On Wed, Aug 23, 2023 at 07:55:17AM -0700, Dmitry Torokhov wrote: > On Wed, Aug 23, 2023 at 03:16:02PM +0200, Maxime Ripard wrote: > > On Wed, Aug 23, 2023 at 05:51:00AM -0700, Dmitry Torokhov wrote: > > > On Wed, Aug 23, 2023 at 09:44:22AM +0200, Maxime Ripard wrote: > > > > On Tue, Aug 22, 2023 at 08:57:41AM -0700, Rahul Rameshbabu wrote: > > > > > On Tue, 22 Aug, 2023 11:12:28 +0200 Maxime Ripard <mripard@xxxxxxxxxx> wrote: > > > > > > Hi, > > > > > > > > > > > > So, we discussed it this morning with Benjamin, and I think the culprit > > > > > > is that the uclogic driver will allocate a char array with devm_kzalloc > > > > > > in uclogic_input_configured() > > > > > > (https://elixir.bootlin.com/linux/latest/source/drivers/hid/hid-uclogic-core.c#L149), > > > > > > and will assign input_dev->name to that pointer. > > > > > > > > > > > > When the device is removed, the devm-allocated array is freed, and the > > > > > > input framework will send a uevent in input_dev_uevent() using the > > > > > > input_dev->name field: > > > > > > > > > > > > https://elixir.bootlin.com/linux/latest/source/drivers/input/input.c#L1688 > > > > > > > > > > > > So it's a classic dangling pointer situation. > > > > > > > > > > > > And even though it was revealed by that patch, I think the issue is > > > > > > unrelated. The fundamental issue seems to be that the usage of devm in > > > > > > that situation is wrong. > > > > > > > > > > > > input_dev->name is accessed by input_dev_uevent, which for KOBJ_UNBIND > > > > > > and KOBJ_REMOVE will be called after remove. > > > > > > > > > > > > For example, in __device_release_driver() (with the driver remove hook > > > > > > being called in device_remove() and devres_release_all() being called in > > > > > > device_unbind_cleanup()): > > > > > > https://elixir.bootlin.com/linux/latest/source/drivers/base/dd.c#L1278 > > > > > > > > > > > > So, it looks to me that, with or without the patch we merged recently, > > > > > > the core has always sent uevent after device-managed resources were > > > > > > freed. Thus, the uclogic (and any other input driver) was wrong in > > > > > > allocating its input_dev name with devm_kzalloc (or the phys and uniq > > > > > > fields in that struct). > > > > > > > > > > > > Note that freeing input_dev->name in remove would have been just as bad. > > > > > > > > > > > > Looking at the code quickly, at least hid-playstation, > > > > > > hid-nvidia-shield, hid-logitech-hidpp, mms114 and tsc200x seem to be > > > > > > affected by the same issue. > > > > > > > > > > I agree with this analysis overall. At least in hid-nvidia-shield, I can > > > > > not use devm for allocating the input name string and explicitly free it > > > > > after calling input_unregister_device. In this scenario, the name string > > > > > would have been freed explicitly after input_put_device was called > > > > > (since the input device is not devres managed). input_put_device would > > > > > drop the reference count to zero and the device would be cleaned up at > > > > > that point triggering KOBJ_REMOVE and firing off that final > > > > > input_dev_uevent. > > > > > > > > > > I think this can be done for a number of the drivers as a workaround > > > > > till this issue is properly resolved. If this seems appropriate, I can > > > > > send out a series later in the day. This is just a workaround till the > > > > > discussion below converges (which I am interested in). > > > > > > > > I'm sorry, I don't know the input framework well enough to understand > > > > what you had in mind exactly. Could you send a patch with your > > > > suggestion for the hid-nvidia-shield so we can discuss this further? > > > > > > > > That being said, I think that the current design around name, phys and > > > > uniq is fairly treacherous to drivers and we should aim for a solution > > > > that prevents that issue from being possible at all. > > > > > > > > I was inclined to go for a char array for each to get rid of the pointer > > > > entirely, but Benjamin raised some concerns over the structure size so > > > > it's probably not a great solution. > > > > > > I think everything is much simpler, with uclogic driver being in the > > > wrong here: devm resource needs to be attached to the right device > > > (instance of HID) rather than to the input device itself (which should > > > never have any driver resources attached since it never has a driver). > > > Something like this: > > > > > > diff --git a/drivers/hid/hid-uclogic-core.c b/drivers/hid/hid-uclogic-core.c > > > index f67835f9ed4c..f234a7c97360 100644 > > > --- a/drivers/hid/hid-uclogic-core.c > > > +++ b/drivers/hid/hid-uclogic-core.c > > > @@ -148,7 +148,7 @@ static int uclogic_input_configured(struct hid_device *hdev, > > > > > > if (suffix) { > > > len = strlen(hdev->name) + 2 + strlen(suffix); > > > - name = devm_kzalloc(&hi->input->dev, len, GFP_KERNEL); > > > + name = devm_kzalloc(&hdev->dev, len, GFP_KERNEL); > > > if (name) { > > > snprintf(name, len, "%s %s", hdev->name, suffix); > > > hi->input->name = name; > > > > > > In general, drivers should attach devm resources they allocate to the > > > instance of device they are binding to, and nothing else. > > > > I'm not sure that's enough unfortunately. The fundamental issue here > > seems to be that input_dev_uevent follows a pointer that can be > > allocated by the driver, and will be free'd before the last call to > > input_dev_uevent. > > Yes, this is a fundamental property of C pointers - you should not free > them before exiting last code section that may reference them. For input > devices it means that pointers should not be freed until after > input_unregister_device() is called. > > I.e. you have sequence like this: > > driver_data = kzalloc(...); > driver_data->input_name = kstrdup(...); > driver_data->input_phys = kstrdup(...); > input = input_allocate_device(); > input->name = driver_data->input_name; > input->phys = driver_data->input_phys; > input_register_device(input); > ... > > input_unregister_device(input); > kfree(driver_data->input_name); > kfree(driver_data->input_phys); > kfree(driver_data); > > devm typically helps with resources being freed at the right time, but > for that the managed resources should be attached to the *correct > device*, with correct device being one the driver is binding to, not any > random device structure. Thanks, I see what you mean now. It looks good to me indeed :) Maxime
Attachment:
signature.asc
Description: PGP signature
