From: "dh10.jung" <dh10.jung@samsung.com>

When DWC3_EP_DELAY_STOP defers ENDXFER, giveback of -ESHUTDOWN requests
arrives after gserial_disconnect(). These stale requests then accumulate
in read_pool via gs_rx_push(). After gadget exit frees the old endpoint,
req->dep becomes a dangling pointer. On the next gserial_connect(),
gs_start_rx() queues these stale requests onto a new endpoint, triggering:

  WARNING: CPU: 0 at drivers/usb/dwc3/gadget.c:1990
  request 00000000e6e350a5 belongs to ''

  Call trace:
   dwc3_gadget_ep_queue+0x158/0x1e8
   usb_ep_queue+0x60/0xe8
   gs_start_rx+0xa4/0x128
   gs_start_io+0x128/0x254
   gserial_connect+0xb4/0x14c
   acm_set_alt+0xa0/0x114
   set_config+0x22c/0x384
   composite_setup+0x37c/0xc7c
   configfs_composite_setup+0x5c/0x88
   dwc3_ep0_interrupt+0x6c8/0xbcc
   dwc3_thread_interrupt+0xa0/0x1284
   irq_thread_fn+0x48/0xa8
   irq_thread+0x150/0x31c
   kthread+0x150/0x27c
   ret_from_fork+0x10/0x20

Fix by discarding -ESHUTDOWN requests immediately in gs_read_complete()
while ep is still valid, preventing stale reqs from entering read_pool.

Fixes: 937ef73d5075 ("USB: serial gadget: rx path data loss fixes")
Signed-off-by: dh10.jung <dh10.jung@samsung.com>
---
 drivers/usb/gadget/function/u_serial.c | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

diff --git a/drivers/usb/gadget/function/u_serial.c b/drivers/usb/gadget/function/u_serial.c
index cdd1dfc666c4..a8f5b0a6df51 100644
--- a/drivers/usb/gadget/function/u_serial.c
+++ b/drivers/usb/gadget/function/u_serial.c
@@ -459,10 +459,22 @@ static void gs_read_complete(struct usb_ep *ep, struct usb_request *req)
 {
 	struct gs_port	*port = ep->driver_data;
 
-	/* Queue all received data until the tty layer is ready for it. */
 	spin_lock(&port->port_lock);
-	list_add_tail(&req->list, &port->read_queue);
-	schedule_delayed_work(&port->push, 0);
+	if (req->status == -ESHUTDOWN) {
+		/*
+		 * Discard shutdown completions here while ep is still valid.
+		 * Returning them to read_pool after gserial_disconnect() has
+		 * reset the counters causes stale reqs (with req->dep pointing
+		 * to the old ep) to be queued onto a new ep at reconnect time,
+		 * triggering a req->dep != dep WARN and kernel panic.
+		 */
+		gs_free_req(ep, req);
+		port->read_started--;
+	} else {
+		/* Queue all received data until the tty layer is ready for it. */
+		list_add_tail(&req->list, &port->read_queue);
+		schedule_delayed_work(&port->push, 0);
+	}
 	spin_unlock(&port->port_lock);
 }
 
-- 
2.34.1

On Tue, Jul 21, 2026 at 10:35:09AM +0900, Daehwan Jung wrote:
> From: "dh10.jung" <dh10.jung@samsung.com>

We need a name, not an email alias.  Same for the changelog, your
internal samsung review should have caught this?

thanks,

greg k-h

On Tue, Jul 21, 2026 at 07:35:26AM +0200, Greg Kroah-Hartman wrote:
> On Tue, Jul 21, 2026 at 10:35:09AM +0900, Daehwan Jung wrote:
> > From: "dh10.jung" <dh10.jung@samsung.com>
> 
> We need a name, not an email alias.  Same for the changelog, your
> internal samsung review should have caught this?
> 
> thanks,
> 
> greg k-h
> 

I apologize for missing it. I'll correct and resend it.

Best Regards,
Daehwan Jung