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