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