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/slab.h>
  15 #include <linux/types.h>
  16 #include <linux/string.h>
  17 #include <linux/errno.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/cgroup.h>
  20 #include <linux/rcupdate.h>
  21 #include <linux/atomic.h>
  22 #include <net/rtnetlink.h>
  23 #include <net/pkt_cls.h>
  24 #include <net/sock.h>
  25 #include <net/netprio_cgroup.h>
  26
  27 #include <linux/fdtable.h>
  28
  29 /*
  30  * netprio allocates per-net_device priomap array which is indexed by
  31  * css->id.  Limiting css ID to 16bits doesn't lose anything.
  32  */
  33 #define NETPRIO_ID_MAX          USHRT_MAX
  34
  35 #define PRIOMAP_MIN_SZ          128
  36
  37 /*
  38  * Extend @dev->priomap so that it's large enough to accommodate
  39  * @target_idx.  @dev->priomap.priomap_len > @target_idx after successful
  40  * return.  Must be called under rtnl lock.
  41  */
  42 static int extend_netdev_table(struct net_device *dev, u32 target_idx)
  43 {
  44         struct netprio_map *old, *new;
  45         size_t new_sz, new_len;
  46
  47         /* is the existing priomap large enough? */
  48         old = rtnl_dereference(dev->priomap);
  49         if (old && old->priomap_len > target_idx)
  50                 return 0;
  51
  52         /*
  53          * Determine the new size.  Let's keep it power-of-two.  We start
  54          * from PRIOMAP_MIN_SZ and double it until it's large enough to
  55          * accommodate @target_idx.
  56          */
  57         new_sz = PRIOMAP_MIN_SZ;
  58         while (true) {
  59                 new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
  60                         sizeof(new->priomap[0]);
  61                 if (new_len > target_idx)
  62                         break;
  63                 new_sz *= 2;
  64                 /* overflowed? */
  65                 if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
  66                         return -ENOSPC;
  67         }
  68
  69         /* allocate & copy */
  70         new = kzalloc(new_sz, GFP_KERNEL);
  71         if (!new)
  72                 return -ENOMEM;
  73
  74         if (old)
  75                 memcpy(new->priomap, old->priomap,
  76                        old->priomap_len * sizeof(old->priomap[0]));
  77
  78         new->priomap_len = new_len;
  79
  80         /* install the new priomap */
  81         rcu_assign_pointer(dev->priomap, new);
  82         if (old)
  83                 kfree_rcu(old, rcu);
  84         return 0;
  85 }
  86
  87 /**
  88  * netprio_prio - return the effective netprio of a cgroup-net_device pair
  89  * @css: css part of the target pair
  90  * @dev: net_device part of the target pair
  91  *
  92  * Should be called under RCU read or rtnl lock.
  93  */
  94 static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
  95 {
  96         struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
  97         int id = css->cgroup->id;
  98
  99         if (map && id < map->priomap_len)
 100                 return map->priomap[id];
 101         return 0;
 102 }
 103
 104 /**
 105  * netprio_set_prio - set netprio on a cgroup-net_device pair
 106  * @css: css part of the target pair
 107  * @dev: net_device part of the target pair
 108  * @prio: prio to set
 109  *
 110  * Set netprio to @prio on @css-@dev pair.  Should be called under rtnl
 111  * lock and may fail under memory pressure for non-zero @prio.
 112  */
 113 static int netprio_set_prio(struct cgroup_subsys_state *css,
 114                             struct net_device *dev, u32 prio)
 115 {
 116         struct netprio_map *map;
 117         int id = css->cgroup->id;
 118         int ret;
 119
 120         /* avoid extending priomap for zero writes */
 121         map = rtnl_dereference(dev->priomap);
 122         if (!prio && (!map || map->priomap_len <= id))
 123                 return 0;
 124
 125         ret = extend_netdev_table(dev, id);
 126         if (ret)
 127                 return ret;
 128
 129         map = rtnl_dereference(dev->priomap);
 130         map->priomap[id] = prio;
 131         return 0;
 132 }
 133
 134 static struct cgroup_subsys_state *
 135 cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
 136 {
 137         struct cgroup_subsys_state *css;
 138
 139         css = kzalloc(sizeof(*css), GFP_KERNEL);
 140         if (!css)
 141                 return ERR_PTR(-ENOMEM);
 142
 143         return css;
 144 }
 145
 146 static int cgrp_css_online(struct cgroup_subsys_state *css)
 147 {
 148         struct cgroup_subsys_state *parent_css = css->parent;
 149         struct net_device *dev;
 150         int ret = 0;
 151
 152         if (css->id > NETPRIO_ID_MAX)
 153                 return -ENOSPC;
 154
 155         if (!parent_css)
 156                 return 0;
 157
 158         rtnl_lock();
 159         /*
 160          * Inherit prios from the parent.  As all prios are set during
 161          * onlining, there is no need to clear them on offline.
 162          */
 163         for_each_netdev(&init_net, dev) {
 164                 u32 prio = netprio_prio(parent_css, dev);
 165
 166                 ret = netprio_set_prio(css, dev, prio);
 167                 if (ret)
 168                         break;
 169         }
 170         rtnl_unlock();
 171         return ret;
 172 }
 173
 174 static void cgrp_css_free(struct cgroup_subsys_state *css)
 175 {
 176         kfree(css);
 177 }
 178
 179 static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
 180 {
 181         return css->cgroup->id;
 182 }
 183
 184 static int read_priomap(struct seq_file *sf, void *v)
 185 {
 186         struct net_device *dev;
 187
 188         rcu_read_lock();
 189         for_each_netdev_rcu(&init_net, dev)
 190                 seq_printf(sf, "%s %u\n", dev->name,
 191                            netprio_prio(seq_css(sf), dev));
 192         rcu_read_unlock();
 193         return 0;
 194 }
 195
 196 static ssize_t write_priomap(struct kernfs_open_file *of,
 197                              char *buf, size_t nbytes, loff_t off)
 198 {
 199         char devname[IFNAMSIZ + 1];
 200         struct net_device *dev;
 201         u32 prio;
 202         int ret;
 203
 204         if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
 205                 return -EINVAL;
 206
 207         dev = dev_get_by_name(&init_net, devname);
 208         if (!dev)
 209                 return -ENODEV;
 210
 211         cgroup_sk_alloc_disable();
 212
 213         rtnl_lock();
 214
 215         ret = netprio_set_prio(of_css(of), dev, prio);
 216
 217         rtnl_unlock();
 218         dev_put(dev);
 219         return ret ?: nbytes;
 220 }
 221
 222 static int update_netprio(const void *v, struct file *file, unsigned n)
 223 {
 224         int err;
 225         struct socket *sock = sock_from_file(file, &err);
 226         if (sock) {
 227                 spin_lock(&cgroup_sk_update_lock);
 228                 sock_cgroup_set_prioidx(&sock->sk->sk_cgrp_data,
 229                                         (unsigned long)v);
 230                 spin_unlock(&cgroup_sk_update_lock);
 231         }
 232         return 0;
 233 }
 234
 235 static void net_prio_attach(struct cgroup_taskset *tset)
 236 {
 237         struct task_struct *p;
 238         struct cgroup_subsys_state *css;
 239
 240         cgroup_taskset_for_each(p, css, tset) {
 241                 void *v = (void *)(unsigned long)css->cgroup->id;
 242
 243                 task_lock(p);
 244                 iterate_fd(p->files, 0, update_netprio, v);
 245                 task_unlock(p);
 246         }
 247 }
 248
 249 static struct cftype ss_files[] = {
 250         {
 251                 .name = "prioidx",
 252                 .read_u64 = read_prioidx,
 253         },
 254         {
 255                 .name = "ifpriomap",
 256                 .seq_show = read_priomap,
 257                 .write = write_priomap,
 258         },
 259         { }     /* terminate */
 260 };
 261
 262 struct cgroup_subsys net_prio_cgrp_subsys = {
 263         .css_alloc      = cgrp_css_alloc,
 264         .css_online     = cgrp_css_online,
 265         .css_free       = cgrp_css_free,
 266         .attach         = net_prio_attach,
 267         .legacy_cftypes = ss_files,
 268 };
 269
 270 static int netprio_device_event(struct notifier_block *unused,
 271                                 unsigned long event, void *ptr)
 272 {
 273         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 274         struct netprio_map *old;
 275
 276         /*
 277          * Note this is called with rtnl_lock held so we have update side
 278          * protection on our rcu assignments
 279          */
 280
 281         switch (event) {
 282         case NETDEV_UNREGISTER:
 283                 old = rtnl_dereference(dev->priomap);
 284                 RCU_INIT_POINTER(dev->priomap, NULL);
 285                 if (old)
 286                         kfree_rcu(old, rcu);
 287                 break;
 288         }
 289         return NOTIFY_DONE;
 290 }
 291
 292 static struct notifier_block netprio_device_notifier = {
 293         .notifier_call = netprio_device_event
 294 };
 295
 296 static int __init init_cgroup_netprio(void)
 297 {
 298         register_netdevice_notifier(&netprio_device_notifier);
 299         return 0;
 300 }
 301
 302 subsys_initcall(init_cgroup_netprio);
 303 MODULE_LICENSE("GPL v2");