[PATCH tip/core/rcu 0/4] kfree_rcu() updates for v5.8

[PATCH tip/core/rcu 0/4] kfree_rcu() updates for v5.8

[Paul E. McKenney]

Hello!

This series updates kfree_rcu() testing and OOM handling.  It also
provides a rcu_gp_might_be_stalled() function that judges whether or
not RCU is being stalled.

1.	Add ability to increase object allocation size, courtesy
	of Joel Fernandes.

2.	rcu/tree: Add a shrinker to prevent OOM due to kfree_rcu()
	batching, courtesy of Joel Fernandes.

3.	rcu/tree: Count number of batched kfree_rcu() locklessly,
	courtesy of Joel Fernandes.

4.	Add rcu_gp_might_be_stalled() to indicate whether or not there
	is currently a good chance that the current RCU grace period
	is being stalled.

							Thanx, Paul

------------------------------------------------------------------------

 include/linux/rcutiny.h |    1 
 include/linux/rcutree.h |    1 
 kernel/rcu/rcuperf.c    |    5 ++-
 kernel/rcu/tree.c       |   70 +++++++++++++++++++++++++++++++++++++++++++-----
 kernel/rcu/tree_stall.h |   40 ++++++++++++++++++++++++---
 5 files changed, 106 insertions(+), 11 deletions(-)

[PATCH tip/core/rcu 1/4] rcuperf: Add ability to increase object allocation size

[paulmck]

From: "Joel Fernandes (Google)" <joel@joelfernandes.org>

This allows us to increase memory pressure dynamically using a new
rcuperf boot command line parameter called 'rcumult'.

Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/rcuperf.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/kernel/rcu/rcuperf.c b/kernel/rcu/rcuperf.c
index a4a8d09..16dd1e6 100644
--- a/kernel/rcu/rcuperf.c
+++ b/kernel/rcu/rcuperf.c
@@ -88,6 +88,7 @@ torture_param(bool, shutdown, RCUPERF_SHUTDOWN,
 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
 torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() perf test?");
+torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
 
 static char *perf_type = "rcu";
 module_param(perf_type, charp, 0444);
@@ -635,7 +636,7 @@ kfree_perf_thread(void *arg)
 		}
 
 		for (i = 0; i < kfree_alloc_num; i++) {
-			alloc_ptr = kmalloc(sizeof(struct kfree_obj), GFP_KERNEL);
+			alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
 			if (!alloc_ptr)
 				return -ENOMEM;
 
@@ -722,6 +723,8 @@ kfree_perf_init(void)
 		schedule_timeout_uninterruptible(1);
 	}
 
+	pr_alert("kfree object size=%lu\n", kfree_mult * sizeof(struct kfree_obj));
+
 	kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
 			       GFP_KERNEL);
 	if (kfree_reader_tasks == NULL) {
-- 
2.9.5

[PATCH tip/core/rcu 2/4] rcu/tree: Add a shrinker to prevent OOM due to kfree_rcu() batching

[paulmck]

From: "Joel Fernandes (Google)" <joel@joelfernandes.org>

To reduce grace periods and improve kfree() performance, we have done
batching recently dramatically bringing down the number of grace periods
while giving us the ability to use kfree_bulk() for efficient kfree'ing.

However, this has increased the likelihood of OOM condition under heavy
kfree_rcu() flood on small memory systems. This patch introduces a
shrinker which starts grace periods right away if the system is under
memory pressure due to existence of objects that have still not started
a grace period.

With this patch, I do not observe an OOM anymore on a system with 512MB
RAM and 8 CPUs, with the following rcuperf options:

rcuperf.kfree_loops=20000 rcuperf.kfree_alloc_num=8000
rcuperf.kfree_rcu_test=1 rcuperf.kfree_mult=2

Otherwise it easily OOMs with the above parameters.

NOTE:
1. On systems with no memory pressure, the patch has no effect as intended.
2. In the future, we can use this same mechanism to prevent grace periods
   from happening even more, by relying on shrinkers carefully.

Cc: urezki@gmail.com
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/tree.c | 60 +++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 60 insertions(+)

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index f13130a..05dcbf8 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2824,6 +2824,8 @@ struct kfree_rcu_cpu {
 	struct delayed_work monitor_work;
 	bool monitor_todo;
 	bool initialized;
+	// Number of objects for which GP not started
+	int count;
 };
 
 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc);
@@ -2937,6 +2939,8 @@ static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
 				krcp->head = NULL;
 			}
 
+			krcp->count = 0;
+
 			/*
 			 * One work is per one batch, so there are two "free channels",
 			 * "bhead_free" and "head_free" the batch can handle. It can be
@@ -3073,6 +3077,8 @@ void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
 		krcp->head = head;
 	}
 
+	krcp->count++;
+
 	// Set timer to drain after KFREE_DRAIN_JIFFIES.
 	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
 	    !krcp->monitor_todo) {
@@ -3087,6 +3093,58 @@ void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
 }
 EXPORT_SYMBOL_GPL(kfree_call_rcu);
 
+static unsigned long
+kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
+{
+	int cpu;
+	unsigned long flags, count = 0;
+
+	/* Snapshot count of all CPUs */
+	for_each_online_cpu(cpu) {
+		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
+
+		spin_lock_irqsave(&krcp->lock, flags);
+		count += krcp->count;
+		spin_unlock_irqrestore(&krcp->lock, flags);
+	}
+
+	return count;
+}
+
+static unsigned long
+kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
+{
+	int cpu, freed = 0;
+	unsigned long flags;
+
+	for_each_online_cpu(cpu) {
+		int count;
+		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
+
+		count = krcp->count;
+		spin_lock_irqsave(&krcp->lock, flags);
+		if (krcp->monitor_todo)
+			kfree_rcu_drain_unlock(krcp, flags);
+		else
+			spin_unlock_irqrestore(&krcp->lock, flags);
+
+		sc->nr_to_scan -= count;
+		freed += count;
+
+		if (sc->nr_to_scan <= 0)
+			break;
+	}
+
+	return freed;
+}
+
+static struct shrinker kfree_rcu_shrinker = {
+	.count_objects = kfree_rcu_shrink_count,
+	.scan_objects = kfree_rcu_shrink_scan,
+	.batch = 0,
+	.seeks = DEFAULT_SEEKS,
+};
+
 void __init kfree_rcu_scheduler_running(void)
 {
 	int cpu;
@@ -4007,6 +4065,8 @@ static void __init kfree_rcu_batch_init(void)
 		INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
 		krcp->initialized = true;
 	}
+	if (register_shrinker(&kfree_rcu_shrinker))
+		pr_err("Failed to register kfree_rcu() shrinker!\n");
 }
 
 void __init rcu_init(void)
-- 
2.9.5

[PATCH tip/core/rcu 3/4] rcu/tree: Count number of batched kfree_rcu() locklessly

[paulmck]

From: "Joel Fernandes (Google)" <joel@joelfernandes.org>

We can relax the correctness of counting of number of queued objects in
favor of not hurting performance, by locklessly sampling per-cpu
counters. This should be Ok since under high memory pressure, it should not
matter if we are off by a few objects while counting. The shrinker will
still do the reclaim.

Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
[ paulmck: Remove unused "flags" variable. ]
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 kernel/rcu/tree.c | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 05dcbf8..aef587e 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2939,7 +2939,7 @@ static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
 				krcp->head = NULL;
 			}
 
-			krcp->count = 0;
+			WRITE_ONCE(krcp->count, 0);
 
 			/*
 			 * One work is per one batch, so there are two "free channels",
@@ -3077,7 +3077,7 @@ void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
 		krcp->head = head;
 	}
 
-	krcp->count++;
+	WRITE_ONCE(krcp->count, krcp->count + 1);
 
 	// Set timer to drain after KFREE_DRAIN_JIFFIES.
 	if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
@@ -3097,15 +3097,13 @@ static unsigned long
 kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
 {
 	int cpu;
-	unsigned long flags, count = 0;
+	unsigned long count = 0;
 
 	/* Snapshot count of all CPUs */
 	for_each_online_cpu(cpu) {
 		struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
 
-		spin_lock_irqsave(&krcp->lock, flags);
-		count += krcp->count;
-		spin_unlock_irqrestore(&krcp->lock, flags);
+		count += READ_ONCE(krcp->count);
 	}
 
 	return count;
-- 
2.9.5

[PATCH tip/core/rcu 4/4] rcu: Add rcu_gp_might_be_stalled()

[paulmck]

From: "Paul E. McKenney" <paulmck@kernel.org>

This commit adds rcu_gp_might_be_stalled(), which returns true if there
is some reason to believe that the RCU grace period is stalled.  The use
case is where an RCU free-memory path needs to allocate memory in order
to free it, a situation that should be avoided where possible.

But where it is necessary, there is always the alternative of using
synchronize_rcu() to wait for a grace period in order to avoid the
allocation.  And if the grace period is stalled, allocating memory to
asynchronously wait for it is a bad idea of epic proportions: Far better
to let others use the memory, because these others might actually be
able to free that memory before the grace period ends.

Thus, rcu_gp_might_be_stalled() can be used to help decide whether
allocating memory on an RCU free path is a semi-reasonable course
of action.

Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Uladzislau Rezki <urezki@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
---
 include/linux/rcutiny.h |  1 +
 include/linux/rcutree.h |  1 +
 kernel/rcu/tree_stall.h | 40 ++++++++++++++++++++++++++++++++++++----
 3 files changed, 38 insertions(+), 4 deletions(-)

diff --git a/include/linux/rcutiny.h b/include/linux/rcutiny.h
index 045c28b..dbf5ac4 100644
--- a/include/linux/rcutiny.h
+++ b/include/linux/rcutiny.h
@@ -87,6 +87,7 @@ static inline bool rcu_inkernel_boot_has_ended(void) { return true; }
 static inline bool rcu_is_watching(void) { return true; }
 static inline void rcu_momentary_dyntick_idle(void) { }
 static inline void kfree_rcu_scheduler_running(void) { }
+static inline bool rcu_gp_might_be_stalled(void) { return false; }
 
 /* Avoid RCU read-side critical sections leaking across. */
 static inline void rcu_all_qs(void) { barrier(); }
diff --git a/include/linux/rcutree.h b/include/linux/rcutree.h
index 45f3f66..fbc2627 100644
--- a/include/linux/rcutree.h
+++ b/include/linux/rcutree.h
@@ -39,6 +39,7 @@ void rcu_barrier(void);
 bool rcu_eqs_special_set(int cpu);
 void rcu_momentary_dyntick_idle(void);
 void kfree_rcu_scheduler_running(void);
+bool rcu_gp_might_be_stalled(void);
 unsigned long get_state_synchronize_rcu(void);
 void cond_synchronize_rcu(unsigned long oldstate);
 
diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
index 119ed6a..4dede00 100644
--- a/kernel/rcu/tree_stall.h
+++ b/kernel/rcu/tree_stall.h
@@ -15,10 +15,12 @@
 int sysctl_panic_on_rcu_stall __read_mostly;
 
 #ifdef CONFIG_PROVE_RCU
-#define RCU_STALL_DELAY_DELTA	       (5 * HZ)
+#define RCU_STALL_DELAY_DELTA		(5 * HZ)
 #else
-#define RCU_STALL_DELAY_DELTA	       0
+#define RCU_STALL_DELAY_DELTA		0
 #endif
+#define RCU_STALL_MIGHT_DIV		8
+#define RCU_STALL_MIGHT_MIN		(2 * HZ)
 
 /* Limit-check stall timeouts specified at boottime and runtime. */
 int rcu_jiffies_till_stall_check(void)
@@ -40,6 +42,36 @@ int rcu_jiffies_till_stall_check(void)
 }
 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
 
+/**
+ * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
+ *
+ * Returns @true if the current grace period is sufficiently old that
+ * it is reasonable to assume that it might be stalled.  This can be
+ * useful when deciding whether to allocate memory to enable RCU-mediated
+ * freeing on the one hand or just invoking synchronize_rcu() on the other.
+ * The latter is preferable when the grace period is stalled.
+ *
+ * Note that sampling of the .gp_start and .gp_seq fields must be done
+ * carefully to avoid false positives at the beginnings and ends of
+ * grace periods.
+ */
+bool rcu_gp_might_be_stalled(void)
+{
+	unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
+	unsigned long j = jiffies;
+
+	if (d < RCU_STALL_MIGHT_MIN)
+		d = RCU_STALL_MIGHT_MIN;
+	smp_mb(); // jiffies before .gp_seq to avoid false positives.
+	if (!rcu_gp_in_progress())
+		return false;
+	// Long delays at this point avoids false positive, but a delay
+	// of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
+	smp_mb(); // .gp_seq before second .gp_start
+	// And ditto here.
+	return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
+}
+
 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
 void rcu_sysrq_start(void)
 {
@@ -104,8 +136,8 @@ static void record_gp_stall_check_time(void)
 
 	WRITE_ONCE(rcu_state.gp_start, j);
 	j1 = rcu_jiffies_till_stall_check();
-	/* Record ->gp_start before ->jiffies_stall. */
-	smp_store_release(&rcu_state.jiffies_stall, j + j1); /* ^^^ */
+	smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
+	WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
 	rcu_state.jiffies_resched = j + j1 / 2;
 	rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
 }
-- 
2.9.5