Snooping the USB Data Stream

Day 1: I open the box to see a small USB device—no bigger than a quarter—a CD and a note from my editor, “Make this work on Linux!” “Okay”, I think, “this should be easy.”

Figure 1. MARX Software Security's CrypToken

I plug the device in to my laptop and run a small program called usbview to learn what the Linux kernel thinks this device is (Figure 2). This device must be calling itself a USB CrypToken, as that is the string contained in the device. Unfortunately, the device name is in red, which means no kernel driver is bound to the device. I either have to write one or find a way to use libusb to talk to the device from user space (see my article “Writing a Real Driver—in User Space”, LJ, June 2004, for more information about libusb).

Figure 2. usbview identifies the device by vendor and product ID.

Not content to rely on pretty GUI programs, I poke around in the /proc/bus/usb/devices file to get the raw device information. The section that describes this device looks like this:

T:  Bus=01 Lev=02 Prnt=03 Port=02 Cnt=01 Dev#=  4 Spd=1.5 MxCh= 0
D:  Ver= 1.00 Cls=ff(vend.) Sub=00 Prot=ff MxPS= 8 #Cfgs=  1
P:  Vendor=0d7a ProdID=0001 Rev= 1.07
S:  Manufacturer=Marx
S:  Product=USB crypToken
C:* #Ifs= 1 Cfg#= 1 Atr=80 MxPwr= 16mA
I:  If#= 0 Alt= 0 #EPs= 0 Cls=ff(vend.) Sub=00 Prot=ff Driver=(none)

Curious to know whether any other Linux user has tried this device, I consult the Linux USB Working Devices List (see the on-line Resources section). Plugging the vendor ID of 0d7a in to the Quick Search field results in no records found. Perhaps this project will take more work than I thought.

Day 2: The CD, where did I toss it? I locate it and put it in the drive and, look, there's a file called linux.txt on it. Wow, a vendor that acknowledges that Linux might be a viable operating system to support—things sure have changed over the years. After poking around further and reading the documentation on the CD, I realize the device is a small crypto token that can be used to do all sorts of fun things, such as read a unique serial number from the device (each device is different), encrypt data through the device with a 128-bit key stored only in the device and save data on the device in a secure storage area.

On the CD is a shared library that can be used to talk to the USB device to allow a program to access the functions provided by this device. Also present is a small test program that shows how the different library functions work. The library uses libusb to talk directly to the device from user space, which means that a kernel driver is not necessary for this device. The library's license does not allow it to be used within a program that would be licensed under the GPL, however, which is unfortunate for many potential uses. I need to find some way to allow GPL programs to talk to the USB device.

Day 3: While rummaging through my old collection of USB patches, I dig up a reference to a developer who modified the kernel usbfs core code to log all data that flows through it. This patch would allow anyone to read the raw USB data for any program that uses usbfs to talk to a USB device. Because libusb uses usbfs to communicate with USB devices, this might offer a way to reverse engineer this device. Unfortunately, the patch wasn't present with the reference, and no amount of digging on the Internet turned up any real code.

Day 4: As there is no available patch to do what I want to do, I might as well do it myself. So, off to grab the latest 2.6 kernel source tree and dive in.

The files inode.c, devices.c and devio.c in the drivers/usb/core/ directory of the kernel source tree implement the usbfs filesystem. The main filesystem code is in the inode.c file. It contains all of the various VFS code that creates a virtual filesystem and the virtual files within it. The file devices.c handles the creation and reading of the /proc/bus/usb/devices file. This file shows all USB devices and information for those devices in the system at the present time.

The file devio.c controls the raw access of USB devices through the usbfs filesystem. For a user-space program to talk to a USB device through usbfs, it needs to use the ioctl() command on a file that represents the USB device. All of the different ioctl messages that can be sent to the USB devices through usbfs are detailed in the include/linux/usbdevfs.h file.

So, in order to log all accesses to all devices through usbfs, the devio.c file should be modified. Digging into the file, the function usbdev_ioctl looked like the proper place to do this logging. It is called for every ioctl call to a usbfs file. Within that function is a big switch statement that calls the proper functions, depending on the different ioctl command. That is the perfect place to log what kind of command was sent to the device. So, I added a simple printk() call to each case statement, causing them to look like this:

...
case USBDEVFS_CLAIMINTERFACE:
  printk("CLAIMINTERFACE\n);
  ret = proc_claiminterface(ps, (void __user *)arg);
  break;

case USBDEVFS_RELEASEINTERFACE:
  printk("RELEASEINTERFACE\n");
  ret = proc_releaseinterface(ps, (void __user *)arg);
  break;
...

A simple compile, install and module load later confirmed that every usbfs access is now logged to the kernel log, which can be seen by running the dmesg program. I determined that running the lsusb program as lsusb -v produced a lot of usbfs accesses as the program retrieves all of the raw USB configuration data from all devices.

Day 5: Now that the different kinds of usbfs accesses can be noticed easily, it is time to log the data these accesses generate. In looking at the description of the device in the /proc/bus/usb/devices file, it appears that I care only about the accesses to the control endpoint, because there are no endpoints assigned to this device.

Digging further into the devio.c file, I determine that the proc_control() function handles all control messages. There, the code determines whether the request is a read or write control message with the code:

if (ctrl.bRequestType & 0x80) {

The USB bRequestType variable is a bitfield, and the uppermost bit determines the direction of the request. So, in the read section of this if statement I add the lines:

printk("control read: "
       "bRequest=%02x bRrequestType=%02x "
       "wValue=%04x wIndex=%04x\n",
       ctrl.bRequest, ctrl.bRequestType,
       ctrl.wValue, ctrl.wIndex);