mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/highmem.h>
  12 #include <linux/nodemask.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/mempolicy.h>
  15 #include <linux/cpuset.h>
  16 #include <linux/mutex.h>
  17
  18 #include <asm/page.h>
  19 #include <asm/pgtable.h>
  20
  21 #include <linux/hugetlb.h>
  22 #include "internal.h"
  23
  24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
  26 static unsigned long surplus_huge_pages;
  27 unsigned long max_huge_pages;
  28 static struct list_head hugepage_freelists[MAX_NUMNODES];
  29 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  30 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  31 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
  32 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  33 unsigned long hugepages_treat_as_movable;
  34 int hugetlb_dynamic_pool;
  35
  36 /*
  37  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  38  */
  39 static DEFINE_SPINLOCK(hugetlb_lock);
  40
  41 static void clear_huge_page(struct page *page, unsigned long addr)
  42 {
  43         int i;
  44
  45         might_sleep();
  46         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  47                 cond_resched();
  48                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
  49         }
  50 }
  51
  52 static void copy_huge_page(struct page *dst, struct page *src,
  53                            unsigned long addr, struct vm_area_struct *vma)
  54 {
  55         int i;
  56
  57         might_sleep();
  58         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  59                 cond_resched();
  60                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  61         }
  62 }
  63
  64 static void enqueue_huge_page(struct page *page)
  65 {
  66         int nid = page_to_nid(page);
  67         list_add(&page->lru, &hugepage_freelists[nid]);
  68         free_huge_pages++;
  69         free_huge_pages_node[nid]++;
  70 }
  71
  72 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  73                                 unsigned long address)
  74 {
  75         int nid;
  76         struct page *page = NULL;
  77         struct mempolicy *mpol;
  78         struct zonelist *zonelist = huge_zonelist(vma, address,
  79                                         htlb_alloc_mask, &mpol);
  80         struct zone **z;
  81
  82         for (z = zonelist->zones; *z; z++) {
  83                 nid = zone_to_nid(*z);
  84                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
  85                     !list_empty(&hugepage_freelists[nid])) {
  86                         page = list_entry(hugepage_freelists[nid].next,
  87                                           struct page, lru);
  88                         list_del(&page->lru);
  89                         free_huge_pages--;
  90                         free_huge_pages_node[nid]--;
  91                         if (vma && vma->vm_flags & VM_MAYSHARE)
  92                                 resv_huge_pages--;
  93                         break;
  94                 }
  95         }
  96         mpol_free(mpol);        /* unref if mpol !NULL */
  97         return page;
  98 }
  99
 100 static void update_and_free_page(struct page *page)
 101 {
 102         int i;
 103         nr_huge_pages--;
 104         nr_huge_pages_node[page_to_nid(page)]--;
 105         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 106                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 107                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 108                                 1 << PG_private | 1<< PG_writeback);
 109         }
 110         set_compound_page_dtor(page, NULL);
 111         set_page_refcounted(page);
 112         __free_pages(page, HUGETLB_PAGE_ORDER);
 113 }
 114
 115 static void free_huge_page(struct page *page)
 116 {
 117         int nid = page_to_nid(page);
 118
 119         BUG_ON(page_count(page));
 120         INIT_LIST_HEAD(&page->lru);
 121
 122         spin_lock(&hugetlb_lock);
 123         if (surplus_huge_pages_node[nid]) {
 124                 update_and_free_page(page);
 125                 surplus_huge_pages--;
 126                 surplus_huge_pages_node[nid]--;
 127         } else {
 128                 enqueue_huge_page(page);
 129         }
 130         spin_unlock(&hugetlb_lock);
 131 }
 132
 133 /*
 134  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 135  * balanced by operating on them in a round-robin fashion.
 136  * Returns 1 if an adjustment was made.
 137  */
 138 static int adjust_pool_surplus(int delta)
 139 {
 140         static int prev_nid;
 141         int nid = prev_nid;
 142         int ret = 0;
 143
 144         VM_BUG_ON(delta != -1 && delta != 1);
 145         do {
 146                 nid = next_node(nid, node_online_map);
 147                 if (nid == MAX_NUMNODES)
 148                         nid = first_node(node_online_map);
 149
 150                 /* To shrink on this node, there must be a surplus page */
 151                 if (delta < 0 && !surplus_huge_pages_node[nid])
 152                         continue;
 153                 /* Surplus cannot exceed the total number of pages */
 154                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 155                                                 nr_huge_pages_node[nid])
 156                         continue;
 157
 158                 surplus_huge_pages += delta;
 159                 surplus_huge_pages_node[nid] += delta;
 160                 ret = 1;
 161                 break;
 162         } while (nid != prev_nid);
 163
 164         prev_nid = nid;
 165         return ret;
 166 }
 167
 168 static int alloc_fresh_huge_page(void)
 169 {
 170         static int prev_nid;
 171         struct page *page;
 172         int nid;
 173
 174         /*
 175          * Copy static prev_nid to local nid, work on that, then copy it
 176          * back to prev_nid afterwards: otherwise there's a window in which
 177          * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node.
 178          * But we don't need to use a spin_lock here: it really doesn't
 179          * matter if occasionally a racer chooses the same nid as we do.
 180          */
 181         nid = next_node(prev_nid, node_online_map);
 182         if (nid == MAX_NUMNODES)
 183                 nid = first_node(node_online_map);
 184         prev_nid = nid;
 185
 186         page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 187                                         HUGETLB_PAGE_ORDER);
 188         if (page) {
 189                 set_compound_page_dtor(page, free_huge_page);
 190                 spin_lock(&hugetlb_lock);
 191                 nr_huge_pages++;
 192                 nr_huge_pages_node[page_to_nid(page)]++;
 193                 spin_unlock(&hugetlb_lock);
 194                 put_page(page); /* free it into the hugepage allocator */
 195                 return 1;
 196         }
 197         return 0;
 198 }
 199
 200 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 201                                                 unsigned long address)
 202 {
 203         struct page *page;
 204
 205         /* Check if the dynamic pool is enabled */
 206         if (!hugetlb_dynamic_pool)
 207                 return NULL;
 208
 209         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 210                                         HUGETLB_PAGE_ORDER);
 211         if (page) {
 212                 set_compound_page_dtor(page, free_huge_page);
 213                 spin_lock(&hugetlb_lock);
 214                 nr_huge_pages++;
 215                 nr_huge_pages_node[page_to_nid(page)]++;
 216                 surplus_huge_pages++;
 217                 surplus_huge_pages_node[page_to_nid(page)]++;
 218                 spin_unlock(&hugetlb_lock);
 219         }
 220
 221         return page;
 222 }
 223
 224 /*
 225  * Increase the hugetlb pool such that it can accomodate a reservation
 226  * of size 'delta'.
 227  */
 228 static int gather_surplus_pages(int delta)
 229 {
 230         struct list_head surplus_list;
 231         struct page *page, *tmp;
 232         int ret, i;
 233         int needed, allocated;
 234
 235         needed = (resv_huge_pages + delta) - free_huge_pages;
 236         if (needed <= 0)
 237                 return 0;
 238
 239         allocated = 0;
 240         INIT_LIST_HEAD(&surplus_list);
 241
 242         ret = -ENOMEM;
 243 retry:
 244         spin_unlock(&hugetlb_lock);
 245         for (i = 0; i < needed; i++) {
 246                 page = alloc_buddy_huge_page(NULL, 0);
 247                 if (!page) {
 248                         /*
 249                          * We were not able to allocate enough pages to
 250                          * satisfy the entire reservation so we free what
 251                          * we've allocated so far.
 252                          */
 253                         spin_lock(&hugetlb_lock);
 254                         needed = 0;
 255                         goto free;
 256                 }
 257
 258                 list_add(&page->lru, &surplus_list);
 259         }
 260         allocated += needed;
 261
 262         /*
 263          * After retaking hugetlb_lock, we need to recalculate 'needed'
 264          * because either resv_huge_pages or free_huge_pages may have changed.
 265          */
 266         spin_lock(&hugetlb_lock);
 267         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
 268         if (needed > 0)
 269                 goto retry;
 270
 271         /*
 272          * The surplus_list now contains _at_least_ the number of extra pages
 273          * needed to accomodate the reservation.  Add the appropriate number
 274          * of pages to the hugetlb pool and free the extras back to the buddy
 275          * allocator.
 276          */
 277         needed += allocated;
 278         ret = 0;
 279 free:
 280         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 281                 list_del(&page->lru);
 282                 if ((--needed) >= 0)
 283                         enqueue_huge_page(page);
 284                 else
 285                         update_and_free_page(page);
 286         }
 287
 288         return ret;
 289 }
 290
 291 /*
 292  * When releasing a hugetlb pool reservation, any surplus pages that were
 293  * allocated to satisfy the reservation must be explicitly freed if they were
 294  * never used.
 295  */
 296 void return_unused_surplus_pages(unsigned long unused_resv_pages)
 297 {
 298         static int nid = -1;
 299         struct page *page;
 300         unsigned long nr_pages;
 301
 302         nr_pages = min(unused_resv_pages, surplus_huge_pages);
 303
 304         while (nr_pages) {
 305                 nid = next_node(nid, node_online_map);
 306                 if (nid == MAX_NUMNODES)
 307                         nid = first_node(node_online_map);
 308
 309                 if (!surplus_huge_pages_node[nid])
 310                         continue;
 311
 312                 if (!list_empty(&hugepage_freelists[nid])) {
 313                         page = list_entry(hugepage_freelists[nid].next,
 314                                           struct page, lru);
 315                         list_del(&page->lru);
 316                         update_and_free_page(page);
 317                         free_huge_pages--;
 318                         free_huge_pages_node[nid]--;
 319                         surplus_huge_pages--;
 320                         surplus_huge_pages_node[nid]--;
 321                         nr_pages--;
 322                 }
 323         }
 324 }
 325
 326 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 327                                     unsigned long addr)
 328 {
 329         struct page *page = NULL;
 330         int use_reserved_page = vma->vm_flags & VM_MAYSHARE;
 331
 332         spin_lock(&hugetlb_lock);
 333         if (!use_reserved_page && (free_huge_pages <= resv_huge_pages))
 334                 goto fail;
 335
 336         page = dequeue_huge_page(vma, addr);
 337         if (!page)
 338                 goto fail;
 339
 340         spin_unlock(&hugetlb_lock);
 341         set_page_refcounted(page);
 342         return page;
 343
 344 fail:
 345         spin_unlock(&hugetlb_lock);
 346
 347         /*
 348          * Private mappings do not use reserved huge pages so the allocation
 349          * may have failed due to an undersized hugetlb pool.  Try to grab a
 350          * surplus huge page from the buddy allocator.
 351          */
 352         if (!use_reserved_page)
 353                 page = alloc_buddy_huge_page(vma, addr);
 354
 355         return page;
 356 }
 357
 358 static int __init hugetlb_init(void)
 359 {
 360         unsigned long i;
 361
 362         if (HPAGE_SHIFT == 0)
 363                 return 0;
 364
 365         for (i = 0; i < MAX_NUMNODES; ++i)
 366                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 367
 368         for (i = 0; i < max_huge_pages; ++i) {
 369                 if (!alloc_fresh_huge_page())
 370                         break;
 371         }
 372         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 373         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 374         return 0;
 375 }
 376 module_init(hugetlb_init);
 377
 378 static int __init hugetlb_setup(char *s)
 379 {
 380         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 381                 max_huge_pages = 0;
 382         return 1;
 383 }
 384 __setup("hugepages=", hugetlb_setup);
 385
 386 static unsigned int cpuset_mems_nr(unsigned int *array)
 387 {
 388         int node;
 389         unsigned int nr = 0;
 390
 391         for_each_node_mask(node, cpuset_current_mems_allowed)
 392                 nr += array[node];
 393
 394         return nr;
 395 }
 396
 397 #ifdef CONFIG_SYSCTL
 398 #ifdef CONFIG_HIGHMEM
 399 static void try_to_free_low(unsigned long count)
 400 {
 401         int i;
 402
 403         for (i = 0; i < MAX_NUMNODES; ++i) {
 404                 struct page *page, *next;
 405                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 406                         if (PageHighMem(page))
 407                                 continue;
 408                         list_del(&page->lru);
 409                         update_and_free_page(page);
 410                         free_huge_pages--;
 411                         free_huge_pages_node[page_to_nid(page)]--;
 412                         if (count >= nr_huge_pages)
 413                                 return;
 414                 }
 415         }
 416 }
 417 #else
 418 static inline void try_to_free_low(unsigned long count)
 419 {
 420 }
 421 #endif
 422
 423 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 424 static unsigned long set_max_huge_pages(unsigned long count)
 425 {
 426         unsigned long min_count, ret;
 427
 428         /*
 429          * Increase the pool size
 430          * First take pages out of surplus state.  Then make up the
 431          * remaining difference by allocating fresh huge pages.
 432          */
 433         spin_lock(&hugetlb_lock);
 434         while (surplus_huge_pages && count > persistent_huge_pages) {
 435                 if (!adjust_pool_surplus(-1))
 436                         break;
 437         }
 438
 439         while (count > persistent_huge_pages) {
 440                 int ret;
 441                 /*
 442                  * If this allocation races such that we no longer need the
 443                  * page, free_huge_page will handle it by freeing the page
 444                  * and reducing the surplus.
 445                  */
 446                 spin_unlock(&hugetlb_lock);
 447                 ret = alloc_fresh_huge_page();
 448                 spin_lock(&hugetlb_lock);
 449                 if (!ret)
 450                         goto out;
 451
 452         }
 453         if (count >= persistent_huge_pages)
 454                 goto out;
 455
 456         /*
 457          * Decrease the pool size
 458          * First return free pages to the buddy allocator (being careful
 459          * to keep enough around to satisfy reservations).  Then place
 460          * pages into surplus state as needed so the pool will shrink
 461          * to the desired size as pages become free.
 462          */
 463         min_count = max(count, resv_huge_pages);
 464         try_to_free_low(min_count);
 465         while (min_count < persistent_huge_pages) {
 466                 struct page *page = dequeue_huge_page(NULL, 0);
 467                 if (!page)
 468                         break;
 469                 update_and_free_page(page);
 470         }
 471         while (count < persistent_huge_pages) {
 472                 if (!adjust_pool_surplus(1))
 473                         break;
 474         }
 475 out:
 476         ret = persistent_huge_pages;
 477         spin_unlock(&hugetlb_lock);
 478         return ret;
 479 }
 480
 481 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 482                            struct file *file, void __user *buffer,
 483                            size_t *length, loff_t *ppos)
 484 {
 485         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 486         max_huge_pages = set_max_huge_pages(max_huge_pages);
 487         return 0;
 488 }
 489
 490 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 491                         struct file *file, void __user *buffer,
 492                         size_t *length, loff_t *ppos)
 493 {
 494         proc_dointvec(table, write, file, buffer, length, ppos);
 495         if (hugepages_treat_as_movable)
 496                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 497         else
 498                 htlb_alloc_mask = GFP_HIGHUSER;
 499         return 0;
 500 }
 501
 502 #endif /* CONFIG_SYSCTL */
 503
 504 int hugetlb_report_meminfo(char *buf)
 505 {
 506         return sprintf(buf,
 507                         "HugePages_Total: %5lu\n"
 508                         "HugePages_Free:  %5lu\n"
 509                         "HugePages_Rsvd:  %5lu\n"
 510                         "HugePages_Surp:  %5lu\n"
 511                         "Hugepagesize:    %5lu kB\n",
 512                         nr_huge_pages,
 513                         free_huge_pages,
 514                         resv_huge_pages,
 515                         surplus_huge_pages,
 516                         HPAGE_SIZE/1024);
 517 }
 518
 519 int hugetlb_report_node_meminfo(int nid, char *buf)
 520 {
 521         return sprintf(buf,
 522                 "Node %d HugePages_Total: %5u\n"
 523                 "Node %d HugePages_Free:  %5u\n",
 524                 nid, nr_huge_pages_node[nid],
 525                 nid, free_huge_pages_node[nid]);
 526 }
 527
 528 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 529 unsigned long hugetlb_total_pages(void)
 530 {
 531         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 532 }
 533
 534 /*
 535  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 536  * handle_mm_fault() to try to instantiate regular-sized pages in the
 537  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 538  * this far.
 539  */
 540 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 541 {
 542         BUG();
 543         return 0;
 544 }
 545
 546 struct vm_operations_struct hugetlb_vm_ops = {
 547         .fault = hugetlb_vm_op_fault,
 548 };
 549
 550 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 551                                 int writable)
 552 {
 553         pte_t entry;
 554
 555         if (writable) {
 556                 entry =
 557                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 558         } else {
 559                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 560         }
 561         entry = pte_mkyoung(entry);
 562         entry = pte_mkhuge(entry);
 563
 564         return entry;
 565 }
 566
 567 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 568                                    unsigned long address, pte_t *ptep)
 569 {
 570         pte_t entry;
 571
 572         entry = pte_mkwrite(pte_mkdirty(*ptep));
 573         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 574                 update_mmu_cache(vma, address, entry);
 575         }
 576 }
 577
 578
 579 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 580                             struct vm_area_struct *vma)
 581 {
 582         pte_t *src_pte, *dst_pte, entry;
 583         struct page *ptepage;
 584         unsigned long addr;
 585         int cow;
 586
 587         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 588
 589         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 590                 src_pte = huge_pte_offset(src, addr);
 591                 if (!src_pte)
 592                         continue;
 593                 dst_pte = huge_pte_alloc(dst, addr);
 594                 if (!dst_pte)
 595                         goto nomem;
 596                 spin_lock(&dst->page_table_lock);
 597                 spin_lock(&src->page_table_lock);
 598                 if (!pte_none(*src_pte)) {
 599                         if (cow)
 600                                 ptep_set_wrprotect(src, addr, src_pte);
 601                         entry = *src_pte;
 602                         ptepage = pte_page(entry);
 603                         get_page(ptepage);
 604                         set_huge_pte_at(dst, addr, dst_pte, entry);
 605                 }
 606                 spin_unlock(&src->page_table_lock);
 607                 spin_unlock(&dst->page_table_lock);
 608         }
 609         return 0;
 610
 611 nomem:
 612         return -ENOMEM;
 613 }
 614
 615 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 616                             unsigned long end)
 617 {
 618         struct mm_struct *mm = vma->vm_mm;
 619         unsigned long address;
 620         pte_t *ptep;
 621         pte_t pte;
 622         struct page *page;
 623         struct page *tmp;
 624         /*
 625          * A page gathering list, protected by per file i_mmap_lock. The
 626          * lock is used to avoid list corruption from multiple unmapping
 627          * of the same page since we are using page->lru.
 628          */
 629         LIST_HEAD(page_list);
 630
 631         WARN_ON(!is_vm_hugetlb_page(vma));
 632         BUG_ON(start & ~HPAGE_MASK);
 633         BUG_ON(end & ~HPAGE_MASK);
 634
 635         spin_lock(&mm->page_table_lock);
 636         for (address = start; address < end; address += HPAGE_SIZE) {
 637                 ptep = huge_pte_offset(mm, address);
 638                 if (!ptep)
 639                         continue;
 640
 641                 if (huge_pmd_unshare(mm, &address, ptep))
 642                         continue;
 643
 644                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 645                 if (pte_none(pte))
 646                         continue;
 647
 648                 page = pte_page(pte);
 649                 if (pte_dirty(pte))
 650                         set_page_dirty(page);
 651                 list_add(&page->lru, &page_list);
 652         }
 653         spin_unlock(&mm->page_table_lock);
 654         flush_tlb_range(vma, start, end);
 655         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 656                 list_del(&page->lru);
 657                 put_page(page);
 658         }
 659 }
 660
 661 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 662                           unsigned long end)
 663 {
 664         /*
 665          * It is undesirable to test vma->vm_file as it should be non-null
 666          * for valid hugetlb area. However, vm_file will be NULL in the error
 667          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 668          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 669          * to clean up. Since no pte has actually been setup, it is safe to
 670          * do nothing in this case.
 671          */
 672         if (vma->vm_file) {
 673                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 674                 __unmap_hugepage_range(vma, start, end);
 675                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 676         }
 677 }
 678
 679 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 680                         unsigned long address, pte_t *ptep, pte_t pte)
 681 {
 682         struct page *old_page, *new_page;
 683         int avoidcopy;
 684
 685         old_page = pte_page(pte);
 686
 687         /* If no-one else is actually using this page, avoid the copy
 688          * and just make the page writable */
 689         avoidcopy = (page_count(old_page) == 1);
 690         if (avoidcopy) {
 691                 set_huge_ptep_writable(vma, address, ptep);
 692                 return 0;
 693         }
 694
 695         page_cache_get(old_page);
 696         new_page = alloc_huge_page(vma, address);
 697
 698         if (!new_page) {
 699                 page_cache_release(old_page);
 700                 return VM_FAULT_OOM;
 701         }
 702
 703         spin_unlock(&mm->page_table_lock);
 704         copy_huge_page(new_page, old_page, address, vma);
 705         spin_lock(&mm->page_table_lock);
 706
 707         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 708         if (likely(pte_same(*ptep, pte))) {
 709                 /* Break COW */
 710                 set_huge_pte_at(mm, address, ptep,
 711                                 make_huge_pte(vma, new_page, 1));
 712                 /* Make the old page be freed below */
 713                 new_page = old_page;
 714         }
 715         page_cache_release(new_page);
 716         page_cache_release(old_page);
 717         return 0;
 718 }
 719
 720 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 721                         unsigned long address, pte_t *ptep, int write_access)
 722 {
 723         int ret = VM_FAULT_SIGBUS;
 724         unsigned long idx;
 725         unsigned long size;
 726         struct page *page;
 727         struct address_space *mapping;
 728         pte_t new_pte;
 729
 730         mapping = vma->vm_file->f_mapping;
 731         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 732                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 733
 734         /*
 735          * Use page lock to guard against racing truncation
 736          * before we get page_table_lock.
 737          */
 738 retry:
 739         page = find_lock_page(mapping, idx);
 740         if (!page) {
 741                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 742                 if (idx >= size)
 743                         goto out;
 744                 if (hugetlb_get_quota(mapping))
 745                         goto out;
 746                 page = alloc_huge_page(vma, address);
 747                 if (!page) {
 748                         hugetlb_put_quota(mapping);
 749                         ret = VM_FAULT_OOM;
 750                         goto out;
 751                 }
 752                 clear_huge_page(page, address);
 753
 754                 if (vma->vm_flags & VM_SHARED) {
 755                         int err;
 756
 757                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 758                         if (err) {
 759                                 put_page(page);
 760                                 hugetlb_put_quota(mapping);
 761                                 if (err == -EEXIST)
 762                                         goto retry;
 763                                 goto out;
 764                         }
 765                 } else
 766                         lock_page(page);
 767         }
 768
 769         spin_lock(&mm->page_table_lock);
 770         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 771         if (idx >= size)
 772                 goto backout;
 773
 774         ret = 0;
 775         if (!pte_none(*ptep))
 776                 goto backout;
 777
 778         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 779                                 && (vma->vm_flags & VM_SHARED)));
 780         set_huge_pte_at(mm, address, ptep, new_pte);
 781
 782         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 783                 /* Optimization, do the COW without a second fault */
 784                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 785         }
 786
 787         spin_unlock(&mm->page_table_lock);
 788         unlock_page(page);
 789 out:
 790         return ret;
 791
 792 backout:
 793         spin_unlock(&mm->page_table_lock);
 794         hugetlb_put_quota(mapping);
 795         unlock_page(page);
 796         put_page(page);
 797         goto out;
 798 }
 799
 800 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 801                         unsigned long address, int write_access)
 802 {
 803         pte_t *ptep;
 804         pte_t entry;
 805         int ret;
 806         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 807
 808         ptep = huge_pte_alloc(mm, address);
 809         if (!ptep)
 810                 return VM_FAULT_OOM;
 811
 812         /*
 813          * Serialize hugepage allocation and instantiation, so that we don't
 814          * get spurious allocation failures if two CPUs race to instantiate
 815          * the same page in the page cache.
 816          */
 817         mutex_lock(&hugetlb_instantiation_mutex);
 818         entry = *ptep;
 819         if (pte_none(entry)) {
 820                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 821                 mutex_unlock(&hugetlb_instantiation_mutex);
 822                 return ret;
 823         }
 824
 825         ret = 0;
 826
 827         spin_lock(&mm->page_table_lock);
 828         /* Check for a racing update before calling hugetlb_cow */
 829         if (likely(pte_same(entry, *ptep)))
 830                 if (write_access && !pte_write(entry))
 831                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 832         spin_unlock(&mm->page_table_lock);
 833         mutex_unlock(&hugetlb_instantiation_mutex);
 834
 835         return ret;
 836 }
 837
 838 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 839                         struct page **pages, struct vm_area_struct **vmas,
 840                         unsigned long *position, int *length, int i)
 841 {
 842         unsigned long pfn_offset;
 843         unsigned long vaddr = *position;
 844         int remainder = *length;
 845
 846         spin_lock(&mm->page_table_lock);
 847         while (vaddr < vma->vm_end && remainder) {
 848                 pte_t *pte;
 849                 struct page *page;
 850
 851                 /*
 852                  * Some archs (sparc64, sh*) have multiple pte_ts to
 853                  * each hugepage.  We have to make * sure we get the
 854                  * first, for the page indexing below to work.
 855                  */
 856                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 857
 858                 if (!pte || pte_none(*pte)) {
 859                         int ret;
 860
 861                         spin_unlock(&mm->page_table_lock);
 862                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 863                         spin_lock(&mm->page_table_lock);
 864                         if (!(ret & VM_FAULT_ERROR))
 865                                 continue;
 866
 867                         remainder = 0;
 868                         if (!i)
 869                                 i = -EFAULT;
 870                         break;
 871                 }
 872
 873                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 874                 page = pte_page(*pte);
 875 same_page:
 876                 if (pages) {
 877                         get_page(page);
 878                         pages[i] = page + pfn_offset;
 879                 }
 880
 881                 if (vmas)
 882                         vmas[i] = vma;
 883
 884                 vaddr += PAGE_SIZE;
 885                 ++pfn_offset;
 886                 --remainder;
 887                 ++i;
 888                 if (vaddr < vma->vm_end && remainder &&
 889                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 890                         /*
 891                          * We use pfn_offset to avoid touching the pageframes
 892                          * of this compound page.
 893                          */
 894                         goto same_page;
 895                 }
 896         }
 897         spin_unlock(&mm->page_table_lock);
 898         *length = remainder;
 899         *position = vaddr;
 900
 901         return i;
 902 }
 903
 904 void hugetlb_change_protection(struct vm_area_struct *vma,
 905                 unsigned long address, unsigned long end, pgprot_t newprot)
 906 {
 907         struct mm_struct *mm = vma->vm_mm;
 908         unsigned long start = address;
 909         pte_t *ptep;
 910         pte_t pte;
 911
 912         BUG_ON(address >= end);
 913         flush_cache_range(vma, address, end);
 914
 915         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 916         spin_lock(&mm->page_table_lock);
 917         for (; address < end; address += HPAGE_SIZE) {
 918                 ptep = huge_pte_offset(mm, address);
 919                 if (!ptep)
 920                         continue;
 921                 if (huge_pmd_unshare(mm, &address, ptep))
 922                         continue;
 923                 if (!pte_none(*ptep)) {
 924                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 925                         pte = pte_mkhuge(pte_modify(pte, newprot));
 926                         set_huge_pte_at(mm, address, ptep, pte);
 927                 }
 928         }
 929         spin_unlock(&mm->page_table_lock);
 930         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 931
 932         flush_tlb_range(vma, start, end);
 933 }
 934
 935 struct file_region {
 936         struct list_head link;
 937         long from;
 938         long to;
 939 };
 940
 941 static long region_add(struct list_head *head, long f, long t)
 942 {
 943         struct file_region *rg, *nrg, *trg;
 944
 945         /* Locate the region we are either in or before. */
 946         list_for_each_entry(rg, head, link)
 947                 if (f <= rg->to)
 948                         break;
 949
 950         /* Round our left edge to the current segment if it encloses us. */
 951         if (f > rg->from)
 952                 f = rg->from;
 953
 954         /* Check for and consume any regions we now overlap with. */
 955         nrg = rg;
 956         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 957                 if (&rg->link == head)
 958                         break;
 959                 if (rg->from > t)
 960                         break;
 961
 962                 /* If this area reaches higher then extend our area to
 963                  * include it completely.  If this is not the first area
 964                  * which we intend to reuse, free it. */
 965                 if (rg->to > t)
 966                         t = rg->to;
 967                 if (rg != nrg) {
 968                         list_del(&rg->link);
 969                         kfree(rg);
 970                 }
 971         }
 972         nrg->from = f;
 973         nrg->to = t;
 974         return 0;
 975 }
 976
 977 static long region_chg(struct list_head *head, long f, long t)
 978 {
 979         struct file_region *rg, *nrg;
 980         long chg = 0;
 981
 982         /* Locate the region we are before or in. */
 983         list_for_each_entry(rg, head, link)
 984                 if (f <= rg->to)
 985                         break;
 986
 987         /* If we are below the current region then a new region is required.
 988          * Subtle, allocate a new region at the position but make it zero
 989          * size such that we can guarentee to record the reservation. */
 990         if (&rg->link == head || t < rg->from) {
 991                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 992                 if (nrg == 0)
 993                         return -ENOMEM;
 994                 nrg->from = f;
 995                 nrg->to   = f;
 996                 INIT_LIST_HEAD(&nrg->link);
 997                 list_add(&nrg->link, rg->link.prev);
 998
 999                 return t - f;
1000         }
1001
1002         /* Round our left edge to the current segment if it encloses us. */
1003         if (f > rg->from)
1004                 f = rg->from;
1005         chg = t - f;
1006
1007         /* Check for and consume any regions we now overlap with. */
1008         list_for_each_entry(rg, rg->link.prev, link) {
1009                 if (&rg->link == head)
1010                         break;
1011                 if (rg->from > t)
1012                         return chg;
1013
1014                 /* We overlap with this area, if it extends futher than
1015                  * us then we must extend ourselves.  Account for its
1016                  * existing reservation. */
1017                 if (rg->to > t) {
1018                         chg += rg->to - t;
1019                         t = rg->to;
1020                 }
1021                 chg -= rg->to - rg->from;
1022         }
1023         return chg;
1024 }
1025
1026 static long region_truncate(struct list_head *head, long end)
1027 {
1028         struct file_region *rg, *trg;
1029         long chg = 0;
1030
1031         /* Locate the region we are either in or before. */
1032         list_for_each_entry(rg, head, link)
1033                 if (end <= rg->to)
1034                         break;
1035         if (&rg->link == head)
1036                 return 0;
1037
1038         /* If we are in the middle of a region then adjust it. */
1039         if (end > rg->from) {
1040                 chg = rg->to - end;
1041                 rg->to = end;
1042                 rg = list_entry(rg->link.next, typeof(*rg), link);
1043         }
1044
1045         /* Drop any remaining regions. */
1046         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1047                 if (&rg->link == head)
1048                         break;
1049                 chg += rg->to - rg->from;
1050                 list_del(&rg->link);
1051                 kfree(rg);
1052         }
1053         return chg;
1054 }
1055
1056 static int hugetlb_acct_memory(long delta)
1057 {
1058         int ret = -ENOMEM;
1059
1060         spin_lock(&hugetlb_lock);
1061         /*
1062          * When cpuset is configured, it breaks the strict hugetlb page
1063          * reservation as the accounting is done on a global variable. Such
1064          * reservation is completely rubbish in the presence of cpuset because
1065          * the reservation is not checked against page availability for the
1066          * current cpuset. Application can still potentially OOM'ed by kernel
1067          * with lack of free htlb page in cpuset that the task is in.
1068          * Attempt to enforce strict accounting with cpuset is almost
1069          * impossible (or too ugly) because cpuset is too fluid that
1070          * task or memory node can be dynamically moved between cpusets.
1071          *
1072          * The change of semantics for shared hugetlb mapping with cpuset is
1073          * undesirable. However, in order to preserve some of the semantics,
1074          * we fall back to check against current free page availability as
1075          * a best attempt and hopefully to minimize the impact of changing
1076          * semantics that cpuset has.
1077          */
1078         if (delta > 0) {
1079                 if (gather_surplus_pages(delta) < 0)
1080                         goto out;
1081
1082                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1083                         goto out;
1084         }
1085
1086         ret = 0;
1087         resv_huge_pages += delta;
1088         if (delta < 0)
1089                 return_unused_surplus_pages((unsigned long) -delta);
1090
1091 out:
1092         spin_unlock(&hugetlb_lock);
1093         return ret;
1094 }
1095
1096 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1097 {
1098         long ret, chg;
1099
1100         chg = region_chg(&inode->i_mapping->private_list, from, to);
1101         if (chg < 0)
1102                 return chg;
1103
1104         ret = hugetlb_acct_memory(chg);
1105         if (ret < 0)
1106                 return ret;
1107         region_add(&inode->i_mapping->private_list, from, to);
1108         return 0;
1109 }
1110
1111 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1112 {
1113         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1114         hugetlb_acct_memory(freed - chg);
1115 }