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 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 166 {
 167         if (active)
 168                 list_move(&pc->lru, &pc->mem_cgroup->active_list);
 169         else
 170                 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 171 }
 172
 173 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
 174 {
 175         int ret;
 176
 177         task_lock(task);
 178         ret = task->mm && mm_cgroup(task->mm) == mem;
 179         task_unlock(task);
 180         return ret;
 181 }
 182
 183 /*
 184  * This routine assumes that the appropriate zone's lru lock is already held
 185  */
 186 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 187 {
 188         struct mem_cgroup *mem;
 189         if (!pc)
 190                 return;
 191
 192         mem = pc->mem_cgroup;
 193
 194         spin_lock(&mem->lru_lock);
 195         __mem_cgroup_move_lists(pc, active);
 196         spin_unlock(&mem->lru_lock);
 197 }
 198
 199 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 200                                         struct list_head *dst,
 201                                         unsigned long *scanned, int order,
 202                                         int mode, struct zone *z,
 203                                         struct mem_cgroup *mem_cont,
 204                                         int active)
 205 {
 206         unsigned long nr_taken = 0;
 207         struct page *page;
 208         unsigned long scan;
 209         LIST_HEAD(pc_list);
 210         struct list_head *src;
 211         struct page_cgroup *pc;
 212
 213         if (active)
 214                 src = &mem_cont->active_list;
 215         else
 216                 src = &mem_cont->inactive_list;
 217
 218         spin_lock(&mem_cont->lru_lock);
 219         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
 220                 pc = list_entry(src->prev, struct page_cgroup, lru);
 221                 page = pc->page;
 222                 VM_BUG_ON(!pc);
 223
 224                 if (PageActive(page) && !active) {
 225                         __mem_cgroup_move_lists(pc, true);
 226                         scan--;
 227                         continue;
 228                 }
 229                 if (!PageActive(page) && active) {
 230                         __mem_cgroup_move_lists(pc, false);
 231                         scan--;
 232                         continue;
 233                 }
 234
 235                 /*
 236                  * Reclaim, per zone
 237                  * TODO: make the active/inactive lists per zone
 238                  */
 239                 if (page_zone(page) != z)
 240                         continue;
 241
 242                 /*
 243                  * Check if the meta page went away from under us
 244                  */
 245                 if (!list_empty(&pc->lru))
 246                         list_move(&pc->lru, &pc_list);
 247                 else
 248                         continue;
 249
 250                 if (__isolate_lru_page(page, mode) == 0) {
 251                         list_move(&page->lru, dst);
 252                         nr_taken++;
 253                 }
 254         }
 255
 256         list_splice(&pc_list, src);
 257         spin_unlock(&mem_cont->lru_lock);
 258
 259         *scanned = scan;
 260         return nr_taken;
 261 }
 262
 263 /*
 264  * Charge the memory controller for page usage.
 265  * Return
 266  * 0 if the charge was successful
 267  * < 0 if the cgroup is over its limit
 268  */
 269 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
 270                                 gfp_t gfp_mask)
 271 {
 272         struct mem_cgroup *mem;
 273         struct page_cgroup *pc, *race_pc;
 274         unsigned long flags;
 275         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 276
 277         /*
 278          * Should page_cgroup's go to their own slab?
 279          * One could optimize the performance of the charging routine
 280          * by saving a bit in the page_flags and using it as a lock
 281          * to see if the cgroup page already has a page_cgroup associated
 282          * with it
 283          */
 284 retry:
 285         lock_page_cgroup(page);
 286         pc = page_get_page_cgroup(page);
 287         /*
 288          * The page_cgroup exists and the page has already been accounted
 289          */
 290         if (pc) {
 291                 if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 292                         /* this page is under being uncharged ? */
 293                         unlock_page_cgroup(page);
 294                         cpu_relax();
 295                         goto retry;
 296                 } else
 297                         goto done;
 298         }
 299
 300         unlock_page_cgroup(page);
 301
 302         pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
 303         if (pc == NULL)
 304                 goto err;
 305
 306         rcu_read_lock();
 307         /*
 308          * We always charge the cgroup the mm_struct belongs to
 309          * the mm_struct's mem_cgroup changes on task migration if the
 310          * thread group leader migrates. It's possible that mm is not
 311          * set, if so charge the init_mm (happens for pagecache usage).
 312          */
 313         if (!mm)
 314                 mm = &init_mm;
 315
 316         mem = rcu_dereference(mm->mem_cgroup);
 317         /*
 318          * For every charge from the cgroup, increment reference
 319          * count
 320          */
 321         css_get(&mem->css);
 322         rcu_read_unlock();
 323
 324         /*
 325          * If we created the page_cgroup, we should free it on exceeding
 326          * the cgroup limit.
 327          */
 328         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 329                 bool is_atomic = gfp_mask & GFP_ATOMIC;
 330                 /*
 331                  * We cannot reclaim under GFP_ATOMIC, fail the charge
 332                  */
 333                 if (is_atomic)
 334                         goto noreclaim;
 335
 336                 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
 337                         continue;
 338
 339                 /*
 340                  * try_to_free_mem_cgroup_pages() might not give us a full
 341                  * picture of reclaim. Some pages are reclaimed and might be
 342                  * moved to swap cache or just unmapped from the cgroup.
 343                  * Check the limit again to see if the reclaim reduced the
 344                  * current usage of the cgroup before giving up
 345                  */
 346                 if (res_counter_check_under_limit(&mem->res))
 347                         continue;
 348                         /*
 349                          * Since we control both RSS and cache, we end up with a
 350                          * very interesting scenario where we end up reclaiming
 351                          * memory (essentially RSS), since the memory is pushed
 352                          * to swap cache, we eventually end up adding those
 353                          * pages back to our list. Hence we give ourselves a
 354                          * few chances before we fail
 355                          */
 356                 else if (nr_retries--) {
 357                         congestion_wait(WRITE, HZ/10);
 358                         continue;
 359                 }
 360 noreclaim:
 361                 css_put(&mem->css);
 362                 if (!is_atomic)
 363                         mem_cgroup_out_of_memory(mem, GFP_KERNEL);
 364                 goto free_pc;
 365         }
 366
 367         lock_page_cgroup(page);
 368         /*
 369          * Check if somebody else beat us to allocating the page_cgroup
 370          */
 371         race_pc = page_get_page_cgroup(page);
 372         if (race_pc) {
 373                 kfree(pc);
 374                 pc = race_pc;
 375                 atomic_inc(&pc->ref_cnt);
 376                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 377                 css_put(&mem->css);
 378                 goto done;
 379         }
 380
 381         atomic_set(&pc->ref_cnt, 1);
 382         pc->mem_cgroup = mem;
 383         pc->page = page;
 384         page_assign_page_cgroup(page, pc);
 385
 386         spin_lock_irqsave(&mem->lru_lock, flags);
 387         list_add(&pc->lru, &mem->active_list);
 388         spin_unlock_irqrestore(&mem->lru_lock, flags);
 389
 390 done:
 391         unlock_page_cgroup(page);
 392         return 0;
 393 free_pc:
 394         kfree(pc);
 395 err:
 396         return -ENOMEM;
 397 }
 398
 399 /*
 400  * See if the cached pages should be charged at all?
 401  */
 402 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
 403                                 gfp_t gfp_mask)
 404 {
 405         struct mem_cgroup *mem;
 406         if (!mm)
 407                 mm = &init_mm;
 408
 409         mem = rcu_dereference(mm->mem_cgroup);
 410         if (mem->control_type == MEM_CGROUP_TYPE_ALL)
 411                 return mem_cgroup_charge(page, mm, gfp_mask);
 412         else
 413                 return 0;
 414 }
 415
 416 /*
 417  * Uncharging is always a welcome operation, we never complain, simply
 418  * uncharge.
 419  */
 420 void mem_cgroup_uncharge(struct page_cgroup *pc)
 421 {
 422         struct mem_cgroup *mem;
 423         struct page *page;
 424         unsigned long flags;
 425
 426         /*
 427          * This can handle cases when a page is not charged at all and we
 428          * are switching between handling the control_type.
 429          */
 430         if (!pc)
 431                 return;
 432
 433         if (atomic_dec_and_test(&pc->ref_cnt)) {
 434                 page = pc->page;
 435                 lock_page_cgroup(page);
 436                 mem = pc->mem_cgroup;
 437                 css_put(&mem->css);
 438                 page_assign_page_cgroup(page, NULL);
 439                 unlock_page_cgroup(page);
 440                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 441
 442                 spin_lock_irqsave(&mem->lru_lock, flags);
 443                 list_del_init(&pc->lru);
 444                 spin_unlock_irqrestore(&mem->lru_lock, flags);
 445                 kfree(pc);
 446         }
 447 }
 448
 449 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 450 {
 451         *tmp = memparse(buf, &buf);
 452         if (*buf != '\0')
 453                 return -EINVAL;
 454
 455         /*
 456          * Round up the value to the closest page size
 457          */
 458         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 459         return 0;
 460 }
 461
 462 static ssize_t mem_cgroup_read(struct cgroup *cont,
 463                         struct cftype *cft, struct file *file,
 464                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 465 {
 466         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 467                                 cft->private, userbuf, nbytes, ppos,
 468                                 NULL);
 469 }
 470
 471 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 472                                 struct file *file, const char __user *userbuf,
 473                                 size_t nbytes, loff_t *ppos)
 474 {
 475         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 476                                 cft->private, userbuf, nbytes, ppos,
 477                                 mem_cgroup_write_strategy);
 478 }
 479
 480 static ssize_t mem_control_type_write(struct cgroup *cont,
 481                         struct cftype *cft, struct file *file,
 482                         const char __user *userbuf,
 483                         size_t nbytes, loff_t *pos)
 484 {
 485         int ret;
 486         char *buf, *end;
 487         unsigned long tmp;
 488         struct mem_cgroup *mem;
 489
 490         mem = mem_cgroup_from_cont(cont);
 491         buf = kmalloc(nbytes + 1, GFP_KERNEL);
 492         ret = -ENOMEM;
 493         if (buf == NULL)
 494                 goto out;
 495
 496         buf[nbytes] = 0;
 497         ret = -EFAULT;
 498         if (copy_from_user(buf, userbuf, nbytes))
 499                 goto out_free;
 500
 501         ret = -EINVAL;
 502         tmp = simple_strtoul(buf, &end, 10);
 503         if (*end != '\0')
 504                 goto out_free;
 505
 506         if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
 507                 goto out_free;
 508
 509         mem->control_type = tmp;
 510         ret = nbytes;
 511 out_free:
 512         kfree(buf);
 513 out:
 514         return ret;
 515 }
 516
 517 static ssize_t mem_control_type_read(struct cgroup *cont,
 518                                 struct cftype *cft,
 519                                 struct file *file, char __user *userbuf,
 520                                 size_t nbytes, loff_t *ppos)
 521 {
 522         unsigned long val;
 523         char buf[64], *s;
 524         struct mem_cgroup *mem;
 525
 526         mem = mem_cgroup_from_cont(cont);
 527         s = buf;
 528         val = mem->control_type;
 529         s += sprintf(s, "%lu\n", val);
 530         return simple_read_from_buffer((void __user *)userbuf, nbytes,
 531                         ppos, buf, s - buf);
 532 }
 533
 534 static struct cftype mem_cgroup_files[] = {
 535         {
 536                 .name = "usage_in_bytes",
 537                 .private = RES_USAGE,
 538                 .read = mem_cgroup_read,
 539         },
 540         {
 541                 .name = "limit_in_bytes",
 542                 .private = RES_LIMIT,
 543                 .write = mem_cgroup_write,
 544                 .read = mem_cgroup_read,
 545         },
 546         {
 547                 .name = "failcnt",
 548                 .private = RES_FAILCNT,
 549                 .read = mem_cgroup_read,
 550         },
 551         {
 552                 .name = "control_type",
 553                 .write = mem_control_type_write,
 554                 .read = mem_control_type_read,
 555         },
 556 };
 557
 558 static struct mem_cgroup init_mem_cgroup;
 559
 560 static struct cgroup_subsys_state *
 561 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 562 {
 563         struct mem_cgroup *mem;
 564
 565         if (unlikely((cont->parent) == NULL)) {
 566                 mem = &init_mem_cgroup;
 567                 init_mm.mem_cgroup = mem;
 568         } else
 569                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
 570
 571         if (mem == NULL)
 572                 return NULL;
 573
 574         res_counter_init(&mem->res);
 575         INIT_LIST_HEAD(&mem->active_list);
 576         INIT_LIST_HEAD(&mem->inactive_list);
 577         spin_lock_init(&mem->lru_lock);
 578         mem->control_type = MEM_CGROUP_TYPE_ALL;
 579         return &mem->css;
 580 }
 581
 582 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
 583                                 struct cgroup *cont)
 584 {
 585         kfree(mem_cgroup_from_cont(cont));
 586 }
 587
 588 static int mem_cgroup_populate(struct cgroup_subsys *ss,
 589                                 struct cgroup *cont)
 590 {
 591         return cgroup_add_files(cont, ss, mem_cgroup_files,
 592                                         ARRAY_SIZE(mem_cgroup_files));
 593 }
 594
 595 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
 596                                 struct cgroup *cont,
 597                                 struct cgroup *old_cont,
 598                                 struct task_struct *p)
 599 {
 600         struct mm_struct *mm;
 601         struct mem_cgroup *mem, *old_mem;
 602
 603         mm = get_task_mm(p);
 604         if (mm == NULL)
 605                 return;
 606
 607         mem = mem_cgroup_from_cont(cont);
 608         old_mem = mem_cgroup_from_cont(old_cont);
 609
 610         if (mem == old_mem)
 611                 goto out;
 612
 613         /*
 614          * Only thread group leaders are allowed to migrate, the mm_struct is
 615          * in effect owned by the leader
 616          */
 617         if (p->tgid != p->pid)
 618                 goto out;
 619
 620         css_get(&mem->css);
 621         rcu_assign_pointer(mm->mem_cgroup, mem);
 622         css_put(&old_mem->css);
 623
 624 out:
 625         mmput(mm);
 626         return;
 627 }
 628
 629 struct cgroup_subsys mem_cgroup_subsys = {
 630         .name = "memory",
 631         .subsys_id = mem_cgroup_subsys_id,
 632         .create = mem_cgroup_create,
 633         .destroy = mem_cgroup_destroy,
 634         .populate = mem_cgroup_populate,
 635         .attach = mem_cgroup_move_task,
 636         .early_init = 1,
 637 };