linux/mm/hugetlb.c
<<
>>
Prefs
   1/*
   2 * Generic hugetlb support.
   3 * (C) Nadia Yvette Chambers, 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 "internal.h"
  34
  35const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  36static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  37unsigned long hugepages_treat_as_movable;
  38
  39int hugetlb_max_hstate __read_mostly;
  40unsigned int default_hstate_idx;
  41struct hstate hstates[HUGE_MAX_HSTATE];
  42
  43__initdata LIST_HEAD(huge_boot_pages);
  44
  45/* for command line parsing */
  46static struct hstate * __initdata parsed_hstate;
  47static unsigned long __initdata default_hstate_max_huge_pages;
  48static unsigned long __initdata default_hstate_size;
  49
  50/*
  51 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  52 */
  53DEFINE_SPINLOCK(hugetlb_lock);
  54
  55static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
  56{
  57        bool free = (spool->count == 0) && (spool->used_hpages == 0);
  58
  59        spin_unlock(&spool->lock);
  60
  61        /* If no pages are used, and no other handles to the subpool
  62         * remain, free the subpool the subpool remain */
  63        if (free)
  64                kfree(spool);
  65}
  66
  67struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
  68{
  69        struct hugepage_subpool *spool;
  70
  71        spool = kmalloc(sizeof(*spool), GFP_KERNEL);
  72        if (!spool)
  73                return NULL;
  74
  75        spin_lock_init(&spool->lock);
  76        spool->count = 1;
  77        spool->max_hpages = nr_blocks;
  78        spool->used_hpages = 0;
  79
  80        return spool;
  81}
  82
  83void hugepage_put_subpool(struct hugepage_subpool *spool)
  84{
  85        spin_lock(&spool->lock);
  86        BUG_ON(!spool->count);
  87        spool->count--;
  88        unlock_or_release_subpool(spool);
  89}
  90
  91static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
  92                                      long delta)
  93{
  94        int ret = 0;
  95
  96        if (!spool)
  97                return 0;
  98
  99        spin_lock(&spool->lock);
 100        if ((spool->used_hpages + delta) <= spool->max_hpages) {
 101                spool->used_hpages += delta;
 102        } else {
 103                ret = -ENOMEM;
 104        }
 105        spin_unlock(&spool->lock);
 106
 107        return ret;
 108}
 109
 110static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
 111                                       long delta)
 112{
 113        if (!spool)
 114                return;
 115
 116        spin_lock(&spool->lock);
 117        spool->used_hpages -= delta;
 118        /* If hugetlbfs_put_super couldn't free spool due to
 119        * an outstanding quota reference, free it now. */
 120        unlock_or_release_subpool(spool);
 121}
 122
 123static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
 124{
 125        return HUGETLBFS_SB(inode->i_sb)->spool;
 126}
 127
 128static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
 129{
 130        return subpool_inode(file_inode(vma->vm_file));
 131}
 132
 133/*
 134 * Region tracking -- allows tracking of reservations and instantiated pages
 135 *                    across the pages in a mapping.
 136 *
 137 * The region data structures are protected by a combination of the mmap_sem
 138 * and the hugetlb_instantion_mutex.  To access or modify a region the caller
 139 * must either hold the mmap_sem for write, or the mmap_sem for read and
 140 * the hugetlb_instantiation mutex:
 141 *
 142 *      down_write(&mm->mmap_sem);
 143 * or
 144 *      down_read(&mm->mmap_sem);
 145 *      mutex_lock(&hugetlb_instantiation_mutex);
 146 */
 147struct file_region {
 148        struct list_head link;
 149        long from;
 150        long to;
 151};
 152
 153static long region_add(struct list_head *head, long f, long t)
 154{
 155        struct file_region *rg, *nrg, *trg;
 156
 157        /* Locate the region we are either in or before. */
 158        list_for_each_entry(rg, head, link)
 159                if (f <= rg->to)
 160                        break;
 161
 162        /* Round our left edge to the current segment if it encloses us. */
 163        if (f > rg->from)
 164                f = rg->from;
 165
 166        /* Check for and consume any regions we now overlap with. */
 167        nrg = rg;
 168        list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 169                if (&rg->link == head)
 170                        break;
 171                if (rg->from > t)
 172                        break;
 173
 174                /* If this area reaches higher then extend our area to
 175                 * include it completely.  If this is not the first area
 176                 * which we intend to reuse, free it. */
 177                if (rg->to > t)
 178                        t = rg->to;
 179                if (rg != nrg) {
 180                        list_del(&rg->link);
 181                        kfree(rg);
 182                }
 183        }
 184        nrg->from = f;
 185        nrg->to = t;
 186        return 0;
 187}
 188
 189static long region_chg(struct list_head *head, long f, long t)
 190{
 191        struct file_region *rg, *nrg;
 192        long chg = 0;
 193
 194        /* Locate the region we are before or in. */
 195        list_for_each_entry(rg, head, link)
 196                if (f <= rg->to)
 197                        break;
 198
 199        /* If we are below the current region then a new region is required.
 200         * Subtle, allocate a new region at the position but make it zero
 201         * size such that we can guarantee to record the reservation. */
 202        if (&rg->link == head || t < rg->from) {
 203                nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 204                if (!nrg)
 205                        return -ENOMEM;
 206                nrg->from = f;
 207                nrg->to   = f;
 208                INIT_LIST_HEAD(&nrg->link);
 209                list_add(&nrg->link, rg->link.prev);
 210
 211                return t - f;
 212        }
 213
 214        /* Round our left edge to the current segment if it encloses us. */
 215        if (f > rg->from)
 216                f = rg->from;
 217        chg = t - f;
 218
 219        /* Check for and consume any regions we now overlap with. */
 220        list_for_each_entry(rg, rg->link.prev, link) {
 221                if (&rg->link == head)
 222                        break;
 223                if (rg->from > t)
 224                        return chg;
 225
 226                /* We overlap with this area, if it extends further than
 227                 * us then we must extend ourselves.  Account for its
 228                 * existing reservation. */
 229                if (rg->to > t) {
 230                        chg += rg->to - t;
 231                        t = rg->to;
 232                }
 233                chg -= rg->to - rg->from;
 234        }
 235        return chg;
 236}
 237
 238static long region_truncate(struct list_head *head, long end)
 239{
 240        struct file_region *rg, *trg;
 241        long chg = 0;
 242
 243        /* Locate the region we are either in or before. */
 244        list_for_each_entry(rg, head, link)
 245                if (end <= rg->to)
 246                        break;
 247        if (&rg->link == head)
 248                return 0;
 249
 250        /* If we are in the middle of a region then adjust it. */
 251        if (end > rg->from) {
 252                chg = rg->to - end;
 253                rg->to = end;
 254                rg = list_entry(rg->link.next, typeof(*rg), link);
 255        }
 256
 257        /* Drop any remaining regions. */
 258        list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 259                if (&rg->link == head)
 260                        break;
 261                chg += rg->to - rg->from;
 262                list_del(&rg->link);
 263                kfree(rg);
 264        }
 265        return chg;
 266}
 267
 268static long region_count(struct list_head *head, long f, long t)
 269{
 270        struct file_region *rg;
 271        long chg = 0;
 272
 273        /* Locate each segment we overlap with, and count that overlap. */
 274        list_for_each_entry(rg, head, link) {
 275                long seg_from;
 276                long seg_to;
 277
 278                if (rg->to <= f)
 279                        continue;
 280                if (rg->from >= t)
 281                        break;
 282
 283                seg_from = max(rg->from, f);
 284                seg_to = min(rg->to, t);
 285
 286                chg += seg_to - seg_from;
 287        }
 288
 289        return chg;
 290}
 291
 292/*
 293 * Convert the address within this vma to the page offset within
 294 * the mapping, in pagecache page units; huge pages here.
 295 */
 296static pgoff_t vma_hugecache_offset(struct hstate *h,
 297                        struct vm_area_struct *vma, unsigned long address)
 298{
 299        return ((address - vma->vm_start) >> huge_page_shift(h)) +
 300                        (vma->vm_pgoff >> huge_page_order(h));
 301}
 302
 303pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
 304                                     unsigned long address)
 305{
 306        return vma_hugecache_offset(hstate_vma(vma), vma, address);
 307}
 308
 309/*
 310 * Return the size of the pages allocated when backing a VMA. In the majority
 311 * cases this will be same size as used by the page table entries.
 312 */
 313unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 314{
 315        struct hstate *hstate;
 316
 317        if (!is_vm_hugetlb_page(vma))
 318                return PAGE_SIZE;
 319
 320        hstate = hstate_vma(vma);
 321
 322        return 1UL << (hstate->order + PAGE_SHIFT);
 323}
 324EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
 325
 326/*
 327 * Return the page size being used by the MMU to back a VMA. In the majority
 328 * of cases, the page size used by the kernel matches the MMU size. On
 329 * architectures where it differs, an architecture-specific version of this
 330 * function is required.
 331 */
 332#ifndef vma_mmu_pagesize
 333unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 334{
 335        return vma_kernel_pagesize(vma);
 336}
 337#endif
 338
 339/*
 340 * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 341 * bits of the reservation map pointer, which are always clear due to
 342 * alignment.
 343 */
 344#define HPAGE_RESV_OWNER    (1UL << 0)
 345#define HPAGE_RESV_UNMAPPED (1UL << 1)
 346#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 347
 348/*
 349 * These helpers are used to track how many pages are reserved for
 350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 351 * is guaranteed to have their future faults succeed.
 352 *
 353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 354 * the reserve counters are updated with the hugetlb_lock held. It is safe
 355 * to reset the VMA at fork() time as it is not in use yet and there is no
 356 * chance of the global counters getting corrupted as a result of the values.
 357 *
 358 * The private mapping reservation is represented in a subtly different
 359 * manner to a shared mapping.  A shared mapping has a region map associated
 360 * with the underlying file, this region map represents the backing file
 361 * pages which have ever had a reservation assigned which this persists even
 362 * after the page is instantiated.  A private mapping has a region map
 363 * associated with the original mmap which is attached to all VMAs which
 364 * reference it, this region map represents those offsets which have consumed
 365 * reservation ie. where pages have been instantiated.
 366 */
 367static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 368{
 369        return (unsigned long)vma->vm_private_data;
 370}
 371
 372static void set_vma_private_data(struct vm_area_struct *vma,
 373                                                        unsigned long value)
 374{
 375        vma->vm_private_data = (void *)value;
 376}
 377
 378struct resv_map {
 379        struct kref refs;
 380        struct list_head regions;
 381};
 382
 383static struct resv_map *resv_map_alloc(void)
 384{
 385        struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 386        if (!resv_map)
 387                return NULL;
 388
 389        kref_init(&resv_map->refs);
 390        INIT_LIST_HEAD(&resv_map->regions);
 391
 392        return resv_map;
 393}
 394
 395static void resv_map_release(struct kref *ref)
 396{
 397        struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 398
 399        /* Clear out any active regions before we release the map. */
 400        region_truncate(&resv_map->regions, 0);
 401        kfree(resv_map);
 402}
 403
 404static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 405{
 406        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 407        if (!(vma->vm_flags & VM_MAYSHARE))
 408                return (struct resv_map *)(get_vma_private_data(vma) &
 409                                                        ~HPAGE_RESV_MASK);
 410        return NULL;
 411}
 412
 413static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 414{
 415        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 416        VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 417
 418        set_vma_private_data(vma, (get_vma_private_data(vma) &
 419                                HPAGE_RESV_MASK) | (unsigned long)map);
 420}
 421
 422static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 423{
 424        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 425        VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 426
 427        set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 428}
 429
 430static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 431{
 432        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 433
 434        return (get_vma_private_data(vma) & flag) != 0;
 435}
 436
 437/* Decrement the reserved pages in the hugepage pool by one */
 438static void decrement_hugepage_resv_vma(struct hstate *h,
 439                        struct vm_area_struct *vma)
 440{
 441        if (vma->vm_flags & VM_NORESERVE)
 442                return;
 443
 444        if (vma->vm_flags & VM_MAYSHARE) {
 445                /* Shared mappings always use reserves */
 446                h->resv_huge_pages--;
 447        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 448                /*
 449                 * Only the process that called mmap() has reserves for
 450                 * private mappings.
 451                 */
 452                h->resv_huge_pages--;
 453        }
 454}
 455
 456/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 457void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 458{
 459        VM_BUG_ON(!is_vm_hugetlb_page(vma));
 460        if (!(vma->vm_flags & VM_MAYSHARE))
 461                vma->vm_private_data = (void *)0;
 462}
 463
 464/* Returns true if the VMA has associated reserve pages */
 465static int vma_has_reserves(struct vm_area_struct *vma)
 466{
 467        if (vma->vm_flags & VM_MAYSHARE)
 468                return 1;
 469        if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 470                return 1;
 471        return 0;
 472}
 473
 474static void copy_gigantic_page(struct page *dst, struct page *src)
 475{
 476        int i;
 477        struct hstate *h = page_hstate(src);
 478        struct page *dst_base = dst;
 479        struct page *src_base = src;
 480
 481        for (i = 0; i < pages_per_huge_page(h); ) {
 482                cond_resched();
 483                copy_highpage(dst, src);
 484
 485                i++;
 486                dst = mem_map_next(dst, dst_base, i);
 487                src = mem_map_next(src, src_base, i);
 488        }
 489}
 490
 491void copy_huge_page(struct page *dst, struct page *src)
 492{
 493        int i;
 494        struct hstate *h = page_hstate(src);
 495
 496        if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
 497                copy_gigantic_page(dst, src);
 498                return;
 499        }
 500
 501        might_sleep();
 502        for (i = 0; i < pages_per_huge_page(h); i++) {
 503                cond_resched();
 504                copy_highpage(dst + i, src + i);
 505        }
 506}
 507
 508static void enqueue_huge_page(struct hstate *h, struct page *page)
 509{
 510        int nid = page_to_nid(page);
 511        list_move(&page->lru, &h->hugepage_freelists[nid]);
 512        h->free_huge_pages++;
 513        h->free_huge_pages_node[nid]++;
 514}
 515
 516static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 517{
 518        struct page *page;
 519
 520        if (list_empty(&h->hugepage_freelists[nid]))
 521                return NULL;
 522        page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
 523        list_move(&page->lru, &h->hugepage_activelist);
 524        set_page_refcounted(page);
 525        h->free_huge_pages--;
 526        h->free_huge_pages_node[nid]--;
 527        return page;
 528}
 529
 530static struct page *dequeue_huge_page_vma(struct hstate *h,
 531                                struct vm_area_struct *vma,
 532                                unsigned long address, int avoid_reserve)
 533{
 534        struct page *page = NULL;
 535        struct mempolicy *mpol;
 536        nodemask_t *nodemask;
 537        struct zonelist *zonelist;
 538        struct zone *zone;
 539        struct zoneref *z;
 540        unsigned int cpuset_mems_cookie;
 541
 542retry_cpuset:
 543        cpuset_mems_cookie = get_mems_allowed();
 544        zonelist = huge_zonelist(vma, address,
 545                                        htlb_alloc_mask, &mpol, &nodemask);
 546        /*
 547         * A child process with MAP_PRIVATE mappings created by their parent
 548         * have no page reserves. This check ensures that reservations are
 549         * not "stolen". The child may still get SIGKILLed
 550         */
 551        if (!vma_has_reserves(vma) &&
 552                        h->free_huge_pages - h->resv_huge_pages == 0)
 553                goto err;
 554
 555        /* If reserves cannot be used, ensure enough pages are in the pool */
 556        if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 557                goto err;
 558
 559        for_each_zone_zonelist_nodemask(zone, z, zonelist,
 560                                                MAX_NR_ZONES - 1, nodemask) {
 561                if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
 562                        page = dequeue_huge_page_node(h, zone_to_nid(zone));
 563                        if (page) {
 564                                if (!avoid_reserve)
 565                                        decrement_hugepage_resv_vma(h, vma);
 566                                break;
 567                        }
 568                }
 569        }
 570
 571        mpol_cond_put(mpol);
 572        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
 573                goto retry_cpuset;
 574        return page;
 575
 576err:
 577        mpol_cond_put(mpol);
 578        return NULL;
 579}
 580
 581static void update_and_free_page(struct hstate *h, struct page *page)
 582{
 583        int i;
 584
 585        VM_BUG_ON(h->order >= MAX_ORDER);
 586
 587        h->nr_huge_pages--;
 588        h->nr_huge_pages_node[page_to_nid(page)]--;
 589        for (i = 0; i < pages_per_huge_page(h); i++) {
 590                page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
 591                                1 << PG_referenced | 1 << PG_dirty |
 592                                1 << PG_active | 1 << PG_reserved |
 593                                1 << PG_private | 1 << PG_writeback);
 594        }
 595        VM_BUG_ON(hugetlb_cgroup_from_page(page));
 596        set_compound_page_dtor(page, NULL);
 597        set_page_refcounted(page);
 598        arch_release_hugepage(page);
 599        __free_pages(page, huge_page_order(h));
 600}
 601
 602struct hstate *size_to_hstate(unsigned long size)
 603{
 604        struct hstate *h;
 605
 606        for_each_hstate(h) {
 607                if (huge_page_size(h) == size)
 608                        return h;
 609        }
 610        return NULL;
 611}
 612
 613static void free_huge_page(struct page *page)
 614{
 615        /*
 616         * Can't pass hstate in here because it is called from the
 617         * compound page destructor.
 618         */
 619        struct hstate *h = page_hstate(page);
 620        int nid = page_to_nid(page);
 621        struct hugepage_subpool *spool =
 622                (struct hugepage_subpool *)page_private(page);
 623
 624        set_page_private(page, 0);
 625        page->mapping = NULL;
 626        BUG_ON(page_count(page));
 627        BUG_ON(page_mapcount(page));
 628
 629        spin_lock(&hugetlb_lock);
 630        hugetlb_cgroup_uncharge_page(hstate_index(h),
 631                                     pages_per_huge_page(h), page);
 632        if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 633                /* remove the page from active list */
 634                list_del(&page->lru);
 635                update_and_free_page(h, page);
 636                h->surplus_huge_pages--;
 637                h->surplus_huge_pages_node[nid]--;
 638        } else {
 639                arch_clear_hugepage_flags(page);
 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
 674/*
 675 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
 676 * transparent huge pages.  See the PageTransHuge() documentation for more
 677 * details.
 678 */
 679int PageHuge(struct page *page)
 680{
 681        compound_page_dtor *dtor;
 682
 683        if (!PageCompound(page))
 684                return 0;
 685
 686        page = compound_head(page);
 687        dtor = get_compound_page_dtor(page);
 688
 689        return dtor == free_huge_page;
 690}
 691EXPORT_SYMBOL_GPL(PageHuge);
 692
 693pgoff_t __basepage_index(struct page *page)
 694{
 695        struct page *page_head = compound_head(page);
 696        pgoff_t index = page_index(page_head);
 697        unsigned long compound_idx;
 698
 699        if (!PageHuge(page_head))
 700                return page_index(page);
 701
 702        if (compound_order(page_head) >= MAX_ORDER)
 703                compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
 704        else
 705                compound_idx = page - page_head;
 706
 707        return (index << compound_order(page_head)) + compound_idx;
 708}
 709
 710static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 711{
 712        struct page *page;
 713
 714        if (h->order >= MAX_ORDER)
 715                return NULL;
 716
 717        page = alloc_pages_exact_node(nid,
 718                htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 719                                                __GFP_REPEAT|__GFP_NOWARN,
 720                huge_page_order(h));
 721        if (page) {
 722                if (arch_prepare_hugepage(page)) {
 723                        __free_pages(page, huge_page_order(h));
 724                        return NULL;
 725                }
 726                prep_new_huge_page(h, page, nid);
 727        }
 728
 729        return page;
 730}
 731
 732/*
 733 * common helper functions for hstate_next_node_to_{alloc|free}.
 734 * We may have allocated or freed a huge page based on a different
 735 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
 736 * be outside of *nodes_allowed.  Ensure that we use an allowed
 737 * node for alloc or free.
 738 */
 739static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 740{
 741        nid = next_node(nid, *nodes_allowed);
 742        if (nid == MAX_NUMNODES)
 743                nid = first_node(*nodes_allowed);
 744        VM_BUG_ON(nid >= MAX_NUMNODES);
 745
 746        return nid;
 747}
 748
 749static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
 750{
 751        if (!node_isset(nid, *nodes_allowed))
 752                nid = next_node_allowed(nid, nodes_allowed);
 753        return nid;
 754}
 755
 756/*
 757 * returns the previously saved node ["this node"] from which to
 758 * allocate a persistent huge page for the pool and advance the
 759 * next node from which to allocate, handling wrap at end of node
 760 * mask.
 761 */
 762static int hstate_next_node_to_alloc(struct hstate *h,
 763                                        nodemask_t *nodes_allowed)
 764{
 765        int nid;
 766
 767        VM_BUG_ON(!nodes_allowed);
 768
 769        nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
 770        h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
 771
 772        return nid;
 773}
 774
 775static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
 776{
 777        struct page *page;
 778        int start_nid;
 779        int next_nid;
 780        int ret = 0;
 781
 782        start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 783        next_nid = start_nid;
 784
 785        do {
 786                page = alloc_fresh_huge_page_node(h, next_nid);
 787                if (page) {
 788                        ret = 1;
 789                        break;
 790                }
 791                next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 792        } while (next_nid != start_nid);
 793
 794        if (ret)
 795                count_vm_event(HTLB_BUDDY_PGALLOC);
 796        else
 797                count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 798
 799        return ret;
 800}
 801
 802/*
 803 * helper for free_pool_huge_page() - return the previously saved
 804 * node ["this node"] from which to free a huge page.  Advance the
 805 * next node id whether or not we find a free huge page to free so
 806 * that the next attempt to free addresses the next node.
 807 */
 808static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
 809{
 810        int nid;
 811
 812        VM_BUG_ON(!nodes_allowed);
 813
 814        nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
 815        h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
 816
 817        return nid;
 818}
 819
 820/*
 821 * Free huge page from pool from next node to free.
 822 * Attempt to keep persistent huge pages more or less
 823 * balanced over allowed nodes.
 824 * Called with hugetlb_lock locked.
 825 */
 826static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
 827                                                         bool acct_surplus)
 828{
 829        int start_nid;
 830        int next_nid;
 831        int ret = 0;
 832
 833        start_nid = hstate_next_node_to_free(h, nodes_allowed);
 834        next_nid = start_nid;
 835
 836        do {
 837                /*
 838                 * If we're returning unused surplus pages, only examine
 839                 * nodes with surplus pages.
 840                 */
 841                if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
 842                    !list_empty(&h->hugepage_freelists[next_nid])) {
 843                        struct page *page =
 844                                list_entry(h->hugepage_freelists[next_nid].next,
 845                                          struct page, lru);
 846                        list_del(&page->lru);
 847                        h->free_huge_pages--;
 848                        h->free_huge_pages_node[next_nid]--;
 849                        if (acct_surplus) {
 850                                h->surplus_huge_pages--;
 851                                h->surplus_huge_pages_node[next_nid]--;
 852                        }
 853                        update_and_free_page(h, page);
 854                        ret = 1;
 855                        break;
 856                }
 857                next_nid = hstate_next_node_to_free(h, nodes_allowed);
 858        } while (next_nid != start_nid);
 859
 860        return ret;
 861}
 862
 863static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 864{
 865        struct page *page;
 866        unsigned int r_nid;
 867
 868        if (h->order >= MAX_ORDER)
 869                return NULL;
 870
 871        /*
 872         * Assume we will successfully allocate the surplus page to
 873         * prevent racing processes from causing the surplus to exceed
 874         * overcommit
 875         *
 876         * This however introduces a different race, where a process B
 877         * tries to grow the static hugepage pool while alloc_pages() is
 878         * called by process A. B will only examine the per-node
 879         * counters in determining if surplus huge pages can be
 880         * converted to normal huge pages in adjust_pool_surplus(). A
 881         * won't be able to increment the per-node counter, until the
 882         * lock is dropped by B, but B doesn't drop hugetlb_lock until
 883         * no more huge pages can be converted from surplus to normal
 884         * state (and doesn't try to convert again). Thus, we have a
 885         * case where a surplus huge page exists, the pool is grown, and
 886         * the surplus huge page still exists after, even though it
 887         * should just have been converted to a normal huge page. This
 888         * does not leak memory, though, as the hugepage will be freed
 889         * once it is out of use. It also does not allow the counters to
 890         * go out of whack in adjust_pool_surplus() as we don't modify
 891         * the node values until we've gotten the hugepage and only the
 892         * per-node value is checked there.
 893         */
 894        spin_lock(&hugetlb_lock);
 895        if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 896                spin_unlock(&hugetlb_lock);
 897                return NULL;
 898        } else {
 899                h->nr_huge_pages++;
 900                h->surplus_huge_pages++;
 901        }
 902        spin_unlock(&hugetlb_lock);
 903
 904        if (nid == NUMA_NO_NODE)
 905                page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 906                                   __GFP_REPEAT|__GFP_NOWARN,
 907                                   huge_page_order(h));
 908        else
 909                page = alloc_pages_exact_node(nid,
 910                        htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 911                        __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
 912
 913        if (page && arch_prepare_hugepage(page)) {
 914                __free_pages(page, huge_page_order(h));
 915                page = NULL;
 916        }
 917
 918        spin_lock(&hugetlb_lock);
 919        if (page) {
 920                INIT_LIST_HEAD(&page->lru);
 921                r_nid = page_to_nid(page);
 922                set_compound_page_dtor(page, free_huge_page);
 923                set_hugetlb_cgroup(page, NULL);
 924                /*
 925                 * We incremented the global counters already
 926                 */
 927                h->nr_huge_pages_node[r_nid]++;
 928                h->surplus_huge_pages_node[r_nid]++;
 929                __count_vm_event(HTLB_BUDDY_PGALLOC);
 930        } else {
 931                h->nr_huge_pages--;
 932                h->surplus_huge_pages--;
 933                __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 934        }
 935        spin_unlock(&hugetlb_lock);
 936
 937        return page;
 938}
 939
 940/*
 941 * This allocation function is useful in the context where vma is irrelevant.
 942 * E.g. soft-offlining uses this function because it only cares physical
 943 * address of error page.
 944 */
 945struct page *alloc_huge_page_node(struct hstate *h, int nid)
 946{
 947        struct page *page;
 948
 949        spin_lock(&hugetlb_lock);
 950        page = dequeue_huge_page_node(h, nid);
 951        spin_unlock(&hugetlb_lock);
 952
 953        if (!page)
 954                page = alloc_buddy_huge_page(h, nid);
 955
 956        return page;
 957}
 958
 959/*
 960 * Increase the hugetlb pool such that it can accommodate a reservation
 961 * of size 'delta'.
 962 */
 963static int gather_surplus_pages(struct hstate *h, int delta)
 964{
 965        struct list_head surplus_list;
 966        struct page *page, *tmp;
 967        int ret, i;
 968        int needed, allocated;
 969        bool alloc_ok = true;
 970
 971        needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
 972        if (needed <= 0) {
 973                h->resv_huge_pages += delta;
 974                return 0;
 975        }
 976
 977        allocated = 0;
 978        INIT_LIST_HEAD(&surplus_list);
 979
 980        ret = -ENOMEM;
 981retry:
 982        spin_unlock(&hugetlb_lock);
 983        for (i = 0; i < needed; i++) {
 984                page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
 985                if (!page) {
 986                        alloc_ok = false;
 987                        break;
 988                }
 989                list_add(&page->lru, &surplus_list);
 990        }
 991        allocated += i;
 992
 993        /*
 994         * After retaking hugetlb_lock, we need to recalculate 'needed'
 995         * because either resv_huge_pages or free_huge_pages may have changed.
 996         */
 997        spin_lock(&hugetlb_lock);
 998        needed = (h->resv_huge_pages + delta) -
 999                        (h->free_huge_pages + allocated);
1000        if (needed > 0) {
1001                if (alloc_ok)
1002                        goto retry;
1003                /*
1004                 * We were not able to allocate enough pages to
1005                 * satisfy the entire reservation so we free what
1006                 * we've allocated so far.
1007                 */
1008                goto free;
1009        }
1010        /*
1011         * The surplus_list now contains _at_least_ the number of extra pages
1012         * needed to accommodate the reservation.  Add the appropriate number
1013         * of pages to the hugetlb pool and free the extras back to the buddy
1014         * allocator.  Commit the entire reservation here to prevent another
1015         * process from stealing the pages as they are added to the pool but
1016         * before they are reserved.
1017         */
1018        needed += allocated;
1019        h->resv_huge_pages += delta;
1020        ret = 0;
1021
1022        /* Free the needed pages to the hugetlb pool */
1023        list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1024                if ((--needed) < 0)
1025                        break;
1026                /*
1027                 * This page is now managed by the hugetlb allocator and has
1028                 * no users -- drop the buddy allocator's reference.
1029                 */
1030                put_page_testzero(page);
1031                VM_BUG_ON(page_count(page));
1032                enqueue_huge_page(h, page);
1033        }
1034free:
1035        spin_unlock(&hugetlb_lock);
1036
1037        /* Free unnecessary surplus pages to the buddy allocator */
1038        if (!list_empty(&surplus_list)) {
1039                list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1040                        put_page(page);
1041                }
1042        }
1043        spin_lock(&hugetlb_lock);
1044
1045        return ret;
1046}
1047
1048/*
1049 * When releasing a hugetlb pool reservation, any surplus pages that were
1050 * allocated to satisfy the reservation must be explicitly freed if they were
1051 * never used.
1052 * Called with hugetlb_lock held.
1053 */
1054static void return_unused_surplus_pages(struct hstate *h,
1055                                        unsigned long unused_resv_pages)
1056{
1057        unsigned long nr_pages;
1058
1059        /* Uncommit the reservation */
1060        h->resv_huge_pages -= unused_resv_pages;
1061
1062        /* Cannot return gigantic pages currently */
1063        if (h->order >= MAX_ORDER)
1064                return;
1065
1066        nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1067
1068        /*
1069         * We want to release as many surplus pages as possible, spread
1070         * evenly across all nodes with memory. Iterate across these nodes
1071         * until we can no longer free unreserved surplus pages. This occurs
1072         * when the nodes with surplus pages have no free pages.
1073         * free_pool_huge_page() will balance the the freed pages across the
1074         * on-line nodes with memory and will handle the hstate accounting.
1075         */
1076        while (nr_pages--) {
1077                if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1078                        break;
1079        }
1080}
1081
1082/*
1083 * Determine if the huge page at addr within the vma has an associated
1084 * reservation.  Where it does not we will need to logically increase
1085 * reservation and actually increase subpool usage before an allocation
1086 * can occur.  Where any new reservation would be required the
1087 * reservation change is prepared, but not committed.  Once the page
1088 * has been allocated from the subpool and instantiated the change should
1089 * be committed via vma_commit_reservation.  No action is required on
1090 * failure.
1091 */
1092static long vma_needs_reservation(struct hstate *h,
1093                        struct vm_area_struct *vma, unsigned long addr)
1094{
1095        struct address_space *mapping = vma->vm_file->f_mapping;
1096        struct inode *inode = mapping->host;
1097
1098        if (vma->vm_flags & VM_MAYSHARE) {
1099                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1100                return region_chg(&inode->i_mapping->private_list,
1101                                                        idx, idx + 1);
1102
1103        } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1104                return 1;
1105
1106        } else  {
1107                long err;
1108                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1109                struct resv_map *reservations = vma_resv_map(vma);
1110
1111                err = region_chg(&reservations->regions, idx, idx + 1);
1112                if (err < 0)
1113                        return err;
1114                return 0;
1115        }
1116}
1117static void vma_commit_reservation(struct hstate *h,
1118                        struct vm_area_struct *vma, unsigned long addr)
1119{
1120        struct address_space *mapping = vma->vm_file->f_mapping;
1121        struct inode *inode = mapping->host;
1122
1123        if (vma->vm_flags & VM_MAYSHARE) {
1124                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1125                region_add(&inode->i_mapping->private_list, idx, idx + 1);
1126
1127        } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1128                pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1129                struct resv_map *reservations = vma_resv_map(vma);
1130
1131                /* Mark this page used in the map. */
1132                region_add(&reservations->regions, idx, idx + 1);
1133        }
1134}
1135
1136static struct page *alloc_huge_page(struct vm_area_struct *vma,
1137                                    unsigned long addr, int avoid_reserve)
1138{
1139        struct hugepage_subpool *spool = subpool_vma(vma);
1140        struct hstate *h = hstate_vma(vma);
1141        struct page *page;
1142        long chg;
1143        int ret, idx;
1144        struct hugetlb_cgroup *h_cg;
1145
1146        idx = hstate_index(h);
1147        /*
1148         * Processes that did not create the mapping will have no
1149         * reserves and will not have accounted against subpool
1150         * limit. Check that the subpool limit can be made before
1151         * satisfying the allocation MAP_NORESERVE mappings may also
1152         * need pages and subpool limit allocated allocated if no reserve
1153         * mapping overlaps.
1154         */
1155        chg = vma_needs_reservation(h, vma, addr);
1156        if (chg < 0)
1157                return ERR_PTR(-ENOMEM);
1158        if (chg)
1159                if (hugepage_subpool_get_pages(spool, chg))
1160                        return ERR_PTR(-ENOSPC);
1161
1162        ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1163        if (ret) {
1164                hugepage_subpool_put_pages(spool, chg);
1165                return ERR_PTR(-ENOSPC);
1166        }
1167        spin_lock(&hugetlb_lock);
1168        page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1169        if (page) {
1170                /* update page cgroup details */
1171                hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1172                                             h_cg, page);
1173                spin_unlock(&hugetlb_lock);
1174        } else {
1175                spin_unlock(&hugetlb_lock);
1176                page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1177                if (!page) {
1178                        hugetlb_cgroup_uncharge_cgroup(idx,
1179                                                       pages_per_huge_page(h),
1180                                                       h_cg);
1181                        hugepage_subpool_put_pages(spool, chg);
1182                        return ERR_PTR(-ENOSPC);
1183                }
1184                spin_lock(&hugetlb_lock);
1185                hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1186                                             h_cg, page);
1187                list_move(&page->lru, &h->hugepage_activelist);
1188                spin_unlock(&hugetlb_lock);
1189        }
1190
1191        set_page_private(page, (unsigned long)spool);
1192
1193        vma_commit_reservation(h, vma, addr);
1194        return page;
1195}
1196
1197int __weak alloc_bootmem_huge_page(struct hstate *h)
1198{
1199        struct huge_bootmem_page *m;
1200        int nr_nodes = nodes_weight(node_states[N_MEMORY]);
1201
1202        while (nr_nodes) {
1203                void *addr;
1204
1205                addr = __alloc_bootmem_node_nopanic(
1206                                NODE_DATA(hstate_next_node_to_alloc(h,
1207                                                &node_states[N_MEMORY])),
1208                                huge_page_size(h), huge_page_size(h), 0);
1209
1210                if (addr) {
1211                        /*
1212                         * Use the beginning of the huge page to store the
1213                         * huge_bootmem_page struct (until gather_bootmem
1214                         * puts them into the mem_map).
1215                         */
1216                        m = addr;
1217                        goto found;
1218                }
1219                nr_nodes--;
1220        }
1221        return 0;
1222
1223found:
1224        BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1225        /* Put them into a private list first because mem_map is not up yet */
1226        list_add(&m->list, &huge_boot_pages);
1227        m->hstate = h;
1228        return 1;
1229}
1230
1231static void prep_compound_huge_page(struct page *page, int order)
1232{
1233        if (unlikely(order > (MAX_ORDER - 1)))
1234                prep_compound_gigantic_page(page, order);
1235        else
1236                prep_compound_page(page, order);
1237}
1238
1239/* Put bootmem huge pages into the standard lists after mem_map is up */
1240static void __init gather_bootmem_prealloc(void)
1241{
1242        struct huge_bootmem_page *m;
1243
1244        list_for_each_entry(m, &huge_boot_pages, list) {
1245                struct hstate *h = m->hstate;
1246                struct page *page;
1247
1248#ifdef CONFIG_HIGHMEM
1249                page = pfn_to_page(m->phys >> PAGE_SHIFT);
1250                free_bootmem_late((unsigned long)m,
1251                                  sizeof(struct huge_bootmem_page));
1252#else
1253                page = virt_to_page(m);
1254#endif
1255                __ClearPageReserved(page);
1256                WARN_ON(page_count(page) != 1);
1257                prep_compound_huge_page(page, h->order);
1258                prep_new_huge_page(h, page, page_to_nid(page));
1259                /*
1260                 * If we had gigantic hugepages allocated at boot time, we need
1261                 * to restore the 'stolen' pages to totalram_pages in order to
1262                 * fix confusing memory reports from free(1) and another
1263                 * side-effects, like CommitLimit going negative.
1264                 */
1265                if (h->order > (MAX_ORDER - 1))
1266                        totalram_pages += 1 << h->order;
1267        }
1268}
1269
1270static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1271{
1272        unsigned long i;
1273
1274        for (i = 0; i < h->max_huge_pages; ++i) {
1275                if (h->order >= MAX_ORDER) {
1276                        if (!alloc_bootmem_huge_page(h))
1277                                break;
1278                } else if (!alloc_fresh_huge_page(h,
1279                                         &node_states[N_MEMORY]))
1280                        break;
1281        }
1282        h->max_huge_pages = i;
1283}
1284
1285static void __init hugetlb_init_hstates(void)
1286{
1287        struct hstate *h;
1288
1289        for_each_hstate(h) {
1290                /* oversize hugepages were init'ed in early boot */
1291                if (h->order < MAX_ORDER)
1292                        hugetlb_hstate_alloc_pages(h);
1293        }
1294}
1295
1296static char * __init memfmt(char *buf, unsigned long n)
1297{
1298        if (n >= (1UL << 30))
1299                sprintf(buf, "%lu GB", n >> 30);
1300        else if (n >= (1UL << 20))
1301                sprintf(buf, "%lu MB", n >> 20);
1302        else
1303                sprintf(buf, "%lu KB", n >> 10);
1304        return buf;
1305}
1306
1307static void __init report_hugepages(void)
1308{
1309        struct hstate *h;
1310
1311        for_each_hstate(h) {
1312                char buf[32];
1313                pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1314                        memfmt(buf, huge_page_size(h)),
1315                        h->free_huge_pages);
1316        }
1317}
1318
1319#ifdef CONFIG_HIGHMEM
1320static void try_to_free_low(struct hstate *h, unsigned long count,
1321                                                nodemask_t *nodes_allowed)
1322{
1323        int i;
1324
1325        if (h->order >= MAX_ORDER)
1326                return;
1327
1328        for_each_node_mask(i, *nodes_allowed) {
1329                struct page *page, *next;
1330                struct list_head *freel = &h->hugepage_freelists[i];
1331                list_for_each_entry_safe(page, next, freel, lru) {
1332                        if (count >= h->nr_huge_pages)
1333                                return;
1334                        if (PageHighMem(page))
1335                                continue;
1336                        list_del(&page->lru);
1337                        update_and_free_page(h, page);
1338                        h->free_huge_pages--;
1339                        h->free_huge_pages_node[page_to_nid(page)]--;
1340                }
1341        }
1342}
1343#else
1344static inline void try_to_free_low(struct hstate *h, unsigned long count,
1345                                                nodemask_t *nodes_allowed)
1346{
1347}
1348#endif
1349
1350/*
1351 * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1352 * balanced by operating on them in a round-robin fashion.
1353 * Returns 1 if an adjustment was made.
1354 */
1355static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1356                                int delta)
1357{
1358        int start_nid, next_nid;
1359        int ret = 0;
1360
1361        VM_BUG_ON(delta != -1 && delta != 1);
1362
1363        if (delta < 0)
1364                start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1365        else
1366                start_nid = hstate_next_node_to_free(h, nodes_allowed);
1367        next_nid = start_nid;
1368
1369        do {
1370                int nid = next_nid;
1371                if (delta < 0)  {
1372                        /*
1373                         * To shrink on this node, there must be a surplus page
1374                         */
1375                        if (!h->surplus_huge_pages_node[nid]) {
1376                                next_nid = hstate_next_node_to_alloc(h,
1377                                                                nodes_allowed);
1378                                continue;
1379                        }
1380                }
1381                if (delta > 0) {
1382                        /*
1383                         * Surplus cannot exceed the total number of pages
1384                         */
1385                        if (h->surplus_huge_pages_node[nid] >=
1386                                                h->nr_huge_pages_node[nid]) {
1387                                next_nid = hstate_next_node_to_free(h,
1388                                                                nodes_allowed);
1389                                continue;
1390                        }
1391                }
1392
1393                h->surplus_huge_pages += delta;
1394                h->surplus_huge_pages_node[nid] += delta;
1395                ret = 1;
1396                break;
1397        } while (next_nid != start_nid);
1398
1399        return ret;
1400}
1401
1402#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1403static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1404                                                nodemask_t *nodes_allowed)
1405{
1406        unsigned long min_count, ret;
1407
1408        if (h->order >= MAX_ORDER)
1409                return h->max_huge_pages;
1410
1411        /*
1412         * Increase the pool size
1413         * First take pages out of surplus state.  Then make up the
1414         * remaining difference by allocating fresh huge pages.
1415         *
1416         * We might race with alloc_buddy_huge_page() here and be unable
1417         * to convert a surplus huge page to a normal huge page. That is
1418         * not critical, though, it just means the overall size of the
1419         * pool might be one hugepage larger than it needs to be, but
1420         * within all the constraints specified by the sysctls.
1421         */
1422        spin_lock(&hugetlb_lock);
1423        while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1424                if (!adjust_pool_surplus(h, nodes_allowed, -1))
1425                        break;
1426        }
1427
1428        while (count > persistent_huge_pages(h)) {
1429                /*
1430                 * If this allocation races such that we no longer need the
1431                 * page, free_huge_page will handle it by freeing the page
1432                 * and reducing the surplus.
1433                 */
1434                spin_unlock(&hugetlb_lock);
1435                ret = alloc_fresh_huge_page(h, nodes_allowed);
1436                spin_lock(&hugetlb_lock);
1437                if (!ret)
1438                        goto out;
1439
1440                /* Bail for signals. Probably ctrl-c from user */
1441                if (signal_pending(current))
1442                        goto out;
1443        }
1444
1445        /*
1446         * Decrease the pool size
1447         * First return free pages to the buddy allocator (being careful
1448         * to keep enough around to satisfy reservations).  Then place
1449         * pages into surplus state as needed so the pool will shrink
1450         * to the desired size as pages become free.
1451         *
1452         * By placing pages into the surplus state independent of the
1453         * overcommit value, we are allowing the surplus pool size to
1454         * exceed overcommit. There are few sane options here. Since
1455         * alloc_buddy_huge_page() is checking the global counter,
1456         * though, we'll note that we're not allowed to exceed surplus
1457         * and won't grow the pool anywhere else. Not until one of the
1458         * sysctls are changed, or the surplus pages go out of use.
1459         */
1460        min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1461        min_count = max(count, min_count);
1462        try_to_free_low(h, min_count, nodes_allowed);
1463        while (min_count < persistent_huge_pages(h)) {
1464                if (!free_pool_huge_page(h, nodes_allowed, 0))
1465                        break;
1466        }
1467        while (count < persistent_huge_pages(h)) {
1468                if (!adjust_pool_surplus(h, nodes_allowed, 1))
1469                        break;
1470        }
1471out:
1472        ret = persistent_huge_pages(h);
1473        spin_unlock(&hugetlb_lock);
1474        return ret;
1475}
1476
1477#define HSTATE_ATTR_RO(_name) \
1478        static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1479
1480#define HSTATE_ATTR(_name) \
1481        static struct kobj_attribute _name##_attr = \
1482                __ATTR(_name, 0644, _name##_show, _name##_store)
1483
1484static struct kobject *hugepages_kobj;
1485static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1486
1487static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1488
1489static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1490{
1491        int i;
1492
1493        for (i = 0; i < HUGE_MAX_HSTATE; i++)
1494                if (hstate_kobjs[i] == kobj) {
1495                        if (nidp)
1496                                *nidp = NUMA_NO_NODE;
1497                        return &hstates[i];
1498                }
1499
1500        return kobj_to_node_hstate(kobj, nidp);
1501}
1502
1503static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1504                                        struct kobj_attribute *attr, char *buf)
1505{
1506        struct hstate *h;
1507        unsigned long nr_huge_pages;
1508        int nid;
1509
1510        h = kobj_to_hstate(kobj, &nid);
1511        if (nid == NUMA_NO_NODE)
1512                nr_huge_pages = h->nr_huge_pages;
1513        else
1514                nr_huge_pages = h->nr_huge_pages_node[nid];
1515
1516        return sprintf(buf, "%lu\n", nr_huge_pages);
1517}
1518
1519static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1520                        struct kobject *kobj, struct kobj_attribute *attr,
1521                        const char *buf, size_t len)
1522{
1523        int err;
1524        int nid;
1525        unsigned long count;
1526        struct hstate *h;
1527        NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1528
1529        err = strict_strtoul(buf, 10, &count);
1530        if (err)
1531                goto out;
1532
1533        h = kobj_to_hstate(kobj, &nid);
1534        if (h->order >= MAX_ORDER) {
1535                err = -EINVAL;
1536                goto out;
1537        }
1538
1539        if (nid == NUMA_NO_NODE) {
1540                /*
1541                 * global hstate attribute
1542                 */
1543                if (!(obey_mempolicy &&
1544                                init_nodemask_of_mempolicy(nodes_allowed))) {
1545                        NODEMASK_FREE(nodes_allowed);
1546                        nodes_allowed = &node_states[N_MEMORY];
1547                }
1548        } else if (nodes_allowed) {
1549                /*
1550                 * per node hstate attribute: adjust count to global,
1551                 * but restrict alloc/free to the specified node.
1552                 */
1553                count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1554                init_nodemask_of_node(nodes_allowed, nid);
1555        } else
1556                nodes_allowed = &node_states[N_MEMORY];
1557
1558        h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1559
1560        if (nodes_allowed != &node_states[N_MEMORY])
1561                NODEMASK_FREE(nodes_allowed);
1562
1563        return len;
1564out:
1565        NODEMASK_FREE(nodes_allowed);
1566        return err;
1567}
1568
1569static ssize_t nr_hugepages_show(struct kobject *kobj,
1570                                       struct kobj_attribute *attr, char *buf)
1571{
1572        return nr_hugepages_show_common(kobj, attr, buf);
1573}
1574
1575static ssize_t nr_hugepages_store(struct kobject *kobj,
1576               struct kobj_attribute *attr, const char *buf, size_t len)
1577{
1578        return nr_hugepages_store_common(false, kobj, attr, buf, len);
1579}
1580HSTATE_ATTR(nr_hugepages);
1581
1582#ifdef CONFIG_NUMA
1583
1584/*
1585 * hstate attribute for optionally mempolicy-based constraint on persistent
1586 * huge page alloc/free.
1587 */
1588static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1589                                       struct kobj_attribute *attr, char *buf)
1590{
1591        return nr_hugepages_show_common(kobj, attr, buf);
1592}
1593
1594static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1595               struct kobj_attribute *attr, const char *buf, size_t len)
1596{
1597        return nr_hugepages_store_common(true, kobj, attr, buf, len);
1598}
1599HSTATE_ATTR(nr_hugepages_mempolicy);
1600#endif
1601
1602
1603static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1604                                        struct kobj_attribute *attr, char *buf)
1605{
1606        struct hstate *h = kobj_to_hstate(kobj, NULL);
1607        return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1608}
1609
1610static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1611                struct kobj_attribute *attr, const char *buf, size_t count)
1612{
1613        int err;
1614        unsigned long input;
1615        struct hstate *h = kobj_to_hstate(kobj, NULL);
1616
1617        if (h->order >= MAX_ORDER)
1618                return -EINVAL;
1619
1620        err = strict_strtoul(buf, 10, &input);
1621        if (err)
1622                return err;
1623
1624        spin_lock(&hugetlb_lock);
1625        h->nr_overcommit_huge_pages = input;
1626        spin_unlock(&hugetlb_lock);
1627
1628        return count;
1629}
1630HSTATE_ATTR(nr_overcommit_hugepages);
1631
1632static ssize_t free_hugepages_show(struct kobject *kobj,
1633                                        struct kobj_attribute *attr, char *buf)
1634{
1635        struct hstate *h;
1636        unsigned long free_huge_pages;
1637        int nid;
1638
1639        h = kobj_to_hstate(kobj, &nid);
1640        if (nid == NUMA_NO_NODE)
1641                free_huge_pages = h->free_huge_pages;
1642        else
1643                free_huge_pages = h->free_huge_pages_node[nid];
1644
1645        return sprintf(buf, "%lu\n", free_huge_pages);
1646}
1647HSTATE_ATTR_RO(free_hugepages);
1648
1649static ssize_t resv_hugepages_show(struct kobject *kobj,
1650                                        struct kobj_attribute *attr, char *buf)
1651{
1652        struct hstate *h = kobj_to_hstate(kobj, NULL);
1653        return sprintf(buf, "%lu\n", h->resv_huge_pages);
1654}
1655HSTATE_ATTR_RO(resv_hugepages);
1656
1657static ssize_t surplus_hugepages_show(struct kobject *kobj,
1658                                        struct kobj_attribute *attr, char *buf)
1659{
1660        struct hstate *h;
1661        unsigned long surplus_huge_pages;
1662        int nid;
1663
1664        h = kobj_to_hstate(kobj, &nid);
1665        if (nid == NUMA_NO_NODE)
1666                surplus_huge_pages = h->surplus_huge_pages;
1667        else
1668                surplus_huge_pages = h->surplus_huge_pages_node[nid];
1669
1670        return sprintf(buf, "%lu\n", surplus_huge_pages);
1671}
1672HSTATE_ATTR_RO(surplus_hugepages);
1673
1674static struct attribute *hstate_attrs[] = {
1675        &nr_hugepages_attr.attr,
1676        &nr_overcommit_hugepages_attr.attr,
1677        &free_hugepages_attr.attr,
1678        &resv_hugepages_attr.attr,
1679        &surplus_hugepages_attr.attr,
1680#ifdef CONFIG_NUMA
1681        &nr_hugepages_mempolicy_attr.attr,
1682#endif
1683        NULL,
1684};
1685
1686static struct attribute_group hstate_attr_group = {
1687        .attrs = hstate_attrs,
1688};
1689
1690static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1691                                    struct kobject **hstate_kobjs,
1692                                    struct attribute_group *hstate_attr_group)
1693{
1694        int retval;
1695        int hi = hstate_index(h);
1696
1697        hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1698        if (!hstate_kobjs[hi])
1699                return -ENOMEM;
1700
1701        retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1702        if (retval)
1703                kobject_put(hstate_kobjs[hi]);
1704
1705        return retval;
1706}
1707
1708static void __init hugetlb_sysfs_init(void)
1709{
1710        struct hstate *h;
1711        int err;
1712
1713        hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1714        if (!hugepages_kobj)
1715                return;
1716
1717        for_each_hstate(h) {
1718                err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1719                                         hstate_kobjs, &hstate_attr_group);
1720                if (err)
1721                        pr_err("Hugetlb: Unable to add hstate %s", h->name);
1722        }
1723}
1724
1725#ifdef CONFIG_NUMA
1726
1727/*
1728 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1729 * with node devices in node_devices[] using a parallel array.  The array
1730 * index of a node device or _hstate == node id.
1731 * This is here to avoid any static dependency of the node device driver, in
1732 * the base kernel, on the hugetlb module.
1733 */
1734struct node_hstate {
1735        struct kobject          *hugepages_kobj;
1736        struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1737};
1738struct node_hstate node_hstates[MAX_NUMNODES];
1739
1740/*
1741 * A subset of global hstate attributes for node devices
1742 */
1743static struct attribute *per_node_hstate_attrs[] = {
1744        &nr_hugepages_attr.attr,
1745        &free_hugepages_attr.attr,
1746        &surplus_hugepages_attr.attr,
1747        NULL,
1748};
1749
1750static struct attribute_group per_node_hstate_attr_group = {
1751        .attrs = per_node_hstate_attrs,
1752};
1753
1754/*
1755 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1756 * Returns node id via non-NULL nidp.
1757 */
1758static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1759{
1760        int nid;
1761
1762        for (nid = 0; nid < nr_node_ids; nid++) {
1763                struct node_hstate *nhs = &node_hstates[nid];
1764                int i;
1765                for (i = 0; i < HUGE_MAX_HSTATE; i++)
1766                        if (nhs->hstate_kobjs[i] == kobj) {
1767                                if (nidp)
1768                                        *nidp = nid;
1769                                return &hstates[i];
1770                        }
1771        }
1772
1773        BUG();
1774        return NULL;
1775}
1776
1777/*
1778 * Unregister hstate attributes from a single node device.
1779 * No-op if no hstate attributes attached.
1780 */
1781static void hugetlb_unregister_node(struct node *node)
1782{
1783        struct hstate *h;
1784        struct node_hstate *nhs = &node_hstates[node->dev.id];
1785
1786        if (!nhs->hugepages_kobj)
1787                return;         /* no hstate attributes */
1788
1789        for_each_hstate(h) {
1790                int idx = hstate_index(h);
1791                if (nhs->hstate_kobjs[idx]) {
1792                        kobject_put(nhs->hstate_kobjs[idx]);
1793                        nhs->hstate_kobjs[idx] = NULL;
1794                }
1795        }
1796
1797        kobject_put(nhs->hugepages_kobj);
1798        nhs->hugepages_kobj = NULL;
1799}
1800
1801/*
1802 * hugetlb module exit:  unregister hstate attributes from node devices
1803 * that have them.
1804 */
1805static void hugetlb_unregister_all_nodes(void)
1806{
1807        int nid;
1808
1809        /*
1810         * disable node device registrations.
1811         */
1812        register_hugetlbfs_with_node(NULL, NULL);
1813
1814        /*
1815         * remove hstate attributes from any nodes that have them.
1816         */
1817        for (nid = 0; nid < nr_node_ids; nid++)
1818                hugetlb_unregister_node(node_devices[nid]);
1819}
1820
1821/*
1822 * Register hstate attributes for a single node device.
1823 * No-op if attributes already registered.
1824 */
1825static void hugetlb_register_node(struct node *node)
1826{
1827        struct hstate *h;
1828        struct node_hstate *nhs = &node_hstates[node->dev.id];
1829        int err;
1830
1831        if (nhs->hugepages_kobj)
1832                return;         /* already allocated */
1833
1834        nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1835                                                        &node->dev.kobj);
1836        if (!nhs->hugepages_kobj)
1837                return;
1838
1839        for_each_hstate(h) {
1840                err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1841                                                nhs->hstate_kobjs,
1842                                                &per_node_hstate_attr_group);
1843                if (err) {
1844                        pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1845                                h->name, node->dev.id);
1846                        hugetlb_unregister_node(node);
1847                        break;
1848                }
1849        }
1850}
1851
1852/*
1853 * hugetlb init time:  register hstate attributes for all registered node
1854 * devices of nodes that have memory.  All on-line nodes should have
1855 * registered their associated device by this time.
1856 */
1857static void hugetlb_register_all_nodes(void)
1858{
1859        int nid;
1860
1861        for_each_node_state(nid, N_MEMORY) {
1862                struct node *node = node_devices[nid];
1863                if (node->dev.id == nid)
1864                        hugetlb_register_node(node);
1865        }
1866
1867        /*
1868         * Let the node device driver know we're here so it can
1869         * [un]register hstate attributes on node hotplug.
1870         */
1871        register_hugetlbfs_with_node(hugetlb_register_node,
1872                                     hugetlb_unregister_node);
1873}
1874#else   /* !CONFIG_NUMA */
1875
1876static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1877{
1878        BUG();
1879        if (nidp)
1880                *nidp = -1;
1881        return NULL;
1882}
1883
1884static void hugetlb_unregister_all_nodes(void) { }
1885
1886static void hugetlb_register_all_nodes(void) { }
1887
1888#endif
1889
1890static void __exit hugetlb_exit(void)
1891{
1892        struct hstate *h;
1893
1894        hugetlb_unregister_all_nodes();
1895
1896        for_each_hstate(h) {
1897                kobject_put(hstate_kobjs[hstate_index(h)]);
1898        }
1899
1900        kobject_put(hugepages_kobj);
1901}
1902module_exit(hugetlb_exit);
1903
1904static int __init hugetlb_init(void)
1905{
1906        /* Some platform decide whether they support huge pages at boot
1907         * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1908         * there is no such support
1909         */
1910        if (HPAGE_SHIFT == 0)
1911                return 0;
1912
1913        if (!size_to_hstate(default_hstate_size)) {
1914                default_hstate_size = HPAGE_SIZE;
1915                if (!size_to_hstate(default_hstate_size))
1916                        hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1917        }
1918        default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1919        if (default_hstate_max_huge_pages)
1920                default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1921
1922        hugetlb_init_hstates();
1923        gather_bootmem_prealloc();
1924        report_hugepages();
1925
1926        hugetlb_sysfs_init();
1927        hugetlb_register_all_nodes();
1928        hugetlb_cgroup_file_init();
1929
1930        return 0;
1931}
1932module_init(hugetlb_init);
1933
1934/* Should be called on processing a hugepagesz=... option */
1935void __init hugetlb_add_hstate(unsigned order)
1936{
1937        struct hstate *h;
1938        unsigned long i;
1939
1940        if (size_to_hstate(PAGE_SIZE << order)) {
1941                pr_warning("hugepagesz= specified twice, ignoring\n");
1942                return;
1943        }
1944        BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1945        BUG_ON(order == 0);
1946        h = &hstates[hugetlb_max_hstate++];
1947        h->order = order;
1948        h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1949        h->nr_huge_pages = 0;
1950        h->free_huge_pages = 0;
1951        for (i = 0; i < MAX_NUMNODES; ++i)
1952                INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1953        INIT_LIST_HEAD(&h->hugepage_activelist);
1954        h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1955        h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1956        snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1957                                        huge_page_size(h)/1024);
1958
1959        parsed_hstate = h;
1960}
1961
1962static int __init hugetlb_nrpages_setup(char *s)
1963{
1964        unsigned long *mhp;
1965        static unsigned long *last_mhp;
1966
1967        /*
1968         * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1969         * so this hugepages= parameter goes to the "default hstate".
1970         */
1971        if (!hugetlb_max_hstate)
1972                mhp = &default_hstate_max_huge_pages;
1973        else
1974                mhp = &parsed_hstate->max_huge_pages;
1975
1976        if (mhp == last_mhp) {
1977                pr_warning("hugepages= specified twice without "
1978                           "interleaving hugepagesz=, ignoring\n");
1979                return 1;
1980        }
1981
1982        if (sscanf(s, "%lu", mhp) <= 0)
1983                *mhp = 0;
1984
1985        /*
1986         * Global state is always initialized later in hugetlb_init.
1987         * But we need to allocate >= MAX_ORDER hstates here early to still
1988         * use the bootmem allocator.
1989         */
1990        if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1991                hugetlb_hstate_alloc_pages(parsed_hstate);
1992
1993        last_mhp = mhp;
1994
1995        return 1;
1996}
1997__setup("hugepages=", hugetlb_nrpages_setup);
1998
1999static int __init hugetlb_default_setup(char *s)
2000{
2001        default_hstate_size = memparse(s, &s);
2002        return 1;
2003}
2004__setup("default_hugepagesz=", hugetlb_default_setup);
2005
2006static unsigned int cpuset_mems_nr(unsigned int *array)
2007{
2008        int node;
2009        unsigned int nr = 0;
2010
2011        for_each_node_mask(node, cpuset_current_mems_allowed)
2012                nr += array[node];
2013
2014        return nr;
2015}
2016
2017#ifdef CONFIG_SYSCTL
2018static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2019                         struct ctl_table *table, int write,
2020                         void __user *buffer, size_t *length, loff_t *ppos)
2021{
2022        struct hstate *h = &default_hstate;
2023        unsigned long tmp;
2024        int ret;
2025
2026        tmp = h->max_huge_pages;
2027
2028        if (write && h->order >= MAX_ORDER)
2029                return -EINVAL;
2030
2031        table->data = &tmp;
2032        table->maxlen = sizeof(unsigned long);
2033        ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2034        if (ret)
2035                goto out;
2036
2037        if (write) {
2038                NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2039                                                GFP_KERNEL | __GFP_NORETRY);
2040                if (!(obey_mempolicy &&
2041                               init_nodemask_of_mempolicy(nodes_allowed))) {
2042                        NODEMASK_FREE(nodes_allowed);
2043                        nodes_allowed = &node_states[N_MEMORY];
2044                }
2045                h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2046
2047                if (nodes_allowed != &node_states[N_MEMORY])
2048                        NODEMASK_FREE(nodes_allowed);
2049        }
2050out:
2051        return ret;
2052}
2053
2054int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2055                          void __user *buffer, size_t *length, loff_t *ppos)
2056{
2057
2058        return hugetlb_sysctl_handler_common(false, table, write,
2059                                                        buffer, length, ppos);
2060}
2061
2062#ifdef CONFIG_NUMA
2063int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2064                          void __user *buffer, size_t *length, loff_t *ppos)
2065{
2066        return hugetlb_sysctl_handler_common(true, table, write,
2067                                                        buffer, length, ppos);
2068}
2069#endif /* CONFIG_NUMA */
2070
2071int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2072                        void __user *buffer,
2073                        size_t *length, loff_t *ppos)
2074{
2075        proc_dointvec(table, write, buffer, length, ppos);
2076        if (hugepages_treat_as_movable)
2077                htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2078        else
2079                htlb_alloc_mask = GFP_HIGHUSER;
2080        return 0;
2081}
2082
2083int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2084                        void __user *buffer,
2085                        size_t *length, loff_t *ppos)
2086{
2087        struct hstate *h = &default_hstate;
2088        unsigned long tmp;
2089        int ret;
2090
2091        tmp = h->nr_overcommit_huge_pages;
2092
2093        if (write && h->order >= MAX_ORDER)
2094                return -EINVAL;
2095
2096        table->data = &tmp;
2097        table->maxlen = sizeof(unsigned long);
2098        ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2099        if (ret)
2100                goto out;
2101
2102        if (write) {
2103                spin_lock(&hugetlb_lock);
2104                h->nr_overcommit_huge_pages = tmp;
2105                spin_unlock(&hugetlb_lock);
2106        }
2107out:
2108        return ret;
2109}
2110
2111#endif /* CONFIG_SYSCTL */
2112
2113void hugetlb_report_meminfo(struct seq_file *m)
2114{
2115        struct hstate *h = &default_hstate;
2116        seq_printf(m,
2117                        "HugePages_Total:   %5lu\n"
2118                        "HugePages_Free:    %5lu\n"
2119                        "HugePages_Rsvd:    %5lu\n"
2120                        "HugePages_Surp:    %5lu\n"
2121                        "Hugepagesize:   %8lu kB\n",
2122                        h->nr_huge_pages,
2123                        h->free_huge_pages,
2124                        h->resv_huge_pages,
2125                        h->surplus_huge_pages,
2126                        1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2127}
2128
2129int hugetlb_report_node_meminfo(int nid, char *buf)
2130{
2131        struct hstate *h = &default_hstate;
2132        return sprintf(buf,
2133                "Node %d HugePages_Total: %5u\n"
2134                "Node %d HugePages_Free:  %5u\n"
2135                "Node %d HugePages_Surp:  %5u\n",
2136                nid, h->nr_huge_pages_node[nid],
2137                nid, h->free_huge_pages_node[nid],
2138                nid, h->surplus_huge_pages_node[nid]);
2139}
2140
2141void hugetlb_show_meminfo(void)
2142{
2143        struct hstate *h;
2144        int nid;
2145
2146        for_each_node_state(nid, N_MEMORY)
2147                for_each_hstate(h)
2148                        pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2149                                nid,
2150                                h->nr_huge_pages_node[nid],
2151                                h->free_huge_pages_node[nid],
2152                                h->surplus_huge_pages_node[nid],
2153                                1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2154}
2155
2156/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2157unsigned long hugetlb_total_pages(void)
2158{
2159        struct hstate *h;
2160        unsigned long nr_total_pages = 0;
2161
2162        for_each_hstate(h)
2163                nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2164        return nr_total_pages;
2165}
2166
2167static int hugetlb_acct_memory(struct hstate *h, long delta)
2168{
2169        int ret = -ENOMEM;
2170
2171        spin_lock(&hugetlb_lock);
2172        /*
2173         * When cpuset is configured, it breaks the strict hugetlb page
2174         * reservation as the accounting is done on a global variable. Such
2175         * reservation is completely rubbish in the presence of cpuset because
2176         * the reservation is not checked against page availability for the
2177         * current cpuset. Application can still potentially OOM'ed by kernel
2178         * with lack of free htlb page in cpuset that the task is in.
2179         * Attempt to enforce strict accounting with cpuset is almost
2180         * impossible (or too ugly) because cpuset is too fluid that
2181         * task or memory node can be dynamically moved between cpusets.
2182         *
2183         * The change of semantics for shared hugetlb mapping with cpuset is
2184         * undesirable. However, in order to preserve some of the semantics,
2185         * we fall back to check against current free page availability as
2186         * a best attempt and hopefully to minimize the impact of changing
2187         * semantics that cpuset has.
2188         */
2189        if (delta > 0) {
2190                if (gather_surplus_pages(h, delta) < 0)
2191                        goto out;
2192
2193                if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2194                        return_unused_surplus_pages(h, delta);
2195                        goto out;
2196                }
2197        }
2198
2199        ret = 0;
2200        if (delta < 0)
2201                return_unused_surplus_pages(h, (unsigned long) -delta);
2202
2203out:
2204        spin_unlock(&hugetlb_lock);
2205        return ret;
2206}
2207
2208static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2209{
2210        struct resv_map *reservations = vma_resv_map(vma);
2211
2212        /*
2213         * This new VMA should share its siblings reservation map if present.
2214         * The VMA will only ever have a valid reservation map pointer where
2215         * it is being copied for another still existing VMA.  As that VMA
2216         * has a reference to the reservation map it cannot disappear until
2217         * after this open call completes.  It is therefore safe to take a
2218         * new reference here without additional locking.
2219         */
2220        if (reservations)
2221                kref_get(&reservations->refs);
2222}
2223
2224static void resv_map_put(struct vm_area_struct *vma)
2225{
2226        struct resv_map *reservations = vma_resv_map(vma);
2227
2228        if (!reservations)
2229                return;
2230        kref_put(&reservations->refs, resv_map_release);
2231}
2232
2233static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2234{
2235        struct hstate *h = hstate_vma(vma);
2236        struct resv_map *reservations = vma_resv_map(vma);
2237        struct hugepage_subpool *spool = subpool_vma(vma);
2238        unsigned long reserve;
2239        unsigned long start;
2240        unsigned long end;
2241
2242        if (reservations) {
2243                start = vma_hugecache_offset(h, vma, vma->vm_start);
2244                end = vma_hugecache_offset(h, vma, vma->vm_end);
2245
2246                reserve = (end - start) -
2247                        region_count(&reservations->regions, start, end);
2248
2249                resv_map_put(vma);
2250
2251                if (reserve) {
2252                        hugetlb_acct_memory(h, -reserve);
2253                        hugepage_subpool_put_pages(spool, reserve);
2254                }
2255        }
2256}
2257
2258/*
2259 * We cannot handle pagefaults against hugetlb pages at all.  They cause
2260 * handle_mm_fault() to try to instantiate regular-sized pages in the
2261 * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2262 * this far.
2263 */
2264static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2265{
2266        BUG();
2267        return 0;
2268}
2269
2270const struct vm_operations_struct hugetlb_vm_ops = {
2271        .fault = hugetlb_vm_op_fault,
2272        .open = hugetlb_vm_op_open,
2273        .close = hugetlb_vm_op_close,
2274};
2275
2276static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2277                                int writable)
2278{
2279        pte_t entry;
2280
2281        if (writable) {
2282                entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2283                                         vma->vm_page_prot)));
2284        } else {
2285                entry = huge_pte_wrprotect(mk_huge_pte(page,
2286                                           vma->vm_page_prot));
2287        }
2288        entry = pte_mkyoung(entry);
2289        entry = pte_mkhuge(entry);
2290        entry = arch_make_huge_pte(entry, vma, page, writable);
2291
2292        return entry;
2293}
2294
2295static void set_huge_ptep_writable(struct vm_area_struct *vma,
2296                                   unsigned long address, pte_t *ptep)
2297{
2298        pte_t entry;
2299
2300        entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2301        if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2302                update_mmu_cache(vma, address, ptep);
2303}
2304
2305
2306int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2307                            struct vm_area_struct *vma)
2308{
2309        pte_t *src_pte, *dst_pte, entry;
2310        struct page *ptepage;
2311        unsigned long addr;
2312        int cow;
2313        struct hstate *h = hstate_vma(vma);
2314        unsigned long sz = huge_page_size(h);
2315
2316        cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2317
2318        for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2319                src_pte = huge_pte_offset(src, addr);
2320                if (!src_pte)
2321                        continue;
2322                dst_pte = huge_pte_alloc(dst, addr, sz);
2323                if (!dst_pte)
2324                        goto nomem;
2325
2326                /* If the pagetables are shared don't copy or take references */
2327                if (dst_pte == src_pte)
2328                        continue;
2329
2330                spin_lock(&dst->page_table_lock);
2331                spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2332                if (!huge_pte_none(huge_ptep_get(src_pte))) {
2333                        if (cow)
2334                                huge_ptep_set_wrprotect(src, addr, src_pte);
2335                        entry = huge_ptep_get(src_pte);
2336                        ptepage = pte_page(entry);
2337                        get_page(ptepage);
2338                        page_dup_rmap(ptepage);
2339                        set_huge_pte_at(dst, addr, dst_pte, entry);
2340                }
2341                spin_unlock(&src->page_table_lock);
2342                spin_unlock(&dst->page_table_lock);
2343        }
2344        return 0;
2345
2346nomem:
2347        return -ENOMEM;
2348}
2349
2350static int is_hugetlb_entry_migration(pte_t pte)
2351{
2352        swp_entry_t swp;
2353
2354        if (huge_pte_none(pte) || pte_present(pte))
2355                return 0;
2356        swp = pte_to_swp_entry(pte);
2357        if (non_swap_entry(swp) && is_migration_entry(swp))
2358                return 1;
2359        else
2360                return 0;
2361}
2362
2363static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2364{
2365        swp_entry_t swp;
2366
2367        if (huge_pte_none(pte) || pte_present(pte))
2368                return 0;
2369        swp = pte_to_swp_entry(pte);
2370        if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2371                return 1;
2372        else
2373                return 0;
2374}
2375
2376void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2377                            unsigned long start, unsigned long end,
2378                            struct page *ref_page)
2379{
2380        int force_flush = 0;
2381        struct mm_struct *mm = vma->vm_mm;
2382        unsigned long address;
2383        pte_t *ptep;
2384        pte_t pte;
2385        struct page *page;
2386        struct hstate *h = hstate_vma(vma);
2387        unsigned long sz = huge_page_size(h);
2388        const unsigned long mmun_start = start; /* For mmu_notifiers */
2389        const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2390
2391        WARN_ON(!is_vm_hugetlb_page(vma));
2392        BUG_ON(start & ~huge_page_mask(h));
2393        BUG_ON(end & ~huge_page_mask(h));
2394
2395        tlb_start_vma(tlb, vma);
2396        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2397again:
2398        spin_lock(&mm->page_table_lock);
2399        for (address = start; address < end; address += sz) {
2400                ptep = huge_pte_offset(mm, address);
2401                if (!ptep)
2402                        continue;
2403
2404                if (huge_pmd_unshare(mm, &address, ptep))
2405                        continue;
2406
2407                pte = huge_ptep_get(ptep);
2408                if (huge_pte_none(pte))
2409                        continue;
2410
2411                /*
2412                 * HWPoisoned hugepage is already unmapped and dropped reference
2413                 */
2414                if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2415                        huge_pte_clear(mm, address, ptep);
2416                        continue;
2417                }
2418
2419                page = pte_page(pte);
2420                /*
2421                 * If a reference page is supplied, it is because a specific
2422                 * page is being unmapped, not a range. Ensure the page we
2423                 * are about to unmap is the actual page of interest.
2424                 */
2425                if (ref_page) {
2426                        if (page != ref_page)
2427                                continue;
2428
2429                        /*
2430                         * Mark the VMA as having unmapped its page so that
2431                         * future faults in this VMA will fail rather than
2432                         * looking like data was lost
2433                         */
2434                        set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2435                }
2436
2437                pte = huge_ptep_get_and_clear(mm, address, ptep);
2438                tlb_remove_tlb_entry(tlb, ptep, address);
2439                if (huge_pte_dirty(pte))
2440                        set_page_dirty(page);
2441
2442                page_remove_rmap(page);
2443                force_flush = !__tlb_remove_page(tlb, page);
2444                if (force_flush)
2445                        break;
2446                /* Bail out after unmapping reference page if supplied */
2447                if (ref_page)
2448                        break;
2449        }
2450        spin_unlock(&mm->page_table_lock);
2451        /*
2452         * mmu_gather ran out of room to batch pages, we break out of
2453         * the PTE lock to avoid doing the potential expensive TLB invalidate
2454         * and page-free while holding it.
2455         */
2456        if (force_flush) {
2457                force_flush = 0;
2458                tlb_flush_mmu(tlb);
2459                if (address < end && !ref_page)
2460                        goto again;
2461        }
2462        mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2463        tlb_end_vma(tlb, vma);
2464}
2465
2466void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2467                          struct vm_area_struct *vma, unsigned long start,
2468                          unsigned long end, struct page *ref_page)
2469{
2470        __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2471
2472        /*
2473         * Clear this flag so that x86's huge_pmd_share page_table_shareable
2474         * test will fail on a vma being torn down, and not grab a page table
2475         * on its way out.  We're lucky that the flag has such an appropriate
2476         * name, and can in fact be safely cleared here. We could clear it
2477         * before the __unmap_hugepage_range above, but all that's necessary
2478         * is to clear it before releasing the i_mmap_mutex. This works
2479         * because in the context this is called, the VMA is about to be
2480         * destroyed and the i_mmap_mutex is held.
2481         */
2482        vma->vm_flags &= ~VM_MAYSHARE;
2483}
2484
2485void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2486                          unsigned long end, struct page *ref_page)
2487{
2488        struct mm_struct *mm;
2489        struct mmu_gather tlb;
2490
2491        mm = vma->vm_mm;
2492
2493        tlb_gather_mmu(&tlb, mm, 0);
2494        __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2495        tlb_finish_mmu(&tlb, start, end);
2496}
2497
2498/*
2499 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2500 * mappping it owns the reserve page for. The intention is to unmap the page
2501 * from other VMAs and let the children be SIGKILLed if they are faulting the
2502 * same region.
2503 */
2504static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2505                                struct page *page, unsigned long address)
2506{
2507        struct hstate *h = hstate_vma(vma);
2508        struct vm_area_struct *iter_vma;
2509        struct address_space *mapping;
2510        pgoff_t pgoff;
2511
2512        /*
2513         * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2514         * from page cache lookup which is in HPAGE_SIZE units.
2515         */
2516        address = address & huge_page_mask(h);
2517        pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2518                        vma->vm_pgoff;
2519        mapping = file_inode(vma->vm_file)->i_mapping;
2520
2521        /*
2522         * Take the mapping lock for the duration of the table walk. As
2523         * this mapping should be shared between all the VMAs,
2524         * __unmap_hugepage_range() is called as the lock is already held
2525         */
2526        mutex_lock(&mapping->i_mmap_mutex);
2527        vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2528                /* Do not unmap the current VMA */
2529                if (iter_vma == vma)
2530                        continue;
2531
2532                /*
2533                 * Unmap the page from other VMAs without their own reserves.
2534                 * They get marked to be SIGKILLed if they fault in these
2535                 * areas. This is because a future no-page fault on this VMA
2536                 * could insert a zeroed page instead of the data existing
2537                 * from the time of fork. This would look like data corruption
2538                 */
2539                if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2540                        unmap_hugepage_range(iter_vma, address,
2541                                             address + huge_page_size(h), page);
2542        }
2543        mutex_unlock(&mapping->i_mmap_mutex);
2544
2545        return 1;
2546}
2547
2548/*
2549 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2550 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2551 * cannot race with other handlers or page migration.
2552 * Keep the pte_same checks anyway to make transition from the mutex easier.
2553 */
2554static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2555                        unsigned long address, pte_t *ptep, pte_t pte,
2556                        struct page *pagecache_page)
2557{
2558        struct hstate *h = hstate_vma(vma);
2559        struct page *old_page, *new_page;
2560        int avoidcopy;
2561        int outside_reserve = 0;
2562        unsigned long mmun_start;       /* For mmu_notifiers */
2563        unsigned long mmun_end;         /* For mmu_notifiers */
2564
2565        old_page = pte_page(pte);
2566
2567retry_avoidcopy:
2568        /* If no-one else is actually using this page, avoid the copy
2569         * and just make the page writable */
2570        avoidcopy = (page_mapcount(old_page) == 1);
2571        if (avoidcopy) {
2572                if (PageAnon(old_page))
2573                        page_move_anon_rmap(old_page, vma, address);
2574                set_huge_ptep_writable(vma, address, ptep);
2575                return 0;
2576        }
2577
2578        /*
2579         * If the process that created a MAP_PRIVATE mapping is about to
2580         * perform a COW due to a shared page count, attempt to satisfy
2581         * the allocation without using the existing reserves. The pagecache
2582         * page is used to determine if the reserve at this address was
2583         * consumed or not. If reserves were used, a partial faulted mapping
2584         * at the time of fork() could consume its reserves on COW instead
2585         * of the full address range.
2586         */
2587        if (!(vma->vm_flags & VM_MAYSHARE) &&
2588                        is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2589                        old_page != pagecache_page)
2590                outside_reserve = 1;
2591
2592        page_cache_get(old_page);
2593
2594        /* Drop page_table_lock as buddy allocator may be called */
2595        spin_unlock(&mm->page_table_lock);
2596        new_page = alloc_huge_page(vma, address, outside_reserve);
2597
2598        if (IS_ERR(new_page)) {
2599                long err = PTR_ERR(new_page);
2600                page_cache_release(old_page);
2601
2602                /*
2603                 * If a process owning a MAP_PRIVATE mapping fails to COW,
2604                 * it is due to references held by a child and an insufficient
2605                 * huge page pool. To guarantee the original mappers
2606                 * reliability, unmap the page from child processes. The child
2607                 * may get SIGKILLed if it later faults.
2608                 */
2609                if (outside_reserve) {
2610                        BUG_ON(huge_pte_none(pte));
2611                        if (unmap_ref_private(mm, vma, old_page, address)) {
2612                                BUG_ON(huge_pte_none(pte));
2613                                spin_lock(&mm->page_table_lock);
2614                                ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2615                                if (likely(pte_same(huge_ptep_get(ptep), pte)))
2616                                        goto retry_avoidcopy;
2617                                /*
2618                                 * race occurs while re-acquiring page_table_lock, and
2619                                 * our job is done.
2620                                 */
2621                                return 0;
2622                        }
2623                        WARN_ON_ONCE(1);
2624                }
2625
2626                /* Caller expects lock to be held */
2627                spin_lock(&mm->page_table_lock);
2628                if (err == -ENOMEM)
2629                        return VM_FAULT_OOM;
2630                else
2631                        return VM_FAULT_SIGBUS;
2632        }
2633
2634        /*
2635         * When the original hugepage is shared one, it does not have
2636         * anon_vma prepared.
2637         */
2638        if (unlikely(anon_vma_prepare(vma))) {
2639                page_cache_release(new_page);
2640                page_cache_release(old_page);
2641                /* Caller expects lock to be held */
2642                spin_lock(&mm->page_table_lock);
2643                return VM_FAULT_OOM;
2644        }
2645
2646        copy_user_huge_page(new_page, old_page, address, vma,
2647                            pages_per_huge_page(h));
2648        __SetPageUptodate(new_page);
2649
2650        mmun_start = address & huge_page_mask(h);
2651        mmun_end = mmun_start + huge_page_size(h);
2652        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2653        /*
2654         * Retake the page_table_lock to check for racing updates
2655         * before the page tables are altered
2656         */
2657        spin_lock(&mm->page_table_lock);
2658        ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2659        if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2660                /* Break COW */
2661                huge_ptep_clear_flush(vma, address, ptep);
2662                set_huge_pte_at(mm, address, ptep,
2663                                make_huge_pte(vma, new_page, 1));
2664                page_remove_rmap(old_page);
2665                hugepage_add_new_anon_rmap(new_page, vma, address);
2666                /* Make the old page be freed below */
2667                new_page = old_page;
2668        }
2669        spin_unlock(&mm->page_table_lock);
2670        mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2671        /* Caller expects lock to be held */
2672        spin_lock(&mm->page_table_lock);
2673        page_cache_release(new_page);
2674        page_cache_release(old_page);
2675        return 0;
2676}
2677
2678/* Return the pagecache page at a given address within a VMA */
2679static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2680                        struct vm_area_struct *vma, unsigned long address)
2681{
2682        struct address_space *mapping;
2683        pgoff_t idx;
2684
2685        mapping = vma->vm_file->f_mapping;
2686        idx = vma_hugecache_offset(h, vma, address);
2687
2688        return find_lock_page(mapping, idx);
2689}
2690
2691/*
2692 * Return whether there is a pagecache page to back given address within VMA.
2693 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2694 */
2695static bool hugetlbfs_pagecache_present(struct hstate *h,
2696                        struct vm_area_struct *vma, unsigned long address)
2697{
2698        struct address_space *mapping;
2699        pgoff_t idx;
2700        struct page *page;
2701
2702        mapping = vma->vm_file->f_mapping;
2703        idx = vma_hugecache_offset(h, vma, address);
2704
2705        page = find_get_page(mapping, idx);
2706        if (page)
2707                put_page(page);
2708        return page != NULL;
2709}
2710
2711static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2712                        unsigned long address, pte_t *ptep, unsigned int flags)
2713{
2714        struct hstate *h = hstate_vma(vma);
2715        int ret = VM_FAULT_SIGBUS;
2716        int anon_rmap = 0;
2717        pgoff_t idx;
2718        unsigned long size;
2719        struct page *page;
2720        struct address_space *mapping;
2721        pte_t new_pte;
2722
2723        /*
2724         * Currently, we are forced to kill the process in the event the
2725         * original mapper has unmapped pages from the child due to a failed
2726         * COW. Warn that such a situation has occurred as it may not be obvious
2727         */
2728        if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2729                pr_warning("PID %d killed due to inadequate hugepage pool\n",
2730                           current->pid);
2731                return ret;
2732        }
2733
2734        mapping = vma->vm_file->f_mapping;
2735        idx = vma_hugecache_offset(h, vma, address);
2736
2737        /*
2738         * Use page lock to guard against racing truncation
2739         * before we get page_table_lock.
2740         */
2741retry:
2742        page = find_lock_page(mapping, idx);
2743        if (!page) {
2744                size = i_size_read(mapping->host) >> huge_page_shift(h);
2745                if (idx >= size)
2746                        goto out;
2747                page = alloc_huge_page(vma, address, 0);
2748                if (IS_ERR(page)) {
2749                        ret = PTR_ERR(page);
2750                        if (ret == -ENOMEM)
2751                                ret = VM_FAULT_OOM;
2752                        else
2753                                ret = VM_FAULT_SIGBUS;
2754                        goto out;
2755                }
2756                clear_huge_page(page, address, pages_per_huge_page(h));
2757                __SetPageUptodate(page);
2758
2759                if (vma->vm_flags & VM_MAYSHARE) {
2760                        int err;
2761                        struct inode *inode = mapping->host;
2762
2763                        err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2764                        if (err) {
2765                                put_page(page);
2766                                if (err == -EEXIST)
2767                                        goto retry;
2768                                goto out;
2769                        }
2770
2771                        spin_lock(&inode->i_lock);
2772                        inode->i_blocks += blocks_per_huge_page(h);
2773                        spin_unlock(&inode->i_lock);
2774                } else {
2775                        lock_page(page);
2776                        if (unlikely(anon_vma_prepare(vma))) {
2777                                ret = VM_FAULT_OOM;
2778                                goto backout_unlocked;
2779                        }
2780                        anon_rmap = 1;
2781                }
2782        } else {
2783                /*
2784                 * If memory error occurs between mmap() and fault, some process
2785                 * don't have hwpoisoned swap entry for errored virtual address.
2786                 * So we need to block hugepage fault by PG_hwpoison bit check.
2787                 */
2788                if (unlikely(PageHWPoison(page))) {
2789                        ret = VM_FAULT_HWPOISON |
2790                                VM_FAULT_SET_HINDEX(hstate_index(h));
2791                        goto backout_unlocked;
2792                }
2793        }
2794
2795        /*
2796         * If we are going to COW a private mapping later, we examine the
2797         * pending reservations for this page now. This will ensure that
2798         * any allocations necessary to record that reservation occur outside
2799         * the spinlock.
2800         */
2801        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2802                if (vma_needs_reservation(h, vma, address) < 0) {
2803                        ret = VM_FAULT_OOM;
2804                        goto backout_unlocked;
2805                }
2806
2807        spin_lock(&mm->page_table_lock);
2808        size = i_size_read(mapping->host) >> huge_page_shift(h);
2809        if (idx >= size)
2810                goto backout;
2811
2812        ret = 0;
2813        if (!huge_pte_none(huge_ptep_get(ptep)))
2814                goto backout;
2815
2816        if (anon_rmap)
2817                hugepage_add_new_anon_rmap(page, vma, address);
2818        else
2819                page_dup_rmap(page);
2820        new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2821                                && (vma->vm_flags & VM_SHARED)));
2822        set_huge_pte_at(mm, address, ptep, new_pte);
2823
2824        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2825                /* Optimization, do the COW without a second fault */
2826                ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2827        }
2828
2829        spin_unlock(&mm->page_table_lock);
2830        unlock_page(page);
2831out:
2832        return ret;
2833
2834backout:
2835        spin_unlock(&mm->page_table_lock);
2836backout_unlocked:
2837        unlock_page(page);
2838        put_page(page);
2839        goto out;
2840}
2841
2842int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2843                        unsigned long address, unsigned int flags)
2844{
2845        pte_t *ptep;
2846        pte_t entry;
2847        int ret;
2848        struct page *page = NULL;
2849        struct page *pagecache_page = NULL;
2850        static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2851        struct hstate *h = hstate_vma(vma);
2852
2853        address &= huge_page_mask(h);
2854
2855        ptep = huge_pte_offset(mm, address);
2856        if (ptep) {
2857                entry = huge_ptep_get(ptep);
2858                if (unlikely(is_hugetlb_entry_migration(entry))) {
2859                        migration_entry_wait_huge(mm, ptep);
2860                        return 0;
2861                } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2862                        return VM_FAULT_HWPOISON_LARGE |
2863                                VM_FAULT_SET_HINDEX(hstate_index(h));
2864        }
2865
2866        ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2867        if (!ptep)
2868                return VM_FAULT_OOM;
2869
2870        /*
2871         * Serialize hugepage allocation and instantiation, so that we don't
2872         * get spurious allocation failures if two CPUs race to instantiate
2873         * the same page in the page cache.
2874         */
2875        mutex_lock(&hugetlb_instantiation_mutex);
2876        entry = huge_ptep_get(ptep);
2877        if (huge_pte_none(entry)) {
2878                ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2879                goto out_mutex;
2880        }
2881
2882        ret = 0;
2883
2884        /*
2885         * If we are going to COW the mapping later, we examine the pending
2886         * reservations for this page now. This will ensure that any
2887         * allocations necessary to record that reservation occur outside the
2888         * spinlock. For private mappings, we also lookup the pagecache
2889         * page now as it is used to determine if a reservation has been
2890         * consumed.
2891         */
2892        if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2893                if (vma_needs_reservation(h, vma, address) < 0) {
2894                        ret = VM_FAULT_OOM;
2895                        goto out_mutex;
2896                }
2897
2898                if (!(vma->vm_flags & VM_MAYSHARE))
2899                        pagecache_page = hugetlbfs_pagecache_page(h,
2900                                                                vma, address);
2901        }
2902
2903        /*
2904         * hugetlb_cow() requires page locks of pte_page(entry) and
2905         * pagecache_page, so here we need take the former one
2906         * when page != pagecache_page or !pagecache_page.
2907         * Note that locking order is always pagecache_page -> page,
2908         * so no worry about deadlock.
2909         */
2910        page = pte_page(entry);
2911        get_page(page);
2912        if (page != pagecache_page)
2913                lock_page(page);
2914
2915        spin_lock(&mm->page_table_lock);
2916        /* Check for a racing update before calling hugetlb_cow */
2917        if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2918                goto out_page_table_lock;
2919
2920
2921        if (flags & FAULT_FLAG_WRITE) {
2922                if (!huge_pte_write(entry)) {
2923                        ret = hugetlb_cow(mm, vma, address, ptep, entry,
2924                                                        pagecache_page);
2925                        goto out_page_table_lock;
2926                }
2927                entry = huge_pte_mkdirty(entry);
2928        }
2929        entry = pte_mkyoung(entry);
2930        if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2931                                                flags & FAULT_FLAG_WRITE))
2932                update_mmu_cache(vma, address, ptep);
2933
2934out_page_table_lock:
2935        spin_unlock(&mm->page_table_lock);
2936
2937        if (pagecache_page) {
2938                unlock_page(pagecache_page);
2939                put_page(pagecache_page);
2940        }
2941        if (page != pagecache_page)
2942                unlock_page(page);
2943        put_page(page);
2944
2945out_mutex:
2946        mutex_unlock(&hugetlb_instantiation_mutex);
2947
2948        return ret;
2949}
2950
2951/* Can be overriden by architectures */
2952__attribute__((weak)) struct page *
2953follow_huge_pud(struct mm_struct *mm, unsigned long address,
2954               pud_t *pud, int write)
2955{
2956        BUG();
2957        return NULL;
2958}
2959
2960long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2961                         struct page **pages, struct vm_area_struct **vmas,
2962                         unsigned long *position, unsigned long *nr_pages,
2963                         long i, unsigned int flags)
2964{
2965        unsigned long pfn_offset;
2966        unsigned long vaddr = *position;
2967        unsigned long remainder = *nr_pages;
2968        struct hstate *h = hstate_vma(vma);
2969
2970        spin_lock(&mm->page_table_lock);
2971        while (vaddr < vma->vm_end && remainder) {
2972                pte_t *pte;
2973                int absent;
2974                struct page *page;
2975
2976                /*
2977                 * Some archs (sparc64, sh*) have multiple pte_ts to
2978                 * each hugepage.  We have to make sure we get the
2979                 * first, for the page indexing below to work.
2980                 */
2981                pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2982                absent = !pte || huge_pte_none(huge_ptep_get(pte));
2983
2984                /*
2985                 * When coredumping, it suits get_dump_page if we just return
2986                 * an error where there's an empty slot with no huge pagecache
2987                 * to back it.  This way, we avoid allocating a hugepage, and
2988                 * the sparse dumpfile avoids allocating disk blocks, but its
2989                 * huge holes still show up with zeroes where they need to be.
2990                 */
2991                if (absent && (flags & FOLL_DUMP) &&
2992                    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2993                        remainder = 0;
2994                        break;
2995                }
2996
2997                /*
2998                 * We need call hugetlb_fault for both hugepages under migration
2999                 * (in which case hugetlb_fault waits for the migration,) and
3000                 * hwpoisoned hugepages (in which case we need to prevent the
3001                 * caller from accessing to them.) In order to do this, we use
3002                 * here is_swap_pte instead of is_hugetlb_entry_migration and
3003                 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3004                 * both cases, and because we can't follow correct pages
3005                 * directly from any kind of swap entries.
3006                 */
3007                if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3008                    ((flags & FOLL_WRITE) &&
3009                      !huge_pte_write(huge_ptep_get(pte)))) {
3010                        int ret;
3011
3012                        spin_unlock(&mm->page_table_lock);
3013                        ret = hugetlb_fault(mm, vma, vaddr,
3014                                (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3015                        spin_lock(&mm->page_table_lock);
3016                        if (!(ret & VM_FAULT_ERROR))
3017                                continue;
3018
3019                        remainder = 0;
3020                        break;
3021                }
3022
3023                pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3024                page = pte_page(huge_ptep_get(pte));
3025same_page:
3026                if (pages) {
3027                        pages[i] = mem_map_offset(page, pfn_offset);
3028                        get_page(pages[i]);
3029                }
3030
3031                if (vmas)
3032                        vmas[i] = vma;
3033
3034                vaddr += PAGE_SIZE;
3035                ++pfn_offset;
3036                --remainder;
3037                ++i;
3038                if (vaddr < vma->vm_end && remainder &&
3039                                pfn_offset < pages_per_huge_page(h)) {
3040                        /*
3041                         * We use pfn_offset to avoid touching the pageframes
3042                         * of this compound page.
3043                         */
3044                        goto same_page;
3045                }
3046        }
3047        spin_unlock(&mm->page_table_lock);
3048        *nr_pages = remainder;
3049        *position = vaddr;
3050
3051        return i ? i : -EFAULT;
3052}
3053
3054unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3055                unsigned long address, unsigned long end, pgprot_t newprot)
3056{
3057        struct mm_struct *mm = vma->vm_mm;
3058        unsigned long start = address;
3059        pte_t *ptep;
3060        pte_t pte;
3061        struct hstate *h = hstate_vma(vma);
3062        unsigned long pages = 0;
3063
3064        BUG_ON(address >= end);
3065        flush_cache_range(vma, address, end);
3066
3067        mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3068        spin_lock(&mm->page_table_lock);
3069        for (; address < end; address += huge_page_size(h)) {
3070                ptep = huge_pte_offset(mm, address);
3071                if (!ptep)
3072                        continue;
3073                if (huge_pmd_unshare(mm, &address, ptep)) {
3074                        pages++;
3075                        continue;
3076                }
3077                if (!huge_pte_none(huge_ptep_get(ptep))) {
3078                        pte = huge_ptep_get_and_clear(mm, address, ptep);
3079                        pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3080                        pte = arch_make_huge_pte(pte, vma, NULL, 0);
3081                        set_huge_pte_at(mm, address, ptep, pte);
3082                        pages++;
3083                }
3084        }
3085        spin_unlock(&mm->page_table_lock);
3086        /*
3087         * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3088         * may have cleared our pud entry and done put_page on the page table:
3089         * once we release i_mmap_mutex, another task can do the final put_page
3090         * and that page table be reused and filled with junk.
3091         */
3092        flush_tlb_range(vma, start, end);
3093        mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3094
3095        return pages << h->order;
3096}
3097
3098int hugetlb_reserve_pages(struct inode *inode,
3099                                        long from, long to,
3100                                        struct vm_area_struct *vma,
3101                                        vm_flags_t vm_flags)
3102{
3103        long ret, chg;
3104        struct hstate *h = hstate_inode(inode);
3105        struct hugepage_subpool *spool = subpool_inode(inode);
3106
3107        /*
3108         * Only apply hugepage reservation if asked. At fault time, an
3109         * attempt will be made for VM_NORESERVE to allocate a page
3110         * without using reserves
3111         */
3112        if (vm_flags & VM_NORESERVE)
3113                return 0;
3114
3115        /*
3116         * Shared mappings base their reservation on the number of pages that
3117         * are already allocated on behalf of the file. Private mappings need
3118         * to reserve the full area even if read-only as mprotect() may be
3119         * called to make the mapping read-write. Assume !vma is a shm mapping
3120         */
3121        if (!vma || vma->vm_flags & VM_MAYSHARE)
3122                chg = region_chg(&inode->i_mapping->private_list, from, to);
3123        else {
3124                struct resv_map *resv_map = resv_map_alloc();
3125                if (!resv_map)
3126                        return -ENOMEM;
3127
3128                chg = to - from;
3129
3130                set_vma_resv_map(vma, resv_map);
3131                set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3132        }
3133
3134        if (chg < 0) {
3135                ret = chg;
3136                goto out_err;
3137        }
3138
3139        /* There must be enough pages in the subpool for the mapping */
3140        if (hugepage_subpool_get_pages(spool, chg)) {
3141                ret = -ENOSPC;
3142                goto out_err;
3143        }
3144
3145        /*
3146         * Check enough hugepages are available for the reservation.
3147         * Hand the pages back to the subpool if there are not
3148         */
3149        ret = hugetlb_acct_memory(h, chg);
3150        if (ret < 0) {
3151                hugepage_subpool_put_pages(spool, chg);
3152                goto out_err;
3153        }
3154
3155        /*
3156         * Account for the reservations made. Shared mappings record regions
3157         * that have reservations as they are shared by multiple VMAs.
3158         * When the last VMA disappears, the region map says how much
3159         * the reservation was and the page cache tells how much of
3160         * the reservation was consumed. Private mappings are per-VMA and
3161         * only the consumed reservations are tracked. When the VMA
3162         * disappears, the original reservation is the VMA size and the
3163         * consumed reservations are stored in the map. Hence, nothing
3164         * else has to be done for private mappings here
3165         */
3166        if (!vma || vma->vm_flags & VM_MAYSHARE)
3167                region_add(&inode->i_mapping->private_list, from, to);
3168        return 0;
3169out_err:
3170        if (vma)
3171                resv_map_put(vma);
3172        return ret;
3173}
3174
3175void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3176{
3177        struct hstate *h = hstate_inode(inode);
3178        long chg = region_truncate(&inode->i_mapping->private_list, offset);
3179        struct hugepage_subpool *spool = subpool_inode(inode);
3180
3181        spin_lock(&inode->i_lock);
3182        inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3183        spin_unlock(&inode->i_lock);
3184
3185        hugepage_subpool_put_pages(spool, (chg - freed));
3186        hugetlb_acct_memory(h, -(chg - freed));
3187}
3188
3189#ifdef CONFIG_MEMORY_FAILURE
3190
3191/* Should be called in hugetlb_lock */
3192static int is_hugepage_on_freelist(struct page *hpage)
3193{
3194        struct page *page;
3195        struct page *tmp;
3196        struct hstate *h = page_hstate(hpage);
3197        int nid = page_to_nid(hpage);
3198
3199        list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3200                if (page == hpage)
3201                        return 1;
3202        return 0;
3203}
3204
3205/*
3206 * This function is called from memory failure code.
3207 * Assume the caller holds page lock of the head page.
3208 */
3209int dequeue_hwpoisoned_huge_page(struct page *hpage)
3210{
3211        struct hstate *h = page_hstate(hpage);
3212        int nid = page_to_nid(hpage);
3213        int ret = -EBUSY;
3214
3215        spin_lock(&hugetlb_lock);
3216        if (is_hugepage_on_freelist(hpage)) {
3217                /*
3218                 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3219                 * but dangling hpage->lru can trigger list-debug warnings
3220                 * (this happens when we call unpoison_memory() on it),
3221                 * so let it point to itself with list_del_init().
3222                 */
3223                list_del_init(&hpage->lru);
3224                set_page_refcounted(hpage);
3225                h->free_huge_pages--;
3226                h->free_huge_pages_node[nid]--;
3227                ret = 0;
3228        }
3229        spin_unlock(&hugetlb_lock);
3230        return ret;
3231}
3232#endif
3233
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.