usb: dwc3: gadget: check if dep->frame_number is still valid

Quoted 51 lines

Quoted 41 lines

Quoted 28 lines

drivers/usb/dwc3/gadget.c +10 -0

index d8c7ad0c22e8..00fe01a01977 100644
@@ -1268,12 +1268,22 @@ static int __dwc3_gadget_get_frame(struct dwc3 *dwc)
 
 static int __dwc3_gadget_start_isoc(struct dwc3_ep *dep)
 {
+	u64 current_timestamp;
+	u64 diff_timestamp;
+	u32 elapsed_frames;
+
 	if (list_empty(&dep->pending_list)) {
 		dep->flags |= DWC3_EP_PENDING_REQUEST;
 		return -EAGAIN;
 	}
 
+	current_timestamp = ktime_get_ns();
+	diff_timestamp = current_timestamp - dep->frame_timestamp;
+	elapsed_frames = DIV_ROUND_UP_ULL(diff_timestamp, 125000);
+
+	dep->frame_number += elapsed_frames;
 	dep->frame_number = DWC3_ALIGN_FRAME(dep);
+
 	return __dwc3_gadget_kick_transfer(dep);
 }
 

drivers/usb/dwc3/gadget.c static void dwc3_gadget_endpoint_frame_from_event(struct dwc3_ep *dep, · line 2336

Quoted 7 lines

drivers/usb/dwc3/gadget.c +1 -0

@@ -2320,6 +2330,7 @@ static void dwc3_gadget_endpoint_frame_from_event(struct dwc3_ep *dep,
 		const struct dwc3_event_depevt *event)
 {
 	dep->frame_number = event->parameters;
+	dep->frame_timestamp = ktime_get_ns();
 }
 
 static void dwc3_gadget_endpoint_transfer_in_progress(struct dwc3_ep *dep,

This may not be enough. The dep->frame_timestamp may not correspond to the frame_number from XferNotReady event. When there's system latency (which is possible when this failure happens), the time the driver handle the event may be a few uframes passed the time the controller's XferInProgress uframe parameter.

Quoted 3 lines

Rather than starting the isoc transfer immediately on the next interval. How about starting the transfer with some minimum buffer uframes just like before? (e.g. frame_number + max(4, interval))

The problem with this is cases with interval of 1ms. This will result in a 4ms delay. I really want to start transfer as soon as possible and the timestamp trick seems to be the best idea so far, without resorting to 4 intervals delay. We do, however, have the possibility that this will start 2 intervals in the future because of the usage of DIV_ROUND_UP_ULL() and because of how DWC3_ALIGN_FRAME() is implemented.

Quoted 3 lines

I understand what you're saying, though, but it seems like we don't have to avoid that case completely. We can only make it less likely.

In the case of interval of 1ms, it will start on the next interval. frame_number + max(4, interval) will start at least 4 uframes in the future.

In any case, what about immediately retry the START_TRANSFER command with a new frame_number + (interval*retry) should it fail with bus-expiry? You can set the number of retries to maybe 5 times. This should remove the need to do time stamping.