Re: [PATCH v2 00/33] HID: logitech-hidpp: G920 force-feedback fixes, rework, and cleanups

From: Gideon Heinz

Date: Thu Nov 12 2026 - 10:11:30 EST


Hi Dmitry,

A follow-up to my last mail. I ran the rmmod test a second time with
the same result: no crash, and rmmod blocked for 3m7s, again with 12
“Failed to send command to device!” lines 15.36 s apart. This time the
kernel printed a hung task warning, which shows where it waits:

INFO: task rmmod:6286 blocked for more than 122 seconds.
Tainted: G S 7.3.4-1-cachyos #1
task:rmmod state:D
Call Trace:
schedule+0x6f/0xb0
schedule_timeout+0x31/0x140
wait_for_completion+0xc7/0x1f0
__flush_workqueue+0x198/0x420
drain_workqueue+0x4f/0x150
destroy_workqueue+0x56/0x390
hidpp_ff_destroy+0x42/0x50 [hid_logitech_hidpp]
input_ff_destroy+0x33/0x70
input_unregister_device+0x48/0x70
hidinput_disconnect+0x45/0xc0
hid_hw_stop+0x8c/0xc0
hidpp_remove+0x39/0x60 [hid_logitech_hidpp]
hid_device_remove+0x4a/0xa0
device_release_driver_internal+0x1c2/0x320
driver_detach+0x8f/0xd0
bus_remove_driver+0x102/0x190
hid_unregister_driver+0x22/0xb0
__se_sys_delete_module+0x291/0x3d0

So with your patch the final teardown now runs from
input_unregister_device() while the module is still loaded, which is
what we wanted. The 3 minute wait is destroy_workqueue() draining the
queued commands, each of which times out after the driver’s 3 x 5 s
retry. I have not tested whether the series’ changes to the send path
and removal order shorten this on this kernel.

Thanks,
Gideon

On Thu, Nov 12, 2026 at 3:30 PM Gideon Heinz <gideonheinz02@xxxxxxxxx> wrote:
>
> Hi Dmitry,
>
> I tested your input core patch (“Input: destroy force-feedback device
> on input unregistration”) on a real kernel: CachyOS 7.3.4 (stable,
> which already has an equivalent G920 init fix) plus only your patch,
> with the stock hid-logitech-hidpp driver. The hidpp series is not
> applied, since it doesn’t apply to 7.3.4 as is.
>
> Same reproduction as before: fftest holding the G920 event node at the
> effect prompt, rmmod hid_logitech_hidpp from another terminal, then
> fftest exiting. The machine no longer goes down. The module unloaded
> cleanly and the system stayed responsive.
>
> One observation: with the stock driver, rmmod blocked for about 3
> minutes. After “Unloading HID++ force feedback.” dmesg shows 12
> “Failed to send command to device!” lines, 15.36 s apart (the 3 x 5 s
> retry per command), then rmmod returned. I assume these are the queued
> effect erases timing out because the transport is already closed, and
> that the series’ changes to hidpp_send_message_sync() and the
> synchronous erase would remove this wait, but I have not tested that
> combination on this kernel.
>
> Tested-by: Gideon Heinz gideonheinz02@xxxxxxxxx
>
> Thanks,
> Gideon
>
> On Thu, Nov 12, 2026 at 11:48 AM Gideon Heinz <gideonheinz02@xxxxxxxxx> wrote:
> >
> > Hi Dmitry,
> >
> > I tested the out-of-tree equivalent you suggested, ff->destroy = NULL;
> > in hidpp_ff_stop(), on top of the 33-patch v2 module (G920 046d:c262,
> > 7.3.1-1-cachyos).
> >
> > Same reproduction as before: fftest holding the event node at the
> > effect prompt, rmmod hid_logitech_hidpp from another terminal, then
> > entering -1 in fftest.
> > This time fftest exits normally, the machine stays up, and there are
> > no errors in dmesg.
> > The only message is the single “Unloading HID++ force feedback.” at rmmod time.
> >
> > I’ll try your input core patch on a rebuilt kernel later and report back
> >
> > Thanks,
> > Gideon
> >
> > On Thu, Nov 12, 2026 at 3:43 AM Dmitry Torokhov
> > <dmitry.torokhov@xxxxxxxxx> wrote:
> > >
> > > Hi Gideon,
> > >
> > > On Thu, Nov 12, 2026 at 02:30:39AM +0100, Gideon Heinz wrote:
> > > > Hi Dmitry,
> > > >
> > > > I tested v2 of the series on my G920 (046d:c262). Everything on my
> > > > list works except one scenario that takes the whole machine down, with
> > > > and without your series (last item).
> > > ...
> > > > - rmmod with fftest holding the device open: rmmod returned after
> > > > 2.99 s with "Unloading HID++ force feedback" and nothing else in
> > > > dmesg. When I then exited fftest (entering -1), the machine went
> > > > down: the first time with a garbled panic screen that I couldn't
> > > > read, and on later attempts as a total hang with no output. The
> > > > same sequence on stock Arch 7.2.11-arch1-1 with the unpatched
> > > > driver also hangs the whole system, so this predates the series.
> > > >
> > > > To reproduce: with the module loaded and the G920 plugged in, start
> > > > fftest on its event node and leave it at the effect prompt, run
> > > > "rmmod hid_logitech_hidpp" from another terminal, then enter -1 in
> > > > fftest.
> > > >
> > > > I could not recover a trace yet and will send one if I manage to.
> > > >
> > > > My guess, from reading the code and not verified by a trace:
> > > > hidpp_ff_create() stores hidpp_ff_destroy() and the other FF
> > > > callbacks in the ff_device as pointers into the module. With fftest
> > > > holding the evdev open, the input_dev outlives the driver, so when
> > > > fftest closes, the final input_dev release calls ff->destroy in
> > > > module code that rmmod has already freed.
> > >
> > > Thank you very much for the thorough testing!
> > >
> > > Your diagnosis is spot on. Opening /dev/input/eventX pins the evdev cdev
> > > (and through it input_dev), but does not hold a module reference on the
> > > underlying HID driver. When rmmod unloads hid_logitech_hidpp while fftest
> > > still has the file descriptor open, __input_unregister_device() disconnects
> > > evdev (flushing effects, closing the device, and marking evdev dead so all
> > > subsequent ioctls and writes return -ENODEV) and runs ff->stop(), after
> > > which rmmod frees the module text. When fftest finally exits and closes the
> > > file descriptor, input_dev_release() calls input_ff_destroy(), which
> > > invokes ff->destroy() pointing into the already-unmapped module text.
> > >
> > > Rather than working around this in individual drivers with a separate
> > > ff->stop() callback, input core should call input_ff_destroy() directly in
> > > __input_unregister_device() once all handlers are disconnected and the
> > > device is removed from sysfs. At that point nothing touches dev->ff
> > > anymore, and calling input_ff_destroy() there guarantees ff->destroy()
> > > always runs while the driver and its module are still loaded (while the
> > > second input_ff_destroy() call in input_dev_release() becomes a no-op once
> > > dev->ff is NULL).
> > >
> > > Could you try the input core patch below (with patch 22/33 "HID:
> > > logitech-hidpp: shut down G920 force feedback on input unregister" dropped
> > > from the hid-logitech-hidpp series, since ff->stop is removed and
> > > hidpp_ff_destroy() will now run on input unregistration)?
> > >
> > > Alternatively, if you prefer to test with your out-of-tree module build
> > > without rebuilding the kernel, you can test the equivalent effect by
> > > setting:
> > >
> > > ff->destroy = NULL;
> > >
> > > inside hidpp_ff_stop() in hid-logitech-hidpp.c.
> > >
> > > Thanks.
> > >
> > > ---
> > > From: Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>
> > > Subject: [PATCH] Input: destroy force-feedback device on input unregistration
> > >
> > > When an input device is unregistered while userspace holds an open file
> > > descriptor to its event node, input_dev_release() is deferred until the
> > > last file descriptor is closed. Because input_ff_destroy() was only
> > > called from input_dev_release(), force-feedback teardown (ff->destroy
> > > and freeing ff->private) could run after the parent driver had already
> > > unbound, freed its device state, or even been unloaded via rmmod,
> > > causing use-after-free errors or page faults in unmapped module text.
> > >
> > > Commit ef166ce08801 ("Input: stop force-feedback timer when
> > > unregistering input devices") added a separate ff->stop callback for
> > > ff-memless, but left ff->destroy invoked from input_dev_release().
> > >
> > > Once __input_unregister_device() has disconnected all input handlers
> > > (which flushes effects, closes the device, and marks evdev dead) and
> > > deleted the device from sysfs, nothing accesses dev->ff anymore. Call
> > > input_ff_destroy() directly in __input_unregister_device() so that
> > > ff->destroy() always runs synchronously while the driver and its module
> > > are still active, and remove the now-redundant ff->stop callback. Keep
> > > input_ff_destroy() in input_dev_release() (where it is a no-op once
> > > dev->ff is NULL) to handle input_free_device() on devices that were
> > > never registered.
> > >
> > > Fixes: 509ca1a93836 ("Input: implement new force feedback interface")
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > Reported-by: Gideon Heinz <gideonheinz02@xxxxxxxxx>
> > > Signed-off-by: Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>
> > > ---
> > > drivers/input/ff-core.c | 2 +-
> > > drivers/input/ff-memless.c | 23 -----------------------
> > > drivers/input/input.c | 5 ++---
> > > include/linux/input.h | 3 ---
> > > 4 files changed, 3 insertions(+), 30 deletions(-)
> > >
> > > diff --git a/drivers/input/ff-core.c b/drivers/input/ff-core.c
> > > index 0e5d1d1ceb46..133b4622dc49 100644
> > > --- a/drivers/input/ff-core.c
> > > +++ b/drivers/input/ff-core.c
> > > @@ -338,7 +338,7 @@ EXPORT_SYMBOL_GPL(input_ff_create);
> > > *
> > > * This function is only needed in error path as input core will
> > > * automatically free force feedback structures when device is
> > > - * destroyed.
> > > + * unregistered or destroyed.
> > > */
> > > void input_ff_destroy(struct input_dev *dev)
> > > {
> > > diff --git a/drivers/input/ff-memless.c b/drivers/input/ff-memless.c
> > > index d1fefd1dfc0d..3069436148ec 100644
> > > --- a/drivers/input/ff-memless.c
> > > +++ b/drivers/input/ff-memless.c
> > > @@ -483,32 +483,10 @@ static void ml_ff_destroy(struct ff_device *ff)
> > > {
> > > struct ml_device *ml = ff->private;
> > >
> > > - /*
> > > - * The timer is normally shut down in ml_ff_stop() when the device
> > > - * is unregistered. However, we still shut it down here as a safety
> > > - * net and for cases where the device was never registered (e.g.
> > > - * error paths during probe).
> > > - */
> > > timer_shutdown_sync(&ml->timer);
> > > -
> > > kfree(ml->private);
> > > }
> > >
> > > -static void ml_ff_stop(struct ff_device *ff)
> > > -{
> > > - struct ml_device *ml = ff->private;
> > > -
> > > - /*
> > > - * Even though we stop all playing effects when tearing down an
> > > - * input device (by the way of evdev calling input_flush_device()
> > > - * that calls into input_ff_flush() that stops and erases all
> > > - * effects), we do not actually shutdown the timer, and therefore
> > > - * we should do it here to prevent it firing after the input
> > > - * device is unregistered and its associated resources are freed.
> > > - */
> > > - timer_shutdown_sync(&ml->timer);
> > > -}
> > > -
> > > /**
> > > * input_ff_create_memless() - create memoryless force-feedback device
> > > * @dev: input device supporting force-feedback
> > > @@ -543,7 +521,6 @@ int input_ff_create_memless(struct input_dev *dev, void *data,
> > > ff->playback = ml_ff_playback;
> > > ff->set_gain = ml_ff_set_gain;
> > > ff->destroy = ml_ff_destroy;
> > > - ff->stop = ml_ff_stop;
> > >
> > > /* we can emulate periodic effects with RUMBLE */
> > > if (test_bit(FF_RUMBLE, ff->ffbit)) {
> > > diff --git a/drivers/input/input.c b/drivers/input/input.c
> > > index 210635ef147e..7be69deed8eb 100644
> > > --- a/drivers/input/input.c
> > > +++ b/drivers/input/input.c
> > > @@ -2286,10 +2286,9 @@ static void __input_unregister_device(struct input_dev *dev)
> > > input_wakeup_procfs_readers();
> > > }
> > >
> > > - if (dev->ff && dev->ff->stop)
> > > - dev->ff->stop(dev->ff);
> > > -
> > > device_del(&dev->dev);
> > > +
> > > + input_ff_destroy(dev);
> > > }
> > >
> > > static void devm_input_device_unregister(struct device *dev, void *res)
> > > diff --git a/include/linux/input.h b/include/linux/input.h
> > > index 3381608127f7..02047cc45d04 100644
> > > --- a/include/linux/input.h
> > > +++ b/include/linux/input.h
> > > @@ -549,8 +549,6 @@ extern const struct class input_class;
> > > * @set_autocenter: Called to auto-center device
> > > * @destroy: called by input core when parent input device is being
> > > * destroyed
> > > - * @stop: called by input core when parent input device is being
> > > - * unregistered
> > > * @private: driver-specific data, will be freed automatically
> > > * @ffbit: bitmap of force feedback capabilities truly supported by
> > > * device (not emulated like ones in input_dev->ffbit)
> > > @@ -579,7 +577,6 @@ struct ff_device {
> > > void (*set_autocenter)(struct input_dev *dev, u16 magnitude);
> > >
> > > void (*destroy)(struct ff_device *);
> > > - void (*stop)(struct ff_device *);
> > >
> > > void *private;
> > >
> > >
> > > --
> > > Dmitry