Date:   Thu, 2 Dec 2021 09:35:31 +0800
From:   Kefeng Wang <wangkefeng.wang@...wei.com>
To:     Mark Rutland <mark.rutland@....com>
CC:     Marco Elver <elver@...gle.com>,
        Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will@...nel.org>,
        <linux-arm-kernel@...ts.infradead.org>,
        <linux-kernel@...r.kernel.org>, <joey.gouly@....com>
Subject: Re: [PATCH v2] arm64: Enable KCSAN


On 2021/12/1 19:53, Mark Rutland wrote:
> Hi Kefeng,
>
> On Mon, Nov 29, 2021 at 10:57:32PM +0800, Kefeng Wang wrote:
>> This patch enables KCSAN for arm64, with updates to build rules
>> to not use KCSAN for several incompatible compilation units.
>>
>> Resent GCC version(at least GCC10) made outline-atomics as the
>> default option(unlike Clang), which will cause linker errors
>> for kernel/kcsan/core.o.
>>
>> Disables the out-of-line atomics by no-outline-atomics to fix
>> the linker errors.
>>
>> Tested selftest and kcsan_test(built with GCC11 and Clang 13),
>> and all passed.
> Nice!
>
> I think there are a few additional bits and pieces we'll need:
>
> * Prior to clang 12.0.0, KCSAN would produce warnings with BTI, as I found in:
>
>    https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/commit/?h=arm64/kcsan&id=2d67c39ae4f619ca94d9790e09186e77922fa826
>
>    Since BTI is in defconfig, I think arm64's Kconfig should require a minimum
>    of clang 12.0.0 to enable KCSAN.

I don't have different clang version to test,  when check KCSAN,

commit eb32f9f990d9 ("kcsan: Improve some Kconfig comments") saids,


     The compiler instruments plain compound read-write operations
     differently (++, --, +=, -=, |=, &=, etc.), which allows KCSAN to
     distinguish them from other plain accesses. This is currently
     supported by Clang 12 or later.

Should we add a  "depends on CLANG_VERSION >= 120000"


>
> * In the past clang did not have an attribute to suppress tsan instrumenation
>    and would instrument noinstr regions. I'm not sure when clang gained the
>    relevant attribute to supress this, but we will need to depend on this
>    existing, either based on the clang version or with a test for the attribute.
>
>    (If we're lucky, clang 12.0.0 is sufficient, and we solve BTI and this in one
>    go).
>
>    I *think* GCC always had an attribute, but I'm not certain.
>
>    Marco, is there an existing dependency somewhere for this to work on x86? I
>    thought there was an objtool pass to NOP this out, but I couldn't find it in
>    mainline. If x86 is implicitly depending on a sufficiently recent version of
>    clang, we add something to the common KCSAN Kconfig for ARCH_WANTS_NO_INSTR?
>
> * There are some latent issues with some code (e.g. alternatives, patching, insn)
>    code being instrumentable even though this is unsound, and depending on
>    compiler choices this can happen to be fine or can result in boot-time
>    failures (I saw lockups when I started trying to add KCSAN for arm64).
>
>    While this isn't just a KCSAN problem, fixing that requires some fairly
>    significant rework to a bunch of code, and until that's done we're on very
>    shaky ground. So I'd like to make KCSAN depend on EXPERT for now.
>
>    I had an initial stab at fixing some of that, e.g.
>
>    https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=arm64/patching/rework
>    
>    Joey has started looking into this too.

Thanks for your information,  I don't know about this. As your say, we 
could add a depend on EXPERT

for now and more explanation into changlog.

>
> * When I last tested, for simple boots I would get frequent KCSAN splats for a
>    few common issues, and those drowned out all other reports.
>
>    One case was manipulation of thread_info::flags, which Thomas Gleixner has
>    queued some fixes at:
>
>    https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/log/?h=core/entry
>   
>    There were some other common failures, e.g. accesses to task_struct::on_cpu,
>    and I hadn't had the chance to investigate/fix those, beyond a (likely
>    unsound) hack:
>
>    https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/commit/?h=arm64/kcsan&id=4fe9d6c2ef85257d80291086e4514eaaebd3504e
>
>    It would be good if we could identify the most frequent problems (e.g. those
>    that will occur when just booting) before we enable this generally, to avoid
>    everyone racing to report/fix those as soon as we enable the feature.
>
>    When you tested, did KCSAN flag anything beyond the selftests?

Yes, there are some KCSAN reports, but this is not only exist on arm64, 
I saw  owner->on_cpu warning

on x86 too, eg, we also hack to disable it via data_race.

Reported by Kernel Concurrency Sanitizer on:
CPU: 7 PID: 2530 Comm: syz-executor.11 Not tainted 5.10.0+ #113
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.104/01/2014
==================================================================
BUG: KCSAN: data-race in rwsem_spin_on_owner+0xf4/0x180

race at unknown origin, with read to 0xffff9767d3becfac of 4 bytes by task 18119 on cpu 0:
  rwsem_spin_on_owner+0xf4/0x180
  rwsem_optimistic_spin+0x48/0x480
  rwsem_down_read_slowpath+0x4a0/0x670
  down_read+0x69/0x190
  process_vm_rw+0x41e/0x840
  __x64_sys_process_vm_writev+0x76/0x90
  do_syscall_64+0x37/0x50
  entry_SYSCALL_64_after_hwframe+0x44/0xa9

>
>> Signed-off-by: Kefeng Wang <wangkefeng.wang@...wei.com>
>> ---
>> Tested on Qemu with clang 13 / gcc 11, based on 5.16-rc3.
>>
>> [    0.221518] kcsan: enabled early
>> [    0.222422] kcsan: strict mode configured
>> ...
>> [    5.839223] kcsan: selftest: 3/3 tests passed
>> ...
>> [  517.895102] # kcsan: pass:24 fail:0 skip:0 total:24
>> [  517.896393] # Totals: pass:168 fail:0 skip:0 total:168
>> [  517.897502] ok 1 - kcsan
>>
>> v2:
>> - tested on GCC11 and disable outline-atomics for kernel/kcsan/core.c
>>    suggested by Marco Elver
>>
>>   arch/arm64/Kconfig               | 1 +
>>   arch/arm64/kernel/vdso/Makefile  | 1 +
>>   arch/arm64/kvm/hyp/nvhe/Makefile | 1 +
>>   kernel/kcsan/Makefile            | 1 +
>>   4 files changed, 4 insertions(+)
> Aside from the `-mno-outline-atomics` portion, this looks basically the same as what I had in:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/commit/?h=arm64/kcsan&id=2d67c39ae4f619ca94d9790e09186e77922fa826
>
> ... so this looks good to me modulo the comments above.
>
> Thanks,
> Mark.

Powered by blists - more mailing lists