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 unsigned long max_huge_pages;
  27 static struct list_head hugepage_freelists[MAX_NUMNODES];
  28 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  29 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  30 /*
  31  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  32  */
  33 static DEFINE_SPINLOCK(hugetlb_lock);
  34
  35 static void clear_huge_page(struct page *page, unsigned long addr)
  36 {
  37         int i;
  38
  39         might_sleep();
  40         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  41                 cond_resched();
  42                 clear_user_highpage(page + i, addr);
  43         }
  44 }
  45
  46 static void copy_huge_page(struct page *dst, struct page *src,
  47                            unsigned long addr, struct vm_area_struct *vma)
  48 {
  49         int i;
  50
  51         might_sleep();
  52         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  53                 cond_resched();
  54                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  55         }
  56 }
  57
  58 static void enqueue_huge_page(struct page *page)
  59 {
  60         int nid = page_to_nid(page);
  61         list_add(&page->lru, &hugepage_freelists[nid]);
  62         free_huge_pages++;
  63         free_huge_pages_node[nid]++;
  64 }
  65
  66 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  67                                 unsigned long address)
  68 {
  69         int nid = numa_node_id();
  70         struct page *page = NULL;
  71         struct zonelist *zonelist = huge_zonelist(vma, address);
  72         struct zone **z;
  73
  74         for (z = zonelist->zones; *z; z++) {
  75                 nid = zone_to_nid(*z);
  76                 if (cpuset_zone_allowed_softwall(*z, GFP_HIGHUSER) &&
  77                     !list_empty(&hugepage_freelists[nid]))
  78                         break;
  79         }
  80
  81         if (*z) {
  82                 page = list_entry(hugepage_freelists[nid].next,
  83                                   struct page, lru);
  84                 list_del(&page->lru);
  85                 free_huge_pages--;
  86                 free_huge_pages_node[nid]--;
  87         }
  88         return page;
  89 }
  90
  91 static void free_huge_page(struct page *page)
  92 {
  93         BUG_ON(page_count(page));
  94
  95         INIT_LIST_HEAD(&page->lru);
  96
  97         spin_lock(&hugetlb_lock);
  98         enqueue_huge_page(page);
  99         spin_unlock(&hugetlb_lock);
 100 }
 101
 102 static int alloc_fresh_huge_page(void)
 103 {
 104         static int nid = 0;
 105         struct page *page;
 106         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
 107                                         HUGETLB_PAGE_ORDER);
 108         nid = next_node(nid, node_online_map);
 109         if (nid == MAX_NUMNODES)
 110                 nid = first_node(node_online_map);
 111         if (page) {
 112                 set_compound_page_dtor(page, free_huge_page);
 113                 spin_lock(&hugetlb_lock);
 114                 nr_huge_pages++;
 115                 nr_huge_pages_node[page_to_nid(page)]++;
 116                 spin_unlock(&hugetlb_lock);
 117                 put_page(page); /* free it into the hugepage allocator */
 118                 return 1;
 119         }
 120         return 0;
 121 }
 122
 123 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 124                                     unsigned long addr)
 125 {
 126         struct page *page;
 127
 128         spin_lock(&hugetlb_lock);
 129         if (vma->vm_flags & VM_MAYSHARE)
 130                 resv_huge_pages--;
 131         else if (free_huge_pages <= resv_huge_pages)
 132                 goto fail;
 133
 134         page = dequeue_huge_page(vma, addr);
 135         if (!page)
 136                 goto fail;
 137
 138         spin_unlock(&hugetlb_lock);
 139         set_page_refcounted(page);
 140         return page;
 141
 142 fail:
 143         if (vma->vm_flags & VM_MAYSHARE)
 144                 resv_huge_pages++;
 145         spin_unlock(&hugetlb_lock);
 146         return NULL;
 147 }
 148
 149 static int __init hugetlb_init(void)
 150 {
 151         unsigned long i;
 152
 153         if (HPAGE_SHIFT == 0)
 154                 return 0;
 155
 156         for (i = 0; i < MAX_NUMNODES; ++i)
 157                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 158
 159         for (i = 0; i < max_huge_pages; ++i) {
 160                 if (!alloc_fresh_huge_page())
 161                         break;
 162         }
 163         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 164         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 165         return 0;
 166 }
 167 module_init(hugetlb_init);
 168
 169 static int __init hugetlb_setup(char *s)
 170 {
 171         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 172                 max_huge_pages = 0;
 173         return 1;
 174 }
 175 __setup("hugepages=", hugetlb_setup);
 176
 177 #ifdef CONFIG_SYSCTL
 178 static void update_and_free_page(struct page *page)
 179 {
 180         int i;
 181         nr_huge_pages--;
 182         nr_huge_pages_node[page_to_nid(page)]--;
 183         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 184                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 185                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 186                                 1 << PG_private | 1<< PG_writeback);
 187         }
 188         page[1].lru.next = NULL;
 189         set_page_refcounted(page);
 190         __free_pages(page, HUGETLB_PAGE_ORDER);
 191 }
 192
 193 #ifdef CONFIG_HIGHMEM
 194 static void try_to_free_low(unsigned long count)
 195 {
 196         int i;
 197
 198         for (i = 0; i < MAX_NUMNODES; ++i) {
 199                 struct page *page, *next;
 200                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 201                         if (PageHighMem(page))
 202                                 continue;
 203                         list_del(&page->lru);
 204                         update_and_free_page(page);
 205                         free_huge_pages--;
 206                         free_huge_pages_node[page_to_nid(page)]--;
 207                         if (count >= nr_huge_pages)
 208                                 return;
 209                 }
 210         }
 211 }
 212 #else
 213 static inline void try_to_free_low(unsigned long count)
 214 {
 215 }
 216 #endif
 217
 218 static unsigned long set_max_huge_pages(unsigned long count)
 219 {
 220         while (count > nr_huge_pages) {
 221                 if (!alloc_fresh_huge_page())
 222                         return nr_huge_pages;
 223         }
 224         if (count >= nr_huge_pages)
 225                 return nr_huge_pages;
 226
 227         spin_lock(&hugetlb_lock);
 228         count = max(count, resv_huge_pages);
 229         try_to_free_low(count);
 230         while (count < nr_huge_pages) {
 231                 struct page *page = dequeue_huge_page(NULL, 0);
 232                 if (!page)
 233                         break;
 234                 update_and_free_page(page);
 235         }
 236         spin_unlock(&hugetlb_lock);
 237         return nr_huge_pages;
 238 }
 239
 240 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 241                            struct file *file, void __user *buffer,
 242                            size_t *length, loff_t *ppos)
 243 {
 244         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 245         max_huge_pages = set_max_huge_pages(max_huge_pages);
 246         return 0;
 247 }
 248 #endif /* CONFIG_SYSCTL */
 249
 250 int hugetlb_report_meminfo(char *buf)
 251 {
 252         return sprintf(buf,
 253                         "HugePages_Total: %5lu\n"
 254                         "HugePages_Free:  %5lu\n"
 255                         "HugePages_Rsvd:  %5lu\n"
 256                         "Hugepagesize:    %5lu kB\n",
 257                         nr_huge_pages,
 258                         free_huge_pages,
 259                         resv_huge_pages,
 260                         HPAGE_SIZE/1024);
 261 }
 262
 263 int hugetlb_report_node_meminfo(int nid, char *buf)
 264 {
 265         return sprintf(buf,
 266                 "Node %d HugePages_Total: %5u\n"
 267                 "Node %d HugePages_Free:  %5u\n",
 268                 nid, nr_huge_pages_node[nid],
 269                 nid, free_huge_pages_node[nid]);
 270 }
 271
 272 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 273 unsigned long hugetlb_total_pages(void)
 274 {
 275         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 276 }
 277
 278 /*
 279  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 280  * handle_mm_fault() to try to instantiate regular-sized pages in the
 281  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 282  * this far.
 283  */
 284 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 285                                 unsigned long address, int *unused)
 286 {
 287         BUG();
 288         return NULL;
 289 }
 290
 291 struct vm_operations_struct hugetlb_vm_ops = {
 292         .nopage = hugetlb_nopage,
 293 };
 294
 295 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 296                                 int writable)
 297 {
 298         pte_t entry;
 299
 300         if (writable) {
 301                 entry =
 302                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 303         } else {
 304                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 305         }
 306         entry = pte_mkyoung(entry);
 307         entry = pte_mkhuge(entry);
 308
 309         return entry;
 310 }
 311
 312 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 313                                    unsigned long address, pte_t *ptep)
 314 {
 315         pte_t entry;
 316
 317         entry = pte_mkwrite(pte_mkdirty(*ptep));
 318         ptep_set_access_flags(vma, address, ptep, entry, 1);
 319         update_mmu_cache(vma, address, entry);
 320         lazy_mmu_prot_update(entry);
 321 }
 322
 323
 324 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 325                             struct vm_area_struct *vma)
 326 {
 327         pte_t *src_pte, *dst_pte, entry;
 328         struct page *ptepage;
 329         unsigned long addr;
 330         int cow;
 331
 332         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 333
 334         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 335                 src_pte = huge_pte_offset(src, addr);
 336                 if (!src_pte)
 337                         continue;
 338                 dst_pte = huge_pte_alloc(dst, addr);
 339                 if (!dst_pte)
 340                         goto nomem;
 341                 spin_lock(&dst->page_table_lock);
 342                 spin_lock(&src->page_table_lock);
 343                 if (!pte_none(*src_pte)) {
 344                         if (cow)
 345                                 ptep_set_wrprotect(src, addr, src_pte);
 346                         entry = *src_pte;
 347                         ptepage = pte_page(entry);
 348                         get_page(ptepage);
 349                         set_huge_pte_at(dst, addr, dst_pte, entry);
 350                 }
 351                 spin_unlock(&src->page_table_lock);
 352                 spin_unlock(&dst->page_table_lock);
 353         }
 354         return 0;
 355
 356 nomem:
 357         return -ENOMEM;
 358 }
 359
 360 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 361                             unsigned long end)
 362 {
 363         struct mm_struct *mm = vma->vm_mm;
 364         unsigned long address;
 365         pte_t *ptep;
 366         pte_t pte;
 367         struct page *page;
 368         struct page *tmp;
 369         /*
 370          * A page gathering list, protected by per file i_mmap_lock. The
 371          * lock is used to avoid list corruption from multiple unmapping
 372          * of the same page since we are using page->lru.
 373          */
 374         LIST_HEAD(page_list);
 375
 376         WARN_ON(!is_vm_hugetlb_page(vma));
 377         BUG_ON(start & ~HPAGE_MASK);
 378         BUG_ON(end & ~HPAGE_MASK);
 379
 380         spin_lock(&mm->page_table_lock);
 381         for (address = start; address < end; address += HPAGE_SIZE) {
 382                 ptep = huge_pte_offset(mm, address);
 383                 if (!ptep)
 384                         continue;
 385
 386                 if (huge_pmd_unshare(mm, &address, ptep))
 387                         continue;
 388
 389                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 390                 if (pte_none(pte))
 391                         continue;
 392
 393                 page = pte_page(pte);
 394                 if (pte_dirty(pte))
 395                         set_page_dirty(page);
 396                 list_add(&page->lru, &page_list);
 397         }
 398         spin_unlock(&mm->page_table_lock);
 399         flush_tlb_range(vma, start, end);
 400         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 401                 list_del(&page->lru);
 402                 put_page(page);
 403         }
 404 }
 405
 406 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 407                           unsigned long end)
 408 {
 409         /*
 410          * It is undesirable to test vma->vm_file as it should be non-null
 411          * for valid hugetlb area. However, vm_file will be NULL in the error
 412          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 413          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 414          * to clean up. Since no pte has actually been setup, it is safe to
 415          * do nothing in this case.
 416          */
 417         if (vma->vm_file) {
 418                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 419                 __unmap_hugepage_range(vma, start, end);
 420                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 421         }
 422 }
 423
 424 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 425                         unsigned long address, pte_t *ptep, pte_t pte)
 426 {
 427         struct page *old_page, *new_page;
 428         int avoidcopy;
 429
 430         old_page = pte_page(pte);
 431
 432         /* If no-one else is actually using this page, avoid the copy
 433          * and just make the page writable */
 434         avoidcopy = (page_count(old_page) == 1);
 435         if (avoidcopy) {
 436                 set_huge_ptep_writable(vma, address, ptep);
 437                 return VM_FAULT_MINOR;
 438         }
 439
 440         page_cache_get(old_page);
 441         new_page = alloc_huge_page(vma, address);
 442
 443         if (!new_page) {
 444                 page_cache_release(old_page);
 445                 return VM_FAULT_OOM;
 446         }
 447
 448         spin_unlock(&mm->page_table_lock);
 449         copy_huge_page(new_page, old_page, address, vma);
 450         spin_lock(&mm->page_table_lock);
 451
 452         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 453         if (likely(pte_same(*ptep, pte))) {
 454                 /* Break COW */
 455                 set_huge_pte_at(mm, address, ptep,
 456                                 make_huge_pte(vma, new_page, 1));
 457                 /* Make the old page be freed below */
 458                 new_page = old_page;
 459         }
 460         page_cache_release(new_page);
 461         page_cache_release(old_page);
 462         return VM_FAULT_MINOR;
 463 }
 464
 465 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 466                         unsigned long address, pte_t *ptep, int write_access)
 467 {
 468         int ret = VM_FAULT_SIGBUS;
 469         unsigned long idx;
 470         unsigned long size;
 471         struct page *page;
 472         struct address_space *mapping;
 473         pte_t new_pte;
 474
 475         mapping = vma->vm_file->f_mapping;
 476         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 477                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 478
 479         /*
 480          * Use page lock to guard against racing truncation
 481          * before we get page_table_lock.
 482          */
 483 retry:
 484         page = find_lock_page(mapping, idx);
 485         if (!page) {
 486                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 487                 if (idx >= size)
 488                         goto out;
 489                 if (hugetlb_get_quota(mapping))
 490                         goto out;
 491                 page = alloc_huge_page(vma, address);
 492                 if (!page) {
 493                         hugetlb_put_quota(mapping);
 494                         ret = VM_FAULT_OOM;
 495                         goto out;
 496                 }
 497                 clear_huge_page(page, address);
 498
 499                 if (vma->vm_flags & VM_SHARED) {
 500                         int err;
 501
 502                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 503                         if (err) {
 504                                 put_page(page);
 505                                 hugetlb_put_quota(mapping);
 506                                 if (err == -EEXIST)
 507                                         goto retry;
 508                                 goto out;
 509                         }
 510                 } else
 511                         lock_page(page);
 512         }
 513
 514         spin_lock(&mm->page_table_lock);
 515         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 516         if (idx >= size)
 517                 goto backout;
 518
 519         ret = VM_FAULT_MINOR;
 520         if (!pte_none(*ptep))
 521                 goto backout;
 522
 523         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 524                                 && (vma->vm_flags & VM_SHARED)));
 525         set_huge_pte_at(mm, address, ptep, new_pte);
 526
 527         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 528                 /* Optimization, do the COW without a second fault */
 529                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 530         }
 531
 532         spin_unlock(&mm->page_table_lock);
 533         unlock_page(page);
 534 out:
 535         return ret;
 536
 537 backout:
 538         spin_unlock(&mm->page_table_lock);
 539         hugetlb_put_quota(mapping);
 540         unlock_page(page);
 541         put_page(page);
 542         goto out;
 543 }
 544
 545 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 546                         unsigned long address, int write_access)
 547 {
 548         pte_t *ptep;
 549         pte_t entry;
 550         int ret;
 551         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 552
 553         ptep = huge_pte_alloc(mm, address);
 554         if (!ptep)
 555                 return VM_FAULT_OOM;
 556
 557         /*
 558          * Serialize hugepage allocation and instantiation, so that we don't
 559          * get spurious allocation failures if two CPUs race to instantiate
 560          * the same page in the page cache.
 561          */
 562         mutex_lock(&hugetlb_instantiation_mutex);
 563         entry = *ptep;
 564         if (pte_none(entry)) {
 565                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 566                 mutex_unlock(&hugetlb_instantiation_mutex);
 567                 return ret;
 568         }
 569
 570         ret = VM_FAULT_MINOR;
 571
 572         spin_lock(&mm->page_table_lock);
 573         /* Check for a racing update before calling hugetlb_cow */
 574         if (likely(pte_same(entry, *ptep)))
 575                 if (write_access && !pte_write(entry))
 576                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 577         spin_unlock(&mm->page_table_lock);
 578         mutex_unlock(&hugetlb_instantiation_mutex);
 579
 580         return ret;
 581 }
 582
 583 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 584                         struct page **pages, struct vm_area_struct **vmas,
 585                         unsigned long *position, int *length, int i)
 586 {
 587         unsigned long pfn_offset;
 588         unsigned long vaddr = *position;
 589         int remainder = *length;
 590
 591         spin_lock(&mm->page_table_lock);
 592         while (vaddr < vma->vm_end && remainder) {
 593                 pte_t *pte;
 594                 struct page *page;
 595
 596                 /*
 597                  * Some archs (sparc64, sh*) have multiple pte_ts to
 598                  * each hugepage.  We have to make * sure we get the
 599                  * first, for the page indexing below to work.
 600                  */
 601                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 602
 603                 if (!pte || pte_none(*pte)) {
 604                         int ret;
 605
 606                         spin_unlock(&mm->page_table_lock);
 607                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 608                         spin_lock(&mm->page_table_lock);
 609                         if (ret == VM_FAULT_MINOR)
 610                                 continue;
 611
 612                         remainder = 0;
 613                         if (!i)
 614                                 i = -EFAULT;
 615                         break;
 616                 }
 617
 618                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 619                 page = pte_page(*pte);
 620 same_page:
 621                 if (pages) {
 622                         get_page(page);
 623                         pages[i] = page + pfn_offset;
 624                 }
 625
 626                 if (vmas)
 627                         vmas[i] = vma;
 628
 629                 vaddr += PAGE_SIZE;
 630                 ++pfn_offset;
 631                 --remainder;
 632                 ++i;
 633                 if (vaddr < vma->vm_end && remainder &&
 634                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 635                         /*
 636                          * We use pfn_offset to avoid touching the pageframes
 637                          * of this compound page.
 638                          */
 639                         goto same_page;
 640                 }
 641         }
 642         spin_unlock(&mm->page_table_lock);
 643         *length = remainder;
 644         *position = vaddr;
 645
 646         return i;
 647 }
 648
 649 void hugetlb_change_protection(struct vm_area_struct *vma,
 650                 unsigned long address, unsigned long end, pgprot_t newprot)
 651 {
 652         struct mm_struct *mm = vma->vm_mm;
 653         unsigned long start = address;
 654         pte_t *ptep;
 655         pte_t pte;
 656
 657         BUG_ON(address >= end);
 658         flush_cache_range(vma, address, end);
 659
 660         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 661         spin_lock(&mm->page_table_lock);
 662         for (; address < end; address += HPAGE_SIZE) {
 663                 ptep = huge_pte_offset(mm, address);
 664                 if (!ptep)
 665                         continue;
 666                 if (huge_pmd_unshare(mm, &address, ptep))
 667                         continue;
 668                 if (!pte_none(*ptep)) {
 669                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 670                         pte = pte_mkhuge(pte_modify(pte, newprot));
 671                         set_huge_pte_at(mm, address, ptep, pte);
 672                         lazy_mmu_prot_update(pte);
 673                 }
 674         }
 675         spin_unlock(&mm->page_table_lock);
 676         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 677
 678         flush_tlb_range(vma, start, end);
 679 }
 680
 681 struct file_region {
 682         struct list_head link;
 683         long from;
 684         long to;
 685 };
 686
 687 static long region_add(struct list_head *head, long f, long t)
 688 {
 689         struct file_region *rg, *nrg, *trg;
 690
 691         /* Locate the region we are either in or before. */
 692         list_for_each_entry(rg, head, link)
 693                 if (f <= rg->to)
 694                         break;
 695
 696         /* Round our left edge to the current segment if it encloses us. */
 697         if (f > rg->from)
 698                 f = rg->from;
 699
 700         /* Check for and consume any regions we now overlap with. */
 701         nrg = rg;
 702         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 703                 if (&rg->link == head)
 704                         break;
 705                 if (rg->from > t)
 706                         break;
 707
 708                 /* If this area reaches higher then extend our area to
 709                  * include it completely.  If this is not the first area
 710                  * which we intend to reuse, free it. */
 711                 if (rg->to > t)
 712                         t = rg->to;
 713                 if (rg != nrg) {
 714                         list_del(&rg->link);
 715                         kfree(rg);
 716                 }
 717         }
 718         nrg->from = f;
 719         nrg->to = t;
 720         return 0;
 721 }
 722
 723 static long region_chg(struct list_head *head, long f, long t)
 724 {
 725         struct file_region *rg, *nrg;
 726         long chg = 0;
 727
 728         /* Locate the region we are before or in. */
 729         list_for_each_entry(rg, head, link)
 730                 if (f <= rg->to)
 731                         break;
 732
 733         /* If we are below the current region then a new region is required.
 734          * Subtle, allocate a new region at the position but make it zero
 735          * size such that we can guarentee to record the reservation. */
 736         if (&rg->link == head || t < rg->from) {
 737                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 738                 if (nrg == 0)
 739                         return -ENOMEM;
 740                 nrg->from = f;
 741                 nrg->to   = f;
 742                 INIT_LIST_HEAD(&nrg->link);
 743                 list_add(&nrg->link, rg->link.prev);
 744
 745                 return t - f;
 746         }
 747
 748         /* Round our left edge to the current segment if it encloses us. */
 749         if (f > rg->from)
 750                 f = rg->from;
 751         chg = t - f;
 752
 753         /* Check for and consume any regions we now overlap with. */
 754         list_for_each_entry(rg, rg->link.prev, link) {
 755                 if (&rg->link == head)
 756                         break;
 757                 if (rg->from > t)
 758                         return chg;
 759
 760                 /* We overlap with this area, if it extends futher than
 761                  * us then we must extend ourselves.  Account for its
 762                  * existing reservation. */
 763                 if (rg->to > t) {
 764                         chg += rg->to - t;
 765                         t = rg->to;
 766                 }
 767                 chg -= rg->to - rg->from;
 768         }
 769         return chg;
 770 }
 771
 772 static long region_truncate(struct list_head *head, long end)
 773 {
 774         struct file_region *rg, *trg;
 775         long chg = 0;
 776
 777         /* Locate the region we are either in or before. */
 778         list_for_each_entry(rg, head, link)
 779                 if (end <= rg->to)
 780                         break;
 781         if (&rg->link == head)
 782                 return 0;
 783
 784         /* If we are in the middle of a region then adjust it. */
 785         if (end > rg->from) {
 786                 chg = rg->to - end;
 787                 rg->to = end;
 788                 rg = list_entry(rg->link.next, typeof(*rg), link);
 789         }
 790
 791         /* Drop any remaining regions. */
 792         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 793                 if (&rg->link == head)
 794                         break;
 795                 chg += rg->to - rg->from;
 796                 list_del(&rg->link);
 797                 kfree(rg);
 798         }
 799         return chg;
 800 }
 801
 802 static int hugetlb_acct_memory(long delta)
 803 {
 804         int ret = -ENOMEM;
 805
 806         spin_lock(&hugetlb_lock);
 807         if ((delta + resv_huge_pages) <= free_huge_pages) {
 808                 resv_huge_pages += delta;
 809                 ret = 0;
 810         }
 811         spin_unlock(&hugetlb_lock);
 812         return ret;
 813 }
 814
 815 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
 816 {
 817         long ret, chg;
 818
 819         chg = region_chg(&inode->i_mapping->private_list, from, to);
 820         if (chg < 0)
 821                 return chg;
 822         ret = hugetlb_acct_memory(chg);
 823         if (ret < 0)
 824                 return ret;
 825         region_add(&inode->i_mapping->private_list, from, to);
 826         return 0;
 827 }
 828
 829 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 830 {
 831         long chg = region_truncate(&inode->i_mapping->private_list, offset);
 832         hugetlb_acct_memory(freed - chg);
 833 }