Re: [PATCH -tip] kprobes: Lock rcu_read_lock() while searching kprobe

[Masami Hiramatsu]

Hi Ingo,

On Wed, 4 Dec 2019 08:12:39 -0800
"Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:

> On Wed, Dec 04, 2019 at 11:05:50AM +0100, Ingo Molnar wrote:
> > 
> > * Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > 
> > > >  * This list-traversal primitive may safely run concurrently with
> > > >  * the _rcu list-mutation primitives such as hlist_add_head_rcu()
> > > >  * as long as the traversal is guarded by rcu_read_lock().
> > > >  */
> > > > #define hlist_for_each_entry_rcu(pos, head, member, cond...)            \
> > > > 
> > > > is actively harmful. Why is it there?
> > > 
> > > For cases where common code might be invoked both from the reader
> > > (with RCU protection) and from the updater (protected by some
> > > lock).  This common code can then use the optional argument to
> > > hlist_for_each_entry_rcu() to truthfully tell lockdep that it might be
> > > called with either form of protection in place.
> > > 
> > > This also combines with the __rcu tag used to mark RCU-protected
> > > pointers, in which case sparse complains when a non-RCU API is applied
> > > to these pointers, to get back to your earlier question about use of
> > > hlist_for_each_entry_rcu() within the update-side lock.
> > > 
> > > But what are you seeing as actively harmful about all of this?
> > > What should we be doing instead?
> > 
> > Yeah, so basically in the write-locked path hlist_for_each_entry() 
> > generates (slightly) more efficient code than hlist_for_each_entry_rcu(), 
> > correct?
> 
> Potentially yes, if the READ_ONCE() constrains the compiler.  Or not,
> depending of course on the compiler and the surrounding code.

For this kprobes case, I can introduce get_kprobe_locked() which uses
 hlist_for_each_entry() instead of hlist_for_each_entry_rcu().

However, this sounds like a bit strange choice, because get_kprobe
(RCU version) should be used on "hot" paths (because it is lock-free),
and get_kprobe_locked() is used on slow paths.
If hlist_for_each_entry() can be more efficient, we will keep unefficient
API for hot paths, but use the efficient one for slow paths.

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>