usb: xhci: add a quirk for Superspeed bulk OUT transfers on VL805

The VL805 has a bug in its internal FIFO space accounting that results
in bulk OUT babble if a TRB in a large multi-element TD has a data
buffer size that is larger than and not a multiple of wMaxPacketSize for
the endpoint. If the downstream USB3.0 link is congested, or latency is
increased through an intermediate hub, then the VL805 enters a suspected
FIFO overflow condition and transmits repeated, garbled data to the
endpoint.

TDs with TRBs of exact multiples of wMaxPacketSize and smaller than
wMaxPacketSize appear to be handled correctly, so split buffers at a
1024-byte length boundary and put the remainder in a separate smaller
TRB.

Signed-off-by: Jonathan Bell <jonathan@raspberrypi.com>

Author: P33M

drivers/usb/host/xhci-pci.c

@@ -298,6 +298,7 @@ static void xhci_pci_quirks(struct device *dev, struct xhci_hcd *xhci)
 		xhci->quirks |= XHCI_EP_CTX_BROKEN_DCS;
 		xhci->quirks |= XHCI_AVOID_DQ_ON_LINK;
 		xhci->quirks |= XHCI_VLI_TRB_CACHE_BUG;
+		xhci->quirks |= XHCI_VLI_SS_BULK_OUT_BUG;
 	}
 
 	if (pdev->vendor == PCI_VENDOR_ID_ASMEDIA &&

drivers/usb/host/xhci-ring.c

@@ -3607,14 +3607,15 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
 	unsigned int num_trbs;
 	unsigned int start_cycle, num_sgs = 0;
 	unsigned int enqd_len, block_len, trb_buff_len, full_len;
-	int sent_len, ret;
-	u32 field, length_field, remainder;
+	int sent_len, ret, vli_quirk = 0;
+	u32 field, length_field, remainder, maxpacket;
 	u64 addr, send_addr;
 
 	ring = xhci_urb_to_transfer_ring(xhci, urb);
 	if (!ring)
 		return -EINVAL;
 
+	maxpacket = usb_endpoint_maxp(&urb->ep->desc);
 	full_len = urb->transfer_buffer_length;
 	/* If we have scatter/gather list, we use it. */
 	if (urb->num_sgs && !(urb->transfer_flags & URB_DMA_MAP_SINGLE)) {
@@ -3651,6 +3652,17 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
 	start_cycle = ring->cycle_state;
 	send_addr = addr;
 
+	if (xhci->quirks & XHCI_VLI_SS_BULK_OUT_BUG &&
+	    !usb_urb_dir_in(urb) && urb->dev->speed >= USB_SPEED_SUPER) {
+		/*
+		 * VL805 - superspeed bulk OUT traffic can cause
+		 * an internal fifo overflow if the TRB buffer is larger
+		 * than wMaxPacket and the length is not an integer
+		 * multiple of wMaxPacket.
+		 */
+		vli_quirk = 1;
+	}
+
 	/* Queue the TRBs, even if they are zero-length */
 	for (enqd_len = 0; first_trb || enqd_len < full_len;
 			enqd_len += trb_buff_len) {
@@ -3663,6 +3675,11 @@ int xhci_queue_bulk_tx(struct xhci_hcd *xhci, gfp_t mem_flags,
 		if (enqd_len + trb_buff_len > full_len)
 			trb_buff_len = full_len - enqd_len;
 
+		if (vli_quirk && trb_buff_len > maxpacket) {
+			/* SS bulk wMaxPacket is 1024B */
+			remainder = trb_buff_len & (maxpacket - 1);
+			trb_buff_len -= remainder;
+		}
 		/* Don't change the cycle bit of the first TRB until later */
 		if (first_trb) {
 			first_trb = false;

drivers/usb/host/xhci.h

@@ -1903,6 +1903,7 @@ struct xhci_hcd {
 #define XHCI_EP_CTX_BROKEN_DCS	BIT_ULL(42)
 #define XHCI_AVOID_DQ_ON_LINK	BIT_ULL(43)
 #define XHCI_VLI_TRB_CACHE_BUG	BIT_ULL(44)
+#define XHCI_VLI_SS_BULK_OUT_BUG	BIT_ULL(45)
 
 	unsigned int		num_active_eps;
 	unsigned int		limit_active_eps;