hugetlb.c revision 53ba51d21d6e048424ab8aadfebdb1f25ae07b60
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#include <linux/bootmem.h> 18#include <linux/sysfs.h> 19 20#include <asm/page.h> 21#include <asm/pgtable.h> 22 23#include <linux/hugetlb.h> 24#include "internal.h" 25 26const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 27static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 28unsigned long hugepages_treat_as_movable; 29 30static int max_hstate; 31unsigned int default_hstate_idx; 32struct hstate hstates[HUGE_MAX_HSTATE]; 33 34__initdata LIST_HEAD(huge_boot_pages); 35 36/* for command line parsing */ 37static struct hstate * __initdata parsed_hstate; 38static unsigned long __initdata default_hstate_max_huge_pages; 39static unsigned long __initdata default_hstate_size; 40 41#define for_each_hstate(h) \ 42 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) 43 44/* 45 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 46 */ 47static DEFINE_SPINLOCK(hugetlb_lock); 48 49/* 50 * Region tracking -- allows tracking of reservations and instantiated pages 51 * across the pages in a mapping. 52 * 53 * The region data structures are protected by a combination of the mmap_sem 54 * and the hugetlb_instantion_mutex. To access or modify a region the caller 55 * must either hold the mmap_sem for write, or the mmap_sem for read and 56 * the hugetlb_instantiation mutex: 57 * 58 * down_write(&mm->mmap_sem); 59 * or 60 * down_read(&mm->mmap_sem); 61 * mutex_lock(&hugetlb_instantiation_mutex); 62 */ 63struct file_region { 64 struct list_head link; 65 long from; 66 long to; 67}; 68 69static long region_add(struct list_head *head, long f, long t) 70{ 71 struct file_region *rg, *nrg, *trg; 72 73 /* Locate the region we are either in or before. */ 74 list_for_each_entry(rg, head, link) 75 if (f <= rg->to) 76 break; 77 78 /* Round our left edge to the current segment if it encloses us. */ 79 if (f > rg->from) 80 f = rg->from; 81 82 /* Check for and consume any regions we now overlap with. */ 83 nrg = rg; 84 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 85 if (&rg->link == head) 86 break; 87 if (rg->from > t) 88 break; 89 90 /* If this area reaches higher then extend our area to 91 * include it completely. If this is not the first area 92 * which we intend to reuse, free it. */ 93 if (rg->to > t) 94 t = rg->to; 95 if (rg != nrg) { 96 list_del(&rg->link); 97 kfree(rg); 98 } 99 } 100 nrg->from = f; 101 nrg->to = t; 102 return 0; 103} 104 105static long region_chg(struct list_head *head, long f, long t) 106{ 107 struct file_region *rg, *nrg; 108 long chg = 0; 109 110 /* Locate the region we are before or in. */ 111 list_for_each_entry(rg, head, link) 112 if (f <= rg->to) 113 break; 114 115 /* If we are below the current region then a new region is required. 116 * Subtle, allocate a new region at the position but make it zero 117 * size such that we can guarantee to record the reservation. */ 118 if (&rg->link == head || t < rg->from) { 119 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 120 if (!nrg) 121 return -ENOMEM; 122 nrg->from = f; 123 nrg->to = f; 124 INIT_LIST_HEAD(&nrg->link); 125 list_add(&nrg->link, rg->link.prev); 126 127 return t - f; 128 } 129 130 /* Round our left edge to the current segment if it encloses us. */ 131 if (f > rg->from) 132 f = rg->from; 133 chg = t - f; 134 135 /* Check for and consume any regions we now overlap with. */ 136 list_for_each_entry(rg, rg->link.prev, link) { 137 if (&rg->link == head) 138 break; 139 if (rg->from > t) 140 return chg; 141 142 /* We overlap with this area, if it extends futher than 143 * us then we must extend ourselves. Account for its 144 * existing reservation. */ 145 if (rg->to > t) { 146 chg += rg->to - t; 147 t = rg->to; 148 } 149 chg -= rg->to - rg->from; 150 } 151 return chg; 152} 153 154static long region_truncate(struct list_head *head, long end) 155{ 156 struct file_region *rg, *trg; 157 long chg = 0; 158 159 /* Locate the region we are either in or before. */ 160 list_for_each_entry(rg, head, link) 161 if (end <= rg->to) 162 break; 163 if (&rg->link == head) 164 return 0; 165 166 /* If we are in the middle of a region then adjust it. */ 167 if (end > rg->from) { 168 chg = rg->to - end; 169 rg->to = end; 170 rg = list_entry(rg->link.next, typeof(*rg), link); 171 } 172 173 /* Drop any remaining regions. */ 174 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 175 if (&rg->link == head) 176 break; 177 chg += rg->to - rg->from; 178 list_del(&rg->link); 179 kfree(rg); 180 } 181 return chg; 182} 183 184static long region_count(struct list_head *head, long f, long t) 185{ 186 struct file_region *rg; 187 long chg = 0; 188 189 /* Locate each segment we overlap with, and count that overlap. */ 190 list_for_each_entry(rg, head, link) { 191 int seg_from; 192 int seg_to; 193 194 if (rg->to <= f) 195 continue; 196 if (rg->from >= t) 197 break; 198 199 seg_from = max(rg->from, f); 200 seg_to = min(rg->to, t); 201 202 chg += seg_to - seg_from; 203 } 204 205 return chg; 206} 207 208/* 209 * Convert the address within this vma to the page offset within 210 * the mapping, in pagecache page units; huge pages here. 211 */ 212static pgoff_t vma_hugecache_offset(struct hstate *h, 213 struct vm_area_struct *vma, unsigned long address) 214{ 215 return ((address - vma->vm_start) >> huge_page_shift(h)) + 216 (vma->vm_pgoff >> huge_page_order(h)); 217} 218 219/* 220 * Flags for MAP_PRIVATE reservations. These are stored in the bottom 221 * bits of the reservation map pointer, which are always clear due to 222 * alignment. 223 */ 224#define HPAGE_RESV_OWNER (1UL << 0) 225#define HPAGE_RESV_UNMAPPED (1UL << 1) 226#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 227 228/* 229 * These helpers are used to track how many pages are reserved for 230 * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 231 * is guaranteed to have their future faults succeed. 232 * 233 * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 234 * the reserve counters are updated with the hugetlb_lock held. It is safe 235 * to reset the VMA at fork() time as it is not in use yet and there is no 236 * chance of the global counters getting corrupted as a result of the values. 237 * 238 * The private mapping reservation is represented in a subtly different 239 * manner to a shared mapping. A shared mapping has a region map associated 240 * with the underlying file, this region map represents the backing file 241 * pages which have ever had a reservation assigned which this persists even 242 * after the page is instantiated. A private mapping has a region map 243 * associated with the original mmap which is attached to all VMAs which 244 * reference it, this region map represents those offsets which have consumed 245 * reservation ie. where pages have been instantiated. 246 */ 247static unsigned long get_vma_private_data(struct vm_area_struct *vma) 248{ 249 return (unsigned long)vma->vm_private_data; 250} 251 252static void set_vma_private_data(struct vm_area_struct *vma, 253 unsigned long value) 254{ 255 vma->vm_private_data = (void *)value; 256} 257 258struct resv_map { 259 struct kref refs; 260 struct list_head regions; 261}; 262 263struct resv_map *resv_map_alloc(void) 264{ 265 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 266 if (!resv_map) 267 return NULL; 268 269 kref_init(&resv_map->refs); 270 INIT_LIST_HEAD(&resv_map->regions); 271 272 return resv_map; 273} 274 275void resv_map_release(struct kref *ref) 276{ 277 struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 278 279 /* Clear out any active regions before we release the map. */ 280 region_truncate(&resv_map->regions, 0); 281 kfree(resv_map); 282} 283 284static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 285{ 286 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 287 if (!(vma->vm_flags & VM_SHARED)) 288 return (struct resv_map *)(get_vma_private_data(vma) & 289 ~HPAGE_RESV_MASK); 290 return 0; 291} 292 293static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 294{ 295 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 296 VM_BUG_ON(vma->vm_flags & VM_SHARED); 297 298 set_vma_private_data(vma, (get_vma_private_data(vma) & 299 HPAGE_RESV_MASK) | (unsigned long)map); 300} 301 302static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 303{ 304 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 305 VM_BUG_ON(vma->vm_flags & VM_SHARED); 306 307 set_vma_private_data(vma, get_vma_private_data(vma) | flags); 308} 309 310static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 311{ 312 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 313 314 return (get_vma_private_data(vma) & flag) != 0; 315} 316 317/* Decrement the reserved pages in the hugepage pool by one */ 318static void decrement_hugepage_resv_vma(struct hstate *h, 319 struct vm_area_struct *vma) 320{ 321 if (vma->vm_flags & VM_NORESERVE) 322 return; 323 324 if (vma->vm_flags & VM_SHARED) { 325 /* Shared mappings always use reserves */ 326 h->resv_huge_pages--; 327 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 328 /* 329 * Only the process that called mmap() has reserves for 330 * private mappings. 331 */ 332 h->resv_huge_pages--; 333 } 334} 335 336/* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 337void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 338{ 339 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 340 if (!(vma->vm_flags & VM_SHARED)) 341 vma->vm_private_data = (void *)0; 342} 343 344/* Returns true if the VMA has associated reserve pages */ 345static int vma_has_private_reserves(struct vm_area_struct *vma) 346{ 347 if (vma->vm_flags & VM_SHARED) 348 return 0; 349 if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 350 return 0; 351 return 1; 352} 353 354static void clear_huge_page(struct page *page, 355 unsigned long addr, unsigned long sz) 356{ 357 int i; 358 359 might_sleep(); 360 for (i = 0; i < sz/PAGE_SIZE; i++) { 361 cond_resched(); 362 clear_user_highpage(page + i, addr + i * PAGE_SIZE); 363 } 364} 365 366static void copy_huge_page(struct page *dst, struct page *src, 367 unsigned long addr, struct vm_area_struct *vma) 368{ 369 int i; 370 struct hstate *h = hstate_vma(vma); 371 372 might_sleep(); 373 for (i = 0; i < pages_per_huge_page(h); i++) { 374 cond_resched(); 375 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); 376 } 377} 378 379static void enqueue_huge_page(struct hstate *h, struct page *page) 380{ 381 int nid = page_to_nid(page); 382 list_add(&page->lru, &h->hugepage_freelists[nid]); 383 h->free_huge_pages++; 384 h->free_huge_pages_node[nid]++; 385} 386 387static struct page *dequeue_huge_page(struct hstate *h) 388{ 389 int nid; 390 struct page *page = NULL; 391 392 for (nid = 0; nid < MAX_NUMNODES; ++nid) { 393 if (!list_empty(&h->hugepage_freelists[nid])) { 394 page = list_entry(h->hugepage_freelists[nid].next, 395 struct page, lru); 396 list_del(&page->lru); 397 h->free_huge_pages--; 398 h->free_huge_pages_node[nid]--; 399 break; 400 } 401 } 402 return page; 403} 404 405static struct page *dequeue_huge_page_vma(struct hstate *h, 406 struct vm_area_struct *vma, 407 unsigned long address, int avoid_reserve) 408{ 409 int nid; 410 struct page *page = NULL; 411 struct mempolicy *mpol; 412 nodemask_t *nodemask; 413 struct zonelist *zonelist = huge_zonelist(vma, address, 414 htlb_alloc_mask, &mpol, &nodemask); 415 struct zone *zone; 416 struct zoneref *z; 417 418 /* 419 * A child process with MAP_PRIVATE mappings created by their parent 420 * have no page reserves. This check ensures that reservations are 421 * not "stolen". The child may still get SIGKILLed 422 */ 423 if (!vma_has_private_reserves(vma) && 424 h->free_huge_pages - h->resv_huge_pages == 0) 425 return NULL; 426 427 /* If reserves cannot be used, ensure enough pages are in the pool */ 428 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 429 return NULL; 430 431 for_each_zone_zonelist_nodemask(zone, z, zonelist, 432 MAX_NR_ZONES - 1, nodemask) { 433 nid = zone_to_nid(zone); 434 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && 435 !list_empty(&h->hugepage_freelists[nid])) { 436 page = list_entry(h->hugepage_freelists[nid].next, 437 struct page, lru); 438 list_del(&page->lru); 439 h->free_huge_pages--; 440 h->free_huge_pages_node[nid]--; 441 442 if (!avoid_reserve) 443 decrement_hugepage_resv_vma(h, vma); 444 445 break; 446 } 447 } 448 mpol_cond_put(mpol); 449 return page; 450} 451 452static void update_and_free_page(struct hstate *h, struct page *page) 453{ 454 int i; 455 456 h->nr_huge_pages--; 457 h->nr_huge_pages_node[page_to_nid(page)]--; 458 for (i = 0; i < pages_per_huge_page(h); i++) { 459 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 460 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 461 1 << PG_private | 1<< PG_writeback); 462 } 463 set_compound_page_dtor(page, NULL); 464 set_page_refcounted(page); 465 arch_release_hugepage(page); 466 __free_pages(page, huge_page_order(h)); 467} 468 469struct hstate *size_to_hstate(unsigned long size) 470{ 471 struct hstate *h; 472 473 for_each_hstate(h) { 474 if (huge_page_size(h) == size) 475 return h; 476 } 477 return NULL; 478} 479 480static void free_huge_page(struct page *page) 481{ 482 /* 483 * Can't pass hstate in here because it is called from the 484 * compound page destructor. 485 */ 486 struct hstate *h = page_hstate(page); 487 int nid = page_to_nid(page); 488 struct address_space *mapping; 489 490 mapping = (struct address_space *) page_private(page); 491 set_page_private(page, 0); 492 BUG_ON(page_count(page)); 493 INIT_LIST_HEAD(&page->lru); 494 495 spin_lock(&hugetlb_lock); 496 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 497 update_and_free_page(h, page); 498 h->surplus_huge_pages--; 499 h->surplus_huge_pages_node[nid]--; 500 } else { 501 enqueue_huge_page(h, page); 502 } 503 spin_unlock(&hugetlb_lock); 504 if (mapping) 505 hugetlb_put_quota(mapping, 1); 506} 507 508/* 509 * Increment or decrement surplus_huge_pages. Keep node-specific counters 510 * balanced by operating on them in a round-robin fashion. 511 * Returns 1 if an adjustment was made. 512 */ 513static int adjust_pool_surplus(struct hstate *h, int delta) 514{ 515 static int prev_nid; 516 int nid = prev_nid; 517 int ret = 0; 518 519 VM_BUG_ON(delta != -1 && delta != 1); 520 do { 521 nid = next_node(nid, node_online_map); 522 if (nid == MAX_NUMNODES) 523 nid = first_node(node_online_map); 524 525 /* To shrink on this node, there must be a surplus page */ 526 if (delta < 0 && !h->surplus_huge_pages_node[nid]) 527 continue; 528 /* Surplus cannot exceed the total number of pages */ 529 if (delta > 0 && h->surplus_huge_pages_node[nid] >= 530 h->nr_huge_pages_node[nid]) 531 continue; 532 533 h->surplus_huge_pages += delta; 534 h->surplus_huge_pages_node[nid] += delta; 535 ret = 1; 536 break; 537 } while (nid != prev_nid); 538 539 prev_nid = nid; 540 return ret; 541} 542 543static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 544{ 545 set_compound_page_dtor(page, free_huge_page); 546 spin_lock(&hugetlb_lock); 547 h->nr_huge_pages++; 548 h->nr_huge_pages_node[nid]++; 549 spin_unlock(&hugetlb_lock); 550 put_page(page); /* free it into the hugepage allocator */ 551} 552 553static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 554{ 555 struct page *page; 556 557 if (h->order >= MAX_ORDER) 558 return NULL; 559 560 page = alloc_pages_node(nid, 561 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 562 __GFP_REPEAT|__GFP_NOWARN, 563 huge_page_order(h)); 564 if (page) { 565 if (arch_prepare_hugepage(page)) { 566 __free_pages(page, HUGETLB_PAGE_ORDER); 567 return NULL; 568 } 569 prep_new_huge_page(h, page, nid); 570 } 571 572 return page; 573} 574 575/* 576 * Use a helper variable to find the next node and then 577 * copy it back to hugetlb_next_nid afterwards: 578 * otherwise there's a window in which a racer might 579 * pass invalid nid MAX_NUMNODES to alloc_pages_node. 580 * But we don't need to use a spin_lock here: it really 581 * doesn't matter if occasionally a racer chooses the 582 * same nid as we do. Move nid forward in the mask even 583 * if we just successfully allocated a hugepage so that 584 * the next caller gets hugepages on the next node. 585 */ 586static int hstate_next_node(struct hstate *h) 587{ 588 int next_nid; 589 next_nid = next_node(h->hugetlb_next_nid, node_online_map); 590 if (next_nid == MAX_NUMNODES) 591 next_nid = first_node(node_online_map); 592 h->hugetlb_next_nid = next_nid; 593 return next_nid; 594} 595 596static int alloc_fresh_huge_page(struct hstate *h) 597{ 598 struct page *page; 599 int start_nid; 600 int next_nid; 601 int ret = 0; 602 603 start_nid = h->hugetlb_next_nid; 604 605 do { 606 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid); 607 if (page) 608 ret = 1; 609 next_nid = hstate_next_node(h); 610 } while (!page && h->hugetlb_next_nid != start_nid); 611 612 if (ret) 613 count_vm_event(HTLB_BUDDY_PGALLOC); 614 else 615 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 616 617 return ret; 618} 619 620static struct page *alloc_buddy_huge_page(struct hstate *h, 621 struct vm_area_struct *vma, unsigned long address) 622{ 623 struct page *page; 624 unsigned int nid; 625 626 if (h->order >= MAX_ORDER) 627 return NULL; 628 629 /* 630 * Assume we will successfully allocate the surplus page to 631 * prevent racing processes from causing the surplus to exceed 632 * overcommit 633 * 634 * This however introduces a different race, where a process B 635 * tries to grow the static hugepage pool while alloc_pages() is 636 * called by process A. B will only examine the per-node 637 * counters in determining if surplus huge pages can be 638 * converted to normal huge pages in adjust_pool_surplus(). A 639 * won't be able to increment the per-node counter, until the 640 * lock is dropped by B, but B doesn't drop hugetlb_lock until 641 * no more huge pages can be converted from surplus to normal 642 * state (and doesn't try to convert again). Thus, we have a 643 * case where a surplus huge page exists, the pool is grown, and 644 * the surplus huge page still exists after, even though it 645 * should just have been converted to a normal huge page. This 646 * does not leak memory, though, as the hugepage will be freed 647 * once it is out of use. It also does not allow the counters to 648 * go out of whack in adjust_pool_surplus() as we don't modify 649 * the node values until we've gotten the hugepage and only the 650 * per-node value is checked there. 651 */ 652 spin_lock(&hugetlb_lock); 653 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 654 spin_unlock(&hugetlb_lock); 655 return NULL; 656 } else { 657 h->nr_huge_pages++; 658 h->surplus_huge_pages++; 659 } 660 spin_unlock(&hugetlb_lock); 661 662 page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 663 __GFP_REPEAT|__GFP_NOWARN, 664 huge_page_order(h)); 665 666 spin_lock(&hugetlb_lock); 667 if (page) { 668 /* 669 * This page is now managed by the hugetlb allocator and has 670 * no users -- drop the buddy allocator's reference. 671 */ 672 put_page_testzero(page); 673 VM_BUG_ON(page_count(page)); 674 nid = page_to_nid(page); 675 set_compound_page_dtor(page, free_huge_page); 676 /* 677 * We incremented the global counters already 678 */ 679 h->nr_huge_pages_node[nid]++; 680 h->surplus_huge_pages_node[nid]++; 681 __count_vm_event(HTLB_BUDDY_PGALLOC); 682 } else { 683 h->nr_huge_pages--; 684 h->surplus_huge_pages--; 685 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 686 } 687 spin_unlock(&hugetlb_lock); 688 689 return page; 690} 691 692/* 693 * Increase the hugetlb pool such that it can accomodate a reservation 694 * of size 'delta'. 695 */ 696static int gather_surplus_pages(struct hstate *h, int delta) 697{ 698 struct list_head surplus_list; 699 struct page *page, *tmp; 700 int ret, i; 701 int needed, allocated; 702 703 needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 704 if (needed <= 0) { 705 h->resv_huge_pages += delta; 706 return 0; 707 } 708 709 allocated = 0; 710 INIT_LIST_HEAD(&surplus_list); 711 712 ret = -ENOMEM; 713retry: 714 spin_unlock(&hugetlb_lock); 715 for (i = 0; i < needed; i++) { 716 page = alloc_buddy_huge_page(h, NULL, 0); 717 if (!page) { 718 /* 719 * We were not able to allocate enough pages to 720 * satisfy the entire reservation so we free what 721 * we've allocated so far. 722 */ 723 spin_lock(&hugetlb_lock); 724 needed = 0; 725 goto free; 726 } 727 728 list_add(&page->lru, &surplus_list); 729 } 730 allocated += needed; 731 732 /* 733 * After retaking hugetlb_lock, we need to recalculate 'needed' 734 * because either resv_huge_pages or free_huge_pages may have changed. 735 */ 736 spin_lock(&hugetlb_lock); 737 needed = (h->resv_huge_pages + delta) - 738 (h->free_huge_pages + allocated); 739 if (needed > 0) 740 goto retry; 741 742 /* 743 * The surplus_list now contains _at_least_ the number of extra pages 744 * needed to accomodate the reservation. Add the appropriate number 745 * of pages to the hugetlb pool and free the extras back to the buddy 746 * allocator. Commit the entire reservation here to prevent another 747 * process from stealing the pages as they are added to the pool but 748 * before they are reserved. 749 */ 750 needed += allocated; 751 h->resv_huge_pages += delta; 752 ret = 0; 753free: 754 /* Free the needed pages to the hugetlb pool */ 755 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 756 if ((--needed) < 0) 757 break; 758 list_del(&page->lru); 759 enqueue_huge_page(h, page); 760 } 761 762 /* Free unnecessary surplus pages to the buddy allocator */ 763 if (!list_empty(&surplus_list)) { 764 spin_unlock(&hugetlb_lock); 765 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 766 list_del(&page->lru); 767 /* 768 * The page has a reference count of zero already, so 769 * call free_huge_page directly instead of using 770 * put_page. This must be done with hugetlb_lock 771 * unlocked which is safe because free_huge_page takes 772 * hugetlb_lock before deciding how to free the page. 773 */ 774 free_huge_page(page); 775 } 776 spin_lock(&hugetlb_lock); 777 } 778 779 return ret; 780} 781 782/* 783 * When releasing a hugetlb pool reservation, any surplus pages that were 784 * allocated to satisfy the reservation must be explicitly freed if they were 785 * never used. 786 */ 787static void return_unused_surplus_pages(struct hstate *h, 788 unsigned long unused_resv_pages) 789{ 790 static int nid = -1; 791 struct page *page; 792 unsigned long nr_pages; 793 794 /* 795 * We want to release as many surplus pages as possible, spread 796 * evenly across all nodes. Iterate across all nodes until we 797 * can no longer free unreserved surplus pages. This occurs when 798 * the nodes with surplus pages have no free pages. 799 */ 800 unsigned long remaining_iterations = num_online_nodes(); 801 802 /* Uncommit the reservation */ 803 h->resv_huge_pages -= unused_resv_pages; 804 805 /* Cannot return gigantic pages currently */ 806 if (h->order >= MAX_ORDER) 807 return; 808 809 nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 810 811 while (remaining_iterations-- && nr_pages) { 812 nid = next_node(nid, node_online_map); 813 if (nid == MAX_NUMNODES) 814 nid = first_node(node_online_map); 815 816 if (!h->surplus_huge_pages_node[nid]) 817 continue; 818 819 if (!list_empty(&h->hugepage_freelists[nid])) { 820 page = list_entry(h->hugepage_freelists[nid].next, 821 struct page, lru); 822 list_del(&page->lru); 823 update_and_free_page(h, page); 824 h->free_huge_pages--; 825 h->free_huge_pages_node[nid]--; 826 h->surplus_huge_pages--; 827 h->surplus_huge_pages_node[nid]--; 828 nr_pages--; 829 remaining_iterations = num_online_nodes(); 830 } 831 } 832} 833 834/* 835 * Determine if the huge page at addr within the vma has an associated 836 * reservation. Where it does not we will need to logically increase 837 * reservation and actually increase quota before an allocation can occur. 838 * Where any new reservation would be required the reservation change is 839 * prepared, but not committed. Once the page has been quota'd allocated 840 * an instantiated the change should be committed via vma_commit_reservation. 841 * No action is required on failure. 842 */ 843static int vma_needs_reservation(struct hstate *h, 844 struct vm_area_struct *vma, unsigned long addr) 845{ 846 struct address_space *mapping = vma->vm_file->f_mapping; 847 struct inode *inode = mapping->host; 848 849 if (vma->vm_flags & VM_SHARED) { 850 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 851 return region_chg(&inode->i_mapping->private_list, 852 idx, idx + 1); 853 854 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 855 return 1; 856 857 } else { 858 int err; 859 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 860 struct resv_map *reservations = vma_resv_map(vma); 861 862 err = region_chg(&reservations->regions, idx, idx + 1); 863 if (err < 0) 864 return err; 865 return 0; 866 } 867} 868static void vma_commit_reservation(struct hstate *h, 869 struct vm_area_struct *vma, unsigned long addr) 870{ 871 struct address_space *mapping = vma->vm_file->f_mapping; 872 struct inode *inode = mapping->host; 873 874 if (vma->vm_flags & VM_SHARED) { 875 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 876 region_add(&inode->i_mapping->private_list, idx, idx + 1); 877 878 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 879 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 880 struct resv_map *reservations = vma_resv_map(vma); 881 882 /* Mark this page used in the map. */ 883 region_add(&reservations->regions, idx, idx + 1); 884 } 885} 886 887static struct page *alloc_huge_page(struct vm_area_struct *vma, 888 unsigned long addr, int avoid_reserve) 889{ 890 struct hstate *h = hstate_vma(vma); 891 struct page *page; 892 struct address_space *mapping = vma->vm_file->f_mapping; 893 struct inode *inode = mapping->host; 894 unsigned int chg; 895 896 /* 897 * Processes that did not create the mapping will have no reserves and 898 * will not have accounted against quota. Check that the quota can be 899 * made before satisfying the allocation 900 * MAP_NORESERVE mappings may also need pages and quota allocated 901 * if no reserve mapping overlaps. 902 */ 903 chg = vma_needs_reservation(h, vma, addr); 904 if (chg < 0) 905 return ERR_PTR(chg); 906 if (chg) 907 if (hugetlb_get_quota(inode->i_mapping, chg)) 908 return ERR_PTR(-ENOSPC); 909 910 spin_lock(&hugetlb_lock); 911 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 912 spin_unlock(&hugetlb_lock); 913 914 if (!page) { 915 page = alloc_buddy_huge_page(h, vma, addr); 916 if (!page) { 917 hugetlb_put_quota(inode->i_mapping, chg); 918 return ERR_PTR(-VM_FAULT_OOM); 919 } 920 } 921 922 set_page_refcounted(page); 923 set_page_private(page, (unsigned long) mapping); 924 925 vma_commit_reservation(h, vma, addr); 926 927 return page; 928} 929 930__attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h) 931{ 932 struct huge_bootmem_page *m; 933 int nr_nodes = nodes_weight(node_online_map); 934 935 while (nr_nodes) { 936 void *addr; 937 938 addr = __alloc_bootmem_node_nopanic( 939 NODE_DATA(h->hugetlb_next_nid), 940 huge_page_size(h), huge_page_size(h), 0); 941 942 if (addr) { 943 /* 944 * Use the beginning of the huge page to store the 945 * huge_bootmem_page struct (until gather_bootmem 946 * puts them into the mem_map). 947 */ 948 m = addr; 949 if (m) 950 goto found; 951 } 952 hstate_next_node(h); 953 nr_nodes--; 954 } 955 return 0; 956 957found: 958 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 959 /* Put them into a private list first because mem_map is not up yet */ 960 list_add(&m->list, &huge_boot_pages); 961 m->hstate = h; 962 return 1; 963} 964 965/* Put bootmem huge pages into the standard lists after mem_map is up */ 966static void __init gather_bootmem_prealloc(void) 967{ 968 struct huge_bootmem_page *m; 969 970 list_for_each_entry(m, &huge_boot_pages, list) { 971 struct page *page = virt_to_page(m); 972 struct hstate *h = m->hstate; 973 __ClearPageReserved(page); 974 WARN_ON(page_count(page) != 1); 975 prep_compound_page(page, h->order); 976 prep_new_huge_page(h, page, page_to_nid(page)); 977 } 978} 979 980static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 981{ 982 unsigned long i; 983 984 for (i = 0; i < h->max_huge_pages; ++i) { 985 if (h->order >= MAX_ORDER) { 986 if (!alloc_bootmem_huge_page(h)) 987 break; 988 } else if (!alloc_fresh_huge_page(h)) 989 break; 990 } 991 h->max_huge_pages = i; 992} 993 994static void __init hugetlb_init_hstates(void) 995{ 996 struct hstate *h; 997 998 for_each_hstate(h) { 999 /* oversize hugepages were init'ed in early boot */ 1000 if (h->order < MAX_ORDER) 1001 hugetlb_hstate_alloc_pages(h); 1002 } 1003} 1004 1005static char * __init memfmt(char *buf, unsigned long n) 1006{ 1007 if (n >= (1UL << 30)) 1008 sprintf(buf, "%lu GB", n >> 30); 1009 else if (n >= (1UL << 20)) 1010 sprintf(buf, "%lu MB", n >> 20); 1011 else 1012 sprintf(buf, "%lu KB", n >> 10); 1013 return buf; 1014} 1015 1016static void __init report_hugepages(void) 1017{ 1018 struct hstate *h; 1019 1020 for_each_hstate(h) { 1021 char buf[32]; 1022 printk(KERN_INFO "HugeTLB registered %s page size, " 1023 "pre-allocated %ld pages\n", 1024 memfmt(buf, huge_page_size(h)), 1025 h->free_huge_pages); 1026 } 1027} 1028 1029#ifdef CONFIG_SYSCTL 1030#ifdef CONFIG_HIGHMEM 1031static void try_to_free_low(struct hstate *h, unsigned long count) 1032{ 1033 int i; 1034 1035 if (h->order >= MAX_ORDER) 1036 return; 1037 1038 for (i = 0; i < MAX_NUMNODES; ++i) { 1039 struct page *page, *next; 1040 struct list_head *freel = &h->hugepage_freelists[i]; 1041 list_for_each_entry_safe(page, next, freel, lru) { 1042 if (count >= h->nr_huge_pages) 1043 return; 1044 if (PageHighMem(page)) 1045 continue; 1046 list_del(&page->lru); 1047 update_and_free_page(h, page); 1048 h->free_huge_pages--; 1049 h->free_huge_pages_node[page_to_nid(page)]--; 1050 } 1051 } 1052} 1053#else 1054static inline void try_to_free_low(struct hstate *h, unsigned long count) 1055{ 1056} 1057#endif 1058 1059#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1060static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count) 1061{ 1062 unsigned long min_count, ret; 1063 1064 if (h->order >= MAX_ORDER) 1065 return h->max_huge_pages; 1066 1067 /* 1068 * Increase the pool size 1069 * First take pages out of surplus state. Then make up the 1070 * remaining difference by allocating fresh huge pages. 1071 * 1072 * We might race with alloc_buddy_huge_page() here and be unable 1073 * to convert a surplus huge page to a normal huge page. That is 1074 * not critical, though, it just means the overall size of the 1075 * pool might be one hugepage larger than it needs to be, but 1076 * within all the constraints specified by the sysctls. 1077 */ 1078 spin_lock(&hugetlb_lock); 1079 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1080 if (!adjust_pool_surplus(h, -1)) 1081 break; 1082 } 1083 1084 while (count > persistent_huge_pages(h)) { 1085 /* 1086 * If this allocation races such that we no longer need the 1087 * page, free_huge_page will handle it by freeing the page 1088 * and reducing the surplus. 1089 */ 1090 spin_unlock(&hugetlb_lock); 1091 ret = alloc_fresh_huge_page(h); 1092 spin_lock(&hugetlb_lock); 1093 if (!ret) 1094 goto out; 1095 1096 } 1097 1098 /* 1099 * Decrease the pool size 1100 * First return free pages to the buddy allocator (being careful 1101 * to keep enough around to satisfy reservations). Then place 1102 * pages into surplus state as needed so the pool will shrink 1103 * to the desired size as pages become free. 1104 * 1105 * By placing pages into the surplus state independent of the 1106 * overcommit value, we are allowing the surplus pool size to 1107 * exceed overcommit. There are few sane options here. Since 1108 * alloc_buddy_huge_page() is checking the global counter, 1109 * though, we'll note that we're not allowed to exceed surplus 1110 * and won't grow the pool anywhere else. Not until one of the 1111 * sysctls are changed, or the surplus pages go out of use. 1112 */ 1113 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 1114 min_count = max(count, min_count); 1115 try_to_free_low(h, min_count); 1116 while (min_count < persistent_huge_pages(h)) { 1117 struct page *page = dequeue_huge_page(h); 1118 if (!page) 1119 break; 1120 update_and_free_page(h, page); 1121 } 1122 while (count < persistent_huge_pages(h)) { 1123 if (!adjust_pool_surplus(h, 1)) 1124 break; 1125 } 1126out: 1127 ret = persistent_huge_pages(h); 1128 spin_unlock(&hugetlb_lock); 1129 return ret; 1130} 1131 1132#define HSTATE_ATTR_RO(_name) \ 1133 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1134 1135#define HSTATE_ATTR(_name) \ 1136 static struct kobj_attribute _name##_attr = \ 1137 __ATTR(_name, 0644, _name##_show, _name##_store) 1138 1139static struct kobject *hugepages_kobj; 1140static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1141 1142static struct hstate *kobj_to_hstate(struct kobject *kobj) 1143{ 1144 int i; 1145 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1146 if (hstate_kobjs[i] == kobj) 1147 return &hstates[i]; 1148 BUG(); 1149 return NULL; 1150} 1151 1152static ssize_t nr_hugepages_show(struct kobject *kobj, 1153 struct kobj_attribute *attr, char *buf) 1154{ 1155 struct hstate *h = kobj_to_hstate(kobj); 1156 return sprintf(buf, "%lu\n", h->nr_huge_pages); 1157} 1158static ssize_t nr_hugepages_store(struct kobject *kobj, 1159 struct kobj_attribute *attr, const char *buf, size_t count) 1160{ 1161 int err; 1162 unsigned long input; 1163 struct hstate *h = kobj_to_hstate(kobj); 1164 1165 err = strict_strtoul(buf, 10, &input); 1166 if (err) 1167 return 0; 1168 1169 h->max_huge_pages = set_max_huge_pages(h, input); 1170 1171 return count; 1172} 1173HSTATE_ATTR(nr_hugepages); 1174 1175static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1176 struct kobj_attribute *attr, char *buf) 1177{ 1178 struct hstate *h = kobj_to_hstate(kobj); 1179 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1180} 1181static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1182 struct kobj_attribute *attr, const char *buf, size_t count) 1183{ 1184 int err; 1185 unsigned long input; 1186 struct hstate *h = kobj_to_hstate(kobj); 1187 1188 err = strict_strtoul(buf, 10, &input); 1189 if (err) 1190 return 0; 1191 1192 spin_lock(&hugetlb_lock); 1193 h->nr_overcommit_huge_pages = input; 1194 spin_unlock(&hugetlb_lock); 1195 1196 return count; 1197} 1198HSTATE_ATTR(nr_overcommit_hugepages); 1199 1200static ssize_t free_hugepages_show(struct kobject *kobj, 1201 struct kobj_attribute *attr, char *buf) 1202{ 1203 struct hstate *h = kobj_to_hstate(kobj); 1204 return sprintf(buf, "%lu\n", h->free_huge_pages); 1205} 1206HSTATE_ATTR_RO(free_hugepages); 1207 1208static ssize_t resv_hugepages_show(struct kobject *kobj, 1209 struct kobj_attribute *attr, char *buf) 1210{ 1211 struct hstate *h = kobj_to_hstate(kobj); 1212 return sprintf(buf, "%lu\n", h->resv_huge_pages); 1213} 1214HSTATE_ATTR_RO(resv_hugepages); 1215 1216static ssize_t surplus_hugepages_show(struct kobject *kobj, 1217 struct kobj_attribute *attr, char *buf) 1218{ 1219 struct hstate *h = kobj_to_hstate(kobj); 1220 return sprintf(buf, "%lu\n", h->surplus_huge_pages); 1221} 1222HSTATE_ATTR_RO(surplus_hugepages); 1223 1224static struct attribute *hstate_attrs[] = { 1225 &nr_hugepages_attr.attr, 1226 &nr_overcommit_hugepages_attr.attr, 1227 &free_hugepages_attr.attr, 1228 &resv_hugepages_attr.attr, 1229 &surplus_hugepages_attr.attr, 1230 NULL, 1231}; 1232 1233static struct attribute_group hstate_attr_group = { 1234 .attrs = hstate_attrs, 1235}; 1236 1237static int __init hugetlb_sysfs_add_hstate(struct hstate *h) 1238{ 1239 int retval; 1240 1241 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name, 1242 hugepages_kobj); 1243 if (!hstate_kobjs[h - hstates]) 1244 return -ENOMEM; 1245 1246 retval = sysfs_create_group(hstate_kobjs[h - hstates], 1247 &hstate_attr_group); 1248 if (retval) 1249 kobject_put(hstate_kobjs[h - hstates]); 1250 1251 return retval; 1252} 1253 1254static void __init hugetlb_sysfs_init(void) 1255{ 1256 struct hstate *h; 1257 int err; 1258 1259 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1260 if (!hugepages_kobj) 1261 return; 1262 1263 for_each_hstate(h) { 1264 err = hugetlb_sysfs_add_hstate(h); 1265 if (err) 1266 printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1267 h->name); 1268 } 1269} 1270 1271static void __exit hugetlb_exit(void) 1272{ 1273 struct hstate *h; 1274 1275 for_each_hstate(h) { 1276 kobject_put(hstate_kobjs[h - hstates]); 1277 } 1278 1279 kobject_put(hugepages_kobj); 1280} 1281module_exit(hugetlb_exit); 1282 1283static int __init hugetlb_init(void) 1284{ 1285 BUILD_BUG_ON(HPAGE_SHIFT == 0); 1286 1287 if (!size_to_hstate(default_hstate_size)) { 1288 default_hstate_size = HPAGE_SIZE; 1289 if (!size_to_hstate(default_hstate_size)) 1290 hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1291 } 1292 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1293 if (default_hstate_max_huge_pages) 1294 default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1295 1296 hugetlb_init_hstates(); 1297 1298 gather_bootmem_prealloc(); 1299 1300 report_hugepages(); 1301 1302 hugetlb_sysfs_init(); 1303 1304 return 0; 1305} 1306module_init(hugetlb_init); 1307 1308/* Should be called on processing a hugepagesz=... option */ 1309void __init hugetlb_add_hstate(unsigned order) 1310{ 1311 struct hstate *h; 1312 unsigned long i; 1313 1314 if (size_to_hstate(PAGE_SIZE << order)) { 1315 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1316 return; 1317 } 1318 BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1319 BUG_ON(order == 0); 1320 h = &hstates[max_hstate++]; 1321 h->order = order; 1322 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 1323 h->nr_huge_pages = 0; 1324 h->free_huge_pages = 0; 1325 for (i = 0; i < MAX_NUMNODES; ++i) 1326 INIT_LIST_HEAD(&h->hugepage_freelists[i]); 1327 h->hugetlb_next_nid = first_node(node_online_map); 1328 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1329 huge_page_size(h)/1024); 1330 1331 parsed_hstate = h; 1332} 1333 1334static int __init hugetlb_nrpages_setup(char *s) 1335{ 1336 unsigned long *mhp; 1337 static unsigned long *last_mhp; 1338 1339 /* 1340 * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1341 * so this hugepages= parameter goes to the "default hstate". 1342 */ 1343 if (!max_hstate) 1344 mhp = &default_hstate_max_huge_pages; 1345 else 1346 mhp = &parsed_hstate->max_huge_pages; 1347 1348 if (mhp == last_mhp) { 1349 printk(KERN_WARNING "hugepages= specified twice without " 1350 "interleaving hugepagesz=, ignoring\n"); 1351 return 1; 1352 } 1353 1354 if (sscanf(s, "%lu", mhp) <= 0) 1355 *mhp = 0; 1356 1357 /* 1358 * Global state is always initialized later in hugetlb_init. 1359 * But we need to allocate >= MAX_ORDER hstates here early to still 1360 * use the bootmem allocator. 1361 */ 1362 if (max_hstate && parsed_hstate->order >= MAX_ORDER) 1363 hugetlb_hstate_alloc_pages(parsed_hstate); 1364 1365 last_mhp = mhp; 1366 1367 return 1; 1368} 1369__setup("hugepages=", hugetlb_nrpages_setup); 1370 1371static int __init hugetlb_default_setup(char *s) 1372{ 1373 default_hstate_size = memparse(s, &s); 1374 return 1; 1375} 1376__setup("default_hugepagesz=", hugetlb_default_setup); 1377 1378static unsigned int cpuset_mems_nr(unsigned int *array) 1379{ 1380 int node; 1381 unsigned int nr = 0; 1382 1383 for_each_node_mask(node, cpuset_current_mems_allowed) 1384 nr += array[node]; 1385 1386 return nr; 1387} 1388 1389int hugetlb_sysctl_handler(struct ctl_table *table, int write, 1390 struct file *file, void __user *buffer, 1391 size_t *length, loff_t *ppos) 1392{ 1393 struct hstate *h = &default_hstate; 1394 unsigned long tmp; 1395 1396 if (!write) 1397 tmp = h->max_huge_pages; 1398 1399 table->data = &tmp; 1400 table->maxlen = sizeof(unsigned long); 1401 proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 1402 1403 if (write) 1404 h->max_huge_pages = set_max_huge_pages(h, tmp); 1405 1406 return 0; 1407} 1408 1409int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 1410 struct file *file, void __user *buffer, 1411 size_t *length, loff_t *ppos) 1412{ 1413 proc_dointvec(table, write, file, buffer, length, ppos); 1414 if (hugepages_treat_as_movable) 1415 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1416 else 1417 htlb_alloc_mask = GFP_HIGHUSER; 1418 return 0; 1419} 1420 1421int hugetlb_overcommit_handler(struct ctl_table *table, int write, 1422 struct file *file, void __user *buffer, 1423 size_t *length, loff_t *ppos) 1424{ 1425 struct hstate *h = &default_hstate; 1426 unsigned long tmp; 1427 1428 if (!write) 1429 tmp = h->nr_overcommit_huge_pages; 1430 1431 table->data = &tmp; 1432 table->maxlen = sizeof(unsigned long); 1433 proc_doulongvec_minmax(table, write, file, buffer, length, ppos); 1434 1435 if (write) { 1436 spin_lock(&hugetlb_lock); 1437 h->nr_overcommit_huge_pages = tmp; 1438 spin_unlock(&hugetlb_lock); 1439 } 1440 1441 return 0; 1442} 1443 1444#endif /* CONFIG_SYSCTL */ 1445 1446int hugetlb_report_meminfo(char *buf) 1447{ 1448 struct hstate *h = &default_hstate; 1449 return sprintf(buf, 1450 "HugePages_Total: %5lu\n" 1451 "HugePages_Free: %5lu\n" 1452 "HugePages_Rsvd: %5lu\n" 1453 "HugePages_Surp: %5lu\n" 1454 "Hugepagesize: %5lu kB\n", 1455 h->nr_huge_pages, 1456 h->free_huge_pages, 1457 h->resv_huge_pages, 1458 h->surplus_huge_pages, 1459 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 1460} 1461 1462int hugetlb_report_node_meminfo(int nid, char *buf) 1463{ 1464 struct hstate *h = &default_hstate; 1465 return sprintf(buf, 1466 "Node %d HugePages_Total: %5u\n" 1467 "Node %d HugePages_Free: %5u\n" 1468 "Node %d HugePages_Surp: %5u\n", 1469 nid, h->nr_huge_pages_node[nid], 1470 nid, h->free_huge_pages_node[nid], 1471 nid, h->surplus_huge_pages_node[nid]); 1472} 1473 1474/* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 1475unsigned long hugetlb_total_pages(void) 1476{ 1477 struct hstate *h = &default_hstate; 1478 return h->nr_huge_pages * pages_per_huge_page(h); 1479} 1480 1481static int hugetlb_acct_memory(struct hstate *h, long delta) 1482{ 1483 int ret = -ENOMEM; 1484 1485 spin_lock(&hugetlb_lock); 1486 /* 1487 * When cpuset is configured, it breaks the strict hugetlb page 1488 * reservation as the accounting is done on a global variable. Such 1489 * reservation is completely rubbish in the presence of cpuset because 1490 * the reservation is not checked against page availability for the 1491 * current cpuset. Application can still potentially OOM'ed by kernel 1492 * with lack of free htlb page in cpuset that the task is in. 1493 * Attempt to enforce strict accounting with cpuset is almost 1494 * impossible (or too ugly) because cpuset is too fluid that 1495 * task or memory node can be dynamically moved between cpusets. 1496 * 1497 * The change of semantics for shared hugetlb mapping with cpuset is 1498 * undesirable. However, in order to preserve some of the semantics, 1499 * we fall back to check against current free page availability as 1500 * a best attempt and hopefully to minimize the impact of changing 1501 * semantics that cpuset has. 1502 */ 1503 if (delta > 0) { 1504 if (gather_surplus_pages(h, delta) < 0) 1505 goto out; 1506 1507 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 1508 return_unused_surplus_pages(h, delta); 1509 goto out; 1510 } 1511 } 1512 1513 ret = 0; 1514 if (delta < 0) 1515 return_unused_surplus_pages(h, (unsigned long) -delta); 1516 1517out: 1518 spin_unlock(&hugetlb_lock); 1519 return ret; 1520} 1521 1522static void hugetlb_vm_op_open(struct vm_area_struct *vma) 1523{ 1524 struct resv_map *reservations = vma_resv_map(vma); 1525 1526 /* 1527 * This new VMA should share its siblings reservation map if present. 1528 * The VMA will only ever have a valid reservation map pointer where 1529 * it is being copied for another still existing VMA. As that VMA 1530 * has a reference to the reservation map it cannot dissappear until 1531 * after this open call completes. It is therefore safe to take a 1532 * new reference here without additional locking. 1533 */ 1534 if (reservations) 1535 kref_get(&reservations->refs); 1536} 1537 1538static void hugetlb_vm_op_close(struct vm_area_struct *vma) 1539{ 1540 struct hstate *h = hstate_vma(vma); 1541 struct resv_map *reservations = vma_resv_map(vma); 1542 unsigned long reserve; 1543 unsigned long start; 1544 unsigned long end; 1545 1546 if (reservations) { 1547 start = vma_hugecache_offset(h, vma, vma->vm_start); 1548 end = vma_hugecache_offset(h, vma, vma->vm_end); 1549 1550 reserve = (end - start) - 1551 region_count(&reservations->regions, start, end); 1552 1553 kref_put(&reservations->refs, resv_map_release); 1554 1555 if (reserve) 1556 hugetlb_acct_memory(h, -reserve); 1557 } 1558} 1559 1560/* 1561 * We cannot handle pagefaults against hugetlb pages at all. They cause 1562 * handle_mm_fault() to try to instantiate regular-sized pages in the 1563 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 1564 * this far. 1565 */ 1566static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 1567{ 1568 BUG(); 1569 return 0; 1570} 1571 1572struct vm_operations_struct hugetlb_vm_ops = { 1573 .fault = hugetlb_vm_op_fault, 1574 .open = hugetlb_vm_op_open, 1575 .close = hugetlb_vm_op_close, 1576}; 1577 1578static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 1579 int writable) 1580{ 1581 pte_t entry; 1582 1583 if (writable) { 1584 entry = 1585 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 1586 } else { 1587 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 1588 } 1589 entry = pte_mkyoung(entry); 1590 entry = pte_mkhuge(entry); 1591 1592 return entry; 1593} 1594 1595static void set_huge_ptep_writable(struct vm_area_struct *vma, 1596 unsigned long address, pte_t *ptep) 1597{ 1598 pte_t entry; 1599 1600 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 1601 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 1602 update_mmu_cache(vma, address, entry); 1603 } 1604} 1605 1606 1607int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 1608 struct vm_area_struct *vma) 1609{ 1610 pte_t *src_pte, *dst_pte, entry; 1611 struct page *ptepage; 1612 unsigned long addr; 1613 int cow; 1614 struct hstate *h = hstate_vma(vma); 1615 unsigned long sz = huge_page_size(h); 1616 1617 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 1618 1619 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 1620 src_pte = huge_pte_offset(src, addr); 1621 if (!src_pte) 1622 continue; 1623 dst_pte = huge_pte_alloc(dst, addr, sz); 1624 if (!dst_pte) 1625 goto nomem; 1626 1627 /* If the pagetables are shared don't copy or take references */ 1628 if (dst_pte == src_pte) 1629 continue; 1630 1631 spin_lock(&dst->page_table_lock); 1632 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 1633 if (!huge_pte_none(huge_ptep_get(src_pte))) { 1634 if (cow) 1635 huge_ptep_set_wrprotect(src, addr, src_pte); 1636 entry = huge_ptep_get(src_pte); 1637 ptepage = pte_page(entry); 1638 get_page(ptepage); 1639 set_huge_pte_at(dst, addr, dst_pte, entry); 1640 } 1641 spin_unlock(&src->page_table_lock); 1642 spin_unlock(&dst->page_table_lock); 1643 } 1644 return 0; 1645 1646nomem: 1647 return -ENOMEM; 1648} 1649 1650void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 1651 unsigned long end, struct page *ref_page) 1652{ 1653 struct mm_struct *mm = vma->vm_mm; 1654 unsigned long address; 1655 pte_t *ptep; 1656 pte_t pte; 1657 struct page *page; 1658 struct page *tmp; 1659 struct hstate *h = hstate_vma(vma); 1660 unsigned long sz = huge_page_size(h); 1661 1662 /* 1663 * A page gathering list, protected by per file i_mmap_lock. The 1664 * lock is used to avoid list corruption from multiple unmapping 1665 * of the same page since we are using page->lru. 1666 */ 1667 LIST_HEAD(page_list); 1668 1669 WARN_ON(!is_vm_hugetlb_page(vma)); 1670 BUG_ON(start & ~huge_page_mask(h)); 1671 BUG_ON(end & ~huge_page_mask(h)); 1672 1673 spin_lock(&mm->page_table_lock); 1674 for (address = start; address < end; address += sz) { 1675 ptep = huge_pte_offset(mm, address); 1676 if (!ptep) 1677 continue; 1678 1679 if (huge_pmd_unshare(mm, &address, ptep)) 1680 continue; 1681 1682 /* 1683 * If a reference page is supplied, it is because a specific 1684 * page is being unmapped, not a range. Ensure the page we 1685 * are about to unmap is the actual page of interest. 1686 */ 1687 if (ref_page) { 1688 pte = huge_ptep_get(ptep); 1689 if (huge_pte_none(pte)) 1690 continue; 1691 page = pte_page(pte); 1692 if (page != ref_page) 1693 continue; 1694 1695 /* 1696 * Mark the VMA as having unmapped its page so that 1697 * future faults in this VMA will fail rather than 1698 * looking like data was lost 1699 */ 1700 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 1701 } 1702 1703 pte = huge_ptep_get_and_clear(mm, address, ptep); 1704 if (huge_pte_none(pte)) 1705 continue; 1706 1707 page = pte_page(pte); 1708 if (pte_dirty(pte)) 1709 set_page_dirty(page); 1710 list_add(&page->lru, &page_list); 1711 } 1712 spin_unlock(&mm->page_table_lock); 1713 flush_tlb_range(vma, start, end); 1714 list_for_each_entry_safe(page, tmp, &page_list, lru) { 1715 list_del(&page->lru); 1716 put_page(page); 1717 } 1718} 1719 1720void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 1721 unsigned long end, struct page *ref_page) 1722{ 1723 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 1724 __unmap_hugepage_range(vma, start, end, ref_page); 1725 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 1726} 1727 1728/* 1729 * This is called when the original mapper is failing to COW a MAP_PRIVATE 1730 * mappping it owns the reserve page for. The intention is to unmap the page 1731 * from other VMAs and let the children be SIGKILLed if they are faulting the 1732 * same region. 1733 */ 1734int unmap_ref_private(struct mm_struct *mm, 1735 struct vm_area_struct *vma, 1736 struct page *page, 1737 unsigned long address) 1738{ 1739 struct vm_area_struct *iter_vma; 1740 struct address_space *mapping; 1741 struct prio_tree_iter iter; 1742 pgoff_t pgoff; 1743 1744 /* 1745 * vm_pgoff is in PAGE_SIZE units, hence the different calculation 1746 * from page cache lookup which is in HPAGE_SIZE units. 1747 */ 1748 address = address & huge_page_mask(hstate_vma(vma)); 1749 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 1750 + (vma->vm_pgoff >> PAGE_SHIFT); 1751 mapping = (struct address_space *)page_private(page); 1752 1753 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 1754 /* Do not unmap the current VMA */ 1755 if (iter_vma == vma) 1756 continue; 1757 1758 /* 1759 * Unmap the page from other VMAs without their own reserves. 1760 * They get marked to be SIGKILLed if they fault in these 1761 * areas. This is because a future no-page fault on this VMA 1762 * could insert a zeroed page instead of the data existing 1763 * from the time of fork. This would look like data corruption 1764 */ 1765 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 1766 unmap_hugepage_range(iter_vma, 1767 address, address + HPAGE_SIZE, 1768 page); 1769 } 1770 1771 return 1; 1772} 1773 1774static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 1775 unsigned long address, pte_t *ptep, pte_t pte, 1776 struct page *pagecache_page) 1777{ 1778 struct hstate *h = hstate_vma(vma); 1779 struct page *old_page, *new_page; 1780 int avoidcopy; 1781 int outside_reserve = 0; 1782 1783 old_page = pte_page(pte); 1784 1785retry_avoidcopy: 1786 /* If no-one else is actually using this page, avoid the copy 1787 * and just make the page writable */ 1788 avoidcopy = (page_count(old_page) == 1); 1789 if (avoidcopy) { 1790 set_huge_ptep_writable(vma, address, ptep); 1791 return 0; 1792 } 1793 1794 /* 1795 * If the process that created a MAP_PRIVATE mapping is about to 1796 * perform a COW due to a shared page count, attempt to satisfy 1797 * the allocation without using the existing reserves. The pagecache 1798 * page is used to determine if the reserve at this address was 1799 * consumed or not. If reserves were used, a partial faulted mapping 1800 * at the time of fork() could consume its reserves on COW instead 1801 * of the full address range. 1802 */ 1803 if (!(vma->vm_flags & VM_SHARED) && 1804 is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 1805 old_page != pagecache_page) 1806 outside_reserve = 1; 1807 1808 page_cache_get(old_page); 1809 new_page = alloc_huge_page(vma, address, outside_reserve); 1810 1811 if (IS_ERR(new_page)) { 1812 page_cache_release(old_page); 1813 1814 /* 1815 * If a process owning a MAP_PRIVATE mapping fails to COW, 1816 * it is due to references held by a child and an insufficient 1817 * huge page pool. To guarantee the original mappers 1818 * reliability, unmap the page from child processes. The child 1819 * may get SIGKILLed if it later faults. 1820 */ 1821 if (outside_reserve) { 1822 BUG_ON(huge_pte_none(pte)); 1823 if (unmap_ref_private(mm, vma, old_page, address)) { 1824 BUG_ON(page_count(old_page) != 1); 1825 BUG_ON(huge_pte_none(pte)); 1826 goto retry_avoidcopy; 1827 } 1828 WARN_ON_ONCE(1); 1829 } 1830 1831 return -PTR_ERR(new_page); 1832 } 1833 1834 spin_unlock(&mm->page_table_lock); 1835 copy_huge_page(new_page, old_page, address, vma); 1836 __SetPageUptodate(new_page); 1837 spin_lock(&mm->page_table_lock); 1838 1839 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 1840 if (likely(pte_same(huge_ptep_get(ptep), pte))) { 1841 /* Break COW */ 1842 huge_ptep_clear_flush(vma, address, ptep); 1843 set_huge_pte_at(mm, address, ptep, 1844 make_huge_pte(vma, new_page, 1)); 1845 /* Make the old page be freed below */ 1846 new_page = old_page; 1847 } 1848 page_cache_release(new_page); 1849 page_cache_release(old_page); 1850 return 0; 1851} 1852 1853/* Return the pagecache page at a given address within a VMA */ 1854static struct page *hugetlbfs_pagecache_page(struct hstate *h, 1855 struct vm_area_struct *vma, unsigned long address) 1856{ 1857 struct address_space *mapping; 1858 pgoff_t idx; 1859 1860 mapping = vma->vm_file->f_mapping; 1861 idx = vma_hugecache_offset(h, vma, address); 1862 1863 return find_lock_page(mapping, idx); 1864} 1865 1866static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 1867 unsigned long address, pte_t *ptep, int write_access) 1868{ 1869 struct hstate *h = hstate_vma(vma); 1870 int ret = VM_FAULT_SIGBUS; 1871 pgoff_t idx; 1872 unsigned long size; 1873 struct page *page; 1874 struct address_space *mapping; 1875 pte_t new_pte; 1876 1877 /* 1878 * Currently, we are forced to kill the process in the event the 1879 * original mapper has unmapped pages from the child due to a failed 1880 * COW. Warn that such a situation has occured as it may not be obvious 1881 */ 1882 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 1883 printk(KERN_WARNING 1884 "PID %d killed due to inadequate hugepage pool\n", 1885 current->pid); 1886 return ret; 1887 } 1888 1889 mapping = vma->vm_file->f_mapping; 1890 idx = vma_hugecache_offset(h, vma, address); 1891 1892 /* 1893 * Use page lock to guard against racing truncation 1894 * before we get page_table_lock. 1895 */ 1896retry: 1897 page = find_lock_page(mapping, idx); 1898 if (!page) { 1899 size = i_size_read(mapping->host) >> huge_page_shift(h); 1900 if (idx >= size) 1901 goto out; 1902 page = alloc_huge_page(vma, address, 0); 1903 if (IS_ERR(page)) { 1904 ret = -PTR_ERR(page); 1905 goto out; 1906 } 1907 clear_huge_page(page, address, huge_page_size(h)); 1908 __SetPageUptodate(page); 1909 1910 if (vma->vm_flags & VM_SHARED) { 1911 int err; 1912 struct inode *inode = mapping->host; 1913 1914 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 1915 if (err) { 1916 put_page(page); 1917 if (err == -EEXIST) 1918 goto retry; 1919 goto out; 1920 } 1921 1922 spin_lock(&inode->i_lock); 1923 inode->i_blocks += blocks_per_huge_page(h); 1924 spin_unlock(&inode->i_lock); 1925 } else 1926 lock_page(page); 1927 } 1928 1929 spin_lock(&mm->page_table_lock); 1930 size = i_size_read(mapping->host) >> huge_page_shift(h); 1931 if (idx >= size) 1932 goto backout; 1933 1934 ret = 0; 1935 if (!huge_pte_none(huge_ptep_get(ptep))) 1936 goto backout; 1937 1938 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 1939 && (vma->vm_flags & VM_SHARED))); 1940 set_huge_pte_at(mm, address, ptep, new_pte); 1941 1942 if (write_access && !(vma->vm_flags & VM_SHARED)) { 1943 /* Optimization, do the COW without a second fault */ 1944 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 1945 } 1946 1947 spin_unlock(&mm->page_table_lock); 1948 unlock_page(page); 1949out: 1950 return ret; 1951 1952backout: 1953 spin_unlock(&mm->page_table_lock); 1954 unlock_page(page); 1955 put_page(page); 1956 goto out; 1957} 1958 1959int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 1960 unsigned long address, int write_access) 1961{ 1962 pte_t *ptep; 1963 pte_t entry; 1964 int ret; 1965 static DEFINE_MUTEX(hugetlb_instantiation_mutex); 1966 struct hstate *h = hstate_vma(vma); 1967 1968 ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 1969 if (!ptep) 1970 return VM_FAULT_OOM; 1971 1972 /* 1973 * Serialize hugepage allocation and instantiation, so that we don't 1974 * get spurious allocation failures if two CPUs race to instantiate 1975 * the same page in the page cache. 1976 */ 1977 mutex_lock(&hugetlb_instantiation_mutex); 1978 entry = huge_ptep_get(ptep); 1979 if (huge_pte_none(entry)) { 1980 ret = hugetlb_no_page(mm, vma, address, ptep, write_access); 1981 mutex_unlock(&hugetlb_instantiation_mutex); 1982 return ret; 1983 } 1984 1985 ret = 0; 1986 1987 spin_lock(&mm->page_table_lock); 1988 /* Check for a racing update before calling hugetlb_cow */ 1989 if (likely(pte_same(entry, huge_ptep_get(ptep)))) 1990 if (write_access && !pte_write(entry)) { 1991 struct page *page; 1992 page = hugetlbfs_pagecache_page(h, vma, address); 1993 ret = hugetlb_cow(mm, vma, address, ptep, entry, page); 1994 if (page) { 1995 unlock_page(page); 1996 put_page(page); 1997 } 1998 } 1999 spin_unlock(&mm->page_table_lock); 2000 mutex_unlock(&hugetlb_instantiation_mutex); 2001 2002 return ret; 2003} 2004 2005/* Can be overriden by architectures */ 2006__attribute__((weak)) struct page * 2007follow_huge_pud(struct mm_struct *mm, unsigned long address, 2008 pud_t *pud, int write) 2009{ 2010 BUG(); 2011 return NULL; 2012} 2013 2014int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 2015 struct page **pages, struct vm_area_struct **vmas, 2016 unsigned long *position, int *length, int i, 2017 int write) 2018{ 2019 unsigned long pfn_offset; 2020 unsigned long vaddr = *position; 2021 int remainder = *length; 2022 struct hstate *h = hstate_vma(vma); 2023 2024 spin_lock(&mm->page_table_lock); 2025 while (vaddr < vma->vm_end && remainder) { 2026 pte_t *pte; 2027 struct page *page; 2028 2029 /* 2030 * Some archs (sparc64, sh*) have multiple pte_ts to 2031 * each hugepage. We have to make * sure we get the 2032 * first, for the page indexing below to work. 2033 */ 2034 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 2035 2036 if (!pte || huge_pte_none(huge_ptep_get(pte)) || 2037 (write && !pte_write(huge_ptep_get(pte)))) { 2038 int ret; 2039 2040 spin_unlock(&mm->page_table_lock); 2041 ret = hugetlb_fault(mm, vma, vaddr, write); 2042 spin_lock(&mm->page_table_lock); 2043 if (!(ret & VM_FAULT_ERROR)) 2044 continue; 2045 2046 remainder = 0; 2047 if (!i) 2048 i = -EFAULT; 2049 break; 2050 } 2051 2052 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 2053 page = pte_page(huge_ptep_get(pte)); 2054same_page: 2055 if (pages) { 2056 get_page(page); 2057 pages[i] = page + pfn_offset; 2058 } 2059 2060 if (vmas) 2061 vmas[i] = vma; 2062 2063 vaddr += PAGE_SIZE; 2064 ++pfn_offset; 2065 --remainder; 2066 ++i; 2067 if (vaddr < vma->vm_end && remainder && 2068 pfn_offset < pages_per_huge_page(h)) { 2069 /* 2070 * We use pfn_offset to avoid touching the pageframes 2071 * of this compound page. 2072 */ 2073 goto same_page; 2074 } 2075 } 2076 spin_unlock(&mm->page_table_lock); 2077 *length = remainder; 2078 *position = vaddr; 2079 2080 return i; 2081} 2082 2083void hugetlb_change_protection(struct vm_area_struct *vma, 2084 unsigned long address, unsigned long end, pgprot_t newprot) 2085{ 2086 struct mm_struct *mm = vma->vm_mm; 2087 unsigned long start = address; 2088 pte_t *ptep; 2089 pte_t pte; 2090 struct hstate *h = hstate_vma(vma); 2091 2092 BUG_ON(address >= end); 2093 flush_cache_range(vma, address, end); 2094 2095 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 2096 spin_lock(&mm->page_table_lock); 2097 for (; address < end; address += huge_page_size(h)) { 2098 ptep = huge_pte_offset(mm, address); 2099 if (!ptep) 2100 continue; 2101 if (huge_pmd_unshare(mm, &address, ptep)) 2102 continue; 2103 if (!huge_pte_none(huge_ptep_get(ptep))) { 2104 pte = huge_ptep_get_and_clear(mm, address, ptep); 2105 pte = pte_mkhuge(pte_modify(pte, newprot)); 2106 set_huge_pte_at(mm, address, ptep, pte); 2107 } 2108 } 2109 spin_unlock(&mm->page_table_lock); 2110 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2111 2112 flush_tlb_range(vma, start, end); 2113} 2114 2115int hugetlb_reserve_pages(struct inode *inode, 2116 long from, long to, 2117 struct vm_area_struct *vma) 2118{ 2119 long ret, chg; 2120 struct hstate *h = hstate_inode(inode); 2121 2122 if (vma && vma->vm_flags & VM_NORESERVE) 2123 return 0; 2124 2125 /* 2126 * Shared mappings base their reservation on the number of pages that 2127 * are already allocated on behalf of the file. Private mappings need 2128 * to reserve the full area even if read-only as mprotect() may be 2129 * called to make the mapping read-write. Assume !vma is a shm mapping 2130 */ 2131 if (!vma || vma->vm_flags & VM_SHARED) 2132 chg = region_chg(&inode->i_mapping->private_list, from, to); 2133 else { 2134 struct resv_map *resv_map = resv_map_alloc(); 2135 if (!resv_map) 2136 return -ENOMEM; 2137 2138 chg = to - from; 2139 2140 set_vma_resv_map(vma, resv_map); 2141 set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 2142 } 2143 2144 if (chg < 0) 2145 return chg; 2146 2147 if (hugetlb_get_quota(inode->i_mapping, chg)) 2148 return -ENOSPC; 2149 ret = hugetlb_acct_memory(h, chg); 2150 if (ret < 0) { 2151 hugetlb_put_quota(inode->i_mapping, chg); 2152 return ret; 2153 } 2154 if (!vma || vma->vm_flags & VM_SHARED) 2155 region_add(&inode->i_mapping->private_list, from, to); 2156 return 0; 2157} 2158 2159void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2160{ 2161 struct hstate *h = hstate_inode(inode); 2162 long chg = region_truncate(&inode->i_mapping->private_list, offset); 2163 2164 spin_lock(&inode->i_lock); 2165 inode->i_blocks -= blocks_per_huge_page(h); 2166 spin_unlock(&inode->i_lock); 2167 2168 hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2169 hugetlb_acct_memory(h, -(chg - freed)); 2170} 2171