Skip to contentSkip to conversation

usb: xhci: Prefer endpoint context dequeue pointer over stopped_trb

PATCH In review Julius Werner 4 participants 7 messages Feb 21, 2014 – Apr 15, 2014

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 usb: xhci: Prefer endpoint context dequeue pointer over stopped_trb

We have observed a rare cycle state desync bug after Set TR Dequeue Pointer commands on Intel LynxPoint xHCs (resulting in an endpoint that doesn't fetch new TRBs and thus an unresponsive USB device). It always triggers when a previous Set TR Dequeue Pointer command has set the pointer to the final Link TRB of a segment, and then another URB gets enqueued and cancelled again before it can be completed. Further investigation showed that the xHC had returned the Link TRB in the TRB Pointer field of the Transfer Event (CC == Stopped -- Length Invalid), but when xhci_find_new_dequeue_state() later accesses the Endpoint Context's TR Dequeue Pointer field it is set to the first TRB of the next segment.

The driver expects those two values to be the same in this situation, and uses the cycle state of the latter together with the address of the former. This should be fine according to the XHCI specification, since the endpoint ring should be stopped when returning the Transfer Event and thus should not advance over the Link TRB before it gets restarted. However, real-world XHCI implementations apparently don't really care that much about these details, so the driver should follow a more defensive approach to try to work around HC spec violations.

This patch removes the stopped_trb variable that had been used to store the TRB Pointer from the last Transfer Event of a stopped TRB. Instead, xhci_find_new_dequeue_state() now relies only on the Endpoint Context, requiring a small amount of additional processing to find the virtual address corresponding to the TR Dequeue Pointer. Some other parts of the function were slightly rearranged to better fit into this model.

This patch should be backported to kernels as old as 2.6.31 that contain the commit ae636747146ea97efa18e04576acd3416e2514f5 "USB: xhci: URB cancellation support."

---
 drivers/usb/host/xhci-ring.c | 66 ++++++++++++++++++++------------------------
 drivers/usb/host/xhci.c      |  1 -
 drivers/usb/host/xhci.h      |  2 --
 3 files changed, 30 insertions(+), 39 deletions(-)
drivers/usb/host/xhci-ring.c +30 -36
index a0b248c..b8277c7 100644
@@ -549,6 +549,7 @@ void xhci_find_new_dequeue_state(struct xhci_hcd *xhci,
 	struct xhci_generic_trb *trb;
 	struct xhci_ep_ctx *ep_ctx;
 	dma_addr_t addr;
+	u64 hw_dequeue;
 
 	ep_ring = xhci_triad_to_transfer_ring(xhci, slot_id,
 			ep_index, stream_id);
@@ -558,56 +559,56 @@ void xhci_find_new_dequeue_state(struct xhci_hcd *xhci,
 				stream_id);
 		return;
 	}
-	state->new_cycle_state = 0;
-	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
-			"Finding segment containing stopped TRB.");
-	state->new_deq_seg = find_trb_seg(cur_td->start_seg,
-			dev->eps[ep_index].stopped_trb,
-			&state->new_cycle_state);
-	if (!state->new_deq_seg) {
-		WARN_ON(1);
-		return;
-	}
 
 	/* Dig out the cycle state saved by the xHC during the stop ep cmd */
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"Finding endpoint context");
 	ep_ctx = xhci_get_ep_ctx(xhci, dev->out_ctx, ep_index);
-	state->new_cycle_state = 0x1 & le64_to_cpu(ep_ctx->deq);
+	hw_dequeue = le64_to_cpu(ep_ctx->deq);
+
+	/* Find virtual address and segment of hardware dequeue pointer */
+	state->new_deq_seg = ep_ring->deq_seg;
+	state->new_deq_ptr = ep_ring->dequeue;
+	while (xhci_trb_virt_to_dma(state->new_deq_seg, state->new_deq_ptr)
+			!= (dma_addr_t)(hw_dequeue & ~0x1)) {
+		next_trb(xhci, ep_ring, &state->new_deq_seg,
+					&state->new_deq_ptr);
+		if (state->new_deq_ptr == ep_ring->dequeue) {
+			WARN_ON(1);
+			return;
+		}
+	}
 
+	/*
+	 * Find cycle state for last_trb, starting at old cycle state of
+	 * hw_dequeue. If there is only one segment ring, find_trb_seg() will
+	 * return immediately and cannot toggle the cycle state if this search
+	 * wraps around, so add one more toggle manually in that case.
+	 */
+	state->new_cycle_state = hw_dequeue & 0x1;
+	if (ep_ring->first_seg == ep_ring->first_seg->next &&
+			cur_td->last_trb < state->new_deq_ptr)
+		state->new_cycle_state ^= 0x1;
 	state->new_deq_ptr = cur_td->last_trb;
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"Finding segment containing last TRB in TD.");
 	state->new_deq_seg = find_trb_seg(state->new_deq_seg,
-			state->new_deq_ptr,
-			&state->new_cycle_state);
+			state->new_deq_ptr, &state->new_cycle_state);
 	if (!state->new_deq_seg) {
 		WARN_ON(1);
 		return;
 	}
 
+	/* Increment to find next TRB after last_trb. Cycle if appropriate. */
 	trb = &state->new_deq_ptr->generic;
 	if (TRB_TYPE_LINK_LE32(trb->field[3]) &&
 	    (trb->field[3] & cpu_to_le32(LINK_TOGGLE)))
 		state->new_cycle_state ^= 0x1;
 	next_trb(xhci, ep_ring, &state->new_deq_seg, &state->new_deq_ptr);
 
-	/*
-	 * If there is only one segment in a ring, find_trb_seg()'s while loop
-	 * will not run, and it will return before it has a chance to see if it
-	 * needs to toggle the cycle bit.  It can't tell if the stalled transfer
-	 * ended just before the link TRB on a one-segment ring, or if the TD
-	 * wrapped around the top of the ring, because it doesn't have the TD in
-	 * question.  Look for the one-segment case where stalled TRB's address
-	 * is greater than the new dequeue pointer address.
-	 */
-	if (ep_ring->first_seg == ep_ring->first_seg->next &&
-			state->new_deq_ptr < dev->eps[ep_index].stopped_trb)
-		state->new_cycle_state ^= 0x1;
+	/* Don't update the ring cycle state for the producer (us). */
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"Cycle state = 0x%x", state->new_cycle_state);
-
-	/* Don't update the ring cycle state for the producer (us). */
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"New dequeue segment = %p (virtual)",
 			state->new_deq_seg);
@@ -790,7 +791,6 @@ static void xhci_handle_cmd_stop_ep(struct xhci_hcd *xhci, int slot_id,
 	if (list_empty(&ep->cancelled_td_list)) {
 		xhci_stop_watchdog_timer_in_irq(xhci, ep);
 		ep->stopped_td = NULL;
-		ep->stopped_trb = NULL;
 		ring_doorbell_for_active_rings(xhci, slot_id, ep_index);
 		return;
 	}
@@ -858,11 +858,9 @@ remove_finished_td:
 		ring_doorbell_for_active_rings(xhci, slot_id, ep_index);
 	}
 
-	/* Clear stopped_td and stopped_trb if endpoint is not halted */
-	if (!(ep->ep_state & EP_HALTED)) {
+	/* Clear stopped_td if endpoint is not halted */
+	if (!(ep->ep_state & EP_HALTED))
 		ep->stopped_td = NULL;
-		ep->stopped_trb = NULL;
-	}
 
 	/*
 	 * Drop the lock and complete the URBs in the cancelled TD list.
@@ -1907,14 +1905,12 @@ static void xhci_cleanup_halted_endpoint(struct xhci_hcd *xhci,
 	struct xhci_virt_ep *ep = &xhci->devs[slot_id]->eps[ep_index];
 	ep->ep_state |= EP_HALTED;
 	ep->stopped_td = td;
-	ep->stopped_trb = event_trb;
 	ep->stopped_stream = stream_id;
 
 	xhci_queue_reset_ep(xhci, slot_id, ep_index);
 	xhci_cleanup_stalled_ring(xhci, td->urb->dev, ep_index);
 
 	ep->stopped_td = NULL;
-	ep->stopped_trb = NULL;
 	ep->stopped_stream = 0;
 
 	xhci_ring_cmd_db(xhci);
@@ -1996,7 +1992,6 @@ static int finish_td(struct xhci_hcd *xhci, struct xhci_td *td,
 		 * the ring dequeue pointer or take this TD off any lists yet.
 		 */
 		ep->stopped_td = td;
-		ep->stopped_trb = event_trb;
 		return 0;
 	} else {
 		if (trb_comp_code == COMP_STALL) {
@@ -2008,7 +2003,6 @@ static int finish_td(struct xhci_hcd *xhci, struct xhci_td *td,
 			 * USB class driver clear the stall later.
 			 */
 			ep->stopped_td = td;
-			ep->stopped_trb = event_trb;
 			ep->stopped_stream = ep_ring->stream_id;
 		} else if (xhci_requires_manual_halt_cleanup(xhci,
 					ep_ctx, trb_comp_code)) {
drivers/usb/host/xhci.c +0 -1
index ad36439..dba5381 100644
@@ -2933,7 +2933,6 @@ void xhci_endpoint_reset(struct usb_hcd *hcd,
 		xhci_ring_cmd_db(xhci);
 	}
 	virt_ep->stopped_td = NULL;
-	virt_ep->stopped_trb = NULL;
 	virt_ep->stopped_stream = 0;
 	spin_unlock_irqrestore(&xhci->lock, flags);
 
drivers/usb/host/xhci.h +0 -2
index f8416639..cf2c5f0 100644
@@ -875,8 +875,6 @@ struct xhci_virt_ep {
 #define EP_GETTING_NO_STREAMS	(1 << 5)
 	/* ----  Related to URB cancellation ---- */
 	struct list_head	cancelled_td_list;
-	/* The TRB that was last reported in a stopped endpoint ring */
-	union xhci_trb		*stopped_trb;
 	struct xhci_td		*stopped_td;
 	unsigned int		stopped_stream;
 	/* Watchdog timer for stop endpoint command to cancel URBs */
Signature · 1 line
1.8.3.2
#1 lore raw
Quoted 12 lines · Julius Werner

We have observed a rare cycle state desync bug after Set TR Dequeue Pointer commands on Intel LynxPoint xHCs (resulting in an endpoint that doesn't fetch new TRBs and thus an unresponsive USB device). It always triggers when a previous Set TR Dequeue Pointer command has set the pointer to the final Link TRB of a segment, and then another URB gets enqueued and cancelled again before it can be completed. Further investigation showed that the xHC had returned the Link TRB in the TRB Pointer field of the Transfer Event (CC == Stopped -- Length Invalid), but when xhci_find_new_dequeue_state() later accesses the Endpoint Context's TR Dequeue Pointer field it is set to the first TRB of the next segment.

Quoted 8 lines

The driver expects those two values to be the same in this situation, and uses the cycle state of the latter together with the address of the former. This should be fine according to the XHCI specification, since the endpoint ring should be stopped when returning the Transfer Event and thus should not advance over the Link TRB before it gets restarted. However, real-world XHCI implementations apparently don't really care that much about these details, so the driver should follow a more defensive approach to try to work around HC spec violations.

Length Invalid can actually be returned when the host is "in between" processing TRBs. Perhaps it had processed the link TRB, but hadn't started processing the TRB at the top of the ring. So it's not a spec violation per say, but definitely a spec ambiguity.

The patch looks fine. Mathias is taking over for xHCI driver maintainership in 3.15. He's currently handling queuing bug fix patches for 3.14 while I finish queueing feature patches for 3.15. Mathias, will you test and queue this up for 3.14?

Quoted 199 lines

This patch removes the stopped_trb variable that had been used to store the TRB Pointer from the last Transfer Event of a stopped TRB. Instead, xhci_find_new_dequeue_state() now relies only on the Endpoint Context, requiring a small amount of additional processing to find the virtual address corresponding to the TR Dequeue Pointer. Some other parts of the function were slightly rearranged to better fit into this model.

This patch should be backported to kernels as old as 2.6.31 that contain the commit ae636747146ea97efa18e04576acd3416e2514f5 "USB: xhci: URB cancellation support."

---
 drivers/usb/host/xhci-ring.c | 66 ++++++++++++++++++++------------------------
 drivers/usb/host/xhci.c      |  1 -
 drivers/usb/host/xhci.h      |  2 --
 3 files changed, 30 insertions(+), 39 deletions(-)
drivers/usb/host/xhci-ring.c +30 -36
index a0b248c..b8277c7 100644
@@ -549,6 +549,7 @@ void xhci_find_new_dequeue_state(struct xhci_hcd *xhci,
 	struct xhci_generic_trb *trb;
 	struct xhci_ep_ctx *ep_ctx;
 	dma_addr_t addr;
+	u64 hw_dequeue;
 
 	ep_ring = xhci_triad_to_transfer_ring(xhci, slot_id,
 			ep_index, stream_id);
@@ -558,56 +559,56 @@ void xhci_find_new_dequeue_state(struct xhci_hcd *xhci,
 				stream_id);
 		return;
 	}
-	state->new_cycle_state = 0;
-	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
-			"Finding segment containing stopped TRB.");
-	state->new_deq_seg = find_trb_seg(cur_td->start_seg,
-			dev->eps[ep_index].stopped_trb,
-			&state->new_cycle_state);
-	if (!state->new_deq_seg) {
-		WARN_ON(1);
-		return;
-	}
 
 	/* Dig out the cycle state saved by the xHC during the stop ep cmd */
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"Finding endpoint context");
 	ep_ctx = xhci_get_ep_ctx(xhci, dev->out_ctx, ep_index);
-	state->new_cycle_state = 0x1 & le64_to_cpu(ep_ctx->deq);
+	hw_dequeue = le64_to_cpu(ep_ctx->deq);
+
+	/* Find virtual address and segment of hardware dequeue pointer */
+	state->new_deq_seg = ep_ring->deq_seg;
+	state->new_deq_ptr = ep_ring->dequeue;
+	while (xhci_trb_virt_to_dma(state->new_deq_seg, state->new_deq_ptr)
+			!= (dma_addr_t)(hw_dequeue & ~0x1)) {
+		next_trb(xhci, ep_ring, &state->new_deq_seg,
+					&state->new_deq_ptr);
+		if (state->new_deq_ptr == ep_ring->dequeue) {
+			WARN_ON(1);
+			return;
+		}
+	}
 
+	/*
+	 * Find cycle state for last_trb, starting at old cycle state of
+	 * hw_dequeue. If there is only one segment ring, find_trb_seg() will
+	 * return immediately and cannot toggle the cycle state if this search
+	 * wraps around, so add one more toggle manually in that case.
+	 */
+	state->new_cycle_state = hw_dequeue & 0x1;
+	if (ep_ring->first_seg == ep_ring->first_seg->next &&
+			cur_td->last_trb < state->new_deq_ptr)
+		state->new_cycle_state ^= 0x1;
 	state->new_deq_ptr = cur_td->last_trb;
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"Finding segment containing last TRB in TD.");
 	state->new_deq_seg = find_trb_seg(state->new_deq_seg,
-			state->new_deq_ptr,
-			&state->new_cycle_state);
+			state->new_deq_ptr, &state->new_cycle_state);
 	if (!state->new_deq_seg) {
 		WARN_ON(1);
 		return;
 	}
 
+	/* Increment to find next TRB after last_trb. Cycle if appropriate. */
 	trb = &state->new_deq_ptr->generic;
 	if (TRB_TYPE_LINK_LE32(trb->field[3]) &&
 	    (trb->field[3] & cpu_to_le32(LINK_TOGGLE)))
 		state->new_cycle_state ^= 0x1;
 	next_trb(xhci, ep_ring, &state->new_deq_seg, &state->new_deq_ptr);
 
-	/*
-	 * If there is only one segment in a ring, find_trb_seg()'s while loop
-	 * will not run, and it will return before it has a chance to see if it
-	 * needs to toggle the cycle bit.  It can't tell if the stalled transfer
-	 * ended just before the link TRB on a one-segment ring, or if the TD
-	 * wrapped around the top of the ring, because it doesn't have the TD in
-	 * question.  Look for the one-segment case where stalled TRB's address
-	 * is greater than the new dequeue pointer address.
-	 */
-	if (ep_ring->first_seg == ep_ring->first_seg->next &&
-			state->new_deq_ptr < dev->eps[ep_index].stopped_trb)
-		state->new_cycle_state ^= 0x1;
+	/* Don't update the ring cycle state for the producer (us). */
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"Cycle state = 0x%x", state->new_cycle_state);
-
-	/* Don't update the ring cycle state for the producer (us). */
 	xhci_dbg_trace(xhci, trace_xhci_dbg_cancel_urb,
 			"New dequeue segment = %p (virtual)",
 			state->new_deq_seg);
@@ -790,7 +791,6 @@ static void xhci_handle_cmd_stop_ep(struct xhci_hcd *xhci, int slot_id,
 	if (list_empty(&ep->cancelled_td_list)) {
 		xhci_stop_watchdog_timer_in_irq(xhci, ep);
 		ep->stopped_td = NULL;
-		ep->stopped_trb = NULL;
 		ring_doorbell_for_active_rings(xhci, slot_id, ep_index);
 		return;
 	}
@@ -858,11 +858,9 @@ remove_finished_td:
 		ring_doorbell_for_active_rings(xhci, slot_id, ep_index);
 	}
 
-	/* Clear stopped_td and stopped_trb if endpoint is not halted */
-	if (!(ep->ep_state & EP_HALTED)) {
+	/* Clear stopped_td if endpoint is not halted */
+	if (!(ep->ep_state & EP_HALTED))
 		ep->stopped_td = NULL;
-		ep->stopped_trb = NULL;
-	}
 
 	/*
 	 * Drop the lock and complete the URBs in the cancelled TD list.
@@ -1907,14 +1905,12 @@ static void xhci_cleanup_halted_endpoint(struct xhci_hcd *xhci,
 	struct xhci_virt_ep *ep = &xhci->devs[slot_id]->eps[ep_index];
 	ep->ep_state |= EP_HALTED;
 	ep->stopped_td = td;
-	ep->stopped_trb = event_trb;
 	ep->stopped_stream = stream_id;
 
 	xhci_queue_reset_ep(xhci, slot_id, ep_index);
 	xhci_cleanup_stalled_ring(xhci, td->urb->dev, ep_index);
 
 	ep->stopped_td = NULL;
-	ep->stopped_trb = NULL;
 	ep->stopped_stream = 0;
 
 	xhci_ring_cmd_db(xhci);
@@ -1996,7 +1992,6 @@ static int finish_td(struct xhci_hcd *xhci, struct xhci_td *td,
 		 * the ring dequeue pointer or take this TD off any lists yet.
 		 */
 		ep->stopped_td = td;
-		ep->stopped_trb = event_trb;
 		return 0;
 	} else {
 		if (trb_comp_code == COMP_STALL) {
@@ -2008,7 +2003,6 @@ static int finish_td(struct xhci_hcd *xhci, struct xhci_td *td,
 			 * USB class driver clear the stall later.
 			 */
 			ep->stopped_td = td;
-			ep->stopped_trb = event_trb;
 			ep->stopped_stream = ep_ring->stream_id;
 		} else if (xhci_requires_manual_halt_cleanup(xhci,
 					ep_ctx, trb_comp_code)) {
drivers/usb/host/xhci.c +0 -1
index ad36439..dba5381 100644
@@ -2933,7 +2933,6 @@ void xhci_endpoint_reset(struct usb_hcd *hcd,
 		xhci_ring_cmd_db(xhci);
 	}
 	virt_ep->stopped_td = NULL;
-	virt_ep->stopped_trb = NULL;
 	virt_ep->stopped_stream = 0;
 	spin_unlock_irqrestore(&xhci->lock, flags);
 
drivers/usb/host/xhci.h +0 -2
index f8416639..cf2c5f0 100644
@@ -875,8 +875,6 @@ struct xhci_virt_ep {
 #define EP_GETTING_NO_STREAMS	(1 << 5)
 	/* ----  Related to URB cancellation ---- */
 	struct list_head	cancelled_td_list;
-	/* The TRB that was last reported in a stopped endpoint ring */
-	union xhci_trb		*stopped_trb;
 	struct xhci_td		*stopped_td;
 	unsigned int		stopped_stream;
 	/* Watchdog timer for stop endpoint command to cancel URBs */
Signature · 6 lines
1.8.3.2

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#2 lore raw

Hi Mathias,

Quoted 4 lines

The patch looks fine. Mathias is taking over for xHCI driver maintainership in 3.15. He's currently handling queuing bug fix patches for 3.14 while I finish queueing feature patches for 3.15. Mathias, will you test and queue this up for 3.14?

Did you pick this patch up anywhere yet or are there still issues with it? I just want to make sure it doesn't slip through the cracks. (Maybe I just didn't see it yet... are you still queueing patches in sarah/xhci.git or do you have your own repository somewhere?)

#3 lore raw

Hi

Quoted 11 lines · Julius Werner

Hi Mathias,

Quoted 4 lines

The patch looks fine. Mathias is taking over for xHCI driver maintainership in 3.15. He's currently handling queuing bug fix patches for 3.14 while I finish queueing feature patches for 3.15. Mathias, will you test and queue this up for 3.14?

Did you pick this patch up anywhere yet or are there still issues with it? I just want to make sure it doesn't slip through the cracks. (Maybe I just didn't see it yet... are you still queueing patches in sarah/xhci.git or do you have your own repository somewhere?)

Patch applies fine and I'll send it forward to Greg once 3.15-rc1 is out.

I'm not using Sarah's tree, but setting up my own tree on kernel.org (Just got access). I got a temporary one on github if you want to take a look, but it's a short term thing and it may change without warning.

[email protected]:matnyman/xhci.git

-Mathias

#4 lore raw
In reply to Julius Werner
Quoted 7 lines · Julius Werner

Hi Mathias,

Quoted 4 lines

The patch looks fine. Mathias is taking over for xHCI driver maintainership in 3.15. He's currently handling queuing bug fix patches for 3.14 while I finish queueing feature patches for 3.15. Mathias, will you test and queue this up for 3.14?

Quoted 9 lines

Did you pick this patch up anywhere yet or are there still issues with it? I just want to make sure it doesn't slip through the cracks. (Maybe I just didn't see it yet... are you still queueing patches in sarah/xhci.git or do you have your own repository somewhere?) --
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

I tried to apply this patch on top of 3.15-rc1, but it fails because of the streams support added to xhci_find_new_dequeue_state()

After some manual editing the interesting parts of xhci_find_new_dequeue_state() looks like this:

@@ -577,46 +568,57 @@ void xhci_find_new_dequeue_state(struct xhci_hcd *xhci,

        if (ep->ep_state & EP_HAS_STREAMS) {
                struct xhci_stream_ctx *ctx =
                        &ep->stream_info->stream_ctx_array[stream_id];
-               state->new_cycle_state = 0x1 & le64_to_cpu(ctx->stream_ring);
+               hw_dequeue = le64_to_cpu(ctx->stream_ring);
        } else {
                struct xhci_ep_ctx *ep_ctx
                        = xhci_get_ep_ctx(xhci, dev->out_ctx, ep_index);
-               state->new_cycle_state = 0x1 & le64_to_cpu(ep_ctx->deq);
+               hw_dequeue = le64_to_cpu(ep_ctx->deq);
        }
+       /* Find virtual address and segment of hardware dequeue pointer */
+       state->new_deq_seg = ep_ring->deq_seg;
+       state->new_deq_ptr = ep_ring->dequeue;
+       while (xhci_trb_virt_to_dma(state->new_deq_seg, state->new_deq_ptr)

+ != (dma_addr_t)(hw_dequeue & ~0x1)) {
+ next_trb(xhci, ep_ring, &state->new_deq_seg,
+ &state->new_deq_ptr);
+ if (state->new_deq_ptr == ep_ring->dequeue) {
+ WARN_ON(1);
+ return;
+ }
+ }

But I'm still concerned about the dequeue pointer in the streams case. streams may be nested, we might be pointing at another stream context instead of the dequeue pointer.

Also the comparison of the dequeue pointers, using (hw_dequeue & ~0x1) might have some troubles with streams. Endpoint context TR dequeue pointer LO field has bits 3:1 reserved (probably zero) but stream context uses those bits. Would it make sense to use (hw_dequeue & ~0xf) here instead?

So there's still some work needed. Are you interested in re-working this to fit on top of 3.15-rc1 or should I add it to my todo list?

-Mathias

#5 lore raw

+hdegoede

Quoted 36 lines

I tried to apply this patch on top of 3.15-rc1, but it fails because of the streams support added to xhci_find_new_dequeue_state()

After some manual editing the interesting parts of xhci_find_new_dequeue_state() looks like this:

(quoted diff) +14 -2
@@ -577,46 +568,57 @@ void xhci_find_new_dequeue_state(struct xhci_hcd
 *xhci,
        if (ep->ep_state & EP_HAS_STREAMS) {
                struct xhci_stream_ctx *ctx =
                        &ep->stream_info->stream_ctx_array[stream_id];
-               state->new_cycle_state = 0x1 &
 le64_to_cpu(ctx->stream_ring);
+               hw_dequeue = le64_to_cpu(ctx->stream_ring);
        } else {
                struct xhci_ep_ctx *ep_ctx
 
                        = xhci_get_ep_ctx(xhci, dev->out_ctx, ep_index);
-               state->new_cycle_state = 0x1 & le64_to_cpu(ep_ctx->deq);
+               hw_dequeue = le64_to_cpu(ep_ctx->deq);
        }
 
+       /* Find virtual address and segment of hardware dequeue pointer */
 
+       state->new_deq_seg = ep_ring->deq_seg;
+       state->new_deq_ptr = ep_ring->dequeue;
+       while (xhci_trb_virt_to_dma(state->new_deq_seg, state->new_deq_ptr)
+                       != (dma_addr_t)(hw_dequeue & ~0x1)) {
+               next_trb(xhci, ep_ring, &state->new_deq_seg,
+                                       &state->new_deq_ptr);
+               if (state->new_deq_ptr == ep_ring->dequeue) {
+                       WARN_ON(1);
+                       return;
+               }
+       }
 
Quoted 4 lines

Also the comparison of the dequeue pointers, using (hw_dequeue & ~0x1) might have some troubles with streams. Endpoint context TR dequeue pointer LO field has bits 3:1 reserved (probably zero) but stream context uses those bits. Would it make sense to use (hw_dequeue & ~0xf) here instead?

Ah, yes, looks like that patch wasn't in Linus' tree yet back when I wrote this. I think your merge looks pretty good... just use (hw_dequeue & ~0xf) instead of (hw_dequeue & ~0x1) to get the pointer as you said, and this should work fine.

Quoted 6 lines

But I'm still concerned about the dequeue pointer in the streams case. streams may be nested, we might be pointing at another stream context instead of the dequeue pointer.

So there's still some work needed. Are you interested in re-working this to fit on top of 3.15-rc1 or should I add it to my todo list?

Hmmm... maybe the stream_id parameter is already pointing to the correct secondary stream (if applicable) so you can rely on having a normal dequeue pointer there? The xhci_triad_to_transfer_ring() function seems to make the same assumption... At any rate, if there is a problem here it would also be in the original c4bedb77e ("xhci: For streams the css flag most be read from the stream-ctx on ep stop") already, so I think you should follow up with that patch's author and fix it in a separate commit if necessary.

I unfortunately don't have a device using streams to test with, so I couldn't do more for this patch than you've already done. I think if you change the hw_dequeue masking as you said that should guarantee that the patch at least won't make things worse for streams, and that should be good enough.

#6 lore raw

Hi,

Quoted 38 lines · Julius Werner

+hdegoede

Quoted 36 lines

I tried to apply this patch on top of 3.15-rc1, but it fails because of the streams support added to xhci_find_new_dequeue_state()

After some manual editing the interesting parts of xhci_find_new_dequeue_state() looks like this:

(quoted diff) +14 -2
@@ -577,46 +568,57 @@ void xhci_find_new_dequeue_state(struct xhci_hcd
 *xhci,
         if (ep->ep_state & EP_HAS_STREAMS) {
                 struct xhci_stream_ctx *ctx =
                         &ep->stream_info->stream_ctx_array[stream_id];
-               state->new_cycle_state = 0x1 &
 le64_to_cpu(ctx->stream_ring);
+               hw_dequeue = le64_to_cpu(ctx->stream_ring);
         } else {
                 struct xhci_ep_ctx *ep_ctx
 
                         = xhci_get_ep_ctx(xhci, dev->out_ctx, ep_index);
-               state->new_cycle_state = 0x1 & le64_to_cpu(ep_ctx->deq);
+               hw_dequeue = le64_to_cpu(ep_ctx->deq);
         }
 
+       /* Find virtual address and segment of hardware dequeue pointer */
 
+       state->new_deq_seg = ep_ring->deq_seg;
+       state->new_deq_ptr = ep_ring->dequeue;
+       while (xhci_trb_virt_to_dma(state->new_deq_seg, state->new_deq_ptr)
+                       != (dma_addr_t)(hw_dequeue & ~0x1)) {
+               next_trb(xhci, ep_ring, &state->new_deq_seg,
+                                       &state->new_deq_ptr);
+               if (state->new_deq_ptr == ep_ring->dequeue) {
+                       WARN_ON(1);
+                       return;
+               }
+       }
 
Quoted 13 lines
Quoted 4 lines

Also the comparison of the dequeue pointers, using (hw_dequeue & ~0x1) might have some troubles with streams. Endpoint context TR dequeue pointer LO field has bits 3:1 reserved (probably zero) but stream context uses those bits. Would it make sense to use (hw_dequeue & ~0xf) here instead?

Ah, yes, looks like that patch wasn't in Linus' tree yet back when I wrote this. I think your merge looks pretty good... just use (hw_dequeue & ~0xf) instead of (hw_dequeue & ~0x1) to get the pointer as you said, and this should work fine.

Quoted 3 lines

But I'm still concerned about the dequeue pointer in the streams case. streams may be nested, we might be pointing at another stream context instead of the dequeue pointer.

Since I've not followed the entire discussion previously to this I cannot really provide any useful feedback on this patch. Other then 2 remarks:

1) We don't use nested streams, so no need to worry about those
2) You're right that for streams to get the dequeue address you need to mask with ~0xf

Regards,

Hans