[PATCH] powerpc: fsl_pci: Fix PCI/PCI-X regression

[Aaron Sierra]

----- Original Message -----
> From: "Scott Wood" <scottwood at freescale.com>
> Sent: Thursday, August 21, 2014 5:01:46 PM
> 
> On Thu, 2014-08-21 at 16:54 -0500, Aaron Sierra wrote:
> > ----- Original Message -----
> > > From: "Scott Wood" <scottwood at freescale.com>
> > > Sent: Thursday, August 21, 2014 4:19:56 PM
> > > 
> > > On Wed, 2014-08-20 at 18:51 -0500, Aaron Sierra wrote:
> > > > @@ -520,9 +520,22 @@ int fsl_add_bridge(struct platform_device *pdev,
> > > > int
> > > > is_primary)
> > > >  			goto no_bridge;
> > > >  
> > > >  	} else {
> > > > -		/* For PCI read PROG to identify controller mode */
> > > > -		early_read_config_byte(hose, 0, 0, PCI_CLASS_PROG, &progif);
> > > > -		if ((progif & 1) == 1)
> > > > +		u16 master;
> > > > +
> > > > +		/*
> > > > +		 * If the controller is PCI-X, then Host mode refers to a
> > > > +		 * bridge that drives the PCI-X initialization pattern to
> > > > +		 * indicate bus operating mode/frequency to devices on the bus.
> > > > +		 * Some hardware (specifically PrPMC modules) are Agents, since
> > > > +		 * the mezzanine carrier is responsible for driving the
> > > > +		 * pattern, but they still may perform bus enumeration.
> > > > +		 *
> > > > +		 * Allow the bridge to be used for enumeration, if hardware
> > > > +		 * strapping (Host mode) or firmware (Agent mode) has enabled
> > > > +		 * bus mastering.
> > > > +		 */
> > > > +		early_read_config_word(hose, 0, 0, PCI_COMMAND, &master);
> > > > +		if (!(master & PCI_COMMAND_MASTER))
> > > >  			goto no_bridge;
> > > >  	}
> > > 
> > > Why wouldn't a normal PCI agent be able to bus master?
> > > 
> > > -Scott
> > > 
> > 
> > Short answer:
> > 
> > Simply because the hardware strapping for Host/Agent determines the
> > default state of the Bus Master bit in the Command register. Without
> > that bit being set, an Agent won't be able to send the PCI cycles
> > necessary to enumerate the bus.
> 
> But what if the host has already set that bit before Linux boots?

That's a very good point. I think that concern can be addressed by looking
for another telltale sign of enumeration, whether an address has been
assigned to the bridge's BAR 0 (PCSRBAR).

> > Long answer:
> > 
> > I think there was an assumption in the patch that introduced the
> > regression that Host and Agent in conventional PCI and PCI-X are more
> > equivalent to PCIe Root Complex and Endpoint than they really are.
> > 
> > I think the purpose of Minghuan's patch was to not attempt to enumerate
> > the bus with a bridge that simply cannot do it. A PCIe Endpoint cannot
> > enumerate the bus and a PCI/PCI-X device that cannot master cycles on the
> > bus will not be able to enumerate the bus either.
> 
> No, the point is also to not enumerate the bus in cases where we
> shouldn't.  We don't want to step on the host's toes.

Agreed.
 
> > The hardware strapping for Host/Agent mode determines the default state
> > of the Bus Master bit in the Command register.
> > 
> > In our more specialized, but still industry standardized, environment our
> > firmware must detect whether our board should be the bus's sole enumerator
> > and set the Bus Master bit accordingly.
> > 
> > My comment in the code still mentions Host and Agent mode, simply because
> > the code it is replacing based its decision on the mode.
> 
> I understand why you need to do this -- I just don't think this is a
> reliable way of detecting that you're in that situation.  How about a
> kernel command line setting?

I'd like to avoid requiring a kernel command-line option for this.

Instead, I propose checking for Agent mode like Minghuan's patch, but
prohibiting enumeration only if Bus Master is not enabled OR the address
portion of PCSRBAR is non-zero (indicating another device has enumerated
the bus).

Pseudo code:
    if (agent && (!master || enumerated))
        goto no_bridge;

-Aaron