[v3,3/6] PCI: generic: Quit clobbering our pci_ops.

From: David Daney <david.daney@cavium.com>

The pci-host-generic driver keeps a global struct pci_ops which it
then patches with the .map_bus method appropriate for the bus device.
A problem arises when the driver is used for two different types of
bus devices, the .map_bus method for the last device probed clobbers
the method for all previous devices.  The result, only the last bus
device probed has the proper .map_bus, and the others fail.

Move the struct pci_ops into the bus specific structure, and
initialize it when the bus device is probed.  Keep a copy of the
gen_pci_cfg_bus_ops structure, instead of a pointer to a global copy,
to future proof against the addition of bus specific elements to
struct pci_ops.

Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: David Daney <david.daney@cavium.com>
---
No change from v1 or v2.

 drivers/pci/host/pci-host-generic.c | 31 +++++++++++++++++--------------
 1 file changed, 17 insertions(+), 14 deletions(-)

On Tuesday 22 September 2015 16:49:14 David Daney wrote:
> From: David Daney <david.daney@cavium.com>
> 
> The pci-host-generic driver keeps a global struct pci_ops which it
> then patches with the .map_bus method appropriate for the bus device.
> A problem arises when the driver is used for two different types of
> bus devices, the .map_bus method for the last device probed clobbers
> the method for all previous devices.  The result, only the last bus
> device probed has the proper .map_bus, and the others fail.
> 
> Move the struct pci_ops into the bus specific structure, and
> initialize it when the bus device is probed.  Keep a copy of the
> gen_pci_cfg_bus_ops structure, instead of a pointer to a global copy,

This is a very useful change.

> to future proof against the addition of bus specific elements to
> struct pci_ops.

but I don't like this part. We should just not have bus specific
elements in pci_ops. We don't really have that here either, except
that the gen_pci driver had a hack for reusing the same operations
for things that are actually different.

It's an established practice that anything named '*_operations' is
meant to be constant and ideally defined as 'static const ... *_ops;'
in the driver. We could try to enforce this better by marking
bus->ops as 'const' and changing all the instances of this structure
accordingly.

> @@ -65,7 +65,11 @@ static void __iomem *gen_pci_map_cfg_bus_cam(struct pci_bus *bus,
>  
>  static struct gen_pci_cfg_bus_ops gen_pci_cfg_cam_bus_ops = {
>  	.bus_shift	= 16,
> -	.map_bus	= gen_pci_map_cfg_bus_cam,
> +	.ops		= {
> +		.map_bus	= gen_pci_map_cfg_bus_cam,
> +		.read		= pci_generic_config_read,
> +		.write		= pci_generic_config_write,
> +	}
>  };

So this is good. We could in theory unify the map_bus functions
like this now:

static void __iomem *gen_pci_map_cfg_bus(struct pci_bus *bus,
                                         unsigned int devfn,
                                         int where)
{
        struct gen_pci *pci = bus->sysdata;
	struct gen_pci_cfg_bus_ops *ops;
        resource_size_t idx;

	ops = container_of(bus->ops, struct gen_pci_cfg_bus_ops, ops);
	idx = bus->number - pci->cfg.bus_range->start;

        return pci->cfg.win[idx] + ((devfn << ops->dev_shift) | where);
}

Not sure if that improves clarity or not, up to Will.

> @@ -234,8 +237,7 @@ static int gen_pci_probe(struct platform_device *pdev)
>  	}
>  
>  	of_id = of_match_node(gen_pci_of_match, np);
> -	pci->cfg.ops = of_id->data;
> +	pci->cfg.ops = *(struct gen_pci_cfg_bus_ops *)of_id->data;

This is the part that grabbed my attention, we should not do it like this.

	Arnd
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On 09/23/2015 01:21 AM, Arnd Bergmann wrote:
> On Tuesday 22 September 2015 16:49:14 David Daney wrote:
>> From: David Daney <david.daney@cavium.com>
>>
>> The pci-host-generic driver keeps a global struct pci_ops which it
>> then patches with the .map_bus method appropriate for the bus device.
>> A problem arises when the driver is used for two different types of
>> bus devices, the .map_bus method for the last device probed clobbers
>> the method for all previous devices.  The result, only the last bus
>> device probed has the proper .map_bus, and the others fail.
>>
>> Move the struct pci_ops into the bus specific structure, and
>> initialize it when the bus device is probed.  Keep a copy of the
>> gen_pci_cfg_bus_ops structure, instead of a pointer to a global copy,
>
> This is a very useful change.
>
>> to future proof against the addition of bus specific elements to
>> struct pci_ops.
>
> but I don't like this part. We should just not have bus specific
> elements in pci_ops. We don't really have that here either, except
> that the gen_pci driver had a hack for reusing the same operations
> for things that are actually different.
>
> It's an established practice that anything named '*_operations' is
> meant to be constant and ideally defined as 'static const ... *_ops;'
> in the driver. We could try to enforce this better by marking
> bus->ops as 'const' and changing all the instances of this structure
> accordingly.
>
>> @@ -65,7 +65,11 @@ static void __iomem *gen_pci_map_cfg_bus_cam(struct pci_bus *bus,
>>
>>   static struct gen_pci_cfg_bus_ops gen_pci_cfg_cam_bus_ops = {
>>   	.bus_shift	= 16,
>> -	.map_bus	= gen_pci_map_cfg_bus_cam,
>> +	.ops		= {
>> +		.map_bus	= gen_pci_map_cfg_bus_cam,
>> +		.read		= pci_generic_config_read,
>> +		.write		= pci_generic_config_write,
>> +	}
>>   };
>
> So this is good. We could in theory unify the map_bus functions
> like this now:
>
> static void __iomem *gen_pci_map_cfg_bus(struct pci_bus *bus,
>                                           unsigned int devfn,
>                                           int where)
> {
>          struct gen_pci *pci = bus->sysdata;
> 	struct gen_pci_cfg_bus_ops *ops;
>          resource_size_t idx;
>
> 	ops = container_of(bus->ops, struct gen_pci_cfg_bus_ops, ops);
> 	idx = bus->number - pci->cfg.bus_range->start;
>
>          return pci->cfg.win[idx] + ((devfn << ops->dev_shift) | where);
> }
>
> Not sure if that improves clarity or not, up to Will.
>
>> @@ -234,8 +237,7 @@ static int gen_pci_probe(struct platform_device *pdev)
>>   	}
>>
>>   	of_id = of_match_node(gen_pci_of_match, np);
>> -	pci->cfg.ops = of_id->data;
>> +	pci->cfg.ops = *(struct gen_pci_cfg_bus_ops *)of_id->data;
>
> This is the part that grabbed my attention, we should not do it like this.
>

I will consider changing this so that a structure copy is not used, 
perhaps as you suggest above.

David Daney.


> 	Arnd
>

--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html