Skip to contentSkip to conversation

USB: Fix device driver race

PATCH In review Bastien Nocera 3 participants 7 messages Jul 25, 2020 – Jul 26, 2020 USB

lore
Share & tools
lore
#0 lore raw

PATCH v5 3/3 USB: Fix device driver race

When a new device with a specialised device driver is plugged in, the new driver will be modprobe()'d but the driver core will attach the "generic" driver to the device.

After that, nothing will trigger a reprobe when the modprobe()'d device driver has finished initialising, as the device has the "generic" driver attached to it.

Trigger a reprobe ourselves when new specialised drivers get registered.

Fixes 88b7381a939d ("USB: Select better matching USB drivers when available")

---
Changes since v4:
- Add commit subject to "fixes" section
- Clarify conditional that checks for generic driver
- Remove check duplicated inside the loop

Changes since v3:
- Only reprobe devices that could use the new driver
- Many code fixes

Changes since v2:
- Fix formatting

Changes since v1:
- Simplified after Alan Stern's comments and some clarifications from Benjamin Tissoires.

 drivers/usb/core/driver.c | 37 +++++++++++++++++++++++++++++++++++--
 1 file changed, 35 insertions(+), 2 deletions(-)
drivers/usb/core/driver.c +35 -2
index f81606c6a35b..7d3878aa8090 100644
@@ -905,6 +905,32 @@ static int usb_uevent(struct device *dev, struct kobj_uevent_env *env)
 	return 0;
 }
 
+static bool is_dev_usb_generic_driver(struct device *dev)
+{
+	struct usb_device_driver *udd = dev->driver ?
+		to_usb_device_driver(dev->driver) : NULL;
+
+	return udd == &usb_generic_driver;
+}
+
+static int __usb_bus_reprobe_drivers(struct device *dev, void *data)
+{
+	struct usb_device_driver *new_udriver = data;
+	struct usb_device *udev;
+
+	if (!is_dev_usb_generic_driver(dev))
+		return 0;
+
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);
+
+	return 0;
+}
+
 /**
  * usb_register_device_driver - register a USB device (not interface) driver
  * @new_udriver: USB operations for the device driver
@@ -934,13 +960,20 @@ int usb_register_device_driver(struct usb_device_driver *new_udriver,
 
 	retval = driver_register(&new_udriver->drvwrap.driver);
 
-	if (!retval)
+	if (!retval) {
 		pr_info("%s: registered new device driver %s\n",
 			usbcore_name, new_udriver->name);
-	else
+		/*
+		 * Check whether any device could be better served with
+		 * this new driver
+		 */
+		bus_for_each_dev(&usb_bus_type, NULL, new_udriver,
+				 __usb_bus_reprobe_drivers);
+	} else {
 		printk(KERN_ERR "%s: error %d registering device "
 			"	driver %s\n",
 			usbcore_name, retval, new_udriver->name);
+	}
 
 	return retval;
 }
Acked-by on v5 3/3
#1 lore raw
Quoted 38 lines · Bastien Nocera

When a new device with a specialised device driver is plugged in, the new driver will be modprobe()'d but the driver core will attach the "generic" driver to the device.

After that, nothing will trigger a reprobe when the modprobe()'d device driver has finished initialising, as the device has the "generic" driver attached to it.

Trigger a reprobe ourselves when new specialised drivers get registered.

Fixes 88b7381a939d ("USB: Select better matching USB drivers when available")

---
Changes since v4:
- Add commit subject to "fixes" section
- Clarify conditional that checks for generic driver
- Remove check duplicated inside the loop

Changes since v3:
- Only reprobe devices that could use the new driver
- Many code fixes

Changes since v2:
- Fix formatting

Changes since v1:
- Simplified after Alan Stern's comments and some clarifications from Benjamin Tissoires.

 drivers/usb/core/driver.c | 37 +++++++++++++++++++++++++++++++++++--
 1 file changed, 35 insertions(+), 2 deletions(-)
drivers/usb/core/driver.c +0 -0
index f81606c6a35b..7d3878aa8090 100644

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 914

Quoted 10 lines
drivers/usb/core/driver.c +7 -0
@@ -905,6 +905,32 @@ static int usb_uevent(struct device *dev, struct kobj_uevent_env *env)
 	return 0;
 }
 
+static bool is_dev_usb_generic_driver(struct device *dev)
+{
+	struct usb_device_driver *udd = dev->driver ?
+		to_usb_device_driver(dev->driver) : NULL;
+
+	return udd == &usb_generic_driver;
+}

Heh... I don't recommend this optimization because it's a little unclear, but the function can be shortened to:

	return dev->driver == &usb_generic_driver.drvwrap.driver;
Quoted 5 lines
drivers/usb/core/driver.c +5 -0
+
+static int __usb_bus_reprobe_drivers(struct device *dev, void *data)
+{
+	struct usb_device_driver *new_udriver = data;
+	struct usb_device *udev;

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 929

Quoted 10 lines
drivers/usb/core/driver.c +10 -0
+
+	if (!is_dev_usb_generic_driver(dev))
+		return 0;
+
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);

What's that '!' doing hiding in there? It doesn't affect the final outcome, but it sure looks weird -- if people notice it at all.

Aside from that,

Alan Stern

Quoted 31 lines
drivers/usb/core/driver.c +13 -2
+
+	return 0;
+}
+
 /**
  * usb_register_device_driver - register a USB device (not interface) driver
  * @new_udriver: USB operations for the device driver
@@ -934,13 +960,20 @@ int usb_register_device_driver(struct usb_device_driver *new_udriver,
 
 	retval = driver_register(&new_udriver->drvwrap.driver);
 
-	if (!retval)
+	if (!retval) {
 		pr_info("%s: registered new device driver %s\n",
 			usbcore_name, new_udriver->name);
-	else
+		/*
+		 * Check whether any device could be better served with
+		 * this new driver
+		 */
+		bus_for_each_dev(&usb_bus_type, NULL, new_udriver,
+				 __usb_bus_reprobe_drivers);
+	} else {
 		printk(KERN_ERR "%s: error %d registering device "
 			"	driver %s\n",
 			usbcore_name, retval, new_udriver->name);
+	}
 
 	return retval;
 }
#2 lore raw

On Sat, 2020-07-25 at 10:59 -0400, Alan Stern wrote:
<snip>

Quoted 9 lines

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 929

Quoted 6 lines
drivers/usb/core/driver.c +6 -0
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);

What's that '!' doing hiding in there? It doesn't affect the final outcome, but it sure looks weird -- if people notice it at all.

It's how we stop gcc from complaining about the warn_unused_result attribute on device_reprobe()... (void) is enough with clang, but not with gcc.

Quoted 3 lines

Aside from that,

Thanks!

#3 lore raw
Quoted 12 lines · Bastien Nocera

On Sat, 2020-07-25 at 10:59 -0400, Alan Stern wrote:
<snip>

Quoted 9 lines

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 929

Quoted 6 lines
drivers/usb/core/driver.c +6 -0
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);

What's that '!' doing hiding in there? It doesn't affect the final outcome, but it sure looks weird -- if people notice it at all.

Quoted 3 lines

It's how we stop gcc from complaining about the warn_unused_result attribute on device_reprobe()... (void) is enough with clang, but not with gcc.

Hmmm. Maybe this is an indication that device_reprobe() doesn't really need to be __must_check.

Greg, do you know why it's annotated this way?

Alan Stern

#4 lore raw
Quoted 13 lines · Alan Stern
Quoted 12 lines · Bastien Nocera

On Sat, 2020-07-25 at 10:59 -0400, Alan Stern wrote:
<snip>

Quoted 9 lines

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 929

Quoted 6 lines
drivers/usb/core/driver.c +6 -0
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);

What's that '!' doing hiding in there? It doesn't affect the final outcome, but it sure looks weird -- if people notice it at all.

Quoted 8 lines
Quoted 3 lines

It's how we stop gcc from complaining about the warn_unused_result attribute on device_reprobe()... (void) is enough with clang, but not with gcc.

Hmmm. Maybe this is an indication that device_reprobe() doesn't really need to be __must_check.

Greg, do you know why it's annotated this way?

Because you really should pass up the return value if an error happens here. Why do we think it is safe to ignore?

And that "(void)!" is not ok, if the annotation is safe to ignore, then we need to remove the annotation, don't work around stuff like this without at the very least, a comment saying why it is ok.

thanks,

greg k-h

#5 lore raw
Quoted 21 lines · Greg Kroah-Hartman
Quoted 20 lines · Alan Stern
Quoted 12 lines · Bastien Nocera

On Sat, 2020-07-25 at 10:59 -0400, Alan Stern wrote:
<snip>

Quoted 9 lines

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 929

Quoted 6 lines
drivers/usb/core/driver.c +6 -0
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);

What's that '!' doing hiding in there? It doesn't affect the final outcome, but it sure looks weird -- if people notice it at all.

Quoted 3 lines

It's how we stop gcc from complaining about the warn_unused_result attribute on device_reprobe()... (void) is enough with clang, but not with gcc.

Hmmm. Maybe this is an indication that device_reprobe() doesn't really need to be __must_check.

Quoted 8 lines
Quoted 1 line

Greg, do you know why it's annotated this way?

Because you really should pass up the return value if an error happens here. Why do we think it is safe to ignore?

And that "(void)!" is not ok, if the annotation is safe to ignore, then we need to remove the annotation, don't work around stuff like this without at the very least, a comment saying why it is ok.

I suppose Bastien could log an error message at that point. There isn't much else to do.

Alan Stern

#6 lore raw
Quoted 27 lines · Alan Stern
Quoted 26 lines · Greg Kroah-Hartman
Quoted 13 lines · Alan Stern
Quoted 12 lines · Bastien Nocera

On Sat, 2020-07-25 at 10:59 -0400, Alan Stern wrote:
<snip>

Quoted 9 lines

drivers/usb/core/driver.c static int usb_uevent(struct device *dev, struct kobj_uevent_env *env) · line 929

Quoted 6 lines
drivers/usb/core/driver.c +6 -0
+	udev = to_usb_device(dev);
+	if (usb_device_match_id(udev, new_udriver->id_table) == NULL &&
+	    (!new_udriver->match || new_udriver->match(udev) != 0))
+		return 0;
+
+	(void)!device_reprobe(dev);

What's that '!' doing hiding in there? It doesn't affect the final outcome, but it sure looks weird -- if people notice it at all.

Quoted 8 lines
Quoted 3 lines

It's how we stop gcc from complaining about the warn_unused_result attribute on device_reprobe()... (void) is enough with clang, but not with gcc.

Hmmm. Maybe this is an indication that device_reprobe() doesn't really need to be __must_check.

Greg, do you know why it's annotated this way?

Because you really should pass up the return value if an error happens here. Why do we think it is safe to ignore?

Quoted 6 lines
Quoted 3 lines

And that "(void)!" is not ok, if the annotation is safe to ignore, then we need to remove the annotation, don't work around stuff like this without at the very least, a comment saying why it is ok.

I suppose Bastien could log an error message at that point. There isn't much else to do.

It looks like no one does anything with the return value of that function, so we should just change it to void and stop checking it entirely :(

greg k-h