[PATCH 2/2] cgroup/cpuset: Make percpu cpuset_rwsem operation depending on DYNMODS state

[Waiman Long]

With commit 6a010a49b63a ("cgroup: Make !percpu threadgroup_rwsem
operations optional"), users can determine if they favor optimizing
for efficiently moving processes between cgroups frequently or for
a more static usage pattern where moving processes among cgroups are
relatively rare.

The percpu cpuset_rwsem is in the same boat as percpu_threadgroup_rwsem
since moving processes among cpusets will have the same latency impact
depending on whether percpu operation in cpuset_rwsem is disabled or not.

Ideally cpuset_bind() is the best place to check if the
cpuset_rwsem should have its reader fast path disabled like
percpu_threadgroup_rwsem so that it gets to be re-evaluated every
time the cpuset is rebound. Unfortunately, cgroup_favor_dynmods()
that sets the CGRP_ROOT_FAVOR_DYNMODS flag is called after the bind()
method call. Instead, the newly added cpuset_check_dynmods() function
is called at the first cpuset_css_online() call after a cpuset_bind()
call when the first child cpuset is created.

Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
 kernel/cgroup/cpuset.c | 36 ++++++++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)

diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index 800c65de5daa..daf8ca948176 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -255,6 +255,7 @@ typedef enum {
 	CS_SCHED_LOAD_BALANCE,
 	CS_SPREAD_PAGE,
 	CS_SPREAD_SLAB,
+	CS_FAVOR_DYNMODS,	/* top_cpuset only */
 } cpuset_flagbits_t;
 
 /* convenient tests for these bits */
@@ -3049,6 +3050,27 @@ static struct cftype dfl_files[] = {
 	{ }	/* terminate */
 };
 
+static bool dynmods_checked __read_mostly;
+static void cpuset_check_dynmods(struct cgroup_root *root)
+{
+	bool favor_dynmods;
+
+	lockdep_assert_held(&cgroup_mutex);
+	percpu_rwsem_assert_held(&cpuset_rwsem);
+
+	/*
+	 * Check the CGRP_ROOT_FAVOR_DYNMODS of the cgroup root to find out
+	 * if we need to enable or disable reader fast path of cpuset_rwsem.
+	 */
+	favor_dynmods = test_bit(CS_FAVOR_DYNMODS, &top_cpuset.flags);
+	if (favor_dynmods && !(root->flags & CGRP_ROOT_FAVOR_DYNMODS)) {
+		rcu_sync_exit(&cpuset_rwsem.rss);
+		clear_bit(CS_FAVOR_DYNMODS, &top_cpuset.flags);
+	} else if (!favor_dynmods && (root->flags & CGRP_ROOT_FAVOR_DYNMODS)) {
+		rcu_sync_enter(&cpuset_rwsem.rss);
+		set_bit(CS_FAVOR_DYNMODS, &top_cpuset.flags);
+	}
+}
 
 /*
  *	cpuset_css_alloc - allocate a cpuset css
@@ -3099,6 +3121,14 @@ static int cpuset_css_online(struct cgroup_subsys_state *css)
 	cpus_read_lock();
 	percpu_down_write(&cpuset_rwsem);
 
+	/*
+	 * Check dynmod state on the first css_online() call.
+	 */
+	if (unlikely(!dynmods_checked)) {
+		cpuset_check_dynmods(cpuset_cgrp_subsys.root);
+		dynmods_checked = true;
+	}
+
 	set_bit(CS_ONLINE, &cs->flags);
 	if (is_spread_page(parent))
 		set_bit(CS_SPREAD_PAGE, &cs->flags);
@@ -3201,6 +3231,12 @@ static void cpuset_css_free(struct cgroup_subsys_state *css)
 
 static void cpuset_bind(struct cgroup_subsys_state *root_css)
 {
+	/*
+	 * Reset dynmods_checked to be evaluated again in the next
+	 * cpuset_css_online()
+	 */
+	dynmods_checked = false;
+
 	percpu_down_write(&cpuset_rwsem);
 	spin_lock_irq(&callback_lock);
 
-- 
2.31.1