Re: [PATCH v2] HID: Replace semaphore driver_lock with mutex

Hey

On Wed, Jun 14, 2017 at 9:20 AM, Arnd Bergmann [off-list ref] wrote:

On Wed, Jun 14, 2017 at 7:22 AM, Binoy Jayan [off-list ref] wrote:

quoted

Hi,

On 14 June 2017 at 01:55, Arnd Bergmann [off-list ref] wrote:

quoted

quoted

The mutex code clearly states mutex_trylock() must not be used in
interrupt context (see kernel/locking/mutex.c), hence we used a
semaphore here. Unless the mutex code is changed to allow this, we
cannot switch away from semaphores.

Right, that makes a lot of sense. I don't think changing the mutex
code is an option here, but I wonder if we can replace the semaphore
with something simpler anyway.

From what I can tell, it currently does two things:

1. it acts as a simple flag to prevent  hid_input_report from derefencing
    the hid->driver pointer during initialization and exit. I think this could
    be done equally well using a simple atomic set_bit()/test_bit() or similar.

2. it prevents the hid->driver pointer from becoming invalid while an
    asynchronous hid_input_report() is in progress. This actually seems to
    be a reference counting problem rather than a locking problem.
    I don't immediately see how to better address it, or how exactly this
    could go wrong in practice, but I would naively expect that either
    hdev->driver->remove() needs to wait for the last user of hdev->driver
    to complete, or we need kref_get/kref_put in hid_input_report()
    to trigger the actual release function.

The HID design is explained in detail in
./Documentation/hid/hid-transport.txt, in case you want some
background information. The problem here is that the low-level
transport driver has a lifetime that is independent of the hid
device-driver. So the transport driver needs to be able to tell the
HID layer about coming/going devices, as well as incoming traffic. At
the same time, the HID layer can bind upper-layer hid device drivers
*anytime* (since it is exposed via the driver core interfaces in /sys
to bind drivers).

The locking architecture is very similar to 's_active' on
super-blocks, or 'active' on kernfs-nodes. However, the big difference
here is that drivers can be rebound. Hence, we're not limited to just
one driver lifetime, which is why we went with a semaphore instead.

Also note that hid_input_report() might be called from interrupt
context, hence it can never call into kref_put() or similar (unless we
want to guarantee that unbinding can run in interrupt context).

If we really want to get rid of the semaphore, a spinlock might do
fine as well. Then again, some hid device drivers might expect their
transport driver to *not* run in irq context, and thus break under a
spinlock. So if you want to fix this, we need to audit the hid device
drivers.

David