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 static unsigned long nr_overcommit_huge_pages;
  28 unsigned long max_huge_pages;
  29 unsigned long sysctl_overcommit_huge_pages;
  30 static struct list_head hugepage_freelists[MAX_NUMNODES];
  31 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  32 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  33 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
  34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  35 unsigned long hugepages_treat_as_movable;
  36 static int hugetlb_next_nid;
  37
  38 /*
  39  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  40  */
  41 static DEFINE_SPINLOCK(hugetlb_lock);
  42
  43 static void clear_huge_page(struct page *page, unsigned long addr)
  44 {
  45         int i;
  46
  47         might_sleep();
  48         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
  49                 cond_resched();
  50                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
  51         }
  52 }
  53
  54 static void copy_huge_page(struct page *dst, struct page *src,
  55                            unsigned long addr, struct vm_area_struct *vma)
  56 {
  57         int i;
  58
  59         might_sleep();
  60         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
  61                 cond_resched();
  62                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
  63         }
  64 }
  65
  66 static void enqueue_huge_page(struct page *page)
  67 {
  68         int nid = page_to_nid(page);
  69         list_add(&page->lru, &hugepage_freelists[nid]);
  70         free_huge_pages++;
  71         free_huge_pages_node[nid]++;
  72 }
  73
  74 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  75                                 unsigned long address)
  76 {
  77         int nid;
  78         struct page *page = NULL;
  79         struct mempolicy *mpol;
  80         struct zonelist *zonelist = huge_zonelist(vma, address,
  81                                         htlb_alloc_mask, &mpol);
  82         struct zone **z;
  83
  84         for (z = zonelist->zones; *z; z++) {
  85                 nid = zone_to_nid(*z);
  86                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
  87                     !list_empty(&hugepage_freelists[nid])) {
  88                         page = list_entry(hugepage_freelists[nid].next,
  89                                           struct page, lru);
  90                         list_del(&page->lru);
  91                         free_huge_pages--;
  92                         free_huge_pages_node[nid]--;
  93                         if (vma && vma->vm_flags & VM_MAYSHARE)
  94                                 resv_huge_pages--;
  95                         break;
  96                 }
  97         }
  98         mpol_free(mpol);        /* unref if mpol !NULL */
  99         return page;
 100 }
 101
 102 static void update_and_free_page(struct page *page)
 103 {
 104         int i;
 105         nr_huge_pages--;
 106         nr_huge_pages_node[page_to_nid(page)]--;
 107         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 108                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 109                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 110                                 1 << PG_private | 1<< PG_writeback);
 111         }
 112         set_compound_page_dtor(page, NULL);
 113         set_page_refcounted(page);
 114         __free_pages(page, HUGETLB_PAGE_ORDER);
 115 }
 116
 117 static void free_huge_page(struct page *page)
 118 {
 119         int nid = page_to_nid(page);
 120         struct address_space *mapping;
 121
 122         mapping = (struct address_space *) page_private(page);
 123         set_page_private(page, 0);
 124         BUG_ON(page_count(page));
 125         INIT_LIST_HEAD(&page->lru);
 126
 127         spin_lock(&hugetlb_lock);
 128         if (surplus_huge_pages_node[nid]) {
 129                 update_and_free_page(page);
 130                 surplus_huge_pages--;
 131                 surplus_huge_pages_node[nid]--;
 132         } else {
 133                 enqueue_huge_page(page);
 134         }
 135         spin_unlock(&hugetlb_lock);
 136         if (mapping)
 137                 hugetlb_put_quota(mapping, 1);
 138 }
 139
 140 /*
 141  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 142  * balanced by operating on them in a round-robin fashion.
 143  * Returns 1 if an adjustment was made.
 144  */
 145 static int adjust_pool_surplus(int delta)
 146 {
 147         static int prev_nid;
 148         int nid = prev_nid;
 149         int ret = 0;
 150
 151         VM_BUG_ON(delta != -1 && delta != 1);
 152         do {
 153                 nid = next_node(nid, node_online_map);
 154                 if (nid == MAX_NUMNODES)
 155                         nid = first_node(node_online_map);
 156
 157                 /* To shrink on this node, there must be a surplus page */
 158                 if (delta < 0 && !surplus_huge_pages_node[nid])
 159                         continue;
 160                 /* Surplus cannot exceed the total number of pages */
 161                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 162                                                 nr_huge_pages_node[nid])
 163                         continue;
 164
 165                 surplus_huge_pages += delta;
 166                 surplus_huge_pages_node[nid] += delta;
 167                 ret = 1;
 168                 break;
 169         } while (nid != prev_nid);
 170
 171         prev_nid = nid;
 172         return ret;
 173 }
 174
 175 static struct page *alloc_fresh_huge_page_node(int nid)
 176 {
 177         struct page *page;
 178
 179         page = alloc_pages_node(nid,
 180                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
 181                 HUGETLB_PAGE_ORDER);
 182         if (page) {
 183                 set_compound_page_dtor(page, free_huge_page);
 184                 spin_lock(&hugetlb_lock);
 185                 nr_huge_pages++;
 186                 nr_huge_pages_node[nid]++;
 187                 spin_unlock(&hugetlb_lock);
 188                 put_page(page); /* free it into the hugepage allocator */
 189         }
 190
 191         return page;
 192 }
 193
 194 static int alloc_fresh_huge_page(void)
 195 {
 196         struct page *page;
 197         int start_nid;
 198         int next_nid;
 199         int ret = 0;
 200
 201         start_nid = hugetlb_next_nid;
 202
 203         do {
 204                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
 205                 if (page)
 206                         ret = 1;
 207                 /*
 208                  * Use a helper variable to find the next node and then
 209                  * copy it back to hugetlb_next_nid afterwards:
 210                  * otherwise there's a window in which a racer might
 211                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 212                  * But we don't need to use a spin_lock here: it really
 213                  * doesn't matter if occasionally a racer chooses the
 214                  * same nid as we do.  Move nid forward in the mask even
 215                  * if we just successfully allocated a hugepage so that
 216                  * the next caller gets hugepages on the next node.
 217                  */
 218                 next_nid = next_node(hugetlb_next_nid, node_online_map);
 219                 if (next_nid == MAX_NUMNODES)
 220                         next_nid = first_node(node_online_map);
 221                 hugetlb_next_nid = next_nid;
 222         } while (!page && hugetlb_next_nid != start_nid);
 223
 224         return ret;
 225 }
 226
 227 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 228                                                 unsigned long address)
 229 {
 230         struct page *page;
 231         unsigned int nid;
 232
 233         /*
 234          * Assume we will successfully allocate the surplus page to
 235          * prevent racing processes from causing the surplus to exceed
 236          * overcommit
 237          *
 238          * This however introduces a different race, where a process B
 239          * tries to grow the static hugepage pool while alloc_pages() is
 240          * called by process A. B will only examine the per-node
 241          * counters in determining if surplus huge pages can be
 242          * converted to normal huge pages in adjust_pool_surplus(). A
 243          * won't be able to increment the per-node counter, until the
 244          * lock is dropped by B, but B doesn't drop hugetlb_lock until
 245          * no more huge pages can be converted from surplus to normal
 246          * state (and doesn't try to convert again). Thus, we have a
 247          * case where a surplus huge page exists, the pool is grown, and
 248          * the surplus huge page still exists after, even though it
 249          * should just have been converted to a normal huge page. This
 250          * does not leak memory, though, as the hugepage will be freed
 251          * once it is out of use. It also does not allow the counters to
 252          * go out of whack in adjust_pool_surplus() as we don't modify
 253          * the node values until we've gotten the hugepage and only the
 254          * per-node value is checked there.
 255          */
 256         spin_lock(&hugetlb_lock);
 257         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
 258                 spin_unlock(&hugetlb_lock);
 259                 return NULL;
 260         } else {
 261                 nr_huge_pages++;
 262                 surplus_huge_pages++;
 263         }
 264         spin_unlock(&hugetlb_lock);
 265
 266         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
 267                                         HUGETLB_PAGE_ORDER);
 268
 269         spin_lock(&hugetlb_lock);
 270         if (page) {
 271                 nid = page_to_nid(page);
 272                 set_compound_page_dtor(page, free_huge_page);
 273                 /*
 274                  * We incremented the global counters already
 275                  */
 276                 nr_huge_pages_node[nid]++;
 277                 surplus_huge_pages_node[nid]++;
 278         } else {
 279                 nr_huge_pages--;
 280                 surplus_huge_pages--;
 281         }
 282         spin_unlock(&hugetlb_lock);
 283
 284         return page;
 285 }
 286
 287 /*
 288  * Increase the hugetlb pool such that it can accomodate a reservation
 289  * of size 'delta'.
 290  */
 291 static int gather_surplus_pages(int delta)
 292 {
 293         struct list_head surplus_list;
 294         struct page *page, *tmp;
 295         int ret, i;
 296         int needed, allocated;
 297
 298         needed = (resv_huge_pages + delta) - free_huge_pages;
 299         if (needed <= 0)
 300                 return 0;
 301
 302         allocated = 0;
 303         INIT_LIST_HEAD(&surplus_list);
 304
 305         ret = -ENOMEM;
 306 retry:
 307         spin_unlock(&hugetlb_lock);
 308         for (i = 0; i < needed; i++) {
 309                 page = alloc_buddy_huge_page(NULL, 0);
 310                 if (!page) {
 311                         /*
 312                          * We were not able to allocate enough pages to
 313                          * satisfy the entire reservation so we free what
 314                          * we've allocated so far.
 315                          */
 316                         spin_lock(&hugetlb_lock);
 317                         needed = 0;
 318                         goto free;
 319                 }
 320
 321                 list_add(&page->lru, &surplus_list);
 322         }
 323         allocated += needed;
 324
 325         /*
 326          * After retaking hugetlb_lock, we need to recalculate 'needed'
 327          * because either resv_huge_pages or free_huge_pages may have changed.
 328          */
 329         spin_lock(&hugetlb_lock);
 330         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
 331         if (needed > 0)
 332                 goto retry;
 333
 334         /*
 335          * The surplus_list now contains _at_least_ the number of extra pages
 336          * needed to accomodate the reservation.  Add the appropriate number
 337          * of pages to the hugetlb pool and free the extras back to the buddy
 338          * allocator.
 339          */
 340         needed += allocated;
 341         ret = 0;
 342 free:
 343         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 344                 list_del(&page->lru);
 345                 if ((--needed) >= 0)
 346                         enqueue_huge_page(page);
 347                 else {
 348                         /*
 349                          * Decrement the refcount and free the page using its
 350                          * destructor.  This must be done with hugetlb_lock
 351                          * unlocked which is safe because free_huge_page takes
 352                          * hugetlb_lock before deciding how to free the page.
 353                          */
 354                         spin_unlock(&hugetlb_lock);
 355                         put_page(page);
 356                         spin_lock(&hugetlb_lock);
 357                 }
 358         }
 359
 360         return ret;
 361 }
 362
 363 /*
 364  * When releasing a hugetlb pool reservation, any surplus pages that were
 365  * allocated to satisfy the reservation must be explicitly freed if they were
 366  * never used.
 367  */
 368 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
 369 {
 370         static int nid = -1;
 371         struct page *page;
 372         unsigned long nr_pages;
 373
 374         nr_pages = min(unused_resv_pages, surplus_huge_pages);
 375
 376         while (nr_pages) {
 377                 nid = next_node(nid, node_online_map);
 378                 if (nid == MAX_NUMNODES)
 379                         nid = first_node(node_online_map);
 380
 381                 if (!surplus_huge_pages_node[nid])
 382                         continue;
 383
 384                 if (!list_empty(&hugepage_freelists[nid])) {
 385                         page = list_entry(hugepage_freelists[nid].next,
 386                                           struct page, lru);
 387                         list_del(&page->lru);
 388                         update_and_free_page(page);
 389                         free_huge_pages--;
 390                         free_huge_pages_node[nid]--;
 391                         surplus_huge_pages--;
 392                         surplus_huge_pages_node[nid]--;
 393                         nr_pages--;
 394                 }
 395         }
 396 }
 397
 398
 399 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
 400                                                 unsigned long addr)
 401 {
 402         struct page *page;
 403
 404         spin_lock(&hugetlb_lock);
 405         page = dequeue_huge_page(vma, addr);
 406         spin_unlock(&hugetlb_lock);
 407         return page ? page : ERR_PTR(-VM_FAULT_OOM);
 408 }
 409
 410 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
 411                                                 unsigned long addr)
 412 {
 413         struct page *page = NULL;
 414
 415         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
 416                 return ERR_PTR(-VM_FAULT_SIGBUS);
 417
 418         spin_lock(&hugetlb_lock);
 419         if (free_huge_pages > resv_huge_pages)
 420                 page = dequeue_huge_page(vma, addr);
 421         spin_unlock(&hugetlb_lock);
 422         if (!page) {
 423                 page = alloc_buddy_huge_page(vma, addr);
 424                 if (!page) {
 425                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
 426                         return ERR_PTR(-VM_FAULT_OOM);
 427                 }
 428         }
 429         return page;
 430 }
 431
 432 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 433                                     unsigned long addr)
 434 {
 435         struct page *page;
 436         struct address_space *mapping = vma->vm_file->f_mapping;
 437
 438         if (vma->vm_flags & VM_MAYSHARE)
 439                 page = alloc_huge_page_shared(vma, addr);
 440         else
 441                 page = alloc_huge_page_private(vma, addr);
 442
 443         if (!IS_ERR(page)) {
 444                 set_page_refcounted(page);
 445                 set_page_private(page, (unsigned long) mapping);
 446         }
 447         return page;
 448 }
 449
 450 static int __init hugetlb_init(void)
 451 {
 452         unsigned long i;
 453
 454         if (HPAGE_SHIFT == 0)
 455                 return 0;
 456
 457         for (i = 0; i < MAX_NUMNODES; ++i)
 458                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 459
 460         hugetlb_next_nid = first_node(node_online_map);
 461
 462         for (i = 0; i < max_huge_pages; ++i) {
 463                 if (!alloc_fresh_huge_page())
 464                         break;
 465         }
 466         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 467         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 468         return 0;
 469 }
 470 module_init(hugetlb_init);
 471
 472 static int __init hugetlb_setup(char *s)
 473 {
 474         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 475                 max_huge_pages = 0;
 476         return 1;
 477 }
 478 __setup("hugepages=", hugetlb_setup);
 479
 480 static unsigned int cpuset_mems_nr(unsigned int *array)
 481 {
 482         int node;
 483         unsigned int nr = 0;
 484
 485         for_each_node_mask(node, cpuset_current_mems_allowed)
 486                 nr += array[node];
 487
 488         return nr;
 489 }
 490
 491 #ifdef CONFIG_SYSCTL
 492 #ifdef CONFIG_HIGHMEM
 493 static void try_to_free_low(unsigned long count)
 494 {
 495         int i;
 496
 497         for (i = 0; i < MAX_NUMNODES; ++i) {
 498                 struct page *page, *next;
 499                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 500                         if (count >= nr_huge_pages)
 501                                 return;
 502                         if (PageHighMem(page))
 503                                 continue;
 504                         list_del(&page->lru);
 505                         update_and_free_page(page);
 506                         free_huge_pages--;
 507                         free_huge_pages_node[page_to_nid(page)]--;
 508                 }
 509         }
 510 }
 511 #else
 512 static inline void try_to_free_low(unsigned long count)
 513 {
 514 }
 515 #endif
 516
 517 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 518 static unsigned long set_max_huge_pages(unsigned long count)
 519 {
 520         unsigned long min_count, ret;
 521
 522         /*
 523          * Increase the pool size
 524          * First take pages out of surplus state.  Then make up the
 525          * remaining difference by allocating fresh huge pages.
 526          *
 527          * We might race with alloc_buddy_huge_page() here and be unable
 528          * to convert a surplus huge page to a normal huge page. That is
 529          * not critical, though, it just means the overall size of the
 530          * pool might be one hugepage larger than it needs to be, but
 531          * within all the constraints specified by the sysctls.
 532          */
 533         spin_lock(&hugetlb_lock);
 534         while (surplus_huge_pages && count > persistent_huge_pages) {
 535                 if (!adjust_pool_surplus(-1))
 536                         break;
 537         }
 538
 539         while (count > persistent_huge_pages) {
 540                 int ret;
 541                 /*
 542                  * If this allocation races such that we no longer need the
 543                  * page, free_huge_page will handle it by freeing the page
 544                  * and reducing the surplus.
 545                  */
 546                 spin_unlock(&hugetlb_lock);
 547                 ret = alloc_fresh_huge_page();
 548                 spin_lock(&hugetlb_lock);
 549                 if (!ret)
 550                         goto out;
 551
 552         }
 553
 554         /*
 555          * Decrease the pool size
 556          * First return free pages to the buddy allocator (being careful
 557          * to keep enough around to satisfy reservations).  Then place
 558          * pages into surplus state as needed so the pool will shrink
 559          * to the desired size as pages become free.
 560          *
 561          * By placing pages into the surplus state independent of the
 562          * overcommit value, we are allowing the surplus pool size to
 563          * exceed overcommit. There are few sane options here. Since
 564          * alloc_buddy_huge_page() is checking the global counter,
 565          * though, we'll note that we're not allowed to exceed surplus
 566          * and won't grow the pool anywhere else. Not until one of the
 567          * sysctls are changed, or the surplus pages go out of use.
 568          */
 569         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
 570         min_count = max(count, min_count);
 571         try_to_free_low(min_count);
 572         while (min_count < persistent_huge_pages) {
 573                 struct page *page = dequeue_huge_page(NULL, 0);
 574                 if (!page)
 575                         break;
 576                 update_and_free_page(page);
 577         }
 578         while (count < persistent_huge_pages) {
 579                 if (!adjust_pool_surplus(1))
 580                         break;
 581         }
 582 out:
 583         ret = persistent_huge_pages;
 584         spin_unlock(&hugetlb_lock);
 585         return ret;
 586 }
 587
 588 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 589                            struct file *file, void __user *buffer,
 590                            size_t *length, loff_t *ppos)
 591 {
 592         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 593         max_huge_pages = set_max_huge_pages(max_huge_pages);
 594         return 0;
 595 }
 596
 597 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 598                         struct file *file, void __user *buffer,
 599                         size_t *length, loff_t *ppos)
 600 {
 601         proc_dointvec(table, write, file, buffer, length, ppos);
 602         if (hugepages_treat_as_movable)
 603                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 604         else
 605                 htlb_alloc_mask = GFP_HIGHUSER;
 606         return 0;
 607 }
 608
 609 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
 610                         struct file *file, void __user *buffer,
 611                         size_t *length, loff_t *ppos)
 612 {
 613         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 614         spin_lock(&hugetlb_lock);
 615         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
 616         spin_unlock(&hugetlb_lock);
 617         return 0;
 618 }
 619
 620 #endif /* CONFIG_SYSCTL */
 621
 622 #ifdef  CONFIG_KDB
 623 #include <linux/kdb.h>
 624 #include <linux/kdbprivate.h>
 625 /* Like hugetlb_report_meminfo() but using kdb_printf() */
 626 void
 627 kdb_hugetlb_report_meminfo(void)
 628 {
 629         kdb_printf(
 630                 "HugePages_Total: %5lu\n"
 631                 "HugePages_Free:  %5lu\n"
 632                 "HugePages_Rsvd:  %5lu\n"
 633                 "Hugepagesize:    %5lu kB\n",
 634                 nr_huge_pages,
 635                 free_huge_pages,
 636                 resv_huge_pages,
 637                 HPAGE_SIZE/1024);
 638 }
 639 #endif  /* CONFIG_KDB */
 640
 641 int hugetlb_report_meminfo(char *buf)
 642 {
 643         return sprintf(buf,
 644                         "HugePages_Total: %5lu\n"
 645                         "HugePages_Free:  %5lu\n"
 646                         "HugePages_Rsvd:  %5lu\n"
 647                         "HugePages_Surp:  %5lu\n"
 648                         "Hugepagesize:    %5lu kB\n",
 649                         nr_huge_pages,
 650                         free_huge_pages,
 651                         resv_huge_pages,
 652                         surplus_huge_pages,
 653                         HPAGE_SIZE/1024);
 654 }
 655
 656 int hugetlb_report_node_meminfo(int nid, char *buf)
 657 {
 658         return sprintf(buf,
 659                 "Node %d HugePages_Total: %5u\n"
 660                 "Node %d HugePages_Free:  %5u\n",
 661                 nid, nr_huge_pages_node[nid],
 662                 nid, free_huge_pages_node[nid]);
 663 }
 664
 665 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 666 unsigned long hugetlb_total_pages(void)
 667 {
 668         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 669 }
 670
 671 /*
 672  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 673  * handle_mm_fault() to try to instantiate regular-sized pages in the
 674  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 675  * this far.
 676  */
 677 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 678 {
 679         BUG();
 680         return 0;
 681 }
 682
 683 struct vm_operations_struct hugetlb_vm_ops = {
 684         .fault = hugetlb_vm_op_fault,
 685 };
 686
 687 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 688                                 int writable)
 689 {
 690         pte_t entry;
 691
 692         if (writable) {
 693                 entry =
 694                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 695         } else {
 696                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 697         }
 698         entry = pte_mkyoung(entry);
 699         entry = pte_mkhuge(entry);
 700
 701         return entry;
 702 }
 703
 704 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 705                                    unsigned long address, pte_t *ptep)
 706 {
 707         pte_t entry;
 708
 709         entry = pte_mkwrite(pte_mkdirty(*ptep));
 710         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 711                 update_mmu_cache(vma, address, entry);
 712         }
 713 }
 714
 715
 716 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 717                             struct vm_area_struct *vma)
 718 {
 719         pte_t *src_pte, *dst_pte, entry;
 720         struct page *ptepage;
 721         unsigned long addr;
 722         int cow;
 723
 724         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 725
 726         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 727                 src_pte = huge_pte_offset(src, addr);
 728                 if (!src_pte)
 729                         continue;
 730                 dst_pte = huge_pte_alloc(dst, addr);
 731                 if (!dst_pte)
 732                         goto nomem;
 733
 734                 /* If the pagetables are shared don't copy or take references */
 735                 if (dst_pte == src_pte)
 736                         continue;
 737
 738                 spin_lock(&dst->page_table_lock);
 739                 spin_lock(&src->page_table_lock);
 740                 if (!pte_none(*src_pte)) {
 741                         if (cow)
 742                                 ptep_set_wrprotect(src, addr, src_pte);
 743                         entry = *src_pte;
 744                         ptepage = pte_page(entry);
 745                         get_page(ptepage);
 746                         set_huge_pte_at(dst, addr, dst_pte, entry);
 747                 }
 748                 spin_unlock(&src->page_table_lock);
 749                 spin_unlock(&dst->page_table_lock);
 750         }
 751         return 0;
 752
 753 nomem:
 754         return -ENOMEM;
 755 }
 756
 757 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 758                             unsigned long end)
 759 {
 760         struct mm_struct *mm = vma->vm_mm;
 761         unsigned long address;
 762         pte_t *ptep;
 763         pte_t pte;
 764         struct page *page;
 765         struct page *tmp;
 766         /*
 767          * A page gathering list, protected by per file i_mmap_lock. The
 768          * lock is used to avoid list corruption from multiple unmapping
 769          * of the same page since we are using page->lru.
 770          */
 771         LIST_HEAD(page_list);
 772
 773         WARN_ON(!is_vm_hugetlb_page(vma));
 774         BUG_ON(start & ~HPAGE_MASK);
 775         BUG_ON(end & ~HPAGE_MASK);
 776
 777         spin_lock(&mm->page_table_lock);
 778         for (address = start; address < end; address += HPAGE_SIZE) {
 779                 ptep = huge_pte_offset(mm, address);
 780                 if (!ptep)
 781                         continue;
 782
 783                 if (huge_pmd_unshare(mm, &address, ptep))
 784                         continue;
 785
 786                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 787                 if (pte_none(pte))
 788                         continue;
 789
 790                 page = pte_page(pte);
 791                 if (pte_dirty(pte))
 792                         set_page_dirty(page);
 793                 list_add(&page->lru, &page_list);
 794         }
 795         spin_unlock(&mm->page_table_lock);
 796         flush_tlb_range(vma, start, end);
 797         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 798                 list_del(&page->lru);
 799                 put_page(page);
 800         }
 801 }
 802
 803 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 804                           unsigned long end)
 805 {
 806         /*
 807          * It is undesirable to test vma->vm_file as it should be non-null
 808          * for valid hugetlb area. However, vm_file will be NULL in the error
 809          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 810          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 811          * to clean up. Since no pte has actually been setup, it is safe to
 812          * do nothing in this case.
 813          */
 814         if (vma->vm_file) {
 815                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 816                 __unmap_hugepage_range(vma, start, end);
 817                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 818         }
 819 }
 820
 821 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 822                         unsigned long address, pte_t *ptep, pte_t pte)
 823 {
 824         struct page *old_page, *new_page;
 825         int avoidcopy;
 826
 827         old_page = pte_page(pte);
 828
 829         /* If no-one else is actually using this page, avoid the copy
 830          * and just make the page writable */
 831         avoidcopy = (page_count(old_page) == 1);
 832         if (avoidcopy) {
 833                 set_huge_ptep_writable(vma, address, ptep);
 834                 return 0;
 835         }
 836
 837         page_cache_get(old_page);
 838         new_page = alloc_huge_page(vma, address);
 839
 840         if (IS_ERR(new_page)) {
 841                 page_cache_release(old_page);
 842                 return -PTR_ERR(new_page);
 843         }
 844
 845         spin_unlock(&mm->page_table_lock);
 846         copy_huge_page(new_page, old_page, address, vma);
 847         __SetPageUptodate(new_page);
 848         spin_lock(&mm->page_table_lock);
 849
 850         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 851         if (likely(pte_same(*ptep, pte))) {
 852                 /* Break COW */
 853                 set_huge_pte_at(mm, address, ptep,
 854                                 make_huge_pte(vma, new_page, 1));
 855                 /* Make the old page be freed below */
 856                 new_page = old_page;
 857         }
 858         page_cache_release(new_page);
 859         page_cache_release(old_page);
 860         return 0;
 861 }
 862
 863 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 864                         unsigned long address, pte_t *ptep, int write_access)
 865 {
 866         int ret = VM_FAULT_SIGBUS;
 867         unsigned long idx;
 868         unsigned long size;
 869         struct page *page;
 870         struct address_space *mapping;
 871         pte_t new_pte;
 872
 873         mapping = vma->vm_file->f_mapping;
 874         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 875                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 876
 877         /*
 878          * Use page lock to guard against racing truncation
 879          * before we get page_table_lock.
 880          */
 881 retry:
 882         page = find_lock_page(mapping, idx);
 883         if (!page) {
 884                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 885                 if (idx >= size)
 886                         goto out;
 887                 page = alloc_huge_page(vma, address);
 888                 if (IS_ERR(page)) {
 889                         ret = -PTR_ERR(page);
 890                         goto out;
 891                 }
 892                 clear_huge_page(page, address);
 893                 __SetPageUptodate(page);
 894
 895                 if (vma->vm_flags & VM_SHARED) {
 896                         int err;
 897                         struct inode *inode = mapping->host;
 898
 899                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 900                         if (err) {
 901                                 put_page(page);
 902                                 if (err == -EEXIST)
 903                                         goto retry;
 904                                 goto out;
 905                         }
 906
 907                         spin_lock(&inode->i_lock);
 908                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
 909                         spin_unlock(&inode->i_lock);
 910                 } else
 911                         lock_page(page);
 912         }
 913
 914         spin_lock(&mm->page_table_lock);
 915         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 916         if (idx >= size)
 917                 goto backout;
 918
 919         ret = 0;
 920         if (!pte_none(*ptep))
 921                 goto backout;
 922
 923         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 924                                 && (vma->vm_flags & VM_SHARED)));
 925         set_huge_pte_at(mm, address, ptep, new_pte);
 926
 927         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 928                 /* Optimization, do the COW without a second fault */
 929                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 930         }
 931
 932         spin_unlock(&mm->page_table_lock);
 933         unlock_page(page);
 934 out:
 935         return ret;
 936
 937 backout:
 938         spin_unlock(&mm->page_table_lock);
 939         unlock_page(page);
 940         put_page(page);
 941         goto out;
 942 }
 943
 944 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 945                         unsigned long address, int write_access)
 946 {
 947         pte_t *ptep;
 948         pte_t entry;
 949         int ret;
 950         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
 951
 952         ptep = huge_pte_alloc(mm, address);
 953         if (!ptep)
 954                 return VM_FAULT_OOM;
 955
 956         /*
 957          * Serialize hugepage allocation and instantiation, so that we don't
 958          * get spurious allocation failures if two CPUs race to instantiate
 959          * the same page in the page cache.
 960          */
 961         mutex_lock(&hugetlb_instantiation_mutex);
 962         entry = *ptep;
 963         if (pte_none(entry)) {
 964                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
 965                 mutex_unlock(&hugetlb_instantiation_mutex);
 966                 return ret;
 967         }
 968
 969         ret = 0;
 970
 971         spin_lock(&mm->page_table_lock);
 972         /* Check for a racing update before calling hugetlb_cow */
 973         if (likely(pte_same(entry, *ptep)))
 974                 if (write_access && !pte_write(entry))
 975                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 976         spin_unlock(&mm->page_table_lock);
 977         mutex_unlock(&hugetlb_instantiation_mutex);
 978
 979         return ret;
 980 }
 981
 982 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 983                         struct page **pages, struct vm_area_struct **vmas,
 984                         unsigned long *position, int *length, int i,
 985                         int write)
 986 {
 987         unsigned long pfn_offset;
 988         unsigned long vaddr = *position;
 989         int remainder = *length;
 990
 991         spin_lock(&mm->page_table_lock);
 992         while (vaddr < vma->vm_end && remainder) {
 993                 pte_t *pte;
 994                 struct page *page;
 995
 996                 /*
 997                  * Some archs (sparc64, sh*) have multiple pte_ts to
 998                  * each hugepage.  We have to make * sure we get the
 999                  * first, for the page indexing below to work.
1000                  */
1001                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1002
1003                 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
1004                         int ret;
1005
1006                         spin_unlock(&mm->page_table_lock);
1007                         ret = hugetlb_fault(mm, vma, vaddr, write);
1008                         spin_lock(&mm->page_table_lock);
1009                         if (!(ret & VM_FAULT_ERROR))
1010                                 continue;
1011
1012                         remainder = 0;
1013                         if (!i)
1014                                 i = -EFAULT;
1015                         break;
1016                 }
1017
1018                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1019                 page = pte_page(*pte);
1020 same_page:
1021                 if (pages) {
1022                         get_page(page);
1023                         pages[i] = page + pfn_offset;
1024                 }
1025
1026                 if (vmas)
1027                         vmas[i] = vma;
1028
1029                 vaddr += PAGE_SIZE;
1030                 ++pfn_offset;
1031                 --remainder;
1032                 ++i;
1033                 if (vaddr < vma->vm_end && remainder &&
1034                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1035                         /*
1036                          * We use pfn_offset to avoid touching the pageframes
1037                          * of this compound page.
1038                          */
1039                         goto same_page;
1040                 }
1041         }
1042         spin_unlock(&mm->page_table_lock);
1043         *length = remainder;
1044         *position = vaddr;
1045
1046         return i;
1047 }
1048
1049 void hugetlb_change_protection(struct vm_area_struct *vma,
1050                 unsigned long address, unsigned long end, pgprot_t newprot)
1051 {
1052         struct mm_struct *mm = vma->vm_mm;
1053         unsigned long start = address;
1054         pte_t *ptep;
1055         pte_t pte;
1056
1057         BUG_ON(address >= end);
1058         flush_cache_range(vma, address, end);
1059
1060         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1061         spin_lock(&mm->page_table_lock);
1062         for (; address < end; address += HPAGE_SIZE) {
1063                 ptep = huge_pte_offset(mm, address);
1064                 if (!ptep)
1065                         continue;
1066                 if (huge_pmd_unshare(mm, &address, ptep))
1067                         continue;
1068                 if (!pte_none(*ptep)) {
1069                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1070                         pte = pte_mkhuge(pte_modify(pte, newprot));
1071                         set_huge_pte_at(mm, address, ptep, pte);
1072                 }
1073         }
1074         spin_unlock(&mm->page_table_lock);
1075         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1076
1077         flush_tlb_range(vma, start, end);
1078 }
1079
1080 struct file_region {
1081         struct list_head link;
1082         long from;
1083         long to;
1084 };
1085
1086 static long region_add(struct list_head *head, long f, long t)
1087 {
1088         struct file_region *rg, *nrg, *trg;
1089
1090         /* Locate the region we are either in or before. */
1091         list_for_each_entry(rg, head, link)
1092                 if (f <= rg->to)
1093                         break;
1094
1095         /* Round our left edge to the current segment if it encloses us. */
1096         if (f > rg->from)
1097                 f = rg->from;
1098
1099         /* Check for and consume any regions we now overlap with. */
1100         nrg = rg;
1101         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1102                 if (&rg->link == head)
1103                         break;
1104                 if (rg->from > t)
1105                         break;
1106
1107                 /* If this area reaches higher then extend our area to
1108                  * include it completely.  If this is not the first area
1109                  * which we intend to reuse, free it. */
1110                 if (rg->to > t)
1111                         t = rg->to;
1112                 if (rg != nrg) {
1113                         list_del(&rg->link);
1114                         kfree(rg);
1115                 }
1116         }
1117         nrg->from = f;
1118         nrg->to = t;
1119         return 0;
1120 }
1121
1122 static long region_chg(struct list_head *head, long f, long t)
1123 {
1124         struct file_region *rg, *nrg;
1125         long chg = 0;
1126
1127         /* Locate the region we are before or in. */
1128         list_for_each_entry(rg, head, link)
1129                 if (f <= rg->to)
1130                         break;
1131
1132         /* If we are below the current region then a new region is required.
1133          * Subtle, allocate a new region at the position but make it zero
1134          * size such that we can guarantee to record the reservation. */
1135         if (&rg->link == head || t < rg->from) {
1136                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1137                 if (!nrg)
1138                         return -ENOMEM;
1139                 nrg->from = f;
1140                 nrg->to   = f;
1141                 INIT_LIST_HEAD(&nrg->link);
1142                 list_add(&nrg->link, rg->link.prev);
1143
1144                 return t - f;
1145         }
1146
1147         /* Round our left edge to the current segment if it encloses us. */
1148         if (f > rg->from)
1149                 f = rg->from;
1150         chg = t - f;
1151
1152         /* Check for and consume any regions we now overlap with. */
1153         list_for_each_entry(rg, rg->link.prev, link) {
1154                 if (&rg->link == head)
1155                         break;
1156                 if (rg->from > t)
1157                         return chg;
1158
1159                 /* We overlap with this area, if it extends futher than
1160                  * us then we must extend ourselves.  Account for its
1161                  * existing reservation. */
1162                 if (rg->to > t) {
1163                         chg += rg->to - t;
1164                         t = rg->to;
1165                 }
1166                 chg -= rg->to - rg->from;
1167         }
1168         return chg;
1169 }
1170
1171 static long region_truncate(struct list_head *head, long end)
1172 {
1173         struct file_region *rg, *trg;
1174         long chg = 0;
1175
1176         /* Locate the region we are either in or before. */
1177         list_for_each_entry(rg, head, link)
1178                 if (end <= rg->to)
1179                         break;
1180         if (&rg->link == head)
1181                 return 0;
1182
1183         /* If we are in the middle of a region then adjust it. */
1184         if (end > rg->from) {
1185                 chg = rg->to - end;
1186                 rg->to = end;
1187                 rg = list_entry(rg->link.next, typeof(*rg), link);
1188         }
1189
1190         /* Drop any remaining regions. */
1191         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1192                 if (&rg->link == head)
1193                         break;
1194                 chg += rg->to - rg->from;
1195                 list_del(&rg->link);
1196                 kfree(rg);
1197         }
1198         return chg;
1199 }
1200
1201 static int hugetlb_acct_memory(long delta)
1202 {
1203         int ret = -ENOMEM;
1204
1205         spin_lock(&hugetlb_lock);
1206         /*
1207          * When cpuset is configured, it breaks the strict hugetlb page
1208          * reservation as the accounting is done on a global variable. Such
1209          * reservation is completely rubbish in the presence of cpuset because
1210          * the reservation is not checked against page availability for the
1211          * current cpuset. Application can still potentially OOM'ed by kernel
1212          * with lack of free htlb page in cpuset that the task is in.
1213          * Attempt to enforce strict accounting with cpuset is almost
1214          * impossible (or too ugly) because cpuset is too fluid that
1215          * task or memory node can be dynamically moved between cpusets.
1216          *
1217          * The change of semantics for shared hugetlb mapping with cpuset is
1218          * undesirable. However, in order to preserve some of the semantics,
1219          * we fall back to check against current free page availability as
1220          * a best attempt and hopefully to minimize the impact of changing
1221          * semantics that cpuset has.
1222          */
1223         if (delta > 0) {
1224                 if (gather_surplus_pages(delta) < 0)
1225                         goto out;
1226
1227                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1228                         goto out;
1229         }
1230
1231         ret = 0;
1232         resv_huge_pages += delta;
1233         if (delta < 0)
1234                 return_unused_surplus_pages((unsigned long) -delta);
1235
1236 out:
1237         spin_unlock(&hugetlb_lock);
1238         return ret;
1239 }
1240
1241 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1242 {
1243         long ret, chg;
1244
1245         chg = region_chg(&inode->i_mapping->private_list, from, to);
1246         if (chg < 0)
1247                 return chg;
1248
1249         if (hugetlb_get_quota(inode->i_mapping, chg))
1250                 return -ENOSPC;
1251         ret = hugetlb_acct_memory(chg);
1252         if (ret < 0) {
1253                 hugetlb_put_quota(inode->i_mapping, chg);
1254                 return ret;
1255         }
1256         region_add(&inode->i_mapping->private_list, from, to);
1257         return 0;
1258 }
1259
1260 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1261 {
1262         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1263
1264         spin_lock(&inode->i_lock);
1265         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1266         spin_unlock(&inode->i_lock);
1267
1268         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1269         hugetlb_acct_memory(-(chg - freed));
1270 }