Re: [PATCH 0/8] Drop host_lock around LLD SHT->queuecommand() caller

["Nicholas A. Bellinger" <nab@linux-iscsi.org>]

On Fri, 2010-09-17 at 14:26 -0400, Jeff Garzik wrote:
> On 09/16/2010 09:31 PM, Nicholas A. Bellinger wrote:
> > On Thu, 2010-09-16 at 18:29 -0700, Tim Chen wrote:
> >> On Thu, 2010-09-16 at 15:35 -0700, Nicholas A. Bellinger wrote:
> >>> From: Nicholas Bellinger<nab@linux-iscsi.org>
> >>>
> >>> Greetings all,
> >>>
> >>> This series contains the first round of a whole-sale conversion for droping
> >>> struct Scsi_Host->host_lock around struct Scsi_Host->queuecommand() within
> >>> scsi_dispatch_cmd().  So with this first patch the only part of scsi_dispatch_cmd()
> >>> that is protected by host_lock is scsi_cmd_get_serial)_.
> >>>
> >>
> >> Maybe we can change the host->cmd_serial_number to atomic and totally
> >> avoid the need to take host_lock.
> >
> > Hmmmm good point, then we would also need an atomic_t for signaling
> > (shost_state == SHOST_DEL) in scsi_dispatch_cmd() to be able to
> > completly drop host_lock usage within scsi_dispatch_cmd().
> >
> > I can take a look at doing this specific part, if you would be willing
> > to have a look at the host->cmd_serial_number conversion piece.  ;)
> 
> But that raises the familiar tale of:  using multiple atomics (w/ their 
> locked instructions) may cost more than a spinlock.

So patch #1 in the v2 series just posted uses the following code:

static inline void scsi_cmd_get_serial(struct Scsi_Host *host, struct scsi_cmnd *cmd)
{
        /*
         * The use of per struct scsi_cmnd->serial_number is disabled by default
         */
        if (!(host->use_serial_number))
                return;
        /*
         * Increment the host->cmd_serial_number by 2 so cmd->serial_number
         * is always odd and wraps to 1 instead of 0.
         */
        cmd->serial_number = atomic_add_return(2, &host->cmd_serial_number);
}


Note that host->use_serial_number is disabled by default, so for the typical case
no atomic_t reading (or writing) is required.

Best,

--nab