hugetlb.c revision adbe8726dc2a3805630d517270db17e3af86e526
1/* 2 * Generic hugetlb support. 3 * (C) William Irwin, April 2004 4 */ 5#include <linux/list.h> 6#include <linux/init.h> 7#include <linux/module.h> 8#include <linux/mm.h> 9#include <linux/seq_file.h> 10#include <linux/sysctl.h> 11#include <linux/highmem.h> 12#include <linux/mmu_notifier.h> 13#include <linux/nodemask.h> 14#include <linux/pagemap.h> 15#include <linux/mempolicy.h> 16#include <linux/cpuset.h> 17#include <linux/mutex.h> 18#include <linux/bootmem.h> 19#include <linux/sysfs.h> 20#include <linux/slab.h> 21#include <linux/rmap.h> 22#include <linux/swap.h> 23#include <linux/swapops.h> 24 25#include <asm/page.h> 26#include <asm/pgtable.h> 27#include <asm/io.h> 28 29#include <linux/hugetlb.h> 30#include <linux/node.h> 31#include "internal.h" 32 33const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 34static gfp_t htlb_alloc_mask = GFP_HIGHUSER; 35unsigned long hugepages_treat_as_movable; 36 37static int max_hstate; 38unsigned int default_hstate_idx; 39struct hstate hstates[HUGE_MAX_HSTATE]; 40 41__initdata LIST_HEAD(huge_boot_pages); 42 43/* for command line parsing */ 44static struct hstate * __initdata parsed_hstate; 45static unsigned long __initdata default_hstate_max_huge_pages; 46static unsigned long __initdata default_hstate_size; 47 48#define for_each_hstate(h) \ 49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) 50 51/* 52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages 53 */ 54static DEFINE_SPINLOCK(hugetlb_lock); 55 56/* 57 * Region tracking -- allows tracking of reservations and instantiated pages 58 * across the pages in a mapping. 59 * 60 * The region data structures are protected by a combination of the mmap_sem 61 * and the hugetlb_instantion_mutex. To access or modify a region the caller 62 * must either hold the mmap_sem for write, or the mmap_sem for read and 63 * the hugetlb_instantiation mutex: 64 * 65 * down_write(&mm->mmap_sem); 66 * or 67 * down_read(&mm->mmap_sem); 68 * mutex_lock(&hugetlb_instantiation_mutex); 69 */ 70struct file_region { 71 struct list_head link; 72 long from; 73 long to; 74}; 75 76static long region_add(struct list_head *head, long f, long t) 77{ 78 struct file_region *rg, *nrg, *trg; 79 80 /* Locate the region we are either in or before. */ 81 list_for_each_entry(rg, head, link) 82 if (f <= rg->to) 83 break; 84 85 /* Round our left edge to the current segment if it encloses us. */ 86 if (f > rg->from) 87 f = rg->from; 88 89 /* Check for and consume any regions we now overlap with. */ 90 nrg = rg; 91 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 92 if (&rg->link == head) 93 break; 94 if (rg->from > t) 95 break; 96 97 /* If this area reaches higher then extend our area to 98 * include it completely. If this is not the first area 99 * which we intend to reuse, free it. */ 100 if (rg->to > t) 101 t = rg->to; 102 if (rg != nrg) { 103 list_del(&rg->link); 104 kfree(rg); 105 } 106 } 107 nrg->from = f; 108 nrg->to = t; 109 return 0; 110} 111 112static long region_chg(struct list_head *head, long f, long t) 113{ 114 struct file_region *rg, *nrg; 115 long chg = 0; 116 117 /* Locate the region we are before or in. */ 118 list_for_each_entry(rg, head, link) 119 if (f <= rg->to) 120 break; 121 122 /* If we are below the current region then a new region is required. 123 * Subtle, allocate a new region at the position but make it zero 124 * size such that we can guarantee to record the reservation. */ 125 if (&rg->link == head || t < rg->from) { 126 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 127 if (!nrg) 128 return -ENOMEM; 129 nrg->from = f; 130 nrg->to = f; 131 INIT_LIST_HEAD(&nrg->link); 132 list_add(&nrg->link, rg->link.prev); 133 134 return t - f; 135 } 136 137 /* Round our left edge to the current segment if it encloses us. */ 138 if (f > rg->from) 139 f = rg->from; 140 chg = t - f; 141 142 /* Check for and consume any regions we now overlap with. */ 143 list_for_each_entry(rg, rg->link.prev, link) { 144 if (&rg->link == head) 145 break; 146 if (rg->from > t) 147 return chg; 148 149 /* We overlap with this area, if it extends futher than 150 * us then we must extend ourselves. Account for its 151 * existing reservation. */ 152 if (rg->to > t) { 153 chg += rg->to - t; 154 t = rg->to; 155 } 156 chg -= rg->to - rg->from; 157 } 158 return chg; 159} 160 161static long region_truncate(struct list_head *head, long end) 162{ 163 struct file_region *rg, *trg; 164 long chg = 0; 165 166 /* Locate the region we are either in or before. */ 167 list_for_each_entry(rg, head, link) 168 if (end <= rg->to) 169 break; 170 if (&rg->link == head) 171 return 0; 172 173 /* If we are in the middle of a region then adjust it. */ 174 if (end > rg->from) { 175 chg = rg->to - end; 176 rg->to = end; 177 rg = list_entry(rg->link.next, typeof(*rg), link); 178 } 179 180 /* Drop any remaining regions. */ 181 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 182 if (&rg->link == head) 183 break; 184 chg += rg->to - rg->from; 185 list_del(&rg->link); 186 kfree(rg); 187 } 188 return chg; 189} 190 191static long region_count(struct list_head *head, long f, long t) 192{ 193 struct file_region *rg; 194 long chg = 0; 195 196 /* Locate each segment we overlap with, and count that overlap. */ 197 list_for_each_entry(rg, head, link) { 198 int seg_from; 199 int seg_to; 200 201 if (rg->to <= f) 202 continue; 203 if (rg->from >= t) 204 break; 205 206 seg_from = max(rg->from, f); 207 seg_to = min(rg->to, t); 208 209 chg += seg_to - seg_from; 210 } 211 212 return chg; 213} 214 215/* 216 * Convert the address within this vma to the page offset within 217 * the mapping, in pagecache page units; huge pages here. 218 */ 219static pgoff_t vma_hugecache_offset(struct hstate *h, 220 struct vm_area_struct *vma, unsigned long address) 221{ 222 return ((address - vma->vm_start) >> huge_page_shift(h)) + 223 (vma->vm_pgoff >> huge_page_order(h)); 224} 225 226pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 227 unsigned long address) 228{ 229 return vma_hugecache_offset(hstate_vma(vma), vma, address); 230} 231 232/* 233 * Return the size of the pages allocated when backing a VMA. In the majority 234 * cases this will be same size as used by the page table entries. 235 */ 236unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 237{ 238 struct hstate *hstate; 239 240 if (!is_vm_hugetlb_page(vma)) 241 return PAGE_SIZE; 242 243 hstate = hstate_vma(vma); 244 245 return 1UL << (hstate->order + PAGE_SHIFT); 246} 247EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 248 249/* 250 * Return the page size being used by the MMU to back a VMA. In the majority 251 * of cases, the page size used by the kernel matches the MMU size. On 252 * architectures where it differs, an architecture-specific version of this 253 * function is required. 254 */ 255#ifndef vma_mmu_pagesize 256unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 257{ 258 return vma_kernel_pagesize(vma); 259} 260#endif 261 262/* 263 * Flags for MAP_PRIVATE reservations. These are stored in the bottom 264 * bits of the reservation map pointer, which are always clear due to 265 * alignment. 266 */ 267#define HPAGE_RESV_OWNER (1UL << 0) 268#define HPAGE_RESV_UNMAPPED (1UL << 1) 269#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 270 271/* 272 * These helpers are used to track how many pages are reserved for 273 * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 274 * is guaranteed to have their future faults succeed. 275 * 276 * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 277 * the reserve counters are updated with the hugetlb_lock held. It is safe 278 * to reset the VMA at fork() time as it is not in use yet and there is no 279 * chance of the global counters getting corrupted as a result of the values. 280 * 281 * The private mapping reservation is represented in a subtly different 282 * manner to a shared mapping. A shared mapping has a region map associated 283 * with the underlying file, this region map represents the backing file 284 * pages which have ever had a reservation assigned which this persists even 285 * after the page is instantiated. A private mapping has a region map 286 * associated with the original mmap which is attached to all VMAs which 287 * reference it, this region map represents those offsets which have consumed 288 * reservation ie. where pages have been instantiated. 289 */ 290static unsigned long get_vma_private_data(struct vm_area_struct *vma) 291{ 292 return (unsigned long)vma->vm_private_data; 293} 294 295static void set_vma_private_data(struct vm_area_struct *vma, 296 unsigned long value) 297{ 298 vma->vm_private_data = (void *)value; 299} 300 301struct resv_map { 302 struct kref refs; 303 struct list_head regions; 304}; 305 306static struct resv_map *resv_map_alloc(void) 307{ 308 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 309 if (!resv_map) 310 return NULL; 311 312 kref_init(&resv_map->refs); 313 INIT_LIST_HEAD(&resv_map->regions); 314 315 return resv_map; 316} 317 318static void resv_map_release(struct kref *ref) 319{ 320 struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 321 322 /* Clear out any active regions before we release the map. */ 323 region_truncate(&resv_map->regions, 0); 324 kfree(resv_map); 325} 326 327static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 328{ 329 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 330 if (!(vma->vm_flags & VM_MAYSHARE)) 331 return (struct resv_map *)(get_vma_private_data(vma) & 332 ~HPAGE_RESV_MASK); 333 return NULL; 334} 335 336static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 337{ 338 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 340 341 set_vma_private_data(vma, (get_vma_private_data(vma) & 342 HPAGE_RESV_MASK) | (unsigned long)map); 343} 344 345static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 346{ 347 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 348 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 349 350 set_vma_private_data(vma, get_vma_private_data(vma) | flags); 351} 352 353static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 354{ 355 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 356 357 return (get_vma_private_data(vma) & flag) != 0; 358} 359 360/* Decrement the reserved pages in the hugepage pool by one */ 361static void decrement_hugepage_resv_vma(struct hstate *h, 362 struct vm_area_struct *vma) 363{ 364 if (vma->vm_flags & VM_NORESERVE) 365 return; 366 367 if (vma->vm_flags & VM_MAYSHARE) { 368 /* Shared mappings always use reserves */ 369 h->resv_huge_pages--; 370 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 371 /* 372 * Only the process that called mmap() has reserves for 373 * private mappings. 374 */ 375 h->resv_huge_pages--; 376 } 377} 378 379/* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 380void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 381{ 382 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 383 if (!(vma->vm_flags & VM_MAYSHARE)) 384 vma->vm_private_data = (void *)0; 385} 386 387/* Returns true if the VMA has associated reserve pages */ 388static int vma_has_reserves(struct vm_area_struct *vma) 389{ 390 if (vma->vm_flags & VM_MAYSHARE) 391 return 1; 392 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 393 return 1; 394 return 0; 395} 396 397static void copy_gigantic_page(struct page *dst, struct page *src) 398{ 399 int i; 400 struct hstate *h = page_hstate(src); 401 struct page *dst_base = dst; 402 struct page *src_base = src; 403 404 for (i = 0; i < pages_per_huge_page(h); ) { 405 cond_resched(); 406 copy_highpage(dst, src); 407 408 i++; 409 dst = mem_map_next(dst, dst_base, i); 410 src = mem_map_next(src, src_base, i); 411 } 412} 413 414void copy_huge_page(struct page *dst, struct page *src) 415{ 416 int i; 417 struct hstate *h = page_hstate(src); 418 419 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { 420 copy_gigantic_page(dst, src); 421 return; 422 } 423 424 might_sleep(); 425 for (i = 0; i < pages_per_huge_page(h); i++) { 426 cond_resched(); 427 copy_highpage(dst + i, src + i); 428 } 429} 430 431static void enqueue_huge_page(struct hstate *h, struct page *page) 432{ 433 int nid = page_to_nid(page); 434 list_add(&page->lru, &h->hugepage_freelists[nid]); 435 h->free_huge_pages++; 436 h->free_huge_pages_node[nid]++; 437} 438 439static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 440{ 441 struct page *page; 442 443 if (list_empty(&h->hugepage_freelists[nid])) 444 return NULL; 445 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); 446 list_del(&page->lru); 447 set_page_refcounted(page); 448 h->free_huge_pages--; 449 h->free_huge_pages_node[nid]--; 450 return page; 451} 452 453static struct page *dequeue_huge_page_vma(struct hstate *h, 454 struct vm_area_struct *vma, 455 unsigned long address, int avoid_reserve) 456{ 457 struct page *page = NULL; 458 struct mempolicy *mpol; 459 nodemask_t *nodemask; 460 struct zonelist *zonelist; 461 struct zone *zone; 462 struct zoneref *z; 463 464 get_mems_allowed(); 465 zonelist = huge_zonelist(vma, address, 466 htlb_alloc_mask, &mpol, &nodemask); 467 /* 468 * A child process with MAP_PRIVATE mappings created by their parent 469 * have no page reserves. This check ensures that reservations are 470 * not "stolen". The child may still get SIGKILLed 471 */ 472 if (!vma_has_reserves(vma) && 473 h->free_huge_pages - h->resv_huge_pages == 0) 474 goto err; 475 476 /* If reserves cannot be used, ensure enough pages are in the pool */ 477 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 478 goto err;; 479 480 for_each_zone_zonelist_nodemask(zone, z, zonelist, 481 MAX_NR_ZONES - 1, nodemask) { 482 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { 483 page = dequeue_huge_page_node(h, zone_to_nid(zone)); 484 if (page) { 485 if (!avoid_reserve) 486 decrement_hugepage_resv_vma(h, vma); 487 break; 488 } 489 } 490 } 491err: 492 mpol_cond_put(mpol); 493 put_mems_allowed(); 494 return page; 495} 496 497static void update_and_free_page(struct hstate *h, struct page *page) 498{ 499 int i; 500 501 VM_BUG_ON(h->order >= MAX_ORDER); 502 503 h->nr_huge_pages--; 504 h->nr_huge_pages_node[page_to_nid(page)]--; 505 for (i = 0; i < pages_per_huge_page(h); i++) { 506 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | 507 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | 508 1 << PG_private | 1<< PG_writeback); 509 } 510 set_compound_page_dtor(page, NULL); 511 set_page_refcounted(page); 512 arch_release_hugepage(page); 513 __free_pages(page, huge_page_order(h)); 514} 515 516struct hstate *size_to_hstate(unsigned long size) 517{ 518 struct hstate *h; 519 520 for_each_hstate(h) { 521 if (huge_page_size(h) == size) 522 return h; 523 } 524 return NULL; 525} 526 527static void free_huge_page(struct page *page) 528{ 529 /* 530 * Can't pass hstate in here because it is called from the 531 * compound page destructor. 532 */ 533 struct hstate *h = page_hstate(page); 534 int nid = page_to_nid(page); 535 struct address_space *mapping; 536 537 mapping = (struct address_space *) page_private(page); 538 set_page_private(page, 0); 539 page->mapping = NULL; 540 BUG_ON(page_count(page)); 541 BUG_ON(page_mapcount(page)); 542 INIT_LIST_HEAD(&page->lru); 543 544 spin_lock(&hugetlb_lock); 545 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 546 update_and_free_page(h, page); 547 h->surplus_huge_pages--; 548 h->surplus_huge_pages_node[nid]--; 549 } else { 550 enqueue_huge_page(h, page); 551 } 552 spin_unlock(&hugetlb_lock); 553 if (mapping) 554 hugetlb_put_quota(mapping, 1); 555} 556 557static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 558{ 559 set_compound_page_dtor(page, free_huge_page); 560 spin_lock(&hugetlb_lock); 561 h->nr_huge_pages++; 562 h->nr_huge_pages_node[nid]++; 563 spin_unlock(&hugetlb_lock); 564 put_page(page); /* free it into the hugepage allocator */ 565} 566 567static void prep_compound_gigantic_page(struct page *page, unsigned long order) 568{ 569 int i; 570 int nr_pages = 1 << order; 571 struct page *p = page + 1; 572 573 /* we rely on prep_new_huge_page to set the destructor */ 574 set_compound_order(page, order); 575 __SetPageHead(page); 576 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 577 __SetPageTail(p); 578 p->first_page = page; 579 } 580} 581 582int PageHuge(struct page *page) 583{ 584 compound_page_dtor *dtor; 585 586 if (!PageCompound(page)) 587 return 0; 588 589 page = compound_head(page); 590 dtor = get_compound_page_dtor(page); 591 592 return dtor == free_huge_page; 593} 594 595EXPORT_SYMBOL_GPL(PageHuge); 596 597static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 598{ 599 struct page *page; 600 601 if (h->order >= MAX_ORDER) 602 return NULL; 603 604 page = alloc_pages_exact_node(nid, 605 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 606 __GFP_REPEAT|__GFP_NOWARN, 607 huge_page_order(h)); 608 if (page) { 609 if (arch_prepare_hugepage(page)) { 610 __free_pages(page, huge_page_order(h)); 611 return NULL; 612 } 613 prep_new_huge_page(h, page, nid); 614 } 615 616 return page; 617} 618 619/* 620 * common helper functions for hstate_next_node_to_{alloc|free}. 621 * We may have allocated or freed a huge page based on a different 622 * nodes_allowed previously, so h->next_node_to_{alloc|free} might 623 * be outside of *nodes_allowed. Ensure that we use an allowed 624 * node for alloc or free. 625 */ 626static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 627{ 628 nid = next_node(nid, *nodes_allowed); 629 if (nid == MAX_NUMNODES) 630 nid = first_node(*nodes_allowed); 631 VM_BUG_ON(nid >= MAX_NUMNODES); 632 633 return nid; 634} 635 636static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 637{ 638 if (!node_isset(nid, *nodes_allowed)) 639 nid = next_node_allowed(nid, nodes_allowed); 640 return nid; 641} 642 643/* 644 * returns the previously saved node ["this node"] from which to 645 * allocate a persistent huge page for the pool and advance the 646 * next node from which to allocate, handling wrap at end of node 647 * mask. 648 */ 649static int hstate_next_node_to_alloc(struct hstate *h, 650 nodemask_t *nodes_allowed) 651{ 652 int nid; 653 654 VM_BUG_ON(!nodes_allowed); 655 656 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 657 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 658 659 return nid; 660} 661 662static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 663{ 664 struct page *page; 665 int start_nid; 666 int next_nid; 667 int ret = 0; 668 669 start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 670 next_nid = start_nid; 671 672 do { 673 page = alloc_fresh_huge_page_node(h, next_nid); 674 if (page) { 675 ret = 1; 676 break; 677 } 678 next_nid = hstate_next_node_to_alloc(h, nodes_allowed); 679 } while (next_nid != start_nid); 680 681 if (ret) 682 count_vm_event(HTLB_BUDDY_PGALLOC); 683 else 684 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 685 686 return ret; 687} 688 689/* 690 * helper for free_pool_huge_page() - return the previously saved 691 * node ["this node"] from which to free a huge page. Advance the 692 * next node id whether or not we find a free huge page to free so 693 * that the next attempt to free addresses the next node. 694 */ 695static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 696{ 697 int nid; 698 699 VM_BUG_ON(!nodes_allowed); 700 701 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 702 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 703 704 return nid; 705} 706 707/* 708 * Free huge page from pool from next node to free. 709 * Attempt to keep persistent huge pages more or less 710 * balanced over allowed nodes. 711 * Called with hugetlb_lock locked. 712 */ 713static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 714 bool acct_surplus) 715{ 716 int start_nid; 717 int next_nid; 718 int ret = 0; 719 720 start_nid = hstate_next_node_to_free(h, nodes_allowed); 721 next_nid = start_nid; 722 723 do { 724 /* 725 * If we're returning unused surplus pages, only examine 726 * nodes with surplus pages. 727 */ 728 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && 729 !list_empty(&h->hugepage_freelists[next_nid])) { 730 struct page *page = 731 list_entry(h->hugepage_freelists[next_nid].next, 732 struct page, lru); 733 list_del(&page->lru); 734 h->free_huge_pages--; 735 h->free_huge_pages_node[next_nid]--; 736 if (acct_surplus) { 737 h->surplus_huge_pages--; 738 h->surplus_huge_pages_node[next_nid]--; 739 } 740 update_and_free_page(h, page); 741 ret = 1; 742 break; 743 } 744 next_nid = hstate_next_node_to_free(h, nodes_allowed); 745 } while (next_nid != start_nid); 746 747 return ret; 748} 749 750static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 751{ 752 struct page *page; 753 unsigned int r_nid; 754 755 if (h->order >= MAX_ORDER) 756 return NULL; 757 758 /* 759 * Assume we will successfully allocate the surplus page to 760 * prevent racing processes from causing the surplus to exceed 761 * overcommit 762 * 763 * This however introduces a different race, where a process B 764 * tries to grow the static hugepage pool while alloc_pages() is 765 * called by process A. B will only examine the per-node 766 * counters in determining if surplus huge pages can be 767 * converted to normal huge pages in adjust_pool_surplus(). A 768 * won't be able to increment the per-node counter, until the 769 * lock is dropped by B, but B doesn't drop hugetlb_lock until 770 * no more huge pages can be converted from surplus to normal 771 * state (and doesn't try to convert again). Thus, we have a 772 * case where a surplus huge page exists, the pool is grown, and 773 * the surplus huge page still exists after, even though it 774 * should just have been converted to a normal huge page. This 775 * does not leak memory, though, as the hugepage will be freed 776 * once it is out of use. It also does not allow the counters to 777 * go out of whack in adjust_pool_surplus() as we don't modify 778 * the node values until we've gotten the hugepage and only the 779 * per-node value is checked there. 780 */ 781 spin_lock(&hugetlb_lock); 782 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 783 spin_unlock(&hugetlb_lock); 784 return NULL; 785 } else { 786 h->nr_huge_pages++; 787 h->surplus_huge_pages++; 788 } 789 spin_unlock(&hugetlb_lock); 790 791 if (nid == NUMA_NO_NODE) 792 page = alloc_pages(htlb_alloc_mask|__GFP_COMP| 793 __GFP_REPEAT|__GFP_NOWARN, 794 huge_page_order(h)); 795 else 796 page = alloc_pages_exact_node(nid, 797 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| 798 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 799 800 if (page && arch_prepare_hugepage(page)) { 801 __free_pages(page, huge_page_order(h)); 802 return NULL; 803 } 804 805 spin_lock(&hugetlb_lock); 806 if (page) { 807 r_nid = page_to_nid(page); 808 set_compound_page_dtor(page, free_huge_page); 809 /* 810 * We incremented the global counters already 811 */ 812 h->nr_huge_pages_node[r_nid]++; 813 h->surplus_huge_pages_node[r_nid]++; 814 __count_vm_event(HTLB_BUDDY_PGALLOC); 815 } else { 816 h->nr_huge_pages--; 817 h->surplus_huge_pages--; 818 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 819 } 820 spin_unlock(&hugetlb_lock); 821 822 return page; 823} 824 825/* 826 * This allocation function is useful in the context where vma is irrelevant. 827 * E.g. soft-offlining uses this function because it only cares physical 828 * address of error page. 829 */ 830struct page *alloc_huge_page_node(struct hstate *h, int nid) 831{ 832 struct page *page; 833 834 spin_lock(&hugetlb_lock); 835 page = dequeue_huge_page_node(h, nid); 836 spin_unlock(&hugetlb_lock); 837 838 if (!page) 839 page = alloc_buddy_huge_page(h, nid); 840 841 return page; 842} 843 844/* 845 * Increase the hugetlb pool such that it can accomodate a reservation 846 * of size 'delta'. 847 */ 848static int gather_surplus_pages(struct hstate *h, int delta) 849{ 850 struct list_head surplus_list; 851 struct page *page, *tmp; 852 int ret, i; 853 int needed, allocated; 854 855 needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 856 if (needed <= 0) { 857 h->resv_huge_pages += delta; 858 return 0; 859 } 860 861 allocated = 0; 862 INIT_LIST_HEAD(&surplus_list); 863 864 ret = -ENOMEM; 865retry: 866 spin_unlock(&hugetlb_lock); 867 for (i = 0; i < needed; i++) { 868 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 869 if (!page) 870 /* 871 * We were not able to allocate enough pages to 872 * satisfy the entire reservation so we free what 873 * we've allocated so far. 874 */ 875 goto free; 876 877 list_add(&page->lru, &surplus_list); 878 } 879 allocated += needed; 880 881 /* 882 * After retaking hugetlb_lock, we need to recalculate 'needed' 883 * because either resv_huge_pages or free_huge_pages may have changed. 884 */ 885 spin_lock(&hugetlb_lock); 886 needed = (h->resv_huge_pages + delta) - 887 (h->free_huge_pages + allocated); 888 if (needed > 0) 889 goto retry; 890 891 /* 892 * The surplus_list now contains _at_least_ the number of extra pages 893 * needed to accomodate the reservation. Add the appropriate number 894 * of pages to the hugetlb pool and free the extras back to the buddy 895 * allocator. Commit the entire reservation here to prevent another 896 * process from stealing the pages as they are added to the pool but 897 * before they are reserved. 898 */ 899 needed += allocated; 900 h->resv_huge_pages += delta; 901 ret = 0; 902 903 spin_unlock(&hugetlb_lock); 904 /* Free the needed pages to the hugetlb pool */ 905 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 906 if ((--needed) < 0) 907 break; 908 list_del(&page->lru); 909 /* 910 * This page is now managed by the hugetlb allocator and has 911 * no users -- drop the buddy allocator's reference. 912 */ 913 put_page_testzero(page); 914 VM_BUG_ON(page_count(page)); 915 enqueue_huge_page(h, page); 916 } 917 918 /* Free unnecessary surplus pages to the buddy allocator */ 919free: 920 if (!list_empty(&surplus_list)) { 921 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 922 list_del(&page->lru); 923 put_page(page); 924 } 925 } 926 spin_lock(&hugetlb_lock); 927 928 return ret; 929} 930 931/* 932 * When releasing a hugetlb pool reservation, any surplus pages that were 933 * allocated to satisfy the reservation must be explicitly freed if they were 934 * never used. 935 * Called with hugetlb_lock held. 936 */ 937static void return_unused_surplus_pages(struct hstate *h, 938 unsigned long unused_resv_pages) 939{ 940 unsigned long nr_pages; 941 942 /* Uncommit the reservation */ 943 h->resv_huge_pages -= unused_resv_pages; 944 945 /* Cannot return gigantic pages currently */ 946 if (h->order >= MAX_ORDER) 947 return; 948 949 nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 950 951 /* 952 * We want to release as many surplus pages as possible, spread 953 * evenly across all nodes with memory. Iterate across these nodes 954 * until we can no longer free unreserved surplus pages. This occurs 955 * when the nodes with surplus pages have no free pages. 956 * free_pool_huge_page() will balance the the freed pages across the 957 * on-line nodes with memory and will handle the hstate accounting. 958 */ 959 while (nr_pages--) { 960 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) 961 break; 962 } 963} 964 965/* 966 * Determine if the huge page at addr within the vma has an associated 967 * reservation. Where it does not we will need to logically increase 968 * reservation and actually increase quota before an allocation can occur. 969 * Where any new reservation would be required the reservation change is 970 * prepared, but not committed. Once the page has been quota'd allocated 971 * an instantiated the change should be committed via vma_commit_reservation. 972 * No action is required on failure. 973 */ 974static long vma_needs_reservation(struct hstate *h, 975 struct vm_area_struct *vma, unsigned long addr) 976{ 977 struct address_space *mapping = vma->vm_file->f_mapping; 978 struct inode *inode = mapping->host; 979 980 if (vma->vm_flags & VM_MAYSHARE) { 981 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 982 return region_chg(&inode->i_mapping->private_list, 983 idx, idx + 1); 984 985 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 986 return 1; 987 988 } else { 989 long err; 990 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 991 struct resv_map *reservations = vma_resv_map(vma); 992 993 err = region_chg(&reservations->regions, idx, idx + 1); 994 if (err < 0) 995 return err; 996 return 0; 997 } 998} 999static void vma_commit_reservation(struct hstate *h, 1000 struct vm_area_struct *vma, unsigned long addr) 1001{ 1002 struct address_space *mapping = vma->vm_file->f_mapping; 1003 struct inode *inode = mapping->host; 1004 1005 if (vma->vm_flags & VM_MAYSHARE) { 1006 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1007 region_add(&inode->i_mapping->private_list, idx, idx + 1); 1008 1009 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1010 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1011 struct resv_map *reservations = vma_resv_map(vma); 1012 1013 /* Mark this page used in the map. */ 1014 region_add(&reservations->regions, idx, idx + 1); 1015 } 1016} 1017 1018static struct page *alloc_huge_page(struct vm_area_struct *vma, 1019 unsigned long addr, int avoid_reserve) 1020{ 1021 struct hstate *h = hstate_vma(vma); 1022 struct page *page; 1023 struct address_space *mapping = vma->vm_file->f_mapping; 1024 struct inode *inode = mapping->host; 1025 long chg; 1026 1027 /* 1028 * Processes that did not create the mapping will have no reserves and 1029 * will not have accounted against quota. Check that the quota can be 1030 * made before satisfying the allocation 1031 * MAP_NORESERVE mappings may also need pages and quota allocated 1032 * if no reserve mapping overlaps. 1033 */ 1034 chg = vma_needs_reservation(h, vma, addr); 1035 if (chg < 0) 1036 return ERR_PTR(chg); 1037 if (chg) 1038 if (hugetlb_get_quota(inode->i_mapping, chg)) 1039 return ERR_PTR(-ENOSPC); 1040 1041 spin_lock(&hugetlb_lock); 1042 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); 1043 spin_unlock(&hugetlb_lock); 1044 1045 if (!page) { 1046 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1047 if (!page) { 1048 hugetlb_put_quota(inode->i_mapping, chg); 1049 return ERR_PTR(-VM_FAULT_SIGBUS); 1050 } 1051 } 1052 1053 set_page_private(page, (unsigned long) mapping); 1054 1055 vma_commit_reservation(h, vma, addr); 1056 1057 return page; 1058} 1059 1060int __weak alloc_bootmem_huge_page(struct hstate *h) 1061{ 1062 struct huge_bootmem_page *m; 1063 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 1064 1065 while (nr_nodes) { 1066 void *addr; 1067 1068 addr = __alloc_bootmem_node_nopanic( 1069 NODE_DATA(hstate_next_node_to_alloc(h, 1070 &node_states[N_HIGH_MEMORY])), 1071 huge_page_size(h), huge_page_size(h), 0); 1072 1073 if (addr) { 1074 /* 1075 * Use the beginning of the huge page to store the 1076 * huge_bootmem_page struct (until gather_bootmem 1077 * puts them into the mem_map). 1078 */ 1079 m = addr; 1080 goto found; 1081 } 1082 nr_nodes--; 1083 } 1084 return 0; 1085 1086found: 1087 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1088 /* Put them into a private list first because mem_map is not up yet */ 1089 list_add(&m->list, &huge_boot_pages); 1090 m->hstate = h; 1091 return 1; 1092} 1093 1094static void prep_compound_huge_page(struct page *page, int order) 1095{ 1096 if (unlikely(order > (MAX_ORDER - 1))) 1097 prep_compound_gigantic_page(page, order); 1098 else 1099 prep_compound_page(page, order); 1100} 1101 1102/* Put bootmem huge pages into the standard lists after mem_map is up */ 1103static void __init gather_bootmem_prealloc(void) 1104{ 1105 struct huge_bootmem_page *m; 1106 1107 list_for_each_entry(m, &huge_boot_pages, list) { 1108 struct page *page = virt_to_page(m); 1109 struct hstate *h = m->hstate; 1110 __ClearPageReserved(page); 1111 WARN_ON(page_count(page) != 1); 1112 prep_compound_huge_page(page, h->order); 1113 prep_new_huge_page(h, page, page_to_nid(page)); 1114 } 1115} 1116 1117static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 1118{ 1119 unsigned long i; 1120 1121 for (i = 0; i < h->max_huge_pages; ++i) { 1122 if (h->order >= MAX_ORDER) { 1123 if (!alloc_bootmem_huge_page(h)) 1124 break; 1125 } else if (!alloc_fresh_huge_page(h, 1126 &node_states[N_HIGH_MEMORY])) 1127 break; 1128 } 1129 h->max_huge_pages = i; 1130} 1131 1132static void __init hugetlb_init_hstates(void) 1133{ 1134 struct hstate *h; 1135 1136 for_each_hstate(h) { 1137 /* oversize hugepages were init'ed in early boot */ 1138 if (h->order < MAX_ORDER) 1139 hugetlb_hstate_alloc_pages(h); 1140 } 1141} 1142 1143static char * __init memfmt(char *buf, unsigned long n) 1144{ 1145 if (n >= (1UL << 30)) 1146 sprintf(buf, "%lu GB", n >> 30); 1147 else if (n >= (1UL << 20)) 1148 sprintf(buf, "%lu MB", n >> 20); 1149 else 1150 sprintf(buf, "%lu KB", n >> 10); 1151 return buf; 1152} 1153 1154static void __init report_hugepages(void) 1155{ 1156 struct hstate *h; 1157 1158 for_each_hstate(h) { 1159 char buf[32]; 1160 printk(KERN_INFO "HugeTLB registered %s page size, " 1161 "pre-allocated %ld pages\n", 1162 memfmt(buf, huge_page_size(h)), 1163 h->free_huge_pages); 1164 } 1165} 1166 1167#ifdef CONFIG_HIGHMEM 1168static void try_to_free_low(struct hstate *h, unsigned long count, 1169 nodemask_t *nodes_allowed) 1170{ 1171 int i; 1172 1173 if (h->order >= MAX_ORDER) 1174 return; 1175 1176 for_each_node_mask(i, *nodes_allowed) { 1177 struct page *page, *next; 1178 struct list_head *freel = &h->hugepage_freelists[i]; 1179 list_for_each_entry_safe(page, next, freel, lru) { 1180 if (count >= h->nr_huge_pages) 1181 return; 1182 if (PageHighMem(page)) 1183 continue; 1184 list_del(&page->lru); 1185 update_and_free_page(h, page); 1186 h->free_huge_pages--; 1187 h->free_huge_pages_node[page_to_nid(page)]--; 1188 } 1189 } 1190} 1191#else 1192static inline void try_to_free_low(struct hstate *h, unsigned long count, 1193 nodemask_t *nodes_allowed) 1194{ 1195} 1196#endif 1197 1198/* 1199 * Increment or decrement surplus_huge_pages. Keep node-specific counters 1200 * balanced by operating on them in a round-robin fashion. 1201 * Returns 1 if an adjustment was made. 1202 */ 1203static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 1204 int delta) 1205{ 1206 int start_nid, next_nid; 1207 int ret = 0; 1208 1209 VM_BUG_ON(delta != -1 && delta != 1); 1210 1211 if (delta < 0) 1212 start_nid = hstate_next_node_to_alloc(h, nodes_allowed); 1213 else 1214 start_nid = hstate_next_node_to_free(h, nodes_allowed); 1215 next_nid = start_nid; 1216 1217 do { 1218 int nid = next_nid; 1219 if (delta < 0) { 1220 /* 1221 * To shrink on this node, there must be a surplus page 1222 */ 1223 if (!h->surplus_huge_pages_node[nid]) { 1224 next_nid = hstate_next_node_to_alloc(h, 1225 nodes_allowed); 1226 continue; 1227 } 1228 } 1229 if (delta > 0) { 1230 /* 1231 * Surplus cannot exceed the total number of pages 1232 */ 1233 if (h->surplus_huge_pages_node[nid] >= 1234 h->nr_huge_pages_node[nid]) { 1235 next_nid = hstate_next_node_to_free(h, 1236 nodes_allowed); 1237 continue; 1238 } 1239 } 1240 1241 h->surplus_huge_pages += delta; 1242 h->surplus_huge_pages_node[nid] += delta; 1243 ret = 1; 1244 break; 1245 } while (next_nid != start_nid); 1246 1247 return ret; 1248} 1249 1250#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1251static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 1252 nodemask_t *nodes_allowed) 1253{ 1254 unsigned long min_count, ret; 1255 1256 if (h->order >= MAX_ORDER) 1257 return h->max_huge_pages; 1258 1259 /* 1260 * Increase the pool size 1261 * First take pages out of surplus state. Then make up the 1262 * remaining difference by allocating fresh huge pages. 1263 * 1264 * We might race with alloc_buddy_huge_page() here and be unable 1265 * to convert a surplus huge page to a normal huge page. That is 1266 * not critical, though, it just means the overall size of the 1267 * pool might be one hugepage larger than it needs to be, but 1268 * within all the constraints specified by the sysctls. 1269 */ 1270 spin_lock(&hugetlb_lock); 1271 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1272 if (!adjust_pool_surplus(h, nodes_allowed, -1)) 1273 break; 1274 } 1275 1276 while (count > persistent_huge_pages(h)) { 1277 /* 1278 * If this allocation races such that we no longer need the 1279 * page, free_huge_page will handle it by freeing the page 1280 * and reducing the surplus. 1281 */ 1282 spin_unlock(&hugetlb_lock); 1283 ret = alloc_fresh_huge_page(h, nodes_allowed); 1284 spin_lock(&hugetlb_lock); 1285 if (!ret) 1286 goto out; 1287 1288 /* Bail for signals. Probably ctrl-c from user */ 1289 if (signal_pending(current)) 1290 goto out; 1291 } 1292 1293 /* 1294 * Decrease the pool size 1295 * First return free pages to the buddy allocator (being careful 1296 * to keep enough around to satisfy reservations). Then place 1297 * pages into surplus state as needed so the pool will shrink 1298 * to the desired size as pages become free. 1299 * 1300 * By placing pages into the surplus state independent of the 1301 * overcommit value, we are allowing the surplus pool size to 1302 * exceed overcommit. There are few sane options here. Since 1303 * alloc_buddy_huge_page() is checking the global counter, 1304 * though, we'll note that we're not allowed to exceed surplus 1305 * and won't grow the pool anywhere else. Not until one of the 1306 * sysctls are changed, or the surplus pages go out of use. 1307 */ 1308 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 1309 min_count = max(count, min_count); 1310 try_to_free_low(h, min_count, nodes_allowed); 1311 while (min_count < persistent_huge_pages(h)) { 1312 if (!free_pool_huge_page(h, nodes_allowed, 0)) 1313 break; 1314 } 1315 while (count < persistent_huge_pages(h)) { 1316 if (!adjust_pool_surplus(h, nodes_allowed, 1)) 1317 break; 1318 } 1319out: 1320 ret = persistent_huge_pages(h); 1321 spin_unlock(&hugetlb_lock); 1322 return ret; 1323} 1324 1325#define HSTATE_ATTR_RO(_name) \ 1326 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1327 1328#define HSTATE_ATTR(_name) \ 1329 static struct kobj_attribute _name##_attr = \ 1330 __ATTR(_name, 0644, _name##_show, _name##_store) 1331 1332static struct kobject *hugepages_kobj; 1333static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1334 1335static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 1336 1337static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1338{ 1339 int i; 1340 1341 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1342 if (hstate_kobjs[i] == kobj) { 1343 if (nidp) 1344 *nidp = NUMA_NO_NODE; 1345 return &hstates[i]; 1346 } 1347 1348 return kobj_to_node_hstate(kobj, nidp); 1349} 1350 1351static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1352 struct kobj_attribute *attr, char *buf) 1353{ 1354 struct hstate *h; 1355 unsigned long nr_huge_pages; 1356 int nid; 1357 1358 h = kobj_to_hstate(kobj, &nid); 1359 if (nid == NUMA_NO_NODE) 1360 nr_huge_pages = h->nr_huge_pages; 1361 else 1362 nr_huge_pages = h->nr_huge_pages_node[nid]; 1363 1364 return sprintf(buf, "%lu\n", nr_huge_pages); 1365} 1366 1367static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 1368 struct kobject *kobj, struct kobj_attribute *attr, 1369 const char *buf, size_t len) 1370{ 1371 int err; 1372 int nid; 1373 unsigned long count; 1374 struct hstate *h; 1375 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1376 1377 err = strict_strtoul(buf, 10, &count); 1378 if (err) { 1379 err = 0; /* This seems wrong */ 1380 goto out; 1381 } 1382 1383 h = kobj_to_hstate(kobj, &nid); 1384 if (h->order >= MAX_ORDER) { 1385 err = -EINVAL; 1386 goto out; 1387 } 1388 1389 if (nid == NUMA_NO_NODE) { 1390 /* 1391 * global hstate attribute 1392 */ 1393 if (!(obey_mempolicy && 1394 init_nodemask_of_mempolicy(nodes_allowed))) { 1395 NODEMASK_FREE(nodes_allowed); 1396 nodes_allowed = &node_states[N_HIGH_MEMORY]; 1397 } 1398 } else if (nodes_allowed) { 1399 /* 1400 * per node hstate attribute: adjust count to global, 1401 * but restrict alloc/free to the specified node. 1402 */ 1403 count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 1404 init_nodemask_of_node(nodes_allowed, nid); 1405 } else 1406 nodes_allowed = &node_states[N_HIGH_MEMORY]; 1407 1408 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1409 1410 if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 1411 NODEMASK_FREE(nodes_allowed); 1412 1413 return len; 1414out: 1415 NODEMASK_FREE(nodes_allowed); 1416 return err; 1417} 1418 1419static ssize_t nr_hugepages_show(struct kobject *kobj, 1420 struct kobj_attribute *attr, char *buf) 1421{ 1422 return nr_hugepages_show_common(kobj, attr, buf); 1423} 1424 1425static ssize_t nr_hugepages_store(struct kobject *kobj, 1426 struct kobj_attribute *attr, const char *buf, size_t len) 1427{ 1428 return nr_hugepages_store_common(false, kobj, attr, buf, len); 1429} 1430HSTATE_ATTR(nr_hugepages); 1431 1432#ifdef CONFIG_NUMA 1433 1434/* 1435 * hstate attribute for optionally mempolicy-based constraint on persistent 1436 * huge page alloc/free. 1437 */ 1438static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 1439 struct kobj_attribute *attr, char *buf) 1440{ 1441 return nr_hugepages_show_common(kobj, attr, buf); 1442} 1443 1444static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 1445 struct kobj_attribute *attr, const char *buf, size_t len) 1446{ 1447 return nr_hugepages_store_common(true, kobj, attr, buf, len); 1448} 1449HSTATE_ATTR(nr_hugepages_mempolicy); 1450#endif 1451 1452 1453static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1454 struct kobj_attribute *attr, char *buf) 1455{ 1456 struct hstate *h = kobj_to_hstate(kobj, NULL); 1457 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1458} 1459 1460static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1461 struct kobj_attribute *attr, const char *buf, size_t count) 1462{ 1463 int err; 1464 unsigned long input; 1465 struct hstate *h = kobj_to_hstate(kobj, NULL); 1466 1467 if (h->order >= MAX_ORDER) 1468 return -EINVAL; 1469 1470 err = strict_strtoul(buf, 10, &input); 1471 if (err) 1472 return 0; 1473 1474 spin_lock(&hugetlb_lock); 1475 h->nr_overcommit_huge_pages = input; 1476 spin_unlock(&hugetlb_lock); 1477 1478 return count; 1479} 1480HSTATE_ATTR(nr_overcommit_hugepages); 1481 1482static ssize_t free_hugepages_show(struct kobject *kobj, 1483 struct kobj_attribute *attr, char *buf) 1484{ 1485 struct hstate *h; 1486 unsigned long free_huge_pages; 1487 int nid; 1488 1489 h = kobj_to_hstate(kobj, &nid); 1490 if (nid == NUMA_NO_NODE) 1491 free_huge_pages = h->free_huge_pages; 1492 else 1493 free_huge_pages = h->free_huge_pages_node[nid]; 1494 1495 return sprintf(buf, "%lu\n", free_huge_pages); 1496} 1497HSTATE_ATTR_RO(free_hugepages); 1498 1499static ssize_t resv_hugepages_show(struct kobject *kobj, 1500 struct kobj_attribute *attr, char *buf) 1501{ 1502 struct hstate *h = kobj_to_hstate(kobj, NULL); 1503 return sprintf(buf, "%lu\n", h->resv_huge_pages); 1504} 1505HSTATE_ATTR_RO(resv_hugepages); 1506 1507static ssize_t surplus_hugepages_show(struct kobject *kobj, 1508 struct kobj_attribute *attr, char *buf) 1509{ 1510 struct hstate *h; 1511 unsigned long surplus_huge_pages; 1512 int nid; 1513 1514 h = kobj_to_hstate(kobj, &nid); 1515 if (nid == NUMA_NO_NODE) 1516 surplus_huge_pages = h->surplus_huge_pages; 1517 else 1518 surplus_huge_pages = h->surplus_huge_pages_node[nid]; 1519 1520 return sprintf(buf, "%lu\n", surplus_huge_pages); 1521} 1522HSTATE_ATTR_RO(surplus_hugepages); 1523 1524static struct attribute *hstate_attrs[] = { 1525 &nr_hugepages_attr.attr, 1526 &nr_overcommit_hugepages_attr.attr, 1527 &free_hugepages_attr.attr, 1528 &resv_hugepages_attr.attr, 1529 &surplus_hugepages_attr.attr, 1530#ifdef CONFIG_NUMA 1531 &nr_hugepages_mempolicy_attr.attr, 1532#endif 1533 NULL, 1534}; 1535 1536static struct attribute_group hstate_attr_group = { 1537 .attrs = hstate_attrs, 1538}; 1539 1540static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 1541 struct kobject **hstate_kobjs, 1542 struct attribute_group *hstate_attr_group) 1543{ 1544 int retval; 1545 int hi = h - hstates; 1546 1547 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 1548 if (!hstate_kobjs[hi]) 1549 return -ENOMEM; 1550 1551 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1552 if (retval) 1553 kobject_put(hstate_kobjs[hi]); 1554 1555 return retval; 1556} 1557 1558static void __init hugetlb_sysfs_init(void) 1559{ 1560 struct hstate *h; 1561 int err; 1562 1563 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1564 if (!hugepages_kobj) 1565 return; 1566 1567 for_each_hstate(h) { 1568 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 1569 hstate_kobjs, &hstate_attr_group); 1570 if (err) 1571 printk(KERN_ERR "Hugetlb: Unable to add hstate %s", 1572 h->name); 1573 } 1574} 1575 1576#ifdef CONFIG_NUMA 1577 1578/* 1579 * node_hstate/s - associate per node hstate attributes, via their kobjects, 1580 * with node sysdevs in node_devices[] using a parallel array. The array 1581 * index of a node sysdev or _hstate == node id. 1582 * This is here to avoid any static dependency of the node sysdev driver, in 1583 * the base kernel, on the hugetlb module. 1584 */ 1585struct node_hstate { 1586 struct kobject *hugepages_kobj; 1587 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1588}; 1589struct node_hstate node_hstates[MAX_NUMNODES]; 1590 1591/* 1592 * A subset of global hstate attributes for node sysdevs 1593 */ 1594static struct attribute *per_node_hstate_attrs[] = { 1595 &nr_hugepages_attr.attr, 1596 &free_hugepages_attr.attr, 1597 &surplus_hugepages_attr.attr, 1598 NULL, 1599}; 1600 1601static struct attribute_group per_node_hstate_attr_group = { 1602 .attrs = per_node_hstate_attrs, 1603}; 1604 1605/* 1606 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj. 1607 * Returns node id via non-NULL nidp. 1608 */ 1609static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 1610{ 1611 int nid; 1612 1613 for (nid = 0; nid < nr_node_ids; nid++) { 1614 struct node_hstate *nhs = &node_hstates[nid]; 1615 int i; 1616 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1617 if (nhs->hstate_kobjs[i] == kobj) { 1618 if (nidp) 1619 *nidp = nid; 1620 return &hstates[i]; 1621 } 1622 } 1623 1624 BUG(); 1625 return NULL; 1626} 1627 1628/* 1629 * Unregister hstate attributes from a single node sysdev. 1630 * No-op if no hstate attributes attached. 1631 */ 1632void hugetlb_unregister_node(struct node *node) 1633{ 1634 struct hstate *h; 1635 struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 1636 1637 if (!nhs->hugepages_kobj) 1638 return; /* no hstate attributes */ 1639 1640 for_each_hstate(h) 1641 if (nhs->hstate_kobjs[h - hstates]) { 1642 kobject_put(nhs->hstate_kobjs[h - hstates]); 1643 nhs->hstate_kobjs[h - hstates] = NULL; 1644 } 1645 1646 kobject_put(nhs->hugepages_kobj); 1647 nhs->hugepages_kobj = NULL; 1648} 1649 1650/* 1651 * hugetlb module exit: unregister hstate attributes from node sysdevs 1652 * that have them. 1653 */ 1654static void hugetlb_unregister_all_nodes(void) 1655{ 1656 int nid; 1657 1658 /* 1659 * disable node sysdev registrations. 1660 */ 1661 register_hugetlbfs_with_node(NULL, NULL); 1662 1663 /* 1664 * remove hstate attributes from any nodes that have them. 1665 */ 1666 for (nid = 0; nid < nr_node_ids; nid++) 1667 hugetlb_unregister_node(&node_devices[nid]); 1668} 1669 1670/* 1671 * Register hstate attributes for a single node sysdev. 1672 * No-op if attributes already registered. 1673 */ 1674void hugetlb_register_node(struct node *node) 1675{ 1676 struct hstate *h; 1677 struct node_hstate *nhs = &node_hstates[node->sysdev.id]; 1678 int err; 1679 1680 if (nhs->hugepages_kobj) 1681 return; /* already allocated */ 1682 1683 nhs->hugepages_kobj = kobject_create_and_add("hugepages", 1684 &node->sysdev.kobj); 1685 if (!nhs->hugepages_kobj) 1686 return; 1687 1688 for_each_hstate(h) { 1689 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 1690 nhs->hstate_kobjs, 1691 &per_node_hstate_attr_group); 1692 if (err) { 1693 printk(KERN_ERR "Hugetlb: Unable to add hstate %s" 1694 " for node %d\n", 1695 h->name, node->sysdev.id); 1696 hugetlb_unregister_node(node); 1697 break; 1698 } 1699 } 1700} 1701 1702/* 1703 * hugetlb init time: register hstate attributes for all registered node 1704 * sysdevs of nodes that have memory. All on-line nodes should have 1705 * registered their associated sysdev by this time. 1706 */ 1707static void hugetlb_register_all_nodes(void) 1708{ 1709 int nid; 1710 1711 for_each_node_state(nid, N_HIGH_MEMORY) { 1712 struct node *node = &node_devices[nid]; 1713 if (node->sysdev.id == nid) 1714 hugetlb_register_node(node); 1715 } 1716 1717 /* 1718 * Let the node sysdev driver know we're here so it can 1719 * [un]register hstate attributes on node hotplug. 1720 */ 1721 register_hugetlbfs_with_node(hugetlb_register_node, 1722 hugetlb_unregister_node); 1723} 1724#else /* !CONFIG_NUMA */ 1725 1726static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 1727{ 1728 BUG(); 1729 if (nidp) 1730 *nidp = -1; 1731 return NULL; 1732} 1733 1734static void hugetlb_unregister_all_nodes(void) { } 1735 1736static void hugetlb_register_all_nodes(void) { } 1737 1738#endif 1739 1740static void __exit hugetlb_exit(void) 1741{ 1742 struct hstate *h; 1743 1744 hugetlb_unregister_all_nodes(); 1745 1746 for_each_hstate(h) { 1747 kobject_put(hstate_kobjs[h - hstates]); 1748 } 1749 1750 kobject_put(hugepages_kobj); 1751} 1752module_exit(hugetlb_exit); 1753 1754static int __init hugetlb_init(void) 1755{ 1756 /* Some platform decide whether they support huge pages at boot 1757 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 1758 * there is no such support 1759 */ 1760 if (HPAGE_SHIFT == 0) 1761 return 0; 1762 1763 if (!size_to_hstate(default_hstate_size)) { 1764 default_hstate_size = HPAGE_SIZE; 1765 if (!size_to_hstate(default_hstate_size)) 1766 hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1767 } 1768 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; 1769 if (default_hstate_max_huge_pages) 1770 default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1771 1772 hugetlb_init_hstates(); 1773 1774 gather_bootmem_prealloc(); 1775 1776 report_hugepages(); 1777 1778 hugetlb_sysfs_init(); 1779 1780 hugetlb_register_all_nodes(); 1781 1782 return 0; 1783} 1784module_init(hugetlb_init); 1785 1786/* Should be called on processing a hugepagesz=... option */ 1787void __init hugetlb_add_hstate(unsigned order) 1788{ 1789 struct hstate *h; 1790 unsigned long i; 1791 1792 if (size_to_hstate(PAGE_SIZE << order)) { 1793 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); 1794 return; 1795 } 1796 BUG_ON(max_hstate >= HUGE_MAX_HSTATE); 1797 BUG_ON(order == 0); 1798 h = &hstates[max_hstate++]; 1799 h->order = order; 1800 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 1801 h->nr_huge_pages = 0; 1802 h->free_huge_pages = 0; 1803 for (i = 0; i < MAX_NUMNODES; ++i) 1804 INIT_LIST_HEAD(&h->hugepage_freelists[i]); 1805 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); 1806 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); 1807 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1808 huge_page_size(h)/1024); 1809 1810 parsed_hstate = h; 1811} 1812 1813static int __init hugetlb_nrpages_setup(char *s) 1814{ 1815 unsigned long *mhp; 1816 static unsigned long *last_mhp; 1817 1818 /* 1819 * !max_hstate means we haven't parsed a hugepagesz= parameter yet, 1820 * so this hugepages= parameter goes to the "default hstate". 1821 */ 1822 if (!max_hstate) 1823 mhp = &default_hstate_max_huge_pages; 1824 else 1825 mhp = &parsed_hstate->max_huge_pages; 1826 1827 if (mhp == last_mhp) { 1828 printk(KERN_WARNING "hugepages= specified twice without " 1829 "interleaving hugepagesz=, ignoring\n"); 1830 return 1; 1831 } 1832 1833 if (sscanf(s, "%lu", mhp) <= 0) 1834 *mhp = 0; 1835 1836 /* 1837 * Global state is always initialized later in hugetlb_init. 1838 * But we need to allocate >= MAX_ORDER hstates here early to still 1839 * use the bootmem allocator. 1840 */ 1841 if (max_hstate && parsed_hstate->order >= MAX_ORDER) 1842 hugetlb_hstate_alloc_pages(parsed_hstate); 1843 1844 last_mhp = mhp; 1845 1846 return 1; 1847} 1848__setup("hugepages=", hugetlb_nrpages_setup); 1849 1850static int __init hugetlb_default_setup(char *s) 1851{ 1852 default_hstate_size = memparse(s, &s); 1853 return 1; 1854} 1855__setup("default_hugepagesz=", hugetlb_default_setup); 1856 1857static unsigned int cpuset_mems_nr(unsigned int *array) 1858{ 1859 int node; 1860 unsigned int nr = 0; 1861 1862 for_each_node_mask(node, cpuset_current_mems_allowed) 1863 nr += array[node]; 1864 1865 return nr; 1866} 1867 1868#ifdef CONFIG_SYSCTL 1869static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 1870 struct ctl_table *table, int write, 1871 void __user *buffer, size_t *length, loff_t *ppos) 1872{ 1873 struct hstate *h = &default_hstate; 1874 unsigned long tmp; 1875 int ret; 1876 1877 if (!write) 1878 tmp = h->max_huge_pages; 1879 1880 if (write && h->order >= MAX_ORDER) 1881 return -EINVAL; 1882 1883 table->data = &tmp; 1884 table->maxlen = sizeof(unsigned long); 1885 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 1886 if (ret) 1887 goto out; 1888 1889 if (write) { 1890 NODEMASK_ALLOC(nodemask_t, nodes_allowed, 1891 GFP_KERNEL | __GFP_NORETRY); 1892 if (!(obey_mempolicy && 1893 init_nodemask_of_mempolicy(nodes_allowed))) { 1894 NODEMASK_FREE(nodes_allowed); 1895 nodes_allowed = &node_states[N_HIGH_MEMORY]; 1896 } 1897 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 1898 1899 if (nodes_allowed != &node_states[N_HIGH_MEMORY]) 1900 NODEMASK_FREE(nodes_allowed); 1901 } 1902out: 1903 return ret; 1904} 1905 1906int hugetlb_sysctl_handler(struct ctl_table *table, int write, 1907 void __user *buffer, size_t *length, loff_t *ppos) 1908{ 1909 1910 return hugetlb_sysctl_handler_common(false, table, write, 1911 buffer, length, ppos); 1912} 1913 1914#ifdef CONFIG_NUMA 1915int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 1916 void __user *buffer, size_t *length, loff_t *ppos) 1917{ 1918 return hugetlb_sysctl_handler_common(true, table, write, 1919 buffer, length, ppos); 1920} 1921#endif /* CONFIG_NUMA */ 1922 1923int hugetlb_treat_movable_handler(struct ctl_table *table, int write, 1924 void __user *buffer, 1925 size_t *length, loff_t *ppos) 1926{ 1927 proc_dointvec(table, write, buffer, length, ppos); 1928 if (hugepages_treat_as_movable) 1929 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; 1930 else 1931 htlb_alloc_mask = GFP_HIGHUSER; 1932 return 0; 1933} 1934 1935int hugetlb_overcommit_handler(struct ctl_table *table, int write, 1936 void __user *buffer, 1937 size_t *length, loff_t *ppos) 1938{ 1939 struct hstate *h = &default_hstate; 1940 unsigned long tmp; 1941 int ret; 1942 1943 if (!write) 1944 tmp = h->nr_overcommit_huge_pages; 1945 1946 if (write && h->order >= MAX_ORDER) 1947 return -EINVAL; 1948 1949 table->data = &tmp; 1950 table->maxlen = sizeof(unsigned long); 1951 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 1952 if (ret) 1953 goto out; 1954 1955 if (write) { 1956 spin_lock(&hugetlb_lock); 1957 h->nr_overcommit_huge_pages = tmp; 1958 spin_unlock(&hugetlb_lock); 1959 } 1960out: 1961 return ret; 1962} 1963 1964#endif /* CONFIG_SYSCTL */ 1965 1966void hugetlb_report_meminfo(struct seq_file *m) 1967{ 1968 struct hstate *h = &default_hstate; 1969 seq_printf(m, 1970 "HugePages_Total: %5lu\n" 1971 "HugePages_Free: %5lu\n" 1972 "HugePages_Rsvd: %5lu\n" 1973 "HugePages_Surp: %5lu\n" 1974 "Hugepagesize: %8lu kB\n", 1975 h->nr_huge_pages, 1976 h->free_huge_pages, 1977 h->resv_huge_pages, 1978 h->surplus_huge_pages, 1979 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 1980} 1981 1982int hugetlb_report_node_meminfo(int nid, char *buf) 1983{ 1984 struct hstate *h = &default_hstate; 1985 return sprintf(buf, 1986 "Node %d HugePages_Total: %5u\n" 1987 "Node %d HugePages_Free: %5u\n" 1988 "Node %d HugePages_Surp: %5u\n", 1989 nid, h->nr_huge_pages_node[nid], 1990 nid, h->free_huge_pages_node[nid], 1991 nid, h->surplus_huge_pages_node[nid]); 1992} 1993 1994/* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 1995unsigned long hugetlb_total_pages(void) 1996{ 1997 struct hstate *h = &default_hstate; 1998 return h->nr_huge_pages * pages_per_huge_page(h); 1999} 2000 2001static int hugetlb_acct_memory(struct hstate *h, long delta) 2002{ 2003 int ret = -ENOMEM; 2004 2005 spin_lock(&hugetlb_lock); 2006 /* 2007 * When cpuset is configured, it breaks the strict hugetlb page 2008 * reservation as the accounting is done on a global variable. Such 2009 * reservation is completely rubbish in the presence of cpuset because 2010 * the reservation is not checked against page availability for the 2011 * current cpuset. Application can still potentially OOM'ed by kernel 2012 * with lack of free htlb page in cpuset that the task is in. 2013 * Attempt to enforce strict accounting with cpuset is almost 2014 * impossible (or too ugly) because cpuset is too fluid that 2015 * task or memory node can be dynamically moved between cpusets. 2016 * 2017 * The change of semantics for shared hugetlb mapping with cpuset is 2018 * undesirable. However, in order to preserve some of the semantics, 2019 * we fall back to check against current free page availability as 2020 * a best attempt and hopefully to minimize the impact of changing 2021 * semantics that cpuset has. 2022 */ 2023 if (delta > 0) { 2024 if (gather_surplus_pages(h, delta) < 0) 2025 goto out; 2026 2027 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2028 return_unused_surplus_pages(h, delta); 2029 goto out; 2030 } 2031 } 2032 2033 ret = 0; 2034 if (delta < 0) 2035 return_unused_surplus_pages(h, (unsigned long) -delta); 2036 2037out: 2038 spin_unlock(&hugetlb_lock); 2039 return ret; 2040} 2041 2042static void hugetlb_vm_op_open(struct vm_area_struct *vma) 2043{ 2044 struct resv_map *reservations = vma_resv_map(vma); 2045 2046 /* 2047 * This new VMA should share its siblings reservation map if present. 2048 * The VMA will only ever have a valid reservation map pointer where 2049 * it is being copied for another still existing VMA. As that VMA 2050 * has a reference to the reservation map it cannot dissappear until 2051 * after this open call completes. It is therefore safe to take a 2052 * new reference here without additional locking. 2053 */ 2054 if (reservations) 2055 kref_get(&reservations->refs); 2056} 2057 2058static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2059{ 2060 struct hstate *h = hstate_vma(vma); 2061 struct resv_map *reservations = vma_resv_map(vma); 2062 unsigned long reserve; 2063 unsigned long start; 2064 unsigned long end; 2065 2066 if (reservations) { 2067 start = vma_hugecache_offset(h, vma, vma->vm_start); 2068 end = vma_hugecache_offset(h, vma, vma->vm_end); 2069 2070 reserve = (end - start) - 2071 region_count(&reservations->regions, start, end); 2072 2073 kref_put(&reservations->refs, resv_map_release); 2074 2075 if (reserve) { 2076 hugetlb_acct_memory(h, -reserve); 2077 hugetlb_put_quota(vma->vm_file->f_mapping, reserve); 2078 } 2079 } 2080} 2081 2082/* 2083 * We cannot handle pagefaults against hugetlb pages at all. They cause 2084 * handle_mm_fault() to try to instantiate regular-sized pages in the 2085 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 2086 * this far. 2087 */ 2088static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 2089{ 2090 BUG(); 2091 return 0; 2092} 2093 2094const struct vm_operations_struct hugetlb_vm_ops = { 2095 .fault = hugetlb_vm_op_fault, 2096 .open = hugetlb_vm_op_open, 2097 .close = hugetlb_vm_op_close, 2098}; 2099 2100static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 2101 int writable) 2102{ 2103 pte_t entry; 2104 2105 if (writable) { 2106 entry = 2107 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); 2108 } else { 2109 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); 2110 } 2111 entry = pte_mkyoung(entry); 2112 entry = pte_mkhuge(entry); 2113 2114 return entry; 2115} 2116 2117static void set_huge_ptep_writable(struct vm_area_struct *vma, 2118 unsigned long address, pte_t *ptep) 2119{ 2120 pte_t entry; 2121 2122 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); 2123 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { 2124 update_mmu_cache(vma, address, ptep); 2125 } 2126} 2127 2128 2129int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 2130 struct vm_area_struct *vma) 2131{ 2132 pte_t *src_pte, *dst_pte, entry; 2133 struct page *ptepage; 2134 unsigned long addr; 2135 int cow; 2136 struct hstate *h = hstate_vma(vma); 2137 unsigned long sz = huge_page_size(h); 2138 2139 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 2140 2141 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2142 src_pte = huge_pte_offset(src, addr); 2143 if (!src_pte) 2144 continue; 2145 dst_pte = huge_pte_alloc(dst, addr, sz); 2146 if (!dst_pte) 2147 goto nomem; 2148 2149 /* If the pagetables are shared don't copy or take references */ 2150 if (dst_pte == src_pte) 2151 continue; 2152 2153 spin_lock(&dst->page_table_lock); 2154 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); 2155 if (!huge_pte_none(huge_ptep_get(src_pte))) { 2156 if (cow) 2157 huge_ptep_set_wrprotect(src, addr, src_pte); 2158 entry = huge_ptep_get(src_pte); 2159 ptepage = pte_page(entry); 2160 get_page(ptepage); 2161 page_dup_rmap(ptepage); 2162 set_huge_pte_at(dst, addr, dst_pte, entry); 2163 } 2164 spin_unlock(&src->page_table_lock); 2165 spin_unlock(&dst->page_table_lock); 2166 } 2167 return 0; 2168 2169nomem: 2170 return -ENOMEM; 2171} 2172 2173static int is_hugetlb_entry_migration(pte_t pte) 2174{ 2175 swp_entry_t swp; 2176 2177 if (huge_pte_none(pte) || pte_present(pte)) 2178 return 0; 2179 swp = pte_to_swp_entry(pte); 2180 if (non_swap_entry(swp) && is_migration_entry(swp)) { 2181 return 1; 2182 } else 2183 return 0; 2184} 2185 2186static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2187{ 2188 swp_entry_t swp; 2189 2190 if (huge_pte_none(pte) || pte_present(pte)) 2191 return 0; 2192 swp = pte_to_swp_entry(pte); 2193 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) { 2194 return 1; 2195 } else 2196 return 0; 2197} 2198 2199void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 2200 unsigned long end, struct page *ref_page) 2201{ 2202 struct mm_struct *mm = vma->vm_mm; 2203 unsigned long address; 2204 pte_t *ptep; 2205 pte_t pte; 2206 struct page *page; 2207 struct page *tmp; 2208 struct hstate *h = hstate_vma(vma); 2209 unsigned long sz = huge_page_size(h); 2210 2211 /* 2212 * A page gathering list, protected by per file i_mmap_lock. The 2213 * lock is used to avoid list corruption from multiple unmapping 2214 * of the same page since we are using page->lru. 2215 */ 2216 LIST_HEAD(page_list); 2217 2218 WARN_ON(!is_vm_hugetlb_page(vma)); 2219 BUG_ON(start & ~huge_page_mask(h)); 2220 BUG_ON(end & ~huge_page_mask(h)); 2221 2222 mmu_notifier_invalidate_range_start(mm, start, end); 2223 spin_lock(&mm->page_table_lock); 2224 for (address = start; address < end; address += sz) { 2225 ptep = huge_pte_offset(mm, address); 2226 if (!ptep) 2227 continue; 2228 2229 if (huge_pmd_unshare(mm, &address, ptep)) 2230 continue; 2231 2232 /* 2233 * If a reference page is supplied, it is because a specific 2234 * page is being unmapped, not a range. Ensure the page we 2235 * are about to unmap is the actual page of interest. 2236 */ 2237 if (ref_page) { 2238 pte = huge_ptep_get(ptep); 2239 if (huge_pte_none(pte)) 2240 continue; 2241 page = pte_page(pte); 2242 if (page != ref_page) 2243 continue; 2244 2245 /* 2246 * Mark the VMA as having unmapped its page so that 2247 * future faults in this VMA will fail rather than 2248 * looking like data was lost 2249 */ 2250 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 2251 } 2252 2253 pte = huge_ptep_get_and_clear(mm, address, ptep); 2254 if (huge_pte_none(pte)) 2255 continue; 2256 2257 /* 2258 * HWPoisoned hugepage is already unmapped and dropped reference 2259 */ 2260 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) 2261 continue; 2262 2263 page = pte_page(pte); 2264 if (pte_dirty(pte)) 2265 set_page_dirty(page); 2266 list_add(&page->lru, &page_list); 2267 } 2268 spin_unlock(&mm->page_table_lock); 2269 flush_tlb_range(vma, start, end); 2270 mmu_notifier_invalidate_range_end(mm, start, end); 2271 list_for_each_entry_safe(page, tmp, &page_list, lru) { 2272 page_remove_rmap(page); 2273 list_del(&page->lru); 2274 put_page(page); 2275 } 2276} 2277 2278void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 2279 unsigned long end, struct page *ref_page) 2280{ 2281 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 2282 __unmap_hugepage_range(vma, start, end, ref_page); 2283 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2284} 2285 2286/* 2287 * This is called when the original mapper is failing to COW a MAP_PRIVATE 2288 * mappping it owns the reserve page for. The intention is to unmap the page 2289 * from other VMAs and let the children be SIGKILLed if they are faulting the 2290 * same region. 2291 */ 2292static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 2293 struct page *page, unsigned long address) 2294{ 2295 struct hstate *h = hstate_vma(vma); 2296 struct vm_area_struct *iter_vma; 2297 struct address_space *mapping; 2298 struct prio_tree_iter iter; 2299 pgoff_t pgoff; 2300 2301 /* 2302 * vm_pgoff is in PAGE_SIZE units, hence the different calculation 2303 * from page cache lookup which is in HPAGE_SIZE units. 2304 */ 2305 address = address & huge_page_mask(h); 2306 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) 2307 + (vma->vm_pgoff >> PAGE_SHIFT); 2308 mapping = (struct address_space *)page_private(page); 2309 2310 /* 2311 * Take the mapping lock for the duration of the table walk. As 2312 * this mapping should be shared between all the VMAs, 2313 * __unmap_hugepage_range() is called as the lock is already held 2314 */ 2315 spin_lock(&mapping->i_mmap_lock); 2316 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 2317 /* Do not unmap the current VMA */ 2318 if (iter_vma == vma) 2319 continue; 2320 2321 /* 2322 * Unmap the page from other VMAs without their own reserves. 2323 * They get marked to be SIGKILLed if they fault in these 2324 * areas. This is because a future no-page fault on this VMA 2325 * could insert a zeroed page instead of the data existing 2326 * from the time of fork. This would look like data corruption 2327 */ 2328 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 2329 __unmap_hugepage_range(iter_vma, 2330 address, address + huge_page_size(h), 2331 page); 2332 } 2333 spin_unlock(&mapping->i_mmap_lock); 2334 2335 return 1; 2336} 2337 2338/* 2339 * Hugetlb_cow() should be called with page lock of the original hugepage held. 2340 */ 2341static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 2342 unsigned long address, pte_t *ptep, pte_t pte, 2343 struct page *pagecache_page) 2344{ 2345 struct hstate *h = hstate_vma(vma); 2346 struct page *old_page, *new_page; 2347 int avoidcopy; 2348 int outside_reserve = 0; 2349 2350 old_page = pte_page(pte); 2351 2352retry_avoidcopy: 2353 /* If no-one else is actually using this page, avoid the copy 2354 * and just make the page writable */ 2355 avoidcopy = (page_mapcount(old_page) == 1); 2356 if (avoidcopy) { 2357 if (PageAnon(old_page)) 2358 page_move_anon_rmap(old_page, vma, address); 2359 set_huge_ptep_writable(vma, address, ptep); 2360 return 0; 2361 } 2362 2363 /* 2364 * If the process that created a MAP_PRIVATE mapping is about to 2365 * perform a COW due to a shared page count, attempt to satisfy 2366 * the allocation without using the existing reserves. The pagecache 2367 * page is used to determine if the reserve at this address was 2368 * consumed or not. If reserves were used, a partial faulted mapping 2369 * at the time of fork() could consume its reserves on COW instead 2370 * of the full address range. 2371 */ 2372 if (!(vma->vm_flags & VM_MAYSHARE) && 2373 is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 2374 old_page != pagecache_page) 2375 outside_reserve = 1; 2376 2377 page_cache_get(old_page); 2378 2379 /* Drop page_table_lock as buddy allocator may be called */ 2380 spin_unlock(&mm->page_table_lock); 2381 new_page = alloc_huge_page(vma, address, outside_reserve); 2382 2383 if (IS_ERR(new_page)) { 2384 page_cache_release(old_page); 2385 2386 /* 2387 * If a process owning a MAP_PRIVATE mapping fails to COW, 2388 * it is due to references held by a child and an insufficient 2389 * huge page pool. To guarantee the original mappers 2390 * reliability, unmap the page from child processes. The child 2391 * may get SIGKILLed if it later faults. 2392 */ 2393 if (outside_reserve) { 2394 BUG_ON(huge_pte_none(pte)); 2395 if (unmap_ref_private(mm, vma, old_page, address)) { 2396 BUG_ON(page_count(old_page) != 1); 2397 BUG_ON(huge_pte_none(pte)); 2398 spin_lock(&mm->page_table_lock); 2399 goto retry_avoidcopy; 2400 } 2401 WARN_ON_ONCE(1); 2402 } 2403 2404 /* Caller expects lock to be held */ 2405 spin_lock(&mm->page_table_lock); 2406 return -PTR_ERR(new_page); 2407 } 2408 2409 /* 2410 * When the original hugepage is shared one, it does not have 2411 * anon_vma prepared. 2412 */ 2413 if (unlikely(anon_vma_prepare(vma))) { 2414 /* Caller expects lock to be held */ 2415 spin_lock(&mm->page_table_lock); 2416 return VM_FAULT_OOM; 2417 } 2418 2419 copy_user_huge_page(new_page, old_page, address, vma, 2420 pages_per_huge_page(h)); 2421 __SetPageUptodate(new_page); 2422 2423 /* 2424 * Retake the page_table_lock to check for racing updates 2425 * before the page tables are altered 2426 */ 2427 spin_lock(&mm->page_table_lock); 2428 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2429 if (likely(pte_same(huge_ptep_get(ptep), pte))) { 2430 /* Break COW */ 2431 mmu_notifier_invalidate_range_start(mm, 2432 address & huge_page_mask(h), 2433 (address & huge_page_mask(h)) + huge_page_size(h)); 2434 huge_ptep_clear_flush(vma, address, ptep); 2435 set_huge_pte_at(mm, address, ptep, 2436 make_huge_pte(vma, new_page, 1)); 2437 page_remove_rmap(old_page); 2438 hugepage_add_new_anon_rmap(new_page, vma, address); 2439 /* Make the old page be freed below */ 2440 new_page = old_page; 2441 mmu_notifier_invalidate_range_end(mm, 2442 address & huge_page_mask(h), 2443 (address & huge_page_mask(h)) + huge_page_size(h)); 2444 } 2445 page_cache_release(new_page); 2446 page_cache_release(old_page); 2447 return 0; 2448} 2449 2450/* Return the pagecache page at a given address within a VMA */ 2451static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2452 struct vm_area_struct *vma, unsigned long address) 2453{ 2454 struct address_space *mapping; 2455 pgoff_t idx; 2456 2457 mapping = vma->vm_file->f_mapping; 2458 idx = vma_hugecache_offset(h, vma, address); 2459 2460 return find_lock_page(mapping, idx); 2461} 2462 2463/* 2464 * Return whether there is a pagecache page to back given address within VMA. 2465 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 2466 */ 2467static bool hugetlbfs_pagecache_present(struct hstate *h, 2468 struct vm_area_struct *vma, unsigned long address) 2469{ 2470 struct address_space *mapping; 2471 pgoff_t idx; 2472 struct page *page; 2473 2474 mapping = vma->vm_file->f_mapping; 2475 idx = vma_hugecache_offset(h, vma, address); 2476 2477 page = find_get_page(mapping, idx); 2478 if (page) 2479 put_page(page); 2480 return page != NULL; 2481} 2482 2483static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2484 unsigned long address, pte_t *ptep, unsigned int flags) 2485{ 2486 struct hstate *h = hstate_vma(vma); 2487 int ret = VM_FAULT_SIGBUS; 2488 pgoff_t idx; 2489 unsigned long size; 2490 struct page *page; 2491 struct address_space *mapping; 2492 pte_t new_pte; 2493 2494 /* 2495 * Currently, we are forced to kill the process in the event the 2496 * original mapper has unmapped pages from the child due to a failed 2497 * COW. Warn that such a situation has occured as it may not be obvious 2498 */ 2499 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2500 printk(KERN_WARNING 2501 "PID %d killed due to inadequate hugepage pool\n", 2502 current->pid); 2503 return ret; 2504 } 2505 2506 mapping = vma->vm_file->f_mapping; 2507 idx = vma_hugecache_offset(h, vma, address); 2508 2509 /* 2510 * Use page lock to guard against racing truncation 2511 * before we get page_table_lock. 2512 */ 2513retry: 2514 page = find_lock_page(mapping, idx); 2515 if (!page) { 2516 size = i_size_read(mapping->host) >> huge_page_shift(h); 2517 if (idx >= size) 2518 goto out; 2519 page = alloc_huge_page(vma, address, 0); 2520 if (IS_ERR(page)) { 2521 ret = -PTR_ERR(page); 2522 goto out; 2523 } 2524 clear_huge_page(page, address, pages_per_huge_page(h)); 2525 __SetPageUptodate(page); 2526 2527 if (vma->vm_flags & VM_MAYSHARE) { 2528 int err; 2529 struct inode *inode = mapping->host; 2530 2531 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 2532 if (err) { 2533 put_page(page); 2534 if (err == -EEXIST) 2535 goto retry; 2536 goto out; 2537 } 2538 2539 spin_lock(&inode->i_lock); 2540 inode->i_blocks += blocks_per_huge_page(h); 2541 spin_unlock(&inode->i_lock); 2542 page_dup_rmap(page); 2543 } else { 2544 lock_page(page); 2545 if (unlikely(anon_vma_prepare(vma))) { 2546 ret = VM_FAULT_OOM; 2547 goto backout_unlocked; 2548 } 2549 hugepage_add_new_anon_rmap(page, vma, address); 2550 } 2551 } else { 2552 /* 2553 * If memory error occurs between mmap() and fault, some process 2554 * don't have hwpoisoned swap entry for errored virtual address. 2555 * So we need to block hugepage fault by PG_hwpoison bit check. 2556 */ 2557 if (unlikely(PageHWPoison(page))) { 2558 ret = VM_FAULT_HWPOISON | 2559 VM_FAULT_SET_HINDEX(h - hstates); 2560 goto backout_unlocked; 2561 } 2562 page_dup_rmap(page); 2563 } 2564 2565 /* 2566 * If we are going to COW a private mapping later, we examine the 2567 * pending reservations for this page now. This will ensure that 2568 * any allocations necessary to record that reservation occur outside 2569 * the spinlock. 2570 */ 2571 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 2572 if (vma_needs_reservation(h, vma, address) < 0) { 2573 ret = VM_FAULT_OOM; 2574 goto backout_unlocked; 2575 } 2576 2577 spin_lock(&mm->page_table_lock); 2578 size = i_size_read(mapping->host) >> huge_page_shift(h); 2579 if (idx >= size) 2580 goto backout; 2581 2582 ret = 0; 2583 if (!huge_pte_none(huge_ptep_get(ptep))) 2584 goto backout; 2585 2586 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 2587 && (vma->vm_flags & VM_SHARED))); 2588 set_huge_pte_at(mm, address, ptep, new_pte); 2589 2590 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 2591 /* Optimization, do the COW without a second fault */ 2592 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); 2593 } 2594 2595 spin_unlock(&mm->page_table_lock); 2596 unlock_page(page); 2597out: 2598 return ret; 2599 2600backout: 2601 spin_unlock(&mm->page_table_lock); 2602backout_unlocked: 2603 unlock_page(page); 2604 put_page(page); 2605 goto out; 2606} 2607 2608int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2609 unsigned long address, unsigned int flags) 2610{ 2611 pte_t *ptep; 2612 pte_t entry; 2613 int ret; 2614 struct page *page = NULL; 2615 struct page *pagecache_page = NULL; 2616 static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2617 struct hstate *h = hstate_vma(vma); 2618 2619 ptep = huge_pte_offset(mm, address); 2620 if (ptep) { 2621 entry = huge_ptep_get(ptep); 2622 if (unlikely(is_hugetlb_entry_migration(entry))) { 2623 migration_entry_wait(mm, (pmd_t *)ptep, address); 2624 return 0; 2625 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2626 return VM_FAULT_HWPOISON_LARGE | 2627 VM_FAULT_SET_HINDEX(h - hstates); 2628 } 2629 2630 ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 2631 if (!ptep) 2632 return VM_FAULT_OOM; 2633 2634 /* 2635 * Serialize hugepage allocation and instantiation, so that we don't 2636 * get spurious allocation failures if two CPUs race to instantiate 2637 * the same page in the page cache. 2638 */ 2639 mutex_lock(&hugetlb_instantiation_mutex); 2640 entry = huge_ptep_get(ptep); 2641 if (huge_pte_none(entry)) { 2642 ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2643 goto out_mutex; 2644 } 2645 2646 ret = 0; 2647 2648 /* 2649 * If we are going to COW the mapping later, we examine the pending 2650 * reservations for this page now. This will ensure that any 2651 * allocations necessary to record that reservation occur outside the 2652 * spinlock. For private mappings, we also lookup the pagecache 2653 * page now as it is used to determine if a reservation has been 2654 * consumed. 2655 */ 2656 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { 2657 if (vma_needs_reservation(h, vma, address) < 0) { 2658 ret = VM_FAULT_OOM; 2659 goto out_mutex; 2660 } 2661 2662 if (!(vma->vm_flags & VM_MAYSHARE)) 2663 pagecache_page = hugetlbfs_pagecache_page(h, 2664 vma, address); 2665 } 2666 2667 /* 2668 * hugetlb_cow() requires page locks of pte_page(entry) and 2669 * pagecache_page, so here we need take the former one 2670 * when page != pagecache_page or !pagecache_page. 2671 * Note that locking order is always pagecache_page -> page, 2672 * so no worry about deadlock. 2673 */ 2674 page = pte_page(entry); 2675 if (page != pagecache_page) 2676 lock_page(page); 2677 2678 spin_lock(&mm->page_table_lock); 2679 /* Check for a racing update before calling hugetlb_cow */ 2680 if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2681 goto out_page_table_lock; 2682 2683 2684 if (flags & FAULT_FLAG_WRITE) { 2685 if (!pte_write(entry)) { 2686 ret = hugetlb_cow(mm, vma, address, ptep, entry, 2687 pagecache_page); 2688 goto out_page_table_lock; 2689 } 2690 entry = pte_mkdirty(entry); 2691 } 2692 entry = pte_mkyoung(entry); 2693 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2694 flags & FAULT_FLAG_WRITE)) 2695 update_mmu_cache(vma, address, ptep); 2696 2697out_page_table_lock: 2698 spin_unlock(&mm->page_table_lock); 2699 2700 if (pagecache_page) { 2701 unlock_page(pagecache_page); 2702 put_page(pagecache_page); 2703 } 2704 if (page != pagecache_page) 2705 unlock_page(page); 2706 2707out_mutex: 2708 mutex_unlock(&hugetlb_instantiation_mutex); 2709 2710 return ret; 2711} 2712 2713/* Can be overriden by architectures */ 2714__attribute__((weak)) struct page * 2715follow_huge_pud(struct mm_struct *mm, unsigned long address, 2716 pud_t *pud, int write) 2717{ 2718 BUG(); 2719 return NULL; 2720} 2721 2722int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 2723 struct page **pages, struct vm_area_struct **vmas, 2724 unsigned long *position, int *length, int i, 2725 unsigned int flags) 2726{ 2727 unsigned long pfn_offset; 2728 unsigned long vaddr = *position; 2729 int remainder = *length; 2730 struct hstate *h = hstate_vma(vma); 2731 2732 spin_lock(&mm->page_table_lock); 2733 while (vaddr < vma->vm_end && remainder) { 2734 pte_t *pte; 2735 int absent; 2736 struct page *page; 2737 2738 /* 2739 * Some archs (sparc64, sh*) have multiple pte_ts to 2740 * each hugepage. We have to make sure we get the 2741 * first, for the page indexing below to work. 2742 */ 2743 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 2744 absent = !pte || huge_pte_none(huge_ptep_get(pte)); 2745 2746 /* 2747 * When coredumping, it suits get_dump_page if we just return 2748 * an error where there's an empty slot with no huge pagecache 2749 * to back it. This way, we avoid allocating a hugepage, and 2750 * the sparse dumpfile avoids allocating disk blocks, but its 2751 * huge holes still show up with zeroes where they need to be. 2752 */ 2753 if (absent && (flags & FOLL_DUMP) && 2754 !hugetlbfs_pagecache_present(h, vma, vaddr)) { 2755 remainder = 0; 2756 break; 2757 } 2758 2759 if (absent || 2760 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { 2761 int ret; 2762 2763 spin_unlock(&mm->page_table_lock); 2764 ret = hugetlb_fault(mm, vma, vaddr, 2765 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 2766 spin_lock(&mm->page_table_lock); 2767 if (!(ret & VM_FAULT_ERROR)) 2768 continue; 2769 2770 remainder = 0; 2771 break; 2772 } 2773 2774 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 2775 page = pte_page(huge_ptep_get(pte)); 2776same_page: 2777 if (pages) { 2778 pages[i] = mem_map_offset(page, pfn_offset); 2779 get_page(pages[i]); 2780 } 2781 2782 if (vmas) 2783 vmas[i] = vma; 2784 2785 vaddr += PAGE_SIZE; 2786 ++pfn_offset; 2787 --remainder; 2788 ++i; 2789 if (vaddr < vma->vm_end && remainder && 2790 pfn_offset < pages_per_huge_page(h)) { 2791 /* 2792 * We use pfn_offset to avoid touching the pageframes 2793 * of this compound page. 2794 */ 2795 goto same_page; 2796 } 2797 } 2798 spin_unlock(&mm->page_table_lock); 2799 *length = remainder; 2800 *position = vaddr; 2801 2802 return i ? i : -EFAULT; 2803} 2804 2805void hugetlb_change_protection(struct vm_area_struct *vma, 2806 unsigned long address, unsigned long end, pgprot_t newprot) 2807{ 2808 struct mm_struct *mm = vma->vm_mm; 2809 unsigned long start = address; 2810 pte_t *ptep; 2811 pte_t pte; 2812 struct hstate *h = hstate_vma(vma); 2813 2814 BUG_ON(address >= end); 2815 flush_cache_range(vma, address, end); 2816 2817 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); 2818 spin_lock(&mm->page_table_lock); 2819 for (; address < end; address += huge_page_size(h)) { 2820 ptep = huge_pte_offset(mm, address); 2821 if (!ptep) 2822 continue; 2823 if (huge_pmd_unshare(mm, &address, ptep)) 2824 continue; 2825 if (!huge_pte_none(huge_ptep_get(ptep))) { 2826 pte = huge_ptep_get_and_clear(mm, address, ptep); 2827 pte = pte_mkhuge(pte_modify(pte, newprot)); 2828 set_huge_pte_at(mm, address, ptep, pte); 2829 } 2830 } 2831 spin_unlock(&mm->page_table_lock); 2832 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); 2833 2834 flush_tlb_range(vma, start, end); 2835} 2836 2837int hugetlb_reserve_pages(struct inode *inode, 2838 long from, long to, 2839 struct vm_area_struct *vma, 2840 int acctflag) 2841{ 2842 long ret, chg; 2843 struct hstate *h = hstate_inode(inode); 2844 2845 /* 2846 * Only apply hugepage reservation if asked. At fault time, an 2847 * attempt will be made for VM_NORESERVE to allocate a page 2848 * and filesystem quota without using reserves 2849 */ 2850 if (acctflag & VM_NORESERVE) 2851 return 0; 2852 2853 /* 2854 * Shared mappings base their reservation on the number of pages that 2855 * are already allocated on behalf of the file. Private mappings need 2856 * to reserve the full area even if read-only as mprotect() may be 2857 * called to make the mapping read-write. Assume !vma is a shm mapping 2858 */ 2859 if (!vma || vma->vm_flags & VM_MAYSHARE) 2860 chg = region_chg(&inode->i_mapping->private_list, from, to); 2861 else { 2862 struct resv_map *resv_map = resv_map_alloc(); 2863 if (!resv_map) 2864 return -ENOMEM; 2865 2866 chg = to - from; 2867 2868 set_vma_resv_map(vma, resv_map); 2869 set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 2870 } 2871 2872 if (chg < 0) 2873 return chg; 2874 2875 /* There must be enough filesystem quota for the mapping */ 2876 if (hugetlb_get_quota(inode->i_mapping, chg)) 2877 return -ENOSPC; 2878 2879 /* 2880 * Check enough hugepages are available for the reservation. 2881 * Hand back the quota if there are not 2882 */ 2883 ret = hugetlb_acct_memory(h, chg); 2884 if (ret < 0) { 2885 hugetlb_put_quota(inode->i_mapping, chg); 2886 return ret; 2887 } 2888 2889 /* 2890 * Account for the reservations made. Shared mappings record regions 2891 * that have reservations as they are shared by multiple VMAs. 2892 * When the last VMA disappears, the region map says how much 2893 * the reservation was and the page cache tells how much of 2894 * the reservation was consumed. Private mappings are per-VMA and 2895 * only the consumed reservations are tracked. When the VMA 2896 * disappears, the original reservation is the VMA size and the 2897 * consumed reservations are stored in the map. Hence, nothing 2898 * else has to be done for private mappings here 2899 */ 2900 if (!vma || vma->vm_flags & VM_MAYSHARE) 2901 region_add(&inode->i_mapping->private_list, from, to); 2902 return 0; 2903} 2904 2905void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 2906{ 2907 struct hstate *h = hstate_inode(inode); 2908 long chg = region_truncate(&inode->i_mapping->private_list, offset); 2909 2910 spin_lock(&inode->i_lock); 2911 inode->i_blocks -= (blocks_per_huge_page(h) * freed); 2912 spin_unlock(&inode->i_lock); 2913 2914 hugetlb_put_quota(inode->i_mapping, (chg - freed)); 2915 hugetlb_acct_memory(h, -(chg - freed)); 2916} 2917 2918#ifdef CONFIG_MEMORY_FAILURE 2919 2920/* Should be called in hugetlb_lock */ 2921static int is_hugepage_on_freelist(struct page *hpage) 2922{ 2923 struct page *page; 2924 struct page *tmp; 2925 struct hstate *h = page_hstate(hpage); 2926 int nid = page_to_nid(hpage); 2927 2928 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 2929 if (page == hpage) 2930 return 1; 2931 return 0; 2932} 2933 2934/* 2935 * This function is called from memory failure code. 2936 * Assume the caller holds page lock of the head page. 2937 */ 2938int dequeue_hwpoisoned_huge_page(struct page *hpage) 2939{ 2940 struct hstate *h = page_hstate(hpage); 2941 int nid = page_to_nid(hpage); 2942 int ret = -EBUSY; 2943 2944 spin_lock(&hugetlb_lock); 2945 if (is_hugepage_on_freelist(hpage)) { 2946 list_del(&hpage->lru); 2947 set_page_refcounted(hpage); 2948 h->free_huge_pages--; 2949 h->free_huge_pages_node[nid]--; 2950 ret = 0; 2951 } 2952 spin_unlock(&hugetlb_lock); 2953 return ret; 2954} 2955#endif 2956