blob: 9043caedcd083bf48836d0a970d48d8c533f86ee [file] [log] [blame]
Neil Horman5bc14212011-11-22 05:10:51 +00001/*
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
Joe Perchese005d192012-05-16 19:58:40 +000012#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13
Neil Horman5bc14212011-11-22 05:10:51 +000014#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
John Fastabend406a3c62012-07-20 10:39:25 +000028#include <linux/fdtable.h>
29
Tejun Heo4a6ee252012-11-22 07:32:46 -080030#define PRIOMAP_MIN_SZ 128
Neil Horman5bc14212011-11-22 05:10:51 +000031
Tejun Heo4a6ee252012-11-22 07:32:46 -080032/*
stephen hemminger8e3bff92013-12-08 12:15:44 -080033 * Extend @dev->priomap so that it's large enough to accommodate
Tejun Heo4a6ee252012-11-22 07:32:46 -080034 * @target_idx. @dev->priomap.priomap_len > @target_idx after successful
35 * return. Must be called under rtnl lock.
36 */
37static int extend_netdev_table(struct net_device *dev, u32 target_idx)
Neil Horman5bc14212011-11-22 05:10:51 +000038{
Tejun Heo4a6ee252012-11-22 07:32:46 -080039 struct netprio_map *old, *new;
40 size_t new_sz, new_len;
Neil Horman5bc14212011-11-22 05:10:51 +000041
Tejun Heo4a6ee252012-11-22 07:32:46 -080042 /* is the existing priomap large enough? */
Tejun Heo52bca932012-11-22 07:32:46 -080043 old = rtnl_dereference(dev->priomap);
Tejun Heo4a6ee252012-11-22 07:32:46 -080044 if (old && old->priomap_len > target_idx)
45 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +000046
Tejun Heo4a6ee252012-11-22 07:32:46 -080047 /*
48 * Determine the new size. Let's keep it power-of-two. We start
49 * from PRIOMAP_MIN_SZ and double it until it's large enough to
50 * accommodate @target_idx.
51 */
52 new_sz = PRIOMAP_MIN_SZ;
53 while (true) {
54 new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
55 sizeof(new->priomap[0]);
56 if (new_len > target_idx)
57 break;
58 new_sz *= 2;
59 /* overflowed? */
60 if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
61 return -ENOSPC;
62 }
63
64 /* allocate & copy */
65 new = kzalloc(new_sz, GFP_KERNEL);
Joe Perches62b59422013-02-04 16:48:16 +000066 if (!new)
Gao fengef209f12012-07-11 21:50:15 +000067 return -ENOMEM;
Neil Horman5bc14212011-11-22 05:10:51 +000068
Tejun Heo52bca932012-11-22 07:32:46 -080069 if (old)
70 memcpy(new->priomap, old->priomap,
71 old->priomap_len * sizeof(old->priomap[0]));
Neil Horman5bc14212011-11-22 05:10:51 +000072
Tejun Heo52bca932012-11-22 07:32:46 -080073 new->priomap_len = new_len;
Neil Horman5bc14212011-11-22 05:10:51 +000074
Tejun Heo4a6ee252012-11-22 07:32:46 -080075 /* install the new priomap */
Tejun Heo52bca932012-11-22 07:32:46 -080076 rcu_assign_pointer(dev->priomap, new);
77 if (old)
78 kfree_rcu(old, rcu);
Gao fengef209f12012-07-11 21:50:15 +000079 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +000080}
81
Tejun Heo666b0eb2012-11-22 07:32:47 -080082/**
83 * netprio_prio - return the effective netprio of a cgroup-net_device pair
Tejun Heo6d37b972013-08-08 20:11:22 -040084 * @css: css part of the target pair
Tejun Heo666b0eb2012-11-22 07:32:47 -080085 * @dev: net_device part of the target pair
86 *
87 * Should be called under RCU read or rtnl lock.
88 */
Tejun Heo6d37b972013-08-08 20:11:22 -040089static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
Tejun Heo666b0eb2012-11-22 07:32:47 -080090{
91 struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
Tejun Heo6d37b972013-08-08 20:11:22 -040092 int id = css->cgroup->id;
Tejun Heo666b0eb2012-11-22 07:32:47 -080093
Tejun Heo6d37b972013-08-08 20:11:22 -040094 if (map && id < map->priomap_len)
95 return map->priomap[id];
Tejun Heo666b0eb2012-11-22 07:32:47 -080096 return 0;
97}
98
99/**
100 * netprio_set_prio - set netprio on a cgroup-net_device pair
Tejun Heo6d37b972013-08-08 20:11:22 -0400101 * @css: css part of the target pair
Tejun Heo666b0eb2012-11-22 07:32:47 -0800102 * @dev: net_device part of the target pair
103 * @prio: prio to set
104 *
Tejun Heo6d37b972013-08-08 20:11:22 -0400105 * Set netprio to @prio on @css-@dev pair. Should be called under rtnl
Tejun Heo666b0eb2012-11-22 07:32:47 -0800106 * lock and may fail under memory pressure for non-zero @prio.
107 */
Tejun Heo6d37b972013-08-08 20:11:22 -0400108static int netprio_set_prio(struct cgroup_subsys_state *css,
109 struct net_device *dev, u32 prio)
Tejun Heo666b0eb2012-11-22 07:32:47 -0800110{
111 struct netprio_map *map;
Tejun Heo6d37b972013-08-08 20:11:22 -0400112 int id = css->cgroup->id;
Tejun Heo666b0eb2012-11-22 07:32:47 -0800113 int ret;
114
115 /* avoid extending priomap for zero writes */
116 map = rtnl_dereference(dev->priomap);
Tejun Heo6d37b972013-08-08 20:11:22 -0400117 if (!prio && (!map || map->priomap_len <= id))
Tejun Heo666b0eb2012-11-22 07:32:47 -0800118 return 0;
119
Tejun Heo6d37b972013-08-08 20:11:22 -0400120 ret = extend_netdev_table(dev, id);
Tejun Heo666b0eb2012-11-22 07:32:47 -0800121 if (ret)
122 return ret;
123
124 map = rtnl_dereference(dev->priomap);
Tejun Heo6d37b972013-08-08 20:11:22 -0400125 map->priomap[id] = prio;
Tejun Heo666b0eb2012-11-22 07:32:47 -0800126 return 0;
127}
128
Tejun Heoeb954192013-08-08 20:11:23 -0400129static struct cgroup_subsys_state *
130cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
Neil Horman5bc14212011-11-22 05:10:51 +0000131{
Tejun Heo6d37b972013-08-08 20:11:22 -0400132 struct cgroup_subsys_state *css;
Tejun Heo88d642f2012-11-22 07:32:47 -0800133
Tejun Heo6d37b972013-08-08 20:11:22 -0400134 css = kzalloc(sizeof(*css), GFP_KERNEL);
135 if (!css)
Neil Horman5bc14212011-11-22 05:10:51 +0000136 return ERR_PTR(-ENOMEM);
137
Tejun Heo6d37b972013-08-08 20:11:22 -0400138 return css;
Neil Horman5bc14212011-11-22 05:10:51 +0000139}
140
Tejun Heoeb954192013-08-08 20:11:23 -0400141static int cgrp_css_online(struct cgroup_subsys_state *css)
Neil Horman5bc14212011-11-22 05:10:51 +0000142{
Tejun Heoeb954192013-08-08 20:11:23 -0400143 struct cgroup_subsys_state *parent_css = css_parent(css);
Neil Horman5bc14212011-11-22 05:10:51 +0000144 struct net_device *dev;
Tejun Heo811d8d62012-11-22 07:32:47 -0800145 int ret = 0;
146
Tejun Heoeb954192013-08-08 20:11:23 -0400147 if (!parent_css)
Tejun Heo811d8d62012-11-22 07:32:47 -0800148 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +0000149
Neil Horman5bc14212011-11-22 05:10:51 +0000150 rtnl_lock();
Tejun Heo811d8d62012-11-22 07:32:47 -0800151 /*
152 * Inherit prios from the parent. As all prios are set during
153 * onlining, there is no need to clear them on offline.
154 */
155 for_each_netdev(&init_net, dev) {
Tejun Heo6d37b972013-08-08 20:11:22 -0400156 u32 prio = netprio_prio(parent_css, dev);
Tejun Heo811d8d62012-11-22 07:32:47 -0800157
Tejun Heo6d37b972013-08-08 20:11:22 -0400158 ret = netprio_set_prio(css, dev, prio);
Tejun Heo811d8d62012-11-22 07:32:47 -0800159 if (ret)
160 break;
161 }
Neil Horman5bc14212011-11-22 05:10:51 +0000162 rtnl_unlock();
Tejun Heo811d8d62012-11-22 07:32:47 -0800163 return ret;
164}
165
Tejun Heoeb954192013-08-08 20:11:23 -0400166static void cgrp_css_free(struct cgroup_subsys_state *css)
Tejun Heo811d8d62012-11-22 07:32:47 -0800167{
Tejun Heoeb954192013-08-08 20:11:23 -0400168 kfree(css);
Neil Horman5bc14212011-11-22 05:10:51 +0000169}
170
Tejun Heo182446d2013-08-08 20:11:24 -0400171static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
Neil Horman5bc14212011-11-22 05:10:51 +0000172{
Tejun Heo182446d2013-08-08 20:11:24 -0400173 return css->cgroup->id;
Neil Horman5bc14212011-11-22 05:10:51 +0000174}
175
Tejun Heo2da8ca82013-12-05 12:28:04 -0500176static int read_priomap(struct seq_file *sf, void *v)
Neil Horman5bc14212011-11-22 05:10:51 +0000177{
178 struct net_device *dev;
Neil Horman5bc14212011-11-22 05:10:51 +0000179
180 rcu_read_lock();
Tejun Heo666b0eb2012-11-22 07:32:47 -0800181 for_each_netdev_rcu(&init_net, dev)
Tejun Heo2da8ca82013-12-05 12:28:04 -0500182 seq_printf(sf, "%s %u\n", dev->name,
183 netprio_prio(seq_css(sf), dev));
Neil Horman5bc14212011-11-22 05:10:51 +0000184 rcu_read_unlock();
185 return 0;
186}
187
Tejun Heo182446d2013-08-08 20:11:24 -0400188static int write_priomap(struct cgroup_subsys_state *css, struct cftype *cft,
Neil Horman5bc14212011-11-22 05:10:51 +0000189 const char *buffer)
190{
Tejun Heo6d5759d2012-11-22 07:32:46 -0800191 char devname[IFNAMSIZ + 1];
Neil Horman5bc14212011-11-22 05:10:51 +0000192 struct net_device *dev;
Tejun Heo6d5759d2012-11-22 07:32:46 -0800193 u32 prio;
194 int ret;
Neil Horman5bc14212011-11-22 05:10:51 +0000195
Tejun Heo6d5759d2012-11-22 07:32:46 -0800196 if (sscanf(buffer, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
197 return -EINVAL;
Neil Horman5bc14212011-11-22 05:10:51 +0000198
199 dev = dev_get_by_name(&init_net, devname);
200 if (!dev)
Tejun Heo6d5759d2012-11-22 07:32:46 -0800201 return -ENODEV;
Neil Horman5bc14212011-11-22 05:10:51 +0000202
John Fastabend476ad152012-08-14 12:34:35 +0000203 rtnl_lock();
Tejun Heo6d5759d2012-11-22 07:32:46 -0800204
Tejun Heo6d37b972013-08-08 20:11:22 -0400205 ret = netprio_set_prio(css, dev, prio);
Gao fengef209f12012-07-11 21:50:15 +0000206
John Fastabend476ad152012-08-14 12:34:35 +0000207 rtnl_unlock();
Neil Horman5bc14212011-11-22 05:10:51 +0000208 dev_put(dev);
Neil Horman5bc14212011-11-22 05:10:51 +0000209 return ret;
210}
211
Al Viroc3c073f2012-08-21 22:32:06 -0400212static int update_netprio(const void *v, struct file *file, unsigned n)
213{
214 int err;
215 struct socket *sock = sock_from_file(file, &err);
216 if (sock)
217 sock->sk->sk_cgrp_prioidx = (u32)(unsigned long)v;
218 return 0;
219}
220
Tejun Heoeb954192013-08-08 20:11:23 -0400221static void net_prio_attach(struct cgroup_subsys_state *css,
222 struct cgroup_taskset *tset)
John Fastabend406a3c62012-07-20 10:39:25 +0000223{
224 struct task_struct *p;
Gao fenge1af5e42013-10-08 11:05:19 +0800225 void *v = (void *)(unsigned long)css->cgroup->id;
John Fastabend406a3c62012-07-20 10:39:25 +0000226
Tejun Heod99c8722013-08-08 20:11:27 -0400227 cgroup_taskset_for_each(p, css, tset) {
John Fastabend406a3c62012-07-20 10:39:25 +0000228 task_lock(p);
Al Viroc3c073f2012-08-21 22:32:06 -0400229 iterate_fd(p->files, 0, update_netprio, v);
John Fastabend406a3c62012-07-20 10:39:25 +0000230 task_unlock(p);
231 }
John Fastabend406a3c62012-07-20 10:39:25 +0000232}
233
Neil Horman5bc14212011-11-22 05:10:51 +0000234static struct cftype ss_files[] = {
235 {
236 .name = "prioidx",
237 .read_u64 = read_prioidx,
238 },
239 {
240 .name = "ifpriomap",
Tejun Heo2da8ca82013-12-05 12:28:04 -0500241 .seq_show = read_priomap,
Neil Horman5bc14212011-11-22 05:10:51 +0000242 .write_string = write_priomap,
243 },
Tejun Heo4baf6e32012-04-01 12:09:55 -0700244 { } /* terminate */
Neil Horman5bc14212011-11-22 05:10:51 +0000245};
246
Tejun Heo676f7c82012-04-01 12:09:55 -0700247struct cgroup_subsys net_prio_subsys = {
248 .name = "net_prio",
Tejun Heo92fb9742012-11-19 08:13:38 -0800249 .css_alloc = cgrp_css_alloc,
Tejun Heo811d8d62012-11-22 07:32:47 -0800250 .css_online = cgrp_css_online,
Tejun Heo92fb9742012-11-19 08:13:38 -0800251 .css_free = cgrp_css_free,
John Fastabend406a3c62012-07-20 10:39:25 +0000252 .attach = net_prio_attach,
Tejun Heo676f7c82012-04-01 12:09:55 -0700253 .subsys_id = net_prio_subsys_id,
Tejun Heo4baf6e32012-04-01 12:09:55 -0700254 .base_cftypes = ss_files,
Tejun Heo8c7f6ed2012-09-13 12:20:58 -0700255 .module = THIS_MODULE,
Tejun Heo676f7c82012-04-01 12:09:55 -0700256};
Neil Horman5bc14212011-11-22 05:10:51 +0000257
258static int netprio_device_event(struct notifier_block *unused,
259 unsigned long event, void *ptr)
260{
Jiri Pirko351638e2013-05-28 01:30:21 +0000261 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
Neil Horman5bc14212011-11-22 05:10:51 +0000262 struct netprio_map *old;
Neil Horman5bc14212011-11-22 05:10:51 +0000263
264 /*
265 * Note this is called with rtnl_lock held so we have update side
266 * protection on our rcu assignments
267 */
268
269 switch (event) {
Neil Horman5bc14212011-11-22 05:10:51 +0000270 case NETDEV_UNREGISTER:
271 old = rtnl_dereference(dev->priomap);
Eric Dumazet2cfa5a02011-11-23 07:09:32 +0000272 RCU_INIT_POINTER(dev->priomap, NULL);
Neil Horman5bc14212011-11-22 05:10:51 +0000273 if (old)
274 kfree_rcu(old, rcu);
275 break;
276 }
277 return NOTIFY_DONE;
278}
279
280static struct notifier_block netprio_device_notifier = {
281 .notifier_call = netprio_device_event
282};
283
284static int __init init_cgroup_netprio(void)
285{
286 int ret;
287
288 ret = cgroup_load_subsys(&net_prio_subsys);
289 if (ret)
290 goto out;
Neil Horman5bc14212011-11-22 05:10:51 +0000291
292 register_netdevice_notifier(&netprio_device_notifier);
293
294out:
295 return ret;
296}
297
298static void __exit exit_cgroup_netprio(void)
299{
300 struct netprio_map *old;
301 struct net_device *dev;
302
303 unregister_netdevice_notifier(&netprio_device_notifier);
304
305 cgroup_unload_subsys(&net_prio_subsys);
306
Neil Horman5bc14212011-11-22 05:10:51 +0000307 rtnl_lock();
308 for_each_netdev(&init_net, dev) {
309 old = rtnl_dereference(dev->priomap);
Eric Dumazet2cfa5a02011-11-23 07:09:32 +0000310 RCU_INIT_POINTER(dev->priomap, NULL);
Neil Horman5bc14212011-11-22 05:10:51 +0000311 if (old)
312 kfree_rcu(old, rcu);
313 }
314 rtnl_unlock();
315}
316
317module_init(init_cgroup_netprio);
318module_exit(exit_cgroup_netprio);
319MODULE_LICENSE("GPL v2");