On Tue, Nov 22, 2022 at 05:56:56PM +0530, Udipto Goswami wrote:
> On 11/22/22 5:17 PM, John Keeping wrote:
> > On Mon, Nov 21, 2022 at 09:52:43AM +0530, Udipto Goswami wrote:
> > > Hi John
> > >
> > > On 11/20/22 11:18 PM, John Keeping wrote:
> > > > On Sun, Nov 20, 2022 at 12:23:50PM +0530, Udipto Goswami wrote:
> > > > > On 11/18/22 9:49 PM, John Keeping wrote:
> > > > > > On Wed, Nov 16, 2022 at 04:49:55PM +0530, Udipto Goswami wrote:
> > > > > > > While performing fast composition switch, there is a possibility that the
> > > > > > > process of ffs_ep0_write/ffs_ep0_read get into a race condition
> > > > > > > due to ep0req being freed up from functionfs_unbind.
> > > > > > >
> > > > > > > Consider the scenario that the ffs_ep0_write calls the ffs_ep0_queue_wait
> > > > > > > by taking a lock &ffs->ev.waitq.lock. However, the functionfs_unbind isn't
> > > > > > > bounded so it can go ahead and mark the ep0req to NULL, and since there
> > > > > > > is no NULL check in ffs_ep0_queue_wait we will end up in use-after-free.
> > > > > > >
> > > > > > > Fix this by making a serialized execution between the two functions using
> > > > > > > a mutex_lock(ffs->mutex). Also, dequeue the ep0req to ensure that no
> > > > > > > other function can use it after the free operation.
> > > > > > >
> > > > > > > Fixes: ddf8abd25994 ("USB: f_fs: the FunctionFS driver")
> > > > > > > Signed-off-by: Udipto Goswami <quic_ugoswami@xxxxxxxxxxx>
> > > > > > > ---
> > > > > > > v2: Replaces spinlock with mutex & added dequeue operation in unbind.
> > > > > > >
> > > > > > > drivers/usb/gadget/function/f_fs.c | 7 +++++++
> > > > > > > 1 file changed, 7 insertions(+)
> > > > > > >
> > > > > > > diff --git a/drivers/usb/gadget/function/f_fs.c b/drivers/usb/gadget/function/f_fs.c
> > > > > > > index 73dc10a77cde..1439449df39a 100644
> > > > > > > --- a/drivers/usb/gadget/function/f_fs.c
> > > > > > > +++ b/drivers/usb/gadget/function/f_fs.c
> > > > > > > @@ -279,6 +279,9 @@ static int __ffs_ep0_queue_wait(struct ffs_data *ffs, char *data, size_t len)
> > > > > > > struct usb_request *req = ffs->ep0req;
> > > > > > > int ret;
> > > > > > > + if (!req)
> > > > > > > + return -EINVAL;
> > > > > > > +
> > > > > > > req->zero = len < le16_to_cpu(ffs->ev.setup.wLength);
> > > > > > > spin_unlock_irq(&ffs->ev.waitq.lock);
> > > > > > > @@ -1892,10 +1895,14 @@ static void functionfs_unbind(struct ffs_data *ffs)
> > > > > > > ENTER();
> > > > > > > if (!WARN_ON(!ffs->gadget)) {
> > > > > > > + mutex_lock(&ffs->mutex);
> > > > > > > + /* dequeue before freeing ep0req */
> > > > > > > + usb_ep_dequeue(ffs->gadget->ep0, ffs->ep0req);
> > > > > > > usb_ep_free_request(ffs->gadget->ep0, ffs->ep0req);
> > > > > > > ffs->ep0req = NULL;
> > > > > > > ffs->gadget = NULL;
> > > > > > > clear_bit(FFS_FL_BOUND, &ffs->flags);
> > > > > > > + mutex_unlock(&ffs->mutex);
> > > > > >
> > > > > > There's now a deadlock here if some other thread is waiting in
> > > > > > __ffs_ep0_queue_wait() on ep0req_completion.
> > > > > >
> > > > > > You need to dequeue before taking the lock.
> > > > > That's a control request right, will it be async?
> > > > >
> > > > > Anyway I see only 2 possible threads ep0_read/ep0_write who calls
> > > > > ep0_queue_wait and waits for the completion of ep0req and both
> > > > > ep0_read/write are prptected by the mutex lock so i guess execution won't
> > > > > reach there right ?
> > > > > Say functionfs_unbind ran first then ep0_read/write had to wait will the
> > > > > functionfs_unbind is completed so ep_dequeue will ran, will get completed,
> > > > > further free the request, mark in NULL. now ep0_read/write will have ep0req
> > > > > as NULL so bail out.
> > > > >
> > > > > Is reverse then functionfs_unbind will wait will the ep0_read/write is
> > > > > completed.
> > > >
> > > > What guarantee is there that the transfer completes?
> > > >
> > > > If there is such a guarantee, then the request will not be queued, so
> > > > why is usb_ep_dequeue() necessary?
> > >
> > > I Agree that we cannot say that for sure, but we see that
> > > wait_for_completion in the ep0_queue_wait is also inside mutex which was
> > > acquired in ep0_read/write right?
> >
> > Correct.
> >
> > > I Though of maintaining the uniformity for the approaches.
> >
> > What uniformity? If one process is blocked waiting for completion and
> > another process wants to cancel the operation, then the cancel
> > (usb_eq_dequeue()) must run concurrently with the wait, otherwise the
> > blocked process will never wake up.
>
> I get that, we want to rely on the dequeue to get us unblocked.
> But this is also true right that doing dequeue outside might cause this?
>
> functionfs_unbind
> ep0_dequeue
> ffs_ep0_read
> mutex_lock()
> giveback ep0_queue
> map request buffer
> unmap buffer
>
> This can affect the controller's list i.e the pending_list for ep0 or might
> also result on controller accessing a stale memory address isn't it ?
>
> Or does the mutex would let the ep0_read execute in atomic context? No
> right. Will it not be able to execute parallely? If not then yah we can do
> dequeue outside mutex for sure.
I would expect that if we're in unbind then any attempt to enqueue a new
request will fail, so if the mutex is taken in the case above ep_queue()
should fail with -ESHUTDOWN.
But I can't actually find an point to any code that ensures that is the
case!
This doesn't matter too much though - it's not going to result in any
access to stale memory because either:
ep0_dequeue
ffs_ep0_read
mutex_lock()
ep0_queue
... wait for response ...
read ep0req->status
mutex_unlock()
mutex_lock()
free_ep0_request
...
or:
ffs_ep0_read
mutex_lock()
ep0_queue
ep0_dequeue
wake up
read ep0req->status
mutex_unlock()
mutex_lock()
free_ep0_request
...
The first case isn't ideal as we don't want to be waiting for a request
while unbinding, but it's not unsafe.
