Re: [RFC PATCH 1/2] pci: don't assume pref memio are 64bit -v3

[Ivan Kokshaysky <ink@jurassic.park.msu.ru>]

On Thu, Apr 23, 2009 at 08:13:04AM -0700, Yinghai Lu wrote:
> at least the one in mm/memory_hotplug.c
> 
> /* add this memory to iomem resource */
> static struct resource *register_memory_resource(u64 start, u64 size)
> {
>         struct resource *res;
>         res = kzalloc(sizeof(struct resource), GFP_KERNEL);
>         BUG_ON(!res);
> 
>         res->name = "System RAM";
>         res->start = start;
>         res->end = start + size - 1;
>         res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
>         if (request_resource(&iomem_resource, res) < 0) {
>                 printk("System RAM resource %llx - %llx cannot be added\n",
>                 (unsigned long long)res->start, (unsigned long long)res->end);
>                 kfree(res);
>                 res = NULL;
>         }
>         return res;
> }

Indeed. It's a really strong argument against mem64 resource approach.

On the other hand, it shows that getting 64-bit allocations right is
indeed a very complex issue - without well defined root bus ranges
there is a high risk of unexpectedly breaking something, like that memory
hotplug. Oh well...

So if the main purpose is to prevent 32-bit allocations in the DAC
area, some mix of your v2 and v3 patches seems to be the way to go.
That is

- keep IORESOURCE_MEM_64 bits from -v3, but drop iomem32_resource
and iomem64_resource things;

- pass "max" argument to allocate_resource() like you did in -v2,
but *only* then allocating from iomem_resource (r == &iomem_resource).
Also, instead of max = 0xffffffff use something like max = PCIBIOS_MAX_MEM_32.

include/linux/pci.h:
#ifndef PCIBIOS_MAX_MEM_32
#define PCIBIOS_MAX_MEM_32	(-1)
#endif

This should preserve the current behaviour of pci_bus_alloc_resource()
on non-x86 arches; overridden in arch/x86/include/asm/pci.h:
#define PCIBIOS_MAX_MEM_32	0xffffffff

> check bits 0-3 and check PCI_PREF_BASE_UPPER32 register being r/w ?

Yes. If these bits are zero, no further checks are needed -
bridge is 32-bit. If they aren't, do additional check for
PCI_PREF_BASE_UPPER32 being non-zero or writable, but *only*
if this prefetch resource is not already allocated (res->parent == NULL),
just for safety reasons - we don't want to disconnect the allocated
range from the bus even for a short time.

Ivan.