Since we use a magic RAM flag to signal to the bootloader there's a risk of the sketch overwriting the magic RAM location before the bootloader starts. By reducing the watchdog timeout we reduce the chance of this happening.
Previously, if avrdude skipped the upload step the read while write memory might never be reenabled. This would prevent the bootloader from ending and the sketch from starting - the bootloader just continued to run even after a successful upload. Now enable RWW when we receive the AVR109 "Exit bootloader" 'E' command. Had to disable lock bit changing support in the bootloader to free up flash for the change - will anyone miss it?
attempt to fix bug reported by Dave Mellis:
"avrdude: ser_close(): can’t reset attributes for device: Invalid argument"
is thrown after attempting an upload in the IDE (though the upload seems to be successful)
Earlier approach had bootloader end as soon as it was told to. On Linux this caused problems with avrdude because the microcontroller never had a chance to send an acknowledgement.
Before the sketch initiates an auto-reset for upload it pokes a magic word into a specific RAM address. On starting the bootloader checks this address. If it finds the magic word it knows the bootloader code should run. If not it jumps straight back to sketch.
Test in a sketch by adding to setup():
wdt_enable(WDTO_2S);
Sketch should upload, start, run for two seconds, WDT, and sketch should restart (not bootloader).
Had to cut out unused descriptor code to make the bootloader still fit in 4k.
Disconnect programmer after burning bootloader to get the full effect. ICSP and JTAG programmers do a hardware reset after any reset condition, confusing the firmware a little.
On Windows COM port changes when board switched between bootloader and sketch. No way to prevent this so now Windows users have to select the upload port separate from the comm port. Also, handling of reset into bootloader was broken on Windows. Would occasionally leave the original COM port completely unusable. Changed the way this reset is initiated.
Finally, had to add upload.disable.flushing=true flag to boards.txt so IDE wouldn't try to flush the original COM port after it disappeared.
