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