fuse: fix use after free issue in fuse_dev_do_read()
authorSahitya Tummala <stummala@codeaurora.org>
Wed, 8 Feb 2017 15:00:56 +0000 (20:30 +0530)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 23 Feb 2017 16:43:09 +0000 (17:43 +0100)
commit 6ba4d2722d06960102c981322035239cd66f7316 upstream.

There is a potential race between fuse_dev_do_write()
and request_wait_answer() contexts as shown below:

TASK 1:
__fuse_request_send():
  |--spin_lock(&fiq->waitq.lock);
  |--queue_request();
  |--spin_unlock(&fiq->waitq.lock);
  |--request_wait_answer():
       |--if (test_bit(FR_SENT, &req->flags))
       <gets pre-empted after it is validated true>
                                   TASK 2:
                                   fuse_dev_do_write():
                                     |--clears bit FR_SENT,
                                     |--request_end():
                                        |--sets bit FR_FINISHED
                                        |--spin_lock(&fiq->waitq.lock);
                                        |--list_del_init(&req->intr_entry);
                                        |--spin_unlock(&fiq->waitq.lock);
                                        |--fuse_put_request();
       |--queue_interrupt();
       <request gets queued to interrupts list>
            |--wake_up_locked(&fiq->waitq);
       |--wait_event_freezable();
       <as FR_FINISHED is set, it returns and then
       the caller frees this request>

Now, the next fuse_dev_do_read(), see interrupts list is not empty
and then calls fuse_read_interrupt() which tries to access the request
which is already free'd and gets the below crash:

[11432.401266] Unable to handle kernel paging request at virtual address
6b6b6b6b6b6b6b6b
...
[11432.418518] Kernel BUG at ffffff80083720e0
[11432.456168] PC is at __list_del_entry+0x6c/0xc4
[11432.463573] LR is at fuse_dev_do_read+0x1ac/0x474
...
[11432.679999] [<ffffff80083720e0>] __list_del_entry+0x6c/0xc4
[11432.687794] [<ffffff80082c65e0>] fuse_dev_do_read+0x1ac/0x474
[11432.693180] [<ffffff80082c6b14>] fuse_dev_read+0x6c/0x78
[11432.699082] [<ffffff80081d5638>] __vfs_read+0xc0/0xe8
[11432.704459] [<ffffff80081d5efc>] vfs_read+0x90/0x108
[11432.709406] [<ffffff80081d67f0>] SyS_read+0x58/0x94

As FR_FINISHED bit is set before deleting the intr_entry with input
queue lock in request completion path, do the testing of this flag and
queueing atomically with the same lock in queue_interrupt().

Signed-off-by: Sahitya Tummala <stummala@codeaurora.org>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Fixes: fd22d62ed0c3 ("fuse: no fc->lock for iqueue parts")
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/fuse/dev.c

index 9096d44eb22155493176c89ede97234ba50e1c06..d0cf1f010fbe7604a781275c01a4f4e50417a13a 100644 (file)
@@ -418,6 +418,10 @@ static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 {
        spin_lock(&fiq->waitq.lock);
+       if (test_bit(FR_FINISHED, &req->flags)) {
+               spin_unlock(&fiq->waitq.lock);
+               return;
+       }
        if (list_empty(&req->intr_entry)) {
                list_add_tail(&req->intr_entry, &fiq->interrupts);
                wake_up_locked(&fiq->waitq);