On 13-12-15 09:52 AM, Stefan Richter wrote:
On Dec 02 2012 Stefan Richter wrote:
On Nov 26 Martin K. Petersen wrote:
"Stefan" == Stefan Richter <stefanr@xxxxxxxxxxxxxxxxx> writes:
Stefan> I did not try "sg_write_same" on any of the devices; but since
Stefan> the two SPC-3 devices are correctly identified as "fully
Stefan> provisioned", won't issue WRITE SAME to them either.
^[the kernel]
What if you have an SSD behind one of them?
At the moment I only have a single old SSD available which does not
implement ATA TRIM as far as I recall.
And the two mentioned OXUF936QSE based SPC-3 devices are four-bay SATA disk
enclosures whose firmwares only support various RAID modes and require at
least two bays to be populated. I.e. I can't test them with the SSD for
now. But I suspect that they don't implement thin provisioning anyway,
particularly translation of WRITE SAME with UNMAP to ATA TRIM.
But now I found another SPC-3 compliant device in my stash; a dual SATA
bridge based on OXUF934DSB which supports JBOD with 1...2 disks
alternatively to striping/ spanning/ mirroring over 2 disks. I attached
the old SSD to it, and its thin_provisioning sysfs attribute was shown as
0 as well. "sg_write_same -U ..." on this device in the 10 and 16 byte
variants ended with Illegal Request/ Invalid command operation code, but
otherwise without discernible malfunction.
Stefan> Hence let's remove the no_report_opcodes and no_write_same
Stefan> blacklist flags so that these commands can be used on
Stefan> respectively capable targets.
I just erred on the side of caution. If you are happy without belt and
suspenders that's perfectly ok with me :)
Blacklisting at first was definitely the right approach. But now that I
looked at a variety of older and newer devices, I am confident that the
general Inquiry_Data.Version >= SPC-3 test keeps the wackier among the
SBP-2 devices safe enough.
(I followed up with https://git.kernel.org/linus/b0ea5f19d3d8.)
Of course it remains to be seen what happens with ATA TRIM enabled SSDs
behind the newer SPC-3 compliant bridges, but at this time the risk with
those seems low.
I now tested
- ONNTO dataTale RSM4QO (OXUF936QSE) with two Samsung 840 Pro in RAID 0,
- ONNTO dataTale RSM4QO (OXUF936QSE) with two Samsung 840 Pro in RAID 1,
- IOI FWBU2-DSATA12 (OXUF934DSB) with one Samsung 840 Pro
and kernel 3.9.
$ grep . /sys/class/scsi_disk/42\:0\:0\:0/*prov*
/sys/class/scsi_disk/42:0:0:0/provisioning_mode:full
/sys/class/scsi_disk/42:0:0:0/thin_provisioning:0
# sg_opcodes /dev/sdg
Ext Hard Disk
Peripheral device type: disk
Report supported operation codes: operation not supported
# sg_write_same --10 --lba=1 /dev/sdg
Write same(10) command not supported
# sg_write_same --16 --lba=1 /dev/sdg
Write same(16) command not supported
# sg_write_same --32 --lba=1 /dev/sdg
Write same: pass through os error: Invalid argument
Write same(32) command failed
The sg driver does not support SCSI commands of greater
than 16 bytes. That is the reason for the "pass-through
os error" on your '--32' variant above. For now you need
to send the '--32' variant to the bsg driver instead.
I have presented a patch to relax the 16 byte cdb limit
on the sg driver:
http://www.spinics.net/lists/linux-scsi/msg70283.html
Like most of my patches that one seems to be residing
in James' /dev/null file for later consideration.
Doug Gilbert
I will send a patch which reverts the drivers/firewire/sbp2.c hunk of
https://git.kernel.org/linus/54b2b50c20a6 "[SCSI] Disable WRITE SAME for
RAID and virtual host adapter drivers". (As an aside, sbp2.c implements a
transport, not a virtual host adapter.)
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
