[Linux-kernel-mentees] [PATCH v1] usbhid: Fix slab-out-of-bounds write in hiddev_ioctl_usage()

Peilin Ye yepeilin.cs at gmail.com
Mon Jul 20 19:05:08 UTC 2020

On Mon, Jul 20, 2020 at 03:12:57PM +0300, Dan Carpenter wrote:
> So another option would be to just add HIDIOCGUSAGE and HIDIOCSUSAGE to
> the earlier check.  That risks breaking userspace.  Another option is to
> just add a check like you did earlier to the HIDIOCGUSAGE case.
> Probably just do option #2 and resend.

Sure, I will just add the same check to the HIDIOCGUSAGE case for the
time being. Thank you for the detailed explanation.

Here's what I found after digging a bit further though:

hid_parser_main() calls different functions in order to process
different type of items:


	static int (*dispatch_type[])(struct hid_parser *parser,
				      struct hid_item *item) = {

In this case, hid_parser_main() calls hid_add_field(), which in turn
calls hid_register_field(), which allocates the `field` object as you


	field = kzalloc((sizeof(struct hid_field) +
			 usages * sizeof(struct hid_usage) +
			 values * sizeof(unsigned)), GFP_KERNEL);

Here, `values` equals to `global.report_count`. See how it is being


	field = hid_register_field(report, usages, parser->global.report_count);

In hid_parser_main(), `global.report_count` can be set by calling

However, the syzkaller reproducer made hid_parser_main() to call
hid_parser_global() __before__ `global.report_count` is properly set. It's
zero. So hid_register_field() allocated `field` with `values` equals to
zero - No room for value[] at all. I believe this caused the bug.

Apparently hid_parser_main() doesn't care about which item (main, local,
global and reserved) gets processed first. I am new to this code and I
don't know whether this is by design, but this arbitrarity is
apparently causing some issues.

As another example, in hid_add_field():


	report->size += parser->global.report_size * parser->global.report_count;

If `global.report_count` is zero, `report->size` gets increased by zero.
Is this working as intended? It seems weird to me.

Thank you,

Peilin Ye

More information about the Linux-kernel-mentees mailing list