Re: Default enable RCU list lockdep debugging with PROVE_RCU

["Paul E. McKenney" <paulmck@kernel.org>]

On Thu, May 14, 2020 at 11:46:23AM -0400, Qian Cai wrote:
> 
> 
> > On May 14, 2020, at 11:34 AM, Paul E. McKenney <paulmck@kernel.org> wrote:
> > 
> > On Thu, May 14, 2020 at 10:03:21AM -0400, Qian Cai wrote:
> >> 
> >> 
> >>> On May 14, 2020, at 9:54 AM, Paul E. McKenney <paulmck@kernel.org> wrote:
> >>> 
> >>> On Thu, May 14, 2020 at 09:44:28AM -0400, Qian Cai wrote:
> >>>> 
> >>>> 
> >>>>> On May 14, 2020, at 9:33 AM, Paul E. McKenney <paulmck@kernel.org> wrote:
> >>>>> 
> >>>>> On Thu, May 14, 2020 at 08:31:13AM -0400, Qian Cai wrote:
> >>>>>> 
> >>>>>> 
> >>>>>>> On May 14, 2020, at 8:25 AM, Stephen Rothwell <sfr@canb.auug.org.au> wrote:
> >>>>>>> 
> >>>>>>> Hi Paul,
> >>>>>>> 
> >>>>>>> This patch in the rcu tree
> >>>>>>> 
> >>>>>>> d13fee049fa8 ("Default enable RCU list lockdep debugging with PROVE_RCU")
> >>>>>>> 
> >>>>>>> is causing whack-a-mole in the syzbot testing of linux-next.  Because
> >>>>>>> they always do a debug build of linux-next, no testing is getting done. :-(
> >>>>>>> 
> >>>>>>> Can we find another way to find all the bugs that are being discovered
> >>>>>>> (very slowly)?
> >>>>>> 
> >>>>>> Alternatively, could syzbot to use PROVE_RCU=n temporarily because it can’t keep up with it? I personally found PROVE_RCU_LIST=y is still useful for my linux-next testing, and don’t want to lose that coverage overnight.
> >>>>> 
> >>>>> The problem is that PROVE_RCU is exactly PROVE_LOCKING, and asking people
> >>>>> to test without PROVE_LOCKING is a no-go in my opinion.  But of course
> >>>>> on the other hand if there is no testing of RCU list lockdep debugging,
> >>>>> those issues will never be found, let alone fixed.
> >>>>> 
> >>>>> One approach would be to do as Stephen asks (either remove d13fee049fa8
> >>>>> or pull it out of -next) and have testers force-enable the RCU list
> >>>>> lockdep debugging.
> >>>>> 
> >>>>> Would that work for you?
> >>>> 
> >>>> Alternatively, how about having
> >>>> 
> >>>> PROVE_RCU_LIST=n if DEBUG_AID_FOR_SYZBOT
> >>>> 
> >>>> since it is only syzbot can’t keep up with it?
> >>> 
> >>> Sound good to me, assuming that this works for the syzkaller guys.
> >>> Or could there be a "select PROVE_RCU_LIST" for the people who would
> >>> like to test it.
> >>> 
> >>> Alternatively, if we revert d13fee049fa8 from -next, I could provide
> >>> you a script that updates your .config to set both RCU_EXPERT and
> >>> PROVE_RCU_LIST.
> >>> 
> >>> There are a lot of ways to appraoch this.
> >>> 
> >>> So what would work best for everyone?
> >> 
> >> 
> >> If PROVE_RCU_LIST=n if DEBUG_AID_FOR_SYZBOT works for syzbot guys, that would be great, so other testing agents could still report/fix those RCU-list bugs and then pave a way for syzbot to return back once all those false positives had been sorted out.
> > 
> > On that, I must defer to the syzbot guys.
> > 
> >> Otherwise,  “select PROVE_RCU_LIST” *might* be better than buried into RCU_EXPERT where we will probably never saw those false positives been addressed since my configs does not cover a wide range of subsystems and probably not many other bots would enable RCU_EXPERT.
> > 
> > Yet another option would be to edit your local kernel/rcu/Kconfig.debug
> > and change the code to the following:
> > 
> > 	config PROVE_RCU_LIST
> > 		def_bool y
> > 		help
> > 		  Enable RCU lockdep checking for list usages. It is default
> > 		  enabled with CONFIG_PROVE_RCU.
> > 
> > Removing the RCU_EXPERT dependency would not go over at all well with
> > some people whose opinions are difficult to ignore.  ;-)
> 
> I am trying to not getting into a game of carrying any custom patch myself.
> 
> Let’s see what syzbot guys will say, and then I’ll enable RCU_EXPERT myself if needed, but again we probably never see PROVE_RCU_LIST to be used again in syzbot for this path. I surely have no cycles to expand the testing coverage for more subsystems at the moment.

Fair enough!  And yes, the Linux kernel is quite large, so I certainly am
not asking you to test the whole thing yourself.

								Thanx, Paul