Message-ID: <Y20EOinwcLSZHmXg@pc638.lan>
Date:   Thu, 10 Nov 2022 15:01:30 +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

> 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:

>From 7ff4038d5dac8d2044b240adeee630ad7944ab8d Mon Sep 17 00:00:00 2001
From: "Uladzislau Rezki (Sony)" <urezki@...il.com>
Date: Wed, 2 Nov 2022 19:26:27 +0100
Subject: [PATCH] rcu: kvfree_rcu: Reduce a memory footptint by using polling
 APIs

Total time taken by all kfree'ers: 6564718459 ns, loops: 10000, batches: 1110, memory footprint: 5057MB
Total time taken by all kfree'ers: 8431051895 ns, loops: 10000, batches: 1109, memory footprint: 2749MB
Total time taken by all kfree'ers: 9477830789 ns, loops: 10000, batches: 1158, memory footprint: 2934MB
Total time taken by all kfree'ers: 9950211144 ns, loops: 10000, batches: 981, memory footprint: 2704MB

with a patch:

Total time taken by all kfree'ers: 7712110118 ns, loops: 10000, batches: 1660, memory footprint: 91MB
Total time taken by all kfree'ers: 7002403664 ns, loops: 10000, batches: 1482, memory footprint: 86MB
Total time taken by all kfree'ers: 7842282319 ns, loops: 10000, batches: 1738, memory footprint: 86MB
Total time taken by all kfree'ers: 7230161977 ns, loops: 10000, batches: 1542, memory footprint: 72MB

Tested with NOCB option, all offloading CPUs:

kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \
  --kconfig CONFIG_NR_CPUS=64 \
  --kconfig CONFIG_RCU_NOCB_CPU=y \
  --kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \
  --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \
  rcuscale.holdoff=20 rcuscale.kfree_loops=10000 torture.disable_onoff_at_boot" --trust-make

According to data there is a big gain in memory footprint with a patch.
It is because of call_rcu() and call_rcu_flush() take more effort and
time to queue a callback and then wait for a gp.

With polling API:
  a) we do not need to queue any callback;
  b) we might not even need wait for a GP completion.

Signed-off-by: Uladzislau Rezki (Sony) <urezki@...il.com>
---
 kernel/rcu/tree.c | 32 ++++++++++++++++++++++++++------
 1 file changed, 26 insertions(+), 6 deletions(-)

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 76973d716921..2be4f80867f2 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2919,18 +2919,20 @@ struct kvfree_rcu_bulk_data {
 	((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
 
 /**
+ * @rcu_work: A work to reclaim a memory after a grace period
  * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
- * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
  * @head_free: List of kfree_rcu() objects waiting for a grace period
  * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
  * @krcp: Pointer to @kfree_rcu_cpu structure
+ * @gp_snap: A snapshot of current grace period
  */
 
 struct kfree_rcu_cpu_work {
-	struct rcu_work rcu_work;
+	struct work_struct rcu_work;
 	struct rcu_head *head_free;
 	struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
 	struct kfree_rcu_cpu *krcp;
+	unsigned long gp_snap;
 };
 
 /**
@@ -3064,10 +3066,11 @@ static void kfree_rcu_work(struct work_struct *work)
 	struct rcu_head *head, *next;
 	struct kfree_rcu_cpu *krcp;
 	struct kfree_rcu_cpu_work *krwp;
+	unsigned long this_krwp_gp_snap;
 	int i, j;
 
-	krwp = container_of(to_rcu_work(work),
-			    struct kfree_rcu_cpu_work, rcu_work);
+	krwp = container_of(work,
+		struct kfree_rcu_cpu_work, rcu_work);
 	krcp = krwp->krcp;
 
 	raw_spin_lock_irqsave(&krcp->lock, flags);
@@ -3080,8 +3083,15 @@ static void kfree_rcu_work(struct work_struct *work)
 	// Channel 3.
 	head = krwp->head_free;
 	krwp->head_free = NULL;
+
+	// Get the latest saved state of grace period
+	// associated with this "krwp" and objects there.
+	this_krwp_gp_snap = krwp->gp_snap;
 	raw_spin_unlock_irqrestore(&krcp->lock, flags);
 
+	// Check the state.
+	cond_synchronize_rcu(this_krwp_gp_snap);
+
 	// Handle the first two channels.
 	for (i = 0; i < FREE_N_CHANNELS; i++) {
 		for (; bkvhead[i]; bkvhead[i] = bnext) {
@@ -3212,12 +3222,22 @@ static void kfree_rcu_monitor(struct work_struct *work)
 
 			WRITE_ONCE(krcp->count, 0);
 
+			// 1. Take a snapshot for this krwp. Please note no
+			// more any objects can be added to channels which
+			// have already been attached.
+			//
+			// 2. Since a "krwp" has several free channels a GP
+			// status of latest attached one is recorded, i.e.
+			// it allows us to maintain a GP status for last in
+			// objects among all channels.
+			krwp->gp_snap = get_state_synchronize_rcu();
+
 			// One work is per one batch, so there are three
 			// "free channels", the batch can handle. It can
 			// be that the work is in the pending state when
 			// channels have been detached following by each
 			// other.
-			queue_rcu_work(system_wq, &krwp->rcu_work);
+			queue_work(system_wq, &krwp->rcu_work);
 		}
 	}
 
@@ -4808,7 +4828,7 @@ static void __init kfree_rcu_batch_init(void)
 		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
 
 		for (i = 0; i < KFREE_N_BATCHES; i++) {
-			INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
+			INIT_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
 			krcp->krw_arr[i].krcp = krcp;
 		}
 
> > > @@ -3217,7 +3220,10 @@ static void kfree_rcu_monitor(struct work_struct
> > *work)
> > >                       // be that the work is in the pending state when
> > >                       // channels have been detached following by each
> > >                       // other.
> > > -                     queue_rcu_work(system_wq, &krwp->rcu_work);
> > > +                     if (poll_state_synchronize_rcu(krcp->gp_snap))
> > > +                             queue_work(system_wq, &krwp->rcu_work.work
> > );
> > > +                     else
> > > +                             queue_rcu_work(system_wq, &krwp->rcu_work);
> > >               }
> > >
> > Why do you want to queue a work over RCU-core?
> >
> > 1.
> > call_rcu()
> >    -> queue_work();
> >       -> do reclaim
> >
> > if it can be improved and simplified as:
> >
> > 2.
> > queue_work();
> >     -> cond_synchronize_rcu(), do reclaim
> 
> 
> The second one blocks, the first doesn’t. So I’m not sure how that’s an
> improvement? I think we don’t want to block in kworker due to possible
> scheduling issues and hurting other kwork during critical operations, if
> possible.
> 
Does the first wait for a full grace period so our kworker is not even run?

--
Uladzislau Rezki

Powered by blists - more mailing lists