1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Process number limiting controller for cgroups.
  4 *
  5 * Used to allow a cgroup hierarchy to stop any new processes from fork()ing
  6 * after a certain limit is reached.
  7 *
  8 * Since it is trivial to hit the task limit without hitting any kmemcg limits
  9 * in place, PIDs are a fundamental resource. As such, PID exhaustion must be
 10 * preventable in the scope of a cgroup hierarchy by allowing resource limiting
 11 * of the number of tasks in a cgroup.
 12 *
 13 * In order to use the `pids` controller, set the maximum number of tasks in
 14 * pids.max (this is not available in the root cgroup for obvious reasons). The
 15 * number of processes currently in the cgroup is given by pids.current.
 16 * Organisational operations are not blocked by cgroup policies, so it is
 17 * possible to have pids.current > pids.max. However, it is not possible to
 18 * violate a cgroup policy through fork(). fork() will return -EAGAIN if forking
 19 * would cause a cgroup policy to be violated.
 20 *
 21 * To set a cgroup to have no limit, set pids.max to "max". This is the default
 22 * for all new cgroups (N.B. that PID limits are hierarchical, so the most
 23 * stringent limit in the hierarchy is followed).
 24 *
 25 * pids.current tracks all child cgroup hierarchies, so parent/pids.current is
 26 * a superset of parent/child/pids.current.
 27 *
 28 * Copyright (C) 2015 Aleksa Sarai <cyphar@cyphar.com>
 29 */
 30
 31#include <linux/kernel.h>
 32#include <linux/threads.h>
 33#include <linux/atomic.h>
 34#include <linux/cgroup.h>
 35#include <linux/slab.h>
 36#include <linux/sched/task.h>
 37
 38#define PIDS_MAX (PID_MAX_LIMIT + 1ULL)
 39#define PIDS_MAX_STR "max"
 40
 41enum pidcg_event {
 42	/* Fork failed in subtree because this pids_cgroup limit was hit. */
 43	PIDCG_MAX,
 44	/* Fork failed in this pids_cgroup because ancestor limit was hit. */
 45	PIDCG_FORKFAIL,
 46	NR_PIDCG_EVENTS,
 47};
 48
 49struct pids_cgroup {
 50	struct cgroup_subsys_state	css;
 51
 52	/*
 53	 * Use 64-bit types so that we can safely represent "max" as
 54	 * %PIDS_MAX = (%PID_MAX_LIMIT + 1).
 55	 */
 56	atomic64_t			counter;
 57	atomic64_t			limit;
 58	int64_t				watermark;
 59
 60	/* Handles for pids.events[.local] */
 61	struct cgroup_file		events_file;
 62	struct cgroup_file		events_local_file;
 63
 64	atomic64_t			events[NR_PIDCG_EVENTS];
 65	atomic64_t			events_local[NR_PIDCG_EVENTS];
 66};
 67
 68static struct pids_cgroup *css_pids(struct cgroup_subsys_state *css)
 69{
 70	return container_of(css, struct pids_cgroup, css);
 71}
 72
 73static struct pids_cgroup *parent_pids(struct pids_cgroup *pids)
 74{
 75	return css_pids(pids->css.parent);
 76}
 77
 78static struct cgroup_subsys_state *
 79pids_css_alloc(struct cgroup_subsys_state *parent)
 80{
 81	struct pids_cgroup *pids;
 82
 83	pids = kzalloc(sizeof(struct pids_cgroup), GFP_KERNEL);
 84	if (!pids)
 85		return ERR_PTR(-ENOMEM);
 86
 87	atomic64_set(&pids->limit, PIDS_MAX);
 88	return &pids->css;
 89}
 90
 91static void pids_css_free(struct cgroup_subsys_state *css)
 92{
 93	kfree(css_pids(css));
 94}
 95
 96static void pids_update_watermark(struct pids_cgroup *p, int64_t nr_pids)
 97{
 98	/*
 99	 * This is racy, but we don't need perfectly accurate tallying of
100	 * the watermark, and this lets us avoid extra atomic overhead.
101	 */
102	if (nr_pids > READ_ONCE(p->watermark))
103		WRITE_ONCE(p->watermark, nr_pids);
104}
105
106/**
107 * pids_cancel - uncharge the local pid count
108 * @pids: the pid cgroup state
109 * @num: the number of pids to cancel
110 *
111 * This function will WARN if the pid count goes under 0, because such a case is
112 * a bug in the pids controller proper.
113 */
114static void pids_cancel(struct pids_cgroup *pids, int num)
115{
116	/*
117	 * A negative count (or overflow for that matter) is invalid,
118	 * and indicates a bug in the `pids` controller proper.
119	 */
120	WARN_ON_ONCE(atomic64_add_negative(-num, &pids->counter));
121}
122
123/**
124 * pids_uncharge - hierarchically uncharge the pid count
125 * @pids: the pid cgroup state
126 * @num: the number of pids to uncharge
127 */
128static void pids_uncharge(struct pids_cgroup *pids, int num)
129{
130	struct pids_cgroup *p;
131
132	for (p = pids; parent_pids(p); p = parent_pids(p))
133		pids_cancel(p, num);
134}
135
136/**
137 * pids_charge - hierarchically charge the pid count
138 * @pids: the pid cgroup state
139 * @num: the number of pids to charge
140 *
141 * This function does *not* follow the pid limit set. It cannot fail and the new
142 * pid count may exceed the limit. This is only used for reverting failed
143 * attaches, where there is no other way out than violating the limit.
144 */
145static void pids_charge(struct pids_cgroup *pids, int num)
146{
147	struct pids_cgroup *p;
148
149	for (p = pids; parent_pids(p); p = parent_pids(p)) {
150		int64_t new = atomic64_add_return(num, &p->counter);
151
152		pids_update_watermark(p, new);
153	}
154}
155
156/**
157 * pids_try_charge - hierarchically try to charge the pid count
158 * @pids: the pid cgroup state
159 * @num: the number of pids to charge
160 * @fail: storage of pid cgroup causing the fail
161 *
162 * This function follows the set limit. It will fail if the charge would cause
163 * the new value to exceed the hierarchical limit. Returns 0 if the charge
164 * succeeded, otherwise -EAGAIN.
165 */
166static int pids_try_charge(struct pids_cgroup *pids, int num, struct pids_cgroup **fail)
167{
168	struct pids_cgroup *p, *q;
169
170	for (p = pids; parent_pids(p); p = parent_pids(p)) {
171		int64_t new = atomic64_add_return(num, &p->counter);
172		int64_t limit = atomic64_read(&p->limit);
173
174		/*
175		 * Since new is capped to the maximum number of pid_t, if
176		 * p->limit is %PIDS_MAX then we know that this test will never
177		 * fail.
178		 */
179		if (new > limit) {
180			*fail = p;
181			goto revert;
182		}
183		/*
184		 * Not technically accurate if we go over limit somewhere up
185		 * the hierarchy, but that's tolerable for the watermark.
186		 */
187		pids_update_watermark(p, new);
188	}
189
190	return 0;
191
192revert:
193	for (q = pids; q != p; q = parent_pids(q))
194		pids_cancel(q, num);
195	pids_cancel(p, num);
196
197	return -EAGAIN;
198}
199
200static int pids_can_attach(struct cgroup_taskset *tset)
201{
202	struct task_struct *task;
203	struct cgroup_subsys_state *dst_css;
204
205	cgroup_taskset_for_each(task, dst_css, tset) {
206		struct pids_cgroup *pids = css_pids(dst_css);
207		struct cgroup_subsys_state *old_css;
208		struct pids_cgroup *old_pids;
209
210		/*
211		 * No need to pin @old_css between here and cancel_attach()
212		 * because cgroup core protects it from being freed before
213		 * the migration completes or fails.
214		 */
215		old_css = task_css(task, pids_cgrp_id);
216		old_pids = css_pids(old_css);
217
218		pids_charge(pids, 1);
219		pids_uncharge(old_pids, 1);
220	}
221
222	return 0;
223}
224
225static void pids_cancel_attach(struct cgroup_taskset *tset)
226{
227	struct task_struct *task;
228	struct cgroup_subsys_state *dst_css;
229
230	cgroup_taskset_for_each(task, dst_css, tset) {
231		struct pids_cgroup *pids = css_pids(dst_css);
232		struct cgroup_subsys_state *old_css;
233		struct pids_cgroup *old_pids;
234
235		old_css = task_css(task, pids_cgrp_id);
236		old_pids = css_pids(old_css);
237
238		pids_charge(old_pids, 1);
239		pids_uncharge(pids, 1);
240	}
241}
242
243static void pids_event(struct pids_cgroup *pids_forking,
244		       struct pids_cgroup *pids_over_limit)
245{
246	struct pids_cgroup *p = pids_forking;
247
248	/* Only log the first time limit is hit. */
249	if (atomic64_inc_return(&p->events_local[PIDCG_FORKFAIL]) == 1) {
250		pr_info("cgroup: fork rejected by pids controller in ");
251		pr_cont_cgroup_path(p->css.cgroup);
252		pr_cont("\n");
253	}
254	if (!cgroup_subsys_on_dfl(pids_cgrp_subsys) ||
255	    cgrp_dfl_root.flags & CGRP_ROOT_PIDS_LOCAL_EVENTS) {
256		cgroup_file_notify(&p->events_local_file);
257		return;
258	}
259
260	atomic64_inc(&pids_over_limit->events_local[PIDCG_MAX]);
261	cgroup_file_notify(&pids_over_limit->events_local_file);
262
263	for (p = pids_over_limit; parent_pids(p); p = parent_pids(p)) {
264		atomic64_inc(&p->events[PIDCG_MAX]);
265		cgroup_file_notify(&p->events_file);
266	}
267}
268
269/*
270 * task_css_check(true) in pids_can_fork() and pids_cancel_fork() relies
271 * on cgroup_threadgroup_change_begin() held by the copy_process().
272 */
273static int pids_can_fork(struct task_struct *task, struct css_set *cset)
274{
275	struct pids_cgroup *pids, *pids_over_limit;
276	int err;
277
278	pids = css_pids(cset->subsys[pids_cgrp_id]);
279	err = pids_try_charge(pids, 1, &pids_over_limit);
280	if (err)
281		pids_event(pids, pids_over_limit);
282
283	return err;
284}
285
286static void pids_cancel_fork(struct task_struct *task, struct css_set *cset)
287{
288	struct pids_cgroup *pids;
289
290	pids = css_pids(cset->subsys[pids_cgrp_id]);
291	pids_uncharge(pids, 1);
292}
293
294static void pids_release(struct task_struct *task)
295{
296	struct pids_cgroup *pids = css_pids(task_css(task, pids_cgrp_id));
297
298	pids_uncharge(pids, 1);
299}
300
301static ssize_t pids_max_write(struct kernfs_open_file *of, char *buf,
302			      size_t nbytes, loff_t off)
303{
304	struct cgroup_subsys_state *css = of_css(of);
305	struct pids_cgroup *pids = css_pids(css);
306	int64_t limit;
307	int err;
308
309	buf = strstrip(buf);
310	if (!strcmp(buf, PIDS_MAX_STR)) {
311		limit = PIDS_MAX;
312		goto set_limit;
313	}
314
315	err = kstrtoll(buf, 0, &limit);
316	if (err)
317		return err;
318
319	if (limit < 0 || limit >= PIDS_MAX)
320		return -EINVAL;
321
322set_limit:
323	/*
324	 * Limit updates don't need to be mutex'd, since it isn't
325	 * critical that any racing fork()s follow the new limit.
326	 */
327	atomic64_set(&pids->limit, limit);
328	return nbytes;
329}
330
331static int pids_max_show(struct seq_file *sf, void *v)
332{
333	struct cgroup_subsys_state *css = seq_css(sf);
334	struct pids_cgroup *pids = css_pids(css);
335	int64_t limit = atomic64_read(&pids->limit);
336
337	if (limit >= PIDS_MAX)
338		seq_printf(sf, "%s\n", PIDS_MAX_STR);
339	else
340		seq_printf(sf, "%lld\n", limit);
341
342	return 0;
343}
344
345static s64 pids_current_read(struct cgroup_subsys_state *css,
346			     struct cftype *cft)
347{
348	struct pids_cgroup *pids = css_pids(css);
349
350	return atomic64_read(&pids->counter);
351}
352
353static s64 pids_peak_read(struct cgroup_subsys_state *css,
354			  struct cftype *cft)
355{
356	struct pids_cgroup *pids = css_pids(css);
357
358	return READ_ONCE(pids->watermark);
359}
360
361static int __pids_events_show(struct seq_file *sf, bool local)
362{
363	struct pids_cgroup *pids = css_pids(seq_css(sf));
364	enum pidcg_event pe = PIDCG_MAX;
365	atomic64_t *events;
366
367	if (!cgroup_subsys_on_dfl(pids_cgrp_subsys) ||
368	    cgrp_dfl_root.flags & CGRP_ROOT_PIDS_LOCAL_EVENTS) {
369		pe = PIDCG_FORKFAIL;
370		local = true;
371	}
372	events = local ? pids->events_local : pids->events;
373
374	seq_printf(sf, "max %lld\n", (s64)atomic64_read(&events[pe]));
375	return 0;
376}
377
378static int pids_events_show(struct seq_file *sf, void *v)
379{
380	__pids_events_show(sf, false);
381	return 0;
382}
383
384static int pids_events_local_show(struct seq_file *sf, void *v)
385{
386	__pids_events_show(sf, true);
387	return 0;
388}
389
390static struct cftype pids_files[] = {
391	{
392		.name = "max",
393		.write = pids_max_write,
394		.seq_show = pids_max_show,
395		.flags = CFTYPE_NOT_ON_ROOT,
396	},
397	{
398		.name = "current",
399		.read_s64 = pids_current_read,
400		.flags = CFTYPE_NOT_ON_ROOT,
401	},
402	{
403		.name = "peak",
404		.flags = CFTYPE_NOT_ON_ROOT,
405		.read_s64 = pids_peak_read,
406	},
407	{
408		.name = "events",
409		.seq_show = pids_events_show,
410		.file_offset = offsetof(struct pids_cgroup, events_file),
411		.flags = CFTYPE_NOT_ON_ROOT,
412	},
413	{
414		.name = "events.local",
415		.seq_show = pids_events_local_show,
416		.file_offset = offsetof(struct pids_cgroup, events_local_file),
417		.flags = CFTYPE_NOT_ON_ROOT,
418	},
419	{ }	/* terminate */
420};
421
422static struct cftype pids_files_legacy[] = {
423	{
424		.name = "max",
425		.write = pids_max_write,
426		.seq_show = pids_max_show,
427		.flags = CFTYPE_NOT_ON_ROOT,
428	},
429	{
430		.name = "current",
431		.read_s64 = pids_current_read,
432		.flags = CFTYPE_NOT_ON_ROOT,
433	},
434	{
435		.name = "peak",
436		.flags = CFTYPE_NOT_ON_ROOT,
437		.read_s64 = pids_peak_read,
438	},
439	{
440		.name = "events",
441		.seq_show = pids_events_show,
442		.file_offset = offsetof(struct pids_cgroup, events_file),
443		.flags = CFTYPE_NOT_ON_ROOT,
444	},
445	{ }	/* terminate */
446};
447
448
449struct cgroup_subsys pids_cgrp_subsys = {
450	.css_alloc	= pids_css_alloc,
451	.css_free	= pids_css_free,
452	.can_attach 	= pids_can_attach,
453	.cancel_attach 	= pids_cancel_attach,
454	.can_fork	= pids_can_fork,
455	.cancel_fork	= pids_cancel_fork,
456	.release	= pids_release,
457	.legacy_cftypes = pids_files_legacy,
458	.dfl_cftypes	= pids_files,
459	.threaded	= true,
460};