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/module.h>
  18 #include <linux/string.h>
  19 #include <linux/errno.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/cgroup.h>
  22 #include <linux/rcupdate.h>
  23 #include <linux/atomic.h>
  24 #include <linux/sched/task.h>
  25
  26 #include <net/rtnetlink.h>
  27 #include <net/pkt_cls.h>
  28 #include <net/sock.h>
  29 #include <net/netprio_cgroup.h>
  30
  31 #include <linux/fdtable.h>
  32
  33 /*
  34  * netprio allocates per-net_device priomap array which is indexed by
  35  * css->id.  Limiting css ID to 16bits doesn't lose anything.
  36  */
  37 #define NETPRIO_ID_MAX          USHRT_MAX
  38
  39 #define PRIOMAP_MIN_SZ          128
  40
  41 /*
  42  * Extend @dev->priomap so that it's large enough to accommodate
  43  * @target_idx.  @dev->priomap.priomap_len > @target_idx after successful
  44  * return.  Must be called under rtnl lock.
  45  */
  46 static int extend_netdev_table(struct net_device *dev, u32 target_idx)
  47 {
  48         struct netprio_map *old, *new;
  49         size_t new_sz, new_len;
  50
  51         /* is the existing priomap large enough? */
  52         old = rtnl_dereference(dev->priomap);
  53         if (old && old->priomap_len > target_idx)
  54                 return 0;
  55
  56         /*
  57          * Determine the new size.  Let's keep it power-of-two.  We start
  58          * from PRIOMAP_MIN_SZ and double it until it's large enough to
  59          * accommodate @target_idx.
  60          */
  61         new_sz = PRIOMAP_MIN_SZ;
  62         while (true) {
  63                 new_len = (new_sz - offsetof(struct netprio_map, priomap)) /
  64                         sizeof(new->priomap[0]);
  65                 if (new_len > target_idx)
  66                         break;
  67                 new_sz *= 2;
  68                 /* overflowed? */
  69                 if (WARN_ON(new_sz < PRIOMAP_MIN_SZ))
  70                         return -ENOSPC;
  71         }
  72
  73         /* allocate & copy */
  74         new = kzalloc(new_sz, GFP_KERNEL);
  75         if (!new)
  76                 return -ENOMEM;
  77
  78         if (old)
  79                 memcpy(new->priomap, old->priomap,
  80                        old->priomap_len * sizeof(old->priomap[0]));
  81
  82         new->priomap_len = new_len;
  83
  84         /* install the new priomap */
  85         rcu_assign_pointer(dev->priomap, new);
  86         if (old)
  87                 kfree_rcu(old, rcu);
  88         return 0;
  89 }
  90
  91 /**
  92  * netprio_prio - return the effective netprio of a cgroup-net_device pair
  93  * @css: css part of the target pair
  94  * @dev: net_device part of the target pair
  95  *
  96  * Should be called under RCU read or rtnl lock.
  97  */
  98 static u32 netprio_prio(struct cgroup_subsys_state *css, struct net_device *dev)
  99 {
 100         struct netprio_map *map = rcu_dereference_rtnl(dev->priomap);
 101         int id = css->cgroup->id;
 102
 103         if (map && id < map->priomap_len)
 104                 return map->priomap[id];
 105         return 0;
 106 }
 107
 108 /**
 109  * netprio_set_prio - set netprio on a cgroup-net_device pair
 110  * @css: css part of the target pair
 111  * @dev: net_device part of the target pair
 112  * @prio: prio to set
 113  *
 114  * Set netprio to @prio on @css-@dev pair.  Should be called under rtnl
 115  * lock and may fail under memory pressure for non-zero @prio.
 116  */
 117 static int netprio_set_prio(struct cgroup_subsys_state *css,
 118                             struct net_device *dev, u32 prio)
 119 {
 120         struct netprio_map *map;
 121         int id = css->cgroup->id;
 122         int ret;
 123
 124         /* avoid extending priomap for zero writes */
 125         map = rtnl_dereference(dev->priomap);
 126         if (!prio && (!map || map->priomap_len <= id))
 127                 return 0;
 128
 129         ret = extend_netdev_table(dev, id);
 130         if (ret)
 131                 return ret;
 132
 133         map = rtnl_dereference(dev->priomap);
 134         map->priomap[id] = prio;
 135         return 0;
 136 }
 137
 138 static struct cgroup_subsys_state *
 139 cgrp_css_alloc(struct cgroup_subsys_state *parent_css)
 140 {
 141         struct cgroup_subsys_state *css;
 142
 143         css = kzalloc(sizeof(*css), GFP_KERNEL);
 144         if (!css)
 145                 return ERR_PTR(-ENOMEM);
 146
 147         return css;
 148 }
 149
 150 static int cgrp_css_online(struct cgroup_subsys_state *css)
 151 {
 152         struct cgroup_subsys_state *parent_css = css->parent;
 153         struct net_device *dev;
 154         int ret = 0;
 155
 156         if (css->id > NETPRIO_ID_MAX)
 157                 return -ENOSPC;
 158
 159         if (!parent_css)
 160                 return 0;
 161
 162         rtnl_lock();
 163         /*
 164          * Inherit prios from the parent.  As all prios are set during
 165          * onlining, there is no need to clear them on offline.
 166          */
 167         for_each_netdev(&init_net, dev) {
 168                 u32 prio = netprio_prio(parent_css, dev);
 169
 170                 ret = netprio_set_prio(css, dev, prio);
 171                 if (ret)
 172                         break;
 173         }
 174         rtnl_unlock();
 175         return ret;
 176 }
 177
 178 static void cgrp_css_free(struct cgroup_subsys_state *css)
 179 {
 180         kfree(css);
 181 }
 182
 183 static u64 read_prioidx(struct cgroup_subsys_state *css, struct cftype *cft)
 184 {
 185         return css->cgroup->id;
 186 }
 187
 188 static int read_priomap(struct seq_file *sf, void *v)
 189 {
 190         struct net_device *dev;
 191
 192         rcu_read_lock();
 193         for_each_netdev_rcu(&init_net, dev)
 194                 seq_printf(sf, "%s %u\n", dev->name,
 195                            netprio_prio(seq_css(sf), dev));
 196         rcu_read_unlock();
 197         return 0;
 198 }
 199
 200 static ssize_t write_priomap(struct kernfs_open_file *of,
 201                              char *buf, size_t nbytes, loff_t off)
 202 {
 203         char devname[IFNAMSIZ + 1];
 204         struct net_device *dev;
 205         u32 prio;
 206         int ret;
 207
 208         if (sscanf(buf, "%"__stringify(IFNAMSIZ)"s %u", devname, &prio) != 2)
 209                 return -EINVAL;
 210
 211         dev = dev_get_by_name(&init_net, devname);
 212         if (!dev)
 213                 return -ENODEV;
 214
 215         cgroup_sk_alloc_disable();
 216
 217         rtnl_lock();
 218
 219         ret = netprio_set_prio(of_css(of), dev, prio);
 220
 221         rtnl_unlock();
 222         dev_put(dev);
 223         return ret ?: nbytes;
 224 }
 225
 226 static int update_netprio(const void *v, struct file *file, unsigned n)
 227 {
 228         int err;
 229         struct socket *sock = sock_from_file(file, &err);
 230         if (sock) {
 231                 spin_lock(&cgroup_sk_update_lock);
 232                 sock_cgroup_set_prioidx(&sock->sk->sk_cgrp_data,
 233                                         (unsigned long)v);
 234                 spin_unlock(&cgroup_sk_update_lock);
 235         }
 236         return 0;
 237 }
 238
 239 static void net_prio_attach(struct cgroup_taskset *tset)
 240 {
 241         struct task_struct *p;
 242         struct cgroup_subsys_state *css;
 243
 244         cgroup_taskset_for_each(p, css, tset) {
 245                 void *v = (void *)(unsigned long)css->cgroup->id;
 246
 247                 task_lock(p);
 248                 iterate_fd(p->files, 0, update_netprio, v);
 249                 task_unlock(p);
 250         }
 251 }
 252
 253 static struct cftype ss_files[] = {
 254         {
 255                 .name = "prioidx",
 256                 .read_u64 = read_prioidx,
 257         },
 258         {
 259                 .name = "ifpriomap",
 260                 .seq_show = read_priomap,
 261                 .write = write_priomap,
 262         },
 263         { }     /* terminate */
 264 };
 265
 266 struct cgroup_subsys net_prio_cgrp_subsys = {
 267         .css_alloc      = cgrp_css_alloc,
 268         .css_online     = cgrp_css_online,
 269         .css_free       = cgrp_css_free,
 270         .attach         = net_prio_attach,
 271         .legacy_cftypes = ss_files,
 272 };
 273
 274 static int netprio_device_event(struct notifier_block *unused,
 275                                 unsigned long event, void *ptr)
 276 {
 277         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 278         struct netprio_map *old;
 279
 280         /*
 281          * Note this is called with rtnl_lock held so we have update side
 282          * protection on our rcu assignments
 283          */
 284
 285         switch (event) {
 286         case NETDEV_UNREGISTER:
 287                 old = rtnl_dereference(dev->priomap);
 288                 RCU_INIT_POINTER(dev->priomap, NULL);
 289                 if (old)
 290                         kfree_rcu(old, rcu);
 291                 break;
 292         }
 293         return NOTIFY_DONE;
 294 }
 295
 296 static struct notifier_block netprio_device_notifier = {
 297         .notifier_call = netprio_device_event
 298 };
 299
 300 static int __init init_cgroup_netprio(void)
 301 {
 302         register_netdevice_notifier(&netprio_device_notifier);
 303         return 0;
 304 }
 305
 306 subsys_initcall(init_cgroup_netprio);
 307 MODULE_LICENSE("GPL v2");