sheaf>,
- cache: Arc<KMemCache<T>>,
+ cache: CacheRef<T>,
dropped: bool,
+ _p: PhantomData<(&'a KMemCache<T>, A)>,
}
// SAFETY: A prefilled sheaf is exclusively owned by the caller and has no
@@ -316,13 +465,13 @@ pub struct Sheaf<T: KMemCacheInit<T>> {
// does not touch percpu state, and `kmem_cache_return_sheaf` reattaches the
// sheaf to the CPU that is current at return time. Thus the sheaf may be sent
// to another thread if the objects it manages can.
-unsafe impl<T: KMemCacheInit<T> + Send> Send for Sheaf<T> {}
+unsafe impl<T: KMemCacheInit<T> + Send, A> Send for Sheaf<'_, T, A> {}
// NOTE: `Sheaf` is deliberately not `Sync`. The C side mutates sheaf state
// without synchronization, relying on the caller's exclusive ownership. The
// mutable receivers of the methods on `Sheaf` enforce this exclusivity.
-impl<T: KMemCacheInit<T>> Sheaf<T> {
+impl<'a, T: KMemCacheInit<T>, A> Sheaf<'a, T, A> {
#[inline]
fn as_raw(&self) -> *mut bindings::slab_sheaf {
self.sheaf.as_ptr()
@@ -346,6 +495,39 @@ pub fn return_refill(mut self, flags: kernel::alloc::Flags) {
drop(self);
}
+ /// Refills the sheaf to at least the specified size.
+ ///
+ /// Replenishes the sheaf by preallocating objects until it contains at
+ /// least `size` objects. If the sheaf already contains `size` or more
+ /// objects, this is a no-op. In practice, the sheaf is refilled to its
+ /// full capacity.
+ ///
+ /// # Arguments
+ ///
+ /// - `flags`: Allocation flags controlling how memory is obtained.
+ /// - `size`: The minimum number of objects the sheaf should contain after
+ /// refilling. If `size` exceeds the cache's `sheaf_capacity`, the sheaf
+ /// may be replaced with a larger one.
+ ///
+ /// # Errors
+ ///
+ /// Returns an error if the objects could not be allocated. If refilling
+ /// fails, the existing sheaf is left intact.
+ pub fn refill(&mut self, flags: kernel::alloc::Flags, size: usize) -> Result {
+ // SAFETY: `self.cache.as_raw()` returns a valid cache pointer and
+ // `&raw mut self.sheaf` points to a valid sheaf per the type invariants.
+ kernel::error::to_result(unsafe {
+ bindings::kmem_cache_refill_sheaf(
+ self.cache.as_raw(),
+ flags.as_raw(),
+ (&raw mut (self.sheaf)).cast(),
+ size.try_into()?,
+ )
+ })
+ }
+}
+
+impl<'a, T: KMemCacheInit<T>> Sheaf<'a, T, Static> {
/// Allocates an object from the sheaf.
///
/// Returns a new [`SBox`] containing an initialized object, or [`None`]
@@ -382,7 +564,44 @@ pub fn alloc(&mut self) -> Option<SBox<T>> {
}
}
-impl<T: KMemCacheInit<T>> Drop for Sheaf<T> {
+impl<'a, T: KMemCacheInit<T>> Sheaf<'a, T, Dynamic> {
+ /// Allocates an object from the sheaf.
+ ///
+ /// Returns a new [`SBox`] containing an initialized object, or [`None`]
+ /// if the sheaf is depleted. Allocations are guaranteed to succeed as
+ /// long as the sheaf contains pre-allocated objects.
+ ///
+ /// The `gfp` flags passed to `kmem_cache_alloc_from_sheaf` are set to zero,
+ /// meaning no additional flags like `__GFP_ZERO` or `__GFP_ACCOUNT` are
+ /// applied.
+ ///
+ /// The returned `T` is initialized as part of this function.
+ pub fn alloc(&mut self) -> Option<SBox<T>> {
+ // SAFETY: `self.cache.as_raw()` and `self.as_raw()` return valid
+ // pointers. The function returns NULL when the sheaf is empty.
+ let ptr = unsafe {
+ bindings::kmem_cache_alloc_from_sheaf_noprof(self.cache.as_raw(), 0, self.as_raw())
+ };
+
+ let ptr = NonNull::new(ptr.cast::<T>())?;
+
+ // SAFETY:
+ // - `ptr` is a valid, non-null pointer as it was just returned by the
+ // cache.
+ // - The initializer is infallible, so an error is never returned.
+ unsafe { T::init().__init(ptr.as_ptr()) }.expect("Initializer is infallible");
+
+ // INVARIANT: `ptr` was returned by `kmem_cache_alloc_from_sheaf_noprof`
+ // and initialized above. `cache` is the cache from which this object
+ // was allocated. The object remains valid until freed in `Drop`.
+ Some(SBox {
+ ptr,
+ cache: self.cache.clone(),
+ })
+ }
+}
+
+impl<'a, T: KMemCacheInit<T>, A> Drop for Sheaf<'a, T, A> {
fn drop(&mut self) {
if !self.dropped {
// SAFETY: `self.cache.as_raw()` and `self.as_raw()` return valid
@@ -399,6 +618,40 @@ fn drop(&mut self) {
}
}
+/// Internal reference to a cache, either static or reference-counted.
+///
+/// # Invariants
+///
+/// - For `CacheRef::Static`: the `NonNull` points to a valid `KMemCache<T>`
+/// with `'static` lifetime, derived from a `&'static KMemCache<T>` reference.
+enum CacheRef<T: KMemCacheInit<T>> {
+ /// A reference-counted handle to a dynamically created cache.
+ Arc(Arc<KMemCacheHandle<T>>),
+ /// A pointer to a static lifetime cache.
+ Static(NonNull<KMemCache<T>>),
+}
+
+impl<T: KMemCacheInit<T>> Clone for CacheRef<T> {
+ fn clone(&self) -> Self {
+ match self {
+ Self::Arc(arg0) => Self::Arc(arg0.clone()),
+ Self::Static(arg0) => Self::Static(*arg0),
+ }
+ }
+}
+
+impl<T: KMemCacheInit<T>> CacheRef<T> {
+ #[inline]
+ fn as_raw(&self) -> *mut bindings::kmem_cache {
+ match self {
+ CacheRef::Arc(handle) => handle.as_raw(),
+ // SAFETY: By type invariant, `ptr` points to a valid `KMemCache<T>`
+ // with `'static` lifetime.
+ CacheRef::Static(ptr) => unsafe { ptr.as_ref() }.as_raw(),
+ }
+ }
+}
+
/// An owned allocation from a cache sheaf.
///
/// `SBox` is similar to `Box` but is backed by a slab cache allocation obtained
@@ -408,10 +661,10 @@ fn drop(&mut self) {
/// The contained `T` is initialized when the `SBox` is returned from alloc and
/// dropped when the `SBox` is dropped.
///
-/// An `SBox` holds a reference to the [`KMemCache`] it was allocated from.
-/// Dropping the `SBox` may thus drop the last reference to the cache and
-/// destroy the cache, which may sleep. See the `# Context` section of
-/// [`KMemCache`].
+/// An `SBox` allocated from a [`KMemCacheHandle`] backed sheaf holds a
+/// reference to the handle. Dropping the `SBox` may thus drop the last
+/// reference to the handle and destroy the cache, which may sleep. See the
+/// `# Context` section of [`KMemCacheHandle`].
///
/// # Invariants
///
@@ -420,7 +673,7 @@ fn drop(&mut self) {
/// - The object remains valid for the lifetime of the `SBox`.
pub struct SBox<T: KMemCacheInit<T>> {
ptr: NonNull<T>,
- cache: Arc<KMemCache<T>>,
+ cache: CacheRef<T>,
}
// SAFETY: `SBox<T>` owns a `T`. Sheaf allocated objects are ordinary slab
@@ -432,6 +685,56 @@ unsafe impl<T: KMemCacheInit<T> + Send> Send for SBox<T> {}
// threads only shares `&T`.
unsafe impl<T: KMemCacheInit<T> + Sync> Sync for SBox<T> {}
+impl<T: KMemCacheInit<T>> SBox<T> {
+ /// Consumes the `SBox` and returns the raw pointer to the contained value.
+ ///
+ /// The caller becomes responsible for freeing the memory. The object is not
+ /// dropped and remains initialized. Use [`static_from_ptr`] to reconstruct
+ /// an `SBox` from the pointer.
+ ///
+ /// This method is only intended for objects allocated from a static cache.
+ /// Calling it on an `SBox` backed by a [`KMemCacheHandle`] leaks a
+ /// reference on the handle, preventing the cache from ever being
+ /// destroyed, as [`static_from_ptr`] cannot restore the reference.
+ ///
+ /// [`static_from_ptr`]: SBox::static_from_ptr
+ pub fn into_ptr(self) -> *mut T {
+ debug_assert!(matches!(self.cache, CacheRef::Static(_)));
+ let ptr = self.ptr.as_ptr();
+ core::mem::forget(self);
+ ptr
+ }
+
+ /// Reconstructs an `SBox` from a raw pointer and cache.
+ ///
+ /// This is intended for use with objects that were previously converted to
+ /// raw pointers via [`into_ptr`], typically for passing through C code.
+ ///
+ /// [`into_ptr`]: SBox::into_ptr
+ ///
+ /// # Safety
+ ///
+ /// - `cache` must be a valid pointer to the `kmem_cache` from which `value`
+ /// was allocated.
+ /// - `cache` must be a statically allocated cache that is never destroyed.
+ /// - `value` must be a valid pointer to an initialized `T` that was
+ /// allocated from `cache`.
+ /// - The caller must ensure that no other `SBox` or reference exists for
+ /// `value`.
+ pub unsafe fn static_from_ptr(cache: *mut bindings::kmem_cache, value: *mut T) -> Self {
+ // INVARIANT: The caller guarantees `value` points to a valid,
+ // initialized `T` allocated from `cache`.
+ Self {
+ // SAFETY: By function safety requirements, `value` is not null.
+ ptr: unsafe { NonNull::new_unchecked(value) },
+ cache: CacheRef::Static(
+ // SAFETY: By function safety requirements, `cache` is not null.
+ unsafe { NonNull::new_unchecked(cache.cast()) },
+ ),
+ }
+ }
+}
+
impl<T: KMemCacheInit<T>> Deref for SBox<T> {
type Target = T;
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* Re: [PATCH v5 09/12] rust: mm: sheaf: allow use of C initialized static caches
2026-09-02 13:26 ` [PATCH v5 09/12] rust: mm: sheaf: allow use of C initialized static caches Andreas Hindborg
@ 2026-09-03 10:11 ` Vlastimil Babka (SUSE)
0 siblings, 0 replies; 15+ messages in thread
From: Vlastimil Babka (SUSE) @ 2026-09-03 10:11 UTC (permalink / raw)
To: Andreas Hindborg, Tamir Duberstein, Miguel Ojeda, Boqun Feng,
Gary Guo, Björn Roy Baron, Benno Lossin, Alice Ryhl,
Trevor Gross, Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Harry Yoo, Hao Li, Christoph Lameter,
David Rientjes, Roman Gushchin
Cc: rust-for-linux, linux-kernel, linux-fsdevel, linux-mm
On 9/2/26 15:26, Andreas Hindborg wrote:
> Extend the sheaf abstraction to support caches initialized by C at kernel
> boot time, in addition to dynamically created Rust caches.
Nit: in C code caches can be also dynamically created :)
> Introduce `KMemCache<T>` as a transparent wrapper around `kmem_cache` for
> static caches with `'static` lifetime. Rename the previous `KMemCache<T>`
> to `KMemCacheHandle<T>` to represent dynamically created, reference-counted
> caches.
>
> Add `Static` and `Dynamic` marker types along with `StaticSheaf` and
> `DynamicSheaf` type aliases to distinguish sheaves from each cache type.
> The `Sheaf` type now carries lifetime and allocation mode type parameters.
>
> Add `SBox::into_ptr()` and `SBox::static_from_ptr()` methods for passing
> allocations through C code via raw pointers.
>
> Add `KMemCache::from_raw()` for wrapping C-initialized static caches and
> `Sheaf::refill()` for replenishing a sheaf to a minimum size.
>
> Export `kmem_cache_prefill_sheaf`, `kmem_cache_return_sheaf`,
> `kmem_cache_refill_sheaf`, and `kmem_cache_alloc_from_sheaf_noprof` to
> allow Rust module code to use the sheaf API.
>
> Cc: Vlastimil Babka <vbabka@suse.cz>
> Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
> Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
> Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> Cc: linux-mm@kvack.org
> Assisted-by: LLM
> Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
Acked-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
^ permalink raw reply [flat|nested] 15+ messages in thread
* [PATCH v5 10/12] xarray, radix-tree: enable sheaf support for kmem_cache
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (8 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 09/12] rust: mm: sheaf: allow use of C initialized static caches Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 11/12] rust: xarray: add preload API Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 12/12] rust: xarray: document `Guard` lock drop semantics Andreas Hindborg
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm
The rust null block driver plans to rely on preloading xarray nodes from
the radix_tree_node_cachep kmem_cache.
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
lib/radix-tree.c | 14 ++++++++++----
1 file changed, 10 insertions(+), 4 deletions(-)
diff --git a/lib/radix-tree.c b/lib/radix-tree.c
index 976b9bd02a1b5..1cf0012b15ade 100644
--- a/lib/radix-tree.c
+++ b/lib/radix-tree.c
@@ -1598,10 +1598,16 @@ void __init radix_tree_init(void)
BUILD_BUG_ON(RADIX_TREE_MAX_TAGS + __GFP_BITS_SHIFT > 32);
BUILD_BUG_ON(ROOT_IS_IDR & ~GFP_ZONEMASK);
BUILD_BUG_ON(XA_CHUNK_SIZE > 255);
- radix_tree_node_cachep = kmem_cache_create("radix_tree_node",
- sizeof(struct radix_tree_node), 0,
- SLAB_PANIC | SLAB_RECLAIM_ACCOUNT,
- radix_tree_node_ctor);
+
+ struct kmem_cache_args args = {
+ .ctor = radix_tree_node_ctor,
+ .sheaf_capacity = 64,
+ };
+
+ radix_tree_node_cachep = kmem_cache_create(
+ "radix_tree_node", sizeof(struct radix_tree_node), &args,
+ SLAB_PANIC | SLAB_RECLAIM_ACCOUNT);
+
ret = cpuhp_setup_state_nocalls(CPUHP_RADIX_DEAD, "lib/radix:dead",
NULL, radix_tree_cpu_dead);
WARN_ON(ret < 0);
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 11/12] rust: xarray: add preload API
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (9 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 10/12] xarray, radix-tree: enable sheaf support for kmem_cache Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
2026-09-02 13:26 ` [PATCH v5 12/12] rust: xarray: document `Guard` lock drop semantics Andreas Hindborg
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm, Liam R. Howlett
Add a preload API that allows preallocating memory for XArray
insertions. This enables insertions to proceed without allocation
failures in contexts where memory allocation is not desirable, such as
in atomic contexts.
The implementation introduces `XArrayNode` representing a single XArray
node and `XArraySheaf` as a type alias for a sheaf of preallocated
nodes.
Add the function `xarray_kmem_cache` to provide access to the global XArray
node cache for creating sheaves.
Update `VacantEntry::insert` and `VacantEntry::insert_entry` to accept
an optional sheaf argument for preloaded memory. Add a new
`Guard::insert_entry` method for inserting with preload support. When an
insertion would fail due to ENOMEM, the XArray state API automatically
consumes a preallocated node from the sheaf if available.
Export `radix_tree_node_ctor` and `radix_tree_node_cachep` from C to
enable Rust code to work with the radix tree node cache.
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
include/linux/radix-tree.h | 3 +
lib/radix-tree.c | 5 +-
rust/bindings/bindings_helper.h | 3 +
rust/kernel/xarray.rs | 214 +++++++++++++++++++++++++++++++++++-----
rust/kernel/xarray/entry.rs | 27 +++--
5 files changed, 215 insertions(+), 37 deletions(-)
diff --git a/include/linux/radix-tree.h b/include/linux/radix-tree.h
index eae67015ce51a..c3699f12b070c 100644
--- a/include/linux/radix-tree.h
+++ b/include/linux/radix-tree.h
@@ -469,4 +469,7 @@ static __always_inline void __rcu **radix_tree_next_slot(void __rcu **slot,
slot = radix_tree_next_slot(slot, iter, \
RADIX_TREE_ITER_TAGGED | tag))
+
+void radix_tree_node_ctor(void *arg);
+
#endif /* _LINUX_RADIX_TREE_H */
diff --git a/lib/radix-tree.c b/lib/radix-tree.c
index 1cf0012b15ade..ddd67ce672f5c 100644
--- a/lib/radix-tree.c
+++ b/lib/radix-tree.c
@@ -33,6 +33,7 @@
* Radix tree node cache.
*/
struct kmem_cache *radix_tree_node_cachep;
+EXPORT_SYMBOL(radix_tree_node_cachep);
/*
* The radix tree is variable-height, so an insert operation not only has
@@ -1566,14 +1567,14 @@ void idr_destroy(struct idr *idr)
}
EXPORT_SYMBOL(idr_destroy);
-static void
-radix_tree_node_ctor(void *arg)
+void radix_tree_node_ctor(void *arg)
{
struct radix_tree_node *node = arg;
memset(node, 0, sizeof(*node));
INIT_LIST_HEAD(&node->private_list);
}
+EXPORT_SYMBOL(radix_tree_node_ctor);
static int radix_tree_cpu_dead(unsigned int cpu)
{
diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h
index 5dda2bb36e3c2..ccbd92880dc88 100644
--- a/rust/bindings/bindings_helper.h
+++ b/rust/bindings/bindings_helper.h
@@ -132,6 +132,9 @@ const gfp_t RUST_CONST_HELPER_XA_FLAGS_ALLOC1 = XA_FLAGS_ALLOC1;
* see https://github.com/rust-lang/rust-bindgen/issues/3347.
*/
const size_t RUST_CONST_HELPER_XAS_RESTART = (size_t)XAS_RESTART;
+const size_t RUST_CONST_HELPER_XA_CHUNK_SHIFT = XA_CHUNK_SHIFT;
+const size_t RUST_CONST_HELPER_XA_CHUNK_SIZE = XA_CHUNK_SIZE;
+extern struct kmem_cache *radix_tree_node_cachep;
const vm_flags_t RUST_CONST_HELPER_VM_MERGEABLE = VM_MERGEABLE;
const vm_flags_t RUST_CONST_HELPER_VM_READ = VM_READ;
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index cf7248bddb8b9..87123ab96a92f 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -5,6 +5,7 @@
//! C header: [`include/linux/xarray.h`](srctree/include/linux/xarray.h)
use core::{
+ convert::Infallible,
iter,
marker::PhantomData,
pin::Pin,
@@ -23,12 +24,18 @@
bindings,
build_assert::build_assert, //
error::{
+ code::*,
to_result,
Error,
Result, //
},
ffi::c_void,
fmt,
+ mm::sheaf::{
+ KMemCache,
+ SBox,
+ StaticSheaf, //
+ },
types::{
ForeignOwnable,
NotThreadSafe,
@@ -36,12 +43,63 @@
},
};
use pin_init::{
+ init,
pin_data,
pin_init,
pinned_drop,
+ Init,
PinInit, //
};
+/// Sheaf of preallocated [`XArray`] nodes.
+pub type XArraySheaf<'a> = StaticSheaf<'a, XArrayNode>;
+
+/// Returns a reference to the global XArray node cache.
+///
+/// This provides access to the kernel's `radix_tree_node_cachep`, which is the
+/// slab cache used for allocating internal XArray nodes. This cache can be used
+/// to create sheaves for preallocating XArray nodes.
+pub fn xarray_kmem_cache() -> &'static KMemCache<XArrayNode> {
+ // SAFETY: `radix_tree_node_cachep` is a valid, statically initialized
+ // kmem_cache that remains valid for the lifetime of the kernel. The cache
+ // is configured for `xa_node` objects which match our `XArrayNode` type.
+ unsafe { KMemCache::from_raw(bindings::radix_tree_node_cachep) }
+}
+
+/// An preallocated XArray node.
+///
+/// This represents a single preallocated internal node for an XArray.
+///
+/// This type is `#[repr(transparent)]` as it is cast to and from pointers to
+/// the inner [`bindings::xa_node`].
+#[repr(transparent)]
+pub struct XArrayNode {
+ node: Opaque<bindings::xa_node>,
+}
+
+// SAFETY: A preallocated `xa_node` is opaque storage for the C XArray
+// implementation, which moves nodes between CPUs freely. It is not tied to
+// the thread that allocated it.
+unsafe impl Send for XArrayNode {}
+
+impl kernel::mm::sheaf::KMemCacheInit<XArrayNode> for XArrayNode {
+ fn init() -> impl Init<Self, Infallible> {
+ init!(Self {
+ // SAFETY:
+ // - This initialization cannot fail and will never return `Err`.
+ // - The xa_node does not move during initialization.
+ node <- unsafe {
+ pin_init::init_from_closure(
+ |place: *mut Opaque<bindings::xa_node>| -> Result<(), Infallible> {
+ bindings::radix_tree_node_ctor(place.cast::<c_void>());
+ Ok(())
+ },
+ )
+ }
+ })
+ }
+}
+
/// An array which efficiently maps sparse integer indices to owned objects.
///
/// This is similar to a [`crate::alloc::kvec::Vec<Option<T>>`], but more efficient when there are
@@ -138,15 +196,22 @@ fn iter(&self) -> impl Iterator<Item = NonNull<c_void>> + '_ {
let mut index = 0;
// SAFETY: `self.xa` is always valid by the type invariant.
- iter::once(unsafe {
- bindings::xa_find(self.xa.get(), &mut index, usize::MAX, bindings::XA_PRESENT)
- })
- .chain(iter::from_fn(move || {
- // SAFETY: `self.xa` is always valid by the type invariant.
- Some(unsafe {
- bindings::xa_find_after(self.xa.get(), &mut index, usize::MAX, bindings::XA_PRESENT)
- })
- }))
+ Iterator::chain(
+ iter::once(unsafe {
+ bindings::xa_find(self.xa.get(), &mut index, usize::MAX, bindings::XA_PRESENT)
+ }),
+ iter::from_fn(move || {
+ // SAFETY: `self.xa` is always valid by the type invariant.
+ Some(unsafe {
+ bindings::xa_find_after(
+ self.xa.get(),
+ &mut index,
+ usize::MAX,
+ bindings::XA_PRESENT,
+ )
+ })
+ }),
+ )
.map_while(|ptr| NonNull::new(ptr.cast()))
}
@@ -167,7 +232,6 @@ pub fn try_lock(&self) -> Option<Guard<'_, T>> {
pub fn lock(&self) -> Guard<'_, T> {
// SAFETY: `self.xa` is always valid by the type invariant.
unsafe { bindings::xa_lock(self.xa.get()) };
-
Guard {
xa: self,
_not_send: NotThreadSafe,
@@ -269,7 +333,7 @@ pub fn get_mut(&mut self, index: usize) -> Option<T::BorrowedMut<'_>> {
///
/// match guard.entry(42) {
/// Entry::Vacant(entry) => {
- /// entry.insert(KBox::new(0x1337u32, GFP_KERNEL)?)?;
+ /// entry.insert(KBox::new(0x1337u32, GFP_ATOMIC)?, None)?;
/// }
/// Entry::Occupied(_) => unreachable!("We did not insert an entry yet"),
/// }
@@ -468,6 +532,45 @@ pub fn store(
Ok(unsafe { T::try_from_foreign(old) })
}
}
+
+ /// Inserts a value and returns an occupied entry for further operations.
+ ///
+ /// If a value is already present, the operation fails.
+ ///
+ /// This method will not drop the XArray lock. If memory allocation is
+ /// required for the operation to succeed, the user should supply memory
+ /// through the `preload` argument.
+ ///
+ /// # Examples
+ ///
+ /// ```
+ /// # use kernel::{prelude::*, xarray::{AllocKind, XArray}};
+ /// let mut xa = KBox::pin_init(XArray::<KBox<u32>>::new(AllocKind::Alloc), GFP_KERNEL)?;
+ /// let mut guard = xa.lock();
+ ///
+ /// assert_eq!(guard.get(42), None);
+ ///
+ /// let value = KBox::new(0x1337u32, GFP_ATOMIC)?;
+ /// let entry = guard.insert_entry(42, value, None)?;
+ /// let borrowed = entry.into_mut();
+ /// assert_eq!(borrowed, &0x1337);
+ ///
+ /// # Ok::<(), kernel::error::Error>(())
+ /// ```
+ pub fn insert_entry<'b>(
+ &'b mut self,
+ index: usize,
+ value: T,
+ preload: Option<&mut XArraySheaf<'_>>,
+ ) -> Result<OccupiedEntry<'a, 'b, T>, StoreError<T>> {
+ match self.entry(index) {
+ Entry::Vacant(entry) => entry.insert_entry(value, preload),
+ Entry::Occupied(_) => Err(StoreError {
+ error: EBUSY,
+ value,
+ }),
+ }
+ }
}
/// Internal state for XArray iteration and entry operations.
@@ -485,6 +588,30 @@ pub(crate) struct XArrayState<R> {
state: bindings::xa_state,
}
+impl<R> Drop for XArrayState<R> {
+ fn drop(&mut self) {
+ free_xa_alloc(&mut self.state);
+ }
+}
+
+fn free_xa_alloc(state: &mut bindings::xa_state) {
+ if !state.xa_alloc.is_null() {
+ // SAFETY:
+ // - `xa_alloc` is only set via `SBox::into_ptr()` in `insert()` where
+ // the node comes from an `XArraySheaf` backed by `radix_tree_node_cachep`.
+ // - `xa_alloc` points to a valid, initialized `XArrayNode`.
+ // - The caller has exclusive ownership of `xa_alloc`, and no other
+ // `SBox` or reference exists for this value.
+ drop(unsafe {
+ SBox::<XArrayNode>::static_from_ptr(
+ bindings::radix_tree_node_cachep,
+ state.xa_alloc.cast(),
+ )
+ });
+ state.xa_alloc = null_mut();
+ }
+}
+
impl<'a, R, T> XArrayState<R>
where
T: ForeignOwnable + 'a,
@@ -595,23 +722,50 @@ fn replace(&mut self, new: *mut c_void) -> *mut c_void {
old
}
- fn insert(&mut self, value: T) -> Result<*mut c_void, StoreError<T>> {
+ fn insert(
+ &mut self,
+ value: T,
+ mut preload: Option<&mut XArraySheaf<'_>>,
+ ) -> Result<*mut c_void, StoreError<T>> {
let new = T::into_foreign(value).cast();
- // SAFETY: `self.state` is a valid `xa_state` by the type invariant. By the same
- // invariant, `self.state.xa` aliases the xarray reachable through `self.guard`,
- // whose lock we hold. `new` came from `T::into_foreign`.
- unsafe { bindings::xas_store(&mut self.state, new) };
-
- // All arrays created by this abstraction have `XA_FLAGS_TRACK_FREE` set, so the
- // free mark must be cleared for a newly occupied index, as `__xa_store` does.
- // This is a no-op if the store above failed.
- //
- // SAFETY: `self.state` is a valid `xa_state` by the type invariant, and we hold
- // the lock on the xarray it refers to.
- unsafe { bindings::xas_clear_mark(&self.state, bindings::XA_FREE_MARK) };
+ loop {
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant. By the same
+ // invariant, `self.state.xa` aliases the xarray reachable through `self.guard`,
+ // whose lock we hold. `new` came from `T::into_foreign`.
+ unsafe { bindings::xas_store(&mut self.state, new) };
- self.status().map(|()| new).map_err(|error| {
+ // All arrays created by this abstraction have `XA_FLAGS_TRACK_FREE` set, so the
+ // free mark must be cleared for a newly occupied index, as `__xa_store` does.
+ // This is a no-op if the store above failed.
+ //
+ // SAFETY: `self.state` is a valid `xa_state` by the type invariant, and we hold
+ // the lock on the xarray it refers to.
+ unsafe { bindings::xas_clear_mark(&self.state, bindings::XA_FREE_MARK) };
+
+ match self.status() {
+ Ok(()) => break Ok(new),
+ Err(ENOMEM) => {
+ debug_assert!(self.state.xa_alloc.is_null());
+ let node = match preload.as_mut().map(|sheaf| sheaf.alloc().ok_or(ENOMEM)) {
+ None => break Err(ENOMEM),
+ Some(Err(e)) => break Err(e),
+ Some(Ok(node)) => node,
+ };
+
+ self.state.xa_alloc = node.into_ptr().cast();
+
+ // On allocation failure, `xas_store` leaves `XA_ERROR(-ENOMEM)` in
+ // `self.state.xa_node`, which makes further operations on the state fail
+ // immediately without consuming `xa_alloc`. Reset the state so the retry
+ // walks the tree again, as `xas_nomem` does.
+ self.restart_at(self.state.xa_index);
+ continue;
+ }
+ Err(e) => break Err(e),
+ }
+ }
+ .map_err(|error| {
// SAFETY: `new` came from `T::into_foreign` and `xas_store` does not take
// ownership of the value on error.
let value = unsafe { T::from_foreign(new) };
@@ -619,10 +773,16 @@ fn insert(&mut self, value: T) -> Result<*mut c_void, StoreError<T>> {
})
}
- /// Consumes `self` and returns the inner `&mut Guard`.
+ /// Consumes `self`, releases any preallocated node held in `xa_alloc`, and
+ /// returns the inner `&mut Guard`.
#[inline]
pub(crate) fn into_guard(self) -> &'b mut Guard<'a, T> {
- self.guard
+ // Suppress the `Drop` impl so we can move `guard` out by hand.
+ let mut this = core::mem::ManuallyDrop::new(self);
+ free_xa_alloc(&mut this.state);
+ // SAFETY: `ManuallyDrop` prevents `Drop::drop` from running, so this is the only place
+ // that consumes `guard`. `state` has no other resources after `free_xa_alloc`.
+ unsafe { core::ptr::read(&this.guard) }
}
}
diff --git a/rust/kernel/xarray/entry.rs b/rust/kernel/xarray/entry.rs
index c6c5385e6b4f9..bd295358d9751 100644
--- a/rust/kernel/xarray/entry.rs
+++ b/rust/kernel/xarray/entry.rs
@@ -3,6 +3,7 @@
use super::{
Guard,
StoreError,
+ XArraySheaf,
XArrayState, //
};
use core::ptr::NonNull;
@@ -75,7 +76,8 @@ pub fn into_guard(self) -> &'b mut Guard<'a, T> {
/// Returns a reference to the newly inserted value.
///
/// - This method will fail if the nodes on the path to the index
- /// represented by this entry are not present in the XArray.
+ /// represented by this entry are not present in the XArray and no memory
+ /// is available via the `preload` argument.
/// - This method will not drop the XArray lock.
///
///
@@ -90,7 +92,7 @@ pub fn into_guard(self) -> &'b mut Guard<'a, T> {
///
/// if let Entry::Vacant(entry) = guard.entry(42) {
/// let value = KBox::new(0x1337u32, GFP_ATOMIC)?;
- /// let borrowed = entry.insert(value)?;
+ /// let borrowed = entry.insert(value, None)?;
/// assert_eq!(*borrowed, 0x1337);
/// }
///
@@ -98,8 +100,12 @@ pub fn into_guard(self) -> &'b mut Guard<'a, T> {
///
/// # Ok::<(), kernel::error::Error>(())
/// ```
- pub fn insert(mut self, value: T) -> Result<T::BorrowedMut<'b>, StoreError<T>> {
- let new = self.state.insert(value)?;
+ pub fn insert(
+ mut self,
+ value: T,
+ preload: Option<&mut XArraySheaf<'_>>,
+ ) -> Result<T::BorrowedMut<'b>, StoreError<T>> {
+ let new = self.state.insert(value, preload)?;
// SAFETY: `new` came from `T::into_foreign`. The entry has exclusive
// ownership of `new` as it holds a mutable reference to `Guard`.
@@ -109,7 +115,8 @@ pub fn insert(mut self, value: T) -> Result<T::BorrowedMut<'b>, StoreError<T>> {
/// Inserts a value and returns an occupied entry representing the newly inserted value.
///
/// - This method will fail if the nodes on the path to the index
- /// represented by this entry are not present in the XArray.
+ /// represented by this entry are not present in the XArray and no memory
+ /// is available via the `preload` argument.
/// - This method will not drop the XArray lock.
///
/// # Examples
@@ -123,7 +130,7 @@ pub fn insert(mut self, value: T) -> Result<T::BorrowedMut<'b>, StoreError<T>> {
///
/// if let Entry::Vacant(entry) = guard.entry(42) {
/// let value = KBox::new(0x1337u32, GFP_ATOMIC)?;
- /// let occupied = entry.insert_entry(value)?;
+ /// let occupied = entry.insert_entry(value, None)?;
/// assert_eq!(occupied.index(), 42);
/// }
///
@@ -131,8 +138,12 @@ pub fn insert(mut self, value: T) -> Result<T::BorrowedMut<'b>, StoreError<T>> {
///
/// # Ok::<(), kernel::error::Error>(())
/// ```
- pub fn insert_entry(mut self, value: T) -> Result<OccupiedEntry<'a, 'b, T>, StoreError<T>> {
- let new = self.state.insert(value)?;
+ pub fn insert_entry(
+ mut self,
+ value: T,
+ preload: Option<&mut XArraySheaf<'_>>,
+ ) -> Result<OccupiedEntry<'a, 'b, T>, StoreError<T>> {
+ let new = self.state.insert(value, preload)?;
Ok(OccupiedEntry::<'a, 'b, T> {
state: self.state,
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread* [PATCH v5 12/12] rust: xarray: document `Guard` lock drop semantics
2026-09-02 13:25 [PATCH v5 00/12] rust: xarray: add entry API with preloading Andreas Hindborg
` (10 preceding siblings ...)
2026-09-02 13:26 ` [PATCH v5 11/12] rust: xarray: add preload API Andreas Hindborg
@ 2026-09-02 13:26 ` Andreas Hindborg
11 siblings, 0 replies; 15+ messages in thread
From: Andreas Hindborg @ 2026-09-02 13:26 UTC (permalink / raw)
To: Tamir Duberstein, Miguel Ojeda, Boqun Feng, Gary Guo,
Björn Roy Baron, Benno Lossin, Alice Ryhl, Trevor Gross,
Danilo Krummrich, Daniel Almeida, Alexandre Courbot,
Onur Özkan, Matthew Wilcox, Andrew Morton, Lorenzo Stoakes,
Liam R. Howlett, Vlastimil Babka, Harry Yoo, Hao Li,
Christoph Lameter, David Rientjes, Roman Gushchin
Cc: Andreas Hindborg, rust-for-linux, linux-kernel, linux-fsdevel,
linux-mm, Sashiko
`Guard::store` calls `__xa_store`, which drops the xarray lock to
allocate memory when called with blocking allocation flags and
reacquires it afterwards. A Rust lock guard is normally expected to
provide continuous mutual exclusion for its entire lifetime, so this
behavior can surprise users: a check-then-act sequence spanning a
blocking `store` call is not atomic.
Document the behavior on `Guard` and expand the `store` docs,
pointing to the entry API with preallocated memory as the way to
modify the array without dropping the lock.
Suggested-by: Sashiko <sashiko-bot@kernel.org>
Assisted-by: LLM
Signed-off-by: Andreas Hindborg <a.hindborg@kernel.org>
---
rust/kernel/xarray.rs | 22 +++++++++++++++++++++-
1 file changed, 21 insertions(+), 1 deletion(-)
diff --git a/rust/kernel/xarray.rs b/rust/kernel/xarray.rs
index 87123ab96a92..a11472acc661 100644
--- a/rust/kernel/xarray.rs
+++ b/rust/kernel/xarray.rs
@@ -242,6 +242,21 @@ pub fn lock(&self) -> Guard<'_, T> {
/// A lock guard.
///
/// The lock is unlocked when the guard goes out of scope.
+///
+/// # Temporary lock drops
+///
+/// Unlike a typical Rust lock guard, holding a `Guard` does not guarantee
+/// continuous mutual exclusion for its entire lifetime: [`store`] may drop and
+/// reacquire the lock to allocate memory when called with blocking allocation
+/// flags. Other threads may lock and modify the array in that window, so a
+/// sequence of operations on the guard that spans such a call is not atomic.
+///
+/// To modify the array without dropping the lock, use the entry API with
+/// preallocated memory, see [`entry`] and [`insert_entry`].
+///
+/// [`store`]: Guard::store
+/// [`entry`]: Guard::entry
+/// [`insert_entry`]: Guard::insert_entry
#[must_use = "the lock unlocks immediately when the guard is unused"]
pub struct Guard<'a, T: ForeignOwnable> {
xa: &'a XArray<T>,
@@ -485,7 +500,12 @@ pub fn remove(&mut self, index: usize) -> Option<T> {
/// Stores an element at the given index.
///
- /// May drop the lock if needed to allocate memory, and then reacquire it afterwards.
+ /// If `gfp` contains blocking allocation flags, this method may drop the
+ /// lock to allocate memory and reacquire it afterwards. Other threads may
+ /// lock and modify the array in that window, so callers must not rely on
+ /// this method being atomic with respect to other operations on the
+ /// guard. To store without dropping the lock, use [`Guard::insert_entry`]
+ /// with preallocated memory.
///
/// On success, returns the element which was previously at the given index.
///
--
2.51.2
^ permalink raw reply related [flat|nested] 15+ messages in thread