1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * net/core/netprio_cgroup.c Priority Control Group
5 * Authors: Neil Horman <nhorman@tuxdriver.com>
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
21 #include <net/rtnetlink.h>
22 #include <net/pkt_cls.h>
24 #include <net/netprio_cgroup.h>
26 #include <linux/fdtable.h>
29 * netprio allocates per-net_device priomap array which is indexed by
30 * css->id. Limiting css ID to 16bits doesn't lose anything.
32 #define NETPRIO_ID_MAX USHRT_MAX
34 #define PRIOMAP_MIN_SZ 128
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.
41 static int extend_netdev_table(struct net_device *dev, u32 target_idx)
43 struct netprio_map *old, *new;
44 size_t new_sz, new_len;
46 /* is the existing priomap large enough? */
47 old = rtnl_dereference(dev->priomap);
48 if (old && old->priomap_len > target_idx)
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.
56 new_sz = PRIOMAP_MIN_SZ;
58 new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
59 sizeof(new->priomap[0]);
60 if (new_len > target_idx)
64 if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
69 new = kzalloc(new_sz, GFP_KERNEL);
74 memcpy(new->priomap, old->priomap,
75 old->priomap_len * sizeof(old->priomap[0]));
77 new->priomap_len = new_len;
79 /* install the new priomap */
80 rcu_assign_pointer(dev->priomap, new);
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
91 * Should be called under RCU read or rtnl lock.
93 static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
95 struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
96 int id = css->cgroup->id;
98 if (map && id < map->priomap_len)
99 return map->priomap[id];
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
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.
112 static int netprio_set_prio(struct cgroup_subsys_state *css,
113 struct net_device *dev, u32 prio)
115 struct netprio_map *map;
116 int id = css->cgroup->id;
119 /* avoid extending priomap for zero writes */
120 map = rtnl_dereference(dev->priomap);
121 if (!prio && (!map || map->priomap_len <= id))
124 ret = extend_netdev_table(dev, id);
128 map = rtnl_dereference(dev->priomap);
129 map->priomap[id] = prio;
133 static struct cgroup_subsys_state *
134 cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
136 struct cgroup_subsys_state *css;
138 css = kzalloc(sizeof(*css), GFP_KERNEL);
140 return ERR_PTR(-ENOMEM);
145 static int cgrp_css_online(struct cgroup_subsys_state *css)
147 struct cgroup_subsys_state *parent_css = css->parent;
148 struct net_device *dev;
151 if (css->id > NETPRIO_ID_MAX)
159 * Inherit prios from the parent. As all prios are set during
160 * onlining, there is no need to clear them on offline.
162 for_each_netdev(&init_net, dev) {
163 u32 prio = netprio_prio(parent_css, dev);
165 ret = netprio_set_prio(css, dev, prio);
173 static void cgrp_css_free(struct cgroup_subsys_state *css)
178 static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
180 return css->cgroup->id;
183 static int read_priomap(struct seq_file *sf, void *v)
185 struct net_device *dev;
188 for_each_netdev_rcu(&init_net, dev)
189 seq_printf(sf, "%s %u\n", dev->name,
190 netprio_prio(seq_css(sf), dev));
195 static ssize_t write_priomap(struct kernfs_open_file *of,
196 char *buf, size_t nbytes, loff_t off)
198 char devname[IFNAMSIZ + 1];
199 struct net_device *dev;
203 if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
206 dev = dev_get_by_name(&init_net, devname);
210 cgroup_sk_alloc_disable();
214 ret = netprio_set_prio(of_css(of), dev, prio);
218 return ret ?: nbytes;
221 static int update_netprio(const void *v, struct file *file, unsigned n)
224 struct socket *sock = sock_from_file(file, &err);
226 spin_lock(&cgroup_sk_update_lock);
227 sock_cgroup_set_prioidx(&sock->sk->sk_cgrp_data,
229 spin_unlock(&cgroup_sk_update_lock);
234 static void net_prio_attach(struct cgroup_taskset *tset)
236 struct task_struct *p;
237 struct cgroup_subsys_state *css;
239 cgroup_sk_alloc_disable();
241 cgroup_taskset_for_each(p, css, tset) {
242 void *v = (void *)(unsigned long)css->cgroup->id;
245 iterate_fd(p->files, 0, update_netprio, v);
250 static struct cftype ss_files[] = {
253 .read_u64 = read_prioidx,
257 .seq_show = read_priomap,
258 .write = write_priomap,
263 struct cgroup_subsys net_prio_cgrp_subsys = {
264 .css_alloc = cgrp_css_alloc,
265 .css_online = cgrp_css_online,
266 .css_free = cgrp_css_free,
267 .attach = net_prio_attach,
268 .legacy_cftypes = ss_files,
271 static int netprio_device_event(struct notifier_block *unused,
272 unsigned long event, void *ptr)
274 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
275 struct netprio_map *old;
278 * Note this is called with rtnl_lock held so we have update side
279 * protection on our rcu assignments
283 case NETDEV_UNREGISTER:
284 old = rtnl_dereference(dev->priomap);
285 RCU_INIT_POINTER(dev->priomap, NULL);
293 static struct notifier_block netprio_device_notifier = {
294 .notifier_call = netprio_device_event
297 static int __init init_cgroup_netprio(void)
299 register_netdevice_notifier(&netprio_device_notifier);
302 subsys_initcall(init_cgroup_netprio);