Re: IOMMU and DMA mode of pata_jmicron

["Roedel, Joerg" <Joerg.Roedel@amd.com>]

On Tue, Dec 28, 2010 at 08:00:12PM -0500, Jeff Garzik wrote:
> (CC'd linux-ide, Joerg)
> 
> On 12/28/2010 06:43 PM, Steffen Moser wrote:
> > I've encountered a problem with my AMD 890FX based system
> > and linux-2.6.35.10 (x86_64 platform).
> >
> > After activating the option "IOMMU" in the mainboard's BIOS
> > setup, the onboard P-ATA controller "JMicron Technology Corp.
> > JMB361 AHCI/IDE (rev 02)" puts itself back in the PIO mode.
> > When loading the module, the kernel reports:
> >
> >    pata_jmicron 0000:06:00.1: BMDMA: failed to set dma mask, \
> >                               falling back to PIO
> [...]
> > As soon as I deactivate IOMMU in BIOS setup, the P-ATA channel
> > runs in the DMA mode again. The S-ATA controllers don't seem
> > to be affected.
> >
> > My question is: Is this a known behavior? Are there any
> > things I have to consider when activating IOMMU (which seems
> > to be AMD-Vi) on an 890FX based system? Is it a problem of
> > the chipset and/or the controller or is it related to a
> > problem in the libata area?
> >
> > Some information about the affected system:
> >
> >  - Mainboard:    ASUS M4A89TD Pro/USB3
> >  - BIOS version: 1101 (most recent version)
> >  - Chipset:      AMD 890FX
> >  - Processor:    AMD Phenom 1090T
> >  - Memory:       4 x 4 GB (Kingston KVR1333D3E9S/4G)
> >  - Distribution: openSUSE 11.3
> >  - Kernel:       2.6.35.10 (from "kernel.org"), x86_64
> >
> 
> This condition occurs when the PCI API cannot set the PCI device's DMA 
> mask to the 32-bit value
> 	#define ATA_DMA_MASK            0xffffffffULL
> 
> That is an unusual failure for such a modern system, which certainly can 
> handle 32-bit masks like that, one would think.

Luckily I have exactly the same motherboard under my desk. It shows the
same issue. The problem is that the PCI device of the PATA controler
(4:00.1 in my case) is not described in the IVRS ACPI table. So the
IOMMU driver does not feel responsible for it and the dma_supported()
call on this device will fail. This also makes pci_set_dma_mask() fail
on this device. So it comes down to a BIOS bug. I look into it to find a
workaround.

	Joerg

-- 
AMD Operating System Research Center

Advanced Micro Devices GmbH Einsteinring 24 85609 Dornach
General Managers: Alberto Bozzo, Andrew Bowd
Registration: Dornach, Landkr. Muenchen; Registerger. Muenchen, HRB Nr. 43632