CVE-2023-52478: Time-of-check Time-of-use (TOCTOU) Race Condition

In the Linux kernel, the following vulnerability has been resolved:

HID: logitech-hidpp: Fix kernel crash on receiver USB disconnect

hidpp_connect_event() has *four* time-of-check vs time-of-use (TOCTOU) races when it races with itself.

hidpp_connect_event() primarily runs from a workqueue but it also runs on probe() and if a "device-connected" packet is received by the hw when the thread running hidpp_connect_event() from probe() is waiting on the hw, then a second thread running hidpp_connect_event() will be started from the workqueue.

This opens the following races (note the below code is simplified):

1. Retrieving + printing the protocol (harmless race):

if (!hidpp->protocol_major) { hidpp_root_get_protocol_version() hidpp->protocol_major = response.rap.params[0]; }

We can actually see this race hit in the dmesg in the abrt output attached to rhbz#2227968:

[ 3064.624215] logitech-hidpp-device 0003:046D:4071.0049: HID++ 4.5 device connected. [ 3064.658184] logitech-hidpp-device 0003:046D:4071.0049: HID++ 4.5 device connected.

Testing with extra logging added has shown that after this the 2 threads take turn grabbing the hw access mutex (send_mutex) so they ping-pong through all the other TOCTOU cases managing to hit all of them:

2. Updating the name to the HIDPP name (harmless race):

if (hidpp->name == hdev->name) { ... hidpp->name = new_name; }

3. Initializing the power_supply class for the battery (problematic!):

hidpp_initialize_battery() { if (hidpp->battery.ps) return 0;

probe_battery(); /* Blocks, threads take turns executing this */

hidpp->battery.desc.properties = devm_kmemdup(dev, hidpp_battery_props, cnt, GFP_KERNEL);

hidpp->battery.ps = devm_power_supply_register(&hidpp->hid_dev->dev, &hidpp->battery.desc, cfg); }

4. Creating delayed input_device (potentially problematic):

if (hidpp->delayed_input) return;

hidpp->delayed_input = hidpp_allocate_input(hdev);

The really big problem here is 3. Hitting the race leads to the following sequence:

hidpp->battery.desc.properties = devm_kmemdup(dev, hidpp_battery_props, cnt, GFP_KERNEL);

hidpp->battery.ps = devm_power_supply_register(&hidpp->hid_dev->dev, &hidpp->battery.desc, cfg);

...

hidpp->battery.desc.properties = devm_kmemdup(dev, hidpp_battery_props, cnt, GFP_KERNEL);

hidpp->battery.ps = devm_power_supply_register(&hidpp->hid_dev->dev, &hidpp->battery.desc, cfg);

So now we have registered 2 power supplies for the same battery, which looks a bit weird from userspace's pov but this is not even the really big problem.

Notice how:

1. This is all devm-maganaged 2. The hidpp->battery.desc struct is shared between the 2 power supplies 3. hidpp->battery.desc.properties points to the result from the second devm_kmemdup()

This causes a use after free scenario on USB disconnect of the receiver: 1. The last registered power supply class device gets unregistered 2. The memory from the last devm_kmemdup() call gets freed, hidpp->battery.desc.properties now points to freed memory 3. The first registered power supply class device gets unregistered, this involves sending a remove uevent to userspace which invokes power_supply_uevent() to fill the uevent data 4. power_supply_uevent() uses hidpp->battery.desc.properties which now points to freed memory leading to backtraces like this one:

Sep 22 20:01:35 eric kernel: BUG: unable to handle page fault for address: ffffb2140e017f08 ... Sep 22 20:01:35 eric kernel: Workqueue: usb_hub_wq hub_event Sep 22 20:01:35 eric kernel: RIP: 0010:power_supply_uevent+0xee/0x1d0 ... Sep 22 20:01:35 eric kernel: ? asm_exc_page_fault+0x26/0x30 Sep 22 20:01:35 eric kernel: ? power_supply_uevent+0xee/0x1d0 Sep 22 20:01:35 eric kernel: ? power_supply_uevent+0x10d/0x1d0 Sep 22 20:01:35 eric kernel: dev_uevent+0x10f/0x2d0 Sep 22 20:01:35 eric kernel: kobject_uevent_env+0x291/0x680 Sep 22 20:01:35 eric kernel: ---truncated---