Ejecting removable disk causes a flood of ALLOW_MEDIUM_REMOVAL

Ejecting removable disk causes a flood of ALLOW_MEDIUM_REMOVAL

[Albert Lee]

Hi Doug,

A problem related to SG_IO.
When testing the 2.6.18-rc1 libata, a flood of ALLOW_MEDIUM_REMOVAL is seen
after ejecting a removable disk drive:

http://bugzilla.kernel.org/show_bug.cgi?id=6799

The interaction to trigger the problem is below:
1. The "eject" utility, using SG_IO, issues ALLOW_MEDIUM_REMOVAL twice to unlock the disk of the device.
2. The disk of the physical device is now unlocked, however, sdev->locked is still 1, inconsistent with the physical device status.
3. The "eject" utility ejects the disk by START_STOP_UNIT.
4. Someone issues TEST_UNIT_READY. The device is not ready since no disk inside. ATAPI check condition returned.
5. 2.6.18-rc1 libata EH triggered in SCSI EH context to request sense.
6. scsi_restart_operations(), seeing sdev->locked == 1, tries to lock the door.
7. Lock failed since no disk inside. ATAPI check condition returned.
8. 2.6.18 libata EH triggered in SCSI EH context. Goto 5 and loop forever until the disk is inserted again.

It looks more like eject/scsi problem. (Libata 2.6.17 EH is not affected by the problem 
because the request sense is done when qc is completed, not in the SCSI EH context.)

Maybe adding a filter to the SCSI SG IO layer to
 - Check incoming SCSI commands
 - Whenever ALLOW_MEDIUM_REMOVAL is seen, update the sdev->locked to make the 
   kernel data structure consistent with the physical device status
can fix the problem?

--
albert

Re: Ejecting removable disk causes a flood of ALLOW_MEDIUM_REMOVAL

[Douglas Gilbert]

Albert Lee wrote:
> Hi Doug,
> 
> A problem related to SG_IO.
> When testing the 2.6.18-rc1 libata, a flood of ALLOW_MEDIUM_REMOVAL is seen
> after ejecting a removable disk drive:
> 
> http://bugzilla.kernel.org/show_bug.cgi?id=6799
> 
> The interaction to trigger the problem is below:
> 1. The "eject" utility, using SG_IO, issues ALLOW_MEDIUM_REMOVAL twice to unlock the disk of the device.
> 2. The disk of the physical device is now unlocked, however, sdev->locked is still 1, inconsistent with the physical device status.
> 3. The "eject" utility ejects the disk by START_STOP_UNIT.
> 4. Someone issues TEST_UNIT_READY. The device is not ready since no disk inside. ATAPI check condition returned.
> 5. 2.6.18-rc1 libata EH triggered in SCSI EH context to request sense.
> 6. scsi_restart_operations(), seeing sdev->locked == 1, tries to lock the door.
> 7. Lock failed since no disk inside. ATAPI check condition returned.
> 8. 2.6.18 libata EH triggered in SCSI EH context. Goto 5 and loop forever until the disk is inserted again.
> 
> It looks more like eject/scsi problem. (Libata 2.6.17 EH is not affected by the problem 
> because the request sense is done when qc is completed, not in the SCSI EH context.)
> 
> Maybe adding a filter to the SCSI SG IO layer to
>  - Check incoming SCSI commands
>  - Whenever ALLOW_MEDIUM_REMOVAL is seen, update the sdev->locked to make the 
>    kernel data structure consistent with the physical device status
> can fix the problem?

Albert,
Just fix libata. My usb external cd/dvd combo drive has a little switch
under it to unlock the door. How could use of that switch be "filtered"
by an OS?

If an OS wants to maintain a state variable for a device that a
user (and perhaps another initiator) can get to then it needs to
be able to cope with that state variable being out
of sync once in a while.

Doug Gilbert