Re: [PATCH v4 1/5] PCI: Add pci_bus_fixup_irqs().

[David Daney <ddaney@caviumnetworks.com>]

On 10/07/2015 12:44 PM, Bjorn Helgaas wrote:
> Hi David,
>
> On Fri, Oct 02, 2015 at 11:43:59AM -0700, David Daney wrote:
>> From: David Daney <david.daney@cavium.com>
>>
>> pci_bus_fixup_irqs() works like pci_fixup_irqs(), except it only does
>> the fixups for devices on the specified bus.
>>
>> Follow-on patch will use the new function.
>>
>> Signed-off-by: David Daney <david.daney@cavium.com>
>> ---
>> No change from v2.
>>
>>   drivers/pci/setup-irq.c | 30 ++++++++++++++++++++++++++++++
>>   include/linux/pci.h     |  4 ++++
>>   2 files changed, 34 insertions(+)
>>
>> diff --git a/drivers/pci/setup-irq.c b/drivers/pci/setup-irq.c
>> index 95c225b..189ad17 100644
>> --- a/drivers/pci/setup-irq.c
>> +++ b/drivers/pci/setup-irq.c
>> @@ -66,3 +66,33 @@ void pci_fixup_irqs(u8 (*swizzle)(struct pci_dev *, u8 *),
>>   		pdev_fixup_irq(dev, swizzle, map_irq);
>>   }
>>   EXPORT_SYMBOL_GPL(pci_fixup_irqs);
>> +
>> +struct pci_bus_fixup_cb_info {
>> +	u8 (*swizzle)(struct pci_dev *, u8 *);
>> +	int (*map_irq)(const struct pci_dev *, u8, u8);
>> +};
>> +
>> +static int pci_bus_fixup_irq_cb(struct pci_dev *dev, void *arg)
>> +{
>> +	struct pci_bus_fixup_cb_info *info = arg;
>> +
>> +	pdev_fixup_irq(dev, info->swizzle, info->map_irq);
>> +	return 0;
>> +}
>> +
>> +/*
>> + * Fixup the irqs only for devices on the given bus using supplied
>> + * swizzle and map_irq function pointers
>> + */
>> +void pci_bus_fixup_irqs(struct pci_bus *bus,
>> +			u8 (*swizzle)(struct pci_dev *, u8 *),
>> +			int (*map_irq)(const struct pci_dev *, u8, u8))
>> +{
>> +	struct pci_bus_fixup_cb_info info;
>> +
>> +	info.swizzle = swizzle;
>> +	info.map_irq = map_irq;
>> +	pci_walk_bus(bus, pci_bus_fixup_irq_cb, &info);
>
> I don't like the existing pci_fixup_irqs(), so by transitivity, I
> don't like pci_bus_fixup_irqs() either.

We are in agreement with respect to this point.


>  The problem is that in both
> cases this is a one-time pass over the tree, so we don't handle
> hot-added devices correctly.
>
> I think we need to get rid of pci_fixup_irqs() and somehow integrate
> it into the pci_device_add() path, where it would be done once for
> every device we enumerate.

I also agree with this point.

>  If we did that, I don't think you would
> need to add pci_bus_fixup_irqs(), would you?

Nope.

However, such a change is essentially untestable by me.  So, I didn't 
attempt it.   pci_fixup_irqs() is used by alpha, arm, m68k, mips, sh, 
sparc, tile, unicore32 and other things as well.  If the core 
pci_device_add() code were to suddenly start doing the fixup, there 
would be the potential to break all these things I cannot test.

The new pci_bus_fixup_irqs() is really an optimization so that if we 
have multiple buses created by pci-host-generic.c, that we only iterate 
over each device once.  I believe that pci-host-generic.c would still 
operate without these patches 1/5 and 2/5, and could test that if you 
are OK with the remaining three patches.  Or we could merge all 5 and 
live a while longer with the ugliness that is already there.

David Daney