Re: [PATCH] Restore PCI IO space behind Freescale pseudo-bridge

Re: [PATCH] Restore PCI IO space behind Freescale pseudo-bridge

[Kumar Gala]

On Jan 27, 2009, at 12:13 PM, Andrew Klossner wrote:

> Kumar, you wrote:
>
>> Any resolution to all this.. Not exactly sure what change impacts
>> things and what part of our screwy HW is causing issues.
>
> The change in question is commit  
> b556151110ff003ce77d84597400c84824690ccf.
>
> The fundamental problem is that the "PCI Bus Command" register, the 16
> bits at offset 4 in config space for both the PCI and PCI-Express bus
> interfaces, does not populate PCI_COMMAND_IO, the least-significant
> bit.  The PCI bridge standard says that software must set this bit to
> 1 before the bridge can respond as a slave to I/O space access,
> including access to devices on the subordinate bus.  The FSL parts
> ignore this requirement and control response to I/O space with the
> (vendor-specific) outbound window attributes registers (POWARn.)
> 2.6.28 differs from 2.6.27 in that it examines this bit and sees that
> it's set to 0, then observes that the I/O space window starts at
> offset 0 and concludes that no I/O space resource should be allocated
> for the bridge.
>
> We are using the patch I emailed.  It's ugly in that it hard-codes the
> Freescale vendor ID, but it works for us.

I guess I'm confused what the actual breakage is beyond some printk  
messages.

We "fixup" the resources in fsl_pcibios_fixup_bus() so Ben's change  
shouldn't have an impact.  Does something actually stop functioning?

> Another option would have been to configure our I/O space window to
> start at an offset other than 0.  This would take me some time to
> achieve, and it feels ugly.


agreed, that's not the right solution.

- k

Re: [PATCH] Restore PCI IO space behind Freescale pseudo-bridge

[Benjamin Herrenschmidt]

> > We are using the patch I emailed.  It's ugly in that it hard-codes the
> > Freescale vendor ID, but it works for us.
> 
> I guess I'm confused what the actual breakage is beyond some printk  
> messages.
> 
> We "fixup" the resources in fsl_pcibios_fixup_bus() so Ben's change  
> shouldn't have an impact.  Does something actually stop functioning?

I think the problem comes from my patch that "detects" closed bridge
windows, which uses the PCI_COMMAND register as a clue, and it seems
that your bridges don't properly support IO enable in there..

Ben.

Re: [PATCH] Restore PCI IO space behind Freescale pseudo-bridge

[Kumar Gala]

On Jan 28, 2009, at 3:31 AM, Benjamin Herrenschmidt wrote:

>
>>> We are using the patch I emailed.  It's ugly in that it hard-codes  
>>> the
>>> Freescale vendor ID, but it works for us.
>>
>> I guess I'm confused what the actual breakage is beyond some printk
>> messages.
>>
>> We "fixup" the resources in fsl_pcibios_fixup_bus() so Ben's change
>> shouldn't have an impact.  Does something actually stop functioning?
>
> I think the problem comes from my patch that "detects" closed bridge
> windows, which uses the PCI_COMMAND register as a clue, and it seems
> that your bridges don't properly support IO enable in there..

Yeah but I don't see why that's an issue.

Because after your code (pcibios_uninitialized_bridge_resource) we'll  
call ppc_md.pcibios_fixup_bus() and that will fix things up.

So I'm wondering if this is really a bug or just annoying printk's  
that are confusing.

- k

Re: [PATCH] Restore PCI IO space behind Freescale pseudo-bridge

[Benjamin Herrenschmidt]

On Wed, 2009-01-14 at 11:48 -0800, Andrew Klossner wrote:
> When I moved from 2.6.27 to 2.6.28, one of my devices didn't work
> until I devised the attached patch.  The kernel disallowed PCI I/O
> space resources behind the pseudo-bridge in the Freescale MPC8545
> because that bridge's config-space registers incorrectly report that
> it doesn't forward I/O space transactions.
> 
> Checking for a specific vendor ID is hacky, but it got my system going
> again with only a few perturbed lines.  Is there a more correct way to
> achieve this?

A few things yes.

First, can't you just set the PCI_COMMAND_IO bit from a quirk or at init
time rather than adding to code to cope with it being cleared later on ?

Also, the fact that it's detected as uninitialized shouldn't have broken
access to your device. If there is need for some IO space, for example
because a device requests some, the kernel should have re-opened that
window of the bridge.

Thus if that failed, it would be useful to track it down a bit more. You
can try enabling DEBUG in pci-common.c and pci-64.c, boot with "debug"
on the kernel command line and and send the resulting dmesg log

Ben.

> diff --git a/arch/powerpc/kernel/pci-common.c b/arch/powerpc/kernel/pci-common.c
> index f36936d..f9f0048 100644
> --- a/arch/powerpc/kernel/pci-common.c
> +++ b/arch/powerpc/kernel/pci-common.c
> @@ -985,6 +985,7 @@ static int __devinit pcibios_uninitialized_bridge_resource(struct pci_bus *bus,
>  	struct pci_dev *dev = bus->self;
>  	resource_size_t offset;
>  	u16 command;
> +	u16 vendor;
>  	int i;
>  
>  	/* We don't do anything if PCI_PROBE_ONLY is set */
> @@ -1030,7 +1031,16 @@ static int __devinit pcibios_uninitialized_bridge_resource(struct pci_bus *bus,
>  		 * starting at low addresses -is- valid. What we do instead if that
>  		 * we consider as unassigned anything that doesn't have IO enabled
>  		 * in the PCI command register, and that's it.
> +		 * However, we don't do that if the bridge is internal to a Freescale
> +		 * CPU, as such bridges break the rules by, for example, not populating
> +		 * the PCI_COMMAND_IO bit.
>  		 */
> +		pci_read_config_word(dev, PCI_VENDOR_ID, &vendor);
> +		if (vendor == PCI_VENDOR_ID_MOTOROLA ||
> +		    vendor == PCI_VENDOR_ID_FREESCALE) {
> +			return 0;
> +		}
> +
>  		pci_read_config_word(dev, PCI_COMMAND, &command);
>  		if (command & PCI_COMMAND_IO)
>  			return 0;
> _______________________________________________
> Linuxppc-dev mailing list
> Linuxppc-dev@ozlabs.org
> https://ozlabs.org/mailman/listinfo/linuxppc-dev

[PATCH] Restore PCI IO space behind Freescale pseudo-bridge

[Andrew Klossner]

When I moved from 2.6.27 to 2.6.28, one of my devices didn't work
until I devised the attached patch.  The kernel disallowed PCI I/O
space resources behind the pseudo-bridge in the Freescale MPC8545
because that bridge's config-space registers incorrectly report that
it doesn't forward I/O space transactions.

Checking for a specific vendor ID is hacky, but it got my system going
again with only a few perturbed lines.  Is there a more correct way to
achieve this?

diff --git a/arch/powerpc/kernel/pci-common.c b/arch/powerpc/kernel/pci-common.c
index f36936d..f9f0048 100644
--- a/arch/powerpc/kernel/pci-common.c
+++ b/arch/powerpc/kernel/pci-common.c
@@ -985,6 +985,7 @@ static int __devinit pcibios_uninitialized_bridge_resource(struct pci_bus *bus,
 	struct pci_dev *dev = bus->self;
 	resource_size_t offset;
 	u16 command;
+	u16 vendor;
 	int i;
 
 	/* We don't do anything if PCI_PROBE_ONLY is set */
@@ -1030,7 +1031,16 @@ static int __devinit pcibios_uninitialized_bridge_resource(struct pci_bus *bus,
 		 * starting at low addresses -is- valid. What we do instead if that
 		 * we consider as unassigned anything that doesn't have IO enabled
 		 * in the PCI command register, and that's it.
+		 * However, we don't do that if the bridge is internal to a Freescale
+		 * CPU, as such bridges break the rules by, for example, not populating
+		 * the PCI_COMMAND_IO bit.
 		 */
+		pci_read_config_word(dev, PCI_VENDOR_ID, &vendor);
+		if (vendor == PCI_VENDOR_ID_MOTOROLA ||
+		    vendor == PCI_VENDOR_ID_FREESCALE) {
+			return 0;
+		}
+
 		pci_read_config_word(dev, PCI_COMMAND, &command);
 		if (command & PCI_COMMAND_IO)
 			return 0;