SAFE public projects git trees. - safe/jmp/linux-2.6/blob - mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/highmem.h>
  12 #include <linux/nodemask.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/mempolicy.h>
  15 #include <linux/cpuset.h>
  16 #include <linux/mutex.h>
  17
  18 #include <asm/page.h>
  19 #include <asm/pgtable.h>
  20
  21 #include <linux/hugetlb.h>
  22 #include "internal.h"
  23
  24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
  26 static unsigned long surplus_huge_pages;
  27 unsigned long max_huge_pages;
  28 static struct list_head hugepage_freelists[MAX_NUMNODES];
  29 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  30 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  31 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
  32 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  33 unsigned long hugepages_treat_as_movable;
  34
  35 /*
  36  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  37  */
  38 static DEFINE_SPINLOCK(hugetlb_lock);
  39
  40 static void clear_huge_page(struct page *page, unsigned long addr)
  41 {
  42         int i;
  43
  44         might_sleep();
  45         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  46                 cond_resched();
  47                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
  48         }
  49 }
  50
  51 static void copy_huge_page(struct page *dst, struct page *src,
  52                            unsigned long addr, struct vm_area_struct *vma)
  53 {
  54         int i;
  55
  56         might_sleep();
  57         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  58                 cond_resched();
  59                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  60         }
  61 }
  62
  63 static void enqueue_huge_page(struct page *page)
  64 {
  65         int nid = page_to_nid(page);
  66         list_add(&page->lru, &hugepage_freelists[nid]);
  67         free_huge_pages++;
  68         free_huge_pages_node[nid]++;
  69 }
  70
  71 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  72                                 unsigned long address)
  73 {
  74         int nid;
  75         struct page *page = NULL;
  76         struct mempolicy *mpol;
  77         struct zonelist *zonelist = huge_zonelist(vma, address,
  78                                         htlb_alloc_mask, &mpol);
  79         struct zone **z;
  80
  81         for (z = zonelist->zones; *z; z++) {
  82                 nid = zone_to_nid(*z);
  83                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
  84                     !list_empty(&hugepage_freelists[nid])) {
  85                         page = list_entry(hugepage_freelists[nid].next,
  86                                           struct page, lru);
  87                         list_del(&page->lru);
  88                         free_huge_pages--;
  89                         free_huge_pages_node[nid]--;
  90                         break;
  91                 }
  92         }
  93         mpol_free(mpol);        /* unref if mpol !NULL */
  94         return page;
  95 }
  96
  97 static void update_and_free_page(struct page *page)
  98 {
  99         int i;
 100         nr_huge_pages--;
 101         nr_huge_pages_node[page_to_nid(page)]--;
 102         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 103                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 104                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 105                                 1 << PG_private | 1<< PG_writeback);
 106         }
 107         set_compound_page_dtor(page, NULL);
 108         set_page_refcounted(page);
 109         __free_pages(page, HUGETLB_PAGE_ORDER);
 110 }
 111
 112 static void free_huge_page(struct page *page)
 113 {
 114         int nid = page_to_nid(page);
 115
 116         BUG_ON(page_count(page));
 117         INIT_LIST_HEAD(&page->lru);
 118
 119         spin_lock(&hugetlb_lock);
 120         if (surplus_huge_pages_node[nid]) {
 121                 update_and_free_page(page);
 122                 surplus_huge_pages--;
 123                 surplus_huge_pages_node[nid]--;
 124         } else {
 125                 enqueue_huge_page(page);
 126         }
 127         spin_unlock(&hugetlb_lock);
 128 }
 129
 130 /*
 131  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 132  * balanced by operating on them in a round-robin fashion.
 133  * Returns 1 if an adjustment was made.
 134  */
 135 static int adjust_pool_surplus(int delta)
 136 {
 137         static int prev_nid;
 138         int nid = prev_nid;
 139         int ret = 0;
 140
 141         VM_BUG_ON(delta != -1 && delta != 1);
 142         do {
 143                 nid = next_node(nid, node_online_map);
 144                 if (nid == MAX_NUMNODES)
 145                         nid = first_node(node_online_map);
 146
 147                 /* To shrink on this node, there must be a surplus page */
 148                 if (delta < 0 && !surplus_huge_pages_node[nid])
 149                         continue;
 150                 /* Surplus cannot exceed the total number of pages */
 151                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 152                                                 nr_huge_pages_node[nid])
 153                         continue;
 154
 155                 surplus_huge_pages += delta;
 156                 surplus_huge_pages_node[nid] += delta;
 157                 ret = 1;
 158                 break;
 159         } while (nid != prev_nid);
 160
 161         prev_nid = nid;
 162         return ret;
 163 }
 164
 165 static int alloc_fresh_huge_page(void)
 166 {
 167         static int prev_nid;
 168         struct page *page;
 169         int nid;
 170
 171         /*
 172          * Copy static prev_nid to local nid, work on that, then copy it
 173          * back to prev_nid afterwards: otherwise there's a window in which
 174          * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node.
 175          * But we don't need to use a spin_lock here: it really doesn't
 176          * matter if occasionally a racer chooses the same nid as we do.
 177          */
 178         nid = next_node(prev_nid, node_online_map);
 179         if (nid == MAX_NUMNODES)
 180                 nid = first_node(node_online_map);
 181         prev_nid = nid;
 182
 183         page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 184                                         HUGETLB_PAGE_ORDER);
 185         if (page) {
 186                 set_compound_page_dtor(page, free_huge_page);
 187                 spin_lock(&hugetlb_lock);
 188                 nr_huge_pages++;
 189                 nr_huge_pages_node[page_to_nid(page)]++;
 190                 spin_unlock(&hugetlb_lock);
 191                 put_page(page); /* free it into the hugepage allocator */
 192                 return 1;
 193         }
 194         return 0;
 195 }
 196
 197 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 198                                                 unsigned long address)
 199 {
 200         struct page *page;
 201
 202         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 203                                         HUGETLB_PAGE_ORDER);
 204         if (page) {
 205                 set_compound_page_dtor(page, free_huge_page);
 206                 spin_lock(&hugetlb_lock);
 207                 nr_huge_pages++;
 208                 nr_huge_pages_node[page_to_nid(page)]++;
 209                 surplus_huge_pages++;
 210                 surplus_huge_pages_node[page_to_nid(page)]++;
 211                 spin_unlock(&hugetlb_lock);
 212         }
 213
 214         return page;
 215 }
 216
 217 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 218                                     unsigned long addr)
 219 {
 220         struct page *page = NULL;
 221
 222         spin_lock(&hugetlb_lock);
 223         if (vma->vm_flags & VM_MAYSHARE)
 224                 resv_huge_pages--;
 225         else if (free_huge_pages <= resv_huge_pages)
 226                 goto fail;
 227
 228         page = dequeue_huge_page(vma, addr);
 229         if (!page)
 230                 goto fail;
 231
 232         spin_unlock(&hugetlb_lock);
 233         set_page_refcounted(page);
 234         return page;
 235
 236 fail:
 237         if (vma->vm_flags & VM_MAYSHARE)
 238                 resv_huge_pages++;
 239         spin_unlock(&hugetlb_lock);
 240
 241         /*
 242          * Private mappings do not use reserved huge pages so the allocation
 243          * may have failed due to an undersized hugetlb pool.  Try to grab a
 244          * surplus huge page from the buddy allocator.
 245          */
 246         if (!(vma->vm_flags & VM_MAYSHARE))
 247                 page = alloc_buddy_huge_page(vma, addr);
 248
 249         return page;
 250 }
 251
 252 static int __init hugetlb_init(void)
 253 {
 254         unsigned long i;
 255
 256         if (HPAGE_SHIFT == 0)
 257                 return 0;
 258
 259         for (i = 0; i < MAX_NUMNODES; ++i)
 260                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 261
 262         for (i = 0; i < max_huge_pages; ++i) {
 263                 if (!alloc_fresh_huge_page())
 264                         break;
 265         }
 266         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 267         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 268         return 0;
 269 }
 270 module_init(hugetlb_init);
 271
 272 static int __init hugetlb_setup(char *s)
 273 {
 274         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 275                 max_huge_pages = 0;
 276         return 1;
 277 }
 278 __setup("hugepages=", hugetlb_setup);
 279
 280 static unsigned int cpuset_mems_nr(unsigned int *array)
 281 {
 282         int node;
 283         unsigned int nr = 0;
 284
 285         for_each_node_mask(node, cpuset_current_mems_allowed)
 286                 nr += array[node];
 287
 288         return nr;
 289 }
 290
 291 #ifdef CONFIG_SYSCTL
 292 #ifdef CONFIG_HIGHMEM
 293 static void try_to_free_low(unsigned long count)
 294 {
 295         int i;
 296
 297         for (i = 0; i < MAX_NUMNODES; ++i) {
 298                 struct page *page, *next;
 299                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 300                         if (PageHighMem(page))
 301                                 continue;
 302                         list_del(&page->lru);
 303                         update_and_free_page(page);
 304                         free_huge_pages--;
 305                         free_huge_pages_node[page_to_nid(page)]--;
 306                         if (count >= nr_huge_pages)
 307                                 return;
 308                 }
 309         }
 310 }
 311 #else
 312 static inline void try_to_free_low(unsigned long count)
 313 {
 314 }
 315 #endif
 316
 317 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 318 static unsigned long set_max_huge_pages(unsigned long count)
 319 {
 320         unsigned long min_count, ret;
 321
 322         /*
 323          * Increase the pool size
 324          * First take pages out of surplus state.  Then make up the
 325          * remaining difference by allocating fresh huge pages.
 326          */
 327         spin_lock(&hugetlb_lock);
 328         while (surplus_huge_pages && count > persistent_huge_pages) {
 329                 if (!adjust_pool_surplus(-1))
 330                         break;
 331         }
 332
 333         while (count > persistent_huge_pages) {
 334                 int ret;
 335                 /*
 336                  * If this allocation races such that we no longer need the
 337                  * page, free_huge_page will handle it by freeing the page
 338                  * and reducing the surplus.
 339                  */
 340                 spin_unlock(&hugetlb_lock);
 341                 ret = alloc_fresh_huge_page();
 342                 spin_lock(&hugetlb_lock);
 343                 if (!ret)
 344                         goto out;
 345
 346         }
 347         if (count >= persistent_huge_pages)
 348                 goto out;
 349
 350         /*
 351          * Decrease the pool size
 352          * First return free pages to the buddy allocator (being careful
 353          * to keep enough around to satisfy reservations).  Then place
 354          * pages into surplus state as needed so the pool will shrink
 355          * to the desired size as pages become free.
 356          */
 357         min_count = max(count, resv_huge_pages);
 358         try_to_free_low(min_count);
 359         while (min_count < persistent_huge_pages) {
 360                 struct page *page = dequeue_huge_page(NULL, 0);
 361                 if (!page)
 362                         break;
 363                 update_and_free_page(page);
 364         }
 365         while (count < persistent_huge_pages) {
 366                 if (!adjust_pool_surplus(1))
 367                         break;
 368         }
 369 out:
 370         ret = persistent_huge_pages;
 371         spin_unlock(&hugetlb_lock);
 372         return ret;
 373 }
 374
 375 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 376                            struct file *file, void __user *buffer,
 377                            size_t *length, loff_t *ppos)
 378 {
 379         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 380         max_huge_pages = set_max_huge_pages(max_huge_pages);
 381         return 0;
 382 }
 383
 384 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 385                         struct file *file, void __user *buffer,
 386                         size_t *length, loff_t *ppos)
 387 {
 388         proc_dointvec(table, write, file, buffer, length, ppos);
 389         if (hugepages_treat_as_movable)
 390                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 391         else
 392                 htlb_alloc_mask = GFP_HIGHUSER;
 393         return 0;
 394 }
 395
 396 #endif /* CONFIG_SYSCTL */
 397
 398 int hugetlb_report_meminfo(char *buf)
 399 {
 400         return sprintf(buf,
 401                         "HugePages_Total: %5lu\n"
 402                         "HugePages_Free:  %5lu\n"
 403                         "HugePages_Rsvd:  %5lu\n"
 404                         "HugePages_Surp:  %5lu\n"
 405                         "Hugepagesize:    %5lu kB\n",
 406                         nr_huge_pages,
 407                         free_huge_pages,
 408                         resv_huge_pages,
 409                         surplus_huge_pages,
 410                         HPAGE_SIZE/1024);
 411 }
 412
 413 int hugetlb_report_node_meminfo(int nid, char *buf)
 414 {
 415         return sprintf(buf,
 416                 "Node %d HugePages_Total: %5u\n"
 417                 "Node %d HugePages_Free:  %5u\n",
 418                 nid, nr_huge_pages_node[nid],
 419                 nid, free_huge_pages_node[nid]);
 420 }
 421
 422 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 423 unsigned long hugetlb_total_pages(void)
 424 {
 425         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 426 }
 427
 428 /*
 429  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 430  * handle_mm_fault() to try to instantiate regular-sized pages in the
 431  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 432  * this far.
 433  */
 434 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 435 {
 436         BUG();
 437         return 0;
 438 }
 439
 440 struct vm_operations_struct hugetlb_vm_ops = {
 441         .fault = hugetlb_vm_op_fault,
 442 };
 443
 444 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 445                                 int writable)
 446 {
 447         pte_t entry;
 448
 449         if (writable) {
 450                 entry =
 451                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 452         } else {
 453                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 454         }
 455         entry = pte_mkyoung(entry);
 456         entry = pte_mkhuge(entry);
 457
 458         return entry;
 459 }
 460
 461 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 462                                    unsigned long address, pte_t *ptep)
 463 {
 464         pte_t entry;
 465
 466         entry = pte_mkwrite(pte_mkdirty(*ptep));
 467         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 468                 update_mmu_cache(vma, address, entry);
 469         }
 470 }
 471
 472
 473 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 474                             struct vm_area_struct *vma)
 475 {
 476         pte_t *src_pte, *dst_pte, entry;
 477         struct page *ptepage;
 478         unsigned long addr;
 479         int cow;
 480
 481         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 482
 483         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 484                 src_pte = huge_pte_offset(src, addr);
 485                 if (!src_pte)
 486                         continue;
 487                 dst_pte = huge_pte_alloc(dst, addr);
 488                 if (!dst_pte)
 489                         goto nomem;
 490                 spin_lock(&dst->page_table_lock);
 491                 spin_lock(&src->page_table_lock);
 492                 if (!pte_none(*src_pte)) {
 493                         if (cow)
 494                                 ptep_set_wrprotect(src, addr, src_pte);
 495                         entry = *src_pte;
 496                         ptepage = pte_page(entry);
 497                         get_page(ptepage);
 498                         set_huge_pte_at(dst, addr, dst_pte, entry);
 499                 }
 500                 spin_unlock(&src->page_table_lock);
 501                 spin_unlock(&dst->page_table_lock);
 502         }
 503         return 0;
 504
 505 nomem:
 506         return -ENOMEM;
 507 }
 508
 509 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 510                             unsigned long end)
 511 {
 512         struct mm_struct *mm = vma->vm_mm;
 513         unsigned long address;
 514         pte_t *ptep;
 515         pte_t pte;
 516         struct page *page;
 517         struct page *tmp;
 518         /*
 519          * A page gathering list, protected by per file i_mmap_lock. The
 520          * lock is used to avoid list corruption from multiple unmapping
 521          * of the same page since we are using page->lru.
 522          */
 523         LIST_HEAD(page_list);
 524
 525         WARN_ON(!is_vm_hugetlb_page(vma));
 526         BUG_ON(start & ~HPAGE_MASK);
 527         BUG_ON(end & ~HPAGE_MASK);
 528
 529         spin_lock(&mm->page_table_lock);
 530         for (address = start; address < end; address += HPAGE_SIZE) {
 531                 ptep = huge_pte_offset(mm, address);
 532                 if (!ptep)
 533                         continue;
 534
 535                 if (huge_pmd_unshare(mm, &address, ptep))
 536                         continue;
 537
 538                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 539                 if (pte_none(pte))
 540                         continue;
 541
 542                 page = pte_page(pte);
 543                 if (pte_dirty(pte))
 544                         set_page_dirty(page);
 545                 list_add(&page->lru, &page_list);
 546         }
 547         spin_unlock(&mm->page_table_lock);
 548         flush_tlb_range(vma, start, end);
 549         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 550                 list_del(&page->lru);
 551                 put_page(page);
 552         }
 553 }
 554
 555 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 556                           unsigned long end)
 557 {
 558         /*
 559          * It is undesirable to test vma->vm_file as it should be non-null
 560          * for valid hugetlb area. However, vm_file will be NULL in the error
 561          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 562          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 563          * to clean up. Since no pte has actually been setup, it is safe to
 564          * do nothing in this case.
 565          */
 566         if (vma->vm_file) {
 567                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 568                 __unmap_hugepage_range(vma, start, end);
 569                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 570         }
 571 }
 572
 573 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 574                         unsigned long address, pte_t *ptep, pte_t pte)
 575 {
 576         struct page *old_page, *new_page;
 577         int avoidcopy;
 578
 579         old_page = pte_page(pte);
 580
 581         /* If no-one else is actually using this page, avoid the copy
 582          * and just make the page writable */
 583         avoidcopy = (page_count(old_page) == 1);
 584         if (avoidcopy) {
 585                 set_huge_ptep_writable(vma, address, ptep);
 586                 return 0;
 587         }
 588
 589         page_cache_get(old_page);
 590         new_page = alloc_huge_page(vma, address);
 591
 592         if (!new_page) {
 593                 page_cache_release(old_page);
 594                 return VM_FAULT_OOM;
 595         }
 596
 597         spin_unlock(&mm->page_table_lock);
 598         copy_huge_page(new_page, old_page, address, vma);
 599         spin_lock(&mm->page_table_lock);
 600
 601         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 602         if (likely(pte_same(*ptep, pte))) {
 603                 /* Break COW */
 604                 set_huge_pte_at(mm, address, ptep,
 605                                 make_huge_pte(vma, new_page, 1));
 606                 /* Make the old page be freed below */
 607                 new_page = old_page;
 608         }
 609         page_cache_release(new_page);
 610         page_cache_release(old_page);
 611         return 0;
 612 }
 613
 614 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 615                         unsigned long address, pte_t *ptep, int write_access)
 616 {
 617         int ret = VM_FAULT_SIGBUS;
 618         unsigned long idx;
 619         unsigned long size;
 620         struct page *page;
 621         struct address_space *mapping;
 622         pte_t new_pte;
 623
 624         mapping = vma->vm_file->f_mapping;
 625         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 626                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 627
 628         /*
 629          * Use page lock to guard against racing truncation
 630          * before we get page_table_lock.
 631          */
 632 retry:
 633         page = find_lock_page(mapping, idx);
 634         if (!page) {
 635                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 636                 if (idx >= size)
 637                         goto out;
 638                 if (hugetlb_get_quota(mapping))
 639                         goto out;
 640                 page = alloc_huge_page(vma, address);
 641                 if (!page) {
 642                         hugetlb_put_quota(mapping);
 643                         ret = VM_FAULT_OOM;
 644                         goto out;
 645                 }
 646                 clear_huge_page(page, address);
 647
 648                 if (vma->vm_flags & VM_SHARED) {
 649                         int err;
 650
 651                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 652                         if (err) {
 653                                 put_page(page);
 654                                 hugetlb_put_quota(mapping);
 655                                 if (err == -EEXIST)
 656                                         goto retry;
 657                                 goto out;
 658                         }
 659                 } else
 660                         lock_page(page);
 661         }
 662
 663         spin_lock(&mm->page_table_lock);
 664         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 665         if (idx >= size)
 666                 goto backout;
 667
 668         ret = 0;
 669         if (!pte_none(*ptep))
 670                 goto backout;
 671
 672         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 673                                 && (vma->vm_flags & VM_SHARED)));
 674         set_huge_pte_at(mm, address, ptep, new_pte);
 675
 676         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 677                 /* Optimization, do the COW without a second fault */
 678                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 679         }
 680
 681         spin_unlock(&mm->page_table_lock);
 682         unlock_page(page);
 683 out:
 684         return ret;
 685
 686 backout:
 687         spin_unlock(&mm->page_table_lock);
 688         hugetlb_put_quota(mapping);
 689         unlock_page(page);
 690         put_page(page);
 691         goto out;
 692 }
 693
 694 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 695                         unsigned long address, int write_access)
 696 {
 697         pte_t *ptep;
 698         pte_t entry;
 699         int ret;
 700         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 701
 702         ptep = huge_pte_alloc(mm, address);
 703         if (!ptep)
 704                 return VM_FAULT_OOM;
 705
 706         /*
 707          * Serialize hugepage allocation and instantiation, so that we don't
 708          * get spurious allocation failures if two CPUs race to instantiate
 709          * the same page in the page cache.
 710          */
 711         mutex_lock(&hugetlb_instantiation_mutex);
 712         entry = *ptep;
 713         if (pte_none(entry)) {
 714                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 715                 mutex_unlock(&hugetlb_instantiation_mutex);
 716                 return ret;
 717         }
 718
 719         ret = 0;
 720
 721         spin_lock(&mm->page_table_lock);
 722         /* Check for a racing update before calling hugetlb_cow */
 723         if (likely(pte_same(entry, *ptep)))
 724                 if (write_access && !pte_write(entry))
 725                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 726         spin_unlock(&mm->page_table_lock);
 727         mutex_unlock(&hugetlb_instantiation_mutex);
 728
 729         return ret;
 730 }
 731
 732 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 733                         struct page **pages, struct vm_area_struct **vmas,
 734                         unsigned long *position, int *length, int i)
 735 {
 736         unsigned long pfn_offset;
 737         unsigned long vaddr = *position;
 738         int remainder = *length;
 739
 740         spin_lock(&mm->page_table_lock);
 741         while (vaddr < vma->vm_end && remainder) {
 742                 pte_t *pte;
 743                 struct page *page;
 744
 745                 /*
 746                  * Some archs (sparc64, sh*) have multiple pte_ts to
 747                  * each hugepage.  We have to make * sure we get the
 748                  * first, for the page indexing below to work.
 749                  */
 750                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 751
 752                 if (!pte || pte_none(*pte)) {
 753                         int ret;
 754
 755                         spin_unlock(&mm->page_table_lock);
 756                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 757                         spin_lock(&mm->page_table_lock);
 758                         if (!(ret & VM_FAULT_ERROR))
 759                                 continue;
 760
 761                         remainder = 0;
 762                         if (!i)
 763                                 i = -EFAULT;
 764                         break;
 765                 }
 766
 767                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 768                 page = pte_page(*pte);
 769 same_page:
 770                 if (pages) {
 771                         get_page(page);
 772                         pages[i] = page + pfn_offset;
 773                 }
 774
 775                 if (vmas)
 776                         vmas[i] = vma;
 777
 778                 vaddr += PAGE_SIZE;
 779                 ++pfn_offset;
 780                 --remainder;
 781                 ++i;
 782                 if (vaddr < vma->vm_end && remainder &&
 783                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 784                         /*
 785                          * We use pfn_offset to avoid touching the pageframes
 786                          * of this compound page.
 787                          */
 788                         goto same_page;
 789                 }
 790         }
 791         spin_unlock(&mm->page_table_lock);
 792         *length = remainder;
 793         *position = vaddr;
 794
 795         return i;
 796 }
 797
 798 void hugetlb_change_protection(struct vm_area_struct *vma,
 799                 unsigned long address, unsigned long end, pgprot_t newprot)
 800 {
 801         struct mm_struct *mm = vma->vm_mm;
 802         unsigned long start = address;
 803         pte_t *ptep;
 804         pte_t pte;
 805
 806         BUG_ON(address >= end);
 807         flush_cache_range(vma, address, end);
 808
 809         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 810         spin_lock(&mm->page_table_lock);
 811         for (; address < end; address += HPAGE_SIZE) {
 812                 ptep = huge_pte_offset(mm, address);
 813                 if (!ptep)
 814                         continue;
 815                 if (huge_pmd_unshare(mm, &address, ptep))
 816                         continue;
 817                 if (!pte_none(*ptep)) {
 818                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 819                         pte = pte_mkhuge(pte_modify(pte, newprot));
 820                         set_huge_pte_at(mm, address, ptep, pte);
 821                 }
 822         }
 823         spin_unlock(&mm->page_table_lock);
 824         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 825
 826         flush_tlb_range(vma, start, end);
 827 }
 828
 829 struct file_region {
 830         struct list_head link;
 831         long from;
 832         long to;
 833 };
 834
 835 static long region_add(struct list_head *head, long f, long t)
 836 {
 837         struct file_region *rg, *nrg, *trg;
 838
 839         /* Locate the region we are either in or before. */
 840         list_for_each_entry(rg, head, link)
 841                 if (f <= rg->to)
 842                         break;
 843
 844         /* Round our left edge to the current segment if it encloses us. */
 845         if (f > rg->from)
 846                 f = rg->from;
 847
 848         /* Check for and consume any regions we now overlap with. */
 849         nrg = rg;
 850         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 851                 if (&rg->link == head)
 852                         break;
 853                 if (rg->from > t)
 854                         break;
 855
 856                 /* If this area reaches higher then extend our area to
 857                  * include it completely.  If this is not the first area
 858                  * which we intend to reuse, free it. */
 859                 if (rg->to > t)
 860                         t = rg->to;
 861                 if (rg != nrg) {
 862                         list_del(&rg->link);
 863                         kfree(rg);
 864                 }
 865         }
 866         nrg->from = f;
 867         nrg->to = t;
 868         return 0;
 869 }
 870
 871 static long region_chg(struct list_head *head, long f, long t)
 872 {
 873         struct file_region *rg, *nrg;
 874         long chg = 0;
 875
 876         /* Locate the region we are before or in. */
 877         list_for_each_entry(rg, head, link)
 878                 if (f <= rg->to)
 879                         break;
 880
 881         /* If we are below the current region then a new region is required.
 882          * Subtle, allocate a new region at the position but make it zero
 883          * size such that we can guarentee to record the reservation. */
 884         if (&rg->link == head || t < rg->from) {
 885                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 886                 if (nrg == 0)
 887                         return -ENOMEM;
 888                 nrg->from = f;
 889                 nrg->to   = f;
 890                 INIT_LIST_HEAD(&nrg->link);
 891                 list_add(&nrg->link, rg->link.prev);
 892
 893                 return t - f;
 894         }
 895
 896         /* Round our left edge to the current segment if it encloses us. */
 897         if (f > rg->from)
 898                 f = rg->from;
 899         chg = t - f;
 900
 901         /* Check for and consume any regions we now overlap with. */
 902         list_for_each_entry(rg, rg->link.prev, link) {
 903                 if (&rg->link == head)
 904                         break;
 905                 if (rg->from > t)
 906                         return chg;
 907
 908                 /* We overlap with this area, if it extends futher than
 909                  * us then we must extend ourselves.  Account for its
 910                  * existing reservation. */
 911                 if (rg->to > t) {
 912                         chg += rg->to - t;
 913                         t = rg->to;
 914                 }
 915                 chg -= rg->to - rg->from;
 916         }
 917         return chg;
 918 }
 919
 920 static long region_truncate(struct list_head *head, long end)
 921 {
 922         struct file_region *rg, *trg;
 923         long chg = 0;
 924
 925         /* Locate the region we are either in or before. */
 926         list_for_each_entry(rg, head, link)
 927                 if (end <= rg->to)
 928                         break;
 929         if (&rg->link == head)
 930                 return 0;
 931
 932         /* If we are in the middle of a region then adjust it. */
 933         if (end > rg->from) {
 934                 chg = rg->to - end;
 935                 rg->to = end;
 936                 rg = list_entry(rg->link.next, typeof(*rg), link);
 937         }
 938
 939         /* Drop any remaining regions. */
 940         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 941                 if (&rg->link == head)
 942                         break;
 943                 chg += rg->to - rg->from;
 944                 list_del(&rg->link);
 945                 kfree(rg);
 946         }
 947         return chg;
 948 }
 949
 950 static int hugetlb_acct_memory(long delta)
 951 {
 952         int ret = -ENOMEM;
 953
 954         spin_lock(&hugetlb_lock);
 955         if ((delta + resv_huge_pages) <= free_huge_pages) {
 956                 resv_huge_pages += delta;
 957                 ret = 0;
 958         }
 959         spin_unlock(&hugetlb_lock);
 960         return ret;
 961 }
 962
 963 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
 964 {
 965         long ret, chg;
 966
 967         chg = region_chg(&inode->i_mapping->private_list, from, to);
 968         if (chg < 0)
 969                 return chg;
 970         /*
 971          * When cpuset is configured, it breaks the strict hugetlb page
 972          * reservation as the accounting is done on a global variable. Such
 973          * reservation is completely rubbish in the presence of cpuset because
 974          * the reservation is not checked against page availability for the
 975          * current cpuset. Application can still potentially OOM'ed by kernel
 976          * with lack of free htlb page in cpuset that the task is in.
 977          * Attempt to enforce strict accounting with cpuset is almost
 978          * impossible (or too ugly) because cpuset is too fluid that
 979          * task or memory node can be dynamically moved between cpusets.
 980          *
 981          * The change of semantics for shared hugetlb mapping with cpuset is
 982          * undesirable. However, in order to preserve some of the semantics,
 983          * we fall back to check against current free page availability as
 984          * a best attempt and hopefully to minimize the impact of changing
 985          * semantics that cpuset has.
 986          */
 987         if (chg > cpuset_mems_nr(free_huge_pages_node))
 988                 return -ENOMEM;
 989
 990         ret = hugetlb_acct_memory(chg);
 991         if (ret < 0)
 992                 return ret;
 993         region_add(&inode->i_mapping->private_list, from, to);
 994         return 0;
 995 }
 996
 997 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 998 {
 999         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1000         hugetlb_acct_memory(freed - chg);
1001 }