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 #ifdef CONFIG_MEMORY_HOTPLUG
 128 static void free_page_cgroup(void *addr)
 129 {
 130         if (is_vmalloc_addr(addr)) {
 131                 vfree(addr);
 132         } else {
 133                 struct page *page = virt_to_page(addr);
 134                 size_t table_size =
 135                         sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 136
 137                 BUG_ON(PageReserved(page));
 138                 free_pages_exact(addr, table_size);
 139         }
 140 }
 141 #endif
 142
 143 static int __meminit init_section_page_cgroup(unsigned long pfn, int nid)
 144 {
 145         struct mem_section *section;
 146         struct page_cgroup *base;
 147         unsigned long table_size;
 148
 149         section = __pfn_to_section(pfn);
 150
 151         if (section->page_cgroup)
 152                 return 0;
 153
 154         table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 155         base = alloc_page_cgroup(table_size, nid);
 156
 157         /*
 158          * The value stored in section->page_cgroup is (base - pfn)
 159          * and it does not point to the memory block allocated above,
 160          * causing kmemleak false positives.
 161          */
 162         kmemleak_not_leak(base);
 163
 164         if (!base) {
 165                 printk(KERN_ERR "page cgroup allocation failure\n");
 166                 return -ENOMEM;
 167         }
 168
 169         /*
 170          * The passed "pfn" may not be aligned to SECTION.  For the calculation
 171          * we need to apply a mask.
 172          */
 173         pfn &= PAGE_SECTION_MASK;
 174         section->page_cgroup = base - pfn;
 175         total_usage += table_size;
 176         return 0;
 177 }
 178 #ifdef CONFIG_MEMORY_HOTPLUG
 179 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 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 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         case MEM_GOING_OFFLINE:
 257                 break;
 258         case MEM_ONLINE:
 259         case MEM_CANCEL_OFFLINE:
 260                 break;
 261         }
 262
 263         return notifier_from_errno(ret);
 264 }
 265
 266 #endif
 267
 268 void __init page_cgroup_init(void)
 269 {
 270         unsigned long pfn;
 271         int nid;
 272
 273         if (mem_cgroup_disabled())
 274                 return;
 275
 276         for_each_node_state(nid, N_HIGH_MEMORY) {
 277                 unsigned long start_pfn, end_pfn;
 278
 279                 start_pfn = node_start_pfn(nid);
 280                 end_pfn = node_end_pfn(nid);
 281                 /*
 282                  * start_pfn and end_pfn may not be aligned to SECTION and the
 283                  * page->flags of out of node pages are not initialized.  So we
 284                  * scan [start_pfn, the biggest section's pfn < end_pfn) here.
 285                  */
 286                 for (pfn = start_pfn;
 287                      pfn < end_pfn;
 288                      pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
 289
 290                         if (!pfn_valid(pfn))
 291                                 continue;
 292                         /*
 293                          * Nodes's pfns can be overlapping.
 294                          * We know some arch can have a nodes layout such as
 295                          * -------------pfn-------------->
 296                          * N0 | N1 | N2 | N0 | N1 | N2|....
 297                          */
 298                         if (pfn_to_nid(pfn) != nid)
 299                                 continue;
 300                         if (init_section_page_cgroup(pfn, nid))
 301                                 goto oom;
 302                 }
 303         }
 304         hotplug_memory_notifier(page_cgroup_callback, 0);
 305         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
 306         printk(KERN_INFO "please try 'cgroup_disable=memory' option if you "
 307                          "don't want memory cgroups\n");
 308         return;
 309 oom:
 310         printk(KERN_CRIT "try 'cgroup_disable=memory' boot option\n");
 311         panic("Out of memory");
 312 }
 313
 314 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
 315 {
 316         return;
 317 }
 318
 319 #endif
 320
 321
 322 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 323
 324 static DEFINE_MUTEX(swap_cgroup_mutex);
 325 struct swap_cgroup_ctrl {
 326         struct page **map;
 327         unsigned long length;
 328         spinlock_t      lock;
 329 };
 330
 331 static struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
 332
 333 struct swap_cgroup {
 334         unsigned short          id;
 335 };
 336 #define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
 337 #define SC_POS_MASK     (SC_PER_PAGE - 1)
 338
 339 /*
 340  * SwapCgroup implements "lookup" and "exchange" operations.
 341  * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
 342  * against SwapCache. At swap_free(), this is accessed directly from swap.
 343  *
 344  * This means,
 345  *  - we have no race in "exchange" when we're accessed via SwapCache because
 346  *    SwapCache(and its swp_entry) is under lock.
 347  *  - When called via swap_free(), there is no user of this entry and no race.
 348  * Then, we don't need lock around "exchange".
 349  *
 350  * TODO: we can push these buffers out to HIGHMEM.
 351  */
 352
 353 /*
 354  * allocate buffer for swap_cgroup.
 355  */
 356 static int swap_cgroup_prepare(int type)
 357 {
 358         struct page *page;
 359         struct swap_cgroup_ctrl *ctrl;
 360         unsigned long idx, max;
 361
 362         ctrl = &swap_cgroup_ctrl[type];
 363
 364         for (idx = 0; idx < ctrl->length; idx++) {
 365                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 366                 if (!page)
 367                         goto not_enough_page;
 368                 ctrl->map[idx] = page;
 369         }
 370         return 0;
 371 not_enough_page:
 372         max = idx;
 373         for (idx = 0; idx < max; idx++)
 374                 __free_page(ctrl->map[idx]);
 375
 376         return -ENOMEM;
 377 }
 378
 379 /**
 380  * swap_cgroup_cmpxchg - cmpxchg mem_cgroup's id for this swp_entry.
 381  * @end: swap entry to be cmpxchged
 382  * @old: old id
 383  * @new: new id
 384  *
 385  * Returns old id at success, 0 at failure.
 386  * (There is no mem_cgroup using 0 as its id)
 387  */
 388 unsigned short swap_cgroup_cmpxchg(swp_entry_t ent,
 389                                         unsigned short old, unsigned short new)
 390 {
 391         int type = swp_type(ent);
 392         unsigned long offset = swp_offset(ent);
 393         unsigned long idx = offset / SC_PER_PAGE;
 394         unsigned long pos = offset & SC_POS_MASK;
 395         struct swap_cgroup_ctrl *ctrl;
 396         struct page *mappage;
 397         struct swap_cgroup *sc;
 398         unsigned long flags;
 399         unsigned short retval;
 400
 401         ctrl = &swap_cgroup_ctrl[type];
 402
 403         mappage = ctrl->map[idx];
 404         sc = page_address(mappage);
 405         sc += pos;
 406         spin_lock_irqsave(&ctrl->lock, flags);
 407         retval = sc->id;
 408         if (retval == old)
 409                 sc->id = new;
 410         else
 411                 retval = 0;
 412         spin_unlock_irqrestore(&ctrl->lock, flags);
 413         return retval;
 414 }
 415
 416 /**
 417  * swap_cgroup_record - record mem_cgroup for this swp_entry.
 418  * @ent: swap entry to be recorded into
 419  * @mem: mem_cgroup to be recorded
 420  *
 421  * Returns old value at success, 0 at failure.
 422  * (Of course, old value can be 0.)
 423  */
 424 unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id)
 425 {
 426         int type = swp_type(ent);
 427         unsigned long offset = swp_offset(ent);
 428         unsigned long idx = offset / SC_PER_PAGE;
 429         unsigned long pos = offset & SC_POS_MASK;
 430         struct swap_cgroup_ctrl *ctrl;
 431         struct page *mappage;
 432         struct swap_cgroup *sc;
 433         unsigned short old;
 434         unsigned long flags;
 435
 436         ctrl = &swap_cgroup_ctrl[type];
 437
 438         mappage = ctrl->map[idx];
 439         sc = page_address(mappage);
 440         sc += pos;
 441         spin_lock_irqsave(&ctrl->lock, flags);
 442         old = sc->id;
 443         sc->id = id;
 444         spin_unlock_irqrestore(&ctrl->lock, flags);
 445
 446         return old;
 447 }
 448
 449 /**
 450  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
 451  * @ent: swap entry to be looked up.
 452  *
 453  * Returns CSS ID of mem_cgroup at success. 0 at failure. (0 is invalid ID)
 454  */
 455 unsigned short lookup_swap_cgroup(swp_entry_t ent)
 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 ret;
 465
 466         ctrl = &swap_cgroup_ctrl[type];
 467         mappage = ctrl->map[idx];
 468         sc = page_address(mappage);
 469         sc += pos;
 470         ret = sc->id;
 471         return ret;
 472 }
 473
 474 int swap_cgroup_swapon(int type, unsigned long max_pages)
 475 {
 476         void *array;
 477         unsigned long array_size;
 478         unsigned long length;
 479         struct swap_cgroup_ctrl *ctrl;
 480
 481         if (!do_swap_account)
 482                 return 0;
 483
 484         length = DIV_ROUND_UP(max_pages, SC_PER_PAGE);
 485         array_size = length * sizeof(void *);
 486
 487         array = vzalloc(array_size);
 488         if (!array)
 489                 goto nomem;
 490
 491         ctrl = &swap_cgroup_ctrl[type];
 492         mutex_lock(&swap_cgroup_mutex);
 493         ctrl->length = length;
 494         ctrl->map = array;
 495         spin_lock_init(&ctrl->lock);
 496         if (swap_cgroup_prepare(type)) {
 497                 /* memory shortage */
 498                 ctrl->map = NULL;
 499                 ctrl->length = 0;
 500                 mutex_unlock(&swap_cgroup_mutex);
 501                 vfree(array);
 502                 goto nomem;
 503         }
 504         mutex_unlock(&swap_cgroup_mutex);
 505
 506         return 0;
 507 nomem:
 508         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
 509         printk(KERN_INFO
 510                 "swap_cgroup can be disabled by swapaccount=0 boot option\n");
 511         return -ENOMEM;
 512 }
 513
 514 void swap_cgroup_swapoff(int type)
 515 {
 516         struct page **map;
 517         unsigned long i, length;
 518         struct swap_cgroup_ctrl *ctrl;
 519
 520         if (!do_swap_account)
 521                 return;
 522
 523         mutex_lock(&swap_cgroup_mutex);
 524         ctrl = &swap_cgroup_ctrl[type];
 525         map = ctrl->map;
 526         length = ctrl->length;
 527         ctrl->map = NULL;
 528         ctrl->length = 0;
 529         mutex_unlock(&swap_cgroup_mutex);
 530
 531         if (map) {
 532                 for (i = 0; i < length; i++) {
 533                         struct page *page = map[i];
 534                         if (page)
 535                                 __free_page(page);
 536                 }
 537                 vfree(map);
 538         }
 539 }
 540
 541 #endif