Skip to contentSkip to conversation

xhci: fix matching completion events with TDs

RFC PATCH Superseded Michał Pecio 1 participant 2 messages Jan 23, 2024 – Jan 26, 2024

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

RFC PATCH 1/1 xhci: fix matching completion events with TDs

A trb_in_td() call is used to determine if a completion event matches one of the TRBs of the currently executing TD. This function is given ep_ring->deq_seg and ep_ring->dequeue as the starting point to search, but these have been observed to be out of sync with td->start_seg and td->first_trb of the current TD under some circumstances when the EP is being stopped.

Consequently, spurious completion events have been observed to become erroneously associated with an unrelated pending TD.

Since the ring is being searched for the specific purpose of finding a match with current TD, always start from current TD's first TRB.

---

Questions:

Is this a good idea? Is it useful? Or am I missing something?

Should a diagnostic be emitted when these pointers don't match? Or when they don't match AND something else is also wrong?

And yes, I am observing unexpected completion events when stopping an isochronous endpoint on NEC uPD720200. They used to retire the current TD and "TRB not in TD" appeared when its true completion arrived, now they correctly give "TRB not in TD" right away. I consider it progress, though the origin of spurious completions is still unknown.

 drivers/usb/host/xhci-ring.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
drivers/usb/host/xhci-ring.c +3 -4
index 9673354d70d5..d9be5023abe6 100644
@@ -2809,7 +2809,7 @@ static int handle_tx_event(struct xhci_hcd *xhci,
 			td_num--;
 
 		/* Is this a TRB in the currently executing TD? */
-		ep_seg = trb_in_td(xhci, ep_ring->deq_seg, ep_ring->dequeue,
+		ep_seg = trb_in_td(xhci, td->start_seg, td->first_trb,
 				td->last_trb, ep_trb_dma, false);
 
 		/*
@@ -2877,9 +2877,8 @@ static int handle_tx_event(struct xhci_hcd *xhci,
 					"part of current TD ep_index %d "
 					"comp_code %u\n", ep_index,
 					trb_comp_code);
-				trb_in_td(xhci, ep_ring->deq_seg,
-					  ep_ring->dequeue, td->last_trb,
-					  ep_trb_dma, true);
+				trb_in_td(xhci, td->start_seg, td->first_trb,
+					  td->last_trb, ep_trb_dma, true);
 				return -ESHUTDOWN;
 			}
 		}
Signature · 1 line
2.43.0
#1 lore raw

PATCH v2 1/1 xhci: fix matching completion events with TDs

A trb_in_td() call is used to determine if a completion event matches any TRB of the currently executing TD. This function is told to start searching right after the last finished TD, which is not at all where the currently expected TD is guaranteed to begin, because some TDs in between may have been cancelled.

Not only is a pointless work performed, but a bug resulting in the HC executing cancelled TDs was seen to trick the driver into associating events from a TD just cancelled with an unrelated future TD.

Since the ring is being traversed for the specific purpose of finding a match with the current TD, always start from its first TRB. This is the most reliable bit of information that we posses.

Tracking of HC's work progress is not affected, except for cases when a misattributed event would have moved dequeue past a pending TD.

---

v2: improved commit message based on new findings

I am now fairly convinced that this is indeed a good idea. Otherwise, certain event abnormalities develop into several further failures:
- completion of TDs not yet completed can be reported to the core
- ... which may conceivably lead even to DMA-after-free
- ring->dequeue is progressed past a TD not yet released by hardware
- diagnostics are printed only on a later, actually correct event

 drivers/usb/host/xhci-ring.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
drivers/usb/host/xhci-ring.c +3 -4
index 9673354d70d5..d9be5023abe6 100644
@@ -2809,7 +2809,7 @@ static int handle_tx_event(struct xhci_hcd *xhci,
 			td_num--;
 
 		/* Is this a TRB in the currently executing TD? */
-		ep_seg = trb_in_td(xhci, ep_ring->deq_seg, ep_ring->dequeue,
+		ep_seg = trb_in_td(xhci, td->start_seg, td->first_trb,
 				td->last_trb, ep_trb_dma, false);
 
 		/*
@@ -2877,9 +2877,8 @@ static int handle_tx_event(struct xhci_hcd *xhci,
 					"part of current TD ep_index %d "
 					"comp_code %u\n", ep_index,
 					trb_comp_code);
-				trb_in_td(xhci, ep_ring->deq_seg,
-					  ep_ring->dequeue, td->last_trb,
-					  ep_trb_dma, true);
+				trb_in_td(xhci, td->start_seg, td->first_trb,
+					  td->last_trb, ep_trb_dma, true);
 				return -ESHUTDOWN;
 			}
 		}
Signature · 1 line
2.43.0