linux-kernel - Re: [PATCH v3] rcu: Remove impossible wakeup rcu GP kthread action from rcu_report_qs

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

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <Y873895orjOvPNg9@lothringen>
Date:   Mon, 23 Jan 2023 22:11:15 +0100
From:   Frederic Weisbecker <frederic@...nel.org>
To:     Joel Fernandes <joel@...lfernandes.org>
Cc:     "Zhang, Qiang1" <qiang1.zhang@...el.com>,
        "Paul E. McKenney" <paulmck@...nel.org>, quic_neeraju@...cinc.com,
        rcu@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v3] rcu: Remove impossible wakeup rcu GP kthread action
 from rcu_report_qs_rdp()

On Mon, Jan 23, 2023 at 03:54:07PM -0500, Joel Fernandes wrote:
> > On Jan 23, 2023, at 11:27 AM, Frederic Weisbecker <frederic@...nel.org> wrote:
> > 
> > On Mon, Jan 23, 2023 at 10:22:19AM -0500, Joel Fernandes wrote:
> >>> What am I missing?
> >> 
> >> That the acceleration is also done by __note_gp_changes() once the
> >> grace period ends anyway, so if any acceleration was missed as you
> >> say, it will be done anyway.
> >> 
> >> Also it is done by scheduler tick raising softirq:
> >> 
> >> rcu_pending() does this:
> >>        /* Has RCU gone idle with this CPU needing another grace period? */
> >>        if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
> >>            !rcu_rdp_is_offloaded(rdp) &&
> >>            !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
> >>                return 1;
> >> 
> >> and rcu_core():
> >>        /* No grace period and unregistered callbacks? */
> >>        if (!rcu_gp_in_progress() &&
> >>            rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
> >>                rcu_nocb_lock_irqsave(rdp, flags);
> >>                if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
> >>                        rcu_accelerate_cbs_unlocked(rnp, rdp);
> >>                rcu_nocb_unlock_irqrestore(rdp, flags);
> >>        }
> >> 
> >> So, I am not sure if you need needacc at all. Those CBs that have not
> >> been assigned grace period numbers will be taken care off :)
> > 
> > But that's only when there is no grace period pending, so it can't happen while
> > we report a QS.
> > 
> > OTOH without the needacc, those callbacks waiting to be accelerated would be
> > eventually processed but only on the next tick following the end of a grace
> > period...if none has started since then. So if someone else starts a new GP
> > before the current CPU, we must wait another GP, etc...
> > 
> > That's potentially dangerous.
> 
> Waiting for just one more GP cannot be dangerous IMO. Anyway there is no
> guarantee that callback will run immediately at end of GP, there may be one or
> more GPs before callback can run, if I remember correctly. That is by
> design.. but please correct me if my understanding is different from yours.

It's not bound to just one GP. If you have N CPUs flooding callbacks for a
long while, your CPU has 1/N chances to be the one starting the next GP on
each turn. Relying on the acceleration to be performed only when no GP is
running may theoretically starve your callbacks forever.

> > And unfortunately we can't do the acceleration from __note_gp_changes() due
> > to lock ordering restrictions: nocb_lock -> rnp_lock
> > 
> 
> Ah. This part I am not sure. Appreciate if point me to any old archive links or documentation detailing that, if possible… 

It's not documented but the code in nocb_gp_wait() or nocb_cb_wait() has that
locking order for example.

Excerpt:

	rcu_nocb_lock_irqsave(rdp, flags);
	if (rcu_segcblist_nextgp(cblist, &cur_gp_seq) &&
	    rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
	    raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
		needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
	}

Thanks.