Skip to contentSkip to conversation

usbip: fix device disconnect loop with isoc endpoints

PATCH In review Sascha Grunert 2 participants 5 messages Jul 1

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover usbip: fix device disconnect loop with isoc endpoints

Forwarding a USB device with isochronous endpoints over USB/IP causes a disconnect/reconnect loop. Hit this with a Turtle Beach Velocity One Flight yoke (10f5:7001) forwarded to a VM.

The first patch fixes a TCP stream desync: when get_pipe() returns -1, the remaining PDU payload stays on the socket and corrupts the next header parse. The second patch prevents activation of alt settings with isoc endpoints, since USB/IP cannot forward them and the failed transfers cascade into a device disconnect.

Sascha Grunert (2):
usbip: drain remaining PDU payload on rejected endpoint
usbip: block SET_INTERFACE for isoc alt settings

 drivers/usb/usbip/stub_rx.c | 96 ++++++++++++++++++++++++++++++++++++-
 1 file changed, 95 insertions(+), 1 deletion(-)
Signature · 1 line
2.52.0
#1 lore raw

PATCH 1/2 usbip: drain remaining PDU payload on rejected endpoint

When get_pipe() returns -1, stub_recv_cmd_submit() bails out without reading the transfer buffer and ISO descriptors that follow the PDU header on the TCP stream. The next recv() parses leftover payload as a PDU header, desyncs the stream, and kills the connection.

Consume those trailing bytes before the early return so the stream stays in sync.

Fixes 635f545a7e8b ("usbip: fix stub_rx: get_pipe() to validate endpoint number")
---
 drivers/usb/usbip/stub_rx.c | 60 ++++++++++++++++++++++++++++++++++++-
 1 file changed, 59 insertions(+), 1 deletion(-)
drivers/usb/usbip/stub_rx.c +59 -1
index 1e9ae57..d0e3d3f 100644
@@ -461,6 +461,62 @@ static int stub_recv_xbuff(struct usbip_device *ud, struct stub_priv *priv)
 	return ret;
 }
 
+/*
+ * When get_pipe() rejects an endpoint (e.g. an isochronous endpoint that
+ * does not exist in the current alt setting), the transfer buffer and ISO
+ * packet descriptors that follow the PDU header on the TCP stream must
+ * still be consumed.  Without this the next recv() interprets leftover
+ * payload bytes as a PDU header, desynchronises the stream, and tears
+ * down the connection.
+ */
+static void stub_recv_cmd_submit_drain(struct usbip_device *ud,
+				       struct usbip_header *pdu)
+{
+	int bufsz, ret, np;
+	void *buf;
+
+	if (pdu->base.direction == USBIP_DIR_OUT) {
+		bufsz = pdu->u.cmd_submit.transfer_buffer_length;
+		if (bufsz > 0) {
+			buf = kzalloc(min_t(int, bufsz, PAGE_SIZE),
+				      GFP_KERNEL);
+			if (!buf) {
+				usbip_event_add(ud, SDEV_EVENT_ERROR_MALLOC);
+				return;
+			}
+			while (bufsz > 0) {
+				int chunk = min_t(int, bufsz, PAGE_SIZE);
+
+				ret = usbip_recv(ud->tcp_socket, buf, chunk);
+				if (ret != chunk) {
+					kfree(buf);
+					usbip_event_add(ud,
+							SDEV_EVENT_ERROR_TCP);
+					return;
+				}
+				bufsz -= chunk;
+			}
+			kfree(buf);
+		}
+	}
+
+	np = pdu->u.cmd_submit.number_of_packets;
+	if (np > 0 && np <= USBIP_MAX_ISO_PACKETS) {
+		bufsz = np * sizeof(struct usbip_iso_packet_descriptor);
+		buf = kzalloc(bufsz, GFP_KERNEL);
+		if (!buf) {
+			usbip_event_add(ud, SDEV_EVENT_ERROR_MALLOC);
+			return;
+		}
+		ret = usbip_recv(ud->tcp_socket, buf, bufsz);
+		kfree(buf);
+		if (ret != bufsz) {
+			usbip_event_add(ud, SDEV_EVENT_ERROR_TCP);
+			return;
+		}
+	}
+}
+
 static void stub_recv_cmd_submit(struct stub_device *sdev,
 				 struct usbip_header *pdu)
 {
@@ -479,8 +535,10 @@ static void stub_recv_cmd_submit(struct stub_device *sdev,
 	int ret, i;
 	int is_tweaked;
 
-	if (pipe == -1)
+	if (pipe == -1) {
+		stub_recv_cmd_submit_drain(ud, pdu);
 		return;
+	}
 
 	/*
 	 * Smatch reported the error case where use_sg is true and buf_len is 0.
Signature · 1 line
2.52.0
#2 lore raw

PATCH 2/2 usbip: block SET_INTERFACE for isoc alt settings

In reply to Sascha Grunert

USB/IP cannot forward isochronous transfers. When the client activates an alt setting with isoc endpoints, the transfers fail with -EPROTO and the resulting usb_clear_halt cascade disconnects the device.

Intercept SET_INTERFACE in tweak_set_interface_cmd() and fake success when the target alt setting contains isoc endpoints, keeping the device at alt 0.

Tested with a Turtle Beach Velocity One Flight yoke (10f5:7001) forwarded to a VM via USB/IP, which previously disconnect-looped every few seconds and now stays connected.

---
 drivers/usb/usbip/stub_rx.c | 36 ++++++++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)
drivers/usb/usbip/stub_rx.c +36 -0
index d0e3d3f..f323b48 100644
@@ -100,6 +100,28 @@ static int tweak_clear_halt_cmd(struct urb *urb)
 	return ret;
 }
 
+static bool altsetting_has_isoc(struct usb_device *udev, __u16 interface,
+				__u16 alternate)
+{
+	struct usb_interface *intf;
+	struct usb_host_interface *alt;
+	int i;
+
+	intf = usb_ifnum_to_if(udev, interface);
+	if (!intf)
+		return false;
+
+	alt = usb_altnum_to_altsetting(intf, alternate);
+	if (!alt)
+		return false;
+
+	for (i = 0; i < alt->desc.bNumEndpoints; i++) {
+		if (usb_endpoint_xfer_isoc(&alt->endpoint[i].desc))
+			return true;
+	}
+	return false;
+}
+
 static int tweak_set_interface_cmd(struct urb *urb)
 {
 	struct usb_ctrlrequest *req;
@@ -111,6 +133,20 @@ static int tweak_set_interface_cmd(struct urb *urb)
 	alternate = le16_to_cpu(req->wValue);
 	interface = le16_to_cpu(req->wIndex);
 
+	/*
+	 * USB/IP cannot forward isochronous transfers.  If the requested
+	 * alt setting activates isochronous endpoints, pretend the switch
+	 * succeeded without touching the device.  This prevents the
+	 * cascade of failed isoc URBs that leads to a device disconnect.
+	 */
+	if (alternate != 0 && altsetting_has_isoc(urb->dev, interface,
+						  alternate)) {
+		dev_info(&urb->dev->dev,
+			 "usb_set_interface blocked: inf %u alt %u (isoc)\n",
+			 interface, alternate);
+		return 0;
+	}
+
 	usbip_dbg_stub_rx("set_interface: inf %u alt %u\n",
 			  interface, alternate);
 
Signature · 1 line
2.52.0
maintainer
#3 lore raw

Re: [PATCH 2/2] usbip: block SET_INTERFACE for isoc alt settings

Quoted 8 lines · Sascha Grunert

USB/IP cannot forward isochronous transfers. When the client activates an alt setting with isoc endpoints, the transfers fail with -EPROTO and the resulting usb_clear_halt cascade disconnects the device.

Intercept SET_INTERFACE in tweak_set_interface_cmd() and fake success when the target alt setting contains isoc endpoints, keeping the device at alt 0.

Quoted 6 lines

Tested with a Turtle Beach Velocity One Flight yoke (10f5:7001) forwarded to a VM via USB/IP, which previously disconnect-looped every few seconds and now stays connected.

What commit id does this fix?

Quoted 42 lines
---
 drivers/usb/usbip/stub_rx.c | 36 ++++++++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)
drivers/usb/usbip/stub_rx.c +23 -0
index d0e3d3f..f323b48 100644
@@ -100,6 +100,28 @@ static int tweak_clear_halt_cmd(struct urb *urb)
 	return ret;
 }
 
+static bool altsetting_has_isoc(struct usb_device *udev, __u16 interface,
+				__u16 alternate)
+{
+	struct usb_interface *intf;
+	struct usb_host_interface *alt;
+	int i;
+
+	intf = usb_ifnum_to_if(udev, interface);
+	if (!intf)
+		return false;
+
+	alt = usb_altnum_to_altsetting(intf, alternate);
+	if (!alt)
+		return false;
+
+	for (i = 0; i < alt->desc.bNumEndpoints; i++) {
+		if (usb_endpoint_xfer_isoc(&alt->endpoint[i].desc))
+			return true;
+	}
+	return false;
+}
+
 static int tweak_set_interface_cmd(struct urb *urb)
 {
 	struct usb_ctrlrequest *req;
@@ -111,6 +133,20 @@ static int tweak_set_interface_cmd(struct urb *urb)
 	alternate = le16_to_cpu(req->wValue);
 	interface = le16_to_cpu(req->wIndex);
 
+	/*

drivers/usb/usbip/stub_rx.c static int tweak_set_interface_cmd(struct urb *urb) · line 146

Quoted 10 lines
drivers/usb/usbip/stub_rx.c +10 -0
+	 * USB/IP cannot forward isochronous transfers.  If the requested
+	 * alt setting activates isochronous endpoints, pretend the switch
+	 * succeeded without touching the device.  This prevents the
+	 * cascade of failed isoc URBs that leads to a device disconnect.
+	 */
+	if (alternate != 0 && altsetting_has_isoc(urb->dev, interface,
+						  alternate)) {
+		dev_info(&urb->dev->dev,
+			 "usb_set_interface blocked: inf %u alt %u (isoc)\n",
+			 interface, alternate);

Why is this not an error? And if a user sees this, what can they do about it?

drivers/usb/usbip/stub_rx.c static int tweak_set_interface_cmd(struct urb *urb) · line 147

Quoted 1 line
drivers/usb/usbip/stub_rx.c +1 -0
+		return 0;

Why isn't this an error?

thanks,

greg k-h

#4 lore raw

Re: [PATCH 2/2] usbip: block SET_INTERFACE for isoc alt settings

Quoted 1 line · Greg Kroah-Hartman

What commit id does this fix?

The limitation goes back to the original import:

Fixes 4d7b5c7f8ad4 ("Staging: USB/IP: add host driver")

tweak_set_interface_cmd() has always forwarded SET_INTERFACE unconditionally, which is fine until an alt setting with isochronous endpoints is activated.

Quoted 6 lines

Why is this not an error? And if a user sees this, what can they do about it?

Quoted 1 line

+ return 0;

Why isn't this an error?

Returning 0 here means "handled" in tweak_special_requests() semantics (!err = 1 = tweaked), so the URB is completed with success and never submitted to the HCD. If we returned an error instead, tweak_special_requests() would return 0 (not tweaked) and the URB goes through to the hardware, the isoc alt setting activates, the transfers fail, and the device disconnects, which is the problem this patch is trying to prevent.

You're right that the user can't act on the log message though. I'll change dev_info to dev_dbg in v2.

Thanks,
Sascha