ffs_epfile_io and ffs_epfile_io_complete runs in different context, but
there is no synchronization between them.
consider the following scenario:
1) ffs_epfile_io interrupted by sigal while
wait_for_completion_interruptible
2) then ffs_epfile_io set ret to -EINTR
3) just before or during usb_ep_dequeue, the request completed
4) ffs_epfile_io return with -EINTR
In this case, ffs_epfile_io tell caller no transfer success but actually
it may has been done. This break the caller's pipe.
Below script can help test it (adbd is the process which lies on f_fs).
while true
do
pkill -19 adbd #SIGSTOP
pkill -18 adbd #SIGCONT
sleep 0.1
done
To avoid this, just dequeue the request first. After usb_ep_dequeue, the
request must be done or canceled.
With this change, we can ensure no race condition in f_fs driver. But
actually I found some of the udc driver has analogical issue in its
dequeue implementation. For example,
1) the dequeue function hold the controller's lock.
2) before driver request controller to stop transfer, a request
completed.
3) the controller trigger a interrupt, but its irq handler need wait
dequeue function to release the lock.
4) dequeue function give back the request with negative status, and
release lock.
5) irq handler get lock but the request has already been given back.
So, the dequeue implementation should take care of this case. IMO, it
can be done as below steps to dequeue a already started request,
1) request controller to stop transfer on the given ep. HW know the
actual transfer status.
2) after hw stop transfer, driver scan if there are any completed one.
3) if found, process it with real status. if no, the request can
canceled.
Signed-off-by: "Du, Changbin" <changbin.du@intel.com>
[mina86@mina86.com: rebased on top of refactoring commits]
Signed-off-by: Michal Nazarewicz <mina86@mina86.com>
Signed-off-by: Felipe Balbi <balbi@kernel.org>
ret = -EINVAL;
} else if (!io_data->aio) {
DECLARE_COMPLETION_ONSTACK(done);
+ bool interrupted = false;
req = ep->req;
req->buf = data;
spin_unlock_irq(&epfile->ffs->eps_lock);
if (unlikely(wait_for_completion_interruptible(&done))) {
- ret = -EINTR;
+ /*
+ * To avoid race condition with ffs_epfile_io_complete,
+ * dequeue the request first then check
+ * status. usb_ep_dequeue API should guarantee no race
+ * condition with req->complete callback.
+ */
usb_ep_dequeue(ep->ep, req);
- goto error_mutex;
+ interrupted = ep->status < 0;
}
/*
* rounded up to maxpacketsize), we may end up with more data
* then user space has space for.
*/
- ret = ep->status;
+ ret = interrupted ? -EINTR : ep->status;
if (io_data->read && ret > 0) {
ret = copy_to_iter(data, ret, &io_data->data);
if (!ret)