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_priomap = kzalloc(new_size, GFP_KERNEL);
  75         struct netprio_map *old_priomap;
  76         int i;
  77
  78         old_priomap  = rtnl_dereference(dev->priomap);
  79
  80         if (!new_priomap) {
  81                 pr_warn("Unable to alloc new priomap!\n");
  82                 return -ENOMEM;
  83         }
  84
  85         for (i = 0;
  86              old_priomap && (i < old_priomap->priomap_len);
  87              i++)
  88                 new_priomap->priomap[i] = old_priomap->priomap[i];
  89
  90         new_priomap->priomap_len = new_len;
  91
  92         rcu_assign_pointer(dev->priomap, new_priomap);
  93         if (old_priomap)
  94                 kfree_rcu(old_priomap, rcu);
  95         return 0;
  96 }
  97
  98 static int write_update_netdev_table(struct net_device *dev)
  99 {
 100         int ret = 0;
 101         u32 max_len;
 102         struct netprio_map *map;
 103
 104         max_len = atomic_read(&max_prioidx) + 1;
 105         map = rtnl_dereference(dev->priomap);
 106         if (!map || map->priomap_len < max_len)
 107                 ret = extend_netdev_table(dev, max_len);
 108
 109         return ret;
 110 }
 111
 112 static int update_netdev_tables(void)
 113 {
 114         int ret = 0;
 115         struct net_device *dev;
 116         u32 max_len;
 117         struct netprio_map *map;
 118
 119         rtnl_lock();
 120         max_len = atomic_read(&max_prioidx) + 1;
 121         for_each_netdev(&init_net, dev) {
 122                 map = rtnl_dereference(dev->priomap);
 123                 /*
 124                  * don't allocate priomap if we didn't
 125                  * change net_prio.ifpriomap (map == NULL),
 126                  * this will speed up skb_update_prio.
 127                  */
 128                 if (map && map->priomap_len < max_len) {
 129                         ret = extend_netdev_table(dev, max_len);
 130                         if (ret < 0)
 131                                 break;
 132                 }
 133         }
 134         rtnl_unlock();
 135         return ret;
 136 }
 137
 138 static struct cgroup_subsys_state *cgrp_create(struct cgroup *cgrp)
 139 {
 140         struct cgroup_netprio_state *cs;
 141         int ret = -EINVAL;
 142
 143         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
 144         if (!cs)
 145                 return ERR_PTR(-ENOMEM);
 146
 147         if (cgrp->parent && cgrp_netprio_state(cgrp->parent)->prioidx)
 148                 goto out;
 149
 150         ret = get_prioidx(&cs->prioidx);
 151         if (ret < 0) {
 152                 pr_warn("No space in priority index array\n");
 153                 goto out;
 154         }
 155
 156         ret = update_netdev_tables();
 157         if (ret < 0) {
 158                 put_prioidx(cs->prioidx);
 159                 goto out;
 160         }
 161
 162         return &cs->css;
 163 out:
 164         kfree(cs);
 165         return ERR_PTR(ret);
 166 }
 167
 168 static void cgrp_destroy(struct cgroup *cgrp)
 169 {
 170         struct cgroup_netprio_state *cs;
 171         struct net_device *dev;
 172         struct netprio_map *map;
 173
 174         cs = cgrp_netprio_state(cgrp);
 175         rtnl_lock();
 176         for_each_netdev(&init_net, dev) {
 177                 map = rtnl_dereference(dev->priomap);
 178                 if (map && cs->prioidx < map->priomap_len)
 179                         map->priomap[cs->prioidx] = 0;
 180         }
 181         rtnl_unlock();
 182         put_prioidx(cs->prioidx);
 183         kfree(cs);
 184 }
 185
 186 static u64 read_prioidx(struct cgroup *cgrp, struct cftype *cft)
 187 {
 188         return (u64)cgrp_netprio_state(cgrp)->prioidx;
 189 }
 190
 191 static int read_priomap(struct cgroup *cont, struct cftype *cft,
 192                         struct cgroup_map_cb *cb)
 193 {
 194         struct net_device *dev;
 195         u32 prioidx = cgrp_netprio_state(cont)->prioidx;
 196         u32 priority;
 197         struct netprio_map *map;
 198
 199         rcu_read_lock();
 200         for_each_netdev_rcu(&init_net, dev) {
 201                 map = rcu_dereference(dev->priomap);
 202                 priority = (map && prioidx < map->priomap_len) ? map->priomap[prioidx] : 0;
 203                 cb->fill(cb, dev->name, priority);
 204         }
 205         rcu_read_unlock();
 206         return 0;
 207 }
 208
 209 static int write_priomap(struct cgroup *cgrp, struct cftype *cft,
 210                          const char *buffer)
 211 {
 212         char *devname = kstrdup(buffer, GFP_KERNEL);
 213         int ret = -EINVAL;
 214         u32 prioidx = cgrp_netprio_state(cgrp)->prioidx;
 215         unsigned long priority;
 216         char *priostr;
 217         struct net_device *dev;
 218         struct netprio_map *map;
 219
 220         if (!devname)
 221                 return -ENOMEM;
 222
 223         /*
 224          * Minimally sized valid priomap string
 225          */
 226         if (strlen(devname) < 3)
 227                 goto out_free_devname;
 228
 229         priostr = strstr(devname, " ");
 230         if (!priostr)
 231                 goto out_free_devname;
 232
 233         /*
 234          *Separate the devname from the associated priority
 235          *and advance the priostr pointer to the priority value
 236          */
 237         *priostr = '\0';
 238         priostr++;
 239
 240         /*
 241          * If the priostr points to NULL, we're at the end of the passed
 242          * in string, and its not a valid write
 243          */
 244         if (*priostr == '\0')
 245                 goto out_free_devname;
 246
 247         ret = kstrtoul(priostr, 10, &priority);
 248         if (ret < 0)
 249                 goto out_free_devname;
 250
 251         ret = -ENODEV;
 252
 253         dev = dev_get_by_name(&init_net, devname);
 254         if (!dev)
 255                 goto out_free_devname;
 256
 257         rtnl_lock();
 258         ret = write_update_netdev_table(dev);
 259         if (ret < 0)
 260                 goto out_put_dev;
 261
 262         map = rtnl_dereference(dev->priomap);
 263         if (map)
 264                 map->priomap[prioidx] = priority;
 265
 266 out_put_dev:
 267         rtnl_unlock();
 268         dev_put(dev);
 269
 270 out_free_devname:
 271         kfree(devname);
 272         return ret;
 273 }
 274
 275 void net_prio_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 276 {
 277         struct task_struct *p;
 278
 279         cgroup_taskset_for_each(p, cgrp, tset) {
 280                 unsigned int fd;
 281                 struct fdtable *fdt;
 282                 struct files_struct *files;
 283
 284                 task_lock(p);
 285                 files = p->files;
 286                 if (!files) {
 287                         task_unlock(p);
 288                         continue;
 289                 }
 290
 291                 spin_lock(&files->file_lock);
 292                 fdt = files_fdtable(files);
 293                 for (fd = 0; fd < fdt->max_fds; fd++) {
 294                         struct file *file;
 295                         struct socket *sock;
 296                         int err;
 297
 298                         file = fcheck_files(files, fd);
 299                         if (!file)
 300                                 continue;
 301
 302                         sock = sock_from_file(file, &err);
 303                         if (sock)
 304                                 sock_update_netprioidx(sock->sk, p);
 305                 }
 306                 spin_unlock(&files->file_lock);
 307                 task_unlock(p);
 308         }
 309 }
 310
 311 static struct cftype ss_files[] = {
 312         {
 313                 .name = "prioidx",
 314                 .read_u64 = read_prioidx,
 315         },
 316         {
 317                 .name = "ifpriomap",
 318                 .read_map = read_priomap,
 319                 .write_string = write_priomap,
 320         },
 321         { }     /* terminate */
 322 };
 323
 324 struct cgroup_subsys net_prio_subsys = {
 325         .name           = "net_prio",
 326         .create         = cgrp_create,
 327         .destroy        = cgrp_destroy,
 328         .attach         = net_prio_attach,
 329         .subsys_id      = net_prio_subsys_id,
 330         .base_cftypes   = ss_files,
 331         .module         = THIS_MODULE,
 332
 333         /*
 334          * net_prio has artificial limit on the number of cgroups and
 335          * disallows nesting making it impossible to co-mount it with other
 336          * hierarchical subsystems.  Remove the artificially low PRIOIDX_SZ
 337          * limit and properly nest configuration such that children follow
 338          * their parents' configurations by default and are allowed to
 339          * override and remove the following.
 340          */
 341         .broken_hierarchy = true,
 342 };
 343
 344 static int netprio_device_event(struct notifier_block *unused,
 345                                 unsigned long event, void *ptr)
 346 {
 347         struct net_device *dev = ptr;
 348         struct netprio_map *old;
 349
 350         /*
 351          * Note this is called with rtnl_lock held so we have update side
 352          * protection on our rcu assignments
 353          */
 354
 355         switch (event) {
 356         case NETDEV_UNREGISTER:
 357                 old = rtnl_dereference(dev->priomap);
 358                 RCU_INIT_POINTER(dev->priomap, NULL);
 359                 if (old)
 360                         kfree_rcu(old, rcu);
 361                 break;
 362         }
 363         return NOTIFY_DONE;
 364 }
 365
 366 static struct notifier_block netprio_device_notifier = {
 367         .notifier_call = netprio_device_event
 368 };
 369
 370 static int __init init_cgroup_netprio(void)
 371 {
 372         int ret;
 373
 374         ret = cgroup_load_subsys(&net_prio_subsys);
 375         if (ret)
 376                 goto out;
 377
 378         register_netdevice_notifier(&netprio_device_notifier);
 379
 380 out:
 381         return ret;
 382 }
 383
 384 static void __exit exit_cgroup_netprio(void)
 385 {
 386         struct netprio_map *old;
 387         struct net_device *dev;
 388
 389         unregister_netdevice_notifier(&netprio_device_notifier);
 390
 391         cgroup_unload_subsys(&net_prio_subsys);
 392
 393         rtnl_lock();
 394         for_each_netdev(&init_net, dev) {
 395                 old = rtnl_dereference(dev->priomap);
 396                 RCU_INIT_POINTER(dev->priomap, NULL);
 397                 if (old)
 398                         kfree_rcu(old, rcu);
 399         }
 400         rtnl_unlock();
 401 }
 402
 403 module_init(init_cgroup_netprio);
 404 module_exit(exit_cgroup_netprio);
 405 MODULE_LICENSE("GPL v2");