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