linux/mm/page_cgroup.c
<<
>>
Prefs
   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
  14static unsigned long total_usage;
  15
  16#if !defined(CONFIG_SPARSEMEM)
  17
  18
  19void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
  20{
  21        pgdat->node_page_cgroup = NULL;
  22}
  23
  24struct 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
  45static 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
  66void __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;
  83fail:
  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
  91struct 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
 108static 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
 127static int __meminit init_section_page_cgroup(unsigned long pfn, int nid)
 128{
 129        struct mem_section *section;
 130        struct page_cgroup *base;
 131        unsigned long table_size;
 132
 133        section = __pfn_to_section(pfn);
 134
 135        if (section->page_cgroup)
 136                return 0;
 137
 138        table_size = sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 139        base = alloc_page_cgroup(table_size, nid);
 140
 141        /*
 142         * The value stored in section->page_cgroup is (base - pfn)
 143         * and it does not point to the memory block allocated above,
 144         * causing kmemleak false positives.
 145         */
 146        kmemleak_not_leak(base);
 147
 148        if (!base) {
 149                printk(KERN_ERR "page cgroup allocation failure\n");
 150                return -ENOMEM;
 151        }
 152
 153        /*
 154         * The passed "pfn" may not be aligned to SECTION.  For the calculation
 155         * we need to apply a mask.
 156         */
 157        pfn &= PAGE_SECTION_MASK;
 158        section->page_cgroup = base - pfn;
 159        total_usage += table_size;
 160        return 0;
 161}
 162#ifdef CONFIG_MEMORY_HOTPLUG
 163static void free_page_cgroup(void *addr)
 164{
 165        if (is_vmalloc_addr(addr)) {
 166                vfree(addr);
 167        } else {
 168                struct page *page = virt_to_page(addr);
 169                size_t table_size =
 170                        sizeof(struct page_cgroup) * PAGES_PER_SECTION;
 171
 172                BUG_ON(PageReserved(page));
 173                free_pages_exact(addr, table_size);
 174        }
 175}
 176
 177void __free_page_cgroup(unsigned long pfn)
 178{
 179        struct mem_section *ms;
 180        struct page_cgroup *base;
 181
 182        ms = __pfn_to_section(pfn);
 183        if (!ms || !ms->page_cgroup)
 184                return;
 185        base = ms->page_cgroup + pfn;
 186        free_page_cgroup(base);
 187        ms->page_cgroup = NULL;
 188}
 189
 190int __meminit online_page_cgroup(unsigned long start_pfn,
 191                        unsigned long nr_pages,
 192                        int nid)
 193{
 194        unsigned long start, end, pfn;
 195        int fail = 0;
 196
 197        start = SECTION_ALIGN_DOWN(start_pfn);
 198        end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 199
 200        if (nid == -1) {
 201                /*
 202                 * In this case, "nid" already exists and contains valid memory.
 203                 * "start_pfn" passed to us is a pfn which is an arg for
 204                 * online__pages(), and start_pfn should exist.
 205                 */
 206                nid = pfn_to_nid(start_pfn);
 207                VM_BUG_ON(!node_state(nid, N_ONLINE));
 208        }
 209
 210        for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
 211                if (!pfn_present(pfn))
 212                        continue;
 213                fail = init_section_page_cgroup(pfn, nid);
 214        }
 215        if (!fail)
 216                return 0;
 217
 218        /* rollback */
 219        for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 220                __free_page_cgroup(pfn);
 221
 222        return -ENOMEM;
 223}
 224
 225int __meminit offline_page_cgroup(unsigned long start_pfn,
 226                unsigned long nr_pages, int nid)
 227{
 228        unsigned long start, end, pfn;
 229
 230        start = SECTION_ALIGN_DOWN(start_pfn);
 231        end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 232
 233        for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 234                __free_page_cgroup(pfn);
 235        return 0;
 236
 237}
 238
 239static int __meminit page_cgroup_callback(struct notifier_block *self,
 240                               unsigned long action, void *arg)
 241{
 242        struct memory_notify *mn = arg;
 243        int ret = 0;
 244        switch (action) {
 245        case MEM_GOING_ONLINE:
 246                ret = online_page_cgroup(mn->start_pfn,
 247                                   mn->nr_pages, mn->status_change_nid);
 248                break;
 249        case MEM_OFFLINE:
 250                offline_page_cgroup(mn->start_pfn,
 251                                mn->nr_pages, mn->status_change_nid);
 252                break;
 253        case MEM_CANCEL_ONLINE:
 254        case MEM_GOING_OFFLINE:
 255                break;
 256        case MEM_ONLINE:
 257        case MEM_CANCEL_OFFLINE:
 258                break;
 259        }
 260
 261        return notifier_from_errno(ret);
 262}
 263
 264#endif
 265
 266void __init page_cgroup_init(void)
 267{
 268        unsigned long pfn;
 269        int nid;
 270
 271        if (mem_cgroup_disabled())
 272                return;
 273
 274        for_each_node_state(nid, N_HIGH_MEMORY) {
 275                unsigned long start_pfn, end_pfn;
 276
 277                start_pfn = node_start_pfn(nid);
 278                end_pfn = node_end_pfn(nid);
 279                /*
 280                 * start_pfn and end_pfn may not be aligned to SECTION and the
 281                 * page->flags of out of node pages are not initialized.  So we
 282                 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
 283                 */
 284                for (pfn = start_pfn;
 285                     pfn < end_pfn;
 286                     pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
 287
 288                        if (!pfn_valid(pfn))
 289                                continue;
 290                        /*
 291                         * Nodes's pfns can be overlapping.
 292                         * We know some arch can have a nodes layout such as
 293                         * -------------pfn-------------->
 294                         * N0 | N1 | N2 | N0 | N1 | N2|....
 295                         */
 296                        if (pfn_to_nid(pfn) != nid)
 297                                continue;
 298                        if (init_section_page_cgroup(pfn, nid))
 299                                goto oom;
 300                }
 301        }
 302        hotplug_memory_notifier(page_cgroup_callback, 0);
 303        printk(KERN_INFO "allocated %ld bytes of page_cgroup\n", total_usage);
 304        printk(KERN_INFO "please try 'cgroup_disable=memory' option if you "
 305                         "don't want memory cgroups\n");
 306        return;
 307oom:
 308        printk(KERN_CRIT "try 'cgroup_disable=memory' boot option\n");
 309        panic("Out of memory");
 310}
 311
 312void __meminit pgdat_page_cgroup_init(struct pglist_data *pgdat)
 313{
 314        return;
 315}
 316
 317#endif
 318
 319
 320#ifdef CONFIG_MEMCG_SWAP
 321
 322static DEFINE_MUTEX(swap_cgroup_mutex);
 323struct swap_cgroup_ctrl {
 324        struct page **map;
 325        unsigned long length;
 326        spinlock_t      lock;
 327};
 328
 329static struct swap_cgroup_ctrl swap_cgroup_ctrl[MAX_SWAPFILES];
 330
 331struct swap_cgroup {
 332        unsigned short          id;
 333};
 334#define SC_PER_PAGE     (PAGE_SIZE/sizeof(struct swap_cgroup))
 335
 336/*
 337 * SwapCgroup implements "lookup" and "exchange" operations.
 338 * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
 339 * against SwapCache. At swap_free(), this is accessed directly from swap.
 340 *
 341 * This means,
 342 *  - we have no race in "exchange" when we're accessed via SwapCache because
 343 *    SwapCache(and its swp_entry) is under lock.
 344 *  - When called via swap_free(), there is no user of this entry and no race.
 345 * Then, we don't need lock around "exchange".
 346 *
 347 * TODO: we can push these buffers out to HIGHMEM.
 348 */
 349
 350/*
 351 * allocate buffer for swap_cgroup.
 352 */
 353static int swap_cgroup_prepare(int type)
 354{
 355        struct page *page;
 356        struct swap_cgroup_ctrl *ctrl;
 357        unsigned long idx, max;
 358
 359        ctrl = &swap_cgroup_ctrl[type];
 360
 361        for (idx = 0; idx < ctrl->length; idx++) {
 362                page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 363                if (!page)
 364                        goto not_enough_page;
 365                ctrl->map[idx] = page;
 366        }
 367        return 0;
 368not_enough_page:
 369        max = idx;
 370        for (idx = 0; idx < max; idx++)
 371                __free_page(ctrl->map[idx]);
 372
 373        return -ENOMEM;
 374}
 375
 376static struct swap_cgroup *lookup_swap_cgroup(swp_entry_t ent,
 377                                        struct swap_cgroup_ctrl **ctrlp)
 378{
 379        pgoff_t offset = swp_offset(ent);
 380        struct swap_cgroup_ctrl *ctrl;
 381        struct page *mappage;
 382        struct swap_cgroup *sc;
 383
 384        ctrl = &swap_cgroup_ctrl[swp_type(ent)];
 385        if (ctrlp)
 386                *ctrlp = ctrl;
 387
 388        mappage = ctrl->map[offset / SC_PER_PAGE];
 389        sc = page_address(mappage);
 390        return sc + offset % SC_PER_PAGE;
 391}
 392
 393/**
 394 * swap_cgroup_cmpxchg - cmpxchg mem_cgroup's id for this swp_entry.
 395 * @ent: swap entry to be cmpxchged
 396 * @old: old id
 397 * @new: new id
 398 *
 399 * Returns old id at success, 0 at failure.
 400 * (There is no mem_cgroup using 0 as its id)
 401 */
 402unsigned short swap_cgroup_cmpxchg(swp_entry_t ent,
 403                                        unsigned short old, unsigned short new)
 404{
 405        struct swap_cgroup_ctrl *ctrl;
 406        struct swap_cgroup *sc;
 407        unsigned long flags;
 408        unsigned short retval;
 409
 410        sc = lookup_swap_cgroup(ent, &ctrl);
 411
 412        spin_lock_irqsave(&ctrl->lock, flags);
 413        retval = sc->id;
 414        if (retval == old)
 415                sc->id = new;
 416        else
 417                retval = 0;
 418        spin_unlock_irqrestore(&ctrl->lock, flags);
 419        return retval;
 420}
 421
 422/**
 423 * swap_cgroup_record - record mem_cgroup for this swp_entry.
 424 * @ent: swap entry to be recorded into
 425 * @id: mem_cgroup to be recorded
 426 *
 427 * Returns old value at success, 0 at failure.
 428 * (Of course, old value can be 0.)
 429 */
 430unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id)
 431{
 432        struct swap_cgroup_ctrl *ctrl;
 433        struct swap_cgroup *sc;
 434        unsigned short old;
 435        unsigned long flags;
 436
 437        sc = lookup_swap_cgroup(ent, &ctrl);
 438
 439        spin_lock_irqsave(&ctrl->lock, flags);
 440        old = sc->id;
 441        sc->id = id;
 442        spin_unlock_irqrestore(&ctrl->lock, flags);
 443
 444        return old;
 445}
 446
 447/**
 448 * lookup_swap_cgroup_id - lookup mem_cgroup id tied to swap entry
 449 * @ent: swap entry to be looked up.
 450 *
 451 * Returns CSS ID of mem_cgroup at success. 0 at failure. (0 is invalid ID)
 452 */
 453unsigned short lookup_swap_cgroup_id(swp_entry_t ent)
 454{
 455        return lookup_swap_cgroup(ent, NULL)->id;
 456}
 457
 458int swap_cgroup_swapon(int type, unsigned long max_pages)
 459{
 460        void *array;
 461        unsigned long array_size;
 462        unsigned long length;
 463        struct swap_cgroup_ctrl *ctrl;
 464
 465        if (!do_swap_account)
 466                return 0;
 467
 468        length = DIV_ROUND_UP(max_pages, SC_PER_PAGE);
 469        array_size = length * sizeof(void *);
 470
 471        array = vzalloc(array_size);
 472        if (!array)
 473                goto nomem;
 474
 475        ctrl = &swap_cgroup_ctrl[type];
 476        mutex_lock(&swap_cgroup_mutex);
 477        ctrl->length = length;
 478        ctrl->map = array;
 479        spin_lock_init(&ctrl->lock);
 480        if (swap_cgroup_prepare(type)) {
 481                /* memory shortage */
 482                ctrl->map = NULL;
 483                ctrl->length = 0;
 484                mutex_unlock(&swap_cgroup_mutex);
 485                vfree(array);
 486                goto nomem;
 487        }
 488        mutex_unlock(&swap_cgroup_mutex);
 489
 490        return 0;
 491nomem:
 492        printk(KERN_INFO "couldn't allocate enough memory for swap_cgroup.\n");
 493        printk(KERN_INFO
 494                "swap_cgroup can be disabled by swapaccount=0 boot option\n");
 495        return -ENOMEM;
 496}
 497
 498void swap_cgroup_swapoff(int type)
 499{
 500        struct page **map;
 501        unsigned long i, length;
 502        struct swap_cgroup_ctrl *ctrl;
 503
 504        if (!do_swap_account)
 505                return;
 506
 507        mutex_lock(&swap_cgroup_mutex);
 508        ctrl = &swap_cgroup_ctrl[type];
 509        map = ctrl->map;
 510        length = ctrl->length;
 511        ctrl->map = NULL;
 512        ctrl->length = 0;
 513        mutex_unlock(&swap_cgroup_mutex);
 514
 515        if (map) {
 516                for (i = 0; i < length; i++) {
 517                        struct page *page = map[i];
 518                        if (page)
 519                                __free_page(page);
 520                }
 521                vfree(map);
 522        }
 523}
 524
 525#endif
 526
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.