Re: [GIT PULL] Block fixes for 6.3-rc3

From: Linus Torvalds <torvalds@linux-foundation.org>
To: Miguel Ojeda <miguel.ojeda.sandonis@gmail.com>
Cc: Kees Cook <keescook@chromium.org>,
	Nick Desaulniers <ndesaulniers@google.com>,
	Jens Axboe <axboe@kernel.dk>,
	Nathan Chancellor <nathan@kernel.org>,
	"linux-block@vger.kernel.org" <linux-block@vger.kernel.org>,
	clang-built-linux <llvm@lists.linux.dev>,
	linux-hardening@vger.kernel.org
Subject: Re: [GIT PULL] Block fixes for 6.3-rc3
Date: Fri, 17 Mar 2023 13:51:17 -0700	[thread overview]
Message-ID: <CAHk-=wgTSdKYbmB1JYM5vmHMcD9J9UZr0mn7BOYM_LudrP+Xvw@mail.gmail.com> (raw)
In-Reply-To: <CANiq72m46OzQPtZbS_VaQGgGknFV-hKvhBw8sVZx9ef=AzupTQ@mail.gmail.com>

On Fri, Mar 17, 2023 at 1:42 PM Miguel Ojeda
<miguel.ojeda.sandonis@gmail.com> wrote:
>
> It is comparing against just the `.code` in the `m5mols_default_ffmt`
> table, i.e. the `MEDIA_BUS_FMT_VYUY8_2X8` (8199 = 0x2007) and
> `MEDIA_BUS_FMT_JPEG_1X8` (16385 = 0x4001), see

Yeah, I see what it's doing.

But:

> If the condition had `++type` instead, it would not be a problem,
> because the loop stops before we go into the out of bounds access thus
> no UB.

Yeah, but clang really should have generated a proper third iteration,
which calls that "out of bounds" case, and then returns, instead fo
falling off the end.

I do think that on the kernel side, the fix is to just change

        } while (type++ != SIZE_DEFAULT_FFMT);

to

        } while (++type != SIZE_DEFAULT_FFMT);

but I would *really* like clang to be fixed to not silently generate
code that does insane things and would be basically impossible to
debug if it ever triggers.

We would have spent a *lot* of time wondering how the heck we Oopsed
in m5mols_get_frame_desc().

             Linus