[PATCH v2 12/13] ANDROID: binder: don't queue async transactions to thread.
From: Martijn Coenen
Date: Thu Aug 31 2017 - 04:05:54 EST
This can cause issues with processes using the poll()
interface:
1) client sends two oneway transactions
2) the second one gets queued on async_todo
(because the server didn't handle the first one
yet)
3) server returns from poll(), picks up the
first transaction and does transaction work
4) server is done with the transaction, sends
BC_FREE_BUFFER, and the second transaction gets
moved to thread->todo
5) libbinder's handlePolledCommands() only handles
the commands in the current data buffer, so
doesn't see the new transaction
6) the server continues running and issues a new
outgoing transaction. Now, it suddenly finds
the incoming oneway transaction on its thread
todo, and returns that to userspace.
7) userspace does not expect this to happen; it
may be holding a lock while making the outgoing
transaction, and if handling the incoming
trasnaction requires taking the same lock,
userspace will deadlock.
By queueing the async transaction to the proc
workqueue, we make sure it's only picked up when
a thread is ready for proc work.
Signed-off-by: Martijn Coenen <maco@xxxxxxxxxxx>
---
drivers/android/binder.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/drivers/android/binder.c b/drivers/android/binder.c
index 2d23f8699d40..7768ba280177 100644
--- a/drivers/android/binder.c
+++ b/drivers/android/binder.c
@@ -3570,11 +3570,13 @@ static int binder_thread_write(struct binder_proc *proc,
BUG_ON(buf_node->proc != proc);
w = binder_dequeue_work_head_ilocked(
&buf_node->async_todo);
- if (!w)
+ if (!w) {
buf_node->has_async_transaction = 0;
- else
+ } else {
binder_enqueue_work_ilocked(
- w, &thread->todo);
+ w, &proc->todo);
+ binder_wakeup_proc_ilocked(proc);
+ }
binder_node_inner_unlock(buf_node);
}
trace_binder_transaction_buffer_release(buffer);
--
2.14.1.581.gf28d330327-goog