[PATCH] fix pci_setup_phb_io_dynamic for pci_iomap

[PATCH] fix pci_setup_phb_io_dynamic for pci_iomap

[Arnd Bergmann]

We had a problem on a system with only dynamically allocated
PCI buses (using of_pci_phb_driver) in combination with libata.

It turns out that pci_setup_phb_io_dynamic does a simple ioremap
instead of an __ioremap_explicit into the reserved I/O space
region, which causes pcim_iomap->pci_iomap->ioport_map to fail
a sanity check for the virtual address of the I/O port.

Also, our setup ended up having no "primary" phb, which means
that pci_io_base never got initialized. This caused the same
bug independently.

This patch fixes both problems.

Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>

---

 arch/powerpc/kernel/of_platform.c |    4 ++--
 arch/powerpc/kernel/pci_64.c      |    6 ++++--
 2 files changed, 6 insertions(+), 4 deletions(-)

Index: linux-2.6.21.ppc64/arch/powerpc/kernel/of_platform.c
===================================================================
--- linux-2.6.21.ppc64.orig/arch/powerpc/kernel/of_platform.c
+++ linux-2.6.21.ppc64/arch/powerpc/kernel/of_platform.c
@@ -426,13 +426,13 @@ static int __devinit of_pci_phb_probe(st
 	}
 
 	/* Process "ranges" property */
-	pci_process_bridge_OF_ranges(phb, dev->node, 0);
+	pci_process_bridge_OF_ranges(phb, dev->node, pci_io_base == 0);
 
 	/* Setup IO space.
 	 * This will not work properly for ISA IOs, something needs to be done
 	 * about it if we ever generalize that way of probing PCI brigdes
 	 */
-	pci_setup_phb_io_dynamic(phb, 0);
+	pci_setup_phb_io_dynamic(phb, pci_io_base == 0);
 
 	/* Init pci_dn data structures */
 	pci_devs_phb_init_dynamic(phb);
Index: linux-2.6.21.ppc64/arch/powerpc/kernel/pci_64.c
===================================================================
--- linux-2.6.21.ppc64.orig/arch/powerpc/kernel/pci_64.c
+++ linux-2.6.21.ppc64/arch/powerpc/kernel/pci_64.c
@@ -1077,8 +1077,7 @@ void __devinit pci_setup_phb_io_dynamic(
 	unsigned long io_virt_offset;
 	struct resource *res;
 
-	hose->io_base_virt = __ioremap(hose->io_base_phys, size,
-					_PAGE_NO_CACHE | _PAGE_GUARDED);
+	hose->io_base_virt = reserve_phb_iospace(size);
 	DBG("phb%d io_base_phys 0x%lx io_base_virt 0x%lx\n",
 		hose->global_number, hose->io_base_phys,
 		(unsigned long) hose->io_base_virt);
@@ -1090,6 +1089,9 @@ void __devinit pci_setup_phb_io_dynamic(
 	res = &hose->io_resource;
 	res->start += io_virt_offset;
 	res->end += io_virt_offset;
+
+	__ioremap_explicit(hose->io_base_phys, (unsigned long)hose->io_base_virt, size,
+			_PAGE_NO_CACHE | _PAGE_GUARDED);
 }
 
 

Re: [PATCH] fix pci_setup_phb_io_dynamic for pci_iomap

[Benjamin Herrenschmidt]

On Thu, 2007-05-03 at 21:42 +0200, Arnd Bergmann wrote:
> We had a problem on a system with only dynamically allocated
> PCI buses (using of_pci_phb_driver) in combination with libata.
> 
> It turns out that pci_setup_phb_io_dynamic does a simple ioremap
> instead of an __ioremap_explicit into the reserved I/O space
> region, which causes pcim_iomap->pci_iomap->ioport_map to fail
> a sanity check for the virtual address of the I/O port.
> 
> Also, our setup ended up having no "primary" phb, which means
> that pci_io_base never got initialized. This caused the same
> bug independently.
> 
> This patch fixes both problems.

Hrm... dynamically added PHBs shouldn't use the reserved region.

We need to fix ioport_map instead I think. I'll have a look.

Ben.