Message-ID: <9cd799fd-e4e9-4a24-9e91-0443592b2960@paulmck-laptop>
Date:   Tue, 7 Nov 2023 08:49:40 -0800
From:   "Paul E. McKenney" <paulmck@...nel.org>
To:     Uladzislau Rezki <urezki@...il.com>
Cc:     RCU <rcu@...r.kernel.org>,
        Neeraj upadhyay <Neeraj.Upadhyay@....com>,
        Boqun Feng <boqun.feng@...il.com>,
        Hillf Danton <hdanton@...a.com>,
        Joel Fernandes <joel@...lfernandes.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Oleksiy Avramchenko <oleksiy.avramchenko@...y.com>,
        Frederic Weisbecker <frederic@...nel.org>
Subject: Re: [PATCH v2 1/3] rcu: Reduce synchronize_rcu() latency

On Tue, Nov 07, 2023 at 03:21:32PM +0100, Uladzislau Rezki wrote:
> On Mon, Nov 06, 2023 at 09:32:00PM -0800, Paul E. McKenney wrote:
> > On Mon, Nov 06, 2023 at 01:04:42PM +0100, Uladzislau Rezki wrote:
> > > On Wed, Nov 01, 2023 at 09:35:30PM -0700, Paul E. McKenney wrote:
> > > > On Mon, Oct 30, 2023 at 02:12:52PM +0100, Uladzislau Rezki (Sony) wrote:
> > > > > A call to a synchronize_rcu() can be optimized from a latency
> > > > > point of view. Workloads which depend on this can benefit of it.
> > > > > 
> > > > > The delay of wakeme_after_rcu() callback, which unblocks a waiter,
> > > > > depends on several factors:
> > > > > 
> > > > > - how fast a process of offloading is started. Combination of:
> > > > >     - !CONFIG_RCU_NOCB_CPU/CONFIG_RCU_NOCB_CPU;
> > > > >     - !CONFIG_RCU_LAZY/CONFIG_RCU_LAZY;
> > > > >     - other.
> > > > > - when started, invoking path is interrupted due to:
> > > > >     - time limit;
> > > > >     - need_resched();
> > > > >     - if limit is reached.
> > > > > - where in a nocb list it is located;
> > > > > - how fast previous callbacks completed;
> > > > > 
> > > > > Example:
> > > > > 
> > > > > 1. On our embedded devices i can easily trigger the scenario when
> > > > > it is a last in the list out of ~3600 callbacks:
> > > > > 
> > > > > <snip>
> > > > >   <...>-29      [001] d..1. 21950.145313: rcu_batch_start: rcu_preempt CBs=3613 bl=28
> > > > > ...
> > > > >   <...>-29      [001] ..... 21950.152578: rcu_invoke_callback: rcu_preempt rhp=00000000b2d6dee8 func=__free_vm_area_struct.cfi_jt
> > > > >   <...>-29      [001] ..... 21950.152579: rcu_invoke_callback: rcu_preempt rhp=00000000a446f607 func=__free_vm_area_struct.cfi_jt
> > > > >   <...>-29      [001] ..... 21950.152580: rcu_invoke_callback: rcu_preempt rhp=00000000a5cab03b func=__free_vm_area_struct.cfi_jt
> > > > >   <...>-29      [001] ..... 21950.152581: rcu_invoke_callback: rcu_preempt rhp=0000000013b7e5ee func=__free_vm_area_struct.cfi_jt
> > > > >   <...>-29      [001] ..... 21950.152582: rcu_invoke_callback: rcu_preempt rhp=000000000a8ca6f9 func=__free_vm_area_struct.cfi_jt
> > > > >   <...>-29      [001] ..... 21950.152583: rcu_invoke_callback: rcu_preempt rhp=000000008f162ca8 func=wakeme_after_rcu.cfi_jt
> > > > >   <...>-29      [001] d..1. 21950.152625: rcu_batch_end: rcu_preempt CBs-invoked=3612 idle=....
> > > > > <snip>
> > > > > 
> > > > > 2. We use cpuset/cgroup to classify tasks and assign them into
> > > > > different cgroups. For example "backgrond" group which binds tasks
> > > > > only to little CPUs or "foreground" which makes use of all CPUs.
> > > > > Tasks can be migrated between groups by a request if an acceleration
> > > > > is needed.
> > > > > 
> > > > > See below an example how "surfaceflinger" task gets migrated.
> > > > > Initially it is located in the "system-background" cgroup which
> > > > > allows to run only on little cores. In order to speed it up it
> > > > > can be temporary moved into "foreground" cgroup which allows
> > > > > to use big/all CPUs:
> > > > > 
> > > > > cgroup_attach_task():
> > > > >  -> cgroup_migrate_execute()
> > > > >    -> cpuset_can_attach()
> > > > >      -> percpu_down_write()
> > > > >        -> rcu_sync_enter()
> > > > >          -> synchronize_rcu()
> > > > >    -> now move tasks to the new cgroup.
> > > > >  -> cgroup_migrate_finish()
> > > > > 
> > > > > <snip>
> > > > >          rcuop/1-29      [000] .....  7030.528570: rcu_invoke_callback: rcu_preempt rhp=00000000461605e0 func=wakeme_after_rcu.cfi_jt
> > > > >     PERFD-SERVER-1855    [000] d..1.  7030.530293: cgroup_attach_task: dst_root=3 dst_id=22 dst_level=1 dst_path=/foreground pid=1900 comm=surfaceflinger
> > > > >    TimerDispatch-2768    [002] d..5.  7030.537542: sched_migrate_task: comm=surfaceflinger pid=1900 prio=98 orig_cpu=0 dest_cpu=4
> > > > > <snip>
> > > > > 
> > > > > "Boosting a task" depends on synchronize_rcu() latency:
> > > > > 
> > > > > - first trace shows a completion of synchronize_rcu();
> > > > > - second shows attaching a task to a new group;
> > > > > - last shows a final step when migration occurs.
> > > > > 
> > > > > 3. To address this drawback, maintain a separate track that consists
> > > > > of synchronize_rcu() callers only. After completion of a grace period
> > > > > users are awaken directly, it is limited by allowed threshold, others
> > > > > are deferred(if still exist) to a worker to complete the rest.
> > > > > 
> > > > > 4. This patch reduces the latency of synchronize_rcu() approximately
> > > > > by ~30-40% on synthetic tests. The real test case, camera launch time,
> > > > > shows(time is in milliseconds):
> > > > > 
> > > > > 1-run 542 vs 489 improvement 9%
> > > > > 2-run 540 vs 466 improvement 13%
> > > > > 3-run 518 vs 468 improvement 9%
> > > > > 4-run 531 vs 457 improvement 13%
> > > > > 5-run 548 vs 475 improvement 13%
> > > > > 6-run 509 vs 484 improvement 4%
> > > > > 
> > > > > Synthetic test:
> > > > > 
> > > > > Hardware: x86_64 64 CPUs, 64GB of memory
> > > > > 
> > > > > - 60K tasks(simultaneous);
> > > > > - each task does(1000 loops)
> > > > >      synchronize_rcu();
> > > > >      kfree(p);
> > > > > 
> > > > > default: CONFIG_RCU_NOCB_CPU: takes 323 seconds to complete all users;
> > > > > patch: CONFIG_RCU_NOCB_CPU: takes 240 seconds to complete all users.
> > > > > 
> > > > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@...il.com>
> > > > 
> > > > This looks pretty close!  Some questions and comments below, much of
> > > > which being what I managed not to write down in earlier discussions.  :-/
> > > > 
> > > Sounds good :)
> > > 
> > > > > ---
> > > > >  kernel/rcu/tree.c     | 154 +++++++++++++++++++++++++++++++++++++++++-
> > > > >  kernel/rcu/tree_exp.h |   2 +-
> > > > >  2 files changed, 154 insertions(+), 2 deletions(-)
> > > > > 
> > > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > > index 78554e7181dd..f04846b543de 100644
> > > > > --- a/kernel/rcu/tree.c
> > > > > +++ b/kernel/rcu/tree.c
> > > > > @@ -1384,6 +1384,125 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
> > > > >  		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > > >  }
> > > > >  
> > > > > +/*
> > > > > + * There are three lists for handling synchronize_rcu() users.
> > > > > + * A first list corresponds to new coming users, second for users
> > > > > + * which wait for a grace period and third is for which a grace
> > > > > + * period is passed.
> > > > > + */
> > > > > +static struct sr_normal_state {
> > > > > +	struct llist_head srs_next;	/* request a GP users. */
> > > > > +	struct llist_head srs_wait;	/* wait for GP users. */
> > > > > +	struct llist_head srs_done;	/* ready for GP users. */
> > > > > +
> > > > > +	/*
> > > > > +	 * In order to add a batch of nodes to already
> > > > > +	 * existing srs-done-list, a tail of srs-wait-list
> > > > > +	 * is maintained.
> > > > > +	 */
> > > > > +	struct llist_node *srs_wait_tail;
> > > > > +} sr;
> > > > 
> > > > It would be good to put these fields into the rcu_state structure.
> > > > Unlike kfree_rcu(), I have no ambitions for the mm guys ever taking
> > > > this one.  ;-)
> > > > 
> > > OK. I will rework it. It is better to keep it in one solid place.
> > 
> > Very good, thank you!
> > 
> > > > > +/* Disabled by default. */
> > > > > +static int rcu_normal_wake_from_gp;
> > > > > +module_param(rcu_normal_wake_from_gp, int, 0644);
> > > > > +
> > > > > +static void rcu_sr_normal_complete(struct llist_node *node)
> > > > > +{
> > > > > +	struct rcu_synchronize *rs = container_of(
> > > > > +		(struct rcu_head *) node, struct rcu_synchronize, head);
> > > > > +	unsigned long oldstate = (unsigned long) rs->head.func;
> > > > > +
> > > > > +	WARN_ONCE(!poll_state_synchronize_rcu(oldstate),
> > > > > +		"A full grace period is not passed yet: %lu",
> > > > > +		rcu_seq_diff(get_state_synchronize_rcu(), oldstate));
> > > > 
> > > > This needs to either:
> > > > 
> > > > 1.	Use poll_state_synchronize_rcu_full(), or
> > > > 
> > > > 2.	Avoid firing unless expedited grace periods have been disabled.
> > > > 	Note that forcing synchronize_rcu() to synchronize_rcu_expedited()
> > > > 	does not help because there might still be call_rcu() invocations
> > > > 	advancing normal grace periods.
> > > > 
> > > > As it stands, you can have false-positive WARN_ONCE()s.  These can happen
> > > > when a normal and an expedited grace period overlap in time.
> > > > 
> > > I prefer an option [2]:
> > > 
> > > <snip>
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index 189975f57e78..85f3e7d3642e 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -1413,7 +1413,7 @@ static void rcu_sr_normal_complete(struct llist_node *node)
> > >                 (struct rcu_head *) node, struct rcu_synchronize, head);
> > >         unsigned long oldstate = (unsigned long) rs->head.func;
> > > 
> > > -       WARN_ONCE(!poll_state_synchronize_rcu(oldstate),
> > > +       WARN_ONCE(!rcu_gp_is_expedited() && !poll_state_synchronize_rcu(oldstate),
> > >                 "A full grace period is not passed yet: %lu",
> > >                 rcu_seq_diff(get_state_synchronize_rcu(), oldstate));
> > > 
> > > <snip>
> > 
> > In this case, don't you instead need rcu_gp_is_normal()?
> > 
> > Ah, but this thing can be changed via sysfs.  For the diagnostic
> > to be reliable, expedited grace periods have to have been disabled
> > for the full time from the start_poll_synchronize_rcu() to the final
> > poll_state_synchronize_rcu().  And userspace can toggle rcu_normal via
> > sysfs as often and as many times as they like.  :-/
> > 
> There is a toggle, indeed. I tried to reproduce such overlap adding
> an extra worker that does start_poll_synchronize_rcu_full() in a tight loop.
> I was not able to trigger that warning.
> 
> Do you have something that can easily trigger it? I mean some proposal
> or steps to test. Probably i should try what you wrote, regarding
> toggling from user space.
> 
> > I can imagine ways around this, but they are a bit ugly.  They end
> > up being things like recording a timestamp on every sysfs change to
> > rcu_normal, and then using that timestamp to deduce whether there could
> > possibly have been sysfs activity on rcu_normal in the meantime.
> > 
> > It feels like it should be so easy...  ;-)
> > 
> Hmm.. Yes it requires more deep analysis :)

Maybe make that WARN_ONCE() condition also test a separate Kconfig
option that depends on both DEBUG_KERNEL and RCU_EXPERT?

> > > > > +	/* Finally. */
> > > > > +	complete(&rs->completion);
> > > > > +}
> > > > > +
> > > > > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
> > > > > +{
> > > > > +	struct llist_node *done, *rcu, *next;
> > > > > +
> > > > > +	done = llist_del_all(&sr.srs_done);
> > > > > +	if (!done)
> > > > > +		return;
> > > > > +
> > > > > +	llist_for_each_safe(rcu, next, done)
> > > > > +		rcu_sr_normal_complete(rcu);
> > > > > +}
> > > > > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work);
> > > > > +
> > > > > +/*
> > > > > + * This is hard-coded and it is a maximum number of
> > > > > + * synchronize_rcu() users(might be +1 extra), which
> > > > > + * are awaken directly by the rcu_gp_kthread(). The
> > > > > + * reset is deferred to a dedicated worker.
> > > > 
> > > > s/reset/rest/
> > > > 
> > > Typo. Thanks!
> > > 
> > > > > + */
> > > > > +#define MAX_SR_WAKE_FROM_GP 5
> > > > > +
> > > > > +/*
> > > > > + * Helper function for rcu_gp_cleanup().
> > > > > + */
> > > > > +static void rcu_sr_normal_gp_cleanup(void)
> > > > > +{
> > > > > +	struct llist_node *head, *tail, *pos;
> > > > > +	int i = 0;
> > > > > +
> > > > > +	if (llist_empty(&sr.srs_wait))
> > > > > +		return;
> > > > > +
> > > > > +	tail = READ_ONCE(sr.srs_wait_tail);
> > > > > +	head = __llist_del_all(&sr.srs_wait);
> > > > > +
> > > > > +	llist_for_each_safe(pos, head, head) {
> > > > > +		rcu_sr_normal_complete(pos);
> > > > > +
> > > > > +		if (++i == MAX_SR_WAKE_FROM_GP) {
> > > > > +			/* If last, process it also. */
> > > > > +			if (head && !head->next)
> > > > > +				continue;
> > > > > +			break;
> > > > 
> > > > Save a line this way?
> > > > 
> > > > 			if (!head || head->next)
> > > > 				break;
> > > I would like to process clients from a GP-kthread but i am not
> > > allowed to offload all by the threshold. If last client is left
> > > i process it also, since we lose nothing and instead of kicking
> > > a worker to do a final job we process it right away.
> > 
> > Unless I blew my de Morgan transformation (which I might well have done),
> > the one-line approach should be functionally identical to your original.
> > 
> Makes sense. After applying, it does absolutely the same job. I need to
> update the comment accordingly.

Thank you!

> > > > > +		}
> > > > > +	}
> > > > > +
> > > > > +	if (head) {
> > > > > +		/* Can be not empty. */
> > > > > +		llist_add_batch(head, tail, &sr.srs_done);
> > > > > +		queue_work(system_highpri_wq, &sr_normal_gp_cleanup);
> > > > > +	}
> > > > > +}
> > > > > +
> > > > > +/*
> > > > > + * Helper function for rcu_gp_init().
> > > > > + */
> > > > > +static void rcu_sr_normal_gp_init(void)
> > > > > +{
> > > > > +	struct llist_node *head, *tail;
> > > > > +
> > > > > +	if (llist_empty(&sr.srs_next))
> > > > > +		return;
> > > > > +
> > > > > +	tail = llist_del_all(&sr.srs_next);
> > > > > +	head = llist_reverse_order(tail);
> > > > 
> > > > Hmmm...  I am not loving this list-reverse operation.  Once someone
> > > > figures out how to generate a long list, it is going to hurt quite badly.
> > > > 
> > > > Except...  Why do we need to reverse the list in the first place?
> > > > It appears that one reason is to be able to get the tail of the list.
> > > > Is it also necessary to do the wakeups in order, or could they be
> > > > reversed?  It seems like they should -- the average latency would remain
> > > > the same.  If so, couldn't we have a single llist with two pointers into
> > > > it (more accurately, to its tail pointers), one for the first done item,
> > > > and the other for the first item waiting on the current grace period?
> > > > 
> > > > Then it would not be necessary to reverse the list, nor would it be
> > > > necessary to move elemetns from one list to another.  Just copy one
> > > > pointer to the next.
> > > > 
> > > > If it ever becomes necessary to put extra elements back, which would be
> > > > challenging if there were no other elements in the list.  The usual way
> > > > to handle this is to have a dummy element to isolate the enqueuers from
> > > > the requeuer.  The GP kthread then enqueues the dummy element if the
> > > > list is empty, which means that enqueue and optimized wakeup are never
> > > > looking at the same pointer.  Alternatively, just use dummy elements to
> > > > mark the segments in the list, with the added pointers always referencing
> > > > these dummy elements.  Might need a VC to make this make sense...
> > > > 
> > > > Or is there some reason that this approach would break things?
> > > > 
> > > Hm.. I need to rework it i agree. Reversing the list is a good thing
> > > if we would like to reduce the worst case, i mean latency. Because we
> > > kick users which waited the most. But it is not critical, it is just
> > > a micro optimization and if we have it - fine, if not - no problem.
> > > 
> > > Can we proceed as it is now? I am asking, because i do not find it too
> > > critical. My tests show only 1% difference doing 60K syncing. I need
> > > some time to rework it more carefully.
> > 
> > I am concerned about latencies.  These sorts of things can bit us
> > pretty hard.
> > 
> OK.
> 
> > > I was thinking about read_lock()/write_lock() since we have many readers
> > > and only one writer. But i do not really like it either.
> > 
> > This might be a hint that we should have multiple lists, perhaps one
> > per CPU.  Or lock contention could be used to trigger the transition
> > from a single list to multiple lists. as is done in SRCU and tasks RCU.
> >
> I do not consider to be a sync call as heavily used as other callbacks
> which require several workers to handle, IMHO. From the other hand my
> experiments show that to handle 60K-100K by NOCB gives even worse results.
> 
> > 
> > But I bet that there are several ways to make things work.
> > 
> Right. The main concern with read_lock()/write_lock() is a PREEMPT_RT
> kernels where it is a rt-mutex. It would be good to avoid of using any
> blocking in the gp-kthread since it is a gp driver.

RCU is pretty low-level, so it is OK with a raw spinlock for the list
manipulation.  But only the list manipulation itself.  Perhaps you are
worried about lock contention, but in that case, there is also the issue
of memory contention for the llist code.

							Thanx, Paul

Powered by blists - more mailing lists