Message-ID: <Y3OPI/pWZ5jf4X9y@pc636>
Date:   Tue, 15 Nov 2022 14:07:47 +0100
From:   Uladzislau Rezki <urezki@...il.com>
To:     Joel Fernandes <joel@...lfernandes.org>
Cc:     Uladzislau Rezki <urezki@...il.com>, linux-kernel@...r.kernel.org,
        paulmck@...nel.org, rcu@...r.kernel.org
Subject: Re: [PATCH v2] rcu/kfree: Do not request RCU when not needed

On Mon, Nov 14, 2022 at 03:49:16PM -0500, Joel Fernandes wrote:
> On Mon, Nov 14, 2022 at 7:20 AM Uladzislau Rezki <urezki@...il.com> wrote:
> >
> > > On Thu, Nov 10, 2022 at 03:01:30PM +0100, Uladzislau Rezki wrote:
> > > > > Hi,
> > > > >
> > > > > On Thu, Nov 10, 2022 at 8:05 AM Uladzislau Rezki <urezki@...il.com> wrote:
> > > > >
> > > > > > > On ChromeOS, using this with the increased timeout, we see that we
> > > > > > almost always
> > > > > > > never need to initiate a new grace period. Testing also shows this frees
> > > > > > large
> > > > > > > amounts of unreclaimed memory, under intense kfree_rcu() pressure.
> > > > > > >
> > > > > > > Signed-off-by: Joel Fernandes (Google) <joel@...lfernandes.org>
> > > > > > > ---
> > > > > > > v1->v2: Same logic but use polled grace periods instead of sampling
> > > > > > gp_seq.
> > > > > > >
> > > > > > >  kernel/rcu/tree.c | 8 +++++++-
> > > > > > >  1 file changed, 7 insertions(+), 1 deletion(-)
> > > > > > >
> > > > > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > > > > index 591187b6352e..ed41243f7a49 100644
> > > > > > > --- a/kernel/rcu/tree.c
> > > > > > > +++ b/kernel/rcu/tree.c
> > > > > > > @@ -2935,6 +2935,7 @@ struct kfree_rcu_cpu_work {
> > > > > > >
> > > > > > >  /**
> > > > > > >   * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace
> > > > > > period
> > > > > > > + * @gp_snap: The GP snapshot recorded at the last scheduling of monitor
> > > > > > work.
> > > > > > >   * @head: List of kfree_rcu() objects not yet waiting for a grace period
> > > > > > >   * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a
> > > > > > grace period
> > > > > > >   * @krw_arr: Array of batches of kfree_rcu() objects waiting for a
> > > > > > grace period
> > > > > > > @@ -2964,6 +2965,7 @@ struct kfree_rcu_cpu {
> > > > > > >       struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
> > > > > > >       raw_spinlock_t lock;
> > > > > > >       struct delayed_work monitor_work;
> > > > > > > +     unsigned long gp_snap;
> > > > > > >       bool initialized;
> > > > > > >       int count;
> > > > > > >
> > > > > > > @@ -3167,6 +3169,7 @@ schedule_delayed_monitor_work(struct kfree_rcu_cpu
> > > > > > *krcp)
> > > > > > >                       mod_delayed_work(system_wq, &krcp->monitor_work,
> > > > > > delay);
> > > > > > >               return;
> > > > > > >       }
> > > > > > > +     krcp->gp_snap = get_state_synchronize_rcu();
> > > > > > >       queue_delayed_work(system_wq, &krcp->monitor_work, delay);
> > > > > > >  }
> > > > > > >
> > > > > > How do you guarantee a full grace period for objects which proceed
> > > > > > to be placed into an input stream that is not yet detached?
> > > > >
> > > > >
> > > > > Just replying from phone as I’m OOO today.
> > > > >
> > > > > Hmm, so you’re saying that objects can be queued after the delayed work has
> > > > > been queued, but processed when the delayed work is run? I’m looking at
> > > > > this code after few years so I may have missed something.
> > > > >
> > > > > That’s a good point and I think I missed that. I think your version did too
> > > > > but I’ll have to double check.
> > > > >
> > > > > The fix then is to sample the clock for the latest object queued, not for
> > > > > when the delayed work is queued.
> > > > >
> > > > The patch i sent gurantee it. Just in case see v2:
> > >
> > > You are right and thank you! CBs can be queued while the monitor timer is in
> > > progress. So we need to sample unconditionally. I think my approach is still
> > > better since I take advantage of multiple seconds (I update snapshot on every
> > > CB queue monitor and sample in the monitor handler).
> > >
> > > Whereas your patch is snapshotting before queuing the regular work and when
> > > the work is executed (This is a much shorter duration and I bet you would be
> > > blocking in cond_synchronize..() more often).
> > >
> > There is a performance test that measures a taken time and memory
> > footprint, so you can do a quick comparison. A "rcutorture" can be
> > run with various parameters to figure out if a patch that is in question
> > makes any difference.
> 
> Yes sure, I am doing a run now with my patch. However, I have a
> question -- why do you feel blocking in the kworker is not an issue?
> You are taking a snapshot before queuing the normal kwork and then
> reading the snapshot when the normal kwork runs. Considering it is a
> high priority queue, the delay between when you are taking the
> snapshot, and reading it is likely small so there is a bigger chance
> of blocking in cond_synchronize_rcu(). Did I miss something?
> 
We can wait indeed in the reclaim worker. But the worker does not do any
nasty or extra work here. If there is a need we block and wait. After a
grace period, we are awoken and proceed.

Therefore i do not see the reason in handling two cases:

if (gp_done)
    queue_work();
else
    queue_rcu_work();

it is the same if we just queue the work and check on entry. The current
scenario is: queue the work after a grace period. This is the difference.

Right if the reclaimer was a high prio kthread a time would be shorter. 

In your scenario the time seems even shorter(i have not checked) because
you update a snapshot of krcp each time a kvfree_rcu() is invoked. So
basically even though you have objects whose grace period is passed you
do not separate it anyhow. Because you update the:

krcp->gp_snap = get_state_synchronize_rcu();

too often.

--
Uladzislau Rezki

Powered by blists - more mailing lists