mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/highmem.h>
  12 #include <linux/nodemask.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/mempolicy.h>
  15 #include <linux/cpuset.h>
  16 #include <linux/mutex.h>
  17
  18 #include <asm/page.h>
  19 #include <asm/pgtable.h>
  20
  21 #include <linux/hugetlb.h>
  22 #include "internal.h"
  23
  24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
  26 static unsigned long surplus_huge_pages;
  27 unsigned long max_huge_pages;
  28 static struct list_head hugepage_freelists[MAX_NUMNODES];
  29 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  30 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  31 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
  32 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  33 unsigned long hugepages_treat_as_movable;
  34 int hugetlb_dynamic_pool;
  35 static int hugetlb_next_nid;
  36
  37 /*
  38  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  39  */
  40 static DEFINE_SPINLOCK(hugetlb_lock);
  41
  42 static void clear_huge_page(struct page *page, unsigned long addr)
  43 {
  44         int i;
  45
  46         might_sleep();
  47         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  48                 cond_resched();
  49                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
  50         }
  51 }
  52
  53 static void copy_huge_page(struct page *dst, struct page *src,
  54                            unsigned long addr, struct vm_area_struct *vma)
  55 {
  56         int i;
  57
  58         might_sleep();
  59         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  60                 cond_resched();
  61                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  62         }
  63 }
  64
  65 static void enqueue_huge_page(struct page *page)
  66 {
  67         int nid = page_to_nid(page);
  68         list_add(&page->lru, &hugepage_freelists[nid]);
  69         free_huge_pages++;
  70         free_huge_pages_node[nid]++;
  71 }
  72
  73 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  74                                 unsigned long address)
  75 {
  76         int nid;
  77         struct page *page = NULL;
  78         struct mempolicy *mpol;
  79         struct zonelist *zonelist = huge_zonelist(vma, address,
  80                                         htlb_alloc_mask, &mpol);
  81         struct zone **z;
  82
  83         for (z = zonelist->zones; *z; z++) {
  84                 nid = zone_to_nid(*z);
  85                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
  86                     !list_empty(&hugepage_freelists[nid])) {
  87                         page = list_entry(hugepage_freelists[nid].next,
  88                                           struct page, lru);
  89                         list_del(&page->lru);
  90                         free_huge_pages--;
  91                         free_huge_pages_node[nid]--;
  92                         if (vma && vma->vm_flags & VM_MAYSHARE)
  93                                 resv_huge_pages--;
  94                         break;
  95                 }
  96         }
  97         mpol_free(mpol);        /* unref if mpol !NULL */
  98         return page;
  99 }
 100
 101 static void update_and_free_page(struct page *page)
 102 {
 103         int i;
 104         nr_huge_pages--;
 105         nr_huge_pages_node[page_to_nid(page)]--;
 106         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 107                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 108                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 109                                 1 << PG_private | 1<< PG_writeback);
 110         }
 111         set_compound_page_dtor(page, NULL);
 112         set_page_refcounted(page);
 113         __free_pages(page, HUGETLB_PAGE_ORDER);
 114 }
 115
 116 static void free_huge_page(struct page *page)
 117 {
 118         int nid = page_to_nid(page);
 119
 120         BUG_ON(page_count(page));
 121         INIT_LIST_HEAD(&page->lru);
 122
 123         spin_lock(&hugetlb_lock);
 124         if (surplus_huge_pages_node[nid]) {
 125                 update_and_free_page(page);
 126                 surplus_huge_pages--;
 127                 surplus_huge_pages_node[nid]--;
 128         } else {
 129                 enqueue_huge_page(page);
 130         }
 131         spin_unlock(&hugetlb_lock);
 132 }
 133
 134 /*
 135  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 136  * balanced by operating on them in a round-robin fashion.
 137  * Returns 1 if an adjustment was made.
 138  */
 139 static int adjust_pool_surplus(int delta)
 140 {
 141         static int prev_nid;
 142         int nid = prev_nid;
 143         int ret = 0;
 144
 145         VM_BUG_ON(delta != -1 && delta != 1);
 146         do {
 147                 nid = next_node(nid, node_online_map);
 148                 if (nid == MAX_NUMNODES)
 149                         nid = first_node(node_online_map);
 150
 151                 /* To shrink on this node, there must be a surplus page */
 152                 if (delta < 0 && !surplus_huge_pages_node[nid])
 153                         continue;
 154                 /* Surplus cannot exceed the total number of pages */
 155                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 156                                                 nr_huge_pages_node[nid])
 157                         continue;
 158
 159                 surplus_huge_pages += delta;
 160                 surplus_huge_pages_node[nid] += delta;
 161                 ret = 1;
 162                 break;
 163         } while (nid != prev_nid);
 164
 165         prev_nid = nid;
 166         return ret;
 167 }
 168
 169 static struct page *alloc_fresh_huge_page_node(int nid)
 170 {
 171         struct page *page;
 172
 173         page = alloc_pages_node(nid,
 174                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
 175                 HUGETLB_PAGE_ORDER);
 176         if (page) {
 177                 set_compound_page_dtor(page, free_huge_page);
 178                 spin_lock(&hugetlb_lock);
 179                 nr_huge_pages++;
 180                 nr_huge_pages_node[nid]++;
 181                 spin_unlock(&hugetlb_lock);
 182                 put_page(page); /* free it into the hugepage allocator */
 183         }
 184
 185         return page;
 186 }
 187
 188 static int alloc_fresh_huge_page(void)
 189 {
 190         struct page *page;
 191         int start_nid;
 192         int next_nid;
 193         int ret = 0;
 194
 195         start_nid = hugetlb_next_nid;
 196
 197         do {
 198                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
 199                 if (page)
 200                         ret = 1;
 201                 /*
 202                  * Use a helper variable to find the next node and then
 203                  * copy it back to hugetlb_next_nid afterwards:
 204                  * otherwise there's a window in which a racer might
 205                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 206                  * But we don't need to use a spin_lock here: it really
 207                  * doesn't matter if occasionally a racer chooses the
 208                  * same nid as we do.  Move nid forward in the mask even
 209                  * if we just successfully allocated a hugepage so that
 210                  * the next caller gets hugepages on the next node.
 211                  */
 212                 next_nid = next_node(hugetlb_next_nid, node_online_map);
 213                 if (next_nid == MAX_NUMNODES)
 214                         next_nid = first_node(node_online_map);
 215                 hugetlb_next_nid = next_nid;
 216         } while (!page && hugetlb_next_nid != start_nid);
 217
 218         return ret;
 219 }
 220
 221 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 222                                                 unsigned long address)
 223 {
 224         struct page *page;
 225
 226         /* Check if the dynamic pool is enabled */
 227         if (!hugetlb_dynamic_pool)
 228                 return NULL;
 229
 230         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 231                                         HUGETLB_PAGE_ORDER);
 232         if (page) {
 233                 set_compound_page_dtor(page, free_huge_page);
 234                 spin_lock(&hugetlb_lock);
 235                 nr_huge_pages++;
 236                 nr_huge_pages_node[page_to_nid(page)]++;
 237                 surplus_huge_pages++;
 238                 surplus_huge_pages_node[page_to_nid(page)]++;
 239                 spin_unlock(&hugetlb_lock);
 240         }
 241
 242         return page;
 243 }
 244
 245 /*
 246  * Increase the hugetlb pool such that it can accomodate a reservation
 247  * of size 'delta'.
 248  */
 249 static int gather_surplus_pages(int delta)
 250 {
 251         struct list_head surplus_list;
 252         struct page *page, *tmp;
 253         int ret, i;
 254         int needed, allocated;
 255
 256         needed = (resv_huge_pages + delta) - free_huge_pages;
 257         if (needed <= 0)
 258                 return 0;
 259
 260         allocated = 0;
 261         INIT_LIST_HEAD(&surplus_list);
 262
 263         ret = -ENOMEM;
 264 retry:
 265         spin_unlock(&hugetlb_lock);
 266         for (i = 0; i < needed; i++) {
 267                 page = alloc_buddy_huge_page(NULL, 0);
 268                 if (!page) {
 269                         /*
 270                          * We were not able to allocate enough pages to
 271                          * satisfy the entire reservation so we free what
 272                          * we've allocated so far.
 273                          */
 274                         spin_lock(&hugetlb_lock);
 275                         needed = 0;
 276                         goto free;
 277                 }
 278
 279                 list_add(&page->lru, &surplus_list);
 280         }
 281         allocated += needed;
 282
 283         /*
 284          * After retaking hugetlb_lock, we need to recalculate 'needed'
 285          * because either resv_huge_pages or free_huge_pages may have changed.
 286          */
 287         spin_lock(&hugetlb_lock);
 288         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
 289         if (needed > 0)
 290                 goto retry;
 291
 292         /*
 293          * The surplus_list now contains _at_least_ the number of extra pages
 294          * needed to accomodate the reservation.  Add the appropriate number
 295          * of pages to the hugetlb pool and free the extras back to the buddy
 296          * allocator.
 297          */
 298         needed += allocated;
 299         ret = 0;
 300 free:
 301         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 302                 list_del(&page->lru);
 303                 if ((--needed) >= 0)
 304                         enqueue_huge_page(page);
 305                 else {
 306                         /*
 307                          * Decrement the refcount and free the page using its
 308                          * destructor.  This must be done with hugetlb_lock
 309                          * unlocked which is safe because free_huge_page takes
 310                          * hugetlb_lock before deciding how to free the page.
 311                          */
 312                         spin_unlock(&hugetlb_lock);
 313                         put_page(page);
 314                         spin_lock(&hugetlb_lock);
 315                 }
 316         }
 317
 318         return ret;
 319 }
 320
 321 /*
 322  * When releasing a hugetlb pool reservation, any surplus pages that were
 323  * allocated to satisfy the reservation must be explicitly freed if they were
 324  * never used.
 325  */
 326 void return_unused_surplus_pages(unsigned long unused_resv_pages)
 327 {
 328         static int nid = -1;
 329         struct page *page;
 330         unsigned long nr_pages;
 331
 332         nr_pages = min(unused_resv_pages, surplus_huge_pages);
 333
 334         while (nr_pages) {
 335                 nid = next_node(nid, node_online_map);
 336                 if (nid == MAX_NUMNODES)
 337                         nid = first_node(node_online_map);
 338
 339                 if (!surplus_huge_pages_node[nid])
 340                         continue;
 341
 342                 if (!list_empty(&hugepage_freelists[nid])) {
 343                         page = list_entry(hugepage_freelists[nid].next,
 344                                           struct page, lru);
 345                         list_del(&page->lru);
 346                         update_and_free_page(page);
 347                         free_huge_pages--;
 348                         free_huge_pages_node[nid]--;
 349                         surplus_huge_pages--;
 350                         surplus_huge_pages_node[nid]--;
 351                         nr_pages--;
 352                 }
 353         }
 354 }
 355
 356 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 357                                     unsigned long addr)
 358 {
 359         struct page *page = NULL;
 360         int use_reserved_page = vma->vm_flags & VM_MAYSHARE;
 361
 362         spin_lock(&hugetlb_lock);
 363         if (!use_reserved_page && (free_huge_pages <= resv_huge_pages))
 364                 goto fail;
 365
 366         page = dequeue_huge_page(vma, addr);
 367         if (!page)
 368                 goto fail;
 369
 370         spin_unlock(&hugetlb_lock);
 371         set_page_refcounted(page);
 372         return page;
 373
 374 fail:
 375         spin_unlock(&hugetlb_lock);
 376
 377         /*
 378          * Private mappings do not use reserved huge pages so the allocation
 379          * may have failed due to an undersized hugetlb pool.  Try to grab a
 380          * surplus huge page from the buddy allocator.
 381          */
 382         if (!use_reserved_page)
 383                 page = alloc_buddy_huge_page(vma, addr);
 384
 385         return page;
 386 }
 387
 388 static int __init hugetlb_init(void)
 389 {
 390         unsigned long i;
 391
 392         if (HPAGE_SHIFT == 0)
 393                 return 0;
 394
 395         for (i = 0; i < MAX_NUMNODES; ++i)
 396                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 397
 398         hugetlb_next_nid = first_node(node_online_map);
 399
 400         for (i = 0; i < max_huge_pages; ++i) {
 401                 if (!alloc_fresh_huge_page())
 402                         break;
 403         }
 404         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 405         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 406         return 0;
 407 }
 408 module_init(hugetlb_init);
 409
 410 static int __init hugetlb_setup(char *s)
 411 {
 412         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 413                 max_huge_pages = 0;
 414         return 1;
 415 }
 416 __setup("hugepages=", hugetlb_setup);
 417
 418 static unsigned int cpuset_mems_nr(unsigned int *array)
 419 {
 420         int node;
 421         unsigned int nr = 0;
 422
 423         for_each_node_mask(node, cpuset_current_mems_allowed)
 424                 nr += array[node];
 425
 426         return nr;
 427 }
 428
 429 #ifdef CONFIG_SYSCTL
 430 #ifdef CONFIG_HIGHMEM
 431 static void try_to_free_low(unsigned long count)
 432 {
 433         int i;
 434
 435         for (i = 0; i < MAX_NUMNODES; ++i) {
 436                 struct page *page, *next;
 437                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 438                         if (count >= nr_huge_pages)
 439                                 return;
 440                         if (PageHighMem(page))
 441                                 continue;
 442                         list_del(&page->lru);
 443                         update_and_free_page(page);
 444                         free_huge_pages--;
 445                         free_huge_pages_node[page_to_nid(page)]--;
 446                 }
 447         }
 448 }
 449 #else
 450 static inline void try_to_free_low(unsigned long count)
 451 {
 452 }
 453 #endif
 454
 455 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 456 static unsigned long set_max_huge_pages(unsigned long count)
 457 {
 458         unsigned long min_count, ret;
 459
 460         /*
 461          * Increase the pool size
 462          * First take pages out of surplus state.  Then make up the
 463          * remaining difference by allocating fresh huge pages.
 464          */
 465         spin_lock(&hugetlb_lock);
 466         while (surplus_huge_pages && count > persistent_huge_pages) {
 467                 if (!adjust_pool_surplus(-1))
 468                         break;
 469         }
 470
 471         while (count > persistent_huge_pages) {
 472                 int ret;
 473                 /*
 474                  * If this allocation races such that we no longer need the
 475                  * page, free_huge_page will handle it by freeing the page
 476                  * and reducing the surplus.
 477                  */
 478                 spin_unlock(&hugetlb_lock);
 479                 ret = alloc_fresh_huge_page();
 480                 spin_lock(&hugetlb_lock);
 481                 if (!ret)
 482                         goto out;
 483
 484         }
 485
 486         /*
 487          * Decrease the pool size
 488          * First return free pages to the buddy allocator (being careful
 489          * to keep enough around to satisfy reservations).  Then place
 490          * pages into surplus state as needed so the pool will shrink
 491          * to the desired size as pages become free.
 492          */
 493         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
 494         min_count = max(count, min_count);
 495         try_to_free_low(min_count);
 496         while (min_count < persistent_huge_pages) {
 497                 struct page *page = dequeue_huge_page(NULL, 0);
 498                 if (!page)
 499                         break;
 500                 update_and_free_page(page);
 501         }
 502         while (count < persistent_huge_pages) {
 503                 if (!adjust_pool_surplus(1))
 504                         break;
 505         }
 506 out:
 507         ret = persistent_huge_pages;
 508         spin_unlock(&hugetlb_lock);
 509         return ret;
 510 }
 511
 512 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 513                            struct file *file, void __user *buffer,
 514                            size_t *length, loff_t *ppos)
 515 {
 516         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 517         max_huge_pages = set_max_huge_pages(max_huge_pages);
 518         return 0;
 519 }
 520
 521 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 522                         struct file *file, void __user *buffer,
 523                         size_t *length, loff_t *ppos)
 524 {
 525         proc_dointvec(table, write, file, buffer, length, ppos);
 526         if (hugepages_treat_as_movable)
 527                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 528         else
 529                 htlb_alloc_mask = GFP_HIGHUSER;
 530         return 0;
 531 }
 532
 533 #endif /* CONFIG_SYSCTL */
 534
 535 int hugetlb_report_meminfo(char *buf)
 536 {
 537         return sprintf(buf,
 538                         "HugePages_Total: %5lu\n"
 539                         "HugePages_Free:  %5lu\n"
 540                         "HugePages_Rsvd:  %5lu\n"
 541                         "HugePages_Surp:  %5lu\n"
 542                         "Hugepagesize:    %5lu kB\n",
 543                         nr_huge_pages,
 544                         free_huge_pages,
 545                         resv_huge_pages,
 546                         surplus_huge_pages,
 547                         HPAGE_SIZE/1024);
 548 }
 549
 550 int hugetlb_report_node_meminfo(int nid, char *buf)
 551 {
 552         return sprintf(buf,
 553                 "Node %d HugePages_Total: %5u\n"
 554                 "Node %d HugePages_Free:  %5u\n",
 555                 nid, nr_huge_pages_node[nid],
 556                 nid, free_huge_pages_node[nid]);
 557 }
 558
 559 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 560 unsigned long hugetlb_total_pages(void)
 561 {
 562         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 563 }
 564
 565 /*
 566  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 567  * handle_mm_fault() to try to instantiate regular-sized pages in the
 568  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 569  * this far.
 570  */
 571 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 572 {
 573         BUG();
 574         return 0;
 575 }
 576
 577 struct vm_operations_struct hugetlb_vm_ops = {
 578         .fault = hugetlb_vm_op_fault,
 579 };
 580
 581 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 582                                 int writable)
 583 {
 584         pte_t entry;
 585
 586         if (writable) {
 587                 entry =
 588                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 589         } else {
 590                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 591         }
 592         entry = pte_mkyoung(entry);
 593         entry = pte_mkhuge(entry);
 594
 595         return entry;
 596 }
 597
 598 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 599                                    unsigned long address, pte_t *ptep)
 600 {
 601         pte_t entry;
 602
 603         entry = pte_mkwrite(pte_mkdirty(*ptep));
 604         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 605                 update_mmu_cache(vma, address, entry);
 606         }
 607 }
 608
 609
 610 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 611                             struct vm_area_struct *vma)
 612 {
 613         pte_t *src_pte, *dst_pte, entry;
 614         struct page *ptepage;
 615         unsigned long addr;
 616         int cow;
 617
 618         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 619
 620         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 621                 src_pte = huge_pte_offset(src, addr);
 622                 if (!src_pte)
 623                         continue;
 624                 dst_pte = huge_pte_alloc(dst, addr);
 625                 if (!dst_pte)
 626                         goto nomem;
 627                 spin_lock(&dst->page_table_lock);
 628                 spin_lock(&src->page_table_lock);
 629                 if (!pte_none(*src_pte)) {
 630                         if (cow)
 631                                 ptep_set_wrprotect(src, addr, src_pte);
 632                         entry = *src_pte;
 633                         ptepage = pte_page(entry);
 634                         get_page(ptepage);
 635                         set_huge_pte_at(dst, addr, dst_pte, entry);
 636                 }
 637                 spin_unlock(&src->page_table_lock);
 638                 spin_unlock(&dst->page_table_lock);
 639         }
 640         return 0;
 641
 642 nomem:
 643         return -ENOMEM;
 644 }
 645
 646 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 647                             unsigned long end)
 648 {
 649         struct mm_struct *mm = vma->vm_mm;
 650         unsigned long address;
 651         pte_t *ptep;
 652         pte_t pte;
 653         struct page *page;
 654         struct page *tmp;
 655         /*
 656          * A page gathering list, protected by per file i_mmap_lock. The
 657          * lock is used to avoid list corruption from multiple unmapping
 658          * of the same page since we are using page->lru.
 659          */
 660         LIST_HEAD(page_list);
 661
 662         WARN_ON(!is_vm_hugetlb_page(vma));
 663         BUG_ON(start & ~HPAGE_MASK);
 664         BUG_ON(end & ~HPAGE_MASK);
 665
 666         spin_lock(&mm->page_table_lock);
 667         for (address = start; address < end; address += HPAGE_SIZE) {
 668                 ptep = huge_pte_offset(mm, address);
 669                 if (!ptep)
 670                         continue;
 671
 672                 if (huge_pmd_unshare(mm, &address, ptep))
 673                         continue;
 674
 675                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 676                 if (pte_none(pte))
 677                         continue;
 678
 679                 page = pte_page(pte);
 680                 if (pte_dirty(pte))
 681                         set_page_dirty(page);
 682                 list_add(&page->lru, &page_list);
 683         }
 684         spin_unlock(&mm->page_table_lock);
 685         flush_tlb_range(vma, start, end);
 686         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 687                 list_del(&page->lru);
 688                 put_page(page);
 689         }
 690 }
 691
 692 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 693                           unsigned long end)
 694 {
 695         /*
 696          * It is undesirable to test vma->vm_file as it should be non-null
 697          * for valid hugetlb area. However, vm_file will be NULL in the error
 698          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 699          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 700          * to clean up. Since no pte has actually been setup, it is safe to
 701          * do nothing in this case.
 702          */
 703         if (vma->vm_file) {
 704                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 705                 __unmap_hugepage_range(vma, start, end);
 706                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 707         }
 708 }
 709
 710 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 711                         unsigned long address, pte_t *ptep, pte_t pte)
 712 {
 713         struct page *old_page, *new_page;
 714         int avoidcopy;
 715
 716         old_page = pte_page(pte);
 717
 718         /* If no-one else is actually using this page, avoid the copy
 719          * and just make the page writable */
 720         avoidcopy = (page_count(old_page) == 1);
 721         if (avoidcopy) {
 722                 set_huge_ptep_writable(vma, address, ptep);
 723                 return 0;
 724         }
 725
 726         page_cache_get(old_page);
 727         new_page = alloc_huge_page(vma, address);
 728
 729         if (!new_page) {
 730                 page_cache_release(old_page);
 731                 return VM_FAULT_OOM;
 732         }
 733
 734         spin_unlock(&mm->page_table_lock);
 735         copy_huge_page(new_page, old_page, address, vma);
 736         spin_lock(&mm->page_table_lock);
 737
 738         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 739         if (likely(pte_same(*ptep, pte))) {
 740                 /* Break COW */
 741                 set_huge_pte_at(mm, address, ptep,
 742                                 make_huge_pte(vma, new_page, 1));
 743                 /* Make the old page be freed below */
 744                 new_page = old_page;
 745         }
 746         page_cache_release(new_page);
 747         page_cache_release(old_page);
 748         return 0;
 749 }
 750
 751 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 752                         unsigned long address, pte_t *ptep, int write_access)
 753 {
 754         int ret = VM_FAULT_SIGBUS;
 755         unsigned long idx;
 756         unsigned long size;
 757         struct page *page;
 758         struct address_space *mapping;
 759         pte_t new_pte;
 760
 761         mapping = vma->vm_file->f_mapping;
 762         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 763                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 764
 765         /*
 766          * Use page lock to guard against racing truncation
 767          * before we get page_table_lock.
 768          */
 769 retry:
 770         page = find_lock_page(mapping, idx);
 771         if (!page) {
 772                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 773                 if (idx >= size)
 774                         goto out;
 775                 if (hugetlb_get_quota(mapping))
 776                         goto out;
 777                 page = alloc_huge_page(vma, address);
 778                 if (!page) {
 779                         hugetlb_put_quota(mapping);
 780                         ret = VM_FAULT_OOM;
 781                         goto out;
 782                 }
 783                 clear_huge_page(page, address);
 784
 785                 if (vma->vm_flags & VM_SHARED) {
 786                         int err;
 787
 788                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 789                         if (err) {
 790                                 put_page(page);
 791                                 hugetlb_put_quota(mapping);
 792                                 if (err == -EEXIST)
 793                                         goto retry;
 794                                 goto out;
 795                         }
 796                 } else
 797                         lock_page(page);
 798         }
 799
 800         spin_lock(&mm->page_table_lock);
 801         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 802         if (idx >= size)
 803                 goto backout;
 804
 805         ret = 0;
 806         if (!pte_none(*ptep))
 807                 goto backout;
 808
 809         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 810                                 && (vma->vm_flags & VM_SHARED)));
 811         set_huge_pte_at(mm, address, ptep, new_pte);
 812
 813         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 814                 /* Optimization, do the COW without a second fault */
 815                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 816         }
 817
 818         spin_unlock(&mm->page_table_lock);
 819         unlock_page(page);
 820 out:
 821         return ret;
 822
 823 backout:
 824         spin_unlock(&mm->page_table_lock);
 825         hugetlb_put_quota(mapping);
 826         unlock_page(page);
 827         put_page(page);
 828         goto out;
 829 }
 830
 831 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 832                         unsigned long address, int write_access)
 833 {
 834         pte_t *ptep;
 835         pte_t entry;
 836         int ret;
 837         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 838
 839         ptep = huge_pte_alloc(mm, address);
 840         if (!ptep)
 841                 return VM_FAULT_OOM;
 842
 843         /*
 844          * Serialize hugepage allocation and instantiation, so that we don't
 845          * get spurious allocation failures if two CPUs race to instantiate
 846          * the same page in the page cache.
 847          */
 848         mutex_lock(&hugetlb_instantiation_mutex);
 849         entry = *ptep;
 850         if (pte_none(entry)) {
 851                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 852                 mutex_unlock(&hugetlb_instantiation_mutex);
 853                 return ret;
 854         }
 855
 856         ret = 0;
 857
 858         spin_lock(&mm->page_table_lock);
 859         /* Check for a racing update before calling hugetlb_cow */
 860         if (likely(pte_same(entry, *ptep)))
 861                 if (write_access && !pte_write(entry))
 862                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 863         spin_unlock(&mm->page_table_lock);
 864         mutex_unlock(&hugetlb_instantiation_mutex);
 865
 866         return ret;
 867 }
 868
 869 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 870                         struct page **pages, struct vm_area_struct **vmas,
 871                         unsigned long *position, int *length, int i)
 872 {
 873         unsigned long pfn_offset;
 874         unsigned long vaddr = *position;
 875         int remainder = *length;
 876
 877         spin_lock(&mm->page_table_lock);
 878         while (vaddr < vma->vm_end && remainder) {
 879                 pte_t *pte;
 880                 struct page *page;
 881
 882                 /*
 883                  * Some archs (sparc64, sh*) have multiple pte_ts to
 884                  * each hugepage.  We have to make * sure we get the
 885                  * first, for the page indexing below to work.
 886                  */
 887                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 888
 889                 if (!pte || pte_none(*pte)) {
 890                         int ret;
 891
 892                         spin_unlock(&mm->page_table_lock);
 893                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 894                         spin_lock(&mm->page_table_lock);
 895                         if (!(ret & VM_FAULT_ERROR))
 896                                 continue;
 897
 898                         remainder = 0;
 899                         if (!i)
 900                                 i = -EFAULT;
 901                         break;
 902                 }
 903
 904                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 905                 page = pte_page(*pte);
 906 same_page:
 907                 if (pages) {
 908                         get_page(page);
 909                         pages[i] = page + pfn_offset;
 910                 }
 911
 912                 if (vmas)
 913                         vmas[i] = vma;
 914
 915                 vaddr += PAGE_SIZE;
 916                 ++pfn_offset;
 917                 --remainder;
 918                 ++i;
 919                 if (vaddr < vma->vm_end && remainder &&
 920                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 921                         /*
 922                          * We use pfn_offset to avoid touching the pageframes
 923                          * of this compound page.
 924                          */
 925                         goto same_page;
 926                 }
 927         }
 928         spin_unlock(&mm->page_table_lock);
 929         *length = remainder;
 930         *position = vaddr;
 931
 932         return i;
 933 }
 934
 935 void hugetlb_change_protection(struct vm_area_struct *vma,
 936                 unsigned long address, unsigned long end, pgprot_t newprot)
 937 {
 938         struct mm_struct *mm = vma->vm_mm;
 939         unsigned long start = address;
 940         pte_t *ptep;
 941         pte_t pte;
 942
 943         BUG_ON(address >= end);
 944         flush_cache_range(vma, address, end);
 945
 946         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 947         spin_lock(&mm->page_table_lock);
 948         for (; address < end; address += HPAGE_SIZE) {
 949                 ptep = huge_pte_offset(mm, address);
 950                 if (!ptep)
 951                         continue;
 952                 if (huge_pmd_unshare(mm, &address, ptep))
 953                         continue;
 954                 if (!pte_none(*ptep)) {
 955                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 956                         pte = pte_mkhuge(pte_modify(pte, newprot));
 957                         set_huge_pte_at(mm, address, ptep, pte);
 958                 }
 959         }
 960         spin_unlock(&mm->page_table_lock);
 961         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 962
 963         flush_tlb_range(vma, start, end);
 964 }
 965
 966 struct file_region {
 967         struct list_head link;
 968         long from;
 969         long to;
 970 };
 971
 972 static long region_add(struct list_head *head, long f, long t)
 973 {
 974         struct file_region *rg, *nrg, *trg;
 975
 976         /* Locate the region we are either in or before. */
 977         list_for_each_entry(rg, head, link)
 978                 if (f <= rg->to)
 979                         break;
 980
 981         /* Round our left edge to the current segment if it encloses us. */
 982         if (f > rg->from)
 983                 f = rg->from;
 984
 985         /* Check for and consume any regions we now overlap with. */
 986         nrg = rg;
 987         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 988                 if (&rg->link == head)
 989                         break;
 990                 if (rg->from > t)
 991                         break;
 992
 993                 /* If this area reaches higher then extend our area to
 994                  * include it completely.  If this is not the first area
 995                  * which we intend to reuse, free it. */
 996                 if (rg->to > t)
 997                         t = rg->to;
 998                 if (rg != nrg) {
 999                         list_del(&rg->link);
1000                         kfree(rg);
1001                 }
1002         }
1003         nrg->from = f;
1004         nrg->to = t;
1005         return 0;
1006 }
1007
1008 static long region_chg(struct list_head *head, long f, long t)
1009 {
1010         struct file_region *rg, *nrg;
1011         long chg = 0;
1012
1013         /* Locate the region we are before or in. */
1014         list_for_each_entry(rg, head, link)
1015                 if (f <= rg->to)
1016                         break;
1017
1018         /* If we are below the current region then a new region is required.
1019          * Subtle, allocate a new region at the position but make it zero
1020          * size such that we can guarentee to record the reservation. */
1021         if (&rg->link == head || t < rg->from) {
1022                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1023                 if (nrg == 0)
1024                         return -ENOMEM;
1025                 nrg->from = f;
1026                 nrg->to   = f;
1027                 INIT_LIST_HEAD(&nrg->link);
1028                 list_add(&nrg->link, rg->link.prev);
1029
1030                 return t - f;
1031         }
1032
1033         /* Round our left edge to the current segment if it encloses us. */
1034         if (f > rg->from)
1035                 f = rg->from;
1036         chg = t - f;
1037
1038         /* Check for and consume any regions we now overlap with. */
1039         list_for_each_entry(rg, rg->link.prev, link) {
1040                 if (&rg->link == head)
1041                         break;
1042                 if (rg->from > t)
1043                         return chg;
1044
1045                 /* We overlap with this area, if it extends futher than
1046                  * us then we must extend ourselves.  Account for its
1047                  * existing reservation. */
1048                 if (rg->to > t) {
1049                         chg += rg->to - t;
1050                         t = rg->to;
1051                 }
1052                 chg -= rg->to - rg->from;
1053         }
1054         return chg;
1055 }
1056
1057 static long region_truncate(struct list_head *head, long end)
1058 {
1059         struct file_region *rg, *trg;
1060         long chg = 0;
1061
1062         /* Locate the region we are either in or before. */
1063         list_for_each_entry(rg, head, link)
1064                 if (end <= rg->to)
1065                         break;
1066         if (&rg->link == head)
1067                 return 0;
1068
1069         /* If we are in the middle of a region then adjust it. */
1070         if (end > rg->from) {
1071                 chg = rg->to - end;
1072                 rg->to = end;
1073                 rg = list_entry(rg->link.next, typeof(*rg), link);
1074         }
1075
1076         /* Drop any remaining regions. */
1077         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1078                 if (&rg->link == head)
1079                         break;
1080                 chg += rg->to - rg->from;
1081                 list_del(&rg->link);
1082                 kfree(rg);
1083         }
1084         return chg;
1085 }
1086
1087 static int hugetlb_acct_memory(long delta)
1088 {
1089         int ret = -ENOMEM;
1090
1091         spin_lock(&hugetlb_lock);
1092         /*
1093          * When cpuset is configured, it breaks the strict hugetlb page
1094          * reservation as the accounting is done on a global variable. Such
1095          * reservation is completely rubbish in the presence of cpuset because
1096          * the reservation is not checked against page availability for the
1097          * current cpuset. Application can still potentially OOM'ed by kernel
1098          * with lack of free htlb page in cpuset that the task is in.
1099          * Attempt to enforce strict accounting with cpuset is almost
1100          * impossible (or too ugly) because cpuset is too fluid that
1101          * task or memory node can be dynamically moved between cpusets.
1102          *
1103          * The change of semantics for shared hugetlb mapping with cpuset is
1104          * undesirable. However, in order to preserve some of the semantics,
1105          * we fall back to check against current free page availability as
1106          * a best attempt and hopefully to minimize the impact of changing
1107          * semantics that cpuset has.
1108          */
1109         if (delta > 0) {
1110                 if (gather_surplus_pages(delta) < 0)
1111                         goto out;
1112
1113                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1114                         goto out;
1115         }
1116
1117         ret = 0;
1118         resv_huge_pages += delta;
1119         if (delta < 0)
1120                 return_unused_surplus_pages((unsigned long) -delta);
1121
1122 out:
1123         spin_unlock(&hugetlb_lock);
1124         return ret;
1125 }
1126
1127 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1128 {
1129         long ret, chg;
1130
1131         chg = region_chg(&inode->i_mapping->private_list, from, to);
1132         if (chg < 0)
1133                 return chg;
1134
1135         ret = hugetlb_acct_memory(chg);
1136         if (ret < 0)
1137                 return ret;
1138         region_add(&inode->i_mapping->private_list, from, to);
1139         return 0;
1140 }
1141
1142 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1143 {
1144         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1145         hugetlb_acct_memory(freed - chg);
1146 }