Re: [PATCH v3 18/77] ncr5380: Eliminate USLEEP_WAITLONG delay

From: Finn Thain
Date: Tue Dec 22 2015 - 07:38:44 EST



On Tue, 22 Dec 2015, Hannes Reinecke wrote:

> On 12/22/2015 02:17 AM, Finn Thain wrote:
> > Linux 2.1.105 introduced the USLEEP_WAITLONG delay, apparently "needed for
> > Mustek scanners". It is intended to stall the issue queue for 5 seconds.
> > There are a number of problems with this.
> >
> > 1. Only g_NCR5380 enables the delay, which implies that the other five
> > drivers using the NCR5380.c core driver remain incompatible with
> > Mustek scanners.
> >
> > 2. The delay is not implemented by atari_NCR5380.c, which is problematic
> > for re-unifying the two core driver forks.
> >
> > 3. The delay is implemented using NCR5380_set_timer() which makes it
> > unreliable. A new command queued by the mid-layer cancels the delay.
> >
> > 4. The delay is applied indiscriminately in several situations in which
> > NCR5380_select() returns -1. These are-- reselection by the target,
> > failure of the target to assert BSY, and failure of the target to
> > assert REQ. It's clear from the comments that USLEEP_WAITLONG is not
> > relevant to the reselection case. And reportedly, these scanners do
> > not disconnect.
> >
> > 5. atari_NCR5380.c was forked before Linux 2.1.105, so it was spared some
> > of the damage done to NCR5380.c. In this case, the atari_NCR5380.c core
> > driver was more standard-compliant and may not have needed any
> > workaround like the USLEEP_WAITLONG kludge. The compliance issue was
> > addressed in the previous patch.
> >
> > If these scanners still don't work, we need a better solution. Retrying
> > selection until EH aborts a command offers equivalent robustness. Bugs in
> > the existing driver prevent EH working correctly but this is addressed in
> > a subsequent patch. Remove USLEEP_WAITLONG.
> >
> > Signed-off-by: Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx>
> >
> > ---
> > drivers/scsi/NCR5380.c | 19 +++++--------------
> > drivers/scsi/g_NCR5380.c | 1 -
> > 2 files changed, 5 insertions(+), 15 deletions(-)
> >
> > Index: linux/drivers/scsi/NCR5380.c
> > ===================================================================
> > --- linux.orig/drivers/scsi/NCR5380.c 2015-12-22 12:15:51.000000000 +1100
> > +++ linux/drivers/scsi/NCR5380.c 2015-12-22 12:15:52.000000000 +1100
> > @@ -468,10 +468,6 @@ static void NCR5380_print_phase(struct S
> > #ifndef USLEEP_POLL
> > #define USLEEP_POLL msecs_to_jiffies(200)
> > #endif
> > -#ifndef USLEEP_WAITLONG
> > -/* RvC: (reasonable time to wait on select error) */
> > -#define USLEEP_WAITLONG USLEEP_SLEEP
> > -#endif
> >
> > /*
> > * Function : int should_disconnect (unsigned char cmd)
> > @@ -619,8 +615,8 @@ static void prepare_info(struct Scsi_Hos
> > "can_queue %d, cmd_per_lun %d, "
> > "sg_tablesize %d, this_id %d, "
> > "flags { %s%s%s%s}, "
> > -#if defined(USLEEP_POLL) && defined(USLEEP_WAITLONG)
> > - "USLEEP_POLL %lu, USLEEP_WAITLONG %lu, "
> > +#if defined(USLEEP_POLL) && defined(USLEEP_SLEEP)
> > + "USLEEP_POLL %lu, USLEEP_SLEEP %lu, "
> > #endif
> > "options { %s} ",
> > instance->hostt->name, instance->io_port, instance->n_io_port,
> > @@ -631,8 +627,8 @@ static void prepare_info(struct Scsi_Hos
> > hostdata->flags & FLAG_DTC3181E ? "DTC3181E " : "",
> > hostdata->flags & FLAG_NO_PSEUDO_DMA ? "NO_PSEUDO_DMA " : "",
> > hostdata->flags & FLAG_TOSHIBA_DELAY ? "TOSHIBA_DELAY " : "",
> > -#if defined(USLEEP_POLL) && defined(USLEEP_WAITLONG)
> > - USLEEP_POLL, USLEEP_WAITLONG,
> > +#if defined(USLEEP_POLL) && defined(USLEEP_SLEEP)
> > + USLEEP_POLL, USLEEP_SLEEP,
> > #endif
> > #ifdef AUTOPROBE_IRQ
> > "AUTOPROBE_IRQ "
> Wouldn't it make more sense to remove the USLEEP_WAITLONG completely?
> From what I can see it is meant to indicate that WAITLONG is enabled,
> but we've just removed that functionality...

Actually, this patch does remove USLEEP_WAITLONG completely. It does not
remove USLEEP_POLL and USLEEP_SLEEP. In patch 25, the USLEEP_POLL and
USLEEP_SLEEP stuff is replaced by an algorithm that sleeps while polling.

You are right that adding USLEEP_SLEEP to this snprintf() is a change that
doesn't really belong here. But since I was changing those lines anyway,
it seemed like a good time to fix a mistake I made when I first added the
snprintf() and wrote "USLEEP_POLL, USLEEP_WAITLONG" instead of
"USLEEP_POLL, USLEEP_SLEEP".

Shall I revise this patch? That will affect patch 25. Or perhaps I should
add the snprintf() change in the commit log?

Thanks for your review.

--

>
> Cheers,
>
> Hannes
>
--
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/