Since commit 54f32fd5dff9 ("HID: Make report_descriptor available for all devices"), the "report_descriptor" sysfs attribute is part of the device's attribute group, so it exists from device_add() until device_del(). It reads hdev->rdesc, but that buffer only lives from hid_open_report() until hid_close_report(), that is, for the duration of a driver binding. For a driver with a report_fixup() callback, rdesc is a private copy that hid_close_report() frees, so a read that races with the driver being unbound accesses freed memory. syzbot reports this as: BUG: KASAN: slab-use-after-free in report_descriptor_read+0xb5/0x100 Read of size 24 at addr ffff888078d61f00 by task fido_id/7281 ... Allocated by task 10: kmemdup_noprof+0x2b/0x70 hid_open_report+0x1fa/0xf00 lg_probe+0x29c/0x890 hid_device_probe+0x46a/0x870 ... Because hid_close_report() also clears rdesc and rsize, the same missing serialization can make the reader see a NULL rdesc with a non-zero rsize. Hold the device lock in report_descriptor_read(). The driver core already holds it across probe and remove, so a read can no longer overlap with them. Reported-by: syzbot+bc537ca7a0efe33988eb@syzkaller.appspotmail.com Tested-by: syzbot+bc537ca7a0efe33988eb@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=bc537ca7a0efe33988eb Fixes: 54f32fd5dff9 ("HID: Make report_descriptor available for all devices") Cc: stable@vger.kernel.org Signed-off-by: Krish Gulati --- drivers/hid/hid-core.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/drivers/hid/hid-core.c b/drivers/hid/hid-core.c index a3ff0514f9cd..824557a02544 100644 --- a/drivers/hid/hid-core.c +++ b/drivers/hid/hid-core.c @@ -2270,6 +2270,8 @@ report_descriptor_read(struct file *filp, struct kobject *kobj, struct device *dev = kobj_to_dev(kobj); struct hid_device *hdev = to_hid_device(dev); + guard(device)(dev); + if (off >= hdev->rsize) return 0; -- 2.56.0