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
  13 static void __meminit
  14 __init_page_cgroup(struct page_cgroup *pc, unsigned long pfn)
  15 {
  16         pc->flags = 0;
  17         pc->mem_cgroup = NULL;
  18         pc->page = pfn_to_page(pfn);
  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 static int __init alloc_node_page_cgroup(int nid)
  46 {
  47         struct page_cgroup *base, *pc;
  48         unsigned long table_size;
  49         unsigned long start_pfn, nr_pages, index;
  50
  51         start_pfn = NODE_DATA(nid)->node_start_pfn;
  52         nr_pages = NODE_DATA(nid)->node_spanned_pages;
  53
  54         if (!nr_pages)
  55                 return 0;
  56
  57         table_size = sizeof(struct page_cgroup) * nr_pages;
  58
  59         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
  60                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
  61         if (!base)
  62                 return -ENOMEM;
  63         for (index = 0; index < nr_pages; index++) {
  64                 pc = base + index;
  65                 __init_page_cgroup(pc, start_pfn + index);
  66         }
  67         NODE_DATA(nid)->node_page_cgroup = base;
  68         total_usage += table_size;
  69         return 0;
  70 }
  71
  72 void __init page_cgroup_init(void)
  73 {
  74
  75         int nid, fail;
  76
  77         if (mem_cgroup_disabled())
  78                 return;
  79
  80         for_each_online_node(nid)  {
  81                 fail = alloc_node_page_cgroup(nid);
  82                 if (fail)
  83                         goto fail;
  84         }
  85         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
  86         printk(KERN_INFO "please try cgroup_disable=memory option if you"
  87         " don't want\n");
  88         return;
  89 fail:
  90         printk(KERN_CRIT "allocation of page_cgroup was failed.\n");
  91         printk(KERN_CRIT "please try cgroup_disable=memory boot option\n");
  92         panic("Out of memory");
  93 }
  94
  95 #else /* CONFIG_FLAT_NODE_MEM_MAP */
  96
  97 struct page_cgroup *lookup_page_cgroup(struct page *page)
  98 {
  99         unsigned long pfn = page_to_pfn(page);
 100         struct mem_section *section = __pfn_to_section(pfn);
 101
 102         return section->page_cgroup + pfn;
 103 }
 104
 105 /* __alloc_bootmem...() is protected by !slab_available() */
 106 static int __init_refok init_section_page_cgroup(unsigned long pfn)
 107 {
 108         struct mem_section *section = __pfn_to_section(pfn);
 109         struct page_cgroup *base, *pc;
 110         unsigned long table_size;
 111         int nid, index;
 112
 113         if (!section->page_cgroup) {
 114                 nid = page_to_nid(pfn_to_page(pfn));
 115                 table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 116                 if (slab_is_available()) {
 117                         base = kmalloc_node(table_size,
 118                                         GFP_KERNEL | __GFP_NOWARN, nid);
 119                         if (!base)
 120                                 base = vmalloc_node(table_size, nid);
 121                 } else {
 122                         base = __alloc_bootmem_node_nopanic(NODE_DATA(nid),
 123                                 table_size,
 124                                 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 125                 }
 126         } else {
 127                 /*
 128                  * We don't have to allocate page_cgroup again, but
 129                  * address of memmap may be changed. So, we have to initialize
 130                  * again.
 131                  */
 132                 base = section->page_cgroup + pfn;
 133                 table_size = 0;
 134                 /* check address of memmap is changed or not. */
 135                 if (base->page == pfn_to_page(pfn))
 136                         return 0;
 137         }
 138
 139         if (!base) {
 140                 printk(KERN_ERR "page cgroup allocation failure\n");
 141                 return -ENOMEM;
 142         }
 143
 144         for (index = 0; index < PAGES_PER_SECTION; index++) {
 145                 pc = base + index;
 146                 __init_page_cgroup(pc, pfn + index);
 147         }
 148
 149         section->page_cgroup = base - pfn;
 150         total_usage += table_size;
 151         return 0;
 152 }
 153 #ifdef CONFIG_MEMORY_HOTPLUG
 154 void __free_page_cgroup(unsigned long pfn)
 155 {
 156         struct mem_section *ms;
 157         struct page_cgroup *base;
 158
 159         ms = __pfn_to_section(pfn);
 160         if (!ms || !ms->page_cgroup)
 161                 return;
 162         base = ms->page_cgroup + pfn;
 163         if (is_vmalloc_addr(base)) {
 164                 vfree(base);
 165                 ms->page_cgroup = NULL;
 166         } else {
 167                 struct page *page = virt_to_page(base);
 168                 if (!PageReserved(page)) { /* Is bootmem ? */
 169                         kfree(base);
 170                         ms->page_cgroup = NULL;
 171                 }
 172         }
 173 }
 174
 175 int __meminit online_page_cgroup(unsigned long start_pfn,
 176                         unsigned long nr_pages,
 177                         int nid)
 178 {
 179         unsigned long start, end, pfn;
 180         int fail = 0;
 181
 182         start = start_pfn & ~(PAGES_PER_SECTION - 1);
 183         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
 184
 185         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
 186                 if (!pfn_present(pfn))
 187                         continue;
 188                 fail = init_section_page_cgroup(pfn);
 189         }
 190         if (!fail)
 191                 return 0;
 192
 193         /* rollback */
 194         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 195                 __free_page_cgroup(pfn);
 196
 197         return -ENOMEM;
 198 }
 199
 200 int __meminit offline_page_cgroup(unsigned long start_pfn,
 201                 unsigned long nr_pages, int nid)
 202 {
 203         unsigned long start, end, pfn;
 204
 205         start = start_pfn & ~(PAGES_PER_SECTION - 1);
 206         end = ALIGN(start_pfn + nr_pages, PAGES_PER_SECTION);
 207
 208         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 209                 __free_page_cgroup(pfn);
 210         return 0;
 211
 212 }
 213
 214 static int __meminit page_cgroup_callback(struct notifier_block *self,
 215                                unsigned long action, void *arg)
 216 {
 217         struct memory_notify *mn = arg;
 218         int ret = 0;
 219         switch (action) {
 220         case MEM_GOING_ONLINE:
 221                 ret = online_page_cgroup(mn->start_pfn,
 222                                    mn->nr_pages, mn->status_change_nid);
 223                 break;
 224         case MEM_OFFLINE:
 225                 offline_page_cgroup(mn->start_pfn,
 226                                 mn->nr_pages, mn->status_change_nid);
 227                 break;
 228         case MEM_CANCEL_ONLINE:
 229         case MEM_GOING_OFFLINE:
 230                 break;
 231         case MEM_ONLINE:
 232         case MEM_CANCEL_OFFLINE:
 233                 break;
 234         }
 235
 236         if (ret)
 237                 ret = notifier_from_errno(ret);
 238         else
 239                 ret = NOTIFY_OK;
 240
 241         return ret;
 242 }
 243
 244 #endif
 245
 246 void __init page_cgroup_init(void)
 247 {
 248         unsigned long pfn;
 249         int fail = 0;
 250
 251         if (mem_cgroup_disabled())
 252                 return;
 253
 254         for (pfn = 0; !fail && pfn < max_pfn; pfn += PAGES_PER_SECTION) {
 255                 if (!pfn_present(pfn))
 256                         continue;
 257                 fail = init_section_page_cgroup(pfn);
 258         }
 259         if (fail) {
 260                 printk(KERN_CRIT "try cgroup_disable=memory boot option\n");
 261                 panic("Out of memory");
 262         } else {
 263                 hotplug_memory_notifier(page_cgroup_callback, 0);
 264         }
 265         printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
 266         printk(KERN_INFO "please try cgroup_disable=memory option if you don't"
 267         " want\n");
 268 }
 269
 270 void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
 271 {
 272         return;
 273 }
 274
 275 #endif
 276
 277
 278 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 279
 280 static DEFINE_MUTEX(swap_cgroup_mutex);
 281 struct swap_cgroup_ctrl {
 282         struct page **map;
 283         unsigned long length;
 284 };
 285
 286 struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
 287
 288 /*
 289  * This 8bytes seems big..maybe we can reduce this when we can use "id" for
 290  * cgroup rather than pointer.
 291  */
 292 struct swap_cgroup {
 293         struct mem_cgroup       *val;
 294 };
 295 #define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
 296 #define SC_POS_MASK     (SC_PER_PAGE - 1)
 297
 298 /*
 299  * SwapCgroup implements "lookup" and "exchange" operations.
 300  * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
 301  * against SwapCache. At swap_free(), this is accessed directly from swap.
 302  *
 303  * This means,
 304  *  - we have no race in "exchange" when we're accessed via SwapCache because
 305  *    SwapCache(and its swp_entry) is under lock.
 306  *  - When called via swap_free(), there is no user of this entry and no race.
 307  * Then, we don't need lock around "exchange".
 308  *
 309  * TODO: we can push these buffers out to HIGHMEM.
 310  */
 311
 312 /*
 313  * allocate buffer for swap_cgroup.
 314  */
 315 static int swap_cgroup_prepare(int type)
 316 {
 317         struct page *page;
 318         struct swap_cgroup_ctrl *ctrl;
 319         unsigned long idx, max;
 320
 321         if (!do_swap_account)
 322                 return 0;
 323         ctrl = &swap_cgroup_ctrl[type];
 324
 325         for (idx = 0; idx < ctrl->length; idx++) {
 326                 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 327                 if (!page)
 328                         goto not_enough_page;
 329                 ctrl->map[idx] = page;
 330         }
 331         return 0;
 332 not_enough_page:
 333         max = idx;
 334         for (idx = 0; idx < max; idx++)
 335                 __free_page(ctrl->map[idx]);
 336
 337         return -ENOMEM;
 338 }
 339
 340 /**
 341  * swap_cgroup_record - record mem_cgroup for this swp_entry.
 342  * @ent: swap entry to be recorded into
 343  * @mem: mem_cgroup to be recorded
 344  *
 345  * Returns old value at success, NULL at failure.
 346  * (Of course, old value can be NULL.)
 347  */
 348 struct mem_cgroup *swap_cgroup_record(swp_entry_t ent, struct mem_cgroup *mem)
 349 {
 350         int type = swp_type(ent);
 351         unsigned long offset = swp_offset(ent);
 352         unsigned long idx = offset / SC_PER_PAGE;
 353         unsigned long pos = offset & SC_POS_MASK;
 354         struct swap_cgroup_ctrl *ctrl;
 355         struct page *mappage;
 356         struct swap_cgroup *sc;
 357         struct mem_cgroup *old;
 358
 359         if (!do_swap_account)
 360                 return NULL;
 361
 362         ctrl = &swap_cgroup_ctrl[type];
 363
 364         mappage = ctrl->map[idx];
 365         sc = page_address(mappage);
 366         sc += pos;
 367         old = sc->val;
 368         sc->val = mem;
 369
 370         return old;
 371 }
 372
 373 /**
 374  * lookup_swap_cgroup - lookup mem_cgroup tied to swap entry
 375  * @ent: swap entry to be looked up.
 376  *
 377  * Returns pointer to mem_cgroup at success. NULL at failure.
 378  */
 379 struct mem_cgroup *lookup_swap_cgroup(swp_entry_t ent)
 380 {
 381         int type = swp_type(ent);
 382         unsigned long offset = swp_offset(ent);
 383         unsigned long idx = offset / SC_PER_PAGE;
 384         unsigned long pos = offset & SC_POS_MASK;
 385         struct swap_cgroup_ctrl *ctrl;
 386         struct page *mappage;
 387         struct swap_cgroup *sc;
 388         struct mem_cgroup *ret;
 389
 390         if (!do_swap_account)
 391                 return NULL;
 392
 393         ctrl = &swap_cgroup_ctrl[type];
 394         mappage = ctrl->map[idx];
 395         sc = page_address(mappage);
 396         sc += pos;
 397         ret = sc->val;
 398         return ret;
 399 }
 400
 401 int swap_cgroup_swapon(int type, unsigned long max_pages)
 402 {
 403         void *array;
 404         unsigned long array_size;
 405         unsigned long length;
 406         struct swap_cgroup_ctrl *ctrl;
 407
 408         if (!do_swap_account)
 409                 return 0;
 410
 411         length = ((max_pages/SC_PER_PAGE) + 1);
 412         array_size = length * sizeof(void *);
 413
 414         array = vmalloc(array_size);
 415         if (!array)
 416                 goto nomem;
 417
 418         memset(array, 0, array_size);
 419         ctrl = &swap_cgroup_ctrl[type];
 420         mutex_lock(&swap_cgroup_mutex);
 421         ctrl->length = length;
 422         ctrl->map = array;
 423         if (swap_cgroup_prepare(type)) {
 424                 /* memory shortage */
 425                 ctrl->map = NULL;
 426                 ctrl->length = 0;
 427                 vfree(array);
 428                 mutex_unlock(&swap_cgroup_mutex);
 429                 goto nomem;
 430         }
 431         mutex_unlock(&swap_cgroup_mutex);
 432
 433         printk(KERN_INFO
 434                 "swap_cgroup: uses %ld bytes of vmalloc for pointer array space"
 435                 " and %ld bytes to hold mem_cgroup pointers on swap\n",
 436                 array_size, length * PAGE_SIZE);
 437         printk(KERN_INFO
 438         "swap_cgroup can be disabled by noswapaccount boot option.\n");
 439
 440         return 0;
 441 nomem:
 442         printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
 443         printk(KERN_INFO
 444                 "swap_cgroup can be disabled by noswapaccount boot option\n");
 445         return -ENOMEM;
 446 }
 447
 448 void swap_cgroup_swapoff(int type)
 449 {
 450         int i;
 451         struct swap_cgroup_ctrl *ctrl;
 452
 453         if (!do_swap_account)
 454                 return;
 455
 456         mutex_lock(&swap_cgroup_mutex);
 457         ctrl = &swap_cgroup_ctrl[type];
 458         if (ctrl->map) {
 459                 for (i = 0; i < ctrl->length; i++) {
 460                         struct page *page = ctrl->map[i];
 461                         if (page)
 462                                 __free_page(page);
 463                 }
 464                 vfree(ctrl->map);
 465                 ctrl->map = NULL;
 466                 ctrl->length = 0;
 467         }
 468         mutex_unlock(&swap_cgroup_mutex);
 469 }
 470
 471 #endif