James Bottomley wrote:
On Sun, 2007-09-23 at 00:04 -0400, Jeff Garzik wrote:
Rather than sitting on this for far too long, I wanted to go ahead and
get this out there. I heard some chips might be trickling out into
public hands.
The first thing to note is about the specs and the pre-production
hardware: the Linux Foundation has mechanism to get both into the hands
of interested developers; if you can point me to contacts, I can at
least get the NDA documentation programme ball rolling.
Well, are there interested, motivated, skilled developers with time? :)
Otherwise it's a moot point.
This is a bare bones Broadcom 8603 SAS+SATA driver, attempting to use
the vaunted libsas. Notes:
I wouldn't call it "vaunted" but it's been a fun project.
That was sarcasm :)
libsas is a big 'ole pain, that I'm finding has many aic94xx-isms buried
in the lib.
* A quick glance at the FIXMEs will tell you obviously doesn't work.
The first thing I really noted is that SMP and STP protocol support is
stubbed out ... you really can't do anything other than direct device
connection without them.
That's sorta the way I read the hardware docs, too.
I have some engineering questions pending with Broadcom, but from my
read, SMP and STP don't seem supported.
* The hardware is quite simple and straightforward and easy to program
in an efficient way: each SAS port has a command queue (DMA ring) and
a response queue (DMA ring). Or if in SATA mode, just a command
queue.
That's not such a bad way of doing it ... it pretty much mimics the wire
protocol, which is simply frame in/frame out for SAS.
Yep. The hardware (on my end of the spectrum) seems to be moving
towards forcing software to generate all "packets," except (a) data
frames [generated via DMA engine] or (b) special frames that need to
modify the software-generate frame.
* The SAS/SATA negotiation is largely out of our hands. The silicon
does its thing, and then tells us what type of device connected. We
are then expected to switch the port to either SAS mode or SATA mode,
accordingly.
* There is no firmware or anything. Just DMA and register bitbanging.
We have plenty of low-level control.
* The state of SAS/SATA integration is perpetually pathetic. Updates
in this area are likely. There's a rumor Brian King @ IBM may look
into this area too.
* This driver pretty much completely lacks exception handling.
I also note there's a slight nomenclature issue which will trip up SAS
people. All through the driver, you seem to use the word "port" to
refer to a physical phy. the struct bs_port seems to actually be a phy
descriptor ... unless there's some missing phy<->port setup logic that
will be in the final driver? The trouble is that phys and ports are
distinct (and not equivalent) objects in SAS.
Nomemclature came straight from the hardware docs, I'm afraid.
Comparing with the Marvell hardware, I can see how (with Marvell) wide
ports can be set up, and the port/phy distinction is easily programmable
depending on the situation.
Not so with BCM8603.
The only places where the docs mention SMP and STP at all is in the SAS
outgoing DMA descriptor docs, when you fill in connection type. The
_only_ other mention of SMP or STP at all is a note saying neither is
supported. So, even the docs contradict themselves, but overall I have
the feeling that SMP/STP are out of my hands.
I wonder if Broadcom's interface is born out of the closed RAID-on-chip
product that this is descended from.
Hopefully more knowledge will be gained soon, as I debug simple SAS and
SATA device plug/unplug, and ask Broadcom questions.
As an aside, I am also writing a driver for Marvell chips that behave
quite similarly to this chip. It seems the future of storage might look
like these Broadcom and Marvell SAS+SATA DMA ring interfaces, in the
volume marketspace at least.
If you have a contact here too, I can get the LF NDA and hardware
programmes rolling.
Same response as at the top :) Marvell is actually better at responding
than Broadcom, but I'm quite reluctant to make /another/ introduction
(already did so for one hacker) that leads nowhere.
Jeff
-
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
