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