Hi Dmitry,
On Thu, 2026-11-12 at 21:50 -0800, Dmitry Torokhov wrote:
Hi Rong,
On Fri, Nov 13, 2026 at 01:53:55AM +0800, Rong Zhang wrote:
quoted
It's painful to have error paths after calling input_ff_create(),
because input_ff_destroy() frees too much things but does not restore
some other things.
On get_device failure, the module's bail-out path calls module_put()
right before hid_haptic_destroy() calling it again, causing refcount
underflow.
Besides, every error path calling input_ff_destroy() frees ff->private,
while the others don't. This divergence has caused quite a lot of
trouble.
Get the module and the device beforehand, so that there is no error path
after input_ff_create(), eliminating the call to input_ff_destroy(). In
this manner, we don't need to restore dev->flush and dev->event either.
The only reason hid_haptic_init() had try_module_get(THIS_MODULE) and
get_device(&hdev->dev) (and their counterparts in hid_haptic_destroy())
was to work around input_ff_destroy() previously being called only from
input_dev_release(), which could be deferred until after hdev was
unbound or the driver module was unloaded if userspace kept the evdev
node open.
Thanks for the clarification. Yeah, it was a painful design...
I am queuing an input core change ("Input: destroy force-feedback device
on input unregistration") [1] that calls input_ff_destroy() directly in
__input_unregister_device() once all handlers are disconnected and the
device is removed from sysfs.
With that change in place, hid_haptic_destroy() is guaranteed to run
synchronously during input_unregister_device() while hdev is still bound
and alive (and dev->dev.parent already holds a reference to &hdev->dev
anyway). That means get_device(&hdev->dev) / put_device(&hdev->dev) and
try_module_get(THIS_MODULE) / module_put(THIS_MODULE) are not needed at
all here (until patch 8/9 where hid-haptic becomes a separate plugin
module and hid_plugin_get() is taken before calling self_plugin->init()),
and simply dropping them eliminates the error path after
input_ff_create(). It also avoids the devres lifetime issue with
put_device(&hdev->dev) mentioned in patch 1/9.
Thanks for fixing the ff core and reviewing this series! The lifecycle of
ff does look much better with your patch.
I will follow your suggestion to simplify the fixes in the series and
resubmit it when your patch lands.
Thanks,
Rong
[1] https://lore.kernel.org/linux-input/2f722967b6db0692d4dd4164aff39b62@gmail.com/ (local)
Thanks.