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 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  31 unsigned long hugepages_treat_as_movable;
  32
  33 /*
  34  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  35  */
  36 static DEFINE_SPINLOCK(hugetlb_lock);
  37
  38 static void clear_huge_page(struct page *page, unsigned long addr)
  39 {
  40         int i;
  41
  42         might_sleep();
  43         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  44                 cond_resched();
  45                 clear_user_highpage(page + i, addr);
  46         }
  47 }
  48
  49 static void copy_huge_page(struct page *dst, struct page *src,
  50                            unsigned long addr, struct vm_area_struct *vma)
  51 {
  52         int i;
  53
  54         might_sleep();
  55         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  56                 cond_resched();
  57                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  58         }
  59 }
  60
  61 static void enqueue_huge_page(struct page *page)
  62 {
  63         int nid = page_to_nid(page);
  64         list_add(&page->lru, &hugepage_freelists[nid]);
  65         free_huge_pages++;
  66         free_huge_pages_node[nid]++;
  67 }
  68
  69 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  70                                 unsigned long address)
  71 {
  72         int nid;
  73         struct page *page = NULL;
  74         struct zonelist *zonelist = huge_zonelist(vma, address,
  75                                                 htlb_alloc_mask);
  76         struct zone **z;
  77
  78         for (z = zonelist->zones; *z; z++) {
  79                 nid = zone_to_nid(*z);
  80                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
  81                     !list_empty(&hugepage_freelists[nid])) {
  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         }
  89         return page;
  90 }
  91
  92 static void free_huge_page(struct page *page)
  93 {
  94         BUG_ON(page_count(page));
  95
  96         INIT_LIST_HEAD(&page->lru);
  97
  98         spin_lock(&hugetlb_lock);
  99         enqueue_huge_page(page);
 100         spin_unlock(&hugetlb_lock);
 101 }
 102
 103 static int alloc_fresh_huge_page(void)
 104 {
 105         static int prev_nid;
 106         struct page *page;
 107         static DEFINE_SPINLOCK(nid_lock);
 108         int nid;
 109
 110         spin_lock(&nid_lock);
 111         nid = next_node(prev_nid, node_online_map);
 112         if (nid == MAX_NUMNODES)
 113                 nid = first_node(node_online_map);
 114         prev_nid = nid;
 115         spin_unlock(&nid_lock);
 116
 117         page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 118                                         HUGETLB_PAGE_ORDER);
 119         if (page) {
 120                 set_compound_page_dtor(page, free_huge_page);
 121                 spin_lock(&hugetlb_lock);
 122                 nr_huge_pages++;
 123                 nr_huge_pages_node[page_to_nid(page)]++;
 124                 spin_unlock(&hugetlb_lock);
 125                 put_page(page); /* free it into the hugepage allocator */
 126                 return 1;
 127         }
 128         return 0;
 129 }
 130
 131 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 132                                     unsigned long addr)
 133 {
 134         struct page *page;
 135
 136         spin_lock(&hugetlb_lock);
 137         if (vma->vm_flags & VM_MAYSHARE)
 138                 resv_huge_pages--;
 139         else if (free_huge_pages <= resv_huge_pages)
 140                 goto fail;
 141
 142         page = dequeue_huge_page(vma, addr);
 143         if (!page)
 144                 goto fail;
 145
 146         spin_unlock(&hugetlb_lock);
 147         set_page_refcounted(page);
 148         return page;
 149
 150 fail:
 151         if (vma->vm_flags & VM_MAYSHARE)
 152                 resv_huge_pages++;
 153         spin_unlock(&hugetlb_lock);
 154         return NULL;
 155 }
 156
 157 static int __init hugetlb_init(void)
 158 {
 159         unsigned long i;
 160
 161         if (HPAGE_SHIFT == 0)
 162                 return 0;
 163
 164         for (i = 0; i < MAX_NUMNODES; ++i)
 165                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 166
 167         for (i = 0; i < max_huge_pages; ++i) {
 168                 if (!alloc_fresh_huge_page())
 169                         break;
 170         }
 171         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 172         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 173         return 0;
 174 }
 175 module_init(hugetlb_init);
 176
 177 static int __init hugetlb_setup(char *s)
 178 {
 179         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 180                 max_huge_pages = 0;
 181         return 1;
 182 }
 183 __setup("hugepages=", hugetlb_setup);
 184
 185 static unsigned int cpuset_mems_nr(unsigned int *array)
 186 {
 187         int node;
 188         unsigned int nr = 0;
 189
 190         for_each_node_mask(node, cpuset_current_mems_allowed)
 191                 nr += array[node];
 192
 193         return nr;
 194 }
 195
 196 #ifdef CONFIG_SYSCTL
 197 static void update_and_free_page(struct page *page)
 198 {
 199         int i;
 200         nr_huge_pages--;
 201         nr_huge_pages_node[page_to_nid(page)]--;
 202         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 203                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 204                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 205                                 1 << PG_private | 1<< PG_writeback);
 206         }
 207         page[1].lru.next = NULL;
 208         set_page_refcounted(page);
 209         __free_pages(page, HUGETLB_PAGE_ORDER);
 210 }
 211
 212 #ifdef CONFIG_HIGHMEM
 213 static void try_to_free_low(unsigned long count)
 214 {
 215         int i;
 216
 217         for (i = 0; i < MAX_NUMNODES; ++i) {
 218                 struct page *page, *next;
 219                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 220                         if (PageHighMem(page))
 221                                 continue;
 222                         list_del(&page->lru);
 223                         update_and_free_page(page);
 224                         free_huge_pages--;
 225                         free_huge_pages_node[page_to_nid(page)]--;
 226                         if (count >= nr_huge_pages)
 227                                 return;
 228                 }
 229         }
 230 }
 231 #else
 232 static inline void try_to_free_low(unsigned long count)
 233 {
 234 }
 235 #endif
 236
 237 static unsigned long set_max_huge_pages(unsigned long count)
 238 {
 239         while (count > nr_huge_pages) {
 240                 if (!alloc_fresh_huge_page())
 241                         return nr_huge_pages;
 242         }
 243         if (count >= nr_huge_pages)
 244                 return nr_huge_pages;
 245
 246         spin_lock(&hugetlb_lock);
 247         count = max(count, resv_huge_pages);
 248         try_to_free_low(count);
 249         while (count < nr_huge_pages) {
 250                 struct page *page = dequeue_huge_page(NULL, 0);
 251                 if (!page)
 252                         break;
 253                 update_and_free_page(page);
 254         }
 255         spin_unlock(&hugetlb_lock);
 256         return nr_huge_pages;
 257 }
 258
 259 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 260                            struct file *file, void __user *buffer,
 261                            size_t *length, loff_t *ppos)
 262 {
 263         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 264         max_huge_pages = set_max_huge_pages(max_huge_pages);
 265         return 0;
 266 }
 267
 268 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 269                         struct file *file, void __user *buffer,
 270                         size_t *length, loff_t *ppos)
 271 {
 272         proc_dointvec(table, write, file, buffer, length, ppos);
 273         if (hugepages_treat_as_movable)
 274                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 275         else
 276                 htlb_alloc_mask = GFP_HIGHUSER;
 277         return 0;
 278 }
 279
 280 #endif /* CONFIG_SYSCTL */
 281
 282 int hugetlb_report_meminfo(char *buf)
 283 {
 284         return sprintf(buf,
 285                         "HugePages_Total: %5lu\n"
 286                         "HugePages_Free:  %5lu\n"
 287                         "HugePages_Rsvd:  %5lu\n"
 288                         "Hugepagesize:    %5lu kB\n",
 289                         nr_huge_pages,
 290                         free_huge_pages,
 291                         resv_huge_pages,
 292                         HPAGE_SIZE/1024);
 293 }
 294
 295 int hugetlb_report_node_meminfo(int nid, char *buf)
 296 {
 297         return sprintf(buf,
 298                 "Node %d HugePages_Total: %5u\n"
 299                 "Node %d HugePages_Free:  %5u\n",
 300                 nid, nr_huge_pages_node[nid],
 301                 nid, free_huge_pages_node[nid]);
 302 }
 303
 304 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 305 unsigned long hugetlb_total_pages(void)
 306 {
 307         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 308 }
 309
 310 /*
 311  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 312  * handle_mm_fault() to try to instantiate regular-sized pages in the
 313  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 314  * this far.
 315  */
 316 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 317 {
 318         BUG();
 319         return 0;
 320 }
 321
 322 struct vm_operations_struct hugetlb_vm_ops = {
 323         .fault = hugetlb_vm_op_fault,
 324 };
 325
 326 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 327                                 int writable)
 328 {
 329         pte_t entry;
 330
 331         if (writable) {
 332                 entry =
 333                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 334         } else {
 335                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 336         }
 337         entry = pte_mkyoung(entry);
 338         entry = pte_mkhuge(entry);
 339
 340         return entry;
 341 }
 342
 343 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 344                                    unsigned long address, pte_t *ptep)
 345 {
 346         pte_t entry;
 347
 348         entry = pte_mkwrite(pte_mkdirty(*ptep));
 349         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 350                 update_mmu_cache(vma, address, entry);
 351                 lazy_mmu_prot_update(entry);
 352         }
 353 }
 354
 355
 356 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 357                             struct vm_area_struct *vma)
 358 {
 359         pte_t *src_pte, *dst_pte, entry;
 360         struct page *ptepage;
 361         unsigned long addr;
 362         int cow;
 363
 364         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 365
 366         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 367                 src_pte = huge_pte_offset(src, addr);
 368                 if (!src_pte)
 369                         continue;
 370                 dst_pte = huge_pte_alloc(dst, addr);
 371                 if (!dst_pte)
 372                         goto nomem;
 373                 spin_lock(&dst->page_table_lock);
 374                 spin_lock(&src->page_table_lock);
 375                 if (!pte_none(*src_pte)) {
 376                         if (cow)
 377                                 ptep_set_wrprotect(src, addr, src_pte);
 378                         entry = *src_pte;
 379                         ptepage = pte_page(entry);
 380                         get_page(ptepage);
 381                         set_huge_pte_at(dst, addr, dst_pte, entry);
 382                 }
 383                 spin_unlock(&src->page_table_lock);
 384                 spin_unlock(&dst->page_table_lock);
 385         }
 386         return 0;
 387
 388 nomem:
 389         return -ENOMEM;
 390 }
 391
 392 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 393                             unsigned long end)
 394 {
 395         struct mm_struct *mm = vma->vm_mm;
 396         unsigned long address;
 397         pte_t *ptep;
 398         pte_t pte;
 399         struct page *page;
 400         struct page *tmp;
 401         /*
 402          * A page gathering list, protected by per file i_mmap_lock. The
 403          * lock is used to avoid list corruption from multiple unmapping
 404          * of the same page since we are using page->lru.
 405          */
 406         LIST_HEAD(page_list);
 407
 408         WARN_ON(!is_vm_hugetlb_page(vma));
 409         BUG_ON(start & ~HPAGE_MASK);
 410         BUG_ON(end & ~HPAGE_MASK);
 411
 412         spin_lock(&mm->page_table_lock);
 413         for (address = start; address < end; address += HPAGE_SIZE) {
 414                 ptep = huge_pte_offset(mm, address);
 415                 if (!ptep)
 416                         continue;
 417
 418                 if (huge_pmd_unshare(mm, &address, ptep))
 419                         continue;
 420
 421                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 422                 if (pte_none(pte))
 423                         continue;
 424
 425                 page = pte_page(pte);
 426                 if (pte_dirty(pte))
 427                         set_page_dirty(page);
 428                 list_add(&page->lru, &page_list);
 429         }
 430         spin_unlock(&mm->page_table_lock);
 431         flush_tlb_range(vma, start, end);
 432         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 433                 list_del(&page->lru);
 434                 put_page(page);
 435         }
 436 }
 437
 438 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 439                           unsigned long end)
 440 {
 441         /*
 442          * It is undesirable to test vma->vm_file as it should be non-null
 443          * for valid hugetlb area. However, vm_file will be NULL in the error
 444          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 445          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 446          * to clean up. Since no pte has actually been setup, it is safe to
 447          * do nothing in this case.
 448          */
 449         if (vma->vm_file) {
 450                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 451                 __unmap_hugepage_range(vma, start, end);
 452                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 453         }
 454 }
 455
 456 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 457                         unsigned long address, pte_t *ptep, pte_t pte)
 458 {
 459         struct page *old_page, *new_page;
 460         int avoidcopy;
 461
 462         old_page = pte_page(pte);
 463
 464         /* If no-one else is actually using this page, avoid the copy
 465          * and just make the page writable */
 466         avoidcopy = (page_count(old_page) == 1);
 467         if (avoidcopy) {
 468                 set_huge_ptep_writable(vma, address, ptep);
 469                 return 0;
 470         }
 471
 472         page_cache_get(old_page);
 473         new_page = alloc_huge_page(vma, address);
 474
 475         if (!new_page) {
 476                 page_cache_release(old_page);
 477                 return VM_FAULT_OOM;
 478         }
 479
 480         spin_unlock(&mm->page_table_lock);
 481         copy_huge_page(new_page, old_page, address, vma);
 482         spin_lock(&mm->page_table_lock);
 483
 484         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 485         if (likely(pte_same(*ptep, pte))) {
 486                 /* Break COW */
 487                 set_huge_pte_at(mm, address, ptep,
 488                                 make_huge_pte(vma, new_page, 1));
 489                 /* Make the old page be freed below */
 490                 new_page = old_page;
 491         }
 492         page_cache_release(new_page);
 493         page_cache_release(old_page);
 494         return 0;
 495 }
 496
 497 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 498                         unsigned long address, pte_t *ptep, int write_access)
 499 {
 500         int ret = VM_FAULT_SIGBUS;
 501         unsigned long idx;
 502         unsigned long size;
 503         struct page *page;
 504         struct address_space *mapping;
 505         pte_t new_pte;
 506
 507         mapping = vma->vm_file->f_mapping;
 508         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 509                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 510
 511         /*
 512          * Use page lock to guard against racing truncation
 513          * before we get page_table_lock.
 514          */
 515 retry:
 516         page = find_lock_page(mapping, idx);
 517         if (!page) {
 518                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 519                 if (idx >= size)
 520                         goto out;
 521                 if (hugetlb_get_quota(mapping))
 522                         goto out;
 523                 page = alloc_huge_page(vma, address);
 524                 if (!page) {
 525                         hugetlb_put_quota(mapping);
 526                         ret = VM_FAULT_OOM;
 527                         goto out;
 528                 }
 529                 clear_huge_page(page, address);
 530
 531                 if (vma->vm_flags & VM_SHARED) {
 532                         int err;
 533
 534                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 535                         if (err) {
 536                                 put_page(page);
 537                                 hugetlb_put_quota(mapping);
 538                                 if (err == -EEXIST)
 539                                         goto retry;
 540                                 goto out;
 541                         }
 542                 } else
 543                         lock_page(page);
 544         }
 545
 546         spin_lock(&mm->page_table_lock);
 547         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 548         if (idx >= size)
 549                 goto backout;
 550
 551         ret = 0;
 552         if (!pte_none(*ptep))
 553                 goto backout;
 554
 555         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 556                                 && (vma->vm_flags & VM_SHARED)));
 557         set_huge_pte_at(mm, address, ptep, new_pte);
 558
 559         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 560                 /* Optimization, do the COW without a second fault */
 561                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 562         }
 563
 564         spin_unlock(&mm->page_table_lock);
 565         unlock_page(page);
 566 out:
 567         return ret;
 568
 569 backout:
 570         spin_unlock(&mm->page_table_lock);
 571         hugetlb_put_quota(mapping);
 572         unlock_page(page);
 573         put_page(page);
 574         goto out;
 575 }
 576
 577 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 578                         unsigned long address, int write_access)
 579 {
 580         pte_t *ptep;
 581         pte_t entry;
 582         int ret;
 583         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 584
 585         ptep = huge_pte_alloc(mm, address);
 586         if (!ptep)
 587                 return VM_FAULT_OOM;
 588
 589         /*
 590          * Serialize hugepage allocation and instantiation, so that we don't
 591          * get spurious allocation failures if two CPUs race to instantiate
 592          * the same page in the page cache.
 593          */
 594         mutex_lock(&hugetlb_instantiation_mutex);
 595         entry = *ptep;
 596         if (pte_none(entry)) {
 597                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 598                 mutex_unlock(&hugetlb_instantiation_mutex);
 599                 return ret;
 600         }
 601
 602         ret = 0;
 603
 604         spin_lock(&mm->page_table_lock);
 605         /* Check for a racing update before calling hugetlb_cow */
 606         if (likely(pte_same(entry, *ptep)))
 607                 if (write_access && !pte_write(entry))
 608                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 609         spin_unlock(&mm->page_table_lock);
 610         mutex_unlock(&hugetlb_instantiation_mutex);
 611
 612         return ret;
 613 }
 614
 615 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 616                         struct page **pages, struct vm_area_struct **vmas,
 617                         unsigned long *position, int *length, int i)
 618 {
 619         unsigned long pfn_offset;
 620         unsigned long vaddr = *position;
 621         int remainder = *length;
 622
 623         spin_lock(&mm->page_table_lock);
 624         while (vaddr < vma->vm_end && remainder) {
 625                 pte_t *pte;
 626                 struct page *page;
 627
 628                 /*
 629                  * Some archs (sparc64, sh*) have multiple pte_ts to
 630                  * each hugepage.  We have to make * sure we get the
 631                  * first, for the page indexing below to work.
 632                  */
 633                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 634
 635                 if (!pte || pte_none(*pte)) {
 636                         int ret;
 637
 638                         spin_unlock(&mm->page_table_lock);
 639                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 640                         spin_lock(&mm->page_table_lock);
 641                         if (!(ret & VM_FAULT_MAJOR))
 642                                 continue;
 643
 644                         remainder = 0;
 645                         if (!i)
 646                                 i = -EFAULT;
 647                         break;
 648                 }
 649
 650                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 651                 page = pte_page(*pte);
 652 same_page:
 653                 if (pages) {
 654                         get_page(page);
 655                         pages[i] = page + pfn_offset;
 656                 }
 657
 658                 if (vmas)
 659                         vmas[i] = vma;
 660
 661                 vaddr += PAGE_SIZE;
 662                 ++pfn_offset;
 663                 --remainder;
 664                 ++i;
 665                 if (vaddr < vma->vm_end && remainder &&
 666                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 667                         /*
 668                          * We use pfn_offset to avoid touching the pageframes
 669                          * of this compound page.
 670                          */
 671                         goto same_page;
 672                 }
 673         }
 674         spin_unlock(&mm->page_table_lock);
 675         *length = remainder;
 676         *position = vaddr;
 677
 678         return i;
 679 }
 680
 681 void hugetlb_change_protection(struct vm_area_struct *vma,
 682                 unsigned long address, unsigned long end, pgprot_t newprot)
 683 {
 684         struct mm_struct *mm = vma->vm_mm;
 685         unsigned long start = address;
 686         pte_t *ptep;
 687         pte_t pte;
 688
 689         BUG_ON(address >= end);
 690         flush_cache_range(vma, address, end);
 691
 692         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 693         spin_lock(&mm->page_table_lock);
 694         for (; address < end; address += HPAGE_SIZE) {
 695                 ptep = huge_pte_offset(mm, address);
 696                 if (!ptep)
 697                         continue;
 698                 if (huge_pmd_unshare(mm, &address, ptep))
 699                         continue;
 700                 if (!pte_none(*ptep)) {
 701                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 702                         pte = pte_mkhuge(pte_modify(pte, newprot));
 703                         set_huge_pte_at(mm, address, ptep, pte);
 704                         lazy_mmu_prot_update(pte);
 705                 }
 706         }
 707         spin_unlock(&mm->page_table_lock);
 708         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 709
 710         flush_tlb_range(vma, start, end);
 711 }
 712
 713 struct file_region {
 714         struct list_head link;
 715         long from;
 716         long to;
 717 };
 718
 719 static long region_add(struct list_head *head, long f, long t)
 720 {
 721         struct file_region *rg, *nrg, *trg;
 722
 723         /* Locate the region we are either in or before. */
 724         list_for_each_entry(rg, head, link)
 725                 if (f <= rg->to)
 726                         break;
 727
 728         /* Round our left edge to the current segment if it encloses us. */
 729         if (f > rg->from)
 730                 f = rg->from;
 731
 732         /* Check for and consume any regions we now overlap with. */
 733         nrg = rg;
 734         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 735                 if (&rg->link == head)
 736                         break;
 737                 if (rg->from > t)
 738                         break;
 739
 740                 /* If this area reaches higher then extend our area to
 741                  * include it completely.  If this is not the first area
 742                  * which we intend to reuse, free it. */
 743                 if (rg->to > t)
 744                         t = rg->to;
 745                 if (rg != nrg) {
 746                         list_del(&rg->link);
 747                         kfree(rg);
 748                 }
 749         }
 750         nrg->from = f;
 751         nrg->to = t;
 752         return 0;
 753 }
 754
 755 static long region_chg(struct list_head *head, long f, long t)
 756 {
 757         struct file_region *rg, *nrg;
 758         long chg = 0;
 759
 760         /* Locate the region we are before or in. */
 761         list_for_each_entry(rg, head, link)
 762                 if (f <= rg->to)
 763                         break;
 764
 765         /* If we are below the current region then a new region is required.
 766          * Subtle, allocate a new region at the position but make it zero
 767          * size such that we can guarentee to record the reservation. */
 768         if (&rg->link == head || t < rg->from) {
 769                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 770                 if (nrg == 0)
 771                         return -ENOMEM;
 772                 nrg->from = f;
 773                 nrg->to   = f;
 774                 INIT_LIST_HEAD(&nrg->link);
 775                 list_add(&nrg->link, rg->link.prev);
 776
 777                 return t - f;
 778         }
 779
 780         /* Round our left edge to the current segment if it encloses us. */
 781         if (f > rg->from)
 782                 f = rg->from;
 783         chg = t - f;
 784
 785         /* Check for and consume any regions we now overlap with. */
 786         list_for_each_entry(rg, rg->link.prev, link) {
 787                 if (&rg->link == head)
 788                         break;
 789                 if (rg->from > t)
 790                         return chg;
 791
 792                 /* We overlap with this area, if it extends futher than
 793                  * us then we must extend ourselves.  Account for its
 794                  * existing reservation. */
 795                 if (rg->to > t) {
 796                         chg += rg->to - t;
 797                         t = rg->to;
 798                 }
 799                 chg -= rg->to - rg->from;
 800         }
 801         return chg;
 802 }
 803
 804 static long region_truncate(struct list_head *head, long end)
 805 {
 806         struct file_region *rg, *trg;
 807         long chg = 0;
 808
 809         /* Locate the region we are either in or before. */
 810         list_for_each_entry(rg, head, link)
 811                 if (end <= rg->to)
 812                         break;
 813         if (&rg->link == head)
 814                 return 0;
 815
 816         /* If we are in the middle of a region then adjust it. */
 817         if (end > rg->from) {
 818                 chg = rg->to - end;
 819                 rg->to = end;
 820                 rg = list_entry(rg->link.next, typeof(*rg), link);
 821         }
 822
 823         /* Drop any remaining regions. */
 824         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 825                 if (&rg->link == head)
 826                         break;
 827                 chg += rg->to - rg->from;
 828                 list_del(&rg->link);
 829                 kfree(rg);
 830         }
 831         return chg;
 832 }
 833
 834 static int hugetlb_acct_memory(long delta)
 835 {
 836         int ret = -ENOMEM;
 837
 838         spin_lock(&hugetlb_lock);
 839         if ((delta + resv_huge_pages) <= free_huge_pages) {
 840                 resv_huge_pages += delta;
 841                 ret = 0;
 842         }
 843         spin_unlock(&hugetlb_lock);
 844         return ret;
 845 }
 846
 847 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
 848 {
 849         long ret, chg;
 850
 851         chg = region_chg(&inode->i_mapping->private_list, from, to);
 852         if (chg < 0)
 853                 return chg;
 854         /*
 855          * When cpuset is configured, it breaks the strict hugetlb page
 856          * reservation as the accounting is done on a global variable. Such
 857          * reservation is completely rubbish in the presence of cpuset because
 858          * the reservation is not checked against page availability for the
 859          * current cpuset. Application can still potentially OOM'ed by kernel
 860          * with lack of free htlb page in cpuset that the task is in.
 861          * Attempt to enforce strict accounting with cpuset is almost
 862          * impossible (or too ugly) because cpuset is too fluid that
 863          * task or memory node can be dynamically moved between cpusets.
 864          *
 865          * The change of semantics for shared hugetlb mapping with cpuset is
 866          * undesirable. However, in order to preserve some of the semantics,
 867          * we fall back to check against current free page availability as
 868          * a best attempt and hopefully to minimize the impact of changing
 869          * semantics that cpuset has.
 870          */
 871         if (chg > cpuset_mems_nr(free_huge_pages_node))
 872                 return -ENOMEM;
 873
 874         ret = hugetlb_acct_memory(chg);
 875         if (ret < 0)
 876                 return ret;
 877         region_add(&inode->i_mapping->private_list, from, to);
 878         return 0;
 879 }
 880
 881 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 882 {
 883         long chg = region_truncate(&inode->i_mapping->private_list, offset);
 884         hugetlb_acct_memory(freed - chg);
 885 }