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         rtnl_lock();
 105         max_len = atomic_read(&max_prioidx) + 1;
 106         map = rtnl_dereference(dev->priomap);
 107         if (!map || map->priomap_len < max_len)
 108                 ret = extend_netdev_table(dev, max_len);
 109         rtnl_unlock();
 110
 111         return ret;
 112 }
 113
 114 static int update_netdev_tables(void)
 115 {
 116         int ret = 0;
 117         struct net_device *dev;
 118         u32 max_len;
 119         struct netprio_map *map;
 120
 121         rtnl_lock();
 122         max_len = atomic_read(&max_prioidx) + 1;
 123         for_each_netdev(&init_net, dev) {
 124                 map = rtnl_dereference(dev->priomap);
 125                 /*
 126                  * don't allocate priomap if we didn't
 127                  * change net_prio.ifpriomap (map == NULL),
 128                  * this will speed up skb_update_prio.
 129                  */
 130                 if (map && map->priomap_len < max_len) {
 131                         ret = extend_netdev_table(dev, max_len);
 132                         if (ret < 0)
 133                                 break;
 134                 }
 135         }
 136         rtnl_unlock();
 137         return ret;
 138 }
 139
 140 static struct cgroup_subsys_state *cgrp_create(struct cgroup *cgrp)
 141 {
 142         struct cgroup_netprio_state *cs;
 143         int ret = -EINVAL;
 144
 145         cs = kzalloc(sizeof(*cs), GFP_KERNEL);
 146         if (!cs)
 147                 return ERR_PTR(-ENOMEM);
 148
 149         if (cgrp->parent && cgrp_netprio_state(cgrp->parent)->prioidx)
 150                 goto out;
 151
 152         ret = get_prioidx(&cs->prioidx);
 153         if (ret < 0) {
 154                 pr_warn("No space in priority index array\n");
 155                 goto out;
 156         }
 157
 158         ret = update_netdev_tables();
 159         if (ret < 0) {
 160                 put_prioidx(cs->prioidx);
 161                 goto out;
 162         }
 163
 164         return &cs->css;
 165 out:
 166         kfree(cs);
 167         return ERR_PTR(ret);
 168 }
 169
 170 static void cgrp_destroy(struct cgroup *cgrp)
 171 {
 172         struct cgroup_netprio_state *cs;
 173         struct net_device *dev;
 174         struct netprio_map *map;
 175
 176         cs = cgrp_netprio_state(cgrp);
 177         rtnl_lock();
 178         for_each_netdev(&init_net, dev) {
 179                 map = rtnl_dereference(dev->priomap);
 180                 if (map && cs->prioidx < map->priomap_len)
 181                         map->priomap[cs->prioidx] = 0;
 182         }
 183         rtnl_unlock();
 184         put_prioidx(cs->prioidx);
 185         kfree(cs);
 186 }
 187
 188 static u64 read_prioidx(struct cgroup *cgrp, struct cftype *cft)
 189 {
 190         return (u64)cgrp_netprio_state(cgrp)->prioidx;
 191 }
 192
 193 static int read_priomap(struct cgroup *cont, struct cftype *cft,
 194                         struct cgroup_map_cb *cb)
 195 {
 196         struct net_device *dev;
 197         u32 prioidx = cgrp_netprio_state(cont)->prioidx;
 198         u32 priority;
 199         struct netprio_map *map;
 200
 201         rcu_read_lock();
 202         for_each_netdev_rcu(&init_net, dev) {
 203                 map = rcu_dereference(dev->priomap);
 204                 priority = (map && prioidx < map->priomap_len) ? map->priomap[prioidx] : 0;
 205                 cb->fill(cb, dev->name, priority);
 206         }
 207         rcu_read_unlock();
 208         return 0;
 209 }
 210
 211 static int write_priomap(struct cgroup *cgrp, struct cftype *cft,
 212                          const char *buffer)
 213 {
 214         char *devname = kstrdup(buffer, GFP_KERNEL);
 215         int ret = -EINVAL;
 216         u32 prioidx = cgrp_netprio_state(cgrp)->prioidx;
 217         unsigned long priority;
 218         char *priostr;
 219         struct net_device *dev;
 220         struct netprio_map *map;
 221
 222         if (!devname)
 223                 return -ENOMEM;
 224
 225         /*
 226          * Minimally sized valid priomap string
 227          */
 228         if (strlen(devname) < 3)
 229                 goto out_free_devname;
 230
 231         priostr = strstr(devname, " ");
 232         if (!priostr)
 233                 goto out_free_devname;
 234
 235         /*
 236          *Separate the devname from the associated priority
 237          *and advance the priostr pointer to the priority value
 238          */
 239         *priostr = '\0';
 240         priostr++;
 241
 242         /*
 243          * If the priostr points to NULL, we're at the end of the passed
 244          * in string, and its not a valid write
 245          */
 246         if (*priostr == '\0')
 247                 goto out_free_devname;
 248
 249         ret = kstrtoul(priostr, 10, &priority);
 250         if (ret < 0)
 251                 goto out_free_devname;
 252
 253         ret = -ENODEV;
 254
 255         dev = dev_get_by_name(&init_net, devname);
 256         if (!dev)
 257                 goto out_free_devname;
 258
 259         ret = write_update_netdev_table(dev);
 260         if (ret < 0)
 261                 goto out_put_dev;
 262
 263         rcu_read_lock();
 264         map = rcu_dereference(dev->priomap);
 265         if (map)
 266                 map->priomap[prioidx] = priority;
 267         rcu_read_unlock();
 268
 269 out_put_dev:
 270         dev_put(dev);
 271
 272 out_free_devname:
 273         kfree(devname);
 274         return ret;
 275 }
 276
 277 void net_prio_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 278 {
 279         struct task_struct *p;
 280         char *tmp = kzalloc(sizeof(char) * PATH_MAX, GFP_KERNEL);
 281
 282         if (!tmp) {
 283                 pr_warn("Unable to attach cgrp due to alloc failure!\n");
 284                 return;
 285         }
 286
 287         cgroup_taskset_for_each(p, cgrp, tset) {
 288                 unsigned int fd;
 289                 struct fdtable *fdt;
 290                 struct files_struct *files;
 291
 292                 task_lock(p);
 293                 files = p->files;
 294                 if (!files) {
 295                         task_unlock(p);
 296                         continue;
 297                 }
 298
 299                 rcu_read_lock();
 300                 fdt = files_fdtable(files);
 301                 for (fd = 0; fd < fdt->max_fds; fd++) {
 302                         char *path;
 303                         struct file *file;
 304                         struct socket *sock;
 305                         unsigned long s;
 306                         int rv, err = 0;
 307
 308                         file = fcheck_files(files, fd);
 309                         if (!file)
 310                                 continue;
 311
 312                         path = d_path(&file->f_path, tmp, PAGE_SIZE);
 313                         rv = sscanf(path, "socket:[%lu]", &s);
 314                         if (rv <= 0)
 315                                 continue;
 316
 317                         sock = sock_from_file(file, &err);
 318                         if (!err)
 319                                 sock_update_netprioidx(sock->sk, p);
 320                 }
 321                 rcu_read_unlock();
 322                 task_unlock(p);
 323         }
 324         kfree(tmp);
 325 }
 326
 327 static struct cftype ss_files[] = {
 328         {
 329                 .name = "prioidx",
 330                 .read_u64 = read_prioidx,
 331         },
 332         {
 333                 .name = "ifpriomap",
 334                 .read_map = read_priomap,
 335                 .write_string = write_priomap,
 336         },
 337         { }     /* terminate */
 338 };
 339
 340 struct cgroup_subsys net_prio_subsys = {
 341         .name           = "net_prio",
 342         .create         = cgrp_create,
 343         .destroy        = cgrp_destroy,
 344         .attach         = net_prio_attach,
 345 #ifdef CONFIG_NETPRIO_CGROUP
 346         .subsys_id      = net_prio_subsys_id,
 347 #endif
 348         .base_cftypes   = ss_files,
 349         .module         = THIS_MODULE
 350 };
 351
 352 static int netprio_device_event(struct notifier_block *unused,
 353                                 unsigned long event, void *ptr)
 354 {
 355         struct net_device *dev = ptr;
 356         struct netprio_map *old;
 357
 358         /*
 359          * Note this is called with rtnl_lock held so we have update side
 360          * protection on our rcu assignments
 361          */
 362
 363         switch (event) {
 364         case NETDEV_UNREGISTER:
 365                 old = rtnl_dereference(dev->priomap);
 366                 RCU_INIT_POINTER(dev->priomap, NULL);
 367                 if (old)
 368                         kfree_rcu(old, rcu);
 369                 break;
 370         }
 371         return NOTIFY_DONE;
 372 }
 373
 374 static struct notifier_block netprio_device_notifier = {
 375         .notifier_call = netprio_device_event
 376 };
 377
 378 static int __init init_cgroup_netprio(void)
 379 {
 380         int ret;
 381
 382         ret = cgroup_load_subsys(&net_prio_subsys);
 383         if (ret)
 384                 goto out;
 385 #ifndef CONFIG_NETPRIO_CGROUP
 386         smp_wmb();
 387         net_prio_subsys_id = net_prio_subsys.subsys_id;
 388 #endif
 389
 390         register_netdevice_notifier(&netprio_device_notifier);
 391
 392 out:
 393         return ret;
 394 }
 395
 396 static void __exit exit_cgroup_netprio(void)
 397 {
 398         struct netprio_map *old;
 399         struct net_device *dev;
 400
 401         unregister_netdevice_notifier(&netprio_device_notifier);
 402
 403         cgroup_unload_subsys(&net_prio_subsys);
 404
 405 #ifndef CONFIG_NETPRIO_CGROUP
 406         net_prio_subsys_id = -1;
 407         synchronize_rcu();
 408 #endif
 409
 410         rtnl_lock();
 411         for_each_netdev(&init_net, dev) {
 412                 old = rtnl_dereference(dev->priomap);
 413                 RCU_INIT_POINTER(dev->priomap, NULL);
 414                 if (old)
 415                         kfree_rcu(old, rcu);
 416         }
 417         rtnl_unlock();
 418 }
 419
 420 module_init(init_cgroup_netprio);
 421 module_exit(exit_cgroup_netprio);
 422 MODULE_LICENSE("GPL v2");