SAFE public projects git trees. - safe/jmp/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 static int hugetlb_next_nid;
  36
  37 /*
  38  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  39  */
  40 static DEFINE_SPINLOCK(hugetlb_lock);
  41
  42 static void clear_huge_page(struct page *page, unsigned long addr)
  43 {
  44         int i;
  45
  46         might_sleep();
  47         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  48                 cond_resched();
  49                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
  50         }
  51 }
  52
  53 static void copy_huge_page(struct page *dst, struct page *src,
  54                            unsigned long addr, struct vm_area_struct *vma)
  55 {
  56         int i;
  57
  58         might_sleep();
  59         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  60                 cond_resched();
  61                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  62         }
  63 }
  64
  65 static void enqueue_huge_page(struct page *page)
  66 {
  67         int nid = page_to_nid(page);
  68         list_add(&page->lru, &hugepage_freelists[nid]);
  69         free_huge_pages++;
  70         free_huge_pages_node[nid]++;
  71 }
  72
  73 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  74                                 unsigned long address)
  75 {
  76         int nid;
  77         struct page *page = NULL;
  78         struct mempolicy *mpol;
  79         struct zonelist *zonelist = huge_zonelist(vma, address,
  80                                         htlb_alloc_mask, &mpol);
  81         struct zone **z;
  82
  83         for (z = zonelist->zones; *z; z++) {
  84                 nid = zone_to_nid(*z);
  85                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
  86                     !list_empty(&hugepage_freelists[nid])) {
  87                         page = list_entry(hugepage_freelists[nid].next,
  88                                           struct page, lru);
  89                         list_del(&page->lru);
  90                         free_huge_pages--;
  91                         free_huge_pages_node[nid]--;
  92                         if (vma && vma->vm_flags & VM_MAYSHARE)
  93                                 resv_huge_pages--;
  94                         break;
  95                 }
  96         }
  97         mpol_free(mpol);        /* unref if mpol !NULL */
  98         return page;
  99 }
 100
 101 static void update_and_free_page(struct page *page)
 102 {
 103         int i;
 104         nr_huge_pages--;
 105         nr_huge_pages_node[page_to_nid(page)]--;
 106         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 107                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 108                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 109                                 1 << PG_private | 1<< PG_writeback);
 110         }
 111         set_compound_page_dtor(page, NULL);
 112         set_page_refcounted(page);
 113         __free_pages(page, HUGETLB_PAGE_ORDER);
 114 }
 115
 116 static void free_huge_page(struct page *page)
 117 {
 118         int nid = page_to_nid(page);
 119
 120         BUG_ON(page_count(page));
 121         INIT_LIST_HEAD(&page->lru);
 122
 123         spin_lock(&hugetlb_lock);
 124         if (surplus_huge_pages_node[nid]) {
 125                 update_and_free_page(page);
 126                 surplus_huge_pages--;
 127                 surplus_huge_pages_node[nid]--;
 128         } else {
 129                 enqueue_huge_page(page);
 130         }
 131         spin_unlock(&hugetlb_lock);
 132 }
 133
 134 /*
 135  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 136  * balanced by operating on them in a round-robin fashion.
 137  * Returns 1 if an adjustment was made.
 138  */
 139 static int adjust_pool_surplus(int delta)
 140 {
 141         static int prev_nid;
 142         int nid = prev_nid;
 143         int ret = 0;
 144
 145         VM_BUG_ON(delta != -1 && delta != 1);
 146         do {
 147                 nid = next_node(nid, node_online_map);
 148                 if (nid == MAX_NUMNODES)
 149                         nid = first_node(node_online_map);
 150
 151                 /* To shrink on this node, there must be a surplus page */
 152                 if (delta < 0 && !surplus_huge_pages_node[nid])
 153                         continue;
 154                 /* Surplus cannot exceed the total number of pages */
 155                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 156                                                 nr_huge_pages_node[nid])
 157                         continue;
 158
 159                 surplus_huge_pages += delta;
 160                 surplus_huge_pages_node[nid] += delta;
 161                 ret = 1;
 162                 break;
 163         } while (nid != prev_nid);
 164
 165         prev_nid = nid;
 166         return ret;
 167 }
 168
 169 static struct page *alloc_fresh_huge_page_node(int nid)
 170 {
 171         struct page *page;
 172
 173         page = alloc_pages_node(nid,
 174                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
 175                 HUGETLB_PAGE_ORDER);
 176         if (page) {
 177                 set_compound_page_dtor(page, free_huge_page);
 178                 spin_lock(&hugetlb_lock);
 179                 nr_huge_pages++;
 180                 nr_huge_pages_node[nid]++;
 181                 spin_unlock(&hugetlb_lock);
 182                 put_page(page); /* free it into the hugepage allocator */
 183         }
 184
 185         return page;
 186 }
 187
 188 static int alloc_fresh_huge_page(void)
 189 {
 190         struct page *page;
 191         int start_nid;
 192         int next_nid;
 193         int ret = 0;
 194
 195         start_nid = hugetlb_next_nid;
 196
 197         do {
 198                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
 199                 if (page)
 200                         ret = 1;
 201                 /*
 202                  * Use a helper variable to find the next node and then
 203                  * copy it back to hugetlb_next_nid afterwards:
 204                  * otherwise there's a window in which a racer might
 205                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 206                  * But we don't need to use a spin_lock here: it really
 207                  * doesn't matter if occasionally a racer chooses the
 208                  * same nid as we do.  Move nid forward in the mask even
 209                  * if we just successfully allocated a hugepage so that
 210                  * the next caller gets hugepages on the next node.
 211                  */
 212                 next_nid = next_node(hugetlb_next_nid, node_online_map);
 213                 if (next_nid == MAX_NUMNODES)
 214                         next_nid = first_node(node_online_map);
 215                 hugetlb_next_nid = next_nid;
 216         } while (!page && hugetlb_next_nid != start_nid);
 217
 218         return ret;
 219 }
 220
 221 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 222                                                 unsigned long address)
 223 {
 224         struct page *page;
 225
 226         /* Check if the dynamic pool is enabled */
 227         if (!hugetlb_dynamic_pool)
 228                 return NULL;
 229
 230         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 231                                         HUGETLB_PAGE_ORDER);
 232         if (page) {
 233                 set_compound_page_dtor(page, free_huge_page);
 234                 spin_lock(&hugetlb_lock);
 235                 nr_huge_pages++;
 236                 nr_huge_pages_node[page_to_nid(page)]++;
 237                 surplus_huge_pages++;
 238                 surplus_huge_pages_node[page_to_nid(page)]++;
 239                 spin_unlock(&hugetlb_lock);
 240         }
 241
 242         return page;
 243 }
 244
 245 /*
 246  * Increase the hugetlb pool such that it can accomodate a reservation
 247  * of size 'delta'.
 248  */
 249 static int gather_surplus_pages(int delta)
 250 {
 251         struct list_head surplus_list;
 252         struct page *page, *tmp;
 253         int ret, i;
 254         int needed, allocated;
 255
 256         needed = (resv_huge_pages + delta) - free_huge_pages;
 257         if (needed <= 0)
 258                 return 0;
 259
 260         allocated = 0;
 261         INIT_LIST_HEAD(&surplus_list);
 262
 263         ret = -ENOMEM;
 264 retry:
 265         spin_unlock(&hugetlb_lock);
 266         for (i = 0; i < needed; i++) {
 267                 page = alloc_buddy_huge_page(NULL, 0);
 268                 if (!page) {
 269                         /*
 270                          * We were not able to allocate enough pages to
 271                          * satisfy the entire reservation so we free what
 272                          * we've allocated so far.
 273                          */
 274                         spin_lock(&hugetlb_lock);
 275                         needed = 0;
 276                         goto free;
 277                 }
 278
 279                 list_add(&page->lru, &surplus_list);
 280         }
 281         allocated += needed;
 282
 283         /*
 284          * After retaking hugetlb_lock, we need to recalculate 'needed'
 285          * because either resv_huge_pages or free_huge_pages may have changed.
 286          */
 287         spin_lock(&hugetlb_lock);
 288         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
 289         if (needed > 0)
 290                 goto retry;
 291
 292         /*
 293          * The surplus_list now contains _at_least_ the number of extra pages
 294          * needed to accomodate the reservation.  Add the appropriate number
 295          * of pages to the hugetlb pool and free the extras back to the buddy
 296          * allocator.
 297          */
 298         needed += allocated;
 299         ret = 0;
 300 free:
 301         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 302                 list_del(&page->lru);
 303                 if ((--needed) >= 0)
 304                         enqueue_huge_page(page);
 305                 else
 306                         update_and_free_page(page);
 307         }
 308
 309         return ret;
 310 }
 311
 312 /*
 313  * When releasing a hugetlb pool reservation, any surplus pages that were
 314  * allocated to satisfy the reservation must be explicitly freed if they were
 315  * never used.
 316  */
 317 void return_unused_surplus_pages(unsigned long unused_resv_pages)
 318 {
 319         static int nid = -1;
 320         struct page *page;
 321         unsigned long nr_pages;
 322
 323         nr_pages = min(unused_resv_pages, surplus_huge_pages);
 324
 325         while (nr_pages) {
 326                 nid = next_node(nid, node_online_map);
 327                 if (nid == MAX_NUMNODES)
 328                         nid = first_node(node_online_map);
 329
 330                 if (!surplus_huge_pages_node[nid])
 331                         continue;
 332
 333                 if (!list_empty(&hugepage_freelists[nid])) {
 334                         page = list_entry(hugepage_freelists[nid].next,
 335                                           struct page, lru);
 336                         list_del(&page->lru);
 337                         update_and_free_page(page);
 338                         free_huge_pages--;
 339                         free_huge_pages_node[nid]--;
 340                         surplus_huge_pages--;
 341                         surplus_huge_pages_node[nid]--;
 342                         nr_pages--;
 343                 }
 344         }
 345 }
 346
 347 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 348                                     unsigned long addr)
 349 {
 350         struct page *page = NULL;
 351         int use_reserved_page = vma->vm_flags & VM_MAYSHARE;
 352
 353         spin_lock(&hugetlb_lock);
 354         if (!use_reserved_page && (free_huge_pages <= resv_huge_pages))
 355                 goto fail;
 356
 357         page = dequeue_huge_page(vma, addr);
 358         if (!page)
 359                 goto fail;
 360
 361         spin_unlock(&hugetlb_lock);
 362         set_page_refcounted(page);
 363         return page;
 364
 365 fail:
 366         spin_unlock(&hugetlb_lock);
 367
 368         /*
 369          * Private mappings do not use reserved huge pages so the allocation
 370          * may have failed due to an undersized hugetlb pool.  Try to grab a
 371          * surplus huge page from the buddy allocator.
 372          */
 373         if (!use_reserved_page)
 374                 page = alloc_buddy_huge_page(vma, addr);
 375
 376         return page;
 377 }
 378
 379 static int __init hugetlb_init(void)
 380 {
 381         unsigned long i;
 382
 383         if (HPAGE_SHIFT == 0)
 384                 return 0;
 385
 386         for (i = 0; i < MAX_NUMNODES; ++i)
 387                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 388
 389         hugetlb_next_nid = first_node(node_online_map);
 390
 391         for (i = 0; i < max_huge_pages; ++i) {
 392                 if (!alloc_fresh_huge_page())
 393                         break;
 394         }
 395         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 396         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 397         return 0;
 398 }
 399 module_init(hugetlb_init);
 400
 401 static int __init hugetlb_setup(char *s)
 402 {
 403         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 404                 max_huge_pages = 0;
 405         return 1;
 406 }
 407 __setup("hugepages=", hugetlb_setup);
 408
 409 static unsigned int cpuset_mems_nr(unsigned int *array)
 410 {
 411         int node;
 412         unsigned int nr = 0;
 413
 414         for_each_node_mask(node, cpuset_current_mems_allowed)
 415                 nr += array[node];
 416
 417         return nr;
 418 }
 419
 420 #ifdef CONFIG_SYSCTL
 421 #ifdef CONFIG_HIGHMEM
 422 static void try_to_free_low(unsigned long count)
 423 {
 424         int i;
 425
 426         for (i = 0; i < MAX_NUMNODES; ++i) {
 427                 struct page *page, *next;
 428                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 429                         if (count >= nr_huge_pages)
 430                                 return;
 431                         if (PageHighMem(page))
 432                                 continue;
 433                         list_del(&page->lru);
 434                         update_and_free_page(page);
 435                         free_huge_pages--;
 436                         free_huge_pages_node[page_to_nid(page)]--;
 437                 }
 438         }
 439 }
 440 #else
 441 static inline void try_to_free_low(unsigned long count)
 442 {
 443 }
 444 #endif
 445
 446 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 447 static unsigned long set_max_huge_pages(unsigned long count)
 448 {
 449         unsigned long min_count, ret;
 450
 451         /*
 452          * Increase the pool size
 453          * First take pages out of surplus state.  Then make up the
 454          * remaining difference by allocating fresh huge pages.
 455          */
 456         spin_lock(&hugetlb_lock);
 457         while (surplus_huge_pages && count > persistent_huge_pages) {
 458                 if (!adjust_pool_surplus(-1))
 459                         break;
 460         }
 461
 462         while (count > persistent_huge_pages) {
 463                 int ret;
 464                 /*
 465                  * If this allocation races such that we no longer need the
 466                  * page, free_huge_page will handle it by freeing the page
 467                  * and reducing the surplus.
 468                  */
 469                 spin_unlock(&hugetlb_lock);
 470                 ret = alloc_fresh_huge_page();
 471                 spin_lock(&hugetlb_lock);
 472                 if (!ret)
 473                         goto out;
 474
 475         }
 476
 477         /*
 478          * Decrease the pool size
 479          * First return free pages to the buddy allocator (being careful
 480          * to keep enough around to satisfy reservations).  Then place
 481          * pages into surplus state as needed so the pool will shrink
 482          * to the desired size as pages become free.
 483          */
 484         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
 485         min_count = max(count, min_count);
 486         try_to_free_low(min_count);
 487         while (min_count < persistent_huge_pages) {
 488                 struct page *page = dequeue_huge_page(NULL, 0);
 489                 if (!page)
 490                         break;
 491                 update_and_free_page(page);
 492         }
 493         while (count < persistent_huge_pages) {
 494                 if (!adjust_pool_surplus(1))
 495                         break;
 496         }
 497 out:
 498         ret = persistent_huge_pages;
 499         spin_unlock(&hugetlb_lock);
 500         return ret;
 501 }
 502
 503 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 504                            struct file *file, void __user *buffer,
 505                            size_t *length, loff_t *ppos)
 506 {
 507         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 508         max_huge_pages = set_max_huge_pages(max_huge_pages);
 509         return 0;
 510 }
 511
 512 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 513                         struct file *file, void __user *buffer,
 514                         size_t *length, loff_t *ppos)
 515 {
 516         proc_dointvec(table, write, file, buffer, length, ppos);
 517         if (hugepages_treat_as_movable)
 518                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 519         else
 520                 htlb_alloc_mask = GFP_HIGHUSER;
 521         return 0;
 522 }
 523
 524 #endif /* CONFIG_SYSCTL */
 525
 526 int hugetlb_report_meminfo(char *buf)
 527 {
 528         return sprintf(buf,
 529                         "HugePages_Total: %5lu\n"
 530                         "HugePages_Free:  %5lu\n"
 531                         "HugePages_Rsvd:  %5lu\n"
 532                         "HugePages_Surp:  %5lu\n"
 533                         "Hugepagesize:    %5lu kB\n",
 534                         nr_huge_pages,
 535                         free_huge_pages,
 536                         resv_huge_pages,
 537                         surplus_huge_pages,
 538                         HPAGE_SIZE/1024);
 539 }
 540
 541 int hugetlb_report_node_meminfo(int nid, char *buf)
 542 {
 543         return sprintf(buf,
 544                 "Node %d HugePages_Total: %5u\n"
 545                 "Node %d HugePages_Free:  %5u\n",
 546                 nid, nr_huge_pages_node[nid],
 547                 nid, free_huge_pages_node[nid]);
 548 }
 549
 550 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 551 unsigned long hugetlb_total_pages(void)
 552 {
 553         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 554 }
 555
 556 /*
 557  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 558  * handle_mm_fault() to try to instantiate regular-sized pages in the
 559  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 560  * this far.
 561  */
 562 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 563 {
 564         BUG();
 565         return 0;
 566 }
 567
 568 struct vm_operations_struct hugetlb_vm_ops = {
 569         .fault = hugetlb_vm_op_fault,
 570 };
 571
 572 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 573                                 int writable)
 574 {
 575         pte_t entry;
 576
 577         if (writable) {
 578                 entry =
 579                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 580         } else {
 581                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 582         }
 583         entry = pte_mkyoung(entry);
 584         entry = pte_mkhuge(entry);
 585
 586         return entry;
 587 }
 588
 589 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 590                                    unsigned long address, pte_t *ptep)
 591 {
 592         pte_t entry;
 593
 594         entry = pte_mkwrite(pte_mkdirty(*ptep));
 595         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 596                 update_mmu_cache(vma, address, entry);
 597         }
 598 }
 599
 600
 601 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 602                             struct vm_area_struct *vma)
 603 {
 604         pte_t *src_pte, *dst_pte, entry;
 605         struct page *ptepage;
 606         unsigned long addr;
 607         int cow;
 608
 609         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 610
 611         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 612                 src_pte = huge_pte_offset(src, addr);
 613                 if (!src_pte)
 614                         continue;
 615                 dst_pte = huge_pte_alloc(dst, addr);
 616                 if (!dst_pte)
 617                         goto nomem;
 618                 spin_lock(&dst->page_table_lock);
 619                 spin_lock(&src->page_table_lock);
 620                 if (!pte_none(*src_pte)) {
 621                         if (cow)
 622                                 ptep_set_wrprotect(src, addr, src_pte);
 623                         entry = *src_pte;
 624                         ptepage = pte_page(entry);
 625                         get_page(ptepage);
 626                         set_huge_pte_at(dst, addr, dst_pte, entry);
 627                 }
 628                 spin_unlock(&src->page_table_lock);
 629                 spin_unlock(&dst->page_table_lock);
 630         }
 631         return 0;
 632
 633 nomem:
 634         return -ENOMEM;
 635 }
 636
 637 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 638                             unsigned long end)
 639 {
 640         struct mm_struct *mm = vma->vm_mm;
 641         unsigned long address;
 642         pte_t *ptep;
 643         pte_t pte;
 644         struct page *page;
 645         struct page *tmp;
 646         /*
 647          * A page gathering list, protected by per file i_mmap_lock. The
 648          * lock is used to avoid list corruption from multiple unmapping
 649          * of the same page since we are using page->lru.
 650          */
 651         LIST_HEAD(page_list);
 652
 653         WARN_ON(!is_vm_hugetlb_page(vma));
 654         BUG_ON(start & ~HPAGE_MASK);
 655         BUG_ON(end & ~HPAGE_MASK);
 656
 657         spin_lock(&mm->page_table_lock);
 658         for (address = start; address < end; address += HPAGE_SIZE) {
 659                 ptep = huge_pte_offset(mm, address);
 660                 if (!ptep)
 661                         continue;
 662
 663                 if (huge_pmd_unshare(mm, &address, ptep))
 664                         continue;
 665
 666                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 667                 if (pte_none(pte))
 668                         continue;
 669
 670                 page = pte_page(pte);
 671                 if (pte_dirty(pte))
 672                         set_page_dirty(page);
 673                 list_add(&page->lru, &page_list);
 674         }
 675         spin_unlock(&mm->page_table_lock);
 676         flush_tlb_range(vma, start, end);
 677         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 678                 list_del(&page->lru);
 679                 put_page(page);
 680         }
 681 }
 682
 683 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 684                           unsigned long end)
 685 {
 686         /*
 687          * It is undesirable to test vma->vm_file as it should be non-null
 688          * for valid hugetlb area. However, vm_file will be NULL in the error
 689          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 690          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 691          * to clean up. Since no pte has actually been setup, it is safe to
 692          * do nothing in this case.
 693          */
 694         if (vma->vm_file) {
 695                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 696                 __unmap_hugepage_range(vma, start, end);
 697                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 698         }
 699 }
 700
 701 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 702                         unsigned long address, pte_t *ptep, pte_t pte)
 703 {
 704         struct page *old_page, *new_page;
 705         int avoidcopy;
 706
 707         old_page = pte_page(pte);
 708
 709         /* If no-one else is actually using this page, avoid the copy
 710          * and just make the page writable */
 711         avoidcopy = (page_count(old_page) == 1);
 712         if (avoidcopy) {
 713                 set_huge_ptep_writable(vma, address, ptep);
 714                 return 0;
 715         }
 716
 717         page_cache_get(old_page);
 718         new_page = alloc_huge_page(vma, address);
 719
 720         if (!new_page) {
 721                 page_cache_release(old_page);
 722                 return VM_FAULT_OOM;
 723         }
 724
 725         spin_unlock(&mm->page_table_lock);
 726         copy_huge_page(new_page, old_page, address, vma);
 727         spin_lock(&mm->page_table_lock);
 728
 729         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 730         if (likely(pte_same(*ptep, pte))) {
 731                 /* Break COW */
 732                 set_huge_pte_at(mm, address, ptep,
 733                                 make_huge_pte(vma, new_page, 1));
 734                 /* Make the old page be freed below */
 735                 new_page = old_page;
 736         }
 737         page_cache_release(new_page);
 738         page_cache_release(old_page);
 739         return 0;
 740 }
 741
 742 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 743                         unsigned long address, pte_t *ptep, int write_access)
 744 {
 745         int ret = VM_FAULT_SIGBUS;
 746         unsigned long idx;
 747         unsigned long size;
 748         struct page *page;
 749         struct address_space *mapping;
 750         pte_t new_pte;
 751
 752         mapping = vma->vm_file->f_mapping;
 753         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 754                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 755
 756         /*
 757          * Use page lock to guard against racing truncation
 758          * before we get page_table_lock.
 759          */
 760 retry:
 761         page = find_lock_page(mapping, idx);
 762         if (!page) {
 763                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 764                 if (idx >= size)
 765                         goto out;
 766                 if (hugetlb_get_quota(mapping))
 767                         goto out;
 768                 page = alloc_huge_page(vma, address);
 769                 if (!page) {
 770                         hugetlb_put_quota(mapping);
 771                         ret = VM_FAULT_OOM;
 772                         goto out;
 773                 }
 774                 clear_huge_page(page, address);
 775
 776                 if (vma->vm_flags & VM_SHARED) {
 777                         int err;
 778
 779                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 780                         if (err) {
 781                                 put_page(page);
 782                                 hugetlb_put_quota(mapping);
 783                                 if (err == -EEXIST)
 784                                         goto retry;
 785                                 goto out;
 786                         }
 787                 } else
 788                         lock_page(page);
 789         }
 790
 791         spin_lock(&mm->page_table_lock);
 792         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 793         if (idx >= size)
 794                 goto backout;
 795
 796         ret = 0;
 797         if (!pte_none(*ptep))
 798                 goto backout;
 799
 800         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 801                                 && (vma->vm_flags & VM_SHARED)));
 802         set_huge_pte_at(mm, address, ptep, new_pte);
 803
 804         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 805                 /* Optimization, do the COW without a second fault */
 806                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 807         }
 808
 809         spin_unlock(&mm->page_table_lock);
 810         unlock_page(page);
 811 out:
 812         return ret;
 813
 814 backout:
 815         spin_unlock(&mm->page_table_lock);
 816         hugetlb_put_quota(mapping);
 817         unlock_page(page);
 818         put_page(page);
 819         goto out;
 820 }
 821
 822 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 823                         unsigned long address, int write_access)
 824 {
 825         pte_t *ptep;
 826         pte_t entry;
 827         int ret;
 828         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 829
 830         ptep = huge_pte_alloc(mm, address);
 831         if (!ptep)
 832                 return VM_FAULT_OOM;
 833
 834         /*
 835          * Serialize hugepage allocation and instantiation, so that we don't
 836          * get spurious allocation failures if two CPUs race to instantiate
 837          * the same page in the page cache.
 838          */
 839         mutex_lock(&hugetlb_instantiation_mutex);
 840         entry = *ptep;
 841         if (pte_none(entry)) {
 842                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 843                 mutex_unlock(&hugetlb_instantiation_mutex);
 844                 return ret;
 845         }
 846
 847         ret = 0;
 848
 849         spin_lock(&mm->page_table_lock);
 850         /* Check for a racing update before calling hugetlb_cow */
 851         if (likely(pte_same(entry, *ptep)))
 852                 if (write_access && !pte_write(entry))
 853                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 854         spin_unlock(&mm->page_table_lock);
 855         mutex_unlock(&hugetlb_instantiation_mutex);
 856
 857         return ret;
 858 }
 859
 860 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 861                         struct page **pages, struct vm_area_struct **vmas,
 862                         unsigned long *position, int *length, int i)
 863 {
 864         unsigned long pfn_offset;
 865         unsigned long vaddr = *position;
 866         int remainder = *length;
 867
 868         spin_lock(&mm->page_table_lock);
 869         while (vaddr < vma->vm_end && remainder) {
 870                 pte_t *pte;
 871                 struct page *page;
 872
 873                 /*
 874                  * Some archs (sparc64, sh*) have multiple pte_ts to
 875                  * each hugepage.  We have to make * sure we get the
 876                  * first, for the page indexing below to work.
 877                  */
 878                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 879
 880                 if (!pte || pte_none(*pte)) {
 881                         int ret;
 882
 883                         spin_unlock(&mm->page_table_lock);
 884                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 885                         spin_lock(&mm->page_table_lock);
 886                         if (!(ret & VM_FAULT_ERROR))
 887                                 continue;
 888
 889                         remainder = 0;
 890                         if (!i)
 891                                 i = -EFAULT;
 892                         break;
 893                 }
 894
 895                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 896                 page = pte_page(*pte);
 897 same_page:
 898                 if (pages) {
 899                         get_page(page);
 900                         pages[i] = page + pfn_offset;
 901                 }
 902
 903                 if (vmas)
 904                         vmas[i] = vma;
 905
 906                 vaddr += PAGE_SIZE;
 907                 ++pfn_offset;
 908                 --remainder;
 909                 ++i;
 910                 if (vaddr < vma->vm_end && remainder &&
 911                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 912                         /*
 913                          * We use pfn_offset to avoid touching the pageframes
 914                          * of this compound page.
 915                          */
 916                         goto same_page;
 917                 }
 918         }
 919         spin_unlock(&mm->page_table_lock);
 920         *length = remainder;
 921         *position = vaddr;
 922
 923         return i;
 924 }
 925
 926 void hugetlb_change_protection(struct vm_area_struct *vma,
 927                 unsigned long address, unsigned long end, pgprot_t newprot)
 928 {
 929         struct mm_struct *mm = vma->vm_mm;
 930         unsigned long start = address;
 931         pte_t *ptep;
 932         pte_t pte;
 933
 934         BUG_ON(address >= end);
 935         flush_cache_range(vma, address, end);
 936
 937         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 938         spin_lock(&mm->page_table_lock);
 939         for (; address < end; address += HPAGE_SIZE) {
 940                 ptep = huge_pte_offset(mm, address);
 941                 if (!ptep)
 942                         continue;
 943                 if (huge_pmd_unshare(mm, &address, ptep))
 944                         continue;
 945                 if (!pte_none(*ptep)) {
 946                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 947                         pte = pte_mkhuge(pte_modify(pte, newprot));
 948                         set_huge_pte_at(mm, address, ptep, pte);
 949                 }
 950         }
 951         spin_unlock(&mm->page_table_lock);
 952         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 953
 954         flush_tlb_range(vma, start, end);
 955 }
 956
 957 struct file_region {
 958         struct list_head link;
 959         long from;
 960         long to;
 961 };
 962
 963 static long region_add(struct list_head *head, long f, long t)
 964 {
 965         struct file_region *rg, *nrg, *trg;
 966
 967         /* Locate the region we are either in or before. */
 968         list_for_each_entry(rg, head, link)
 969                 if (f <= rg->to)
 970                         break;
 971
 972         /* Round our left edge to the current segment if it encloses us. */
 973         if (f > rg->from)
 974                 f = rg->from;
 975
 976         /* Check for and consume any regions we now overlap with. */
 977         nrg = rg;
 978         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 979                 if (&rg->link == head)
 980                         break;
 981                 if (rg->from > t)
 982                         break;
 983
 984                 /* If this area reaches higher then extend our area to
 985                  * include it completely.  If this is not the first area
 986                  * which we intend to reuse, free it. */
 987                 if (rg->to > t)
 988                         t = rg->to;
 989                 if (rg != nrg) {
 990                         list_del(&rg->link);
 991                         kfree(rg);
 992                 }
 993         }
 994         nrg->from = f;
 995         nrg->to = t;
 996         return 0;
 997 }
 998
 999 static long region_chg(struct list_head *head, long f, long t)
1000 {
1001         struct file_region *rg, *nrg;
1002         long chg = 0;
1003
1004         /* Locate the region we are before or in. */
1005         list_for_each_entry(rg, head, link)
1006                 if (f <= rg->to)
1007                         break;
1008
1009         /* If we are below the current region then a new region is required.
1010          * Subtle, allocate a new region at the position but make it zero
1011          * size such that we can guarentee to record the reservation. */
1012         if (&rg->link == head || t < rg->from) {
1013                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1014                 if (nrg == 0)
1015                         return -ENOMEM;
1016                 nrg->from = f;
1017                 nrg->to   = f;
1018                 INIT_LIST_HEAD(&nrg->link);
1019                 list_add(&nrg->link, rg->link.prev);
1020
1021                 return t - f;
1022         }
1023
1024         /* Round our left edge to the current segment if it encloses us. */
1025         if (f > rg->from)
1026                 f = rg->from;
1027         chg = t - f;
1028
1029         /* Check for and consume any regions we now overlap with. */
1030         list_for_each_entry(rg, rg->link.prev, link) {
1031                 if (&rg->link == head)
1032                         break;
1033                 if (rg->from > t)
1034                         return chg;
1035
1036                 /* We overlap with this area, if it extends futher than
1037                  * us then we must extend ourselves.  Account for its
1038                  * existing reservation. */
1039                 if (rg->to > t) {
1040                         chg += rg->to - t;
1041                         t = rg->to;
1042                 }
1043                 chg -= rg->to - rg->from;
1044         }
1045         return chg;
1046 }
1047
1048 static long region_truncate(struct list_head *head, long end)
1049 {
1050         struct file_region *rg, *trg;
1051         long chg = 0;
1052
1053         /* Locate the region we are either in or before. */
1054         list_for_each_entry(rg, head, link)
1055                 if (end <= rg->to)
1056                         break;
1057         if (&rg->link == head)
1058                 return 0;
1059
1060         /* If we are in the middle of a region then adjust it. */
1061         if (end > rg->from) {
1062                 chg = rg->to - end;
1063                 rg->to = end;
1064                 rg = list_entry(rg->link.next, typeof(*rg), link);
1065         }
1066
1067         /* Drop any remaining regions. */
1068         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1069                 if (&rg->link == head)
1070                         break;
1071                 chg += rg->to - rg->from;
1072                 list_del(&rg->link);
1073                 kfree(rg);
1074         }
1075         return chg;
1076 }
1077
1078 static int hugetlb_acct_memory(long delta)
1079 {
1080         int ret = -ENOMEM;
1081
1082         spin_lock(&hugetlb_lock);
1083         /*
1084          * When cpuset is configured, it breaks the strict hugetlb page
1085          * reservation as the accounting is done on a global variable. Such
1086          * reservation is completely rubbish in the presence of cpuset because
1087          * the reservation is not checked against page availability for the
1088          * current cpuset. Application can still potentially OOM'ed by kernel
1089          * with lack of free htlb page in cpuset that the task is in.
1090          * Attempt to enforce strict accounting with cpuset is almost
1091          * impossible (or too ugly) because cpuset is too fluid that
1092          * task or memory node can be dynamically moved between cpusets.
1093          *
1094          * The change of semantics for shared hugetlb mapping with cpuset is
1095          * undesirable. However, in order to preserve some of the semantics,
1096          * we fall back to check against current free page availability as
1097          * a best attempt and hopefully to minimize the impact of changing
1098          * semantics that cpuset has.
1099          */
1100         if (delta > 0) {
1101                 if (gather_surplus_pages(delta) < 0)
1102                         goto out;
1103
1104                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1105                         goto out;
1106         }
1107
1108         ret = 0;
1109         resv_huge_pages += delta;
1110         if (delta < 0)
1111                 return_unused_surplus_pages((unsigned long) -delta);
1112
1113 out:
1114         spin_unlock(&hugetlb_lock);
1115         return ret;
1116 }
1117
1118 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1119 {
1120         long ret, chg;
1121
1122         chg = region_chg(&inode->i_mapping->private_list, from, to);
1123         if (chg < 0)
1124                 return chg;
1125
1126         ret = hugetlb_acct_memory(chg);
1127         if (ret < 0)
1128                 return ret;
1129         region_add(&inode->i_mapping->private_list, from, to);
1130         return 0;
1131 }
1132
1133 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1134 {
1135         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1136         hugetlb_acct_memory(freed - chg);
1137 }