err.no Git - linux-2.6/blob - 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 (count >= nr_huge_pages)
 407                                 return;
 408                         if (PageHighMem(page))
 409                                 continue;
 410                         list_del(&page->lru);
 411                         update_and_free_page(page);
 412                         free_huge_pages--;
 413                         free_huge_pages_node[page_to_nid(page)]--;
 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
 454         /*
 455          * Decrease the pool size
 456          * First return free pages to the buddy allocator (being careful
 457          * to keep enough around to satisfy reservations).  Then place
 458          * pages into surplus state as needed so the pool will shrink
 459          * to the desired size as pages become free.
 460          */
 461         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
 462         min_count = max(count, min_count);
 463         try_to_free_low(min_count);
 464         while (min_count < persistent_huge_pages) {
 465                 struct page *page = dequeue_huge_page(NULL, 0);
 466                 if (!page)
 467                         break;
 468                 update_and_free_page(page);
 469         }
 470         while (count < persistent_huge_pages) {
 471                 if (!adjust_pool_surplus(1))
 472                         break;
 473         }
 474 out:
 475         ret = persistent_huge_pages;
 476         spin_unlock(&hugetlb_lock);
 477         return ret;
 478 }
 479
 480 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 481                            struct file *file, void __user *buffer,
 482                            size_t *length, loff_t *ppos)
 483 {
 484         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 485         max_huge_pages = set_max_huge_pages(max_huge_pages);
 486         return 0;
 487 }
 488
 489 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 490                         struct file *file, void __user *buffer,
 491                         size_t *length, loff_t *ppos)
 492 {
 493         proc_dointvec(table, write, file, buffer, length, ppos);
 494         if (hugepages_treat_as_movable)
 495                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 496         else
 497                 htlb_alloc_mask = GFP_HIGHUSER;
 498         return 0;
 499 }
 500
 501 #endif /* CONFIG_SYSCTL */
 502
 503 int hugetlb_report_meminfo(char *buf)
 504 {
 505         return sprintf(buf,
 506                         "HugePages_Total: %5lu\n"
 507                         "HugePages_Free:  %5lu\n"
 508                         "HugePages_Rsvd:  %5lu\n"
 509                         "HugePages_Surp:  %5lu\n"
 510                         "Hugepagesize:    %5lu kB\n",
 511                         nr_huge_pages,
 512                         free_huge_pages,
 513                         resv_huge_pages,
 514                         surplus_huge_pages,
 515                         HPAGE_SIZE/1024);
 516 }
 517
 518 int hugetlb_report_node_meminfo(int nid, char *buf)
 519 {
 520         return sprintf(buf,
 521                 "Node %d HugePages_Total: %5u\n"
 522                 "Node %d HugePages_Free:  %5u\n",
 523                 nid, nr_huge_pages_node[nid],
 524                 nid, free_huge_pages_node[nid]);
 525 }
 526
 527 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 528 unsigned long hugetlb_total_pages(void)
 529 {
 530         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 531 }
 532
 533 /*
 534  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 535  * handle_mm_fault() to try to instantiate regular-sized pages in the
 536  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 537  * this far.
 538  */
 539 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 540 {
 541         BUG();
 542         return 0;
 543 }
 544
 545 struct vm_operations_struct hugetlb_vm_ops = {
 546         .fault = hugetlb_vm_op_fault,
 547 };
 548
 549 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 550                                 int writable)
 551 {
 552         pte_t entry;
 553
 554         if (writable) {
 555                 entry =
 556                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 557         } else {
 558                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 559         }
 560         entry = pte_mkyoung(entry);
 561         entry = pte_mkhuge(entry);
 562
 563         return entry;
 564 }
 565
 566 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 567                                    unsigned long address, pte_t *ptep)
 568 {
 569         pte_t entry;
 570
 571         entry = pte_mkwrite(pte_mkdirty(*ptep));
 572         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 573                 update_mmu_cache(vma, address, entry);
 574         }
 575 }
 576
 577
 578 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 579                             struct vm_area_struct *vma)
 580 {
 581         pte_t *src_pte, *dst_pte, entry;
 582         struct page *ptepage;
 583         unsigned long addr;
 584         int cow;
 585
 586         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 587
 588         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 589                 src_pte = huge_pte_offset(src, addr);
 590                 if (!src_pte)
 591                         continue;
 592                 dst_pte = huge_pte_alloc(dst, addr);
 593                 if (!dst_pte)
 594                         goto nomem;
 595                 spin_lock(&dst->page_table_lock);
 596                 spin_lock(&src->page_table_lock);
 597                 if (!pte_none(*src_pte)) {
 598                         if (cow)
 599                                 ptep_set_wrprotect(src, addr, src_pte);
 600                         entry = *src_pte;
 601                         ptepage = pte_page(entry);
 602                         get_page(ptepage);
 603                         set_huge_pte_at(dst, addr, dst_pte, entry);
 604                 }
 605                 spin_unlock(&src->page_table_lock);
 606                 spin_unlock(&dst->page_table_lock);
 607         }
 608         return 0;
 609
 610 nomem:
 611         return -ENOMEM;
 612 }
 613
 614 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 615                             unsigned long end)
 616 {
 617         struct mm_struct *mm = vma->vm_mm;
 618         unsigned long address;
 619         pte_t *ptep;
 620         pte_t pte;
 621         struct page *page;
 622         struct page *tmp;
 623         /*
 624          * A page gathering list, protected by per file i_mmap_lock. The
 625          * lock is used to avoid list corruption from multiple unmapping
 626          * of the same page since we are using page->lru.
 627          */
 628         LIST_HEAD(page_list);
 629
 630         WARN_ON(!is_vm_hugetlb_page(vma));
 631         BUG_ON(start & ~HPAGE_MASK);
 632         BUG_ON(end & ~HPAGE_MASK);
 633
 634         spin_lock(&mm->page_table_lock);
 635         for (address = start; address < end; address += HPAGE_SIZE) {
 636                 ptep = huge_pte_offset(mm, address);
 637                 if (!ptep)
 638                         continue;
 639
 640                 if (huge_pmd_unshare(mm, &address, ptep))
 641                         continue;
 642
 643                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 644                 if (pte_none(pte))
 645                         continue;
 646
 647                 page = pte_page(pte);
 648                 if (pte_dirty(pte))
 649                         set_page_dirty(page);
 650                 list_add(&page->lru, &page_list);
 651         }
 652         spin_unlock(&mm->page_table_lock);
 653         flush_tlb_range(vma, start, end);
 654         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 655                 list_del(&page->lru);
 656                 put_page(page);
 657         }
 658 }
 659
 660 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 661                           unsigned long end)
 662 {
 663         /*
 664          * It is undesirable to test vma->vm_file as it should be non-null
 665          * for valid hugetlb area. However, vm_file will be NULL in the error
 666          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 667          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 668          * to clean up. Since no pte has actually been setup, it is safe to
 669          * do nothing in this case.
 670          */
 671         if (vma->vm_file) {
 672                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 673                 __unmap_hugepage_range(vma, start, end);
 674                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 675         }
 676 }
 677
 678 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 679                         unsigned long address, pte_t *ptep, pte_t pte)
 680 {
 681         struct page *old_page, *new_page;
 682         int avoidcopy;
 683
 684         old_page = pte_page(pte);
 685
 686         /* If no-one else is actually using this page, avoid the copy
 687          * and just make the page writable */
 688         avoidcopy = (page_count(old_page) == 1);
 689         if (avoidcopy) {
 690                 set_huge_ptep_writable(vma, address, ptep);
 691                 return 0;
 692         }
 693
 694         page_cache_get(old_page);
 695         new_page = alloc_huge_page(vma, address);
 696
 697         if (!new_page) {
 698                 page_cache_release(old_page);
 699                 return VM_FAULT_OOM;
 700         }
 701
 702         spin_unlock(&mm->page_table_lock);
 703         copy_huge_page(new_page, old_page, address, vma);
 704         spin_lock(&mm->page_table_lock);
 705
 706         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 707         if (likely(pte_same(*ptep, pte))) {
 708                 /* Break COW */
 709                 set_huge_pte_at(mm, address, ptep,
 710                                 make_huge_pte(vma, new_page, 1));
 711                 /* Make the old page be freed below */
 712                 new_page = old_page;
 713         }
 714         page_cache_release(new_page);
 715         page_cache_release(old_page);
 716         return 0;
 717 }
 718
 719 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 720                         unsigned long address, pte_t *ptep, int write_access)
 721 {
 722         int ret = VM_FAULT_SIGBUS;
 723         unsigned long idx;
 724         unsigned long size;
 725         struct page *page;
 726         struct address_space *mapping;
 727         pte_t new_pte;
 728
 729         mapping = vma->vm_file->f_mapping;
 730         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 731                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 732
 733         /*
 734          * Use page lock to guard against racing truncation
 735          * before we get page_table_lock.
 736          */
 737 retry:
 738         page = find_lock_page(mapping, idx);
 739         if (!page) {
 740                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 741                 if (idx >= size)
 742                         goto out;
 743                 if (hugetlb_get_quota(mapping))
 744                         goto out;
 745                 page = alloc_huge_page(vma, address);
 746                 if (!page) {
 747                         hugetlb_put_quota(mapping);
 748                         ret = VM_FAULT_OOM;
 749                         goto out;
 750                 }
 751                 clear_huge_page(page, address);
 752
 753                 if (vma->vm_flags & VM_SHARED) {
 754                         int err;
 755
 756                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 757                         if (err) {
 758                                 put_page(page);
 759                                 hugetlb_put_quota(mapping);
 760                                 if (err == -EEXIST)
 761                                         goto retry;
 762                                 goto out;
 763                         }
 764                 } else
 765                         lock_page(page);
 766         }
 767
 768         spin_lock(&mm->page_table_lock);
 769         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 770         if (idx >= size)
 771                 goto backout;
 772
 773         ret = 0;
 774         if (!pte_none(*ptep))
 775                 goto backout;
 776
 777         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 778                                 && (vma->vm_flags & VM_SHARED)));
 779         set_huge_pte_at(mm, address, ptep, new_pte);
 780
 781         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 782                 /* Optimization, do the COW without a second fault */
 783                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 784         }
 785
 786         spin_unlock(&mm->page_table_lock);
 787         unlock_page(page);
 788 out:
 789         return ret;
 790
 791 backout:
 792         spin_unlock(&mm->page_table_lock);
 793         hugetlb_put_quota(mapping);
 794         unlock_page(page);
 795         put_page(page);
 796         goto out;
 797 }
 798
 799 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 800                         unsigned long address, int write_access)
 801 {
 802         pte_t *ptep;
 803         pte_t entry;
 804         int ret;
 805         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 806
 807         ptep = huge_pte_alloc(mm, address);
 808         if (!ptep)
 809                 return VM_FAULT_OOM;
 810
 811         /*
 812          * Serialize hugepage allocation and instantiation, so that we don't
 813          * get spurious allocation failures if two CPUs race to instantiate
 814          * the same page in the page cache.
 815          */
 816         mutex_lock(&hugetlb_instantiation_mutex);
 817         entry = *ptep;
 818         if (pte_none(entry)) {
 819                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 820                 mutex_unlock(&hugetlb_instantiation_mutex);
 821                 return ret;
 822         }
 823
 824         ret = 0;
 825
 826         spin_lock(&mm->page_table_lock);
 827         /* Check for a racing update before calling hugetlb_cow */
 828         if (likely(pte_same(entry, *ptep)))
 829                 if (write_access && !pte_write(entry))
 830                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 831         spin_unlock(&mm->page_table_lock);
 832         mutex_unlock(&hugetlb_instantiation_mutex);
 833
 834         return ret;
 835 }
 836
 837 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 838                         struct page **pages, struct vm_area_struct **vmas,
 839                         unsigned long *position, int *length, int i)
 840 {
 841         unsigned long pfn_offset;
 842         unsigned long vaddr = *position;
 843         int remainder = *length;
 844
 845         spin_lock(&mm->page_table_lock);
 846         while (vaddr < vma->vm_end && remainder) {
 847                 pte_t *pte;
 848                 struct page *page;
 849
 850                 /*
 851                  * Some archs (sparc64, sh*) have multiple pte_ts to
 852                  * each hugepage.  We have to make * sure we get the
 853                  * first, for the page indexing below to work.
 854                  */
 855                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 856
 857                 if (!pte || pte_none(*pte)) {
 858                         int ret;
 859
 860                         spin_unlock(&mm->page_table_lock);
 861                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 862                         spin_lock(&mm->page_table_lock);
 863                         if (!(ret & VM_FAULT_ERROR))
 864                                 continue;
 865
 866                         remainder = 0;
 867                         if (!i)
 868                                 i = -EFAULT;
 869                         break;
 870                 }
 871
 872                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 873                 page = pte_page(*pte);
 874 same_page:
 875                 if (pages) {
 876                         get_page(page);
 877                         pages[i] = page + pfn_offset;
 878                 }
 879
 880                 if (vmas)
 881                         vmas[i] = vma;
 882
 883                 vaddr += PAGE_SIZE;
 884                 ++pfn_offset;
 885                 --remainder;
 886                 ++i;
 887                 if (vaddr < vma->vm_end && remainder &&
 888                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 889                         /*
 890                          * We use pfn_offset to avoid touching the pageframes
 891                          * of this compound page.
 892                          */
 893                         goto same_page;
 894                 }
 895         }
 896         spin_unlock(&mm->page_table_lock);
 897         *length = remainder;
 898         *position = vaddr;
 899
 900         return i;
 901 }
 902
 903 void hugetlb_change_protection(struct vm_area_struct *vma,
 904                 unsigned long address, unsigned long end, pgprot_t newprot)
 905 {
 906         struct mm_struct *mm = vma->vm_mm;
 907         unsigned long start = address;
 908         pte_t *ptep;
 909         pte_t pte;
 910
 911         BUG_ON(address >= end);
 912         flush_cache_range(vma, address, end);
 913
 914         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 915         spin_lock(&mm->page_table_lock);
 916         for (; address < end; address += HPAGE_SIZE) {
 917                 ptep = huge_pte_offset(mm, address);
 918                 if (!ptep)
 919                         continue;
 920                 if (huge_pmd_unshare(mm, &address, ptep))
 921                         continue;
 922                 if (!pte_none(*ptep)) {
 923                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 924                         pte = pte_mkhuge(pte_modify(pte, newprot));
 925                         set_huge_pte_at(mm, address, ptep, pte);
 926                 }
 927         }
 928         spin_unlock(&mm->page_table_lock);
 929         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 930
 931         flush_tlb_range(vma, start, end);
 932 }
 933
 934 struct file_region {
 935         struct list_head link;
 936         long from;
 937         long to;
 938 };
 939
 940 static long region_add(struct list_head *head, long f, long t)
 941 {
 942         struct file_region *rg, *nrg, *trg;
 943
 944         /* Locate the region we are either in or before. */
 945         list_for_each_entry(rg, head, link)
 946                 if (f <= rg->to)
 947                         break;
 948
 949         /* Round our left edge to the current segment if it encloses us. */
 950         if (f > rg->from)
 951                 f = rg->from;
 952
 953         /* Check for and consume any regions we now overlap with. */
 954         nrg = rg;
 955         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 956                 if (&rg->link == head)
 957                         break;
 958                 if (rg->from > t)
 959                         break;
 960
 961                 /* If this area reaches higher then extend our area to
 962                  * include it completely.  If this is not the first area
 963                  * which we intend to reuse, free it. */
 964                 if (rg->to > t)
 965                         t = rg->to;
 966                 if (rg != nrg) {
 967                         list_del(&rg->link);
 968                         kfree(rg);
 969                 }
 970         }
 971         nrg->from = f;
 972         nrg->to = t;
 973         return 0;
 974 }
 975
 976 static long region_chg(struct list_head *head, long f, long t)
 977 {
 978         struct file_region *rg, *nrg;
 979         long chg = 0;
 980
 981         /* Locate the region we are before or in. */
 982         list_for_each_entry(rg, head, link)
 983                 if (f <= rg->to)
 984                         break;
 985
 986         /* If we are below the current region then a new region is required.
 987          * Subtle, allocate a new region at the position but make it zero
 988          * size such that we can guarentee to record the reservation. */
 989         if (&rg->link == head || t < rg->from) {
 990                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 991                 if (nrg == 0)
 992                         return -ENOMEM;
 993                 nrg->from = f;
 994                 nrg->to   = f;
 995                 INIT_LIST_HEAD(&nrg->link);
 996                 list_add(&nrg->link, rg->link.prev);
 997
 998                 return t - f;
 999         }
1000
1001         /* Round our left edge to the current segment if it encloses us. */
1002         if (f > rg->from)
1003                 f = rg->from;
1004         chg = t - f;
1005
1006         /* Check for and consume any regions we now overlap with. */
1007         list_for_each_entry(rg, rg->link.prev, link) {
1008                 if (&rg->link == head)
1009                         break;
1010                 if (rg->from > t)
1011                         return chg;
1012
1013                 /* We overlap with this area, if it extends futher than
1014                  * us then we must extend ourselves.  Account for its
1015                  * existing reservation. */
1016                 if (rg->to > t) {
1017                         chg += rg->to - t;
1018                         t = rg->to;
1019                 }
1020                 chg -= rg->to - rg->from;
1021         }
1022         return chg;
1023 }
1024
1025 static long region_truncate(struct list_head *head, long end)
1026 {
1027         struct file_region *rg, *trg;
1028         long chg = 0;
1029
1030         /* Locate the region we are either in or before. */
1031         list_for_each_entry(rg, head, link)
1032                 if (end <= rg->to)
1033                         break;
1034         if (&rg->link == head)
1035                 return 0;
1036
1037         /* If we are in the middle of a region then adjust it. */
1038         if (end > rg->from) {
1039                 chg = rg->to - end;
1040                 rg->to = end;
1041                 rg = list_entry(rg->link.next, typeof(*rg), link);
1042         }
1043
1044         /* Drop any remaining regions. */
1045         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1046                 if (&rg->link == head)
1047                         break;
1048                 chg += rg->to - rg->from;
1049                 list_del(&rg->link);
1050                 kfree(rg);
1051         }
1052         return chg;
1053 }
1054
1055 static int hugetlb_acct_memory(long delta)
1056 {
1057         int ret = -ENOMEM;
1058
1059         spin_lock(&hugetlb_lock);
1060         /*
1061          * When cpuset is configured, it breaks the strict hugetlb page
1062          * reservation as the accounting is done on a global variable. Such
1063          * reservation is completely rubbish in the presence of cpuset because
1064          * the reservation is not checked against page availability for the
1065          * current cpuset. Application can still potentially OOM'ed by kernel
1066          * with lack of free htlb page in cpuset that the task is in.
1067          * Attempt to enforce strict accounting with cpuset is almost
1068          * impossible (or too ugly) because cpuset is too fluid that
1069          * task or memory node can be dynamically moved between cpusets.
1070          *
1071          * The change of semantics for shared hugetlb mapping with cpuset is
1072          * undesirable. However, in order to preserve some of the semantics,
1073          * we fall back to check against current free page availability as
1074          * a best attempt and hopefully to minimize the impact of changing
1075          * semantics that cpuset has.
1076          */
1077         if (delta > 0) {
1078                 if (gather_surplus_pages(delta) < 0)
1079                         goto out;
1080
1081                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1082                         goto out;
1083         }
1084
1085         ret = 0;
1086         resv_huge_pages += delta;
1087         if (delta < 0)
1088                 return_unused_surplus_pages((unsigned long) -delta);
1089
1090 out:
1091         spin_unlock(&hugetlb_lock);
1092         return ret;
1093 }
1094
1095 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1096 {
1097         long ret, chg;
1098
1099         chg = region_chg(&inode->i_mapping->private_list, from, to);
1100         if (chg < 0)
1101                 return chg;
1102
1103         ret = hugetlb_acct_memory(chg);
1104         if (ret < 0)
1105                 return ret;
1106         region_add(&inode->i_mapping->private_list, from, to);
1107         return 0;
1108 }
1109
1110 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1111 {
1112         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1113         hugetlb_acct_memory(freed - chg);
1114 }