mm/page_cgroup.c

   1 #include <linux/mm.h>
   2 #include <linux/mmzone.h>
   3 #include <linux/bootmem.h>
   4 #include <linux/bit_spinlock.h>
   5 #include <linux/page_cgroup.h>
   6 #include <linux/hash.h>
   7 #include <linux/slab.h>
   8 #include <linux/memory.h>
   9 #include <linux/vmalloc.h>
  10 #include <linux/cgroup.h>
  11 #include <linux/swapops.h>
  12 #include <linux/kmemleak.h>
  13
  14 static void __meminit init_page_cgroup(struct page_cgroup *pc, unsigned long id)
  15 {
  16         pc->flags = 0;
  17         set_page_cgroup_array_id(pc, id);
  18         pc->mem_cgroup = NULL;
  19         INIT_LIST_HEAD(&pc->lru);
  20 }
  21 static unsigned long total_usage;
  22
  23 #if !defined(CONFIG_SPARSEMEM)
  24
  25
  26 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
  27 {
  28         pgdat->node_page_cgroup = NULL;
  29 }
  30
  31 struct page_cgroup *lookup_page_cgroup(struct page *page)
  32 {
  33         unsigned long pfn = page_to_pfn(page);
  34         unsigned long offset;
  35         struct page_cgroup *base;
  36
  37         base = NODE_DATA(page_to_nid(page))->node_page_cgroup;
  38         if (unlikely(!base))
  39                 return NULL;
  40
  41         offset = pfn - NODE_DATA(page_to_nid(page))->node_start_pfn;
  42         return base + offset;
  43 }
  44
  45 struct page *lookup_cgroup_page(struct page_cgroup *pc)
  46 {
  47         unsigned long pfn;
  48         struct page *page;
  49         pg_data_t *pgdat;
  50
  51         pgdat = NODE_DATA(page_cgroup_array_id(pc));
  52         pfn = pc - pgdat->node_page_cgroup + pgdat->node_start_pfn;
  53         page = pfn_to_page(pfn);
  54         VM_BUG_ON(pc != lookup_page_cgroup(page));
  55         return page;
  56 }
  57
  58 static int __init alloc_node_page_cgroup(int nid)
  59 {
  60         struct page_cgroup *base, *pc;
  61         unsigned long table_size;
  62         unsigned long start_pfn, nr_pages, index;
  63
  64         start_pfn = NODE_DATA(nid)->node_start_pfn;
  65         nr_pages = NODE_DATA(nid)->node_spanned_pages;
  66
  67         if (!nr_pages)
  68                 return 0;
  69
  70         table_size = sizeof(struct page_cgroup) * nr_pages;
  71
  72         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
  73                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
  74         if (!base)
  75                 return -ENOMEM;
  76         for (index = 0; index < nr_pages; index++) {
  77                 pc = base + index;
  78                 init_page_cgroup(pc, nid);
  79         }
  80         NODE_DATA(nid)->node_page_cgroup = base;
  81         total_usage += table_size;
  82         return 0;
  83 }
  84
  85 void __init page_cgroup_init_flatmem(void)
  86 {
  87
  88         int nid, fail;
  89
  90         if (mem_cgroup_disabled())
  91                 return;
  92
  93         for_each_online_node(nid)  {
  94                 fail = alloc_node_page_cgroup(nid);
  95                 if (fail)
  96                         goto fail;
  97         }
  98         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
  99         printk(KERN_INFO "please try 'cgroup_disable=memory' option if you"
 100         " don't want memory cgroups\n");
 101         return;
 102 fail:
 103         printk(KERN_CRIT "allocation of page_cgroup failed.\n");
 104         printk(KERN_CRIT "please try 'cgroup_disable=memory' boot option\n");
 105         panic("Out of memory");
 106 }
 107
 108 #else /* CONFIG_FLAT_NODE_MEM_MAP */
 109
 110 struct page_cgroup *lookup_page_cgroup(struct page *page)
 111 {
 112         unsigned long pfn = page_to_pfn(page);
 113         struct mem_section *section = __pfn_to_section(pfn);
 114
 115         if (!section->page_cgroup)
 116                 return NULL;
 117         return section->page_cgroup + pfn;
 118 }
 119
 120 struct page *lookup_cgroup_page(struct page_cgroup *pc)
 121 {
 122         struct mem_section *section;
 123         struct page *page;
 124         unsigned long nr;
 125
 126         nr = page_cgroup_array_id(pc);
 127         section = __nr_to_section(nr);
 128         page = pfn_to_page(pc - section->page_cgroup);
 129         VM_BUG_ON(pc != lookup_page_cgroup(page));
 130         return page;
 131 }
 132
 133 static void *__meminit alloc_page_cgroup(size_t size, int nid)
 134 {
 135         void *addr = NULL;
 136         gfp_t flags = GFP_KERNEL | __GFP_NOWARN;
 137
 138         addr = alloc_pages_exact_nid(nid, size, flags);
 139         if (addr) {
 140                 kmemleak_alloc(addr, size, 1, flags);
 141                 return addr;
 142         }
 143
 144         if (node_state(nid, N_HIGH_MEMORY))
 145                 addr = vmalloc_node(size, nid);
 146         else
 147                 addr = vmalloc(size);
 148
 149         return addr;
 150 }
 151
 152 #ifdef CONFIG_MEMORY_HOTPLUG
 153 static void free_page_cgroup(void *addr)
 154 {
 155         if (is_vmalloc_addr(addr)) {
 156                 vfree(addr);
 157         } else {
 158                 struct page *page = virt_to_page(addr);
 159                 size_t table_size =
 160                         sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 161
 162                 BUG_ON(PageReserved(page));
 163                 free_pages_exact(addr, table_size);
 164         }
 165 }
 166 #endif
 167
 168 static int __meminit init_section_page_cgroup(unsigned long pfn, int nid)
 169 {
 170         struct page_cgroup *base, *pc;
 171         struct mem_section *section;
 172         unsigned long table_size;
 173         unsigned long nr;
 174         int index;
 175
 176         nr = pfn_to_section_nr(pfn);
 177         section = __nr_to_section(nr);
 178
 179         if (section->page_cgroup)
 180                 return 0;
 181
 182         table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 183         base = alloc_page_cgroup(table_size, nid);
 184
 185         /*
 186          * The value stored in section->page_cgroup is (base - pfn)
 187          * and it does not point to the memory block allocated above,
 188          * causing kmemleak false positives.
 189          */
 190         kmemleak_not_leak(base);
 191
 192         if (!base) {
 193                 printk(KERN_ERR "page cgroup allocation failure\n");
 194                 return -ENOMEM;
 195         }
 196
 197         for (index = 0; index < PAGES_PER_SECTION; index++) {
 198                 pc = base + index;
 199                 init_page_cgroup(pc, nr);
 200         }
 201         /*
 202          * The passed "pfn" may not be aligned to SECTION.  For the calculation
 203          * we need to apply a mask.
 204          */
 205         pfn &= PAGE_SECTION_MASK;
 206         section->page_cgroup = base - pfn;
 207         total_usage += table_size;
 208         return 0;
 209 }
 210 #ifdef CONFIG_MEMORY_HOTPLUG
 211 void __free_page_cgroup(unsigned long pfn)
 212 {
 213         struct mem_section *ms;
 214         struct page_cgroup *base;
 215
 216         ms = __pfn_to_section(pfn);
 217         if (!ms || !ms->page_cgroup)
 218                 return;
 219         base = ms->page_cgroup + pfn;
 220         free_page_cgroup(base);
 221         ms->page_cgroup = NULL;
 222 }
 223
 224 int __meminit online_page_cgroup(unsigned long start_pfn,
 225                         unsigned long nr_pages,
 226                         int nid)
 227 {
 228         unsigned long start, end, pfn;
 229         int fail = 0;
 230
 231         start = SECTION_ALIGN_DOWN(start_pfn);
 232         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 233
 234         if (nid == -1) {
 235                 /*
 236                  * In this case, "nid" already exists and contains valid memory.
 237                  * "start_pfn" passed to us is a pfn which is an arg for
 238                  * online__pages(), and start_pfn should exist.
 239                  */
 240                 nid = pfn_to_nid(start_pfn);
 241                 VM_BUG_ON(!node_state(nid, N_ONLINE));
 242         }
 243
 244         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
 245                 if (!pfn_present(pfn))
 246                         continue;
 247                 fail = init_section_page_cgroup(pfn, nid);
 248         }
 249         if (!fail)
 250                 return 0;
 251
 252         /* rollback */
 253         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 254                 __free_page_cgroup(pfn);
 255
 256         return -ENOMEM;
 257 }
 258
 259 int __meminit offline_page_cgroup(unsigned long start_pfn,
 260                 unsigned long nr_pages, int nid)
 261 {
 262         unsigned long start, end, pfn;
 263
 264         start = SECTION_ALIGN_DOWN(start_pfn);
 265         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 266
 267         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 268                 __free_page_cgroup(pfn);
 269         return 0;
 270
 271 }
 272
 273 static int __meminit page_cgroup_callback(struct notifier_block *self,
 274                                unsigned long action, void *arg)
 275 {
 276         struct memory_notify *mn = arg;
 277         int ret = 0;
 278         switch (action) {
 279         case MEM_GOING_ONLINE:
 280                 ret = online_page_cgroup(mn->start_pfn,
 281                                    mn->nr_pages, mn->status_change_nid);
 282                 break;
 283         case MEM_OFFLINE:
 284                 offline_page_cgroup(mn->start_pfn,
 285                                 mn->nr_pages, mn->status_change_nid);
 286                 break;
 287         case MEM_CANCEL_ONLINE:
 288         case MEM_GOING_OFFLINE:
 289                 break;
 290         case MEM_ONLINE:
 291         case MEM_CANCEL_OFFLINE:
 292                 break;
 293         }
 294
 295         return notifier_from_errno(ret);
 296 }
 297
 298 #endif
 299
 300 void __init page_cgroup_init(void)
 301 {
 302         unsigned long pfn;
 303         int nid;
 304
 305         if (mem_cgroup_disabled())
 306                 return;
 307
 308         for_each_node_state(nid, N_HIGH_MEMORY) {
 309                 unsigned long start_pfn, end_pfn;
 310
 311                 start_pfn = node_start_pfn(nid);
 312                 end_pfn = node_end_pfn(nid);
 313                 /*
 314                  * start_pfn and end_pfn may not be aligned to SECTION and the
 315                  * page->flags of out of node pages are not initialized.  So we
 316                  * scan [start_pfn, the biggest section's pfn < end_pfn) here.
 317                  */
 318                 for (pfn = start_pfn;
 319                      pfn < end_pfn;
 320                      pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
 321
 322                         if (!pfn_valid(pfn))
 323                                 continue;
 324                         /*
 325                          * Nodes's pfns can be overlapping.
 326                          * We know some arch can have a nodes layout such as
 327                          * -------------pfn-------------->
 328                          * N0 | N1 | N2 | N0 | N1 | N2|....
 329                          */
 330                         if (pfn_to_nid(pfn) != nid)
 331                                 continue;
 332                         if (init_section_page_cgroup(pfn, nid))
 333                                 goto oom;
 334                 }
 335         }
 336         hotplug_memory_notifier(page_cgroup_callback, 0);
 337         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
 338         printk(KERN_INFO "please try 'cgroup_disable=memory' option if you "
 339                          "don't want memory cgroups\n");
 340         return;
 341 oom:
 342         printk(KERN_CRIT "try 'cgroup_disable=memory' boot option\n");
 343         panic("Out of memory");
 344 }
 345
 346 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
 347 {
 348         return;
 349 }
 350
 351 #endif
 352
 353
 354 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 355
 356 static DEFINE_MUTEX(swap_cgroup_mutex);
 357 struct swap_cgroup_ctrl {
 358         struct page **map;
 359         unsigned long length;
 360         spinlock_t      lock;
 361 };
 362
 363 static struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
 364
 365 struct swap_cgroup {
 366         unsigned short          id;
 367 };
 368 #define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
 369 #define SC_POS_MASK     (SC_PER_PAGE - 1)
 370
 371 /*
 372  * SwapCgroup implements "lookup" and "exchange" operations.
 373  * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
 374  * against SwapCache. At swap_free(), this is accessed directly from swap.
 375  *
 376  * This means,
 377  *  - we have no race in "exchange" when we're accessed via SwapCache because
 378  *    SwapCache(and its swp_entry) is under lock.
 379  *  - When called via swap_free(), there is no user of this entry and no race.
 380  * Then, we don't need lock around "exchange".
 381  *
 382  * TODO: we can push these buffers out to HIGHMEM.
 383  */
 384
 385 /*
 386  * allocate buffer for swap_cgroup.
 387  */
 388 static int swap_cgroup_prepare(int type)
 389 {
 390         struct page *page;
 391         struct swap_cgroup_ctrl *ctrl;
 392         unsigned long idx, max;
 393
 394         ctrl = &swap_cgroup_ctrl[type];
 395
 396         for (idx = 0; idx < ctrl->length; idx++) {
 397                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 398                 if (!page)
 399                         goto not_enough_page;
 400                 ctrl->map[idx] = page;
 401         }
 402         return 0;
 403 not_enough_page:
 404         max = idx;
 405         for (idx = 0; idx < max; idx++)
 406                 __free_page(ctrl->map[idx]);
 407
 408         return -ENOMEM;
 409 }
 410
 411 /**
 412  * swap_cgroup_cmpxchg - cmpxchg mem_cgroup's id for this swp_entry.
 413  * @end: swap entry to be cmpxchged
 414  * @old: old id
 415  * @new: new id
 416  *
 417  * Returns old id at success, 0 at failure.
 418  * (There is no mem_cgroup using 0 as its id)
 419  */
 420 unsigned short swap_cgroup_cmpxchg(swp_entry_t ent,
 421                                         unsigned short old, unsigned short new)
 422 {
 423         int type = swp_type(ent);
 424         unsigned long offset = swp_offset(ent);
 425         unsigned long idx = offset / SC_PER_PAGE;
 426         unsigned long pos = offset & SC_POS_MASK;
 427         struct swap_cgroup_ctrl *ctrl;
 428         struct page *mappage;
 429         struct swap_cgroup *sc;
 430         unsigned long flags;
 431         unsigned short retval;
 432
 433         ctrl = &swap_cgroup_ctrl[type];
 434
 435         mappage = ctrl->map[idx];
 436         sc = page_address(mappage);
 437         sc += pos;
 438         spin_lock_irqsave(&ctrl->lock, flags);
 439         retval = sc->id;
 440         if (retval == old)
 441                 sc->id = new;
 442         else
 443                 retval = 0;
 444         spin_unlock_irqrestore(&ctrl->lock, flags);
 445         return retval;
 446 }
 447
 448 /**
 449  * swap_cgroup_record - record mem_cgroup for this swp_entry.
 450  * @ent: swap entry to be recorded into
 451  * @mem: mem_cgroup to be recorded
 452  *
 453  * Returns old value at success, 0 at failure.
 454  * (Of course, old value can be 0.)
 455  */
 456 unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id)
 457 {
 458         int type = swp_type(ent);
 459         unsigned long offset = swp_offset(ent);
 460         unsigned long idx = offset / SC_PER_PAGE;
 461         unsigned long pos = offset & SC_POS_MASK;
 462         struct swap_cgroup_ctrl *ctrl;
 463         struct page *mappage;
 464         struct swap_cgroup *sc;
 465         unsigned short old;
 466         unsigned long flags;
 467
 468         ctrl = &swap_cgroup_ctrl[type];
 469
 470         mappage = ctrl->map[idx];
 471         sc = page_address(mappage);
 472         sc += pos;
 473         spin_lock_irqsave(&ctrl->lock, flags);
 474         old = sc->id;
 475         sc->id = id;
 476         spin_unlock_irqrestore(&ctrl->lock, flags);
 477
 478         return old;
 479 }
 480
 481 /**
 482  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
 483  * @ent: swap entry to be looked up.
 484  *
 485  * Returns CSS ID of mem_cgroup at success. 0 at failure. (0 is invalid ID)
 486  */
 487 unsigned short lookup_swap_cgroup(swp_entry_t ent)
 488 {
 489         int type = swp_type(ent);
 490         unsigned long offset = swp_offset(ent);
 491         unsigned long idx = offset / SC_PER_PAGE;
 492         unsigned long pos = offset & SC_POS_MASK;
 493         struct swap_cgroup_ctrl *ctrl;
 494         struct page *mappage;
 495         struct swap_cgroup *sc;
 496         unsigned short ret;
 497
 498         ctrl = &swap_cgroup_ctrl[type];
 499         mappage = ctrl->map[idx];
 500         sc = page_address(mappage);
 501         sc += pos;
 502         ret = sc->id;
 503         return ret;
 504 }
 505
 506 int swap_cgroup_swapon(int type, unsigned long max_pages)
 507 {
 508         void *array;
 509         unsigned long array_size;
 510         unsigned long length;
 511         struct swap_cgroup_ctrl *ctrl;
 512
 513         if (!do_swap_account)
 514                 return 0;
 515
 516         length = DIV_ROUND_UP(max_pages, SC_PER_PAGE);
 517         array_size = length * sizeof(void *);
 518
 519         array = vzalloc(array_size);
 520         if (!array)
 521                 goto nomem;
 522
 523         ctrl = &swap_cgroup_ctrl[type];
 524         mutex_lock(&swap_cgroup_mutex);
 525         ctrl->length = length;
 526         ctrl->map = array;
 527         spin_lock_init(&ctrl->lock);
 528         if (swap_cgroup_prepare(type)) {
 529                 /* memory shortage */
 530                 ctrl->map = NULL;
 531                 ctrl->length = 0;
 532                 mutex_unlock(&swap_cgroup_mutex);
 533                 vfree(array);
 534                 goto nomem;
 535         }
 536         mutex_unlock(&swap_cgroup_mutex);
 537
 538         return 0;
 539 nomem:
 540         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
 541         printk(KERN_INFO
 542                 "swap_cgroup can be disabled by swapaccount=0 boot option\n");
 543         return -ENOMEM;
 544 }
 545
 546 void swap_cgroup_swapoff(int type)
 547 {
 548         struct page **map;
 549         unsigned long i, length;
 550         struct swap_cgroup_ctrl *ctrl;
 551
 552         if (!do_swap_account)
 553                 return;
 554
 555         mutex_lock(&swap_cgroup_mutex);
 556         ctrl = &swap_cgroup_ctrl[type];
 557         map = ctrl->map;
 558         length = ctrl->length;
 559         ctrl->map = NULL;
 560         ctrl->length = 0;
 561         mutex_unlock(&swap_cgroup_mutex);
 562
 563         if (map) {
 564                 for (i = 0; i < length; i++) {
 565                         struct page *page = map[i];
 566                         if (page)
 567                                 __free_page(page);
 568                 }
 569                 vfree(map);
 570         }
 571 }
 572
 573 #endif