net/ipv4/tcp_memcontrol.c

   1 #include <net/tcp.h>
   2 #include <net/tcp_memcontrol.h>
   3 #include <net/sock.h>
   4 #include <net/ip.h>
   5 #include <linux/nsproxy.h>
   6 #include <linux/memcontrol.h>
   7 #include <linux/module.h>
   8
   9 static void memcg_tcp_enter_memory_pressure(struct sock *sk)
  10 {
  11         if (sk->sk_cgrp->memory_pressure)
  12                 sk->sk_cgrp->memory_pressure = 1;
  13 }
  14 EXPORT_SYMBOL(memcg_tcp_enter_memory_pressure);
  15
  16 int tcp_init_cgroup(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
  17 {
  18         /*
  19          * The root cgroup does not use res_counters, but rather,
  20          * rely on the data already collected by the network
  21          * subsystem
  22          */
  23         struct res_counter *res_parent = NULL;
  24         struct cg_proto *cg_proto, *parent_cg;
  25         struct mem_cgroup *parent = parent_mem_cgroup(memcg);
  26
  27         cg_proto = tcp_prot.proto_cgroup(memcg);
  28         if (!cg_proto)
  29                 return 0;
  30
  31         cg_proto->sysctl_mem[0] = sysctl_tcp_mem[0];
  32         cg_proto->sysctl_mem[1] = sysctl_tcp_mem[1];
  33         cg_proto->sysctl_mem[2] = sysctl_tcp_mem[2];
  34         cg_proto->memory_pressure = 0;
  35         cg_proto->memcg = memcg;
  36
  37         parent_cg = tcp_prot.proto_cgroup(parent);
  38         if (parent_cg)
  39                 res_parent = &parent_cg->memory_allocated;
  40
  41         res_counter_init(&cg_proto->memory_allocated, res_parent);
  42         percpu_counter_init(&cg_proto->sockets_allocated, 0);
  43
  44         return 0;
  45 }
  46 EXPORT_SYMBOL(tcp_init_cgroup);
  47
  48 void tcp_destroy_cgroup(struct mem_cgroup *memcg)
  49 {
  50         struct cg_proto *cg_proto;
  51
  52         cg_proto = tcp_prot.proto_cgroup(memcg);
  53         if (!cg_proto)
  54                 return;
  55
  56         percpu_counter_destroy(&cg_proto->sockets_allocated);
  57 }
  58 EXPORT_SYMBOL(tcp_destroy_cgroup);
  59
  60 static int tcp_update_limit(struct mem_cgroup *memcg, u64 val)
  61 {
  62         struct cg_proto *cg_proto;
  63         u64 old_lim;
  64         int i;
  65         int ret;
  66
  67         cg_proto = tcp_prot.proto_cgroup(memcg);
  68         if (!cg_proto)
  69                 return -EINVAL;
  70
  71         if (val > RES_COUNTER_MAX)
  72                 val = RES_COUNTER_MAX;
  73
  74         old_lim = res_counter_read_u64(&cg_proto->memory_allocated, RES_LIMIT);
  75         ret = res_counter_set_limit(&cg_proto->memory_allocated, val);
  76         if (ret)
  77                 return ret;
  78
  79         for (i = 0; i < 3; i++)
  80                 cg_proto->sysctl_mem[i] = min_t(long, val >> PAGE_SHIFT,
  81                                                 sysctl_tcp_mem[i]);
  82
  83         if (val == RES_COUNTER_MAX)
  84                 clear_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags);
  85         else if (val != RES_COUNTER_MAX) {
  86                 /*
  87                  * The active bit needs to be written after the static_key
  88                  * update. This is what guarantees that the socket activation
  89                  * function is the last one to run. See sock_update_memcg() for
  90                  * details, and note that we don't mark any socket as belonging
  91                  * to this memcg until that flag is up.
  92                  *
  93                  * We need to do this, because static_keys will span multiple
  94                  * sites, but we can't control their order. If we mark a socket
  95                  * as accounted, but the accounting functions are not patched in
  96                  * yet, we'll lose accounting.
  97                  *
  98                  * We never race with the readers in sock_update_memcg(),
  99                  * because when this value change, the code to process it is not
 100                  * patched in yet.
 101                  *
 102                  * The activated bit is used to guarantee that no two writers
 103                  * will do the update in the same memcg. Without that, we can't
 104                  * properly shutdown the static key.
 105                  */
 106                 if (!test_and_set_bit(MEMCG_SOCK_ACTIVATED, &cg_proto->flags))
 107                         static_key_slow_inc(&memcg_socket_limit_enabled);
 108                 set_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags);
 109         }
 110
 111         return 0;
 112 }
 113
 114 static int tcp_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
 115                             const char *buffer)
 116 {
 117         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
 118         unsigned long long val;
 119         int ret = 0;
 120
 121         switch (cft->private) {
 122         case RES_LIMIT:
 123                 /* see memcontrol.c */
 124                 ret = res_counter_memparse_write_strategy(buffer, &val);
 125                 if (ret)
 126                         break;
 127                 ret = tcp_update_limit(memcg, val);
 128                 break;
 129         default:
 130                 ret = -EINVAL;
 131                 break;
 132         }
 133         return ret;
 134 }
 135
 136 static u64 tcp_read_stat(struct mem_cgroup *memcg, int type, u64 default_val)
 137 {
 138         struct cg_proto *cg_proto;
 139
 140         cg_proto = tcp_prot.proto_cgroup(memcg);
 141         if (!cg_proto)
 142                 return default_val;
 143
 144         return res_counter_read_u64(&cg_proto->memory_allocated, type);
 145 }
 146
 147 static u64 tcp_read_usage(struct mem_cgroup *memcg)
 148 {
 149         struct cg_proto *cg_proto;
 150
 151         cg_proto = tcp_prot.proto_cgroup(memcg);
 152         if (!cg_proto)
 153                 return atomic_long_read(&tcp_memory_allocated) << PAGE_SHIFT;
 154
 155         return res_counter_read_u64(&cg_proto->memory_allocated, RES_USAGE);
 156 }
 157
 158 static u64 tcp_cgroup_read(struct cgroup_subsys_state *css, struct cftype *cft)
 159 {
 160         struct mem_cgroup *memcg = mem_cgroup_from_css(css);
 161         u64 val;
 162
 163         switch (cft->private) {
 164         case RES_LIMIT:
 165                 val = tcp_read_stat(memcg, RES_LIMIT, RES_COUNTER_MAX);
 166                 break;
 167         case RES_USAGE:
 168                 val = tcp_read_usage(memcg);
 169                 break;
 170         case RES_FAILCNT:
 171         case RES_MAX_USAGE:
 172                 val = tcp_read_stat(memcg, cft->private, 0);
 173                 break;
 174         default:
 175                 BUG();
 176         }
 177         return val;
 178 }
 179
 180 static int tcp_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
 181 {
 182         struct mem_cgroup *memcg;
 183         struct cg_proto *cg_proto;
 184
 185         memcg = mem_cgroup_from_css(css);
 186         cg_proto = tcp_prot.proto_cgroup(memcg);
 187         if (!cg_proto)
 188                 return 0;
 189
 190         switch (event) {
 191         case RES_MAX_USAGE:
 192                 res_counter_reset_max(&cg_proto->memory_allocated);
 193                 break;
 194         case RES_FAILCNT:
 195                 res_counter_reset_failcnt(&cg_proto->memory_allocated);
 196                 break;
 197         }
 198
 199         return 0;
 200 }
 201
 202 static struct cftype tcp_files[] = {
 203         {
 204                 .name = "kmem.tcp.limit_in_bytes",
 205                 .write_string = tcp_cgroup_write,
 206                 .read_u64 = tcp_cgroup_read,
 207                 .private = RES_LIMIT,
 208         },
 209         {
 210                 .name = "kmem.tcp.usage_in_bytes",
 211                 .read_u64 = tcp_cgroup_read,
 212                 .private = RES_USAGE,
 213         },
 214         {
 215                 .name = "kmem.tcp.failcnt",
 216                 .private = RES_FAILCNT,
 217                 .trigger = tcp_cgroup_reset,
 218                 .read_u64 = tcp_cgroup_read,
 219         },
 220         {
 221                 .name = "kmem.tcp.max_usage_in_bytes",
 222                 .private = RES_MAX_USAGE,
 223                 .trigger = tcp_cgroup_reset,
 224                 .read_u64 = tcp_cgroup_read,
 225         },
 226         { }     /* terminate */
 227 };
 228
 229 static int __init tcp_memcontrol_init(void)
 230 {
 231         WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, tcp_files));
 232         return 0;
 233 }
 234 __initcall(tcp_memcontrol_init);