On 12/13/2012 09:31 PM, Bjorn Helgaas wrote:

[+cc Betty]

On Thu, Dec 13, 2012 at 4:04 PM, Lucas Kannebley Tavares lucaskt@linux.vnet.ibm.com wrote:

...

On architectures such as ppc64, there is no root bus device (it belongs to the hypervisor). DRM attempted to get one, causing a null-pointer dereference.

In addition to ppc64, at least ia64 and parisc have the same situation of the PCI host bridge not appearing as a PCI device itself.

...

Signed-off-by: Lucas Kannebley Tavareslucaskt@linux.vnet.ibm.com

-- diff --git a/arch/powerpc/platforms/pseries/Makefile b/arch/powerpc/platforms/pseries/Makefile index 890622b..ddfdda8 100644 --- a/arch/powerpc/platforms/pseries/Makefile +++ b/arch/powerpc/platforms/pseries/Makefile @@ -1,6 +1,8 @@ ccflags-$(CONFIG_PPC64) := -mno-minimal-toc ccflags-$(CONFIG_PPC_PSERIES_DEBUG) += -DDEBUG

+drm-y += drm_pci.o

• obj-y := lpar.o hvCall.o nvram.o reconfig.o \ setup.o iommu.o event_sources.o ras.o \ firmware.o power.o dlpar.o mobility.o

diff --git a/arch/powerpc/platforms/pseries/drm_pci.c b/arch/powerpc/platforms/pseries/drm_pci.c new file mode 100644 index 0000000..da6675e --- /dev/null +++ b/arch/powerpc/platforms/pseries/drm_pci.c @@ -0,0 +1,24 @@ +/*

• • Copyright (C) 2012 Lucas Kannebley Tavares, IBM Corporation
• • pSeries specific routines for DRM.
• • This program is free software; you can redistribute it and/or modify
• • it under the terms of the GNU General Public License as published by
• • the Free Software Foundation; either version 2 of the License, or
• • (at your option) any later version.
• • This program is distributed in the hope that it will be useful,
• • but WITHOUT ANY WARRANTY; without even the implied warranty of
• • MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
• • GNU General Public License for more details.
• • You should have received a copy of the GNU General Public License
• • along with this program; if not, write to the Free Software
• • Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
• */

+inline struct pci_device *drm_get_parent_device(struct drm_device *dev) {

•   return (dev->pdev->bus->self == NULL) ? dev->pdev :
  

dev->pdev->bus->self;

So for DRM devices on a root bus, the parent is the DRM device itself, while for DRM devices deeper in the hierarchy, the parent is the upstream P2P bridge? That doesn't really make sense to me. If the caller operates on the DRM device in some cases and on the bridge in other cases, it's going to need to know the difference, so hiding the difference in this wrapper seems counterproductive.

...

+}

diff --git a/drivers/gpu/drm/drm_pci.c b/drivers/gpu/drm/drm_pci.c index eb37466..5a8a4f5 100644 --- a/drivers/gpu/drm/drm_pci.c +++ b/drivers/gpu/drm/drm_pci.c @@ -466,6 +466,10 @@ void drm_pci_exit(struct drm_driver *driver, struct pci_driver *pdriver) } EXPORT_SYMBOL(drm_pci_exit);

+inline __weak struct pci_device *drm_get_parent_device(struct drm_device *dev) {

•   return dev->pdev->bus->self;
  

+}

• int drm_pcie_get_speed_cap_mask(struct drm_device *dev, u32 *mask) { struct pci_dev *root;

@@ -479,7 +483,7 @@ int drm_pcie_get_speed_cap_mask(struct drm_device *dev, u32 *mask) return -EINVAL;

     // find PCI device for capabilities

•   root = dev->pdev->bus->self;
  

•   root = drm_get_parent_device(dev);
  
     // some architectures might not have host bridges as PCI devices
     if (root == NULL)
  

What tree does this apply to? Upstream doesn't have the "if (root == NULL)" check yet. That check looks like the sort of thing you'd need to avoid the null pointer dereference. So maybe adding that check and the associated code is enough to fix the problem, even without adding drm_get_parent_device().

With the code in the tree, it looks like you'd dereference a null pointer in pci_pcie_cap(root), so I assume that's what you tripped over.

I'm not really sure that code outside the PCI core should be looking at capabilities of upstream devices like this. It seems like the sort of thing where the core might need to provide better interfaces.

Bjorn

Ok Bjorn, thanks for the comments, indeed I had a dirty tree here and didn't realize it, sorry. Either way I'm then sending the "if (root == NULL)" patch as a reply to this. I'm sending it along with another independent patch (they are NOT a series) that changes pci_read_config_dword calls to pci_capability_read_dword ones on the drm driver. There were only a couple of those to start with.

On Sat, Dec 22, 2012 at 12:01 PM, Lucas Kannebley Tavares lucaskt@linux.vnet.ibm.com wrote:

During the process of obtaining the speed cap for the device, it attempts go get the PCI Host bus. However on architectures such as PPC or IA64, those do not appear as devices.

Signed-off-by: Lucas Kannebley Tavares lucaskt@linux.vnet.ibm.com

drivers/gpu/drm/drm_pci.c | 5 +++++ 1 files changed, 5 insertions(+), 0 deletions(-)

diff --git a/drivers/gpu/drm/drm_pci.c b/drivers/gpu/drm/drm_pci.c index 754bc96..ea41234 100644 --- a/drivers/gpu/drm/drm_pci.c +++ b/drivers/gpu/drm/drm_pci.c @@ -479,8 +479,13 @@ int drm_pcie_get_speed_cap_mask(struct drm_device *dev, u32 *mask) if (!pci_is_pcie(dev->pdev)) return -EINVAL;

•   // find PCI device for capabilities
    root = dev->pdev->bus->self;
  
  

•   // some architectures might not have host bridges as PCI devices
  

•   if (root == NULL)
  

•           root = dev->pdev;
  

You didn't address my question about this. Obviously this will avoid a null pointer dereference. But you have to also explain why this change is correct.

If it's good enough to just look at the capabilities of the DRM device (not the upstream bridge) on PPC and ia64, why not do that everywhere and forget about the bridge completely?

    pos = pci_pcie_cap(root);
    if (!pos)
            return -EINVAL;

-- Lucas Kannebley Tavares Software Engineer IBM Linux Technology Center

Replacing these calls avoids compatibility problems with PCIe v1/v2 Capability structures.

Signed-off-by: Lucas Kannebley Tavares lucaskt@linux.vnet.ibm.com --- drivers/gpu/drm/drm_pci.c | 8 ++------ 1 files changed, 2 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/drm_pci.c b/drivers/gpu/drm/drm_pci.c index ea41234..b824d4c 100644 --- a/drivers/gpu/drm/drm_pci.c +++ b/drivers/gpu/drm/drm_pci.c @@ -486,17 +486,13 @@ int drm_pcie_get_speed_cap_mask(struct drm_device *dev, u32 *mask) if (root == NULL) root = dev->pdev;

- pos = pci_pcie_cap(root); - if (!pos) - return -EINVAL; - /* we've been informed via and serverworks don't make the cut */ if (root->vendor == PCI_VENDOR_ID_VIA || root->vendor == PCI_VENDOR_ID_SERVERWORKS) return -EINVAL;

- pci_read_config_dword(root, pos + PCI_EXP_LNKCAP, &lnkcap); - pci_read_config_dword(root, pos + PCI_EXP_LNKCAP2, &lnkcap2); + pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap); + pcie_capability_read_dword(root, PCI_EXP_LNKCAP, &lnkcap2);

lnkcap &= PCI_EXP_LNKCAP_SLS; lnkcap2 &= 0xfe;