[PATCH 0/2] [tgt]: Add proper SG_IO struct cmd CDB passthrough

[PATCH 0/2] [tgt]: Add proper SG_IO struct cmd CDB passthrough

[Nicholas A. Bellinger]

From: Nicholas Bellinger <nab@linux-iscsi.org>

Greetings Tomo and Mike,

These are two patches to allow the STGT SG_IO backstore to function using a
new struct scsi_cmd->sg_iovec member with sg_io_hdr->iovec_count=1, and to update
usr/scsi.c:scsi_cmd_perform() to allow for a hand off incoming descriptors into the
struct device_type_template->cmd_passthrough() function pointer.

I initially updated usr/sbc.c to set sbc_rw() as it's CDB passthrough function for
TYPE_DISK, but the logic should apply to any device types *_rw() functions.  Please let
me know if this makes sense, and I will look at updating the other stgt device types.

So far with this patch, I am able to successfully mkfs new iSCSI TYPE_DISK LUNS with
Open-iSCSI to the userspace STGT iSCSI fabric module + SG_IO into kernel level TCM_Loop
virtual iSCSI target ports on a x86 HVM, and bulk 512KB DATA_SG_IO payloads between STGT
and TCM_Loop are now functioning as expected.

There is one unresolved issue that I ran into while running initially with TCM_Loop
and struct scsi_device->max_sectors=256.  SG_IO writes larger than the
struct scsi_device->max_sectors are being returned with -EINVAL, but it appears that
the CHECK_CONDITION is not actually being propigated back into STGT/iSCSI.

After increasing the TCM_Loop ports to struct scsi_device->max_sectors=1024 following
the Open-iSCSI Initiator default, the 512KB paylods begin work as expected.  I will need
to have a look at the case for CDB length > max_sectors, which means that bs_sg.c getting
the correct scsi_status and sense data out of the kernel might still need more work.

In the mean time, please have a look at applying these patches.

Best,

Signed-off-by: Nicholas A. Bellinger <nab@linux-iscsi.org>

Nicholas Bellinger (2):
  [tgt]: Add proper CDB passthrough for SG_IO backstores
  [tgt]: Add support for SG_IO CDB passthrough in scsi_cmd_perform()

 usr/bs_sg.c     |   31 ++++++++++++++++++++++++-------
 usr/sbc.c       |    1 +
 usr/scsi.c      |   11 +++++++++++
 usr/scsi_cmnd.h |    4 ++++
 usr/tgtd.h      |    2 ++
 5 files changed, 42 insertions(+), 7 deletions(-)

Re: [PATCH 0/2] [tgt]: Add proper SG_IO struct cmd CDB passthrough

[FUJITA Tomonori]

On Mon, 17 May 2010 22:12:54 -0700
"Nicholas A. Bellinger" <nab@linux-iscsi.org> wrote:

> These are two patches to allow the STGT SG_IO backstore to function using a
> new struct scsi_cmd->sg_iovec member with sg_io_hdr->iovec_count=1, and to update
> usr/scsi.c:scsi_cmd_perform() to allow for a hand off incoming descriptors into the
> struct device_type_template->cmd_passthrough() function pointer.
> 
> I initially updated usr/sbc.c to set sbc_rw() as it's CDB passthrough function for
> TYPE_DISK, but the logic should apply to any device types *_rw() functions.  Please let
> me know if this makes sense, and I will look at updating the other stgt device types.
> 
> So far with this patch, I am able to successfully mkfs new iSCSI TYPE_DISK LUNS with
> Open-iSCSI to the userspace STGT iSCSI fabric module + SG_IO into kernel level TCM_Loop
> virtual iSCSI target ports on a x86 HVM, and bulk 512KB DATA_SG_IO payloads between STGT
> and TCM_Loop are now functioning as expected.
> 
> There is one unresolved issue that I ran into while running initially with TCM_Loop
> and struct scsi_device->max_sectors=256.  SG_IO writes larger than the
> struct scsi_device->max_sectors are being returned with -EINVAL, but it appears that
> the CHECK_CONDITION is not actually being propigated back into STGT/iSCSI.
> 
> After increasing the TCM_Loop ports to struct scsi_device->max_sectors=1024 following
> the Open-iSCSI Initiator default, the 512KB paylods begin work as expected.  I will need
> to have a look at the case for CDB length > max_sectors, which means that bs_sg.c getting
> the correct scsi_status and sense data out of the kernel might still need more work.
> 
> In the mean time, please have a look at applying these patches.
> 
> Best,
> 
> Signed-off-by: Nicholas A. Bellinger <nab@linux-iscsi.org>
> 
> Nicholas Bellinger (2):
>   [tgt]: Add proper CDB passthrough for SG_IO backstores
>   [tgt]: Add support for SG_IO CDB passthrough in scsi_cmd_perform()
> 
>  usr/bs_sg.c     |   31 ++++++++++++++++++++++++-------
>  usr/sbc.c       |    1 +
>  usr/scsi.c      |   11 +++++++++++
>  usr/scsi_cmnd.h |    4 ++++
>  usr/tgtd.h      |    2 ++
>  5 files changed, 42 insertions(+), 7 deletions(-)

Thanks! Surely, the first patch looks fine.

I thought that you want to send all SCSI commands to your in-kernel
target code via SG_IO. That's, you want the proper 'pass-through' code
since tgt's pass-through code is broken.

However, looks like non I/O commands are still performed in user-space
tgt code and only I/O commands goes to kernel via SG_IO. Hmm, that's
what you want to do?

Re: [PATCH 0/2] [tgt]: Add proper SG_IO struct cmd CDB passthrough

[Nicholas A. Bellinger]

On Tue, 2010-05-18 at 13:39 +0900, FUJITA Tomonori wrote:
> On Mon, 17 May 2010 22:12:54 -0700
> "Nicholas A. Bellinger" <nab@linux-iscsi.org> wrote:
> 
> > These are two patches to allow the STGT SG_IO backstore to function using a
> > new struct scsi_cmd->sg_iovec member with sg_io_hdr->iovec_count=1, and to update
> > usr/scsi.c:scsi_cmd_perform() to allow for a hand off incoming descriptors into the
> > struct device_type_template->cmd_passthrough() function pointer.
> > 
> > I initially updated usr/sbc.c to set sbc_rw() as it's CDB passthrough function for
> > TYPE_DISK, but the logic should apply to any device types *_rw() functions.  Please let
> > me know if this makes sense, and I will look at updating the other stgt device types.
> > 
> > So far with this patch, I am able to successfully mkfs new iSCSI TYPE_DISK LUNS with
> > Open-iSCSI to the userspace STGT iSCSI fabric module + SG_IO into kernel level TCM_Loop
> > virtual iSCSI target ports on a x86 HVM, and bulk 512KB DATA_SG_IO payloads between STGT
> > and TCM_Loop are now functioning as expected.
> > 
> > There is one unresolved issue that I ran into while running initially with TCM_Loop
> > and struct scsi_device->max_sectors=256.  SG_IO writes larger than the
> > struct scsi_device->max_sectors are being returned with -EINVAL, but it appears that
> > the CHECK_CONDITION is not actually being propigated back into STGT/iSCSI.
> > 
> > After increasing the TCM_Loop ports to struct scsi_device->max_sectors=1024 following
> > the Open-iSCSI Initiator default, the 512KB paylods begin work as expected.  I will need
> > to have a look at the case for CDB length > max_sectors, which means that bs_sg.c getting
> > the correct scsi_status and sense data out of the kernel might still need more work.
> > 
> > In the mean time, please have a look at applying these patches.
> > 
> > Best,
> > 
> > Signed-off-by: Nicholas A. Bellinger <nab@linux-iscsi.org>
> > 
> > Nicholas Bellinger (2):
> >   [tgt]: Add proper CDB passthrough for SG_IO backstores
> >   [tgt]: Add support for SG_IO CDB passthrough in scsi_cmd_perform()
> > 
> >  usr/bs_sg.c     |   31 ++++++++++++++++++++++++-------
> >  usr/sbc.c       |    1 +
> >  usr/scsi.c      |   11 +++++++++++
> >  usr/scsi_cmnd.h |    4 ++++
> >  usr/tgtd.h      |    2 ++
> >  5 files changed, 42 insertions(+), 7 deletions(-)
> 
> Thanks! Surely, the first patch looks fine.
> 

Great!   Please let me know if I should also generate patches for
conversion of remaining STGT struct device_type_templates for proper
SG_IO passthrough support.

> I thought that you want to send all SCSI commands to your in-kernel
> target code via SG_IO. That's, you want the proper 'pass-through' code
> since tgt's pass-through code is broken.
> 
> However, looks like non I/O commands are still performed in user-space
> tgt code and only I/O commands goes to kernel via SG_IO. Hmm, that's
> what you want to do?

Hmmm, not sure what you mean here..

AFAICT using the newly provided struct
device_type_template->cmd_passthrough() instead CDB emulation specific
callers in usr/scsi.c:scsi_cmd_perform() with device type template
   
	cmd->dev->dev_type_template.ops[op].cmd_perform(()

will always capture incoming struct scsi_cmnd for SG_IO passthrough
before the device type CDB userspace STGT emulation execution and hand
off into .cmd_perform()..?

This is why the patches to use sbc_rw() as device type passthrough works
for TYPE_DISK into SG_IO backstores with [READ,WRITE]_[6,10,12,16])
CDBs, yes?

Best,

--nab

Re: [PATCH 0/2] [tgt]: Add proper SG_IO struct cmd CDB passthrough

[Nicholas A. Bellinger]

On Mon, 2010-05-17 at 22:12 -0700, Nicholas A. Bellinger wrote:
> From: Nicholas Bellinger <nab@linux-iscsi.org>
> 
> Greetings Tomo and Mike,
> 
> These are two patches to allow the STGT SG_IO backstore to function using a
> new struct scsi_cmd->sg_iovec member with sg_io_hdr->iovec_count=1, and to update
> usr/scsi.c:scsi_cmd_perform() to allow for a hand off incoming descriptors into the
> struct device_type_template->cmd_passthrough() function pointer.
> 
> I initially updated usr/sbc.c to set sbc_rw() as it's CDB passthrough function for
> TYPE_DISK, but the logic should apply to any device types *_rw() functions.  Please let
> me know if this makes sense, and I will look at updating the other stgt device types.
> 
> So far with this patch, I am able to successfully mkfs new iSCSI TYPE_DISK LUNS with
> Open-iSCSI to the userspace STGT iSCSI fabric module + SG_IO into kernel level TCM_Loop
> virtual iSCSI target ports on a x86 HVM, and bulk 512KB DATA_SG_IO payloads between STGT
> and TCM_Loop are now functioning as expected.
> 
> There is one unresolved issue that I ran into while running initially with TCM_Loop
> and struct scsi_device->max_sectors=256.  SG_IO writes larger than the
> struct scsi_device->max_sectors are being returned with -EINVAL, but it appears that
> the CHECK_CONDITION is not actually being propigated back into STGT/iSCSI.
> 
> After increasing the TCM_Loop ports to struct scsi_device->max_sectors=1024 following
> the Open-iSCSI Initiator default, the 512KB paylods begin work as expected.  I will need
> to have a look at the case for CDB length > max_sectors, which means that bs_sg.c getting
> the correct scsi_status and sense data out of the kernel might still need more work.
> 
> In the mean time, please have a look at applying these patches.
> 
> Best,
> 
> Signed-off-by: Nicholas A. Bellinger <nab@linux-iscsi.org>
> 
> Nicholas Bellinger (2):
>   [tgt]: Add proper CDB passthrough for SG_IO backstores
>   [tgt]: Add support for SG_IO CDB passthrough in scsi_cmd_perform()
> 
>  usr/bs_sg.c     |   31 ++++++++++++++++++++++++-------
>  usr/sbc.c       |    1 +
>  usr/scsi.c      |   11 +++++++++++
>  usr/scsi_cmnd.h |    4 ++++
>  usr/tgtd.h      |    2 ++
>  5 files changed, 42 insertions(+), 7 deletions(-)
> 

Hi Tomo-san,

Ping on these two patches for SG_IO backstores with STGT.  ;)

Please let me know if these are OK for you or if you would like me to
add or change anything and resubmit.

Best,

--nab