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 #include <linux/module.h>
  13 #include <linux/slab.h>
  14 #include <linux/types.h>
  15 #include <linux/string.h>
  16 #include <linux/errno.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/cgroup.h>
  19 #include <linux/rcupdate.h>
  20 #include <linux/atomic.h>
  21 #include <net/rtnetlink.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/sock.h>
  24 #include <net/netprio_cgroup.h>
  25
  26 static struct cgroup_subsys_state *cgrp_create(struct cgroup *cgrp);
  27 static void cgrp_destroy(struct cgroup *cgrp);
  28 static int cgrp_populate(struct cgroup_subsys *ss, struct cgroup *cgrp);
  29
  30 struct cgroup_subsys net_prio_subsys = {
  31         .name           = "net_prio",
  32         .create         = cgrp_create,
  33         .destroy        = cgrp_destroy,
  34         .populate       = cgrp_populate,
  35 #ifdef CONFIG_NETPRIO_CGROUP
  36         .subsys_id      = net_prio_subsys_id,
  37 #endif
  38         .module         = THIS_MODULE
  39 };
  40
  41 #define PRIOIDX_SZ 128
  42
  43 static unsigned long prioidx_map[PRIOIDX_SZ];
  44 static DEFINE_SPINLOCK(prioidx_map_lock);
  45 static atomic_t max_prioidx = ATOMIC_INIT(0);
  46
  47 static inline struct cgroup_netprio_state *cgrp_netprio_state(struct cgroup *cgrp)
  48 {
  49         return container_of(cgroup_subsys_state(cgrp, net_prio_subsys_id),
  50                             struct cgroup_netprio_state, css);
  51 }
  52
  53 static int get_prioidx(u32 *prio)
  54 {
  55         unsigned long flags;
  56         u32 prioidx;
  57
  58         spin_lock_irqsave(&prioidx_map_lock, flags);
  59         prioidx = find_first_zero_bit(prioidx_map, sizeof(unsigned long) * PRIOIDX_SZ);
  60         set_bit(prioidx, prioidx_map);
  61         spin_unlock_irqrestore(&prioidx_map_lock, flags);
  62         if (prioidx == sizeof(unsigned long) * PRIOIDX_SZ)
  63                 return -ENOSPC;
  64
  65         atomic_set(&max_prioidx, prioidx);
  66         *prio = prioidx;
  67         return 0;
  68 }
  69
  70 static void put_prioidx(u32 idx)
  71 {
  72         unsigned long flags;
  73
  74         spin_lock_irqsave(&prioidx_map_lock, flags);
  75         clear_bit(idx, prioidx_map);
  76         spin_unlock_irqrestore(&prioidx_map_lock, flags);
  77 }
  78
  79 static void extend_netdev_table(struct net_device *dev, u32 new_len)
  80 {
  81         size_t new_size = sizeof(struct netprio_map) +
  82                            ((sizeof(u32) * new_len));
  83         struct netprio_map *new_priomap = kzalloc(new_size, GFP_KERNEL);
  84         struct netprio_map *old_priomap;
  85         int i;
  86
  87         old_priomap  = rtnl_dereference(dev->priomap);
  88
  89         if (!new_priomap) {
  90                 printk(KERN_WARNING "Unable to alloc new priomap!\n");
  91                 return;
  92         }
  93
  94         for (i = 0;
  95              old_priomap && (i < old_priomap->priomap_len);
  96              i++)
  97                 new_priomap->priomap[i] = old_priomap->priomap[i];
  98
  99         new_priomap->priomap_len = new_len;
 100
 101         rcu_assign_pointer(dev->priomap, new_priomap);
 102         if (old_priomap)
 103                 kfree_rcu(old_priomap, rcu);
 104 }
 105
 106 static void update_netdev_tables(void)
 107 {
 108         struct net_device *dev;
 109         u32 max_len = atomic_read(&max_prioidx);
 110         struct netprio_map *map;
 111
 112         rtnl_lock();
 113         for_each_netdev(&init_net, dev) {
 114                 map = rtnl_dereference(dev->priomap);
 115                 if ((!map) ||
 116                     (map->priomap_len < max_len))
 117                         extend_netdev_table(dev, max_len);
 118         }
 119         rtnl_unlock();
 120 }
 121
 122 static struct cgroup_subsys_state *cgrp_create(struct cgroup *cgrp)
 123 {
 124         struct cgroup_netprio_state *cs;
 125         int ret;
 126
 127         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
 128         if (!cs)
 129                 return ERR_PTR(-ENOMEM);
 130
 131         if (cgrp->parent && cgrp_netprio_state(cgrp->parent)->prioidx) {
 132                 kfree(cs);
 133                 return ERR_PTR(-EINVAL);
 134         }
 135
 136         ret = get_prioidx(&cs->prioidx);
 137         if (ret != 0) {
 138                 printk(KERN_WARNING "No space in priority index array\n");
 139                 kfree(cs);
 140                 return ERR_PTR(ret);
 141         }
 142
 143         return &cs->css;
 144 }
 145
 146 static void cgrp_destroy(struct cgroup *cgrp)
 147 {
 148         struct cgroup_netprio_state *cs;
 149         struct net_device *dev;
 150         struct netprio_map *map;
 151
 152         cs = cgrp_netprio_state(cgrp);
 153         rtnl_lock();
 154         for_each_netdev(&init_net, dev) {
 155                 map = rtnl_dereference(dev->priomap);
 156                 if (map)
 157                         map->priomap[cs->prioidx] = 0;
 158         }
 159         rtnl_unlock();
 160         put_prioidx(cs->prioidx);
 161         kfree(cs);
 162 }
 163
 164 static u64 read_prioidx(struct cgroup *cgrp, struct cftype *cft)
 165 {
 166         return (u64)cgrp_netprio_state(cgrp)->prioidx;
 167 }
 168
 169 static int read_priomap(struct cgroup *cont, struct cftype *cft,
 170                         struct cgroup_map_cb *cb)
 171 {
 172         struct net_device *dev;
 173         u32 prioidx = cgrp_netprio_state(cont)->prioidx;
 174         u32 priority;
 175         struct netprio_map *map;
 176
 177         rcu_read_lock();
 178         for_each_netdev_rcu(&init_net, dev) {
 179                 map = rcu_dereference(dev->priomap);
 180                 priority = map ? map->priomap[prioidx] : 0;
 181                 cb->fill(cb, dev->name, priority);
 182         }
 183         rcu_read_unlock();
 184         return 0;
 185 }
 186
 187 static int write_priomap(struct cgroup *cgrp, struct cftype *cft,
 188                          const char *buffer)
 189 {
 190         char *devname = kstrdup(buffer, GFP_KERNEL);
 191         int ret = -EINVAL;
 192         u32 prioidx = cgrp_netprio_state(cgrp)->prioidx;
 193         unsigned long priority;
 194         char *priostr;
 195         struct net_device *dev;
 196         struct netprio_map *map;
 197
 198         if (!devname)
 199                 return -ENOMEM;
 200
 201         /*
 202          * Minimally sized valid priomap string
 203          */
 204         if (strlen(devname) < 3)
 205                 goto out_free_devname;
 206
 207         priostr = strstr(devname, " ");
 208         if (!priostr)
 209                 goto out_free_devname;
 210
 211         /*
 212          *Separate the devname from the associated priority
 213          *and advance the priostr poitner to the priority value
 214          */
 215         *priostr = '\0';
 216         priostr++;
 217
 218         /*
 219          * If the priostr points to NULL, we're at the end of the passed
 220          * in string, and its not a valid write
 221          */
 222         if (*priostr == '\0')
 223                 goto out_free_devname;
 224
 225         ret = kstrtoul(priostr, 10, &priority);
 226         if (ret < 0)
 227                 goto out_free_devname;
 228
 229         ret = -ENODEV;
 230
 231         dev = dev_get_by_name(&init_net, devname);
 232         if (!dev)
 233                 goto out_free_devname;
 234
 235         update_netdev_tables();
 236         ret = 0;
 237         rcu_read_lock();
 238         map = rcu_dereference(dev->priomap);
 239         if (map)
 240                 map->priomap[prioidx] = priority;
 241         rcu_read_unlock();
 242         dev_put(dev);
 243
 244 out_free_devname:
 245         kfree(devname);
 246         return ret;
 247 }
 248
 249 static struct cftype ss_files[] = {
 250         {
 251                 .name = "prioidx",
 252                 .read_u64 = read_prioidx,
 253         },
 254         {
 255                 .name = "ifpriomap",
 256                 .read_map = read_priomap,
 257                 .write_string = write_priomap,
 258         },
 259 };
 260
 261 static int cgrp_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
 262 {
 263         return cgroup_add_files(cgrp, ss, ss_files, ARRAY_SIZE(ss_files));
 264 }
 265
 266 static int netprio_device_event(struct notifier_block *unused,
 267                                 unsigned long event, void *ptr)
 268 {
 269         struct net_device *dev = ptr;
 270         struct netprio_map *old;
 271         u32 max_len = atomic_read(&max_prioidx);
 272
 273         /*
 274          * Note this is called with rtnl_lock held so we have update side
 275          * protection on our rcu assignments
 276          */
 277
 278         switch (event) {
 279
 280         case NETDEV_REGISTER:
 281                 if (max_len)
 282                         extend_netdev_table(dev, max_len);
 283                 break;
 284         case NETDEV_UNREGISTER:
 285                 old = rtnl_dereference(dev->priomap);
 286                 RCU_INIT_POINTER(dev->priomap, NULL);
 287                 if (old)
 288                         kfree_rcu(old, rcu);
 289                 break;
 290         }
 291         return NOTIFY_DONE;
 292 }
 293
 294 static struct notifier_block netprio_device_notifier = {
 295         .notifier_call = netprio_device_event
 296 };
 297
 298 static int __init init_cgroup_netprio(void)
 299 {
 300         int ret;
 301
 302         ret = cgroup_load_subsys(&net_prio_subsys);
 303         if (ret)
 304                 goto out;
 305 #ifndef CONFIG_NETPRIO_CGROUP
 306         smp_wmb();
 307         net_prio_subsys_id = net_prio_subsys.subsys_id;
 308 #endif
 309
 310         register_netdevice_notifier(&netprio_device_notifier);
 311
 312 out:
 313         return ret;
 314 }
 315
 316 static void __exit exit_cgroup_netprio(void)
 317 {
 318         struct netprio_map *old;
 319         struct net_device *dev;
 320
 321         unregister_netdevice_notifier(&netprio_device_notifier);
 322
 323         cgroup_unload_subsys(&net_prio_subsys);
 324
 325 #ifndef CONFIG_NETPRIO_CGROUP
 326         net_prio_subsys_id = -1;
 327         synchronize_rcu();
 328 #endif
 329
 330         rtnl_lock();
 331         for_each_netdev(&init_net, dev) {
 332                 old = rtnl_dereference(dev->priomap);
 333                 RCU_INIT_POINTER(dev->priomap, NULL);
 334                 if (old)
 335                         kfree_rcu(old, rcu);
 336         }
 337         rtnl_unlock();
 338 }
 339
 340 module_init(init_cgroup_netprio);
 341 module_exit(exit_cgroup_netprio);
 342 MODULE_LICENSE("GPL v2");