[PATCH v12 0/2] rust: Add support for reserving of ranges of IDs

[PATCH v12 0/2] rust: Add support for reserving of ranges of IDs

[Eliot Courtney]

Add support for reserving of ranges of IDs, with a usage in nova-core
for channel IDs.

The bitmap and `IdPool` parts from earlier versions are in v7.4-rc1
now, and this version has the remaining patches: exact `IdPool`
capacities and `ChannelIdPool`.

Channel ID tracking is needed for allotting ranges of channel IDs to
vGPU guests, and later for regular host channel ID reservation.
nova-core needs allocation of a contiguous sequence of IDs with a
specific length and sometimes a specific alignment [1].

About the tradeoffs between different data structures:
- IDA/xarray do not support allocating a contiguous sequence of IDs
  (ida_alloc_range() allocates a single ID within a range, not a contiguous
  sequence).
- A maple tree works, but is not as good a fit. The ID space is small
  (limited to 2048) and aligned allocation needs an alloc_range()+erase() retry
  loop (plus a Mutex around it, or new mas_empty_area() bindings) that
  essentially reimplements bitmap_find_next_zero_area(). See the maple tree
  version at [2]. For 2048 IDs a bitmap is also considerably faster and smaller
  [3].
- The bitmap API natively supports aligned contiguous area allocation
  (bitmap_find_next_zero_area()).

This is based on drm-rust-next.

[1]: https://lore.kernel.org/all/84bc8bd2-e292-4b84-9580-a1b5df4c5bdc@nvidia.com/
[2]: https://lore.kernel.org/all/20260710-chid-maple-v1-1-4ee869055268@nvidia.com/
[3]: https://lore.kernel.org/all/20260717053241.916441-1-ynorov@nvidia.com/

Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
---
Changes in v12:
- Drop the patches that were applied to drm-rust-next and driver-core
- Rebase on drm-rust-next
- Drop the zero capacity part of the patch #1 commit message (Yury)
- Add reviewed-by tags
- Link to v11: https://patch.msgid.link/20261109-chid-v11-0-082aa9bf5a91@nvidia.com

Changes in v11:
- Rebase on drm-rust-next (pre-reqs are merged now)
- Split patch #1 by subsystem (Danilo). Just splitting with no actual
  changes, so I kept the reviewed-by tags on the new patches
- Add reviewed-by tags
- Link to v10: https://patch.msgid.link/20261019-chid-v10-0-67de9180b4ed@nvidia.com

Changes in v10:
- Drop the patches that were applied to rust-next
- Rebase on drm-rust-next (plus pre-reqs)
- Use a capacity of 1 instead of 0 in the grow_request() example in
  patch #2 (Yury)
- Convert the remaining Alignment::new() users in nova-core in patch #1
- Link to v9: https://patch.msgid.link/20260930-chid-v9-0-0d6cca376cff@nvidia.com

Changes in v9:
- Rebase on drm-rust-next
- Moved cv! to a separate series which this now depends on
- Reference Link: tag from the text + indent code in patch #4 commit
  message (Miguel)
- Add reviewed-by tags
- Link to v8: https://patch.msgid.link/20260827-chid-v8-0-bc74c77d0214@nvidia.com

Changes in v8:
- Clarify overflow cases in patch #2 (Alex)
- Replace nz! with generalized cv! macro (Gary)
- Add cv! macro to prelude
- Split SizeConstants and adding new constants patch (Alex)
- Move alignment size constants conersions patch (Alex)
- Drop NonZero on IdPool, fix grow_request instead (Alice, Alex)
- Add shrink/grow doctests (Burak)
- Link to v7: https://patch.msgid.link/20260817-chid-v7-0-a5872e64d8f4@nvidia.com

Changes in v7:
- Add nz! macro for compile time NonZero
- Implement SizeConstants for Alignment + add sub 1k size constants
- Use Alignment constants + nz! in this series
- Add final patch converting existing callers to use Alignment constants
- Rename alloc_area/release_area -> reserve_ids / release_ids
- Link to v6: https://patch.msgid.link/20260813-chid-v6-0-160be5dfb5bd@nvidia.com

Changes in v6:
- Take a NonZero nbits in `next_zero_area_off`, `set` and `clear` (Yury)
- Remove `UnusedArea`, `IdPool::alloc_area` now allocates directly (Yury)
- Add patch: take a NonZero capacity in `IdPool::with_capacity()`
- Add patch: do not round the capacity up to `BitmapVec::MAX_INLINE_LEN`,
  which also removes the bounds check in `ChannelIdPool::alloc_area`
- Add more testing of Drop for `ChannelIdPool` (Yury)
- Link to v5: https://patch.msgid.link/20260812-chid-v5-0-6c767770b3f4@nvidia.com

Changes in v5:
- `bitmap_assert!` i32::MAX length for Bitmap::from_raw* (Yury)
- Only run overflow check on 32-bit (Yury)
- Link to v4: https://patch.msgid.link/20260810-chid-v4-0-c9f206fdcb97@nvidia.com

Changes in v4:
- Add `next_zero_area_off` to match C code (Yury)
- Replace overflow checks to match C code in bitmap-for-next.
- Tighten `Bitmap` unsafe contract to disallow Bitmaps larger than i32::MAX
- Link to v3: https://patch.msgid.link/20260729-chid-v3-0-20cc08032bbc@nvidia.com

Changes in v3:
- Use `Alignment` type in id_pool and bitmap (Alice)
- Remove hang check on the basis that it's extraordinarily rare.
- Link to v2: https://patch.msgid.link/20260723-chid-v2-0-c35e5e9fb3d9@nvidia.com

Changes in v2:
- Collected Alice's Reviewed-by on patch 1.
- Address Yury's comments w.r.t. using __bitmap_set etc directly.
- Address Yury's comments w.r.t. following the C names
- Additionally check for an overflow case that causes a hang
- Added more info to cover letter + patch 4 w.r.t. channel ID allottment
  requirements
- Add align parameter to ChannelIdPool::alloc_area() plus an aligned
  allocation test
- Add missing INVARIANT comment when constructing UnusedArea
- Link to v1:
  https://patch.msgid.link/20260703-chid-v1-0-84fe8259e46e@nvidia.com

---
Eliot Courtney (2):
      rust: id_pool: do not round capacity up to BitmapVec::MAX_INLINE_LEN
      gpu: nova-core: add ChannelIdPool

 drivers/gpu/nova-core/gpu.rs         |   2 +
 drivers/gpu/nova-core/gpu/channel.rs | 198 +++++++++++++++++++++++++++++++++++
 rust/kernel/id_pool.rs               |  38 +++++--
 3 files changed, 232 insertions(+), 6 deletions(-)
---
base-commit: 6e557330b874b0e1fdd03e8fe4c438f36c93eefa
change-id: 20260608-chid-18fa943c6d6c

Best regards,
--  
Eliot Courtney <ecourtney@nvidia.com>

[PATCH v12 1/2] rust: id_pool: do not round capacity up to BitmapVec::MAX_INLINE_LEN

[Eliot Courtney]

Current code in IdPool::with_capacity rounds the capacity up to
BitmapVec::MAX_INLINE_LEN, but BitmapVec::new works fine with values
smaller than this and still uses an inline representation. Remove this
behaviour.

Change `grow_request` to grow to at least `BitmapVec::MAX_INLINE_LEN`
instead, mirroring the capacity floor in `shrink_request`.

Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Reviewed-by: Burak Emir <burak.emir@gmail.com>
Reviewed-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
---
 rust/kernel/id_pool.rs | 38 ++++++++++++++++++++++++++++++++------
 1 file changed, 32 insertions(+), 6 deletions(-)

diff --git a/rust/kernel/id_pool.rs b/rust/kernel/id_pool.rs
index 92eeb7caa243..f77bd90325ff 100644
--- a/rust/kernel/id_pool.rs
+++ b/rust/kernel/id_pool.rs
@@ -112,13 +112,8 @@ pub fn new() -> Self {
     }
 
     /// Constructs a new [`IdPool`] with space for a specific number of bits.
-    ///
-    /// A capacity below [`MAX_INLINE_LEN`] is adjusted to [`MAX_INLINE_LEN`].
-    ///
-    /// [`MAX_INLINE_LEN`]: BitmapVec::MAX_INLINE_LEN
     #[inline]
     pub fn with_capacity(num_ids: usize, flags: Flags) -> Result<Self, AllocError> {
-        let num_ids = usize::max(num_ids, BitmapVec::MAX_INLINE_LEN);
         let map = BitmapVec::new(num_ids, flags)?;
         Ok(Self { map })
     }
@@ -152,6 +147,13 @@ pub fn capacity(&self) -> usize {
     /// let resizer = alloc_request.realloc(GFP_KERNEL)?;
     /// pool.shrink(resizer);
     /// assert_eq!(pool.capacity(), BitmapVec::MAX_INLINE_LEN);
+    ///
+    /// // A pool at the `MAX_INLINE_LEN` floor cannot shrink further.
+    /// assert!(pool.shrink_request().is_none());
+    ///
+    /// // Neither can a pool with a capacity below `MAX_INLINE_LEN`.
+    /// let small = IdPool::with_capacity(8, GFP_KERNEL)?;
+    /// assert!(small.shrink_request().is_none());
     /// # Ok::<(), AllocError>(())
     /// ```
     #[inline]
@@ -198,12 +200,36 @@ pub fn shrink(&mut self, mut resizer: PoolResizer) {
 
     /// Returns a [`ReallocRequest`] for growing this [`IdPool`], if possible.
     ///
+    /// Grows to at least [`MAX_INLINE_LEN`].
     /// The capacity of an [`IdPool`] cannot be grown above [`MAX_LEN`].
     ///
+    /// [`MAX_INLINE_LEN`]: BitmapVec::MAX_INLINE_LEN
     /// [`MAX_LEN`]: BitmapVec::MAX_LEN
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use kernel::{
+    ///     alloc::AllocError,
+    ///     bitmap::BitmapVec,
+    ///     id_pool::IdPool, //
+    /// };
+    ///
+    /// // Grow goes to at least BitmapVec::MAX_INLINE_LEN.
+    /// let mut pool = IdPool::with_capacity(1, GFP_KERNEL)?;
+    /// let resizer = pool.grow_request().ok_or(AllocError)?.realloc(GFP_KERNEL)?;
+    /// pool.grow(resizer);
+    /// assert_eq!(pool.capacity(), BitmapVec::MAX_INLINE_LEN);
+    ///
+    /// // Grow doubles if at least BitmapVec::MAX_INLINE_LEN.
+    /// let resizer = pool.grow_request().ok_or(AllocError)?.realloc(GFP_KERNEL)?;
+    /// pool.grow(resizer);
+    /// assert_eq!(pool.capacity(), 2 * BitmapVec::MAX_INLINE_LEN);
+    /// # Ok::<(), AllocError>(())
+    /// ```
     #[inline]
     pub fn grow_request(&self) -> Option<ReallocRequest> {
-        let num_ids = self.capacity() * 2;
+        let num_ids = usize::max(BitmapVec::MAX_INLINE_LEN, self.capacity() * 2);
         if num_ids > BitmapVec::MAX_LEN {
             return None;
         }

-- 
2.56.0

[PATCH v12 2/2] gpu: nova-core: add ChannelIdPool

[Eliot Courtney]

Add `ChannelIdPool` which adds automatic tracking and releasing of
channel IDs on top of `IdPool`. This is necessary for apportioning
ranges of channel IDs to be used in e.g. vGPU.

Channel IDs are allocated as a contiguous sequence with a specific
length and sometimes a specific alignment [1] for vGPU. The ID space is
small (limited to 2048) and allocation is not on a hot path, so a
bitmap-backed `IdPool` is a better fit than IDA/xarray (which allocate a
single ID within a range, not a contiguous sequence) or a maple tree
(where aligned allocation needs an alloc_range()+erase() retry loop that
essentially reimplements bitmap_find_next_zero_area()) [2]. It is
also faster than maple tree [3].

Link: https://lore.kernel.org/all/84bc8bd2-e292-4b84-9580-a1b5df4c5bdc@nvidia.com/ # [1]
Link: https://lore.kernel.org/all/20260710-chid-maple-v1-1-4ee869055268@nvidia.com/ # [2]
Link: https://lore.kernel.org/all/20260717053241.916441-1-ynorov@nvidia.com/ # [3]
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
---
 drivers/gpu/nova-core/gpu.rs         |   2 +
 drivers/gpu/nova-core/gpu/channel.rs | 198 +++++++++++++++++++++++++++++++++++
 2 files changed, 200 insertions(+)

diff --git a/drivers/gpu/nova-core/gpu.rs b/drivers/gpu/nova-core/gpu.rs
index d185e927233f..90b1d24cdf00 100644
--- a/drivers/gpu/nova-core/gpu.rs
+++ b/drivers/gpu/nova-core/gpu.rs
@@ -53,6 +53,8 @@
     vgpu::VgpuManager, //
 };
 
+#[cfg_attr(not(CONFIG_KUNIT = "y"), expect(dead_code))]
+mod channel;
 mod hal;
 mod regs;
 
diff --git a/drivers/gpu/nova-core/gpu/channel.rs b/drivers/gpu/nova-core/gpu/channel.rs
new file mode 100644
index 000000000000..485efaba059d
--- /dev/null
+++ b/drivers/gpu/nova-core/gpu/channel.rs
@@ -0,0 +1,198 @@
+// SPDX-License-Identifier: GPL-2.0
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+
+//! Channel ID allocation.
+
+use core::{
+    num::NonZero,
+    ops::{
+        Deref,
+        Range, //
+    }, //
+};
+
+use kernel::{
+    id_pool::IdPool,
+    prelude::*,
+    ptr::Alignment,
+    sync::{
+        new_mutex,
+        Mutex, //
+    }, //
+};
+
+/// Pool for tracking reservations of channel IDs.
+#[pin_data]
+pub(crate) struct ChannelIdPool {
+    #[pin]
+    inner: Mutex<IdPool>,
+}
+
+impl ChannelIdPool {
+    /// Creates a pool managing `num_chids` channel IDs.
+    pub(crate) fn new(num_chids: NonZero<usize>) -> impl PinInit<Self, Error> {
+        try_pin_init!(Self {
+            inner <- new_mutex!(IdPool::with_capacity(num_chids.get(), GFP_KERNEL)?),
+        })
+    }
+
+    /// Reserves a contiguous area of `count` channel IDs starting at a multiple of `align`,
+    /// returning a guard that releases the area on drop.
+    pub(crate) fn reserve_ids(
+        &self,
+        count: NonZero<usize>,
+        align: Alignment,
+    ) -> Result<ChannelIdReservation<'_>> {
+        let mut ids = self.inner.lock();
+        let range = ids.reserve_ids(0, count, align).ok_or(ENOSPC)?;
+        Ok(ChannelIdReservation { pool: self, range })
+    }
+}
+
+/// A reserved contiguous area of channel IDs.
+///
+/// Releases the whole area back to its [`ChannelIdPool`] when dropped. Releasing locks a
+/// sleeping [`Mutex`], so the area must be dropped in a context that is allowed to sleep.
+#[must_use = "the channel ID reservation is released immediately when unused"]
+pub(crate) struct ChannelIdReservation<'a> {
+    pool: &'a ChannelIdPool,
+    range: Range<usize>,
+}
+
+impl Drop for ChannelIdReservation<'_> {
+    fn drop(&mut self) {
+        self.pool.inner.lock().release_ids(&self.range);
+    }
+}
+
+impl Deref for ChannelIdReservation<'_> {
+    type Target = Range<usize>;
+
+    fn deref(&self) -> &Self::Target {
+        &self.range
+    }
+}
+
+#[kunit_tests(nova_core_channel)]
+mod tests {
+    use super::*;
+    use kernel::sizes::SizeConstants;
+
+    #[test]
+    fn chid_reservation() -> Result {
+        let pool = KBox::pin_init(ChannelIdPool::new(cv!(2048)), GFP_KERNEL)?;
+
+        let first = pool.reserve_ids(cv!(48), Alignment::SZ_1)?;
+        assert_eq!(0, first.start);
+        assert_eq!(48, first.len());
+        assert_eq!(48, first.end);
+
+        let second = pool.reserve_ids(cv!(48), Alignment::SZ_1)?;
+        assert!(first.end <= second.start || second.end <= first.start);
+
+        let first_start = first.start;
+        drop(first);
+        assert_eq!(
+            first_start,
+            pool.reserve_ids(cv!(48), Alignment::SZ_1)?.start
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn chid_reservation_drop() -> Result {
+        let pool = KBox::pin_init(ChannelIdPool::new(cv!(8)), GFP_KERNEL)?;
+
+        let a = pool.reserve_ids(cv!(3), Alignment::SZ_1)?;
+        let b = pool.reserve_ids(cv!(3), Alignment::SZ_1)?;
+        let c = pool.reserve_ids(cv!(2), Alignment::SZ_1)?;
+        assert_eq!(0, a.start);
+        assert_eq!(3, b.start);
+        assert_eq!(6, c.start);
+
+        drop(b);
+
+        // Only have space for 3 IDs right now.
+        assert_eq!(
+            Err(ENOSPC),
+            pool.reserve_ids(cv!(4), Alignment::SZ_1).map(|_| ())
+        );
+        let b = pool.reserve_ids(cv!(3), Alignment::SZ_1)?;
+        assert_eq!(3, b.start);
+
+        drop(a);
+        drop(c);
+        drop(b);
+
+        // Everything was dropped so the pool should be empty.
+        assert_eq!(0, pool.reserve_ids(cv!(8), Alignment::SZ_1)?.start);
+        Ok(())
+    }
+
+    #[test]
+    fn chid_bounded_by_num_chids() -> Result {
+        let pool = KBox::pin_init(ChannelIdPool::new(cv!(4)), GFP_KERNEL)?;
+
+        {
+            let a = pool.reserve_ids(cv!(1), Alignment::SZ_1)?;
+            let b = pool.reserve_ids(cv!(1), Alignment::SZ_1)?;
+            let c = pool.reserve_ids(cv!(1), Alignment::SZ_1)?;
+            let d = pool.reserve_ids(cv!(1), Alignment::SZ_1)?;
+            assert_eq!(0, a.start);
+            assert_eq!(1, b.start);
+            assert_eq!(2, c.start);
+            assert_eq!(3, d.start);
+            assert_eq!(
+                Err(ENOSPC),
+                pool.reserve_ids(cv!(1), Alignment::SZ_1).map(|_| ())
+            );
+        }
+
+        assert_eq!(0, pool.reserve_ids(cv!(4), Alignment::SZ_1)?.start);
+        assert_eq!(
+            Err(ENOSPC),
+            pool.reserve_ids(cv!(5), Alignment::SZ_1).map(|_| ())
+        );
+
+        let head = pool.reserve_ids(cv!(3), Alignment::SZ_1)?;
+        assert_eq!(0, head.start);
+        assert_eq!(
+            Err(ENOSPC),
+            pool.reserve_ids(cv!(2), Alignment::SZ_1).map(|_| ())
+        );
+        assert_eq!(3, pool.reserve_ids(cv!(1), Alignment::SZ_1)?.start);
+        Ok(())
+    }
+
+    #[test]
+    fn chid_reservation_aligned() -> Result {
+        let pool = KBox::pin_init(ChannelIdPool::new(cv!(16)), GFP_KERNEL)?;
+
+        // Alloc 0 so the first fit for the next area is unaligned.
+        let pad = pool.reserve_ids(cv!(1), Alignment::SZ_1)?;
+        assert_eq!(0, pad.start);
+
+        let a = pool.reserve_ids(cv!(4), Alignment::SZ_4)?;
+        assert_eq!(4, a.start);
+
+        // The area skipped over by the aligned allocation should still be available.
+        let b = pool.reserve_ids(cv!(1), Alignment::SZ_1)?;
+        assert_eq!(1, b.start);
+
+        let c = pool.reserve_ids(cv!(8), Alignment::SZ_8)?;
+        assert_eq!(8, c.start);
+
+        // Only 2 IDs left.
+        assert_eq!(
+            Err(ENOSPC),
+            pool.reserve_ids(cv!(4), Alignment::SZ_4).map(|_| ())
+        );
+        assert_eq!(
+            Err(ENOSPC),
+            pool.reserve_ids(cv!(1), Alignment::SZ_32).map(|_| ())
+        );
+
+        assert_eq!(2, pool.reserve_ids(cv!(2), Alignment::SZ_1)?.start);
+        Ok(())
+    }
+}

-- 
2.56.0