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/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 #include <linux/fdtable.h>
  29
  30 #define PRIOIDX_SZ 128
  31
  32 static unsigned long prioidx_map[PRIOIDX_SZ];
  33 static DEFINE_SPINLOCK(prioidx_map_lock);
  34 static atomic_t max_prioidx = ATOMIC_INIT(0);
  35
  36 static inline struct cgroup_netprio_state *cgrp_netprio_state(struct cgroup *cgrp)
  37 {
  38         return container_of(cgroup_subsys_state(cgrp, net_prio_subsys_id),
  39                             struct cgroup_netprio_state, css);
  40 }
  41
  42 static int get_prioidx(u32 *prio)
  43 {
  44         unsigned long flags;
  45         u32 prioidx;
  46
  47         spin_lock_irqsave(&prioidx_map_lock, flags);
  48         prioidx = find_first_zero_bit(prioidx_map, sizeof(unsigned long) * PRIOIDX_SZ);
  49         if (prioidx == sizeof(unsigned long) * PRIOIDX_SZ) {
  50                 spin_unlock_irqrestore(&prioidx_map_lock, flags);
  51                 return -ENOSPC;
  52         }
  53         set_bit(prioidx, prioidx_map);
  54         if (atomic_read(&max_prioidx) < prioidx)
  55                 atomic_set(&max_prioidx, prioidx);
  56         spin_unlock_irqrestore(&prioidx_map_lock, flags);
  57         *prio = prioidx;
  58         return 0;
  59 }
  60
  61 static void put_prioidx(u32 idx)
  62 {
  63         unsigned long flags;
  64
  65         spin_lock_irqsave(&prioidx_map_lock, flags);
  66         clear_bit(idx, prioidx_map);
  67         spin_unlock_irqrestore(&prioidx_map_lock, flags);
  68 }
  69
  70 static int extend_netdev_table(struct net_device *dev, u32 new_len)
  71 {
  72         size_t new_size = sizeof(struct netprio_map) +
  73                            ((sizeof(u32) * new_len));
  74         struct netprio_map *new = kzalloc(new_size, GFP_KERNEL);
  75         struct netprio_map *old;
  76
  77         old = rtnl_dereference(dev->priomap);
  78
  79         if (!new) {
  80                 pr_warn("Unable to alloc new priomap!\n");
  81                 return -ENOMEM;
  82         }
  83
  84         if (old)
  85                 memcpy(new->priomap, old->priomap,
  86                        old->priomap_len * sizeof(old->priomap[0]));
  87
  88         new->priomap_len = new_len;
  89
  90         rcu_assign_pointer(dev->priomap, new);
  91         if (old)
  92                 kfree_rcu(old, rcu);
  93         return 0;
  94 }
  95
  96 static 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         max_len = atomic_read(&max_prioidx) + 1;
 103         map = rtnl_dereference(dev->priomap);
 104         if (!map || map->priomap_len < max_len)
 105                 ret = extend_netdev_table(dev, max_len);
 106
 107         return ret;
 108 }
 109
 110 static struct cgroup_subsys_state *cgrp_css_alloc(struct cgroup *cgrp)
 111 {
 112         struct cgroup_netprio_state *cs;
 113         int ret = -EINVAL;
 114
 115         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
 116         if (!cs)
 117                 return ERR_PTR(-ENOMEM);
 118
 119         if (cgrp->parent && cgrp_netprio_state(cgrp->parent)->prioidx)
 120                 goto out;
 121
 122         ret = get_prioidx(&cs->prioidx);
 123         if (ret < 0) {
 124                 pr_warn("No space in priority index array\n");
 125                 goto out;
 126         }
 127
 128         return &cs->css;
 129 out:
 130         kfree(cs);
 131         return ERR_PTR(ret);
 132 }
 133
 134 static void cgrp_css_free(struct cgroup *cgrp)
 135 {
 136         struct cgroup_netprio_state *cs;
 137         struct net_device *dev;
 138         struct netprio_map *map;
 139
 140         cs = cgrp_netprio_state(cgrp);
 141         rtnl_lock();
 142         for_each_netdev(&init_net, dev) {
 143                 map = rtnl_dereference(dev->priomap);
 144                 if (map && cs->prioidx < map->priomap_len)
 145                         map->priomap[cs->prioidx] = 0;
 146         }
 147         rtnl_unlock();
 148         put_prioidx(cs->prioidx);
 149         kfree(cs);
 150 }
 151
 152 static u64 read_prioidx(struct cgroup *cgrp, struct cftype *cft)
 153 {
 154         return (u64)cgrp_netprio_state(cgrp)->prioidx;
 155 }
 156
 157 static int read_priomap(struct cgroup *cont, struct cftype *cft,
 158                         struct cgroup_map_cb *cb)
 159 {
 160         struct net_device *dev;
 161         u32 prioidx = cgrp_netprio_state(cont)->prioidx;
 162         u32 priority;
 163         struct netprio_map *map;
 164
 165         rcu_read_lock();
 166         for_each_netdev_rcu(&init_net, dev) {
 167                 map = rcu_dereference(dev->priomap);
 168                 priority = (map && prioidx < map->priomap_len) ? map->priomap[prioidx] : 0;
 169                 cb->fill(cb, dev->name, priority);
 170         }
 171         rcu_read_unlock();
 172         return 0;
 173 }
 174
 175 static int write_priomap(struct cgroup *cgrp, struct cftype *cft,
 176                          const char *buffer)
 177 {
 178         u32 prioidx = cgrp_netprio_state(cgrp)->prioidx;
 179         char devname[IFNAMSIZ + 1];
 180         struct net_device *dev;
 181         struct netprio_map *map;
 182         u32 prio;
 183         int ret;
 184
 185         if (sscanf(buffer, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
 186                 return -EINVAL;
 187
 188         dev = dev_get_by_name(&init_net, devname);
 189         if (!dev)
 190                 return -ENODEV;
 191
 192         rtnl_lock();
 193
 194         ret = write_update_netdev_table(dev);
 195         if (ret)
 196                 goto out_unlock;
 197
 198         map = rtnl_dereference(dev->priomap);
 199         if (map)
 200                 map->priomap[prioidx] = prio;
 201 out_unlock:
 202         rtnl_unlock();
 203         dev_put(dev);
 204         return ret;
 205 }
 206
 207 static int update_netprio(const void *v, struct file *file, unsigned n)
 208 {
 209         int err;
 210         struct socket *sock = sock_from_file(file, &err);
 211         if (sock)
 212                 sock->sk->sk_cgrp_prioidx = (u32)(unsigned long)v;
 213         return 0;
 214 }
 215
 216 void net_prio_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 217 {
 218         struct task_struct *p;
 219         void *v;
 220
 221         cgroup_taskset_for_each(p, cgrp, tset) {
 222                 task_lock(p);
 223                 v = (void *)(unsigned long)task_netprioidx(p);
 224                 iterate_fd(p->files, 0, update_netprio, v);
 225                 task_unlock(p);
 226         }
 227 }
 228
 229 static struct cftype ss_files[] = {
 230         {
 231                 .name = "prioidx",
 232                 .read_u64 = read_prioidx,
 233         },
 234         {
 235                 .name = "ifpriomap",
 236                 .read_map = read_priomap,
 237                 .write_string = write_priomap,
 238         },
 239         { }     /* terminate */
 240 };
 241
 242 struct cgroup_subsys net_prio_subsys = {
 243         .name           = "net_prio",
 244         .css_alloc      = cgrp_css_alloc,
 245         .css_free       = cgrp_css_free,
 246         .attach         = net_prio_attach,
 247         .subsys_id      = net_prio_subsys_id,
 248         .base_cftypes   = ss_files,
 249         .module         = THIS_MODULE,
 250
 251         /*
 252          * net_prio has artificial limit on the number of cgroups and
 253          * disallows nesting making it impossible to co-mount it with other
 254          * hierarchical subsystems.  Remove the artificially low PRIOIDX_SZ
 255          * limit and properly nest configuration such that children follow
 256          * their parents' configurations by default and are allowed to
 257          * override and remove the following.
 258          */
 259         .broken_hierarchy = true,
 260 };
 261
 262 static int netprio_device_event(struct notifier_block *unused,
 263                                 unsigned long event, void *ptr)
 264 {
 265         struct net_device *dev = ptr;
 266         struct netprio_map *old;
 267
 268         /*
 269          * Note this is called with rtnl_lock held so we have update side
 270          * protection on our rcu assignments
 271          */
 272
 273         switch (event) {
 274         case NETDEV_UNREGISTER:
 275                 old = rtnl_dereference(dev->priomap);
 276                 RCU_INIT_POINTER(dev->priomap, NULL);
 277                 if (old)
 278                         kfree_rcu(old, rcu);
 279                 break;
 280         }
 281         return NOTIFY_DONE;
 282 }
 283
 284 static struct notifier_block netprio_device_notifier = {
 285         .notifier_call = netprio_device_event
 286 };
 287
 288 static int __init init_cgroup_netprio(void)
 289 {
 290         int ret;
 291
 292         ret = cgroup_load_subsys(&net_prio_subsys);
 293         if (ret)
 294                 goto out;
 295
 296         register_netdevice_notifier(&netprio_device_notifier);
 297
 298 out:
 299         return ret;
 300 }
 301
 302 static void __exit exit_cgroup_netprio(void)
 303 {
 304         struct netprio_map *old;
 305         struct net_device *dev;
 306
 307         unregister_netdevice_notifier(&netprio_device_notifier);
 308
 309         cgroup_unload_subsys(&net_prio_subsys);
 310
 311         rtnl_lock();
 312         for_each_netdev(&init_net, dev) {
 313                 old = rtnl_dereference(dev->priomap);
 314                 RCU_INIT_POINTER(dev->priomap, NULL);
 315                 if (old)
 316                         kfree_rcu(old, rcu);
 317         }
 318         rtnl_unlock();
 319 }
 320
 321 module_init(init_cgroup_netprio);
 322 module_exit(exit_cgroup_netprio);
 323 MODULE_LICENSE("GPL v2");