[PATCH v14 0/4] cgroup: add PIDs subsystem

[PATCH v14 0/4] cgroup: add PIDs subsystem

[Aleksa Sarai]

This is another small update to v13 of the pids patchset[1] (half of
which has already been merged to Tejun's tree). The main changes are:

* Fix uninitialised iterator in for_each_subsys_which when
  CGROUP_SUBSYS_COUNT = 0.

* A few small (mainly stylistic) fixes to kernel/cgroup_pids.c and
  kernel/cgroup.c.

[1]: https://lkml.org/lkml/2015/6/5/857

Aleksa Sarai (3):
  cgroup: fix uninitialised iterator in for_each_subsys_which
  cgroup: allow a cgroup subsystem to reject a fork
  cgroup: implement the PIDs subsystem

Tejun Heo (1):
  cgroup, block: implement task_get_css()

 CREDITS                       |   5 +
 include/linux/cgroup-defs.h   |  10 +-
 include/linux/cgroup.h        |  40 ++++-
 include/linux/cgroup_subsys.h |  28 ++++
 init/Kconfig                  |  16 ++
 kernel/Makefile               |   1 +
 kernel/cgroup.c               |  75 ++++++++-
 kernel/cgroup_freezer.c       |   2 +-
 kernel/cgroup_pids.c          | 366 ++++++++++++++++++++++++++++++++++++++++++
 kernel/fork.c                 |  17 +-
 kernel/sched/core.c           |   2 +-
 11 files changed, 552 insertions(+), 10 deletions(-)
 create mode 100644 kernel/cgroup_pids.c

-- 
2.4.2

[PATCH v14 1/4] cgroup: fix uninitialised iterator in for_each_subsys_which

[Aleksa Sarai]

Fix the fact that @ssid is uninitialised in the case where
CGROUP_SUBSYS_COUNT = 0 by setting ssid to 0.

Fixes: cb4a31675270 ("cgroup: use bitmask to filter for_each_subsys")
Signed-off-by: Aleksa Sarai <cyphar-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>
---
 kernel/cgroup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 6d6d3c7..12b580f 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -424,7 +424,7 @@ static int notify_on_release(const struct cgroup *cgrp)
  */
 #define for_each_subsys_which(ss, ssid, ss_maskp)			\
 	if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */	\
-		;							\
+		(ssid) = 0;						\
 	else								\
 		for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT)	\
 			if (((ss) = cgroup_subsys[ssid]) && false)	\
-- 
2.4.2

Re: [PATCH v14 1/4] cgroup: fix uninitialised iterator in for_each_subsys_which

[Tejun Heo]

On Tue, Jun 09, 2015 at 09:32:07PM +1000, Aleksa Sarai wrote:
> Fix the fact that @ssid is uninitialised in the case where
> CGROUP_SUBSYS_COUNT = 0 by setting ssid to 0.
> 
> Fixes: cb4a31675270 ("cgroup: use bitmask to filter for_each_subsys")
> Signed-off-by: Aleksa Sarai <cyphar-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>

Applied to cgroup/for-4.2.

Thanks.

-- 
tejun

[PATCH v14 2/4] cgroup, block: implement task_get_css()

[Aleksa Sarai]

From: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>

Implement task_get_css() which find and pins the css for the specified
task and subsys.  As a task is always associated with an online css
for every subsystem except while the css_set update is propagating,
task_get_css() retries till css_tryget_online() succeeds.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Cc: Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>
---
 include/linux/cgroup.h | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
index 82319fb..a593e29 100644
--- a/include/linux/cgroup.h
+++ b/include/linux/cgroup.h
@@ -354,6 +354,31 @@ static inline struct cgroup_subsys_state *task_css(struct task_struct *task,
 }
 
 /**
+ * task_get_css - find and get the css for (task, subsys)
+ * @task: the target task
+ * @subsys_id: the target subsystem ID
+ *
+ * Find the css for the (@task, @subsys_id) combination, increment a
+ * reference on and return it.  This function is guaranteed to return a
+ * valid css.
+ */
+static inline struct cgroup_subsys_state *
+task_get_css(struct task_struct *task, int subsys_id)
+{
+	struct cgroup_subsys_state *css;
+
+	rcu_read_lock();
+	while (true) {
+		css = task_css(task, subsys_id);
+		if (likely(css_tryget_online(css)))
+			break;
+		cpu_relax();
+	}
+	rcu_read_unlock();
+	return css;
+}
+
+/**
  * task_css_is_root - test whether a task belongs to the root css
  * @task: the target task
  * @subsys_id: the target subsystem ID
-- 
2.4.2

[PATCH v14 3/4] cgroup: allow a cgroup subsystem to reject a fork

[Aleksa Sarai]

Add a new cgroup subsystem callback can_fork that conditionally
states whether or not the fork is accepted or rejected by a cgroup
policy. In addition, add a cancel_fork callback so that if an error
occurs later in the forking process, any state modified by can_fork can
be reverted.

Allow for a private opaque pointer to be passed from cgroup_can_fork to
cgroup_post_fork, allowing for the fork state to be stored by each
subsystem separately.

Also add a tagging system for cgroup_subsys.h to allow for CGROUP_<TAG>
enumerations to be be defined and used. In addition, explicitly add a
CGROUP_CANFORK_COUNT macro to make arrays easier to define.

This is in preparation for implementing the pids cgroup subsystem.

Signed-off-by: Aleksa Sarai <cyphar-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>
---
 include/linux/cgroup-defs.h   | 10 +++++-
 include/linux/cgroup.h        | 15 +++++++--
 include/linux/cgroup_subsys.h | 23 ++++++++++++++
 kernel/cgroup.c               | 73 +++++++++++++++++++++++++++++++++++++++++--
 kernel/cgroup_freezer.c       |  2 +-
 kernel/fork.c                 | 17 ++++++++--
 kernel/sched/core.c           |  2 +-
 7 files changed, 133 insertions(+), 9 deletions(-)

diff --git a/include/linux/cgroup-defs.h b/include/linux/cgroup-defs.h
index c5588c4..2009ceb 100644
--- a/include/linux/cgroup-defs.h
+++ b/include/linux/cgroup-defs.h
@@ -34,12 +34,17 @@ struct seq_file;
 
 /* define the enumeration of all cgroup subsystems */
 #define SUBSYS(_x) _x ## _cgrp_id,
+#define SUBSYS_TAG(_t) CGROUP_ ## _t, \
+	__unused_tag_ ## _t = CGROUP_ ## _t - 1,
 enum cgroup_subsys_id {
 #include <linux/cgroup_subsys.h>
 	CGROUP_SUBSYS_COUNT,
 };
+#undef SUBSYS_TAG
 #undef SUBSYS
 
+#define CGROUP_CANFORK_COUNT (CGROUP_CANFORK_END - CGROUP_CANFORK_START)
+
 /* bits in struct cgroup_subsys_state flags field */
 enum {
 	CSS_NO_REF	= (1 << 0), /* no reference counting for this css */
@@ -405,7 +410,9 @@ struct cgroup_subsys {
 			      struct cgroup_taskset *tset);
 	void (*attach)(struct cgroup_subsys_state *css,
 		       struct cgroup_taskset *tset);
-	void (*fork)(struct task_struct *task);
+	int (*can_fork)(struct task_struct *task, void **priv_p);
+	void (*cancel_fork)(struct task_struct *task, void *priv);
+	void (*fork)(struct task_struct *task, void *priv);
 	void (*exit)(struct cgroup_subsys_state *css,
 		     struct cgroup_subsys_state *old_css,
 		     struct task_struct *task);
@@ -490,6 +497,7 @@ static inline void cgroup_threadgroup_change_end(struct task_struct *tsk)
 
 #else	/* CONFIG_CGROUPS */
 
+#define CGROUP_CANFORK_COUNT 0
 #define CGROUP_SUBSYS_COUNT 0
 
 static inline void cgroup_threadgroup_change_begin(struct task_struct *tsk) {}
diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
index a593e29..a71fe2a 100644
--- a/include/linux/cgroup.h
+++ b/include/linux/cgroup.h
@@ -62,7 +62,12 @@ int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
 		     struct pid *pid, struct task_struct *tsk);
 
 void cgroup_fork(struct task_struct *p);
-void cgroup_post_fork(struct task_struct *p);
+extern int cgroup_can_fork(struct task_struct *p,
+			   void *ss_priv[CGROUP_CANFORK_COUNT]);
+extern void cgroup_cancel_fork(struct task_struct *p,
+			       void *ss_priv[CGROUP_CANFORK_COUNT]);
+extern void cgroup_post_fork(struct task_struct *p,
+			     void *old_ss_priv[CGROUP_CANFORK_COUNT]);
 void cgroup_exit(struct task_struct *p);
 
 int cgroup_init_early(void);
@@ -524,7 +529,13 @@ static inline int cgroupstats_build(struct cgroupstats *stats,
 				    struct dentry *dentry) { return -EINVAL; }
 
 static inline void cgroup_fork(struct task_struct *p) {}
-static inline void cgroup_post_fork(struct task_struct *p) {}
+static inline int cgroup_can_fork(struct task_struct *p,
+				  void *ss_priv[CGROUP_CANFORK_COUNT])
+{ return 0; }
+static inline void cgroup_cancel_fork(struct task_struct *p,
+				      void *ss_priv[CGROUP_CANFORK_COUNT]) {}
+static inline void cgroup_post_fork(struct task_struct *p,
+				    void *ss_priv[CGROUP_CANFORK_COUNT]) {}
 static inline void cgroup_exit(struct task_struct *p) {}
 
 static inline int cgroup_init_early(void) { return 0; }
diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h
index e4a96fb..ec43bce 100644
--- a/include/linux/cgroup_subsys.h
+++ b/include/linux/cgroup_subsys.h
@@ -3,6 +3,17 @@
  *
  * DO NOT ADD ANY SUBSYSTEM WITHOUT EXPLICIT ACKS FROM CGROUP MAINTAINERS.
  */
+
+/*
+ * This file *must* be included with SUBSYS() defined.
+ * SUBSYS_TAG() is a noop if undefined.
+ */
+
+#ifndef SUBSYS_TAG
+#define __TMP_SUBSYS_TAG
+#define SUBSYS_TAG(_x)
+#endif
+
 #if IS_ENABLED(CONFIG_CPUSETS)
 SUBSYS(cpuset)
 #endif
@@ -48,11 +59,23 @@ SUBSYS(hugetlb)
 #endif
 
 /*
+ * Subsystems that implement the can_fork() family of callbacks.
+ */
+SUBSYS_TAG(CANFORK_START)
+SUBSYS_TAG(CANFORK_END)
+
+/*
  * The following subsystems are not supported on the default hierarchy.
  */
 #if IS_ENABLED(CONFIG_CGROUP_DEBUG)
 SUBSYS(debug)
 #endif
+
+#ifdef __TMP_SUBSYS_TAG
+#undef __TMP_SUBSYS_TAG
+#undef SUBSYS_TAG
+#endif
+
 /*
  * DO NOT ADD ANY SUBSYSTEM WITHOUT EXPLICIT ACKS FROM CGROUP MAINTAINERS.
  */
diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 12b580f..ee63ea2 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -186,6 +186,9 @@ static u64 css_serial_nr_next = 1;
 static unsigned long have_fork_callback __read_mostly;
 static unsigned long have_exit_callback __read_mostly;
 
+/* Ditto for the can_fork callback. */
+static unsigned long have_canfork_callback __read_mostly;
+
 static struct cftype cgroup_dfl_base_files[];
 static struct cftype cgroup_legacy_base_files[];
 
@@ -4923,6 +4926,7 @@ static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
 
 	have_fork_callback |= (bool)ss->fork << ss->id;
 	have_exit_callback |= (bool)ss->exit << ss->id;
+	have_canfork_callback |= (bool)ss->can_fork << ss->id;
 
 	/* At system boot, before all subsystems have been
 	 * registered, no tasks have been forked, so we don't
@@ -5165,6 +5169,19 @@ static const struct file_operations proc_cgroupstats_operations = {
 	.release = single_release,
 };
 
+static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
+{
+	if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
+		return &ss_priv[i - CGROUP_CANFORK_START];
+	return NULL;
+}
+
+static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
+{
+	void **private = subsys_canfork_priv_p(ss_priv, i);
+	return private ? *private : NULL;
+}
+
 /**
  * cgroup_fork - initialize cgroup related fields during copy_process()
  * @child: pointer to task_struct of forking parent process.
@@ -5180,6 +5197,57 @@ void cgroup_fork(struct task_struct *child)
 }
 
 /**
+ * cgroup_can_fork - called on a new task before the process is exposed
+ * @child: the task in question.
+ *
+ * This calls the subsystem can_fork() callbacks. If the can_fork() callback
+ * returns an error, the fork aborts with that error code. This allows for
+ * a cgroup subsystem to conditionally allow or deny new forks.
+ */
+int cgroup_can_fork(struct task_struct *child,
+		    void *ss_priv[CGROUP_CANFORK_COUNT])
+{
+	struct cgroup_subsys *ss;
+	int i, j, ret;
+
+	for_each_subsys_which(ss, i, &have_canfork_callback) {
+		ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
+		if (ret)
+			goto out_revert;
+	}
+
+	return 0;
+
+out_revert:
+	for_each_subsys(ss, j) {
+		if (j >= i)
+			break;
+		if (ss->cancel_fork)
+			ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
+	}
+
+	return ret;
+}
+
+/**
+ * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
+ * @child: the task in question
+ *
+ * This calls the cancel_fork() callbacks if a fork failed *after*
+ * cgroup_can_fork() succeded.
+ */
+void cgroup_cancel_fork(struct task_struct *child,
+			void *ss_priv[CGROUP_CANFORK_COUNT])
+{
+	struct cgroup_subsys *ss;
+	int i;
+
+	for_each_subsys(ss, i)
+		if (ss->cancel_fork)
+			ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
+}
+
+/**
  * cgroup_post_fork - called on a new task after adding it to the task list
  * @child: the task in question
  *
@@ -5189,7 +5257,8 @@ void cgroup_fork(struct task_struct *child)
  * cgroup_task_iter_start() - to guarantee that the new task ends up on its
  * list.
  */
-void cgroup_post_fork(struct task_struct *child)
+void cgroup_post_fork(struct task_struct *child,
+		      void *old_ss_priv[CGROUP_CANFORK_COUNT])
 {
 	struct cgroup_subsys *ss;
 	int i;
@@ -5234,7 +5303,7 @@ void cgroup_post_fork(struct task_struct *child)
 	 * and addition to css_set.
 	 */
 	for_each_subsys_which(ss, i, &have_fork_callback)
-		ss->fork(child);
+		ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
 }
 
 /**
diff --git a/kernel/cgroup_freezer.c b/kernel/cgroup_freezer.c
index 92b98cc..f1b30ad 100644
--- a/kernel/cgroup_freezer.c
+++ b/kernel/cgroup_freezer.c
@@ -203,7 +203,7 @@ static void freezer_attach(struct cgroup_subsys_state *new_css,
  * to do anything as freezer_attach() will put @task into the appropriate
  * state.
  */
-static void freezer_fork(struct task_struct *task)
+static void freezer_fork(struct task_struct *task, void *private)
 {
 	struct freezer *freezer;
 
diff --git a/kernel/fork.c b/kernel/fork.c
index 9531275..de7ff51 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -1241,6 +1241,7 @@ static struct task_struct *copy_process(unsigned long clone_flags,
 {
 	int retval;
 	struct task_struct *p;
+	void *cgrp_ss_priv[CGROUP_CANFORK_COUNT] = {};
 
 	if ((clone_flags & (CLONE_NEWNS|CLONE_FS)) == (CLONE_NEWNS|CLONE_FS))
 		return ERR_PTR(-EINVAL);
@@ -1512,6 +1513,16 @@ static struct task_struct *copy_process(unsigned long clone_flags,
 	p->task_works = NULL;
 
 	/*
+	 * Ensure that the cgroup subsystem policies allow the new process to be
+	 * forked. It should be noted the the new process's css_set can be changed
+	 * between here and cgroup_post_fork() if an organisation operation is in
+	 * progress.
+	 */
+	retval = cgroup_can_fork(p, cgrp_ss_priv);
+	if (retval)
+		goto bad_fork_free_pid;
+
+	/*
 	 * Make it visible to the rest of the system, but dont wake it up yet.
 	 * Need tasklist lock for parent etc handling!
 	 */
@@ -1547,7 +1558,7 @@ static struct task_struct *copy_process(unsigned long clone_flags,
 		spin_unlock(&current->sighand->siglock);
 		write_unlock_irq(&tasklist_lock);
 		retval = -ERESTARTNOINTR;
-		goto bad_fork_free_pid;
+		goto bad_fork_cancel_cgroup;
 	}
 
 	if (likely(p->pid)) {
@@ -1589,7 +1600,7 @@ static struct task_struct *copy_process(unsigned long clone_flags,
 	write_unlock_irq(&tasklist_lock);
 
 	proc_fork_connector(p);
-	cgroup_post_fork(p);
+	cgroup_post_fork(p, cgrp_ss_priv);
 	if (clone_flags & CLONE_THREAD)
 		threadgroup_change_end(current);
 	perf_event_fork(p);
@@ -1599,6 +1610,8 @@ static struct task_struct *copy_process(unsigned long clone_flags,
 
 	return p;
 
+bad_fork_cancel_cgroup:
+	cgroup_cancel_fork(p, cgrp_ss_priv);
 bad_fork_free_pid:
 	if (pid != &init_struct_pid)
 		free_pid(pid);
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index f9123a8..050936e 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -8007,7 +8007,7 @@ static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
 	sched_offline_group(tg);
 }
 
-static void cpu_cgroup_fork(struct task_struct *task)
+static void cpu_cgroup_fork(struct task_struct *task, void *private)
 {
 	sched_move_task(task);
 }
-- 
2.4.2

[PATCH v14 4/4] cgroup: 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 -- even if migrating
mid-fork it must be able to fork in the parent first).

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-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>
---
 CREDITS                       |   5 +
 include/linux/cgroup_subsys.h |   5 +
 init/Kconfig                  |  16 ++
 kernel/Makefile               |   1 +
 kernel/cgroup_pids.c          | 366 ++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 393 insertions(+)
 create mode 100644 kernel/cgroup_pids.c

diff --git a/CREDITS b/CREDITS
index 40cc4bf..0727426 100644
--- a/CREDITS
+++ b/CREDITS
@@ -3215,6 +3215,11 @@ S: 69 rue Dunois
 S: 75013 Paris
 S: France
 
+N: Aleksa Sarai
+E: cyphar-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org
+W: https://www.cyphar.com/
+D: `pids` cgroup subsystem
+
 N: Dipankar Sarma
 E: dipankar-xthvdsQ13ZrQT0dZR+AlfA@public.gmane.org
 D: RCU
diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h
index ec43bce..1f36945 100644
--- a/include/linux/cgroup_subsys.h
+++ b/include/linux/cgroup_subsys.h
@@ -62,6 +62,11 @@ SUBSYS(hugetlb)
  * Subsystems that implement the can_fork() family of callbacks.
  */
 SUBSYS_TAG(CANFORK_START)
+
+#if IS_ENABLED(CONFIG_CGROUP_PIDS)
+SUBSYS(pids)
+#endif
+
 SUBSYS_TAG(CANFORK_END)
 
 /*
diff --git a/init/Kconfig b/init/Kconfig
index b9b824b..f4e4918 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -968,6 +968,22 @@ 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 0f8f8b0..df5406c 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -55,6 +55,7 @@ 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..d754888
--- /dev/null
+++ b/kernel/cgroup_pids.c
@@ -0,0 +1,366 @@
+/*
+ * 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, it is not possible to
+ * violate a cgroup policy through fork(). fork() will return -EAGAIN if forking
+ * would cause a cgroup policy to be violated.
+ *
+ * To set a cgroup to have no limit, set pids.max to "max". This is the default
+ * for all new cgroups (N.B. that PID limits are hierarchical, so the most
+ * stringent limit in the hierarchy is followed).
+ *
+ * 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-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>
+ *
+ * This file is subject to the terms and conditions of version 2 of the GNU
+ * General Public License.  See the file COPYING in the main directory of the
+ * Linux distribution for more details.
+ */
+
+#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
+	 * %PIDS_MAX = (%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. This is only used for reverting failed
+ * attaches, where there is no other way out than violating the limit.
+ */
+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;
+
+	cgroup_taskset_for_each(task, tset) {
+		struct cgroup_subsys_state *old_css;
+		struct pids_cgroup *old_pids;
+
+		/*
+		 * Grab a ref to each task's css. We don't drop the ref until
+		 * we either fail and hit ->cancel_attach() or succeed and hit
+		 * ->attach().
+		 */
+		old_css = task_get_css(task, pids_cgrp_id);
+		old_pids = css_pids(old_css);
+
+		pids_charge(pids, 1);
+		pids_uncharge(old_pids, 1);
+	}
+
+	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;
+
+	cgroup_taskset_for_each(task, tset) {
+		struct cgroup_subsys_state *old_css;
+		struct pids_cgroup *old_pids;
+
+		old_css = task_css(task, pids_cgrp_id);
+		old_pids = css_pids(old_css);
+
+		pids_charge(old_pids, 1);
+		pids_uncharge(pids, 1);
+		css_put(old_css);
+	}
+}
+
+static void pids_attach(struct cgroup_subsys_state *css,
+			struct cgroup_taskset *tset)
+{
+	struct task_struct *task;
+
+	cgroup_taskset_for_each(task, tset)
+		css_put(task_css(task, pids_cgrp_id));
+}
+
+static int pids_can_fork(struct task_struct *task, void **priv_p)
+{
+	struct cgroup_subsys_state *css;
+	struct pids_cgroup *pids;
+	int err;
+
+	/*
+	 * 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.
+	 */
+	css = task_get_css(current, pids_cgrp_id);
+	pids = css_pids(css);
+
+	err = pids_try_charge(pids, 1);
+	if (err)
+		goto err_css_put;
+
+	*priv_p = css;
+	return 0;
+
+err_css_put:
+	css_put(css);
+	return err;
+}
+
+static void pids_cancel_fork(struct task_struct *task, void *priv)
+{
+	struct cgroup_subsys_state *css = priv;
+	struct pids_cgroup *pids = css_pids(css);
+
+	pids_uncharge(pids, 1);
+	css_put(css);
+}
+
+static void pids_fork(struct task_struct *task, void *priv)
+{
+	struct cgroup_subsys_state *css;
+	struct cgroup_subsys_state *old_css = priv;
+	struct pids_cgroup *pids;
+	struct pids_cgroup *old_pids = css_pids(old_css);
+
+	css = task_get_css(task, pids_cgrp_id);
+	pids = css_pids(css);
+
+	/*
+	 * If 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;
+
+	if (limit < 0 || limit >= PIDS_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 pids_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,
+	.attach		= pids_attach,
+	.can_attach 	= pids_can_attach,
+	.cancel_attach 	= pids_cancel_attach,
+	.can_fork	= pids_can_fork,
+	.cancel_fork	= pids_cancel_fork,
+	.fork		= pids_fork,
+	.exit		= pids_exit,
+	.legacy_cftypes	= pids_files,
+	.dfl_cftypes	= pids_files,
+};
-- 
2.4.2

Re: [PATCH v14 4/4] cgroup: implement the PIDs subsystem

[Tejun Heo]

On Tue, Jun 09, 2015 at 09:32:10PM +1000, Aleksa Sarai wrote:
> 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 -- even if migrating
> mid-fork it must be able to fork in the parent first).
> 
> 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.

Patches 3-4 look good to me.  Will apply once v4.3 dev window opens.

Thanks.

-- 
tejun

Re: [PATCH v14 4/4] cgroup: implement the PIDs subsystem

[Aleksa Sarai]

Hi Tejun,

> Patches 3-4 look good to me.  Will apply once v4.3 dev window opens.

Do you want me to update Documentation/cgroups/ to include a
description of this?

--
Aleksa Sarai (cyphar)
www.cyphar.com

[PATCH] cgroup: add documentation for the PIDs controller

[Aleksa Sarai]

The attached patch adds documentation concerning the PIDs controller.
This should be applied alongside the rest of this patchset[1].

[1]: https://lkml.org/lkml/2015/6/9/320

8<-----------------------------------------------------------------------------

Add documentation derived from kernel/cgroup_pids.c to the relevant
Documentation/ directory, along with a few examples of how to use the
PIDs controller as well an explanation of its peculiarities.

Signed-off-by: Aleksa Sarai <cyphar-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>
---
 Documentation/cgroups/00-INDEX |  2 +
 Documentation/cgroups/pids.txt | 85 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 87 insertions(+)
 create mode 100644 Documentation/cgroups/pids.txt

diff --git a/Documentation/cgroups/00-INDEX b/Documentation/cgroups/00-INDEX
index 96ce071..3f5a40f 100644
--- a/Documentation/cgroups/00-INDEX
+++ b/Documentation/cgroups/00-INDEX
@@ -22,6 +22,8 @@ net_cls.txt
 	- Network classifier cgroups details and usages.
 net_prio.txt
 	- Network priority cgroups details and usages.
+pids.txt
+	- Process number cgroups details and usages.
 resource_counter.txt
 	- Resource Counter API.
 unified-hierarchy.txt
diff --git a/Documentation/cgroups/pids.txt b/Documentation/cgroups/pids.txt
new file mode 100644
index 0000000..1a078b5
--- /dev/null
+++ b/Documentation/cgroups/pids.txt
@@ -0,0 +1,85 @@
+						   Process Number Controller
+						   =========================
+
+Abstract
+--------
+
+The process number controller is used to allow a cgroup hierarchy to stop any
+new tasks from being fork()'d or clone()'d 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.
+
+Usage
+-----
+
+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. This can be done by either setting the limit to
+be smaller than pids.current, or attaching enough processes to the cgroup such
+that pids.current > pids.max. However, it is not possible to violate a cgroup
+policy through fork() or clone(). fork() and clone() will return -EAGAIN if the
+creation of a new process would cause a cgroup policy to be violated.
+
+To set a cgroup to have no limit, set pids.max to "max". This is the default for
+all new cgroups (N.B. that PID limits are hierarchical, so the most stringent
+limit in the hierarchy is followed).
+
+pids.current tracks all child cgroup hierarchies, so parent/pids.current is a
+superset of parent/child/pids.current.
+
+Example
+-------
+
+First, we mount the pids controller:
+# mkdir -p /sys/fs/cgroup/pids
+# mount -t cgroup -o pids none /sys/fs/cgroup/pids
+
+Then we create a hierarchy, set limits and attach processes to it:
+# mkdir -p /sys/fs/cgroup/pids/parent/child
+# echo 2 > /sys/fs/cgroup/pids/parent/pids.max
+# echo $$ > /sys/fs/cgroup/pids/parent/cgroup.procs
+# cat /sys/fs/cgroup/pids/parent/pids.current
+2
+#
+
+It should be noted that attempts to overcome the set limit (2 in this case) will
+fail:
+
+# cat /sys/fs/cgroup/pids/parent/pids.current
+2
+# ( /bin/echo "Here's some processes for you." | cat )
+sh: fork: Resource temporary unavailable
+#
+
+Even if we migrate to a child cgroup (which doesn't have a set limit), we will
+not be able to overcome the most stringent limit in the hierarchy (in this case,
+parent's):
+
+# echo $$ > /sys/fs/cgroup/pids/parent/child/cgroup.procs
+# cat /sys/fs/cgroup/pids/parent/pids.current
+2
+# cat /sys/fs/cgroup/pids/parent/child/pids.current
+2
+# cat /sys/fs/cgroup/pids/parent/child/pids.max
+max
+# ( /bin/echo "Here's some processes for you." | cat )
+sh: fork: Resource temporary unavailable
+#
+
+We can set a limit that is smaller than pids.current, which will stop any new
+processes from being forked at all (note that the shell itself counts towards
+pids.current):
+
+# echo 1 > /sys/fs/cgroup/pids/parent/pids.max
+# /bin/echo "We can't even spawn a single process now."
+sh: fork: Resource temporary unavailable
+# echo 0 > /sys/fs/cgroup/pids/parent/pids.max
+# /bin/echo "We can't even spawn a single process now."
+sh: fork: Resource temporary unavailable
+#
-- 
2.4.2

Re: [PATCH] cgroup: add documentation for the PIDs controller

[Tejun Heo]

On Sat, Jun 13, 2015 at 03:21:58AM +1000, Aleksa Sarai wrote:
> The attached patch adds documentation concerning the PIDs controller.
> This should be applied alongside the rest of this patchset[1].
> 
> [1]: https://lkml.org/lkml/2015/6/9/320
> 
> 8<-----------------------------------------------------------------------------
> 
> Add documentation derived from kernel/cgroup_pids.c to the relevant
> Documentation/ directory, along with a few examples of how to use the
> PIDs controller as well an explanation of its peculiarities.
> 
> Signed-off-by: Aleksa Sarai <cyphar-gVpy/LI/lHzQT0dZR+AlfA@public.gmane.org>

Applied to cgroup/for-4.3.

Thanks.

-- 
tejun

Re: [PATCH v14 4/4] cgroup: implement the PIDs subsystem

[Tejun Heo]

On Wed, Jun 10, 2015 at 01:53:04PM +0900, Tejun Heo wrote:
> On Tue, Jun 09, 2015 at 09:32:10PM +1000, Aleksa Sarai wrote:
> > 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 -- even if migrating
> > mid-fork it must be able to fork in the parent first).
> > 
> > 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.
> 
> Patches 3-4 look good to me.  Will apply once v4.3 dev window opens.

Applied 3-4 to cgroup/for-4.3.

Thanks.

-- 
tejun