RE: iSCSI: Out Of Sequence due to null sequence with multiple connections.

["Douglas Otis" <dotis@sanlight.net>]

Sandeep,

I understand your concern as you view the sequencer as potentially holding
an inordinate amount of commands.  Flow control should help keep this queue
lean.  Consider the consequence of a connection failure creating a series of
holes in the sequence.  The alternative to an immediate rejection of
commands may be to time-out on each hole in the sequencer.  In this case,
resending a list of rejected commands is a better solution than an endless
sequence of timeouts if this queue is really that large.  Data wise, this
list should represent only a small amount to be resent.  If the CmdSN was
not session but LUN wide, then there would be less recovery.  I do not see
this happening often enough for there to be a concern about the number of
commands rejected.

Doug


> Black_David@emc.com wrote:
> >
> > Sandeep,
> >
> > > Still wading thru related emails but I believe that if a refCmdSN
> > > is added to the task management PDU (not present currently but
> > > could be added for task-related management commands), then it
> > > might fix the above-mentioned flaws and allow for safe execution
> > > and immediate delivery of the abort task to the target.
> >
> > I think this is a functional subset of Doug Otis's suggestion to always
> > use a CmdSN and add a header flag indicating that the command should
> > be executed immediately at the target rather than waiting for those with
> > prior CmdSNs to arrive.  Doug's suggestion also consumes less space
> > in the header.
>
> Yes, i now understand Doug's proposal.  The "PDU sequencer" was
> new terminology which baffled me some.. :-)
>
> His solution does seem like a cleaner approach (no target state)
>
> In some ways, it reminds one of the M_FLUSH function used
> in the STREAMS layer.   If we stand back a bit, we see that the
> scenario is essentially the same -> Stream module processing is
> normally asynchronous.   With M_FLUSH, the stream head desires to
> clean out all the downstream module queues.
>
> The only concern with that solution is that it may delete
> a whole lot of unrelated SCSI I/O tasks as the sequencer cmdSN
> moves way up to the cmdSN of the task_mgmt command.  "Abort task"
> as such will not exist in the iSCSI world.   These unrelated
> commands will now have to be retried by the initiator.
>
> Do we have any numbers on how often ABORT_TASK/ABORT_TASK_SET actually
> occurs in filesystem/SCSI workloads ?
>
> regards,
> -Sandeep
>