Re: [PATCH] sd: do not set changed flag on all unit attention conditions

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Tue, 2012-07-17 at 10:34 +0200, Paolo Bonzini wrote:
> Il 17/07/2012 09:45, James Bottomley ha scritto:
> > On Mon, 2012-07-16 at 19:20 +0200, Paolo Bonzini wrote:
> >> Il 16/07/2012 18:18, James Bottomley ha scritto:
> >>>>> diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> >>>>> index b583277..6d8ca08 100644
> >>>>> --- a/drivers/scsi/scsi_lib.c
> >>>>> +++ b/drivers/scsi/scsi_lib.c
> >>>>> @@ -843,8 +843,11 @@ void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
> >>>>>  	} else if (sense_valid && !sense_deferred) {
> >>>>>  		switch (sshdr.sense_key) {
> >>>>>  		case UNIT_ATTENTION:
> >>>>> -			if (cmd->device->removable) {
> >>>>> -				/* Detected disc change.  Set a bit
> >>>>> +			if (cmd->device->removable &&
> >>>>> +			    (sshdr.asc == 0x3a ||
> >>>>> +			     (sshdr.asc == 0x28 && sshdr.ascq == 0x00))) {
> >>>>> +				/* "No medium" or "Medium may have changed."
> >>>>> +				 * This means a disc change.  Set a bit
> >>> This type of change would likely cause a huge cascade of errors in real
> >>> removable media devices.  Under the MMC standards, which a lot of the
> >>> older removable discs seem to follow, UNIT ATTENTION indicates either
> >>> medium change or device reset (which we check for and eat lower down);
> >>> we can't rely on them giving proper SBC-2 sense codes.  If you want to
> >>> pretend to be removable media, you have to conform to its standards.
> >>
> >> Would you accept a patch doing the opposite, i.e. passing some sense
> >> codes such as PARAMETERS CHANGED and TARGET OPERATING CONDITIONS HAVE
> >> CHANGED?
> > 
> > Could you explain what the problem actually is?  It looks like you had a
> > reason to mark virtio-scsi as removable, even though it isn't, and now
> > you want to add further hacks because being removable doesn't quite
> > work.
> 
> It's not specific to virtio-scsi, in fact I expect that virtio-scsi will
> be almost always used with non-removable disks.
> 
> However, QEMU's SCSI target is not used just for virtio-scsi (for
> example it can be used for USB storage), and it lets you mark a disk as
> removable---why? because there exists real hardware that presents itself
> as an SBC removable disk.  The only thing that is specific to
> virtualization, is support for online resizing (which generates a unit
> attention condition CAPACITY DATA HAS CHANGED).

So what's the problem?  If you're doing pass through of a physical disk,
we pick up removable from its inquiry string ... a physical removable
device doesn't get resized.  If you have a virtual disk you want to
resize, you don't set the removable flag in the inquiry data.

James