x86/kprobes: Fix crash when probing CS CALL instructionsgrafted

When using eBPF to probe CS CALL instructions within a function,
a crash can be triggered.

The eBPF tool probes offset 257 of the __hrtimer_run_queues()
function:

<__hrtimer_run_queues+249>:  nopl   0x0(%rax,%rax,1)
<__hrtimer_run_queues+254>:  mov    %r14,%rdi
<__hrtimer_run_queues+257>:  cs call <__x86_indirect_thunk_r12>
<__hrtimer_run_queues+263>:  mov    %eax,%r12d
<__hrtimer_run_queues+266>:  xchg   %ax,%ax
<__hrtimer_run_queues+268>:  mov    %r13,%rdi

Which triggers this crash:

  BUG: unable to handle page fault for address: 00000000000f41c9
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP NOPTI
  CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: P
  RIP: 0010:__hrtimer_run_queues+0x106/0x230

Note that __hrtimer_run_queues+0x106 is __hrtimer_run_queues+262, which is
at the 6th byte of the above CS CALL instruction. Since the CS CALL
instruction occupies 6 bytes, the exception occurred in the middle of that
call instruction.

The root cause is that when using eBPF tools to probe in the middle of a
function, a kprobe with INT3 is used as the underlying implementation.

During single-step emulation of the original CALL instruction,
int3_emulate_call() assumes that the probed CALL instruction is 5 bytes
long. However, the actual CS-prefixed CALL instruction occupies 6 bytes,
so it constructs an incorrect exception return address. When the CPU
returns from the kprobe handler, the next instruction to be executed is at
the address of the last byte of that CS CALL instruction. Coincidentally,
starting from that address, the CPU fetches and decodes a completely
different instruction, which ultimately triggers a kernel crash.

Fix the issue by using the actual instruction length obtained from
the instruction decoder when constructing the exception return
address, rather than relying on the hardcoded CALL_INSN_SIZE macro.

[ mingo: Refined the changelog ]

Fixes: 6256e668b7af ("x86/kprobes: Use int3 instead of debug trap for single-step")
Suggested-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Jinke Han <jinkehan@didiglobal.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Acked-by: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://patch.msgid.link/20260908073742.GA10517@didi-ThinkCentre-M920t-N000

1 files changed, 612 insertions, 0 deletions

diff --git a/drivers/android/binder/allocation.rs b/drivers/android/binder/allocation.rs
new file mode 100644
index 000000000..165cb797e
--- /dev/null
+++ b/drivers/android/binder/allocation.rs
@@ -0,0 +1,612 @@
+// SPDX-License-Identifier: GPL-2.0
+
+// Copyright (C) 2025 Google LLC.
+
+use core::mem::{size_of, size_of_val, MaybeUninit};
+use core::ops::Range;
+
+use kernel::{
+    bindings,
+    fs::file::{File, FileDescriptorReservation},
+    prelude::*,
+    sync::{aref::ARef, Arc},
+    transmute::{AsBytes, FromBytes},
+    uaccess::UserSliceReader,
+    uapi,
+};
+
+use crate::{
+    deferred_close::DeferredFdCloser,
+    defs::*,
+    node::{Node, NodeRef},
+    process::Process,
+    DArc,
+};
+
+#[derive(Default)]
+pub(crate) struct AllocationInfo {
+    /// Range within the allocation where we can find the offsets to the object descriptors.
+    pub(crate) offsets: Option<Range<usize>>,
+    /// The target node of the transaction this allocation is associated to.
+    /// Not set for replies.
+    pub(crate) target_node: Option<NodeRef>,
+    /// When this allocation is dropped, call `pending_oneway_finished` on the node.
+    ///
+    /// This is used to serialize oneway transaction on the same node. Binder guarantees that
+    /// oneway transactions to the same node are delivered sequentially in the order they are sent.
+    pub(crate) oneway_node: Option<DArc<Node>>,
+    /// Zero the data in the buffer on free.
+    pub(crate) clear_on_free: bool,
+    /// List of files embedded in this transaction.
+    file_list: FileList,
+}
+
+/// Represents an allocation that the kernel is currently using.
+///
+/// When allocations are idle, the range allocator holds the data related to them.
+///
+/// # Invariants
+///
+/// This allocation corresponds to an allocation in the range allocator, so the relevant pages are
+/// marked in use in the page range.
+pub(crate) struct Allocation {
+    pub(crate) offset: usize,
+    size: usize,
+    pub(crate) ptr: usize,
+    pub(crate) process: Arc<Process>,
+    allocation_info: Option<AllocationInfo>,
+    free_on_drop: bool,
+    #[allow(dead_code)]
+    pub(crate) debug_id: usize,
+}
+
+impl Allocation {
+    pub(crate) fn new(
+        process: Arc<Process>,
+        debug_id: usize,
+        offset: usize,
+        size: usize,
+        ptr: usize,
+    ) -> Self {
+        Self {
+            process,
+            offset,
+            size,
+            ptr,
+            debug_id,
+            allocation_info: None,
+            free_on_drop: true,
+        }
+    }
+
+    fn size_check(&self, offset: usize, size: usize) -> Result {
+        let overflow_fail = offset.checked_add(size).is_none();
+        let cmp_size_fail = offset.wrapping_add(size) > self.size;
+        if overflow_fail || cmp_size_fail {
+            return Err(EFAULT);
+        }
+        Ok(())
+    }
+
+    pub(crate) fn copy_into(
+        &self,
+        reader: &mut UserSliceReader,
+        offset: usize,
+        size: usize,
+    ) -> Result {
+        self.size_check(offset, size)?;
+
+        // SAFETY: While this object exists, the range allocator will keep the range allocated, and
+        // in turn, the pages will be marked as in use.
+        unsafe {
+            self.process
+                .pages
+                .copy_from_user_slice(reader, self.offset + offset, size)
+        }
+    }
+
+    pub(crate) fn read<T: FromBytes>(&self, offset: usize) -> Result<T> {
+        self.size_check(offset, size_of::<T>())?;
+
+        // SAFETY: While this object exists, the range allocator will keep the range allocated, and
+        // in turn, the pages will be marked as in use.
+        unsafe { self.process.pages.read(self.offset + offset) }
+    }
+
+    pub(crate) fn write<T: ?Sized>(&self, offset: usize, obj: &T) -> Result {
+        self.size_check(offset, size_of_val::<T>(obj))?;
+
+        // SAFETY: While this object exists, the range allocator will keep the range allocated, and
+        // in turn, the pages will be marked as in use.
+        unsafe { self.process.pages.write(self.offset + offset, obj) }
+    }
+
+    pub(crate) fn fill_zero(&self) -> Result {
+        // SAFETY: While this object exists, the range allocator will keep the range allocated, and
+        // in turn, the pages will be marked as in use.
+        unsafe { self.process.pages.fill_zero(self.offset, self.size) }
+    }
+
+    pub(crate) fn keep_alive(mut self) {
+        self.process
+            .buffer_make_freeable(self.offset, self.allocation_info.take());
+        self.free_on_drop = false;
+    }
+
+    pub(crate) fn set_info(&mut self, info: AllocationInfo) {
+        self.allocation_info = Some(info);
+    }
+
+    pub(crate) fn get_or_init_info(&mut self) -> &mut AllocationInfo {
+        self.allocation_info.get_or_insert_with(Default::default)
+    }
+
+    pub(crate) fn set_info_offsets(&mut self, offsets: Range<usize>) {
+        self.get_or_init_info().offsets = Some(offsets);
+    }
+
+    pub(crate) fn set_info_oneway_node(&mut self, oneway_node: DArc<Node>) {
+        self.get_or_init_info().oneway_node = Some(oneway_node);
+    }
+
+    pub(crate) fn set_info_clear_on_drop(&mut self) {
+        self.get_or_init_info().clear_on_free = true;
+    }
+
+    pub(crate) fn set_info_target_node(&mut self, target_node: NodeRef) {
+        self.get_or_init_info().target_node = Some(target_node);
+    }
+
+    pub(crate) fn take_oneway_node(&mut self) -> Option<DArc<Node>> {
+        if let Some(info) = self.allocation_info.as_mut() {
+            info.oneway_node.take()
+        } else {
+            None
+        }
+    }
+
+    /// Reserve enough space to push at least `num_fds` fds.
+    pub(crate) fn info_add_fd_reserve(&mut self, num_fds: usize) -> Result {
+        self.get_or_init_info()
+            .file_list
+            .files_to_translate
+            .reserve(num_fds, GFP_KERNEL)?;
+
+        Ok(())
+    }
+
+    pub(crate) fn info_add_fd(
+        &mut self,
+        file: ARef<File>,
+        buffer_offset: usize,
+        close_on_free: bool,
+    ) -> Result {
+        self.get_or_init_info().file_list.files_to_translate.push(
+            FileEntry {
+                file,
+                buffer_offset,
+                close_on_free,
+            },
+            GFP_KERNEL,
+        )?;
+
+        Ok(())
+    }
+
+    pub(crate) fn set_info_close_on_free(&mut self, cof: FdsCloseOnFree) {
+        self.get_or_init_info().file_list.close_on_free = cof.0;
+    }
+
+    pub(crate) fn translate_fds(&mut self) -> Result<TranslatedFds> {
+        let file_list = match self.allocation_info.as_mut() {
+            Some(info) => &mut info.file_list,
+            None => return Ok(TranslatedFds::new()),
+        };
+
+        let files = core::mem::take(&mut file_list.files_to_translate);
+
+        let num_close_on_free = files.iter().filter(|entry| entry.close_on_free).count();
+        let mut close_on_free = KVec::with_capacity(num_close_on_free, GFP_KERNEL)?;
+
+        let mut reservations = KVec::with_capacity(files.len(), GFP_KERNEL)?;
+        for file_info in files {
+            let res = FileDescriptorReservation::get_unused_fd_flags(bindings::O_CLOEXEC)?;
+            let fd = res.reserved_fd();
+            self.write::<u32>(file_info.buffer_offset, &fd)?;
+            crate::trace::trace_transaction_fd_recv(self.debug_id, fd, file_info.buffer_offset);
+
+            reservations.push(
+                Reservation {
+                    res,
+                    file: file_info.file,
+                },
+                GFP_KERNEL,
+            )?;
+            if file_info.close_on_free {
+                close_on_free.push(fd, GFP_KERNEL)?;
+            }
+        }
+
+        Ok(TranslatedFds {
+            reservations,
+            close_on_free: FdsCloseOnFree(close_on_free),
+        })
+    }
+
+    /// Should the looper return to userspace when freeing this allocation?
+    pub(crate) fn looper_need_return_on_free(&self) -> bool {
+        // Closing fds involves pushing task_work for execution when we return to userspace. Hence,
+        // we should return to userspace asap if we are closing fds.
+        match self.allocation_info {
+            Some(ref info) => !info.file_list.close_on_free.is_empty(),
+            None => false,
+        }
+    }
+}
+
+impl Drop for Allocation {
+    fn drop(&mut self) {
+        if !self.free_on_drop {
+            return;
+        }
+
+        if let Some(mut info) = self.allocation_info.take() {
+            if let Some(oneway_node) = info.oneway_node.as_ref() {
+                oneway_node.pending_oneway_finished();
+            }
+
+            info.target_node = None;
+
+            if let Some(offsets) = info.offsets.clone() {
+                let view = AllocationView::new(self, offsets.start);
+                for i in offsets.step_by(size_of::<u64>()) {
+                    if view.cleanup_object(i).is_err() {
+                        pr_warn!("Error cleaning up object at offset {}\n", i)
+                    }
+                }
+            }
+
+            if self.process.task == kernel::current!().group_leader() {
+                for &fd in &info.file_list.close_on_free {
+                    let closer = match DeferredFdCloser::new(GFP_KERNEL) {
+                        Ok(closer) => closer,
+                        Err(kernel::alloc::AllocError) => {
+                            // Ignore allocation failures.
+                            break;
+                        }
+                    };
+
+                    // Here, we ignore errors. The operation can fail if the fd is not valid, or if
+                    // the method is called from a kthread. However, this is always called from a
+                    // syscall, so the latter case cannot happen, and we don't care about the first
+                    // case.
+                    let _ = closer.close_fd(fd);
+                }
+            }
+
+            if info.clear_on_free {
+                if let Err(e) = self.fill_zero() {
+                    pr_warn!("Failed to clear data on free: {:?}", e);
+                }
+            }
+        }
+
+        self.process.buffer_raw_free(self.ptr);
+    }
+}
+
+/// A wrapper around `Allocation` that is being created.
+///
+/// If the allocation is destroyed while wrapped in this wrapper, then the allocation will be
+/// considered to be part of a failed transaction. Successful transactions avoid that by calling
+/// `success`, which skips the destructor.
+#[repr(transparent)]
+pub(crate) struct NewAllocation(pub(crate) Allocation);
+
+impl NewAllocation {
+    pub(crate) fn success(self) -> Allocation {
+        // This skips the destructor.
+        //
+        // SAFETY: This type is `#[repr(transparent)]`, so the layout matches.
+        unsafe { core::mem::transmute(self) }
+    }
+}
+
+impl core::ops::Deref for NewAllocation {
+    type Target = Allocation;
+    fn deref(&self) -> &Allocation {
+        &self.0
+    }
+}
+
+impl core::ops::DerefMut for NewAllocation {
+    fn deref_mut(&mut self) -> &mut Allocation {
+        &mut self.0
+    }
+}
+
+/// A view into the beginning of an allocation.
+///
+/// All attempts to read or write outside of the view will fail. To intentionally access outside of
+/// this view, use the `alloc` field of this struct directly.
+pub(crate) struct AllocationView<'a> {
+    pub(crate) alloc: &'a mut Allocation,
+    limit: usize,
+}
+
+impl<'a> AllocationView<'a> {
+    pub(crate) fn new(alloc: &'a mut Allocation, limit: usize) -> Self {
+        AllocationView { alloc, limit }
+    }
+
+    pub(crate) fn read<T: FromBytes>(&self, offset: usize) -> Result<T> {
+        if offset.checked_add(size_of::<T>()).ok_or(EINVAL)? > self.limit {
+            return Err(EINVAL);
+        }
+        self.alloc.read(offset)
+    }
+
+    pub(crate) fn write<T: AsBytes>(&self, offset: usize, obj: &T) -> Result {
+        if offset.checked_add(size_of::<T>()).ok_or(EINVAL)? > self.limit {
+            return Err(EINVAL);
+        }
+        self.alloc.write(offset, obj)
+    }
+
+    pub(crate) fn copy_into(
+        &self,
+        reader: &mut UserSliceReader,
+        offset: usize,
+        size: usize,
+    ) -> Result {
+        if offset.checked_add(size).ok_or(EINVAL)? > self.limit {
+            return Err(EINVAL);
+        }
+        self.alloc.copy_into(reader, offset, size)
+    }
+
+    pub(crate) fn transfer_binder_object(
+        &self,
+        offset: usize,
+        obj: &uapi::flat_binder_object,
+        strong: bool,
+        node_ref: NodeRef,
+    ) -> Result {
+        let mut newobj = FlatBinderObject::default();
+        let node = node_ref.node.clone();
+        if Arc::ptr_eq(&node_ref.node.owner, &self.alloc.process) {
+            // The receiving process is the owner of the node, so send it a binder object (instead
+            // of a handle).
+            let (ptr, cookie) = node.get_id();
+            newobj.hdr.type_ = if strong {
+                BINDER_TYPE_BINDER
+            } else {
+                BINDER_TYPE_WEAK_BINDER
+            };
+            newobj.flags = obj.flags;
+            newobj.__bindgen_anon_1.binder = ptr as uapi::binder_uintptr_t;
+            newobj.cookie = cookie as uapi::binder_uintptr_t;
+            self.write(offset, &newobj)?;
+            // Increment the user ref count on the node. It will be decremented as part of the
+            // destruction of the buffer, when we see a binder or weak-binder object.
+            node.update_refcount(true, 1, strong);
+        } else {
+            // The receiving process is different from the owner, so we need to insert a handle to
+            // the binder object.
+            let handle = self
+                .alloc
+                .process
+                .as_arc_borrow()
+                .insert_or_update_handle(node_ref, false)?;
+            newobj.hdr.type_ = if strong {
+                BINDER_TYPE_HANDLE
+            } else {
+                BINDER_TYPE_WEAK_HANDLE
+            };
+            newobj.flags = obj.flags;
+            newobj.__bindgen_anon_1.handle = handle;
+            if self.write(offset, &newobj).is_err() {
+                // Decrement ref count on the handle we just created.
+                let _ = self
+                    .alloc
+                    .process
+                    .as_arc_borrow()
+                    .update_ref(handle, false, strong);
+                return Err(EINVAL);
+            }
+        }
+
+        Ok(())
+    }
+
+    fn cleanup_object(&self, index_offset: usize) -> Result {
+        let offset = self.alloc.read::<u64>(index_offset)?;
+        let offset: usize = offset.try_into().map_err(|_| EINVAL)?;
+        let header = self.read::<BinderObjectHeader>(offset)?;
+        match header.type_ {
+            BINDER_TYPE_WEAK_BINDER | BINDER_TYPE_BINDER => {
+                let obj = self.read::<FlatBinderObject>(offset)?;
+                let strong = header.type_ == BINDER_TYPE_BINDER;
+                // SAFETY: The type is `BINDER_TYPE_{WEAK_}BINDER`, so the `binder` field is
+                // populated.
+                let ptr = unsafe { obj.__bindgen_anon_1.binder };
+                let cookie = obj.cookie;
+                self.alloc.process.update_node(ptr, cookie, strong);
+                Ok(())
+            }
+            BINDER_TYPE_WEAK_HANDLE | BINDER_TYPE_HANDLE => {
+                let obj = self.read::<FlatBinderObject>(offset)?;
+                let strong = header.type_ == BINDER_TYPE_HANDLE;
+                // SAFETY: The type is `BINDER_TYPE_{WEAK_}HANDLE`, so the `handle` field is
+                // populated.
+                let handle = unsafe { obj.__bindgen_anon_1.handle };
+                self.alloc
+                    .process
+                    .as_arc_borrow()
+                    .update_ref(handle, false, strong)
+            }
+            _ => Ok(()),
+        }
+    }
+}
+
+/// A binder object as it is serialized.
+///
+/// # Invariants
+///
+/// All bytes must be initialized, and the value of `self.hdr.type_` must be one of the allowed
+/// types.
+#[repr(C)]
+pub(crate) union BinderObject {
+    hdr: uapi::binder_object_header,
+    fbo: uapi::flat_binder_object,
+    fdo: uapi::binder_fd_object,
+    bbo: uapi::binder_buffer_object,
+    fdao: uapi::binder_fd_array_object,
+}
+
+/// A view into a `BinderObject` that can be used in a match statement.
+pub(crate) enum BinderObjectRef<'a> {
+    Binder(&'a mut uapi::flat_binder_object),
+    Handle(&'a mut uapi::flat_binder_object),
+    Fd(&'a mut uapi::binder_fd_object),
+    Ptr(&'a mut uapi::binder_buffer_object),
+    Fda(&'a mut uapi::binder_fd_array_object),
+}
+
+impl BinderObject {
+    pub(crate) fn read_from(reader: &mut UserSliceReader) -> Result<BinderObject> {
+        let object = Self::read_from_inner(|slice| {
+            let read_len = usize::min(slice.len(), reader.len());
+            reader.clone_reader().read_slice(&mut slice[..read_len])?;
+            Ok(())
+        })?;
+
+        // If we used a object type smaller than the largest object size, then we've read more
+        // bytes than we needed to. However, we used `.clone_reader()` to avoid advancing the
+        // original reader. Now, we call `skip` so that the caller's reader is advanced by the
+        // right amount.
+        //
+        // The `skip` call fails if the reader doesn't have `size` bytes available. This could
+        // happen if the type header corresponds to an object type that is larger than the rest of
+        // the reader.
+        //
+        // Any extra bytes beyond the size of the object are inaccessible after this call, so
+        // reading them again from the `reader` later does not result in TOCTOU bugs.
+        reader.skip(object.size())?;
+
+        Ok(object)
+    }
+
+    /// Use the provided reader closure to construct a `BinderObject`.
+    ///
+    /// The closure should write the bytes for the object into the provided slice.
+    pub(crate) fn read_from_inner<R>(reader: R) -> Result<BinderObject>
+    where
+        R: FnOnce(&mut [u8; size_of::<BinderObject>()]) -> Result<()>,
+    {
+        let mut obj = MaybeUninit::<BinderObject>::zeroed();
+
+        // SAFETY: The lengths of `BinderObject` and `[u8; size_of::<BinderObject>()]` are equal,
+        // and the byte array has an alignment requirement of one, so the pointer cast is okay.
+        // Additionally, `obj` was initialized to zeros, so the byte array will not be
+        // uninitialized.
+        (reader)(unsafe { &mut *obj.as_mut_ptr().cast() })?;
+
+        // SAFETY: The entire object is initialized, so accessing this field is safe.
+        let type_ = unsafe { obj.assume_init_ref().hdr.type_ };
+        if Self::type_to_size(type_).is_none() {
+            // The value of `obj.hdr_type_` was invalid.
+            return Err(EINVAL);
+        }
+
+        // SAFETY: All bytes are initialized (since we zeroed them at the start) and we checked
+        // that `self.hdr.type_` is one of the allowed types, so the type invariants are satisfied.
+        unsafe { Ok(obj.assume_init()) }
+    }
+
+    pub(crate) fn as_ref(&mut self) -> BinderObjectRef<'_> {
+        use BinderObjectRef::*;
+        // SAFETY: The constructor ensures that all bytes of `self` are initialized, and all
+        // variants of this union accept all initialized bit patterns.
+        unsafe {
+            match self.hdr.type_ {
+                BINDER_TYPE_WEAK_BINDER | BINDER_TYPE_BINDER => Binder(&mut self.fbo),
+                BINDER_TYPE_WEAK_HANDLE | BINDER_TYPE_HANDLE => Handle(&mut self.fbo),
+                BINDER_TYPE_FD => Fd(&mut self.fdo),
+                BINDER_TYPE_PTR => Ptr(&mut self.bbo),
+                BINDER_TYPE_FDA => Fda(&mut self.fdao),
+                // SAFETY: By the type invariant, the value of `self.hdr.type_` cannot have any
+                // other value than the ones checked above.
+                _ => core::hint::unreachable_unchecked(),
+            }
+        }
+    }
+
+    pub(crate) fn size(&self) -> usize {
+        // SAFETY: The entire object is initialized, so accessing this field is safe.
+        let type_ = unsafe { self.hdr.type_ };
+
+        // SAFETY: The type invariants guarantee that the type field is correct.
+        unsafe { Self::type_to_size(type_).unwrap_unchecked() }
+    }
+
+    fn type_to_size(type_: u32) -> Option<usize> {
+        match type_ {
+            BINDER_TYPE_WEAK_BINDER => Some(size_of::<uapi::flat_binder_object>()),
+            BINDER_TYPE_BINDER => Some(size_of::<uapi::flat_binder_object>()),
+            BINDER_TYPE_WEAK_HANDLE => Some(size_of::<uapi::flat_binder_object>()),
+            BINDER_TYPE_HANDLE => Some(size_of::<uapi::flat_binder_object>()),
+            BINDER_TYPE_FD => Some(size_of::<uapi::binder_fd_object>()),
+            BINDER_TYPE_PTR => Some(size_of::<uapi::binder_buffer_object>()),
+            BINDER_TYPE_FDA => Some(size_of::<uapi::binder_fd_array_object>()),
+            _ => None,
+        }
+    }
+}
+
+#[derive(Default)]
+struct FileList {
+    files_to_translate: KVec<FileEntry>,
+    close_on_free: KVec<u32>,
+}
+
+struct FileEntry {
+    /// The file for which a descriptor will be created in the recipient process.
+    file: ARef<File>,
+    /// The offset in the buffer where the file descriptor is stored.
+    buffer_offset: usize,
+    /// Whether this fd should be closed when the allocation is freed.
+    close_on_free: bool,
+}
+
+pub(crate) struct TranslatedFds {
+    reservations: KVec<Reservation>,
+    /// If commit is called, then these fds should be closed. (If commit is not called, then they
+    /// shouldn't be closed.)
+    close_on_free: FdsCloseOnFree,
+}
+
+struct Reservation {
+    res: FileDescriptorReservation,
+    file: ARef<File>,
+}
+
+impl TranslatedFds {
+    pub(crate) fn new() -> Self {
+        Self {
+            reservations: KVec::new(),
+            close_on_free: FdsCloseOnFree(KVec::new()),
+        }
+    }
+
+    pub(crate) fn commit(self) -> FdsCloseOnFree {
+        for entry in self.reservations {
+            entry.res.fd_install(entry.file);
+        }
+
+        self.close_on_free
+    }
+}
+
+pub(crate) struct FdsCloseOnFree(KVec<u32>);