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

[Masami Hiramatsu]

On Mon, 13 Jan 2020 11:23:31 -0800
"Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:

> On Mon, Jan 13, 2020 at 10:09:53PM +0900, Masami Hiramatsu wrote:
> > Hi Joel,
> > 
> > On Mon, 13 Jan 2020 12:16:40 +0900
> > Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> > > > > > 
> > > > > > 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...
> 
> The correct answer will be different in different situations.
> For example, code that might be called either with the mutex held or
> within an RCU read-side critical section will definitely need the _rcu()
> and the lockdep_is_held().  Code that looks OK to call from within
> RCU readers, but must not be (e.g., because it sleeps), will just as
> definitely need to avoid _rcu().

I see. So the patch [2/2] is just removing useless rcu_read_lock()
and use non RCU api for kprobe_table, because those code never be
called from rcu read-side critical section. (It makes a critical section
only for using RCU list operation)

>  (If the lack of _rcu() proves confusing,
> maybe list_for_each_entry() needs to grow an optional lockdep expression?)

That is OK for me, anyway the [2/2] also introduces some lockdep_assert_held()
instead of rcu_read_lock() so that lockdep can check sanity.

> 
> I am therefore personally OK with either approach, though in confusing
> cases a comment might help.
> 
> > I meant, do we always need to do something like below?
> > 
> > {
> > 	mutex_lock(&lock);
> > 	list_for_each_entry_rcu(list, ..., lockdep_is_held(&lock)) {
> > 		...
> > 	}
> > 	mutex_unlock(&lock);
> > }
> > 
> > BTW, I found another problem on this policy, since we don't have
> > list_for_each_*_safe() equivalents for RCU, we can not do a safe
> > loop on it. Should we call a find function for each time?
> 
> Good point.
> 
> RCU readers don't need _safe() because RCU grace periods provide this
> for free within RCU read-side critical sections.
> 
> So agreed, if you need _safe() on the update side, you would need to
> call list_for_each_entry_safe().  If this proves confusing due to RCU
> readers, maybe it should grow a lockdep expression?  In the meantime,
> lockdep_assert_held() could be used if needed to let people know that
> this should not be used in an RCU reader.

I think lockdep_assert_held() is enough.

> 
> Does that work, or am I missing part of the problem?
> 
> 							Thanx, Paul

Thank you,


-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>