FCP RECOVERY ABORT OPERATIONS
GFRAZIER at ausvm6.vnet.ibm.com
GFRAZIER at ausvm6.vnet.ibm.com
Mon Aug 1 13:59:15 PDT 1994
In the continuing discussion about FCP recovery aborts, consider
the following potential solution.
If FCP allows an initiator to wait RA_TOV before reusing an XID
as an alternative to requiring targets/initiators to abort
ambiguous exchanges, AND if RA_TOV is set for *zero* (for nondelaying
fabrics) or to the loop circuit time (on a loop), then perhaps
we have a solution which allows the optimization which Kurt needs
for the loop.
Giles Frazier
gfrazier at ausvm6.vnet.ibm.com
IBM Austin
**********LEASE RECENTLY WRITTEN REFERENCED NOTE************************
| The way I look at it is as follows: In class 1 and 2,
| potentially hundreds of ABTS/LS operations do not need to be done if
| FCP remains unchaged. This is because a very large number of exchanges
| may exist but a very small number of sequences may be ambiguous at the
| target. (Please refer to FCP 8b for definition of ambiguous.)
| This situation occurs when the initiator has a very
| large number of commands (open exchanges) at the target but the target
| only has a single ambiguous exchange. (e.g. one of the exchanges has
| an unacknowledged frame but all the rest of the exchanges are
| quiescient--the target has acknowledged the command but is accessing
| the data.)
I agree with Giles on this point. There are interlock points where
one side owns Sequence Initiative and the other side knows it owns it.
If a Task Management flag is received (or a Unit Attention condition
established) at these points, no Recovery Abort is needed. In class
3, three out of the 7 possible Sequence launch points fall into this
category as noted by my previous memo.
This is a SEPARATE issue from whether or not Targets are required to
initiate ABTS for all the other cases where the Exchange is "ambiguous".
| There is another solution which may pertain more to class 3 but could
| also be used: The initiator may wait RA_TOV before reusing any of the
| XID's.
|
| This may be the only reliable method for class 3 anyway.
Not true. Class 3 local loops or fabrics which cannot delay frames do
not require the N_Ports to wait R_A_TOV - the behavior is identical to
class 1 in such cases. FCP should not rule out optimizations for
these topologies.
*********************END OF REFERENCED NOTE
More information about the T10
mailing list