binder: Prevent context manager from incrementing ref 0
authorJann Horn <jannh@google.com>
Mon, 27 Jul 2020 12:04:24 +0000 (14:04 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 21 Aug 2020 09:01:55 +0000 (11:01 +0200)
commit 4b836a1426cb0f1ef2a6e211d7e553221594f8fc upstream.

Binder is designed such that a binder_proc never has references to
itself. If this rule is violated, memory corruption can occur when a
process sends a transaction to itself; see e.g.
<https://syzkaller.appspot.com/bug?extid=09e05aba06723a94d43d>.

There is a remaining edgecase through which such a transaction-to-self
can still occur from the context of a task with BINDER_SET_CONTEXT_MGR
access:

 - task A opens /dev/binder twice, creating binder_proc instances P1
   and P2
 - P1 becomes context manager
 - P2 calls ACQUIRE on the magic handle 0, allocating index 0 in its
   handle table
 - P1 dies (by closing the /dev/binder fd and waiting a bit)
 - P2 becomes context manager
 - P2 calls ACQUIRE on the magic handle 0, allocating index 1 in its
   handle table
   [this triggers a warning: "binder: 1974:1974 tried to acquire
   reference to desc 0, got 1 instead"]
 - task B opens /dev/binder once, creating binder_proc instance P3
 - P3 calls P2 (via magic handle 0) with (void*)1 as argument (two-way
   transaction)
 - P2 receives the handle and uses it to call P3 (two-way transaction)
 - P3 calls P2 (via magic handle 0) (two-way transaction)
 - P2 calls P2 (via handle 1) (two-way transaction)

And then, if P2 does *NOT* accept the incoming transaction work, but
instead closes the binder fd, we get a crash.

Solve it by preventing the context manager from using ACQUIRE on ref 0.
There shouldn't be any legitimate reason for the context manager to do
that.

Additionally, print a warning if someone manages to find another way to
trigger a transaction-to-self bug in the future.

Cc: stable@vger.kernel.org
Fixes: 457b9a6f09f0 ("Staging: android: add binder driver")
Acked-by: Todd Kjos <tkjos@google.com>
Signed-off-by: Jann Horn <jannh@google.com>
Reviewed-by: Martijn Coenen <maco@android.com>
Link: https://lore.kernel.org/r/20200727120424.1627555-1-jannh@google.com
[manual backport: remove fine-grained locking and error reporting that
                  don't exist in <=4.9]
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/android/binder.c

index e12288c..f4c0b62 100644 (file)
@@ -1427,6 +1427,10 @@ static void binder_transaction(struct binder_proc *proc,
                        return_error = BR_DEAD_REPLY;
                        goto err_dead_binder;
                }
+               if (WARN_ON(proc == target_proc)) {
+                       return_error = BR_FAILED_REPLY;
+                       goto err_invalid_target_handle;
+               }
                if (security_binder_transaction(proc->tsk,
                                                target_proc->tsk) < 0) {
                        return_error = BR_FAILED_REPLY;
@@ -1830,6 +1834,11 @@ static int binder_thread_write(struct binder_proc *proc,
                        ptr += sizeof(uint32_t);
                        if (target == 0 && binder_context_mgr_node &&
                            (cmd == BC_INCREFS || cmd == BC_ACQUIRE)) {
+                               if (binder_context_mgr_node->proc == proc) {
+                                       binder_user_error("%d:%d context manager tried to acquire desc 0\n",
+                                                         proc->pid, thread->pid);
+                                       return -EINVAL;
+                               }
                                ref = binder_get_ref_for_node(proc,
                                               binder_context_mgr_node);
                                if (ref->desc != target) {