Short description: An absolute pointing device using USB HID defines a LOGICAL_MINIMUM and a LOGICAL_MAXIMUM for X, Y axes, and then sends a HID report containing values outside that range. Linux kernel should ignore values outside that range, as they are not meaningful. Just for comparison, Windows ignores such values. (and I hate this kind of comparison) Long description: I'm building a homebrew USB Absolute Pointing Device using an AVR ATmega8 and the V-USB firmware. This device will work almost like a mouse, with one (obvious) difference: while a mouse sends "relative" movements (move the pointer x,y units, relative to its current position), an absolute pointing device sends "absolute" positions (move the mouse exactly to this x,y location). This device is using two 16-bit fields to report the X and Y axes, and limit the meaningful values to those within the [0..32767] range. This is the relevant part of the USB HID Descriptor: 0x09, 0x30, // USAGE (X) 0x09, 0x31, // USAGE (Y) 0x15, 0x00, // LOGICAL_MINIMUM (0) 0x26, 0xff, 0x7f, // LOGICAL_MAXIMUM (32767) 0x75, 0x10, // REPORT_SIZE (16) 0x95, 0x02, // REPORT_COUNT (2) 0x81, 0x42, // INPUT (Data,Var,Abs,Null) However, Linux kernel is not behaving as expected when receiving out-of-range values. Instead of ignoring such coordinates, it moves the pointer to the bottom-right corner of the screen. Each HID report for this device has 1 byte for the Report ID (with value 02), 2 bytes for the X position, 2 bytes for the Y position, and finally 1 byte for the button presses. They shows up as "02XXXXYY YYbb" in usbmon. Look at this usbmon output: ffff88000c9a0600 1465973871 C Ii:5:093:1 0:8 6 = 02ff0fff 0f00 ffff88000c9a0600 1465973936 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1467317888 C Ii:5:093:1 0:8 6 = 02ff1fff 1f00 ffff88000c9a0600 1467317950 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1467725888 C Ii:5:093:1 0:8 6 = 02ff2fff 2f00 ffff88000c9a0600 1467725931 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1468661905 C Ii:5:093:1 0:8 6 = 02ff3fff 3f00 ffff88000c9a0600 1468661961 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1469429914 C Ii:5:093:1 0:8 6 = 02ff4fff 4f00 ffff88000c9a0600 1469429969 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1470133907 C Ii:5:093:1 0:8 6 = 02ff5fff 5f00 ffff88000c9a0600 1470133948 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1470837931 C Ii:5:093:1 0:8 6 = 02ff6fff 6f00 ffff88000c9a0600 1470837986 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1471445923 C Ii:5:093:1 0:8 6 = 02ff7fff 7f00 ffff88000c9a0600 1471445981 S Ii:5:093:1 -115:8 6 < Here, the pointer is moving from 0x0fff, 0x0fff (near the top-left of the screen) up to the maximum possible value: 0x7fff, 0x7fff (the bottom-right corner). When receiving these reports, the pointer was correctly moving diagonally across my screen. Up until here, no problem yet. The problem starts now: ffff88000c9a0600 1472261948 C Ii:5:093:1 0:8 6 = 02ff8fff 8f00 ffff88000c9a0600 1472262007 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1472965956 C Ii:5:093:1 0:8 6 = 02ff9fff 9f00 ffff88000c9a0600 1472966010 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1473621964 C Ii:5:093:1 0:8 6 = 02ffafff af00 ffff88000c9a0600 1473622019 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1474285972 C Ii:5:093:1 0:8 6 = 02ffbfff bf00 ffff88000c9a0600 1474286027 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1474885980 C Ii:5:093:1 0:8 6 = 02ffcfff cf00 ffff88000c9a0600 1474886034 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1475517975 C Ii:5:093:1 0:8 6 = 02ffdfff df00 ffff88000c9a0600 1475518040 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1476333997 C Ii:5:093:1 0:8 6 = 02ffefff ef00 ffff88000c9a0600 1476334055 S Ii:5:093:1 -115:8 6 < ffff88000c9a0600 1477366011 C Ii:5:093:1 0:8 6 = 02ffffff ff00 ffff88000c9a0600 1477366069 S Ii:5:093:1 -115:8 6 < These reports send values outside the range defined in the HID Descriptor. The values here go from 0x8fff up to 0xffff. All of these values are invalid and should have been ignored by the Linux kernel. Instead, what happened was that the kernel insisted into moving the pointer to the bottom-right corner of the screen. Just to be sure about what I'm saying, I used my touchpad to move the pointer away from the corner before each one of those reports, and upon receiving them, the pointer went back to the corner. It is worth noting that changing this: 0x81, 0x42, // INPUT (Data,Var,Abs,Null) to this: 0x81, 0x02, // INPUT (Data,Var,Abs) makes no difference at all. I've also tried this device on a Windows XP system, and it works as expected there (out-of-range values are ignored and don't move the pointer). It's also important to remember that an out-of-range value should be ignored, but not the full report. For instance, the device could have sent "02 ffff ffff 01", meaning an invalid X,Y position (which should be ignored), but a valid button press of the first button (which should be processed). In this case, the kernel should send a button_press event, but not a motion event. It may even happen to send an out-of-range value for one axis, but a valid value for another axis. The code should be prepared for that (ignore one, but keep the other). Found this bug in Linux kernel 2.6.38 If someone wants to build the hardware to reproduce this bug, go to the revision 662baa542e07 of this project: https://bitbucket.org/denilsonsa/atmega8-magnetometer-usb-mouse/ It will probably need a few tweaks, though. Basically, after removing the non-relevant portions of the code (sensor, menu, keyemu), it should work using a fairly minimalistic ATmega8 circuit. -- Denilson Figueiredo de Sá Rio de Janeiro - Brasil -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
