lkml.org 
[lkml]   [2019]   [Sep]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Kernel Concurrency Sanitizer (KCSAN)
On Fri, Sep 20, 2019 at 04:18:57PM +0200, Marco Elver wrote:
> Hi all,

Hi,

> We would like to share a new data-race detector for the Linux kernel:
> Kernel Concurrency Sanitizer (KCSAN) --
> https://github.com/google/ktsan/wiki/KCSAN (Details:
> https://github.com/google/ktsan/blob/kcsan/Documentation/dev-tools/kcsan.rst)

Nice!

BTW kcsan_atomic_next() is missing a stub definition in <linux/kcsan.h>
when !CONFIG_KCSAN:

https://github.com/google/ktsan/commit/a22a093a0f0d0b582c82cdbac4f133a3f61d207c#diff-19d7c475b4b92aab8ba440415ab786ec

... and I think the kcsan_{begin,end}_atomic() stubs need to be static
inline too.

It looks like this is easy enough to enable on arm64, with the only real
special case being secondary_start_kernel() which we might want to
refactor to allow some portions to be instrumented.

I pushed the trivial patches I needed to get arm64 booting to my arm64/kcsan
branch:

git://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git arm64/kcsan

We have some interesting splats at boot time in stop_machine, which
don't seem to have been hit/fixed on x86 yet in the kcsan-with-fixes
branch, e.g.

[ 0.237939] ==================================================================
[ 0.239431] BUG: KCSAN: data-race in multi_cpu_stop+0xa8/0x198 and set_state+0x80/0xb0
[ 0.241189]
[ 0.241606] write to 0xffff00001003bd00 of 4 bytes by task 24 on cpu 3:
[ 0.243435] set_state+0x80/0xb0
[ 0.244328] multi_cpu_stop+0x16c/0x198
[ 0.245406] cpu_stopper_thread+0x170/0x298
[ 0.246565] smpboot_thread_fn+0x40c/0x560
[ 0.247696] kthread+0x1a8/0x1b0
[ 0.248586] ret_from_fork+0x10/0x18
[ 0.249589]
[ 0.250006] read to 0xffff00001003bd00 of 4 bytes by task 14 on cpu 1:
[ 0.251804] multi_cpu_stop+0xa8/0x198
[ 0.252851] cpu_stopper_thread+0x170/0x298
[ 0.254008] smpboot_thread_fn+0x40c/0x560
[ 0.255135] kthread+0x1a8/0x1b0
[ 0.256027] ret_from_fork+0x10/0x18
[ 0.257036]
[ 0.257449] Reported by Kernel Concurrency Sanitizer on:
[ 0.258918] CPU: 1 PID: 14 Comm: migration/1 Not tainted 5.3.0-00007-g67ab35a199f4-dirty #3
[ 0.261241] Hardware name: linux,dummy-virt (DT)
[ 0.262517] ==================================================================

> To those of you who we mentioned at LPC that we're working on a
> watchpoint-based KTSAN inspired by DataCollider [1], this is it (we
> renamed it to KCSAN to avoid confusion with KTSAN).
> [1] http://usenix.org/legacy/events/osdi10/tech/full_papers/Erickson.pdf
>
> In the coming weeks we're planning to:
> * Set up a syzkaller instance.
> * Share the dashboard so that you can see the races that are found.
> * Attempt to send fixes for some races upstream (if you find that the
> kcsan-with-fixes branch contains an important fix, please feel free to
> point it out and we'll prioritize that).
>
> There are a few open questions:
> * The big one: most of the reported races are due to unmarked
> accesses; prioritization or pruning of races to focus initial efforts
> to fix races might be required. Comments on how best to proceed are
> welcome. We're aware that these are issues that have recently received
> attention in the context of the LKMM
> (https://lwn.net/Articles/793253/).

I think the big risk here is drive-by "fixes" masking the warnings
rather than fixing the actual issue. It's easy for people to suppress a
warning with {READ,WRITE}_ONCE(), so they're liable to do that even the
resulting race isn't benign.

I don't have a clue how to prevent that, though.

> * How/when to upstream KCSAN?

I would love to see this soon!

Thanks,
Mark.

\
 
 \ /
  Last update: 2019-09-20 18:32    [W:0.065 / U:3.992 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site