Re: [PATCH] PCI: endpoint: handle probable NULL pointer access

[Pankaj Dubey]

On 25 October 2017 at 17:32, Kishon Vijay Abraham I <kishon@xxxxxx> wrote:
> Hi,
>
> On Wednesday 25 October 2017 01:32 AM, Bjorn Helgaas wrote:
>> On Thu, Oct 12, 2017 at 09:27:57AM +0530, Pankaj Dubey wrote:
>>> controller_group allocation in pci_ep_cfs_init function can fail
>>> so we should have a check while using it in pci_ep_cfs_add_epc_group
>>> for registering group, else we will hit NULL pointer access.
>>>
>>> This patch adds required check for the same and returns -EPROBE_DEFER,
>>> so that endpoint controller driver probe can be reattempted later
>>> in case controller_group is not allocated because pci_ep_cfs_init is
>>> not yet called.
>>>
>>> Signed-off-by: Pankaj Dubey <pankaj.dubey@xxxxxxxxxxx>
>>
>> Looking for Kishon's ack here.
>>
>>> ---
>>>  drivers/pci/endpoint/pci-ep-cfs.c   | 7 ++++++-
>>>  drivers/pci/endpoint/pci-epc-core.c | 4 ++++
>>>  2 files changed, 10 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/pci/endpoint/pci-ep-cfs.c b/drivers/pci/endpoint/pci-ep-cfs.c
>>> index 424fdd6..3cac818 100644
>>> --- a/drivers/pci/endpoint/pci-ep-cfs.c
>>> +++ b/drivers/pci/endpoint/pci-ep-cfs.c
>>> @@ -172,7 +172,12 @@ struct config_group *pci_ep_cfs_add_epc_group(const char *name)
>>>      group = &epc_group->group;
>>>
>>>      config_group_init_type_name(group, name, &pci_epc_type);
>>> -    ret = configfs_register_group(controllers_group, group);
>>> +
>>> +    if (controllers_group)
>>> +            ret = configfs_register_group(controllers_group, group);
>>> +    else
>>> +            ret = -EPROBE_DEFER;
>>> +
>
> Do you ever face this issue?

Yes, I was adding support for PCIe endpoint in Exynos driver and if we
see pci-exynos.c,
platform_driver_probe is called via subsys_initcall, which will happen
much before that module_init
and during endpoint probe sequence I got kernel panic for NULL pointer access.

> Ideally controllers_group should always be
> initialized if the dependencies are modeled properly.

Ideally Yes.

But we can't ignore error cases. Even though dependencies are modeled properly,
this check is mandatory, if we see pci_ep_cfs_init function where
"controllers_group" is suppose
to be allocated via call to "configfs_register_default_group", is
prone to failure as allocated via
kzalloc. We are handling error condition in pci_ep_cfs_init if it
fails to allocate "controllers_group"
but during EP initialization sequence, there is no check on
"controllers_group" pointer in
"configfs_register_default_group" function. So I feel it should have a
check for valid pointer.


>>>      if (ret) {
>>>              pr_err("failed to register configfs group for %s\n", name);
>>>              goto err_register_group;
>>> diff --git a/drivers/pci/endpoint/pci-epc-core.c b/drivers/pci/endpoint/pci-epc-core.c
>>> index 42c2a11..d327a2a 100644
>>> --- a/drivers/pci/endpoint/pci-epc-core.c
>>> +++ b/drivers/pci/endpoint/pci-epc-core.c
>>> @@ -518,6 +518,10 @@ __pci_epc_create(struct device *dev, const struct pci_epc_ops *ops,
>>>              goto put_dev;
>>>
>>>      epc->group = pci_ep_cfs_add_epc_group(dev_name(dev));
>>> +    if (IS_ERR(epc->group)) {
>>> +            ret = -EPROBE_DEFER;
>
> should use the return value of pci_ep_cfs_add_epc_group().
>

OK. Will modify in next version.

> However I don't think this is required since drivers might not actually need cfs.

Ok, we can avoid propagating error to the caller here, but in case if
ERR_PTR there should be
at least one warn message. What do you say?


Thanks,
Pankaj Dubey
>
> Thanks
> Kishon