[POWERPC] PCI Bug fix for MRM failure in PPC 440EPx

Problem Description and Fix : Memory Read Multiples(MRM) do not work
correctly on PPC 440EPX based systems. A PCI driver determines whether
MRMs are supported by reading the PCI cache line size register. If this
value is zero then MRMs are not supported. However some PCI drivers
write to the PCI cache line size register on initialization. This
results in MRMs being sent to system memory on 440EPX based systems.
Since MRMs do not work correctly in 440EPX based systems this may cause
system hang. This patch solves this problem by modifying the PPC
platform specific PCI configuration register write function, by forcing
any value written to PCI_CACHE_LINE_SIZE register to be 0. This fix was
tested on different PCI cards : i.e. Silicon Image ATA card and Intel
E1000 GIGE card. On Silicon Image ATA card without this fix in place
creating a filesystem on IDE drive "mke2fs /dev/hda" was hanging the
system. MRMs issued by the PCI card were seen on the PCI analyzer since
the Silicon Image driver was setting the PCI_CACHE_LINE_SIZE register to
255. With this patch the PCI_CACHE_LINE_SIZE register was 0 and only
Memory Reads were seen on PCI analyzer.

Signed-off-by: Pravin M. Bathija <redacted>
Signed-off-by: Stefan Roese <sr@denx.de>

---
I know this patch is a little "dirty", but perhaps somebody has a better
idea to fix this problem. Thanks.

commit bae603c964831f26d9816e85ffc923afc3275e44
tree 6fb93d22bcb56670a8e6f344afbe8a4a8438692c
parent b07d68b5ca4d55a16fab223d63d5fb36f89ff42f
author Stefan Roese [off-list ref] Thu, 30 Aug 2007 10:30:09 +0200
committer Stefan Roese [off-list ref] Thu, 30 Aug 2007 10:30:09 +0200

 arch/powerpc/sysdev/indirect_pci.c |    7 +++++++
 1 files changed, 7 insertions(+), 0 deletions(-)

diff --git a/arch/powerpc/sysdev/indirect_pci.c b/arch/powerpc/sysdev/indirect_pci.c
index 5294560..d694acc 100644
--- a/arch/powerpc/sysdev/indirect_pci.c
+++ b/arch/powerpc/sysdev/indirect_pci.c

@@ -128,6 +128,13 @@ indirect_write_config(struct pci_bus *bus, unsigned int devfn, int offset,
 	 * suitably aligned and that len is 1, 2 or 4.
 	 */
 	cfg_data = hose->cfg_data + (offset & 3);
+
+#if defined(CONFIG_440EPX) || defined(CONFIG_440GRX)
+	/* Workaround for MRM failure in 440EPx/GRx */
+	if (offset == PCI_CACHE_LINE_SIZE)
+		val = 0;
+#endif
+
 	switch (len) {
 	case 1:
 		out_8(cfg_data, val);

On Thu, 30 Aug 2007 10:33:28 +0200
Stefan Roese [off-list ref] wrote:

Problem Description and Fix : Memory Read Multiples(MRM) do not work
correctly on PPC 440EPX based systems. A PCI driver determines whether
MRMs are supported by reading the PCI cache line size register. If this
value is zero then MRMs are not supported. However some PCI drivers
write to the PCI cache line size register on initialization. This
results in MRMs being sent to system memory on 440EPX based systems.
Since MRMs do not work correctly in 440EPX based systems this may cause
system hang. This patch solves this problem by modifying the PPC
platform specific PCI configuration register write function, by forcing
any value written to PCI_CACHE_LINE_SIZE register to be 0. This fix was
tested on different PCI cards : i.e. Silicon Image ATA card and Intel
E1000 GIGE card. On Silicon Image ATA card without this fix in place
creating a filesystem on IDE drive "mke2fs /dev/hda" was hanging the
system. MRMs issued by the PCI card were seen on the PCI analyzer since
the Silicon Image driver was setting the PCI_CACHE_LINE_SIZE register to
255. With this patch the PCI_CACHE_LINE_SIZE register was 0 and only
Memory Reads were seen on PCI analyzer.

Signed-off-by: Pravin M. Bathija <redacted>
Signed-off-by: Stefan Roese <sr@denx.de>

---
I know this patch is a little "dirty", but perhaps somebody has a better
idea to fix this problem. Thanks.

For the peanut gallery, Stefan and I discussed this a bit on IRC today
and a different approach for arch/powerpc is going to be looked at
instead.  Namely, introducing a new flag for indirect_type in the
pci_controller structure to key off of instead of having ifdefs.

josh

On Sep 12, 2007, at 9:45 AM, Josh Boyer wrote:

On Thu, 30 Aug 2007 10:33:28 +0200
Stefan Roese [off-list ref] wrote:

quoted

Problem Description and Fix : Memory Read Multiples(MRM) do not work
correctly on PPC 440EPX based systems. A PCI driver determines  
whether
MRMs are supported by reading the PCI cache line size register. If  
this
value is zero then MRMs are not supported. However some PCI drivers
write to the PCI cache line size register on initialization. This
results in MRMs being sent to system memory on 440EPX based systems.
Since MRMs do not work correctly in 440EPX based systems this may  
cause
system hang. This patch solves this problem by modifying the PPC
platform specific PCI configuration register write function, by  
forcing
any value written to PCI_CACHE_LINE_SIZE register to be 0. This  
fix was
tested on different PCI cards : i.e. Silicon Image ATA card and Intel
E1000 GIGE card. On Silicon Image ATA card without this fix in place
creating a filesystem on IDE drive "mke2fs /dev/hda" was hanging the
system. MRMs issued by the PCI card were seen on the PCI analyzer  
since
the Silicon Image driver was setting the PCI_CACHE_LINE_SIZE  
register to
255. With this patch the PCI_CACHE_LINE_SIZE register was 0 and only
Memory Reads were seen on PCI analyzer.

Signed-off-by: Pravin M. Bathija <redacted>
Signed-off-by: Stefan Roese <sr@denx.de>

---
I know this patch is a little "dirty", but perhaps somebody has a  
better
idea to fix this problem. Thanks.

For the peanut gallery, Stefan and I discussed this a bit on IRC today
and a different approach for arch/powerpc is going to be looked at
instead.  Namely, introducing a new flag for indirect_type in the
pci_controller structure to key off of instead of having ifdefs.

I assume you'll be adding to the 'quirk' flags that we already have  
in place.

- k