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                 spin_unlock(&nlru->lock);
 121                 return true;
 122         }
 123         spin_unlock(&nlru->lock);
 124         return false;
 125 }
 126 EXPORT_SYMBOL_GPL(list_lru_add);
 127
 128 bool list_lru_del(struct list_lru *lru, struct list_head *item)
 129 {
 130         int nid = page_to_nid(virt_to_page(item));
 131         struct list_lru_node *nlru = &lru->node[nid];
 132         struct list_lru_one *l;
 133
 134         spin_lock(&nlru->lock);
 135         if (!list_empty(item)) {
 136                 l = list_lru_from_kmem(nlru, item);
 137                 list_del_init(item);
 138                 l->nr_items--;
 139                 spin_unlock(&nlru->lock);
 140                 return true;
 141         }
 142         spin_unlock(&nlru->lock);
 143         return false;
 144 }
 145 EXPORT_SYMBOL_GPL(list_lru_del);
 146
 147 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
 148 {
 149         list_del_init(item);
 150         list->nr_items--;
 151 }
 152 EXPORT_SYMBOL_GPL(list_lru_isolate);
 153
 154 void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
 155                            struct list_head *head)
 156 {
 157         list_move(item, head);
 158         list->nr_items--;
 159 }
 160 EXPORT_SYMBOL_GPL(list_lru_isolate_move);
 161
 162 static unsigned long __list_lru_count_one(struct list_lru *lru,
 163                                           int nid, int memcg_idx)
 164 {
 165         struct list_lru_node *nlru = &lru->node[nid];
 166         struct list_lru_one *l;
 167         unsigned long count;
 168
 169         spin_lock(&nlru->lock);
 170         l = list_lru_from_memcg_idx(nlru, memcg_idx);
 171         count = l->nr_items;
 172         spin_unlock(&nlru->lock);
 173
 174         return count;
 175 }
 176
 177 unsigned long list_lru_count_one(struct list_lru *lru,
 178                                  int nid, struct mem_cgroup *memcg)
 179 {
 180         return __list_lru_count_one(lru, nid, memcg_cache_id(memcg));
 181 }
 182 EXPORT_SYMBOL_GPL(list_lru_count_one);
 183
 184 unsigned long list_lru_count_node(struct list_lru *lru, int nid)
 185 {
 186         long count = 0;
 187         int memcg_idx;
 188
 189         count += __list_lru_count_one(lru, nid, -1);
 190         if (list_lru_memcg_aware(lru)) {
 191                 for_each_memcg_cache_index(memcg_idx)
 192                         count += __list_lru_count_one(lru, nid, memcg_idx);
 193         }
 194         return count;
 195 }
 196 EXPORT_SYMBOL_GPL(list_lru_count_node);
 197
 198 static unsigned long
 199 __list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
 200                     list_lru_walk_cb isolate, void *cb_arg,
 201                     unsigned long *nr_to_walk)
 202 {
 203
 204         struct list_lru_node *nlru = &lru->node[nid];
 205         struct list_lru_one *l;
 206         struct list_head *item, *n;
 207         unsigned long isolated = 0;
 208
 209         spin_lock(&nlru->lock);
 210         l = list_lru_from_memcg_idx(nlru, memcg_idx);
 211 restart:
 212         list_for_each_safe(item, n, &l->list) {
 213                 enum lru_status ret;
 214
 215                 /*
 216                  * decrement nr_to_walk first so that we don't livelock if we
 217                  * get stuck on large numbesr of LRU_RETRY items
 218                  */
 219                 if (!*nr_to_walk)
 220                         break;
 221                 --*nr_to_walk;
 222
 223                 ret = isolate(item, l, &nlru->lock, cb_arg);
 224                 switch (ret) {
 225                 case LRU_REMOVED_RETRY:
 226                         assert_spin_locked(&nlru->lock);
 227                 case LRU_REMOVED:
 228                         isolated++;
 229                         /*
 230                          * If the lru lock has been dropped, our list
 231                          * traversal is now invalid and so we have to
 232                          * restart from scratch.
 233                          */
 234                         if (ret == LRU_REMOVED_RETRY)
 235                                 goto restart;
 236                         break;
 237                 case LRU_ROTATE:
 238                         list_move_tail(item, &l->list);
 239                         break;
 240                 case LRU_SKIP:
 241                         break;
 242                 case LRU_RETRY:
 243                         /*
 244                          * The lru lock has been dropped, our list traversal is
 245                          * now invalid and so we have to restart from scratch.
 246                          */
 247                         assert_spin_locked(&nlru->lock);
 248                         goto restart;
 249                 default:
 250                         BUG();
 251                 }
 252         }
 253
 254         spin_unlock(&nlru->lock);
 255         return isolated;
 256 }
 257
 258 unsigned long
 259 list_lru_walk_one(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
 260                   list_lru_walk_cb isolate, void *cb_arg,
 261                   unsigned long *nr_to_walk)
 262 {
 263         return __list_lru_walk_one(lru, nid, memcg_cache_id(memcg),
 264                                    isolate, cb_arg, nr_to_walk);
 265 }
 266 EXPORT_SYMBOL_GPL(list_lru_walk_one);
 267
 268 unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
 269                                  list_lru_walk_cb isolate, void *cb_arg,
 270                                  unsigned long *nr_to_walk)
 271 {
 272         long isolated = 0;
 273         int memcg_idx;
 274
 275         isolated += __list_lru_walk_one(lru, nid, -1, isolate, cb_arg,
 276                                         nr_to_walk);
 277         if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
 278                 for_each_memcg_cache_index(memcg_idx) {
 279                         isolated += __list_lru_walk_one(lru, nid, memcg_idx,
 280                                                 isolate, cb_arg, nr_to_walk);
 281                         if (*nr_to_walk <= 0)
 282                                 break;
 283                 }
 284         }
 285         return isolated;
 286 }
 287 EXPORT_SYMBOL_GPL(list_lru_walk_node);
 288
 289 static void init_one_lru(struct list_lru_one *l)
 290 {
 291         INIT_LIST_HEAD(&l->list);
 292         l->nr_items = 0;
 293 }
 294
 295 #if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
 296 static void __memcg_destroy_list_lru_node(struct list_lru_memcg *memcg_lrus,
 297                                           int begin, int end)
 298 {
 299         int i;
 300
 301         for (i = begin; i < end; i++)
 302                 kfree(memcg_lrus->lru[i]);
 303 }
 304
 305 static int __memcg_init_list_lru_node(struct list_lru_memcg *memcg_lrus,
 306                                       int begin, int end)
 307 {
 308         int i;
 309
 310         for (i = begin; i < end; i++) {
 311                 struct list_lru_one *l;
 312
 313                 l = kmalloc(sizeof(struct list_lru_one), GFP_KERNEL);
 314                 if (!l)
 315                         goto fail;
 316
 317                 init_one_lru(l);
 318                 memcg_lrus->lru[i] = l;
 319         }
 320         return 0;
 321 fail:
 322         __memcg_destroy_list_lru_node(memcg_lrus, begin, i - 1);
 323         return -ENOMEM;
 324 }
 325
 326 static int memcg_init_list_lru_node(struct list_lru_node *nlru)
 327 {
 328         int size = memcg_nr_cache_ids;
 329
 330         nlru->memcg_lrus = kmalloc(size * sizeof(void *), GFP_KERNEL);
 331         if (!nlru->memcg_lrus)
 332                 return -ENOMEM;
 333
 334         if (__memcg_init_list_lru_node(nlru->memcg_lrus, 0, size)) {
 335                 kfree(nlru->memcg_lrus);
 336                 return -ENOMEM;
 337         }
 338
 339         return 0;
 340 }
 341
 342 static void memcg_destroy_list_lru_node(struct list_lru_node *nlru)
 343 {
 344         __memcg_destroy_list_lru_node(nlru->memcg_lrus, 0, memcg_nr_cache_ids);
 345         kfree(nlru->memcg_lrus);
 346 }
 347
 348 static int memcg_update_list_lru_node(struct list_lru_node *nlru,
 349                                       int old_size, int new_size)
 350 {
 351         struct list_lru_memcg *old, *new;
 352
 353         BUG_ON(old_size > new_size);
 354
 355         old = nlru->memcg_lrus;
 356         new = kmalloc(new_size * sizeof(void *), GFP_KERNEL);
 357         if (!new)
 358                 return -ENOMEM;
 359
 360         if (__memcg_init_list_lru_node(new, old_size, new_size)) {
 361                 kfree(new);
 362                 return -ENOMEM;
 363         }
 364
 365         memcpy(new, old, old_size * sizeof(void *));
 366
 367         /*
 368          * The lock guarantees that we won't race with a reader
 369          * (see list_lru_from_memcg_idx).
 370          *
 371          * Since list_lru_{add,del} may be called under an IRQ-safe lock,
 372          * we have to use IRQ-safe primitives here to avoid deadlock.
 373          */
 374         spin_lock_irq(&nlru->lock);
 375         nlru->memcg_lrus = new;
 376         spin_unlock_irq(&nlru->lock);
 377
 378         kfree(old);
 379         return 0;
 380 }
 381
 382 static void memcg_cancel_update_list_lru_node(struct list_lru_node *nlru,
 383                                               int old_size, int new_size)
 384 {
 385         /* do not bother shrinking the array back to the old size, because we
 386          * cannot handle allocation failures here */
 387         __memcg_destroy_list_lru_node(nlru->memcg_lrus, old_size, new_size);
 388 }
 389
 390 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
 391 {
 392         int i;
 393
 394         if (!memcg_aware)
 395                 return 0;
 396
 397         for_each_node(i) {
 398                 if (memcg_init_list_lru_node(&lru->node[i]))
 399                         goto fail;
 400         }
 401         return 0;
 402 fail:
 403         for (i = i - 1; i >= 0; i--) {
 404                 if (!lru->node[i].memcg_lrus)
 405                         continue;
 406                 memcg_destroy_list_lru_node(&lru->node[i]);
 407         }
 408         return -ENOMEM;
 409 }
 410
 411 static void memcg_destroy_list_lru(struct list_lru *lru)
 412 {
 413         int i;
 414
 415         if (!list_lru_memcg_aware(lru))
 416                 return;
 417
 418         for_each_node(i)
 419                 memcg_destroy_list_lru_node(&lru->node[i]);
 420 }
 421
 422 static int memcg_update_list_lru(struct list_lru *lru,
 423                                  int old_size, int new_size)
 424 {
 425         int i;
 426
 427         if (!list_lru_memcg_aware(lru))
 428                 return 0;
 429
 430         for_each_node(i) {
 431                 if (memcg_update_list_lru_node(&lru->node[i],
 432                                                old_size, new_size))
 433                         goto fail;
 434         }
 435         return 0;
 436 fail:
 437         for (i = i - 1; i >= 0; i--) {
 438                 if (!lru->node[i].memcg_lrus)
 439                         continue;
 440
 441                 memcg_cancel_update_list_lru_node(&lru->node[i],
 442                                                   old_size, new_size);
 443         }
 444         return -ENOMEM;
 445 }
 446
 447 static void memcg_cancel_update_list_lru(struct list_lru *lru,
 448                                          int old_size, int new_size)
 449 {
 450         int i;
 451
 452         if (!list_lru_memcg_aware(lru))
 453                 return;
 454
 455         for_each_node(i)
 456                 memcg_cancel_update_list_lru_node(&lru->node[i],
 457                                                   old_size, new_size);
 458 }
 459
 460 int memcg_update_all_list_lrus(int new_size)
 461 {
 462         int ret = 0;
 463         struct list_lru *lru;
 464         int old_size = memcg_nr_cache_ids;
 465
 466         mutex_lock(&list_lrus_mutex);
 467         list_for_each_entry(lru, &list_lrus, list) {
 468                 ret = memcg_update_list_lru(lru, old_size, new_size);
 469                 if (ret)
 470                         goto fail;
 471         }
 472 out:
 473         mutex_unlock(&list_lrus_mutex);
 474         return ret;
 475 fail:
 476         list_for_each_entry_continue_reverse(lru, &list_lrus, list)
 477                 memcg_cancel_update_list_lru(lru, old_size, new_size);
 478         goto out;
 479 }
 480
 481 static void memcg_drain_list_lru_node(struct list_lru_node *nlru,
 482                                       int src_idx, int dst_idx)
 483 {
 484         struct list_lru_one *src, *dst;
 485
 486         /*
 487          * Since list_lru_{add,del} may be called under an IRQ-safe lock,
 488          * we have to use IRQ-safe primitives here to avoid deadlock.
 489          */
 490         spin_lock_irq(&nlru->lock);
 491
 492         src = list_lru_from_memcg_idx(nlru, src_idx);
 493         dst = list_lru_from_memcg_idx(nlru, dst_idx);
 494
 495         list_splice_init(&src->list, &dst->list);
 496         dst->nr_items += src->nr_items;
 497         src->nr_items = 0;
 498
 499         spin_unlock_irq(&nlru->lock);
 500 }
 501
 502 static void memcg_drain_list_lru(struct list_lru *lru,
 503                                  int src_idx, int dst_idx)
 504 {
 505         int i;
 506
 507         if (!list_lru_memcg_aware(lru))
 508                 return;
 509
 510         for_each_node(i)
 511                 memcg_drain_list_lru_node(&lru->node[i], src_idx, dst_idx);
 512 }
 513
 514 void memcg_drain_all_list_lrus(int src_idx, int dst_idx)
 515 {
 516         struct list_lru *lru;
 517
 518         mutex_lock(&list_lrus_mutex);
 519         list_for_each_entry(lru, &list_lrus, list)
 520                 memcg_drain_list_lru(lru, src_idx, dst_idx);
 521         mutex_unlock(&list_lrus_mutex);
 522 }
 523 #else
 524 static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
 525 {
 526         return 0;
 527 }
 528
 529 static void memcg_destroy_list_lru(struct list_lru *lru)
 530 {
 531 }
 532 #endif /* CONFIG_MEMCG && !CONFIG_SLOB */
 533
 534 int __list_lru_init(struct list_lru *lru, bool memcg_aware,
 535                     struct lock_class_key *key)
 536 {
 537         int i;
 538         size_t size = sizeof(*lru->node) * nr_node_ids;
 539         int err = -ENOMEM;
 540
 541         memcg_get_cache_ids();
 542
 543         lru->node = kzalloc(size, GFP_KERNEL);
 544         if (!lru->node)
 545                 goto out;
 546
 547         for_each_node(i) {
 548                 spin_lock_init(&lru->node[i].lock);
 549                 if (key)
 550                         lockdep_set_class(&lru->node[i].lock, key);
 551                 init_one_lru(&lru->node[i].lru);
 552         }
 553
 554         err = memcg_init_list_lru(lru, memcg_aware);
 555         if (err) {
 556                 kfree(lru->node);
 557                 goto out;
 558         }
 559
 560         list_lru_register(lru);
 561 out:
 562         memcg_put_cache_ids();
 563         return err;
 564 }
 565 EXPORT_SYMBOL_GPL(__list_lru_init);
 566
 567 void list_lru_destroy(struct list_lru *lru)
 568 {
 569         /* Already destroyed or not yet initialized? */
 570         if (!lru->node)
 571                 return;
 572
 573         memcg_get_cache_ids();
 574
 575         list_lru_unregister(lru);
 576
 577         memcg_destroy_list_lru(lru);
 578         kfree(lru->node);
 579         lru->node = NULL;
 580
 581         memcg_put_cache_ids();
 582 }
 583 EXPORT_SYMBOL_GPL(list_lru_destroy);