在 2016/6/16 6:39, Jann Horn 写道:
On Thu, Jun 16, 2016 at 12:31 AM, Arve Hjønnevåg <arve@xxxxxxxxxxx> wrote:
On Wed, Jun 15, 2016 at 3:09 PM, Jann Horn <jannh@xxxxxxxxxx> wrote:
If /dev/binder is opened and the opener process then e.g. calls execve,
proc->vma_vm_mm will still point to the location of the now-freed
mm_struct. If the process then calls ioctl(binder_fd, ...), the dangling
proc->vma_vm_mm pointer will be compared to current->mm.
Let the binder take a reference to the mm_struct to avoid this.
In the current code, the BUG_ON() will never trigger; it's just there
in case someone changes the conditions under which get_task_mm() can
fail. Just WARN_ON() wouldn't work because of the mmput(), and I don't
want to complicate the code with a dead error handling code path.
(If the BUG_ON() is unacceptable, I'd replace the get_task_mm() with
a direct refcount increment or just omit the BUG_ON().)
This shouldn't cause additional memory usage in a well-behaved system
because you're not supposed to keep binder fds open over fork() or
execve().
Fixes: a906d6931f3ccaf7de805643190765ddd7378e27
Signed-off-by: Jann Horn <jannh@xxxxxxxxxx>
---
drivers/android/binder.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/drivers/android/binder.c b/drivers/android/binder.c
index 16288e7..4e9131d 100644
--- a/drivers/android/binder.c
+++ b/drivers/android/binder.c
@@ -2962,7 +2962,8 @@ static int binder_open(struct inode *nodp, struct file *filp)
return -ENOMEM;
get_task_struct(current);
proc->tsk = current;
- proc->vma_vm_mm = current->mm;
+ proc->vma_vm_mm = get_task_mm(current);
+ BUG_ON(proc->vma_vm_mm == NULL);
INIT_LIST_HEAD(&proc->todo);
init_waitqueue_head(&proc->wait);
proc->default_priority = task_nice(current);
@@ -3167,6 +3168,7 @@ static void binder_deferred_release(struct binder_proc *proc)
vfree(proc->buffer);
}
+ mmput(proc->vma_vm_mm);
put_task_struct(proc->tsk);
binder_debug(BINDER_DEBUG_OPEN_CLOSE,
--
2.8.0.rc3.226.g39d4020
Does this work? In the past, holding a reference to a task's mm while
the driver is open would prevent vma close which in turn would prevent
release of the fd.
Ah, right, mm_struct has two refcounters. I think that should be
atomic_inc(¤t->mm->mm_count) / mmdrop() instead.
.
Hi Jann Horn:
In android platform, the libbinder use the O_CLOEXEC flag:
static int open_driver()
{
int fd = open("/dev/binder", O_RDWR | O_CLOEXEC);
...
return fd;
}
So I think you don't need to consider the case:
"If the process then calls ioctl(binder_fd, ...), the dangling
proc->vma_vm_mm pointer will be compared to current->mm."
Thanks
Zhaojunmin
_______________________________________________
devel mailing list
devel@xxxxxxxxxxxxxxxxxxxxxx
http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
