Re: Intel SATA combined mode quirk broken for SCSI_SATA=m

[Jeff Garzik]

Jesse Barnes wrote:
> Back in July, Adrian Bunk sent in a patch to make SCSI_SATA tristate.  This 
> prevents the intel_ide_combined quirk in drivers/pci/quirks.c from working if 
> SCSI_SATA=m, which is the case for Fedora kernels (my motivation for tracking 
> this down).
>
> In my configuration, not running the quirk causes the ata_piix driver (the 
> libata driver for my IDE controller) to fail to attach to the device, since 
> the legacy IDE driver has already claimed the ports.  Unfortunately, the AHCI 
> driver also tries to mess with the device, and ends up disabling its 
> interrupts before aborting its load, causing the IDE layer to complain loudly 
> that hda is losing interrupts.

Thanks a ton for figuring this out!


> So what should be done?  Ideally, libata would fully support ATAPI and then I 
> wouldn't need the legacy IDE drivers at all on this box, making the quirk 
> moot, but that won't happen for 2.6.14, so we'll need something else.  
> Unconditionally enabling the quirk will cause at least one of the ports to be 
> reserved for the SATA driver, which may never load.  And obviously not 
> running the quirk leads to the situation described above.
>
> A hack that might be suitable for 2.6.14 is to make the quirk depend on either 
> CONFIG_SCSI_SATA or CONFIG_SCSI_SATA_MODULE.  Then the quirk could be removed 
> entirely when ATAPI support for libata is merged.

Overall the quirk is a hack until ATAPI is supported -- but even after 
ATAPI is supported, we need to figure out some way to keep IDE driver 
from stealing the legacy IDE ports before libata can touch them :(

However, when ATAPI is supported, that at least means that both PATA and 
SATA can run at full speed.

Your patch is correct, and should go into 2.6.14-rc post-haste.

I continue to grumble at the Kconfig annoyance:  CONFIG_SCSI_SATA is 
fundamentally a yes/no question.  CONFIG_SCSI_SATA=m makes no sense at 
all -- and causes problems, as we see -- but is required in order to do 
Kconfig dependencies correctly AFAIK.

Maybe 'if SCSI_SATA' is needed instead?  A lame way to implement 
dependencies, but if there is no other way...

	Jeff


-
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/