Problem: security locked disk and partition table read..

I was sorting through drives here today, and found one that was "security locked" (had a user passwd set on it), and for which I did not know the password.

So I hotplugged it to my SiI-3132 card, with the intent of simply erasing the drive (thereby clearing the passwd) using hdparm.

Except.. the kernel (?) sat there for a very long time trying (and failing) over and over and over and over and over and over to read sector-0 (the partition table). Of course the reads each failed, and then got retried 5-times by SCSI, and then retried again a zillion times at a higher level.

There's no hope of a read (or write) ever succeeding on a security locked drive. So why do we even bother allowing them?

To fix the problem at hand, I patched libata-scsi.c to simply fail any command other that ATA_12 or ATA_16 (passthru). This resulted in 56 occurences of this in syslog when the drive was subsequently powered on:

[   11.050986] ata_scsi_translate: blocking SCSI op 0x28 ATA op 0x60
[   11.050991] sd 2:0:0:0: [sdb]  Result: hostbyte=0x00 driverbyte=0x08
[   11.050994] sd 2:0:0:0: [sdb]  Sense Key : 0x5 [current]
[   11.050998] sd 2:0:0:0: [sdb]  ASC=0x21 ASCQ=0x0
[   11.051001] sd 2:0:0:0: [sdb] CDB: cdb[0]=0x28: 28 00 00 00 00 00 00 00 08 00
[   11.051008] end_request: I/O error, dev sdb, sector 0

Why so many attempts to read the partition table???

Anyway, here's the hack I used:

@@ -1819,7 +1819,15 @@
 
 	if (xlat_func(qc))
 		goto early_finish;
-
+	/* don't flog a dead horse, err.. drive, if it is security-locked */
+	if (cmd->cmnd[0] != ATA_12 && cmd->cmnd[0] != ATA_16) {
+		if (dev->id[128] & (1<<2)) { /* locked? */
+			printk(KERN_INFO "%s: blocking SCSI op 0x%02x ATA op 0x%02x\n", __func__,

cmd->cmnd[0], qc->tf.command);
+ ata_scsi_set_sense(cmd, ILLEGAL_REQUEST, 0x21, 0x0);
+ goto early_finish;
+ }
+ }
+
if (ap->ops->qc_defer) {
if ((rc = ap->ops->qc_defer(qc)))
goto defer;

Ideally, that should get re-coded to block only READ/WRITE accesses, and allow all other commands through, but it worked well enough for the situation at hand. After the quick failure of the 56 attempts to read the partition table, I was then able to use hdparm to issue the security-erase command to clear out the drive.

Suggestions?
Anyone want to gain fame and fortune (well, fame anyway) by implementing this more properly?

Cheers

I think SCSI and libata are behaving okay here. The root cause is how the partition code handles IO errors. Different partition table type probes do IOs separately. It's somewhat reasonable given that they may poke at different sectors which sometimes could be near the end of the device failure of which might not necessarily mean the whole device is inaccessible, but anyways the end result is that the each partition table type pokes the drive regardless of how the previous attempt went. And yeah we probably need some improvements there.

Thanks.

tejun

It gets prodded by the various partition formats, the userspace hotplug code and so on, not always for block 0.

Each partitioning format does its own requests. Normally that is fine but in this case its bad news.

Presumably we should set a per device flag of 'security locked' and fail normal I/O to it but allow the other paths ? The 'proper' way would then be to use the key rings and also a kernel event to poke userspace into handling it.

Alan