mm/list_lru.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2013 Red Hat, Inc. and Parallels Inc. All rights reserved.
   4  * Authors: David Chinner and Glauber Costa
   5  *
   6  * Generic LRU infrastructure
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/mm.h>
  11 #include <linux/list_lru.h>
  12 #include <linux/slab.h>
  13 #include <linux/mutex.h>
  14 #include <linux/memcontrol.h>
  15 #include "slab.h"
  16 #include "internal.h"
  17
  18 #ifdef CONFIG_MEMCG_KMEM
  19 static LIST_HEAD(memcg_list_lrus);
  20 static DEFINE_MUTEX(list_lrus_mutex);
  21
  22 static inline bool list_lru_memcg_aware(struct list_lru *lru)
  23 {
  24         return lru->memcg_aware;
  25 }
  26
  27 static void list_lru_register(struct list_lru *lru)
  28 {
  29         if (!list_lru_memcg_aware(lru))
  30                 return;
  31
  32         mutex_lock(&list_lrus_mutex);
  33         list_add(&lru->list, &memcg_list_lrus);
  34         mutex_unlock(&list_lrus_mutex);
  35 }
  36
  37 static void list_lru_unregister(struct list_lru *lru)
  38 {
  39         if (!list_lru_memcg_aware(lru))
  40                 return;
  41
  42         mutex_lock(&list_lrus_mutex);
  43         list_del(&lru->list);
  44         mutex_unlock(&list_lrus_mutex);
  45 }
  46
  47 static int lru_shrinker_id(struct list_lru *lru)
  48 {
  49         return lru->shrinker_id;
  50 }
  51
  52 static inline struct list_lru_one *
  53 list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
  54 {
  55         if (list_lru_memcg_aware(lru) && idx >= 0) {
  56                 struct list_lru_memcg *mlru = xa_load(&lru->xa, idx);
  57
  58                 return mlru ? &mlru->node[nid] : NULL;
  59         }
  60         return &lru->node[nid].lru;
  61 }
  62 #else
  63 static void list_lru_register(struct list_lru *lru)
  64 {
  65 }
  66
  67 static void list_lru_unregister(struct list_lru *lru)
  68 {
  69 }
  70
  71 static int lru_shrinker_id(struct list_lru *lru)
  72 {
  73         return -1;
  74 }
  75
  76 static inline bool list_lru_memcg_aware(struct list_lru *lru)
  77 {
  78         return false;
  79 }
  80
  81 static inline struct list_lru_one *
  82 list_lru_from_memcg_idx(struct list_lru *lru, int nid, int idx)
  83 {
  84         return &lru->node[nid].lru;
  85 }
  86 #endif /* CONFIG_MEMCG_KMEM */
  87
  88 bool list_lru_add(struct list_lru *lru, struct list_head *item, int nid,
  89                     struct mem_cgroup *memcg)
  90 {
  91         struct list_lru_node *nlru = &lru->node[nid];
  92         struct list_lru_one *l;
  93
  94         spin_lock(&nlru->lock);
  95         if (list_empty(item)) {
  96                 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
  97                 list_add_tail(item, &l->list);
  98                 /* Set shrinker bit if the first element was added */
  99                 if (!l->nr_items++)
 100                         set_shrinker_bit(memcg, nid, lru_shrinker_id(lru));
 101                 nlru->nr_items++;
 102                 spin_unlock(&nlru->lock);
 103                 return true;
 104         }
 105         spin_unlock(&nlru->lock);
 106         return false;
 107 }
 108 EXPORT_SYMBOL_GPL(list_lru_add);
 109
 110 bool list_lru_add_obj(struct list_lru *lru, struct list_head *item)
 111 {
 112         int nid = page_to_nid(virt_to_page(item));
 113         struct mem_cgroup *memcg = list_lru_memcg_aware(lru) ?
 114                 mem_cgroup_from_slab_obj(item) : NULL;
 115
 116         return list_lru_add(lru, item, nid, memcg);
 117 }
 118 EXPORT_SYMBOL_GPL(list_lru_add_obj);
 119
 120 bool list_lru_del(struct list_lru *lru, struct list_head *item, int nid,
 121                     struct mem_cgroup *memcg)
 122 {
 123         struct list_lru_node *nlru = &lru->node[nid];
 124         struct list_lru_one *l;
 125
 126         spin_lock(&nlru->lock);
 127         if (!list_empty(item)) {
 128                 l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
 129                 list_del_init(item);
 130                 l->nr_items--;
 131                 nlru->nr_items--;
 132                 spin_unlock(&nlru->lock);
 133                 return true;
 134         }
 135         spin_unlock(&nlru->lock);
 136         return false;
 137 }
 138 EXPORT_SYMBOL_GPL(list_lru_del);
 139
 140 bool list_lru_del_obj(struct list_lru *lru, struct list_head *item)
 141 {
 142         int nid = page_to_nid(virt_to_page(item));
 143         struct mem_cgroup *memcg = list_lru_memcg_aware(lru) ?
 144                 mem_cgroup_from_slab_obj(item) : NULL;
 145
 146         return list_lru_del(lru, item, nid, memcg);
 147 }
 148 EXPORT_SYMBOL_GPL(list_lru_del_obj);
 149
 150 void list_lru_isolate(struct list_lru_one *list, struct list_head *item)
 151 {
 152         list_del_init(item);
 153         list->nr_items--;
 154 }
 155 EXPORT_SYMBOL_GPL(list_lru_isolate);
 156
 157 void list_lru_isolate_move(struct list_lru_one *list, struct list_head *item,
 158                            struct list_head *head)
 159 {
 160         list_move(item, head);
 161         list->nr_items--;
 162 }
 163 EXPORT_SYMBOL_GPL(list_lru_isolate_move);
 164
 165 unsigned long list_lru_count_one(struct list_lru *lru,
 166                                  int nid, struct mem_cgroup *memcg)
 167 {
 168         struct list_lru_one *l;
 169         long count;
 170
 171         rcu_read_lock();
 172         l = list_lru_from_memcg_idx(lru, nid, memcg_kmem_id(memcg));
 173         count = l ? READ_ONCE(l->nr_items) : 0;
 174         rcu_read_unlock();
 175
 176         if (unlikely(count < 0))
 177                 count = 0;
 178
 179         return count;
 180 }
 181 EXPORT_SYMBOL_GPL(list_lru_count_one);
 182
 183 unsigned long list_lru_count_node(struct list_lru *lru, int nid)
 184 {
 185         struct list_lru_node *nlru;
 186
 187         nlru = &lru->node[nid];
 188         return nlru->nr_items;
 189 }
 190 EXPORT_SYMBOL_GPL(list_lru_count_node);
 191
 192 static unsigned long
 193 __list_lru_walk_one(struct list_lru *lru, int nid, int memcg_idx,
 194                     list_lru_walk_cb isolate, void *cb_arg,
 195                     unsigned long *nr_to_walk)
 196 {
 197         struct list_lru_node *nlru = &lru->node[nid];
 198         struct list_lru_one *l;
 199         struct list_head *item, *n;
 200         unsigned long isolated = 0;
 201
 202 restart:
 203         l = list_lru_from_memcg_idx(lru, nid, memcg_idx);
 204         if (!l)
 205                 goto out;
 206
 207         list_for_each_safe(item, n, &l->list) {
 208                 enum lru_status ret;
 209
 210                 /*
 211                  * decrement nr_to_walk first so that we don't livelock if we
 212                  * get stuck on large numbers of LRU_RETRY items
 213                  */
 214                 if (!*nr_to_walk)
 215                         break;
 216                 --*nr_to_walk;
 217
 218                 ret = isolate(item, l, &nlru->lock, cb_arg);
 219                 switch (ret) {
 220                 case LRU_REMOVED_RETRY:
 221                         assert_spin_locked(&nlru->lock);
 222                         fallthrough;
 223                 case LRU_REMOVED:
 224                         isolated++;
 225                         nlru->nr_items--;
 226                         /*
 227                          * If the lru lock has been dropped, our list
 228                          * traversal is now invalid and so we have to
 229                          * restart from scratch.
 230                          */
 231                         if (ret == LRU_REMOVED_RETRY)
 232                                 goto restart;
 233                         break;
 234                 case LRU_ROTATE:
 235                         list_move_tail(item, &l->list);
 236                         break;
 237                 case LRU_SKIP:
 238                         break;
 239                 case LRU_RETRY:
 240                         /*
 241                          * The lru lock has been dropped, our list traversal is
 242                          * now invalid and so we have to restart from scratch.
 243                          */
 244                         assert_spin_locked(&nlru->lock);
 245                         goto restart;
 246                 case LRU_STOP:
 247                         assert_spin_locked(&nlru->lock);
 248                         goto out;
 249                 default:
 250                         BUG();
 251                 }
 252         }
 253 out:
 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         struct list_lru_node *nlru = &lru->node[nid];
 263         unsigned long ret;
 264
 265         spin_lock(&nlru->lock);
 266         ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
 267                                   cb_arg, nr_to_walk);
 268         spin_unlock(&nlru->lock);
 269         return ret;
 270 }
 271 EXPORT_SYMBOL_GPL(list_lru_walk_one);
 272
 273 unsigned long
 274 list_lru_walk_one_irq(struct list_lru *lru, int nid, struct mem_cgroup *memcg,
 275                       list_lru_walk_cb isolate, void *cb_arg,
 276                       unsigned long *nr_to_walk)
 277 {
 278         struct list_lru_node *nlru = &lru->node[nid];
 279         unsigned long ret;
 280
 281         spin_lock_irq(&nlru->lock);
 282         ret = __list_lru_walk_one(lru, nid, memcg_kmem_id(memcg), isolate,
 283                                   cb_arg, nr_to_walk);
 284         spin_unlock_irq(&nlru->lock);
 285         return ret;
 286 }
 287
 288 unsigned long list_lru_walk_node(struct list_lru *lru, int nid,
 289                                  list_lru_walk_cb isolate, void *cb_arg,
 290                                  unsigned long *nr_to_walk)
 291 {
 292         long isolated = 0;
 293
 294         isolated += list_lru_walk_one(lru, nid, NULL, isolate, cb_arg,
 295                                       nr_to_walk);
 296
 297 #ifdef CONFIG_MEMCG_KMEM
 298         if (*nr_to_walk > 0 && list_lru_memcg_aware(lru)) {
 299                 struct list_lru_memcg *mlru;
 300                 unsigned long index;
 301
 302                 xa_for_each(&lru->xa, index, mlru) {
 303                         struct list_lru_node *nlru = &lru->node[nid];
 304
 305                         spin_lock(&nlru->lock);
 306                         isolated += __list_lru_walk_one(lru, nid, index,
 307                                                         isolate, cb_arg,
 308                                                         nr_to_walk);
 309                         spin_unlock(&nlru->lock);
 310
 311                         if (*nr_to_walk <= 0)
 312                                 break;
 313                 }
 314         }
 315 #endif
 316
 317         return isolated;
 318 }
 319 EXPORT_SYMBOL_GPL(list_lru_walk_node);
 320
 321 static void init_one_lru(struct list_lru_one *l)
 322 {
 323         INIT_LIST_HEAD(&l->list);
 324         l->nr_items = 0;
 325 }
 326
 327 #ifdef CONFIG_MEMCG_KMEM
 328 static struct list_lru_memcg *memcg_init_list_lru_one(gfp_t gfp)
 329 {
 330         int nid;
 331         struct list_lru_memcg *mlru;
 332
 333         mlru = kmalloc(struct_size(mlru, node, nr_node_ids), gfp);
 334         if (!mlru)
 335                 return NULL;
 336
 337         for_each_node(nid)
 338                 init_one_lru(&mlru->node[nid]);
 339
 340         return mlru;
 341 }
 342
 343 static void memcg_list_lru_free(struct list_lru *lru, int src_idx)
 344 {
 345         struct list_lru_memcg *mlru = xa_erase_irq(&lru->xa, src_idx);
 346
 347         /*
 348          * The __list_lru_walk_one() can walk the list of this node.
 349          * We need kvfree_rcu() here. And the walking of the list
 350          * is under lru->node[nid]->lock, which can serve as a RCU
 351          * read-side critical section.
 352          */
 353         if (mlru)
 354                 kvfree_rcu(mlru, rcu);
 355 }
 356
 357 static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
 358 {
 359         if (memcg_aware)
 360                 xa_init_flags(&lru->xa, XA_FLAGS_LOCK_IRQ);
 361         lru->memcg_aware = memcg_aware;
 362 }
 363
 364 static void memcg_destroy_list_lru(struct list_lru *lru)
 365 {
 366         XA_STATE(xas, &lru->xa, 0);
 367         struct list_lru_memcg *mlru;
 368
 369         if (!list_lru_memcg_aware(lru))
 370                 return;
 371
 372         xas_lock_irq(&xas);
 373         xas_for_each(&xas, mlru, ULONG_MAX) {
 374                 kfree(mlru);
 375                 xas_store(&xas, NULL);
 376         }
 377         xas_unlock_irq(&xas);
 378 }
 379
 380 static void memcg_reparent_list_lru_node(struct list_lru *lru, int nid,
 381                                          int src_idx, struct mem_cgroup *dst_memcg)
 382 {
 383         struct list_lru_node *nlru = &lru->node[nid];
 384         int dst_idx = dst_memcg->kmemcg_id;
 385         struct list_lru_one *src, *dst;
 386
 387         /*
 388          * Since list_lru_{add,del} may be called under an IRQ-safe lock,
 389          * we have to use IRQ-safe primitives here to avoid deadlock.
 390          */
 391         spin_lock_irq(&nlru->lock);
 392
 393         src = list_lru_from_memcg_idx(lru, nid, src_idx);
 394         if (!src)
 395                 goto out;
 396         dst = list_lru_from_memcg_idx(lru, nid, dst_idx);
 397
 398         list_splice_init(&src->list, &dst->list);
 399
 400         if (src->nr_items) {
 401                 dst->nr_items += src->nr_items;
 402                 set_shrinker_bit(dst_memcg, nid, lru_shrinker_id(lru));
 403                 src->nr_items = 0;
 404         }
 405 out:
 406         spin_unlock_irq(&nlru->lock);
 407 }
 408
 409 static void memcg_reparent_list_lru(struct list_lru *lru,
 410                                     int src_idx, struct mem_cgroup *dst_memcg)
 411 {
 412         int i;
 413
 414         for_each_node(i)
 415                 memcg_reparent_list_lru_node(lru, i, src_idx, dst_memcg);
 416
 417         memcg_list_lru_free(lru, src_idx);
 418 }
 419
 420 void memcg_reparent_list_lrus(struct mem_cgroup *memcg, struct mem_cgroup *parent)
 421 {
 422         struct cgroup_subsys_state *css;
 423         struct list_lru *lru;
 424         int src_idx = memcg->kmemcg_id;
 425
 426         /*
 427          * Change kmemcg_id of this cgroup and all its descendants to the
 428          * parent's id, and then move all entries from this cgroup's list_lrus
 429          * to ones of the parent.
 430          *
 431          * After we have finished, all list_lrus corresponding to this cgroup
 432          * are guaranteed to remain empty. So we can safely free this cgroup's
 433          * list lrus in memcg_list_lru_free().
 434          *
 435          * Changing ->kmemcg_id to the parent can prevent memcg_list_lru_alloc()
 436          * from allocating list lrus for this cgroup after memcg_list_lru_free()
 437          * call.
 438          */
 439         rcu_read_lock();
 440         css_for_each_descendant_pre(css, &memcg->css) {
 441                 struct mem_cgroup *child;
 442
 443                 child = mem_cgroup_from_css(css);
 444                 WRITE_ONCE(child->kmemcg_id, parent->kmemcg_id);
 445         }
 446         rcu_read_unlock();
 447
 448         mutex_lock(&list_lrus_mutex);
 449         list_for_each_entry(lru, &memcg_list_lrus, list)
 450                 memcg_reparent_list_lru(lru, src_idx, parent);
 451         mutex_unlock(&list_lrus_mutex);
 452 }
 453
 454 static inline bool memcg_list_lru_allocated(struct mem_cgroup *memcg,
 455                                             struct list_lru *lru)
 456 {
 457         int idx = memcg->kmemcg_id;
 458
 459         return idx < 0 || xa_load(&lru->xa, idx);
 460 }
 461
 462 int memcg_list_lru_alloc(struct mem_cgroup *memcg, struct list_lru *lru,
 463                          gfp_t gfp)
 464 {
 465         int i;
 466         unsigned long flags;
 467         struct list_lru_memcg_table {
 468                 struct list_lru_memcg *mlru;
 469                 struct mem_cgroup *memcg;
 470         } *table;
 471         XA_STATE(xas, &lru->xa, 0);
 472
 473         if (!list_lru_memcg_aware(lru) || memcg_list_lru_allocated(memcg, lru))
 474                 return 0;
 475
 476         gfp &= GFP_RECLAIM_MASK;
 477         table = kmalloc_array(memcg->css.cgroup->level, sizeof(*table), gfp);
 478         if (!table)
 479                 return -ENOMEM;
 480
 481         /*
 482          * Because the list_lru can be reparented to the parent cgroup's
 483          * list_lru, we should make sure that this cgroup and all its
 484          * ancestors have allocated list_lru_memcg.
 485          */
 486         for (i = 0; memcg; memcg = parent_mem_cgroup(memcg), i++) {
 487                 if (memcg_list_lru_allocated(memcg, lru))
 488                         break;
 489
 490                 table[i].memcg = memcg;
 491                 table[i].mlru = memcg_init_list_lru_one(gfp);
 492                 if (!table[i].mlru) {
 493                         while (i--)
 494                                 kfree(table[i].mlru);
 495                         kfree(table);
 496                         return -ENOMEM;
 497                 }
 498         }
 499
 500         xas_lock_irqsave(&xas, flags);
 501         while (i--) {
 502                 int index = READ_ONCE(table[i].memcg->kmemcg_id);
 503                 struct list_lru_memcg *mlru = table[i].mlru;
 504
 505                 xas_set(&xas, index);
 506 retry:
 507                 if (unlikely(index < 0 || xas_error(&xas) || xas_load(&xas))) {
 508                         kfree(mlru);
 509                 } else {
 510                         xas_store(&xas, mlru);
 511                         if (xas_error(&xas) == -ENOMEM) {
 512                                 xas_unlock_irqrestore(&xas, flags);
 513                                 if (xas_nomem(&xas, gfp))
 514                                         xas_set_err(&xas, 0);
 515                                 xas_lock_irqsave(&xas, flags);
 516                                 /*
 517                                  * The xas lock has been released, this memcg
 518                                  * can be reparented before us. So reload
 519                                  * memcg id. More details see the comments
 520                                  * in memcg_reparent_list_lrus().
 521                                  */
 522                                 index = READ_ONCE(table[i].memcg->kmemcg_id);
 523                                 if (index < 0)
 524                                         xas_set_err(&xas, 0);
 525                                 else if (!xas_error(&xas) && index != xas.xa_index)
 526                                         xas_set(&xas, index);
 527                                 goto retry;
 528                         }
 529                 }
 530         }
 531         /* xas_nomem() is used to free memory instead of memory allocation. */
 532         if (xas.xa_alloc)
 533                 xas_nomem(&xas, gfp);
 534         xas_unlock_irqrestore(&xas, flags);
 535         kfree(table);
 536
 537         return xas_error(&xas);
 538 }
 539 #else
 540 static inline void memcg_init_list_lru(struct list_lru *lru, bool memcg_aware)
 541 {
 542 }
 543
 544 static void memcg_destroy_list_lru(struct list_lru *lru)
 545 {
 546 }
 547 #endif /* CONFIG_MEMCG_KMEM */
 548
 549 int __list_lru_init(struct list_lru *lru, bool memcg_aware,
 550                     struct lock_class_key *key, struct shrinker *shrinker)
 551 {
 552         int i;
 553
 554 #ifdef CONFIG_MEMCG_KMEM
 555         if (shrinker)
 556                 lru->shrinker_id = shrinker->id;
 557         else
 558                 lru->shrinker_id = -1;
 559
 560         if (mem_cgroup_kmem_disabled())
 561                 memcg_aware = false;
 562 #endif
 563
 564         lru->node = kcalloc(nr_node_ids, sizeof(*lru->node), GFP_KERNEL);
 565         if (!lru->node)
 566                 return -ENOMEM;
 567
 568         for_each_node(i) {
 569                 spin_lock_init(&lru->node[i].lock);
 570                 if (key)
 571                         lockdep_set_class(&lru->node[i].lock, key);
 572                 init_one_lru(&lru->node[i].lru);
 573         }
 574
 575         memcg_init_list_lru(lru, memcg_aware);
 576         list_lru_register(lru);
 577
 578         return 0;
 579 }
 580 EXPORT_SYMBOL_GPL(__list_lru_init);
 581
 582 void list_lru_destroy(struct list_lru *lru)
 583 {
 584         /* Already destroyed or not yet initialized? */
 585         if (!lru->node)
 586                 return;
 587
 588         list_lru_unregister(lru);
 589
 590         memcg_destroy_list_lru(lru);
 591         kfree(lru->node);
 592         lru->node = NULL;
 593
 594 #ifdef CONFIG_MEMCG_KMEM
 595         lru->shrinker_id = -1;
 596 #endif
 597 }
 598 EXPORT_SYMBOL_GPL(list_lru_destroy);