Skip to contentSkip to conversation

rust_binder: check current before closing fds

PATCH Superseded Alice Ryhl 2 participants 10 messages Feb 27

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Alice Ryhl

author maintainer
#0 lore raw

v2 cover rust_binder: check current before closing fds

Please see the last patch in this series for details.

---
Changes in v2:
- Note that ARef<U> == U can't be implemented due to impl overlap.
- Add #[inline] in == for Task.
- Add a patch that makes use of the new == impl for Task.
- Link to v1: https://lore.kernel.org/r/[email protected]

---
Alice Ryhl (4):

      rust: sync: implement == operator for ARef
      rust: task: implement == operator for Task
      rust_binder: make use of == for Task
      rust_binder: check current before closing fds
 drivers/android/binder/allocation.rs |  4 ++++
 drivers/android/binder/process.rs    |  4 ++--
 rust/kernel/sync/aref.rs             | 22 ++++++++++++++++++++++
 rust/kernel/task.rs                  |  9 +++++++++
 4 files changed, 37 insertions(+), 2 deletions(-)

---
base-commit: 6de23f81a5e08be8fbf5e8d7e9febc72a5b5f27f
change-id: 20260219-close-fd-check-current-98e3ee59880a

Best regards,

Signature · 1 line
Alice Ryhl <[email protected]>
author maintainer
#1 lore raw

PATCH v2 1/4 rust: sync: implement == operator for ARef

Rust Binder wants to perform a comparison between ARef<Task> and &Task, so define the == operator for ARef<_> when compared with another ARef<_> or just a reference. The operator is implemented in terms of the same operator applied to the inner type.

Note that PartialEq<U> cannot be implemented because it would overlap with the impl for ARef<U>.

---
 rust/kernel/sync/aref.rs | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
rust/kernel/sync/aref.rs +22 -0
index 0616c0353c2b..9989f56d0605 100644
@@ -170,3 +170,25 @@ fn drop(&mut self) {
         unsafe { T::dec_ref(self.ptr) };
     }
 }
+
+impl<T, U> PartialEq<ARef<U>> for ARef<T>
+where
+    T: AlwaysRefCounted + PartialEq<U>,
+    U: AlwaysRefCounted,
+{
+    #[inline]
+    fn eq(&self, other: &ARef<U>) -> bool {
+        T::eq(&**self, &**other)
+    }
+}
+impl<T: AlwaysRefCounted + Eq> Eq for ARef<T> {}
+
+impl<T, U> PartialEq<&'_ U> for ARef<T>
+where
+    T: AlwaysRefCounted + PartialEq<U>,
+{
+    #[inline]
+    fn eq(&self, other: &&U) -> bool {
+        T::eq(&**self, other)
+    }
+}
Signature · 1 line
2.53.0.473.g4a7958ca14-goog
maintainer Reviewed-by on v2 1/4
#2 lore raw

Re: [PATCH v2 1/4] rust: sync: implement == operator for ARef

Quoted 9 lines · Alice Ryhl

Rust Binder wants to perform a comparison between ARef<Task> and &Task, so define the == operator for ARef<_> when compared with another ARef<_> or just a reference. The operator is implemented in terms of the same operator applied to the inner type.

Note that PartialEq<U> cannot be implemented because it would overlap with the impl for ARef<U>.

Quoted 3 lines
---
 rust/kernel/sync/aref.rs | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)
author maintainer
#3 lore raw

PATCH v2 2/4 rust: task: implement == operator for Task

In reply to Alice Ryhl

It's useful to compare if two tasks are the same task or not. Rust Binder wants this to check if a certain task is equal to the group leader of current.

---
 rust/kernel/task.rs | 9 +++++++++
 1 file changed, 9 insertions(+)
rust/kernel/task.rs +9 -0
index cc907fb531bc..deb948d99921 100644
@@ -362,6 +362,15 @@ unsafe fn dec_ref(obj: ptr::NonNull<Self>) {
     }
 }
 
+impl PartialEq for Task {
+    #[inline]
+    fn eq(&self, other: &Self) -> bool {
+        ptr::eq(self.as_ptr(), other.as_ptr())
+    }
+}
+
+impl Eq for Task {}
+
 impl Kuid {
     /// Get the current euid.
     #[inline]
Signature · 1 line
2.53.0.473.g4a7958ca14-goog
maintainer Reviewed-by on v2 2/4
#4 lore raw

Re: [PATCH v2 2/4] rust: task: implement == operator for Task

Quoted 5 lines · Alice Ryhl

It's useful to compare if two tasks are the same task or not. Rust Binder wants this to check if a certain task is equal to the group leader of current.

Quoted 3 lines
---
 rust/kernel/task.rs | 9 +++++++++
 1 file changed, 9 insertions(+)
author maintainer
#5 lore raw

PATCH v2 3/4 rust_binder: make use of == for Task

In reply to Alice Ryhl

Now that we have implemented the == operator for Task, replace the two raw pointer comparisons in Binder with the == operator.

---
 drivers/android/binder/process.rs | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/android/binder/process.rs +2 -2
index 41de5593197c..b78c1490e654 100644
@@ -681,7 +681,7 @@ pub(crate) fn get_work_or_register<'a>(
     fn get_current_thread(self: ArcBorrow<'_, Self>) -> Result<Arc<Thread>> {
         let id = {
             let current = kernel::current!();
-            if !core::ptr::eq(current.group_leader(), &*self.task) {
+            if self.task != current.group_leader() {
                 pr_err!("get_current_thread was called from the wrong process.");
                 return Err(EINVAL);
             }
@@ -1669,7 +1669,7 @@ pub(crate) fn mmap(
         vma: &mm::virt::VmaNew,
     ) -> Result {
         // We don't allow mmap to be used in a different process.
-        if !core::ptr::eq(kernel::current!().group_leader(), &*this.task) {
+        if this.task != kernel::current!().group_leader() {
             return Err(EINVAL);
         }
         if vma.start() == 0 {
Signature · 1 line
2.53.0.473.g4a7958ca14-goog
maintainer Reviewed-by on v2 3/4
#6 lore raw

Re: [PATCH v2 3/4] rust_binder: make use of == for Task

Quoted 4 lines · Alice Ryhl

Now that we have implemented the == operator for Task, replace the two raw pointer comparisons in Binder with the == operator.

Quoted 3 lines
---
 drivers/android/binder/process.rs | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
author maintainer
#7 lore raw

PATCH v2 4/4 rust_binder: check current before closing fds

In reply to Alice Ryhl

This list gets populated once the transaction is delivered to the target process, at which point it's not touched again except in BC_FREE_BUFFER and process exit, so if the list has been populated then this code should not run in the context of the wrong userspace process.

However, why tempt fate? The function itself can run in the context of both the sender and receiver, and if someone can engineer a scenario where it runs in the sender and this list is non-empty (or future Rust Binder changes make such a scenario possible), then that'd be a problem because we'd be closing random unrelated fds in the wrong process.

---
 drivers/android/binder/allocation.rs | 4 ++++
 1 file changed, 4 insertions(+)
drivers/android/binder/allocation.rs +4 -0
index 7f65a9c3a0e5..31a42738a99d 100644
@@ -260,6 +260,10 @@ fn drop(&mut self) {
                 }
             }
 
+            if self.process.task != kernel::current!().group_leader() {
+                // Called from wrong task, so do not free fds.
+                info.file_list.close_on_free.clear();
+            }
             for &fd in &info.file_list.close_on_free {
                 let closer = match DeferredFdCloser::new(GFP_KERNEL) {
                     Ok(closer) => closer,
Signature · 1 line
2.53.0.473.g4a7958ca14-goog
maintainer
#8 lore raw
Quoted 22 lines · Alice Ryhl

This list gets populated once the transaction is delivered to the target process, at which point it's not touched again except in BC_FREE_BUFFER and process exit, so if the list has been populated then this code should not run in the context of the wrong userspace process.

However, why tempt fate? The function itself can run in the context of both the sender and receiver, and if someone can engineer a scenario where it runs in the sender and this list is non-empty (or future Rust Binder changes make such a scenario possible), then that'd be a problem because we'd be closing random unrelated fds in the wrong process.

---
 drivers/android/binder/allocation.rs | 4 ++++
 1 file changed, 4 insertions(+)
drivers/android/binder/allocation.rs +0 -0
index 7f65a9c3a0e5..31a42738a99d 100644

drivers/android/binder/allocation.rs fn drop(&mut self) { · line 266

Quoted 7 lines
drivers/android/binder/allocation.rs +4 -0
@@ -260,6 +260,10 @@ fn drop(&mut self) {
                 }
             }
 
+            if self.process.task != kernel::current!().group_leader() {
+                // Called from wrong task, so do not free fds.
+                info.file_list.close_on_free.clear();
+            }

If you're sure that this won't actually happen, perhaps print a warning if it's called from a different task but the list is not empty?

Also, I think this can be

if ... {
    for &fd in ... {
    }
}

rather than .clear() and then iterate.

Best,
Gary

Quoted 3 lines
drivers/android/binder/allocation.rs +0 -0
             for &fd in &info.file_list.close_on_free {
                 let closer = match DeferredFdCloser::new(GFP_KERNEL) {
                     Ok(closer) => closer,
author maintainer
#9 lore raw
Quoted 34 lines · Gary Guo
Quoted 22 lines · Alice Ryhl

This list gets populated once the transaction is delivered to the target process, at which point it's not touched again except in BC_FREE_BUFFER and process exit, so if the list has been populated then this code should not run in the context of the wrong userspace process.

However, why tempt fate? The function itself can run in the context of both the sender and receiver, and if someone can engineer a scenario where it runs in the sender and this list is non-empty (or future Rust Binder changes make such a scenario possible), then that'd be a problem because we'd be closing random unrelated fds in the wrong process.

---
 drivers/android/binder/allocation.rs | 4 ++++
 1 file changed, 4 insertions(+)
drivers/android/binder/allocation.rs +0 -0
index 7f65a9c3a0e5..31a42738a99d 100644

drivers/android/binder/allocation.rs fn drop(&mut self) { · line 266

Quoted 7 lines
drivers/android/binder/allocation.rs +4 -0
@@ -260,6 +260,10 @@ fn drop(&mut self) {
                 }
             }
 
+            if self.process.task != kernel::current!().group_leader() {
+                // Called from wrong task, so do not free fds.
+                info.file_list.close_on_free.clear();
+            }

If you're sure that this won't actually happen, perhaps print a warning if it's called from a different task but the list is not empty?

Quoted 8 lines

Also, I think this can be

if ... {
    for &fd in ... {
    }
}

rather than .clear() and then iterate.

Actually, I guess there is one case. When the binder fd is closed, this code may run from deferred_release() in workqueue context. Now, the fd close logic is a no-op from workqueue context, so this patch still doesn't change behavior, but it means the fds wont get closed.

That said, this actually matches C Binder behavior. It also does not close BINDER_TYPE_FDA fds when you close the Binder fd without first issuing the cleanup command for each live BINDER_TYPE_FDA object. Probably not intentional, though.

Alice