Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Jeff Garzik]

Mark Lord wrote:
> Alan Cox wrote:
>> Something like this then ?
>>
>>
>> The legacy old IDE API for this is a bit primitive so we try and map 
>> stuff
>> sensibly onto it.
>>
>> - Add flags2 as we ran out of flags
>> - Set PIO over DMA devices to report 32bit
>> - Add ability to change the PIO32 settings if the controller permits it
>> - Add that functionality into the sff drivers
>> - Add that functionality into the VLB legacy driver
>> - Turn on the 32bit PIO on the ninja32 and add support there
>>
>> Signed-off-by: Alan Cox <alan@linux.intel.com>
> ..
> 
> By all means go ahead with something like this,
> and thanks for doing it, Alan.
> 
> But.. these do look rather suspiciously like sysfs attrs to me.
> 
> We didn't have a sysfs in the IDE days (so Gadi added the klunky
> procfs "settings" thing for there), but we do now.
> 
> I'm happy to update hdparm to check sysfs if that's where this
> ends up, or to do almost nothing for it if we stick with the ioctls.

hmmm, Tejun's storage summit presentation would be quite relevant at 
this juncture :)

The short answer:  we need a control mechanism for tasks such SATA PHY 
twiddling, host controller reset, host controller configuration 
(set/clear 32-bit I/O in this example), ...

sysfs attrs are desired for this, but we've held off largely because 
this is all intertwined with the long term direction of "storage model", 
"transport protocol" and libata's eventual move away from strict SCSI 
dependency.

I think the best place for sysfs attr attachment is an ata_transport 
module, something that's needed to be written for quite a while.  You 
can see some of the eventual direction in scsi_transport_*.[ch] and 
raid_class.c.

Eventually I hope to reach a point where the current crop of SATA+SAS 
controllers out in the field (e.g. mvsas) will attach to 
scsi_transport_sas or ata_transport, depending on controller port 
configuration and attached device(s).  And for the sake of users' 
sanity, a single utility should be used to configure a SATA+SAS 
controller port.  Requiring use of hdparm when the port is in SATA mode, 
and sg_* when the port is in SAS mode, is a bit annoying.

	Jeff

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Mark Lord]

Jeff Garzik wrote:
> a single utility should be used to configure a SATA+SAS 
> controller port.  Requiring use of hdparm when the port is in SATA mode, 
> and sg_* when the port is in SAS mode, is a bit annoying.
..

That's why hdparm uses sg_* whenever possible already.

-ml

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Jeff Garzik]

Mark Lord wrote:
> Jeff Garzik wrote:
>> a single utility should be used to configure a SATA+SAS controller 
>> port.  Requiring use of hdparm when the port is in SATA mode, and sg_* 
>> when the port is in SAS mode, is a bit annoying.
> ..
> 
> That's why hdparm uses sg_* whenever possible already.

Does hdparm really call out to external applications sg_senddiag, 
sg_opcodes, sg_inq, sg_turs, etc.?

	Jeff

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Douglas Gilbert]

Jeff Garzik wrote:
> Mark Lord wrote:
>> Alan Cox wrote:
>>> Something like this then ?
>>>
>>>
>>> The legacy old IDE API for this is a bit primitive so we try and map 
>>> stuff
>>> sensibly onto it.
>>>
>>> - Add flags2 as we ran out of flags
>>> - Set PIO over DMA devices to report 32bit
>>> - Add ability to change the PIO32 settings if the controller permits it
>>> - Add that functionality into the sff drivers
>>> - Add that functionality into the VLB legacy driver
>>> - Turn on the 32bit PIO on the ninja32 and add support there
>>>
>>> Signed-off-by: Alan Cox <alan@linux.intel.com>
>> ..
>>
>> By all means go ahead with something like this,
>> and thanks for doing it, Alan.
>>
>> But.. these do look rather suspiciously like sysfs attrs to me.
>>
>> We didn't have a sysfs in the IDE days (so Gadi added the klunky
>> procfs "settings" thing for there), but we do now.
>>
>> I'm happy to update hdparm to check sysfs if that's where this
>> ends up, or to do almost nothing for it if we stick with the ioctls.
> 
> hmmm, Tejun's storage summit presentation would be quite relevant at 
> this juncture :)
> 
> The short answer:  we need a control mechanism for tasks such SATA PHY 
> twiddling, host controller reset, host controller configuration 
> (set/clear 32-bit I/O in this example), ...
> 
> sysfs attrs are desired for this, but we've held off largely because 
> this is all intertwined with the long term direction of "storage model", 
> "transport protocol" and libata's eventual move away from strict SCSI 
> dependency.
> 
> I think the best place for sysfs attr attachment is an ata_transport 
> module, something that's needed to be written for quite a while.  You 
> can see some of the eventual direction in scsi_transport_*.[ch] and 
> raid_class.c.
> 
> Eventually I hope to reach a point where the current crop of SATA+SAS 
> controllers out in the field (e.g. mvsas) will attach to 
> scsi_transport_sas or ata_transport, depending on controller port 
> configuration and attached device(s).  And for the sake of users' 
> sanity, a single utility should be used to configure a SATA+SAS 
> controller port.  Requiring use of hdparm when the port is in SATA mode, 
> and sg_* when the port is in SAS mode, is a bit annoying.

The original question seemed to be about PATA. sysfs
attributes would be one solution. Another solution is
for libata to implement the SAT (ref: sat2r06.pdf at
www.t10.org) specific mode page called "PATA Control"
(see section 12.3.2). The latter solution may well work
with FCoE, iSCSI, FC, USB, 1394 and SAS. Currently there
is no SATA equivalent to that mode page.
In the PATA case surprise me and choose the better solution.

Getting to the host phy directly connected to a SATA disk
is a particularly thorny problem and a no go area for
sysfs in all but the direct connect case. For example
a SATA "host" phy connected to a SATA disk attached to
a SAS expander can be controlled via the SAS SMP protocol.

hdparm assumes the device it is talking to is a PATA or
SATA disk (with some support for the ATA part of an
ATAPI device). hdparm is now capable of talking to
such devices irrespective of whether the OS sees a
ATA transport (PATA or SATA) or one of those other (evil)
SCSI friendly transports (same list as above). My
sg3_utils are lower level utilities that mostly map
directly to SCSI or ATA commands. Feel free to write
a utility that unifies the attributes of
PATA/SATA/FC/SCSI/SAS/logical disks and OSD devices :-)

As for the freestanding ata_transport idea which transport
would a mvsas host adapter belong to which has one phy
connected to a SATA disk, another connected to a SAS disk and
a third one connected to a SAS expander with lots of SATA
disks the other side of that?

Doug Gilbert

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Jeff Garzik]

Douglas Gilbert wrote:
> As for the freestanding ata_transport idea which transport
> would a mvsas host adapter belong to which has one phy
> connected to a SATA disk, another connected to a SAS disk and
> a third one connected to a SAS expander with lots of SATA
> disks the other side of that?

Only the direct attach example would use ata_transport.

	Jeff

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Jeff Garzik]

Alan Cox wrote:
>> But.. these do look rather suspiciously like sysfs attrs to me.
> 
> I'd rather any future sysfs interface actually exposed something more
> meaningful. This is a legacy fixup 8)
> 
>> We didn't have a sysfs in the IDE days (so Gadi added the klunky
>> procfs "settings" thing for there), but we do now.
>>
>> I'm happy to update hdparm to check sysfs if that's where this
>> ends up, or to do almost nothing for it if we stick with the ioctls.
> 
> I'm not sure sysfs helps much anyway - you have to open the device file
> and keep it open while accessing the sysfs nodes anyway (something huge
> numbers of apps hopelessly fail to do so)

FWIW...  here is the sysfs work I referred to (in a message sent several 
days ago in this thread)...

http://lwn.net/Articles/294608/

	Jeff

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Alan Cox]

On Mon, 13 Apr 2009 12:32:57 -0400
Jeff Garzik <jeff@garzik.org> wrote:

> > I'm not sure sysfs helps much anyway - you have to open the device file
> > and keep it open while accessing the sysfs nodes anyway (something huge
> > numbers of apps hopelessly fail to do so)  
> 
> FWIW...  here is the sysfs work I referred to (in a message sent several 
> days ago in this thread)...
> 
> http://lwn.net/Articles/294608/

Which indeed shows the same problems. There is nothing to stop changes in
the rest of the topology from causing me to write to the sysfs at the
wrong moment and reconfigure/misconfigure a different object to the one
intended.

Alan

Re: [PROPOSED] ata: Report 16/32bit PIO as best we can

[Jeff Garzik]

Alan Cox wrote:
> On Mon, 13 Apr 2009 12:32:57 -0400
> Jeff Garzik <jeff@garzik.org> wrote:
> 
>>> I'm not sure sysfs helps much anyway - you have to open the device file
>>> and keep it open while accessing the sysfs nodes anyway (something huge
>>> numbers of apps hopelessly fail to do so)  
>> FWIW...  here is the sysfs work I referred to (in a message sent several 
>> days ago in this thread)...
>>
>> http://lwn.net/Articles/294608/
> 
> Which indeed shows the same problems. There is nothing to stop changes in
> the rest of the topology from causing me to write to the sysfs at the
> wrong moment and reconfigure/misconfigure a different object to the one
> intended.

The horse has already left the barn, on that one...

Google's ata transport class is consistent with existing transport class 
work in the kernel.  It is also consistent with recent admonitions in 
the osdblk thread, regarding the "one piece of data per sysfs file" rule.

Personally I think a netlink-like approach to managing and controlling 
SAS and ATA would be better, but that's not what gets merged...

	Jeff