| 1 | /* memcontrol.c - Memory Controller |
| 2 | * |
| 3 | * Copyright IBM Corporation, 2007 |
| 4 | * Author Balbir Singh <balbir@linux.vnet.ibm.com> |
| 5 | * |
| 6 | * Copyright 2007 OpenVZ SWsoft Inc |
| 7 | * Author: Pavel Emelianov <xemul@openvz.org> |
| 8 | * |
| 9 | * Memory thresholds |
| 10 | * Copyright (C) 2009 Nokia Corporation |
| 11 | * Author: Kirill A. Shutemov |
| 12 | * |
| 13 | * This program is free software; you can redistribute it and/or modify |
| 14 | * it under the terms of the GNU General Public License as published by |
| 15 | * the Free Software Foundation; either version 2 of the License, or |
| 16 | * (at your option) any later version. |
| 17 | * |
| 18 | * This program is distributed in the hope that it will be useful, |
| 19 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 20 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 21 | * GNU General Public License for more details. |
| 22 | */ |
| 23 | |
| 24 | #include <linux/res_counter.h> |
| 25 | #include <linux/memcontrol.h> |
| 26 | #include <linux/cgroup.h> |
| 27 | #include <linux/mm.h> |
| 28 | #include <linux/hugetlb.h> |
| 29 | #include <linux/pagemap.h> |
| 30 | #include <linux/smp.h> |
| 31 | #include <linux/page-flags.h> |
| 32 | #include <linux/backing-dev.h> |
| 33 | #include <linux/bit_spinlock.h> |
| 34 | #include <linux/rcupdate.h> |
| 35 | #include <linux/limits.h> |
| 36 | #include <linux/export.h> |
| 37 | #include <linux/mutex.h> |
| 38 | #include <linux/rbtree.h> |
| 39 | #include <linux/slab.h> |
| 40 | #include <linux/swap.h> |
| 41 | #include <linux/swapops.h> |
| 42 | #include <linux/spinlock.h> |
| 43 | #include <linux/eventfd.h> |
| 44 | #include <linux/sort.h> |
| 45 | #include <linux/fs.h> |
| 46 | #include <linux/seq_file.h> |
| 47 | #include <linux/vmalloc.h> |
| 48 | #include <linux/mm_inline.h> |
| 49 | #include <linux/page_cgroup.h> |
| 50 | #include <linux/cpu.h> |
| 51 | #include <linux/oom.h> |
| 52 | #include "internal.h" |
| 53 | #include <net/sock.h> |
| 54 | #include <net/tcp_memcontrol.h> |
| 55 | |
| 56 | #include <asm/uaccess.h> |
| 57 | |
| 58 | #include <trace/events/vmscan.h> |
| 59 | |
| 60 | struct cgroup_subsys mem_cgroup_subsys __read_mostly; |
| 61 | #define MEM_CGROUP_RECLAIM_RETRIES 5 |
| 62 | struct mem_cgroup *root_mem_cgroup __read_mostly; |
| 63 | |
| 64 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
| 65 | /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ |
| 66 | int do_swap_account __read_mostly; |
| 67 | |
| 68 | /* for remember boot option*/ |
| 69 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED |
| 70 | static int really_do_swap_account __initdata = 1; |
| 71 | #else |
| 72 | static int really_do_swap_account __initdata = 0; |
| 73 | #endif |
| 74 | |
| 75 | #else |
| 76 | #define do_swap_account (0) |
| 77 | #endif |
| 78 | |
| 79 | |
| 80 | /* |
| 81 | * Statistics for memory cgroup. |
| 82 | */ |
| 83 | enum mem_cgroup_stat_index { |
| 84 | /* |
| 85 | * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. |
| 86 | */ |
| 87 | MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ |
| 88 | MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ |
| 89 | MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ |
| 90 | MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ |
| 91 | MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ |
| 92 | MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */ |
| 93 | MEM_CGROUP_STAT_NSTATS, |
| 94 | }; |
| 95 | |
| 96 | enum mem_cgroup_events_index { |
| 97 | MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ |
| 98 | MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ |
| 99 | MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */ |
| 100 | MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */ |
| 101 | MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */ |
| 102 | MEM_CGROUP_EVENTS_NSTATS, |
| 103 | }; |
| 104 | /* |
| 105 | * Per memcg event counter is incremented at every pagein/pageout. With THP, |
| 106 | * it will be incremated by the number of pages. This counter is used for |
| 107 | * for trigger some periodic events. This is straightforward and better |
| 108 | * than using jiffies etc. to handle periodic memcg event. |
| 109 | */ |
| 110 | enum mem_cgroup_events_target { |
| 111 | MEM_CGROUP_TARGET_THRESH, |
| 112 | MEM_CGROUP_TARGET_SOFTLIMIT, |
| 113 | MEM_CGROUP_TARGET_NUMAINFO, |
| 114 | MEM_CGROUP_NTARGETS, |
| 115 | }; |
| 116 | #define THRESHOLDS_EVENTS_TARGET (128) |
| 117 | #define SOFTLIMIT_EVENTS_TARGET (1024) |
| 118 | #define NUMAINFO_EVENTS_TARGET (1024) |
| 119 | |
| 120 | struct mem_cgroup_stat_cpu { |
| 121 | long count[MEM_CGROUP_STAT_NSTATS]; |
| 122 | unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; |
| 123 | unsigned long targets[MEM_CGROUP_NTARGETS]; |
| 124 | }; |
| 125 | |
| 126 | struct mem_cgroup_reclaim_iter { |
| 127 | /* css_id of the last scanned hierarchy member */ |
| 128 | int position; |
| 129 | /* scan generation, increased every round-trip */ |
| 130 | unsigned int generation; |
| 131 | }; |
| 132 | |
| 133 | /* |
| 134 | * per-zone information in memory controller. |
| 135 | */ |
| 136 | struct mem_cgroup_per_zone { |
| 137 | struct lruvec lruvec; |
| 138 | unsigned long lru_size[NR_LRU_LISTS]; |
| 139 | |
| 140 | struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1]; |
| 141 | |
| 142 | struct zone_reclaim_stat reclaim_stat; |
| 143 | struct rb_node tree_node; /* RB tree node */ |
| 144 | unsigned long long usage_in_excess;/* Set to the value by which */ |
| 145 | /* the soft limit is exceeded*/ |
| 146 | bool on_tree; |
| 147 | struct mem_cgroup *memcg; /* Back pointer, we cannot */ |
| 148 | /* use container_of */ |
| 149 | }; |
| 150 | |
| 151 | struct mem_cgroup_per_node { |
| 152 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; |
| 153 | }; |
| 154 | |
| 155 | struct mem_cgroup_lru_info { |
| 156 | struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; |
| 157 | }; |
| 158 | |
| 159 | /* |
| 160 | * Cgroups above their limits are maintained in a RB-Tree, independent of |
| 161 | * their hierarchy representation |
| 162 | */ |
| 163 | |
| 164 | struct mem_cgroup_tree_per_zone { |
| 165 | struct rb_root rb_root; |
| 166 | spinlock_t lock; |
| 167 | }; |
| 168 | |
| 169 | struct mem_cgroup_tree_per_node { |
| 170 | struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; |
| 171 | }; |
| 172 | |
| 173 | struct mem_cgroup_tree { |
| 174 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; |
| 175 | }; |
| 176 | |
| 177 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; |
| 178 | |
| 179 | struct mem_cgroup_threshold { |
| 180 | struct eventfd_ctx *eventfd; |
| 181 | u64 threshold; |
| 182 | }; |
| 183 | |
| 184 | /* For threshold */ |
| 185 | struct mem_cgroup_threshold_ary { |
| 186 | /* An array index points to threshold just below usage. */ |
| 187 | int current_threshold; |
| 188 | /* Size of entries[] */ |
| 189 | unsigned int size; |
| 190 | /* Array of thresholds */ |
| 191 | struct mem_cgroup_threshold entries[0]; |
| 192 | }; |
| 193 | |
| 194 | struct mem_cgroup_thresholds { |
| 195 | /* Primary thresholds array */ |
| 196 | struct mem_cgroup_threshold_ary *primary; |
| 197 | /* |
| 198 | * Spare threshold array. |
| 199 | * This is needed to make mem_cgroup_unregister_event() "never fail". |
| 200 | * It must be able to store at least primary->size - 1 entries. |
| 201 | */ |
| 202 | struct mem_cgroup_threshold_ary *spare; |
| 203 | }; |
| 204 | |
| 205 | /* for OOM */ |
| 206 | struct mem_cgroup_eventfd_list { |
| 207 | struct list_head list; |
| 208 | struct eventfd_ctx *eventfd; |
| 209 | }; |
| 210 | |
| 211 | static void mem_cgroup_threshold(struct mem_cgroup *memcg); |
| 212 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg); |
| 213 | |
| 214 | /* |
| 215 | * The memory controller data structure. The memory controller controls both |
| 216 | * page cache and RSS per cgroup. We would eventually like to provide |
| 217 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, |
| 218 | * to help the administrator determine what knobs to tune. |
| 219 | * |
| 220 | * TODO: Add a water mark for the memory controller. Reclaim will begin when |
| 221 | * we hit the water mark. May be even add a low water mark, such that |
| 222 | * no reclaim occurs from a cgroup at it's low water mark, this is |
| 223 | * a feature that will be implemented much later in the future. |
| 224 | */ |
| 225 | struct mem_cgroup { |
| 226 | struct cgroup_subsys_state css; |
| 227 | /* |
| 228 | * the counter to account for memory usage |
| 229 | */ |
| 230 | struct res_counter res; |
| 231 | |
| 232 | union { |
| 233 | /* |
| 234 | * the counter to account for mem+swap usage. |
| 235 | */ |
| 236 | struct res_counter memsw; |
| 237 | |
| 238 | /* |
| 239 | * rcu_freeing is used only when freeing struct mem_cgroup, |
| 240 | * so put it into a union to avoid wasting more memory. |
| 241 | * It must be disjoint from the css field. It could be |
| 242 | * in a union with the res field, but res plays a much |
| 243 | * larger part in mem_cgroup life than memsw, and might |
| 244 | * be of interest, even at time of free, when debugging. |
| 245 | * So share rcu_head with the less interesting memsw. |
| 246 | */ |
| 247 | struct rcu_head rcu_freeing; |
| 248 | /* |
| 249 | * But when using vfree(), that cannot be done at |
| 250 | * interrupt time, so we must then queue the work. |
| 251 | */ |
| 252 | struct work_struct work_freeing; |
| 253 | }; |
| 254 | |
| 255 | /* |
| 256 | * Per cgroup active and inactive list, similar to the |
| 257 | * per zone LRU lists. |
| 258 | */ |
| 259 | struct mem_cgroup_lru_info info; |
| 260 | int last_scanned_node; |
| 261 | #if MAX_NUMNODES > 1 |
| 262 | nodemask_t scan_nodes; |
| 263 | atomic_t numainfo_events; |
| 264 | atomic_t numainfo_updating; |
| 265 | #endif |
| 266 | /* |
| 267 | * Should the accounting and control be hierarchical, per subtree? |
| 268 | */ |
| 269 | bool use_hierarchy; |
| 270 | |
| 271 | bool oom_lock; |
| 272 | atomic_t under_oom; |
| 273 | |
| 274 | atomic_t refcnt; |
| 275 | |
| 276 | int swappiness; |
| 277 | /* OOM-Killer disable */ |
| 278 | int oom_kill_disable; |
| 279 | |
| 280 | /* set when res.limit == memsw.limit */ |
| 281 | bool memsw_is_minimum; |
| 282 | |
| 283 | /* protect arrays of thresholds */ |
| 284 | struct mutex thresholds_lock; |
| 285 | |
| 286 | /* thresholds for memory usage. RCU-protected */ |
| 287 | struct mem_cgroup_thresholds thresholds; |
| 288 | |
| 289 | /* thresholds for mem+swap usage. RCU-protected */ |
| 290 | struct mem_cgroup_thresholds memsw_thresholds; |
| 291 | |
| 292 | /* For oom notifier event fd */ |
| 293 | struct list_head oom_notify; |
| 294 | |
| 295 | /* |
| 296 | * Should we move charges of a task when a task is moved into this |
| 297 | * mem_cgroup ? And what type of charges should we move ? |
| 298 | */ |
| 299 | unsigned long move_charge_at_immigrate; |
| 300 | /* |
| 301 | * percpu counter. |
| 302 | */ |
| 303 | struct mem_cgroup_stat_cpu *stat; |
| 304 | /* |
| 305 | * used when a cpu is offlined or other synchronizations |
| 306 | * See mem_cgroup_read_stat(). |
| 307 | */ |
| 308 | struct mem_cgroup_stat_cpu nocpu_base; |
| 309 | spinlock_t pcp_counter_lock; |
| 310 | |
| 311 | #ifdef CONFIG_INET |
| 312 | struct tcp_memcontrol tcp_mem; |
| 313 | #endif |
| 314 | }; |
| 315 | |
| 316 | /* Stuffs for move charges at task migration. */ |
| 317 | /* |
| 318 | * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a |
| 319 | * left-shifted bitmap of these types. |
| 320 | */ |
| 321 | enum move_type { |
| 322 | MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ |
| 323 | MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ |
| 324 | NR_MOVE_TYPE, |
| 325 | }; |
| 326 | |
| 327 | /* "mc" and its members are protected by cgroup_mutex */ |
| 328 | static struct move_charge_struct { |
| 329 | spinlock_t lock; /* for from, to */ |
| 330 | struct mem_cgroup *from; |
| 331 | struct mem_cgroup *to; |
| 332 | unsigned long precharge; |
| 333 | unsigned long moved_charge; |
| 334 | unsigned long moved_swap; |
| 335 | struct task_struct *moving_task; /* a task moving charges */ |
| 336 | wait_queue_head_t waitq; /* a waitq for other context */ |
| 337 | } mc = { |
| 338 | .lock = __SPIN_LOCK_UNLOCKED(mc.lock), |
| 339 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), |
| 340 | }; |
| 341 | |
| 342 | static bool move_anon(void) |
| 343 | { |
| 344 | return test_bit(MOVE_CHARGE_TYPE_ANON, |
| 345 | &mc.to->move_charge_at_immigrate); |
| 346 | } |
| 347 | |
| 348 | static bool move_file(void) |
| 349 | { |
| 350 | return test_bit(MOVE_CHARGE_TYPE_FILE, |
| 351 | &mc.to->move_charge_at_immigrate); |
| 352 | } |
| 353 | |
| 354 | /* |
| 355 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft |
| 356 | * limit reclaim to prevent infinite loops, if they ever occur. |
| 357 | */ |
| 358 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) |
| 359 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) |
| 360 | |
| 361 | enum charge_type { |
| 362 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, |
| 363 | MEM_CGROUP_CHARGE_TYPE_MAPPED, |
| 364 | MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ |
| 365 | MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ |
| 366 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ |
| 367 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ |
| 368 | NR_CHARGE_TYPE, |
| 369 | }; |
| 370 | |
| 371 | /* for encoding cft->private value on file */ |
| 372 | #define _MEM (0) |
| 373 | #define _MEMSWAP (1) |
| 374 | #define _OOM_TYPE (2) |
| 375 | #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) |
| 376 | #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) |
| 377 | #define MEMFILE_ATTR(val) ((val) & 0xffff) |
| 378 | /* Used for OOM nofiier */ |
| 379 | #define OOM_CONTROL (0) |
| 380 | |
| 381 | /* |
| 382 | * Reclaim flags for mem_cgroup_hierarchical_reclaim |
| 383 | */ |
| 384 | #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 |
| 385 | #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) |
| 386 | #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 |
| 387 | #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) |
| 388 | |
| 389 | static void mem_cgroup_get(struct mem_cgroup *memcg); |
| 390 | static void mem_cgroup_put(struct mem_cgroup *memcg); |
| 391 | |
| 392 | /* Writing them here to avoid exposing memcg's inner layout */ |
| 393 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM |
| 394 | #include <net/sock.h> |
| 395 | #include <net/ip.h> |
| 396 | |
| 397 | static bool mem_cgroup_is_root(struct mem_cgroup *memcg); |
| 398 | void sock_update_memcg(struct sock *sk) |
| 399 | { |
| 400 | if (mem_cgroup_sockets_enabled) { |
| 401 | struct mem_cgroup *memcg; |
| 402 | |
| 403 | BUG_ON(!sk->sk_prot->proto_cgroup); |
| 404 | |
| 405 | /* Socket cloning can throw us here with sk_cgrp already |
| 406 | * filled. It won't however, necessarily happen from |
| 407 | * process context. So the test for root memcg given |
| 408 | * the current task's memcg won't help us in this case. |
| 409 | * |
| 410 | * Respecting the original socket's memcg is a better |
| 411 | * decision in this case. |
| 412 | */ |
| 413 | if (sk->sk_cgrp) { |
| 414 | BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg)); |
| 415 | mem_cgroup_get(sk->sk_cgrp->memcg); |
| 416 | return; |
| 417 | } |
| 418 | |
| 419 | rcu_read_lock(); |
| 420 | memcg = mem_cgroup_from_task(current); |
| 421 | if (!mem_cgroup_is_root(memcg)) { |
| 422 | mem_cgroup_get(memcg); |
| 423 | sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg); |
| 424 | } |
| 425 | rcu_read_unlock(); |
| 426 | } |
| 427 | } |
| 428 | EXPORT_SYMBOL(sock_update_memcg); |
| 429 | |
| 430 | void sock_release_memcg(struct sock *sk) |
| 431 | { |
| 432 | if (mem_cgroup_sockets_enabled && sk->sk_cgrp) { |
| 433 | struct mem_cgroup *memcg; |
| 434 | WARN_ON(!sk->sk_cgrp->memcg); |
| 435 | memcg = sk->sk_cgrp->memcg; |
| 436 | mem_cgroup_put(memcg); |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | #ifdef CONFIG_INET |
| 441 | struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg) |
| 442 | { |
| 443 | if (!memcg || mem_cgroup_is_root(memcg)) |
| 444 | return NULL; |
| 445 | |
| 446 | return &memcg->tcp_mem.cg_proto; |
| 447 | } |
| 448 | EXPORT_SYMBOL(tcp_proto_cgroup); |
| 449 | #endif /* CONFIG_INET */ |
| 450 | #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */ |
| 451 | |
| 452 | static void drain_all_stock_async(struct mem_cgroup *memcg); |
| 453 | |
| 454 | static struct mem_cgroup_per_zone * |
| 455 | mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid) |
| 456 | { |
| 457 | return &memcg->info.nodeinfo[nid]->zoneinfo[zid]; |
| 458 | } |
| 459 | |
| 460 | struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg) |
| 461 | { |
| 462 | return &memcg->css; |
| 463 | } |
| 464 | |
| 465 | static struct mem_cgroup_per_zone * |
| 466 | page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page) |
| 467 | { |
| 468 | int nid = page_to_nid(page); |
| 469 | int zid = page_zonenum(page); |
| 470 | |
| 471 | return mem_cgroup_zoneinfo(memcg, nid, zid); |
| 472 | } |
| 473 | |
| 474 | static struct mem_cgroup_tree_per_zone * |
| 475 | soft_limit_tree_node_zone(int nid, int zid) |
| 476 | { |
| 477 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; |
| 478 | } |
| 479 | |
| 480 | static struct mem_cgroup_tree_per_zone * |
| 481 | soft_limit_tree_from_page(struct page *page) |
| 482 | { |
| 483 | int nid = page_to_nid(page); |
| 484 | int zid = page_zonenum(page); |
| 485 | |
| 486 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; |
| 487 | } |
| 488 | |
| 489 | static void |
| 490 | __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg, |
| 491 | struct mem_cgroup_per_zone *mz, |
| 492 | struct mem_cgroup_tree_per_zone *mctz, |
| 493 | unsigned long long new_usage_in_excess) |
| 494 | { |
| 495 | struct rb_node **p = &mctz->rb_root.rb_node; |
| 496 | struct rb_node *parent = NULL; |
| 497 | struct mem_cgroup_per_zone *mz_node; |
| 498 | |
| 499 | if (mz->on_tree) |
| 500 | return; |
| 501 | |
| 502 | mz->usage_in_excess = new_usage_in_excess; |
| 503 | if (!mz->usage_in_excess) |
| 504 | return; |
| 505 | while (*p) { |
| 506 | parent = *p; |
| 507 | mz_node = rb_entry(parent, struct mem_cgroup_per_zone, |
| 508 | tree_node); |
| 509 | if (mz->usage_in_excess < mz_node->usage_in_excess) |
| 510 | p = &(*p)->rb_left; |
| 511 | /* |
| 512 | * We can't avoid mem cgroups that are over their soft |
| 513 | * limit by the same amount |
| 514 | */ |
| 515 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) |
| 516 | p = &(*p)->rb_right; |
| 517 | } |
| 518 | rb_link_node(&mz->tree_node, parent, p); |
| 519 | rb_insert_color(&mz->tree_node, &mctz->rb_root); |
| 520 | mz->on_tree = true; |
| 521 | } |
| 522 | |
| 523 | static void |
| 524 | __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, |
| 525 | struct mem_cgroup_per_zone *mz, |
| 526 | struct mem_cgroup_tree_per_zone *mctz) |
| 527 | { |
| 528 | if (!mz->on_tree) |
| 529 | return; |
| 530 | rb_erase(&mz->tree_node, &mctz->rb_root); |
| 531 | mz->on_tree = false; |
| 532 | } |
| 533 | |
| 534 | static void |
| 535 | mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, |
| 536 | struct mem_cgroup_per_zone *mz, |
| 537 | struct mem_cgroup_tree_per_zone *mctz) |
| 538 | { |
| 539 | spin_lock(&mctz->lock); |
| 540 | __mem_cgroup_remove_exceeded(memcg, mz, mctz); |
| 541 | spin_unlock(&mctz->lock); |
| 542 | } |
| 543 | |
| 544 | |
| 545 | static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page) |
| 546 | { |
| 547 | unsigned long long excess; |
| 548 | struct mem_cgroup_per_zone *mz; |
| 549 | struct mem_cgroup_tree_per_zone *mctz; |
| 550 | int nid = page_to_nid(page); |
| 551 | int zid = page_zonenum(page); |
| 552 | mctz = soft_limit_tree_from_page(page); |
| 553 | |
| 554 | /* |
| 555 | * Necessary to update all ancestors when hierarchy is used. |
| 556 | * because their event counter is not touched. |
| 557 | */ |
| 558 | for (; memcg; memcg = parent_mem_cgroup(memcg)) { |
| 559 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
| 560 | excess = res_counter_soft_limit_excess(&memcg->res); |
| 561 | /* |
| 562 | * We have to update the tree if mz is on RB-tree or |
| 563 | * mem is over its softlimit. |
| 564 | */ |
| 565 | if (excess || mz->on_tree) { |
| 566 | spin_lock(&mctz->lock); |
| 567 | /* if on-tree, remove it */ |
| 568 | if (mz->on_tree) |
| 569 | __mem_cgroup_remove_exceeded(memcg, mz, mctz); |
| 570 | /* |
| 571 | * Insert again. mz->usage_in_excess will be updated. |
| 572 | * If excess is 0, no tree ops. |
| 573 | */ |
| 574 | __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess); |
| 575 | spin_unlock(&mctz->lock); |
| 576 | } |
| 577 | } |
| 578 | } |
| 579 | |
| 580 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg) |
| 581 | { |
| 582 | int node, zone; |
| 583 | struct mem_cgroup_per_zone *mz; |
| 584 | struct mem_cgroup_tree_per_zone *mctz; |
| 585 | |
| 586 | for_each_node(node) { |
| 587 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
| 588 | mz = mem_cgroup_zoneinfo(memcg, node, zone); |
| 589 | mctz = soft_limit_tree_node_zone(node, zone); |
| 590 | mem_cgroup_remove_exceeded(memcg, mz, mctz); |
| 591 | } |
| 592 | } |
| 593 | } |
| 594 | |
| 595 | static struct mem_cgroup_per_zone * |
| 596 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) |
| 597 | { |
| 598 | struct rb_node *rightmost = NULL; |
| 599 | struct mem_cgroup_per_zone *mz; |
| 600 | |
| 601 | retry: |
| 602 | mz = NULL; |
| 603 | rightmost = rb_last(&mctz->rb_root); |
| 604 | if (!rightmost) |
| 605 | goto done; /* Nothing to reclaim from */ |
| 606 | |
| 607 | mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); |
| 608 | /* |
| 609 | * Remove the node now but someone else can add it back, |
| 610 | * we will to add it back at the end of reclaim to its correct |
| 611 | * position in the tree. |
| 612 | */ |
| 613 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); |
| 614 | if (!res_counter_soft_limit_excess(&mz->memcg->res) || |
| 615 | !css_tryget(&mz->memcg->css)) |
| 616 | goto retry; |
| 617 | done: |
| 618 | return mz; |
| 619 | } |
| 620 | |
| 621 | static struct mem_cgroup_per_zone * |
| 622 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) |
| 623 | { |
| 624 | struct mem_cgroup_per_zone *mz; |
| 625 | |
| 626 | spin_lock(&mctz->lock); |
| 627 | mz = __mem_cgroup_largest_soft_limit_node(mctz); |
| 628 | spin_unlock(&mctz->lock); |
| 629 | return mz; |
| 630 | } |
| 631 | |
| 632 | /* |
| 633 | * Implementation Note: reading percpu statistics for memcg. |
| 634 | * |
| 635 | * Both of vmstat[] and percpu_counter has threshold and do periodic |
| 636 | * synchronization to implement "quick" read. There are trade-off between |
| 637 | * reading cost and precision of value. Then, we may have a chance to implement |
| 638 | * a periodic synchronizion of counter in memcg's counter. |
| 639 | * |
| 640 | * But this _read() function is used for user interface now. The user accounts |
| 641 | * memory usage by memory cgroup and he _always_ requires exact value because |
| 642 | * he accounts memory. Even if we provide quick-and-fuzzy read, we always |
| 643 | * have to visit all online cpus and make sum. So, for now, unnecessary |
| 644 | * synchronization is not implemented. (just implemented for cpu hotplug) |
| 645 | * |
| 646 | * If there are kernel internal actions which can make use of some not-exact |
| 647 | * value, and reading all cpu value can be performance bottleneck in some |
| 648 | * common workload, threashold and synchonization as vmstat[] should be |
| 649 | * implemented. |
| 650 | */ |
| 651 | static long mem_cgroup_read_stat(struct mem_cgroup *memcg, |
| 652 | enum mem_cgroup_stat_index idx) |
| 653 | { |
| 654 | long val = 0; |
| 655 | int cpu; |
| 656 | |
| 657 | get_online_cpus(); |
| 658 | for_each_online_cpu(cpu) |
| 659 | val += per_cpu(memcg->stat->count[idx], cpu); |
| 660 | #ifdef CONFIG_HOTPLUG_CPU |
| 661 | spin_lock(&memcg->pcp_counter_lock); |
| 662 | val += memcg->nocpu_base.count[idx]; |
| 663 | spin_unlock(&memcg->pcp_counter_lock); |
| 664 | #endif |
| 665 | put_online_cpus(); |
| 666 | return val; |
| 667 | } |
| 668 | |
| 669 | static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg, |
| 670 | bool charge) |
| 671 | { |
| 672 | int val = (charge) ? 1 : -1; |
| 673 | this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); |
| 674 | } |
| 675 | |
| 676 | static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg, |
| 677 | enum mem_cgroup_events_index idx) |
| 678 | { |
| 679 | unsigned long val = 0; |
| 680 | int cpu; |
| 681 | |
| 682 | for_each_online_cpu(cpu) |
| 683 | val += per_cpu(memcg->stat->events[idx], cpu); |
| 684 | #ifdef CONFIG_HOTPLUG_CPU |
| 685 | spin_lock(&memcg->pcp_counter_lock); |
| 686 | val += memcg->nocpu_base.events[idx]; |
| 687 | spin_unlock(&memcg->pcp_counter_lock); |
| 688 | #endif |
| 689 | return val; |
| 690 | } |
| 691 | |
| 692 | static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, |
| 693 | bool anon, int nr_pages) |
| 694 | { |
| 695 | preempt_disable(); |
| 696 | |
| 697 | /* |
| 698 | * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is |
| 699 | * counted as CACHE even if it's on ANON LRU. |
| 700 | */ |
| 701 | if (anon) |
| 702 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS], |
| 703 | nr_pages); |
| 704 | else |
| 705 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE], |
| 706 | nr_pages); |
| 707 | |
| 708 | /* pagein of a big page is an event. So, ignore page size */ |
| 709 | if (nr_pages > 0) |
| 710 | __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); |
| 711 | else { |
| 712 | __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); |
| 713 | nr_pages = -nr_pages; /* for event */ |
| 714 | } |
| 715 | |
| 716 | __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); |
| 717 | |
| 718 | preempt_enable(); |
| 719 | } |
| 720 | |
| 721 | unsigned long |
| 722 | mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid, |
| 723 | unsigned int lru_mask) |
| 724 | { |
| 725 | struct mem_cgroup_per_zone *mz; |
| 726 | enum lru_list lru; |
| 727 | unsigned long ret = 0; |
| 728 | |
| 729 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
| 730 | |
| 731 | for_each_lru(lru) { |
| 732 | if (BIT(lru) & lru_mask) |
| 733 | ret += mz->lru_size[lru]; |
| 734 | } |
| 735 | return ret; |
| 736 | } |
| 737 | |
| 738 | static unsigned long |
| 739 | mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, |
| 740 | int nid, unsigned int lru_mask) |
| 741 | { |
| 742 | u64 total = 0; |
| 743 | int zid; |
| 744 | |
| 745 | for (zid = 0; zid < MAX_NR_ZONES; zid++) |
| 746 | total += mem_cgroup_zone_nr_lru_pages(memcg, |
| 747 | nid, zid, lru_mask); |
| 748 | |
| 749 | return total; |
| 750 | } |
| 751 | |
| 752 | static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, |
| 753 | unsigned int lru_mask) |
| 754 | { |
| 755 | int nid; |
| 756 | u64 total = 0; |
| 757 | |
| 758 | for_each_node_state(nid, N_HIGH_MEMORY) |
| 759 | total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask); |
| 760 | return total; |
| 761 | } |
| 762 | |
| 763 | static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, |
| 764 | enum mem_cgroup_events_target target) |
| 765 | { |
| 766 | unsigned long val, next; |
| 767 | |
| 768 | val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]); |
| 769 | next = __this_cpu_read(memcg->stat->targets[target]); |
| 770 | /* from time_after() in jiffies.h */ |
| 771 | if ((long)next - (long)val < 0) { |
| 772 | switch (target) { |
| 773 | case MEM_CGROUP_TARGET_THRESH: |
| 774 | next = val + THRESHOLDS_EVENTS_TARGET; |
| 775 | break; |
| 776 | case MEM_CGROUP_TARGET_SOFTLIMIT: |
| 777 | next = val + SOFTLIMIT_EVENTS_TARGET; |
| 778 | break; |
| 779 | case MEM_CGROUP_TARGET_NUMAINFO: |
| 780 | next = val + NUMAINFO_EVENTS_TARGET; |
| 781 | break; |
| 782 | default: |
| 783 | break; |
| 784 | } |
| 785 | __this_cpu_write(memcg->stat->targets[target], next); |
| 786 | return true; |
| 787 | } |
| 788 | return false; |
| 789 | } |
| 790 | |
| 791 | /* |
| 792 | * Check events in order. |
| 793 | * |
| 794 | */ |
| 795 | static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) |
| 796 | { |
| 797 | preempt_disable(); |
| 798 | /* threshold event is triggered in finer grain than soft limit */ |
| 799 | if (unlikely(mem_cgroup_event_ratelimit(memcg, |
| 800 | MEM_CGROUP_TARGET_THRESH))) { |
| 801 | bool do_softlimit; |
| 802 | bool do_numainfo __maybe_unused; |
| 803 | |
| 804 | do_softlimit = mem_cgroup_event_ratelimit(memcg, |
| 805 | MEM_CGROUP_TARGET_SOFTLIMIT); |
| 806 | #if MAX_NUMNODES > 1 |
| 807 | do_numainfo = mem_cgroup_event_ratelimit(memcg, |
| 808 | MEM_CGROUP_TARGET_NUMAINFO); |
| 809 | #endif |
| 810 | preempt_enable(); |
| 811 | |
| 812 | mem_cgroup_threshold(memcg); |
| 813 | if (unlikely(do_softlimit)) |
| 814 | mem_cgroup_update_tree(memcg, page); |
| 815 | #if MAX_NUMNODES > 1 |
| 816 | if (unlikely(do_numainfo)) |
| 817 | atomic_inc(&memcg->numainfo_events); |
| 818 | #endif |
| 819 | } else |
| 820 | preempt_enable(); |
| 821 | } |
| 822 | |
| 823 | struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) |
| 824 | { |
| 825 | return container_of(cgroup_subsys_state(cont, |
| 826 | mem_cgroup_subsys_id), struct mem_cgroup, |
| 827 | css); |
| 828 | } |
| 829 | |
| 830 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) |
| 831 | { |
| 832 | /* |
| 833 | * mm_update_next_owner() may clear mm->owner to NULL |
| 834 | * if it races with swapoff, page migration, etc. |
| 835 | * So this can be called with p == NULL. |
| 836 | */ |
| 837 | if (unlikely(!p)) |
| 838 | return NULL; |
| 839 | |
| 840 | return container_of(task_subsys_state(p, mem_cgroup_subsys_id), |
| 841 | struct mem_cgroup, css); |
| 842 | } |
| 843 | |
| 844 | struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) |
| 845 | { |
| 846 | struct mem_cgroup *memcg = NULL; |
| 847 | |
| 848 | if (!mm) |
| 849 | return NULL; |
| 850 | /* |
| 851 | * Because we have no locks, mm->owner's may be being moved to other |
| 852 | * cgroup. We use css_tryget() here even if this looks |
| 853 | * pessimistic (rather than adding locks here). |
| 854 | */ |
| 855 | rcu_read_lock(); |
| 856 | do { |
| 857 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); |
| 858 | if (unlikely(!memcg)) |
| 859 | break; |
| 860 | } while (!css_tryget(&memcg->css)); |
| 861 | rcu_read_unlock(); |
| 862 | return memcg; |
| 863 | } |
| 864 | |
| 865 | /** |
| 866 | * mem_cgroup_iter - iterate over memory cgroup hierarchy |
| 867 | * @root: hierarchy root |
| 868 | * @prev: previously returned memcg, NULL on first invocation |
| 869 | * @reclaim: cookie for shared reclaim walks, NULL for full walks |
| 870 | * |
| 871 | * Returns references to children of the hierarchy below @root, or |
| 872 | * @root itself, or %NULL after a full round-trip. |
| 873 | * |
| 874 | * Caller must pass the return value in @prev on subsequent |
| 875 | * invocations for reference counting, or use mem_cgroup_iter_break() |
| 876 | * to cancel a hierarchy walk before the round-trip is complete. |
| 877 | * |
| 878 | * Reclaimers can specify a zone and a priority level in @reclaim to |
| 879 | * divide up the memcgs in the hierarchy among all concurrent |
| 880 | * reclaimers operating on the same zone and priority. |
| 881 | */ |
| 882 | struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, |
| 883 | struct mem_cgroup *prev, |
| 884 | struct mem_cgroup_reclaim_cookie *reclaim) |
| 885 | { |
| 886 | struct mem_cgroup *memcg = NULL; |
| 887 | int id = 0; |
| 888 | |
| 889 | if (mem_cgroup_disabled()) |
| 890 | return NULL; |
| 891 | |
| 892 | if (!root) |
| 893 | root = root_mem_cgroup; |
| 894 | |
| 895 | if (prev && !reclaim) |
| 896 | id = css_id(&prev->css); |
| 897 | |
| 898 | if (prev && prev != root) |
| 899 | css_put(&prev->css); |
| 900 | |
| 901 | if (!root->use_hierarchy && root != root_mem_cgroup) { |
| 902 | if (prev) |
| 903 | return NULL; |
| 904 | return root; |
| 905 | } |
| 906 | |
| 907 | while (!memcg) { |
| 908 | struct mem_cgroup_reclaim_iter *uninitialized_var(iter); |
| 909 | struct cgroup_subsys_state *css; |
| 910 | |
| 911 | if (reclaim) { |
| 912 | int nid = zone_to_nid(reclaim->zone); |
| 913 | int zid = zone_idx(reclaim->zone); |
| 914 | struct mem_cgroup_per_zone *mz; |
| 915 | |
| 916 | mz = mem_cgroup_zoneinfo(root, nid, zid); |
| 917 | iter = &mz->reclaim_iter[reclaim->priority]; |
| 918 | if (prev && reclaim->generation != iter->generation) |
| 919 | return NULL; |
| 920 | id = iter->position; |
| 921 | } |
| 922 | |
| 923 | rcu_read_lock(); |
| 924 | css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id); |
| 925 | if (css) { |
| 926 | if (css == &root->css || css_tryget(css)) |
| 927 | memcg = container_of(css, |
| 928 | struct mem_cgroup, css); |
| 929 | } else |
| 930 | id = 0; |
| 931 | rcu_read_unlock(); |
| 932 | |
| 933 | if (reclaim) { |
| 934 | iter->position = id; |
| 935 | if (!css) |
| 936 | iter->generation++; |
| 937 | else if (!prev && memcg) |
| 938 | reclaim->generation = iter->generation; |
| 939 | } |
| 940 | |
| 941 | if (prev && !css) |
| 942 | return NULL; |
| 943 | } |
| 944 | return memcg; |
| 945 | } |
| 946 | |
| 947 | /** |
| 948 | * mem_cgroup_iter_break - abort a hierarchy walk prematurely |
| 949 | * @root: hierarchy root |
| 950 | * @prev: last visited hierarchy member as returned by mem_cgroup_iter() |
| 951 | */ |
| 952 | void mem_cgroup_iter_break(struct mem_cgroup *root, |
| 953 | struct mem_cgroup *prev) |
| 954 | { |
| 955 | if (!root) |
| 956 | root = root_mem_cgroup; |
| 957 | if (prev && prev != root) |
| 958 | css_put(&prev->css); |
| 959 | } |
| 960 | |
| 961 | /* |
| 962 | * Iteration constructs for visiting all cgroups (under a tree). If |
| 963 | * loops are exited prematurely (break), mem_cgroup_iter_break() must |
| 964 | * be used for reference counting. |
| 965 | */ |
| 966 | #define for_each_mem_cgroup_tree(iter, root) \ |
| 967 | for (iter = mem_cgroup_iter(root, NULL, NULL); \ |
| 968 | iter != NULL; \ |
| 969 | iter = mem_cgroup_iter(root, iter, NULL)) |
| 970 | |
| 971 | #define for_each_mem_cgroup(iter) \ |
| 972 | for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ |
| 973 | iter != NULL; \ |
| 974 | iter = mem_cgroup_iter(NULL, iter, NULL)) |
| 975 | |
| 976 | static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg) |
| 977 | { |
| 978 | return (memcg == root_mem_cgroup); |
| 979 | } |
| 980 | |
| 981 | void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) |
| 982 | { |
| 983 | struct mem_cgroup *memcg; |
| 984 | |
| 985 | if (!mm) |
| 986 | return; |
| 987 | |
| 988 | rcu_read_lock(); |
| 989 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); |
| 990 | if (unlikely(!memcg)) |
| 991 | goto out; |
| 992 | |
| 993 | switch (idx) { |
| 994 | case PGFAULT: |
| 995 | this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]); |
| 996 | break; |
| 997 | case PGMAJFAULT: |
| 998 | this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]); |
| 999 | break; |
| 1000 | default: |
| 1001 | BUG(); |
| 1002 | } |
| 1003 | out: |
| 1004 | rcu_read_unlock(); |
| 1005 | } |
| 1006 | EXPORT_SYMBOL(mem_cgroup_count_vm_event); |
| 1007 | |
| 1008 | /** |
| 1009 | * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg |
| 1010 | * @zone: zone of the wanted lruvec |
| 1011 | * @mem: memcg of the wanted lruvec |
| 1012 | * |
| 1013 | * Returns the lru list vector holding pages for the given @zone and |
| 1014 | * @mem. This can be the global zone lruvec, if the memory controller |
| 1015 | * is disabled. |
| 1016 | */ |
| 1017 | struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone, |
| 1018 | struct mem_cgroup *memcg) |
| 1019 | { |
| 1020 | struct mem_cgroup_per_zone *mz; |
| 1021 | |
| 1022 | if (mem_cgroup_disabled()) |
| 1023 | return &zone->lruvec; |
| 1024 | |
| 1025 | mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone)); |
| 1026 | return &mz->lruvec; |
| 1027 | } |
| 1028 | |
| 1029 | /* |
| 1030 | * Following LRU functions are allowed to be used without PCG_LOCK. |
| 1031 | * Operations are called by routine of global LRU independently from memcg. |
| 1032 | * What we have to take care of here is validness of pc->mem_cgroup. |
| 1033 | * |
| 1034 | * Changes to pc->mem_cgroup happens when |
| 1035 | * 1. charge |
| 1036 | * 2. moving account |
| 1037 | * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. |
| 1038 | * It is added to LRU before charge. |
| 1039 | * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. |
| 1040 | * When moving account, the page is not on LRU. It's isolated. |
| 1041 | */ |
| 1042 | |
| 1043 | /** |
| 1044 | * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec |
| 1045 | * @zone: zone of the page |
| 1046 | * @page: the page |
| 1047 | * @lru: current lru |
| 1048 | * |
| 1049 | * This function accounts for @page being added to @lru, and returns |
| 1050 | * the lruvec for the given @zone and the memcg @page is charged to. |
| 1051 | * |
| 1052 | * The callsite is then responsible for physically linking the page to |
| 1053 | * the returned lruvec->lists[@lru]. |
| 1054 | */ |
| 1055 | struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page, |
| 1056 | enum lru_list lru) |
| 1057 | { |
| 1058 | struct mem_cgroup_per_zone *mz; |
| 1059 | struct mem_cgroup *memcg; |
| 1060 | struct page_cgroup *pc; |
| 1061 | |
| 1062 | if (mem_cgroup_disabled()) |
| 1063 | return &zone->lruvec; |
| 1064 | |
| 1065 | pc = lookup_page_cgroup(page); |
| 1066 | memcg = pc->mem_cgroup; |
| 1067 | |
| 1068 | /* |
| 1069 | * Surreptitiously switch any uncharged page to root: |
| 1070 | * an uncharged page off lru does nothing to secure |
| 1071 | * its former mem_cgroup from sudden removal. |
| 1072 | * |
| 1073 | * Our caller holds lru_lock, and PageCgroupUsed is updated |
| 1074 | * under page_cgroup lock: between them, they make all uses |
| 1075 | * of pc->mem_cgroup safe. |
| 1076 | */ |
| 1077 | if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup) |
| 1078 | pc->mem_cgroup = memcg = root_mem_cgroup; |
| 1079 | |
| 1080 | mz = page_cgroup_zoneinfo(memcg, page); |
| 1081 | /* compound_order() is stabilized through lru_lock */ |
| 1082 | mz->lru_size[lru] += 1 << compound_order(page); |
| 1083 | return &mz->lruvec; |
| 1084 | } |
| 1085 | |
| 1086 | /** |
| 1087 | * mem_cgroup_lru_del_list - account for removing an lru page |
| 1088 | * @page: the page |
| 1089 | * @lru: target lru |
| 1090 | * |
| 1091 | * This function accounts for @page being removed from @lru. |
| 1092 | * |
| 1093 | * The callsite is then responsible for physically unlinking |
| 1094 | * @page->lru. |
| 1095 | */ |
| 1096 | void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru) |
| 1097 | { |
| 1098 | struct mem_cgroup_per_zone *mz; |
| 1099 | struct mem_cgroup *memcg; |
| 1100 | struct page_cgroup *pc; |
| 1101 | |
| 1102 | if (mem_cgroup_disabled()) |
| 1103 | return; |
| 1104 | |
| 1105 | pc = lookup_page_cgroup(page); |
| 1106 | memcg = pc->mem_cgroup; |
| 1107 | VM_BUG_ON(!memcg); |
| 1108 | mz = page_cgroup_zoneinfo(memcg, page); |
| 1109 | /* huge page split is done under lru_lock. so, we have no races. */ |
| 1110 | VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page))); |
| 1111 | mz->lru_size[lru] -= 1 << compound_order(page); |
| 1112 | } |
| 1113 | |
| 1114 | void mem_cgroup_lru_del(struct page *page) |
| 1115 | { |
| 1116 | mem_cgroup_lru_del_list(page, page_lru(page)); |
| 1117 | } |
| 1118 | |
| 1119 | /** |
| 1120 | * mem_cgroup_lru_move_lists - account for moving a page between lrus |
| 1121 | * @zone: zone of the page |
| 1122 | * @page: the page |
| 1123 | * @from: current lru |
| 1124 | * @to: target lru |
| 1125 | * |
| 1126 | * This function accounts for @page being moved between the lrus @from |
| 1127 | * and @to, and returns the lruvec for the given @zone and the memcg |
| 1128 | * @page is charged to. |
| 1129 | * |
| 1130 | * The callsite is then responsible for physically relinking |
| 1131 | * @page->lru to the returned lruvec->lists[@to]. |
| 1132 | */ |
| 1133 | struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone, |
| 1134 | struct page *page, |
| 1135 | enum lru_list from, |
| 1136 | enum lru_list to) |
| 1137 | { |
| 1138 | /* XXX: Optimize this, especially for @from == @to */ |
| 1139 | mem_cgroup_lru_del_list(page, from); |
| 1140 | return mem_cgroup_lru_add_list(zone, page, to); |
| 1141 | } |
| 1142 | |
| 1143 | /* |
| 1144 | * Checks whether given mem is same or in the root_mem_cgroup's |
| 1145 | * hierarchy subtree |
| 1146 | */ |
| 1147 | static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, |
| 1148 | struct mem_cgroup *memcg) |
| 1149 | { |
| 1150 | if (root_memcg != memcg) { |
| 1151 | return (root_memcg->use_hierarchy && |
| 1152 | css_is_ancestor(&memcg->css, &root_memcg->css)); |
| 1153 | } |
| 1154 | |
| 1155 | return true; |
| 1156 | } |
| 1157 | |
| 1158 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) |
| 1159 | { |
| 1160 | int ret; |
| 1161 | struct mem_cgroup *curr = NULL; |
| 1162 | struct task_struct *p; |
| 1163 | |
| 1164 | p = find_lock_task_mm(task); |
| 1165 | if (p) { |
| 1166 | curr = try_get_mem_cgroup_from_mm(p->mm); |
| 1167 | task_unlock(p); |
| 1168 | } else { |
| 1169 | /* |
| 1170 | * All threads may have already detached their mm's, but the oom |
| 1171 | * killer still needs to detect if they have already been oom |
| 1172 | * killed to prevent needlessly killing additional tasks. |
| 1173 | */ |
| 1174 | task_lock(task); |
| 1175 | curr = mem_cgroup_from_task(task); |
| 1176 | if (curr) |
| 1177 | css_get(&curr->css); |
| 1178 | task_unlock(task); |
| 1179 | } |
| 1180 | if (!curr) |
| 1181 | return 0; |
| 1182 | /* |
| 1183 | * We should check use_hierarchy of "memcg" not "curr". Because checking |
| 1184 | * use_hierarchy of "curr" here make this function true if hierarchy is |
| 1185 | * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* |
| 1186 | * hierarchy(even if use_hierarchy is disabled in "memcg"). |
| 1187 | */ |
| 1188 | ret = mem_cgroup_same_or_subtree(memcg, curr); |
| 1189 | css_put(&curr->css); |
| 1190 | return ret; |
| 1191 | } |
| 1192 | |
| 1193 | int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) |
| 1194 | { |
| 1195 | unsigned long inactive_ratio; |
| 1196 | int nid = zone_to_nid(zone); |
| 1197 | int zid = zone_idx(zone); |
| 1198 | unsigned long inactive; |
| 1199 | unsigned long active; |
| 1200 | unsigned long gb; |
| 1201 | |
| 1202 | inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
| 1203 | BIT(LRU_INACTIVE_ANON)); |
| 1204 | active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
| 1205 | BIT(LRU_ACTIVE_ANON)); |
| 1206 | |
| 1207 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
| 1208 | if (gb) |
| 1209 | inactive_ratio = int_sqrt(10 * gb); |
| 1210 | else |
| 1211 | inactive_ratio = 1; |
| 1212 | |
| 1213 | return inactive * inactive_ratio < active; |
| 1214 | } |
| 1215 | |
| 1216 | int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) |
| 1217 | { |
| 1218 | unsigned long active; |
| 1219 | unsigned long inactive; |
| 1220 | int zid = zone_idx(zone); |
| 1221 | int nid = zone_to_nid(zone); |
| 1222 | |
| 1223 | inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
| 1224 | BIT(LRU_INACTIVE_FILE)); |
| 1225 | active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
| 1226 | BIT(LRU_ACTIVE_FILE)); |
| 1227 | |
| 1228 | return (active > inactive); |
| 1229 | } |
| 1230 | |
| 1231 | struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, |
| 1232 | struct zone *zone) |
| 1233 | { |
| 1234 | int nid = zone_to_nid(zone); |
| 1235 | int zid = zone_idx(zone); |
| 1236 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
| 1237 | |
| 1238 | return &mz->reclaim_stat; |
| 1239 | } |
| 1240 | |
| 1241 | struct zone_reclaim_stat * |
| 1242 | mem_cgroup_get_reclaim_stat_from_page(struct page *page) |
| 1243 | { |
| 1244 | struct page_cgroup *pc; |
| 1245 | struct mem_cgroup_per_zone *mz; |
| 1246 | |
| 1247 | if (mem_cgroup_disabled()) |
| 1248 | return NULL; |
| 1249 | |
| 1250 | pc = lookup_page_cgroup(page); |
| 1251 | if (!PageCgroupUsed(pc)) |
| 1252 | return NULL; |
| 1253 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ |
| 1254 | smp_rmb(); |
| 1255 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); |
| 1256 | return &mz->reclaim_stat; |
| 1257 | } |
| 1258 | |
| 1259 | #define mem_cgroup_from_res_counter(counter, member) \ |
| 1260 | container_of(counter, struct mem_cgroup, member) |
| 1261 | |
| 1262 | /** |
| 1263 | * mem_cgroup_margin - calculate chargeable space of a memory cgroup |
| 1264 | * @mem: the memory cgroup |
| 1265 | * |
| 1266 | * Returns the maximum amount of memory @mem can be charged with, in |
| 1267 | * pages. |
| 1268 | */ |
| 1269 | static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) |
| 1270 | { |
| 1271 | unsigned long long margin; |
| 1272 | |
| 1273 | margin = res_counter_margin(&memcg->res); |
| 1274 | if (do_swap_account) |
| 1275 | margin = min(margin, res_counter_margin(&memcg->memsw)); |
| 1276 | return margin >> PAGE_SHIFT; |
| 1277 | } |
| 1278 | |
| 1279 | int mem_cgroup_swappiness(struct mem_cgroup *memcg) |
| 1280 | { |
| 1281 | struct cgroup *cgrp = memcg->css.cgroup; |
| 1282 | |
| 1283 | /* root ? */ |
| 1284 | if (cgrp->parent == NULL) |
| 1285 | return vm_swappiness; |
| 1286 | |
| 1287 | return memcg->swappiness; |
| 1288 | } |
| 1289 | |
| 1290 | static void mem_cgroup_start_move(struct mem_cgroup *memcg) |
| 1291 | { |
| 1292 | int cpu; |
| 1293 | |
| 1294 | get_online_cpus(); |
| 1295 | spin_lock(&memcg->pcp_counter_lock); |
| 1296 | for_each_online_cpu(cpu) |
| 1297 | per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; |
| 1298 | memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; |
| 1299 | spin_unlock(&memcg->pcp_counter_lock); |
| 1300 | put_online_cpus(); |
| 1301 | |
| 1302 | synchronize_rcu(); |
| 1303 | } |
| 1304 | |
| 1305 | static void mem_cgroup_end_move(struct mem_cgroup *memcg) |
| 1306 | { |
| 1307 | int cpu; |
| 1308 | |
| 1309 | if (!memcg) |
| 1310 | return; |
| 1311 | get_online_cpus(); |
| 1312 | spin_lock(&memcg->pcp_counter_lock); |
| 1313 | for_each_online_cpu(cpu) |
| 1314 | per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; |
| 1315 | memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; |
| 1316 | spin_unlock(&memcg->pcp_counter_lock); |
| 1317 | put_online_cpus(); |
| 1318 | } |
| 1319 | /* |
| 1320 | * 2 routines for checking "mem" is under move_account() or not. |
| 1321 | * |
| 1322 | * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used |
| 1323 | * for avoiding race in accounting. If true, |
| 1324 | * pc->mem_cgroup may be overwritten. |
| 1325 | * |
| 1326 | * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or |
| 1327 | * under hierarchy of moving cgroups. This is for |
| 1328 | * waiting at hith-memory prressure caused by "move". |
| 1329 | */ |
| 1330 | |
| 1331 | static bool mem_cgroup_stealed(struct mem_cgroup *memcg) |
| 1332 | { |
| 1333 | VM_BUG_ON(!rcu_read_lock_held()); |
| 1334 | return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0; |
| 1335 | } |
| 1336 | |
| 1337 | static bool mem_cgroup_under_move(struct mem_cgroup *memcg) |
| 1338 | { |
| 1339 | struct mem_cgroup *from; |
| 1340 | struct mem_cgroup *to; |
| 1341 | bool ret = false; |
| 1342 | /* |
| 1343 | * Unlike task_move routines, we access mc.to, mc.from not under |
| 1344 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. |
| 1345 | */ |
| 1346 | spin_lock(&mc.lock); |
| 1347 | from = mc.from; |
| 1348 | to = mc.to; |
| 1349 | if (!from) |
| 1350 | goto unlock; |
| 1351 | |
| 1352 | ret = mem_cgroup_same_or_subtree(memcg, from) |
| 1353 | || mem_cgroup_same_or_subtree(memcg, to); |
| 1354 | unlock: |
| 1355 | spin_unlock(&mc.lock); |
| 1356 | return ret; |
| 1357 | } |
| 1358 | |
| 1359 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) |
| 1360 | { |
| 1361 | if (mc.moving_task && current != mc.moving_task) { |
| 1362 | if (mem_cgroup_under_move(memcg)) { |
| 1363 | DEFINE_WAIT(wait); |
| 1364 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); |
| 1365 | /* moving charge context might have finished. */ |
| 1366 | if (mc.moving_task) |
| 1367 | schedule(); |
| 1368 | finish_wait(&mc.waitq, &wait); |
| 1369 | return true; |
| 1370 | } |
| 1371 | } |
| 1372 | return false; |
| 1373 | } |
| 1374 | |
| 1375 | /** |
| 1376 | * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. |
| 1377 | * @memcg: The memory cgroup that went over limit |
| 1378 | * @p: Task that is going to be killed |
| 1379 | * |
| 1380 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is |
| 1381 | * enabled |
| 1382 | */ |
| 1383 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) |
| 1384 | { |
| 1385 | struct cgroup *task_cgrp; |
| 1386 | struct cgroup *mem_cgrp; |
| 1387 | /* |
| 1388 | * Need a buffer in BSS, can't rely on allocations. The code relies |
| 1389 | * on the assumption that OOM is serialized for memory controller. |
| 1390 | * If this assumption is broken, revisit this code. |
| 1391 | */ |
| 1392 | static char memcg_name[PATH_MAX]; |
| 1393 | int ret; |
| 1394 | |
| 1395 | if (!memcg || !p) |
| 1396 | return; |
| 1397 | |
| 1398 | rcu_read_lock(); |
| 1399 | |
| 1400 | mem_cgrp = memcg->css.cgroup; |
| 1401 | task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); |
| 1402 | |
| 1403 | ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); |
| 1404 | if (ret < 0) { |
| 1405 | /* |
| 1406 | * Unfortunately, we are unable to convert to a useful name |
| 1407 | * But we'll still print out the usage information |
| 1408 | */ |
| 1409 | rcu_read_unlock(); |
| 1410 | goto done; |
| 1411 | } |
| 1412 | rcu_read_unlock(); |
| 1413 | |
| 1414 | printk(KERN_INFO "Task in %s killed", memcg_name); |
| 1415 | |
| 1416 | rcu_read_lock(); |
| 1417 | ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); |
| 1418 | if (ret < 0) { |
| 1419 | rcu_read_unlock(); |
| 1420 | goto done; |
| 1421 | } |
| 1422 | rcu_read_unlock(); |
| 1423 | |
| 1424 | /* |
| 1425 | * Continues from above, so we don't need an KERN_ level |
| 1426 | */ |
| 1427 | printk(KERN_CONT " as a result of limit of %s\n", memcg_name); |
| 1428 | done: |
| 1429 | |
| 1430 | printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", |
| 1431 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, |
| 1432 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, |
| 1433 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); |
| 1434 | printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " |
| 1435 | "failcnt %llu\n", |
| 1436 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, |
| 1437 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, |
| 1438 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); |
| 1439 | } |
| 1440 | |
| 1441 | /* |
| 1442 | * This function returns the number of memcg under hierarchy tree. Returns |
| 1443 | * 1(self count) if no children. |
| 1444 | */ |
| 1445 | static int mem_cgroup_count_children(struct mem_cgroup *memcg) |
| 1446 | { |
| 1447 | int num = 0; |
| 1448 | struct mem_cgroup *iter; |
| 1449 | |
| 1450 | for_each_mem_cgroup_tree(iter, memcg) |
| 1451 | num++; |
| 1452 | return num; |
| 1453 | } |
| 1454 | |
| 1455 | /* |
| 1456 | * Return the memory (and swap, if configured) limit for a memcg. |
| 1457 | */ |
| 1458 | u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) |
| 1459 | { |
| 1460 | u64 limit; |
| 1461 | u64 memsw; |
| 1462 | |
| 1463 | limit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
| 1464 | limit += total_swap_pages << PAGE_SHIFT; |
| 1465 | |
| 1466 | memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
| 1467 | /* |
| 1468 | * If memsw is finite and limits the amount of swap space available |
| 1469 | * to this memcg, return that limit. |
| 1470 | */ |
| 1471 | return min(limit, memsw); |
| 1472 | } |
| 1473 | |
| 1474 | static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, |
| 1475 | gfp_t gfp_mask, |
| 1476 | unsigned long flags) |
| 1477 | { |
| 1478 | unsigned long total = 0; |
| 1479 | bool noswap = false; |
| 1480 | int loop; |
| 1481 | |
| 1482 | if (flags & MEM_CGROUP_RECLAIM_NOSWAP) |
| 1483 | noswap = true; |
| 1484 | if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) |
| 1485 | noswap = true; |
| 1486 | |
| 1487 | for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { |
| 1488 | if (loop) |
| 1489 | drain_all_stock_async(memcg); |
| 1490 | total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); |
| 1491 | /* |
| 1492 | * Allow limit shrinkers, which are triggered directly |
| 1493 | * by userspace, to catch signals and stop reclaim |
| 1494 | * after minimal progress, regardless of the margin. |
| 1495 | */ |
| 1496 | if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) |
| 1497 | break; |
| 1498 | if (mem_cgroup_margin(memcg)) |
| 1499 | break; |
| 1500 | /* |
| 1501 | * If nothing was reclaimed after two attempts, there |
| 1502 | * may be no reclaimable pages in this hierarchy. |
| 1503 | */ |
| 1504 | if (loop && !total) |
| 1505 | break; |
| 1506 | } |
| 1507 | return total; |
| 1508 | } |
| 1509 | |
| 1510 | /** |
| 1511 | * test_mem_cgroup_node_reclaimable |
| 1512 | * @mem: the target memcg |
| 1513 | * @nid: the node ID to be checked. |
| 1514 | * @noswap : specify true here if the user wants flle only information. |
| 1515 | * |
| 1516 | * This function returns whether the specified memcg contains any |
| 1517 | * reclaimable pages on a node. Returns true if there are any reclaimable |
| 1518 | * pages in the node. |
| 1519 | */ |
| 1520 | static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, |
| 1521 | int nid, bool noswap) |
| 1522 | { |
| 1523 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) |
| 1524 | return true; |
| 1525 | if (noswap || !total_swap_pages) |
| 1526 | return false; |
| 1527 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) |
| 1528 | return true; |
| 1529 | return false; |
| 1530 | |
| 1531 | } |
| 1532 | #if MAX_NUMNODES > 1 |
| 1533 | |
| 1534 | /* |
| 1535 | * Always updating the nodemask is not very good - even if we have an empty |
| 1536 | * list or the wrong list here, we can start from some node and traverse all |
| 1537 | * nodes based on the zonelist. So update the list loosely once per 10 secs. |
| 1538 | * |
| 1539 | */ |
| 1540 | static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) |
| 1541 | { |
| 1542 | int nid; |
| 1543 | /* |
| 1544 | * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET |
| 1545 | * pagein/pageout changes since the last update. |
| 1546 | */ |
| 1547 | if (!atomic_read(&memcg->numainfo_events)) |
| 1548 | return; |
| 1549 | if (atomic_inc_return(&memcg->numainfo_updating) > 1) |
| 1550 | return; |
| 1551 | |
| 1552 | /* make a nodemask where this memcg uses memory from */ |
| 1553 | memcg->scan_nodes = node_states[N_HIGH_MEMORY]; |
| 1554 | |
| 1555 | for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { |
| 1556 | |
| 1557 | if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) |
| 1558 | node_clear(nid, memcg->scan_nodes); |
| 1559 | } |
| 1560 | |
| 1561 | atomic_set(&memcg->numainfo_events, 0); |
| 1562 | atomic_set(&memcg->numainfo_updating, 0); |
| 1563 | } |
| 1564 | |
| 1565 | /* |
| 1566 | * Selecting a node where we start reclaim from. Because what we need is just |
| 1567 | * reducing usage counter, start from anywhere is O,K. Considering |
| 1568 | * memory reclaim from current node, there are pros. and cons. |
| 1569 | * |
| 1570 | * Freeing memory from current node means freeing memory from a node which |
| 1571 | * we'll use or we've used. So, it may make LRU bad. And if several threads |
| 1572 | * hit limits, it will see a contention on a node. But freeing from remote |
| 1573 | * node means more costs for memory reclaim because of memory latency. |
| 1574 | * |
| 1575 | * Now, we use round-robin. Better algorithm is welcomed. |
| 1576 | */ |
| 1577 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) |
| 1578 | { |
| 1579 | int node; |
| 1580 | |
| 1581 | mem_cgroup_may_update_nodemask(memcg); |
| 1582 | node = memcg->last_scanned_node; |
| 1583 | |
| 1584 | node = next_node(node, memcg->scan_nodes); |
| 1585 | if (node == MAX_NUMNODES) |
| 1586 | node = first_node(memcg->scan_nodes); |
| 1587 | /* |
| 1588 | * We call this when we hit limit, not when pages are added to LRU. |
| 1589 | * No LRU may hold pages because all pages are UNEVICTABLE or |
| 1590 | * memcg is too small and all pages are not on LRU. In that case, |
| 1591 | * we use curret node. |
| 1592 | */ |
| 1593 | if (unlikely(node == MAX_NUMNODES)) |
| 1594 | node = numa_node_id(); |
| 1595 | |
| 1596 | memcg->last_scanned_node = node; |
| 1597 | return node; |
| 1598 | } |
| 1599 | |
| 1600 | /* |
| 1601 | * Check all nodes whether it contains reclaimable pages or not. |
| 1602 | * For quick scan, we make use of scan_nodes. This will allow us to skip |
| 1603 | * unused nodes. But scan_nodes is lazily updated and may not cotain |
| 1604 | * enough new information. We need to do double check. |
| 1605 | */ |
| 1606 | bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) |
| 1607 | { |
| 1608 | int nid; |
| 1609 | |
| 1610 | /* |
| 1611 | * quick check...making use of scan_node. |
| 1612 | * We can skip unused nodes. |
| 1613 | */ |
| 1614 | if (!nodes_empty(memcg->scan_nodes)) { |
| 1615 | for (nid = first_node(memcg->scan_nodes); |
| 1616 | nid < MAX_NUMNODES; |
| 1617 | nid = next_node(nid, memcg->scan_nodes)) { |
| 1618 | |
| 1619 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) |
| 1620 | return true; |
| 1621 | } |
| 1622 | } |
| 1623 | /* |
| 1624 | * Check rest of nodes. |
| 1625 | */ |
| 1626 | for_each_node_state(nid, N_HIGH_MEMORY) { |
| 1627 | if (node_isset(nid, memcg->scan_nodes)) |
| 1628 | continue; |
| 1629 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) |
| 1630 | return true; |
| 1631 | } |
| 1632 | return false; |
| 1633 | } |
| 1634 | |
| 1635 | #else |
| 1636 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) |
| 1637 | { |
| 1638 | return 0; |
| 1639 | } |
| 1640 | |
| 1641 | bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) |
| 1642 | { |
| 1643 | return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); |
| 1644 | } |
| 1645 | #endif |
| 1646 | |
| 1647 | static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, |
| 1648 | struct zone *zone, |
| 1649 | gfp_t gfp_mask, |
| 1650 | unsigned long *total_scanned) |
| 1651 | { |
| 1652 | struct mem_cgroup *victim = NULL; |
| 1653 | int total = 0; |
| 1654 | int loop = 0; |
| 1655 | unsigned long excess; |
| 1656 | unsigned long nr_scanned; |
| 1657 | struct mem_cgroup_reclaim_cookie reclaim = { |
| 1658 | .zone = zone, |
| 1659 | .priority = 0, |
| 1660 | }; |
| 1661 | |
| 1662 | excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; |
| 1663 | |
| 1664 | while (1) { |
| 1665 | victim = mem_cgroup_iter(root_memcg, victim, &reclaim); |
| 1666 | if (!victim) { |
| 1667 | loop++; |
| 1668 | if (loop >= 2) { |
| 1669 | /* |
| 1670 | * If we have not been able to reclaim |
| 1671 | * anything, it might because there are |
| 1672 | * no reclaimable pages under this hierarchy |
| 1673 | */ |
| 1674 | if (!total) |
| 1675 | break; |
| 1676 | /* |
| 1677 | * We want to do more targeted reclaim. |
| 1678 | * excess >> 2 is not to excessive so as to |
| 1679 | * reclaim too much, nor too less that we keep |
| 1680 | * coming back to reclaim from this cgroup |
| 1681 | */ |
| 1682 | if (total >= (excess >> 2) || |
| 1683 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) |
| 1684 | break; |
| 1685 | } |
| 1686 | continue; |
| 1687 | } |
| 1688 | if (!mem_cgroup_reclaimable(victim, false)) |
| 1689 | continue; |
| 1690 | total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, |
| 1691 | zone, &nr_scanned); |
| 1692 | *total_scanned += nr_scanned; |
| 1693 | if (!res_counter_soft_limit_excess(&root_memcg->res)) |
| 1694 | break; |
| 1695 | } |
| 1696 | mem_cgroup_iter_break(root_memcg, victim); |
| 1697 | return total; |
| 1698 | } |
| 1699 | |
| 1700 | /* |
| 1701 | * Check OOM-Killer is already running under our hierarchy. |
| 1702 | * If someone is running, return false. |
| 1703 | * Has to be called with memcg_oom_lock |
| 1704 | */ |
| 1705 | static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) |
| 1706 | { |
| 1707 | struct mem_cgroup *iter, *failed = NULL; |
| 1708 | |
| 1709 | for_each_mem_cgroup_tree(iter, memcg) { |
| 1710 | if (iter->oom_lock) { |
| 1711 | /* |
| 1712 | * this subtree of our hierarchy is already locked |
| 1713 | * so we cannot give a lock. |
| 1714 | */ |
| 1715 | failed = iter; |
| 1716 | mem_cgroup_iter_break(memcg, iter); |
| 1717 | break; |
| 1718 | } else |
| 1719 | iter->oom_lock = true; |
| 1720 | } |
| 1721 | |
| 1722 | if (!failed) |
| 1723 | return true; |
| 1724 | |
| 1725 | /* |
| 1726 | * OK, we failed to lock the whole subtree so we have to clean up |
| 1727 | * what we set up to the failing subtree |
| 1728 | */ |
| 1729 | for_each_mem_cgroup_tree(iter, memcg) { |
| 1730 | if (iter == failed) { |
| 1731 | mem_cgroup_iter_break(memcg, iter); |
| 1732 | break; |
| 1733 | } |
| 1734 | iter->oom_lock = false; |
| 1735 | } |
| 1736 | return false; |
| 1737 | } |
| 1738 | |
| 1739 | /* |
| 1740 | * Has to be called with memcg_oom_lock |
| 1741 | */ |
| 1742 | static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) |
| 1743 | { |
| 1744 | struct mem_cgroup *iter; |
| 1745 | |
| 1746 | for_each_mem_cgroup_tree(iter, memcg) |
| 1747 | iter->oom_lock = false; |
| 1748 | return 0; |
| 1749 | } |
| 1750 | |
| 1751 | static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) |
| 1752 | { |
| 1753 | struct mem_cgroup *iter; |
| 1754 | |
| 1755 | for_each_mem_cgroup_tree(iter, memcg) |
| 1756 | atomic_inc(&iter->under_oom); |
| 1757 | } |
| 1758 | |
| 1759 | static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) |
| 1760 | { |
| 1761 | struct mem_cgroup *iter; |
| 1762 | |
| 1763 | /* |
| 1764 | * When a new child is created while the hierarchy is under oom, |
| 1765 | * mem_cgroup_oom_lock() may not be called. We have to use |
| 1766 | * atomic_add_unless() here. |
| 1767 | */ |
| 1768 | for_each_mem_cgroup_tree(iter, memcg) |
| 1769 | atomic_add_unless(&iter->under_oom, -1, 0); |
| 1770 | } |
| 1771 | |
| 1772 | static DEFINE_SPINLOCK(memcg_oom_lock); |
| 1773 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); |
| 1774 | |
| 1775 | struct oom_wait_info { |
| 1776 | struct mem_cgroup *memcg; |
| 1777 | wait_queue_t wait; |
| 1778 | }; |
| 1779 | |
| 1780 | static int memcg_oom_wake_function(wait_queue_t *wait, |
| 1781 | unsigned mode, int sync, void *arg) |
| 1782 | { |
| 1783 | struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg; |
| 1784 | struct mem_cgroup *oom_wait_memcg; |
| 1785 | struct oom_wait_info *oom_wait_info; |
| 1786 | |
| 1787 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); |
| 1788 | oom_wait_memcg = oom_wait_info->memcg; |
| 1789 | |
| 1790 | /* |
| 1791 | * Both of oom_wait_info->memcg and wake_memcg are stable under us. |
| 1792 | * Then we can use css_is_ancestor without taking care of RCU. |
| 1793 | */ |
| 1794 | if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) |
| 1795 | && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) |
| 1796 | return 0; |
| 1797 | return autoremove_wake_function(wait, mode, sync, arg); |
| 1798 | } |
| 1799 | |
| 1800 | static void memcg_wakeup_oom(struct mem_cgroup *memcg) |
| 1801 | { |
| 1802 | /* for filtering, pass "memcg" as argument. */ |
| 1803 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); |
| 1804 | } |
| 1805 | |
| 1806 | static void memcg_oom_recover(struct mem_cgroup *memcg) |
| 1807 | { |
| 1808 | if (memcg && atomic_read(&memcg->under_oom)) |
| 1809 | memcg_wakeup_oom(memcg); |
| 1810 | } |
| 1811 | |
| 1812 | /* |
| 1813 | * try to call OOM killer. returns false if we should exit memory-reclaim loop. |
| 1814 | */ |
| 1815 | bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order) |
| 1816 | { |
| 1817 | struct oom_wait_info owait; |
| 1818 | bool locked, need_to_kill; |
| 1819 | |
| 1820 | owait.memcg = memcg; |
| 1821 | owait.wait.flags = 0; |
| 1822 | owait.wait.func = memcg_oom_wake_function; |
| 1823 | owait.wait.private = current; |
| 1824 | INIT_LIST_HEAD(&owait.wait.task_list); |
| 1825 | need_to_kill = true; |
| 1826 | mem_cgroup_mark_under_oom(memcg); |
| 1827 | |
| 1828 | /* At first, try to OOM lock hierarchy under memcg.*/ |
| 1829 | spin_lock(&memcg_oom_lock); |
| 1830 | locked = mem_cgroup_oom_lock(memcg); |
| 1831 | /* |
| 1832 | * Even if signal_pending(), we can't quit charge() loop without |
| 1833 | * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL |
| 1834 | * under OOM is always welcomed, use TASK_KILLABLE here. |
| 1835 | */ |
| 1836 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); |
| 1837 | if (!locked || memcg->oom_kill_disable) |
| 1838 | need_to_kill = false; |
| 1839 | if (locked) |
| 1840 | mem_cgroup_oom_notify(memcg); |
| 1841 | spin_unlock(&memcg_oom_lock); |
| 1842 | |
| 1843 | if (need_to_kill) { |
| 1844 | finish_wait(&memcg_oom_waitq, &owait.wait); |
| 1845 | mem_cgroup_out_of_memory(memcg, mask, order); |
| 1846 | } else { |
| 1847 | schedule(); |
| 1848 | finish_wait(&memcg_oom_waitq, &owait.wait); |
| 1849 | } |
| 1850 | spin_lock(&memcg_oom_lock); |
| 1851 | if (locked) |
| 1852 | mem_cgroup_oom_unlock(memcg); |
| 1853 | memcg_wakeup_oom(memcg); |
| 1854 | spin_unlock(&memcg_oom_lock); |
| 1855 | |
| 1856 | mem_cgroup_unmark_under_oom(memcg); |
| 1857 | |
| 1858 | if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) |
| 1859 | return false; |
| 1860 | /* Give chance to dying process */ |
| 1861 | schedule_timeout_uninterruptible(1); |
| 1862 | return true; |
| 1863 | } |
| 1864 | |
| 1865 | /* |
| 1866 | * Currently used to update mapped file statistics, but the routine can be |
| 1867 | * generalized to update other statistics as well. |
| 1868 | * |
| 1869 | * Notes: Race condition |
| 1870 | * |
| 1871 | * We usually use page_cgroup_lock() for accessing page_cgroup member but |
| 1872 | * it tends to be costly. But considering some conditions, we doesn't need |
| 1873 | * to do so _always_. |
| 1874 | * |
| 1875 | * Considering "charge", lock_page_cgroup() is not required because all |
| 1876 | * file-stat operations happen after a page is attached to radix-tree. There |
| 1877 | * are no race with "charge". |
| 1878 | * |
| 1879 | * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup |
| 1880 | * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even |
| 1881 | * if there are race with "uncharge". Statistics itself is properly handled |
| 1882 | * by flags. |
| 1883 | * |
| 1884 | * Considering "move", this is an only case we see a race. To make the race |
| 1885 | * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are |
| 1886 | * possibility of race condition. If there is, we take a lock. |
| 1887 | */ |
| 1888 | |
| 1889 | void mem_cgroup_update_page_stat(struct page *page, |
| 1890 | enum mem_cgroup_page_stat_item idx, int val) |
| 1891 | { |
| 1892 | struct mem_cgroup *memcg; |
| 1893 | struct page_cgroup *pc = lookup_page_cgroup(page); |
| 1894 | bool need_unlock = false; |
| 1895 | unsigned long uninitialized_var(flags); |
| 1896 | |
| 1897 | if (mem_cgroup_disabled()) |
| 1898 | return; |
| 1899 | |
| 1900 | rcu_read_lock(); |
| 1901 | memcg = pc->mem_cgroup; |
| 1902 | if (unlikely(!memcg || !PageCgroupUsed(pc))) |
| 1903 | goto out; |
| 1904 | /* pc->mem_cgroup is unstable ? */ |
| 1905 | if (unlikely(mem_cgroup_stealed(memcg))) { |
| 1906 | /* take a lock against to access pc->mem_cgroup */ |
| 1907 | move_lock_page_cgroup(pc, &flags); |
| 1908 | need_unlock = true; |
| 1909 | memcg = pc->mem_cgroup; |
| 1910 | if (!memcg || !PageCgroupUsed(pc)) |
| 1911 | goto out; |
| 1912 | } |
| 1913 | |
| 1914 | switch (idx) { |
| 1915 | case MEMCG_NR_FILE_MAPPED: |
| 1916 | if (val > 0) |
| 1917 | SetPageCgroupFileMapped(pc); |
| 1918 | else if (!page_mapped(page)) |
| 1919 | ClearPageCgroupFileMapped(pc); |
| 1920 | idx = MEM_CGROUP_STAT_FILE_MAPPED; |
| 1921 | break; |
| 1922 | default: |
| 1923 | BUG(); |
| 1924 | } |
| 1925 | |
| 1926 | this_cpu_add(memcg->stat->count[idx], val); |
| 1927 | |
| 1928 | out: |
| 1929 | if (unlikely(need_unlock)) |
| 1930 | move_unlock_page_cgroup(pc, &flags); |
| 1931 | rcu_read_unlock(); |
| 1932 | } |
| 1933 | |
| 1934 | /* |
| 1935 | * size of first charge trial. "32" comes from vmscan.c's magic value. |
| 1936 | * TODO: maybe necessary to use big numbers in big irons. |
| 1937 | */ |
| 1938 | #define CHARGE_BATCH 32U |
| 1939 | struct memcg_stock_pcp { |
| 1940 | struct mem_cgroup *cached; /* this never be root cgroup */ |
| 1941 | unsigned int nr_pages; |
| 1942 | struct work_struct work; |
| 1943 | unsigned long flags; |
| 1944 | #define FLUSHING_CACHED_CHARGE (0) |
| 1945 | }; |
| 1946 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); |
| 1947 | static DEFINE_MUTEX(percpu_charge_mutex); |
| 1948 | |
| 1949 | /* |
| 1950 | * Try to consume stocked charge on this cpu. If success, one page is consumed |
| 1951 | * from local stock and true is returned. If the stock is 0 or charges from a |
| 1952 | * cgroup which is not current target, returns false. This stock will be |
| 1953 | * refilled. |
| 1954 | */ |
| 1955 | static bool consume_stock(struct mem_cgroup *memcg) |
| 1956 | { |
| 1957 | struct memcg_stock_pcp *stock; |
| 1958 | bool ret = true; |
| 1959 | |
| 1960 | stock = &get_cpu_var(memcg_stock); |
| 1961 | if (memcg == stock->cached && stock->nr_pages) |
| 1962 | stock->nr_pages--; |
| 1963 | else /* need to call res_counter_charge */ |
| 1964 | ret = false; |
| 1965 | put_cpu_var(memcg_stock); |
| 1966 | return ret; |
| 1967 | } |
| 1968 | |
| 1969 | /* |
| 1970 | * Returns stocks cached in percpu to res_counter and reset cached information. |
| 1971 | */ |
| 1972 | static void drain_stock(struct memcg_stock_pcp *stock) |
| 1973 | { |
| 1974 | struct mem_cgroup *old = stock->cached; |
| 1975 | |
| 1976 | if (stock->nr_pages) { |
| 1977 | unsigned long bytes = stock->nr_pages * PAGE_SIZE; |
| 1978 | |
| 1979 | res_counter_uncharge(&old->res, bytes); |
| 1980 | if (do_swap_account) |
| 1981 | res_counter_uncharge(&old->memsw, bytes); |
| 1982 | stock->nr_pages = 0; |
| 1983 | } |
| 1984 | stock->cached = NULL; |
| 1985 | } |
| 1986 | |
| 1987 | /* |
| 1988 | * This must be called under preempt disabled or must be called by |
| 1989 | * a thread which is pinned to local cpu. |
| 1990 | */ |
| 1991 | static void drain_local_stock(struct work_struct *dummy) |
| 1992 | { |
| 1993 | struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); |
| 1994 | drain_stock(stock); |
| 1995 | clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); |
| 1996 | } |
| 1997 | |
| 1998 | /* |
| 1999 | * Cache charges(val) which is from res_counter, to local per_cpu area. |
| 2000 | * This will be consumed by consume_stock() function, later. |
| 2001 | */ |
| 2002 | static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) |
| 2003 | { |
| 2004 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); |
| 2005 | |
| 2006 | if (stock->cached != memcg) { /* reset if necessary */ |
| 2007 | drain_stock(stock); |
| 2008 | stock->cached = memcg; |
| 2009 | } |
| 2010 | stock->nr_pages += nr_pages; |
| 2011 | put_cpu_var(memcg_stock); |
| 2012 | } |
| 2013 | |
| 2014 | /* |
| 2015 | * Drains all per-CPU charge caches for given root_memcg resp. subtree |
| 2016 | * of the hierarchy under it. sync flag says whether we should block |
| 2017 | * until the work is done. |
| 2018 | */ |
| 2019 | static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) |
| 2020 | { |
| 2021 | int cpu, curcpu; |
| 2022 | |
| 2023 | /* Notify other cpus that system-wide "drain" is running */ |
| 2024 | get_online_cpus(); |
| 2025 | curcpu = get_cpu(); |
| 2026 | for_each_online_cpu(cpu) { |
| 2027 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); |
| 2028 | struct mem_cgroup *memcg; |
| 2029 | |
| 2030 | memcg = stock->cached; |
| 2031 | if (!memcg || !stock->nr_pages) |
| 2032 | continue; |
| 2033 | if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) |
| 2034 | continue; |
| 2035 | if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { |
| 2036 | if (cpu == curcpu) |
| 2037 | drain_local_stock(&stock->work); |
| 2038 | else |
| 2039 | schedule_work_on(cpu, &stock->work); |
| 2040 | } |
| 2041 | } |
| 2042 | put_cpu(); |
| 2043 | |
| 2044 | if (!sync) |
| 2045 | goto out; |
| 2046 | |
| 2047 | for_each_online_cpu(cpu) { |
| 2048 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); |
| 2049 | if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) |
| 2050 | flush_work(&stock->work); |
| 2051 | } |
| 2052 | out: |
| 2053 | put_online_cpus(); |
| 2054 | } |
| 2055 | |
| 2056 | /* |
| 2057 | * Tries to drain stocked charges in other cpus. This function is asynchronous |
| 2058 | * and just put a work per cpu for draining localy on each cpu. Caller can |
| 2059 | * expects some charges will be back to res_counter later but cannot wait for |
| 2060 | * it. |
| 2061 | */ |
| 2062 | static void drain_all_stock_async(struct mem_cgroup *root_memcg) |
| 2063 | { |
| 2064 | /* |
| 2065 | * If someone calls draining, avoid adding more kworker runs. |
| 2066 | */ |
| 2067 | if (!mutex_trylock(&percpu_charge_mutex)) |
| 2068 | return; |
| 2069 | drain_all_stock(root_memcg, false); |
| 2070 | mutex_unlock(&percpu_charge_mutex); |
| 2071 | } |
| 2072 | |
| 2073 | /* This is a synchronous drain interface. */ |
| 2074 | static void drain_all_stock_sync(struct mem_cgroup *root_memcg) |
| 2075 | { |
| 2076 | /* called when force_empty is called */ |
| 2077 | mutex_lock(&percpu_charge_mutex); |
| 2078 | drain_all_stock(root_memcg, true); |
| 2079 | mutex_unlock(&percpu_charge_mutex); |
| 2080 | } |
| 2081 | |
| 2082 | /* |
| 2083 | * This function drains percpu counter value from DEAD cpu and |
| 2084 | * move it to local cpu. Note that this function can be preempted. |
| 2085 | */ |
| 2086 | static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) |
| 2087 | { |
| 2088 | int i; |
| 2089 | |
| 2090 | spin_lock(&memcg->pcp_counter_lock); |
| 2091 | for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { |
| 2092 | long x = per_cpu(memcg->stat->count[i], cpu); |
| 2093 | |
| 2094 | per_cpu(memcg->stat->count[i], cpu) = 0; |
| 2095 | memcg->nocpu_base.count[i] += x; |
| 2096 | } |
| 2097 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { |
| 2098 | unsigned long x = per_cpu(memcg->stat->events[i], cpu); |
| 2099 | |
| 2100 | per_cpu(memcg->stat->events[i], cpu) = 0; |
| 2101 | memcg->nocpu_base.events[i] += x; |
| 2102 | } |
| 2103 | /* need to clear ON_MOVE value, works as a kind of lock. */ |
| 2104 | per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; |
| 2105 | spin_unlock(&memcg->pcp_counter_lock); |
| 2106 | } |
| 2107 | |
| 2108 | static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu) |
| 2109 | { |
| 2110 | int idx = MEM_CGROUP_ON_MOVE; |
| 2111 | |
| 2112 | spin_lock(&memcg->pcp_counter_lock); |
| 2113 | per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx]; |
| 2114 | spin_unlock(&memcg->pcp_counter_lock); |
| 2115 | } |
| 2116 | |
| 2117 | static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, |
| 2118 | unsigned long action, |
| 2119 | void *hcpu) |
| 2120 | { |
| 2121 | int cpu = (unsigned long)hcpu; |
| 2122 | struct memcg_stock_pcp *stock; |
| 2123 | struct mem_cgroup *iter; |
| 2124 | |
| 2125 | if ((action == CPU_ONLINE)) { |
| 2126 | for_each_mem_cgroup(iter) |
| 2127 | synchronize_mem_cgroup_on_move(iter, cpu); |
| 2128 | return NOTIFY_OK; |
| 2129 | } |
| 2130 | |
| 2131 | if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) |
| 2132 | return NOTIFY_OK; |
| 2133 | |
| 2134 | for_each_mem_cgroup(iter) |
| 2135 | mem_cgroup_drain_pcp_counter(iter, cpu); |
| 2136 | |
| 2137 | stock = &per_cpu(memcg_stock, cpu); |
| 2138 | drain_stock(stock); |
| 2139 | return NOTIFY_OK; |
| 2140 | } |
| 2141 | |
| 2142 | |
| 2143 | /* See __mem_cgroup_try_charge() for details */ |
| 2144 | enum { |
| 2145 | CHARGE_OK, /* success */ |
| 2146 | CHARGE_RETRY, /* need to retry but retry is not bad */ |
| 2147 | CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ |
| 2148 | CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ |
| 2149 | CHARGE_OOM_DIE, /* the current is killed because of OOM */ |
| 2150 | }; |
| 2151 | |
| 2152 | static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, |
| 2153 | unsigned int nr_pages, bool oom_check) |
| 2154 | { |
| 2155 | unsigned long csize = nr_pages * PAGE_SIZE; |
| 2156 | struct mem_cgroup *mem_over_limit; |
| 2157 | struct res_counter *fail_res; |
| 2158 | unsigned long flags = 0; |
| 2159 | int ret; |
| 2160 | |
| 2161 | ret = res_counter_charge(&memcg->res, csize, &fail_res); |
| 2162 | |
| 2163 | if (likely(!ret)) { |
| 2164 | if (!do_swap_account) |
| 2165 | return CHARGE_OK; |
| 2166 | ret = res_counter_charge(&memcg->memsw, csize, &fail_res); |
| 2167 | if (likely(!ret)) |
| 2168 | return CHARGE_OK; |
| 2169 | |
| 2170 | res_counter_uncharge(&memcg->res, csize); |
| 2171 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); |
| 2172 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; |
| 2173 | } else |
| 2174 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); |
| 2175 | /* |
| 2176 | * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch |
| 2177 | * of regular pages (CHARGE_BATCH), or a single regular page (1). |
| 2178 | * |
| 2179 | * Never reclaim on behalf of optional batching, retry with a |
| 2180 | * single page instead. |
| 2181 | */ |
| 2182 | if (nr_pages == CHARGE_BATCH) |
| 2183 | return CHARGE_RETRY; |
| 2184 | |
| 2185 | if (!(gfp_mask & __GFP_WAIT)) |
| 2186 | return CHARGE_WOULDBLOCK; |
| 2187 | |
| 2188 | ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); |
| 2189 | if (mem_cgroup_margin(mem_over_limit) >= nr_pages) |
| 2190 | return CHARGE_RETRY; |
| 2191 | /* |
| 2192 | * Even though the limit is exceeded at this point, reclaim |
| 2193 | * may have been able to free some pages. Retry the charge |
| 2194 | * before killing the task. |
| 2195 | * |
| 2196 | * Only for regular pages, though: huge pages are rather |
| 2197 | * unlikely to succeed so close to the limit, and we fall back |
| 2198 | * to regular pages anyway in case of failure. |
| 2199 | */ |
| 2200 | if (nr_pages == 1 && ret) |
| 2201 | return CHARGE_RETRY; |
| 2202 | |
| 2203 | /* |
| 2204 | * At task move, charge accounts can be doubly counted. So, it's |
| 2205 | * better to wait until the end of task_move if something is going on. |
| 2206 | */ |
| 2207 | if (mem_cgroup_wait_acct_move(mem_over_limit)) |
| 2208 | return CHARGE_RETRY; |
| 2209 | |
| 2210 | /* If we don't need to call oom-killer at el, return immediately */ |
| 2211 | if (!oom_check) |
| 2212 | return CHARGE_NOMEM; |
| 2213 | /* check OOM */ |
| 2214 | if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize))) |
| 2215 | return CHARGE_OOM_DIE; |
| 2216 | |
| 2217 | return CHARGE_RETRY; |
| 2218 | } |
| 2219 | |
| 2220 | /* |
| 2221 | * __mem_cgroup_try_charge() does |
| 2222 | * 1. detect memcg to be charged against from passed *mm and *ptr, |
| 2223 | * 2. update res_counter |
| 2224 | * 3. call memory reclaim if necessary. |
| 2225 | * |
| 2226 | * In some special case, if the task is fatal, fatal_signal_pending() or |
| 2227 | * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup |
| 2228 | * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon |
| 2229 | * as possible without any hazards. 2: all pages should have a valid |
| 2230 | * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg |
| 2231 | * pointer, that is treated as a charge to root_mem_cgroup. |
| 2232 | * |
| 2233 | * So __mem_cgroup_try_charge() will return |
| 2234 | * 0 ... on success, filling *ptr with a valid memcg pointer. |
| 2235 | * -ENOMEM ... charge failure because of resource limits. |
| 2236 | * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup. |
| 2237 | * |
| 2238 | * Unlike the exported interface, an "oom" parameter is added. if oom==true, |
| 2239 | * the oom-killer can be invoked. |
| 2240 | */ |
| 2241 | static int __mem_cgroup_try_charge(struct mm_struct *mm, |
| 2242 | gfp_t gfp_mask, |
| 2243 | unsigned int nr_pages, |
| 2244 | struct mem_cgroup **ptr, |
| 2245 | bool oom) |
| 2246 | { |
| 2247 | unsigned int batch = max(CHARGE_BATCH, nr_pages); |
| 2248 | int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; |
| 2249 | struct mem_cgroup *memcg = NULL; |
| 2250 | int ret; |
| 2251 | |
| 2252 | /* |
| 2253 | * Unlike gloval-vm's OOM-kill, we're not in memory shortage |
| 2254 | * in system level. So, allow to go ahead dying process in addition to |
| 2255 | * MEMDIE process. |
| 2256 | */ |
| 2257 | if (unlikely(test_thread_flag(TIF_MEMDIE) |
| 2258 | || fatal_signal_pending(current))) |
| 2259 | goto bypass; |
| 2260 | |
| 2261 | /* |
| 2262 | * We always charge the cgroup the mm_struct belongs to. |
| 2263 | * The mm_struct's mem_cgroup changes on task migration if the |
| 2264 | * thread group leader migrates. It's possible that mm is not |
| 2265 | * set, if so charge the init_mm (happens for pagecache usage). |
| 2266 | */ |
| 2267 | if (!*ptr && !mm) |
| 2268 | *ptr = root_mem_cgroup; |
| 2269 | again: |
| 2270 | if (*ptr) { /* css should be a valid one */ |
| 2271 | memcg = *ptr; |
| 2272 | VM_BUG_ON(css_is_removed(&memcg->css)); |
| 2273 | if (mem_cgroup_is_root(memcg)) |
| 2274 | goto done; |
| 2275 | if (nr_pages == 1 && consume_stock(memcg)) |
| 2276 | goto done; |
| 2277 | css_get(&memcg->css); |
| 2278 | } else { |
| 2279 | struct task_struct *p; |
| 2280 | |
| 2281 | rcu_read_lock(); |
| 2282 | p = rcu_dereference(mm->owner); |
| 2283 | /* |
| 2284 | * Because we don't have task_lock(), "p" can exit. |
| 2285 | * In that case, "memcg" can point to root or p can be NULL with |
| 2286 | * race with swapoff. Then, we have small risk of mis-accouning. |
| 2287 | * But such kind of mis-account by race always happens because |
| 2288 | * we don't have cgroup_mutex(). It's overkill and we allo that |
| 2289 | * small race, here. |
| 2290 | * (*) swapoff at el will charge against mm-struct not against |
| 2291 | * task-struct. So, mm->owner can be NULL. |
| 2292 | */ |
| 2293 | memcg = mem_cgroup_from_task(p); |
| 2294 | if (!memcg) |
| 2295 | memcg = root_mem_cgroup; |
| 2296 | if (mem_cgroup_is_root(memcg)) { |
| 2297 | rcu_read_unlock(); |
| 2298 | goto done; |
| 2299 | } |
| 2300 | if (nr_pages == 1 && consume_stock(memcg)) { |
| 2301 | /* |
| 2302 | * It seems dagerous to access memcg without css_get(). |
| 2303 | * But considering how consume_stok works, it's not |
| 2304 | * necessary. If consume_stock success, some charges |
| 2305 | * from this memcg are cached on this cpu. So, we |
| 2306 | * don't need to call css_get()/css_tryget() before |
| 2307 | * calling consume_stock(). |
| 2308 | */ |
| 2309 | rcu_read_unlock(); |
| 2310 | goto done; |
| 2311 | } |
| 2312 | /* after here, we may be blocked. we need to get refcnt */ |
| 2313 | if (!css_tryget(&memcg->css)) { |
| 2314 | rcu_read_unlock(); |
| 2315 | goto again; |
| 2316 | } |
| 2317 | rcu_read_unlock(); |
| 2318 | } |
| 2319 | |
| 2320 | do { |
| 2321 | bool oom_check; |
| 2322 | |
| 2323 | /* If killed, bypass charge */ |
| 2324 | if (fatal_signal_pending(current)) { |
| 2325 | css_put(&memcg->css); |
| 2326 | goto bypass; |
| 2327 | } |
| 2328 | |
| 2329 | oom_check = false; |
| 2330 | if (oom && !nr_oom_retries) { |
| 2331 | oom_check = true; |
| 2332 | nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; |
| 2333 | } |
| 2334 | |
| 2335 | ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); |
| 2336 | switch (ret) { |
| 2337 | case CHARGE_OK: |
| 2338 | break; |
| 2339 | case CHARGE_RETRY: /* not in OOM situation but retry */ |
| 2340 | batch = nr_pages; |
| 2341 | css_put(&memcg->css); |
| 2342 | memcg = NULL; |
| 2343 | goto again; |
| 2344 | case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ |
| 2345 | css_put(&memcg->css); |
| 2346 | goto nomem; |
| 2347 | case CHARGE_NOMEM: /* OOM routine works */ |
| 2348 | if (!oom) { |
| 2349 | css_put(&memcg->css); |
| 2350 | goto nomem; |
| 2351 | } |
| 2352 | /* If oom, we never return -ENOMEM */ |
| 2353 | nr_oom_retries--; |
| 2354 | break; |
| 2355 | case CHARGE_OOM_DIE: /* Killed by OOM Killer */ |
| 2356 | css_put(&memcg->css); |
| 2357 | goto bypass; |
| 2358 | } |
| 2359 | } while (ret != CHARGE_OK); |
| 2360 | |
| 2361 | if (batch > nr_pages) |
| 2362 | refill_stock(memcg, batch - nr_pages); |
| 2363 | css_put(&memcg->css); |
| 2364 | done: |
| 2365 | *ptr = memcg; |
| 2366 | return 0; |
| 2367 | nomem: |
| 2368 | *ptr = NULL; |
| 2369 | return -ENOMEM; |
| 2370 | bypass: |
| 2371 | *ptr = root_mem_cgroup; |
| 2372 | return -EINTR; |
| 2373 | } |
| 2374 | |
| 2375 | /* |
| 2376 | * Somemtimes we have to undo a charge we got by try_charge(). |
| 2377 | * This function is for that and do uncharge, put css's refcnt. |
| 2378 | * gotten by try_charge(). |
| 2379 | */ |
| 2380 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, |
| 2381 | unsigned int nr_pages) |
| 2382 | { |
| 2383 | if (!mem_cgroup_is_root(memcg)) { |
| 2384 | unsigned long bytes = nr_pages * PAGE_SIZE; |
| 2385 | |
| 2386 | res_counter_uncharge(&memcg->res, bytes); |
| 2387 | if (do_swap_account) |
| 2388 | res_counter_uncharge(&memcg->memsw, bytes); |
| 2389 | } |
| 2390 | } |
| 2391 | |
| 2392 | /* |
| 2393 | * A helper function to get mem_cgroup from ID. must be called under |
| 2394 | * rcu_read_lock(). The caller must check css_is_removed() or some if |
| 2395 | * it's concern. (dropping refcnt from swap can be called against removed |
| 2396 | * memcg.) |
| 2397 | */ |
| 2398 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) |
| 2399 | { |
| 2400 | struct cgroup_subsys_state *css; |
| 2401 | |
| 2402 | /* ID 0 is unused ID */ |
| 2403 | if (!id) |
| 2404 | return NULL; |
| 2405 | css = css_lookup(&mem_cgroup_subsys, id); |
| 2406 | if (!css) |
| 2407 | return NULL; |
| 2408 | return container_of(css, struct mem_cgroup, css); |
| 2409 | } |
| 2410 | |
| 2411 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) |
| 2412 | { |
| 2413 | struct mem_cgroup *memcg = NULL; |
| 2414 | struct page_cgroup *pc; |
| 2415 | unsigned short id; |
| 2416 | swp_entry_t ent; |
| 2417 | |
| 2418 | VM_BUG_ON(!PageLocked(page)); |
| 2419 | |
| 2420 | pc = lookup_page_cgroup(page); |
| 2421 | lock_page_cgroup(pc); |
| 2422 | if (PageCgroupUsed(pc)) { |
| 2423 | memcg = pc->mem_cgroup; |
| 2424 | if (memcg && !css_tryget(&memcg->css)) |
| 2425 | memcg = NULL; |
| 2426 | } else if (PageSwapCache(page)) { |
| 2427 | ent.val = page_private(page); |
| 2428 | id = lookup_swap_cgroup_id(ent); |
| 2429 | rcu_read_lock(); |
| 2430 | memcg = mem_cgroup_lookup(id); |
| 2431 | if (memcg && !css_tryget(&memcg->css)) |
| 2432 | memcg = NULL; |
| 2433 | rcu_read_unlock(); |
| 2434 | } |
| 2435 | unlock_page_cgroup(pc); |
| 2436 | return memcg; |
| 2437 | } |
| 2438 | |
| 2439 | static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, |
| 2440 | struct page *page, |
| 2441 | unsigned int nr_pages, |
| 2442 | struct page_cgroup *pc, |
| 2443 | enum charge_type ctype, |
| 2444 | bool lrucare) |
| 2445 | { |
| 2446 | struct zone *uninitialized_var(zone); |
| 2447 | bool was_on_lru = false; |
| 2448 | bool anon; |
| 2449 | |
| 2450 | lock_page_cgroup(pc); |
| 2451 | if (unlikely(PageCgroupUsed(pc))) { |
| 2452 | unlock_page_cgroup(pc); |
| 2453 | __mem_cgroup_cancel_charge(memcg, nr_pages); |
| 2454 | return; |
| 2455 | } |
| 2456 | /* |
| 2457 | * we don't need page_cgroup_lock about tail pages, becase they are not |
| 2458 | * accessed by any other context at this point. |
| 2459 | */ |
| 2460 | |
| 2461 | /* |
| 2462 | * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page |
| 2463 | * may already be on some other mem_cgroup's LRU. Take care of it. |
| 2464 | */ |
| 2465 | if (lrucare) { |
| 2466 | zone = page_zone(page); |
| 2467 | spin_lock_irq(&zone->lru_lock); |
| 2468 | if (PageLRU(page)) { |
| 2469 | ClearPageLRU(page); |
| 2470 | del_page_from_lru_list(zone, page, page_lru(page)); |
| 2471 | was_on_lru = true; |
| 2472 | } |
| 2473 | } |
| 2474 | |
| 2475 | pc->mem_cgroup = memcg; |
| 2476 | /* |
| 2477 | * We access a page_cgroup asynchronously without lock_page_cgroup(). |
| 2478 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup |
| 2479 | * is accessed after testing USED bit. To make pc->mem_cgroup visible |
| 2480 | * before USED bit, we need memory barrier here. |
| 2481 | * See mem_cgroup_add_lru_list(), etc. |
| 2482 | */ |
| 2483 | smp_wmb(); |
| 2484 | SetPageCgroupUsed(pc); |
| 2485 | |
| 2486 | if (lrucare) { |
| 2487 | if (was_on_lru) { |
| 2488 | VM_BUG_ON(PageLRU(page)); |
| 2489 | SetPageLRU(page); |
| 2490 | add_page_to_lru_list(zone, page, page_lru(page)); |
| 2491 | } |
| 2492 | spin_unlock_irq(&zone->lru_lock); |
| 2493 | } |
| 2494 | |
| 2495 | if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) |
| 2496 | anon = true; |
| 2497 | else |
| 2498 | anon = false; |
| 2499 | |
| 2500 | mem_cgroup_charge_statistics(memcg, anon, nr_pages); |
| 2501 | unlock_page_cgroup(pc); |
| 2502 | |
| 2503 | /* |
| 2504 | * "charge_statistics" updated event counter. Then, check it. |
| 2505 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. |
| 2506 | * if they exceeds softlimit. |
| 2507 | */ |
| 2508 | memcg_check_events(memcg, page); |
| 2509 | } |
| 2510 | |
| 2511 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 2512 | |
| 2513 | #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ |
| 2514 | (1 << PCG_MIGRATION)) |
| 2515 | /* |
| 2516 | * Because tail pages are not marked as "used", set it. We're under |
| 2517 | * zone->lru_lock, 'splitting on pmd' and compound_lock. |
| 2518 | * charge/uncharge will be never happen and move_account() is done under |
| 2519 | * compound_lock(), so we don't have to take care of races. |
| 2520 | */ |
| 2521 | void mem_cgroup_split_huge_fixup(struct page *head) |
| 2522 | { |
| 2523 | struct page_cgroup *head_pc = lookup_page_cgroup(head); |
| 2524 | struct page_cgroup *pc; |
| 2525 | int i; |
| 2526 | |
| 2527 | if (mem_cgroup_disabled()) |
| 2528 | return; |
| 2529 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
| 2530 | pc = head_pc + i; |
| 2531 | pc->mem_cgroup = head_pc->mem_cgroup; |
| 2532 | smp_wmb();/* see __commit_charge() */ |
| 2533 | pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; |
| 2534 | } |
| 2535 | } |
| 2536 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |
| 2537 | |
| 2538 | /** |
| 2539 | * mem_cgroup_move_account - move account of the page |
| 2540 | * @page: the page |
| 2541 | * @nr_pages: number of regular pages (>1 for huge pages) |
| 2542 | * @pc: page_cgroup of the page. |
| 2543 | * @from: mem_cgroup which the page is moved from. |
| 2544 | * @to: mem_cgroup which the page is moved to. @from != @to. |
| 2545 | * @uncharge: whether we should call uncharge and css_put against @from. |
| 2546 | * |
| 2547 | * The caller must confirm following. |
| 2548 | * - page is not on LRU (isolate_page() is useful.) |
| 2549 | * - compound_lock is held when nr_pages > 1 |
| 2550 | * |
| 2551 | * This function doesn't do "charge" nor css_get to new cgroup. It should be |
| 2552 | * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is |
| 2553 | * true, this function does "uncharge" from old cgroup, but it doesn't if |
| 2554 | * @uncharge is false, so a caller should do "uncharge". |
| 2555 | */ |
| 2556 | static int mem_cgroup_move_account(struct page *page, |
| 2557 | unsigned int nr_pages, |
| 2558 | struct page_cgroup *pc, |
| 2559 | struct mem_cgroup *from, |
| 2560 | struct mem_cgroup *to, |
| 2561 | bool uncharge) |
| 2562 | { |
| 2563 | unsigned long flags; |
| 2564 | int ret; |
| 2565 | bool anon = PageAnon(page); |
| 2566 | |
| 2567 | VM_BUG_ON(from == to); |
| 2568 | VM_BUG_ON(PageLRU(page)); |
| 2569 | /* |
| 2570 | * The page is isolated from LRU. So, collapse function |
| 2571 | * will not handle this page. But page splitting can happen. |
| 2572 | * Do this check under compound_page_lock(). The caller should |
| 2573 | * hold it. |
| 2574 | */ |
| 2575 | ret = -EBUSY; |
| 2576 | if (nr_pages > 1 && !PageTransHuge(page)) |
| 2577 | goto out; |
| 2578 | |
| 2579 | lock_page_cgroup(pc); |
| 2580 | |
| 2581 | ret = -EINVAL; |
| 2582 | if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) |
| 2583 | goto unlock; |
| 2584 | |
| 2585 | move_lock_page_cgroup(pc, &flags); |
| 2586 | |
| 2587 | if (PageCgroupFileMapped(pc)) { |
| 2588 | /* Update mapped_file data for mem_cgroup */ |
| 2589 | preempt_disable(); |
| 2590 | __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
| 2591 | __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
| 2592 | preempt_enable(); |
| 2593 | } |
| 2594 | mem_cgroup_charge_statistics(from, anon, -nr_pages); |
| 2595 | if (uncharge) |
| 2596 | /* This is not "cancel", but cancel_charge does all we need. */ |
| 2597 | __mem_cgroup_cancel_charge(from, nr_pages); |
| 2598 | |
| 2599 | /* caller should have done css_get */ |
| 2600 | pc->mem_cgroup = to; |
| 2601 | mem_cgroup_charge_statistics(to, anon, nr_pages); |
| 2602 | /* |
| 2603 | * We charges against "to" which may not have any tasks. Then, "to" |
| 2604 | * can be under rmdir(). But in current implementation, caller of |
| 2605 | * this function is just force_empty() and move charge, so it's |
| 2606 | * guaranteed that "to" is never removed. So, we don't check rmdir |
| 2607 | * status here. |
| 2608 | */ |
| 2609 | move_unlock_page_cgroup(pc, &flags); |
| 2610 | ret = 0; |
| 2611 | unlock: |
| 2612 | unlock_page_cgroup(pc); |
| 2613 | /* |
| 2614 | * check events |
| 2615 | */ |
| 2616 | memcg_check_events(to, page); |
| 2617 | memcg_check_events(from, page); |
| 2618 | out: |
| 2619 | return ret; |
| 2620 | } |
| 2621 | |
| 2622 | /* |
| 2623 | * move charges to its parent. |
| 2624 | */ |
| 2625 | |
| 2626 | static int mem_cgroup_move_parent(struct page *page, |
| 2627 | struct page_cgroup *pc, |
| 2628 | struct mem_cgroup *child, |
| 2629 | gfp_t gfp_mask) |
| 2630 | { |
| 2631 | struct cgroup *cg = child->css.cgroup; |
| 2632 | struct cgroup *pcg = cg->parent; |
| 2633 | struct mem_cgroup *parent; |
| 2634 | unsigned int nr_pages; |
| 2635 | unsigned long uninitialized_var(flags); |
| 2636 | int ret; |
| 2637 | |
| 2638 | /* Is ROOT ? */ |
| 2639 | if (!pcg) |
| 2640 | return -EINVAL; |
| 2641 | |
| 2642 | ret = -EBUSY; |
| 2643 | if (!get_page_unless_zero(page)) |
| 2644 | goto out; |
| 2645 | if (isolate_lru_page(page)) |
| 2646 | goto put; |
| 2647 | |
| 2648 | nr_pages = hpage_nr_pages(page); |
| 2649 | |
| 2650 | parent = mem_cgroup_from_cont(pcg); |
| 2651 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); |
| 2652 | if (ret) |
| 2653 | goto put_back; |
| 2654 | |
| 2655 | if (nr_pages > 1) |
| 2656 | flags = compound_lock_irqsave(page); |
| 2657 | |
| 2658 | ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); |
| 2659 | if (ret) |
| 2660 | __mem_cgroup_cancel_charge(parent, nr_pages); |
| 2661 | |
| 2662 | if (nr_pages > 1) |
| 2663 | compound_unlock_irqrestore(page, flags); |
| 2664 | put_back: |
| 2665 | putback_lru_page(page); |
| 2666 | put: |
| 2667 | put_page(page); |
| 2668 | out: |
| 2669 | return ret; |
| 2670 | } |
| 2671 | |
| 2672 | /* |
| 2673 | * Charge the memory controller for page usage. |
| 2674 | * Return |
| 2675 | * 0 if the charge was successful |
| 2676 | * < 0 if the cgroup is over its limit |
| 2677 | */ |
| 2678 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, |
| 2679 | gfp_t gfp_mask, enum charge_type ctype) |
| 2680 | { |
| 2681 | struct mem_cgroup *memcg = NULL; |
| 2682 | unsigned int nr_pages = 1; |
| 2683 | struct page_cgroup *pc; |
| 2684 | bool oom = true; |
| 2685 | int ret; |
| 2686 | |
| 2687 | if (PageTransHuge(page)) { |
| 2688 | nr_pages <<= compound_order(page); |
| 2689 | VM_BUG_ON(!PageTransHuge(page)); |
| 2690 | /* |
| 2691 | * Never OOM-kill a process for a huge page. The |
| 2692 | * fault handler will fall back to regular pages. |
| 2693 | */ |
| 2694 | oom = false; |
| 2695 | } |
| 2696 | |
| 2697 | pc = lookup_page_cgroup(page); |
| 2698 | ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); |
| 2699 | if (ret == -ENOMEM) |
| 2700 | return ret; |
| 2701 | __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype, false); |
| 2702 | return 0; |
| 2703 | } |
| 2704 | |
| 2705 | int mem_cgroup_newpage_charge(struct page *page, |
| 2706 | struct mm_struct *mm, gfp_t gfp_mask) |
| 2707 | { |
| 2708 | if (mem_cgroup_disabled()) |
| 2709 | return 0; |
| 2710 | VM_BUG_ON(page_mapped(page)); |
| 2711 | VM_BUG_ON(page->mapping && !PageAnon(page)); |
| 2712 | VM_BUG_ON(!mm); |
| 2713 | return mem_cgroup_charge_common(page, mm, gfp_mask, |
| 2714 | MEM_CGROUP_CHARGE_TYPE_MAPPED); |
| 2715 | } |
| 2716 | |
| 2717 | static void |
| 2718 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, |
| 2719 | enum charge_type ctype); |
| 2720 | |
| 2721 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, |
| 2722 | gfp_t gfp_mask) |
| 2723 | { |
| 2724 | struct mem_cgroup *memcg = NULL; |
| 2725 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; |
| 2726 | int ret; |
| 2727 | |
| 2728 | if (mem_cgroup_disabled()) |
| 2729 | return 0; |
| 2730 | if (PageCompound(page)) |
| 2731 | return 0; |
| 2732 | |
| 2733 | if (unlikely(!mm)) |
| 2734 | mm = &init_mm; |
| 2735 | if (!page_is_file_cache(page)) |
| 2736 | type = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
| 2737 | |
| 2738 | if (!PageSwapCache(page)) |
| 2739 | ret = mem_cgroup_charge_common(page, mm, gfp_mask, type); |
| 2740 | else { /* page is swapcache/shmem */ |
| 2741 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); |
| 2742 | if (!ret) |
| 2743 | __mem_cgroup_commit_charge_swapin(page, memcg, type); |
| 2744 | } |
| 2745 | return ret; |
| 2746 | } |
| 2747 | |
| 2748 | /* |
| 2749 | * While swap-in, try_charge -> commit or cancel, the page is locked. |
| 2750 | * And when try_charge() successfully returns, one refcnt to memcg without |
| 2751 | * struct page_cgroup is acquired. This refcnt will be consumed by |
| 2752 | * "commit()" or removed by "cancel()" |
| 2753 | */ |
| 2754 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, |
| 2755 | struct page *page, |
| 2756 | gfp_t mask, struct mem_cgroup **memcgp) |
| 2757 | { |
| 2758 | struct mem_cgroup *memcg; |
| 2759 | int ret; |
| 2760 | |
| 2761 | *memcgp = NULL; |
| 2762 | |
| 2763 | if (mem_cgroup_disabled()) |
| 2764 | return 0; |
| 2765 | |
| 2766 | if (!do_swap_account) |
| 2767 | goto charge_cur_mm; |
| 2768 | /* |
| 2769 | * A racing thread's fault, or swapoff, may have already updated |
| 2770 | * the pte, and even removed page from swap cache: in those cases |
| 2771 | * do_swap_page()'s pte_same() test will fail; but there's also a |
| 2772 | * KSM case which does need to charge the page. |
| 2773 | */ |
| 2774 | if (!PageSwapCache(page)) |
| 2775 | goto charge_cur_mm; |
| 2776 | memcg = try_get_mem_cgroup_from_page(page); |
| 2777 | if (!memcg) |
| 2778 | goto charge_cur_mm; |
| 2779 | *memcgp = memcg; |
| 2780 | ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true); |
| 2781 | css_put(&memcg->css); |
| 2782 | if (ret == -EINTR) |
| 2783 | ret = 0; |
| 2784 | return ret; |
| 2785 | charge_cur_mm: |
| 2786 | if (unlikely(!mm)) |
| 2787 | mm = &init_mm; |
| 2788 | ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true); |
| 2789 | if (ret == -EINTR) |
| 2790 | ret = 0; |
| 2791 | return ret; |
| 2792 | } |
| 2793 | |
| 2794 | static void |
| 2795 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, |
| 2796 | enum charge_type ctype) |
| 2797 | { |
| 2798 | struct page_cgroup *pc; |
| 2799 | |
| 2800 | if (mem_cgroup_disabled()) |
| 2801 | return; |
| 2802 | if (!memcg) |
| 2803 | return; |
| 2804 | cgroup_exclude_rmdir(&memcg->css); |
| 2805 | |
| 2806 | pc = lookup_page_cgroup(page); |
| 2807 | __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype, true); |
| 2808 | /* |
| 2809 | * Now swap is on-memory. This means this page may be |
| 2810 | * counted both as mem and swap....double count. |
| 2811 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable |
| 2812 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() |
| 2813 | * may call delete_from_swap_cache() before reach here. |
| 2814 | */ |
| 2815 | if (do_swap_account && PageSwapCache(page)) { |
| 2816 | swp_entry_t ent = {.val = page_private(page)}; |
| 2817 | struct mem_cgroup *swap_memcg; |
| 2818 | unsigned short id; |
| 2819 | |
| 2820 | id = swap_cgroup_record(ent, 0); |
| 2821 | rcu_read_lock(); |
| 2822 | swap_memcg = mem_cgroup_lookup(id); |
| 2823 | if (swap_memcg) { |
| 2824 | /* |
| 2825 | * This recorded memcg can be obsolete one. So, avoid |
| 2826 | * calling css_tryget |
| 2827 | */ |
| 2828 | if (!mem_cgroup_is_root(swap_memcg)) |
| 2829 | res_counter_uncharge(&swap_memcg->memsw, |
| 2830 | PAGE_SIZE); |
| 2831 | mem_cgroup_swap_statistics(swap_memcg, false); |
| 2832 | mem_cgroup_put(swap_memcg); |
| 2833 | } |
| 2834 | rcu_read_unlock(); |
| 2835 | } |
| 2836 | /* |
| 2837 | * At swapin, we may charge account against cgroup which has no tasks. |
| 2838 | * So, rmdir()->pre_destroy() can be called while we do this charge. |
| 2839 | * In that case, we need to call pre_destroy() again. check it here. |
| 2840 | */ |
| 2841 | cgroup_release_and_wakeup_rmdir(&memcg->css); |
| 2842 | } |
| 2843 | |
| 2844 | void mem_cgroup_commit_charge_swapin(struct page *page, |
| 2845 | struct mem_cgroup *memcg) |
| 2846 | { |
| 2847 | __mem_cgroup_commit_charge_swapin(page, memcg, |
| 2848 | MEM_CGROUP_CHARGE_TYPE_MAPPED); |
| 2849 | } |
| 2850 | |
| 2851 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) |
| 2852 | { |
| 2853 | if (mem_cgroup_disabled()) |
| 2854 | return; |
| 2855 | if (!memcg) |
| 2856 | return; |
| 2857 | __mem_cgroup_cancel_charge(memcg, 1); |
| 2858 | } |
| 2859 | |
| 2860 | static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, |
| 2861 | unsigned int nr_pages, |
| 2862 | const enum charge_type ctype) |
| 2863 | { |
| 2864 | struct memcg_batch_info *batch = NULL; |
| 2865 | bool uncharge_memsw = true; |
| 2866 | |
| 2867 | /* If swapout, usage of swap doesn't decrease */ |
| 2868 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) |
| 2869 | uncharge_memsw = false; |
| 2870 | |
| 2871 | batch = ¤t->memcg_batch; |
| 2872 | /* |
| 2873 | * In usual, we do css_get() when we remember memcg pointer. |
| 2874 | * But in this case, we keep res->usage until end of a series of |
| 2875 | * uncharges. Then, it's ok to ignore memcg's refcnt. |
| 2876 | */ |
| 2877 | if (!batch->memcg) |
| 2878 | batch->memcg = memcg; |
| 2879 | /* |
| 2880 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. |
| 2881 | * In those cases, all pages freed continuously can be expected to be in |
| 2882 | * the same cgroup and we have chance to coalesce uncharges. |
| 2883 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) |
| 2884 | * because we want to do uncharge as soon as possible. |
| 2885 | */ |
| 2886 | |
| 2887 | if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) |
| 2888 | goto direct_uncharge; |
| 2889 | |
| 2890 | if (nr_pages > 1) |
| 2891 | goto direct_uncharge; |
| 2892 | |
| 2893 | /* |
| 2894 | * In typical case, batch->memcg == mem. This means we can |
| 2895 | * merge a series of uncharges to an uncharge of res_counter. |
| 2896 | * If not, we uncharge res_counter ony by one. |
| 2897 | */ |
| 2898 | if (batch->memcg != memcg) |
| 2899 | goto direct_uncharge; |
| 2900 | /* remember freed charge and uncharge it later */ |
| 2901 | batch->nr_pages++; |
| 2902 | if (uncharge_memsw) |
| 2903 | batch->memsw_nr_pages++; |
| 2904 | return; |
| 2905 | direct_uncharge: |
| 2906 | res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); |
| 2907 | if (uncharge_memsw) |
| 2908 | res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); |
| 2909 | if (unlikely(batch->memcg != memcg)) |
| 2910 | memcg_oom_recover(memcg); |
| 2911 | } |
| 2912 | |
| 2913 | /* |
| 2914 | * uncharge if !page_mapped(page) |
| 2915 | */ |
| 2916 | static struct mem_cgroup * |
| 2917 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) |
| 2918 | { |
| 2919 | struct mem_cgroup *memcg = NULL; |
| 2920 | unsigned int nr_pages = 1; |
| 2921 | struct page_cgroup *pc; |
| 2922 | bool anon; |
| 2923 | |
| 2924 | if (mem_cgroup_disabled()) |
| 2925 | return NULL; |
| 2926 | |
| 2927 | if (PageSwapCache(page)) |
| 2928 | return NULL; |
| 2929 | |
| 2930 | if (PageTransHuge(page)) { |
| 2931 | nr_pages <<= compound_order(page); |
| 2932 | VM_BUG_ON(!PageTransHuge(page)); |
| 2933 | } |
| 2934 | /* |
| 2935 | * Check if our page_cgroup is valid |
| 2936 | */ |
| 2937 | pc = lookup_page_cgroup(page); |
| 2938 | if (unlikely(!PageCgroupUsed(pc))) |
| 2939 | return NULL; |
| 2940 | |
| 2941 | lock_page_cgroup(pc); |
| 2942 | |
| 2943 | memcg = pc->mem_cgroup; |
| 2944 | |
| 2945 | if (!PageCgroupUsed(pc)) |
| 2946 | goto unlock_out; |
| 2947 | |
| 2948 | anon = PageAnon(page); |
| 2949 | |
| 2950 | switch (ctype) { |
| 2951 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: |
| 2952 | anon = true; |
| 2953 | /* fallthrough */ |
| 2954 | case MEM_CGROUP_CHARGE_TYPE_DROP: |
| 2955 | /* See mem_cgroup_prepare_migration() */ |
| 2956 | if (page_mapped(page) || PageCgroupMigration(pc)) |
| 2957 | goto unlock_out; |
| 2958 | break; |
| 2959 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: |
| 2960 | if (!PageAnon(page)) { /* Shared memory */ |
| 2961 | if (page->mapping && !page_is_file_cache(page)) |
| 2962 | goto unlock_out; |
| 2963 | } else if (page_mapped(page)) /* Anon */ |
| 2964 | goto unlock_out; |
| 2965 | break; |
| 2966 | default: |
| 2967 | break; |
| 2968 | } |
| 2969 | |
| 2970 | mem_cgroup_charge_statistics(memcg, anon, -nr_pages); |
| 2971 | |
| 2972 | ClearPageCgroupUsed(pc); |
| 2973 | /* |
| 2974 | * pc->mem_cgroup is not cleared here. It will be accessed when it's |
| 2975 | * freed from LRU. This is safe because uncharged page is expected not |
| 2976 | * to be reused (freed soon). Exception is SwapCache, it's handled by |
| 2977 | * special functions. |
| 2978 | */ |
| 2979 | |
| 2980 | unlock_page_cgroup(pc); |
| 2981 | /* |
| 2982 | * even after unlock, we have memcg->res.usage here and this memcg |
| 2983 | * will never be freed. |
| 2984 | */ |
| 2985 | memcg_check_events(memcg, page); |
| 2986 | if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { |
| 2987 | mem_cgroup_swap_statistics(memcg, true); |
| 2988 | mem_cgroup_get(memcg); |
| 2989 | } |
| 2990 | if (!mem_cgroup_is_root(memcg)) |
| 2991 | mem_cgroup_do_uncharge(memcg, nr_pages, ctype); |
| 2992 | |
| 2993 | return memcg; |
| 2994 | |
| 2995 | unlock_out: |
| 2996 | unlock_page_cgroup(pc); |
| 2997 | return NULL; |
| 2998 | } |
| 2999 | |
| 3000 | void mem_cgroup_uncharge_page(struct page *page) |
| 3001 | { |
| 3002 | /* early check. */ |
| 3003 | if (page_mapped(page)) |
| 3004 | return; |
| 3005 | VM_BUG_ON(page->mapping && !PageAnon(page)); |
| 3006 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); |
| 3007 | } |
| 3008 | |
| 3009 | void mem_cgroup_uncharge_cache_page(struct page *page) |
| 3010 | { |
| 3011 | VM_BUG_ON(page_mapped(page)); |
| 3012 | VM_BUG_ON(page->mapping); |
| 3013 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); |
| 3014 | } |
| 3015 | |
| 3016 | /* |
| 3017 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. |
| 3018 | * In that cases, pages are freed continuously and we can expect pages |
| 3019 | * are in the same memcg. All these calls itself limits the number of |
| 3020 | * pages freed at once, then uncharge_start/end() is called properly. |
| 3021 | * This may be called prural(2) times in a context, |
| 3022 | */ |
| 3023 | |
| 3024 | void mem_cgroup_uncharge_start(void) |
| 3025 | { |
| 3026 | current->memcg_batch.do_batch++; |
| 3027 | /* We can do nest. */ |
| 3028 | if (current->memcg_batch.do_batch == 1) { |
| 3029 | current->memcg_batch.memcg = NULL; |
| 3030 | current->memcg_batch.nr_pages = 0; |
| 3031 | current->memcg_batch.memsw_nr_pages = 0; |
| 3032 | } |
| 3033 | } |
| 3034 | |
| 3035 | void mem_cgroup_uncharge_end(void) |
| 3036 | { |
| 3037 | struct memcg_batch_info *batch = ¤t->memcg_batch; |
| 3038 | |
| 3039 | if (!batch->do_batch) |
| 3040 | return; |
| 3041 | |
| 3042 | batch->do_batch--; |
| 3043 | if (batch->do_batch) /* If stacked, do nothing. */ |
| 3044 | return; |
| 3045 | |
| 3046 | if (!batch->memcg) |
| 3047 | return; |
| 3048 | /* |
| 3049 | * This "batch->memcg" is valid without any css_get/put etc... |
| 3050 | * bacause we hide charges behind us. |
| 3051 | */ |
| 3052 | if (batch->nr_pages) |
| 3053 | res_counter_uncharge(&batch->memcg->res, |
| 3054 | batch->nr_pages * PAGE_SIZE); |
| 3055 | if (batch->memsw_nr_pages) |
| 3056 | res_counter_uncharge(&batch->memcg->memsw, |
| 3057 | batch->memsw_nr_pages * PAGE_SIZE); |
| 3058 | memcg_oom_recover(batch->memcg); |
| 3059 | /* forget this pointer (for sanity check) */ |
| 3060 | batch->memcg = NULL; |
| 3061 | } |
| 3062 | |
| 3063 | #ifdef CONFIG_SWAP |
| 3064 | /* |
| 3065 | * called after __delete_from_swap_cache() and drop "page" account. |
| 3066 | * memcg information is recorded to swap_cgroup of "ent" |
| 3067 | */ |
| 3068 | void |
| 3069 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) |
| 3070 | { |
| 3071 | struct mem_cgroup *memcg; |
| 3072 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; |
| 3073 | |
| 3074 | if (!swapout) /* this was a swap cache but the swap is unused ! */ |
| 3075 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; |
| 3076 | |
| 3077 | memcg = __mem_cgroup_uncharge_common(page, ctype); |
| 3078 | |
| 3079 | /* |
| 3080 | * record memcg information, if swapout && memcg != NULL, |
| 3081 | * mem_cgroup_get() was called in uncharge(). |
| 3082 | */ |
| 3083 | if (do_swap_account && swapout && memcg) |
| 3084 | swap_cgroup_record(ent, css_id(&memcg->css)); |
| 3085 | } |
| 3086 | #endif |
| 3087 | |
| 3088 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
| 3089 | /* |
| 3090 | * called from swap_entry_free(). remove record in swap_cgroup and |
| 3091 | * uncharge "memsw" account. |
| 3092 | */ |
| 3093 | void mem_cgroup_uncharge_swap(swp_entry_t ent) |
| 3094 | { |
| 3095 | struct mem_cgroup *memcg; |
| 3096 | unsigned short id; |
| 3097 | |
| 3098 | if (!do_swap_account) |
| 3099 | return; |
| 3100 | |
| 3101 | id = swap_cgroup_record(ent, 0); |
| 3102 | rcu_read_lock(); |
| 3103 | memcg = mem_cgroup_lookup(id); |
| 3104 | if (memcg) { |
| 3105 | /* |
| 3106 | * We uncharge this because swap is freed. |
| 3107 | * This memcg can be obsolete one. We avoid calling css_tryget |
| 3108 | */ |
| 3109 | if (!mem_cgroup_is_root(memcg)) |
| 3110 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); |
| 3111 | mem_cgroup_swap_statistics(memcg, false); |
| 3112 | mem_cgroup_put(memcg); |
| 3113 | } |
| 3114 | rcu_read_unlock(); |
| 3115 | } |
| 3116 | |
| 3117 | /** |
| 3118 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. |
| 3119 | * @entry: swap entry to be moved |
| 3120 | * @from: mem_cgroup which the entry is moved from |
| 3121 | * @to: mem_cgroup which the entry is moved to |
| 3122 | * @need_fixup: whether we should fixup res_counters and refcounts. |
| 3123 | * |
| 3124 | * It succeeds only when the swap_cgroup's record for this entry is the same |
| 3125 | * as the mem_cgroup's id of @from. |
| 3126 | * |
| 3127 | * Returns 0 on success, -EINVAL on failure. |
| 3128 | * |
| 3129 | * The caller must have charged to @to, IOW, called res_counter_charge() about |
| 3130 | * both res and memsw, and called css_get(). |
| 3131 | */ |
| 3132 | static int mem_cgroup_move_swap_account(swp_entry_t entry, |
| 3133 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) |
| 3134 | { |
| 3135 | unsigned short old_id, new_id; |
| 3136 | |
| 3137 | old_id = css_id(&from->css); |
| 3138 | new_id = css_id(&to->css); |
| 3139 | |
| 3140 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { |
| 3141 | mem_cgroup_swap_statistics(from, false); |
| 3142 | mem_cgroup_swap_statistics(to, true); |
| 3143 | /* |
| 3144 | * This function is only called from task migration context now. |
| 3145 | * It postpones res_counter and refcount handling till the end |
| 3146 | * of task migration(mem_cgroup_clear_mc()) for performance |
| 3147 | * improvement. But we cannot postpone mem_cgroup_get(to) |
| 3148 | * because if the process that has been moved to @to does |
| 3149 | * swap-in, the refcount of @to might be decreased to 0. |
| 3150 | */ |
| 3151 | mem_cgroup_get(to); |
| 3152 | if (need_fixup) { |
| 3153 | if (!mem_cgroup_is_root(from)) |
| 3154 | res_counter_uncharge(&from->memsw, PAGE_SIZE); |
| 3155 | mem_cgroup_put(from); |
| 3156 | /* |
| 3157 | * we charged both to->res and to->memsw, so we should |
| 3158 | * uncharge to->res. |
| 3159 | */ |
| 3160 | if (!mem_cgroup_is_root(to)) |
| 3161 | res_counter_uncharge(&to->res, PAGE_SIZE); |
| 3162 | } |
| 3163 | return 0; |
| 3164 | } |
| 3165 | return -EINVAL; |
| 3166 | } |
| 3167 | #else |
| 3168 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, |
| 3169 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) |
| 3170 | { |
| 3171 | return -EINVAL; |
| 3172 | } |
| 3173 | #endif |
| 3174 | |
| 3175 | /* |
| 3176 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old |
| 3177 | * page belongs to. |
| 3178 | */ |
| 3179 | int mem_cgroup_prepare_migration(struct page *page, |
| 3180 | struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask) |
| 3181 | { |
| 3182 | struct mem_cgroup *memcg = NULL; |
| 3183 | struct page_cgroup *pc; |
| 3184 | enum charge_type ctype; |
| 3185 | int ret = 0; |
| 3186 | |
| 3187 | *memcgp = NULL; |
| 3188 | |
| 3189 | VM_BUG_ON(PageTransHuge(page)); |
| 3190 | if (mem_cgroup_disabled()) |
| 3191 | return 0; |
| 3192 | |
| 3193 | pc = lookup_page_cgroup(page); |
| 3194 | lock_page_cgroup(pc); |
| 3195 | if (PageCgroupUsed(pc)) { |
| 3196 | memcg = pc->mem_cgroup; |
| 3197 | css_get(&memcg->css); |
| 3198 | /* |
| 3199 | * At migrating an anonymous page, its mapcount goes down |
| 3200 | * to 0 and uncharge() will be called. But, even if it's fully |
| 3201 | * unmapped, migration may fail and this page has to be |
| 3202 | * charged again. We set MIGRATION flag here and delay uncharge |
| 3203 | * until end_migration() is called |
| 3204 | * |
| 3205 | * Corner Case Thinking |
| 3206 | * A) |
| 3207 | * When the old page was mapped as Anon and it's unmap-and-freed |
| 3208 | * while migration was ongoing. |
| 3209 | * If unmap finds the old page, uncharge() of it will be delayed |
| 3210 | * until end_migration(). If unmap finds a new page, it's |
| 3211 | * uncharged when it make mapcount to be 1->0. If unmap code |
| 3212 | * finds swap_migration_entry, the new page will not be mapped |
| 3213 | * and end_migration() will find it(mapcount==0). |
| 3214 | * |
| 3215 | * B) |
| 3216 | * When the old page was mapped but migraion fails, the kernel |
| 3217 | * remaps it. A charge for it is kept by MIGRATION flag even |
| 3218 | * if mapcount goes down to 0. We can do remap successfully |
| 3219 | * without charging it again. |
| 3220 | * |
| 3221 | * C) |
| 3222 | * The "old" page is under lock_page() until the end of |
| 3223 | * migration, so, the old page itself will not be swapped-out. |
| 3224 | * If the new page is swapped out before end_migraton, our |
| 3225 | * hook to usual swap-out path will catch the event. |
| 3226 | */ |
| 3227 | if (PageAnon(page)) |
| 3228 | SetPageCgroupMigration(pc); |
| 3229 | } |
| 3230 | unlock_page_cgroup(pc); |
| 3231 | /* |
| 3232 | * If the page is not charged at this point, |
| 3233 | * we return here. |
| 3234 | */ |
| 3235 | if (!memcg) |
| 3236 | return 0; |
| 3237 | |
| 3238 | *memcgp = memcg; |
| 3239 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false); |
| 3240 | css_put(&memcg->css);/* drop extra refcnt */ |
| 3241 | if (ret) { |
| 3242 | if (PageAnon(page)) { |
| 3243 | lock_page_cgroup(pc); |
| 3244 | ClearPageCgroupMigration(pc); |
| 3245 | unlock_page_cgroup(pc); |
| 3246 | /* |
| 3247 | * The old page may be fully unmapped while we kept it. |
| 3248 | */ |
| 3249 | mem_cgroup_uncharge_page(page); |
| 3250 | } |
| 3251 | /* we'll need to revisit this error code (we have -EINTR) */ |
| 3252 | return -ENOMEM; |
| 3253 | } |
| 3254 | /* |
| 3255 | * We charge new page before it's used/mapped. So, even if unlock_page() |
| 3256 | * is called before end_migration, we can catch all events on this new |
| 3257 | * page. In the case new page is migrated but not remapped, new page's |
| 3258 | * mapcount will be finally 0 and we call uncharge in end_migration(). |
| 3259 | */ |
| 3260 | pc = lookup_page_cgroup(newpage); |
| 3261 | if (PageAnon(page)) |
| 3262 | ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; |
| 3263 | else if (page_is_file_cache(page)) |
| 3264 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; |
| 3265 | else |
| 3266 | ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
| 3267 | __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype, false); |
| 3268 | return ret; |
| 3269 | } |
| 3270 | |
| 3271 | /* remove redundant charge if migration failed*/ |
| 3272 | void mem_cgroup_end_migration(struct mem_cgroup *memcg, |
| 3273 | struct page *oldpage, struct page *newpage, bool migration_ok) |
| 3274 | { |
| 3275 | struct page *used, *unused; |
| 3276 | struct page_cgroup *pc; |
| 3277 | bool anon; |
| 3278 | |
| 3279 | if (!memcg) |
| 3280 | return; |
| 3281 | /* blocks rmdir() */ |
| 3282 | cgroup_exclude_rmdir(&memcg->css); |
| 3283 | if (!migration_ok) { |
| 3284 | used = oldpage; |
| 3285 | unused = newpage; |
| 3286 | } else { |
| 3287 | used = newpage; |
| 3288 | unused = oldpage; |
| 3289 | } |
| 3290 | /* |
| 3291 | * We disallowed uncharge of pages under migration because mapcount |
| 3292 | * of the page goes down to zero, temporarly. |
| 3293 | * Clear the flag and check the page should be charged. |
| 3294 | */ |
| 3295 | pc = lookup_page_cgroup(oldpage); |
| 3296 | lock_page_cgroup(pc); |
| 3297 | ClearPageCgroupMigration(pc); |
| 3298 | unlock_page_cgroup(pc); |
| 3299 | anon = PageAnon(used); |
| 3300 | __mem_cgroup_uncharge_common(unused, |
| 3301 | anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED |
| 3302 | : MEM_CGROUP_CHARGE_TYPE_CACHE); |
| 3303 | |
| 3304 | /* |
| 3305 | * If a page is a file cache, radix-tree replacement is very atomic |
| 3306 | * and we can skip this check. When it was an Anon page, its mapcount |
| 3307 | * goes down to 0. But because we added MIGRATION flage, it's not |
| 3308 | * uncharged yet. There are several case but page->mapcount check |
| 3309 | * and USED bit check in mem_cgroup_uncharge_page() will do enough |
| 3310 | * check. (see prepare_charge() also) |
| 3311 | */ |
| 3312 | if (anon) |
| 3313 | mem_cgroup_uncharge_page(used); |
| 3314 | /* |
| 3315 | * At migration, we may charge account against cgroup which has no |
| 3316 | * tasks. |
| 3317 | * So, rmdir()->pre_destroy() can be called while we do this charge. |
| 3318 | * In that case, we need to call pre_destroy() again. check it here. |
| 3319 | */ |
| 3320 | cgroup_release_and_wakeup_rmdir(&memcg->css); |
| 3321 | } |
| 3322 | |
| 3323 | /* |
| 3324 | * At replace page cache, newpage is not under any memcg but it's on |
| 3325 | * LRU. So, this function doesn't touch res_counter but handles LRU |
| 3326 | * in correct way. Both pages are locked so we cannot race with uncharge. |
| 3327 | */ |
| 3328 | void mem_cgroup_replace_page_cache(struct page *oldpage, |
| 3329 | struct page *newpage) |
| 3330 | { |
| 3331 | struct mem_cgroup *memcg; |
| 3332 | struct page_cgroup *pc; |
| 3333 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; |
| 3334 | |
| 3335 | if (mem_cgroup_disabled()) |
| 3336 | return; |
| 3337 | |
| 3338 | pc = lookup_page_cgroup(oldpage); |
| 3339 | /* fix accounting on old pages */ |
| 3340 | lock_page_cgroup(pc); |
| 3341 | memcg = pc->mem_cgroup; |
| 3342 | mem_cgroup_charge_statistics(memcg, false, -1); |
| 3343 | ClearPageCgroupUsed(pc); |
| 3344 | unlock_page_cgroup(pc); |
| 3345 | |
| 3346 | if (PageSwapBacked(oldpage)) |
| 3347 | type = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
| 3348 | |
| 3349 | /* |
| 3350 | * Even if newpage->mapping was NULL before starting replacement, |
| 3351 | * the newpage may be on LRU(or pagevec for LRU) already. We lock |
| 3352 | * LRU while we overwrite pc->mem_cgroup. |
| 3353 | */ |
| 3354 | __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type, true); |
| 3355 | } |
| 3356 | |
| 3357 | #ifdef CONFIG_DEBUG_VM |
| 3358 | static struct page_cgroup *lookup_page_cgroup_used(struct page *page) |
| 3359 | { |
| 3360 | struct page_cgroup *pc; |
| 3361 | |
| 3362 | pc = lookup_page_cgroup(page); |
| 3363 | /* |
| 3364 | * Can be NULL while feeding pages into the page allocator for |
| 3365 | * the first time, i.e. during boot or memory hotplug; |
| 3366 | * or when mem_cgroup_disabled(). |
| 3367 | */ |
| 3368 | if (likely(pc) && PageCgroupUsed(pc)) |
| 3369 | return pc; |
| 3370 | return NULL; |
| 3371 | } |
| 3372 | |
| 3373 | bool mem_cgroup_bad_page_check(struct page *page) |
| 3374 | { |
| 3375 | if (mem_cgroup_disabled()) |
| 3376 | return false; |
| 3377 | |
| 3378 | return lookup_page_cgroup_used(page) != NULL; |
| 3379 | } |
| 3380 | |
| 3381 | void mem_cgroup_print_bad_page(struct page *page) |
| 3382 | { |
| 3383 | struct page_cgroup *pc; |
| 3384 | |
| 3385 | pc = lookup_page_cgroup_used(page); |
| 3386 | if (pc) { |
| 3387 | printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n", |
| 3388 | pc, pc->flags, pc->mem_cgroup); |
| 3389 | } |
| 3390 | } |
| 3391 | #endif |
| 3392 | |
| 3393 | static DEFINE_MUTEX(set_limit_mutex); |
| 3394 | |
| 3395 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, |
| 3396 | unsigned long long val) |
| 3397 | { |
| 3398 | int retry_count; |
| 3399 | u64 memswlimit, memlimit; |
| 3400 | int ret = 0; |
| 3401 | int children = mem_cgroup_count_children(memcg); |
| 3402 | u64 curusage, oldusage; |
| 3403 | int enlarge; |
| 3404 | |
| 3405 | /* |
| 3406 | * For keeping hierarchical_reclaim simple, how long we should retry |
| 3407 | * is depends on callers. We set our retry-count to be function |
| 3408 | * of # of children which we should visit in this loop. |
| 3409 | */ |
| 3410 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; |
| 3411 | |
| 3412 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); |
| 3413 | |
| 3414 | enlarge = 0; |
| 3415 | while (retry_count) { |
| 3416 | if (signal_pending(current)) { |
| 3417 | ret = -EINTR; |
| 3418 | break; |
| 3419 | } |
| 3420 | /* |
| 3421 | * Rather than hide all in some function, I do this in |
| 3422 | * open coded manner. You see what this really does. |
| 3423 | * We have to guarantee memcg->res.limit < memcg->memsw.limit. |
| 3424 | */ |
| 3425 | mutex_lock(&set_limit_mutex); |
| 3426 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
| 3427 | if (memswlimit < val) { |
| 3428 | ret = -EINVAL; |
| 3429 | mutex_unlock(&set_limit_mutex); |
| 3430 | break; |
| 3431 | } |
| 3432 | |
| 3433 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
| 3434 | if (memlimit < val) |
| 3435 | enlarge = 1; |
| 3436 | |
| 3437 | ret = res_counter_set_limit(&memcg->res, val); |
| 3438 | if (!ret) { |
| 3439 | if (memswlimit == val) |
| 3440 | memcg->memsw_is_minimum = true; |
| 3441 | else |
| 3442 | memcg->memsw_is_minimum = false; |
| 3443 | } |
| 3444 | mutex_unlock(&set_limit_mutex); |
| 3445 | |
| 3446 | if (!ret) |
| 3447 | break; |
| 3448 | |
| 3449 | mem_cgroup_reclaim(memcg, GFP_KERNEL, |
| 3450 | MEM_CGROUP_RECLAIM_SHRINK); |
| 3451 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); |
| 3452 | /* Usage is reduced ? */ |
| 3453 | if (curusage >= oldusage) |
| 3454 | retry_count--; |
| 3455 | else |
| 3456 | oldusage = curusage; |
| 3457 | } |
| 3458 | if (!ret && enlarge) |
| 3459 | memcg_oom_recover(memcg); |
| 3460 | |
| 3461 | return ret; |
| 3462 | } |
| 3463 | |
| 3464 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, |
| 3465 | unsigned long long val) |
| 3466 | { |
| 3467 | int retry_count; |
| 3468 | u64 memlimit, memswlimit, oldusage, curusage; |
| 3469 | int children = mem_cgroup_count_children(memcg); |
| 3470 | int ret = -EBUSY; |
| 3471 | int enlarge = 0; |
| 3472 | |
| 3473 | /* see mem_cgroup_resize_res_limit */ |
| 3474 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; |
| 3475 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); |
| 3476 | while (retry_count) { |
| 3477 | if (signal_pending(current)) { |
| 3478 | ret = -EINTR; |
| 3479 | break; |
| 3480 | } |
| 3481 | /* |
| 3482 | * Rather than hide all in some function, I do this in |
| 3483 | * open coded manner. You see what this really does. |
| 3484 | * We have to guarantee memcg->res.limit < memcg->memsw.limit. |
| 3485 | */ |
| 3486 | mutex_lock(&set_limit_mutex); |
| 3487 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
| 3488 | if (memlimit > val) { |
| 3489 | ret = -EINVAL; |
| 3490 | mutex_unlock(&set_limit_mutex); |
| 3491 | break; |
| 3492 | } |
| 3493 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
| 3494 | if (memswlimit < val) |
| 3495 | enlarge = 1; |
| 3496 | ret = res_counter_set_limit(&memcg->memsw, val); |
| 3497 | if (!ret) { |
| 3498 | if (memlimit == val) |
| 3499 | memcg->memsw_is_minimum = true; |
| 3500 | else |
| 3501 | memcg->memsw_is_minimum = false; |
| 3502 | } |
| 3503 | mutex_unlock(&set_limit_mutex); |
| 3504 | |
| 3505 | if (!ret) |
| 3506 | break; |
| 3507 | |
| 3508 | mem_cgroup_reclaim(memcg, GFP_KERNEL, |
| 3509 | MEM_CGROUP_RECLAIM_NOSWAP | |
| 3510 | MEM_CGROUP_RECLAIM_SHRINK); |
| 3511 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); |
| 3512 | /* Usage is reduced ? */ |
| 3513 | if (curusage >= oldusage) |
| 3514 | retry_count--; |
| 3515 | else |
| 3516 | oldusage = curusage; |
| 3517 | } |
| 3518 | if (!ret && enlarge) |
| 3519 | memcg_oom_recover(memcg); |
| 3520 | return ret; |
| 3521 | } |
| 3522 | |
| 3523 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, |
| 3524 | gfp_t gfp_mask, |
| 3525 | unsigned long *total_scanned) |
| 3526 | { |
| 3527 | unsigned long nr_reclaimed = 0; |
| 3528 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; |
| 3529 | unsigned long reclaimed; |
| 3530 | int loop = 0; |
| 3531 | struct mem_cgroup_tree_per_zone *mctz; |
| 3532 | unsigned long long excess; |
| 3533 | unsigned long nr_scanned; |
| 3534 | |
| 3535 | if (order > 0) |
| 3536 | return 0; |
| 3537 | |
| 3538 | mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); |
| 3539 | /* |
| 3540 | * This loop can run a while, specially if mem_cgroup's continuously |
| 3541 | * keep exceeding their soft limit and putting the system under |
| 3542 | * pressure |
| 3543 | */ |
| 3544 | do { |
| 3545 | if (next_mz) |
| 3546 | mz = next_mz; |
| 3547 | else |
| 3548 | mz = mem_cgroup_largest_soft_limit_node(mctz); |
| 3549 | if (!mz) |
| 3550 | break; |
| 3551 | |
| 3552 | nr_scanned = 0; |
| 3553 | reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone, |
| 3554 | gfp_mask, &nr_scanned); |
| 3555 | nr_reclaimed += reclaimed; |
| 3556 | *total_scanned += nr_scanned; |
| 3557 | spin_lock(&mctz->lock); |
| 3558 | |
| 3559 | /* |
| 3560 | * If we failed to reclaim anything from this memory cgroup |
| 3561 | * it is time to move on to the next cgroup |
| 3562 | */ |
| 3563 | next_mz = NULL; |
| 3564 | if (!reclaimed) { |
| 3565 | do { |
| 3566 | /* |
| 3567 | * Loop until we find yet another one. |
| 3568 | * |
| 3569 | * By the time we get the soft_limit lock |
| 3570 | * again, someone might have aded the |
| 3571 | * group back on the RB tree. Iterate to |
| 3572 | * make sure we get a different mem. |
| 3573 | * mem_cgroup_largest_soft_limit_node returns |
| 3574 | * NULL if no other cgroup is present on |
| 3575 | * the tree |
| 3576 | */ |
| 3577 | next_mz = |
| 3578 | __mem_cgroup_largest_soft_limit_node(mctz); |
| 3579 | if (next_mz == mz) |
| 3580 | css_put(&next_mz->memcg->css); |
| 3581 | else /* next_mz == NULL or other memcg */ |
| 3582 | break; |
| 3583 | } while (1); |
| 3584 | } |
| 3585 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); |
| 3586 | excess = res_counter_soft_limit_excess(&mz->memcg->res); |
| 3587 | /* |
| 3588 | * One school of thought says that we should not add |
| 3589 | * back the node to the tree if reclaim returns 0. |
| 3590 | * But our reclaim could return 0, simply because due |
| 3591 | * to priority we are exposing a smaller subset of |
| 3592 | * memory to reclaim from. Consider this as a longer |
| 3593 | * term TODO. |
| 3594 | */ |
| 3595 | /* If excess == 0, no tree ops */ |
| 3596 | __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess); |
| 3597 | spin_unlock(&mctz->lock); |
| 3598 | css_put(&mz->memcg->css); |
| 3599 | loop++; |
| 3600 | /* |
| 3601 | * Could not reclaim anything and there are no more |
| 3602 | * mem cgroups to try or we seem to be looping without |
| 3603 | * reclaiming anything. |
| 3604 | */ |
| 3605 | if (!nr_reclaimed && |
| 3606 | (next_mz == NULL || |
| 3607 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) |
| 3608 | break; |
| 3609 | } while (!nr_reclaimed); |
| 3610 | if (next_mz) |
| 3611 | css_put(&next_mz->memcg->css); |
| 3612 | return nr_reclaimed; |
| 3613 | } |
| 3614 | |
| 3615 | /* |
| 3616 | * This routine traverse page_cgroup in given list and drop them all. |
| 3617 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. |
| 3618 | */ |
| 3619 | static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, |
| 3620 | int node, int zid, enum lru_list lru) |
| 3621 | { |
| 3622 | struct mem_cgroup_per_zone *mz; |
| 3623 | unsigned long flags, loop; |
| 3624 | struct list_head *list; |
| 3625 | struct page *busy; |
| 3626 | struct zone *zone; |
| 3627 | int ret = 0; |
| 3628 | |
| 3629 | zone = &NODE_DATA(node)->node_zones[zid]; |
| 3630 | mz = mem_cgroup_zoneinfo(memcg, node, zid); |
| 3631 | list = &mz->lruvec.lists[lru]; |
| 3632 | |
| 3633 | loop = mz->lru_size[lru]; |
| 3634 | /* give some margin against EBUSY etc...*/ |
| 3635 | loop += 256; |
| 3636 | busy = NULL; |
| 3637 | while (loop--) { |
| 3638 | struct page_cgroup *pc; |
| 3639 | struct page *page; |
| 3640 | |
| 3641 | ret = 0; |
| 3642 | spin_lock_irqsave(&zone->lru_lock, flags); |
| 3643 | if (list_empty(list)) { |
| 3644 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
| 3645 | break; |
| 3646 | } |
| 3647 | page = list_entry(list->prev, struct page, lru); |
| 3648 | if (busy == page) { |
| 3649 | list_move(&page->lru, list); |
| 3650 | busy = NULL; |
| 3651 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
| 3652 | continue; |
| 3653 | } |
| 3654 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
| 3655 | |
| 3656 | pc = lookup_page_cgroup(page); |
| 3657 | |
| 3658 | ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); |
| 3659 | if (ret == -ENOMEM || ret == -EINTR) |
| 3660 | break; |
| 3661 | |
| 3662 | if (ret == -EBUSY || ret == -EINVAL) { |
| 3663 | /* found lock contention or "pc" is obsolete. */ |
| 3664 | busy = page; |
| 3665 | cond_resched(); |
| 3666 | } else |
| 3667 | busy = NULL; |
| 3668 | } |
| 3669 | |
| 3670 | if (!ret && !list_empty(list)) |
| 3671 | return -EBUSY; |
| 3672 | return ret; |
| 3673 | } |
| 3674 | |
| 3675 | /* |
| 3676 | * make mem_cgroup's charge to be 0 if there is no task. |
| 3677 | * This enables deleting this mem_cgroup. |
| 3678 | */ |
| 3679 | static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) |
| 3680 | { |
| 3681 | int ret; |
| 3682 | int node, zid, shrink; |
| 3683 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; |
| 3684 | struct cgroup *cgrp = memcg->css.cgroup; |
| 3685 | |
| 3686 | css_get(&memcg->css); |
| 3687 | |
| 3688 | shrink = 0; |
| 3689 | /* should free all ? */ |
| 3690 | if (free_all) |
| 3691 | goto try_to_free; |
| 3692 | move_account: |
| 3693 | do { |
| 3694 | ret = -EBUSY; |
| 3695 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) |
| 3696 | goto out; |
| 3697 | ret = -EINTR; |
| 3698 | if (signal_pending(current)) |
| 3699 | goto out; |
| 3700 | /* This is for making all *used* pages to be on LRU. */ |
| 3701 | lru_add_drain_all(); |
| 3702 | drain_all_stock_sync(memcg); |
| 3703 | ret = 0; |
| 3704 | mem_cgroup_start_move(memcg); |
| 3705 | for_each_node_state(node, N_HIGH_MEMORY) { |
| 3706 | for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { |
| 3707 | enum lru_list lru; |
| 3708 | for_each_lru(lru) { |
| 3709 | ret = mem_cgroup_force_empty_list(memcg, |
| 3710 | node, zid, lru); |
| 3711 | if (ret) |
| 3712 | break; |
| 3713 | } |
| 3714 | } |
| 3715 | if (ret) |
| 3716 | break; |
| 3717 | } |
| 3718 | mem_cgroup_end_move(memcg); |
| 3719 | memcg_oom_recover(memcg); |
| 3720 | /* it seems parent cgroup doesn't have enough mem */ |
| 3721 | if (ret == -ENOMEM) |
| 3722 | goto try_to_free; |
| 3723 | cond_resched(); |
| 3724 | /* "ret" should also be checked to ensure all lists are empty. */ |
| 3725 | } while (memcg->res.usage > 0 || ret); |
| 3726 | out: |
| 3727 | css_put(&memcg->css); |
| 3728 | return ret; |
| 3729 | |
| 3730 | try_to_free: |
| 3731 | /* returns EBUSY if there is a task or if we come here twice. */ |
| 3732 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { |
| 3733 | ret = -EBUSY; |
| 3734 | goto out; |
| 3735 | } |
| 3736 | /* we call try-to-free pages for make this cgroup empty */ |
| 3737 | lru_add_drain_all(); |
| 3738 | /* try to free all pages in this cgroup */ |
| 3739 | shrink = 1; |
| 3740 | while (nr_retries && memcg->res.usage > 0) { |
| 3741 | int progress; |
| 3742 | |
| 3743 | if (signal_pending(current)) { |
| 3744 | ret = -EINTR; |
| 3745 | goto out; |
| 3746 | } |
| 3747 | progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, |
| 3748 | false); |
| 3749 | if (!progress) { |
| 3750 | nr_retries--; |
| 3751 | /* maybe some writeback is necessary */ |
| 3752 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
| 3753 | } |
| 3754 | |
| 3755 | } |
| 3756 | lru_add_drain(); |
| 3757 | /* try move_account...there may be some *locked* pages. */ |
| 3758 | goto move_account; |
| 3759 | } |
| 3760 | |
| 3761 | int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) |
| 3762 | { |
| 3763 | return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); |
| 3764 | } |
| 3765 | |
| 3766 | |
| 3767 | static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) |
| 3768 | { |
| 3769 | return mem_cgroup_from_cont(cont)->use_hierarchy; |
| 3770 | } |
| 3771 | |
| 3772 | static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, |
| 3773 | u64 val) |
| 3774 | { |
| 3775 | int retval = 0; |
| 3776 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 3777 | struct cgroup *parent = cont->parent; |
| 3778 | struct mem_cgroup *parent_memcg = NULL; |
| 3779 | |
| 3780 | if (parent) |
| 3781 | parent_memcg = mem_cgroup_from_cont(parent); |
| 3782 | |
| 3783 | cgroup_lock(); |
| 3784 | /* |
| 3785 | * If parent's use_hierarchy is set, we can't make any modifications |
| 3786 | * in the child subtrees. If it is unset, then the change can |
| 3787 | * occur, provided the current cgroup has no children. |
| 3788 | * |
| 3789 | * For the root cgroup, parent_mem is NULL, we allow value to be |
| 3790 | * set if there are no children. |
| 3791 | */ |
| 3792 | if ((!parent_memcg || !parent_memcg->use_hierarchy) && |
| 3793 | (val == 1 || val == 0)) { |
| 3794 | if (list_empty(&cont->children)) |
| 3795 | memcg->use_hierarchy = val; |
| 3796 | else |
| 3797 | retval = -EBUSY; |
| 3798 | } else |
| 3799 | retval = -EINVAL; |
| 3800 | cgroup_unlock(); |
| 3801 | |
| 3802 | return retval; |
| 3803 | } |
| 3804 | |
| 3805 | |
| 3806 | static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, |
| 3807 | enum mem_cgroup_stat_index idx) |
| 3808 | { |
| 3809 | struct mem_cgroup *iter; |
| 3810 | long val = 0; |
| 3811 | |
| 3812 | /* Per-cpu values can be negative, use a signed accumulator */ |
| 3813 | for_each_mem_cgroup_tree(iter, memcg) |
| 3814 | val += mem_cgroup_read_stat(iter, idx); |
| 3815 | |
| 3816 | if (val < 0) /* race ? */ |
| 3817 | val = 0; |
| 3818 | return val; |
| 3819 | } |
| 3820 | |
| 3821 | static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) |
| 3822 | { |
| 3823 | u64 val; |
| 3824 | |
| 3825 | if (!mem_cgroup_is_root(memcg)) { |
| 3826 | if (!swap) |
| 3827 | return res_counter_read_u64(&memcg->res, RES_USAGE); |
| 3828 | else |
| 3829 | return res_counter_read_u64(&memcg->memsw, RES_USAGE); |
| 3830 | } |
| 3831 | |
| 3832 | val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); |
| 3833 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); |
| 3834 | |
| 3835 | if (swap) |
| 3836 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); |
| 3837 | |
| 3838 | return val << PAGE_SHIFT; |
| 3839 | } |
| 3840 | |
| 3841 | static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) |
| 3842 | { |
| 3843 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 3844 | u64 val; |
| 3845 | int type, name; |
| 3846 | |
| 3847 | type = MEMFILE_TYPE(cft->private); |
| 3848 | name = MEMFILE_ATTR(cft->private); |
| 3849 | switch (type) { |
| 3850 | case _MEM: |
| 3851 | if (name == RES_USAGE) |
| 3852 | val = mem_cgroup_usage(memcg, false); |
| 3853 | else |
| 3854 | val = res_counter_read_u64(&memcg->res, name); |
| 3855 | break; |
| 3856 | case _MEMSWAP: |
| 3857 | if (name == RES_USAGE) |
| 3858 | val = mem_cgroup_usage(memcg, true); |
| 3859 | else |
| 3860 | val = res_counter_read_u64(&memcg->memsw, name); |
| 3861 | break; |
| 3862 | default: |
| 3863 | BUG(); |
| 3864 | break; |
| 3865 | } |
| 3866 | return val; |
| 3867 | } |
| 3868 | /* |
| 3869 | * The user of this function is... |
| 3870 | * RES_LIMIT. |
| 3871 | */ |
| 3872 | static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, |
| 3873 | const char *buffer) |
| 3874 | { |
| 3875 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 3876 | int type, name; |
| 3877 | unsigned long long val; |
| 3878 | int ret; |
| 3879 | |
| 3880 | type = MEMFILE_TYPE(cft->private); |
| 3881 | name = MEMFILE_ATTR(cft->private); |
| 3882 | switch (name) { |
| 3883 | case RES_LIMIT: |
| 3884 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ |
| 3885 | ret = -EINVAL; |
| 3886 | break; |
| 3887 | } |
| 3888 | /* This function does all necessary parse...reuse it */ |
| 3889 | ret = res_counter_memparse_write_strategy(buffer, &val); |
| 3890 | if (ret) |
| 3891 | break; |
| 3892 | if (type == _MEM) |
| 3893 | ret = mem_cgroup_resize_limit(memcg, val); |
| 3894 | else |
| 3895 | ret = mem_cgroup_resize_memsw_limit(memcg, val); |
| 3896 | break; |
| 3897 | case RES_SOFT_LIMIT: |
| 3898 | ret = res_counter_memparse_write_strategy(buffer, &val); |
| 3899 | if (ret) |
| 3900 | break; |
| 3901 | /* |
| 3902 | * For memsw, soft limits are hard to implement in terms |
| 3903 | * of semantics, for now, we support soft limits for |
| 3904 | * control without swap |
| 3905 | */ |
| 3906 | if (type == _MEM) |
| 3907 | ret = res_counter_set_soft_limit(&memcg->res, val); |
| 3908 | else |
| 3909 | ret = -EINVAL; |
| 3910 | break; |
| 3911 | default: |
| 3912 | ret = -EINVAL; /* should be BUG() ? */ |
| 3913 | break; |
| 3914 | } |
| 3915 | return ret; |
| 3916 | } |
| 3917 | |
| 3918 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, |
| 3919 | unsigned long long *mem_limit, unsigned long long *memsw_limit) |
| 3920 | { |
| 3921 | struct cgroup *cgroup; |
| 3922 | unsigned long long min_limit, min_memsw_limit, tmp; |
| 3923 | |
| 3924 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
| 3925 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
| 3926 | cgroup = memcg->css.cgroup; |
| 3927 | if (!memcg->use_hierarchy) |
| 3928 | goto out; |
| 3929 | |
| 3930 | while (cgroup->parent) { |
| 3931 | cgroup = cgroup->parent; |
| 3932 | memcg = mem_cgroup_from_cont(cgroup); |
| 3933 | if (!memcg->use_hierarchy) |
| 3934 | break; |
| 3935 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); |
| 3936 | min_limit = min(min_limit, tmp); |
| 3937 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
| 3938 | min_memsw_limit = min(min_memsw_limit, tmp); |
| 3939 | } |
| 3940 | out: |
| 3941 | *mem_limit = min_limit; |
| 3942 | *memsw_limit = min_memsw_limit; |
| 3943 | } |
| 3944 | |
| 3945 | static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) |
| 3946 | { |
| 3947 | struct mem_cgroup *memcg; |
| 3948 | int type, name; |
| 3949 | |
| 3950 | memcg = mem_cgroup_from_cont(cont); |
| 3951 | type = MEMFILE_TYPE(event); |
| 3952 | name = MEMFILE_ATTR(event); |
| 3953 | switch (name) { |
| 3954 | case RES_MAX_USAGE: |
| 3955 | if (type == _MEM) |
| 3956 | res_counter_reset_max(&memcg->res); |
| 3957 | else |
| 3958 | res_counter_reset_max(&memcg->memsw); |
| 3959 | break; |
| 3960 | case RES_FAILCNT: |
| 3961 | if (type == _MEM) |
| 3962 | res_counter_reset_failcnt(&memcg->res); |
| 3963 | else |
| 3964 | res_counter_reset_failcnt(&memcg->memsw); |
| 3965 | break; |
| 3966 | } |
| 3967 | |
| 3968 | return 0; |
| 3969 | } |
| 3970 | |
| 3971 | static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, |
| 3972 | struct cftype *cft) |
| 3973 | { |
| 3974 | return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; |
| 3975 | } |
| 3976 | |
| 3977 | #ifdef CONFIG_MMU |
| 3978 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, |
| 3979 | struct cftype *cft, u64 val) |
| 3980 | { |
| 3981 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 3982 | |
| 3983 | if (val >= (1 << NR_MOVE_TYPE)) |
| 3984 | return -EINVAL; |
| 3985 | /* |
| 3986 | * We check this value several times in both in can_attach() and |
| 3987 | * attach(), so we need cgroup lock to prevent this value from being |
| 3988 | * inconsistent. |
| 3989 | */ |
| 3990 | cgroup_lock(); |
| 3991 | memcg->move_charge_at_immigrate = val; |
| 3992 | cgroup_unlock(); |
| 3993 | |
| 3994 | return 0; |
| 3995 | } |
| 3996 | #else |
| 3997 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, |
| 3998 | struct cftype *cft, u64 val) |
| 3999 | { |
| 4000 | return -ENOSYS; |
| 4001 | } |
| 4002 | #endif |
| 4003 | |
| 4004 | |
| 4005 | /* For read statistics */ |
| 4006 | enum { |
| 4007 | MCS_CACHE, |
| 4008 | MCS_RSS, |
| 4009 | MCS_FILE_MAPPED, |
| 4010 | MCS_PGPGIN, |
| 4011 | MCS_PGPGOUT, |
| 4012 | MCS_SWAP, |
| 4013 | MCS_PGFAULT, |
| 4014 | MCS_PGMAJFAULT, |
| 4015 | MCS_INACTIVE_ANON, |
| 4016 | MCS_ACTIVE_ANON, |
| 4017 | MCS_INACTIVE_FILE, |
| 4018 | MCS_ACTIVE_FILE, |
| 4019 | MCS_UNEVICTABLE, |
| 4020 | NR_MCS_STAT, |
| 4021 | }; |
| 4022 | |
| 4023 | struct mcs_total_stat { |
| 4024 | s64 stat[NR_MCS_STAT]; |
| 4025 | }; |
| 4026 | |
| 4027 | struct { |
| 4028 | char *local_name; |
| 4029 | char *total_name; |
| 4030 | } memcg_stat_strings[NR_MCS_STAT] = { |
| 4031 | {"cache", "total_cache"}, |
| 4032 | {"rss", "total_rss"}, |
| 4033 | {"mapped_file", "total_mapped_file"}, |
| 4034 | {"pgpgin", "total_pgpgin"}, |
| 4035 | {"pgpgout", "total_pgpgout"}, |
| 4036 | {"swap", "total_swap"}, |
| 4037 | {"pgfault", "total_pgfault"}, |
| 4038 | {"pgmajfault", "total_pgmajfault"}, |
| 4039 | {"inactive_anon", "total_inactive_anon"}, |
| 4040 | {"active_anon", "total_active_anon"}, |
| 4041 | {"inactive_file", "total_inactive_file"}, |
| 4042 | {"active_file", "total_active_file"}, |
| 4043 | {"unevictable", "total_unevictable"} |
| 4044 | }; |
| 4045 | |
| 4046 | |
| 4047 | static void |
| 4048 | mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) |
| 4049 | { |
| 4050 | s64 val; |
| 4051 | |
| 4052 | /* per cpu stat */ |
| 4053 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); |
| 4054 | s->stat[MCS_CACHE] += val * PAGE_SIZE; |
| 4055 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); |
| 4056 | s->stat[MCS_RSS] += val * PAGE_SIZE; |
| 4057 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); |
| 4058 | s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; |
| 4059 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); |
| 4060 | s->stat[MCS_PGPGIN] += val; |
| 4061 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); |
| 4062 | s->stat[MCS_PGPGOUT] += val; |
| 4063 | if (do_swap_account) { |
| 4064 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); |
| 4065 | s->stat[MCS_SWAP] += val * PAGE_SIZE; |
| 4066 | } |
| 4067 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); |
| 4068 | s->stat[MCS_PGFAULT] += val; |
| 4069 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); |
| 4070 | s->stat[MCS_PGMAJFAULT] += val; |
| 4071 | |
| 4072 | /* per zone stat */ |
| 4073 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); |
| 4074 | s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; |
| 4075 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); |
| 4076 | s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; |
| 4077 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); |
| 4078 | s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; |
| 4079 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); |
| 4080 | s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; |
| 4081 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); |
| 4082 | s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; |
| 4083 | } |
| 4084 | |
| 4085 | static void |
| 4086 | mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) |
| 4087 | { |
| 4088 | struct mem_cgroup *iter; |
| 4089 | |
| 4090 | for_each_mem_cgroup_tree(iter, memcg) |
| 4091 | mem_cgroup_get_local_stat(iter, s); |
| 4092 | } |
| 4093 | |
| 4094 | #ifdef CONFIG_NUMA |
| 4095 | static int mem_control_numa_stat_show(struct seq_file *m, void *arg) |
| 4096 | { |
| 4097 | int nid; |
| 4098 | unsigned long total_nr, file_nr, anon_nr, unevictable_nr; |
| 4099 | unsigned long node_nr; |
| 4100 | struct cgroup *cont = m->private; |
| 4101 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 4102 | |
| 4103 | total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL); |
| 4104 | seq_printf(m, "total=%lu", total_nr); |
| 4105 | for_each_node_state(nid, N_HIGH_MEMORY) { |
| 4106 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL); |
| 4107 | seq_printf(m, " N%d=%lu", nid, node_nr); |
| 4108 | } |
| 4109 | seq_putc(m, '\n'); |
| 4110 | |
| 4111 | file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE); |
| 4112 | seq_printf(m, "file=%lu", file_nr); |
| 4113 | for_each_node_state(nid, N_HIGH_MEMORY) { |
| 4114 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, |
| 4115 | LRU_ALL_FILE); |
| 4116 | seq_printf(m, " N%d=%lu", nid, node_nr); |
| 4117 | } |
| 4118 | seq_putc(m, '\n'); |
| 4119 | |
| 4120 | anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON); |
| 4121 | seq_printf(m, "anon=%lu", anon_nr); |
| 4122 | for_each_node_state(nid, N_HIGH_MEMORY) { |
| 4123 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, |
| 4124 | LRU_ALL_ANON); |
| 4125 | seq_printf(m, " N%d=%lu", nid, node_nr); |
| 4126 | } |
| 4127 | seq_putc(m, '\n'); |
| 4128 | |
| 4129 | unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); |
| 4130 | seq_printf(m, "unevictable=%lu", unevictable_nr); |
| 4131 | for_each_node_state(nid, N_HIGH_MEMORY) { |
| 4132 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, |
| 4133 | BIT(LRU_UNEVICTABLE)); |
| 4134 | seq_printf(m, " N%d=%lu", nid, node_nr); |
| 4135 | } |
| 4136 | seq_putc(m, '\n'); |
| 4137 | return 0; |
| 4138 | } |
| 4139 | #endif /* CONFIG_NUMA */ |
| 4140 | |
| 4141 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, |
| 4142 | struct cgroup_map_cb *cb) |
| 4143 | { |
| 4144 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 4145 | struct mcs_total_stat mystat; |
| 4146 | int i; |
| 4147 | |
| 4148 | memset(&mystat, 0, sizeof(mystat)); |
| 4149 | mem_cgroup_get_local_stat(memcg, &mystat); |
| 4150 | |
| 4151 | |
| 4152 | for (i = 0; i < NR_MCS_STAT; i++) { |
| 4153 | if (i == MCS_SWAP && !do_swap_account) |
| 4154 | continue; |
| 4155 | cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); |
| 4156 | } |
| 4157 | |
| 4158 | /* Hierarchical information */ |
| 4159 | { |
| 4160 | unsigned long long limit, memsw_limit; |
| 4161 | memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit); |
| 4162 | cb->fill(cb, "hierarchical_memory_limit", limit); |
| 4163 | if (do_swap_account) |
| 4164 | cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); |
| 4165 | } |
| 4166 | |
| 4167 | memset(&mystat, 0, sizeof(mystat)); |
| 4168 | mem_cgroup_get_total_stat(memcg, &mystat); |
| 4169 | for (i = 0; i < NR_MCS_STAT; i++) { |
| 4170 | if (i == MCS_SWAP && !do_swap_account) |
| 4171 | continue; |
| 4172 | cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); |
| 4173 | } |
| 4174 | |
| 4175 | #ifdef CONFIG_DEBUG_VM |
| 4176 | { |
| 4177 | int nid, zid; |
| 4178 | struct mem_cgroup_per_zone *mz; |
| 4179 | unsigned long recent_rotated[2] = {0, 0}; |
| 4180 | unsigned long recent_scanned[2] = {0, 0}; |
| 4181 | |
| 4182 | for_each_online_node(nid) |
| 4183 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
| 4184 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
| 4185 | |
| 4186 | recent_rotated[0] += |
| 4187 | mz->reclaim_stat.recent_rotated[0]; |
| 4188 | recent_rotated[1] += |
| 4189 | mz->reclaim_stat.recent_rotated[1]; |
| 4190 | recent_scanned[0] += |
| 4191 | mz->reclaim_stat.recent_scanned[0]; |
| 4192 | recent_scanned[1] += |
| 4193 | mz->reclaim_stat.recent_scanned[1]; |
| 4194 | } |
| 4195 | cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); |
| 4196 | cb->fill(cb, "recent_rotated_file", recent_rotated[1]); |
| 4197 | cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); |
| 4198 | cb->fill(cb, "recent_scanned_file", recent_scanned[1]); |
| 4199 | } |
| 4200 | #endif |
| 4201 | |
| 4202 | return 0; |
| 4203 | } |
| 4204 | |
| 4205 | static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) |
| 4206 | { |
| 4207 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4208 | |
| 4209 | return mem_cgroup_swappiness(memcg); |
| 4210 | } |
| 4211 | |
| 4212 | static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, |
| 4213 | u64 val) |
| 4214 | { |
| 4215 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4216 | struct mem_cgroup *parent; |
| 4217 | |
| 4218 | if (val > 100) |
| 4219 | return -EINVAL; |
| 4220 | |
| 4221 | if (cgrp->parent == NULL) |
| 4222 | return -EINVAL; |
| 4223 | |
| 4224 | parent = mem_cgroup_from_cont(cgrp->parent); |
| 4225 | |
| 4226 | cgroup_lock(); |
| 4227 | |
| 4228 | /* If under hierarchy, only empty-root can set this value */ |
| 4229 | if ((parent->use_hierarchy) || |
| 4230 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { |
| 4231 | cgroup_unlock(); |
| 4232 | return -EINVAL; |
| 4233 | } |
| 4234 | |
| 4235 | memcg->swappiness = val; |
| 4236 | |
| 4237 | cgroup_unlock(); |
| 4238 | |
| 4239 | return 0; |
| 4240 | } |
| 4241 | |
| 4242 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) |
| 4243 | { |
| 4244 | struct mem_cgroup_threshold_ary *t; |
| 4245 | u64 usage; |
| 4246 | int i; |
| 4247 | |
| 4248 | rcu_read_lock(); |
| 4249 | if (!swap) |
| 4250 | t = rcu_dereference(memcg->thresholds.primary); |
| 4251 | else |
| 4252 | t = rcu_dereference(memcg->memsw_thresholds.primary); |
| 4253 | |
| 4254 | if (!t) |
| 4255 | goto unlock; |
| 4256 | |
| 4257 | usage = mem_cgroup_usage(memcg, swap); |
| 4258 | |
| 4259 | /* |
| 4260 | * current_threshold points to threshold just below usage. |
| 4261 | * If it's not true, a threshold was crossed after last |
| 4262 | * call of __mem_cgroup_threshold(). |
| 4263 | */ |
| 4264 | i = t->current_threshold; |
| 4265 | |
| 4266 | /* |
| 4267 | * Iterate backward over array of thresholds starting from |
| 4268 | * current_threshold and check if a threshold is crossed. |
| 4269 | * If none of thresholds below usage is crossed, we read |
| 4270 | * only one element of the array here. |
| 4271 | */ |
| 4272 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) |
| 4273 | eventfd_signal(t->entries[i].eventfd, 1); |
| 4274 | |
| 4275 | /* i = current_threshold + 1 */ |
| 4276 | i++; |
| 4277 | |
| 4278 | /* |
| 4279 | * Iterate forward over array of thresholds starting from |
| 4280 | * current_threshold+1 and check if a threshold is crossed. |
| 4281 | * If none of thresholds above usage is crossed, we read |
| 4282 | * only one element of the array here. |
| 4283 | */ |
| 4284 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) |
| 4285 | eventfd_signal(t->entries[i].eventfd, 1); |
| 4286 | |
| 4287 | /* Update current_threshold */ |
| 4288 | t->current_threshold = i - 1; |
| 4289 | unlock: |
| 4290 | rcu_read_unlock(); |
| 4291 | } |
| 4292 | |
| 4293 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) |
| 4294 | { |
| 4295 | while (memcg) { |
| 4296 | __mem_cgroup_threshold(memcg, false); |
| 4297 | if (do_swap_account) |
| 4298 | __mem_cgroup_threshold(memcg, true); |
| 4299 | |
| 4300 | memcg = parent_mem_cgroup(memcg); |
| 4301 | } |
| 4302 | } |
| 4303 | |
| 4304 | static int compare_thresholds(const void *a, const void *b) |
| 4305 | { |
| 4306 | const struct mem_cgroup_threshold *_a = a; |
| 4307 | const struct mem_cgroup_threshold *_b = b; |
| 4308 | |
| 4309 | return _a->threshold - _b->threshold; |
| 4310 | } |
| 4311 | |
| 4312 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) |
| 4313 | { |
| 4314 | struct mem_cgroup_eventfd_list *ev; |
| 4315 | |
| 4316 | list_for_each_entry(ev, &memcg->oom_notify, list) |
| 4317 | eventfd_signal(ev->eventfd, 1); |
| 4318 | return 0; |
| 4319 | } |
| 4320 | |
| 4321 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) |
| 4322 | { |
| 4323 | struct mem_cgroup *iter; |
| 4324 | |
| 4325 | for_each_mem_cgroup_tree(iter, memcg) |
| 4326 | mem_cgroup_oom_notify_cb(iter); |
| 4327 | } |
| 4328 | |
| 4329 | static int mem_cgroup_usage_register_event(struct cgroup *cgrp, |
| 4330 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) |
| 4331 | { |
| 4332 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4333 | struct mem_cgroup_thresholds *thresholds; |
| 4334 | struct mem_cgroup_threshold_ary *new; |
| 4335 | int type = MEMFILE_TYPE(cft->private); |
| 4336 | u64 threshold, usage; |
| 4337 | int i, size, ret; |
| 4338 | |
| 4339 | ret = res_counter_memparse_write_strategy(args, &threshold); |
| 4340 | if (ret) |
| 4341 | return ret; |
| 4342 | |
| 4343 | mutex_lock(&memcg->thresholds_lock); |
| 4344 | |
| 4345 | if (type == _MEM) |
| 4346 | thresholds = &memcg->thresholds; |
| 4347 | else if (type == _MEMSWAP) |
| 4348 | thresholds = &memcg->memsw_thresholds; |
| 4349 | else |
| 4350 | BUG(); |
| 4351 | |
| 4352 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); |
| 4353 | |
| 4354 | /* Check if a threshold crossed before adding a new one */ |
| 4355 | if (thresholds->primary) |
| 4356 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); |
| 4357 | |
| 4358 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; |
| 4359 | |
| 4360 | /* Allocate memory for new array of thresholds */ |
| 4361 | new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), |
| 4362 | GFP_KERNEL); |
| 4363 | if (!new) { |
| 4364 | ret = -ENOMEM; |
| 4365 | goto unlock; |
| 4366 | } |
| 4367 | new->size = size; |
| 4368 | |
| 4369 | /* Copy thresholds (if any) to new array */ |
| 4370 | if (thresholds->primary) { |
| 4371 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * |
| 4372 | sizeof(struct mem_cgroup_threshold)); |
| 4373 | } |
| 4374 | |
| 4375 | /* Add new threshold */ |
| 4376 | new->entries[size - 1].eventfd = eventfd; |
| 4377 | new->entries[size - 1].threshold = threshold; |
| 4378 | |
| 4379 | /* Sort thresholds. Registering of new threshold isn't time-critical */ |
| 4380 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), |
| 4381 | compare_thresholds, NULL); |
| 4382 | |
| 4383 | /* Find current threshold */ |
| 4384 | new->current_threshold = -1; |
| 4385 | for (i = 0; i < size; i++) { |
| 4386 | if (new->entries[i].threshold < usage) { |
| 4387 | /* |
| 4388 | * new->current_threshold will not be used until |
| 4389 | * rcu_assign_pointer(), so it's safe to increment |
| 4390 | * it here. |
| 4391 | */ |
| 4392 | ++new->current_threshold; |
| 4393 | } |
| 4394 | } |
| 4395 | |
| 4396 | /* Free old spare buffer and save old primary buffer as spare */ |
| 4397 | kfree(thresholds->spare); |
| 4398 | thresholds->spare = thresholds->primary; |
| 4399 | |
| 4400 | rcu_assign_pointer(thresholds->primary, new); |
| 4401 | |
| 4402 | /* To be sure that nobody uses thresholds */ |
| 4403 | synchronize_rcu(); |
| 4404 | |
| 4405 | unlock: |
| 4406 | mutex_unlock(&memcg->thresholds_lock); |
| 4407 | |
| 4408 | return ret; |
| 4409 | } |
| 4410 | |
| 4411 | static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, |
| 4412 | struct cftype *cft, struct eventfd_ctx *eventfd) |
| 4413 | { |
| 4414 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4415 | struct mem_cgroup_thresholds *thresholds; |
| 4416 | struct mem_cgroup_threshold_ary *new; |
| 4417 | int type = MEMFILE_TYPE(cft->private); |
| 4418 | u64 usage; |
| 4419 | int i, j, size; |
| 4420 | |
| 4421 | mutex_lock(&memcg->thresholds_lock); |
| 4422 | if (type == _MEM) |
| 4423 | thresholds = &memcg->thresholds; |
| 4424 | else if (type == _MEMSWAP) |
| 4425 | thresholds = &memcg->memsw_thresholds; |
| 4426 | else |
| 4427 | BUG(); |
| 4428 | |
| 4429 | /* |
| 4430 | * Something went wrong if we trying to unregister a threshold |
| 4431 | * if we don't have thresholds |
| 4432 | */ |
| 4433 | BUG_ON(!thresholds); |
| 4434 | |
| 4435 | if (!thresholds->primary) |
| 4436 | goto unlock; |
| 4437 | |
| 4438 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); |
| 4439 | |
| 4440 | /* Check if a threshold crossed before removing */ |
| 4441 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); |
| 4442 | |
| 4443 | /* Calculate new number of threshold */ |
| 4444 | size = 0; |
| 4445 | for (i = 0; i < thresholds->primary->size; i++) { |
| 4446 | if (thresholds->primary->entries[i].eventfd != eventfd) |
| 4447 | size++; |
| 4448 | } |
| 4449 | |
| 4450 | new = thresholds->spare; |
| 4451 | |
| 4452 | /* Set thresholds array to NULL if we don't have thresholds */ |
| 4453 | if (!size) { |
| 4454 | kfree(new); |
| 4455 | new = NULL; |
| 4456 | goto swap_buffers; |
| 4457 | } |
| 4458 | |
| 4459 | new->size = size; |
| 4460 | |
| 4461 | /* Copy thresholds and find current threshold */ |
| 4462 | new->current_threshold = -1; |
| 4463 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { |
| 4464 | if (thresholds->primary->entries[i].eventfd == eventfd) |
| 4465 | continue; |
| 4466 | |
| 4467 | new->entries[j] = thresholds->primary->entries[i]; |
| 4468 | if (new->entries[j].threshold < usage) { |
| 4469 | /* |
| 4470 | * new->current_threshold will not be used |
| 4471 | * until rcu_assign_pointer(), so it's safe to increment |
| 4472 | * it here. |
| 4473 | */ |
| 4474 | ++new->current_threshold; |
| 4475 | } |
| 4476 | j++; |
| 4477 | } |
| 4478 | |
| 4479 | swap_buffers: |
| 4480 | /* Swap primary and spare array */ |
| 4481 | thresholds->spare = thresholds->primary; |
| 4482 | rcu_assign_pointer(thresholds->primary, new); |
| 4483 | |
| 4484 | /* To be sure that nobody uses thresholds */ |
| 4485 | synchronize_rcu(); |
| 4486 | unlock: |
| 4487 | mutex_unlock(&memcg->thresholds_lock); |
| 4488 | } |
| 4489 | |
| 4490 | static int mem_cgroup_oom_register_event(struct cgroup *cgrp, |
| 4491 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) |
| 4492 | { |
| 4493 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4494 | struct mem_cgroup_eventfd_list *event; |
| 4495 | int type = MEMFILE_TYPE(cft->private); |
| 4496 | |
| 4497 | BUG_ON(type != _OOM_TYPE); |
| 4498 | event = kmalloc(sizeof(*event), GFP_KERNEL); |
| 4499 | if (!event) |
| 4500 | return -ENOMEM; |
| 4501 | |
| 4502 | spin_lock(&memcg_oom_lock); |
| 4503 | |
| 4504 | event->eventfd = eventfd; |
| 4505 | list_add(&event->list, &memcg->oom_notify); |
| 4506 | |
| 4507 | /* already in OOM ? */ |
| 4508 | if (atomic_read(&memcg->under_oom)) |
| 4509 | eventfd_signal(eventfd, 1); |
| 4510 | spin_unlock(&memcg_oom_lock); |
| 4511 | |
| 4512 | return 0; |
| 4513 | } |
| 4514 | |
| 4515 | static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, |
| 4516 | struct cftype *cft, struct eventfd_ctx *eventfd) |
| 4517 | { |
| 4518 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4519 | struct mem_cgroup_eventfd_list *ev, *tmp; |
| 4520 | int type = MEMFILE_TYPE(cft->private); |
| 4521 | |
| 4522 | BUG_ON(type != _OOM_TYPE); |
| 4523 | |
| 4524 | spin_lock(&memcg_oom_lock); |
| 4525 | |
| 4526 | list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { |
| 4527 | if (ev->eventfd == eventfd) { |
| 4528 | list_del(&ev->list); |
| 4529 | kfree(ev); |
| 4530 | } |
| 4531 | } |
| 4532 | |
| 4533 | spin_unlock(&memcg_oom_lock); |
| 4534 | } |
| 4535 | |
| 4536 | static int mem_cgroup_oom_control_read(struct cgroup *cgrp, |
| 4537 | struct cftype *cft, struct cgroup_map_cb *cb) |
| 4538 | { |
| 4539 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4540 | |
| 4541 | cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); |
| 4542 | |
| 4543 | if (atomic_read(&memcg->under_oom)) |
| 4544 | cb->fill(cb, "under_oom", 1); |
| 4545 | else |
| 4546 | cb->fill(cb, "under_oom", 0); |
| 4547 | return 0; |
| 4548 | } |
| 4549 | |
| 4550 | static int mem_cgroup_oom_control_write(struct cgroup *cgrp, |
| 4551 | struct cftype *cft, u64 val) |
| 4552 | { |
| 4553 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
| 4554 | struct mem_cgroup *parent; |
| 4555 | |
| 4556 | /* cannot set to root cgroup and only 0 and 1 are allowed */ |
| 4557 | if (!cgrp->parent || !((val == 0) || (val == 1))) |
| 4558 | return -EINVAL; |
| 4559 | |
| 4560 | parent = mem_cgroup_from_cont(cgrp->parent); |
| 4561 | |
| 4562 | cgroup_lock(); |
| 4563 | /* oom-kill-disable is a flag for subhierarchy. */ |
| 4564 | if ((parent->use_hierarchy) || |
| 4565 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { |
| 4566 | cgroup_unlock(); |
| 4567 | return -EINVAL; |
| 4568 | } |
| 4569 | memcg->oom_kill_disable = val; |
| 4570 | if (!val) |
| 4571 | memcg_oom_recover(memcg); |
| 4572 | cgroup_unlock(); |
| 4573 | return 0; |
| 4574 | } |
| 4575 | |
| 4576 | #ifdef CONFIG_NUMA |
| 4577 | static const struct file_operations mem_control_numa_stat_file_operations = { |
| 4578 | .read = seq_read, |
| 4579 | .llseek = seq_lseek, |
| 4580 | .release = single_release, |
| 4581 | }; |
| 4582 | |
| 4583 | static int mem_control_numa_stat_open(struct inode *unused, struct file *file) |
| 4584 | { |
| 4585 | struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; |
| 4586 | |
| 4587 | file->f_op = &mem_control_numa_stat_file_operations; |
| 4588 | return single_open(file, mem_control_numa_stat_show, cont); |
| 4589 | } |
| 4590 | #endif /* CONFIG_NUMA */ |
| 4591 | |
| 4592 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM |
| 4593 | static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) |
| 4594 | { |
| 4595 | /* |
| 4596 | * Part of this would be better living in a separate allocation |
| 4597 | * function, leaving us with just the cgroup tree population work. |
| 4598 | * We, however, depend on state such as network's proto_list that |
| 4599 | * is only initialized after cgroup creation. I found the less |
| 4600 | * cumbersome way to deal with it to defer it all to populate time |
| 4601 | */ |
| 4602 | return mem_cgroup_sockets_init(cont, ss); |
| 4603 | }; |
| 4604 | |
| 4605 | static void kmem_cgroup_destroy(struct cgroup *cont) |
| 4606 | { |
| 4607 | mem_cgroup_sockets_destroy(cont); |
| 4608 | } |
| 4609 | #else |
| 4610 | static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) |
| 4611 | { |
| 4612 | return 0; |
| 4613 | } |
| 4614 | |
| 4615 | static void kmem_cgroup_destroy(struct cgroup *cont) |
| 4616 | { |
| 4617 | } |
| 4618 | #endif |
| 4619 | |
| 4620 | static struct cftype mem_cgroup_files[] = { |
| 4621 | { |
| 4622 | .name = "usage_in_bytes", |
| 4623 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), |
| 4624 | .read_u64 = mem_cgroup_read, |
| 4625 | .register_event = mem_cgroup_usage_register_event, |
| 4626 | .unregister_event = mem_cgroup_usage_unregister_event, |
| 4627 | }, |
| 4628 | { |
| 4629 | .name = "max_usage_in_bytes", |
| 4630 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), |
| 4631 | .trigger = mem_cgroup_reset, |
| 4632 | .read_u64 = mem_cgroup_read, |
| 4633 | }, |
| 4634 | { |
| 4635 | .name = "limit_in_bytes", |
| 4636 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), |
| 4637 | .write_string = mem_cgroup_write, |
| 4638 | .read_u64 = mem_cgroup_read, |
| 4639 | }, |
| 4640 | { |
| 4641 | .name = "soft_limit_in_bytes", |
| 4642 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), |
| 4643 | .write_string = mem_cgroup_write, |
| 4644 | .read_u64 = mem_cgroup_read, |
| 4645 | }, |
| 4646 | { |
| 4647 | .name = "failcnt", |
| 4648 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), |
| 4649 | .trigger = mem_cgroup_reset, |
| 4650 | .read_u64 = mem_cgroup_read, |
| 4651 | }, |
| 4652 | { |
| 4653 | .name = "stat", |
| 4654 | .read_map = mem_control_stat_show, |
| 4655 | }, |
| 4656 | { |
| 4657 | .name = "force_empty", |
| 4658 | .trigger = mem_cgroup_force_empty_write, |
| 4659 | }, |
| 4660 | { |
| 4661 | .name = "use_hierarchy", |
| 4662 | .write_u64 = mem_cgroup_hierarchy_write, |
| 4663 | .read_u64 = mem_cgroup_hierarchy_read, |
| 4664 | }, |
| 4665 | { |
| 4666 | .name = "swappiness", |
| 4667 | .read_u64 = mem_cgroup_swappiness_read, |
| 4668 | .write_u64 = mem_cgroup_swappiness_write, |
| 4669 | }, |
| 4670 | { |
| 4671 | .name = "move_charge_at_immigrate", |
| 4672 | .read_u64 = mem_cgroup_move_charge_read, |
| 4673 | .write_u64 = mem_cgroup_move_charge_write, |
| 4674 | }, |
| 4675 | { |
| 4676 | .name = "oom_control", |
| 4677 | .read_map = mem_cgroup_oom_control_read, |
| 4678 | .write_u64 = mem_cgroup_oom_control_write, |
| 4679 | .register_event = mem_cgroup_oom_register_event, |
| 4680 | .unregister_event = mem_cgroup_oom_unregister_event, |
| 4681 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), |
| 4682 | }, |
| 4683 | #ifdef CONFIG_NUMA |
| 4684 | { |
| 4685 | .name = "numa_stat", |
| 4686 | .open = mem_control_numa_stat_open, |
| 4687 | .mode = S_IRUGO, |
| 4688 | }, |
| 4689 | #endif |
| 4690 | }; |
| 4691 | |
| 4692 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
| 4693 | static struct cftype memsw_cgroup_files[] = { |
| 4694 | { |
| 4695 | .name = "memsw.usage_in_bytes", |
| 4696 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), |
| 4697 | .read_u64 = mem_cgroup_read, |
| 4698 | .register_event = mem_cgroup_usage_register_event, |
| 4699 | .unregister_event = mem_cgroup_usage_unregister_event, |
| 4700 | }, |
| 4701 | { |
| 4702 | .name = "memsw.max_usage_in_bytes", |
| 4703 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), |
| 4704 | .trigger = mem_cgroup_reset, |
| 4705 | .read_u64 = mem_cgroup_read, |
| 4706 | }, |
| 4707 | { |
| 4708 | .name = "memsw.limit_in_bytes", |
| 4709 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), |
| 4710 | .write_string = mem_cgroup_write, |
| 4711 | .read_u64 = mem_cgroup_read, |
| 4712 | }, |
| 4713 | { |
| 4714 | .name = "memsw.failcnt", |
| 4715 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), |
| 4716 | .trigger = mem_cgroup_reset, |
| 4717 | .read_u64 = mem_cgroup_read, |
| 4718 | }, |
| 4719 | }; |
| 4720 | |
| 4721 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) |
| 4722 | { |
| 4723 | if (!do_swap_account) |
| 4724 | return 0; |
| 4725 | return cgroup_add_files(cont, ss, memsw_cgroup_files, |
| 4726 | ARRAY_SIZE(memsw_cgroup_files)); |
| 4727 | }; |
| 4728 | #else |
| 4729 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) |
| 4730 | { |
| 4731 | return 0; |
| 4732 | } |
| 4733 | #endif |
| 4734 | |
| 4735 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) |
| 4736 | { |
| 4737 | struct mem_cgroup_per_node *pn; |
| 4738 | struct mem_cgroup_per_zone *mz; |
| 4739 | enum lru_list lru; |
| 4740 | int zone, tmp = node; |
| 4741 | /* |
| 4742 | * This routine is called against possible nodes. |
| 4743 | * But it's BUG to call kmalloc() against offline node. |
| 4744 | * |
| 4745 | * TODO: this routine can waste much memory for nodes which will |
| 4746 | * never be onlined. It's better to use memory hotplug callback |
| 4747 | * function. |
| 4748 | */ |
| 4749 | if (!node_state(node, N_NORMAL_MEMORY)) |
| 4750 | tmp = -1; |
| 4751 | pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); |
| 4752 | if (!pn) |
| 4753 | return 1; |
| 4754 | |
| 4755 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
| 4756 | mz = &pn->zoneinfo[zone]; |
| 4757 | for_each_lru(lru) |
| 4758 | INIT_LIST_HEAD(&mz->lruvec.lists[lru]); |
| 4759 | mz->usage_in_excess = 0; |
| 4760 | mz->on_tree = false; |
| 4761 | mz->memcg = memcg; |
| 4762 | } |
| 4763 | memcg->info.nodeinfo[node] = pn; |
| 4764 | return 0; |
| 4765 | } |
| 4766 | |
| 4767 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) |
| 4768 | { |
| 4769 | kfree(memcg->info.nodeinfo[node]); |
| 4770 | } |
| 4771 | |
| 4772 | static struct mem_cgroup *mem_cgroup_alloc(void) |
| 4773 | { |
| 4774 | struct mem_cgroup *memcg; |
| 4775 | int size = sizeof(struct mem_cgroup); |
| 4776 | |
| 4777 | /* Can be very big if MAX_NUMNODES is very big */ |
| 4778 | if (size < PAGE_SIZE) |
| 4779 | memcg = kzalloc(size, GFP_KERNEL); |
| 4780 | else |
| 4781 | memcg = vzalloc(size); |
| 4782 | |
| 4783 | if (!memcg) |
| 4784 | return NULL; |
| 4785 | |
| 4786 | memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu); |
| 4787 | if (!memcg->stat) |
| 4788 | goto out_free; |
| 4789 | spin_lock_init(&memcg->pcp_counter_lock); |
| 4790 | return memcg; |
| 4791 | |
| 4792 | out_free: |
| 4793 | if (size < PAGE_SIZE) |
| 4794 | kfree(memcg); |
| 4795 | else |
| 4796 | vfree(memcg); |
| 4797 | return NULL; |
| 4798 | } |
| 4799 | |
| 4800 | /* |
| 4801 | * Helpers for freeing a vzalloc()ed mem_cgroup by RCU, |
| 4802 | * but in process context. The work_freeing structure is overlaid |
| 4803 | * on the rcu_freeing structure, which itself is overlaid on memsw. |
| 4804 | */ |
| 4805 | static void vfree_work(struct work_struct *work) |
| 4806 | { |
| 4807 | struct mem_cgroup *memcg; |
| 4808 | |
| 4809 | memcg = container_of(work, struct mem_cgroup, work_freeing); |
| 4810 | vfree(memcg); |
| 4811 | } |
| 4812 | static void vfree_rcu(struct rcu_head *rcu_head) |
| 4813 | { |
| 4814 | struct mem_cgroup *memcg; |
| 4815 | |
| 4816 | memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing); |
| 4817 | INIT_WORK(&memcg->work_freeing, vfree_work); |
| 4818 | schedule_work(&memcg->work_freeing); |
| 4819 | } |
| 4820 | |
| 4821 | /* |
| 4822 | * At destroying mem_cgroup, references from swap_cgroup can remain. |
| 4823 | * (scanning all at force_empty is too costly...) |
| 4824 | * |
| 4825 | * Instead of clearing all references at force_empty, we remember |
| 4826 | * the number of reference from swap_cgroup and free mem_cgroup when |
| 4827 | * it goes down to 0. |
| 4828 | * |
| 4829 | * Removal of cgroup itself succeeds regardless of refs from swap. |
| 4830 | */ |
| 4831 | |
| 4832 | static void __mem_cgroup_free(struct mem_cgroup *memcg) |
| 4833 | { |
| 4834 | int node; |
| 4835 | |
| 4836 | mem_cgroup_remove_from_trees(memcg); |
| 4837 | free_css_id(&mem_cgroup_subsys, &memcg->css); |
| 4838 | |
| 4839 | for_each_node(node) |
| 4840 | free_mem_cgroup_per_zone_info(memcg, node); |
| 4841 | |
| 4842 | free_percpu(memcg->stat); |
| 4843 | if (sizeof(struct mem_cgroup) < PAGE_SIZE) |
| 4844 | kfree_rcu(memcg, rcu_freeing); |
| 4845 | else |
| 4846 | call_rcu(&memcg->rcu_freeing, vfree_rcu); |
| 4847 | } |
| 4848 | |
| 4849 | static void mem_cgroup_get(struct mem_cgroup *memcg) |
| 4850 | { |
| 4851 | atomic_inc(&memcg->refcnt); |
| 4852 | } |
| 4853 | |
| 4854 | static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) |
| 4855 | { |
| 4856 | if (atomic_sub_and_test(count, &memcg->refcnt)) { |
| 4857 | struct mem_cgroup *parent = parent_mem_cgroup(memcg); |
| 4858 | __mem_cgroup_free(memcg); |
| 4859 | if (parent) |
| 4860 | mem_cgroup_put(parent); |
| 4861 | } |
| 4862 | } |
| 4863 | |
| 4864 | static void mem_cgroup_put(struct mem_cgroup *memcg) |
| 4865 | { |
| 4866 | __mem_cgroup_put(memcg, 1); |
| 4867 | } |
| 4868 | |
| 4869 | /* |
| 4870 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. |
| 4871 | */ |
| 4872 | struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) |
| 4873 | { |
| 4874 | if (!memcg->res.parent) |
| 4875 | return NULL; |
| 4876 | return mem_cgroup_from_res_counter(memcg->res.parent, res); |
| 4877 | } |
| 4878 | EXPORT_SYMBOL(parent_mem_cgroup); |
| 4879 | |
| 4880 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
| 4881 | static void __init enable_swap_cgroup(void) |
| 4882 | { |
| 4883 | if (!mem_cgroup_disabled() && really_do_swap_account) |
| 4884 | do_swap_account = 1; |
| 4885 | } |
| 4886 | #else |
| 4887 | static void __init enable_swap_cgroup(void) |
| 4888 | { |
| 4889 | } |
| 4890 | #endif |
| 4891 | |
| 4892 | static int mem_cgroup_soft_limit_tree_init(void) |
| 4893 | { |
| 4894 | struct mem_cgroup_tree_per_node *rtpn; |
| 4895 | struct mem_cgroup_tree_per_zone *rtpz; |
| 4896 | int tmp, node, zone; |
| 4897 | |
| 4898 | for_each_node(node) { |
| 4899 | tmp = node; |
| 4900 | if (!node_state(node, N_NORMAL_MEMORY)) |
| 4901 | tmp = -1; |
| 4902 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); |
| 4903 | if (!rtpn) |
| 4904 | goto err_cleanup; |
| 4905 | |
| 4906 | soft_limit_tree.rb_tree_per_node[node] = rtpn; |
| 4907 | |
| 4908 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
| 4909 | rtpz = &rtpn->rb_tree_per_zone[zone]; |
| 4910 | rtpz->rb_root = RB_ROOT; |
| 4911 | spin_lock_init(&rtpz->lock); |
| 4912 | } |
| 4913 | } |
| 4914 | return 0; |
| 4915 | |
| 4916 | err_cleanup: |
| 4917 | for_each_node(node) { |
| 4918 | if (!soft_limit_tree.rb_tree_per_node[node]) |
| 4919 | break; |
| 4920 | kfree(soft_limit_tree.rb_tree_per_node[node]); |
| 4921 | soft_limit_tree.rb_tree_per_node[node] = NULL; |
| 4922 | } |
| 4923 | return 1; |
| 4924 | |
| 4925 | } |
| 4926 | |
| 4927 | static struct cgroup_subsys_state * __ref |
| 4928 | mem_cgroup_create(struct cgroup *cont) |
| 4929 | { |
| 4930 | struct mem_cgroup *memcg, *parent; |
| 4931 | long error = -ENOMEM; |
| 4932 | int node; |
| 4933 | |
| 4934 | memcg = mem_cgroup_alloc(); |
| 4935 | if (!memcg) |
| 4936 | return ERR_PTR(error); |
| 4937 | |
| 4938 | for_each_node(node) |
| 4939 | if (alloc_mem_cgroup_per_zone_info(memcg, node)) |
| 4940 | goto free_out; |
| 4941 | |
| 4942 | /* root ? */ |
| 4943 | if (cont->parent == NULL) { |
| 4944 | int cpu; |
| 4945 | enable_swap_cgroup(); |
| 4946 | parent = NULL; |
| 4947 | if (mem_cgroup_soft_limit_tree_init()) |
| 4948 | goto free_out; |
| 4949 | root_mem_cgroup = memcg; |
| 4950 | for_each_possible_cpu(cpu) { |
| 4951 | struct memcg_stock_pcp *stock = |
| 4952 | &per_cpu(memcg_stock, cpu); |
| 4953 | INIT_WORK(&stock->work, drain_local_stock); |
| 4954 | } |
| 4955 | hotcpu_notifier(memcg_cpu_hotplug_callback, 0); |
| 4956 | } else { |
| 4957 | parent = mem_cgroup_from_cont(cont->parent); |
| 4958 | memcg->use_hierarchy = parent->use_hierarchy; |
| 4959 | memcg->oom_kill_disable = parent->oom_kill_disable; |
| 4960 | } |
| 4961 | |
| 4962 | if (parent && parent->use_hierarchy) { |
| 4963 | res_counter_init(&memcg->res, &parent->res); |
| 4964 | res_counter_init(&memcg->memsw, &parent->memsw); |
| 4965 | /* |
| 4966 | * We increment refcnt of the parent to ensure that we can |
| 4967 | * safely access it on res_counter_charge/uncharge. |
| 4968 | * This refcnt will be decremented when freeing this |
| 4969 | * mem_cgroup(see mem_cgroup_put). |
| 4970 | */ |
| 4971 | mem_cgroup_get(parent); |
| 4972 | } else { |
| 4973 | res_counter_init(&memcg->res, NULL); |
| 4974 | res_counter_init(&memcg->memsw, NULL); |
| 4975 | } |
| 4976 | memcg->last_scanned_node = MAX_NUMNODES; |
| 4977 | INIT_LIST_HEAD(&memcg->oom_notify); |
| 4978 | |
| 4979 | if (parent) |
| 4980 | memcg->swappiness = mem_cgroup_swappiness(parent); |
| 4981 | atomic_set(&memcg->refcnt, 1); |
| 4982 | memcg->move_charge_at_immigrate = 0; |
| 4983 | mutex_init(&memcg->thresholds_lock); |
| 4984 | return &memcg->css; |
| 4985 | free_out: |
| 4986 | __mem_cgroup_free(memcg); |
| 4987 | return ERR_PTR(error); |
| 4988 | } |
| 4989 | |
| 4990 | static int mem_cgroup_pre_destroy(struct cgroup *cont) |
| 4991 | { |
| 4992 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 4993 | |
| 4994 | return mem_cgroup_force_empty(memcg, false); |
| 4995 | } |
| 4996 | |
| 4997 | static void mem_cgroup_destroy(struct cgroup *cont) |
| 4998 | { |
| 4999 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
| 5000 | |
| 5001 | kmem_cgroup_destroy(cont); |
| 5002 | |
| 5003 | mem_cgroup_put(memcg); |
| 5004 | } |
| 5005 | |
| 5006 | static int mem_cgroup_populate(struct cgroup_subsys *ss, |
| 5007 | struct cgroup *cont) |
| 5008 | { |
| 5009 | int ret; |
| 5010 | |
| 5011 | ret = cgroup_add_files(cont, ss, mem_cgroup_files, |
| 5012 | ARRAY_SIZE(mem_cgroup_files)); |
| 5013 | |
| 5014 | if (!ret) |
| 5015 | ret = register_memsw_files(cont, ss); |
| 5016 | |
| 5017 | if (!ret) |
| 5018 | ret = register_kmem_files(cont, ss); |
| 5019 | |
| 5020 | return ret; |
| 5021 | } |
| 5022 | |
| 5023 | #ifdef CONFIG_MMU |
| 5024 | /* Handlers for move charge at task migration. */ |
| 5025 | #define PRECHARGE_COUNT_AT_ONCE 256 |
| 5026 | static int mem_cgroup_do_precharge(unsigned long count) |
| 5027 | { |
| 5028 | int ret = 0; |
| 5029 | int batch_count = PRECHARGE_COUNT_AT_ONCE; |
| 5030 | struct mem_cgroup *memcg = mc.to; |
| 5031 | |
| 5032 | if (mem_cgroup_is_root(memcg)) { |
| 5033 | mc.precharge += count; |
| 5034 | /* we don't need css_get for root */ |
| 5035 | return ret; |
| 5036 | } |
| 5037 | /* try to charge at once */ |
| 5038 | if (count > 1) { |
| 5039 | struct res_counter *dummy; |
| 5040 | /* |
| 5041 | * "memcg" cannot be under rmdir() because we've already checked |
| 5042 | * by cgroup_lock_live_cgroup() that it is not removed and we |
| 5043 | * are still under the same cgroup_mutex. So we can postpone |
| 5044 | * css_get(). |
| 5045 | */ |
| 5046 | if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) |
| 5047 | goto one_by_one; |
| 5048 | if (do_swap_account && res_counter_charge(&memcg->memsw, |
| 5049 | PAGE_SIZE * count, &dummy)) { |
| 5050 | res_counter_uncharge(&memcg->res, PAGE_SIZE * count); |
| 5051 | goto one_by_one; |
| 5052 | } |
| 5053 | mc.precharge += count; |
| 5054 | return ret; |
| 5055 | } |
| 5056 | one_by_one: |
| 5057 | /* fall back to one by one charge */ |
| 5058 | while (count--) { |
| 5059 | if (signal_pending(current)) { |
| 5060 | ret = -EINTR; |
| 5061 | break; |
| 5062 | } |
| 5063 | if (!batch_count--) { |
| 5064 | batch_count = PRECHARGE_COUNT_AT_ONCE; |
| 5065 | cond_resched(); |
| 5066 | } |
| 5067 | ret = __mem_cgroup_try_charge(NULL, |
| 5068 | GFP_KERNEL, 1, &memcg, false); |
| 5069 | if (ret) |
| 5070 | /* mem_cgroup_clear_mc() will do uncharge later */ |
| 5071 | return ret; |
| 5072 | mc.precharge++; |
| 5073 | } |
| 5074 | return ret; |
| 5075 | } |
| 5076 | |
| 5077 | /** |
| 5078 | * is_target_pte_for_mc - check a pte whether it is valid for move charge |
| 5079 | * @vma: the vma the pte to be checked belongs |
| 5080 | * @addr: the address corresponding to the pte to be checked |
| 5081 | * @ptent: the pte to be checked |
| 5082 | * @target: the pointer the target page or swap ent will be stored(can be NULL) |
| 5083 | * |
| 5084 | * Returns |
| 5085 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. |
| 5086 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for |
| 5087 | * move charge. if @target is not NULL, the page is stored in target->page |
| 5088 | * with extra refcnt got(Callers should handle it). |
| 5089 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a |
| 5090 | * target for charge migration. if @target is not NULL, the entry is stored |
| 5091 | * in target->ent. |
| 5092 | * |
| 5093 | * Called with pte lock held. |
| 5094 | */ |
| 5095 | union mc_target { |
| 5096 | struct page *page; |
| 5097 | swp_entry_t ent; |
| 5098 | }; |
| 5099 | |
| 5100 | enum mc_target_type { |
| 5101 | MC_TARGET_NONE, /* not used */ |
| 5102 | MC_TARGET_PAGE, |
| 5103 | MC_TARGET_SWAP, |
| 5104 | }; |
| 5105 | |
| 5106 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, |
| 5107 | unsigned long addr, pte_t ptent) |
| 5108 | { |
| 5109 | struct page *page = vm_normal_page(vma, addr, ptent); |
| 5110 | |
| 5111 | if (!page || !page_mapped(page)) |
| 5112 | return NULL; |
| 5113 | if (PageAnon(page)) { |
| 5114 | /* we don't move shared anon */ |
| 5115 | if (!move_anon() || page_mapcount(page) > 2) |
| 5116 | return NULL; |
| 5117 | } else if (!move_file()) |
| 5118 | /* we ignore mapcount for file pages */ |
| 5119 | return NULL; |
| 5120 | if (!get_page_unless_zero(page)) |
| 5121 | return NULL; |
| 5122 | |
| 5123 | return page; |
| 5124 | } |
| 5125 | |
| 5126 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, |
| 5127 | unsigned long addr, pte_t ptent, swp_entry_t *entry) |
| 5128 | { |
| 5129 | int usage_count; |
| 5130 | struct page *page = NULL; |
| 5131 | swp_entry_t ent = pte_to_swp_entry(ptent); |
| 5132 | |
| 5133 | if (!move_anon() || non_swap_entry(ent)) |
| 5134 | return NULL; |
| 5135 | usage_count = mem_cgroup_count_swap_user(ent, &page); |
| 5136 | if (usage_count > 1) { /* we don't move shared anon */ |
| 5137 | if (page) |
| 5138 | put_page(page); |
| 5139 | return NULL; |
| 5140 | } |
| 5141 | if (do_swap_account) |
| 5142 | entry->val = ent.val; |
| 5143 | |
| 5144 | return page; |
| 5145 | } |
| 5146 | |
| 5147 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, |
| 5148 | unsigned long addr, pte_t ptent, swp_entry_t *entry) |
| 5149 | { |
| 5150 | struct page *page = NULL; |
| 5151 | struct inode *inode; |
| 5152 | struct address_space *mapping; |
| 5153 | pgoff_t pgoff; |
| 5154 | |
| 5155 | if (!vma->vm_file) /* anonymous vma */ |
| 5156 | return NULL; |
| 5157 | if (!move_file()) |
| 5158 | return NULL; |
| 5159 | |
| 5160 | inode = vma->vm_file->f_path.dentry->d_inode; |
| 5161 | mapping = vma->vm_file->f_mapping; |
| 5162 | if (pte_none(ptent)) |
| 5163 | pgoff = linear_page_index(vma, addr); |
| 5164 | else /* pte_file(ptent) is true */ |
| 5165 | pgoff = pte_to_pgoff(ptent); |
| 5166 | |
| 5167 | /* page is moved even if it's not RSS of this task(page-faulted). */ |
| 5168 | page = find_get_page(mapping, pgoff); |
| 5169 | |
| 5170 | #ifdef CONFIG_SWAP |
| 5171 | /* shmem/tmpfs may report page out on swap: account for that too. */ |
| 5172 | if (radix_tree_exceptional_entry(page)) { |
| 5173 | swp_entry_t swap = radix_to_swp_entry(page); |
| 5174 | if (do_swap_account) |
| 5175 | *entry = swap; |
| 5176 | page = find_get_page(&swapper_space, swap.val); |
| 5177 | } |
| 5178 | #endif |
| 5179 | return page; |
| 5180 | } |
| 5181 | |
| 5182 | static int is_target_pte_for_mc(struct vm_area_struct *vma, |
| 5183 | unsigned long addr, pte_t ptent, union mc_target *target) |
| 5184 | { |
| 5185 | struct page *page = NULL; |
| 5186 | struct page_cgroup *pc; |
| 5187 | int ret = 0; |
| 5188 | swp_entry_t ent = { .val = 0 }; |
| 5189 | |
| 5190 | if (pte_present(ptent)) |
| 5191 | page = mc_handle_present_pte(vma, addr, ptent); |
| 5192 | else if (is_swap_pte(ptent)) |
| 5193 | page = mc_handle_swap_pte(vma, addr, ptent, &ent); |
| 5194 | else if (pte_none(ptent) || pte_file(ptent)) |
| 5195 | page = mc_handle_file_pte(vma, addr, ptent, &ent); |
| 5196 | |
| 5197 | if (!page && !ent.val) |
| 5198 | return 0; |
| 5199 | if (page) { |
| 5200 | pc = lookup_page_cgroup(page); |
| 5201 | /* |
| 5202 | * Do only loose check w/o page_cgroup lock. |
| 5203 | * mem_cgroup_move_account() checks the pc is valid or not under |
| 5204 | * the lock. |
| 5205 | */ |
| 5206 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { |
| 5207 | ret = MC_TARGET_PAGE; |
| 5208 | if (target) |
| 5209 | target->page = page; |
| 5210 | } |
| 5211 | if (!ret || !target) |
| 5212 | put_page(page); |
| 5213 | } |
| 5214 | /* There is a swap entry and a page doesn't exist or isn't charged */ |
| 5215 | if (ent.val && !ret && |
| 5216 | css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) { |
| 5217 | ret = MC_TARGET_SWAP; |
| 5218 | if (target) |
| 5219 | target->ent = ent; |
| 5220 | } |
| 5221 | return ret; |
| 5222 | } |
| 5223 | |
| 5224 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, |
| 5225 | unsigned long addr, unsigned long end, |
| 5226 | struct mm_walk *walk) |
| 5227 | { |
| 5228 | struct vm_area_struct *vma = walk->private; |
| 5229 | pte_t *pte; |
| 5230 | spinlock_t *ptl; |
| 5231 | |
| 5232 | split_huge_page_pmd(walk->mm, pmd); |
| 5233 | if (pmd_trans_unstable(pmd)) |
| 5234 | return 0; |
| 5235 | |
| 5236 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
| 5237 | for (; addr != end; pte++, addr += PAGE_SIZE) |
| 5238 | if (is_target_pte_for_mc(vma, addr, *pte, NULL)) |
| 5239 | mc.precharge++; /* increment precharge temporarily */ |
| 5240 | pte_unmap_unlock(pte - 1, ptl); |
| 5241 | cond_resched(); |
| 5242 | |
| 5243 | return 0; |
| 5244 | } |
| 5245 | |
| 5246 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) |
| 5247 | { |
| 5248 | unsigned long precharge; |
| 5249 | struct vm_area_struct *vma; |
| 5250 | |
| 5251 | down_read(&mm->mmap_sem); |
| 5252 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
| 5253 | struct mm_walk mem_cgroup_count_precharge_walk = { |
| 5254 | .pmd_entry = mem_cgroup_count_precharge_pte_range, |
| 5255 | .mm = mm, |
| 5256 | .private = vma, |
| 5257 | }; |
| 5258 | if (is_vm_hugetlb_page(vma)) |
| 5259 | continue; |
| 5260 | walk_page_range(vma->vm_start, vma->vm_end, |
| 5261 | &mem_cgroup_count_precharge_walk); |
| 5262 | } |
| 5263 | up_read(&mm->mmap_sem); |
| 5264 | |
| 5265 | precharge = mc.precharge; |
| 5266 | mc.precharge = 0; |
| 5267 | |
| 5268 | return precharge; |
| 5269 | } |
| 5270 | |
| 5271 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) |
| 5272 | { |
| 5273 | unsigned long precharge = mem_cgroup_count_precharge(mm); |
| 5274 | |
| 5275 | VM_BUG_ON(mc.moving_task); |
| 5276 | mc.moving_task = current; |
| 5277 | return mem_cgroup_do_precharge(precharge); |
| 5278 | } |
| 5279 | |
| 5280 | /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ |
| 5281 | static void __mem_cgroup_clear_mc(void) |
| 5282 | { |
| 5283 | struct mem_cgroup *from = mc.from; |
| 5284 | struct mem_cgroup *to = mc.to; |
| 5285 | |
| 5286 | /* we must uncharge all the leftover precharges from mc.to */ |
| 5287 | if (mc.precharge) { |
| 5288 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); |
| 5289 | mc.precharge = 0; |
| 5290 | } |
| 5291 | /* |
| 5292 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so |
| 5293 | * we must uncharge here. |
| 5294 | */ |
| 5295 | if (mc.moved_charge) { |
| 5296 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); |
| 5297 | mc.moved_charge = 0; |
| 5298 | } |
| 5299 | /* we must fixup refcnts and charges */ |
| 5300 | if (mc.moved_swap) { |
| 5301 | /* uncharge swap account from the old cgroup */ |
| 5302 | if (!mem_cgroup_is_root(mc.from)) |
| 5303 | res_counter_uncharge(&mc.from->memsw, |
| 5304 | PAGE_SIZE * mc.moved_swap); |
| 5305 | __mem_cgroup_put(mc.from, mc.moved_swap); |
| 5306 | |
| 5307 | if (!mem_cgroup_is_root(mc.to)) { |
| 5308 | /* |
| 5309 | * we charged both to->res and to->memsw, so we should |
| 5310 | * uncharge to->res. |
| 5311 | */ |
| 5312 | res_counter_uncharge(&mc.to->res, |
| 5313 | PAGE_SIZE * mc.moved_swap); |
| 5314 | } |
| 5315 | /* we've already done mem_cgroup_get(mc.to) */ |
| 5316 | mc.moved_swap = 0; |
| 5317 | } |
| 5318 | memcg_oom_recover(from); |
| 5319 | memcg_oom_recover(to); |
| 5320 | wake_up_all(&mc.waitq); |
| 5321 | } |
| 5322 | |
| 5323 | static void mem_cgroup_clear_mc(void) |
| 5324 | { |
| 5325 | struct mem_cgroup *from = mc.from; |
| 5326 | |
| 5327 | /* |
| 5328 | * we must clear moving_task before waking up waiters at the end of |
| 5329 | * task migration. |
| 5330 | */ |
| 5331 | mc.moving_task = NULL; |
| 5332 | __mem_cgroup_clear_mc(); |
| 5333 | spin_lock(&mc.lock); |
| 5334 | mc.from = NULL; |
| 5335 | mc.to = NULL; |
| 5336 | spin_unlock(&mc.lock); |
| 5337 | mem_cgroup_end_move(from); |
| 5338 | } |
| 5339 | |
| 5340 | static int mem_cgroup_can_attach(struct cgroup *cgroup, |
| 5341 | struct cgroup_taskset *tset) |
| 5342 | { |
| 5343 | struct task_struct *p = cgroup_taskset_first(tset); |
| 5344 | int ret = 0; |
| 5345 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); |
| 5346 | |
| 5347 | if (memcg->move_charge_at_immigrate) { |
| 5348 | struct mm_struct *mm; |
| 5349 | struct mem_cgroup *from = mem_cgroup_from_task(p); |
| 5350 | |
| 5351 | VM_BUG_ON(from == memcg); |
| 5352 | |
| 5353 | mm = get_task_mm(p); |
| 5354 | if (!mm) |
| 5355 | return 0; |
| 5356 | /* We move charges only when we move a owner of the mm */ |
| 5357 | if (mm->owner == p) { |
| 5358 | VM_BUG_ON(mc.from); |
| 5359 | VM_BUG_ON(mc.to); |
| 5360 | VM_BUG_ON(mc.precharge); |
| 5361 | VM_BUG_ON(mc.moved_charge); |
| 5362 | VM_BUG_ON(mc.moved_swap); |
| 5363 | mem_cgroup_start_move(from); |
| 5364 | spin_lock(&mc.lock); |
| 5365 | mc.from = from; |
| 5366 | mc.to = memcg; |
| 5367 | spin_unlock(&mc.lock); |
| 5368 | /* We set mc.moving_task later */ |
| 5369 | |
| 5370 | ret = mem_cgroup_precharge_mc(mm); |
| 5371 | if (ret) |
| 5372 | mem_cgroup_clear_mc(); |
| 5373 | } |
| 5374 | mmput(mm); |
| 5375 | } |
| 5376 | return ret; |
| 5377 | } |
| 5378 | |
| 5379 | static void mem_cgroup_cancel_attach(struct cgroup *cgroup, |
| 5380 | struct cgroup_taskset *tset) |
| 5381 | { |
| 5382 | mem_cgroup_clear_mc(); |
| 5383 | } |
| 5384 | |
| 5385 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, |
| 5386 | unsigned long addr, unsigned long end, |
| 5387 | struct mm_walk *walk) |
| 5388 | { |
| 5389 | int ret = 0; |
| 5390 | struct vm_area_struct *vma = walk->private; |
| 5391 | pte_t *pte; |
| 5392 | spinlock_t *ptl; |
| 5393 | |
| 5394 | split_huge_page_pmd(walk->mm, pmd); |
| 5395 | if (pmd_trans_unstable(pmd)) |
| 5396 | return 0; |
| 5397 | retry: |
| 5398 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
| 5399 | for (; addr != end; addr += PAGE_SIZE) { |
| 5400 | pte_t ptent = *(pte++); |
| 5401 | union mc_target target; |
| 5402 | int type; |
| 5403 | struct page *page; |
| 5404 | struct page_cgroup *pc; |
| 5405 | swp_entry_t ent; |
| 5406 | |
| 5407 | if (!mc.precharge) |
| 5408 | break; |
| 5409 | |
| 5410 | type = is_target_pte_for_mc(vma, addr, ptent, &target); |
| 5411 | switch (type) { |
| 5412 | case MC_TARGET_PAGE: |
| 5413 | page = target.page; |
| 5414 | if (isolate_lru_page(page)) |
| 5415 | goto put; |
| 5416 | pc = lookup_page_cgroup(page); |
| 5417 | if (!mem_cgroup_move_account(page, 1, pc, |
| 5418 | mc.from, mc.to, false)) { |
| 5419 | mc.precharge--; |
| 5420 | /* we uncharge from mc.from later. */ |
| 5421 | mc.moved_charge++; |
| 5422 | } |
| 5423 | putback_lru_page(page); |
| 5424 | put: /* is_target_pte_for_mc() gets the page */ |
| 5425 | put_page(page); |
| 5426 | break; |
| 5427 | case MC_TARGET_SWAP: |
| 5428 | ent = target.ent; |
| 5429 | if (!mem_cgroup_move_swap_account(ent, |
| 5430 | mc.from, mc.to, false)) { |
| 5431 | mc.precharge--; |
| 5432 | /* we fixup refcnts and charges later. */ |
| 5433 | mc.moved_swap++; |
| 5434 | } |
| 5435 | break; |
| 5436 | default: |
| 5437 | break; |
| 5438 | } |
| 5439 | } |
| 5440 | pte_unmap_unlock(pte - 1, ptl); |
| 5441 | cond_resched(); |
| 5442 | |
| 5443 | if (addr != end) { |
| 5444 | /* |
| 5445 | * We have consumed all precharges we got in can_attach(). |
| 5446 | * We try charge one by one, but don't do any additional |
| 5447 | * charges to mc.to if we have failed in charge once in attach() |
| 5448 | * phase. |
| 5449 | */ |
| 5450 | ret = mem_cgroup_do_precharge(1); |
| 5451 | if (!ret) |
| 5452 | goto retry; |
| 5453 | } |
| 5454 | |
| 5455 | return ret; |
| 5456 | } |
| 5457 | |
| 5458 | static void mem_cgroup_move_charge(struct mm_struct *mm) |
| 5459 | { |
| 5460 | struct vm_area_struct *vma; |
| 5461 | |
| 5462 | lru_add_drain_all(); |
| 5463 | retry: |
| 5464 | if (unlikely(!down_read_trylock(&mm->mmap_sem))) { |
| 5465 | /* |
| 5466 | * Someone who are holding the mmap_sem might be waiting in |
| 5467 | * waitq. So we cancel all extra charges, wake up all waiters, |
| 5468 | * and retry. Because we cancel precharges, we might not be able |
| 5469 | * to move enough charges, but moving charge is a best-effort |
| 5470 | * feature anyway, so it wouldn't be a big problem. |
| 5471 | */ |
| 5472 | __mem_cgroup_clear_mc(); |
| 5473 | cond_resched(); |
| 5474 | goto retry; |
| 5475 | } |
| 5476 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
| 5477 | int ret; |
| 5478 | struct mm_walk mem_cgroup_move_charge_walk = { |
| 5479 | .pmd_entry = mem_cgroup_move_charge_pte_range, |
| 5480 | .mm = mm, |
| 5481 | .private = vma, |
| 5482 | }; |
| 5483 | if (is_vm_hugetlb_page(vma)) |
| 5484 | continue; |
| 5485 | ret = walk_page_range(vma->vm_start, vma->vm_end, |
| 5486 | &mem_cgroup_move_charge_walk); |
| 5487 | if (ret) |
| 5488 | /* |
| 5489 | * means we have consumed all precharges and failed in |
| 5490 | * doing additional charge. Just abandon here. |
| 5491 | */ |
| 5492 | break; |
| 5493 | } |
| 5494 | up_read(&mm->mmap_sem); |
| 5495 | } |
| 5496 | |
| 5497 | static void mem_cgroup_move_task(struct cgroup *cont, |
| 5498 | struct cgroup_taskset *tset) |
| 5499 | { |
| 5500 | struct task_struct *p = cgroup_taskset_first(tset); |
| 5501 | struct mm_struct *mm = get_task_mm(p); |
| 5502 | |
| 5503 | if (mm) { |
| 5504 | if (mc.to) |
| 5505 | mem_cgroup_move_charge(mm); |
| 5506 | put_swap_token(mm); |
| 5507 | mmput(mm); |
| 5508 | } |
| 5509 | if (mc.to) |
| 5510 | mem_cgroup_clear_mc(); |
| 5511 | } |
| 5512 | #else /* !CONFIG_MMU */ |
| 5513 | static int mem_cgroup_can_attach(struct cgroup *cgroup, |
| 5514 | struct cgroup_taskset *tset) |
| 5515 | { |
| 5516 | return 0; |
| 5517 | } |
| 5518 | static void mem_cgroup_cancel_attach(struct cgroup *cgroup, |
| 5519 | struct cgroup_taskset *tset) |
| 5520 | { |
| 5521 | } |
| 5522 | static void mem_cgroup_move_task(struct cgroup *cont, |
| 5523 | struct cgroup_taskset *tset) |
| 5524 | { |
| 5525 | } |
| 5526 | #endif |
| 5527 | |
| 5528 | struct cgroup_subsys mem_cgroup_subsys = { |
| 5529 | .name = "memory", |
| 5530 | .subsys_id = mem_cgroup_subsys_id, |
| 5531 | .create = mem_cgroup_create, |
| 5532 | .pre_destroy = mem_cgroup_pre_destroy, |
| 5533 | .destroy = mem_cgroup_destroy, |
| 5534 | .populate = mem_cgroup_populate, |
| 5535 | .can_attach = mem_cgroup_can_attach, |
| 5536 | .cancel_attach = mem_cgroup_cancel_attach, |
| 5537 | .attach = mem_cgroup_move_task, |
| 5538 | .early_init = 0, |
| 5539 | .use_id = 1, |
| 5540 | }; |
| 5541 | |
| 5542 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
| 5543 | static int __init enable_swap_account(char *s) |
| 5544 | { |
| 5545 | /* consider enabled if no parameter or 1 is given */ |
| 5546 | if (!strcmp(s, "1")) |
| 5547 | really_do_swap_account = 1; |
| 5548 | else if (!strcmp(s, "0")) |
| 5549 | really_do_swap_account = 0; |
| 5550 | return 1; |
| 5551 | } |
| 5552 | __setup("swapaccount=", enable_swap_account); |
| 5553 | |
| 5554 | #endif |