Re: [PATCH v4 34/78] atari_NCR5380: Use arbitration timeout

[Finn Thain]

On Sun, 24 Jan 2016, Geert Uytterhoeven wrote:

> Hi Finn,
> 
> On Sun, Jan 3, 2016 at 6:05 AM, Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx> wrote:
> > Allow target selection to fail with a timeout instead of waiting in
> > infinite loops. This gets rid of the unused NCR_TIMEOUT macro, it is more
> > defensive and has proved helpful in debugging.
> >
> > Signed-off-by: Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx>
> > Reviewed-by: Hannes Reinecke <hare@xxxxxxxx>
> > Tested-by: Ondrej Zary <linux@xxxxxxxxxxxxxxxxxxxx>
> > Tested-by: Michael Schmitz <schmitzmic@xxxxxxxxx>
> 
> This patch (commit 55500d9b08295e3b6016b53879dea1cb7787f1b0) causes a hang
> on ARAnyM with atari_defconfig after:
> 
>     scsi host0: Atari native SCSI, io_port 0x0, n_io_port 0, base 0x0,
> irq 15, can_queue 8, cmd_per_lun 1, sg_tablesize 0, this_id 7, flags {
> }, options { REAL_DMA SUPPORT_TAGS }
>     blk_queue_max_segments: set to minimum 1
> 
> > --- linux.orig/drivers/scsi/atari_NCR5380.c     2016-01-03 16:03:43.000000000 +1100
> > +++ linux/drivers/scsi/atari_NCR5380.c  2016-01-03 16:03:44.000000000 +1100
> > @@ -1436,42 +1437,28 @@ static int NCR5380_select(struct Scsi_Ho
> >         NCR5380_write(OUTPUT_DATA_REG, hostdata->id_mask);
> >         NCR5380_write(MODE_REG, MR_ARBITRATE);
> >
> > -       local_irq_restore(flags);
> > +       /* The chip now waits for BUS FREE phase. Then after the 800 ns
> > +        * Bus Free Delay, arbitration will begin.
> > +        */
> >
> > -       /* Wait for arbitration logic to complete */
> > -#if defined(NCR_TIMEOUT)
> > -       {
> > -               unsigned long timeout = jiffies + 2*NCR_TIMEOUT;
> > -
> > -               while (!(NCR5380_read(INITIATOR_COMMAND_REG) & ICR_ARBITRATION_PROGRESS) &&
> > -                      time_before(jiffies, timeout) && !hostdata->connected)
> > -                       ;
> > -               if (time_after_eq(jiffies, timeout)) {
> > -                       printk("scsi : arbitration timeout at %d\n", __LINE__);
> > +       local_irq_restore(flags);
> > +       timeout = jiffies + HZ;
> > +       while (1) {
> > +               if (time_is_before_jiffies(timeout)) {
> >                         NCR5380_write(MODE_REG, MR_BASE);
> > -                       NCR5380_write(SELECT_ENABLE_REG, hostdata->id_mask);
> > +                       shost_printk(KERN_ERR, instance,
> > +                                    "select: arbitration timeout\n");
> >                         return -1;
> >                 }
> > +               if (!(NCR5380_read(MODE_REG) & MR_ARBITRATE)) {
> 
> This newly added check always triggers, causing an infinite loop calling
> NCR5380_select().

If you bisected and got a failure here, it would not be surprising because 
some of the remaining patches fix bugs in the exception handlers.

But if you test the entire patch series and get a hang (after waiting for 
the command to timeout and abort etc) it could be caused by a known bug in 
the abort handler. I will be sending a patch for that.

It can be helpful to enable scsi error recovery logging. E.g.
# scsi_logging_level -s -E 5
and/or assign the value from /proc/sys/dev/scsi/logging_level to the
scsi_logging_level kernel parameter.

> Perhaps this is an ARAnyM quirk?

I'd say this is an ARAnyM bug. atari_scsi has been tested on an actual 
Atari Falcon, hence Michael's tested-by tag.

> If not, does it trigger (on some hardware) with drivers/scsi/NCR5380.c, 
> too?

For the arbitration and selection phases, there is no difference between 
NCR5380.c and atari_NCR5380.c. That's one of the benefits of my patches.

That means this code was tested on silicon from NCR (53C400), Symbios 
Logic (53C400A), AMD (Am85C80), Domex Technology Corp (DTC-536, DTC-436) 
and LOGIC Devices (L5380).

> 
> > +                       /* Reselection interrupt */
> > +                       return -1;
> > +               }
> > +               if (NCR5380_read(INITIATOR_COMMAND_REG) & ICR_ARBITRATION_PROGRESS)
> > +                       break;
> >         }
> > -#else /* NCR_TIMEOUT */
> > -       while (!(NCR5380_read(INITIATOR_COMMAND_REG) & ICR_ARBITRATION_PROGRESS) &&
> > -              !hostdata->connected)
> > -               ;
> > -#endif
> > -
> > -       dprintk(NDEBUG_ARBITRATION, "scsi%d: arbitration complete\n", HOSTNO);
> > -
> > -       if (hostdata->connected) {
> > -               NCR5380_write(MODE_REG, MR_BASE);
> > -               return -1;
> > -       }
> > -       /*
> > -        * The arbitration delay is 2.2us, but this is a minimum and there is
> > -        * no maximum so we can safely sleep for ceil(2.2) usecs to accommodate
> > -        * the integral nature of udelay().
> > -        *
> > -        */
> >
> > +       /* The SCSI-2 arbitration delay is 2.4 us */
> >         udelay(3);
> >
> >         /* Check for lost arbitration */
> 
> On current mainline, this (whitespace-damaged) patch fixed the issue for me:
> 
> --- a/drivers/scsi/atari_NCR5380.c
> +++ b/drivers/scsi/atari_NCR5380.c
> @@ -1253,10 +1253,6 @@ static struct scsi_cmnd *NCR5380_select(struct
> Scsi_Host *instance,
>                         INITIATOR_COMMAND_REG, ICR_ARBITRATION_PROGRESS,
>                                                ICR_ARBITRATION_PROGRESS, HZ);
>         spin_lock_irq(&hostdata->lock);
> -       if (!(NCR5380_read(MODE_REG) & MR_ARBITRATE)) {
> -               /* Reselection interrupt */
> -               goto out;
> -       }
>         if (err < 0) {
>                 NCR5380_write(MODE_REG, MR_BASE);
>                 shost_printk(KERN_ERR, instance,
> @@ -1297,10 +1293,6 @@ static struct scsi_cmnd *NCR5380_select(struct
> Scsi_Host *instance,
> 
>         spin_lock_irq(&hostdata->lock);
> 
> -       /* NCR5380_reselect() clears MODE_REG after a reselection interrupt */
> -       if (!(NCR5380_read(MODE_REG) & MR_ARBITRATE))
> -               goto out;
> -
>         if (!hostdata->selecting) {
>                 NCR5380_write(MODE_REG, MR_BASE);
>                 NCR5380_write(INITIATOR_COMMAND_REG, ICR_BASE);
> 

The MR_ARBITRATE bit should remain set until the driver clears it (or the 
reset logic clears it). But it looks like aranym simply discards writes to 
the mode register, such that reads always return 0.

Compare
  http://sourceforge.net/p/aranym/code/ci/master/tree/src/ncr5380.cpp
with the MAME/MESS emulated device
  https://github.com/mamedev/mame/blob/master/src/devices/machine/ncr5380.cpp

I don't know what the Hatari emulator does.

In principle I think that Linux drivers should not carry workarounds for 
emulators.

-- 

> Gr{oetje,eeting}s,
> 
>                         Geert
>