Skip to contentSkip to conversation

rust_binder: use a u64 stride when cleaning up the offsets array

PATCH In review Hyunwoo Kim 3 participants 3 messages May 31 – Jul 2

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 rust_binder: use a u64 stride when cleaning up the offsets array

Allocation's Drop walks the offsets array (binder_size_t = u64 entries), cleaning up the objects, but it used usize instead of u64 for both the stride and the per-entry read.

On 64-bit kernels (usize == u64) this is harmless, but on 32-bit kernels it walks the 8-byte entries in 4-byte steps, iterating an N-entry array 2N times, and reads the always-zero high word as offset 0, cleaning up the object at offset 0 N extra times. As a result the referenced node or handle ends up with a lower reference count than it actually has (a refcount over-decrement), and binder's reference accounting is corrupted; for example, the owner can be notified of a strong reference release (BR_RELEASE) even though references still remain.

Change the stride to u64, and read each entry as a u64, narrowing it to usize with try_into().

On 32-bit ARM, when this over-decrement would drive a count below zero, the driver's existing refcount guard refuses it and fires:

rust_binder: Failure: refcount underflow!

Fixes eafedbc7c050 ("rust_binder: add Rust Binder driver")

---
Changes in v2:
- reformat to satisfy rustfmt, as pointed out by the kernel test robot
- v1: https://lore.kernel.org/all/ahjpn-3WQTywTdyj@v4bel/

---
 drivers/android/binder/allocation.rs | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/android/binder/allocation.rs +3 -2
index b7b05e72970a..ea5846e4da16 100644
@@ -259,7 +259,7 @@ fn drop(&mut self) {
 
             if let Some(offsets) = info.offsets.clone() {
                 let view = AllocationView::new(self, offsets.start);
-                for i in offsets.step_by(size_of::<usize>()) {
+                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)
                     }
@@ -420,7 +420,8 @@ pub(crate) fn transfer_binder_object(
     }
 
     fn cleanup_object(&self, index_offset: usize) -> Result {
-        let offset = self.alloc.read(index_offset)?;
+        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 => {
Signature · 1 line
2.43.0
maintainer Acked-by on v2 1/1
#1 lore raw
Quoted 58 lines · Hyunwoo Kim

Allocation's Drop walks the offsets array (binder_size_t = u64 entries), cleaning up the objects, but it used usize instead of u64 for both the stride and the per-entry read.

On 64-bit kernels (usize == u64) this is harmless, but on 32-bit kernels it walks the 8-byte entries in 4-byte steps, iterating an N-entry array 2N times, and reads the always-zero high word as offset 0, cleaning up the object at offset 0 N extra times. As a result the referenced node or handle ends up with a lower reference count than it actually has (a refcount over-decrement), and binder's reference accounting is corrupted; for example, the owner can be notified of a strong reference release (BR_RELEASE) even though references still remain.

Change the stride to u64, and read each entry as a u64, narrowing it to usize with try_into().

On 32-bit ARM, when this over-decrement would drive a count below zero, the driver's existing refcount guard refuses it and fires:

rust_binder: Failure: refcount underflow!

Fixes eafedbc7c050 ("rust_binder: add Rust Binder driver")

---
Changes in v2:
- reformat to satisfy rustfmt, as pointed out by the kernel test robot
- v1: https://lore.kernel.org/all/ahjpn-3WQTywTdyj@v4bel/

---
 drivers/android/binder/allocation.rs | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/android/binder/allocation.rs +3 -2
index b7b05e72970a..ea5846e4da16 100644
@@ -259,7 +259,7 @@ fn drop(&mut self) {
 
             if let Some(offsets) = info.offsets.clone() {
                 let view = AllocationView::new(self, offsets.start);
-                for i in offsets.step_by(size_of::<usize>()) {
+                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)
                     }
@@ -420,7 +420,8 @@ pub(crate) fn transfer_binder_object(
     }
 
     fn cleanup_object(&self, index_offset: usize) -> Result {
-        let offset = self.alloc.read(index_offset)?;
+        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 => {
Signature · 1 line
2.43.0

Hey Alice, have you seen this? This looks correct to me so,

maintainer Reviewed-by on v2 1/1
#2 lore raw
In reply to Hyunwoo Kim
Quoted 16 lines · Hyunwoo Kim

Allocation's Drop walks the offsets array (binder_size_t = u64 entries), cleaning up the objects, but it used usize instead of u64 for both the stride and the per-entry read.

On 64-bit kernels (usize == u64) this is harmless, but on 32-bit kernels it walks the 8-byte entries in 4-byte steps, iterating an N-entry array 2N times, and reads the always-zero high word as offset 0, cleaning up the object at offset 0 N extra times. As a result the referenced node or handle ends up with a lower reference count than it actually has (a refcount over-decrement), and binder's reference accounting is corrupted; for example, the owner can be notified of a strong reference release (BR_RELEASE) even though references still remain.

Change the stride to u64, and read each entry as a u64, narrowing it to usize with try_into().

Quoted 8 lines

On 32-bit ARM, when this over-decrement would drive a count below zero, the driver's existing refcount guard refuses it and fires:

rust_binder: Failure: refcount underflow!

Fixes eafedbc7c050 ("rust_binder: add Rust Binder driver")

Thanks!