Merge tag 'cgroup-for-5.20' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup

Pull cgroup updates from Tejun Heo:
 "Several core optimizations:

   - threadgroup_rwsem write locking is skipped when configuring
     controllers in empty subtrees.

     Combined with CLONE_INTO_CGROUP, this allows the common static
     usage pattern to not grab threadgroup_rwsem at all (glibc still
     doesn't seem ready for CLONE_INTO_CGROUP unfortunately).

   - threadgroup_rwsem used to be put into non-percpu mode by default
     due to latency concerns in specific use cases. There's no reason
     for everyone else to pay for it. Make the behavior optional.

   - psi no longer allocates memory when disabled.

  ... along with some code cleanups"

* tag 'cgroup-for-5.20' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
  cgroup: Skip subtree root in cgroup_update_dfl_csses()
  cgroup: remove "no" prefixed mount options
  cgroup: Make !percpu threadgroup_rwsem operations optional
  cgroup: Add "no" prefixed mount options
  cgroup: Elide write-locking threadgroup_rwsem when updating csses on an empty subtree
  cgroup.c: remove redundant check for mixable cgroup in cgroup_migrate_vet_dst
  cgroup.c: add helper __cset_cgroup_from_root to cleanup duplicated codes
  psi: dont alloc memory for psi by default

Documentation/admin-guide/cgroup-v2.rst

@@ -184,6 +184,14 @@ cgroup v2 currently supports the following mount options.
 	ignored on non-init namespace mounts.  Please refer to the
 	Delegation section for details.
 
+  favordynmods
+        Reduce the latencies of dynamic cgroup modifications such as
+        task migrations and controller on/offs at the cost of making
+        hot path operations such as forks and exits more expensive.
+        The static usage pattern of creating a cgroup, enabling
+        controllers, and then seeding it with CLONE_INTO_CGROUP is
+        not affected by this option.
+
   memory_localevents
         Only populate memory.events with data for the current cgroup,
         and not any subtrees. This is legacy behaviour, the default

include/linux/cgroup-defs.h

@@ -88,20 +88,33 @@ enum {
 	 */
 	CGRP_ROOT_NS_DELEGATE	= (1 << 3),
 
+	/*
+	 * Reduce latencies on dynamic cgroup modifications such as task
+	 * migrations and controller on/offs by disabling percpu operation on
+	 * cgroup_threadgroup_rwsem. This makes hot path operations such as
+	 * forks and exits into the slow path and more expensive.
+	 *
+	 * The static usage pattern of creating a cgroup, enabling controllers,
+	 * and then seeding it with CLONE_INTO_CGROUP doesn't require write
+	 * locking cgroup_threadgroup_rwsem and thus doesn't benefit from
+	 * favordynmod.
+	 */
+	CGRP_ROOT_FAVOR_DYNMODS = (1 << 4),
+
 	/*
 	 * Enable cpuset controller in v1 cgroup to use v2 behavior.
 	 */
-	CGRP_ROOT_CPUSET_V2_MODE = (1 << 4),
+	CGRP_ROOT_CPUSET_V2_MODE = (1 << 16),
 
 	/*
 	 * Enable legacy local memory.events.
 	 */
-	CGRP_ROOT_MEMORY_LOCAL_EVENTS = (1 << 5),
+	CGRP_ROOT_MEMORY_LOCAL_EVENTS = (1 << 17),
 
 	/*
 	 * Enable recursive subtree protection
 	 */
-	CGRP_ROOT_MEMORY_RECURSIVE_PROT = (1 << 6),
+	CGRP_ROOT_MEMORY_RECURSIVE_PROT = (1 << 18),
 };
 
 /* cftype->flags */
@@ -480,7 +493,7 @@ struct cgroup {
 	struct work_struct release_agent_work;
 
 	/* used to track pressure stalls */
-	struct psi_group psi;
+	struct psi_group *psi;
 
 	/* used to store eBPF programs */
 	struct cgroup_bpf bpf;

include/linux/cgroup.h

@@ -674,7 +674,7 @@ static inline void pr_cont_cgroup_path(struct cgroup *cgrp)
 
 static inline struct psi_group *cgroup_psi(struct cgroup *cgrp)
 {
-	return &cgrp->psi;
+	return cgrp->psi;
 }
 
 bool cgroup_psi_enabled(void);

init/Kconfig

@@ -945,6 +945,16 @@ if CGROUPS
 config PAGE_COUNTER
 	bool
 
+config CGROUP_FAVOR_DYNMODS
+        bool "Favor dynamic modification latency reduction by default"
+        help
+          This option enables the "favordynmods" mount option by default
+          which reduces the latencies of dynamic cgroup modifications such
+          as task migrations and controller on/offs at the cost of making
+          hot path operations such as forks and exits more expensive.
+
+          Say N if unsure.
+
 config MEMCG
 	bool "Memory controller"
 	select PAGE_COUNTER

kernel/cgroup/cgroup-internal.h

@@ -233,6 +233,7 @@ void cgroup_kn_unlock(struct kernfs_node *kn);
 int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
 			  struct cgroup_namespace *ns);
 
+void cgroup_favor_dynmods(struct cgroup_root *root, bool favor);
 void cgroup_free_root(struct cgroup_root *root);
 void init_cgroup_root(struct cgroup_fs_context *ctx);
 int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask);

kernel/cgroup/cgroup-v1.c

@@ -875,6 +875,8 @@ static int cgroup1_show_options(struct seq_file *seq, struct kernfs_root *kf_roo
 		seq_puts(seq, ",xattr");
 	if (root->flags & CGRP_ROOT_CPUSET_V2_MODE)
 		seq_puts(seq, ",cpuset_v2_mode");
+	if (root->flags & CGRP_ROOT_FAVOR_DYNMODS)
+		seq_puts(seq, ",favordynmods");
 
 	spin_lock(&release_agent_path_lock);
 	if (strlen(root->release_agent_path))
@@ -898,6 +900,8 @@ enum cgroup1_param {
 	Opt_noprefix,
 	Opt_release_agent,
 	Opt_xattr,
+	Opt_favordynmods,
+	Opt_nofavordynmods,
 };
 
 const struct fs_parameter_spec cgroup1_fs_parameters[] = {
@@ -909,6 +913,8 @@ const struct fs_parameter_spec cgroup1_fs_parameters[] = {
 	fsparam_flag  ("noprefix",	Opt_noprefix),
 	fsparam_string("release_agent",	Opt_release_agent),
 	fsparam_flag  ("xattr",		Opt_xattr),
+	fsparam_flag  ("favordynmods",	Opt_favordynmods),
+	fsparam_flag  ("nofavordynmods", Opt_nofavordynmods),
 	{}
 };
 
@@ -960,6 +966,12 @@ int cgroup1_parse_param(struct fs_context *fc, struct fs_parameter *param)
 	case Opt_xattr:
 		ctx->flags |= CGRP_ROOT_XATTR;
 		break;
+	case Opt_favordynmods:
+		ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
+		break;
+	case Opt_nofavordynmods:
+		ctx->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
+		break;
 	case Opt_release_agent:
 		/* Specifying two release agents is forbidden */
 		if (ctx->release_agent)
@@ -1211,8 +1223,11 @@ static int cgroup1_root_to_use(struct fs_context *fc)
 	init_cgroup_root(ctx);
 
 	ret = cgroup_setup_root(root, ctx->subsys_mask);
-	if (ret)
+	if (!ret)
+		cgroup_favor_dynmods(root, ctx->flags & CGRP_ROOT_FAVOR_DYNMODS);
+	else
 		cgroup_free_root(root);
+
 	return ret;
 }
 

kernel/cgroup/cgroup.c

@@ -279,8 +279,6 @@ bool cgroup_ssid_enabled(int ssid)
  *
  * - When mounting an existing superblock, mount options should match.
  *
- * - Remount is disallowed.
- *
  * - rename(2) is disallowed.
  *
  * - "tasks" is removed.  Everything should be at process granularity.  Use
@@ -1309,6 +1307,20 @@ struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
 	return root_cgrp->root;
 }
 
+void cgroup_favor_dynmods(struct cgroup_root *root, bool favor)
+{
+	bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS;
+
+	/* see the comment above CGRP_ROOT_FAVOR_DYNMODS definition */
+	if (favor && !favoring) {
+		rcu_sync_enter(&cgroup_threadgroup_rwsem.rss);
+		root->flags |= CGRP_ROOT_FAVOR_DYNMODS;
+	} else if (!favor && favoring) {
+		rcu_sync_exit(&cgroup_threadgroup_rwsem.rss);
+		root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
+	}
+}
+
 static int cgroup_init_root_id(struct cgroup_root *root)
 {
 	int id;
@@ -1369,6 +1381,7 @@ static void cgroup_destroy_root(struct cgroup_root *root)
 		cgroup_root_count--;
 	}
 
+	cgroup_favor_dynmods(root, false);
 	cgroup_exit_root_id(root);
 
 	mutex_unlock(&cgroup_mutex);
@@ -1378,6 +1391,31 @@ static void cgroup_destroy_root(struct cgroup_root *root)
 	cgroup_free_root(root);
 }
 
+static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset,
+					    struct cgroup_root *root)
+{
+	struct cgroup *res_cgroup = NULL;
+
+	if (cset == &init_css_set) {
+		res_cgroup = &root->cgrp;
+	} else if (root == &cgrp_dfl_root) {
+		res_cgroup = cset->dfl_cgrp;
+	} else {
+		struct cgrp_cset_link *link;
+
+		list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
+			struct cgroup *c = link->cgrp;
+
+			if (c->root == root) {
+				res_cgroup = c;
+				break;
+			}
+		}
+	}
+
+	return res_cgroup;
+}
+
 /*
  * look up cgroup associated with current task's cgroup namespace on the
  * specified hierarchy
@@ -1393,22 +1431,8 @@ current_cgns_cgroup_from_root(struct cgroup_root *root)
 	rcu_read_lock();
 
 	cset = current->nsproxy->cgroup_ns->root_cset;
-	if (cset == &init_css_set) {
-		res = &root->cgrp;
-	} else if (root == &cgrp_dfl_root) {
-		res = cset->dfl_cgrp;
-	} else {
-		struct cgrp_cset_link *link;
-
-		list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
-			struct cgroup *c = link->cgrp;
+	res = __cset_cgroup_from_root(cset, root);
 
-			if (c->root == root) {
-				res = c;
-				break;
-			}
-		}
-	}
 	rcu_read_unlock();
 
 	BUG_ON(!res);
@@ -1424,22 +1448,7 @@ static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
 	lockdep_assert_held(&cgroup_mutex);
 	lockdep_assert_held(&css_set_lock);
 
-	if (cset == &init_css_set) {
-		res = &root->cgrp;
-	} else if (root == &cgrp_dfl_root) {
-		res = cset->dfl_cgrp;
-	} else {
-		struct cgrp_cset_link *link;
-
-		list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
-			struct cgroup *c = link->cgrp;
-
-			if (c->root == root) {
-				res = c;
-				break;
-			}
-		}
-	}
+	res = __cset_cgroup_from_root(cset, root);
 
 	BUG_ON(!res);
 	return res;
@@ -1866,6 +1875,7 @@ int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
 
 enum cgroup2_param {
 	Opt_nsdelegate,
+	Opt_favordynmods,
 	Opt_memory_localevents,
 	Opt_memory_recursiveprot,
 	nr__cgroup2_params
@@ -1873,6 +1883,7 @@ enum cgroup2_param {
 
 static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
 	fsparam_flag("nsdelegate",		Opt_nsdelegate),
+	fsparam_flag("favordynmods",		Opt_favordynmods),
 	fsparam_flag("memory_localevents",	Opt_memory_localevents),
 	fsparam_flag("memory_recursiveprot",	Opt_memory_recursiveprot),
 	{}
@@ -1892,6 +1903,9 @@ static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param
 	case Opt_nsdelegate:
 		ctx->flags |= CGRP_ROOT_NS_DELEGATE;
 		return 0;
+	case Opt_favordynmods:
+		ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
+		return 0;
 	case Opt_memory_localevents:
 		ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
 		return 0;
@@ -1910,6 +1924,9 @@ static void apply_cgroup_root_flags(unsigned int root_flags)
 		else
 			cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
 
+		cgroup_favor_dynmods(&cgrp_dfl_root,
+				     root_flags & CGRP_ROOT_FAVOR_DYNMODS);
+
 		if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
 			cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
 		else
@@ -1926,6 +1943,8 @@ static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root
 {
 	if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
 		seq_puts(seq, ",nsdelegate");
+	if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS)
+		seq_puts(seq, ",favordynmods");
 	if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
 		seq_puts(seq, ",memory_localevents");
 	if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
@@ -1976,7 +1995,8 @@ void init_cgroup_root(struct cgroup_fs_context *ctx)
 	cgrp->root = root;
 	init_cgroup_housekeeping(cgrp);
 
-	root->flags = ctx->flags;
+	/* DYNMODS must be modified through cgroup_favor_dynmods() */
+	root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS;
 	if (ctx->release_agent)
 		strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
 	if (ctx->name)
@@ -2198,6 +2218,10 @@ static int cgroup_init_fs_context(struct fs_context *fc)
 	put_user_ns(fc->user_ns);
 	fc->user_ns = get_user_ns(ctx->ns->user_ns);
 	fc->global = true;
+
+#ifdef CONFIG_CGROUP_FAVOR_DYNMODS
+	ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
+#endif
 	return 0;
 }
 
@@ -2572,10 +2596,6 @@ int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
 	if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
 		return -EOPNOTSUPP;
 
-	/* mixables don't care */
-	if (cgroup_is_mixable(dst_cgrp))
-		return 0;
-
 	/*
 	 * If @dst_cgrp is already or can become a thread root or is
 	 * threaded, it doesn't matter.
@@ -2949,22 +2969,40 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
 	struct cgroup_subsys_state *d_css;
 	struct cgroup *dsct;
 	struct css_set *src_cset;
+	bool has_tasks;
 	int ret;
 
 	lockdep_assert_held(&cgroup_mutex);
 
-	percpu_down_write(&cgroup_threadgroup_rwsem);
-
 	/* look up all csses currently attached to @cgrp's subtree */
 	spin_lock_irq(&css_set_lock);
 	cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
 		struct cgrp_cset_link *link;
 
+		/*
+		 * As cgroup_update_dfl_csses() is only called by
+		 * cgroup_apply_control(). The csses associated with the
+		 * given cgrp will not be affected by changes made to
+		 * its subtree_control file. We can skip them.
+		 */
+		if (dsct == cgrp)
+			continue;
+
 		list_for_each_entry(link, &dsct->cset_links, cset_link)
 			cgroup_migrate_add_src(link->cset, dsct, &mgctx);
 	}
 	spin_unlock_irq(&css_set_lock);
 
+	/*
+	 * We need to write-lock threadgroup_rwsem while migrating tasks.
+	 * However, if there are no source csets for @cgrp, changing its
+	 * controllers isn't gonna produce any task migrations and the
+	 * write-locking can be skipped safely.
+	 */
+	has_tasks = !list_empty(&mgctx.preloaded_src_csets);
+	if (has_tasks)
+		percpu_down_write(&cgroup_threadgroup_rwsem);
+
 	/* NULL dst indicates self on default hierarchy */
 	ret = cgroup_migrate_prepare_dst(&mgctx);
 	if (ret)
@@ -2984,6 +3022,7 @@ static int cgroup_update_dfl_csses(struct cgroup *cgrp)
 	ret = cgroup_migrate_execute(&mgctx);
 out_finish:
 	cgroup_migrate_finish(&mgctx);
+	if (has_tasks)
 		percpu_up_write(&cgroup_threadgroup_rwsem);
 	return ret;
 }
@@ -3618,21 +3657,21 @@ static int cpu_stat_show(struct seq_file *seq, void *v)
 static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
 {
 	struct cgroup *cgrp = seq_css(seq)->cgroup;
-	struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
+	struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : cgrp->psi;
 
 	return psi_show(seq, psi, PSI_IO);
 }
 static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
 {
 	struct cgroup *cgrp = seq_css(seq)->cgroup;
-	struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
+	struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : cgrp->psi;
 
 	return psi_show(seq, psi, PSI_MEM);
 }
 static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
 {
 	struct cgroup *cgrp = seq_css(seq)->cgroup;
-	struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
+	struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : cgrp->psi;
 
 	return psi_show(seq, psi, PSI_CPU);
 }
@@ -3658,7 +3697,7 @@ static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
 		return -EBUSY;
 	}
 
-	psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
+	psi = cgroup_ino(cgrp) == 1 ? &psi_system : cgrp->psi;
 	new = psi_trigger_create(psi, buf, nbytes, res);
 	if (IS_ERR(new)) {
 		cgroup_put(cgrp);
@@ -5851,12 +5890,6 @@ int __init cgroup_init(void)
 
 	cgroup_rstat_boot();
 
-	/*
-	 * The latency of the synchronize_rcu() is too high for cgroups,
-	 * avoid it at the cost of forcing all readers into the slow path.
-	 */
-	rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
-
 	get_user_ns(init_cgroup_ns.user_ns);
 
 	mutex_lock(&cgroup_mutex);
@@ -6768,6 +6801,7 @@ static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr,
 {
 	return snprintf(buf, PAGE_SIZE,
 			"nsdelegate\n"
+			"favordynmods\n"
 			"memory_localevents\n"
 			"memory_recursiveprot\n");
 }

kernel/sched/psi.c

@@ -957,10 +957,16 @@ int psi_cgroup_alloc(struct cgroup *cgroup)
 	if (static_branch_likely(&psi_disabled))
 		return 0;
 
-	cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu);
-	if (!cgroup->psi.pcpu)
+	cgroup->psi = kmalloc(sizeof(struct psi_group), GFP_KERNEL);
+	if (!cgroup->psi)
 		return -ENOMEM;
-	group_init(&cgroup->psi);
+
+	cgroup->psi->pcpu = alloc_percpu(struct psi_group_cpu);
+	if (!cgroup->psi->pcpu) {
+		kfree(cgroup->psi);
+		return -ENOMEM;
+	}
+	group_init(cgroup->psi);
 	return 0;
 }
 
@@ -969,10 +975,11 @@ void psi_cgroup_free(struct cgroup *cgroup)
 	if (static_branch_likely(&psi_disabled))
 		return;
 
-	cancel_delayed_work_sync(&cgroup->psi.avgs_work);
-	free_percpu(cgroup->psi.pcpu);
+	cancel_delayed_work_sync(&cgroup->psi->avgs_work);
+	free_percpu(cgroup->psi->pcpu);
 	/* All triggers must be removed by now */
-	WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n");
+	WARN_ONCE(cgroup->psi->poll_states, "psi: trigger leak\n");
+	kfree(cgroup->psi);
 }
 
 /**