On Thu, 2010-11-11 at 09:53 -0600, James Bottomley wrote:
quoted text
> On Thu, 2010-11-11 at 02:46 -0800, Nicholas A. Bellinger wrote:
> > From: Nicholas Bellinger <nab@linux-iscsi.org>
> > 
> > This patch converts struct scsi_host->cmd_serial_number to a 'odd incremented by 2'
> > atomic_t counter in scsi_cmd_get_serial(), and moves the host_lock held call in
> > jgarzik's DEF_SCSI_QCMD() wrapper back into it's original location in scsi_dispatch_cmd().
> 
> Actually, no ... this isn't really a good idea; you're lengthening our
> critical path here (an atomic costs a lot more than a simple add under
> spinlock).
> 

Fair enough, but this is only less expensive with a spinlock w/p
interrupts disabled, yes..?

quoted text
> There are only a few drivers left that actually make use of a serial
> number.  Of those, the only modern ones are qla4, lpfc, mpt2sas and
> megaraid.
> 

I am not exactly sure that qla4 or lpfc is on the list of LLDs that use
and still need a cmd->serial_number for anything beyond simple printk()
purposes..  

quoted text
> So the next logical step seems to be eliminate the overloading of the
> serial number zero value, which removes the last mid-layer use (dpt_i2o
> seems to abuse this unnecessarily as well), then the serial number code
> can be pushed down into the queuecommand routines of only those drivers
> that actually use it.

Correct, this is what we where originally doing in the
host_lock-less-for-37-v6 series here:

http://git.kernel.org/?p=linux/kernel/git/nab/lio-4.0.git;a=commitdiff;h=72a7203366150...

The LLDs that ended up requring the explict calls for:

mpt2sas: Add scsi_cmd_get_serial() call and set SHT->queuecommand_unlocked()
mpt/fusion: Add scsi_cmd_get_serial() call and set SHT->queuecommand_unlocked()
dpt_i2o: Add scsi_cmd_get_serial() call
eata: Add scsi_cmd_get_serial() call
u14-34f: Add scsi_cmd_get_serial() call

quoted text
> None of the modern ones seems to have a
> legitimate use, so I think their uses can mostly be eliminated.

quoted text
>From the above 5 LLDs, they all appear to be releated to using
cmd->serial_number in their internal error recovery handling.

quoted text
> Thus,
> we might be able to get away with a simple queuecommand push down and
> never bother with atomics for this (since it's unlikely the legacy users
> would convert away from a lock wrapping their queuecommand routines).
> 

Sounds good to me, but you will recall the last attempt to make
scsi_cmd_get_serial() optional for the special case LLDs, that we
started running quickly in the legacy usage of cmd->serial_number in
scsi_softirq_done() and the side effects in scsi_try_to_abort_cmd(), who
use is complex enough that we have not found a proper resolution
sufficent to andmike discussed here:

http://marc.info/?l=linux-scsi&m=128535319915212&w=2
http://marc.info/?l=linux-scsi&m=128820726325009&w=2

Any takers..?  ;)

--nab

quoted text
> James
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/