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