blob: 2b9159b7a28a43f1772963a9df361f49cc7f6a63 [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 Heo297dbde2015-12-07 17:38:51 -050030/*
31 * netprio allocates per-net_device priomap array which is indexed by
32 * css->id. Limiting css ID to 16bits doesn't lose anything.
33 */
34#define NETPRIO_ID_MAX USHRT_MAX
35
Tejun Heo4a6ee252012-11-22 07:32:46 -080036#define PRIOMAP_MIN_SZ 128
Neil Horman5bc14212011-11-22 05:10:51 +000037
Tejun Heo4a6ee252012-11-22 07:32:46 -080038/*
stephen hemminger8e3bff92013-12-08 12:15:44 -080039 * Extend @dev->priomap so that it's large enough to accommodate
Tejun Heo4a6ee252012-11-22 07:32:46 -080040 * @target_idx. @dev->priomap.priomap_len > @target_idx after successful
41 * return. Must be called under rtnl lock.
42 */
43static int extend_netdev_table(struct net_device *dev, u32 target_idx)
Neil Horman5bc14212011-11-22 05:10:51 +000044{
Tejun Heo4a6ee252012-11-22 07:32:46 -080045 struct netprio_map *old, *new;
46 size_t new_sz, new_len;
Neil Horman5bc14212011-11-22 05:10:51 +000047
Tejun Heo4a6ee252012-11-22 07:32:46 -080048 /* is the existing priomap large enough? */
Tejun Heo52bca932012-11-22 07:32:46 -080049 old = rtnl_dereference(dev->priomap);
Tejun Heo4a6ee252012-11-22 07:32:46 -080050 if (old && old->priomap_len > target_idx)
51 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +000052
Tejun Heo4a6ee252012-11-22 07:32:46 -080053 /*
54 * Determine the new size. Let's keep it power-of-two. We start
55 * from PRIOMAP_MIN_SZ and double it until it's large enough to
56 * accommodate @target_idx.
57 */
58 new_sz = PRIOMAP_MIN_SZ;
59 while (true) {
60 new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
61 sizeof(new->priomap[0]);
62 if (new_len > target_idx)
63 break;
64 new_sz *= 2;
65 /* overflowed? */
66 if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
67 return -ENOSPC;
68 }
69
70 /* allocate & copy */
71 new = kzalloc(new_sz, GFP_KERNEL);
Joe Perches62b59422013-02-04 16:48:16 +000072 if (!new)
Gao fengef209f12012-07-11 21:50:15 +000073 return -ENOMEM;
Neil Horman5bc14212011-11-22 05:10:51 +000074
Tejun Heo52bca932012-11-22 07:32:46 -080075 if (old)
76 memcpy(new->priomap, old->priomap,
77 old->priomap_len * sizeof(old->priomap[0]));
Neil Horman5bc14212011-11-22 05:10:51 +000078
Tejun Heo52bca932012-11-22 07:32:46 -080079 new->priomap_len = new_len;
Neil Horman5bc14212011-11-22 05:10:51 +000080
Tejun Heo4a6ee252012-11-22 07:32:46 -080081 /* install the new priomap */
Tejun Heo52bca932012-11-22 07:32:46 -080082 rcu_assign_pointer(dev->priomap, new);
83 if (old)
84 kfree_rcu(old, rcu);
Gao fengef209f12012-07-11 21:50:15 +000085 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +000086}
87
Tejun Heo666b0eb2012-11-22 07:32:47 -080088/**
89 * netprio_prio - return the effective netprio of a cgroup-net_device pair
Tejun Heo6d37b972013-08-08 20:11:22 -040090 * @css: css part of the target pair
Tejun Heo666b0eb2012-11-22 07:32:47 -080091 * @dev: net_device part of the target pair
92 *
93 * Should be called under RCU read or rtnl lock.
94 */
Tejun Heo6d37b972013-08-08 20:11:22 -040095static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
Tejun Heo666b0eb2012-11-22 07:32:47 -080096{
97 struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
Tejun Heo6d37b972013-08-08 20:11:22 -040098 int id = css->cgroup->id;
Tejun Heo666b0eb2012-11-22 07:32:47 -080099
Tejun Heo6d37b972013-08-08 20:11:22 -0400100 if (map && id < map->priomap_len)
101 return map->priomap[id];
Tejun Heo666b0eb2012-11-22 07:32:47 -0800102 return 0;
103}
104
105/**
106 * netprio_set_prio - set netprio on a cgroup-net_device pair
Tejun Heo6d37b972013-08-08 20:11:22 -0400107 * @css: css part of the target pair
Tejun Heo666b0eb2012-11-22 07:32:47 -0800108 * @dev: net_device part of the target pair
109 * @prio: prio to set
110 *
Tejun Heo6d37b972013-08-08 20:11:22 -0400111 * Set netprio to @prio on @css-@dev pair. Should be called under rtnl
Tejun Heo666b0eb2012-11-22 07:32:47 -0800112 * lock and may fail under memory pressure for non-zero @prio.
113 */
Tejun Heo6d37b972013-08-08 20:11:22 -0400114static int netprio_set_prio(struct cgroup_subsys_state *css,
115 struct net_device *dev, u32 prio)
Tejun Heo666b0eb2012-11-22 07:32:47 -0800116{
117 struct netprio_map *map;
Tejun Heo6d37b972013-08-08 20:11:22 -0400118 int id = css->cgroup->id;
Tejun Heo666b0eb2012-11-22 07:32:47 -0800119 int ret;
120
121 /* avoid extending priomap for zero writes */
122 map = rtnl_dereference(dev->priomap);
Tejun Heo6d37b972013-08-08 20:11:22 -0400123 if (!prio && (!map || map->priomap_len <= id))
Tejun Heo666b0eb2012-11-22 07:32:47 -0800124 return 0;
125
Tejun Heo6d37b972013-08-08 20:11:22 -0400126 ret = extend_netdev_table(dev, id);
Tejun Heo666b0eb2012-11-22 07:32:47 -0800127 if (ret)
128 return ret;
129
130 map = rtnl_dereference(dev->priomap);
Tejun Heo6d37b972013-08-08 20:11:22 -0400131 map->priomap[id] = prio;
Tejun Heo666b0eb2012-11-22 07:32:47 -0800132 return 0;
133}
134
Tejun Heoeb954192013-08-08 20:11:23 -0400135static struct cgroup_subsys_state *
136cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
Neil Horman5bc14212011-11-22 05:10:51 +0000137{
Tejun Heo6d37b972013-08-08 20:11:22 -0400138 struct cgroup_subsys_state *css;
Tejun Heo88d642f2012-11-22 07:32:47 -0800139
Tejun Heo6d37b972013-08-08 20:11:22 -0400140 css = kzalloc(sizeof(*css), GFP_KERNEL);
141 if (!css)
Neil Horman5bc14212011-11-22 05:10:51 +0000142 return ERR_PTR(-ENOMEM);
143
Tejun Heo6d37b972013-08-08 20:11:22 -0400144 return css;
Neil Horman5bc14212011-11-22 05:10:51 +0000145}
146
Tejun Heoeb954192013-08-08 20:11:23 -0400147static int cgrp_css_online(struct cgroup_subsys_state *css)
Neil Horman5bc14212011-11-22 05:10:51 +0000148{
Tejun Heo5c9d5352014-05-16 13:22:48 -0400149 struct cgroup_subsys_state *parent_css = css->parent;
Neil Horman5bc14212011-11-22 05:10:51 +0000150 struct net_device *dev;
Tejun Heo811d8d62012-11-22 07:32:47 -0800151 int ret = 0;
152
Tejun Heo297dbde2015-12-07 17:38:51 -0500153 if (css->id > NETPRIO_ID_MAX)
154 return -ENOSPC;
155
Tejun Heoeb954192013-08-08 20:11:23 -0400156 if (!parent_css)
Tejun Heo811d8d62012-11-22 07:32:47 -0800157 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +0000158
Neil Horman5bc14212011-11-22 05:10:51 +0000159 rtnl_lock();
Tejun Heo811d8d62012-11-22 07:32:47 -0800160 /*
161 * Inherit prios from the parent. As all prios are set during
162 * onlining, there is no need to clear them on offline.
163 */
164 for_each_netdev(&init_net, dev) {
Tejun Heo6d37b972013-08-08 20:11:22 -0400165 u32 prio = netprio_prio(parent_css, dev);
Tejun Heo811d8d62012-11-22 07:32:47 -0800166
Tejun Heo6d37b972013-08-08 20:11:22 -0400167 ret = netprio_set_prio(css, dev, prio);
Tejun Heo811d8d62012-11-22 07:32:47 -0800168 if (ret)
169 break;
170 }
Neil Horman5bc14212011-11-22 05:10:51 +0000171 rtnl_unlock();
Tejun Heo811d8d62012-11-22 07:32:47 -0800172 return ret;
173}
174
Tejun Heoeb954192013-08-08 20:11:23 -0400175static void cgrp_css_free(struct cgroup_subsys_state *css)
Tejun Heo811d8d62012-11-22 07:32:47 -0800176{
Tejun Heoeb954192013-08-08 20:11:23 -0400177 kfree(css);
Neil Horman5bc14212011-11-22 05:10:51 +0000178}
179
Tejun Heo182446d2013-08-08 20:11:24 -0400180static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
Neil Horman5bc14212011-11-22 05:10:51 +0000181{
Tejun Heo182446d2013-08-08 20:11:24 -0400182 return css->cgroup->id;
Neil Horman5bc14212011-11-22 05:10:51 +0000183}
184
Tejun Heo2da8ca82013-12-05 12:28:04 -0500185static int read_priomap(struct seq_file *sf, void *v)
Neil Horman5bc14212011-11-22 05:10:51 +0000186{
187 struct net_device *dev;
Neil Horman5bc14212011-11-22 05:10:51 +0000188
189 rcu_read_lock();
Tejun Heo666b0eb2012-11-22 07:32:47 -0800190 for_each_netdev_rcu(&init_net, dev)
Tejun Heo2da8ca82013-12-05 12:28:04 -0500191 seq_printf(sf, "%s %u\n", dev->name,
192 netprio_prio(seq_css(sf), dev));
Neil Horman5bc14212011-11-22 05:10:51 +0000193 rcu_read_unlock();
194 return 0;
195}
196
Tejun Heo451af502014-05-13 12:16:21 -0400197static ssize_t write_priomap(struct kernfs_open_file *of,
198 char *buf, size_t nbytes, loff_t off)
Neil Horman5bc14212011-11-22 05:10:51 +0000199{
Tejun Heo6d5759d2012-11-22 07:32:46 -0800200 char devname[IFNAMSIZ + 1];
Neil Horman5bc14212011-11-22 05:10:51 +0000201 struct net_device *dev;
Tejun Heo6d5759d2012-11-22 07:32:46 -0800202 u32 prio;
203 int ret;
Neil Horman5bc14212011-11-22 05:10:51 +0000204
Tejun Heo451af502014-05-13 12:16:21 -0400205 if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
Tejun Heo6d5759d2012-11-22 07:32:46 -0800206 return -EINVAL;
Neil Horman5bc14212011-11-22 05:10:51 +0000207
208 dev = dev_get_by_name(&init_net, devname);
209 if (!dev)
Tejun Heo6d5759d2012-11-22 07:32:46 -0800210 return -ENODEV;
Neil Horman5bc14212011-11-22 05:10:51 +0000211
John Fastabend476ad152012-08-14 12:34:35 +0000212 rtnl_lock();
Tejun Heo6d5759d2012-11-22 07:32:46 -0800213
Tejun Heo451af502014-05-13 12:16:21 -0400214 ret = netprio_set_prio(of_css(of), dev, prio);
Gao fengef209f12012-07-11 21:50:15 +0000215
John Fastabend476ad152012-08-14 12:34:35 +0000216 rtnl_unlock();
Neil Horman5bc14212011-11-22 05:10:51 +0000217 dev_put(dev);
Tejun Heo451af502014-05-13 12:16:21 -0400218 return ret ?: nbytes;
Neil Horman5bc14212011-11-22 05:10:51 +0000219}
220
Al Viroc3c073f2012-08-21 22:32:06 -0400221static int update_netprio(const void *v, struct file *file, unsigned n)
222{
223 int err;
224 struct socket *sock = sock_from_file(file, &err);
225 if (sock)
226 sock->sk->sk_cgrp_prioidx = (u32)(unsigned long)v;
227 return 0;
228}
229
Tejun Heoeb954192013-08-08 20:11:23 -0400230static void net_prio_attach(struct cgroup_subsys_state *css,
231 struct cgroup_taskset *tset)
John Fastabend406a3c62012-07-20 10:39:25 +0000232{
233 struct task_struct *p;
Gao fenge1af5e42013-10-08 11:05:19 +0800234 void *v = (void *)(unsigned long)css->cgroup->id;
John Fastabend406a3c62012-07-20 10:39:25 +0000235
Tejun Heo924f0d92014-02-13 06:58:41 -0500236 cgroup_taskset_for_each(p, tset) {
John Fastabend406a3c62012-07-20 10:39:25 +0000237 task_lock(p);
Al Viroc3c073f2012-08-21 22:32:06 -0400238 iterate_fd(p->files, 0, update_netprio, v);
John Fastabend406a3c62012-07-20 10:39:25 +0000239 task_unlock(p);
240 }
John Fastabend406a3c62012-07-20 10:39:25 +0000241}
242
Neil Horman5bc14212011-11-22 05:10:51 +0000243static struct cftype ss_files[] = {
244 {
245 .name = "prioidx",
246 .read_u64 = read_prioidx,
247 },
248 {
249 .name = "ifpriomap",
Tejun Heo2da8ca82013-12-05 12:28:04 -0500250 .seq_show = read_priomap,
Tejun Heo451af502014-05-13 12:16:21 -0400251 .write = write_priomap,
Neil Horman5bc14212011-11-22 05:10:51 +0000252 },
Tejun Heo4baf6e32012-04-01 12:09:55 -0700253 { } /* terminate */
Neil Horman5bc14212011-11-22 05:10:51 +0000254};
255
Tejun Heo073219e2014-02-08 10:36:58 -0500256struct cgroup_subsys net_prio_cgrp_subsys = {
Tejun Heo92fb9742012-11-19 08:13:38 -0800257 .css_alloc = cgrp_css_alloc,
Tejun Heo811d8d62012-11-22 07:32:47 -0800258 .css_online = cgrp_css_online,
Tejun Heo92fb9742012-11-19 08:13:38 -0800259 .css_free = cgrp_css_free,
John Fastabend406a3c62012-07-20 10:39:25 +0000260 .attach = net_prio_attach,
Tejun Heo55779642014-07-15 11:05:09 -0400261 .legacy_cftypes = ss_files,
Tejun Heo676f7c82012-04-01 12:09:55 -0700262};
Neil Horman5bc14212011-11-22 05:10:51 +0000263
264static int netprio_device_event(struct notifier_block *unused,
265 unsigned long event, void *ptr)
266{
Jiri Pirko351638e2013-05-28 01:30:21 +0000267 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
Neil Horman5bc14212011-11-22 05:10:51 +0000268 struct netprio_map *old;
Neil Horman5bc14212011-11-22 05:10:51 +0000269
270 /*
271 * Note this is called with rtnl_lock held so we have update side
272 * protection on our rcu assignments
273 */
274
275 switch (event) {
Neil Horman5bc14212011-11-22 05:10:51 +0000276 case NETDEV_UNREGISTER:
277 old = rtnl_dereference(dev->priomap);
Eric Dumazet2cfa5a02011-11-23 07:09:32 +0000278 RCU_INIT_POINTER(dev->priomap, NULL);
Neil Horman5bc14212011-11-22 05:10:51 +0000279 if (old)
280 kfree_rcu(old, rcu);
281 break;
282 }
283 return NOTIFY_DONE;
284}
285
286static struct notifier_block netprio_device_notifier = {
287 .notifier_call = netprio_device_event
288};
289
290static int __init init_cgroup_netprio(void)
291{
Neil Horman5bc14212011-11-22 05:10:51 +0000292 register_netdevice_notifier(&netprio_device_notifier);
Tejun Heoaf636332014-02-08 10:36:58 -0500293 return 0;
Neil Horman5bc14212011-11-22 05:10:51 +0000294}
295
Tejun Heoaf636332014-02-08 10:36:58 -0500296subsys_initcall(init_cgroup_netprio);
Neil Horman5bc14212011-11-22 05:10:51 +0000297MODULE_LICENSE("GPL v2");