Without item d) taken into consideration, this an example sequence I am trying to describe:
 
Received Dwords              DWS state    SP state 
invalid                   DWS0        SP29
TRAIN_DONE(3rd)               
TRAIN_DONE(4th)           DWS1 
TRAIN_DONE(5th)           DWS2
TRAIN_DONE(6th)           DWS3  -->   SP30 (dword sync acquired here)
zero scrambled(58 times)
random scrambled "forever"
         
 
This is how I interpret the state machine flow without the new d) item.  I think where we may be seeing things differently is based on your comment about being in SP29 for the duration of the training pattern.  My assumption was that SP transmitter enforces the training pattern "wholeness", and perhaps your assumption is that the SP state machine enforces the "wholeness" of the pattern. (Not I'm kinda guessing here at where we're not together in thought)
I think I would agree with your concept of nothing getting missed if the SP state machine only transitioned at the pattern boundaries... but that's not how I read the wording.  Am I forgetting/missing a statement in the spec that indicates that?
 
Brian
 

From: Bill.Martin@Emulex.Com [mailto:Bill.Martin@Emulex.Com]
Sent: Wednesday, November 05, 2008 3:43 PM
To: Day, Brian; t10@t10.org
Subject: RE: Response to 08-425r0

Brian:

 

The TRAIN_DONE received message is sent by the SP Receiver to the SP state machine.  The SP Receiver is responsible for recognizing the three primitives, so this will either be seen in the SP29 or SP30 state of the SP state machine, but it will not be missed since detecting three is not up to the SP state machine.  The device will be in the SP29 state for at least one full TRAIN or TRAIN_DONE pattern in order to complete training.  If any time during that process a TRAIN_DONE received message is received it will set the variable that is sent to the SP30 state.  I do not see any way that this message will be missed.  The previous problem was that SP29 was not looking for the TRAIN_DONE received message, so if it was sent just before the transition to SP30 then there would not be another one.

 

Bill Martin

Emulex
Office of Technology
Industry Standards
916 772-3658
916 765-6875 (Cell)
bill.martin@emulex.com

From: Day, Brian [mailto:Brian.Day@lsi.com]
Sent: Wednesday, November 05, 2008 2:28 PM
To: Martin, Bill; t10@t10.org
Subject: RE: Response to 08-425r0

 

Hi Bill...

 

I'm suggesting the new item d) is required so we don't have the exact same race condition issue when the "slower" trained device acquires dword sync on the 4th, 5th, or 6th dword of a TRAIN_DONE (i.e. there aren't enough of the TRAIN_DONE dwords left to detect 3 out of 6 redundant rule in the primitive sequence)

 

Without item d), it would be possible to then transition to SP30 without detecting the TRAIN_DONE, and have the faster device move to SAS_PHY_Ready, and we've got the same problem again.

 

Essentially, I'm suggesting a device doesn't say TRAIN_DONE until it's reached the point where it's actually seeing/decoding the other devices primitives, instead of just acquiring dword sync as the gating factor.

 

Brian

 

From: Bill.Martin@Emulex.Com [mailto:Bill.Martin@Emulex.Com]
Sent: Wednesday, November 05, 2008 12:16 PM
To: Day, Brian; t10@t10.org
Subject: Response to 08-425r0

Brian:

 

I understood the need to detect the TRAIN_DONE primitive in the training state; however, I do not see the reason that a TRAIN or TRAIN_DONE has to be received before exiting this state.  The device will continue sending TRAIN_DONE until a TRAIN_DONE is received, so the device that is later entering the training sequence will see a TRAIN or TRAIN_DONE sequence during its training process, and should be able to train on either pattern.

 

I would like to see the new item d) in 6.8.4.12.5.

 

I am sorry that I did not catch this part of your proposal in our discussions on Monday.

 

Bill Martin

Emulex
Office of Technology
Industry Standards
916 772-3658
916 765-6875 (Cell)
bill.martin@emulex.com