Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * net/core/netprio_cgroup.c Priority Control Group
4 *
5 * Authors: Neil Horman <nhorman@tuxdriver.com>
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/module.h>
11#include <linux/slab.h>
12#include <linux/types.h>
13#include <linux/string.h>
14#include <linux/errno.h>
15#include <linux/skbuff.h>
16#include <linux/cgroup.h>
17#include <linux/rcupdate.h>
18#include <linux/atomic.h>
19#include <linux/sched/task.h>
20
21#include <net/rtnetlink.h>
22#include <net/pkt_cls.h>
23#include <net/sock.h>
24#include <net/netprio_cgroup.h>
25
26#include <linux/fdtable.h>
27
28/*
29 * netprio allocates per-net_device priomap array which is indexed by
30 * css->id. Limiting css ID to 16bits doesn't lose anything.
31 */
32#define NETPRIO_ID_MAX USHRT_MAX
33
34#define PRIOMAP_MIN_SZ 128
35
36/*
37 * Extend @dev->priomap so that it's large enough to accommodate
38 * @target_idx. @dev->priomap.priomap_len > @target_idx after successful
39 * return. Must be called under rtnl lock.
40 */
41static int extend_netdev_table(struct net_device *dev, u32 target_idx)
42{
43 struct netprio_map *old, *new;
44 size_t new_sz, new_len;
45
46 /* is the existing priomap large enough? */
47 old = rtnl_dereference(dev->priomap);
48 if (old && old->priomap_len > target_idx)
49 return 0;
50
51 /*
52 * Determine the new size. Let's keep it power-of-two. We start
53 * from PRIOMAP_MIN_SZ and double it until it's large enough to
54 * accommodate @target_idx.
55 */
56 new_sz = PRIOMAP_MIN_SZ;
57 while (true) {
58 new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
59 sizeof(new->priomap[0]);
60 if (new_len > target_idx)
61 break;
62 new_sz *= 2;
63 /* overflowed? */
64 if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
65 return -ENOSPC;
66 }
67
68 /* allocate & copy */
69 new = kzalloc(new_sz, GFP_KERNEL);
70 if (!new)
71 return -ENOMEM;
72
73 if (old)
74 memcpy(new->priomap, old->priomap,
75 old->priomap_len * sizeof(old->priomap[0]));
76
77 new->priomap_len = new_len;
78
79 /* install the new priomap */
80 rcu_assign_pointer(dev->priomap, new);
81 if (old)
82 kfree_rcu(old, rcu);
83 return 0;
84}
85
86/**
87 * netprio_prio - return the effective netprio of a cgroup-net_device pair
88 * @css: css part of the target pair
89 * @dev: net_device part of the target pair
90 *
91 * Should be called under RCU read or rtnl lock.
92 */
93static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
94{
95 struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
96 int id = css->id;
97
98 if (map && id < map->priomap_len)
99 return map->priomap[id];
100 return 0;
101}
102
103/**
104 * netprio_set_prio - set netprio on a cgroup-net_device pair
105 * @css: css part of the target pair
106 * @dev: net_device part of the target pair
107 * @prio: prio to set
108 *
109 * Set netprio to @prio on @css-@dev pair. Should be called under rtnl
110 * lock and may fail under memory pressure for non-zero @prio.
111 */
112static int netprio_set_prio(struct cgroup_subsys_state *css,
113 struct net_device *dev, u32 prio)
114{
115 struct netprio_map *map;
116 int id = css->id;
117 int ret;
118
119 /* avoid extending priomap for zero writes */
120 map = rtnl_dereference(dev->priomap);
121 if (!prio && (!map || map->priomap_len <= id))
122 return 0;
123
124 ret = extend_netdev_table(dev, id);
125 if (ret)
126 return ret;
127
128 map = rtnl_dereference(dev->priomap);
129 map->priomap[id] = prio;
130 return 0;
131}
132
133static struct cgroup_subsys_state *
134cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
135{
136 struct cgroup_subsys_state *css;
137
138 css = kzalloc(sizeof(*css), GFP_KERNEL);
139 if (!css)
140 return ERR_PTR(-ENOMEM);
141
142 return css;
143}
144
145static int cgrp_css_online(struct cgroup_subsys_state *css)
146{
147 struct cgroup_subsys_state *parent_css = css->parent;
148 struct net_device *dev;
149 int ret = 0;
150
151 if (css->id > NETPRIO_ID_MAX)
152 return -ENOSPC;
153
154 if (!parent_css)
155 return 0;
156
157 rtnl_lock();
158 /*
159 * Inherit prios from the parent. As all prios are set during
160 * onlining, there is no need to clear them on offline.
161 */
162 for_each_netdev(&init_net, dev) {
163 u32 prio = netprio_prio(parent_css, dev);
164
165 ret = netprio_set_prio(css, dev, prio);
166 if (ret)
167 break;
168 }
169 rtnl_unlock();
170 return ret;
171}
172
173static void cgrp_css_free(struct cgroup_subsys_state *css)
174{
175 kfree(css);
176}
177
178static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
179{
180 return css->id;
181}
182
183static int read_priomap(struct seq_file *sf, void *v)
184{
185 struct net_device *dev;
186
187 rcu_read_lock();
188 for_each_netdev_rcu(&init_net, dev)
189 seq_printf(sf, "%s %u\n", dev->name,
190 netprio_prio(seq_css(sf), dev));
191 rcu_read_unlock();
192 return 0;
193}
194
195static ssize_t write_priomap(struct kernfs_open_file *of,
196 char *buf, size_t nbytes, loff_t off)
197{
198 char devname[IFNAMSIZ + 1];
199 struct net_device *dev;
200 u32 prio;
201 int ret;
202
203 if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
204 return -EINVAL;
205
206 dev = dev_get_by_name(&init_net, devname);
207 if (!dev)
208 return -ENODEV;
209
210 cgroup_sk_alloc_disable();
211
212 rtnl_lock();
213
214 ret = netprio_set_prio(of_css(of), dev, prio);
215
216 rtnl_unlock();
217 dev_put(dev);
218 return ret ?: nbytes;
219}
220
221static int update_netprio(const void *v, struct file *file, unsigned n)
222{
223 struct socket *sock = sock_from_file(file);
224 if (sock) {
225 spin_lock(&cgroup_sk_update_lock);
226 sock_cgroup_set_prioidx(&sock->sk->sk_cgrp_data,
227 (unsigned long)v);
228 spin_unlock(&cgroup_sk_update_lock);
229 }
230 return 0;
231}
232
233static void net_prio_attach(struct cgroup_taskset *tset)
234{
235 struct task_struct *p;
236 struct cgroup_subsys_state *css;
237
238 cgroup_sk_alloc_disable();
239
240 cgroup_taskset_for_each(p, css, tset) {
241 void *v = (void *)(unsigned long)css->id;
242
243 task_lock(p);
244 iterate_fd(p->files, 0, update_netprio, v);
245 task_unlock(p);
246 }
247}
248
249static struct cftype ss_files[] = {
250 {
251 .name = "prioidx",
252 .read_u64 = read_prioidx,
253 },
254 {
255 .name = "ifpriomap",
256 .seq_show = read_priomap,
257 .write = write_priomap,
258 },
259 { } /* terminate */
260};
261
262struct cgroup_subsys net_prio_cgrp_subsys = {
263 .css_alloc = cgrp_css_alloc,
264 .css_online = cgrp_css_online,
265 .css_free = cgrp_css_free,
266 .attach = net_prio_attach,
267 .legacy_cftypes = ss_files,
268};
269
270static int netprio_device_event(struct notifier_block *unused,
271 unsigned long event, void *ptr)
272{
273 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
274 struct netprio_map *old;
275
276 /*
277 * Note this is called with rtnl_lock held so we have update side
278 * protection on our rcu assignments
279 */
280
281 switch (event) {
282 case NETDEV_UNREGISTER:
283 old = rtnl_dereference(dev->priomap);
284 RCU_INIT_POINTER(dev->priomap, NULL);
285 if (old)
286 kfree_rcu(old, rcu);
287 break;
288 }
289 return NOTIFY_DONE;
290}
291
292static struct notifier_block netprio_device_notifier = {
293 .notifier_call = netprio_device_event
294};
295
296static int __init init_cgroup_netprio(void)
297{
298 register_netdevice_notifier(&netprio_device_notifier);
299 return 0;
300}
301subsys_initcall(init_cgroup_netprio);
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 <net/rtnetlink.h>
24#include <net/pkt_cls.h>
25#include <net/sock.h>
26#include <net/netprio_cgroup.h>
27
28#define PRIOIDX_SZ 128
29
30static unsigned long prioidx_map[PRIOIDX_SZ];
31static DEFINE_SPINLOCK(prioidx_map_lock);
32static atomic_t max_prioidx = ATOMIC_INIT(0);
33
34static inline struct cgroup_netprio_state *cgrp_netprio_state(struct cgroup *cgrp)
35{
36 return container_of(cgroup_subsys_state(cgrp, net_prio_subsys_id),
37 struct cgroup_netprio_state, css);
38}
39
40static int get_prioidx(u32 *prio)
41{
42 unsigned long flags;
43 u32 prioidx;
44
45 spin_lock_irqsave(&prioidx_map_lock, flags);
46 prioidx = find_first_zero_bit(prioidx_map, sizeof(unsigned long) * PRIOIDX_SZ);
47 if (prioidx == sizeof(unsigned long) * PRIOIDX_SZ) {
48 spin_unlock_irqrestore(&prioidx_map_lock, flags);
49 return -ENOSPC;
50 }
51 set_bit(prioidx, prioidx_map);
52 if (atomic_read(&max_prioidx) < prioidx)
53 atomic_set(&max_prioidx, prioidx);
54 spin_unlock_irqrestore(&prioidx_map_lock, flags);
55 *prio = prioidx;
56 return 0;
57}
58
59static void put_prioidx(u32 idx)
60{
61 unsigned long flags;
62
63 spin_lock_irqsave(&prioidx_map_lock, flags);
64 clear_bit(idx, prioidx_map);
65 spin_unlock_irqrestore(&prioidx_map_lock, flags);
66}
67
68static int extend_netdev_table(struct net_device *dev, u32 new_len)
69{
70 size_t new_size = sizeof(struct netprio_map) +
71 ((sizeof(u32) * new_len));
72 struct netprio_map *new_priomap = kzalloc(new_size, GFP_KERNEL);
73 struct netprio_map *old_priomap;
74 int i;
75
76 old_priomap = rtnl_dereference(dev->priomap);
77
78 if (!new_priomap) {
79 pr_warn("Unable to alloc new priomap!\n");
80 return -ENOMEM;
81 }
82
83 for (i = 0;
84 old_priomap && (i < old_priomap->priomap_len);
85 i++)
86 new_priomap->priomap[i] = old_priomap->priomap[i];
87
88 new_priomap->priomap_len = new_len;
89
90 rcu_assign_pointer(dev->priomap, new_priomap);
91 if (old_priomap)
92 kfree_rcu(old_priomap, rcu);
93 return 0;
94}
95
96static int write_update_netdev_table(struct net_device *dev)
97{
98 int ret = 0;
99 u32 max_len;
100 struct netprio_map *map;
101
102 rtnl_lock();
103 max_len = atomic_read(&max_prioidx) + 1;
104 map = rtnl_dereference(dev->priomap);
105 if (!map || map->priomap_len < max_len)
106 ret = extend_netdev_table(dev, max_len);
107 rtnl_unlock();
108
109 return ret;
110}
111
112static int update_netdev_tables(void)
113{
114 int ret = 0;
115 struct net_device *dev;
116 u32 max_len;
117 struct netprio_map *map;
118
119 rtnl_lock();
120 max_len = atomic_read(&max_prioidx) + 1;
121 for_each_netdev(&init_net, dev) {
122 map = rtnl_dereference(dev->priomap);
123 /*
124 * don't allocate priomap if we didn't
125 * change net_prio.ifpriomap (map == NULL),
126 * this will speed up skb_update_prio.
127 */
128 if (map && map->priomap_len < max_len) {
129 ret = extend_netdev_table(dev, max_len);
130 if (ret < 0)
131 break;
132 }
133 }
134 rtnl_unlock();
135 return ret;
136}
137
138static struct cgroup_subsys_state *cgrp_create(struct cgroup *cgrp)
139{
140 struct cgroup_netprio_state *cs;
141 int ret = -EINVAL;
142
143 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
144 if (!cs)
145 return ERR_PTR(-ENOMEM);
146
147 if (cgrp->parent && cgrp_netprio_state(cgrp->parent)->prioidx)
148 goto out;
149
150 ret = get_prioidx(&cs->prioidx);
151 if (ret < 0) {
152 pr_warn("No space in priority index array\n");
153 goto out;
154 }
155
156 ret = update_netdev_tables();
157 if (ret < 0) {
158 put_prioidx(cs->prioidx);
159 goto out;
160 }
161
162 return &cs->css;
163out:
164 kfree(cs);
165 return ERR_PTR(ret);
166}
167
168static void cgrp_destroy(struct cgroup *cgrp)
169{
170 struct cgroup_netprio_state *cs;
171 struct net_device *dev;
172 struct netprio_map *map;
173
174 cs = cgrp_netprio_state(cgrp);
175 rtnl_lock();
176 for_each_netdev(&init_net, dev) {
177 map = rtnl_dereference(dev->priomap);
178 if (map && cs->prioidx < map->priomap_len)
179 map->priomap[cs->prioidx] = 0;
180 }
181 rtnl_unlock();
182 put_prioidx(cs->prioidx);
183 kfree(cs);
184}
185
186static u64 read_prioidx(struct cgroup *cgrp, struct cftype *cft)
187{
188 return (u64)cgrp_netprio_state(cgrp)->prioidx;
189}
190
191static int read_priomap(struct cgroup *cont, struct cftype *cft,
192 struct cgroup_map_cb *cb)
193{
194 struct net_device *dev;
195 u32 prioidx = cgrp_netprio_state(cont)->prioidx;
196 u32 priority;
197 struct netprio_map *map;
198
199 rcu_read_lock();
200 for_each_netdev_rcu(&init_net, dev) {
201 map = rcu_dereference(dev->priomap);
202 priority = (map && prioidx < map->priomap_len) ? map->priomap[prioidx] : 0;
203 cb->fill(cb, dev->name, priority);
204 }
205 rcu_read_unlock();
206 return 0;
207}
208
209static int write_priomap(struct cgroup *cgrp, struct cftype *cft,
210 const char *buffer)
211{
212 char *devname = kstrdup(buffer, GFP_KERNEL);
213 int ret = -EINVAL;
214 u32 prioidx = cgrp_netprio_state(cgrp)->prioidx;
215 unsigned long priority;
216 char *priostr;
217 struct net_device *dev;
218 struct netprio_map *map;
219
220 if (!devname)
221 return -ENOMEM;
222
223 /*
224 * Minimally sized valid priomap string
225 */
226 if (strlen(devname) < 3)
227 goto out_free_devname;
228
229 priostr = strstr(devname, " ");
230 if (!priostr)
231 goto out_free_devname;
232
233 /*
234 *Separate the devname from the associated priority
235 *and advance the priostr poitner to the priority value
236 */
237 *priostr = '\0';
238 priostr++;
239
240 /*
241 * If the priostr points to NULL, we're at the end of the passed
242 * in string, and its not a valid write
243 */
244 if (*priostr == '\0')
245 goto out_free_devname;
246
247 ret = kstrtoul(priostr, 10, &priority);
248 if (ret < 0)
249 goto out_free_devname;
250
251 ret = -ENODEV;
252
253 dev = dev_get_by_name(&init_net, devname);
254 if (!dev)
255 goto out_free_devname;
256
257 ret = write_update_netdev_table(dev);
258 if (ret < 0)
259 goto out_put_dev;
260
261 rcu_read_lock();
262 map = rcu_dereference(dev->priomap);
263 if (map)
264 map->priomap[prioidx] = priority;
265 rcu_read_unlock();
266
267out_put_dev:
268 dev_put(dev);
269
270out_free_devname:
271 kfree(devname);
272 return ret;
273}
274
275static struct cftype ss_files[] = {
276 {
277 .name = "prioidx",
278 .read_u64 = read_prioidx,
279 },
280 {
281 .name = "ifpriomap",
282 .read_map = read_priomap,
283 .write_string = write_priomap,
284 },
285 { } /* terminate */
286};
287
288struct cgroup_subsys net_prio_subsys = {
289 .name = "net_prio",
290 .create = cgrp_create,
291 .destroy = cgrp_destroy,
292#ifdef CONFIG_NETPRIO_CGROUP
293 .subsys_id = net_prio_subsys_id,
294#endif
295 .base_cftypes = ss_files,
296 .module = THIS_MODULE
297};
298
299static int netprio_device_event(struct notifier_block *unused,
300 unsigned long event, void *ptr)
301{
302 struct net_device *dev = ptr;
303 struct netprio_map *old;
304
305 /*
306 * Note this is called with rtnl_lock held so we have update side
307 * protection on our rcu assignments
308 */
309
310 switch (event) {
311 case NETDEV_UNREGISTER:
312 old = rtnl_dereference(dev->priomap);
313 RCU_INIT_POINTER(dev->priomap, NULL);
314 if (old)
315 kfree_rcu(old, rcu);
316 break;
317 }
318 return NOTIFY_DONE;
319}
320
321static struct notifier_block netprio_device_notifier = {
322 .notifier_call = netprio_device_event
323};
324
325static int __init init_cgroup_netprio(void)
326{
327 int ret;
328
329 ret = cgroup_load_subsys(&net_prio_subsys);
330 if (ret)
331 goto out;
332#ifndef CONFIG_NETPRIO_CGROUP
333 smp_wmb();
334 net_prio_subsys_id = net_prio_subsys.subsys_id;
335#endif
336
337 register_netdevice_notifier(&netprio_device_notifier);
338
339out:
340 return ret;
341}
342
343static void __exit exit_cgroup_netprio(void)
344{
345 struct netprio_map *old;
346 struct net_device *dev;
347
348 unregister_netdevice_notifier(&netprio_device_notifier);
349
350 cgroup_unload_subsys(&net_prio_subsys);
351
352#ifndef CONFIG_NETPRIO_CGROUP
353 net_prio_subsys_id = -1;
354 synchronize_rcu();
355#endif
356
357 rtnl_lock();
358 for_each_netdev(&init_net, dev) {
359 old = rtnl_dereference(dev->priomap);
360 RCU_INIT_POINTER(dev->priomap, NULL);
361 if (old)
362 kfree_rcu(old, rcu);
363 }
364 rtnl_unlock();
365}
366
367module_init(init_cgroup_netprio);
368module_exit(exit_cgroup_netprio);
369MODULE_LICENSE("GPL v2");