-
Alan Stern authored
A recent commit expanding the scope of the udc_lock mutex in the gadget core managed to cause an obscure and slightly bizarre lockdep violation. In abbreviated form: ====================================================== WARNING: possible circular locking dependency detected 5.19.0-rc7+ #12510 Not tainted ------------------------------------------------------ udevadm/312 is trying to acquire lock: ffff80000aae1058 (udc_lock){+.+.}-{3:3}, at: usb_udc_uevent+0x54/0xe0 but task is already holding lock: ffff000002277548 (kn->active#4){++++}-{0:0}, at: kernfs_seq_start+0x34/0xe0 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #3 (kn->active#4){++++}-{0:0}: lock_acquire+0x68/0x84 __kernfs_remove+0x268/0x380 kernfs_remove_by_name_ns+0x58/0xac sysfs_remove_file_ns+0x18/0x24 device_del+0x15c/0x440 -> #2 (device_links_lock){+.+.}-{3:3}: lock_acquire+0x68/0x84 __mutex_lock+0x9c/0x430 mutex_lock_nested+0x38/0x64 device_link_remove+0x3c/0xa0 _regulator_put.part.0+0x168/0x190 regulator_put+0x3c/0x54 devm_regulator_release+0x14/0x20 -> #1 (regulator_list_mutex){+.+.}-{3:3}: lock_acquire+0x68/0x84 __mutex_lock+0x9c/0x430 mutex_lock_nested+0x38/0x64 regulator_lock_dependent+0x54/0x284 regulator_enable+0x34/0x80 phy_power_on+0x24/0x130 __dwc2_lowlevel_hw_enable+0x100/0x130 dwc2_lowlevel_hw_enable+0x18/0x40 dwc2_hsotg_udc_start+0x6c/0x2f0 gadget_bind_driver+0x124/0x1f4 -> #0 (udc_lock){+.+.}-{3:3}: __lock_acquire+0x1298/0x20cc lock_acquire.part.0+0xe0/0x230 lock_acquire+0x68/0x84 __mutex_lock+0x9c/0x430 mutex_lock_nested+0x38/0x64 usb_udc_uevent+0x54/0xe0 Evidently this was caused by the scope of udc_mutex being too large. The mutex is only meant to protect udc->driver along with a few other things. As far as I can tell, there's no reason for the mutex to be held while the gadget core calls a gadget driver's ->bind or ->unbind routine, or while a UDC is being started or stopped. (This accounts for link #1 in the chain above, where the mutex is held while the dwc2_hsotg_udc is started as part of driver probing.) Gadget drivers' ->disconnect callbacks are problematic. Even though usb_gadget_disconnect() will now acquire the udc_mutex, there's a window in usb_gadget_bind_driver() between the times when the mutex is released and the ->bind callback is invoked. If a disconnect occurred during that window, we could call the driver's ->disconnect routine before its ->bind routine. To prevent this from happening, it will be necessary to prevent a UDC from connecting while it has no gadget driver. This should be done already but it doesn't seem to be; currently usb_gadget_connect() has no check for this. Such a check will have to be added later. Some degree of mutual exclusion is required in soft_connect_store(), which can dereference udc->driver at arbitrary times since it is a sysfs callback. The solution here is to acquire the gadget's device lock rather than the udc_mutex. Since the driver core guarantees that the device lock is always held during driver binding and unbinding, this will make the accesses in soft_connect_store() mutually exclusive with any changes to udc->driver. Lastly, it turns out there is one place which should hold the udc_mutex but currently does not: The function_show() routine needs protection while it dereferences udc->driver. The missing lock and unlock calls are added. Link: https://lore.kernel.org/all/b2ba4245-9917-e399-94c8-03a383e7070e@samsung.com/ Fixes: 2191c008 ("USB: gadget: Fix use-after-free Read in usb_udc_uevent()") Cc: Felipe Balbi <balbi@kernel.org> Cc: stable@vger.kernel.org Reported-by:Marek Szyprowski <m.szyprowski@samsung.com> Tested-by:
Alan Stern <stern@rowland.harvard.edu> Link: https://lore.kernel.org/r/YwkfhdxA/I2nOcK7@rowland.harvard.eduSigned-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
1016fc0c
