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                 kmemleak_free(addr);
 164                 free_pages_exact(addr, table_size);
 165         }
 166 }
 167 #endif
 168
 169 static int __meminit init_section_page_cgroup(unsigned long pfn, int nid)
 170 {
 171         struct page_cgroup *base, *pc;
 172         struct mem_section *section;
 173         unsigned long table_size;
 174         unsigned long nr;
 175         int index;
 176
 177         nr = pfn_to_section_nr(pfn);
 178         section = __nr_to_section(nr);
 179
 180         if (section->page_cgroup)
 181                 return 0;
 182
 183         table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 184         base = alloc_page_cgroup(table_size, nid);
 185
 186         /*
 187          * The value stored in section->page_cgroup is (base - pfn)
 188          * and it does not point to the memory block allocated above,
 189          * causing kmemleak false positives.
 190          */
 191         kmemleak_not_leak(base);
 192
 193         if (!base) {
 194                 printk(KERN_ERR "page cgroup allocation failure\n");
 195                 return -ENOMEM;
 196         }
 197
 198         for (index = 0; index < PAGES_PER_SECTION; index++) {
 199                 pc = base + index;
 200                 init_page_cgroup(pc, nr);
 201         }
 202         /*
 203          * The passed "pfn" may not be aligned to SECTION.  For the calculation
 204          * we need to apply a mask.
 205          */
 206         pfn &= PAGE_SECTION_MASK;
 207         section->page_cgroup = base - pfn;
 208         total_usage += table_size;
 209         return 0;
 210 }
 211 #ifdef CONFIG_MEMORY_HOTPLUG
 212 void __free_page_cgroup(unsigned long pfn)
 213 {
 214         struct mem_section *ms;
 215         struct page_cgroup *base;
 216
 217         ms = __pfn_to_section(pfn);
 218         if (!ms || !ms->page_cgroup)
 219                 return;
 220         base = ms->page_cgroup + pfn;
 221         free_page_cgroup(base);
 222         ms->page_cgroup = NULL;
 223 }
 224
 225 int __meminit online_page_cgroup(unsigned long start_pfn,
 226                         unsigned long nr_pages,
 227                         int nid)
 228 {
 229         unsigned long start, end, pfn;
 230         int fail = 0;
 231
 232         start = SECTION_ALIGN_DOWN(start_pfn);
 233         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 234
 235         if (nid == -1) {
 236                 /*
 237                  * In this case, "nid" already exists and contains valid memory.
 238                  * "start_pfn" passed to us is a pfn which is an arg for
 239                  * online__pages(), and start_pfn should exist.
 240                  */
 241                 nid = pfn_to_nid(start_pfn);
 242                 VM_BUG_ON(!node_state(nid, N_ONLINE));
 243         }
 244
 245         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
 246                 if (!pfn_present(pfn))
 247                         continue;
 248                 fail = init_section_page_cgroup(pfn, nid);
 249         }
 250         if (!fail)
 251                 return 0;
 252
 253         /* rollback */
 254         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 255                 __free_page_cgroup(pfn);
 256
 257         return -ENOMEM;
 258 }
 259
 260 int __meminit offline_page_cgroup(unsigned long start_pfn,
 261                 unsigned long nr_pages, int nid)
 262 {
 263         unsigned long start, end, pfn;
 264
 265         start = SECTION_ALIGN_DOWN(start_pfn);
 266         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 267
 268         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 269                 __free_page_cgroup(pfn);
 270         return 0;
 271
 272 }
 273
 274 static int __meminit page_cgroup_callback(struct notifier_block *self,
 275                                unsigned long action, void *arg)
 276 {
 277         struct memory_notify *mn = arg;
 278         int ret = 0;
 279         switch (action) {
 280         case MEM_GOING_ONLINE:
 281                 ret = online_page_cgroup(mn->start_pfn,
 282                                    mn->nr_pages, mn->status_change_nid);
 283                 break;
 284         case MEM_OFFLINE:
 285                 offline_page_cgroup(mn->start_pfn,
 286                                 mn->nr_pages, mn->status_change_nid);
 287                 break;
 288         case MEM_CANCEL_ONLINE:
 289         case MEM_GOING_OFFLINE:
 290                 break;
 291         case MEM_ONLINE:
 292         case MEM_CANCEL_OFFLINE:
 293                 break;
 294         }
 295
 296         return notifier_from_errno(ret);
 297 }
 298
 299 #endif
 300
 301 void __init page_cgroup_init(void)
 302 {
 303         unsigned long pfn;
 304         int nid;
 305
 306         if (mem_cgroup_disabled())
 307                 return;
 308
 309         for_each_node_state(nid, N_HIGH_MEMORY) {
 310                 unsigned long start_pfn, end_pfn;
 311
 312                 start_pfn = node_start_pfn(nid);
 313                 end_pfn = node_end_pfn(nid);
 314                 /*
 315                  * start_pfn and end_pfn may not be aligned to SECTION and the
 316                  * page->flags of out of node pages are not initialized.  So we
 317                  * scan [start_pfn, the biggest section's pfn < end_pfn) here.
 318                  */
 319                 for (pfn = start_pfn;
 320                      pfn < end_pfn;
 321                      pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
 322
 323                         if (!pfn_valid(pfn))
 324                                 continue;
 325                         /*
 326                          * Nodes's pfns can be overlapping.
 327                          * We know some arch can have a nodes layout such as
 328                          * -------------pfn-------------->
 329                          * N0 | N1 | N2 | N0 | N1 | N2|....
 330                          */
 331                         if (pfn_to_nid(pfn) != nid)
 332                                 continue;
 333                         if (init_section_page_cgroup(pfn, nid))
 334                                 goto oom;
 335                 }
 336         }
 337         hotplug_memory_notifier(page_cgroup_callback, 0);
 338         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
 339         printk(KERN_INFO "please try 'cgroup_disable=memory' option if you "
 340                          "don't want memory cgroups\n");
 341         return;
 342 oom:
 343         printk(KERN_CRIT "try 'cgroup_disable=memory' boot option\n");
 344         panic("Out of memory");
 345 }
 346
 347 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
 348 {
 349         return;
 350 }
 351
 352 #endif
 353
 354
 355 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 356
 357 static DEFINE_MUTEX(swap_cgroup_mutex);
 358 struct swap_cgroup_ctrl {
 359         struct page **map;
 360         unsigned long length;
 361         spinlock_t      lock;
 362 };
 363
 364 static struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
 365
 366 struct swap_cgroup {
 367         unsigned short          id;
 368 };
 369 #define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
 370 #define SC_POS_MASK     (SC_PER_PAGE - 1)
 371
 372 /*
 373  * SwapCgroup implements "lookup" and "exchange" operations.
 374  * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
 375  * against SwapCache. At swap_free(), this is accessed directly from swap.
 376  *
 377  * This means,
 378  *  - we have no race in "exchange" when we're accessed via SwapCache because
 379  *    SwapCache(and its swp_entry) is under lock.
 380  *  - When called via swap_free(), there is no user of this entry and no race.
 381  * Then, we don't need lock around "exchange".
 382  *
 383  * TODO: we can push these buffers out to HIGHMEM.
 384  */
 385
 386 /*
 387  * allocate buffer for swap_cgroup.
 388  */
 389 static int swap_cgroup_prepare(int type)
 390 {
 391         struct page *page;
 392         struct swap_cgroup_ctrl *ctrl;
 393         unsigned long idx, max;
 394
 395         ctrl = &swap_cgroup_ctrl[type];
 396
 397         for (idx = 0; idx < ctrl->length; idx++) {
 398                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 399                 if (!page)
 400                         goto not_enough_page;
 401                 ctrl->map[idx] = page;
 402         }
 403         return 0;
 404 not_enough_page:
 405         max = idx;
 406         for (idx = 0; idx < max; idx++)
 407                 __free_page(ctrl->map[idx]);
 408
 409         return -ENOMEM;
 410 }
 411
 412 /**
 413  * swap_cgroup_cmpxchg - cmpxchg mem_cgroup's id for this swp_entry.
 414  * @end: swap entry to be cmpxchged
 415  * @old: old id
 416  * @new: new id
 417  *
 418  * Returns old id at success, 0 at failure.
 419  * (There is no mem_cgroup using 0 as its id)
 420  */
 421 unsigned short swap_cgroup_cmpxchg(swp_entry_t ent,
 422                                         unsigned short old, unsigned short new)
 423 {
 424         int type = swp_type(ent);
 425         unsigned long offset = swp_offset(ent);
 426         unsigned long idx = offset / SC_PER_PAGE;
 427         unsigned long pos = offset & SC_POS_MASK;
 428         struct swap_cgroup_ctrl *ctrl;
 429         struct page *mappage;
 430         struct swap_cgroup *sc;
 431         unsigned long flags;
 432         unsigned short retval;
 433
 434         ctrl = &swap_cgroup_ctrl[type];
 435
 436         mappage = ctrl->map[idx];
 437         sc = page_address(mappage);
 438         sc += pos;
 439         spin_lock_irqsave(&ctrl->lock, flags);
 440         retval = sc->id;
 441         if (retval == old)
 442                 sc->id = new;
 443         else
 444                 retval = 0;
 445         spin_unlock_irqrestore(&ctrl->lock, flags);
 446         return retval;
 447 }
 448
 449 /**
 450  * swap_cgroup_record - record mem_cgroup for this swp_entry.
 451  * @ent: swap entry to be recorded into
 452  * @mem: mem_cgroup to be recorded
 453  *
 454  * Returns old value at success, 0 at failure.
 455  * (Of course, old value can be 0.)
 456  */
 457 unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id)
 458 {
 459         int type = swp_type(ent);
 460         unsigned long offset = swp_offset(ent);
 461         unsigned long idx = offset / SC_PER_PAGE;
 462         unsigned long pos = offset & SC_POS_MASK;
 463         struct swap_cgroup_ctrl *ctrl;
 464         struct page *mappage;
 465         struct swap_cgroup *sc;
 466         unsigned short old;
 467         unsigned long flags;
 468
 469         ctrl = &swap_cgroup_ctrl[type];
 470
 471         mappage = ctrl->map[idx];
 472         sc = page_address(mappage);
 473         sc += pos;
 474         spin_lock_irqsave(&ctrl->lock, flags);
 475         old = sc->id;
 476         sc->id = id;
 477         spin_unlock_irqrestore(&ctrl->lock, flags);
 478
 479         return old;
 480 }
 481
 482 /**
 483  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
 484  * @ent: swap entry to be looked up.
 485  *
 486  * Returns CSS ID of mem_cgroup at success. 0 at failure. (0 is invalid ID)
 487  */
 488 unsigned short lookup_swap_cgroup(swp_entry_t ent)
 489 {
 490         int type = swp_type(ent);
 491         unsigned long offset = swp_offset(ent);
 492         unsigned long idx = offset / SC_PER_PAGE;
 493         unsigned long pos = offset & SC_POS_MASK;
 494         struct swap_cgroup_ctrl *ctrl;
 495         struct page *mappage;
 496         struct swap_cgroup *sc;
 497         unsigned short ret;
 498
 499         ctrl = &swap_cgroup_ctrl[type];
 500         mappage = ctrl->map[idx];
 501         sc = page_address(mappage);
 502         sc += pos;
 503         ret = sc->id;
 504         return ret;
 505 }
 506
 507 int swap_cgroup_swapon(int type, unsigned long max_pages)
 508 {
 509         void *array;
 510         unsigned long array_size;
 511         unsigned long length;
 512         struct swap_cgroup_ctrl *ctrl;
 513
 514         if (!do_swap_account)
 515                 return 0;
 516
 517         length = DIV_ROUND_UP(max_pages, SC_PER_PAGE);
 518         array_size = length * sizeof(void *);
 519
 520         array = vzalloc(array_size);
 521         if (!array)
 522                 goto nomem;
 523
 524         ctrl = &swap_cgroup_ctrl[type];
 525         mutex_lock(&swap_cgroup_mutex);
 526         ctrl->length = length;
 527         ctrl->map = array;
 528         spin_lock_init(&ctrl->lock);
 529         if (swap_cgroup_prepare(type)) {
 530                 /* memory shortage */
 531                 ctrl->map = NULL;
 532                 ctrl->length = 0;
 533                 mutex_unlock(&swap_cgroup_mutex);
 534                 vfree(array);
 535                 goto nomem;
 536         }
 537         mutex_unlock(&swap_cgroup_mutex);
 538
 539         return 0;
 540 nomem:
 541         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
 542         printk(KERN_INFO
 543                 "swap_cgroup can be disabled by swapaccount=0 boot option\n");
 544         return -ENOMEM;
 545 }
 546
 547 void swap_cgroup_swapoff(int type)
 548 {
 549         struct page **map;
 550         unsigned long i, length;
 551         struct swap_cgroup_ctrl *ctrl;
 552
 553         if (!do_swap_account)
 554                 return;
 555
 556         mutex_lock(&swap_cgroup_mutex);
 557         ctrl = &swap_cgroup_ctrl[type];
 558         map = ctrl->map;
 559         length = ctrl->length;
 560         ctrl->map = NULL;
 561         ctrl->length = 0;
 562         mutex_unlock(&swap_cgroup_mutex);
 563
 564         if (map) {
 565                 for (i = 0; i < length; i++) {
 566                         struct page *page = map[i];
 567                         if (page)
 568                                 __free_page(page);
 569                 }
 570                 vfree(map);
 571         }
 572 }
 573
 574 #endif