mm/list_lru.c

   1 /*
   2  * Copyright (c) 2013 Red Hat, Inc. and Parallels Inc. All rights reserved.
   3  * Authors: David Chinner and Glauber Costa
   4  *
   5  * Generic LRU infrastructure
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/list_lru.h>
  11 #include <linux/slab.h>
  12 #include <linux/mutex.h>
  13 #include <linux/memcontrol.h>
  14
  15 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
  16 static LIST_HEAD(list_lrus);
  17 static DEFINE_MUTEX(list_lrus_mutex);
  18
  19 static void list_lru_register(struct list_lru *lru)
  20 {
  21         mutex_lock(&list_lrus_mutex);
  22         list_add(&lru->list, &list_lrus);
  23         mutex_unlock(&list_lrus_mutex);
  24 }
  25
  26 static void list_lru_unregister(struct list_lru *lru)
  27 {
  28         mutex_lock(&list_lrus_mutex);
  29         list_del(&lru->list);
  30         mutex_unlock(&list_lrus_mutex);
  31 }
  32 #else
  33 static void list_lru_register(struct list_lru *lru)
  34 {
  35 }
  36
  37 static void list_lru_unregister(struct list_lru *lru)
  38 {
  39 }
  40 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
  41
  42 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
  43 static inline bool list_lru_memcg_aware(struct list_lru *lru)
  44 {
  45         /*
  46          * This needs node 0 to be always present, even
  47          * in the systems supporting sparse numa ids.
  48          */
  49         return !!lru->node[0].memcg_lrus;
  50 }
  51
  52 static inline struct list_lru_one *
  53 list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
  54 {
  55         /*
  56          * The lock protects the array of per cgroup lists from relocation
  57          * (see memcg_update_list_lru_node).
  58          */
  59         lockdep_assert_held(&nlru->lock);
  60         if (nlru->memcg_lrus && idx >= 0)
  61                 return nlru->memcg_lrus->lru[idx];
  62
  63         return &nlru->lru;
  64 }
  65
  66 static __always_inline struct mem_cgroup *mem_cgroup_from_kmem(void *ptr)
  67 {
  68         struct page *page;
  69
  70         if (!memcg_kmem_enabled())
  71                 return NULL;
  72         page = virt_to_head_page(ptr);
  73         return page->mem_cgroup;
  74 }
  75
  76 static inline struct list_lru_one *
  77 list_lru_from_kmem(struct list_lru_node *nlru, void *ptr)
  78 {
  79         struct mem_cgroup *memcg;
  80
  81         if (!nlru->memcg_lrus)
  82                 return &nlru->lru;
  83
  84         memcg = mem_cgroup_from_kmem(ptr);
  85         if (!memcg)
  86                 return &nlru->lru;
  87
  88         return list_lru_from_memcg_idx(nlru, memcg_cache_id(memcg));
  89 }
  90 #else
  91 static inline bool list_lru_memcg_aware(struct list_lru *lru)
  92 {
  93         return false;
  94 }
  95
  96 static inline struct list_lru_one *
  97 list_lru_from_memcg_idx(struct list_lru_node *nlru, int idx)
  98 {
  99         return &nlru->lru;
 100 }
 101
 102 static inline struct list_lru_one *
 103 list_lru_from_kmem(struct list_lru_node *nlru, void *ptr)
 104 {
 105         return &nlru->lru;
 106 }
 107 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
 108
 109 bool list_lru_add(struct list_lru *lru, struct list_head *item)
 110 {
 111         int nid = page_to_nid(virt_to_page(item));
 112         struct list_lru_node *nlru = &lru->node[nid];
 113         struct list_lru_one *l;
 114
 115         spin_lock(&nlru->lock);
 116         if (list_empty(item)) {
 117                 l = list_lru_from_kmem(nlru, item);
 118                 list_add_tail(item, &l->list);
 119                 l->nr_items++;
 120                 nlru->nr_items++;
 121                 spin_unlock(&nlru->lock);
 122                 return true;
 123         }
 124         spin_unlock(&nlru->lock);
 125         return false;
 126 }
 127 EXPORT_SYMBOL_GPL(list_lru_add);
 128
 129 bool list_lru_del(struct list_lru *lru, struct list_head *item)
 130 {
 131         int nid = page_to_nid(virt_to_page(item));
 132         struct list_lru_node *nlru = &lru->node[nid];
 133         struct list_lru_one *l;
 134
 135         spin_lock(&nlru->lock);
 136         if (!list_empty(item)) {
 137                 l = list_lru_from_kmem(nlru, item);
 138                 list_del_init(item);
 139                 l->nr_items--;
 140                 nlru->nr_items--;
 141                 spin_unlock(&nlru->lock);
 142                 return true;
 143         }
 144         spin_unlock(&nlru->lock);
 145         return false;
 146 }
 147 EXPORT_SYMBOL_GPL(list_lru_del);
 148
 149 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
 150 {
 151         list_del_init(item);
 152         list->nr_items--;
 153 }
 154 EXPORT_SYMBOL_GPL(list_lru_isolate);
 155
 156 void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
 157                            struct list_head *head)
 158 {
 159         list_move(item, head);
 160         list->nr_items--;
 161 }
 162 EXPORT_SYMBOL_GPL(list_lru_isolate_move);
 163
 164 static unsigned long __list_lru_count_one(struct list_lru *lru,
 165                                           int nid, int memcg_idx)
 166 {
 167         struct list_lru_node *nlru = &lru->node[nid];
 168         struct list_lru_one *l;
 169         unsigned long count;
 170
 171         spin_lock(&nlru->lock);
 172         l = list_lru_from_memcg_idx(nlru, memcg_idx);
 173         count = l->nr_items;
 174         spin_unlock(&nlru->lock);
 175
 176         return count;
 177 }
 178
 179 unsigned long list_lru_count_one(struct list_lru *lru,
 180                                  int nid, struct mem_cgroup *memcg)
 181 {
 182         return __list_lru_count_one(lru, nid, memcg_cache_id(memcg));
 183 }
 184 EXPORT_SYMBOL_GPL(list_lru_count_one);
 185
 186 unsigned long list_lru_count_node(struct list_lru *lru, int nid)
 187 {
 188         struct list_lru_node *nlru;
 189
 190         nlru = &lru->node[nid];
 191         return nlru->nr_items;
 192 }
 193 EXPORT_SYMBOL_GPL(list_lru_count_node);
 194
 195 static unsigned long
 196 __list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
 197                     list_lru_walk_cb isolate, void *cb_arg,
 198                     unsigned long *nr_to_walk)
 199 {
 200
 201         struct list_lru_node *nlru = &lru->node[nid];
 202         struct list_lru_one *l;
 203         struct list_head *item, *n;
 204         unsigned long isolated = 0;
 205
 206         spin_lock(&nlru->lock);
 207         l = list_lru_from_memcg_idx(nlru, memcg_idx);
 208 restart:
 209         list_for_each_safe(item, n, &l->list) {
 210                 enum lru_status ret;
 211
 212                 /*
 213                  * decrement nr_to_walk first so that we don't livelock if we
 214                  * get stuck on large numbesr of LRU_RETRY items
 215                  */
 216                 if (!*nr_to_walk)
 217                         break;
 218                 --*nr_to_walk;
 219
 220                 ret = isolate(item, l, &nlru->lock, cb_arg);
 221                 switch (ret) {
 222                 case LRU_REMOVED_RETRY:
 223                         assert_spin_locked(&nlru->lock);
 224                         /* fall through */
 225                 case LRU_REMOVED:
 226                         isolated++;
 227                         nlru->nr_items--;
 228                         /*
 229                          * If the lru lock has been dropped, our list
 230                          * traversal is now invalid and so we have to
 231                          * restart from scratch.
 232                          */
 233                         if (ret == LRU_REMOVED_RETRY)
 234                                 goto restart;
 235                         break;
 236                 case LRU_ROTATE:
 237                         list_move_tail(item, &l->list);
 238                         break;
 239                 case LRU_SKIP:
 240                         break;
 241                 case LRU_RETRY:
 242                         /*
 243                          * The lru lock has been dropped, our list traversal is
 244                          * now invalid and so we have to restart from scratch.
 245                          */
 246                         assert_spin_locked(&nlru->lock);
 247                         goto restart;
 248                 default:
 249                         BUG();
 250                 }
 251         }
 252
 253         spin_unlock(&nlru->lock);
 254         return isolated;
 255 }
 256
 257 unsigned long
 258 list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
 259                   list_lru_walk_cb isolate, void *cb_arg,
 260                   unsigned long *nr_to_walk)
 261 {
 262         return __list_lru_walk_one(lru, nid, memcg_cache_id(memcg),
 263                                    isolate, cb_arg, nr_to_walk);
 264 }
 265 EXPORT_SYMBOL_GPL(list_lru_walk_one);
 266
 267 unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
 268                                  list_lru_walk_cb isolate, void *cb_arg,
 269                                  unsigned long *nr_to_walk)
 270 {
 271         long isolated = 0;
 272         int memcg_idx;
 273
 274         isolated += __list_lru_walk_one(lru, nid, -1, isolate, cb_arg,
 275                                         nr_to_walk);
 276         if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
 277                 for_each_memcg_cache_index(memcg_idx) {
 278                         isolated += __list_lru_walk_one(lru, nid, memcg_idx,
 279                                                 isolate, cb_arg, nr_to_walk);
 280                         if (*nr_to_walk <= 0)
 281                                 break;
 282                 }
 283         }
 284         return isolated;
 285 }
 286 EXPORT_SYMBOL_GPL(list_lru_walk_node);
 287
 288 static void init_one_lru(struct list_lru_one *l)
 289 {
 290         INIT_LIST_HEAD(&l->list);
 291         l->nr_items = 0;
 292 }
 293
 294 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
 295 static void __memcg_destroy_list_lru_node(struct list_lru_memcg *memcg_lrus,
 296                                           int begin, int end)
 297 {
 298         int i;
 299
 300         for (i = begin; i < end; i++)
 301                 kfree(memcg_lrus->lru[i]);
 302 }
 303
 304 static int __memcg_init_list_lru_node(struct list_lru_memcg *memcg_lrus,
 305                                       int begin, int end)
 306 {
 307         int i;
 308
 309         for (i = begin; i < end; i++) {
 310                 struct list_lru_one *l;
 311
 312                 l = kmalloc(sizeof(struct list_lru_one), GFP_KERNEL);
 313                 if (!l)
 314                         goto fail;
 315
 316                 init_one_lru(l);
 317                 memcg_lrus->lru[i] = l;
 318         }
 319         return 0;
 320 fail:
 321         __memcg_destroy_list_lru_node(memcg_lrus, begin, i - 1);
 322         return -ENOMEM;
 323 }
 324
 325 static int memcg_init_list_lru_node(struct list_lru_node *nlru)
 326 {
 327         int size = memcg_nr_cache_ids;
 328
 329         nlru->memcg_lrus = kvmalloc(size * sizeof(void *), GFP_KERNEL);
 330         if (!nlru->memcg_lrus)
 331                 return -ENOMEM;
 332
 333         if (__memcg_init_list_lru_node(nlru->memcg_lrus, 0, size)) {
 334                 kvfree(nlru->memcg_lrus);
 335                 return -ENOMEM;
 336         }
 337
 338         return 0;
 339 }
 340
 341 static void memcg_destroy_list_lru_node(struct list_lru_node *nlru)
 342 {
 343         __memcg_destroy_list_lru_node(nlru->memcg_lrus, 0, memcg_nr_cache_ids);
 344         kvfree(nlru->memcg_lrus);
 345 }
 346
 347 static int memcg_update_list_lru_node(struct list_lru_node *nlru,
 348                                       int old_size, int new_size)
 349 {
 350         struct list_lru_memcg *old, *new;
 351
 352         BUG_ON(old_size > new_size);
 353
 354         old = nlru->memcg_lrus;
 355         new = kvmalloc(new_size * sizeof(void *), GFP_KERNEL);
 356         if (!new)
 357                 return -ENOMEM;
 358
 359         if (__memcg_init_list_lru_node(new, old_size, new_size)) {
 360                 kvfree(new);
 361                 return -ENOMEM;
 362         }
 363
 364         memcpy(new, old, old_size * sizeof(void *));
 365
 366         /*
 367          * The lock guarantees that we won't race with a reader
 368          * (see list_lru_from_memcg_idx).
 369          *
 370          * Since list_lru_{add,del} may be called under an IRQ-safe lock,
 371          * we have to use IRQ-safe primitives here to avoid deadlock.
 372          */
 373         spin_lock_irq(&nlru->lock);
 374         nlru->memcg_lrus = new;
 375         spin_unlock_irq(&nlru->lock);
 376
 377         kvfree(old);
 378         return 0;
 379 }
 380
 381 static void memcg_cancel_update_list_lru_node(struct list_lru_node *nlru,
 382                                               int old_size, int new_size)
 383 {
 384         /* do not bother shrinking the array back to the old size, because we
 385          * cannot handle allocation failures here */
 386         __memcg_destroy_list_lru_node(nlru->memcg_lrus, old_size, new_size);
 387 }
 388
 389 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
 390 {
 391         int i;
 392
 393         if (!memcg_aware)
 394                 return 0;
 395
 396         for_each_node(i) {
 397                 if (memcg_init_list_lru_node(&lru->node[i]))
 398                         goto fail;
 399         }
 400         return 0;
 401 fail:
 402         for (i = i - 1; i >= 0; i--) {
 403                 if (!lru->node[i].memcg_lrus)
 404                         continue;
 405                 memcg_destroy_list_lru_node(&lru->node[i]);
 406         }
 407         return -ENOMEM;
 408 }
 409
 410 static void memcg_destroy_list_lru(struct list_lru *lru)
 411 {
 412         int i;
 413
 414         if (!list_lru_memcg_aware(lru))
 415                 return;
 416
 417         for_each_node(i)
 418                 memcg_destroy_list_lru_node(&lru->node[i]);
 419 }
 420
 421 static int memcg_update_list_lru(struct list_lru *lru,
 422                                  int old_size, int new_size)
 423 {
 424         int i;
 425
 426         if (!list_lru_memcg_aware(lru))
 427                 return 0;
 428
 429         for_each_node(i) {
 430                 if (memcg_update_list_lru_node(&lru->node[i],
 431                                                old_size, new_size))
 432                         goto fail;
 433         }
 434         return 0;
 435 fail:
 436         for (i = i - 1; i >= 0; i--) {
 437                 if (!lru->node[i].memcg_lrus)
 438                         continue;
 439
 440                 memcg_cancel_update_list_lru_node(&lru->node[i],
 441                                                   old_size, new_size);
 442         }
 443         return -ENOMEM;
 444 }
 445
 446 static void memcg_cancel_update_list_lru(struct list_lru *lru,
 447                                          int old_size, int new_size)
 448 {
 449         int i;
 450
 451         if (!list_lru_memcg_aware(lru))
 452                 return;
 453
 454         for_each_node(i)
 455                 memcg_cancel_update_list_lru_node(&lru->node[i],
 456                                                   old_size, new_size);
 457 }
 458
 459 int memcg_update_all_list_lrus(int new_size)
 460 {
 461         int ret = 0;
 462         struct list_lru *lru;
 463         int old_size = memcg_nr_cache_ids;
 464
 465         mutex_lock(&list_lrus_mutex);
 466         list_for_each_entry(lru, &list_lrus, list) {
 467                 ret = memcg_update_list_lru(lru, old_size, new_size);
 468                 if (ret)
 469                         goto fail;
 470         }
 471 out:
 472         mutex_unlock(&list_lrus_mutex);
 473         return ret;
 474 fail:
 475         list_for_each_entry_continue_reverse(lru, &list_lrus, list)
 476                 memcg_cancel_update_list_lru(lru, old_size, new_size);
 477         goto out;
 478 }
 479
 480 static void memcg_drain_list_lru_node(struct list_lru_node *nlru,
 481                                       int src_idx, int dst_idx)
 482 {
 483         struct list_lru_one *src, *dst;
 484
 485         /*
 486          * Since list_lru_{add,del} may be called under an IRQ-safe lock,
 487          * we have to use IRQ-safe primitives here to avoid deadlock.
 488          */
 489         spin_lock_irq(&nlru->lock);
 490
 491         src = list_lru_from_memcg_idx(nlru, src_idx);
 492         dst = list_lru_from_memcg_idx(nlru, dst_idx);
 493
 494         list_splice_init(&src->list, &dst->list);
 495         dst->nr_items += src->nr_items;
 496         src->nr_items = 0;
 497
 498         spin_unlock_irq(&nlru->lock);
 499 }
 500
 501 static void memcg_drain_list_lru(struct list_lru *lru,
 502                                  int src_idx, int dst_idx)
 503 {
 504         int i;
 505
 506         if (!list_lru_memcg_aware(lru))
 507                 return;
 508
 509         for_each_node(i)
 510                 memcg_drain_list_lru_node(&lru->node[i], src_idx, dst_idx);
 511 }
 512
 513 void memcg_drain_all_list_lrus(int src_idx, int dst_idx)
 514 {
 515         struct list_lru *lru;
 516
 517         mutex_lock(&list_lrus_mutex);
 518         list_for_each_entry(lru, &list_lrus, list)
 519                 memcg_drain_list_lru(lru, src_idx, dst_idx);
 520         mutex_unlock(&list_lrus_mutex);
 521 }
 522 #else
 523 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
 524 {
 525         return 0;
 526 }
 527
 528 static void memcg_destroy_list_lru(struct list_lru *lru)
 529 {
 530 }
 531 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
 532
 533 int __list_lru_init(struct list_lru *lru, bool memcg_aware,
 534                     struct lock_class_key *key)
 535 {
 536         int i;
 537         size_t size = sizeof(*lru->node) * nr_node_ids;
 538         int err = -ENOMEM;
 539
 540         memcg_get_cache_ids();
 541
 542         lru->node = kzalloc(size, GFP_KERNEL);
 543         if (!lru->node)
 544                 goto out;
 545
 546         for_each_node(i) {
 547                 spin_lock_init(&lru->node[i].lock);
 548                 if (key)
 549                         lockdep_set_class(&lru->node[i].lock, key);
 550                 init_one_lru(&lru->node[i].lru);
 551         }
 552
 553         err = memcg_init_list_lru(lru, memcg_aware);
 554         if (err) {
 555                 kfree(lru->node);
 556                 /* Do this so a list_lru_destroy() doesn't crash: */
 557                 lru->node = NULL;
 558                 goto out;
 559         }
 560
 561         list_lru_register(lru);
 562 out:
 563         memcg_put_cache_ids();
 564         return err;
 565 }
 566 EXPORT_SYMBOL_GPL(__list_lru_init);
 567
 568 void list_lru_destroy(struct list_lru *lru)
 569 {
 570         /* Already destroyed or not yet initialized? */
 571         if (!lru->node)
 572                 return;
 573
 574         memcg_get_cache_ids();
 575
 576         list_lru_unregister(lru);
 577
 578         memcg_destroy_list_lru(lru);
 579         kfree(lru->node);
 580         lru->node = NULL;
 581
 582         memcg_put_cache_ids();
 583 }
 584 EXPORT_SYMBOL_GPL(list_lru_destroy);