Re: [PATCH] pci/pcie: Avoid unnecessary PCIe link retrains

[Kenji Kaneshige <kaneshige.kenji@jp.fujitsu.com>]

Bjorn Helgaas wrote:
> On Thursday 05 November 2009 12:07:07 pm Matthew Wilcox wrote:
>> On Thu, Nov 05, 2009 at 12:59:25PM -0600, Bjorn Helgaas wrote:
>>> Here's another possibility, the idea being to collect all the PCIe
>>> stuff in one place.  This would require a lot of changes in the PCIe
>>> driver code, but most of them would be trivial.
>> I don't like the idea of kmallocing a 6- or 10-byte data structure
>> ... better to keep it in the pci_dev.  Maybe embedding a pcie_dev inside
>> the pci_dev would be a good idea, but unless there're more things to
>> move to it, this seems like a net loss to me.
> 
> That's true, it's not worth it for such a small structure.  I figured
> there would probably be more PCIe-related stuff that could go there,
> e.g., embedding the link_state directly.  But I haven't looked to
> see whether there actually is enough PCIe stuff to make it worthwhile.
> 

In my understanding, there are the following PCIe specific data in
struct pci_dev. It is not many.

	u8 pcie_cap;			*I added this time
	u8 pcie_type;
	struct pcie_link_state *link_state;
	unsigned int ari_enabled:1;
	unsigned int is_pcie:1;		*No longer needed and will be removed.
	unsigned int aer_firmware_first:1;

How about keeping cap offset in pci_dev so far and introducing the
following wrapper function?

static inline pcie_cap(struct pci_dev *pdev)
{
        return pdev->pcie_cap;
}

When we actually need a new data structure for PCIe-related stuff in the
future, all we need to do would be changing this wrapper like below.

static inline pcie_cap(struct pci_dev *pdev)
{
	struct pcie_dev *pcie = pdev->pcie;
	if (pcie)
		return pcie->cap;
	return 0;
}

Thanks,
Kenji Kaneshige