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 inline struct mem_cgroup *mm_cgroup(struct mm_struct *mm)
 114 {
 115         return rcu_dereference(mm->mem_cgroup);
 116 }
 117
 118 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
 119 {
 120         struct mem_cgroup *mem;
 121
 122         mem = mem_cgroup_from_task(p);
 123         css_get(&mem->css);
 124         mm->mem_cgroup = mem;
 125 }
 126
 127 void mm_free_cgroup(struct mm_struct *mm)
 128 {
 129         css_put(&mm->mem_cgroup->css);
 130 }
 131
 132 static inline int page_cgroup_locked(struct page *page)
 133 {
 134         return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
 135                                         &page->page_cgroup);
 136 }
 137
 138 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
 139 {
 140         int locked;
 141
 142         /*
 143          * While resetting the page_cgroup we might not hold the
 144          * page_cgroup lock. free_hot_cold_page() is an example
 145          * of such a scenario
 146          */
 147         if (pc)
 148                 VM_BUG_ON(!page_cgroup_locked(page));
 149         locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
 150         page->page_cgroup = ((unsigned long)pc | locked);
 151 }
 152
 153 struct page_cgroup *page_get_page_cgroup(struct page *page)
 154 {
 155         return (struct page_cgroup *)
 156                 (page->page_cgroup & ~PAGE_CGROUP_LOCK);
 157 }
 158
 159 static void __always_inline lock_page_cgroup(struct page *page)
 160 {
 161         bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 162         VM_BUG_ON(!page_cgroup_locked(page));
 163 }
 164
 165 static void __always_inline unlock_page_cgroup(struct page *page)
 166 {
 167         bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
 168 }
 169
 170 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 171 {
 172         if (active)
 173                 list_move(&pc->lru, &pc->mem_cgroup->active_list);
 174         else
 175                 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
 176 }
 177
 178 /*
 179  * This routine assumes that the appropriate zone's lru lock is already held
 180  */
 181 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
 182 {
 183         struct mem_cgroup *mem;
 184         if (!pc)
 185                 return;
 186
 187         mem = pc->mem_cgroup;
 188
 189         spin_lock(&mem->lru_lock);
 190         __mem_cgroup_move_lists(pc, active);
 191         spin_unlock(&mem->lru_lock);
 192 }
 193
 194 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
 195                                         struct list_head *dst,
 196                                         unsigned long *scanned, int order,
 197                                         int mode, struct zone *z,
 198                                         struct mem_cgroup *mem_cont,
 199                                         int active)
 200 {
 201         unsigned long nr_taken = 0;
 202         struct page *page;
 203         unsigned long scan;
 204         LIST_HEAD(pc_list);
 205         struct list_head *src;
 206         struct page_cgroup *pc;
 207
 208         if (active)
 209                 src = &mem_cont->active_list;
 210         else
 211                 src = &mem_cont->inactive_list;
 212
 213         spin_lock(&mem_cont->lru_lock);
 214         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
 215                 pc = list_entry(src->prev, struct page_cgroup, lru);
 216                 page = pc->page;
 217                 VM_BUG_ON(!pc);
 218
 219                 if (PageActive(page) && !active) {
 220                         __mem_cgroup_move_lists(pc, true);
 221                         scan--;
 222                         continue;
 223                 }
 224                 if (!PageActive(page) && active) {
 225                         __mem_cgroup_move_lists(pc, false);
 226                         scan--;
 227                         continue;
 228                 }
 229
 230                 /*
 231                  * Reclaim, per zone
 232                  * TODO: make the active/inactive lists per zone
 233                  */
 234                 if (page_zone(page) != z)
 235                         continue;
 236
 237                 /*
 238                  * Check if the meta page went away from under us
 239                  */
 240                 if (!list_empty(&pc->lru))
 241                         list_move(&pc->lru, &pc_list);
 242                 else
 243                         continue;
 244
 245                 if (__isolate_lru_page(page, mode) == 0) {
 246                         list_move(&page->lru, dst);
 247                         nr_taken++;
 248                 }
 249         }
 250
 251         list_splice(&pc_list, src);
 252         spin_unlock(&mem_cont->lru_lock);
 253
 254         *scanned = scan;
 255         return nr_taken;
 256 }
 257
 258 /*
 259  * Charge the memory controller for page usage.
 260  * Return
 261  * 0 if the charge was successful
 262  * < 0 if the cgroup is over its limit
 263  */
 264 int mem_cgroup_charge(struct page *page, struct mm_struct *mm)
 265 {
 266         struct mem_cgroup *mem;
 267         struct page_cgroup *pc, *race_pc;
 268         unsigned long flags;
 269         unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
 270
 271         /*
 272          * Should page_cgroup's go to their own slab?
 273          * One could optimize the performance of the charging routine
 274          * by saving a bit in the page_flags and using it as a lock
 275          * to see if the cgroup page already has a page_cgroup associated
 276          * with it
 277          */
 278 retry:
 279         lock_page_cgroup(page);
 280         pc = page_get_page_cgroup(page);
 281         /*
 282          * The page_cgroup exists and the page has already been accounted
 283          */
 284         if (pc) {
 285                 if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
 286                         /* this page is under being uncharged ? */
 287                         unlock_page_cgroup(page);
 288                         cpu_relax();
 289                         goto retry;
 290                 } else
 291                         goto done;
 292         }
 293
 294         unlock_page_cgroup(page);
 295
 296         pc = kzalloc(sizeof(struct page_cgroup), GFP_KERNEL);
 297         if (pc == NULL)
 298                 goto err;
 299
 300         rcu_read_lock();
 301         /*
 302          * We always charge the cgroup the mm_struct belongs to
 303          * the mm_struct's mem_cgroup changes on task migration if the
 304          * thread group leader migrates. It's possible that mm is not
 305          * set, if so charge the init_mm (happens for pagecache usage).
 306          */
 307         if (!mm)
 308                 mm = &init_mm;
 309
 310         mem = rcu_dereference(mm->mem_cgroup);
 311         /*
 312          * For every charge from the cgroup, increment reference
 313          * count
 314          */
 315         css_get(&mem->css);
 316         rcu_read_unlock();
 317
 318         /*
 319          * If we created the page_cgroup, we should free it on exceeding
 320          * the cgroup limit.
 321          */
 322         while (res_counter_charge(&mem->res, PAGE_SIZE)) {
 323                 if (try_to_free_mem_cgroup_pages(mem))
 324                         continue;
 325
 326                 /*
 327                  * try_to_free_mem_cgroup_pages() might not give us a full
 328                  * picture of reclaim. Some pages are reclaimed and might be
 329                  * moved to swap cache or just unmapped from the cgroup.
 330                  * Check the limit again to see if the reclaim reduced the
 331                  * current usage of the cgroup before giving up
 332                  */
 333                 if (res_counter_check_under_limit(&mem->res))
 334                         continue;
 335                         /*
 336                          * Since we control both RSS and cache, we end up with a
 337                          * very interesting scenario where we end up reclaiming
 338                          * memory (essentially RSS), since the memory is pushed
 339                          * to swap cache, we eventually end up adding those
 340                          * pages back to our list. Hence we give ourselves a
 341                          * few chances before we fail
 342                          */
 343                 else if (nr_retries--) {
 344                         congestion_wait(WRITE, HZ/10);
 345                         continue;
 346                 }
 347
 348                 css_put(&mem->css);
 349                 mem_cgroup_out_of_memory(mem, GFP_KERNEL);
 350                 goto free_pc;
 351         }
 352
 353         lock_page_cgroup(page);
 354         /*
 355          * Check if somebody else beat us to allocating the page_cgroup
 356          */
 357         race_pc = page_get_page_cgroup(page);
 358         if (race_pc) {
 359                 kfree(pc);
 360                 pc = race_pc;
 361                 atomic_inc(&pc->ref_cnt);
 362                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 363                 css_put(&mem->css);
 364                 goto done;
 365         }
 366
 367         atomic_set(&pc->ref_cnt, 1);
 368         pc->mem_cgroup = mem;
 369         pc->page = page;
 370         page_assign_page_cgroup(page, pc);
 371
 372         spin_lock_irqsave(&mem->lru_lock, flags);
 373         list_add(&pc->lru, &mem->active_list);
 374         spin_unlock_irqrestore(&mem->lru_lock, flags);
 375
 376 done:
 377         unlock_page_cgroup(page);
 378         return 0;
 379 free_pc:
 380         kfree(pc);
 381 err:
 382         return -ENOMEM;
 383 }
 384
 385 /*
 386  * See if the cached pages should be charged at all?
 387  */
 388 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm)
 389 {
 390         struct mem_cgroup *mem;
 391         if (!mm)
 392                 mm = &init_mm;
 393
 394         mem = rcu_dereference(mm->mem_cgroup);
 395         if (mem->control_type == MEM_CGROUP_TYPE_ALL)
 396                 return mem_cgroup_charge(page, mm);
 397         else
 398                 return 0;
 399 }
 400
 401 /*
 402  * Uncharging is always a welcome operation, we never complain, simply
 403  * uncharge.
 404  */
 405 void mem_cgroup_uncharge(struct page_cgroup *pc)
 406 {
 407         struct mem_cgroup *mem;
 408         struct page *page;
 409         unsigned long flags;
 410
 411         /*
 412          * This can handle cases when a page is not charged at all and we
 413          * are switching between handling the control_type.
 414          */
 415         if (!pc)
 416                 return;
 417
 418         if (atomic_dec_and_test(&pc->ref_cnt)) {
 419                 page = pc->page;
 420                 lock_page_cgroup(page);
 421                 mem = pc->mem_cgroup;
 422                 css_put(&mem->css);
 423                 page_assign_page_cgroup(page, NULL);
 424                 unlock_page_cgroup(page);
 425                 res_counter_uncharge(&mem->res, PAGE_SIZE);
 426
 427                 spin_lock_irqsave(&mem->lru_lock, flags);
 428                 list_del_init(&pc->lru);
 429                 spin_unlock_irqrestore(&mem->lru_lock, flags);
 430                 kfree(pc);
 431         }
 432 }
 433
 434 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 435 {
 436         *tmp = memparse(buf, &buf);
 437         if (*buf != '\0')
 438                 return -EINVAL;
 439
 440         /*
 441          * Round up the value to the closest page size
 442          */
 443         *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
 444         return 0;
 445 }
 446
 447 static ssize_t mem_cgroup_read(struct cgroup *cont,
 448                         struct cftype *cft, struct file *file,
 449                         char __user *userbuf, size_t nbytes, loff_t *ppos)
 450 {
 451         return res_counter_read(&mem_cgroup_from_cont(cont)->res,
 452                                 cft->private, userbuf, nbytes, ppos,
 453                                 NULL);
 454 }
 455
 456 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 457                                 struct file *file, const char __user *userbuf,
 458                                 size_t nbytes, loff_t *ppos)
 459 {
 460         return res_counter_write(&mem_cgroup_from_cont(cont)->res,
 461                                 cft->private, userbuf, nbytes, ppos,
 462                                 mem_cgroup_write_strategy);
 463 }
 464
 465 static ssize_t mem_control_type_write(struct cgroup *cont,
 466                         struct cftype *cft, struct file *file,
 467                         const char __user *userbuf,
 468                         size_t nbytes, loff_t *pos)
 469 {
 470         int ret;
 471         char *buf, *end;
 472         unsigned long tmp;
 473         struct mem_cgroup *mem;
 474
 475         mem = mem_cgroup_from_cont(cont);
 476         buf = kmalloc(nbytes + 1, GFP_KERNEL);
 477         ret = -ENOMEM;
 478         if (buf == NULL)
 479                 goto out;
 480
 481         buf[nbytes] = 0;
 482         ret = -EFAULT;
 483         if (copy_from_user(buf, userbuf, nbytes))
 484                 goto out_free;
 485
 486         ret = -EINVAL;
 487         tmp = simple_strtoul(buf, &end, 10);
 488         if (*end != '\0')
 489                 goto out_free;
 490
 491         if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
 492                 goto out_free;
 493
 494         mem->control_type = tmp;
 495         ret = nbytes;
 496 out_free:
 497         kfree(buf);
 498 out:
 499         return ret;
 500 }
 501
 502 static ssize_t mem_control_type_read(struct cgroup *cont,
 503                                 struct cftype *cft,
 504                                 struct file *file, char __user *userbuf,
 505                                 size_t nbytes, loff_t *ppos)
 506 {
 507         unsigned long val;
 508         char buf[64], *s;
 509         struct mem_cgroup *mem;
 510
 511         mem = mem_cgroup_from_cont(cont);
 512         s = buf;
 513         val = mem->control_type;
 514         s += sprintf(s, "%lu\n", val);
 515         return simple_read_from_buffer((void __user *)userbuf, nbytes,
 516                         ppos, buf, s - buf);
 517 }
 518
 519 static struct cftype mem_cgroup_files[] = {
 520         {
 521                 .name = "usage_in_bytes",
 522                 .private = RES_USAGE,
 523                 .read = mem_cgroup_read,
 524         },
 525         {
 526                 .name = "limit_in_bytes",
 527                 .private = RES_LIMIT,
 528                 .write = mem_cgroup_write,
 529                 .read = mem_cgroup_read,
 530         },
 531         {
 532                 .name = "failcnt",
 533                 .private = RES_FAILCNT,
 534                 .read = mem_cgroup_read,
 535         },
 536         {
 537                 .name = "control_type",
 538                 .write = mem_control_type_write,
 539                 .read = mem_control_type_read,
 540         },
 541 };
 542
 543 static struct mem_cgroup init_mem_cgroup;
 544
 545 static struct cgroup_subsys_state *
 546 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
 547 {
 548         struct mem_cgroup *mem;
 549
 550         if (unlikely((cont->parent) == NULL)) {
 551                 mem = &init_mem_cgroup;
 552                 init_mm.mem_cgroup = mem;
 553         } else
 554                 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
 555
 556         if (mem == NULL)
 557                 return NULL;
 558
 559         res_counter_init(&mem->res);
 560         INIT_LIST_HEAD(&mem->active_list);
 561         INIT_LIST_HEAD(&mem->inactive_list);
 562         spin_lock_init(&mem->lru_lock);
 563         mem->control_type = MEM_CGROUP_TYPE_ALL;
 564         return &mem->css;
 565 }
 566
 567 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
 568                                 struct cgroup *cont)
 569 {
 570         kfree(mem_cgroup_from_cont(cont));
 571 }
 572
 573 static int mem_cgroup_populate(struct cgroup_subsys *ss,
 574                                 struct cgroup *cont)
 575 {
 576         return cgroup_add_files(cont, ss, mem_cgroup_files,
 577                                         ARRAY_SIZE(mem_cgroup_files));
 578 }
 579
 580 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
 581                                 struct cgroup *cont,
 582                                 struct cgroup *old_cont,
 583                                 struct task_struct *p)
 584 {
 585         struct mm_struct *mm;
 586         struct mem_cgroup *mem, *old_mem;
 587
 588         mm = get_task_mm(p);
 589         if (mm == NULL)
 590                 return;
 591
 592         mem = mem_cgroup_from_cont(cont);
 593         old_mem = mem_cgroup_from_cont(old_cont);
 594
 595         if (mem == old_mem)
 596                 goto out;
 597
 598         /*
 599          * Only thread group leaders are allowed to migrate, the mm_struct is
 600          * in effect owned by the leader
 601          */
 602         if (p->tgid != p->pid)
 603                 goto out;
 604
 605         css_get(&mem->css);
 606         rcu_assign_pointer(mm->mem_cgroup, mem);
 607         css_put(&old_mem->css);
 608
 609 out:
 610         mmput(mm);
 611         return;
 612 }
 613
 614 struct cgroup_subsys mem_cgroup_subsys = {
 615         .name = "memory",
 616         .subsys_id = mem_cgroup_subsys_id,
 617         .create = mem_cgroup_create,
 618         .destroy = mem_cgroup_destroy,
 619         .populate = mem_cgroup_populate,
 620         .attach = mem_cgroup_move_task,
 621         .early_init = 1,
 622 };