Suspend/Resume bug in PCI layer wrt quirks

[Arjan van de Ven <arjan@infradead.org>]

Hi,

the following is a bug in the PCI layer wrt suspend/resume and quirks:

pci_device_resume_early does the following:

static int pci_device_resume_early(struct device * dev)
{
        int error = 0;
        struct pci_dev * pci_dev = to_pci_dev(dev);
        struct pci_driver * drv = pci_dev->driver; 

        pci_fixup_device(pci_fixup_resume, pci_dev);

eg it calls the resume quirks.

However, one of these quirks is 
DECLARE_PCI_FIXUP_RESUME(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_ICH6_1,     asus_hides_smbus_lpc_ich6);

and asus_hides_smbus_lpc_ich6 does

        base = ioremap_nocache(rcba & 0xFFFFC000, 0x4000); /* use bits 31:14, 16 kB aligned */
        if (base == NULL) return;
        val=readl(base + 0x3418); /* read the Function Disable register, dword mode only */
        writel(val & 0xFFFFFFF7, base + 0x3418); /* enable the SMBus device */
        iounmap(base);

internally. ioremap()/iounmap() does a global tlb flush of course, which includes doing an IPI to other cpus.
However, during early resume, interrupts are disabled, and it's not allowed to do a IPI/global tlbflush with interrupts
off...

This is hitting quite a few people now:
http://www.kerneloops.org/searchweek.php?search=smp_call_function

do we need a DECLARE_PCI_FIXUP_RESUME_LATE() and use that for this one?