Re: [PATCH 0/5] kcov: PREEMPT_RT fixup + misc

From: Marco Elver <elver@google.com>
To: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: kasan-dev@googlegroups.com, linux-kernel@vger.kernel.org,
	Dmitry Vyukov <dvyukov@google.com>,
	Andrey Konovalov <andreyknvl@gmail.com>,
	Thomas Gleixner <tglx@linutronix.de>,
	Steven Rostedt <rostedt@goodmis.org>,
	Clark Williams <williams@redhat.com>,
	Andrew Morton <akpm@linux-foundation.org>
Subject: Re: [PATCH 0/5] kcov: PREEMPT_RT fixup + misc
Date: Mon, 6 Sep 2021 18:13:11 +0200	[thread overview]
Message-ID: <CANpmjNPZMVkr5BpywHTY_m+ndLTeWrMLTog=yGG=VLg_miqUvQ@mail.gmail.com> (raw)
In-Reply-To: <20210830172627.267989-1-bigeasy@linutronix.de>

On Mon, 30 Aug 2021 at 19:26, Sebastian Andrzej Siewior
<bigeasy@linutronix.de> wrote:
> The last patch in series is follow-up to address the PREEMPT_RT issue
> within in kcov reported by Clark [0].
> Patches 1-3 are smaller things that I noticed while staring at it.
> Patch 4 is small change which makes replacement in #5 simpler / more
> obvious.
> I tested this with the three examples in the documentation folder and I
> didn't notice higher latency with kcov enabled. Debug or not, I don't
> see a reason to make the lock a raw_spin_lock_t annd it would complicate
> memory allocation as mentioned in #5.

Thanks for sorting this out. Given syzkaller is exercising all of
KCOV's feature, I let syzkaller run for a few hours with PROVE_LOCKING
(and PROVE_RAW_LOCK_NESTING) on, and looks fine:

    Acked-by: Marco Elver <elver@google.com>
    Tested-by: Marco Elver <elver@google.com>

> One thing I noticed and have no idea if this is right or not:
> The code seems to mix long and uint64_t for the reported instruction
> pointer / position in the buffer. For instance
> __sanitizer_cov_trace_pc() refers to a 64bit pointer (in the comment)
> while the area pointer itself is (long *). The problematic part is that
> a 32bit application on a 64bit pointer will expect a four byte pointer
> while kernel uses an eight byte pointer.

I think the code is consistent in using 'unsigned long' for writing
regular pos/IP (except write_comp_data(), which has a comment about
it). The mentions of 64-bit in comments might be inaccurate though.
But I think it's working as expected:

- on 64-bit kernels, pos/IP can be up to 64-bit;
- on 32-bit kernels, pos/IP can only be up to 32-bit.

User space necessarily has to know about the bit-ness of its kernel,
because the coverage information is entirely dependent on the kernel
image. I think the examples in documentation weren't exhaustive in
this regard. At least that's my take -- Dmitry or Andrey would know
for sure (Dmitry is currently on vacation, but hopefully can clarify
next week).

Thanks,
-- Marco