Re: [PATCH -tip V2 0/2] kprobes: Fix RCU warning and cleanup

[Masami Hiramatsu]

On Sat, 11 Jan 2020 21:05:37 -0500
Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:

> On Sat, Jan 11, 2020 at 08:35:07AM +0900, Masami Hiramatsu wrote:
> > Hi Joel and Paul,
> > 
> > On Fri, 10 Jan 2020 16:14:38 -0500
> > Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> > 
> > > On Tue, Jan 07, 2020 at 09:15:35PM +0900, Masami Hiramatsu wrote:
> > > > Hello,
> > > > 
> > > > Anyone have any comment on this series?
> > > > Without this series, I still see the suspicious RCU warning for kprobe on -tip tree.
> > > 
> > > +Paul since RCU.
> > > 
> > > Hi Masami,
> > > 
> > > I believe I had commented before that I don't agree with this patch:
> > > https://lore.kernel.org/lkml/157535318870.16485.6366477974356032624.stgit@devnote2/
> > > 
> > > The rationale you used is to replace RCU-api with non-RCU api just to avoid
> > > warnings. I think a better approach is to use RCU api and pass the optional
> > > expression to silence the false-positive warnings by informing the RCU API
> > > about the fact that locks are held (similar to what we do for
> > > rcu_dereference_protected()). The RCU API will do additional checking
> > > (such as making sure preemption is disabled for safe RCU usage etc) as well.
> > 
> > Yes, that is what I did in [1/2] for get_kprobe().
> > Let me clarify the RCU list usage in [2/2].
> > 
> > With the careful check, other list traversals never be done in non-sleepable
> > context, those are always runs with kprobe_mutex held.
> > If I correctly understand the Documentation/RCU/listRCU.rst, we should/can use
> > non-RCU api for those cases, or do I miss something?
> 
> Yes, that is fine. However personally I prefer not to mix usage of
> list_for_each_entry_rcu() and list_for_each_entry() on the same pointer
> (kprobe_table). I think it is more confusing and error prone. Just use
> list_for_each_entry_rcu() everywhere and pass the appropriate lockdep
> expression, instead of calling lockdep_assert_held() independently. Is this
> not doable?

Hmm, but isn't it more confusing that user just take a mutex but
no rcu_read_lock() with list_for_each_entry_rcu()? In that case,
sometimes it might sleep inside list_for_each_entry_rcu(), I thought
that might be more confusing mind model for users...

Anyway, if so, please update Documentation/RCU/listRCU.rst too.

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>