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