On Wed, Nov 02, 2022 at 04:22:32PM -0700, Kees Cook wrote:
> On Wed, Nov 02, 2022 at 01:23:51PM -0700, Nathan Chancellor wrote:
> > On Wed, Nov 02, 2022 at 12:21:23PM -0700, Kees Cook wrote:
> > > On Wed, Nov 02, 2022 at 10:22:14AM -0700, Nathan Chancellor wrote:
> > > > The ->signal_u32_read(), ->count_u32_read(), and ->count_u32_write()
> > > > callbacks in 'struct counter_comp' expect the final parameter to have a
> > > > type of 'u32' or 'u32 *' but the ops functions that are being assigned
> > > > to those callbacks have an enumerated type as the final parameter. While
> > > > these are compatible from an ABI perspective, they will fail the
> > > > aforementioned CFI checks.
> > > >
> > > > Adjust the type of the final parameter in the ->signal_read(),
> > > > ->function_read(), and ->function_write() callbacks in 'struct
> > > > counter_ops' and their implementations to match the prototypes in
> > > > 'struct counter_comp' to clear up these warnings and CFI failures.
> > >
> > > I don't understand these changes. Where do 'struct counter_comp'
> > > and 'struct counter_ops' get confused? I can only find matching
> > > ops/assignments/calls, so I must be missing something. This looks like
> > > a loss of CFI granularity instead of having wrappers added if there is
> > > an enum/u32 conversion needed somewhere.
> >
> > Right, I am not the biggest fan of this change myself and it is entirely
> > possible that I am misreading the warnings from the commit message but I
> > do not see how
> >
> > comp_node.comp.signal_u32_read = counter->ops->signal_read;
> >
> > and
> >
> > comp_node.comp.count_u32_read = counter->ops->function_read;
> >
> > in counter_add_watch(),
> >
> > comp.signal_u32_read = counter->ops->signal_read;
> >
> > in counter_signal_attrs_create(), and
> >
> > comp.count_u32_read = counter->ops->function_read;
> > comp.count_u32_write = counter->ops->function_write;
> >
> > in counter_count_attrs_create() are currently safe under kCFI, since the
> > final parameter type of the prototypes in 'struct counter_ops' does not
> > match the final parameter type of the prototypes in 'struct
> > counter_comp'. I would expect the indirect calls in counter_get_data()
> > and counter_comp_u32_show() to fail currently.
> >
> > I briefly looked at making the 'struct counter_comp' callbacks match the
> > 'struct counter_ops' ones but the COUNTER_COMP macros in
> > include/linux/counter.h made it seem like these callbacks might be used
> > by implementations that might use different enumerated types as the
> > final parameter. I can look a little closer to see if we can make
> > everything match.
> >
> > I am not sure how wrappers would work here, I can take a look into how
> > feasible that is.
>
> How about this? (I only did signal_read -- similar changes are needed
> for function_read and function_write:
The reason for the u32 type is that all the Counter enum components can
make use of the same *_u32_read/*_u32_write calls; in other words, we
don't have to handle each Counter enum read/write as unique.
If you want to get rid of that design, then you'll need to create
respective code paths for each Counter enum. See the comments inline
below.
>
> diff --git a/drivers/counter/counter-chrdev.c b/drivers/counter/counter-chrdev.c
> index 80acdf62794a..cb391b2498a6 100644
> --- a/drivers/counter/counter-chrdev.c
> +++ b/drivers/counter/counter-chrdev.c
> @@ -38,6 +38,7 @@ struct counter_comp_node {
> a.device_u32_read == b.device_u32_read || \
> a.count_u32_read == b.count_u32_read || \
> a.signal_u32_read == b.signal_u32_read || \
> + a.signal_read == b.signal_read || \
> a.device_u64_read == b.device_u64_read || \
> a.count_u64_read == b.count_u64_read || \
> a.signal_u64_read == b.signal_u64_read || \
> @@ -54,6 +55,7 @@ struct counter_comp_node {
> comp.device_u32_read || \
> comp.count_u32_read || \
> comp.signal_u32_read || \
> + comp.signal_read || \
> comp.device_u64_read || \
> comp.count_u64_read || \
> comp.signal_u64_read || \
> @@ -320,7 +322,7 @@ static int counter_add_watch(struct counter_device *const counter,
> return -EINVAL;
>
> comp_node.comp.type = COUNTER_COMP_SIGNAL_LEVEL;
> - comp_node.comp.signal_u32_read = counter->ops->signal_read;
> + comp_node.comp.signal_read = counter->ops->signal_read;
> break;
> case COUNTER_COMPONENT_COUNT:
> if (watch.component.scope != COUNTER_SCOPE_COUNT)
> @@ -530,6 +532,7 @@ static int counter_get_data(struct counter_device *const counter,
> const size_t id = comp_node->component.id;
> struct counter_signal *const signal = comp_node->parent;
> struct counter_count *const count = comp_node->parent;
> + enum counter_signal_level level = 0;
> u8 value_u8 = 0;
> u32 value_u32 = 0;
> const struct counter_comp *ext;
> @@ -569,8 +572,8 @@ static int counter_get_data(struct counter_device *const counter,
> ret = comp->device_u32_read(counter, &value_u32);
> break;
> case COUNTER_SCOPE_SIGNAL:
> - ret = comp->signal_u32_read(counter, signal,
> - &value_u32);
> + ret = comp->signal_read(counter, signal, &level);
> + value_u32 = level;
This code path is for all Counter enum types currently; changing
signal_u32_read to signal_read here will work for
COUNTER_COMP_SIGNAL_LEVEL but break all the other Counter Signal enum
types. Instead, you should duplicate this code with proper adjustments
for each of the Counter enums in their own respective case blocks.
> break;
> case COUNTER_SCOPE_COUNT:
> ret = comp->count_u32_read(counter, count, &value_u32);
> diff --git a/drivers/counter/counter-sysfs.c b/drivers/counter/counter-sysfs.c
> index b9efe66f9f8d..07ce2543b70d 100644
> --- a/drivers/counter/counter-sysfs.c
> +++ b/drivers/counter/counter-sysfs.c
> @@ -170,6 +170,7 @@ static ssize_t counter_comp_u32_show(struct device *dev,
> const struct counter_attribute *const a = to_counter_attribute(attr);
> struct counter_device *const counter = counter_from_dev(dev);
> const struct counter_available *const avail = a->comp.priv;
> + enum counter_signal_level level = 0;
> int err;
> u32 data = 0;
>
> @@ -178,7 +179,8 @@ static ssize_t counter_comp_u32_show(struct device *dev,
> err = a->comp.device_u32_read(counter, &data);
> break;
> case COUNTER_SCOPE_SIGNAL:
> - err = a->comp.signal_u32_read(counter, a->parent, &data);
> + err = a->comp.signal_read(counter, a->parent, &level);
> + data = level;
Same issue as comment above: Counter Signal enums besides
COUNTER_COMP_SIGNAL_LEVEL are possible.
> break;
> case COUNTER_SCOPE_COUNT:
> if (a->comp.type == COUNTER_COMP_SYNAPSE_ACTION)
> @@ -842,7 +844,7 @@ static int counter_signal_attrs_create(struct counter_device *const counter,
>
> /* Create main Signal attribute */
> comp = counter_signal_comp;
> - comp.signal_u32_read = counter->ops->signal_read;
> + comp.signal_read = counter->ops->signal_read;
> err = counter_attr_create(dev, cattr_group, &comp, scope, signal);
> if (err < 0)
> return err;
> diff --git a/include/linux/counter.h b/include/linux/counter.h
> index c41fa602ed28..3f1516076f20 100644
> --- a/include/linux/counter.h
> +++ b/include/linux/counter.h
> @@ -169,6 +169,9 @@ struct counter_comp {
> struct counter_count *count, u32 *val);
> int (*signal_u32_read)(struct counter_device *counter,
> struct counter_signal *signal, u32 *val);
> + int (*signal_read)(struct counter_device *counter,
> + struct counter_signal *signal,
> + enum counter_signal_level *level);
> int (*device_u64_read)(struct counter_device *counter,
> u64 *val);
> int (*count_u64_read)(struct counter_device *counter,
>
> --
> Kees Cook
I'm not familiar with kCFI, but a better solution if possible would be
to hint to the compiler the intention of the code here -- that we're
intentionally type punning the last parameter in order to avoid
duplicate code for each Counter enum because they are all handled the
same way.
William Breathitt Gray
Attachment:
signature.asc
Description: PGP signature
