dd-parallel-posix occasionally corrupts data about dd-parallel HOT 2 CLOSED

Investigation revealed that the corruption always starts from the very first mebibyte, and consists of swapping of evens and odds: so instead of mebibytes 0, 1, 2, 3, …, the output contains mebibytes 1, 0, 3, 2, ….

from dd-parallel.

Further investigation determined that the cause was the reader thread racing ahead of the writer thread initially, and the writer thread starting from the most-recently-read buffer rather than the first buffer. The reader thread racing ahead causes the most-recently-read buffer to be the second buffer by the time the writer thread picks it up.

The fix is for the writer thread to always start with the first buffer, same as the reader thread does. I've been testing this fix for the last 12 hours (where previously the bug would occur after about an hour of repeated runs) and haven't had corruption once.

from dd-parallel.