Re: [PATCH v2 3/3] [SCSI] Make LBP quirk skip lbpme checks tests

[James Bottomley]

On Fri, 2014-07-25 at 16:47 +0000, KY Srinivasan wrote:
> 
> > -----Original Message-----
> > From: Martin K. Petersen [mailto:martin.petersen@xxxxxxxxxx]
> > Sent: Thursday, July 24, 2014 8:54 AM
> > To: Sitsofe Wheeler
> > Cc: Martin K. Petersen; Christoph Hellwig; KY Srinivasan;
> > gregkh@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > devel@xxxxxxxxxxxxxxxxxxxxxx; ohering@xxxxxxxx; apw@xxxxxxxxxxxxx;
> > jasowang@xxxxxxxxxx; jbottomley@xxxxxxxxxxxxx; linux-
> > scsi@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH v2 3/3] [SCSI] Make LBP quirk skip lbpme checks tests
> > 
> > >>>>> "Sitsofe" == Sitsofe Wheeler <sitsofe@xxxxxxxxx> writes:
> > 
> > Sitsofe> So we can see it is really a SATA device that announces discard
> > Sitsofe> correctly and supports discard through WRITE_SAME(16).
> > 
> > No, that's the SATA device that announces support for DSM TRIM, and as a
> > result the Linux SATL reports support for WRITE SAME(16) w. the UNMAP bit
> > set and LBPME.
> > 
> > Sitsofe> It is the act of passing it through Hyper-V that turned it into
> > Sitsofe> a SCSI device that supports UNMAP (but not WRITE_SAME(16)),
> > Sitsofe> doesn't announce its SCSI conformance number and doesn't
> > Sitsofe> correctly announce which features it supports. Surely in this
> > Sitsofe> case it's reasonable to quirk our way around the problem?
> > 
> > No. That's an issue in Hyper-V that'll you'll have to take up with Microsoft. I
> > don't know what their passthrough limitations are for SCSI-ATA translation.
> > Maybe K. Y. has some insight into this?
> 
> For the pass through case, the host validates the request and passes
> the request to the device.
> However, not all scsi commands are passed through even though the
> device it is being passed through
> may support the command. WRITE_SAME is one such command. Consequently,
> in the EVPD page,
> we will set state indicating that WRITE_SAME is not supported (even if
> the device supports it).

I think you haven't appreciated the problem: He's passing a SATA SSD via
the SCSI hyper-v interface.  That means that the windows host is doing
SCSI<->ATA translation.  The problem is that the Windows translation
layer (SATL) looks to be incomplete and it's not correctly translating
the IDENTIFY bit that corresponds to TRIM to the correct VPD pages so
consequently, Linux  won't send UNMAP commands to the device (to be
translated back to TRIM).

We already know this is a bug in the Windows SATL which needs fixing (if
you could report it and get a fix, that would be great) and that we're
not going to be able to work around this automatically in Linux because
the proposed patch would have us unconditionally try UNMAP for all
Hyper-V devices.  The current proposed fix is to enable UNMAP manually
via sysfs in the guest boot scripts, but obviously that means that
Hyper-V guests with direct pass through of SSDs need operator
intervention to turn on TRIM.

James