Bug 1325654

md-raid6's "repair" scan mode is documented as possibly repairing raid6 mismatches amongst the component drives.  But this is implemented by assuming that if a verification scan failed, the parity drives must have been both erroneous, and thus the parity drives are rewritten.  If the actual bit-rot was on one of the data drives, this then ***propagates that data corruption*** irreversibly.

md-raid6 has enough redundancy to correct any one drive's worth of bitrot.  "repair" mode should be changed to exploit that redundancy: it should attempt to rewrite exactly the bad areas - maybe even on a byte-by-byte basis - not necessarily the parity drives.

md-raid6 has probably enough redundancy to detect two drives' worth of overlapping bitrot errors, which it could signal, and refuse to propagate / make-worse.  More than two drives' worth of overlapping errors are probably not reliably diagnosable.

This change would make md-raid6 a reasonable defence against bit-rot, even with overlying filesystems that have no data checksumming features, and with normal applications that cannot do error detection/correction on their files.