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
41struct pids_cgroup {
42	struct cgroup_subsys_state	css;
43
44	/*
45	 * Use 64-bit types so that we can safely represent "max" as
46	 * %PIDS_MAX = (%PID_MAX_LIMIT + 1).
47	 */
48	atomic64_t			counter;
49	atomic64_t			limit;
50	int64_t				watermark;
51
52	/* Handle for "pids.events" */
53	struct cgroup_file		events_file;
54
55	/* Number of times fork failed because limit was hit. */
56	atomic64_t			events_limit;
57};
58
59static struct pids_cgroup *css_pids(struct cgroup_subsys_state *css)
60{
61	return container_of(css, struct pids_cgroup, css);
62}
63
64static struct pids_cgroup *parent_pids(struct pids_cgroup *pids)
65{
66	return css_pids(pids->css.parent);
67}
68
69static struct cgroup_subsys_state *
70pids_css_alloc(struct cgroup_subsys_state *parent)
71{
72	struct pids_cgroup *pids;
73
74	pids = kzalloc(sizeof(struct pids_cgroup), GFP_KERNEL);
75	if (!pids)
76		return ERR_PTR(-ENOMEM);
77
78	atomic64_set(&pids->counter, 0);
79	atomic64_set(&pids->limit, PIDS_MAX);
80	atomic64_set(&pids->events_limit, 0);
81	return &pids->css;
82}
83
84static void pids_css_free(struct cgroup_subsys_state *css)
85{
86	kfree(css_pids(css));
87}
88
89static void pids_update_watermark(struct pids_cgroup *p, int64_t nr_pids)
90{
91	/*
92	 * This is racy, but we don't need perfectly accurate tallying of
93	 * the watermark, and this lets us avoid extra atomic overhead.
94	 */
95	if (nr_pids > READ_ONCE(p->watermark))
96		WRITE_ONCE(p->watermark, nr_pids);
97}
98
99/**
100 * pids_cancel - uncharge the local pid count
101 * @pids: the pid cgroup state
102 * @num: the number of pids to cancel
103 *
104 * This function will WARN if the pid count goes under 0, because such a case is
105 * a bug in the pids controller proper.
106 */
107static void pids_cancel(struct pids_cgroup *pids, int num)
108{
109	/*
110	 * A negative count (or overflow for that matter) is invalid,
111	 * and indicates a bug in the `pids` controller proper.
112	 */
113	WARN_ON_ONCE(atomic64_add_negative(-num, &pids->counter));
114}
115
116/**
117 * pids_uncharge - hierarchically uncharge the pid count
118 * @pids: the pid cgroup state
119 * @num: the number of pids to uncharge
120 */
121static void pids_uncharge(struct pids_cgroup *pids, int num)
122{
123	struct pids_cgroup *p;
124
125	for (p = pids; parent_pids(p); p = parent_pids(p))
126		pids_cancel(p, num);
127}
128
129/**
130 * pids_charge - hierarchically charge the pid count
131 * @pids: the pid cgroup state
132 * @num: the number of pids to charge
133 *
134 * This function does *not* follow the pid limit set. It cannot fail and the new
135 * pid count may exceed the limit. This is only used for reverting failed
136 * attaches, where there is no other way out than violating the limit.
137 */
138static void pids_charge(struct pids_cgroup *pids, int num)
139{
140	struct pids_cgroup *p;
141
142	for (p = pids; parent_pids(p); p = parent_pids(p)) {
143		int64_t new = atomic64_add_return(num, &p->counter);
144
145		pids_update_watermark(p, new);
146	}
147}
148
149/**
150 * pids_try_charge - hierarchically try to charge the pid count
151 * @pids: the pid cgroup state
152 * @num: the number of pids to charge
153 *
154 * This function follows the set limit. It will fail if the charge would cause
155 * the new value to exceed the hierarchical limit. Returns 0 if the charge
156 * succeeded, otherwise -EAGAIN.
157 */
158static int pids_try_charge(struct pids_cgroup *pids, int num)
159{
160	struct pids_cgroup *p, *q;
161
162	for (p = pids; parent_pids(p); p = parent_pids(p)) {
163		int64_t new = atomic64_add_return(num, &p->counter);
164		int64_t limit = atomic64_read(&p->limit);
165
166		/*
167		 * Since new is capped to the maximum number of pid_t, if
168		 * p->limit is %PIDS_MAX then we know that this test will never
169		 * fail.
170		 */
171		if (new > limit)
172			goto revert;
173
174		/*
175		 * Not technically accurate if we go over limit somewhere up
176		 * the hierarchy, but that's tolerable for the watermark.
177		 */
178		pids_update_watermark(p, new);
179	}
180
181	return 0;
182
183revert:
184	for (q = pids; q != p; q = parent_pids(q))
185		pids_cancel(q, num);
186	pids_cancel(p, num);
187
188	return -EAGAIN;
189}
190
191static int pids_can_attach(struct cgroup_taskset *tset)
192{
193	struct task_struct *task;
194	struct cgroup_subsys_state *dst_css;
195
196	cgroup_taskset_for_each(task, dst_css, tset) {
197		struct pids_cgroup *pids = css_pids(dst_css);
198		struct cgroup_subsys_state *old_css;
199		struct pids_cgroup *old_pids;
200
201		/*
202		 * No need to pin @old_css between here and cancel_attach()
203		 * because cgroup core protects it from being freed before
204		 * the migration completes or fails.
205		 */
206		old_css = task_css(task, pids_cgrp_id);
207		old_pids = css_pids(old_css);
208
209		pids_charge(pids, 1);
210		pids_uncharge(old_pids, 1);
211	}
212
213	return 0;
214}
215
216static void pids_cancel_attach(struct cgroup_taskset *tset)
217{
218	struct task_struct *task;
219	struct cgroup_subsys_state *dst_css;
220
221	cgroup_taskset_for_each(task, dst_css, tset) {
222		struct pids_cgroup *pids = css_pids(dst_css);
223		struct cgroup_subsys_state *old_css;
224		struct pids_cgroup *old_pids;
225
226		old_css = task_css(task, pids_cgrp_id);
227		old_pids = css_pids(old_css);
228
229		pids_charge(old_pids, 1);
230		pids_uncharge(pids, 1);
231	}
232}
233
234/*
235 * task_css_check(true) in pids_can_fork() and pids_cancel_fork() relies
236 * on cgroup_threadgroup_change_begin() held by the copy_process().
237 */
238static int pids_can_fork(struct task_struct *task, struct css_set *cset)
239{
240	struct cgroup_subsys_state *css;
241	struct pids_cgroup *pids;
242	int err;
243
244	if (cset)
245		css = cset->subsys[pids_cgrp_id];
246	else
247		css = task_css_check(current, pids_cgrp_id, true);
248	pids = css_pids(css);
249	err = pids_try_charge(pids, 1);
250	if (err) {
251		/* Only log the first time events_limit is incremented. */
252		if (atomic64_inc_return(&pids->events_limit) == 1) {
253			pr_info("cgroup: fork rejected by pids controller in ");
254			pr_cont_cgroup_path(css->cgroup);
255			pr_cont("\n");
256		}
257		cgroup_file_notify(&pids->events_file);
258	}
259	return err;
260}
261
262static void pids_cancel_fork(struct task_struct *task, struct css_set *cset)
263{
264	struct cgroup_subsys_state *css;
265	struct pids_cgroup *pids;
266
267	if (cset)
268		css = cset->subsys[pids_cgrp_id];
269	else
270		css = task_css_check(current, pids_cgrp_id, true);
271	pids = css_pids(css);
272	pids_uncharge(pids, 1);
273}
274
275static void pids_release(struct task_struct *task)
276{
277	struct pids_cgroup *pids = css_pids(task_css(task, pids_cgrp_id));
278
279	pids_uncharge(pids, 1);
280}
281
282static ssize_t pids_max_write(struct kernfs_open_file *of, char *buf,
283			      size_t nbytes, loff_t off)
284{
285	struct cgroup_subsys_state *css = of_css(of);
286	struct pids_cgroup *pids = css_pids(css);
287	int64_t limit;
288	int err;
289
290	buf = strstrip(buf);
291	if (!strcmp(buf, PIDS_MAX_STR)) {
292		limit = PIDS_MAX;
293		goto set_limit;
294	}
295
296	err = kstrtoll(buf, 0, &limit);
297	if (err)
298		return err;
299
300	if (limit < 0 || limit >= PIDS_MAX)
301		return -EINVAL;
302
303set_limit:
304	/*
305	 * Limit updates don't need to be mutex'd, since it isn't
306	 * critical that any racing fork()s follow the new limit.
307	 */
308	atomic64_set(&pids->limit, limit);
309	return nbytes;
310}
311
312static int pids_max_show(struct seq_file *sf, void *v)
313{
314	struct cgroup_subsys_state *css = seq_css(sf);
315	struct pids_cgroup *pids = css_pids(css);
316	int64_t limit = atomic64_read(&pids->limit);
317
318	if (limit >= PIDS_MAX)
319		seq_printf(sf, "%s\n", PIDS_MAX_STR);
320	else
321		seq_printf(sf, "%lld\n", limit);
322
323	return 0;
324}
325
326static s64 pids_current_read(struct cgroup_subsys_state *css,
327			     struct cftype *cft)
328{
329	struct pids_cgroup *pids = css_pids(css);
330
331	return atomic64_read(&pids->counter);
332}
333
334static s64 pids_peak_read(struct cgroup_subsys_state *css,
335			  struct cftype *cft)
336{
337	struct pids_cgroup *pids = css_pids(css);
338
339	return READ_ONCE(pids->watermark);
340}
341
342static int pids_events_show(struct seq_file *sf, void *v)
343{
344	struct pids_cgroup *pids = css_pids(seq_css(sf));
345
346	seq_printf(sf, "max %lld\n", (s64)atomic64_read(&pids->events_limit));
347	return 0;
348}
349
350static struct cftype pids_files[] = {
351	{
352		.name = "max",
353		.write = pids_max_write,
354		.seq_show = pids_max_show,
355		.flags = CFTYPE_NOT_ON_ROOT,
356	},
357	{
358		.name = "current",
359		.read_s64 = pids_current_read,
360		.flags = CFTYPE_NOT_ON_ROOT,
361	},
362	{
363		.name = "peak",
364		.flags = CFTYPE_NOT_ON_ROOT,
365		.read_s64 = pids_peak_read,
366	},
367	{
368		.name = "events",
369		.seq_show = pids_events_show,
370		.file_offset = offsetof(struct pids_cgroup, events_file),
371		.flags = CFTYPE_NOT_ON_ROOT,
372	},
373	{ }	/* terminate */
374};
375
376struct cgroup_subsys pids_cgrp_subsys = {
377	.css_alloc	= pids_css_alloc,
378	.css_free	= pids_css_free,
379	.can_attach 	= pids_can_attach,
380	.cancel_attach 	= pids_cancel_attach,
381	.can_fork	= pids_can_fork,
382	.cancel_fork	= pids_cancel_fork,
383	.release	= pids_release,
384	.legacy_cftypes	= pids_files,
385	.dfl_cftypes	= pids_files,
386	.threaded	= true,
387};
388