Re: [PATCH] usb: gadget: raw_gadget: fix dangling ep->desc on enable failure

[Liu Chao <liuc63@xiaopeng.com>]

On Tue, Oct 27, 2026 at 06:58:44AM +0100, Greg KH wrote:
> How was this found and tested?

By code audit of the raw-gadget endpoint ownership, not from a crash report
or a fuzzer.

Reaching the failing path from userspace is straightforward.
usb_gadget_ep_match_desc() does not look at bDescriptorType, but dummy_hcd's
dummy_enable() does and rejects anything that is not USB_DT_ENDPOINT with
-EINVAL. So a single USB_RAW_IOCTL_EP_ENABLE with an otherwise valid
descriptor (0x81, bulk, 512 bytes, i.e. ep1in-bulk at high speed) and
bDescriptorType = 0 makes usb_ep_enable() fail and drops us into out_free:

  raw-gadget.0 gadget.0: fail, usb_ep_enable returned -22

one line before kfree(desc), with ep->ep->desc still pointing at it.

Tested on mainline with CONFIG_USB_RAW_GADGET=y, CONFIG_USB_DUMMY_HCD=y and
KASAN, userspace side running in a QEMU guest: the descriptor gets freed and
ep->ep->desc keeps the stale value.

I could not find any in-tree reader of that pointer along this path, though.
usb_ep_enable() only reads a descriptor assigned right before it,
raw_process_ep_io() and the halt/wedge ioctls return early on the endpoint
state, and dummy_hcd keeps its own private descriptor pointer which
dummy_disable() clears, so the value left behind is only the core struct
usb_ep.desc. So a stock kernel stays silent under KASAN: as of today this is
a latent dangling reference, not a reachable crash. Reading the pointer from
a temporary probe patch (not part of the submission) right after the kfree()
gives:

  BUG: KASAN: slab-use-after-free in raw_ioctl
  Read of size 1 at 0xffff88800a7425a0 by task repro/71
  ...
  Allocated by task 71:
    memdup_user+0x26/0x80
    raw_ioctl+0xbfb/0x1ea0
  Freed by task 71:
    kfree+0x189/0x3b0
    raw_ioctl+0x1648/0x1ea0

and with the patch applied there is nothing left to read.

What makes it worth fixing anyway is that the stale pointer outlives the
file descriptor: a UDC's struct usb_ep is owned by the UDC, not by
raw_gadget, so the dangling value survives dev_free() and is still there for
whichever gadget driver binds that UDC next. The core keeps the opposite
invariant elsewhere -- usb_ep_autoconfig_ss() sets ep->desc = NULL,
composite.c assigns _ep->desc before enabling, and f_fs reads ep->ep->desc.

thanks,