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