On Tue, Jan 21, 2025 at 02:49:13PM +0100, Vlastimil Babka wrote:
> On 1/21/25 2:33 PM, Uladzislau Rezki wrote:
> > On Mon, Jan 20, 2025 at 11:06:13PM +0100, Vlastimil Babka wrote:
> >> On 12/16/24 17:46, Paul E. McKenney wrote:
> >>> On Mon, Dec 16, 2024 at 04:55:06PM +0100, Uladzislau Rezki wrote:
> >>>> On Mon, Dec 16, 2024 at 04:44:41PM +0100, Vlastimil Babka wrote:
> >>>>> On 12/16/24 16:41, Uladzislau Rezki wrote:
> >>>>>> On Mon, Dec 16, 2024 at 03:20:44PM +0100, Vlastimil Babka wrote:
> >>>>>>> On 12/16/24 12:03, Uladzislau Rezki wrote:
> >>>>>>>> On Sun, Dec 15, 2024 at 06:30:02PM +0100, Vlastimil Babka wrote:
> >>>>>>>>
> >>>>>>>>> Also how about a followup patch moving the rcu-tiny implementation of
> >>>>>>>>> kvfree_call_rcu()?
> >>>>>>>>>
> >>>>>>>> As, Paul already noted, it would make sense. Or just remove a tiny
> >>>>>>>> implementation.
> >>>>>>>
> >>>>>>> AFAICS tiny rcu is for !SMP systems. Do they benefit from the "full"
> >>>>>>> implementation with all the batching etc or would that be unnecessary overhead?
> >>>>>>>
> >>>>>> Yes, it is for a really small systems with low amount of memory. I see
> >>>>>> only one overhead it is about driving objects in pages. For a small
> >>>>>> system it can be critical because we allocate.
> >>>>>>
> >>>>>> From the other hand, for a tiny variant we can modify the normal variant
> >>>>>> by bypassing batching logic, thus do not consume memory(for Tiny case)
> >>>>>> i.e. merge it to a normal kvfree_rcu() path.
> >>>>>
> >>>>> Maybe we could change it to use CONFIG_SLUB_TINY as that has similar use
> >>>>> case (less memory usage on low memory system, tradeoff for worse performance).
> >>>>>
> >>>> Yep, i also was thinking about that without saying it :)
> >>>
> >>> Works for me as well!
> >>
> >> Hi, so I tried looking at this. First I just moved the code to slab as seen
> >> in the top-most commit here [1]. Hope the non-inlined __kvfree_call_rcu() is
> >> not a show-stopper here.
> >>
> >> Then I wanted to switch the #ifdefs from CONFIG_TINY_RCU to CONFIG_SLUB_TINY
> >> to control whether we use the full blown batching implementation or the
> >> simple call_rcu() implmentation, and realized it's not straightforward and
> >> reveals there are still some subtle dependencies of kvfree_rcu() on RCU
> >> internals :)
> >>
> >> Problem 1: !CONFIG_SLUB_TINY with CONFIG_TINY_RCU
> >>
> >> AFAICS the batching implementation includes kfree_rcu_scheduler_running()
> >> which is called from rcu_set_runtime_mode() but only on TREE_RCU. Perhaps
> >> there are other facilities the batching implementation needs that only
> >> exists in the TREE_RCU implementation
> >>
> >> Possible solution: batching implementation depends on both !CONFIG_SLUB_TINY
> >> and !CONFIG_TINY_RCU. I think it makes sense as both !SMP systems and small
> >> memory systems are fine with the simple implementation.
> >>
> >> Problem 2: CONFIG_TREE_RCU with !CONFIG_SLUB_TINY
> >>
> >> AFAICS I can't just make the simple implementation do call_rcu() on
> >> CONFIG_TREE_RCU, because call_rcu() no longer knows how to handle the fake
> >> callback (__is_kvfree_rcu_offset()) - I see how rcu_reclaim_tiny() does that
> >> but no such equivalent exists in TREE_RCU. Am I right?
> >>
> >> Possible solution: teach TREE_RCU callback invocation to handle
> >> __is_kvfree_rcu_offset() again, perhaps hide that branch behind #ifndef
> >> CONFIG_SLUB_TINY to avoid overhead if the batching implementation is used.
> >> Downside: we visibly demonstrate how kvfree_rcu() is not purely a slab thing
> >> but RCU has to special case it still.
> >>
> >> Possible solution 2: instead of the special offset handling, SLUB provides a
> >> callback function, which will determine pointer to the object from the
> >> pointer to a middle of it without knowing the rcu_head offset.
> >> Downside: this will have some overhead, but SLUB_TINY is not meant to be
> >> performant anyway so we might not care.
> >> Upside: we can remove __is_kvfree_rcu_offset() from TINY_RCU as well
> >>
> >> Thoughts?
> >>
> > For the call_rcu() and to be able to reclaim over it we need to patch the
> > tree.c(please note TINY already works):
> >
> > <snip>
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index b1f883fcd918..ab24229dfa73 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -2559,13 +2559,19 @@ static void rcu_do_batch(struct rcu_data *rdp)
> > debug_rcu_head_unqueue(rhp);
> >
> > rcu_lock_acquire(&rcu_callback_map);
> > - trace_rcu_invoke_callback(rcu_state.name, rhp);
> >
> > f = rhp->func;
> > - debug_rcu_head_callback(rhp);
> > - WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
> > - f(rhp);
> >
> > + if (__is_kvfree_rcu_offset((unsigned long) f)) {
> > + trace_rcu_invoke_kvfree_callback("", rhp, (unsigned long) f);
> > + kvfree((void *) rhp - (unsigned long) f);
> > + } else {
> > + trace_rcu_invoke_callback(rcu_state.name, rhp);
> > + debug_rcu_head_callback(rhp);
> > + WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
> > + f(rhp);
> > + }
> > rcu_lock_release(&rcu_callback_map);
>
> Right so that's the first Possible solution, but without the #ifdef. So
> there's an overhead of checking __is_kvfree_rcu_offset() even if the
> batching is done in slab and this function is never called with an offset.
>
Or fulfilling a missing functionality? TREE is broken in that sense
whereas a TINY handles it without any issues.
It can be called for SLUB_TINY option, just call_rcu() instead of
batching layer. And yes, kvfree_rcu_barrier() switches to rcu_barrier().
>
> After coming up with Possible solution 2, I've started liking the idea
> more as RCU could then forget about the __is_kvfree_rcu_offset()
> "callbacks" completely, and the performant case of TREE_RCU + batching
> would be unaffected.
>
I doubt it is a performance issue :)
>
> I'm speculating perhaps if there was not CONFIG_SLOB in the past, the
> __is_kvfree_rcu_offset() would never exist in the first place? SLAB and
> SLUB both can determine start of the object from a pointer to the middle
> of it, while SLOB couldn't.
>
We needed just to reclaim over RCU. So, i do not know. Paul probably
knows more then me :)
> > /*
> > <snip>
> >
> > Mixing up CONFIG_SLUB_TINY with CONFIG_TINY_RCU in the slab_common.c
> > should be avoided, i.e. if we can, we should eliminate a dependency on
> > TREE_RCU or TINY_RCU in a slab. As much as possible.
> >
> > So, it requires a more closer look for sure :)
>
> That requires solving Problem 1 above, but question is if it's worth the
> trouble. Systems running TINY_RCU are unlikely to benefit from the batching?
>
> But sure there's also possibility to hide these dependencies in KConfig,
> so the slab code would only consider a single (for example) #ifdef
> CONFIG_KVFREE_RCU_BATCHING that would be set automatically depending on
> TREE_RCU and !SLUB_TINY.
>
It is for small systems. We can use TINY or !SMP. We covered this AFAIR
that a single CPU system should not go with batching:
#ifdef SLUB_TINY || !SMP || TINY_RCU
or:
config TINY_RCU
bool
default y if !PREEMPT_RCU && !SMP
+ select SLUB_TINY
Paul, more input?
--
Uladzislau Rezki
