net/core/netprio_cgroup.c

   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/slab.h>
  15 #include <linux/types.h>
  16 #include <linux/module.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 #include <linux/fdtable.h>
  29
  30 /*
  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
  36 #define PRIOMAP_MIN_SZ          128
  37
  38 /*
  39  * Extend @dev->priomap so that it's large enough to accommodate
  40  * @target_idx.  @dev->priomap.priomap_len > @target_idx after successful
  41  * return.  Must be called under rtnl lock.
  42  */
  43 static int extend_netdev_table(struct net_device *dev, u32 target_idx)
  44 {
  45         struct netprio_map *old, *new;
  46         size_t new_sz, new_len;
  47
  48         /* is the existing priomap large enough? */
  49         old = rtnl_dereference(dev->priomap);
  50         if (old && old->priomap_len > target_idx)
  51                 return 0;
  52
  53         /*
  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);
  72         if (!new)
  73                 return -ENOMEM;
  74
  75         if (old)
  76                 memcpy(new->priomap, old->priomap,
  77                        old->priomap_len * sizeof(old->priomap[0]));
  78
  79         new->priomap_len = new_len;
  80
  81         /* install the new priomap */
  82         rcu_assign_pointer(dev->priomap, new);
  83         if (old)
  84                 kfree_rcu(old, rcu);
  85         return 0;
  86 }
  87
  88 /**
  89  * netprio_prio - return the effective netprio of a cgroup-net_device pair
  90  * @css: css part of the target pair
  91  * @dev: net_device part of the target pair
  92  *
  93  * Should be called under RCU read or rtnl lock.
  94  */
  95 static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
  96 {
  97         struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
  98         int id = css->cgroup->id;
  99
 100         if (map && id < map->priomap_len)
 101                 return map->priomap[id];
 102         return 0;
 103 }
 104
 105 /**
 106  * netprio_set_prio - set netprio on a cgroup-net_device pair
 107  * @css: css part of the target pair
 108  * @dev: net_device part of the target pair
 109  * @prio: prio to set
 110  *
 111  * Set netprio to @prio on @css-@dev pair.  Should be called under rtnl
 112  * lock and may fail under memory pressure for non-zero @prio.
 113  */
 114 static int netprio_set_prio(struct cgroup_subsys_state *css,
 115                             struct net_device *dev, u32 prio)
 116 {
 117         struct netprio_map *map;
 118         int id = css->cgroup->id;
 119         int ret;
 120
 121         /* avoid extending priomap for zero writes */
 122         map = rtnl_dereference(dev->priomap);
 123         if (!prio && (!map || map->priomap_len <= id))
 124                 return 0;
 125
 126         ret = extend_netdev_table(dev, id);
 127         if (ret)
 128                 return ret;
 129
 130         map = rtnl_dereference(dev->priomap);
 131         map->priomap[id] = prio;
 132         return 0;
 133 }
 134
 135 static struct cgroup_subsys_state *
 136 cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
 137 {
 138         struct cgroup_subsys_state *css;
 139
 140         css = kzalloc(sizeof(*css), GFP_KERNEL);
 141         if (!css)
 142                 return ERR_PTR(-ENOMEM);
 143
 144         return css;
 145 }
 146
 147 static int cgrp_css_online(struct cgroup_subsys_state *css)
 148 {
 149         struct cgroup_subsys_state *parent_css = css->parent;
 150         struct net_device *dev;
 151         int ret = 0;
 152
 153         if (css->id > NETPRIO_ID_MAX)
 154                 return -ENOSPC;
 155
 156         if (!parent_css)
 157                 return 0;
 158
 159         rtnl_lock();
 160         /*
 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) {
 165                 u32 prio = netprio_prio(parent_css, dev);
 166
 167                 ret = netprio_set_prio(css, dev, prio);
 168                 if (ret)
 169                         break;
 170         }
 171         rtnl_unlock();
 172         return ret;
 173 }
 174
 175 static void cgrp_css_free(struct cgroup_subsys_state *css)
 176 {
 177         kfree(css);
 178 }
 179
 180 static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
 181 {
 182         return css->cgroup->id;
 183 }
 184
 185 static int read_priomap(struct seq_file *sf, void *v)
 186 {
 187         struct net_device *dev;
 188
 189         rcu_read_lock();
 190         for_each_netdev_rcu(&init_net, dev)
 191                 seq_printf(sf, "%s %u\n", dev->name,
 192                            netprio_prio(seq_css(sf), dev));
 193         rcu_read_unlock();
 194         return 0;
 195 }
 196
 197 static ssize_t write_priomap(struct kernfs_open_file *of,
 198                              char *buf, size_t nbytes, loff_t off)
 199 {
 200         char devname[IFNAMSIZ + 1];
 201         struct net_device *dev;
 202         u32 prio;
 203         int ret;
 204
 205         if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
 206                 return -EINVAL;
 207
 208         dev = dev_get_by_name(&init_net, devname);
 209         if (!dev)
 210                 return -ENODEV;
 211
 212         cgroup_sk_alloc_disable();
 213
 214         rtnl_lock();
 215
 216         ret = netprio_set_prio(of_css(of), dev, prio);
 217
 218         rtnl_unlock();
 219         dev_put(dev);
 220         return ret ?: nbytes;
 221 }
 222
 223 static int update_netprio(const void *v, struct file *file, unsigned n)
 224 {
 225         int err;
 226         struct socket *sock = sock_from_file(file, &err);
 227         if (sock) {
 228                 spin_lock(&cgroup_sk_update_lock);
 229                 sock_cgroup_set_prioidx(&sock->sk->sk_cgrp_data,
 230                                         (unsigned long)v);
 231                 spin_unlock(&cgroup_sk_update_lock);
 232         }
 233         return 0;
 234 }
 235
 236 static void net_prio_attach(struct cgroup_taskset *tset)
 237 {
 238         struct task_struct *p;
 239         struct cgroup_subsys_state *css;
 240
 241         cgroup_taskset_for_each(p, css, tset) {
 242                 void *v = (void *)(unsigned long)css->cgroup->id;
 243
 244                 task_lock(p);
 245                 iterate_fd(p->files, 0, update_netprio, v);
 246                 task_unlock(p);
 247         }
 248 }
 249
 250 static struct cftype ss_files[] = {
 251         {
 252                 .name = "prioidx",
 253                 .read_u64 = read_prioidx,
 254         },
 255         {
 256                 .name = "ifpriomap",
 257                 .seq_show = read_priomap,
 258                 .write = write_priomap,
 259         },
 260         { }     /* terminate */
 261 };
 262
 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,
 269 };
 270
 271 static int netprio_device_event(struct notifier_block *unused,
 272                                 unsigned long event, void *ptr)
 273 {
 274         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 275         struct netprio_map *old;
 276
 277         /*
 278          * Note this is called with rtnl_lock held so we have update side
 279          * protection on our rcu assignments
 280          */
 281
 282         switch (event) {
 283         case NETDEV_UNREGISTER:
 284                 old = rtnl_dereference(dev->priomap);
 285                 RCU_INIT_POINTER(dev->priomap, NULL);
 286                 if (old)
 287                         kfree_rcu(old, rcu);
 288                 break;
 289         }
 290         return NOTIFY_DONE;
 291 }
 292
 293 static struct notifier_block netprio_device_notifier = {
 294         .notifier_call = netprio_device_event
 295 };
 296
 297 static int __init init_cgroup_netprio(void)
 298 {
 299         register_netdevice_notifier(&netprio_device_notifier);
 300         return 0;
 301 }
 302
 303 subsys_initcall(init_cgroup_netprio);
 304 MODULE_LICENSE("GPL v2");