Messages in this thread |  | | From | Dmitry Vyukov <> | Date | Tue, 18 Aug 2020 14:02:38 +0200 | Subject | Re: [PATCH] lockdep: Introduce CONFIG_LOCKDEP_LARGE |
| |
On Tue, Aug 18, 2020 at 1:07 PM Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp> wrote: > > On 2020/08/18 18:57, Dmitry Vyukov wrote: > > On Tue, Aug 4, 2020 at 4:36 AM Tetsuo Handa > > <penguin-kernel@i-love.sakura.ne.jp> wrote: > >> > >> Hello, Peter, Ingo and Will. > >> > >> (Q1) Can we change the capacity using kernel config? > >> > >> (Q2) If we can change the capacity, is it OK to specify these constants > >> independently? (In other words, is there inter-dependency among > >> these constants?) > > > > > > I think we should do this. > > syzbot uses a very beefy kernel config and very broad load. > > We are hitting "BUG: MAX_LOCKDEP_ENTRIES too low!" for the part 428 > > days and already hit it 96K times. It's just harming overall kernel > > testing: > > https://syzkaller.appspot.com/bug?id=3d97ba93fb3566000c1c59691ea427370d33ea1b > > > > I think it's better if exact values are not hardcoded, but rather > > specified in the config. Today we are switching from 4K to 8K, but as > > we enable more configs and learn to reach more code, we may need 16K. > > For short term, increasing the capacity would be fine. But for long term, I doubt. > > Learning more locks being held within one boot by enabling more configs, I suspect > that it becomes more and more timing dependent and difficult to hold all locks that > can generate a lockdep warning. > > > > > > >> (Q3) Do you think that we can extend lockdep to be used as a tool for auditing > >> locks held in kernel space and rebuilding lock dependency map in user space? > > > > This looks like lots of work. Also unpleasant dependencies on > > user-space. If there is a user-space component, it will need to be > > deployed to _all_ of kernel testing systems and for all users of > > syzkaller. And it will also be a dependency for reproducers. Currently > > one can run a C reproducer and get the errors message from LOCKDEP. It > > seems that a user-space component will make it significantly more > > complicated. > > My suggestion is to detach lockdep warning from realtime alarming. > > Since not all locks are always held (e.g. some locks are held only if exceeding > some threshold), requiring all locks being held within one boot sounds difficult. > Such requirement results in flaky bisection like "Fix bisection: failed" in > https://syzkaller.appspot.com/bug?id=b23ec126241ad0d86628de6eb5c1cff57d282632 . > > Then, I'm wishing that we could build non-realtime alarming based on all locks held > across all boots on each vmlinux file.
Unless I am missing something, deployment/maintenance story for this for syzbot, syzkaller users, other kernel testing, reproducer extraction, bisection, resproducer hermeticity is quite complicated. I don't see it outweighing any potential benefit in reporting quality.
I also don't see how it will improve reproducer/bisection quality: to confirm presence of a bug we still need to trigger all cycle edges within a single run anyway, it does not have to be a single VM, but still needs to be a single test case. And this "having all edges within a single test case" seems to be the root problem. I don't see how this proposal addresses this problem.
> >> On 2020/07/25 14:23, Tetsuo Handa wrote: > >>>> Also somebody may use it to _reduce_ size of the table for a smaller kernel. > >>> > >>> Maybe. But my feeling is that it is very rare that the kernel actually deadlocks > >>> as soon as lockdep warned the possibility of deadlock. > >>> > >>> Since syzbot runs many instances in parallel, a lot of CPU resource is spent for > >>> checking the same dependency tree. However, the possibility of deadlock can be > >>> warned for only locks held within each kernel boot, and it is impossible to hold > >>> all locks with one kernel boot. > >>> > >>> Then, it might be nice if lockdep can audit only "which lock was held from which > >>> context and what backtrace" and export that log like KCOV data (instead of evaluating > >>> the possibility of deadlock), and rebuild the whole dependency (and evaluate the > >>> possibility of deadlock) across multiple kernel boots in userspace. > >> >
|  |