1/*
  2 * net/core/netprio_cgroup.c	Priority Control Group
  3 *
  4 *		This program is free software; you can redistribute it and/or
  5 *		modify it under the terms of the GNU General Public License
  6 *		as published by the Free Software Foundation; either version
  7 *		2 of the License, or (at your option) any later version.
  8 *
  9 * Authors:	Neil Horman <nhorman@tuxdriver.com>
 10 */
 11
 12#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 13
 14#include <linux/module.h>
 15#include <linux/slab.h>
 16#include <linux/types.h>
 17#include <linux/string.h>
 18#include <linux/errno.h>
 19#include <linux/skbuff.h>
 20#include <linux/cgroup.h>
 21#include <linux/rcupdate.h>
 22#include <linux/atomic.h>
 23#include <linux/sched/task.h>
 24
 25#include <net/rtnetlink.h>
 26#include <net/pkt_cls.h>
 27#include <net/sock.h>
 28#include <net/netprio_cgroup.h>
 29
 30#include <linux/fdtable.h>
 31
 32/*
 33 * netprio allocates per-net_device priomap array which is indexed by
 34 * css->id.  Limiting css ID to 16bits doesn't lose anything.
 35 */
 36#define NETPRIO_ID_MAX		USHRT_MAX
 37
 38#define PRIOMAP_MIN_SZ		128
 39
 40/*
 41 * Extend @dev->priomap so that it's large enough to accommodate
 42 * @target_idx.  @dev->priomap.priomap_len > @target_idx after successful
 43 * return.  Must be called under rtnl lock.
 44 */
 45static int extend_netdev_table(struct net_device *dev, u32 target_idx)
 46{
 47	struct netprio_map *old, *new;
 48	size_t new_sz, new_len;
 49
 50	/* is the existing priomap large enough? */
 51	old = rtnl_dereference(dev->priomap);
 52	if (old && old->priomap_len > target_idx)
 53		return 0;
 54
 55	/*
 56	 * Determine the new size.  Let's keep it power-of-two.  We start
 57	 * from PRIOMAP_MIN_SZ and double it until it's large enough to
 58	 * accommodate @target_idx.
 59	 */
 60	new_sz = PRIOMAP_MIN_SZ;
 61	while (true) {
 62		new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
 63			sizeof(new->priomap[0]);
 64		if (new_len > target_idx)
 65			break;
 66		new_sz *= 2;
 67		/* overflowed? */
 68		if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
 69			return -ENOSPC;
 70	}
 71
 72	/* allocate & copy */
 73	new = kzalloc(new_sz, GFP_KERNEL);
 74	if (!new)
 75		return -ENOMEM;
 76
 77	if (old)
 78		memcpy(new->priomap, old->priomap,
 79		       old->priomap_len * sizeof(old->priomap[0]));
 80
 81	new->priomap_len = new_len;
 82
 83	/* install the new priomap */
 84	rcu_assign_pointer(dev->priomap, new);
 85	if (old)
 86		kfree_rcu(old, rcu);
 87	return 0;
 88}
 89
 90/**
 91 * netprio_prio - return the effective netprio of a cgroup-net_device pair
 92 * @css: css part of the target pair
 93 * @dev: net_device part of the target pair
 94 *
 95 * Should be called under RCU read or rtnl lock.
 96 */
 97static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
 98{
 99	struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
100	int id = css->cgroup->id;
101
102	if (map && id < map->priomap_len)
103		return map->priomap[id];
104	return 0;
105}
106
107/**
108 * netprio_set_prio - set netprio on a cgroup-net_device pair
109 * @css: css part of the target pair
110 * @dev: net_device part of the target pair
111 * @prio: prio to set
112 *
113 * Set netprio to @prio on @css-@dev pair.  Should be called under rtnl
114 * lock and may fail under memory pressure for non-zero @prio.
115 */
116static int netprio_set_prio(struct cgroup_subsys_state *css,
117			    struct net_device *dev, u32 prio)
118{
119	struct netprio_map *map;
120	int id = css->cgroup->id;
121	int ret;
122
123	/* avoid extending priomap for zero writes */
124	map = rtnl_dereference(dev->priomap);
125	if (!prio && (!map || map->priomap_len <= id))
126		return 0;
127
128	ret = extend_netdev_table(dev, id);
129	if (ret)
130		return ret;
131
132	map = rtnl_dereference(dev->priomap);
133	map->priomap[id] = prio;
134	return 0;
135}
136
137static struct cgroup_subsys_state *
138cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
139{
140	struct cgroup_subsys_state *css;
141
142	css = kzalloc(sizeof(*css), GFP_KERNEL);
143	if (!css)
144		return ERR_PTR(-ENOMEM);
145
146	return css;
147}
148
149static int cgrp_css_online(struct cgroup_subsys_state *css)
150{
151	struct cgroup_subsys_state *parent_css = css->parent;
152	struct net_device *dev;
153	int ret = 0;
154
155	if (css->id > NETPRIO_ID_MAX)
156		return -ENOSPC;
157
158	if (!parent_css)
159		return 0;
160
161	rtnl_lock();
162	/*
163	 * Inherit prios from the parent.  As all prios are set during
164	 * onlining, there is no need to clear them on offline.
165	 */
166	for_each_netdev(&init_net, dev) {
167		u32 prio = netprio_prio(parent_css, dev);
168
169		ret = netprio_set_prio(css, dev, prio);
170		if (ret)
171			break;
172	}
173	rtnl_unlock();
174	return ret;
175}
176
177static void cgrp_css_free(struct cgroup_subsys_state *css)
178{
179	kfree(css);
180}
181
182static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
183{
184	return css->cgroup->id;
185}
186
187static int read_priomap(struct seq_file *sf, void *v)
188{
189	struct net_device *dev;
190
191	rcu_read_lock();
192	for_each_netdev_rcu(&init_net, dev)
193		seq_printf(sf, "%s %u\n", dev->name,
194			   netprio_prio(seq_css(sf), dev));
195	rcu_read_unlock();
196	return 0;
197}
198
199static ssize_t write_priomap(struct kernfs_open_file *of,
200			     char *buf, size_t nbytes, loff_t off)
201{
202	char devname[IFNAMSIZ + 1];
203	struct net_device *dev;
204	u32 prio;
205	int ret;
206
207	if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
208		return -EINVAL;
209
210	dev = dev_get_by_name(&init_net, devname);
211	if (!dev)
212		return -ENODEV;
213
214	cgroup_sk_alloc_disable();
215
216	rtnl_lock();
217
218	ret = netprio_set_prio(of_css(of), dev, prio);
219
220	rtnl_unlock();
221	dev_put(dev);
222	return ret ?: nbytes;
223}
224
225static int update_netprio(const void *v, struct file *file, unsigned n)
226{
227	int err;
228	struct socket *sock = sock_from_file(file, &err);
229	if (sock) {
230		spin_lock(&cgroup_sk_update_lock);
231		sock_cgroup_set_prioidx(&sock->sk->sk_cgrp_data,
232					(unsigned long)v);
233		spin_unlock(&cgroup_sk_update_lock);
234	}
235	return 0;
236}
237
238static void net_prio_attach(struct cgroup_taskset *tset)
239{
240	struct task_struct *p;
241	struct cgroup_subsys_state *css;
242
243	cgroup_taskset_for_each(p, css, tset) {
244		void *v = (void *)(unsigned long)css->cgroup->id;
245
246		task_lock(p);
247		iterate_fd(p->files, 0, update_netprio, v);
248		task_unlock(p);
249	}
250}
251
252static struct cftype ss_files[] = {
253	{
254		.name = "prioidx",
255		.read_u64 = read_prioidx,
256	},
257	{
258		.name = "ifpriomap",
259		.seq_show = read_priomap,
260		.write = write_priomap,
261	},
262	{ }	/* terminate */
263};
264
265struct cgroup_subsys net_prio_cgrp_subsys = {
266	.css_alloc	= cgrp_css_alloc,
267	.css_online	= cgrp_css_online,
268	.css_free	= cgrp_css_free,
269	.attach		= net_prio_attach,
270	.legacy_cftypes	= ss_files,
271};
272
273static int netprio_device_event(struct notifier_block *unused,
274				unsigned long event, void *ptr)
275{
276	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
277	struct netprio_map *old;
278
279	/*
280	 * Note this is called with rtnl_lock held so we have update side
281	 * protection on our rcu assignments
282	 */
283
284	switch (event) {
285	case NETDEV_UNREGISTER:
286		old = rtnl_dereference(dev->priomap);
287		RCU_INIT_POINTER(dev->priomap, NULL);
288		if (old)
289			kfree_rcu(old, rcu);
290		break;
291	}
292	return NOTIFY_DONE;
293}
294
295static struct notifier_block netprio_device_notifier = {
296	.notifier_call = netprio_device_event
297};
298
299static int __init init_cgroup_netprio(void)
300{
301	register_netdevice_notifier(&netprio_device_notifier);
302	return 0;
303}
304
305subsys_initcall(init_cgroup_netprio);
306MODULE_LICENSE("GPL v2");