Symptoms: The device appears to work fine, however there is a stealthy data corruption, attempting to do a Gentoo Linux installation failed 3 times in a row until I figured out how the ‘disk’ was getting corrupted. Because Linux uses a write-cache, some failed writes dont appear to be failed until you either reboot, or hard unmount the device and force clear the cache. Then on subsequent re-reads of the data you will see subtle bit-wise errors.

I’ve experienced this with my Marvel based design, if the board designer does not correctly match wire trace-lengths on the PWB amongst the ATA address/data/control lines, during high speed operation such as UDMA modes, there are race conditions leading to errors. The problem is generally that the Marvel chip in combination with some CF card devices will claim to perform CRC checking, but will actually not, so the host never knows that the data payload got corrupted on the way from the SATA controller to the CF card media.

Possible ways to overcome the issue, turn off all higher speed operations, limit operation to PIO0 or PIO1 mode (very slow), or choose a CF card that correctly implements the ATA CRC functionality.