rust_binder: add Rust Binder driver

We're generally not proponents of rewrites (nasty uncomfortable things
that make you late for dinner!). So why rewrite Binder?

Binder has been evolving over the past 15+ years to meet the evolving
needs of Android. Its responsibilities, expectations, and complexity
have grown considerably during that time. While we expect Binder to
continue to evolve along with Android, there are a number of factors
that currently constrain our ability to develop/maintain it. Briefly
those are:

1. Complexity: Binder is at the intersection of everything in Android and
   fulfills many responsibilities beyond IPC. It has become many things
   to many people, and due to its many features and their interactions
   with each other, its complexity is quite high. In just 6kLOC it must
   deliver transactions to the right threads. It must correctly parse
   and translate the contents of transactions, which can contain several
   objects of different types (e.g., pointers, fds) that can interact
   with each other. It controls the size of thread pools in userspace,
   and ensures that transactions are assigned to threads in ways that
   avoid deadlocks where the threadpool has run out of threads. It must
   track refcounts of objects that are shared by several processes by
   forwarding refcount changes between the processes correctly.  It must
   handle numerous error scenarios and it combines/nests 13 different
   locks, 7 reference counters, and atomic variables. Finally, It must
   do all of this as fast and efficiently as possible. Minor performance
   regressions can cause a noticeably degraded user experience.

2. Things to improve: Thousand-line functions [1], error-prone error
   handling [2], and confusing structure can occur as a code base grows
   organically. After more than a decade of development, this codebase
   could use an overhaul.

[1]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/android/binder.c?h=v6.5#n2896
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/android/binder.c?h=v6.5#n3658

3. Security critical: Binder is a critical part of Android's sandboxing
   strategy. Even Android's most de-privileged sandboxes (e.g. the
   Chrome renderer, or SW Codec) have direct access to Binder. More than
   just about any other component, it's important that Binder provide
   robust security, and itself be robust against security
   vulnerabilities.

It's #1 (high complexity) that has made continuing to evolve Binder and
resolving #2 (tech debt) exceptionally difficult without causing #3
(security issues). For Binder to continue to meet Android's needs, we
need better ways to manage (and reduce!) complexity without increasing
the risk.

The biggest change is obviously the choice of programming language. We
decided to use Rust because it directly addresses a number of the
challenges within Binder that we have faced during the last years. It
prevents mistakes with ref counting, locking, bounds checking, and also
does a lot to reduce the complexity of error handling. Additionally,
we've been able to use the more expressive type system to encode the
ownership semantics of the various structs and pointers, which takes the
complexity of managing object lifetimes out of the hands of the
programmer, reducing the risk of use-after-frees and similar problems.

Rust has many different pointer types that it uses to encode ownership
semantics into the type system, and this is probably one of the most
important aspects of how it helps in Binder. The Binder driver has a lot
of different objects that have complex ownership semantics; some
pointers own a refcount, some pointers have exclusive ownership, and
some pointers just reference the object and it is kept alive in some
other manner. With Rust, we can use a different pointer type for each
kind of pointer, which enables the compiler to enforce that the
ownership semantics are implemented correctly.

Another useful feature is Rust's error handling. Rust allows for more
simplified error handling with features such as destructors, and you get
compilation failures if errors are not properly handled. This means that
even though Rust requires you to spend more lines of code than C on
things such as writing down invariants that are left implicit in C, the
Rust driver is still slightly smaller than C binder: Rust is 5.5kLOC and
C is 5.8kLOC. (These numbers are excluding blank lines, comments,
binderfs, and any debugging facilities in C that are not yet implemented
in the Rust driver. The numbers include abstractions in rust/kernel/
that are unlikely to be used by other drivers than Binder.)

Although this rewrite completely rethinks how the code is structured and
how assumptions are enforced, we do not fundamentally change *how* the
driver does the things it does. A lot of careful thought has gone into
the existing design. The rewrite is aimed rather at improving code
health, structure, readability, robustness, security, maintainability
and extensibility. We also include more inline documentation, and
improve how assumptions in the code are enforced. Furthermore, all
unsafe code is annotated with a SAFETY comment that explains why it is
correct.

We have left the binderfs filesystem component in C. Rewriting it in
Rust would be a large amount of work and requires a lot of bindings to
the file system interfaces. Binderfs has not historically had the same
challenges with security and complexity, so rewriting binderfs seems to
have lower value than the rest of Binder.

Correctness and feature parity
------------------------------

Rust binder passes all tests that validate the correctness of Binder in
the Android Open Source Project. We can boot a device, and run a variety
of apps and functionality without issues. We have performed this both on
the Cuttlefish Android emulator device, and on a Pixel 6 Pro.

As for feature parity, Rust binder currently implements all features
that C binder supports, with the exception of some debugging facilities.
The missing debugging facilities will be added before we submit the Rust
implementation upstream.

Tracepoints
-----------

I did not include all of the tracepoints as I felt that the mechansim
for making C access fields of Rust structs should be discussed on list
separately. I also did not include the support for building Rust Binder
as a module since that requires exporting a bunch of additional symbols
on the C side.

Original RFC Link with old benchmark numbers:
	https://lore.kernel.org/r/20231101-rust-binder-v1-0-08ba9197f637@google.com

Co-developed-by: Wedson Almeida Filho <wedsonaf@gmail.com>
Signed-off-by: Wedson Almeida Filho <wedsonaf@gmail.com>
Co-developed-by: Matt Gilbride <mattgilbride@google.com>
Signed-off-by: Matt Gilbride <mattgilbride@google.com>
Acked-by: Carlos Llamas <cmllamas@google.com>
Acked-by: Paul Moore <paul@paul-moore.com>
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
Link: https://lore.kernel.org/r/20250919-rust-binder-v2-1-a384b09f28dd@google.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

1 files changed, 1131 insertions, 0 deletions

diff --git a/drivers/android/binder/node.rs b/drivers/android/binder/node.rs
new file mode 100644
index 000000000000..ade895ef791e
--- /dev/null
+++ b/drivers/android/binder/node.rs
@@ -0,0 +1,1131 @@
+// SPDX-License-Identifier: GPL-2.0
+
+// Copyright (C) 2025 Google LLC.
+
+use kernel::{
+    list::{AtomicTracker, List, ListArc, ListLinks, TryNewListArc},
+    prelude::*,
+    seq_file::SeqFile,
+    seq_print,
+    sync::lock::{spinlock::SpinLockBackend, Guard},
+    sync::{Arc, LockedBy, SpinLock},
+};
+
+use crate::{
+    defs::*,
+    error::BinderError,
+    process::{NodeRefInfo, Process, ProcessInner},
+    thread::Thread,
+    transaction::Transaction,
+    BinderReturnWriter, DArc, DLArc, DTRWrap, DeliverToRead,
+};
+
+use core::mem;
+
+mod wrapper;
+pub(crate) use self::wrapper::CritIncrWrapper;
+
+#[derive(Debug)]
+pub(crate) struct CouldNotDeliverCriticalIncrement;
+
+/// Keeps track of how this node is scheduled.
+///
+/// There are two ways to schedule a node to a work list. Just schedule the node itself, or
+/// allocate a wrapper that references the node and schedule the wrapper. These wrappers exists to
+/// make it possible to "move" a node from one list to another - when `do_work` is called directly
+/// on the `Node`, then it's a no-op if there's also a pending wrapper.
+///
+/// Wrappers are generally only needed for zero-to-one refcount increments, and there are two cases
+/// of this: weak increments and strong increments. We call such increments "critical" because it
+/// is critical that they are delivered to the thread doing the increment. Some examples:
+///
+/// * One thread makes a zero-to-one strong increment, and another thread makes a zero-to-one weak
+///   increment. Delivering the node to the thread doing the weak increment is wrong, since the
+///   thread doing the strong increment may have ended a long time ago when the command is actually
+///   processed by userspace.
+///
+/// * We have a weak reference and are about to drop it on one thread. But then another thread does
+///   a zero-to-one strong increment. If the strong increment gets sent to the thread that was
+///   about to drop the weak reference, then the strong increment could be processed after the
+///   other thread has already exited, which would be too late.
+///
+/// Note that trying to create a `ListArc` to the node can succeed even if `has_normal_push` is
+/// set. This is because another thread might just have popped the node from a todo list, but not
+/// yet called `do_work`. However, if `has_normal_push` is false, then creating a `ListArc` should
+/// always succeed.
+///
+/// Like the other fields in `NodeInner`, the delivery state is protected by the process lock.
+struct DeliveryState {
+    /// Is the `Node` currently scheduled?
+    has_pushed_node: bool,
+
+    /// Is a wrapper currently scheduled?
+    ///
+    /// The wrapper is used only for strong zero2one increments.
+    has_pushed_wrapper: bool,
+
+    /// Is the currently scheduled `Node` scheduled due to a weak zero2one increment?
+    ///
+    /// Weak zero2one operations are always scheduled using the `Node`.
+    has_weak_zero2one: bool,
+
+    /// Is the currently scheduled wrapper/`Node` scheduled due to a strong zero2one increment?
+    ///
+    /// If `has_pushed_wrapper` is set, then the strong zero2one increment was scheduled using the
+    /// wrapper. Otherwise, `has_pushed_node` must be set and it was scheduled using the `Node`.
+    has_strong_zero2one: bool,
+}
+
+impl DeliveryState {
+    fn should_normal_push(&self) -> bool {
+        !self.has_pushed_node && !self.has_pushed_wrapper
+    }
+
+    fn did_normal_push(&mut self) {
+        assert!(self.should_normal_push());
+        self.has_pushed_node = true;
+    }
+
+    fn should_push_weak_zero2one(&self) -> bool {
+        !self.has_weak_zero2one && !self.has_strong_zero2one
+    }
+
+    fn can_push_weak_zero2one_normally(&self) -> bool {
+        !self.has_pushed_node
+    }
+
+    fn did_push_weak_zero2one(&mut self) {
+        assert!(self.should_push_weak_zero2one());
+        assert!(self.can_push_weak_zero2one_normally());
+        self.has_pushed_node = true;
+        self.has_weak_zero2one = true;
+    }
+
+    fn should_push_strong_zero2one(&self) -> bool {
+        !self.has_strong_zero2one
+    }
+
+    fn can_push_strong_zero2one_normally(&self) -> bool {
+        !self.has_pushed_node
+    }
+
+    fn did_push_strong_zero2one(&mut self) {
+        assert!(self.should_push_strong_zero2one());
+        assert!(self.can_push_strong_zero2one_normally());
+        self.has_pushed_node = true;
+        self.has_strong_zero2one = true;
+    }
+
+    fn did_push_strong_zero2one_wrapper(&mut self) {
+        assert!(self.should_push_strong_zero2one());
+        assert!(!self.can_push_strong_zero2one_normally());
+        self.has_pushed_wrapper = true;
+        self.has_strong_zero2one = true;
+    }
+}
+
+struct CountState {
+    /// The reference count.
+    count: usize,
+    /// Whether the process that owns this node thinks that we hold a refcount on it. (Note that
+    /// even if count is greater than one, we only increment it once in the owning process.)
+    has_count: bool,
+}
+
+impl CountState {
+    fn new() -> Self {
+        Self {
+            count: 0,
+            has_count: false,
+        }
+    }
+}
+
+struct NodeInner {
+    /// Strong refcounts held on this node by `NodeRef` objects.
+    strong: CountState,
+    /// Weak refcounts held on this node by `NodeRef` objects.
+    weak: CountState,
+    delivery_state: DeliveryState,
+    /// The binder driver guarantees that oneway transactions sent to the same node are serialized,
+    /// that is, userspace will not be given the next one until it has finished processing the
+    /// previous oneway transaction. This is done to avoid the case where two oneway transactions
+    /// arrive in opposite order from the order in which they were sent. (E.g., they could be
+    /// delivered to two different threads, which could appear as-if they were sent in opposite
+    /// order.)
+    ///
+    /// To fix that, we store pending oneway transactions in a separate list in the node, and don't
+    /// deliver the next oneway transaction until userspace signals that it has finished processing
+    /// the previous oneway transaction by calling the `BC_FREE_BUFFER` ioctl.
+    oneway_todo: List<DTRWrap<Transaction>>,
+    /// Keeps track of whether this node has a pending oneway transaction.
+    ///
+    /// When this is true, incoming oneway transactions are stored in `oneway_todo`, instead of
+    /// being delivered directly to the process.
+    has_oneway_transaction: bool,
+    /// List of processes to deliver a notification to when this node is destroyed (usually due to
+    /// the process dying).
+    death_list: List<DTRWrap<NodeDeath>, 1>,
+    /// List of processes to deliver freeze notifications to.
+    freeze_list: KVVec<Arc<Process>>,
+    /// The number of active BR_INCREFS or BR_ACQUIRE operations. (should be maximum two)
+    ///
+    /// If this is non-zero, then we postpone any BR_RELEASE or BR_DECREFS notifications until the
+    /// active operations have ended. This avoids the situation an increment and decrement get
+    /// reordered from userspace's perspective.
+    active_inc_refs: u8,
+    /// List of `NodeRefInfo` objects that reference this node.
+    refs: List<NodeRefInfo, { NodeRefInfo::LIST_NODE }>,
+}
+
+#[pin_data]
+pub(crate) struct Node {
+    pub(crate) debug_id: usize,
+    ptr: u64,
+    pub(crate) cookie: u64,
+    pub(crate) flags: u32,
+    pub(crate) owner: Arc<Process>,
+    inner: LockedBy<NodeInner, ProcessInner>,
+    #[pin]
+    links_track: AtomicTracker,
+}
+
+kernel::list::impl_list_arc_safe! {
+    impl ListArcSafe<0> for Node {
+        tracked_by links_track: AtomicTracker;
+    }
+}
+
+// Make `oneway_todo` work.
+kernel::list::impl_list_item! {
+    impl ListItem<0> for DTRWrap<Transaction> {
+        using ListLinks { self.links.inner };
+    }
+}
+
+impl Node {
+    pub(crate) fn new(
+        ptr: u64,
+        cookie: u64,
+        flags: u32,
+        owner: Arc<Process>,
+    ) -> impl PinInit<Self> {
+        pin_init!(Self {
+            inner: LockedBy::new(
+                &owner.inner,
+                NodeInner {
+                    strong: CountState::new(),
+                    weak: CountState::new(),
+                    delivery_state: DeliveryState {
+                        has_pushed_node: false,
+                        has_pushed_wrapper: false,
+                        has_weak_zero2one: false,
+                        has_strong_zero2one: false,
+                    },
+                    death_list: List::new(),
+                    oneway_todo: List::new(),
+                    freeze_list: KVVec::new(),
+                    has_oneway_transaction: false,
+                    active_inc_refs: 0,
+                    refs: List::new(),
+                },
+            ),
+            debug_id: super::next_debug_id(),
+            ptr,
+            cookie,
+            flags,
+            owner,
+            links_track <- AtomicTracker::new(),
+        })
+    }
+
+    pub(crate) fn has_oneway_transaction(&self, owner_inner: &mut ProcessInner) -> bool {
+        let inner = self.inner.access_mut(owner_inner);
+        inner.has_oneway_transaction
+    }
+
+    #[inline(never)]
+    pub(crate) fn full_debug_print(
+        &self,
+        m: &SeqFile,
+        owner_inner: &mut ProcessInner,
+    ) -> Result<()> {
+        let inner = self.inner.access_mut(owner_inner);
+        seq_print!(
+            m,
+            "  node {}: u{:016x} c{:016x} hs {} hw {} cs {} cw {}",
+            self.debug_id,
+            self.ptr,
+            self.cookie,
+            inner.strong.has_count,
+            inner.weak.has_count,
+            inner.strong.count,
+            inner.weak.count,
+        );
+        if !inner.refs.is_empty() {
+            seq_print!(m, " proc");
+            for node_ref in &inner.refs {
+                seq_print!(m, " {}", node_ref.process.task.pid());
+            }
+        }
+        seq_print!(m, "\n");
+        for t in &inner.oneway_todo {
+            t.debug_print_inner(m, "    pending async transaction ");
+        }
+        Ok(())
+    }
+
+    /// Insert the `NodeRef` into this `refs` list.
+    ///
+    /// # Safety
+    ///
+    /// It must be the case that `info.node_ref.node` is this node.
+    pub(crate) unsafe fn insert_node_info(
+        &self,
+        info: ListArc<NodeRefInfo, { NodeRefInfo::LIST_NODE }>,
+    ) {
+        self.inner
+            .access_mut(&mut self.owner.inner.lock())
+            .refs
+            .push_front(info);
+    }
+
+    /// Insert the `NodeRef` into this `refs` list.
+    ///
+    /// # Safety
+    ///
+    /// It must be the case that `info.node_ref.node` is this node.
+    pub(crate) unsafe fn remove_node_info(
+        &self,
+        info: &NodeRefInfo,
+    ) -> Option<ListArc<NodeRefInfo, { NodeRefInfo::LIST_NODE }>> {
+        // SAFETY: We always insert `NodeRefInfo` objects into the `refs` list of the node that it
+        // references in `info.node_ref.node`. That is this node, so `info` cannot possibly be in
+        // the `refs` list of another node.
+        unsafe {
+            self.inner
+                .access_mut(&mut self.owner.inner.lock())
+                .refs
+                .remove(info)
+        }
+    }
+
+    /// An id that is unique across all binder nodes on the system. Used as the key in the
+    /// `by_node` map.
+    pub(crate) fn global_id(&self) -> usize {
+        self as *const Node as usize
+    }
+
+    pub(crate) fn get_id(&self) -> (u64, u64) {
+        (self.ptr, self.cookie)
+    }
+
+    pub(crate) fn add_death(
+        &self,
+        death: ListArc<DTRWrap<NodeDeath>, 1>,
+        guard: &mut Guard<'_, ProcessInner, SpinLockBackend>,
+    ) {
+        self.inner.access_mut(guard).death_list.push_back(death);
+    }
+
+    pub(crate) fn inc_ref_done_locked(
+        self: &DArc<Node>,
+        _strong: bool,
+        owner_inner: &mut ProcessInner,
+    ) -> Option<DLArc<Node>> {
+        let inner = self.inner.access_mut(owner_inner);
+        if inner.active_inc_refs == 0 {
+            pr_err!("inc_ref_done called when no active inc_refs");
+            return None;
+        }
+
+        inner.active_inc_refs -= 1;
+        if inner.active_inc_refs == 0 {
+            // Having active inc_refs can inhibit dropping of ref-counts. Calculate whether we
+            // would send a refcount decrement, and if so, tell the caller to schedule us.
+            let strong = inner.strong.count > 0;
+            let has_strong = inner.strong.has_count;
+            let weak = strong || inner.weak.count > 0;
+            let has_weak = inner.weak.has_count;
+
+            let should_drop_weak = !weak && has_weak;
+            let should_drop_strong = !strong && has_strong;
+
+            // If we want to drop the ref-count again, tell the caller to schedule a work node for
+            // that.
+            let need_push = should_drop_weak || should_drop_strong;
+
+            if need_push && inner.delivery_state.should_normal_push() {
+                let list_arc = ListArc::try_from_arc(self.clone()).ok().unwrap();
+                inner.delivery_state.did_normal_push();
+                Some(list_arc)
+            } else {
+                None
+            }
+        } else {
+            None
+        }
+    }
+
+    pub(crate) fn update_refcount_locked(
+        self: &DArc<Node>,
+        inc: bool,
+        strong: bool,
+        count: usize,
+        owner_inner: &mut ProcessInner,
+    ) -> Option<DLArc<Node>> {
+        let is_dead = owner_inner.is_dead;
+        let inner = self.inner.access_mut(owner_inner);
+
+        // Get a reference to the state we'll update.
+        let state = if strong {
+            &mut inner.strong
+        } else {
+            &mut inner.weak
+        };
+
+        // Update the count and determine whether we need to push work.
+        let need_push = if inc {
+            state.count += count;
+            // TODO: This method shouldn't be used for zero-to-one increments.
+            !is_dead && !state.has_count
+        } else {
+            if state.count < count {
+                pr_err!("Failure: refcount underflow!");
+                return None;
+            }
+            state.count -= count;
+            !is_dead && state.count == 0 && state.has_count
+        };
+
+        if need_push && inner.delivery_state.should_normal_push() {
+            let list_arc = ListArc::try_from_arc(self.clone()).ok().unwrap();
+            inner.delivery_state.did_normal_push();
+            Some(list_arc)
+        } else {
+            None
+        }
+    }
+
+    pub(crate) fn incr_refcount_allow_zero2one(
+        self: &DArc<Self>,
+        strong: bool,
+        owner_inner: &mut ProcessInner,
+    ) -> Result<Option<DLArc<Node>>, CouldNotDeliverCriticalIncrement> {
+        let is_dead = owner_inner.is_dead;
+        let inner = self.inner.access_mut(owner_inner);
+
+        // Get a reference to the state we'll update.
+        let state = if strong {
+            &mut inner.strong
+        } else {
+            &mut inner.weak
+        };
+
+        // Update the count and determine whether we need to push work.
+        state.count += 1;
+        if is_dead || state.has_count {
+            return Ok(None);
+        }
+
+        // Userspace needs to be notified of this.
+        if !strong && inner.delivery_state.should_push_weak_zero2one() {
+            assert!(inner.delivery_state.can_push_weak_zero2one_normally());
+            let list_arc = ListArc::try_from_arc(self.clone()).ok().unwrap();
+            inner.delivery_state.did_push_weak_zero2one();
+            Ok(Some(list_arc))
+        } else if strong && inner.delivery_state.should_push_strong_zero2one() {
+            if inner.delivery_state.can_push_strong_zero2one_normally() {
+                let list_arc = ListArc::try_from_arc(self.clone()).ok().unwrap();
+                inner.delivery_state.did_push_strong_zero2one();
+                Ok(Some(list_arc))
+            } else {
+                state.count -= 1;
+                Err(CouldNotDeliverCriticalIncrement)
+            }
+        } else {
+            // Work is already pushed, and we don't need to push again.
+            Ok(None)
+        }
+    }
+
+    pub(crate) fn incr_refcount_allow_zero2one_with_wrapper(
+        self: &DArc<Self>,
+        strong: bool,
+        wrapper: CritIncrWrapper,
+        owner_inner: &mut ProcessInner,
+    ) -> Option<DLArc<dyn DeliverToRead>> {
+        match self.incr_refcount_allow_zero2one(strong, owner_inner) {
+            Ok(Some(node)) => Some(node as _),
+            Ok(None) => None,
+            Err(CouldNotDeliverCriticalIncrement) => {
+                assert!(strong);
+                let inner = self.inner.access_mut(owner_inner);
+                inner.strong.count += 1;
+                inner.delivery_state.did_push_strong_zero2one_wrapper();
+                Some(wrapper.init(self.clone()))
+            }
+        }
+    }
+
+    pub(crate) fn update_refcount(self: &DArc<Self>, inc: bool, count: usize, strong: bool) {
+        self.owner
+            .inner
+            .lock()
+            .update_node_refcount(self, inc, strong, count, None);
+    }
+
+    pub(crate) fn populate_counts(
+        &self,
+        out: &mut BinderNodeInfoForRef,
+        guard: &Guard<'_, ProcessInner, SpinLockBackend>,
+    ) {
+        let inner = self.inner.access(guard);
+        out.strong_count = inner.strong.count as _;
+        out.weak_count = inner.weak.count as _;
+    }
+
+    pub(crate) fn populate_debug_info(
+        &self,
+        out: &mut BinderNodeDebugInfo,
+        guard: &Guard<'_, ProcessInner, SpinLockBackend>,
+    ) {
+        out.ptr = self.ptr as _;
+        out.cookie = self.cookie as _;
+        let inner = self.inner.access(guard);
+        if inner.strong.has_count {
+            out.has_strong_ref = 1;
+        }
+        if inner.weak.has_count {
+            out.has_weak_ref = 1;
+        }
+    }
+
+    pub(crate) fn force_has_count(&self, guard: &mut Guard<'_, ProcessInner, SpinLockBackend>) {
+        let inner = self.inner.access_mut(guard);
+        inner.strong.has_count = true;
+        inner.weak.has_count = true;
+    }
+
+    fn write(&self, writer: &mut BinderReturnWriter<'_>, code: u32) -> Result {
+        writer.write_code(code)?;
+        writer.write_payload(&self.ptr)?;
+        writer.write_payload(&self.cookie)?;
+        Ok(())
+    }
+
+    pub(crate) fn submit_oneway(
+        &self,
+        transaction: DLArc<Transaction>,
+        guard: &mut Guard<'_, ProcessInner, SpinLockBackend>,
+    ) -> Result<(), (BinderError, DLArc<dyn DeliverToRead>)> {
+        if guard.is_dead {
+            return Err((BinderError::new_dead(), transaction));
+        }
+
+        let inner = self.inner.access_mut(guard);
+        if inner.has_oneway_transaction {
+            inner.oneway_todo.push_back(transaction);
+        } else {
+            inner.has_oneway_transaction = true;
+            guard.push_work(transaction)?;
+        }
+        Ok(())
+    }
+
+    pub(crate) fn release(&self) {
+        let mut guard = self.owner.inner.lock();
+        while let Some(work) = self.inner.access_mut(&mut guard).oneway_todo.pop_front() {
+            drop(guard);
+            work.into_arc().cancel();
+            guard = self.owner.inner.lock();
+        }
+
+        let death_list = core::mem::take(&mut self.inner.access_mut(&mut guard).death_list);
+        drop(guard);
+        for death in death_list {
+            death.into_arc().set_dead();
+        }
+    }
+
+    pub(crate) fn pending_oneway_finished(&self) {
+        let mut guard = self.owner.inner.lock();
+        if guard.is_dead {
+            // Cleanup will happen in `Process::deferred_release`.
+            return;
+        }
+
+        let inner = self.inner.access_mut(&mut guard);
+
+        let transaction = inner.oneway_todo.pop_front();
+        inner.has_oneway_transaction = transaction.is_some();
+        if let Some(transaction) = transaction {
+            match guard.push_work(transaction) {
+                Ok(()) => {}
+                Err((_err, work)) => {
+                    // Process is dead.
+                    // This shouldn't happen due to the `is_dead` check, but if it does, just drop
+                    // the transaction and return.
+                    drop(guard);
+                    drop(work);
+                }
+            }
+        }
+    }
+
+    /// Finds an outdated transaction that the given transaction can replace.
+    ///
+    /// If one is found, it is removed from the list and returned.
+    pub(crate) fn take_outdated_transaction(
+        &self,
+        new: &Transaction,
+        guard: &mut Guard<'_, ProcessInner, SpinLockBackend>,
+    ) -> Option<DLArc<Transaction>> {
+        let inner = self.inner.access_mut(guard);
+        let mut cursor = inner.oneway_todo.cursor_front();
+        while let Some(next) = cursor.peek_next() {
+            if new.can_replace(&next) {
+                return Some(next.remove());
+            }
+            cursor.move_next();
+        }
+        None
+    }
+
+    /// This is split into a separate function since it's called by both `Node::do_work` and
+    /// `NodeWrapper::do_work`.
+    fn do_work_locked(
+        &self,
+        writer: &mut BinderReturnWriter<'_>,
+        mut guard: Guard<'_, ProcessInner, SpinLockBackend>,
+    ) -> Result<bool> {
+        let inner = self.inner.access_mut(&mut guard);
+        let strong = inner.strong.count > 0;
+        let has_strong = inner.strong.has_count;
+        let weak = strong || inner.weak.count > 0;
+        let has_weak = inner.weak.has_count;
+
+        if weak && !has_weak {
+            inner.weak.has_count = true;
+            inner.active_inc_refs += 1;
+        }
+
+        if strong && !has_strong {
+            inner.strong.has_count = true;
+            inner.active_inc_refs += 1;
+        }
+
+        let no_active_inc_refs = inner.active_inc_refs == 0;
+        let should_drop_weak = no_active_inc_refs && (!weak && has_weak);
+        let should_drop_strong = no_active_inc_refs && (!strong && has_strong);
+        if should_drop_weak {
+            inner.weak.has_count = false;
+        }
+        if should_drop_strong {
+            inner.strong.has_count = false;
+        }
+        if no_active_inc_refs && !weak {
+            // Remove the node if there are no references to it.
+            guard.remove_node(self.ptr);
+        }
+        drop(guard);
+
+        if weak && !has_weak {
+            self.write(writer, BR_INCREFS)?;
+        }
+        if strong && !has_strong {
+            self.write(writer, BR_ACQUIRE)?;
+        }
+        if should_drop_strong {
+            self.write(writer, BR_RELEASE)?;
+        }
+        if should_drop_weak {
+            self.write(writer, BR_DECREFS)?;
+        }
+
+        Ok(true)
+    }
+
+    pub(crate) fn add_freeze_listener(
+        &self,
+        process: &Arc<Process>,
+        flags: kernel::alloc::Flags,
+    ) -> Result {
+        let mut vec_alloc = KVVec::<Arc<Process>>::new();
+        loop {
+            let mut guard = self.owner.inner.lock();
+            // Do not check for `guard.dead`. The `dead` flag that matters here is the owner of the
+            // listener, no the target.
+            let inner = self.inner.access_mut(&mut guard);
+            let len = inner.freeze_list.len();
+            if len >= inner.freeze_list.capacity() {
+                if len >= vec_alloc.capacity() {
+                    drop(guard);
+                    vec_alloc = KVVec::with_capacity((1 + len).next_power_of_two(), flags)?;
+                    continue;
+                }
+                mem::swap(&mut inner.freeze_list, &mut vec_alloc);
+                for elem in vec_alloc.drain_all() {
+                    inner.freeze_list.push_within_capacity(elem)?;
+                }
+            }
+            inner.freeze_list.push_within_capacity(process.clone())?;
+            return Ok(());
+        }
+    }
+
+    pub(crate) fn remove_freeze_listener(&self, p: &Arc<Process>) {
+        let _unused_capacity;
+        let mut guard = self.owner.inner.lock();
+        let inner = self.inner.access_mut(&mut guard);
+        let len = inner.freeze_list.len();
+        inner.freeze_list.retain(|proc| !Arc::ptr_eq(proc, p));
+        if len == inner.freeze_list.len() {
+            pr_warn!(
+                "Could not remove freeze listener for {}\n",
+                p.pid_in_current_ns()
+            );
+        }
+        if inner.freeze_list.is_empty() {
+            _unused_capacity = mem::replace(&mut inner.freeze_list, KVVec::new());
+        }
+    }
+
+    pub(crate) fn freeze_list<'a>(&'a self, guard: &'a ProcessInner) -> &'a [Arc<Process>] {
+        &self.inner.access(guard).freeze_list
+    }
+}
+
+impl DeliverToRead for Node {
+    fn do_work(
+        self: DArc<Self>,
+        _thread: &Thread,
+        writer: &mut BinderReturnWriter<'_>,
+    ) -> Result<bool> {
+        let mut owner_inner = self.owner.inner.lock();
+        let inner = self.inner.access_mut(&mut owner_inner);
+
+        assert!(inner.delivery_state.has_pushed_node);
+        if inner.delivery_state.has_pushed_wrapper {
+            // If the wrapper is scheduled, then we are either a normal push or weak zero2one
+            // increment, and the wrapper is a strong zero2one increment, so the wrapper always
+            // takes precedence over us.
+            assert!(inner.delivery_state.has_strong_zero2one);
+            inner.delivery_state.has_pushed_node = false;
+            inner.delivery_state.has_weak_zero2one = false;
+            return Ok(true);
+        }
+
+        inner.delivery_state.has_pushed_node = false;
+        inner.delivery_state.has_weak_zero2one = false;
+        inner.delivery_state.has_strong_zero2one = false;
+
+        self.do_work_locked(writer, owner_inner)
+    }
+
+    fn cancel(self: DArc<Self>) {}
+
+    fn should_sync_wakeup(&self) -> bool {
+        false
+    }
+
+    #[inline(never)]
+    fn debug_print(&self, m: &SeqFile, prefix: &str, _tprefix: &str) -> Result<()> {
+        seq_print!(
+            m,
+            "{}node work {}: u{:016x} c{:016x}\n",
+            prefix,
+            self.debug_id,
+            self.ptr,
+            self.cookie,
+        );
+        Ok(())
+    }
+}
+
+/// Represents something that holds one or more ref-counts to a `Node`.
+///
+/// Whenever process A holds a refcount to a node owned by a different process B, then process A
+/// will store a `NodeRef` that refers to the `Node` in process B. When process A releases the
+/// refcount, we destroy the NodeRef, which decrements the ref-count in process A.
+///
+/// This type is also used for some other cases. For example, a transaction allocation holds a
+/// refcount on the target node, and this is implemented by storing a `NodeRef` in the allocation
+/// so that the destructor of the allocation will drop a refcount of the `Node`.
+pub(crate) struct NodeRef {
+    pub(crate) node: DArc<Node>,
+    /// How many times does this NodeRef hold a refcount on the Node?
+    strong_node_count: usize,
+    weak_node_count: usize,
+    /// How many times does userspace hold a refcount on this NodeRef?
+    strong_count: usize,
+    weak_count: usize,
+}
+
+impl NodeRef {
+    pub(crate) fn new(node: DArc<Node>, strong_count: usize, weak_count: usize) -> Self {
+        Self {
+            node,
+            strong_node_count: strong_count,
+            weak_node_count: weak_count,
+            strong_count,
+            weak_count,
+        }
+    }
+
+    pub(crate) fn absorb(&mut self, mut other: Self) {
+        assert!(
+            Arc::ptr_eq(&self.node, &other.node),
+            "absorb called with differing nodes"
+        );
+        self.strong_node_count += other.strong_node_count;
+        self.weak_node_count += other.weak_node_count;
+        self.strong_count += other.strong_count;
+        self.weak_count += other.weak_count;
+        other.strong_count = 0;
+        other.weak_count = 0;
+        other.strong_node_count = 0;
+        other.weak_node_count = 0;
+
+        if self.strong_node_count >= 2 || self.weak_node_count >= 2 {
+            let mut guard = self.node.owner.inner.lock();
+            let inner = self.node.inner.access_mut(&mut guard);
+
+            if self.strong_node_count >= 2 {
+                inner.strong.count -= self.strong_node_count - 1;
+                self.strong_node_count = 1;
+                assert_ne!(inner.strong.count, 0);
+            }
+            if self.weak_node_count >= 2 {
+                inner.weak.count -= self.weak_node_count - 1;
+                self.weak_node_count = 1;
+                assert_ne!(inner.weak.count, 0);
+            }
+        }
+    }
+
+    pub(crate) fn get_count(&self) -> (usize, usize) {
+        (self.strong_count, self.weak_count)
+    }
+
+    pub(crate) fn clone(&self, strong: bool) -> Result<NodeRef> {
+        if strong && self.strong_count == 0 {
+            return Err(EINVAL);
+        }
+        Ok(self
+            .node
+            .owner
+            .inner
+            .lock()
+            .new_node_ref(self.node.clone(), strong, None))
+    }
+
+    /// Updates (increments or decrements) the number of references held against the node. If the
+    /// count being updated transitions from 0 to 1 or from 1 to 0, the node is notified by having
+    /// its `update_refcount` function called.
+    ///
+    /// Returns whether `self` should be removed (when both counts are zero).
+    pub(crate) fn update(&mut self, inc: bool, strong: bool) -> bool {
+        if strong && self.strong_count == 0 {
+            return false;
+        }
+        let (count, node_count, other_count) = if strong {
+            (
+                &mut self.strong_count,
+                &mut self.strong_node_count,
+                self.weak_count,
+            )
+        } else {
+            (
+                &mut self.weak_count,
+                &mut self.weak_node_count,
+                self.strong_count,
+            )
+        };
+        if inc {
+            if *count == 0 {
+                *node_count = 1;
+                self.node.update_refcount(true, 1, strong);
+            }
+            *count += 1;
+        } else {
+            if *count == 0 {
+                pr_warn!(
+                    "pid {} performed invalid decrement on ref\n",
+                    kernel::current!().pid()
+                );
+                return false;
+            }
+            *count -= 1;
+            if *count == 0 {
+                self.node.update_refcount(false, *node_count, strong);
+                *node_count = 0;
+                return other_count == 0;
+            }
+        }
+        false
+    }
+}
+
+impl Drop for NodeRef {
+    // This destructor is called conditionally from `Allocation::drop`. That branch is often
+    // mispredicted. Inlining this method call reduces the cost of those branch mispredictions.
+    #[inline(always)]
+    fn drop(&mut self) {
+        if self.strong_node_count > 0 {
+            self.node
+                .update_refcount(false, self.strong_node_count, true);
+        }
+        if self.weak_node_count > 0 {
+            self.node
+                .update_refcount(false, self.weak_node_count, false);
+        }
+    }
+}
+
+struct NodeDeathInner {
+    dead: bool,
+    cleared: bool,
+    notification_done: bool,
+    /// Indicates whether the normal flow was interrupted by removing the handle. In this case, we
+    /// need behave as if the death notification didn't exist (i.e., we don't deliver anything to
+    /// the user.
+    aborted: bool,
+}
+
+/// Used to deliver notifications when a process dies.
+///
+/// A process can request to be notified when a process dies using `BC_REQUEST_DEATH_NOTIFICATION`.
+/// This will make the driver send a `BR_DEAD_BINDER` to userspace when the process dies (or
+/// immediately if it is already dead). Userspace is supposed to respond with `BC_DEAD_BINDER_DONE`
+/// once it has processed the notification.
+///
+/// Userspace can unregister from death notifications using the `BC_CLEAR_DEATH_NOTIFICATION`
+/// command. In this case, the kernel will respond with `BR_CLEAR_DEATH_NOTIFICATION_DONE` once the
+/// notification has been removed. Note that if the remote process dies before the kernel has
+/// responded with `BR_CLEAR_DEATH_NOTIFICATION_DONE`, then the kernel will still send a
+/// `BR_DEAD_BINDER`, which userspace must be able to process. In this case, the kernel will wait
+/// for the `BC_DEAD_BINDER_DONE` command before it sends `BR_CLEAR_DEATH_NOTIFICATION_DONE`.
+///
+/// Note that even if the kernel sends a `BR_DEAD_BINDER`, this does not remove the death
+/// notification. Userspace must still remove it manually using `BC_CLEAR_DEATH_NOTIFICATION`.
+///
+/// If a process uses `BC_RELEASE` to destroy its last refcount on a node that has an active death
+/// registration, then the death registration is immediately deleted (we implement this using the
+/// `aborted` field). However, userspace is not supposed to delete a `NodeRef` without first
+/// deregistering death notifications, so this codepath is not executed under normal circumstances.
+#[pin_data]
+pub(crate) struct NodeDeath {
+    node: DArc<Node>,
+    process: Arc<Process>,
+    pub(crate) cookie: u64,
+    #[pin]
+    links_track: AtomicTracker<0>,
+    /// Used by the owner `Node` to store a list of registered death notifications.
+    ///
+    /// # Invariants
+    ///
+    /// Only ever used with the `death_list` list of `self.node`.
+    #[pin]
+    death_links: ListLinks<1>,
+    /// Used by the process to keep track of the death notifications for which we have sent a
+    /// `BR_DEAD_BINDER` but not yet received a `BC_DEAD_BINDER_DONE`.
+    ///
+    /// # Invariants
+    ///
+    /// Only ever used with the `delivered_deaths` list of `self.process`.
+    #[pin]
+    delivered_links: ListLinks<2>,
+    #[pin]
+    delivered_links_track: AtomicTracker<2>,
+    #[pin]
+    inner: SpinLock<NodeDeathInner>,
+}
+
+impl NodeDeath {
+    /// Constructs a new node death notification object.
+    pub(crate) fn new(
+        node: DArc<Node>,
+        process: Arc<Process>,
+        cookie: u64,
+    ) -> impl PinInit<DTRWrap<Self>> {
+        DTRWrap::new(pin_init!(
+            Self {
+                node,
+                process,
+                cookie,
+                links_track <- AtomicTracker::new(),
+                death_links <- ListLinks::new(),
+                delivered_links <- ListLinks::new(),
+                delivered_links_track <- AtomicTracker::new(),
+                inner <- kernel::new_spinlock!(NodeDeathInner {
+                    dead: false,
+                    cleared: false,
+                    notification_done: false,
+                    aborted: false,
+                }, "NodeDeath::inner"),
+            }
+        ))
+    }
+
+    /// Sets the cleared flag to `true`.
+    ///
+    /// It removes `self` from the node's death notification list if needed.
+    ///
+    /// Returns whether it needs to be queued.
+    pub(crate) fn set_cleared(self: &DArc<Self>, abort: bool) -> bool {
+        let (needs_removal, needs_queueing) = {
+            // Update state and determine if we need to queue a work item. We only need to do it
+            // when the node is not dead or if the user already completed the death notification.
+            let mut inner = self.inner.lock();
+            if abort {
+                inner.aborted = true;
+            }
+            if inner.cleared {
+                // Already cleared.
+                return false;
+            }
+            inner.cleared = true;
+            (!inner.dead, !inner.dead || inner.notification_done)
+        };
+
+        // Remove death notification from node.
+        if needs_removal {
+            let mut owner_inner = self.node.owner.inner.lock();
+            let node_inner = self.node.inner.access_mut(&mut owner_inner);
+            // SAFETY: A `NodeDeath` is never inserted into the death list of any node other than
+            // its owner, so it is either in this death list or in no death list.
+            unsafe { node_inner.death_list.remove(self) };
+        }
+        needs_queueing
+    }
+
+    /// Sets the 'notification done' flag to `true`.
+    pub(crate) fn set_notification_done(self: DArc<Self>, thread: &Thread) {
+        let needs_queueing = {
+            let mut inner = self.inner.lock();
+            inner.notification_done = true;
+            inner.cleared
+        };
+        if needs_queueing {
+            if let Some(death) = ListArc::try_from_arc_or_drop(self) {
+                let _ = thread.push_work_if_looper(death);
+            }
+        }
+    }
+
+    /// Sets the 'dead' flag to `true` and queues work item if needed.
+    pub(crate) fn set_dead(self: DArc<Self>) {
+        let needs_queueing = {
+            let mut inner = self.inner.lock();
+            if inner.cleared {
+                false
+            } else {
+                inner.dead = true;
+                true
+            }
+        };
+        if needs_queueing {
+            // Push the death notification to the target process. There is nothing else to do if
+            // it's already dead.
+            if let Some(death) = ListArc::try_from_arc_or_drop(self) {
+                let process = death.process.clone();
+                let _ = process.push_work(death);
+            }
+        }
+    }
+}
+
+kernel::list::impl_list_arc_safe! {
+    impl ListArcSafe<0> for NodeDeath {
+        tracked_by links_track: AtomicTracker;
+    }
+}
+
+kernel::list::impl_list_arc_safe! {
+    impl ListArcSafe<1> for DTRWrap<NodeDeath> { untracked; }
+}
+kernel::list::impl_list_item! {
+    impl ListItem<1> for DTRWrap<NodeDeath> {
+        using ListLinks { self.wrapped.death_links };
+    }
+}
+
+kernel::list::impl_list_arc_safe! {
+    impl ListArcSafe<2> for DTRWrap<NodeDeath> {
+        tracked_by wrapped: NodeDeath;
+    }
+}
+kernel::list::impl_list_arc_safe! {
+    impl ListArcSafe<2> for NodeDeath {
+        tracked_by delivered_links_track: AtomicTracker<2>;
+    }
+}
+kernel::list::impl_list_item! {
+    impl ListItem<2> for DTRWrap<NodeDeath> {
+        using ListLinks { self.wrapped.delivered_links };
+    }
+}
+
+impl DeliverToRead for NodeDeath {
+    fn do_work(
+        self: DArc<Self>,
+        _thread: &Thread,
+        writer: &mut BinderReturnWriter<'_>,
+    ) -> Result<bool> {
+        let done = {
+            let inner = self.inner.lock();
+            if inner.aborted {
+                return Ok(true);
+            }
+            inner.cleared && (!inner.dead || inner.notification_done)
+        };
+
+        let cookie = self.cookie;
+        let cmd = if done {
+            BR_CLEAR_DEATH_NOTIFICATION_DONE
+        } else {
+            let process = self.process.clone();
+            let mut process_inner = process.inner.lock();
+            let inner = self.inner.lock();
+            if inner.aborted {
+                return Ok(true);
+            }
+            // We're still holding the inner lock, so it cannot be aborted while we insert it into
+            // the delivered list.
+            process_inner.death_delivered(self.clone());
+            BR_DEAD_BINDER
+        };
+
+        writer.write_code(cmd)?;
+        writer.write_payload(&cookie)?;
+        // DEAD_BINDER notifications can cause transactions, so stop processing work items when we
+        // get to a death notification.
+        Ok(cmd != BR_DEAD_BINDER)
+    }
+
+    fn cancel(self: DArc<Self>) {}
+
+    fn should_sync_wakeup(&self) -> bool {
+        false
+    }
+
+    #[inline(never)]
+    fn debug_print(&self, m: &SeqFile, prefix: &str, _tprefix: &str) -> Result<()> {
+        let inner = self.inner.lock();
+
+        let dead_binder = inner.dead && !inner.notification_done;
+
+        if dead_binder {
+            if inner.cleared {
+                seq_print!(m, "{}has cleared dead binder\n", prefix);
+            } else {
+                seq_print!(m, "{}has dead binder\n", prefix);
+            }
+        } else {
+            seq_print!(m, "{}has cleared death notification\n", prefix);
+        }
+
+        Ok(())
+    }
+}