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(void)
  75 {
  76         int nid;
  77         struct page *page = NULL;
  78
  79         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
  80                 if (!list_empty(&hugepage_freelists[nid])) {
  81                         page = list_entry(hugepage_freelists[nid].next,
  82                                           struct page, lru);
  83                         list_del(&page->lru);
  84                         free_huge_pages--;
  85                         free_huge_pages_node[nid]--;
  86                         break;
  87                 }
  88         }
  89         return page;
  90 }
  91
  92 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
  93                                 unsigned long address)
  94 {
  95         int nid;
  96         struct page *page = NULL;
  97         struct mempolicy *mpol;
  98         nodemask_t *nodemask;
  99         struct zonelist *zonelist = huge_zonelist(vma, address,
 100                                         htlb_alloc_mask, &mpol, &nodemask);
 101         struct zone *zone;
 102         struct zoneref *z;
 103
 104         for_each_zone_zonelist_nodemask(zone, z, zonelist,
 105                                                 MAX_NR_ZONES - 1, nodemask) {
 106                 nid = zone_to_nid(zone);
 107                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
 108                     !list_empty(&hugepage_freelists[nid])) {
 109                         page = list_entry(hugepage_freelists[nid].next,
 110                                           struct page, lru);
 111                         list_del(&page->lru);
 112                         free_huge_pages--;
 113                         free_huge_pages_node[nid]--;
 114                         if (vma && vma->vm_flags & VM_MAYSHARE)
 115                                 resv_huge_pages--;
 116                         break;
 117                 }
 118         }
 119         mpol_cond_put(mpol);
 120         return page;
 121 }
 122
 123 static void update_and_free_page(struct page *page)
 124 {
 125         int i;
 126         nr_huge_pages--;
 127         nr_huge_pages_node[page_to_nid(page)]--;
 128         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 129                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 130                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 131                                 1 << PG_private | 1<< PG_writeback);
 132         }
 133         set_compound_page_dtor(page, NULL);
 134         set_page_refcounted(page);
 135         arch_release_hugepage(page);
 136         __free_pages(page, HUGETLB_PAGE_ORDER);
 137 }
 138
 139 static void free_huge_page(struct page *page)
 140 {
 141         int nid = page_to_nid(page);
 142         struct address_space *mapping;
 143
 144         mapping = (struct address_space *) page_private(page);
 145         set_page_private(page, 0);
 146         BUG_ON(page_count(page));
 147         INIT_LIST_HEAD(&page->lru);
 148
 149         spin_lock(&hugetlb_lock);
 150         if (surplus_huge_pages_node[nid]) {
 151                 update_and_free_page(page);
 152                 surplus_huge_pages--;
 153                 surplus_huge_pages_node[nid]--;
 154         } else {
 155                 enqueue_huge_page(page);
 156         }
 157         spin_unlock(&hugetlb_lock);
 158         if (mapping)
 159                 hugetlb_put_quota(mapping, 1);
 160 }
 161
 162 /*
 163  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
 164  * balanced by operating on them in a round-robin fashion.
 165  * Returns 1 if an adjustment was made.
 166  */
 167 static int adjust_pool_surplus(int delta)
 168 {
 169         static int prev_nid;
 170         int nid = prev_nid;
 171         int ret = 0;
 172
 173         VM_BUG_ON(delta != -1 && delta != 1);
 174         do {
 175                 nid = next_node(nid, node_online_map);
 176                 if (nid == MAX_NUMNODES)
 177                         nid = first_node(node_online_map);
 178
 179                 /* To shrink on this node, there must be a surplus page */
 180                 if (delta < 0 && !surplus_huge_pages_node[nid])
 181                         continue;
 182                 /* Surplus cannot exceed the total number of pages */
 183                 if (delta > 0 && surplus_huge_pages_node[nid] >=
 184                                                 nr_huge_pages_node[nid])
 185                         continue;
 186
 187                 surplus_huge_pages += delta;
 188                 surplus_huge_pages_node[nid] += delta;
 189                 ret = 1;
 190                 break;
 191         } while (nid != prev_nid);
 192
 193         prev_nid = nid;
 194         return ret;
 195 }
 196
 197 static struct page *alloc_fresh_huge_page_node(int nid)
 198 {
 199         struct page *page;
 200
 201         page = alloc_pages_node(nid,
 202                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 203                                                 __GFP_REPEAT|__GFP_NOWARN,
 204                 HUGETLB_PAGE_ORDER);
 205         if (page) {
 206                 if (arch_prepare_hugepage(page)) {
 207                         __free_pages(page, HUGETLB_PAGE_ORDER);
 208                         return NULL;
 209                 }
 210                 set_compound_page_dtor(page, free_huge_page);
 211                 spin_lock(&hugetlb_lock);
 212                 nr_huge_pages++;
 213                 nr_huge_pages_node[nid]++;
 214                 spin_unlock(&hugetlb_lock);
 215                 put_page(page); /* free it into the hugepage allocator */
 216         }
 217
 218         return page;
 219 }
 220
 221 static int alloc_fresh_huge_page(void)
 222 {
 223         struct page *page;
 224         int start_nid;
 225         int next_nid;
 226         int ret = 0;
 227
 228         start_nid = hugetlb_next_nid;
 229
 230         do {
 231                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
 232                 if (page)
 233                         ret = 1;
 234                 /*
 235                  * Use a helper variable to find the next node and then
 236                  * copy it back to hugetlb_next_nid afterwards:
 237                  * otherwise there's a window in which a racer might
 238                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
 239                  * But we don't need to use a spin_lock here: it really
 240                  * doesn't matter if occasionally a racer chooses the
 241                  * same nid as we do.  Move nid forward in the mask even
 242                  * if we just successfully allocated a hugepage so that
 243                  * the next caller gets hugepages on the next node.
 244                  */
 245                 next_nid = next_node(hugetlb_next_nid, node_online_map);
 246                 if (next_nid == MAX_NUMNODES)
 247                         next_nid = first_node(node_online_map);
 248                 hugetlb_next_nid = next_nid;
 249         } while (!page && hugetlb_next_nid != start_nid);
 250
 251         if (ret)
 252                 count_vm_event(HTLB_BUDDY_PGALLOC);
 253         else
 254                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 255
 256         return ret;
 257 }
 258
 259 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
 260                                                 unsigned long address)
 261 {
 262         struct page *page;
 263         unsigned int nid;
 264
 265         /*
 266          * Assume we will successfully allocate the surplus page to
 267          * prevent racing processes from causing the surplus to exceed
 268          * overcommit
 269          *
 270          * This however introduces a different race, where a process B
 271          * tries to grow the static hugepage pool while alloc_pages() is
 272          * called by process A. B will only examine the per-node
 273          * counters in determining if surplus huge pages can be
 274          * converted to normal huge pages in adjust_pool_surplus(). A
 275          * won't be able to increment the per-node counter, until the
 276          * lock is dropped by B, but B doesn't drop hugetlb_lock until
 277          * no more huge pages can be converted from surplus to normal
 278          * state (and doesn't try to convert again). Thus, we have a
 279          * case where a surplus huge page exists, the pool is grown, and
 280          * the surplus huge page still exists after, even though it
 281          * should just have been converted to a normal huge page. This
 282          * does not leak memory, though, as the hugepage will be freed
 283          * once it is out of use. It also does not allow the counters to
 284          * go out of whack in adjust_pool_surplus() as we don't modify
 285          * the node values until we've gotten the hugepage and only the
 286          * per-node value is checked there.
 287          */
 288         spin_lock(&hugetlb_lock);
 289         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
 290                 spin_unlock(&hugetlb_lock);
 291                 return NULL;
 292         } else {
 293                 nr_huge_pages++;
 294                 surplus_huge_pages++;
 295         }
 296         spin_unlock(&hugetlb_lock);
 297
 298         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 299                                         __GFP_REPEAT|__GFP_NOWARN,
 300                                         HUGETLB_PAGE_ORDER);
 301
 302         spin_lock(&hugetlb_lock);
 303         if (page) {
 304                 /*
 305                  * This page is now managed by the hugetlb allocator and has
 306                  * no users -- drop the buddy allocator's reference.
 307                  */
 308                 put_page_testzero(page);
 309                 VM_BUG_ON(page_count(page));
 310                 nid = page_to_nid(page);
 311                 set_compound_page_dtor(page, free_huge_page);
 312                 /*
 313                  * We incremented the global counters already
 314                  */
 315                 nr_huge_pages_node[nid]++;
 316                 surplus_huge_pages_node[nid]++;
 317                 __count_vm_event(HTLB_BUDDY_PGALLOC);
 318         } else {
 319                 nr_huge_pages--;
 320                 surplus_huge_pages--;
 321                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 322         }
 323         spin_unlock(&hugetlb_lock);
 324
 325         return page;
 326 }
 327
 328 /*
 329  * Increase the hugetlb pool such that it can accomodate a reservation
 330  * of size 'delta'.
 331  */
 332 static int gather_surplus_pages(int delta)
 333 {
 334         struct list_head surplus_list;
 335         struct page *page, *tmp;
 336         int ret, i;
 337         int needed, allocated;
 338
 339         needed = (resv_huge_pages + delta) - free_huge_pages;
 340         if (needed <= 0) {
 341                 resv_huge_pages += delta;
 342                 return 0;
 343         }
 344
 345         allocated = 0;
 346         INIT_LIST_HEAD(&surplus_list);
 347
 348         ret = -ENOMEM;
 349 retry:
 350         spin_unlock(&hugetlb_lock);
 351         for (i = 0; i < needed; i++) {
 352                 page = alloc_buddy_huge_page(NULL, 0);
 353                 if (!page) {
 354                         /*
 355                          * We were not able to allocate enough pages to
 356                          * satisfy the entire reservation so we free what
 357                          * we've allocated so far.
 358                          */
 359                         spin_lock(&hugetlb_lock);
 360                         needed = 0;
 361                         goto free;
 362                 }
 363
 364                 list_add(&page->lru, &surplus_list);
 365         }
 366         allocated += needed;
 367
 368         /*
 369          * After retaking hugetlb_lock, we need to recalculate 'needed'
 370          * because either resv_huge_pages or free_huge_pages may have changed.
 371          */
 372         spin_lock(&hugetlb_lock);
 373         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
 374         if (needed > 0)
 375                 goto retry;
 376
 377         /*
 378          * The surplus_list now contains _at_least_ the number of extra pages
 379          * needed to accomodate the reservation.  Add the appropriate number
 380          * of pages to the hugetlb pool and free the extras back to the buddy
 381          * allocator.  Commit the entire reservation here to prevent another
 382          * process from stealing the pages as they are added to the pool but
 383          * before they are reserved.
 384          */
 385         needed += allocated;
 386         resv_huge_pages += delta;
 387         ret = 0;
 388 free:
 389         /* Free the needed pages to the hugetlb pool */
 390         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 391                 if ((--needed) < 0)
 392                         break;
 393                 list_del(&page->lru);
 394                 enqueue_huge_page(page);
 395         }
 396
 397         /* Free unnecessary surplus pages to the buddy allocator */
 398         if (!list_empty(&surplus_list)) {
 399                 spin_unlock(&hugetlb_lock);
 400                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 401                         list_del(&page->lru);
 402                         /*
 403                          * The page has a reference count of zero already, so
 404                          * call free_huge_page directly instead of using
 405                          * put_page.  This must be done with hugetlb_lock
 406                          * unlocked which is safe because free_huge_page takes
 407                          * hugetlb_lock before deciding how to free the page.
 408                          */
 409                         free_huge_page(page);
 410                 }
 411                 spin_lock(&hugetlb_lock);
 412         }
 413
 414         return ret;
 415 }
 416
 417 /*
 418  * When releasing a hugetlb pool reservation, any surplus pages that were
 419  * allocated to satisfy the reservation must be explicitly freed if they were
 420  * never used.
 421  */
 422 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
 423 {
 424         static int nid = -1;
 425         struct page *page;
 426         unsigned long nr_pages;
 427
 428         /*
 429          * We want to release as many surplus pages as possible, spread
 430          * evenly across all nodes. Iterate across all nodes until we
 431          * can no longer free unreserved surplus pages. This occurs when
 432          * the nodes with surplus pages have no free pages.
 433          */
 434         unsigned long remaining_iterations = num_online_nodes();
 435
 436         /* Uncommit the reservation */
 437         resv_huge_pages -= unused_resv_pages;
 438
 439         nr_pages = min(unused_resv_pages, surplus_huge_pages);
 440
 441         while (remaining_iterations-- && nr_pages) {
 442                 nid = next_node(nid, node_online_map);
 443                 if (nid == MAX_NUMNODES)
 444                         nid = first_node(node_online_map);
 445
 446                 if (!surplus_huge_pages_node[nid])
 447                         continue;
 448
 449                 if (!list_empty(&hugepage_freelists[nid])) {
 450                         page = list_entry(hugepage_freelists[nid].next,
 451                                           struct page, lru);
 452                         list_del(&page->lru);
 453                         update_and_free_page(page);
 454                         free_huge_pages--;
 455                         free_huge_pages_node[nid]--;
 456                         surplus_huge_pages--;
 457                         surplus_huge_pages_node[nid]--;
 458                         nr_pages--;
 459                         remaining_iterations = num_online_nodes();
 460                 }
 461         }
 462 }
 463
 464
 465 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
 466                                                 unsigned long addr)
 467 {
 468         struct page *page;
 469
 470         spin_lock(&hugetlb_lock);
 471         page = dequeue_huge_page_vma(vma, addr);
 472         spin_unlock(&hugetlb_lock);
 473         return page ? page : ERR_PTR(-VM_FAULT_OOM);
 474 }
 475
 476 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
 477                                                 unsigned long addr)
 478 {
 479         struct page *page = NULL;
 480
 481         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
 482                 return ERR_PTR(-VM_FAULT_SIGBUS);
 483
 484         spin_lock(&hugetlb_lock);
 485         if (free_huge_pages > resv_huge_pages)
 486                 page = dequeue_huge_page_vma(vma, addr);
 487         spin_unlock(&hugetlb_lock);
 488         if (!page) {
 489                 page = alloc_buddy_huge_page(vma, addr);
 490                 if (!page) {
 491                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
 492                         return ERR_PTR(-VM_FAULT_OOM);
 493                 }
 494         }
 495         return page;
 496 }
 497
 498 static struct page *alloc_huge_page(struct vm_area_struct *vma,
 499                                     unsigned long addr)
 500 {
 501         struct page *page;
 502         struct address_space *mapping = vma->vm_file->f_mapping;
 503
 504         if (vma->vm_flags & VM_MAYSHARE)
 505                 page = alloc_huge_page_shared(vma, addr);
 506         else
 507                 page = alloc_huge_page_private(vma, addr);
 508
 509         if (!IS_ERR(page)) {
 510                 set_page_refcounted(page);
 511                 set_page_private(page, (unsigned long) mapping);
 512         }
 513         return page;
 514 }
 515
 516 static int __init hugetlb_init(void)
 517 {
 518         unsigned long i;
 519
 520         if (HPAGE_SHIFT == 0)
 521                 return 0;
 522
 523         for (i = 0; i < MAX_NUMNODES; ++i)
 524                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 525
 526         hugetlb_next_nid = first_node(node_online_map);
 527
 528         for (i = 0; i < max_huge_pages; ++i) {
 529                 if (!alloc_fresh_huge_page())
 530                         break;
 531         }
 532         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 533         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 534         return 0;
 535 }
 536 module_init(hugetlb_init);
 537
 538 static int __init hugetlb_setup(char *s)
 539 {
 540         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 541                 max_huge_pages = 0;
 542         return 1;
 543 }
 544 __setup("hugepages=", hugetlb_setup);
 545
 546 static unsigned int cpuset_mems_nr(unsigned int *array)
 547 {
 548         int node;
 549         unsigned int nr = 0;
 550
 551         for_each_node_mask(node, cpuset_current_mems_allowed)
 552                 nr += array[node];
 553
 554         return nr;
 555 }
 556
 557 #ifdef CONFIG_SYSCTL
 558 #ifdef CONFIG_HIGHMEM
 559 static void try_to_free_low(unsigned long count)
 560 {
 561         int i;
 562
 563         for (i = 0; i < MAX_NUMNODES; ++i) {
 564                 struct page *page, *next;
 565                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 566                         if (count >= nr_huge_pages)
 567                                 return;
 568                         if (PageHighMem(page))
 569                                 continue;
 570                         list_del(&page->lru);
 571                         update_and_free_page(page);
 572                         free_huge_pages--;
 573                         free_huge_pages_node[page_to_nid(page)]--;
 574                 }
 575         }
 576 }
 577 #else
 578 static inline void try_to_free_low(unsigned long count)
 579 {
 580 }
 581 #endif
 582
 583 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 584 static unsigned long set_max_huge_pages(unsigned long count)
 585 {
 586         unsigned long min_count, ret;
 587
 588         /*
 589          * Increase the pool size
 590          * First take pages out of surplus state.  Then make up the
 591          * remaining difference by allocating fresh huge pages.
 592          *
 593          * We might race with alloc_buddy_huge_page() here and be unable
 594          * to convert a surplus huge page to a normal huge page. That is
 595          * not critical, though, it just means the overall size of the
 596          * pool might be one hugepage larger than it needs to be, but
 597          * within all the constraints specified by the sysctls.
 598          */
 599         spin_lock(&hugetlb_lock);
 600         while (surplus_huge_pages && count > persistent_huge_pages) {
 601                 if (!adjust_pool_surplus(-1))
 602                         break;
 603         }
 604
 605         while (count > persistent_huge_pages) {
 606                 int ret;
 607                 /*
 608                  * If this allocation races such that we no longer need the
 609                  * page, free_huge_page will handle it by freeing the page
 610                  * and reducing the surplus.
 611                  */
 612                 spin_unlock(&hugetlb_lock);
 613                 ret = alloc_fresh_huge_page();
 614                 spin_lock(&hugetlb_lock);
 615                 if (!ret)
 616                         goto out;
 617
 618         }
 619
 620         /*
 621          * Decrease the pool size
 622          * First return free pages to the buddy allocator (being careful
 623          * to keep enough around to satisfy reservations).  Then place
 624          * pages into surplus state as needed so the pool will shrink
 625          * to the desired size as pages become free.
 626          *
 627          * By placing pages into the surplus state independent of the
 628          * overcommit value, we are allowing the surplus pool size to
 629          * exceed overcommit. There are few sane options here. Since
 630          * alloc_buddy_huge_page() is checking the global counter,
 631          * though, we'll note that we're not allowed to exceed surplus
 632          * and won't grow the pool anywhere else. Not until one of the
 633          * sysctls are changed, or the surplus pages go out of use.
 634          */
 635         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
 636         min_count = max(count, min_count);
 637         try_to_free_low(min_count);
 638         while (min_count < persistent_huge_pages) {
 639                 struct page *page = dequeue_huge_page();
 640                 if (!page)
 641                         break;
 642                 update_and_free_page(page);
 643         }
 644         while (count < persistent_huge_pages) {
 645                 if (!adjust_pool_surplus(1))
 646                         break;
 647         }
 648 out:
 649         ret = persistent_huge_pages;
 650         spin_unlock(&hugetlb_lock);
 651         return ret;
 652 }
 653
 654 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 655                            struct file *file, void __user *buffer,
 656                            size_t *length, loff_t *ppos)
 657 {
 658         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 659         max_huge_pages = set_max_huge_pages(max_huge_pages);
 660         return 0;
 661 }
 662
 663 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
 664                         struct file *file, void __user *buffer,
 665                         size_t *length, loff_t *ppos)
 666 {
 667         proc_dointvec(table, write, file, buffer, length, ppos);
 668         if (hugepages_treat_as_movable)
 669                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
 670         else
 671                 htlb_alloc_mask = GFP_HIGHUSER;
 672         return 0;
 673 }
 674
 675 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
 676                         struct file *file, void __user *buffer,
 677                         size_t *length, loff_t *ppos)
 678 {
 679         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 680         spin_lock(&hugetlb_lock);
 681         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
 682         spin_unlock(&hugetlb_lock);
 683         return 0;
 684 }
 685
 686 #endif /* CONFIG_SYSCTL */
 687
 688 int hugetlb_report_meminfo(char *buf)
 689 {
 690         return sprintf(buf,
 691                         "HugePages_Total: %5lu\n"
 692                         "HugePages_Free:  %5lu\n"
 693                         "HugePages_Rsvd:  %5lu\n"
 694                         "HugePages_Surp:  %5lu\n"
 695                         "Hugepagesize:    %5lu kB\n",
 696                         nr_huge_pages,
 697                         free_huge_pages,
 698                         resv_huge_pages,
 699                         surplus_huge_pages,
 700                         HPAGE_SIZE/1024);
 701 }
 702
 703 int hugetlb_report_node_meminfo(int nid, char *buf)
 704 {
 705         return sprintf(buf,
 706                 "Node %d HugePages_Total: %5u\n"
 707                 "Node %d HugePages_Free:  %5u\n"
 708                 "Node %d HugePages_Surp:  %5u\n",
 709                 nid, nr_huge_pages_node[nid],
 710                 nid, free_huge_pages_node[nid],
 711                 nid, surplus_huge_pages_node[nid]);
 712 }
 713
 714 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 715 unsigned long hugetlb_total_pages(void)
 716 {
 717         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 718 }
 719
 720 /*
 721  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 722  * handle_mm_fault() to try to instantiate regular-sized pages in the
 723  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 724  * this far.
 725  */
 726 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 727 {
 728         BUG();
 729         return 0;
 730 }
 731
 732 struct vm_operations_struct hugetlb_vm_ops = {
 733         .fault = hugetlb_vm_op_fault,
 734 };
 735
 736 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 737                                 int writable)
 738 {
 739         pte_t entry;
 740
 741         if (writable) {
 742                 entry =
 743                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 744         } else {
 745                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 746         }
 747         entry = pte_mkyoung(entry);
 748         entry = pte_mkhuge(entry);
 749
 750         return entry;
 751 }
 752
 753 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 754                                    unsigned long address, pte_t *ptep)
 755 {
 756         pte_t entry;
 757
 758         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
 759         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
 760                 update_mmu_cache(vma, address, entry);
 761         }
 762 }
 763
 764
 765 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 766                             struct vm_area_struct *vma)
 767 {
 768         pte_t *src_pte, *dst_pte, entry;
 769         struct page *ptepage;
 770         unsigned long addr;
 771         int cow;
 772
 773         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 774
 775         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 776                 src_pte = huge_pte_offset(src, addr);
 777                 if (!src_pte)
 778                         continue;
 779                 dst_pte = huge_pte_alloc(dst, addr);
 780                 if (!dst_pte)
 781                         goto nomem;
 782
 783                 /* If the pagetables are shared don't copy or take references */
 784                 if (dst_pte == src_pte)
 785                         continue;
 786
 787                 spin_lock(&dst->page_table_lock);
 788                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
 789                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
 790                         if (cow)
 791                                 huge_ptep_set_wrprotect(src, addr, src_pte);
 792                         entry = huge_ptep_get(src_pte);
 793                         ptepage = pte_page(entry);
 794                         get_page(ptepage);
 795                         set_huge_pte_at(dst, addr, dst_pte, entry);
 796                 }
 797                 spin_unlock(&src->page_table_lock);
 798                 spin_unlock(&dst->page_table_lock);
 799         }
 800         return 0;
 801
 802 nomem:
 803         return -ENOMEM;
 804 }
 805
 806 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 807                             unsigned long end)
 808 {
 809         struct mm_struct *mm = vma->vm_mm;
 810         unsigned long address;
 811         pte_t *ptep;
 812         pte_t pte;
 813         struct page *page;
 814         struct page *tmp;
 815         /*
 816          * A page gathering list, protected by per file i_mmap_lock. The
 817          * lock is used to avoid list corruption from multiple unmapping
 818          * of the same page since we are using page->lru.
 819          */
 820         LIST_HEAD(page_list);
 821
 822         WARN_ON(!is_vm_hugetlb_page(vma));
 823         BUG_ON(start & ~HPAGE_MASK);
 824         BUG_ON(end & ~HPAGE_MASK);
 825
 826         spin_lock(&mm->page_table_lock);
 827         for (address = start; address < end; address += HPAGE_SIZE) {
 828                 ptep = huge_pte_offset(mm, address);
 829                 if (!ptep)
 830                         continue;
 831
 832                 if (huge_pmd_unshare(mm, &address, ptep))
 833                         continue;
 834
 835                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 836                 if (huge_pte_none(pte))
 837                         continue;
 838
 839                 page = pte_page(pte);
 840                 if (pte_dirty(pte))
 841                         set_page_dirty(page);
 842                 list_add(&page->lru, &page_list);
 843         }
 844         spin_unlock(&mm->page_table_lock);
 845         flush_tlb_range(vma, start, end);
 846         list_for_each_entry_safe(page, tmp, &page_list, lru) {
 847                 list_del(&page->lru);
 848                 put_page(page);
 849         }
 850 }
 851
 852 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 853                           unsigned long end)
 854 {
 855         /*
 856          * It is undesirable to test vma->vm_file as it should be non-null
 857          * for valid hugetlb area. However, vm_file will be NULL in the error
 858          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
 859          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
 860          * to clean up. Since no pte has actually been setup, it is safe to
 861          * do nothing in this case.
 862          */
 863         if (vma->vm_file) {
 864                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
 865                 __unmap_hugepage_range(vma, start, end);
 866                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 867         }
 868 }
 869
 870 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 871                         unsigned long address, pte_t *ptep, pte_t pte)
 872 {
 873         struct page *old_page, *new_page;
 874         int avoidcopy;
 875
 876         old_page = pte_page(pte);
 877
 878         /* If no-one else is actually using this page, avoid the copy
 879          * and just make the page writable */
 880         avoidcopy = (page_count(old_page) == 1);
 881         if (avoidcopy) {
 882                 set_huge_ptep_writable(vma, address, ptep);
 883                 return 0;
 884         }
 885
 886         page_cache_get(old_page);
 887         new_page = alloc_huge_page(vma, address);
 888
 889         if (IS_ERR(new_page)) {
 890                 page_cache_release(old_page);
 891                 return -PTR_ERR(new_page);
 892         }
 893
 894         spin_unlock(&mm->page_table_lock);
 895         copy_huge_page(new_page, old_page, address, vma);
 896         __SetPageUptodate(new_page);
 897         spin_lock(&mm->page_table_lock);
 898
 899         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 900         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
 901                 /* Break COW */
 902                 huge_ptep_clear_flush(vma, address, ptep);
 903                 set_huge_pte_at(mm, address, ptep,
 904                                 make_huge_pte(vma, new_page, 1));
 905                 /* Make the old page be freed below */
 906                 new_page = old_page;
 907         }
 908         page_cache_release(new_page);
 909         page_cache_release(old_page);
 910         return 0;
 911 }
 912
 913 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 914                         unsigned long address, pte_t *ptep, int write_access)
 915 {
 916         int ret = VM_FAULT_SIGBUS;
 917         unsigned long idx;
 918         unsigned long size;
 919         struct page *page;
 920         struct address_space *mapping;
 921         pte_t new_pte;
 922
 923         mapping = vma->vm_file->f_mapping;
 924         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 925                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 926
 927         /*
 928          * Use page lock to guard against racing truncation
 929          * before we get page_table_lock.
 930          */
 931 retry:
 932         page = find_lock_page(mapping, idx);
 933         if (!page) {
 934                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 935                 if (idx >= size)
 936                         goto out;
 937                 page = alloc_huge_page(vma, address);
 938                 if (IS_ERR(page)) {
 939                         ret = -PTR_ERR(page);
 940                         goto out;
 941                 }
 942                 clear_huge_page(page, address);
 943                 __SetPageUptodate(page);
 944
 945                 if (vma->vm_flags & VM_SHARED) {
 946                         int err;
 947                         struct inode *inode = mapping->host;
 948
 949                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 950                         if (err) {
 951                                 put_page(page);
 952                                 if (err == -EEXIST)
 953                                         goto retry;
 954                                 goto out;
 955                         }
 956
 957                         spin_lock(&inode->i_lock);
 958                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
 959                         spin_unlock(&inode->i_lock);
 960                 } else
 961                         lock_page(page);
 962         }
 963
 964         spin_lock(&mm->page_table_lock);
 965         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 966         if (idx >= size)
 967                 goto backout;
 968
 969         ret = 0;
 970         if (!huge_pte_none(huge_ptep_get(ptep)))
 971                 goto backout;
 972
 973         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 974                                 && (vma->vm_flags & VM_SHARED)));
 975         set_huge_pte_at(mm, address, ptep, new_pte);
 976
 977         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 978                 /* Optimization, do the COW without a second fault */
 979                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 980         }
 981
 982         spin_unlock(&mm->page_table_lock);
 983         unlock_page(page);
 984 out:
 985         return ret;
 986
 987 backout:
 988         spin_unlock(&mm->page_table_lock);
 989         unlock_page(page);
 990         put_page(page);
 991         goto out;
 992 }
 993
 994 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 995                         unsigned long address, int write_access)
 996 {
 997         pte_t *ptep;
 998         pte_t entry;
 999         int ret;
1000         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1001
1002         ptep = huge_pte_alloc(mm, address);
1003         if (!ptep)
1004                 return VM_FAULT_OOM;
1005
1006         /*
1007          * Serialize hugepage allocation and instantiation, so that we don't
1008          * get spurious allocation failures if two CPUs race to instantiate
1009          * the same page in the page cache.
1010          */
1011         mutex_lock(&hugetlb_instantiation_mutex);
1012         entry = huge_ptep_get(ptep);
1013         if (huge_pte_none(entry)) {
1014                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1015                 mutex_unlock(&hugetlb_instantiation_mutex);
1016                 return ret;
1017         }
1018
1019         ret = 0;
1020
1021         spin_lock(&mm->page_table_lock);
1022         /* Check for a racing update before calling hugetlb_cow */
1023         if (likely(pte_same(entry, huge_ptep_get(ptep))))
1024                 if (write_access && !pte_write(entry))
1025                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
1026         spin_unlock(&mm->page_table_lock);
1027         mutex_unlock(&hugetlb_instantiation_mutex);
1028
1029         return ret;
1030 }
1031
1032 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1033                         struct page **pages, struct vm_area_struct **vmas,
1034                         unsigned long *position, int *length, int i,
1035                         int write)
1036 {
1037         unsigned long pfn_offset;
1038         unsigned long vaddr = *position;
1039         int remainder = *length;
1040
1041         spin_lock(&mm->page_table_lock);
1042         while (vaddr < vma->vm_end && remainder) {
1043                 pte_t *pte;
1044                 struct page *page;
1045
1046                 /*
1047                  * Some archs (sparc64, sh*) have multiple pte_ts to
1048                  * each hugepage.  We have to make * sure we get the
1049                  * first, for the page indexing below to work.
1050                  */
1051                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1052
1053                 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1054                     (write && !pte_write(huge_ptep_get(pte)))) {
1055                         int ret;
1056
1057                         spin_unlock(&mm->page_table_lock);
1058                         ret = hugetlb_fault(mm, vma, vaddr, write);
1059                         spin_lock(&mm->page_table_lock);
1060                         if (!(ret & VM_FAULT_ERROR))
1061                                 continue;
1062
1063                         remainder = 0;
1064                         if (!i)
1065                                 i = -EFAULT;
1066                         break;
1067                 }
1068
1069                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1070                 page = pte_page(huge_ptep_get(pte));
1071 same_page:
1072                 if (pages) {
1073                         get_page(page);
1074                         pages[i] = page + pfn_offset;
1075                 }
1076
1077                 if (vmas)
1078                         vmas[i] = vma;
1079
1080                 vaddr += PAGE_SIZE;
1081                 ++pfn_offset;
1082                 --remainder;
1083                 ++i;
1084                 if (vaddr < vma->vm_end && remainder &&
1085                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1086                         /*
1087                          * We use pfn_offset to avoid touching the pageframes
1088                          * of this compound page.
1089                          */
1090                         goto same_page;
1091                 }
1092         }
1093         spin_unlock(&mm->page_table_lock);
1094         *length = remainder;
1095         *position = vaddr;
1096
1097         return i;
1098 }
1099
1100 void hugetlb_change_protection(struct vm_area_struct *vma,
1101                 unsigned long address, unsigned long end, pgprot_t newprot)
1102 {
1103         struct mm_struct *mm = vma->vm_mm;
1104         unsigned long start = address;
1105         pte_t *ptep;
1106         pte_t pte;
1107
1108         BUG_ON(address >= end);
1109         flush_cache_range(vma, address, end);
1110
1111         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1112         spin_lock(&mm->page_table_lock);
1113         for (; address < end; address += HPAGE_SIZE) {
1114                 ptep = huge_pte_offset(mm, address);
1115                 if (!ptep)
1116                         continue;
1117                 if (huge_pmd_unshare(mm, &address, ptep))
1118                         continue;
1119                 if (!huge_pte_none(huge_ptep_get(ptep))) {
1120                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1121                         pte = pte_mkhuge(pte_modify(pte, newprot));
1122                         set_huge_pte_at(mm, address, ptep, pte);
1123                 }
1124         }
1125         spin_unlock(&mm->page_table_lock);
1126         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1127
1128         flush_tlb_range(vma, start, end);
1129 }
1130
1131 struct file_region {
1132         struct list_head link;
1133         long from;
1134         long to;
1135 };
1136
1137 static long region_add(struct list_head *head, long f, long t)
1138 {
1139         struct file_region *rg, *nrg, *trg;
1140
1141         /* Locate the region we are either in or before. */
1142         list_for_each_entry(rg, head, link)
1143                 if (f <= rg->to)
1144                         break;
1145
1146         /* Round our left edge to the current segment if it encloses us. */
1147         if (f > rg->from)
1148                 f = rg->from;
1149
1150         /* Check for and consume any regions we now overlap with. */
1151         nrg = rg;
1152         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1153                 if (&rg->link == head)
1154                         break;
1155                 if (rg->from > t)
1156                         break;
1157
1158                 /* If this area reaches higher then extend our area to
1159                  * include it completely.  If this is not the first area
1160                  * which we intend to reuse, free it. */
1161                 if (rg->to > t)
1162                         t = rg->to;
1163                 if (rg != nrg) {
1164                         list_del(&rg->link);
1165                         kfree(rg);
1166                 }
1167         }
1168         nrg->from = f;
1169         nrg->to = t;
1170         return 0;
1171 }
1172
1173 static long region_chg(struct list_head *head, long f, long t)
1174 {
1175         struct file_region *rg, *nrg;
1176         long chg = 0;
1177
1178         /* Locate the region we are before or in. */
1179         list_for_each_entry(rg, head, link)
1180                 if (f <= rg->to)
1181                         break;
1182
1183         /* If we are below the current region then a new region is required.
1184          * Subtle, allocate a new region at the position but make it zero
1185          * size such that we can guarantee to record the reservation. */
1186         if (&rg->link == head || t < rg->from) {
1187                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1188                 if (!nrg)
1189                         return -ENOMEM;
1190                 nrg->from = f;
1191                 nrg->to   = f;
1192                 INIT_LIST_HEAD(&nrg->link);
1193                 list_add(&nrg->link, rg->link.prev);
1194
1195                 return t - f;
1196         }
1197
1198         /* Round our left edge to the current segment if it encloses us. */
1199         if (f > rg->from)
1200                 f = rg->from;
1201         chg = t - f;
1202
1203         /* Check for and consume any regions we now overlap with. */
1204         list_for_each_entry(rg, rg->link.prev, link) {
1205                 if (&rg->link == head)
1206                         break;
1207                 if (rg->from > t)
1208                         return chg;
1209
1210                 /* We overlap with this area, if it extends futher than
1211                  * us then we must extend ourselves.  Account for its
1212                  * existing reservation. */
1213                 if (rg->to > t) {
1214                         chg += rg->to - t;
1215                         t = rg->to;
1216                 }
1217                 chg -= rg->to - rg->from;
1218         }
1219         return chg;
1220 }
1221
1222 static long region_truncate(struct list_head *head, long end)
1223 {
1224         struct file_region *rg, *trg;
1225         long chg = 0;
1226
1227         /* Locate the region we are either in or before. */
1228         list_for_each_entry(rg, head, link)
1229                 if (end <= rg->to)
1230                         break;
1231         if (&rg->link == head)
1232                 return 0;
1233
1234         /* If we are in the middle of a region then adjust it. */
1235         if (end > rg->from) {
1236                 chg = rg->to - end;
1237                 rg->to = end;
1238                 rg = list_entry(rg->link.next, typeof(*rg), link);
1239         }
1240
1241         /* Drop any remaining regions. */
1242         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1243                 if (&rg->link == head)
1244                         break;
1245                 chg += rg->to - rg->from;
1246                 list_del(&rg->link);
1247                 kfree(rg);
1248         }
1249         return chg;
1250 }
1251
1252 static int hugetlb_acct_memory(long delta)
1253 {
1254         int ret = -ENOMEM;
1255
1256         spin_lock(&hugetlb_lock);
1257         /*
1258          * When cpuset is configured, it breaks the strict hugetlb page
1259          * reservation as the accounting is done on a global variable. Such
1260          * reservation is completely rubbish in the presence of cpuset because
1261          * the reservation is not checked against page availability for the
1262          * current cpuset. Application can still potentially OOM'ed by kernel
1263          * with lack of free htlb page in cpuset that the task is in.
1264          * Attempt to enforce strict accounting with cpuset is almost
1265          * impossible (or too ugly) because cpuset is too fluid that
1266          * task or memory node can be dynamically moved between cpusets.
1267          *
1268          * The change of semantics for shared hugetlb mapping with cpuset is
1269          * undesirable. However, in order to preserve some of the semantics,
1270          * we fall back to check against current free page availability as
1271          * a best attempt and hopefully to minimize the impact of changing
1272          * semantics that cpuset has.
1273          */
1274         if (delta > 0) {
1275                 if (gather_surplus_pages(delta) < 0)
1276                         goto out;
1277
1278                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1279                         return_unused_surplus_pages(delta);
1280                         goto out;
1281                 }
1282         }
1283
1284         ret = 0;
1285         if (delta < 0)
1286                 return_unused_surplus_pages((unsigned long) -delta);
1287
1288 out:
1289         spin_unlock(&hugetlb_lock);
1290         return ret;
1291 }
1292
1293 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1294 {
1295         long ret, chg;
1296
1297         chg = region_chg(&inode->i_mapping->private_list, from, to);
1298         if (chg < 0)
1299                 return chg;
1300
1301         if (hugetlb_get_quota(inode->i_mapping, chg))
1302                 return -ENOSPC;
1303         ret = hugetlb_acct_memory(chg);
1304         if (ret < 0) {
1305                 hugetlb_put_quota(inode->i_mapping, chg);
1306                 return ret;
1307         }
1308         region_add(&inode->i_mapping->private_list, from, to);
1309         return 0;
1310 }
1311
1312 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1313 {
1314         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1315
1316         spin_lock(&inode->i_lock);
1317         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1318         spin_unlock(&inode->i_lock);
1319
1320         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1321         hugetlb_acct_memory(-(chg - freed));
1322 }