| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <ZUIpfzEt_rpTAS8T@pc636>
Date: Wed, 1 Nov 2023 11:33:35 +0100
From: Uladzislau Rezki <urezki@...il.com>
To: Boqun Feng <boqun.feng@...il.com>
Cc: Uladzislau Rezki <urezki@...il.com>,
"Paul E . McKenney" <paulmck@...nel.org>,
RCU <rcu@...r.kernel.org>,
Neeraj upadhyay <Neeraj.Upadhyay@....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 1/3] rcu: Reduce synchronize_rcu() waiting time
On Sun, Oct 29, 2023 at 11:21:11AM -0700, Boqun Feng wrote:
> On Thu, Oct 26, 2023 at 03:09:02PM +0200, Uladzislau Rezki wrote:
> [...]
> > > Late to the party, but I kinda wonder whether we can resolve it by:
> > >
> > > 1) either introduce a separate seglist that only contains callbacks
> > > queued by call_rcu_hurry(), and whenever after an GP and callbacks are
> > > ready, call_rcu_hurry() callbacks will be called first.
> > >
> > > 2) or make call_rcu_hurry() callbacks always inserted at the head of the
> > > NEXT list instead of the tail, e.g. (untested code):
> > >
> > > diff --git a/kernel/rcu/rcu_segcblist.c b/kernel/rcu/rcu_segcblist.c
> > > index f71fac422c8f..89a875f8ecc7 100644
> > > --- a/kernel/rcu/rcu_segcblist.c
> > > +++ b/kernel/rcu/rcu_segcblist.c
> > > @@ -338,13 +338,21 @@ bool rcu_segcblist_nextgp(struct rcu_segcblist *rsclp, unsigned long *lp)
> > > * absolutely not OK for it to ever miss posting a callback.
> > > */
> > > void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
> > > - struct rcu_head *rhp)
> > > + struct rcu_head *rhp,
> > > + bool is_lazy)
> > > {
> > > rcu_segcblist_inc_len(rsclp);
> > > rcu_segcblist_inc_seglen(rsclp, RCU_NEXT_TAIL);
> > > - rhp->next = NULL;
> > > - WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
> > > - WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], &rhp->next);
> > > + /* If hurry and the list is not empty, put it in the front */
> > > + if (!is_lazy && rcu_segcblist_get_seglen(rsclp, RCU_NEXT_TAIL) > 1) {
> > > + // hurry callback, queued at front
> > > + rhp->next = READ_ONCE(*rsclp->tails[RCU_NEXT_READY_TAIL]);
> > > + WRITE_ONCE(*rsclp->tails[RCU_NEXT_READY_TAIL], rhp);
> > > + } else {
> > > + rhp->next = NULL;
> > > + WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
> > > + WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], &rhp->next);
> > > + }
> > > }
> > >
> > > /*
> > > diff --git a/kernel/rcu/rcu_segcblist.h b/kernel/rcu/rcu_segcblist.h
> > > index 4fe877f5f654..459475bb8df9 100644
> > > --- a/kernel/rcu/rcu_segcblist.h
> > > +++ b/kernel/rcu/rcu_segcblist.h
> > > @@ -136,7 +136,8 @@ struct rcu_head *rcu_segcblist_first_cb(struct rcu_segcblist *rsclp);
> > > struct rcu_head *rcu_segcblist_first_pend_cb(struct rcu_segcblist *rsclp);
> > > bool rcu_segcblist_nextgp(struct rcu_segcblist *rsclp, unsigned long *lp);
> > > void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
> > > - struct rcu_head *rhp);
> > > + struct rcu_head *rhp,
> > > + bool is_lazy);
> > > bool rcu_segcblist_entrain(struct rcu_segcblist *rsclp,
> > > struct rcu_head *rhp);
> > > void rcu_segcblist_extract_done_cbs(struct rcu_segcblist *rsclp,
> > > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> > > index 20d7a238d675..53adf5ab9c9f 100644
> > > --- a/kernel/rcu/srcutree.c
> > > +++ b/kernel/rcu/srcutree.c
> > > @@ -1241,7 +1241,7 @@ static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
> > > sdp = raw_cpu_ptr(ssp->sda);
> > > spin_lock_irqsave_sdp_contention(sdp, &flags);
> > > if (rhp)
> > > - rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp);
> > > + rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp, true);
> > > rcu_segcblist_advance(&sdp->srcu_cblist,
> > > rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq));
> > > s = rcu_seq_snap(&ssp->srcu_sup->srcu_gp_seq);
> > > diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
> > > index 8d65f7d576a3..7dec7c68f88f 100644
> > > --- a/kernel/rcu/tasks.h
> > > +++ b/kernel/rcu/tasks.h
> > > @@ -362,7 +362,7 @@ static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
> > > }
> > > if (needwake)
> > > rtpcp->urgent_gp = 3;
> > > - rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
> > > + rcu_segcblist_enqueue(&rtpcp->cblist, rhp, true);
> > > raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
> > > if (unlikely(needadjust)) {
> > > raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index cb1caefa8bd0..e05cbff40dc7 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -2670,7 +2670,7 @@ __call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in)
> > > if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags, lazy))
> > > return; // Enqueued onto ->nocb_bypass, so just leave.
> > > // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
> > > - rcu_segcblist_enqueue(&rdp->cblist, head);
> > > + rcu_segcblist_enqueue(&rdp->cblist, head, lazy_in);
> > > if (__is_kvfree_rcu_offset((unsigned long)func))
> > > trace_rcu_kvfree_callback(rcu_state.name, head,
> > > (unsigned long)func,
> > >
>
> Surprisingly, this survives from a whole rcutorture run ;-)
>
> > > Sure, there may be some corner cases I'm missing, but I think overall
> > > this is better than (sorta) duplicating the logic of seglist (the llist
> > > in sr_normal_state) or the logic of wake_rcu_gp()
> > > (synchronize_rcu_normal).
> > >
> > > Anyway, these are just if-you-have-time-to-try options ;-)
> > >
> > Hm.. You still mix callbacks and there is a dependency in order
> > of execution. The callback process time also might be varied from
> > one callback to another.
> >
> > If you have many *_hurry() calls we end in the same situation. Apart
>
> I plan to resolve that by only puting a call_rcu_hurry(wakeme_after_gp)
> in the front of the list.
>
> > of that we also have !CONFIG_RCU_NOCB_CPU path that is also covered
> > by the patch that is in question.
>
> I don't see why the above approach doesn't work for
> !CONFIG_RCU_NOCB_CPU, but I maybe miss something here.
>
Basically it does not work, because you do not fix the mixing "issue".
I have been working on it and we agreed to separate it. Because it is
just makes sense. The reason and the problem i see, i described in the
commit message of v2.
>
> Do you have a benchmark I can try out to see if my diff can achieve the
> similar result? Thanks!
>
There is no a good benchmark. But you can write it for sure. I tested
three scenarios:
- Run a camera app on our Android devices. Measuring app launch in
milliseconds;
- Doing synchronize_rcu() and kfree(ptr) simultaneously by 10K/etc
workers. It is important test case because we have a fallback to
this scenario for our kvfree_rcu_mightslepp() API.
- I had a look at time delta of loading 100 kernel modules.
That were my main test cases.
--
Uladzislau Rezki
Powered by blists - more mailing lists