]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
USB: dummy-hcd: Fix locking/synchronization error
authorAlan Stern <stern@rowland.harvard.edu>
Sun, 15 Mar 2026 18:30:43 +0000 (14:30 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 18 Mar 2026 15:16:29 +0000 (16:16 +0100)
Syzbot testing was able to provoke an addressing exception and crash
in the usb_gadget_udc_reset() routine in
drivers/usb/gadgets/udc/core.c, resulting from the fact that the
routine was called with a second ("driver") argument of NULL.  The bad
caller was set_link_state() in dummy_hcd.c, and the problem arose
because of a race between a USB reset and driver unbind.

These sorts of races were not supposed to be possible; commit
7dbd8f4cabd9 ("USB: dummy-hcd: Fix erroneous synchronization change"),
along with a few followup commits, was written specifically to prevent
them.  As it turns out, there are (at least) two errors remaining in
the code.  Another patch will address the second error; this one is
concerned with the first.

The error responsible for the syzbot crash occurred because the
stop_activity() routine will sometimes drop and then re-acquire the
dum->lock spinlock.  A call to stop_activity() occurs in
set_link_state() when handling an emulated USB reset, after the test
of dum->ints_enabled and before the increment of dum->callback_usage.
This allowed another thread (doing a driver unbind) to sneak in and
grab the spinlock, and then clear dum->ints_enabled and dum->driver.
Normally this other thread would have to wait for dum->callback_usage
to go down to 0 before it would clear dum->driver, but in this case it
didn't have to wait since dum->callback_usage had not yet been
incremented.

The fix is to increment dum->callback_usage _before_ calling
stop_activity() instead of after.  Then the thread doing the unbind
will not clear dum->driver until after the call to
usb_gadget_udc_reset() safely returns and dum->callback_usage has been
decremented again.

Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Reported-by: syzbot+19bed92c97bee999e5db@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/linux-usb/68fc7c9c.050a0220.346f24.023c.GAE@google.com/
Tested-by: syzbot+19bed92c97bee999e5db@syzkaller.appspotmail.com
Fixes: 7dbd8f4cabd9 ("USB: dummy-hcd: Fix erroneous synchronization change")
Cc: stable <stable@kernel.org>
Link: https://patch.msgid.link/46135f42-fdbe-46b5-aac0-6ca70492af15@rowland.harvard.edu
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/usb/gadget/udc/dummy_hcd.c

index c9eca90376e2b0e7986d263eb0836843c840dc1d..860ee7d3ef982fb9091df65a2e8c7fd0d5c1f5e6 100644 (file)
@@ -462,8 +462,13 @@ static void set_link_state(struct dummy_hcd *dum_hcd)
 
                /* Report reset and disconnect events to the driver */
                if (dum->ints_enabled && (disconnect || reset)) {
-                       stop_activity(dum);
                        ++dum->callback_usage;
+                       /*
+                        * stop_activity() can drop dum->lock, so it must
+                        * not come between the dum->ints_enabled test
+                        * and the ++dum->callback_usage.
+                        */
+                       stop_activity(dum);
                        spin_unlock(&dum->lock);
                        if (reset)
                                usb_gadget_udc_reset(&dum->gadget, dum->driver);