[PATCH v9 4/4] cgroups: implement the PIDs subsystem

[Aleksa Sarai]

Adds a new single-purpose PIDs subsystem to limit the number of
tasks that can be forked inside a cgroup. Essentially this is an
implementation of RLIMIT_NPROC that applies to a cgroup rather than a
process tree.

However, it should be noted that organisational operations (adding and
removing tasks from a PIDs hierarchy) will *not* be prevented. Rather,
the number of tasks in the hierarchy cannot exceed the limit through
forking. This is due to the fact that, in the unified hierarchy, attach
cannot fail (and it is not possible for a task to overcome its PIDs
cgroup policy limit by attaching to a child cgroup).

PIDs are fundamentally a global resource, and it is possible to reach
PID exhaustion inside a cgroup without hitting any reasonable kmemcg
policy. Once you've hit PID exhaustion, you're only in a marginally
better state than OOM. This subsystem allows PID exhaustion inside a
cgroup to be prevented.

Signed-off-by: Aleksa Sarai <cyphar@xxxxxxxxxx>
---
 include/linux/cgroup_subsys.h |   5 +
 init/Kconfig                  |  16 ++
 kernel/Makefile               |   1 +
 kernel/cgroup_pids.c          | 367 ++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 389 insertions(+)
 create mode 100644 kernel/cgroup_pids.c

diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h
index 1a5a9aa..eb8bfb4 100644
--- a/include/linux/cgroup_subsys.h
+++ b/include/linux/cgroup_subsys.h
@@ -66,6 +66,11 @@ SUBSYS(hugetlb)
  * Subsystems that implement the can_fork() family of callbacks.
  */
 TAG(PREFORK_START)
+
+#if IS_ENABLED(CONFIG_CGROUP_PIDS)
+SUBSYS(pids)
+#endif
+
 TAG(PREFORK_END)
 COUNT(PREFORK)
 
diff --git a/init/Kconfig b/init/Kconfig
index f5dbc6d..1f135b7 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -952,6 +952,22 @@ config CGROUP_FREEZER
 	  Provides a way to freeze and unfreeze all tasks in a
 	  cgroup.
 
+config CGROUP_PIDS
+	bool "PIDs cgroup subsystem"
+	help
+	  Provides enforcement of process number limits in the scope of a
+	  cgroup. Any attempt to fork more processes than is allowed in the
+	  cgroup will fail. PIDs are fundamentally a global resource because it
+	  is fairly trivial to reach PID exhaustion before you reach even a
+	  conservative kmemcg limit. As a result, it is possible to grind a
+	  system to halt without being limited by other cgroup policies. The
+	  PIDs cgroup subsystem is designed to stop this from happening.
+
+	  It should be noted that organisational operations (such as attaching
+	  to a cgroup hierarchy will *not* be blocked by the PIDs subsystem),
+	  since the PIDs limit only affects a process's ability to fork, not to
+	  attach to a cgroup.
+
 config CGROUP_DEVICE
 	bool "Device controller for cgroups"
 	help
diff --git a/kernel/Makefile b/kernel/Makefile
index 1408b33..e823592 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -53,6 +53,7 @@ obj-$(CONFIG_BACKTRACE_SELF_TEST) += backtracetest.o
 obj-$(CONFIG_COMPAT) += compat.o
 obj-$(CONFIG_CGROUPS) += cgroup.o
 obj-$(CONFIG_CGROUP_FREEZER) += cgroup_freezer.o
+obj-$(CONFIG_CGROUP_PIDS) += cgroup_pids.o
 obj-$(CONFIG_CPUSETS) += cpuset.o
 obj-$(CONFIG_UTS_NS) += utsname.o
 obj-$(CONFIG_USER_NS) += user_namespace.o
diff --git a/kernel/cgroup_pids.c b/kernel/cgroup_pids.c
new file mode 100644
index 0000000..5e244b3
--- /dev/null
+++ b/kernel/cgroup_pids.c
@@ -0,0 +1,367 @@
+/*
+ * Process number limiting controller for cgroups.
+ *
+ * Used to allow a cgroup hierarchy to stop any new processes
+ * from fork()ing after a certain limit is reached.
+ *
+ * Since it is trivial to hit the task limit without hitting
+ * any kmemcg limits in place, PIDs are a fundamental resource.
+ * As such, PID exhaustion must be preventable in the scope of
+ * a cgroup hierarchy by allowing resource limiting of the
+ * number of tasks in a cgroup.
+ *
+ * In order to use the `pids` controller, set the maximum number
+ * of tasks in pids.max (this is not available in the root cgroup
+ * for obvious reasons). The number of processes currently
+ * in the cgroup is given by pids.current. Organisational operations
+ * are not blocked by cgroup policies, so it is possible to have
+ * pids.current > pids.max. However, fork()s will still not work.
+ *
+ * To set a cgroup to have no limit, set pids.max to "max". fork()
+ * will return -EBUSY if forking would cause a cgroup policy to be
+ * violated.
+ *
+ * pids.current tracks all child cgroup hierarchies, so
+ * parent/pids.current is a superset of parent/child/pids.current.
+ *
+ * Copyright (C) 2015 Aleksa Sarai <cyphar@xxxxxxxxxx>
+ *
+ */
+
+#include <linux/kernel.h>
+#include <linux/threads.h>
+#include <linux/atomic.h>
+#include <linux/cgroup.h>
+#include <linux/slab.h>
+
+#define PIDS_MAX (PID_MAX_LIMIT + 1ULL)
+#define PIDS_MAX_STR "max"
+
+struct pids_cgroup {
+	struct cgroup_subsys_state	css;
+
+	/*
+	 * Use 64-bit types so that we can safely represent "max" as
+	 * (PID_MAX_LIMIT + 1).
+	 */
+	atomic64_t			counter;
+	int64_t				limit;
+};
+
+static struct pids_cgroup *css_pids(struct cgroup_subsys_state *css)
+{
+	return container_of(css, struct pids_cgroup, css);
+}
+
+static struct pids_cgroup *parent_pids(struct pids_cgroup *pids)
+{
+	return css_pids(pids->css.parent);
+}
+
+static struct cgroup_subsys_state *
+pids_css_alloc(struct cgroup_subsys_state *parent)
+{
+	struct pids_cgroup *pids;
+
+	pids = kzalloc(sizeof(struct pids_cgroup), GFP_KERNEL);
+	if (!pids)
+		return ERR_PTR(-ENOMEM);
+
+	pids->limit = PIDS_MAX;
+	atomic64_set(&pids->counter, 0);
+	return &pids->css;
+}
+
+static void pids_css_free(struct cgroup_subsys_state *css)
+{
+	kfree(css_pids(css));
+}
+
+/**
+ * pids_cancel - uncharge the local pid count
+ * @pids: the pid cgroup state
+ * @num: the number of pids to cancel
+ *
+ * This function will WARN if the pid count goes under 0,
+ * because such a case is a bug in the pids controller proper.
+ */
+static void pids_cancel(struct pids_cgroup *pids, int num)
+{
+	/*
+	 * A negative count (or overflow for that matter) is invalid,
+	 * and indicates a bug in the pids controller proper.
+	 */
+	WARN_ON_ONCE(atomic64_add_negative(-num, &pids->counter));
+}
+
+/**
+ * pids_uncharge - hierarchically uncharge the pid count
+ * @pids: the pid cgroup state
+ * @num: the number of pids to uncharge
+ */
+static void pids_uncharge(struct pids_cgroup *pids, int num)
+{
+	struct pids_cgroup *p;
+
+	for (p = pids; p; p = parent_pids(p))
+		pids_cancel(p, num);
+}
+
+/**
+ * pids_charge - hierarchically charge the pid count
+ * @pids: the pid cgroup state
+ * @num: the number of pids to charge
+ *
+ * This function does *not* follow the pid limit set. It cannot
+ * fail and the new pid count may exceed the limit, because
+ * organisational operations cannot fail in the unified hierarchy.
+ */
+static void pids_charge(struct pids_cgroup *pids, int num)
+{
+	struct pids_cgroup *p;
+
+	for (p = pids; p; p = parent_pids(p))
+		atomic64_add(num, &p->counter);
+}
+
+/**
+ * pids_try_charge - hierarchically try to charge the pid count
+ * @pids: the pid cgroup state
+ * @num: the number of pids to charge
+ *
+ * This function follows the set limit. It will fail if the charge
+ * would cause the new value to exceed the hierarchical limit.
+ * Returns 0 if the charge succeded, otherwise -EAGAIN.
+ */
+static int pids_try_charge(struct pids_cgroup *pids, int num)
+{
+	struct pids_cgroup *p, *q;
+
+	for (p = pids; p; p = parent_pids(p)) {
+		int64_t new = atomic64_add_return(num, &p->counter);
+
+		/*
+		 * Since new is capped to the maximum number of pid_t, if p->limit is
+		 * %PIDS_MAX then we know that this test will never fail.
+		 */
+		if (new > p->limit)
+			goto revert;
+	}
+
+	return 0;
+
+revert:
+	for (q = pids; q != p; q = parent_pids(q))
+		pids_cancel(q, num);
+	pids_cancel(p, num);
+
+	return -EAGAIN;
+}
+
+static int pids_can_attach(struct cgroup_subsys_state *css,
+			   struct cgroup_taskset *tset)
+{
+	struct pids_cgroup *pids = css_pids(css);
+	struct task_struct *task;
+	int64_t num = 0;
+
+	cgroup_taskset_for_each(task, tset)
+		num++;
+
+	/*
+	 * Attaching to a cgroup is allowed to overcome the
+	 * the PID limit, so that organisation operations aren't
+	 * blocked by the `pids` cgroup controller.
+	 */
+	pids_charge(pids, num);
+	return 0;
+}
+
+static void pids_cancel_attach(struct cgroup_subsys_state *css,
+			       struct cgroup_taskset *tset)
+{
+	struct pids_cgroup *pids = css_pids(css);
+	struct task_struct *task;
+	int64_t num = 0;
+
+	cgroup_taskset_for_each(task, tset)
+		num++;
+
+	pids_uncharge(pids, num);
+}
+
+static void pids_detach(struct cgroup_subsys_state *old_css,
+			struct task_struct *task)
+{
+	struct pids_cgroup *old_pids = css_pids(old_css);
+
+	pids_uncharge(old_pids, 1);
+}
+
+static int pids_can_fork(struct task_struct *task, void **private)
+{
+	struct cgroup_subsys_state *css;
+	struct pids_cgroup *pids;
+	int retval;
+
+	/*
+	 * Use the "current" task_css for the pids subsystem as the tentative
+	 * css. It is possible we will charge the wrong hierarchy, in which
+	 * case we will forcefully revert/reapply the charge on the right
+	 * hierarchy after it is committed to the task proper.
+	 */
+	rcu_read_lock();
+	css = task_css(current, pids_cgrp_id);
+	if (!css_tryget_online(css)) {
+		retval = -EBUSY;
+		goto err_rcu_unlock;
+	}
+	rcu_read_unlock();
+	pids = css_pids(css);
+
+	retval = pids_try_charge(pids, 1);
+	if (retval)
+		goto err_css_put;
+
+	*private = css;
+	return 0;
+
+err_rcu_unlock:
+	rcu_read_unlock();
+err_css_put:
+	css_put(css);
+	return retval;
+}
+
+static void pids_cancel_fork(struct task_struct *task, void *private)
+{
+	struct cgroup_subsys_state *css = private;
+	struct pids_cgroup *pids = css_pids(css);
+
+	pids_uncharge(pids, 1);
+	css_put(css);
+}
+
+static void pids_fork(struct task_struct *task, void *private)
+{
+	struct cgroup_subsys_state *css;
+	struct cgroup_subsys_state *old_css = private;
+	struct pids_cgroup *pids;
+	struct pids_cgroup *old_pids = css_pids(old_css);
+
+	/*
+	 * Get the current task css. Since the task has already been exposed to the
+	 * system and had its cg_list updated, we know that we already have an
+	 * implicit reference through task.
+	 */
+	rcu_read_lock();
+	css = task_css(task, pids_cgrp_id);
+	css_get(css);
+	rcu_read_unlock();
+
+	pids = css_pids(css);
+
+	/*
+	 * The association has changed, we have to revert and reapply the
+	 * charge/uncharge on the wrong hierarchy to the current one. Since
+	 * the association can only change due to an organisation event, its
+	 * okay for us to ignore the limit in this case.
+	 */
+	if(pids != old_pids) {
+		pids_uncharge(old_pids, 1);
+		pids_charge(pids, 1);
+	}
+
+	css_put(css);
+	css_put(old_css);
+}
+
+static void pids_exit(struct cgroup_subsys_state *css,
+		      struct cgroup_subsys_state *old_css,
+		      struct task_struct *task)
+{
+	struct pids_cgroup *pids = css_pids(old_css);
+
+	pids_uncharge(pids, 1);
+}
+
+static ssize_t pids_max_write(struct kernfs_open_file *of, char *buf,
+			      size_t nbytes, loff_t off)
+{
+	struct cgroup_subsys_state *css = of_css(of);
+	struct pids_cgroup *pids = css_pids(css);
+	int64_t limit;
+	int err;
+
+	buf = strstrip(buf);
+	if (!strcmp(buf, PIDS_MAX_STR)) {
+		limit = PIDS_MAX;
+		goto set_limit;
+	}
+
+	err = kstrtoll(buf, 0, &limit);
+	if (err)
+		return err;
+
+	/* We use INT_MAX as the maximum value of pid_t. */
+	if (limit < 0 || limit > INT_MAX)
+		return -EINVAL;
+
+set_limit:
+	/*
+	 * Limit updates don't need to be mutex'd, since it isn't
+	 * critical that any racing fork()s follow the new limit.
+	 */
+	pids->limit = limit;
+	return nbytes;
+}
+
+static int pids_max_show(struct seq_file *sf, void *v)
+{
+	struct cgroup_subsys_state *css = seq_css(sf);
+	struct pids_cgroup *pids = css_pids(css);
+	int64_t limit = pids->limit;
+
+	if (limit == PIDS_MAX)
+		seq_printf(sf, "%s\n", PIDS_MAX_STR);
+	else
+		seq_printf(sf, "%lld\n", limit);
+
+	return 0;
+}
+
+static s64 pids_current_read(struct cgroup_subsys_state *css,
+			     struct cftype *cft)
+{
+	struct pids_cgroup *pids = css_pids(css);
+
+	return atomic64_read(&pids->counter);
+}
+
+static struct cftype files[] = {
+	{
+		.name = "max",
+		.write = pids_max_write,
+		.seq_show = pids_max_show,
+		.flags = CFTYPE_NOT_ON_ROOT,
+	},
+	{
+		.name = "current",
+		.read_s64 = pids_current_read,
+	},
+	{ }	/* terminate */
+};
+
+struct cgroup_subsys pids_cgrp_subsys = {
+	.css_alloc	= pids_css_alloc,
+	.css_free	= pids_css_free,
+	.can_attach	= pids_can_attach,
+	.cancel_attach	= pids_cancel_attach,
+	.detach		= pids_detach,
+	.can_fork	= pids_can_fork,
+	.cancel_fork	= pids_cancel_fork,
+	.fork		= pids_fork,
+	.exit		= pids_exit,
+	.legacy_cftypes	= files,
+	.early_init	= 0,
+};
-- 
2.3.5

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/