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