Re: CFI violation in drivers/infiniband/core/sysfs.c

From: Jason Gunthorpe
Date: Wed May 05 2021 - 13:57:53 EST

On Wed, May 05, 2021 at 06:26:06PM +0200, Greg KH wrote:
> > They are in many places, for instance.
> >
> > int device_create_file(struct device *dev,
> > const struct device_attribute *attr)
> >
> > We loose the type safety when working with attribute arrays, and
> > people can just bypass the "proper" APIs to raw sysfs ones whenever
> > they like.
> >
> > It is fundamentally completely wrong to attach a 'struct
> > kobject_attribute' to a 'struct device' kobject.
> But it works because we are using C and we don't have RTTI :)
> Yes, it's horrid, but we do it because we "know" the real type that is
> being called here. That was an explicit design decision at the time.

I think it is beyond horrid. Just so everyone is clear on what is
happening here..

RDMA has this:

struct hw_stats_attribute {
struct attribute attr;
ssize_t (*show)(struct kobject *kobj,
struct attribute *attr, char *buf);

And it has two kobject types, a struct device kobject and a ib_port

When the user invokes show on the struct device sysfs we have this
call path:

struct device_attribute *dev_attr = to_dev_attr(attr);
ret = dev_attr->show(dev, dev_attr, buf);
struct hw_stats_attribute *hsa = container_of(attr, struct hw_stats_attribute, attr)

And from the ib_port kobject we have this one:

struct port_attribute *port_attr =
container_of(attr, struct port_attribute, attr);
return port_attr->show(p, port_attr, buf);
struct hw_stats_attribute *hsa = container_of(attr, struct hw_stats_attribute, attr)

Then show_hw_stats() goes on to detect which call chain it uses so it
can apply the proper container of to the kobj:

if (!hsa->port_num)
dev = container_of((struct device *)kobj,
struct ib_device, dev);
port = container_of(kobj, struct ib_port, kobj);

There are several nasty defeats of the C typing system here:

- A hw_stats_attribute is casted to device_attribute hidden inside

- The 'show' function pointer is being casted from from a
(*show)(kobject,attr,buf) to (*show)(device,device_attr,buf)
This cast is hidden by the above wrong use of container_of()

- The dev_attr 'struct device_attribute *' is casted directly to a
'struct attribute *' and this cast is hidden because of the wrongly type
function pointer

- The dev 'struct device *' is casted directly to a 'struct kobject *'
and like above this is hidden inside the wrongly typed function

- All of the above is true again when talking about port_attribute
and struct ib_port.

This all implicitly relies on the following unchecked and undocumated
- struct device's kobject is the first member in the struct
- struct ib_port's kobject is the first member in the struct
- The attr, show and store members are at the same offset
in struct device_attribute and struct hw_stats_attribute

None of this is even slightly clear from the code. If Nathan hadn't
pointed it out I don't think anyone would have known..

> If that was a good decision or not, I don't know, but it's served us
> well for the past 20 years or so...

I agree with Kees, "my mind rebelled". I don't think it aligned with
the modern kernel style. If tooling starts to shine light on these
bast casts I feel it would only improve code quality.

For instance the patch Kees pointed at e6d701dca989 ("ACPI: sysfs: Fix
pm_profile_attr type")

This is a legitimate typing bug. ACPI should not have been using
struct device_attribute with a kobject creted by

acpi_kobj = kobject_create_and_add("acpi", firmware_kobj);

Certainly this RDMA code has no buisness being written like this
either, it nets out to saving about 50 lines of straightforward
duplicated code for a lot of worse junk.