Re: [PATCH v2 2/2] VMCI: Fix memcpy() run-time warning in dg_dispatch_as_host()

[Kees Cook <keescook@chromium.org>]

On Thu, Jan 11, 2024 at 10:15:40AM +0300, Dan Carpenter wrote:
> On Wed, Jan 10, 2024 at 04:03:28PM -0800, Kees Cook wrote:
> > 
> > Oops, yes, thanks for fixing my confusion. Right, this is a direct write
> > across members into the flex array, not a composite destination. Yay
> > all the corner cases. :P
> 
> Is there a document somewhere which explains what will trigger a runtime
> warning?  For example, if you write across members but it's not into a
> flex array does that cause an issue?  Or if you read across members?

There isn't a good place to find this. There are some code comments near
the memcpy macros, but that's not really sufficient.

At present FORTIFY is only being pedantic about _writes_, as that's
generally a higher priority problem. The implemented restriction is that
the destination buffer must be a single addressable struct member. That
destination can be a composite member (i.e. an entire substruct), but
going beyond a single member in a single memcpy() is no longer allowed
because the compiler cannot reason about whether such a copy is
"intentional".

> For example, this line reads from bulletin->vlan and
> bulletin->vlan_padding.  This causes a compiler warning in Clang and
> GCC depending on the options but does it also trigger a runtime warning?
> https://elixir.bootlin.com/linux/latest/source/drivers/net/ethernet/broadcom/bnx2x/bnx2x_sriov.c#L2655

Right, the -Wstringop-overread and related compiler flags are doing the
source buffer size checking.

Note that for the compile-time warnings, GCC has the capacity to be much
more strict than the FORTIFY checks because it can perform value _range_
tracking, where as FORTIFY compile-time checks are limited to having the
copy size being a constant expression. (i.e. GCC may produce compile
time warnings for cases that FORTIFY will only warn about at runtime if
the range is violated.)

> (I wrote a patch for this a few months back but didn't send it because
> of the merge window.  I forgot about it until now that we're in a merge
> window again...  :P)

			memcpy(&ivi->vlan, &bulletin->vlan, VLAN_HLEN);

#define VLAN_HLEN 4
ivi->vlan is u32
bulletin has:
	u16 vlan;
	u8 vlan_padding[6];

yeah, ew. Should it even be reading padding? i.e. should this be:

		ivi->vlan = bulletin->vlan << 16;

?

Or should bulletin be:

	union {
		struct {
			u16 vlan;
			u8 vlan_padding[6];
		};
		struct {
			u32 vlan_header;
			u8 vlan_header_padding[4];
		};
	};

with:

		ivi->vlan = bulletin->vlan_header;

?

I've been finding that almost all memcpy()s and memset()s into non-array
types are better just rewritten as a direct assignment. :P

-- 
Kees Cook