Re: [PATCH] scsi: Convert scsi_host->cmd_serial_number to oddnumbered atomic_t counter

[Nicholas A. Bellinger]

On Thu, 2010-11-11 at 15:55 -0600, James Bottomley wrote:
> On Thu, 2010-11-11 at 13:37 -0800, Nicholas A. Bellinger wrote:
> > On Thu, 2010-11-11 at 09:53 -0600, James Bottomley wrote:
> > > On Thu, 2010-11-11 at 02:46 -0800, Nicholas A. Bellinger wrote:
> > > > From: Nicholas Bellinger <nab@xxxxxxxxxxxxxxx>
> > > > 
> > > > 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..?
> 
> It's less expensive than the explicit lock around just the serial
> number, yes ... but if all use cases use an explicit lock, we can just
> use a simple inc in there without bothering with atomics.
> 

Sure, that makes sense for the legacy drivers running in host_lock
mode..  But with the fully lock-less LLDs (like the freshly minted
tcm_loop patch to go fully lock-less), this is a easier transitional
step from:

full host_lock push down ->

	lock_less w/ atomic_t host serial_number counter ->

		lock_less w/o atomic cmd->serial_number counter + 
                scsi_error.c changes

So my tenative plan is to enable the host_lock less LLDs we know about
(include tcm_loop and core-iscsi) with the atomic_t serial number
counter, and continue getting polling feedback from various LLD
maintainers..  This part will be tagged as a for-38 branch.

> > > 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..  
> > 
> > > 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=72a72033661506ead54e0f366218fd0aef2e5339
> > 
> > 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
> > 
> > > None of the modern ones seems to have a
> > > legitimate use, so I think their uses can mostly be eliminated.
> > 
> > >From the above 5 LLDs, they all appear to be releated to using
> > cmd->serial_number in their internal error recovery handling.
> > 
> > > 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:
> 
> Yes, that's what I meant by "eliminate the overloading of the serial
> number zero value" above.  This needs fixing before the serial number
> can be dumped for fast hba drivers.
> 

Understood, so at least for the moment this is not an immediate
addressable item unless andmike, hch or someone else wants to take
another look at extracting the ghosts from this code.

Until then I will keep the atomic_t serial_number counter in
lio-core-2.6.git code, and plan to push the usage back down for the 5
legacy LLDs once scsi_error.c can be sorted out.

Best,

--nab



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