On Fri, May 05, 2023 at 04:24:16PM -0700, Stephen Boyd wrote:
>
...
> Unfortunately, the hid google hammer driver hand rolls a devm function
> to call hid_hw_stop() when the driver is unbound and implements an
> hid_driver::remove() function. The driver core doesn't call the devm
> release functions until _after_ the bus unbinds the driver, so the order
> of operations is like this:
Excellent analysis, but the problem is not limited to the hammer driver
(potentially) and shalt be dealt with appropriately, at the HID bus
level.
Actually, it is not even limited to HID, but exists in most buses with
non-trivial ->remove() implementation. For example I fixed similar issue
in I2C in 5b5475826c52 ("i2c: ensure timely release of driver-allocated
resources"). I tried fixing it in SPI but Mark has some objections, and
wanted to fix it in the driver core, so I was thinking about it and then
dropped the ball. At this time I do not think fixing it at driver core
makes logic any clearer, so I think we just need to fix a handful of
buses.
Anyway, I'll CC you on an alternative patch.
Thanks.
--
Dmitry
