linux/mm/hugetlb.c
<<
>>
Prefs
   1/*
   2 * Generic hugetlb support.
   3 * (C) William Irwin, April 2004
   4 */
   5#include <linux/list.h>
   6#include <linux/init.h>
   7#include <linux/module.h>
   8#include <linux/mm.h>
   9#include <linux/seq_file.h>
  10#include <linux/sysctl.h>
  11#include <linux/highmem.h>
  12#include <linux/mmu_notifier.h>
  13#include <linux/nodemask.h>
  14#include <linux/pagemap.h>
  15#include <linux/mempolicy.h>
  16#include <linux/cpuset.h>
  17#include <linux/mutex.h>
  18#include <linux/bootmem.h>
  19#include <linux/sysfs.h>
  20#include <linux/slab.h>
  21#include <linux/rmap.h>
  22#include <linux/swap.h>
  23#include <linux/swapops.h>
  24
  25#include <asm/page.h>
  26#include <asm/pgtable.h>
  27#include <asm/tlb.h>
  28
  29#include <linux/io.h>
  30#include <linux/hugetlb.h>
  31#include <linux/hugetlb_cgroup.h>
  32#include <linux/node.h>
  33#include <linux/hugetlb_cgroup.h>
  34#include "internal.h"
  35
  36const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  37static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  38unsigned long hugepages_treat_as_movable;
  39
  40int hugetlb_max_hstate __read_mostly;
  41unsigned int default_hstate_idx;
  42struct hstate hstates[HUGE_MAX_HSTATE];
  43
  44__initdata LIST_HEAD(huge_boot_pages);
  45
  46/* for command line parsing */
  47static struct hstate * __initdata parsed_hstate;
  48static unsigned long __initdata default_hstate_max_huge_pages;
  49static unsigned long __initdata default_hstate_size;
  50
  51/*
  52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  53 */
  54DEFINE_SPINLOCK(hugetlb_lock);
  55
  56static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
  57{
  58        bool free = (spool->count == 0) && (spool->used_hpages == 0);
  59
  60        spin_unlock(&spool->lock);
  61
  62        /* If no pages are used, and no other handles to the subpool
  63         * remain, free the subpool the subpool remain */
  64        if (free)
  65                kfree(spool);
  66}
  67
  68struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
  69{
  70        struct hugepage_subpool *spool;
  71
  72        spool = kmalloc(sizeof(*spool), GFP_KERNEL);
  73        if (!spool)
  74                return NULL;
  75
  76        spin_lock_init(&spool->lock);
  77        spool->count = 1;
  78        spool->max_hpages = nr_blocks;
  79        spool->used_hpages = 0;
  80
  81        return spool;
  82}
  83
  84void hugepage_put_subpool(struct hugepage_subpool *spool)
  85{
  86        spin_lock(&spool->lock);
  87        BUG_ON(!spool->count);
  88        spool->count--;
  89        unlock_or_release_subpool(spool);
  90}
  91
  92static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
  93                                      long delta)
  94{
  95        int ret = 0;
  96
  97        if (!spool)
  98                return 0;
  99
 100        spin_lock(&spool->lock);
 101        if ((spool->used_hpages + delta) <= spool->max_hpages) {
 102                spool->used_hpages += delta;
 103        } else {
 104                ret = -ENOMEM;
 105        }
 106        spin_unlock(&spool->lock);
 107
 108        return ret;
 109}
 110
 111static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
 112                                       long delta)
 113{
 114        if (!spool)
 115                return;
 116
 117        spin_lock(&spool->lock);
 118        spool->used_hpages -= delta;
 119        /* If hugetlbfs_put_super couldn't free spool due to
 120        * an outstanding quota reference, free it now. */
 121        unlock_or_release_subpool(spool);
 122}
 123
 124static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
 125{
 126        return HUGETLBFS_SB(inode->i_sb)->spool;
 127}
 128
 129static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
 130{
 131        return subpool_inode(vma->vm_file->f_dentry->d_inode);
 132}
 133
 134/*
 135 * Region tracking -- allows tracking of reservations and instantiated pages
 136 *                    across the pages in a mapping.
 137 *
 138 * The region data structures are protected by a combination of the mmap_sem
 139 * and the hugetlb_instantion_mutex.  To access or modify a region the caller
 140 * must either hold the mmap_sem for write, or the mmap_sem for read and
 141 * the hugetlb_instantiation mutex:
 142 *
 143 *      down_write(&mm->mmap_sem);
 144 * or
 145 *      down_read(&mm->mmap_sem);
 146 *      mutex_lock(&hugetlb_instantiation_mutex);
 147 */
 148struct file_region {
 149        struct list_head link;
 150        long from;
 151        long to;
 152};
 153
 154static long region_add(struct list_head *head, long f, long t)
 155{
 156        struct file_region *rg, *nrg, *trg;
 157
 158        /* Locate the region we are either in or before. */
 159        list_for_each_entry(rg, head, link)
 160                if (f <= rg->to)
 161                        break;
 162
 163        /* Round our left edge to the current segment if it encloses us. */
 164        if (f > rg->from)
 165                f = rg->from;
 166
 167        /* Check for and consume any regions we now overlap with. */
 168        nrg = rg;
 169        list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 170                if (&rg->link == head)
 171                        break;
 172                if (rg->from > t)
 173                        break;
 174
 175                /* If this area reaches higher then extend our area to
 176                 * include it completely.  If this is not the first area
 177                 * which we intend to reuse, free it. */
 178                if (rg->to > t)
 179                        t = rg->to;
 180                if (rg != nrg) {
 181                        list_del(&rg->link);
 182                        kfree(rg);
 183                }
 184        }
 185        nrg->from = f;
 186        nrg->to = t;
 187        return 0;
 188}
 189
 190static long region_chg(struct list_head *head, long f, long t)
 191{
 192        struct file_region *rg, *nrg;
 193        long chg = 0;
 194
 195        /* Locate the region we are before or in. */
 196        list_for_each_entry(rg, head, link)
 197                if (f <= rg->to)
 198                        break;
 199
 200        /* If we are below the current region then a new region is required.
 201         * Subtle, allocate a new region at the position but make it zero
 202         * size such that we can guarantee to record the reservation. */
 203        if (&rg->link == head || t < rg->from) {
 204                nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 205                if (!nrg)
 206                        return -ENOMEM;
 207                nrg->from = f;
 208                nrg->to   = f;
 209                INIT_LIST_HEAD(&nrg->link);
 210                list_add(&nrg->link, rg->link.prev);
 211
 212                return t - f;
 213        }
 214
 215        /* Round our left edge to the current segment if it encloses us. */
 216        if (f > rg->from)
 217                f = rg->from;
 218        chg = t - f;
 219
 220        /* Check for and consume any regions we now overlap with. */
 221        list_for_each_entry(rg, rg->link.prev, link) {
 222                if (&rg->link == head)
 223                        break;
 224                if (rg->from > t)
 225                        return chg;
 226
 227                /* We overlap with this area, if it extends further than
 228                 * us then we must extend ourselves.  Account for its
 229                 * existing reservation. */
 230                if (rg->to > t) {
 231                        chg += rg->to - t;
 232                        t = rg->to;
 233                }
 234                chg -= rg->to - rg->from;
 235        }
 236        return chg;
 237}
 238
 239static long region_truncate(struct list_head *head, long end)
 240{
 241        struct file_region *rg, *trg;
 242        long chg = 0;
 243
 244        /* Locate the region we are either in or before. */
 245        list_for_each_entry(rg, head, link)
 246                if (end <= rg->to)
 247                        break;
 248        if (&rg->link == head)
 249                return 0;
 250
 251        /* If we are in the middle of a region then adjust it. */
 252        if (end > rg->from) {
 253                chg = rg->to - end;
 254                rg->to = end;
 255                rg = list_entry(rg->link.next, typeof(*rg), link);
 256        }
 257
 258        /* Drop any remaining regions. */
 259        list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 260                if (&rg->link == head)
 261                        break;
 262                chg += rg->to - rg->from;
 263                list_del(&rg->link);
 264                kfree(rg);
 265        }
 266        return chg;
 267}
 268
 269static long region_count(struct list_head *head, long f, long t)
 270{
 271        struct file_region *rg;
 272        long chg = 0;
 273
 274        /* Locate each segment we overlap with, and count that overlap. */
 275        list_for_each_entry(rg, head, link) {
 276                long seg_from;
 277                long seg_to;
 278
 279                if (rg->to <= f)
 280                        continue;
 281                if (rg->from >= t)
 282                        break;
 283
 284                seg_from = max(rg->from, f);
 285                seg_to = min(rg->to, t);
 286
 287                chg += seg_to - seg_from;
 288        }
 289
 290        return chg;
 291}
 292
 293/*
 294 * Convert the address within this vma to the page offset within
 295 * the mapping, in pagecache page units; huge pages here.
 296 */
 297static pgoff_t vma_hugecache_offset(struct hstate *h,
 298                        struct vm_area_struct *vma, unsigned long address)
 299{
 300        return ((address - vma->vm_start) >> huge_page_shift(h)) +
 301                        (vma->vm_pgoff >> huge_page_order(h));
 302}
 303
 304pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
 305                                     unsigned long address)
 306{
 307        return vma_hugecache_offset(hstate_vma(vma), vma, address);
 308}
 309
 310/*
 311 * Return the size of the pages allocated when backing a VMA. In the majority
 312 * cases this will be same size as used by the page table entries.
 313 */
 314unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 315{
 316        struct hstate *hstate;
 317
 318        if (!is_vm_hugetlb_page(vma))
 319                return PAGE_SIZE;
 320
 321        hstate = hstate_vma(vma);
 322
 323        return 1UL << (hstate->order + PAGE_SHIFT);
 324}
 325EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
 326
 327/*
 328 * Return the page size being used by the MMU to back a VMA. In the majority
 329 * of cases, the page size used by the kernel matches the MMU size. On
 330 * architectures where it differs, an architecture-specific version of this
 331 * function is required.
 332 */
 333#ifndef vma_mmu_pagesize
 334unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 335{
 336        return vma_kernel_pagesize(vma);
 337}
 338#endif
 339
 340/*
 341 * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 342 * bits of the reservation map pointer, which are always clear due to
 343 * alignment.
 344 */
 345#define HPAGE_RESV_OWNER    (1UL << 0)
 346#define HPAGE_RESV_UNMAPPED (1UL << 1)
 347#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 348
 349/*
 350 * These helpers are used to track how many pages are reserved for
 351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 352 * is guaranteed to have their future faults succeed.
 353 *
 354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 355 * the reserve counters are updated with the hugetlb_lock held. It is safe
 356 * to reset the VMA at fork() time as it is not in use yet and there is no
 357 * chance of the global counters getting corrupted as a result of the values.
 358 *
 359 * The private mapping reservation is represented in a subtly different
 360 * manner to a shared mapping.  A shared mapping has a region map associated
 361 * with the underlying file, this region map represents the backing file
 362 * pages which have ever had a reservation assigned which this persists even
 363 * after the page is instantiated.  A private mapping has a region map
 364 * associated with the original mmap which is attached to all VMAs which
 365 * reference it, this region map represents those offsets which have consumed
 366 * reservation ie. where pages have been instantiated.
 367 */
 368static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 369{
 370        return (unsigned long)vma->vm_private_data;
 371}
 372
 373static void set_vma_private_data(struct vm_area_struct *vma,
 374                                                        unsigned long value)
 375{
 376        vma->vm_private_data = (void *)value;
 377}
 378
 379struct resv_map {
 380        struct kref refs;
 381        struct list_head regions;
 382};
 383
 384static struct resv_map *resv_map_alloc(void)
 385{
 386        struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 387        if (!resv_map)
 388                return NULL;
 389
 390        kref_init(&resv_map->refs);
 391        INIT_LIST_HEAD(&resv_map->regions);
 392
 393        return resv_map;
 394}
 395
 396static void resv_map_release(struct kref *ref)
 397{
 398        struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 399
 400        /* Clear out any active regions before we release the map. */
 401        region_truncate(&resv_map->regions, 0);
 402        kfree(resv_map);
 403}
 404
 405static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 406{
 407        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 408        if (!(vma->vm_flags & VM_MAYSHARE))
 409                return (struct resv_map *)(get_vma_private_data(vma) &
 410                                                        ~HPAGE_RESV_MASK);
 411        return NULL;
 412}
 413
 414static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 415{
 416        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 417        VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 418
 419        set_vma_private_data(vma, (get_vma_private_data(vma) &
 420                                HPAGE_RESV_MASK) | (unsigned long)map);
 421}
 422
 423static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 424{
 425        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 426        VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 427
 428        set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 429}
 430
 431static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 432{
 433        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 434
 435        return (get_vma_private_data(vma) & flag) != 0;
 436}
 437
 438/* Decrement the reserved pages in the hugepage pool by one */
 439static void decrement_hugepage_resv_vma(struct hstate *h,
 440                        struct vm_area_struct *vma)
 441{
 442        if (vma->vm_flags & VM_NORESERVE)
 443                return;
 444
 445        if (vma->vm_flags & VM_MAYSHARE) {
 446                /* Shared mappings always use reserves */
 447                h->resv_huge_pages--;
 448        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 449                /*
 450                 * Only the process that called mmap() has reserves for
 451                 * private mappings.
 452                 */
 453                h->resv_huge_pages--;
 454        }
 455}
 456
 457/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 458void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 459{
 460        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 461        if (!(vma->vm_flags & VM_MAYSHARE))
 462                vma->vm_private_data = (void *)0;
 463}
 464
 465/* Returns true if the VMA has associated reserve pages */
 466static int vma_has_reserves(struct vm_area_struct *vma)
 467{
 468        if (vma->vm_flags & VM_MAYSHARE)
 469                return 1;
 470        if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 471                return 1;
 472        return 0;
 473}
 474
 475static void copy_gigantic_page(struct page *dst, struct page *src)
 476{
 477        int i;
 478        struct hstate *h = page_hstate(src);
 479        struct page *dst_base = dst;
 480        struct page *src_base = src;
 481
 482        for (i = 0; i < pages_per_huge_page(h); ) {
 483                cond_resched();
 484                copy_highpage(dst, src);
 485
 486                i++;
 487                dst = mem_map_next(dst, dst_base, i);
 488                src = mem_map_next(src, src_base, i);
 489        }
 490}
 491
 492void copy_huge_page(struct page *dst, struct page *src)
 493{
 494        int i;
 495        struct hstate *h = page_hstate(src);
 496
 497        if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
 498                copy_gigantic_page(dst, src);
 499                return;
 500        }
 501
 502        might_sleep();
 503        for (i = 0; i < pages_per_huge_page(h); i++) {
 504                cond_resched();
 505                copy_highpage(dst + i, src + i);
 506        }
 507}
 508
 509static void enqueue_huge_page(struct hstate *h, struct page *page)
 510{
 511        int nid = page_to_nid(page);
 512        list_move(&page->lru, &h->hugepage_freelists[nid]);
 513        h->free_huge_pages++;
 514        h->free_huge_pages_node[nid]++;
 515}
 516
 517static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 518{
 519        struct page *page;
 520
 521        if (list_empty(&h->hugepage_freelists[nid]))
 522                return NULL;
 523        page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
 524        list_move(&page->lru, &h->hugepage_activelist);
 525        set_page_refcounted(page);
 526        h->free_huge_pages--;
 527        h->free_huge_pages_node[nid]--;
 528        return page;
 529}
 530
 531static struct page *dequeue_huge_page_vma(struct hstate *h,
 532                                struct vm_area_struct *vma,
 533                                unsigned long address, int avoid_reserve)
 534{
 535        struct page *page = NULL;
 536        struct mempolicy *mpol;
 537        nodemask_t *nodemask;
 538        struct zonelist *zonelist;
 539        struct zone *zone;
 540        struct zoneref *z;
 541        unsigned int cpuset_mems_cookie;
 542
 543retry_cpuset:
 544        cpuset_mems_cookie = get_mems_allowed();
 545        zonelist = huge_zonelist(vma, address,
 546                                        htlb_alloc_mask, &mpol, &nodemask);
 547        /*
 548         * A child process with MAP_PRIVATE mappings created by their parent
 549         * have no page reserves. This check ensures that reservations are
 550         * not "stolen". The child may still get SIGKILLed
 551         */
 552        if (!vma_has_reserves(vma) &&
 553                        h->free_huge_pages - h->resv_huge_pages == 0)
 554                goto err;
 555
 556        /* If reserves cannot be used, ensure enough pages are in the pool */
 557        if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 558                goto err;
 559
 560        for_each_zone_zonelist_nodemask(zone, z, zonelist,
 561                                                MAX_NR_ZONES - 1, nodemask) {
 562                if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
 563                        page = dequeue_huge_page_node(h, zone_to_nid(zone));
 564                        if (page) {
 565                                if (!avoid_reserve)
 566                                        decrement_hugepage_resv_vma(h, vma);
 567                                break;
 568                        }
 569                }
 570        }
 571
 572        mpol_cond_put(mpol);
 573        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
 574                goto retry_cpuset;
 575        return page;
 576
 577err:
 578        mpol_cond_put(mpol);
 579        return NULL;
 580}
 581
 582static void update_and_free_page(struct hstate *h, struct page *page)
 583{
 584        int i;
 585
 586        VM_BUG_ON(h->order >= MAX_ORDER);
 587
 588        h->nr_huge_pages--;
 589        h->nr_huge_pages_node[page_to_nid(page)]--;
 590        for (i = 0; i < pages_per_huge_page(h); i++) {
 591                page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
 592                                1 << PG_referenced | 1 << PG_dirty |
 593                                1 << PG_active | 1 << PG_reserved |
 594                                1 << PG_private | 1 << PG_writeback);
 595        }
 596        VM_BUG_ON(hugetlb_cgroup_from_page(page));
 597        set_compound_page_dtor(page, NULL);
 598        set_page_refcounted(page);
 599        arch_release_hugepage(page);
 600        __free_pages(page, huge_page_order(h));
 601}
 602
 603struct hstate *size_to_hstate(unsigned long size)
 604{
 605        struct hstate *h;
 606
 607        for_each_hstate(h) {
 608                if (huge_page_size(h) == size)
 609                        return h;
 610        }
 611        return NULL;
 612}
 613
 614static void free_huge_page(struct page *page)
 615{
 616        /*
 617         * Can't pass hstate in here because it is called from the
 618         * compound page destructor.
 619         */
 620        struct hstate *h = page_hstate(page);
 621        int nid = page_to_nid(page);
 622        struct hugepage_subpool *spool =
 623                (struct hugepage_subpool *)page_private(page);
 624
 625        set_page_private(page, 0);
 626        page->mapping = NULL;
 627        BUG_ON(page_count(page));
 628        BUG_ON(page_mapcount(page));
 629
 630        spin_lock(&hugetlb_lock);
 631        hugetlb_cgroup_uncharge_page(hstate_index(h),
 632                                     pages_per_huge_page(h), page);
 633        if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 634                /* remove the page from active list */
 635                list_del(&page->lru);
 636                update_and_free_page(h, page);
 637                h->surplus_huge_pages--;
 638                h->surplus_huge_pages_node[nid]--;
 639        } else {
 640                enqueue_huge_page(h, page);
 641        }
 642        spin_unlock(&hugetlb_lock);
 643        hugepage_subpool_put_pages(spool, 1);
 644}
 645
 646static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 647{
 648        INIT_LIST_HEAD(&page->lru);
 649        set_compound_page_dtor(page, free_huge_page);
 650        spin_lock(&hugetlb_lock);
 651        set_hugetlb_cgroup(page, NULL);
 652        h->nr_huge_pages++;
 653        h->nr_huge_pages_node[nid]++;
 654        spin_unlock(&hugetlb_lock);
 655        put_page(page); /* free it into the hugepage allocator */
 656}
 657
 658static void prep_compound_gigantic_page(struct page *page, unsigned long order)
 659{
 660        int i;
 661        int nr_pages = 1 << order;
 662        struct page *p = page + 1;
 663
 664        /* we rely on prep_new_huge_page to set the destructor */
 665        set_compound_order(page, order);
 666        __SetPageHead(page);
 667        for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
 668                __SetPageTail(p);
 669                set_page_count(p, 0);
 670                p->first_page = page;
 671        }
 672}
 673
 674int PageHuge(struct page *page)
 675{
 676        compound_page_dtor *dtor;
 677
 678        if (!PageCompound(page))
 679                return 0;
 680
 681        page = compound_head(page);
 682        dtor = get_compound_page_dtor(page);
 683
 684        return dtor == free_huge_page;
 685}
 686EXPORT_SYMBOL_GPL(PageHuge);
 687
 688static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 689{
 690        struct page *page;
 691
 692        if (h->order >= MAX_ORDER)
 693                return NULL;
 694
 695        page = alloc_pages_exact_node(nid,
 696                htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 697                                                __GFP_REPEAT|__GFP_NOWARN,
 698                huge_page_order(h));
 699        if (page) {
 700                if (arch_prepare_hugepage(page)) {
 701                        __free_pages(page, huge_page_order(h));
 702                        return NULL;
 703                }
 704                prep_new_huge_page(h, page, nid);
 705        }
 706
 707        return page;
 708}
 709
 710/*
 711 * common helper functions for hstate_next_node_to_{alloc|free}.
 712 * We may have allocated or freed a huge page based on a different
 713 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
 714 * be outside of *nodes_allowed.  Ensure that we use an allowed
 715 * node for alloc or free.
 716 */
 717static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 718{
 719        nid = next_node(nid, *nodes_allowed);
 720        if (nid == MAX_NUMNODES)
 721                nid = first_node(*nodes_allowed);
 722        VM_BUG_ON(nid >= MAX_NUMNODES);
 723
 724        return nid;
 725}
 726
 727static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
 728{
 729        if (!node_isset(nid, *nodes_allowed))
 730                nid = next_node_allowed(nid, nodes_allowed);
 731        return nid;
 732}
 733
 734/*
 735 * returns the previously saved node ["this node"] from which to
 736 * allocate a persistent huge page for the pool and advance the
 737 * next node from which to allocate, handling wrap at end of node
 738 * mask.
 739 */
 740static int hstate_next_node_to_alloc(struct hstate *h,
 741                                        nodemask_t *nodes_allowed)
 742{
 743        int nid;
 744
 745        VM_BUG_ON(!nodes_allowed);
 746
 747        nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
 748        h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
 749
 750        return nid;
 751}
 752
 753static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
 754{
 755        struct page *page;
 756        int start_nid;
 757        int next_nid;
 758        int ret = 0;
 759
 760        start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 761        next_nid = start_nid;
 762
 763        do {
 764                page = alloc_fresh_huge_page_node(h, next_nid);
 765                if (page) {
 766                        ret = 1;
 767                        break;
 768                }
 769                next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 770        } while (next_nid != start_nid);
 771
 772        if (ret)
 773                count_vm_event(HTLB_BUDDY_PGALLOC);
 774        else
 775                count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 776
 777        return ret;
 778}
 779
 780/*
 781 * helper for free_pool_huge_page() - return the previously saved
 782 * node ["this node"] from which to free a huge page.  Advance the
 783 * next node id whether or not we find a free huge page to free so
 784 * that the next attempt to free addresses the next node.
 785 */
 786static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
 787{
 788        int nid;
 789
 790        VM_BUG_ON(!nodes_allowed);
 791
 792        nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
 793        h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
 794
 795        return nid;
 796}
 797
 798/*
 799 * Free huge page from pool from next node to free.
 800 * Attempt to keep persistent huge pages more or less
 801 * balanced over allowed nodes.
 802 * Called with hugetlb_lock locked.
 803 */
 804static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
 805                                                         bool acct_surplus)
 806{
 807        int start_nid;
 808        int next_nid;
 809        int ret = 0;
 810
 811        start_nid = hstate_next_node_to_free(h, nodes_allowed);
 812        next_nid = start_nid;
 813
 814        do {
 815                /*
 816                 * If we're returning unused surplus pages, only examine
 817                 * nodes with surplus pages.
 818                 */
 819                if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
 820                    !list_empty(&h->hugepage_freelists[next_nid])) {
 821                        struct page *page =
 822                                list_entry(h->hugepage_freelists[next_nid].next,
 823                                          struct page, lru);
 824                        list_del(&page->lru);
 825                        h->free_huge_pages--;
 826                        h->free_huge_pages_node[next_nid]--;
 827                        if (acct_surplus) {
 828                                h->surplus_huge_pages--;
 829                                h->surplus_huge_pages_node[next_nid]--;
 830                        }
 831                        update_and_free_page(h, page);
 832                        ret = 1;
 833                        break;
 834                }
 835                next_nid = hstate_next_node_to_free(h, nodes_allowed);
 836        } while (next_nid != start_nid);
 837
 838        return ret;
 839}
 840
 841static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 842{
 843        struct page *page;
 844        unsigned int r_nid;
 845
 846        if (h->order >= MAX_ORDER)
 847                return NULL;
 848
 849        /*
 850         * Assume we will successfully allocate the surplus page to
 851         * prevent racing processes from causing the surplus to exceed
 852         * overcommit
 853         *
 854         * This however introduces a different race, where a process B
 855         * tries to grow the static hugepage pool while alloc_pages() is
 856         * called by process A. B will only examine the per-node
 857         * counters in determining if surplus huge pages can be
 858         * converted to normal huge pages in adjust_pool_surplus(). A
 859         * won't be able to increment the per-node counter, until the
 860         * lock is dropped by B, but B doesn't drop hugetlb_lock until
 861         * no more huge pages can be converted from surplus to normal
 862         * state (and doesn't try to convert again). Thus, we have a
 863         * case where a surplus huge page exists, the pool is grown, and
 864         * the surplus huge page still exists after, even though it
 865         * should just have been converted to a normal huge page. This
 866         * does not leak memory, though, as the hugepage will be freed
 867         * once it is out of use. It also does not allow the counters to
 868         * go out of whack in adjust_pool_surplus() as we don't modify
 869         * the node values until we've gotten the hugepage and only the
 870         * per-node value is checked there.
 871         */
 872        spin_lock(&hugetlb_lock);
 873        if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 874                spin_unlock(&hugetlb_lock);
 875                return NULL;
 876        } else {
 877                h->nr_huge_pages++;
 878                h->surplus_huge_pages++;
 879        }
 880        spin_unlock(&hugetlb_lock);
 881
 882        if (nid == NUMA_NO_NODE)
 883                page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 884                                   __GFP_REPEAT|__GFP_NOWARN,
 885                                   huge_page_order(h));
 886        else
 887                page = alloc_pages_exact_node(nid,
 888                        htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 889                        __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
 890
 891        if (page && arch_prepare_hugepage(page)) {
 892                __free_pages(page, huge_page_order(h));
 893                page = NULL;
 894        }
 895
 896        spin_lock(&hugetlb_lock);
 897        if (page) {
 898                INIT_LIST_HEAD(&page->lru);
 899                r_nid = page_to_nid(page);
 900                set_compound_page_dtor(page, free_huge_page);
 901                set_hugetlb_cgroup(page, NULL);
 902                /*
 903                 * We incremented the global counters already
 904                 */
 905                h->nr_huge_pages_node[r_nid]++;
 906                h->surplus_huge_pages_node[r_nid]++;
 907                __count_vm_event(HTLB_BUDDY_PGALLOC);
 908        } else {
 909                h->nr_huge_pages--;
 910                h->surplus_huge_pages--;
 911                __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 912        }
 913        spin_unlock(&hugetlb_lock);
 914
 915        return page;
 916}
 917
 918/*
 919 * This allocation function is useful in the context where vma is irrelevant.
 920 * E.g. soft-offlining uses this function because it only cares physical
 921 * address of error page.
 922 */
 923struct page *alloc_huge_page_node(struct hstate *h, int nid)
 924{
 925        struct page *page;
 926
 927        spin_lock(&hugetlb_lock);
 928        page = dequeue_huge_page_node(h, nid);
 929        spin_unlock(&hugetlb_lock);
 930
 931        if (!page)
 932                page = alloc_buddy_huge_page(h, nid);
 933
 934        return page;
 935}
 936
 937/*
 938 * Increase the hugetlb pool such that it can accommodate a reservation
 939 * of size 'delta'.
 940 */
 941static int gather_surplus_pages(struct hstate *h, int delta)
 942{
 943        struct list_head surplus_list;
 944        struct page *page, *tmp;
 945        int ret, i;
 946        int needed, allocated;
 947        bool alloc_ok = true;
 948
 949        needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
 950        if (needed <= 0) {
 951                h->resv_huge_pages += delta;
 952                return 0;
 953        }
 954
 955        allocated = 0;
 956        INIT_LIST_HEAD(&surplus_list);
 957
 958        ret = -ENOMEM;
 959retry:
 960        spin_unlock(&hugetlb_lock);
 961        for (i = 0; i < needed; i++) {
 962                page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
 963                if (!page) {
 964                        alloc_ok = false;
 965                        break;
 966                }
 967                list_add(&page->lru, &surplus_list);
 968        }
 969        allocated += i;
 970
 971        /*
 972         * After retaking hugetlb_lock, we need to recalculate 'needed'
 973         * because either resv_huge_pages or free_huge_pages may have changed.
 974         */
 975        spin_lock(&hugetlb_lock);
 976        needed = (h->resv_huge_pages + delta) -
 977                        (h->free_huge_pages + allocated);
 978        if (needed > 0) {
 979                if (alloc_ok)
 980                        goto retry;
 981                /*
 982                 * We were not able to allocate enough pages to
 983                 * satisfy the entire reservation so we free what
 984                 * we've allocated so far.
 985                 */
 986                goto free;
 987        }
 988        /*
 989         * The surplus_list now contains _at_least_ the number of extra pages
 990         * needed to accommodate the reservation.  Add the appropriate number
 991         * of pages to the hugetlb pool and free the extras back to the buddy
 992         * allocator.  Commit the entire reservation here to prevent another
 993         * process from stealing the pages as they are added to the pool but
 994         * before they are reserved.
 995         */
 996        needed += allocated;
 997        h->resv_huge_pages += delta;
 998        ret = 0;
 999
1000        /* Free the needed pages to the hugetlb pool */
1001        list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1002                if ((--needed) < 0)
1003                        break;
1004                /*
1005                 * This page is now managed by the hugetlb allocator and has
1006                 * no users -- drop the buddy allocator's reference.
1007                 */
1008                put_page_testzero(page);
1009                VM_BUG_ON(page_count(page));
1010                enqueue_huge_page(h, page);
1011        }
1012free:
1013        spin_unlock(&hugetlb_lock);
1014
1015        /* Free unnecessary surplus pages to the buddy allocator */
1016        if (!list_empty(&surplus_list)) {
1017                list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1018                        put_page(page);
1019                }
1020        }
1021        spin_lock(&hugetlb_lock);
1022
1023        return ret;
1024}
1025
1026/*
1027 * When releasing a hugetlb pool reservation, any surplus pages that were
1028 * allocated to satisfy the reservation must be explicitly freed if they were
1029 * never used.
1030 * Called with hugetlb_lock held.
1031 */
1032static void return_unused_surplus_pages(struct hstate *h,
1033                                        unsigned long unused_resv_pages)
1034{
1035        unsigned long nr_pages;
1036
1037        /* Uncommit the reservation */
1038        h->resv_huge_pages -= unused_resv_pages;
1039
1040        /* Cannot return gigantic pages currently */
1041        if (h->order >= MAX_ORDER)
1042                return;
1043
1044        nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1045
1046        /*
1047         * We want to release as many surplus pages as possible, spread
1048         * evenly across all nodes with memory. Iterate across these nodes
1049         * until we can no longer free unreserved surplus pages. This occurs
1050         * when the nodes with surplus pages have no free pages.
1051         * free_pool_huge_page() will balance the the freed pages across the
1052         * on-line nodes with memory and will handle the hstate accounting.
1053         */
1054        while (nr_pages--) {
1055                if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1056                        break;
1057        }
1058}
1059
1060/*
1061 * Determine if the huge page at addr within the vma has an associated
1062 * reservation.  Where it does not we will need to logically increase
1063 * reservation and actually increase subpool usage before an allocation
1064 * can occur.  Where any new reservation would be required the
1065 * reservation change is prepared, but not committed.  Once the page
1066 * has been allocated from the subpool and instantiated the change should
1067 * be committed via vma_commit_reservation.  No action is required on
1068 * failure.
1069 */
1070static long vma_needs_reservation(struct hstate *h,
1071                        struct vm_area_struct *vma, unsigned long addr)
1072{
1073        struct address_space *mapping = vma->vm_file->f_mapping;
1074        struct inode *inode = mapping->host;
1075
1076        if (vma->vm_flags & VM_MAYSHARE) {
1077                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1078                return region_chg(&inode->i_mapping->private_list,
1079                                                        idx, idx + 1);
1080
1081        } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1082                return 1;
1083
1084        } else  {
1085                long err;
1086                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1087                struct resv_map *reservations = vma_resv_map(vma);
1088
1089                err = region_chg(&reservations->regions, idx, idx + 1);
1090                if (err < 0)
1091                        return err;
1092                return 0;
1093        }
1094}
1095static void vma_commit_reservation(struct hstate *h,
1096                        struct vm_area_struct *vma, unsigned long addr)
1097{
1098        struct address_space *mapping = vma->vm_file->f_mapping;
1099        struct inode *inode = mapping->host;
1100
1101        if (vma->vm_flags & VM_MAYSHARE) {
1102                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1103                region_add(&inode->i_mapping->private_list, idx, idx + 1);
1104
1105        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1106                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1107                struct resv_map *reservations = vma_resv_map(vma);
1108
1109                /* Mark this page used in the map. */
1110                region_add(&reservations->regions, idx, idx + 1);
1111        }
1112}
1113
1114static struct page *alloc_huge_page(struct vm_area_struct *vma,
1115                                    unsigned long addr, int avoid_reserve)
1116{
1117        struct hugepage_subpool *spool = subpool_vma(vma);
1118        struct hstate *h = hstate_vma(vma);
1119        struct page *page;
1120        long chg;
1121        int ret, idx;
1122        struct hugetlb_cgroup *h_cg;
1123
1124        idx = hstate_index(h);
1125        /*
1126         * Processes that did not create the mapping will have no
1127         * reserves and will not have accounted against subpool
1128         * limit. Check that the subpool limit can be made before
1129         * satisfying the allocation MAP_NORESERVE mappings may also
1130         * need pages and subpool limit allocated allocated if no reserve
1131         * mapping overlaps.
1132         */
1133        chg = vma_needs_reservation(h, vma, addr);
1134        if (chg < 0)
1135                return ERR_PTR(-ENOMEM);
1136        if (chg)
1137                if (hugepage_subpool_get_pages(spool, chg))
1138                        return ERR_PTR(-ENOSPC);
1139
1140        ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1141        if (ret) {
1142                hugepage_subpool_put_pages(spool, chg);
1143                return ERR_PTR(-ENOSPC);
1144        }
1145        spin_lock(&hugetlb_lock);
1146        page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1147        if (page) {
1148                /* update page cgroup details */
1149                hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1150                                             h_cg, page);
1151                spin_unlock(&hugetlb_lock);
1152        } else {
1153                spin_unlock(&hugetlb_lock);
1154                page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1155                if (!page) {
1156                        hugetlb_cgroup_uncharge_cgroup(idx,
1157                                                       pages_per_huge_page(h),
1158                                                       h_cg);
1159                        hugepage_subpool_put_pages(spool, chg);
1160                        return ERR_PTR(-ENOSPC);
1161                }
1162                spin_lock(&hugetlb_lock);
1163                hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1164                                             h_cg, page);
1165                list_move(&page->lru, &h->hugepage_activelist);
1166                spin_unlock(&hugetlb_lock);
1167        }
1168
1169        set_page_private(page, (unsigned long)spool);
1170
1171        vma_commit_reservation(h, vma, addr);
1172        return page;
1173}
1174
1175int __weak alloc_bootmem_huge_page(struct hstate *h)
1176{
1177        struct huge_bootmem_page *m;
1178        int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1179
1180        while (nr_nodes) {
1181                void *addr;
1182
1183                addr = __alloc_bootmem_node_nopanic(
1184                                NODE_DATA(hstate_next_node_to_alloc(h,
1185                                                &node_states[N_HIGH_MEMORY])),
1186                                huge_page_size(h), huge_page_size(h), 0);
1187
1188                if (addr) {
1189                        /*
1190                         * Use the beginning of the huge page to store the
1191                         * huge_bootmem_page struct (until gather_bootmem
1192                         * puts them into the mem_map).
1193                         */
1194                        m = addr;
1195                        goto found;
1196                }
1197                nr_nodes--;
1198        }
1199        return 0;
1200
1201found:
1202        BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1203        /* Put them into a private list first because mem_map is not up yet */
1204        list_add(&m->list, &huge_boot_pages);
1205        m->hstate = h;
1206        return 1;
1207}
1208
1209static void prep_compound_huge_page(struct page *page, int order)
1210{
1211        if (unlikely(order > (MAX_ORDER - 1)))
1212                prep_compound_gigantic_page(page, order);
1213        else
1214                prep_compound_page(page, order);
1215}
1216
1217/* Put bootmem huge pages into the standard lists after mem_map is up */
1218static void __init gather_bootmem_prealloc(void)
1219{
1220        struct huge_bootmem_page *m;
1221
1222        list_for_each_entry(m, &huge_boot_pages, list) {
1223                struct hstate *h = m->hstate;
1224                struct page *page;
1225
1226#ifdef CONFIG_HIGHMEM
1227                page = pfn_to_page(m->phys >> PAGE_SHIFT);
1228                free_bootmem_late((unsigned long)m,
1229                                  sizeof(struct huge_bootmem_page));
1230#else
1231                page = virt_to_page(m);
1232#endif
1233                __ClearPageReserved(page);
1234                WARN_ON(page_count(page) != 1);
1235                prep_compound_huge_page(page, h->order);
1236                prep_new_huge_page(h, page, page_to_nid(page));
1237                /*
1238                 * If we had gigantic hugepages allocated at boot time, we need
1239                 * to restore the 'stolen' pages to totalram_pages in order to
1240                 * fix confusing memory reports from free(1) and another
1241                 * side-effects, like CommitLimit going negative.
1242                 */
1243                if (h->order > (MAX_ORDER - 1))
1244                        totalram_pages += 1 << h->order;
1245        }
1246}
1247
1248static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1249{
1250        unsigned long i;
1251
1252        for (i = 0; i < h->max_huge_pages; ++i) {
1253                if (h->order >= MAX_ORDER) {
1254                        if (!alloc_bootmem_huge_page(h))
1255                                break;
1256                } else if (!alloc_fresh_huge_page(h,
1257                                         &node_states[N_HIGH_MEMORY]))
1258                        break;
1259        }
1260        h->max_huge_pages = i;
1261}
1262
1263static void __init hugetlb_init_hstates(void)
1264{
1265        struct hstate *h;
1266
1267        for_each_hstate(h) {
1268                /* oversize hugepages were init'ed in early boot */
1269                if (h->order < MAX_ORDER)
1270                        hugetlb_hstate_alloc_pages(h);
1271        }
1272}
1273
1274static char * __init memfmt(char *buf, unsigned long n)
1275{
1276        if (n >= (1UL << 30))
1277                sprintf(buf, "%lu GB", n >> 30);
1278        else if (n >= (1UL << 20))
1279                sprintf(buf, "%lu MB", n >> 20);
1280        else
1281                sprintf(buf, "%lu KB", n >> 10);
1282        return buf;
1283}
1284
1285static void __init report_hugepages(void)
1286{
1287        struct hstate *h;
1288
1289        for_each_hstate(h) {
1290                char buf[32];
1291                printk(KERN_INFO "HugeTLB registered %s page size, "
1292                                 "pre-allocated %ld pages\n",
1293                        memfmt(buf, huge_page_size(h)),
1294                        h->free_huge_pages);
1295        }
1296}
1297
1298#ifdef CONFIG_HIGHMEM
1299static void try_to_free_low(struct hstate *h, unsigned long count,
1300                                                nodemask_t *nodes_allowed)
1301{
1302        int i;
1303
1304        if (h->order >= MAX_ORDER)
1305                return;
1306
1307        for_each_node_mask(i, *nodes_allowed) {
1308                struct page *page, *next;
1309                struct list_head *freel = &h->hugepage_freelists[i];
1310                list_for_each_entry_safe(page, next, freel, lru) {
1311                        if (count >= h->nr_huge_pages)
1312                                return;
1313                        if (PageHighMem(page))
1314                                continue;
1315                        list_del(&page->lru);
1316                        update_and_free_page(h, page);
1317                        h->free_huge_pages--;
1318                        h->free_huge_pages_node[page_to_nid(page)]--;
1319                }
1320        }
1321}
1322#else
1323static inline void try_to_free_low(struct hstate *h, unsigned long count,
1324                                                nodemask_t *nodes_allowed)
1325{
1326}
1327#endif
1328
1329/*
1330 * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1331 * balanced by operating on them in a round-robin fashion.
1332 * Returns 1 if an adjustment was made.
1333 */
1334static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1335                                int delta)
1336{
1337        int start_nid, next_nid;
1338        int ret = 0;
1339
1340        VM_BUG_ON(delta != -1 && delta != 1);
1341
1342        if (delta < 0)
1343                start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1344        else
1345                start_nid = hstate_next_node_to_free(h, nodes_allowed);
1346        next_nid = start_nid;
1347
1348        do {
1349                int nid = next_nid;
1350                if (delta < 0)  {
1351                        /*
1352                         * To shrink on this node, there must be a surplus page
1353                         */
1354                        if (!h->surplus_huge_pages_node[nid]) {
1355                                next_nid = hstate_next_node_to_alloc(h,
1356                                                                nodes_allowed);
1357                                continue;
1358                        }
1359                }
1360                if (delta > 0) {
1361                        /*
1362                         * Surplus cannot exceed the total number of pages
1363                         */
1364                        if (h->surplus_huge_pages_node[nid] >=
1365                                                h->nr_huge_pages_node[nid]) {
1366                                next_nid = hstate_next_node_to_free(h,
1367                                                                nodes_allowed);
1368                                continue;
1369                        }
1370                }
1371
1372                h->surplus_huge_pages += delta;
1373                h->surplus_huge_pages_node[nid] += delta;
1374                ret = 1;
1375                break;
1376        } while (next_nid != start_nid);
1377
1378        return ret;
1379}
1380
1381#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1382static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1383                                                nodemask_t *nodes_allowed)
1384{
1385        unsigned long min_count, ret;
1386
1387        if (h->order >= MAX_ORDER)
1388                return h->max_huge_pages;
1389
1390        /*
1391         * Increase the pool size
1392         * First take pages out of surplus state.  Then make up the
1393         * remaining difference by allocating fresh huge pages.
1394         *
1395         * We might race with alloc_buddy_huge_page() here and be unable
1396         * to convert a surplus huge page to a normal huge page. That is
1397         * not critical, though, it just means the overall size of the
1398         * pool might be one hugepage larger than it needs to be, but
1399         * within all the constraints specified by the sysctls.
1400         */
1401        spin_lock(&hugetlb_lock);
1402        while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1403                if (!adjust_pool_surplus(h, nodes_allowed, -1))
1404                        break;
1405        }
1406
1407        while (count > persistent_huge_pages(h)) {
1408                /*
1409                 * If this allocation races such that we no longer need the
1410                 * page, free_huge_page will handle it by freeing the page
1411                 * and reducing the surplus.
1412                 */
1413                spin_unlock(&hugetlb_lock);
1414                ret = alloc_fresh_huge_page(h, nodes_allowed);
1415                spin_lock(&hugetlb_lock);
1416                if (!ret)
1417                        goto out;
1418
1419                /* Bail for signals. Probably ctrl-c from user */
1420                if (signal_pending(current))
1421                        goto out;
1422        }
1423
1424        /*
1425         * Decrease the pool size
1426         * First return free pages to the buddy allocator (being careful
1427         * to keep enough around to satisfy reservations).  Then place
1428         * pages into surplus state as needed so the pool will shrink
1429         * to the desired size as pages become free.
1430         *
1431         * By placing pages into the surplus state independent of the
1432         * overcommit value, we are allowing the surplus pool size to
1433         * exceed overcommit. There are few sane options here. Since
1434         * alloc_buddy_huge_page() is checking the global counter,
1435         * though, we'll note that we're not allowed to exceed surplus
1436         * and won't grow the pool anywhere else. Not until one of the
1437         * sysctls are changed, or the surplus pages go out of use.
1438         */
1439        min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1440        min_count = max(count, min_count);
1441        try_to_free_low(h, min_count, nodes_allowed);
1442        while (min_count < persistent_huge_pages(h)) {
1443                if (!free_pool_huge_page(h, nodes_allowed, 0))
1444                        break;
1445        }
1446        while (count < persistent_huge_pages(h)) {
1447                if (!adjust_pool_surplus(h, nodes_allowed, 1))
1448                        break;
1449        }
1450out:
1451        ret = persistent_huge_pages(h);
1452        spin_unlock(&hugetlb_lock);
1453        return ret;
1454}
1455
1456#define HSTATE_ATTR_RO(_name) \
1457        static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1458
1459#define HSTATE_ATTR(_name) \
1460        static struct kobj_attribute _name##_attr = \
1461                __ATTR(_name, 0644, _name##_show, _name##_store)
1462
1463static struct kobject *hugepages_kobj;
1464static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1465
1466static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1467
1468static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1469{
1470        int i;
1471
1472        for (i = 0; i < HUGE_MAX_HSTATE; i++)
1473                if (hstate_kobjs[i] == kobj) {
1474                        if (nidp)
1475                                *nidp = NUMA_NO_NODE;
1476                        return &hstates[i];
1477                }
1478
1479        return kobj_to_node_hstate(kobj, nidp);
1480}
1481
1482static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1483                                        struct kobj_attribute *attr, char *buf)
1484{
1485        struct hstate *h;
1486        unsigned long nr_huge_pages;
1487        int nid;
1488
1489        h = kobj_to_hstate(kobj, &nid);
1490        if (nid == NUMA_NO_NODE)
1491                nr_huge_pages = h->nr_huge_pages;
1492        else
1493                nr_huge_pages = h->nr_huge_pages_node[nid];
1494
1495        return sprintf(buf, "%lu\n", nr_huge_pages);
1496}
1497
1498static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1499                        struct kobject *kobj, struct kobj_attribute *attr,
1500                        const char *buf, size_t len)
1501{
1502        int err;
1503        int nid;
1504        unsigned long count;
1505        struct hstate *h;
1506        NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1507
1508        err = strict_strtoul(buf, 10, &count);
1509        if (err)
1510                goto out;
1511
1512        h = kobj_to_hstate(kobj, &nid);
1513        if (h->order >= MAX_ORDER) {
1514                err = -EINVAL;
1515                goto out;
1516        }
1517
1518        if (nid == NUMA_NO_NODE) {
1519                /*
1520                 * global hstate attribute
1521                 */
1522                if (!(obey_mempolicy &&
1523                                init_nodemask_of_mempolicy(nodes_allowed))) {
1524                        NODEMASK_FREE(nodes_allowed);
1525                        nodes_allowed = &node_states[N_HIGH_MEMORY];
1526                }
1527        } else if (nodes_allowed) {
1528                /*
1529                 * per node hstate attribute: adjust count to global,
1530                 * but restrict alloc/free to the specified node.
1531                 */
1532                count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1533                init_nodemask_of_node(nodes_allowed, nid);
1534        } else
1535                nodes_allowed = &node_states[N_HIGH_MEMORY];
1536
1537        h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1538
1539        if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1540                NODEMASK_FREE(nodes_allowed);
1541
1542        return len;
1543out:
1544        NODEMASK_FREE(nodes_allowed);
1545        return err;
1546}
1547
1548static ssize_t nr_hugepages_show(struct kobject *kobj,
1549                                       struct kobj_attribute *attr, char *buf)
1550{
1551        return nr_hugepages_show_common(kobj, attr, buf);
1552}
1553
1554static ssize_t nr_hugepages_store(struct kobject *kobj,
1555               struct kobj_attribute *attr, const char *buf, size_t len)
1556{
1557        return nr_hugepages_store_common(false, kobj, attr, buf, len);
1558}
1559HSTATE_ATTR(nr_hugepages);
1560
1561#ifdef CONFIG_NUMA
1562
1563/*
1564 * hstate attribute for optionally mempolicy-based constraint on persistent
1565 * huge page alloc/free.
1566 */
1567static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1568                                       struct kobj_attribute *attr, char *buf)
1569{
1570        return nr_hugepages_show_common(kobj, attr, buf);
1571}
1572
1573static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1574               struct kobj_attribute *attr, const char *buf, size_t len)
1575{
1576        return nr_hugepages_store_common(true, kobj, attr, buf, len);
1577}
1578HSTATE_ATTR(nr_hugepages_mempolicy);
1579#endif
1580
1581
1582static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1583                                        struct kobj_attribute *attr, char *buf)
1584{
1585        struct hstate *h = kobj_to_hstate(kobj, NULL);
1586        return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1587}
1588
1589static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1590                struct kobj_attribute *attr, const char *buf, size_t count)
1591{
1592        int err;
1593        unsigned long input;
1594        struct hstate *h = kobj_to_hstate(kobj, NULL);
1595
1596        if (h->order >= MAX_ORDER)
1597                return -EINVAL;
1598
1599        err = strict_strtoul(buf, 10, &input);
1600        if (err)
1601                return err;
1602
1603        spin_lock(&hugetlb_lock);
1604        h->nr_overcommit_huge_pages = input;
1605        spin_unlock(&hugetlb_lock);
1606
1607        return count;
1608}
1609HSTATE_ATTR(nr_overcommit_hugepages);
1610
1611static ssize_t free_hugepages_show(struct kobject *kobj,
1612                                        struct kobj_attribute *attr, char *buf)
1613{
1614        struct hstate *h;
1615        unsigned long free_huge_pages;
1616        int nid;
1617
1618        h = kobj_to_hstate(kobj, &nid);
1619        if (nid == NUMA_NO_NODE)
1620                free_huge_pages = h->free_huge_pages;
1621        else
1622                free_huge_pages = h->free_huge_pages_node[nid];
1623
1624        return sprintf(buf, "%lu\n", free_huge_pages);
1625}
1626HSTATE_ATTR_RO(free_hugepages);
1627
1628static ssize_t resv_hugepages_show(struct kobject *kobj,
1629                                        struct kobj_attribute *attr, char *buf)
1630{
1631        struct hstate *h = kobj_to_hstate(kobj, NULL);
1632        return sprintf(buf, "%lu\n", h->resv_huge_pages);
1633}
1634HSTATE_ATTR_RO(resv_hugepages);
1635
1636static ssize_t surplus_hugepages_show(struct kobject *kobj,
1637                                        struct kobj_attribute *attr, char *buf)
1638{
1639        struct hstate *h;
1640        unsigned long surplus_huge_pages;
1641        int nid;
1642
1643        h = kobj_to_hstate(kobj, &nid);
1644        if (nid == NUMA_NO_NODE)
1645                surplus_huge_pages = h->surplus_huge_pages;
1646        else
1647                surplus_huge_pages = h->surplus_huge_pages_node[nid];
1648
1649        return sprintf(buf, "%lu\n", surplus_huge_pages);
1650}
1651HSTATE_ATTR_RO(surplus_hugepages);
1652
1653static struct attribute *hstate_attrs[] = {
1654        &nr_hugepages_attr.attr,
1655        &nr_overcommit_hugepages_attr.attr,
1656        &free_hugepages_attr.attr,
1657        &resv_hugepages_attr.attr,
1658        &surplus_hugepages_attr.attr,
1659#ifdef CONFIG_NUMA
1660        &nr_hugepages_mempolicy_attr.attr,
1661#endif
1662        NULL,
1663};
1664
1665static struct attribute_group hstate_attr_group = {
1666        .attrs = hstate_attrs,
1667};
1668
1669static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1670                                    struct kobject **hstate_kobjs,
1671                                    struct attribute_group *hstate_attr_group)
1672{
1673        int retval;
1674        int hi = hstate_index(h);
1675
1676        hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1677        if (!hstate_kobjs[hi])
1678                return -ENOMEM;
1679
1680        retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1681        if (retval)
1682                kobject_put(hstate_kobjs[hi]);
1683
1684        return retval;
1685}
1686
1687static void __init hugetlb_sysfs_init(void)
1688{
1689        struct hstate *h;
1690        int err;
1691
1692        hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1693        if (!hugepages_kobj)
1694                return;
1695
1696        for_each_hstate(h) {
1697                err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1698                                         hstate_kobjs, &hstate_attr_group);
1699                if (err)
1700                        printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1701                                                                h->name);
1702        }
1703}
1704
1705#ifdef CONFIG_NUMA
1706
1707/*
1708 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1709 * with node devices in node_devices[] using a parallel array.  The array
1710 * index of a node device or _hstate == node id.
1711 * This is here to avoid any static dependency of the node device driver, in
1712 * the base kernel, on the hugetlb module.
1713 */
1714struct node_hstate {
1715        struct kobject          *hugepages_kobj;
1716        struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1717};
1718struct node_hstate node_hstates[MAX_NUMNODES];
1719
1720/*
1721 * A subset of global hstate attributes for node devices
1722 */
1723static struct attribute *per_node_hstate_attrs[] = {
1724        &nr_hugepages_attr.attr,
1725        &free_hugepages_attr.attr,
1726        &surplus_hugepages_attr.attr,
1727        NULL,
1728};
1729
1730static struct attribute_group per_node_hstate_attr_group = {
1731        .attrs = per_node_hstate_attrs,
1732};
1733
1734/*
1735 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1736 * Returns node id via non-NULL nidp.
1737 */
1738static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1739{
1740        int nid;
1741
1742        for (nid = 0; nid < nr_node_ids; nid++) {
1743                struct node_hstate *nhs = &node_hstates[nid];
1744                int i;
1745                for (i = 0; i < HUGE_MAX_HSTATE; i++)
1746                        if (nhs->hstate_kobjs[i] == kobj) {
1747                                if (nidp)
1748                                        *nidp = nid;
1749                                return &hstates[i];
1750                        }
1751        }
1752
1753        BUG();
1754        return NULL;
1755}
1756
1757/*
1758 * Unregister hstate attributes from a single node device.
1759 * No-op if no hstate attributes attached.
1760 */
1761void hugetlb_unregister_node(struct node *node)
1762{
1763        struct hstate *h;
1764        struct node_hstate *nhs = &node_hstates[node->dev.id];
1765
1766        if (!nhs->hugepages_kobj)
1767                return;         /* no hstate attributes */
1768
1769        for_each_hstate(h) {
1770                int idx = hstate_index(h);
1771                if (nhs->hstate_kobjs[idx]) {
1772                        kobject_put(nhs->hstate_kobjs[idx]);
1773                        nhs->hstate_kobjs[idx] = NULL;
1774                }
1775        }
1776
1777        kobject_put(nhs->hugepages_kobj);
1778        nhs->hugepages_kobj = NULL;
1779}
1780
1781/*
1782 * hugetlb module exit:  unregister hstate attributes from node devices
1783 * that have them.
1784 */
1785static void hugetlb_unregister_all_nodes(void)
1786{
1787        int nid;
1788
1789        /*
1790         * disable node device registrations.
1791         */
1792        register_hugetlbfs_with_node(NULL, NULL);
1793
1794        /*
1795         * remove hstate attributes from any nodes that have them.
1796         */
1797        for (nid = 0; nid < nr_node_ids; nid++)
1798                hugetlb_unregister_node(&node_devices[nid]);
1799}
1800
1801/*
1802 * Register hstate attributes for a single node device.
1803 * No-op if attributes already registered.
1804 */
1805void hugetlb_register_node(struct node *node)
1806{
1807        struct hstate *h;
1808        struct node_hstate *nhs = &node_hstates[node->dev.id];
1809        int err;
1810
1811        if (nhs->hugepages_kobj)
1812                return;         /* already allocated */
1813
1814        nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1815                                                        &node->dev.kobj);
1816        if (!nhs->hugepages_kobj)
1817                return;
1818
1819        for_each_hstate(h) {
1820                err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1821                                                nhs->hstate_kobjs,
1822                                                &per_node_hstate_attr_group);
1823                if (err) {
1824                        printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1825                                        " for node %d\n",
1826                                                h->name, node->dev.id);
1827                        hugetlb_unregister_node(node);
1828                        break;
1829                }
1830        }
1831}
1832
1833/*
1834 * hugetlb init time:  register hstate attributes for all registered node
1835 * devices of nodes that have memory.  All on-line nodes should have
1836 * registered their associated device by this time.
1837 */
1838static void hugetlb_register_all_nodes(void)
1839{
1840        int nid;
1841
1842        for_each_node_state(nid, N_HIGH_MEMORY) {
1843                struct node *node = &node_devices[nid];
1844                if (node->dev.id == nid)
1845                        hugetlb_register_node(node);
1846        }
1847
1848        /*
1849         * Let the node device driver know we're here so it can
1850         * [un]register hstate attributes on node hotplug.
1851         */
1852        register_hugetlbfs_with_node(hugetlb_register_node,
1853                                     hugetlb_unregister_node);
1854}
1855#else   /* !CONFIG_NUMA */
1856
1857static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1858{
1859        BUG();
1860        if (nidp)
1861                *nidp = -1;
1862        return NULL;
1863}
1864
1865static void hugetlb_unregister_all_nodes(void) { }
1866
1867static void hugetlb_register_all_nodes(void) { }
1868
1869#endif
1870
1871static void __exit hugetlb_exit(void)
1872{
1873        struct hstate *h;
1874
1875        hugetlb_unregister_all_nodes();
1876
1877        for_each_hstate(h) {
1878                kobject_put(hstate_kobjs[hstate_index(h)]);
1879        }
1880
1881        kobject_put(hugepages_kobj);
1882}
1883module_exit(hugetlb_exit);
1884
1885static int __init hugetlb_init(void)
1886{
1887        /* Some platform decide whether they support huge pages at boot
1888         * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1889         * there is no such support
1890         */
1891        if (HPAGE_SHIFT == 0)
1892                return 0;
1893
1894        if (!size_to_hstate(default_hstate_size)) {
1895                default_hstate_size = HPAGE_SIZE;
1896                if (!size_to_hstate(default_hstate_size))
1897                        hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1898        }
1899        default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1900        if (default_hstate_max_huge_pages)
1901                default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1902
1903        hugetlb_init_hstates();
1904
1905        gather_bootmem_prealloc();
1906
1907        report_hugepages();
1908
1909        hugetlb_sysfs_init();
1910
1911        hugetlb_register_all_nodes();
1912
1913        return 0;
1914}
1915module_init(hugetlb_init);
1916
1917/* Should be called on processing a hugepagesz=... option */
1918void __init hugetlb_add_hstate(unsigned order)
1919{
1920        struct hstate *h;
1921        unsigned long i;
1922
1923        if (size_to_hstate(PAGE_SIZE << order)) {
1924                printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1925                return;
1926        }
1927        BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1928        BUG_ON(order == 0);
1929        h = &hstates[hugetlb_max_hstate++];
1930        h->order = order;
1931        h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1932        h->nr_huge_pages = 0;
1933        h->free_huge_pages = 0;
1934        for (i = 0; i < MAX_NUMNODES; ++i)
1935                INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1936        INIT_LIST_HEAD(&h->hugepage_activelist);
1937        h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1938        h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1939        snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1940                                        huge_page_size(h)/1024);
1941        /*
1942         * Add cgroup control files only if the huge page consists
1943         * of more than two normal pages. This is because we use
1944         * page[2].lru.next for storing cgoup details.
1945         */
1946        if (order >= HUGETLB_CGROUP_MIN_ORDER)
1947                hugetlb_cgroup_file_init(hugetlb_max_hstate - 1);
1948
1949        parsed_hstate = h;
1950}
1951
1952static int __init hugetlb_nrpages_setup(char *s)
1953{
1954        unsigned long *mhp;
1955        static unsigned long *last_mhp;
1956
1957        /*
1958         * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1959         * so this hugepages= parameter goes to the "default hstate".
1960         */
1961        if (!hugetlb_max_hstate)
1962                mhp = &default_hstate_max_huge_pages;
1963        else
1964                mhp = &parsed_hstate->max_huge_pages;
1965
1966        if (mhp == last_mhp) {
1967                printk(KERN_WARNING "hugepages= specified twice without "
1968                        "interleaving hugepagesz=, ignoring\n");
1969                return 1;
1970        }
1971
1972        if (sscanf(s, "%lu", mhp) <= 0)
1973                *mhp = 0;
1974
1975        /*
1976         * Global state is always initialized later in hugetlb_init.
1977         * But we need to allocate >= MAX_ORDER hstates here early to still
1978         * use the bootmem allocator.
1979         */
1980        if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1981                hugetlb_hstate_alloc_pages(parsed_hstate);
1982
1983        last_mhp = mhp;
1984
1985        return 1;
1986}
1987__setup("hugepages=", hugetlb_nrpages_setup);
1988
1989static int __init hugetlb_default_setup(char *s)
1990{
1991        default_hstate_size = memparse(s, &s);
1992        return 1;
1993}
1994__setup("default_hugepagesz=", hugetlb_default_setup);
1995
1996static unsigned int cpuset_mems_nr(unsigned int *array)
1997{
1998        int node;
1999        unsigned int nr = 0;
2000
2001        for_each_node_mask(node, cpuset_current_mems_allowed)
2002                nr += array[node];
2003
2004        return nr;
2005}
2006
2007#ifdef CONFIG_SYSCTL
2008static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2009                         struct ctl_table *table, int write,
2010                         void __user *buffer, size_t *length, loff_t *ppos)
2011{
2012        struct hstate *h = &default_hstate;
2013        unsigned long tmp;
2014        int ret;
2015
2016        tmp = h->max_huge_pages;
2017
2018        if (write && h->order >= MAX_ORDER)
2019                return -EINVAL;
2020
2021        table->data = &tmp;
2022        table->maxlen = sizeof(unsigned long);
2023        ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2024        if (ret)
2025                goto out;
2026
2027        if (write) {
2028                NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2029                                                GFP_KERNEL | __GFP_NORETRY);
2030                if (!(obey_mempolicy &&
2031                               init_nodemask_of_mempolicy(nodes_allowed))) {
2032                        NODEMASK_FREE(nodes_allowed);
2033                        nodes_allowed = &node_states[N_HIGH_MEMORY];
2034                }
2035                h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2036
2037                if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2038                        NODEMASK_FREE(nodes_allowed);
2039        }
2040out:
2041        return ret;
2042}
2043
2044int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2045                          void __user *buffer, size_t *length, loff_t *ppos)
2046{
2047
2048        return hugetlb_sysctl_handler_common(false, table, write,
2049                                                        buffer, length, ppos);
2050}
2051
2052#ifdef CONFIG_NUMA
2053int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2054                          void __user *buffer, size_t *length, loff_t *ppos)
2055{
2056        return hugetlb_sysctl_handler_common(true, table, write,
2057                                                        buffer, length, ppos);
2058}
2059#endif /* CONFIG_NUMA */
2060
2061int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2062                        void __user *buffer,
2063                        size_t *length, loff_t *ppos)
2064{
2065        proc_dointvec(table, write, buffer, length, ppos);
2066        if (hugepages_treat_as_movable)
2067                htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2068        else
2069                htlb_alloc_mask = GFP_HIGHUSER;
2070        return 0;
2071}
2072
2073int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2074                        void __user *buffer,
2075                        size_t *length, loff_t *ppos)
2076{
2077        struct hstate *h = &default_hstate;
2078        unsigned long tmp;
2079        int ret;
2080
2081        tmp = h->nr_overcommit_huge_pages;
2082
2083        if (write && h->order >= MAX_ORDER)
2084                return -EINVAL;
2085
2086        table->data = &tmp;
2087        table->maxlen = sizeof(unsigned long);
2088        ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2089        if (ret)
2090                goto out;
2091
2092        if (write) {
2093                spin_lock(&hugetlb_lock);
2094                h->nr_overcommit_huge_pages = tmp;
2095                spin_unlock(&hugetlb_lock);
2096        }
2097out:
2098        return ret;
2099}
2100
2101#endif /* CONFIG_SYSCTL */
2102
2103void hugetlb_report_meminfo(struct seq_file *m)
2104{
2105        struct hstate *h = &default_hstate;
2106        seq_printf(m,
2107                        "HugePages_Total:   %5lu\n"
2108                        "HugePages_Free:    %5lu\n"
2109                        "HugePages_Rsvd:    %5lu\n"
2110                        "HugePages_Surp:    %5lu\n"
2111                        "Hugepagesize:   %8lu kB\n",
2112                        h->nr_huge_pages,
2113                        h->free_huge_pages,
2114                        h->resv_huge_pages,
2115                        h->surplus_huge_pages,
2116                        1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2117}
2118
2119int hugetlb_report_node_meminfo(int nid, char *buf)
2120{
2121        struct hstate *h = &default_hstate;
2122        return sprintf(buf,
2123                "Node %d HugePages_Total: %5u\n"
2124                "Node %d HugePages_Free:  %5u\n"
2125                "Node %d HugePages_Surp:  %5u\n",
2126                nid, h->nr_huge_pages_node[nid],
2127                nid, h->free_huge_pages_node[nid],
2128                nid, h->surplus_huge_pages_node[nid]);
2129}
2130
2131/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2132unsigned long hugetlb_total_pages(void)
2133{
2134        struct hstate *h = &default_hstate;
2135        return h->nr_huge_pages * pages_per_huge_page(h);
2136}
2137
2138static int hugetlb_acct_memory(struct hstate *h, long delta)
2139{
2140        int ret = -ENOMEM;
2141
2142        spin_lock(&hugetlb_lock);
2143        /*
2144         * When cpuset is configured, it breaks the strict hugetlb page
2145         * reservation as the accounting is done on a global variable. Such
2146         * reservation is completely rubbish in the presence of cpuset because
2147         * the reservation is not checked against page availability for the
2148         * current cpuset. Application can still potentially OOM'ed by kernel
2149         * with lack of free htlb page in cpuset that the task is in.
2150         * Attempt to enforce strict accounting with cpuset is almost
2151         * impossible (or too ugly) because cpuset is too fluid that
2152         * task or memory node can be dynamically moved between cpusets.
2153         *
2154         * The change of semantics for shared hugetlb mapping with cpuset is
2155         * undesirable. However, in order to preserve some of the semantics,
2156         * we fall back to check against current free page availability as
2157         * a best attempt and hopefully to minimize the impact of changing
2158         * semantics that cpuset has.
2159         */
2160        if (delta > 0) {
2161                if (gather_surplus_pages(h, delta) < 0)
2162                        goto out;
2163
2164                if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2165                        return_unused_surplus_pages(h, delta);
2166                        goto out;
2167                }
2168        }
2169
2170        ret = 0;
2171        if (delta < 0)
2172                return_unused_surplus_pages(h, (unsigned long) -delta);
2173
2174out:
2175        spin_unlock(&hugetlb_lock);
2176        return ret;
2177}
2178
2179static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2180{
2181        struct resv_map *reservations = vma_resv_map(vma);
2182
2183        /*
2184         * This new VMA should share its siblings reservation map if present.
2185         * The VMA will only ever have a valid reservation map pointer where
2186         * it is being copied for another still existing VMA.  As that VMA
2187         * has a reference to the reservation map it cannot disappear until
2188         * after this open call completes.  It is therefore safe to take a
2189         * new reference here without additional locking.
2190         */
2191        if (reservations)
2192                kref_get(&reservations->refs);
2193}
2194
2195static void resv_map_put(struct vm_area_struct *vma)
2196{
2197        struct resv_map *reservations = vma_resv_map(vma);
2198
2199        if (!reservations)
2200                return;
2201        kref_put(&reservations->refs, resv_map_release);
2202}
2203
2204static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2205{
2206        struct hstate *h = hstate_vma(vma);
2207        struct resv_map *reservations = vma_resv_map(vma);
2208        struct hugepage_subpool *spool = subpool_vma(vma);
2209        unsigned long reserve;
2210        unsigned long start;
2211        unsigned long end;
2212
2213        if (reservations) {
2214                start = vma_hugecache_offset(h, vma, vma->vm_start);
2215                end = vma_hugecache_offset(h, vma, vma->vm_end);
2216
2217                reserve = (end - start) -
2218                        region_count(&reservations->regions, start, end);
2219
2220                resv_map_put(vma);
2221
2222                if (reserve) {
2223                        hugetlb_acct_memory(h, -reserve);
2224                        hugepage_subpool_put_pages(spool, reserve);
2225                }
2226        }
2227}
2228
2229/*
2230 * We cannot handle pagefaults against hugetlb pages at all.  They cause
2231 * handle_mm_fault() to try to instantiate regular-sized pages in the
2232 * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2233 * this far.
2234 */
2235static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2236{
2237        BUG();
2238        return 0;
2239}
2240
2241const struct vm_operations_struct hugetlb_vm_ops = {
2242        .fault = hugetlb_vm_op_fault,
2243        .open = hugetlb_vm_op_open,
2244        .close = hugetlb_vm_op_close,
2245};
2246
2247static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2248                                int writable)
2249{
2250        pte_t entry;
2251
2252        if (writable) {
2253                entry =
2254                    pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2255        } else {
2256                entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2257        }
2258        entry = pte_mkyoung(entry);
2259        entry = pte_mkhuge(entry);
2260        entry = arch_make_huge_pte(entry, vma, page, writable);
2261
2262        return entry;
2263}
2264
2265static void set_huge_ptep_writable(struct vm_area_struct *vma,
2266                                   unsigned long address, pte_t *ptep)
2267{
2268        pte_t entry;
2269
2270        entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2271        if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2272                update_mmu_cache(vma, address, ptep);
2273}
2274
2275
2276int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2277                            struct vm_area_struct *vma)
2278{
2279        pte_t *src_pte, *dst_pte, entry;
2280        struct page *ptepage;
2281        unsigned long addr;
2282        int cow;
2283        struct hstate *h = hstate_vma(vma);
2284        unsigned long sz = huge_page_size(h);
2285
2286        cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2287
2288        for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2289                src_pte = huge_pte_offset(src, addr);
2290                if (!src_pte)
2291                        continue;
2292                dst_pte = huge_pte_alloc(dst, addr, sz);
2293                if (!dst_pte)
2294                        goto nomem;
2295
2296                /* If the pagetables are shared don't copy or take references */
2297                if (dst_pte == src_pte)
2298                        continue;
2299
2300                spin_lock(&dst->page_table_lock);
2301                spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2302                if (!huge_pte_none(huge_ptep_get(src_pte))) {
2303                        if (cow)
2304                                huge_ptep_set_wrprotect(src, addr, src_pte);
2305                        entry = huge_ptep_get(src_pte);
2306                        ptepage = pte_page(entry);
2307                        get_page(ptepage);
2308                        page_dup_rmap(ptepage);
2309                        set_huge_pte_at(dst, addr, dst_pte, entry);
2310                }
2311                spin_unlock(&src->page_table_lock);
2312                spin_unlock(&dst->page_table_lock);
2313        }
2314        return 0;
2315
2316nomem:
2317        return -ENOMEM;
2318}
2319
2320static int is_hugetlb_entry_migration(pte_t pte)
2321{
2322        swp_entry_t swp;
2323
2324        if (huge_pte_none(pte) || pte_present(pte))
2325                return 0;
2326        swp = pte_to_swp_entry(pte);
2327        if (non_swap_entry(swp) && is_migration_entry(swp))
2328                return 1;
2329        else
2330                return 0;
2331}
2332
2333static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2334{
2335        swp_entry_t swp;
2336
2337        if (huge_pte_none(pte) || pte_present(pte))
2338                return 0;
2339        swp = pte_to_swp_entry(pte);
2340        if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2341                return 1;
2342        else
2343                return 0;
2344}
2345
2346void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2347                            unsigned long start, unsigned long end,
2348                            struct page *ref_page)
2349{
2350        int force_flush = 0;
2351        struct mm_struct *mm = vma->vm_mm;
2352        unsigned long address;
2353        pte_t *ptep;
2354        pte_t pte;
2355        struct page *page;
2356        struct hstate *h = hstate_vma(vma);
2357        unsigned long sz = huge_page_size(h);
2358
2359        WARN_ON(!is_vm_hugetlb_page(vma));
2360        BUG_ON(start & ~huge_page_mask(h));
2361        BUG_ON(end & ~huge_page_mask(h));
2362
2363        tlb_start_vma(tlb, vma);
2364        mmu_notifier_invalidate_range_start(mm, start, end);
2365again:
2366        spin_lock(&mm->page_table_lock);
2367        for (address = start; address < end; address += sz) {
2368                ptep = huge_pte_offset(mm, address);
2369                if (!ptep)
2370                        continue;
2371
2372                if (huge_pmd_unshare(mm, &address, ptep))
2373                        continue;
2374
2375                pte = huge_ptep_get(ptep);
2376                if (huge_pte_none(pte))
2377                        continue;
2378
2379                /*
2380                 * HWPoisoned hugepage is already unmapped and dropped reference
2381                 */
2382                if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2383                        continue;
2384
2385                page = pte_page(pte);
2386                /*
2387                 * If a reference page is supplied, it is because a specific
2388                 * page is being unmapped, not a range. Ensure the page we
2389                 * are about to unmap is the actual page of interest.
2390                 */
2391                if (ref_page) {
2392                        if (page != ref_page)
2393                                continue;
2394
2395                        /*
2396                         * Mark the VMA as having unmapped its page so that
2397                         * future faults in this VMA will fail rather than
2398                         * looking like data was lost
2399                         */
2400                        set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2401                }
2402
2403                pte = huge_ptep_get_and_clear(mm, address, ptep);
2404                tlb_remove_tlb_entry(tlb, ptep, address);
2405                if (pte_dirty(pte))
2406                        set_page_dirty(page);
2407
2408                page_remove_rmap(page);
2409                force_flush = !__tlb_remove_page(tlb, page);
2410                if (force_flush)
2411                        break;
2412                /* Bail out after unmapping reference page if supplied */
2413                if (ref_page)
2414                        break;
2415        }
2416        spin_unlock(&mm->page_table_lock);
2417        /*
2418         * mmu_gather ran out of room to batch pages, we break out of
2419         * the PTE lock to avoid doing the potential expensive TLB invalidate
2420         * and page-free while holding it.
2421         */
2422        if (force_flush) {
2423                force_flush = 0;
2424                tlb_flush_mmu(tlb);
2425                if (address < end && !ref_page)
2426                        goto again;
2427        }
2428        mmu_notifier_invalidate_range_end(mm, start, end);
2429        tlb_end_vma(tlb, vma);
2430}
2431
2432void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2433                          struct vm_area_struct *vma, unsigned long start,
2434                          unsigned long end, struct page *ref_page)
2435{
2436        __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2437
2438        /*
2439         * Clear this flag so that x86's huge_pmd_share page_table_shareable
2440         * test will fail on a vma being torn down, and not grab a page table
2441         * on its way out.  We're lucky that the flag has such an appropriate
2442         * name, and can in fact be safely cleared here. We could clear it
2443         * before the __unmap_hugepage_range above, but all that's necessary
2444         * is to clear it before releasing the i_mmap_mutex. This works
2445         * because in the context this is called, the VMA is about to be
2446         * destroyed and the i_mmap_mutex is held.
2447         */
2448        vma->vm_flags &= ~VM_MAYSHARE;
2449}
2450
2451void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2452                          unsigned long end, struct page *ref_page)
2453{
2454        struct mm_struct *mm;
2455        struct mmu_gather tlb;
2456
2457        mm = vma->vm_mm;
2458
2459        tlb_gather_mmu(&tlb, mm, 0);
2460        __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2461        tlb_finish_mmu(&tlb, start, end);
2462}
2463
2464/*
2465 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2466 * mappping it owns the reserve page for. The intention is to unmap the page
2467 * from other VMAs and let the children be SIGKILLed if they are faulting the
2468 * same region.
2469 */
2470static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2471                                struct page *page, unsigned long address)
2472{
2473        struct hstate *h = hstate_vma(vma);
2474        struct vm_area_struct *iter_vma;
2475        struct address_space *mapping;
2476        struct prio_tree_iter iter;
2477        pgoff_t pgoff;
2478
2479        /*
2480         * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2481         * from page cache lookup which is in HPAGE_SIZE units.
2482         */
2483        address = address & huge_page_mask(h);
2484        pgoff = vma_hugecache_offset(h, vma, address);
2485        mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2486
2487        /*
2488         * Take the mapping lock for the duration of the table walk. As
2489         * this mapping should be shared between all the VMAs,
2490         * __unmap_hugepage_range() is called as the lock is already held
2491         */
2492        mutex_lock(&mapping->i_mmap_mutex);
2493        vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2494                /* Do not unmap the current VMA */
2495                if (iter_vma == vma)
2496                        continue;
2497
2498                /*
2499                 * Unmap the page from other VMAs without their own reserves.
2500                 * They get marked to be SIGKILLed if they fault in these
2501                 * areas. This is because a future no-page fault on this VMA
2502                 * could insert a zeroed page instead of the data existing
2503                 * from the time of fork. This would look like data corruption
2504                 */
2505                if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2506                        unmap_hugepage_range(iter_vma, address,
2507                                             address + huge_page_size(h), page);
2508        }
2509        mutex_unlock(&mapping->i_mmap_mutex);
2510
2511        return 1;
2512}
2513
2514/*
2515 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2516 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2517 * cannot race with other handlers or page migration.
2518 * Keep the pte_same checks anyway to make transition from the mutex easier.
2519 */
2520static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2521                        unsigned long address, pte_t *ptep, pte_t pte,
2522                        struct page *pagecache_page)
2523{
2524        struct hstate *h = hstate_vma(vma);
2525        struct page *old_page, *new_page;
2526        int avoidcopy;
2527        int outside_reserve = 0;
2528
2529        old_page = pte_page(pte);
2530
2531retry_avoidcopy:
2532        /* If no-one else is actually using this page, avoid the copy
2533         * and just make the page writable */
2534        avoidcopy = (page_mapcount(old_page) == 1);
2535        if (avoidcopy) {
2536                if (PageAnon(old_page))
2537                        page_move_anon_rmap(old_page, vma, address);
2538                set_huge_ptep_writable(vma, address, ptep);
2539                return 0;
2540        }
2541
2542        /*
2543         * If the process that created a MAP_PRIVATE mapping is about to
2544         * perform a COW due to a shared page count, attempt to satisfy
2545         * the allocation without using the existing reserves. The pagecache
2546         * page is used to determine if the reserve at this address was
2547         * consumed or not. If reserves were used, a partial faulted mapping
2548         * at the time of fork() could consume its reserves on COW instead
2549         * of the full address range.
2550         */
2551        if (!(vma->vm_flags & VM_MAYSHARE) &&
2552                        is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2553                        old_page != pagecache_page)
2554                outside_reserve = 1;
2555
2556        page_cache_get(old_page);
2557
2558        /* Drop page_table_lock as buddy allocator may be called */
2559        spin_unlock(&mm->page_table_lock);
2560        new_page = alloc_huge_page(vma, address, outside_reserve);
2561
2562        if (IS_ERR(new_page)) {
2563                long err = PTR_ERR(new_page);
2564                page_cache_release(old_page);
2565
2566                /*
2567                 * If a process owning a MAP_PRIVATE mapping fails to COW,
2568                 * it is due to references held by a child and an insufficient
2569                 * huge page pool. To guarantee the original mappers
2570                 * reliability, unmap the page from child processes. The child
2571                 * may get SIGKILLed if it later faults.
2572                 */
2573                if (outside_reserve) {
2574                        BUG_ON(huge_pte_none(pte));
2575                        if (unmap_ref_private(mm, vma, old_page, address)) {
2576                                BUG_ON(huge_pte_none(pte));
2577                                spin_lock(&mm->page_table_lock);
2578                                ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2579                                if (likely(pte_same(huge_ptep_get(ptep), pte)))
2580                                        goto retry_avoidcopy;
2581                                /*
2582                                 * race occurs while re-acquiring page_table_lock, and
2583                                 * our job is done.
2584                                 */
2585                                return 0;
2586                        }
2587                        WARN_ON_ONCE(1);
2588                }
2589
2590                /* Caller expects lock to be held */
2591                spin_lock(&mm->page_table_lock);
2592                if (err == -ENOMEM)
2593                        return VM_FAULT_OOM;
2594                else
2595                        return VM_FAULT_SIGBUS;
2596        }
2597
2598        /*
2599         * When the original hugepage is shared one, it does not have
2600         * anon_vma prepared.
2601         */
2602        if (unlikely(anon_vma_prepare(vma))) {
2603                page_cache_release(new_page);
2604                page_cache_release(old_page);
2605                /* Caller expects lock to be held */
2606                spin_lock(&mm->page_table_lock);
2607                return VM_FAULT_OOM;
2608        }
2609
2610        copy_user_huge_page(new_page, old_page, address, vma,
2611                            pages_per_huge_page(h));
2612        __SetPageUptodate(new_page);
2613
2614        /*
2615         * Retake the page_table_lock to check for racing updates
2616         * before the page tables are altered
2617         */
2618        spin_lock(&mm->page_table_lock);
2619        ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2620        if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2621                /* Break COW */
2622                mmu_notifier_invalidate_range_start(mm,
2623                        address & huge_page_mask(h),
2624                        (address & huge_page_mask(h)) + huge_page_size(h));
2625                huge_ptep_clear_flush(vma, address, ptep);
2626                set_huge_pte_at(mm, address, ptep,
2627                                make_huge_pte(vma, new_page, 1));
2628                page_remove_rmap(old_page);
2629                hugepage_add_new_anon_rmap(new_page, vma, address);
2630                /* Make the old page be freed below */
2631                new_page = old_page;
2632                mmu_notifier_invalidate_range_end(mm,
2633                        address & huge_page_mask(h),
2634                        (address & huge_page_mask(h)) + huge_page_size(h));
2635        }
2636        page_cache_release(new_page);
2637        page_cache_release(old_page);
2638        return 0;
2639}
2640
2641/* Return the pagecache page at a given address within a VMA */
2642static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2643                        struct vm_area_struct *vma, unsigned long address)
2644{
2645        struct address_space *mapping;
2646        pgoff_t idx;
2647
2648        mapping = vma->vm_file->f_mapping;
2649        idx = vma_hugecache_offset(h, vma, address);
2650
2651        return find_lock_page(mapping, idx);
2652}
2653
2654/*
2655 * Return whether there is a pagecache page to back given address within VMA.
2656 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2657 */
2658static bool hugetlbfs_pagecache_present(struct hstate *h,
2659                        struct vm_area_struct *vma, unsigned long address)
2660{
2661        struct address_space *mapping;
2662        pgoff_t idx;
2663        struct page *page;
2664
2665        mapping = vma->vm_file->f_mapping;
2666        idx = vma_hugecache_offset(h, vma, address);
2667
2668        page = find_get_page(mapping, idx);
2669        if (page)
2670                put_page(page);
2671        return page != NULL;
2672}
2673
2674static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2675                        unsigned long address, pte_t *ptep, unsigned int flags)
2676{
2677        struct hstate *h = hstate_vma(vma);
2678        int ret = VM_FAULT_SIGBUS;
2679        int anon_rmap = 0;
2680        pgoff_t idx;
2681        unsigned long size;
2682        struct page *page;
2683        struct address_space *mapping;
2684        pte_t new_pte;
2685
2686        /*
2687         * Currently, we are forced to kill the process in the event the
2688         * original mapper has unmapped pages from the child due to a failed
2689         * COW. Warn that such a situation has occurred as it may not be obvious
2690         */
2691        if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2692                printk(KERN_WARNING
2693                        "PID %d killed due to inadequate hugepage pool\n",
2694                        current->pid);
2695                return ret;
2696        }
2697
2698        mapping = vma->vm_file->f_mapping;
2699        idx = vma_hugecache_offset(h, vma, address);
2700
2701        /*
2702         * Use page lock to guard against racing truncation
2703         * before we get page_table_lock.
2704         */
2705retry:
2706        page = find_lock_page(mapping, idx);
2707        if (!page) {
2708                size = i_size_read(mapping->host) >> huge_page_shift(h);
2709                if (idx >= size)
2710                        goto out;
2711                page = alloc_huge_page(vma, address, 0);
2712                if (IS_ERR(page)) {
2713                        ret = PTR_ERR(page);
2714                        if (ret == -ENOMEM)
2715                                ret = VM_FAULT_OOM;
2716                        else
2717                                ret = VM_FAULT_SIGBUS;
2718                        goto out;
2719                }
2720                clear_huge_page(page, address, pages_per_huge_page(h));
2721                __SetPageUptodate(page);
2722
2723                if (vma->vm_flags & VM_MAYSHARE) {
2724                        int err;
2725                        struct inode *inode = mapping->host;
2726
2727                        err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2728                        if (err) {
2729                                put_page(page);
2730                                if (err == -EEXIST)
2731                                        goto retry;
2732                                goto out;
2733                        }
2734
2735                        spin_lock(&inode->i_lock);
2736                        inode->i_blocks += blocks_per_huge_page(h);
2737                        spin_unlock(&inode->i_lock);
2738                } else {
2739                        lock_page(page);
2740                        if (unlikely(anon_vma_prepare(vma))) {
2741                                ret = VM_FAULT_OOM;
2742                                goto backout_unlocked;
2743                        }
2744                        anon_rmap = 1;
2745                }
2746        } else {
2747                /*
2748                 * If memory error occurs between mmap() and fault, some process
2749                 * don't have hwpoisoned swap entry for errored virtual address.
2750                 * So we need to block hugepage fault by PG_hwpoison bit check.
2751                 */
2752                if (unlikely(PageHWPoison(page))) {
2753                        ret = VM_FAULT_HWPOISON |
2754                                VM_FAULT_SET_HINDEX(hstate_index(h));
2755                        goto backout_unlocked;
2756                }
2757        }
2758
2759        /*
2760         * If we are going to COW a private mapping later, we examine the
2761         * pending reservations for this page now. This will ensure that
2762         * any allocations necessary to record that reservation occur outside
2763         * the spinlock.
2764         */
2765        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2766                if (vma_needs_reservation(h, vma, address) < 0) {
2767                        ret = VM_FAULT_OOM;
2768                        goto backout_unlocked;
2769                }
2770
2771        spin_lock(&mm->page_table_lock);
2772        size = i_size_read(mapping->host) >> huge_page_shift(h);
2773        if (idx >= size)
2774                goto backout;
2775
2776        ret = 0;
2777        if (!huge_pte_none(huge_ptep_get(ptep)))
2778                goto backout;
2779
2780        if (anon_rmap)
2781                hugepage_add_new_anon_rmap(page, vma, address);
2782        else
2783                page_dup_rmap(page);
2784        new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2785                                && (vma->vm_flags & VM_SHARED)));
2786        set_huge_pte_at(mm, address, ptep, new_pte);
2787
2788        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2789                /* Optimization, do the COW without a second fault */
2790                ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2791        }
2792
2793        spin_unlock(&mm->page_table_lock);
2794        unlock_page(page);
2795out:
2796        return ret;
2797
2798backout:
2799        spin_unlock(&mm->page_table_lock);
2800backout_unlocked:
2801        unlock_page(page);
2802        put_page(page);
2803        goto out;
2804}
2805
2806int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2807                        unsigned long address, unsigned int flags)
2808{
2809        pte_t *ptep;
2810        pte_t entry;
2811        int ret;
2812        struct page *page = NULL;
2813        struct page *pagecache_page = NULL;
2814        static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2815        struct hstate *h = hstate_vma(vma);
2816
2817        address &= huge_page_mask(h);
2818
2819        ptep = huge_pte_offset(mm, address);
2820        if (ptep) {
2821                entry = huge_ptep_get(ptep);
2822                if (unlikely(is_hugetlb_entry_migration(entry))) {
2823                        migration_entry_wait(mm, (pmd_t *)ptep, address);
2824                        return 0;
2825                } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2826                        return VM_FAULT_HWPOISON_LARGE |
2827                                VM_FAULT_SET_HINDEX(hstate_index(h));
2828        }
2829
2830        ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2831        if (!ptep)
2832                return VM_FAULT_OOM;
2833
2834        /*
2835         * Serialize hugepage allocation and instantiation, so that we don't
2836         * get spurious allocation failures if two CPUs race to instantiate
2837         * the same page in the page cache.
2838         */
2839        mutex_lock(&hugetlb_instantiation_mutex);
2840        entry = huge_ptep_get(ptep);
2841        if (huge_pte_none(entry)) {
2842                ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2843                goto out_mutex;
2844        }
2845
2846        ret = 0;
2847
2848        /*
2849         * If we are going to COW the mapping later, we examine the pending
2850         * reservations for this page now. This will ensure that any
2851         * allocations necessary to record that reservation occur outside the
2852         * spinlock. For private mappings, we also lookup the pagecache
2853         * page now as it is used to determine if a reservation has been
2854         * consumed.
2855         */
2856        if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2857                if (vma_needs_reservation(h, vma, address) < 0) {
2858                        ret = VM_FAULT_OOM;
2859                        goto out_mutex;
2860                }
2861
2862                if (!(vma->vm_flags & VM_MAYSHARE))
2863                        pagecache_page = hugetlbfs_pagecache_page(h,
2864                                                                vma, address);
2865        }
2866
2867        /*
2868         * hugetlb_cow() requires page locks of pte_page(entry) and
2869         * pagecache_page, so here we need take the former one
2870         * when page != pagecache_page or !pagecache_page.
2871         * Note that locking order is always pagecache_page -> page,
2872         * so no worry about deadlock.
2873         */
2874        page = pte_page(entry);
2875        get_page(page);
2876        if (page != pagecache_page)
2877                lock_page(page);
2878
2879        spin_lock(&mm->page_table_lock);
2880        /* Check for a racing update before calling hugetlb_cow */
2881        if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2882                goto out_page_table_lock;
2883
2884
2885        if (flags & FAULT_FLAG_WRITE) {
2886                if (!pte_write(entry)) {
2887                        ret = hugetlb_cow(mm, vma, address, ptep, entry,
2888                                                        pagecache_page);
2889                        goto out_page_table_lock;
2890                }
2891                entry = pte_mkdirty(entry);
2892        }
2893        entry = pte_mkyoung(entry);
2894        if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2895                                                flags & FAULT_FLAG_WRITE))
2896                update_mmu_cache(vma, address, ptep);
2897
2898out_page_table_lock:
2899        spin_unlock(&mm->page_table_lock);
2900
2901        if (pagecache_page) {
2902                unlock_page(pagecache_page);
2903                put_page(pagecache_page);
2904        }
2905        if (page != pagecache_page)
2906                unlock_page(page);
2907        put_page(page);
2908
2909out_mutex:
2910        mutex_unlock(&hugetlb_instantiation_mutex);
2911
2912        return ret;
2913}
2914
2915/* Can be overriden by architectures */
2916__attribute__((weak)) struct page *
2917follow_huge_pud(struct mm_struct *mm, unsigned long address,
2918               pud_t *pud, int write)
2919{
2920        BUG();
2921        return NULL;
2922}
2923
2924int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2925                        struct page **pages, struct vm_area_struct **vmas,
2926                        unsigned long *position, int *length, int i,
2927                        unsigned int flags)
2928{
2929        unsigned long pfn_offset;
2930        unsigned long vaddr = *position;
2931        int remainder = *length;
2932        struct hstate *h = hstate_vma(vma);
2933
2934        spin_lock(&mm->page_table_lock);
2935        while (vaddr < vma->vm_end && remainder) {
2936                pte_t *pte;
2937                int absent;
2938                struct page *page;
2939
2940                /*
2941                 * Some archs (sparc64, sh*) have multiple pte_ts to
2942                 * each hugepage.  We have to make sure we get the
2943                 * first, for the page indexing below to work.
2944                 */
2945                pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2946                absent = !pte || huge_pte_none(huge_ptep_get(pte));
2947
2948                /*
2949                 * When coredumping, it suits get_dump_page if we just return
2950                 * an error where there's an empty slot with no huge pagecache
2951                 * to back it.  This way, we avoid allocating a hugepage, and
2952                 * the sparse dumpfile avoids allocating disk blocks, but its
2953                 * huge holes still show up with zeroes where they need to be.
2954                 */
2955                if (absent && (flags & FOLL_DUMP) &&
2956                    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2957                        remainder = 0;
2958                        break;
2959                }
2960
2961                if (absent ||
2962                    ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2963                        int ret;
2964
2965                        spin_unlock(&mm->page_table_lock);
2966                        ret = hugetlb_fault(mm, vma, vaddr,
2967                                (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2968                        spin_lock(&mm->page_table_lock);
2969                        if (!(ret & VM_FAULT_ERROR))
2970                                continue;
2971
2972                        remainder = 0;
2973                        break;
2974                }
2975
2976                pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2977                page = pte_page(huge_ptep_get(pte));
2978same_page:
2979                if (pages) {
2980                        pages[i] = mem_map_offset(page, pfn_offset);
2981                        get_page(pages[i]);
2982                }
2983
2984                if (vmas)
2985                        vmas[i] = vma;
2986
2987                vaddr += PAGE_SIZE;
2988                ++pfn_offset;
2989                --remainder;
2990                ++i;
2991                if (vaddr < vma->vm_end && remainder &&
2992                                pfn_offset < pages_per_huge_page(h)) {
2993                        /*
2994                         * We use pfn_offset to avoid touching the pageframes
2995                         * of this compound page.
2996                         */
2997                        goto same_page;
2998                }
2999        }
3000        spin_unlock(&mm->page_table_lock);
3001        *length = remainder;
3002        *position = vaddr;
3003
3004        return i ? i : -EFAULT;
3005}
3006
3007void hugetlb_change_protection(struct vm_area_struct *vma,
3008                unsigned long address, unsigned long end, pgprot_t newprot)
3009{
3010        struct mm_struct *mm = vma->vm_mm;
3011        unsigned long start = address;
3012        pte_t *ptep;
3013        pte_t pte;
3014        struct hstate *h = hstate_vma(vma);
3015
3016        BUG_ON(address >= end);
3017        flush_cache_range(vma, address, end);
3018
3019        mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3020        spin_lock(&mm->page_table_lock);
3021        for (; address < end; address += huge_page_size(h)) {
3022                ptep = huge_pte_offset(mm, address);
3023                if (!ptep)
3024                        continue;
3025                if (huge_pmd_unshare(mm, &address, ptep))
3026                        continue;
3027                if (!huge_pte_none(huge_ptep_get(ptep))) {
3028                        pte = huge_ptep_get_and_clear(mm, address, ptep);
3029                        pte = pte_mkhuge(pte_modify(pte, newprot));
3030                        set_huge_pte_at(mm, address, ptep, pte);
3031                }
3032        }
3033        spin_unlock(&mm->page_table_lock);
3034        /*
3035         * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3036         * may have cleared our pud entry and done put_page on the page table:
3037         * once we release i_mmap_mutex, another task can do the final put_page
3038         * and that page table be reused and filled with junk.
3039         */
3040        flush_tlb_range(vma, start, end);
3041        mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3042}
3043
3044int hugetlb_reserve_pages(struct inode *inode,
3045                                        long from, long to,
3046                                        struct vm_area_struct *vma,
3047                                        vm_flags_t vm_flags)
3048{
3049        long ret, chg;
3050        struct hstate *h = hstate_inode(inode);
3051        struct hugepage_subpool *spool = subpool_inode(inode);
3052
3053        /*
3054         * Only apply hugepage reservation if asked. At fault time, an
3055         * attempt will be made for VM_NORESERVE to allocate a page
3056         * without using reserves
3057         */
3058        if (vm_flags & VM_NORESERVE)
3059                return 0;
3060
3061        /*
3062         * Shared mappings base their reservation on the number of pages that
3063         * are already allocated on behalf of the file. Private mappings need
3064         * to reserve the full area even if read-only as mprotect() may be
3065         * called to make the mapping read-write. Assume !vma is a shm mapping
3066         */
3067        if (!vma || vma->vm_flags & VM_MAYSHARE)
3068                chg = region_chg(&inode->i_mapping->private_list, from, to);
3069        else {
3070                struct resv_map *resv_map = resv_map_alloc();
3071                if (!resv_map)
3072                        return -ENOMEM;
3073
3074                chg = to - from;
3075
3076                set_vma_resv_map(vma, resv_map);
3077                set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3078        }
3079
3080        if (chg < 0) {
3081                ret = chg;
3082                goto out_err;
3083        }
3084
3085        /* There must be enough pages in the subpool for the mapping */
3086        if (hugepage_subpool_get_pages(spool, chg)) {
3087                ret = -ENOSPC;
3088                goto out_err;
3089        }
3090
3091        /*
3092         * Check enough hugepages are available for the reservation.
3093         * Hand the pages back to the subpool if there are not
3094         */
3095        ret = hugetlb_acct_memory(h, chg);
3096        if (ret < 0) {
3097                hugepage_subpool_put_pages(spool, chg);
3098                goto out_err;
3099        }
3100
3101        /*
3102         * Account for the reservations made. Shared mappings record regions
3103         * that have reservations as they are shared by multiple VMAs.
3104         * When the last VMA disappears, the region map says how much
3105         * the reservation was and the page cache tells how much of
3106         * the reservation was consumed. Private mappings are per-VMA and
3107         * only the consumed reservations are tracked. When the VMA
3108         * disappears, the original reservation is the VMA size and the
3109         * consumed reservations are stored in the map. Hence, nothing
3110         * else has to be done for private mappings here
3111         */
3112        if (!vma || vma->vm_flags & VM_MAYSHARE)
3113                region_add(&inode->i_mapping->private_list, from, to);
3114        return 0;
3115out_err:
3116        if (vma)
3117                resv_map_put(vma);
3118        return ret;
3119}
3120
3121void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3122{
3123        struct hstate *h = hstate_inode(inode);
3124        long chg = region_truncate(&inode->i_mapping->private_list, offset);
3125        struct hugepage_subpool *spool = subpool_inode(inode);
3126
3127        spin_lock(&inode->i_lock);
3128        inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3129        spin_unlock(&inode->i_lock);
3130
3131        hugepage_subpool_put_pages(spool, (chg - freed));
3132        hugetlb_acct_memory(h, -(chg - freed));
3133}
3134
3135#ifdef CONFIG_MEMORY_FAILURE
3136
3137/* Should be called in hugetlb_lock */
3138static int is_hugepage_on_freelist(struct page *hpage)
3139{
3140        struct page *page;
3141        struct page *tmp;
3142        struct hstate *h = page_hstate(hpage);
3143        int nid = page_to_nid(hpage);
3144
3145        list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3146                if (page == hpage)
3147                        return 1;
3148        return 0;
3149}
3150
3151/*
3152 * This function is called from memory failure code.
3153 * Assume the caller holds page lock of the head page.
3154 */
3155int dequeue_hwpoisoned_huge_page(struct page *hpage)
3156{
3157        struct hstate *h = page_hstate(hpage);
3158        int nid = page_to_nid(hpage);
3159        int ret = -EBUSY;
3160
3161        spin_lock(&hugetlb_lock);
3162        if (is_hugepage_on_freelist(hpage)) {
3163                list_del(&hpage->lru);
3164                set_page_refcounted(hpage);
3165                h->free_huge_pages--;
3166                h->free_huge_pages_node[nid]--;
3167                ret = 0;
3168        }
3169        spin_unlock(&hugetlb_lock);
3170        return ret;
3171}
3172#endif
3173
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.