Re: [tip: locking/kcsan] READ_ONCE: Use data_race() to avoid KCSAN instrumentation

[Nathan Chancellor]

On Thu, May 21, 2020 at 12:17:12AM +0200, Borislav Petkov wrote:
> Hi,
> 
> On Tue, May 12, 2020 at 02:36:53PM -0000, tip-bot2 for Will Deacon wrote:
> > The following commit has been merged into the locking/kcsan branch of tip:
> > 
> > Commit-ID:     cdd28ad2d8110099e43527e96d059c5639809680
> > Gitweb:        https://git.kernel.org/tip/cdd28ad2d8110099e43527e96d059c5639809680
> > Author:        Will Deacon <will@xxxxxxxxxx>
> > AuthorDate:    Mon, 11 May 2020 21:41:49 +01:00
> > Committer:     Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > CommitterDate: Tue, 12 May 2020 11:04:17 +02:00
> > 
> > READ_ONCE: Use data_race() to avoid KCSAN instrumentation
> > 
> > Rather then open-code the disabling/enabling of KCSAN across the guts of
> > {READ,WRITE}_ONCE(), defer to the data_race() macro instead.
> > 
> > Signed-off-by: Will Deacon <will@xxxxxxxxxx>
> > Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Cc: Marco Elver <elver@xxxxxxxxxx>
> > Link: https://lkml.kernel.org/r/20200511204150.27858-18-will@xxxxxxxxxx
> 
> so this commit causes a kernel build slowdown depending on the .config
> of between 50% and over 100%. I just bisected locking/kcsan and got
> 
> NOT_OK:	cdd28ad2d811 READ_ONCE: Use data_race() to avoid KCSAN instrumentation
> OK:	88f1be32068d kcsan: Rework data_race() so that it can be used by READ_ONCE()
> 
> with a simple:
> 
> $ git clean -dqfx && mk defconfig
> $ time make -j<NUM_CORES+1>
> 
> I'm not even booting the kernels - simply checking out the above commits
> and building the target kernels. I.e., something in that commit is
> making gcc go nuts in the compilation phases.
> 
> -- 
> Regards/Gruss,
>     Boris.
> 
> https://people.kernel.org/tglx/notes-about-netiquette

For what it's worth, I also noticed the same thing with clang. I only
verified the issue in one of my first build targets, an arm defconfig
build, which regressed from 2.5 minutes to 10+ minutes.

More details available on our issue tracker (Nick did some more
profiling on other configs with both clang and gcc):

https://github.com/ClangBuiltLinux/linux/issues/1032

More than happy to do further triage as time permits. I do note Marco's
message about the upcoming series to eliminate this but it would be nice
if this did not regress in the meantime.

Cheers,
Nathan