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)
  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);
  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(*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                 page[1].lru.next = (void *)free_huge_page;      /* dtor */
 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         spin_unlock(&hugetlb_lock);
 144         return NULL;
 145 }
 146
 147 static int __init hugetlb_init(void)
 148 {
 149         unsigned long i;
 150
 151         if (HPAGE_SHIFT == 0)
 152                 return 0;
 153
 154         for (i = 0; i < MAX_NUMNODES; ++i)
 155                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 156
 157         for (i = 0; i < max_huge_pages; ++i) {
 158                 if (!alloc_fresh_huge_page())
 159                         break;
 160         }
 161         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 162         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 163         return 0;
 164 }
 165 module_init(hugetlb_init);
 166
 167 static int __init hugetlb_setup(char *s)
 168 {
 169         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 170                 max_huge_pages = 0;
 171         return 1;
 172 }
 173 __setup("hugepages=", hugetlb_setup);
 174
 175 #ifdef CONFIG_SYSCTL
 176 static void update_and_free_page(struct page *page)
 177 {
 178         int i;
 179         nr_huge_pages--;
 180         nr_huge_pages_node[page_to_nid(page)]--;
 181         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 182                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 183                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 184                                 1 << PG_private | 1<< PG_writeback);
 185         }
 186         page[1].lru.next = NULL;
 187         set_page_refcounted(page);
 188         __free_pages(page, HUGETLB_PAGE_ORDER);
 189 }
 190
 191 #ifdef CONFIG_HIGHMEM
 192 static void try_to_free_low(unsigned long count)
 193 {
 194         int i;
 195
 196         for (i = 0; i < MAX_NUMNODES; ++i) {
 197                 struct page *page, *next;
 198                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 199                         if (PageHighMem(page))
 200                                 continue;
 201                         list_del(&page->lru);
 202                         update_and_free_page(page);
 203                         free_huge_pages--;
 204                         free_huge_pages_node[page_to_nid(page)]--;
 205                         if (count >= nr_huge_pages)
 206                                 return;
 207                 }
 208         }
 209 }
 210 #else
 211 static inline void try_to_free_low(unsigned long count)
 212 {
 213 }
 214 #endif
 215
 216 static unsigned long set_max_huge_pages(unsigned long count)
 217 {
 218         while (count > nr_huge_pages) {
 219                 if (!alloc_fresh_huge_page())
 220                         return nr_huge_pages;
 221         }
 222         if (count >= nr_huge_pages)
 223                 return nr_huge_pages;
 224
 225         spin_lock(&hugetlb_lock);
 226         count = max(count, resv_huge_pages);
 227         try_to_free_low(count);
 228         while (count < nr_huge_pages) {
 229                 struct page *page = dequeue_huge_page(NULL, 0);
 230                 if (!page)
 231                         break;
 232                 update_and_free_page(page);
 233         }
 234         spin_unlock(&hugetlb_lock);
 235         return nr_huge_pages;
 236 }
 237
 238 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 239                            struct file *file, void __user *buffer,
 240                            size_t *length, loff_t *ppos)
 241 {
 242         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 243         max_huge_pages = set_max_huge_pages(max_huge_pages);
 244         return 0;
 245 }
 246 #endif /* CONFIG_SYSCTL */
 247
 248 int hugetlb_report_meminfo(char *buf)
 249 {
 250         return sprintf(buf,
 251                         "HugePages_Total: %5lu\n"
 252                         "HugePages_Free:  %5lu\n"
 253                         "HugePages_Rsvd:  %5lu\n"
 254                         "Hugepagesize:    %5lu kB\n",
 255                         nr_huge_pages,
 256                         free_huge_pages,
 257                         resv_huge_pages,
 258                         HPAGE_SIZE/1024);
 259 }
 260
 261 int hugetlb_report_node_meminfo(int nid, char *buf)
 262 {
 263         return sprintf(buf,
 264                 "Node %d HugePages_Total: %5u\n"
 265                 "Node %d HugePages_Free:  %5u\n",
 266                 nid, nr_huge_pages_node[nid],
 267                 nid, free_huge_pages_node[nid]);
 268 }
 269
 270 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 271 unsigned long hugetlb_total_pages(void)
 272 {
 273         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 274 }
 275
 276 /*
 277  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 278  * handle_mm_fault() to try to instantiate regular-sized pages in the
 279  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 280  * this far.
 281  */
 282 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 283                                 unsigned long address, int *unused)
 284 {
 285         BUG();
 286         return NULL;
 287 }
 288
 289 struct vm_operations_struct hugetlb_vm_ops = {
 290         .nopage = hugetlb_nopage,
 291 };
 292
 293 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 294                                 int writable)
 295 {
 296         pte_t entry;
 297
 298         if (writable) {
 299                 entry =
 300                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 301         } else {
 302                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 303         }
 304         entry = pte_mkyoung(entry);
 305         entry = pte_mkhuge(entry);
 306
 307         return entry;
 308 }
 309
 310 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 311                                    unsigned long address, pte_t *ptep)
 312 {
 313         pte_t entry;
 314
 315         entry = pte_mkwrite(pte_mkdirty(*ptep));
 316         ptep_set_access_flags(vma, address, ptep, entry, 1);
 317         update_mmu_cache(vma, address, entry);
 318         lazy_mmu_prot_update(entry);
 319 }
 320
 321
 322 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 323                             struct vm_area_struct *vma)
 324 {
 325         pte_t *src_pte, *dst_pte, entry;
 326         struct page *ptepage;
 327         unsigned long addr;
 328         int cow;
 329
 330         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 331
 332         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 333                 src_pte = huge_pte_offset(src, addr);
 334                 if (!src_pte)
 335                         continue;
 336                 dst_pte = huge_pte_alloc(dst, addr);
 337                 if (!dst_pte)
 338                         goto nomem;
 339                 spin_lock(&dst->page_table_lock);
 340                 spin_lock(&src->page_table_lock);
 341                 if (!pte_none(*src_pte)) {
 342                         if (cow)
 343                                 ptep_set_wrprotect(src, addr, src_pte);
 344                         entry = *src_pte;
 345                         ptepage = pte_page(entry);
 346                         get_page(ptepage);
 347                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
 348                         set_huge_pte_at(dst, addr, dst_pte, entry);
 349                 }
 350                 spin_unlock(&src->page_table_lock);
 351                 spin_unlock(&dst->page_table_lock);
 352         }
 353         return 0;
 354
 355 nomem:
 356         return -ENOMEM;
 357 }
 358
 359 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 360                             unsigned long end)
 361 {
 362         struct mm_struct *mm = vma->vm_mm;
 363         unsigned long address;
 364         pte_t *ptep;
 365         pte_t pte;
 366         struct page *page;
 367         struct page *tmp;
 368         /*
 369          * A page gathering list, protected by per file i_mmap_lock. The
 370          * lock is used to avoid list corruption from multiple unmapping
 371          * of the same page since we are using page->lru.
 372          */
 373         LIST_HEAD(page_list);
 374
 375         WARN_ON(!is_vm_hugetlb_page(vma));
 376         BUG_ON(start & ~HPAGE_MASK);
 377         BUG_ON(end & ~HPAGE_MASK);
 378
 379         spin_lock(&mm->page_table_lock);
 380
 381         /* Update high watermark before we lower rss */
 382         update_hiwater_rss(mm);
 383
 384         for (address = start; address < end; address += HPAGE_SIZE) {
 385                 ptep = huge_pte_offset(mm, address);
 386                 if (!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                 list_add(&page->lru, &page_list);
 395                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
 396         }
 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);
 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         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
 524         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 525                                 && (vma->vm_flags & VM_SHARED)));
 526         set_huge_pte_at(mm, address, ptep, new_pte);
 527
 528         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 529                 /* Optimization, do the COW without a second fault */
 530                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 531         }
 532
 533         spin_unlock(&mm->page_table_lock);
 534         unlock_page(page);
 535 out:
 536         return ret;
 537
 538 backout:
 539         spin_unlock(&mm->page_table_lock);
 540         hugetlb_put_quota(mapping);
 541         unlock_page(page);
 542         put_page(page);
 543         goto out;
 544 }
 545
 546 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 547                         unsigned long address, int write_access)
 548 {
 549         pte_t *ptep;
 550         pte_t entry;
 551         int ret;
 552         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 553
 554         ptep = huge_pte_alloc(mm, address);
 555         if (!ptep)
 556                 return VM_FAULT_OOM;
 557
 558         /*
 559          * Serialize hugepage allocation and instantiation, so that we don't
 560          * get spurious allocation failures if two CPUs race to instantiate
 561          * the same page in the page cache.
 562          */
 563         mutex_lock(&hugetlb_instantiation_mutex);
 564         entry = *ptep;
 565         if (pte_none(entry)) {
 566                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 567                 mutex_unlock(&hugetlb_instantiation_mutex);
 568                 return ret;
 569         }
 570
 571         ret = VM_FAULT_MINOR;
 572
 573         spin_lock(&mm->page_table_lock);
 574         /* Check for a racing update before calling hugetlb_cow */
 575         if (likely(pte_same(entry, *ptep)))
 576                 if (write_access && !pte_write(entry))
 577                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 578         spin_unlock(&mm->page_table_lock);
 579         mutex_unlock(&hugetlb_instantiation_mutex);
 580
 581         return ret;
 582 }
 583
 584 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 585                         struct page **pages, struct vm_area_struct **vmas,
 586                         unsigned long *position, int *length, int i)
 587 {
 588         unsigned long pfn_offset;
 589         unsigned long vaddr = *position;
 590         int remainder = *length;
 591
 592         spin_lock(&mm->page_table_lock);
 593         while (vaddr < vma->vm_end && remainder) {
 594                 pte_t *pte;
 595                 struct page *page;
 596
 597                 /*
 598                  * Some archs (sparc64, sh*) have multiple pte_ts to
 599                  * each hugepage.  We have to make * sure we get the
 600                  * first, for the page indexing below to work.
 601                  */
 602                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 603
 604                 if (!pte || pte_none(*pte)) {
 605                         int ret;
 606
 607                         spin_unlock(&mm->page_table_lock);
 608                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 609                         spin_lock(&mm->page_table_lock);
 610                         if (ret == VM_FAULT_MINOR)
 611                                 continue;
 612
 613                         remainder = 0;
 614                         if (!i)
 615                                 i = -EFAULT;
 616                         break;
 617                 }
 618
 619                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 620                 page = pte_page(*pte);
 621 same_page:
 622                 if (pages) {
 623                         get_page(page);
 624                         pages[i] = page + pfn_offset;
 625                 }
 626
 627                 if (vmas)
 628                         vmas[i] = vma;
 629
 630                 vaddr += PAGE_SIZE;
 631                 ++pfn_offset;
 632                 --remainder;
 633                 ++i;
 634                 if (vaddr < vma->vm_end && remainder &&
 635                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 636                         /*
 637                          * We use pfn_offset to avoid touching the pageframes
 638                          * of this compound page.
 639                          */
 640                         goto same_page;
 641                 }
 642         }
 643         spin_unlock(&mm->page_table_lock);
 644         *length = remainder;
 645         *position = vaddr;
 646
 647         return i;
 648 }
 649
 650 void hugetlb_change_protection(struct vm_area_struct *vma,
 651                 unsigned long address, unsigned long end, pgprot_t newprot)
 652 {
 653         struct mm_struct *mm = vma->vm_mm;
 654         unsigned long start = address;
 655         pte_t *ptep;
 656         pte_t pte;
 657
 658         BUG_ON(address >= end);
 659         flush_cache_range(vma, address, end);
 660
 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 (!pte_none(*ptep)) {
 667                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 668                         pte = pte_mkhuge(pte_modify(pte, newprot));
 669                         set_huge_pte_at(mm, address, ptep, pte);
 670                         lazy_mmu_prot_update(pte);
 671                 }
 672         }
 673         spin_unlock(&mm->page_table_lock);
 674
 675         flush_tlb_range(vma, start, end);
 676 }
 677
 678 struct file_region {
 679         struct list_head link;
 680         long from;
 681         long to;
 682 };
 683
 684 static long region_add(struct list_head *head, long f, long t)
 685 {
 686         struct file_region *rg, *nrg, *trg;
 687
 688         /* Locate the region we are either in or before. */
 689         list_for_each_entry(rg, head, link)
 690                 if (f <= rg->to)
 691                         break;
 692
 693         /* Round our left edge to the current segment if it encloses us. */
 694         if (f > rg->from)
 695                 f = rg->from;
 696
 697         /* Check for and consume any regions we now overlap with. */
 698         nrg = rg;
 699         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 700                 if (&rg->link == head)
 701                         break;
 702                 if (rg->from > t)
 703                         break;
 704
 705                 /* If this area reaches higher then extend our area to
 706                  * include it completely.  If this is not the first area
 707                  * which we intend to reuse, free it. */
 708                 if (rg->to > t)
 709                         t = rg->to;
 710                 if (rg != nrg) {
 711                         list_del(&rg->link);
 712                         kfree(rg);
 713                 }
 714         }
 715         nrg->from = f;
 716         nrg->to = t;
 717         return 0;
 718 }
 719
 720 static long region_chg(struct list_head *head, long f, long t)
 721 {
 722         struct file_region *rg, *nrg;
 723         long chg = 0;
 724
 725         /* Locate the region we are before or in. */
 726         list_for_each_entry(rg, head, link)
 727                 if (f <= rg->to)
 728                         break;
 729
 730         /* If we are below the current region then a new region is required.
 731          * Subtle, allocate a new region at the position but make it zero
 732          * size such that we can guarentee to record the reservation. */
 733         if (&rg->link == head || t < rg->from) {
 734                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 735                 if (nrg == 0)
 736                         return -ENOMEM;
 737                 nrg->from = f;
 738                 nrg->to   = f;
 739                 INIT_LIST_HEAD(&nrg->link);
 740                 list_add(&nrg->link, rg->link.prev);
 741
 742                 return t - f;
 743         }
 744
 745         /* Round our left edge to the current segment if it encloses us. */
 746         if (f > rg->from)
 747                 f = rg->from;
 748         chg = t - f;
 749
 750         /* Check for and consume any regions we now overlap with. */
 751         list_for_each_entry(rg, rg->link.prev, link) {
 752                 if (&rg->link == head)
 753                         break;
 754                 if (rg->from > t)
 755                         return chg;
 756
 757                 /* We overlap with this area, if it extends futher than
 758                  * us then we must extend ourselves.  Account for its
 759                  * existing reservation. */
 760                 if (rg->to > t) {
 761                         chg += rg->to - t;
 762                         t = rg->to;
 763                 }
 764                 chg -= rg->to - rg->from;
 765         }
 766         return chg;
 767 }
 768
 769 static long region_truncate(struct list_head *head, long end)
 770 {
 771         struct file_region *rg, *trg;
 772         long chg = 0;
 773
 774         /* Locate the region we are either in or before. */
 775         list_for_each_entry(rg, head, link)
 776                 if (end <= rg->to)
 777                         break;
 778         if (&rg->link == head)
 779                 return 0;
 780
 781         /* If we are in the middle of a region then adjust it. */
 782         if (end > rg->from) {
 783                 chg = rg->to - end;
 784                 rg->to = end;
 785                 rg = list_entry(rg->link.next, typeof(*rg), link);
 786         }
 787
 788         /* Drop any remaining regions. */
 789         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 790                 if (&rg->link == head)
 791                         break;
 792                 chg += rg->to - rg->from;
 793                 list_del(&rg->link);
 794                 kfree(rg);
 795         }
 796         return chg;
 797 }
 798
 799 static int hugetlb_acct_memory(long delta)
 800 {
 801         int ret = -ENOMEM;
 802
 803         spin_lock(&hugetlb_lock);
 804         if ((delta + resv_huge_pages) <= free_huge_pages) {
 805                 resv_huge_pages += delta;
 806                 ret = 0;
 807         }
 808         spin_unlock(&hugetlb_lock);
 809         return ret;
 810 }
 811
 812 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
 813 {
 814         long ret, chg;
 815
 816         chg = region_chg(&inode->i_mapping->private_list, from, to);
 817         if (chg < 0)
 818                 return chg;
 819         ret = hugetlb_acct_memory(chg);
 820         if (ret < 0)
 821                 return ret;
 822         region_add(&inode->i_mapping->private_list, from, to);
 823         return 0;
 824 }
 825
 826 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 827 {
 828         long chg = region_truncate(&inode->i_mapping->private_list, offset);
 829         hugetlb_acct_memory(freed - chg);
 830 }