git.oblomov.eu Git - 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 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         struct address_space *mapping;
 120
 121         mapping = (struct address_space *) page_private(page);
 122         BUG_ON(page_count(page));
 123         INIT_LIST_HEAD(&page->lru);
 124
 125         spin_lock(&hugetlb_lock);
 126         if (surplus_huge_pages_node[nid]) {
 127                 update_and_free_page(page);
 128                 surplus_huge_pages--;
 129                 surplus_huge_pages_node[nid]--;
 130         } else {
 131                 enqueue_huge_page(page);
 132         }
 133         spin_unlock(&hugetlb_lock);
 134         if (mapping)
 135                 hugetlb_put_quota(mapping, 1);
 136         set_page_private(page, 0);
 137 }
 138
 139 /*
 140  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 141  * balanced by operating on them in a round-robin fashion.
 142  * Returns 1 if an adjustment was made.
 143  */
 144 static int adjust_pool_surplus(int delta)
 145 {
 146         static int prev_nid;
 147         int nid = prev_nid;
 148         int ret = 0;
 149
 150         VM_BUG_ON(delta != -1 && delta != 1);
 151         do {
 152                 nid = next_node(nid, node_online_map);
 153                 if (nid == MAX_NUMNODES)
 154                         nid = first_node(node_online_map);
 155
 156                 /* To shrink on this node, there must be a surplus page */
 157                 if (delta < 0 && !surplus_huge_pages_node[nid])
 158                         continue;
 159                 /* Surplus cannot exceed the total number of pages */
 160                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 161                                                 nr_huge_pages_node[nid])
 162                         continue;
 163
 164                 surplus_huge_pages += delta;
 165                 surplus_huge_pages_node[nid] += delta;
 166                 ret = 1;
 167                 break;
 168         } while (nid != prev_nid);
 169
 170         prev_nid = nid;
 171         return ret;
 172 }
 173
 174 static struct page *alloc_fresh_huge_page_node(int nid)
 175 {
 176         struct page *page;
 177
 178         page = alloc_pages_node(nid,
 179                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
 180                 HUGETLB_PAGE_ORDER);
 181         if (page) {
 182                 set_compound_page_dtor(page, free_huge_page);
 183                 spin_lock(&hugetlb_lock);
 184                 nr_huge_pages++;
 185                 nr_huge_pages_node[nid]++;
 186                 spin_unlock(&hugetlb_lock);
 187                 put_page(page); /* free it into the hugepage allocator */
 188         }
 189
 190         return page;
 191 }
 192
 193 static int alloc_fresh_huge_page(void)
 194 {
 195         struct page *page;
 196         int start_nid;
 197         int next_nid;
 198         int ret = 0;
 199
 200         start_nid = hugetlb_next_nid;
 201
 202         do {
 203                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
 204                 if (page)
 205                         ret = 1;
 206                 /*
 207                  * Use a helper variable to find the next node and then
 208                  * copy it back to hugetlb_next_nid afterwards:
 209                  * otherwise there's a window in which a racer might
 210                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 211                  * But we don't need to use a spin_lock here: it really
 212                  * doesn't matter if occasionally a racer chooses the
 213                  * same nid as we do.  Move nid forward in the mask even
 214                  * if we just successfully allocated a hugepage so that
 215                  * the next caller gets hugepages on the next node.
 216                  */
 217                 next_nid = next_node(hugetlb_next_nid, node_online_map);
 218                 if (next_nid == MAX_NUMNODES)
 219                         next_nid = first_node(node_online_map);
 220                 hugetlb_next_nid = next_nid;
 221         } while (!page && hugetlb_next_nid != start_nid);
 222
 223         return ret;
 224 }
 225
 226 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 227                                                 unsigned long address)
 228 {
 229         struct page *page;
 230
 231         /* Check if the dynamic pool is enabled */
 232         if (!hugetlb_dynamic_pool)
 233                 return NULL;
 234
 235         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 236                                         HUGETLB_PAGE_ORDER);
 237         if (page) {
 238                 set_compound_page_dtor(page, free_huge_page);
 239                 spin_lock(&hugetlb_lock);
 240                 nr_huge_pages++;
 241                 nr_huge_pages_node[page_to_nid(page)]++;
 242                 surplus_huge_pages++;
 243                 surplus_huge_pages_node[page_to_nid(page)]++;
 244                 spin_unlock(&hugetlb_lock);
 245         }
 246
 247         return page;
 248 }
 249
 250 /*
 251  * Increase the hugetlb pool such that it can accomodate a reservation
 252  * of size 'delta'.
 253  */
 254 static int gather_surplus_pages(int delta)
 255 {
 256         struct list_head surplus_list;
 257         struct page *page, *tmp;
 258         int ret, i;
 259         int needed, allocated;
 260
 261         needed = (resv_huge_pages + delta) - free_huge_pages;
 262         if (needed <= 0)
 263                 return 0;
 264
 265         allocated = 0;
 266         INIT_LIST_HEAD(&surplus_list);
 267
 268         ret = -ENOMEM;
 269 retry:
 270         spin_unlock(&hugetlb_lock);
 271         for (i = 0; i < needed; i++) {
 272                 page = alloc_buddy_huge_page(NULL, 0);
 273                 if (!page) {
 274                         /*
 275                          * We were not able to allocate enough pages to
 276                          * satisfy the entire reservation so we free what
 277                          * we've allocated so far.
 278                          */
 279                         spin_lock(&hugetlb_lock);
 280                         needed = 0;
 281                         goto free;
 282                 }
 283
 284                 list_add(&page->lru, &surplus_list);
 285         }
 286         allocated += needed;
 287
 288         /*
 289          * After retaking hugetlb_lock, we need to recalculate 'needed'
 290          * because either resv_huge_pages or free_huge_pages may have changed.
 291          */
 292         spin_lock(&hugetlb_lock);
 293         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
 294         if (needed > 0)
 295                 goto retry;
 296
 297         /*
 298          * The surplus_list now contains _at_least_ the number of extra pages
 299          * needed to accomodate the reservation.  Add the appropriate number
 300          * of pages to the hugetlb pool and free the extras back to the buddy
 301          * allocator.
 302          */
 303         needed += allocated;
 304         ret = 0;
 305 free:
 306         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 307                 list_del(&page->lru);
 308                 if ((--needed) >= 0)
 309                         enqueue_huge_page(page);
 310                 else {
 311                         /*
 312                          * Decrement the refcount and free the page using its
 313                          * destructor.  This must be done with hugetlb_lock
 314                          * unlocked which is safe because free_huge_page takes
 315                          * hugetlb_lock before deciding how to free the page.
 316                          */
 317                         spin_unlock(&hugetlb_lock);
 318                         put_page(page);
 319                         spin_lock(&hugetlb_lock);
 320                 }
 321         }
 322
 323         return ret;
 324 }
 325
 326 /*
 327  * When releasing a hugetlb pool reservation, any surplus pages that were
 328  * allocated to satisfy the reservation must be explicitly freed if they were
 329  * never used.
 330  */
 331 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
 332 {
 333         static int nid = -1;
 334         struct page *page;
 335         unsigned long nr_pages;
 336
 337         nr_pages = min(unused_resv_pages, surplus_huge_pages);
 338
 339         while (nr_pages) {
 340                 nid = next_node(nid, node_online_map);
 341                 if (nid == MAX_NUMNODES)
 342                         nid = first_node(node_online_map);
 343
 344                 if (!surplus_huge_pages_node[nid])
 345                         continue;
 346
 347                 if (!list_empty(&hugepage_freelists[nid])) {
 348                         page = list_entry(hugepage_freelists[nid].next,
 349                                           struct page, lru);
 350                         list_del(&page->lru);
 351                         update_and_free_page(page);
 352                         free_huge_pages--;
 353                         free_huge_pages_node[nid]--;
 354                         surplus_huge_pages--;
 355                         surplus_huge_pages_node[nid]--;
 356                         nr_pages--;
 357                 }
 358         }
 359 }
 360
 361
 362 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
 363                                                 unsigned long addr)
 364 {
 365         struct page *page;
 366
 367         spin_lock(&hugetlb_lock);
 368         page = dequeue_huge_page(vma, addr);
 369         spin_unlock(&hugetlb_lock);
 370         return page ? page : ERR_PTR(-VM_FAULT_OOM);
 371 }
 372
 373 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
 374                                                 unsigned long addr)
 375 {
 376         struct page *page = NULL;
 377
 378         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
 379                 return ERR_PTR(-VM_FAULT_SIGBUS);
 380
 381         spin_lock(&hugetlb_lock);
 382         if (free_huge_pages > resv_huge_pages)
 383                 page = dequeue_huge_page(vma, addr);
 384         spin_unlock(&hugetlb_lock);
 385         if (!page)
 386                 page = alloc_buddy_huge_page(vma, addr);
 387         return page ? page : ERR_PTR(-VM_FAULT_OOM);
 388 }
 389
 390 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 391                                     unsigned long addr)
 392 {
 393         struct page *page;
 394         struct address_space *mapping = vma->vm_file->f_mapping;
 395
 396         if (vma->vm_flags & VM_MAYSHARE)
 397                 page = alloc_huge_page_shared(vma, addr);
 398         else
 399                 page = alloc_huge_page_private(vma, addr);
 400
 401         if (!IS_ERR(page)) {
 402                 set_page_refcounted(page);
 403                 set_page_private(page, (unsigned long) mapping);
 404         }
 405         return page;
 406 }
 407
 408 static int __init hugetlb_init(void)
 409 {
 410         unsigned long i;
 411
 412         if (HPAGE_SHIFT == 0)
 413                 return 0;
 414
 415         for (i = 0; i < MAX_NUMNODES; ++i)
 416                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 417
 418         hugetlb_next_nid = first_node(node_online_map);
 419
 420         for (i = 0; i < max_huge_pages; ++i) {
 421                 if (!alloc_fresh_huge_page())
 422                         break;
 423         }
 424         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 425         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 426         return 0;
 427 }
 428 module_init(hugetlb_init);
 429
 430 static int __init hugetlb_setup(char *s)
 431 {
 432         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 433                 max_huge_pages = 0;
 434         return 1;
 435 }
 436 __setup("hugepages=", hugetlb_setup);
 437
 438 static unsigned int cpuset_mems_nr(unsigned int *array)
 439 {
 440         int node;
 441         unsigned int nr = 0;
 442
 443         for_each_node_mask(node, cpuset_current_mems_allowed)
 444                 nr += array[node];
 445
 446         return nr;
 447 }
 448
 449 #ifdef CONFIG_SYSCTL
 450 #ifdef CONFIG_HIGHMEM
 451 static void try_to_free_low(unsigned long count)
 452 {
 453         int i;
 454
 455         for (i = 0; i < MAX_NUMNODES; ++i) {
 456                 struct page *page, *next;
 457                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 458                         if (count >= nr_huge_pages)
 459                                 return;
 460                         if (PageHighMem(page))
 461                                 continue;
 462                         list_del(&page->lru);
 463                         update_and_free_page(page);
 464                         free_huge_pages--;
 465                         free_huge_pages_node[page_to_nid(page)]--;
 466                 }
 467         }
 468 }
 469 #else
 470 static inline void try_to_free_low(unsigned long count)
 471 {
 472 }
 473 #endif
 474
 475 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 476 static unsigned long set_max_huge_pages(unsigned long count)
 477 {
 478         unsigned long min_count, ret;
 479
 480         /*
 481          * Increase the pool size
 482          * First take pages out of surplus state.  Then make up the
 483          * remaining difference by allocating fresh huge pages.
 484          */
 485         spin_lock(&hugetlb_lock);
 486         while (surplus_huge_pages && count > persistent_huge_pages) {
 487                 if (!adjust_pool_surplus(-1))
 488                         break;
 489         }
 490
 491         while (count > persistent_huge_pages) {
 492                 int ret;
 493                 /*
 494                  * If this allocation races such that we no longer need the
 495                  * page, free_huge_page will handle it by freeing the page
 496                  * and reducing the surplus.
 497                  */
 498                 spin_unlock(&hugetlb_lock);
 499                 ret = alloc_fresh_huge_page();
 500                 spin_lock(&hugetlb_lock);
 501                 if (!ret)
 502                         goto out;
 503
 504         }
 505
 506         /*
 507          * Decrease the pool size
 508          * First return free pages to the buddy allocator (being careful
 509          * to keep enough around to satisfy reservations).  Then place
 510          * pages into surplus state as needed so the pool will shrink
 511          * to the desired size as pages become free.
 512          */
 513         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
 514         min_count = max(count, min_count);
 515         try_to_free_low(min_count);
 516         while (min_count < persistent_huge_pages) {
 517                 struct page *page = dequeue_huge_page(NULL, 0);
 518                 if (!page)
 519                         break;
 520                 update_and_free_page(page);
 521         }
 522         while (count < persistent_huge_pages) {
 523                 if (!adjust_pool_surplus(1))
 524                         break;
 525         }
 526 out:
 527         ret = persistent_huge_pages;
 528         spin_unlock(&hugetlb_lock);
 529         return ret;
 530 }
 531
 532 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 533                            struct file *file, void __user *buffer,
 534                            size_t *length, loff_t *ppos)
 535 {
 536         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 537         max_huge_pages = set_max_huge_pages(max_huge_pages);
 538         return 0;
 539 }
 540
 541 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 542                         struct file *file, void __user *buffer,
 543                         size_t *length, loff_t *ppos)
 544 {
 545         proc_dointvec(table, write, file, buffer, length, ppos);
 546         if (hugepages_treat_as_movable)
 547                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 548         else
 549                 htlb_alloc_mask = GFP_HIGHUSER;
 550         return 0;
 551 }
 552
 553 #endif /* CONFIG_SYSCTL */
 554
 555 int hugetlb_report_meminfo(char *buf)
 556 {
 557         return sprintf(buf,
 558                         "HugePages_Total: %5lu\n"
 559                         "HugePages_Free:  %5lu\n"
 560                         "HugePages_Rsvd:  %5lu\n"
 561                         "HugePages_Surp:  %5lu\n"
 562                         "Hugepagesize:    %5lu kB\n",
 563                         nr_huge_pages,
 564                         free_huge_pages,
 565                         resv_huge_pages,
 566                         surplus_huge_pages,
 567                         HPAGE_SIZE/1024);
 568 }
 569
 570 int hugetlb_report_node_meminfo(int nid, char *buf)
 571 {
 572         return sprintf(buf,
 573                 "Node %d HugePages_Total: %5u\n"
 574                 "Node %d HugePages_Free:  %5u\n",
 575                 nid, nr_huge_pages_node[nid],
 576                 nid, free_huge_pages_node[nid]);
 577 }
 578
 579 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 580 unsigned long hugetlb_total_pages(void)
 581 {
 582         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 583 }
 584
 585 /*
 586  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 587  * handle_mm_fault() to try to instantiate regular-sized pages in the
 588  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 589  * this far.
 590  */
 591 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 592 {
 593         BUG();
 594         return 0;
 595 }
 596
 597 struct vm_operations_struct hugetlb_vm_ops = {
 598         .fault = hugetlb_vm_op_fault,
 599 };
 600
 601 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 602                                 int writable)
 603 {
 604         pte_t entry;
 605
 606         if (writable) {
 607                 entry =
 608                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 609         } else {
 610                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 611         }
 612         entry = pte_mkyoung(entry);
 613         entry = pte_mkhuge(entry);
 614
 615         return entry;
 616 }
 617
 618 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 619                                    unsigned long address, pte_t *ptep)
 620 {
 621         pte_t entry;
 622
 623         entry = pte_mkwrite(pte_mkdirty(*ptep));
 624         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 625                 update_mmu_cache(vma, address, entry);
 626         }
 627 }
 628
 629
 630 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 631                             struct vm_area_struct *vma)
 632 {
 633         pte_t *src_pte, *dst_pte, entry;
 634         struct page *ptepage;
 635         unsigned long addr;
 636         int cow;
 637
 638         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 639
 640         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 641                 src_pte = huge_pte_offset(src, addr);
 642                 if (!src_pte)
 643                         continue;
 644                 dst_pte = huge_pte_alloc(dst, addr);
 645                 if (!dst_pte)
 646                         goto nomem;
 647                 spin_lock(&dst->page_table_lock);
 648                 spin_lock(&src->page_table_lock);
 649                 if (!pte_none(*src_pte)) {
 650                         if (cow)
 651                                 ptep_set_wrprotect(src, addr, src_pte);
 652                         entry = *src_pte;
 653                         ptepage = pte_page(entry);
 654                         get_page(ptepage);
 655                         set_huge_pte_at(dst, addr, dst_pte, entry);
 656                 }
 657                 spin_unlock(&src->page_table_lock);
 658                 spin_unlock(&dst->page_table_lock);
 659         }
 660         return 0;
 661
 662 nomem:
 663         return -ENOMEM;
 664 }
 665
 666 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 667                             unsigned long end)
 668 {
 669         struct mm_struct *mm = vma->vm_mm;
 670         unsigned long address;
 671         pte_t *ptep;
 672         pte_t pte;
 673         struct page *page;
 674         struct page *tmp;
 675         /*
 676          * A page gathering list, protected by per file i_mmap_lock. The
 677          * lock is used to avoid list corruption from multiple unmapping
 678          * of the same page since we are using page->lru.
 679          */
 680         LIST_HEAD(page_list);
 681
 682         WARN_ON(!is_vm_hugetlb_page(vma));
 683         BUG_ON(start & ~HPAGE_MASK);
 684         BUG_ON(end & ~HPAGE_MASK);
 685
 686         spin_lock(&mm->page_table_lock);
 687         for (address = start; address < end; address += HPAGE_SIZE) {
 688                 ptep = huge_pte_offset(mm, address);
 689                 if (!ptep)
 690                         continue;
 691
 692                 if (huge_pmd_unshare(mm, &address, ptep))
 693                         continue;
 694
 695                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 696                 if (pte_none(pte))
 697                         continue;
 698
 699                 page = pte_page(pte);
 700                 if (pte_dirty(pte))
 701                         set_page_dirty(page);
 702                 list_add(&page->lru, &page_list);
 703         }
 704         spin_unlock(&mm->page_table_lock);
 705         flush_tlb_range(vma, start, end);
 706         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 707                 list_del(&page->lru);
 708                 put_page(page);
 709         }
 710 }
 711
 712 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 713                           unsigned long end)
 714 {
 715         /*
 716          * It is undesirable to test vma->vm_file as it should be non-null
 717          * for valid hugetlb area. However, vm_file will be NULL in the error
 718          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 719          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 720          * to clean up. Since no pte has actually been setup, it is safe to
 721          * do nothing in this case.
 722          */
 723         if (vma->vm_file) {
 724                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 725                 __unmap_hugepage_range(vma, start, end);
 726                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 727         }
 728 }
 729
 730 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 731                         unsigned long address, pte_t *ptep, pte_t pte)
 732 {
 733         struct page *old_page, *new_page;
 734         int avoidcopy;
 735
 736         old_page = pte_page(pte);
 737
 738         /* If no-one else is actually using this page, avoid the copy
 739          * and just make the page writable */
 740         avoidcopy = (page_count(old_page) == 1);
 741         if (avoidcopy) {
 742                 set_huge_ptep_writable(vma, address, ptep);
 743                 return 0;
 744         }
 745
 746         page_cache_get(old_page);
 747         new_page = alloc_huge_page(vma, address);
 748
 749         if (IS_ERR(new_page)) {
 750                 page_cache_release(old_page);
 751                 return -PTR_ERR(new_page);
 752         }
 753
 754         spin_unlock(&mm->page_table_lock);
 755         copy_huge_page(new_page, old_page, address, vma);
 756         spin_lock(&mm->page_table_lock);
 757
 758         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 759         if (likely(pte_same(*ptep, pte))) {
 760                 /* Break COW */
 761                 set_huge_pte_at(mm, address, ptep,
 762                                 make_huge_pte(vma, new_page, 1));
 763                 /* Make the old page be freed below */
 764                 new_page = old_page;
 765         }
 766         page_cache_release(new_page);
 767         page_cache_release(old_page);
 768         return 0;
 769 }
 770
 771 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 772                         unsigned long address, pte_t *ptep, int write_access)
 773 {
 774         int ret = VM_FAULT_SIGBUS;
 775         unsigned long idx;
 776         unsigned long size;
 777         struct page *page;
 778         struct address_space *mapping;
 779         pte_t new_pte;
 780
 781         mapping = vma->vm_file->f_mapping;
 782         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 783                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 784
 785         /*
 786          * Use page lock to guard against racing truncation
 787          * before we get page_table_lock.
 788          */
 789 retry:
 790         page = find_lock_page(mapping, idx);
 791         if (!page) {
 792                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 793                 if (idx >= size)
 794                         goto out;
 795                 page = alloc_huge_page(vma, address);
 796                 if (IS_ERR(page)) {
 797                         ret = -PTR_ERR(page);
 798                         goto out;
 799                 }
 800                 clear_huge_page(page, address);
 801
 802                 if (vma->vm_flags & VM_SHARED) {
 803                         int err;
 804                         struct inode *inode = mapping->host;
 805
 806                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 807                         if (err) {
 808                                 put_page(page);
 809                                 if (err == -EEXIST)
 810                                         goto retry;
 811                                 goto out;
 812                         }
 813
 814                         spin_lock(&inode->i_lock);
 815                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
 816                         spin_unlock(&inode->i_lock);
 817                 } else
 818                         lock_page(page);
 819         }
 820
 821         spin_lock(&mm->page_table_lock);
 822         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 823         if (idx >= size)
 824                 goto backout;
 825
 826         ret = 0;
 827         if (!pte_none(*ptep))
 828                 goto backout;
 829
 830         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 831                                 && (vma->vm_flags & VM_SHARED)));
 832         set_huge_pte_at(mm, address, ptep, new_pte);
 833
 834         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 835                 /* Optimization, do the COW without a second fault */
 836                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 837         }
 838
 839         spin_unlock(&mm->page_table_lock);
 840         unlock_page(page);
 841 out:
 842         return ret;
 843
 844 backout:
 845         spin_unlock(&mm->page_table_lock);
 846         unlock_page(page);
 847         put_page(page);
 848         goto out;
 849 }
 850
 851 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 852                         unsigned long address, int write_access)
 853 {
 854         pte_t *ptep;
 855         pte_t entry;
 856         int ret;
 857         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 858
 859         ptep = huge_pte_alloc(mm, address);
 860         if (!ptep)
 861                 return VM_FAULT_OOM;
 862
 863         /*
 864          * Serialize hugepage allocation and instantiation, so that we don't
 865          * get spurious allocation failures if two CPUs race to instantiate
 866          * the same page in the page cache.
 867          */
 868         mutex_lock(&hugetlb_instantiation_mutex);
 869         entry = *ptep;
 870         if (pte_none(entry)) {
 871                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 872                 mutex_unlock(&hugetlb_instantiation_mutex);
 873                 return ret;
 874         }
 875
 876         ret = 0;
 877
 878         spin_lock(&mm->page_table_lock);
 879         /* Check for a racing update before calling hugetlb_cow */
 880         if (likely(pte_same(entry, *ptep)))
 881                 if (write_access && !pte_write(entry))
 882                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 883         spin_unlock(&mm->page_table_lock);
 884         mutex_unlock(&hugetlb_instantiation_mutex);
 885
 886         return ret;
 887 }
 888
 889 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 890                         struct page **pages, struct vm_area_struct **vmas,
 891                         unsigned long *position, int *length, int i,
 892                         int write)
 893 {
 894         unsigned long pfn_offset;
 895         unsigned long vaddr = *position;
 896         int remainder = *length;
 897
 898         spin_lock(&mm->page_table_lock);
 899         while (vaddr < vma->vm_end && remainder) {
 900                 pte_t *pte;
 901                 struct page *page;
 902
 903                 /*
 904                  * Some archs (sparc64, sh*) have multiple pte_ts to
 905                  * each hugepage.  We have to make * sure we get the
 906                  * first, for the page indexing below to work.
 907                  */
 908                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 909
 910                 if (!pte || pte_none(*pte)) {
 911                         int ret;
 912
 913                         spin_unlock(&mm->page_table_lock);
 914                         ret = hugetlb_fault(mm, vma, vaddr, write);
 915                         spin_lock(&mm->page_table_lock);
 916                         if (!(ret & VM_FAULT_ERROR))
 917                                 continue;
 918
 919                         remainder = 0;
 920                         if (!i)
 921                                 i = -EFAULT;
 922                         break;
 923                 }
 924
 925                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
 926                 page = pte_page(*pte);
 927 same_page:
 928                 if (pages) {
 929                         get_page(page);
 930                         pages[i] = page + pfn_offset;
 931                 }
 932
 933                 if (vmas)
 934                         vmas[i] = vma;
 935
 936                 vaddr += PAGE_SIZE;
 937                 ++pfn_offset;
 938                 --remainder;
 939                 ++i;
 940                 if (vaddr < vma->vm_end && remainder &&
 941                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
 942                         /*
 943                          * We use pfn_offset to avoid touching the pageframes
 944                          * of this compound page.
 945                          */
 946                         goto same_page;
 947                 }
 948         }
 949         spin_unlock(&mm->page_table_lock);
 950         *length = remainder;
 951         *position = vaddr;
 952
 953         return i;
 954 }
 955
 956 void hugetlb_change_protection(struct vm_area_struct *vma,
 957                 unsigned long address, unsigned long end, pgprot_t newprot)
 958 {
 959         struct mm_struct *mm = vma->vm_mm;
 960         unsigned long start = address;
 961         pte_t *ptep;
 962         pte_t pte;
 963
 964         BUG_ON(address >= end);
 965         flush_cache_range(vma, address, end);
 966
 967         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 968         spin_lock(&mm->page_table_lock);
 969         for (; address < end; address += HPAGE_SIZE) {
 970                 ptep = huge_pte_offset(mm, address);
 971                 if (!ptep)
 972                         continue;
 973                 if (huge_pmd_unshare(mm, &address, ptep))
 974                         continue;
 975                 if (!pte_none(*ptep)) {
 976                         pte = huge_ptep_get_and_clear(mm, address, ptep);
 977                         pte = pte_mkhuge(pte_modify(pte, newprot));
 978                         set_huge_pte_at(mm, address, ptep, pte);
 979                 }
 980         }
 981         spin_unlock(&mm->page_table_lock);
 982         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 983
 984         flush_tlb_range(vma, start, end);
 985 }
 986
 987 struct file_region {
 988         struct list_head link;
 989         long from;
 990         long to;
 991 };
 992
 993 static long region_add(struct list_head *head, long f, long t)
 994 {
 995         struct file_region *rg, *nrg, *trg;
 996
 997         /* Locate the region we are either in or before. */
 998         list_for_each_entry(rg, head, link)
 999                 if (f <= rg->to)
1000                         break;
1001
1002         /* Round our left edge to the current segment if it encloses us. */
1003         if (f > rg->from)
1004                 f = rg->from;
1005
1006         /* Check for and consume any regions we now overlap with. */
1007         nrg = rg;
1008         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1009                 if (&rg->link == head)
1010                         break;
1011                 if (rg->from > t)
1012                         break;
1013
1014                 /* If this area reaches higher then extend our area to
1015                  * include it completely.  If this is not the first area
1016                  * which we intend to reuse, free it. */
1017                 if (rg->to > t)
1018                         t = rg->to;
1019                 if (rg != nrg) {
1020                         list_del(&rg->link);
1021                         kfree(rg);
1022                 }
1023         }
1024         nrg->from = f;
1025         nrg->to = t;
1026         return 0;
1027 }
1028
1029 static long region_chg(struct list_head *head, long f, long t)
1030 {
1031         struct file_region *rg, *nrg;
1032         long chg = 0;
1033
1034         /* Locate the region we are before or in. */
1035         list_for_each_entry(rg, head, link)
1036                 if (f <= rg->to)
1037                         break;
1038
1039         /* If we are below the current region then a new region is required.
1040          * Subtle, allocate a new region at the position but make it zero
1041          * size such that we can guarantee to record the reservation. */
1042         if (&rg->link == head || t < rg->from) {
1043                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1044                 if (!nrg)
1045                         return -ENOMEM;
1046                 nrg->from = f;
1047                 nrg->to   = f;
1048                 INIT_LIST_HEAD(&nrg->link);
1049                 list_add(&nrg->link, rg->link.prev);
1050
1051                 return t - f;
1052         }
1053
1054         /* Round our left edge to the current segment if it encloses us. */
1055         if (f > rg->from)
1056                 f = rg->from;
1057         chg = t - f;
1058
1059         /* Check for and consume any regions we now overlap with. */
1060         list_for_each_entry(rg, rg->link.prev, link) {
1061                 if (&rg->link == head)
1062                         break;
1063                 if (rg->from > t)
1064                         return chg;
1065
1066                 /* We overlap with this area, if it extends futher than
1067                  * us then we must extend ourselves.  Account for its
1068                  * existing reservation. */
1069                 if (rg->to > t) {
1070                         chg += rg->to - t;
1071                         t = rg->to;
1072                 }
1073                 chg -= rg->to - rg->from;
1074         }
1075         return chg;
1076 }
1077
1078 static long region_truncate(struct list_head *head, long end)
1079 {
1080         struct file_region *rg, *trg;
1081         long chg = 0;
1082
1083         /* Locate the region we are either in or before. */
1084         list_for_each_entry(rg, head, link)
1085                 if (end <= rg->to)
1086                         break;
1087         if (&rg->link == head)
1088                 return 0;
1089
1090         /* If we are in the middle of a region then adjust it. */
1091         if (end > rg->from) {
1092                 chg = rg->to - end;
1093                 rg->to = end;
1094                 rg = list_entry(rg->link.next, typeof(*rg), link);
1095         }
1096
1097         /* Drop any remaining regions. */
1098         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1099                 if (&rg->link == head)
1100                         break;
1101                 chg += rg->to - rg->from;
1102                 list_del(&rg->link);
1103                 kfree(rg);
1104         }
1105         return chg;
1106 }
1107
1108 static int hugetlb_acct_memory(long delta)
1109 {
1110         int ret = -ENOMEM;
1111
1112         spin_lock(&hugetlb_lock);
1113         /*
1114          * When cpuset is configured, it breaks the strict hugetlb page
1115          * reservation as the accounting is done on a global variable. Such
1116          * reservation is completely rubbish in the presence of cpuset because
1117          * the reservation is not checked against page availability for the
1118          * current cpuset. Application can still potentially OOM'ed by kernel
1119          * with lack of free htlb page in cpuset that the task is in.
1120          * Attempt to enforce strict accounting with cpuset is almost
1121          * impossible (or too ugly) because cpuset is too fluid that
1122          * task or memory node can be dynamically moved between cpusets.
1123          *
1124          * The change of semantics for shared hugetlb mapping with cpuset is
1125          * undesirable. However, in order to preserve some of the semantics,
1126          * we fall back to check against current free page availability as
1127          * a best attempt and hopefully to minimize the impact of changing
1128          * semantics that cpuset has.
1129          */
1130         if (delta > 0) {
1131                 if (gather_surplus_pages(delta) < 0)
1132                         goto out;
1133
1134                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1135                         goto out;
1136         }
1137
1138         ret = 0;
1139         resv_huge_pages += delta;
1140         if (delta < 0)
1141                 return_unused_surplus_pages((unsigned long) -delta);
1142
1143 out:
1144         spin_unlock(&hugetlb_lock);
1145         return ret;
1146 }
1147
1148 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1149 {
1150         long ret, chg;
1151
1152         chg = region_chg(&inode->i_mapping->private_list, from, to);
1153         if (chg < 0)
1154                 return chg;
1155
1156         if (hugetlb_get_quota(inode->i_mapping, chg))
1157                 return -ENOSPC;
1158         ret = hugetlb_acct_memory(chg);
1159         if (ret < 0)
1160                 return ret;
1161         region_add(&inode->i_mapping->private_list, from, to);
1162         return 0;
1163 }
1164
1165 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1166 {
1167         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1168
1169         spin_lock(&inode->i_lock);
1170         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1171         spin_unlock(&inode->i_lock);
1172
1173         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1174         hugetlb_acct_memory(-(chg - freed));
1175 }