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