Media error or link error? What SCSI sense codes and Linux ATA log lines can and cannot tell apart

The first thing to read in a failing SATA read is not the sense key but the ATA error register. UNC (0x40) is reported by the drive only after it has used its own retries, and libata turns it into 03h MEDIUM ERROR with 11h/04h; ICRC with ABRT (0x84) is a transfer that arrived damaged and becomes 0Bh ABORTED COMMAND with 47h/00h. The kernel wiki reads these the same way: UNC as 'often due to bad sectors on the disk', ICRC as 'often either a bad cable or power problem'. The two sources agree on the split but both say 'often', so it is a pointer, not a diagnosis.

Three details limit the split. First, the SCSI side loses information: libata maps every interface CRC to 47h/00h SCSI PARITY ERROR, although T10 lists data-phase CRC (47h/01h), iuCRC (47h/03h) and UDMA CRC (08h/03h) as separate codes, so a tool that only reads sense data cannot tell which link layer was involved. Second, when a command ends with an ATA bus error, libata-eh.c clears the device, media and invalid-argument bits of that command; a media problem on the same command is therefore not logged beside the bus error. Third, IDNF is shown as 05h ILLEGAL REQUEST with 21h/00h (LBA out of range), a name that points to addressing, not to a damaged sector.

A timeout (Emask 0x4) and an HSM violation (0x2) are the least informative lines. The kernel wiki says a timeout may have 'any number of causes', and the same lines drive the speed-down counters as bus errors do. The T10 code 5Dh/00h, FAILURE PREDICTION THRESHOLD EXCEEDED, is a drive-side flag, not an observed fault, and it appears in neither the libata translation table nor the kernel's sense-key handling read for this note.

The kernel reacts differently to the two classes. After more than 3 bus or timeout errors in 10 minutes it lowers the link speed, and after more than 6 combined errors in 5 minutes it falls back to PIO. Media errors get no such treatment: the scsi_error.c handler returns a final disposition for a MEDIUM ERROR with ASC 11h, 13h or 14h and marks it as a medium error, and tries again only for other ASC values. The practical reading is that a log full of bus errors with a link that drops speed is a different situation from a log with scattered UNC lines, and the code treats them as such; whether a given drive's cause is a cable, a backplane, power or the drive's own interface is not something these fields settle.