mm/memcontrol.c

   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  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  */
  19
  20 #include <linux/res_counter.h>
  21 #include <linux/memcontrol.h>
  22 #include <linux/cgroup.h>
  23 #include <linux/mm.h>
  24 #include <linux/page-flags.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/bit_spinlock.h>
  27 #include <linux/rcupdate.h>
  28 #include <linux/swap.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/fs.h>
  31
  32 #include <asm/uaccess.h>
  33
  34 struct cgroup_subsys mem_cgroup_subsys;
  35 static const int MEM_CGROUP_RECLAIM_RETRIES = 5;
  36
  37 /*
  38  * The memory controller data structure. The memory controller controls both
  39  * page cache and RSS per cgroup. We would eventually like to provide
  40  * statistics based on the statistics developed by Rik Van Riel for clock-pro,
  41  * to help the administrator determine what knobs to tune.
  42  *
  43  * TODO: Add a water mark for the memory controller. Reclaim will begin when
  44  * we hit the water mark. May be even add a low water mark, such that
  45  * no reclaim occurs from a cgroup at it's low water mark, this is
  46  * a feature that will be implemented much later in the future.
  47  */
  48 struct mem_cgroup {
  49         struct cgroup_subsys_state css;
  50         /*
  51          * the counter to account for memory usage
  52          */
  53         struct res_counter res;
  54         /*
  55          * Per cgroup active and inactive list, similar to the
  56          * per zone LRU lists.
  57          * TODO: Consider making these lists per zone
  58          */
  59         struct list_head active_list;
  60         struct list_head inactive_list;
  61         /*
  62          * spin_lock to protect the per cgroup LRU
  63          */
  64         spinlock_t lru_lock;
  65         unsigned long control_type;     /* control RSS or RSS+Pagecache */
  66 };
  67
  68 /*
  69  * We use the lower bit of the page->page_cgroup pointer as a bit spin
  70  * lock. We need to ensure that page->page_cgroup is atleast two
  71  * byte aligned (based on comments from Nick Piggin)
  72  */
  73 #define PAGE_CGROUP_LOCK_BIT    0x0
  74 #define PAGE_CGROUP_LOCK                (1 << PAGE_CGROUP_LOCK_BIT)
  75
  76 /*
  77  * A page_cgroup page is associated with every page descriptor. The
  78  * page_cgroup helps us identify information about the cgroup
  79  */
  80 struct page_cgroup {
  81         struct list_head lru;           /* per cgroup LRU list */
  82         struct page *page;
  83         struct mem_cgroup *mem_cgroup;
  84         atomic_t ref_cnt;               /* Helpful when pages move b/w  */
  85                                         /* mapped and cached states     */
  86 };
  87
  88 enum {
  89         MEM_CGROUP_TYPE_UNSPEC = 0,
  90         MEM_CGROUP_TYPE_MAPPED,
  91         MEM_CGROUP_TYPE_CACHED,
  92         MEM_CGROUP_TYPE_ALL,
  93         MEM_CGROUP_TYPE_MAX,
  94 };
  95
  96 static struct mem_cgroup init_mem_cgroup;
  97
  98 static inline
  99 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 100 {
 101         return container_of(cgroup_subsys_state(cont,
 102                                 mem_cgroup_subsys_id), struct mem_cgroup,
 103                                 css);
 104 }
 105
 106 static inline
 107 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 108 {
 109         return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
 110                                 struct mem_cgroup, css);
 111 }
 112
 113 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
 114 {
 115         struct mem_cgroup *mem;
 116
 117         mem = mem_cgroup_from_task(p);
 118         css_get(&mem->css);
 119         mm->mem_cgroup = mem;
 120 }
 121
 122 void mm_free_cgroup(struct mm_struct *mm)
 123 {
 124         css_put(&mm->mem_cgroup->css);
 125 }
 126
 127 static inline int page_cgroup_locked(struct page *page)
 128 {
 129         return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
 130                                         &page->page_cgroup);
 131 }
 132
 133 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 134 {
 135         int locked;
 136
 137         /*
 138          * While resetting the page_cgroup we might not hold the
 139          * page_cgroup lock. free_hot_cold_page() is an example
 140          * of such a scenario
 141          */
 142         if (pc)
 143                 VM_BUG_ON(!page_cgroup_locked(page));
 144         locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
 145         page->page_cgroup = ((unsigned long)pc | locked);
 146 }
 147
 148 struct page_cgroup *page_get_page_cgroup(struct page *page)
 149 {
 150         return (struct page_cgroup *)
 151                 (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 152 }
 153
 154 static void __always_inline lock_page_cgroup(struct page *page)
 155 {
 156         bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 157         VM_BUG_ON(!page_cgroup_locked(page));
 158 }
 159
 160 static void __always_inline unlock_page_cgroup(struct page *page)
 161 {
 162         bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 163 }
 164
 165 /*
 166  * Tie new page_cgroup to struct page under lock_page_cgroup()
 167  * This can fail if the page has been tied to a page_cgroup.
 168  * If success, returns 0.
 169  */
 170 static inline int
 171 page_cgroup_assign_new_page_cgroup(struct page *page, struct page_cgroup *pc)
 172 {
 173         int ret = 0;
 174
 175         lock_page_cgroup(page);
 176         if (!page_get_page_cgroup(page))
 177                 page_assign_page_cgroup(page, pc);
 178         else /* A page is tied to other pc. */
 179                 ret = 1;
 180         unlock_page_cgroup(page);
 181         return ret;
 182 }
 183
 184 /*
 185  * Clear page->page_cgroup member under lock_page_cgroup().
 186  * If given "pc" value is different from one page->page_cgroup,
 187  * page->cgroup is not cleared.
 188  * Returns a value of page->page_cgroup at lock taken.
 189  * A can can detect failure of clearing by following
 190  *  clear_page_cgroup(page, pc) == pc
 191  */
 192
 193 static inline struct page_cgroup *
 194 clear_page_cgroup(struct page *page, struct page_cgroup *pc)
 195 {
 196         struct page_cgroup *ret;
 197         /* lock and clear */
 198         lock_page_cgroup(page);
 199         ret = page_get_page_cgroup(page);
 200         if (likely(ret == pc))
 201                 page_assign_page_cgroup(page, NULL);
 202         unlock_page_cgroup(page);
 203         return ret;
 204 }
 205
 206
 207 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 208 {
 209         if (active)
 210                 list_move(&pc->lru, &pc->mem_cgroup->active_list);
 211         else
 212                 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 213 }
 214
 215 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 216 {
 217         int ret;
 218
 219         task_lock(task);
 220         ret = task->mm && mm_cgroup(task->mm) == mem;
 221         task_unlock(task);
 222         return ret;
 223 }
 224
 225 /*
 226  * This routine assumes that the appropriate zone's lru lock is already held
 227  */
 228 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 229 {
 230         struct mem_cgroup *mem;
 231         if (!pc)
 232                 return;
 233
 234         mem = pc->mem_cgroup;
 235
 236         spin_lock(&mem->lru_lock);
 237         __mem_cgroup_move_lists(pc, active);
 238         spin_unlock(&mem->lru_lock);
 239 }
 240
 241 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 242                                         struct list_head *dst,
 243                                         unsigned long *scanned, int order,
 244                                         int mode, struct zone *z,
 245                                         struct mem_cgroup *mem_cont,
 246                                         int active)
 247 {
 248         unsigned long nr_taken = 0;
 249         struct page *page;
 250         unsigned long scan;
 251         LIST_HEAD(pc_list);
 252         struct list_head *src;
 253         struct page_cgroup *pc;
 254
 255         if (active)
 256                 src = &mem_cont->active_list;
 257         else
 258                 src = &mem_cont->inactive_list;
 259
 260         spin_lock(&mem_cont->lru_lock);
 261         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
 262                 pc = list_entry(src->prev, struct page_cgroup, lru);
 263                 page = pc->page;
 264                 VM_BUG_ON(!pc);
 265
 266                 if (PageActive(page) && !active) {
 267                         __mem_cgroup_move_lists(pc, true);
 268                         scan--;
 269                         continue;
 270                 }
 271                 if (!PageActive(page) && active) {
 272                         __mem_cgroup_move_lists(pc, false);
 273                         scan--;
 274                         continue;
 275                 }
 276
 277                 /*
 278                  * Reclaim, per zone
 279                  * TODO: make the active/inactive lists per zone
 280                  */
 281                 if (page_zone(page) != z)
 282                         continue;
 283
 284                 /*
 285                  * Check if the meta page went away from under us
 286                  */
 287                 if (!list_empty(&pc->lru))
 288                         list_move(&pc->lru, &pc_list);
 289                 else
 290                         continue;
 291
 292                 if (__isolate_lru_page(page, mode) == 0) {
 293                         list_move(&page->lru, dst);
 294                         nr_taken++;
 295                 }
 296         }
 297
 298         list_splice(&pc_list, src);
 299         spin_unlock(&mem_cont->lru_lock);
 300
 301         *scanned = scan;
 302         return nr_taken;
 303 }
 304
 305 /*
 306  * Charge the memory controller for page usage.
 307  * Return
 308  * 0 if the charge was successful
 309  * < 0 if the cgroup is over its limit
 310  */
 311 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
 312                                 gfp_t gfp_mask)
 313 {
 314         struct mem_cgroup *mem;
 315         struct page_cgroup *pc;
 316         unsigned long flags;
 317         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 318
 319         /*
 320          * Should page_cgroup's go to their own slab?
 321          * One could optimize the performance of the charging routine
 322          * by saving a bit in the page_flags and using it as a lock
 323          * to see if the cgroup page already has a page_cgroup associated
 324          * with it
 325          */
 326 retry:
 327         lock_page_cgroup(page);
 328         pc = page_get_page_cgroup(page);
 329         /*
 330          * The page_cgroup exists and the page has already been accounted
 331          */
 332         if (pc) {
 333                 if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 334                         /* this page is under being uncharged ? */
 335                         unlock_page_cgroup(page);
 336                         cpu_relax();
 337                         goto retry;
 338                 } else {
 339                         unlock_page_cgroup(page);
 340                         goto done;
 341                 }
 342         }
 343
 344         unlock_page_cgroup(page);
 345
 346         pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
 347         if (pc == NULL)
 348                 goto err;
 349
 350         rcu_read_lock();
 351         /*
 352          * We always charge the cgroup the mm_struct belongs to
 353          * the mm_struct's mem_cgroup changes on task migration if the
 354          * thread group leader migrates. It's possible that mm is not
 355          * set, if so charge the init_mm (happens for pagecache usage).
 356          */
 357         if (!mm)
 358                 mm = &init_mm;
 359
 360         mem = rcu_dereference(mm->mem_cgroup);
 361         /*
 362          * For every charge from the cgroup, increment reference
 363          * count
 364          */
 365         css_get(&mem->css);
 366         rcu_read_unlock();
 367
 368         /*
 369          * If we created the page_cgroup, we should free it on exceeding
 370          * the cgroup limit.
 371          */
 372         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 373                 bool is_atomic = gfp_mask & GFP_ATOMIC;
 374                 /*
 375                  * We cannot reclaim under GFP_ATOMIC, fail the charge
 376                  */
 377                 if (is_atomic)
 378                         goto noreclaim;
 379
 380                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 381                         continue;
 382
 383                 /*
 384                  * try_to_free_mem_cgroup_pages() might not give us a full
 385                  * picture of reclaim. Some pages are reclaimed and might be
 386                  * moved to swap cache or just unmapped from the cgroup.
 387                  * Check the limit again to see if the reclaim reduced the
 388                  * current usage of the cgroup before giving up
 389                  */
 390                 if (res_counter_check_under_limit(&mem->res))
 391                         continue;
 392                         /*
 393                          * Since we control both RSS and cache, we end up with a
 394                          * very interesting scenario where we end up reclaiming
 395                          * memory (essentially RSS), since the memory is pushed
 396                          * to swap cache, we eventually end up adding those
 397                          * pages back to our list. Hence we give ourselves a
 398                          * few chances before we fail
 399                          */
 400                 else if (nr_retries--) {
 401                         congestion_wait(WRITE, HZ/10);
 402                         continue;
 403                 }
 404 noreclaim:
 405                 css_put(&mem->css);
 406                 if (!is_atomic)
 407                         mem_cgroup_out_of_memory(mem, GFP_KERNEL);
 408                 goto free_pc;
 409         }
 410
 411         atomic_set(&pc->ref_cnt, 1);
 412         pc->mem_cgroup = mem;
 413         pc->page = page;
 414         if (page_cgroup_assign_new_page_cgroup(page, pc)) {
 415                 /*
 416                  * an another charge is added to this page already.
 417                  * we do take lock_page_cgroup(page) again and read
 418                  * page->cgroup, increment refcnt.... just retry is OK.
 419                  */
 420                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 421                 css_put(&mem->css);
 422                 kfree(pc);
 423                 goto retry;
 424         }
 425
 426         spin_lock_irqsave(&mem->lru_lock, flags);
 427         list_add(&pc->lru, &mem->active_list);
 428         spin_unlock_irqrestore(&mem->lru_lock, flags);
 429
 430 done:
 431         return 0;
 432 free_pc:
 433         kfree(pc);
 434 err:
 435         return -ENOMEM;
 436 }
 437
 438 /*
 439  * See if the cached pages should be charged at all?
 440  */
 441 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 442                                 gfp_t gfp_mask)
 443 {
 444         struct mem_cgroup *mem;
 445         if (!mm)
 446                 mm = &init_mm;
 447
 448         mem = rcu_dereference(mm->mem_cgroup);
 449         if (mem->control_type == MEM_CGROUP_TYPE_ALL)
 450                 return mem_cgroup_charge(page, mm, gfp_mask);
 451         else
 452                 return 0;
 453 }
 454
 455 /*
 456  * Uncharging is always a welcome operation, we never complain, simply
 457  * uncharge.
 458  */
 459 void mem_cgroup_uncharge(struct page_cgroup *pc)
 460 {
 461         struct mem_cgroup *mem;
 462         struct page *page;
 463         unsigned long flags;
 464
 465         /*
 466          * This can handle cases when a page is not charged at all and we
 467          * are switching between handling the control_type.
 468          */
 469         if (!pc)
 470                 return;
 471
 472         if (atomic_dec_and_test(&pc->ref_cnt)) {
 473                 page = pc->page;
 474                 /*
 475                  * get page->cgroup and clear it under lock.
 476                  */
 477                 if (clear_page_cgroup(page, pc) == pc) {
 478                         mem = pc->mem_cgroup;
 479                         css_put(&mem->css);
 480                         res_counter_uncharge(&mem->res, PAGE_SIZE);
 481                         spin_lock_irqsave(&mem->lru_lock, flags);
 482                         list_del_init(&pc->lru);
 483                         spin_unlock_irqrestore(&mem->lru_lock, flags);
 484                         kfree(pc);
 485                 } else {
 486                         /*
 487                          * Note:This will be removed when force-empty patch is
 488                          * applied. just show warning here.
 489                          */
 490                         printk(KERN_ERR "Race in mem_cgroup_uncharge() ?");
 491                         dump_stack();
 492                 }
 493         }
 494 }
 495
 496 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 497 {
 498         *tmp = memparse(buf, &buf);
 499         if (*buf != '\0')
 500                 return -EINVAL;
 501
 502         /*
 503          * Round up the value to the closest page size
 504          */
 505         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 506         return 0;
 507 }
 508
 509 static ssize_t mem_cgroup_read(struct cgroup *cont,
 510                         struct cftype *cft, struct file *file,
 511                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 512 {
 513         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 514                                 cft->private, userbuf, nbytes, ppos,
 515                                 NULL);
 516 }
 517
 518 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 519                                 struct file *file, const char __user *userbuf,
 520                                 size_t nbytes, loff_t *ppos)
 521 {
 522         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 523                                 cft->private, userbuf, nbytes, ppos,
 524                                 mem_cgroup_write_strategy);
 525 }
 526
 527 static ssize_t mem_control_type_write(struct cgroup *cont,
 528                         struct cftype *cft, struct file *file,
 529                         const char __user *userbuf,
 530                         size_t nbytes, loff_t *pos)
 531 {
 532         int ret;
 533         char *buf, *end;
 534         unsigned long tmp;
 535         struct mem_cgroup *mem;
 536
 537         mem = mem_cgroup_from_cont(cont);
 538         buf = kmalloc(nbytes + 1, GFP_KERNEL);
 539         ret = -ENOMEM;
 540         if (buf == NULL)
 541                 goto out;
 542
 543         buf[nbytes] = 0;
 544         ret = -EFAULT;
 545         if (copy_from_user(buf, userbuf, nbytes))
 546                 goto out_free;
 547
 548         ret = -EINVAL;
 549         tmp = simple_strtoul(buf, &end, 10);
 550         if (*end != '\0')
 551                 goto out_free;
 552
 553         if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
 554                 goto out_free;
 555
 556         mem->control_type = tmp;
 557         ret = nbytes;
 558 out_free:
 559         kfree(buf);
 560 out:
 561         return ret;
 562 }
 563
 564 static ssize_t mem_control_type_read(struct cgroup *cont,
 565                                 struct cftype *cft,
 566                                 struct file *file, char __user *userbuf,
 567                                 size_t nbytes, loff_t *ppos)
 568 {
 569         unsigned long val;
 570         char buf[64], *s;
 571         struct mem_cgroup *mem;
 572
 573         mem = mem_cgroup_from_cont(cont);
 574         s = buf;
 575         val = mem->control_type;
 576         s += sprintf(s, "%lu\n", val);
 577         return simple_read_from_buffer((void __user *)userbuf, nbytes,
 578                         ppos, buf, s - buf);
 579 }
 580
 581 static struct cftype mem_cgroup_files[] = {
 582         {
 583                 .name = "usage_in_bytes",
 584                 .private = RES_USAGE,
 585                 .read = mem_cgroup_read,
 586         },
 587         {
 588                 .name = "limit_in_bytes",
 589                 .private = RES_LIMIT,
 590                 .write = mem_cgroup_write,
 591                 .read = mem_cgroup_read,
 592         },
 593         {
 594                 .name = "failcnt",
 595                 .private = RES_FAILCNT,
 596                 .read = mem_cgroup_read,
 597         },
 598         {
 599                 .name = "control_type",
 600                 .write = mem_control_type_write,
 601                 .read = mem_control_type_read,
 602         },
 603 };
 604
 605 static struct mem_cgroup init_mem_cgroup;
 606
 607 static struct cgroup_subsys_state *
 608 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 609 {
 610         struct mem_cgroup *mem;
 611
 612         if (unlikely((cont->parent) == NULL)) {
 613                 mem = &init_mem_cgroup;
 614                 init_mm.mem_cgroup = mem;
 615         } else
 616                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
 617
 618         if (mem == NULL)
 619                 return NULL;
 620
 621         res_counter_init(&mem->res);
 622         INIT_LIST_HEAD(&mem->active_list);
 623         INIT_LIST_HEAD(&mem->inactive_list);
 624         spin_lock_init(&mem->lru_lock);
 625         mem->control_type = MEM_CGROUP_TYPE_ALL;
 626         return &mem->css;
 627 }
 628
 629 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
 630                                 struct cgroup *cont)
 631 {
 632         kfree(mem_cgroup_from_cont(cont));
 633 }
 634
 635 static int mem_cgroup_populate(struct cgroup_subsys *ss,
 636                                 struct cgroup *cont)
 637 {
 638         return cgroup_add_files(cont, ss, mem_cgroup_files,
 639                                         ARRAY_SIZE(mem_cgroup_files));
 640 }
 641
 642 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
 643                                 struct cgroup *cont,
 644                                 struct cgroup *old_cont,
 645                                 struct task_struct *p)
 646 {
 647         struct mm_struct *mm;
 648         struct mem_cgroup *mem, *old_mem;
 649
 650         mm = get_task_mm(p);
 651         if (mm == NULL)
 652                 return;
 653
 654         mem = mem_cgroup_from_cont(cont);
 655         old_mem = mem_cgroup_from_cont(old_cont);
 656
 657         if (mem == old_mem)
 658                 goto out;
 659
 660         /*
 661          * Only thread group leaders are allowed to migrate, the mm_struct is
 662          * in effect owned by the leader
 663          */
 664         if (p->tgid != p->pid)
 665                 goto out;
 666
 667         css_get(&mem->css);
 668         rcu_assign_pointer(mm->mem_cgroup, mem);
 669         css_put(&old_mem->css);
 670
 671 out:
 672         mmput(mm);
 673         return;
 674 }
 675
 676 struct cgroup_subsys mem_cgroup_subsys = {
 677         .name = "memory",
 678         .subsys_id = mem_cgroup_subsys_id,
 679         .create = mem_cgroup_create,
 680         .destroy = mem_cgroup_destroy,
 681         .populate = mem_cgroup_populate,
 682         .attach = mem_cgroup_move_task,
 683         .early_init = 1,
 684 };