1/* 2 * Generic hugetlb support. 3 * (C) Nadia Yvette Chambers, 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/compiler.h> 17#include <linux/cpuset.h> 18#include <linux/mutex.h> 19#include <linux/bootmem.h> 20#include <linux/sysfs.h> 21#include <linux/slab.h> 22#include <linux/rmap.h> 23#include <linux/swap.h> 24#include <linux/swapops.h> 25#include <linux/page-isolation.h> 26#include <linux/jhash.h> 27 28#include <asm/page.h> 29#include <asm/pgtable.h> 30#include <asm/tlb.h> 31 32#include <linux/io.h> 33#include <linux/hugetlb.h> 34#include <linux/hugetlb_cgroup.h> 35#include <linux/node.h> 36#include "internal.h" 37 38int hugepages_treat_as_movable; 39 40int hugetlb_max_hstate __read_mostly; 41unsigned int default_hstate_idx; 42struct hstate hstates[HUGE_MAX_HSTATE]; 43/* 44 * Minimum page order among possible hugepage sizes, set to a proper value 45 * at boot time. 46 */ 47static unsigned int minimum_order __read_mostly = UINT_MAX; 48 49__initdata LIST_HEAD(huge_boot_pages); 50 51/* for command line parsing */ 52static struct hstate * __initdata parsed_hstate; 53static unsigned long __initdata default_hstate_max_huge_pages; 54static unsigned long __initdata default_hstate_size; 55 56/* 57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 58 * free_huge_pages, and surplus_huge_pages. 59 */ 60DEFINE_SPINLOCK(hugetlb_lock); 61 62/* 63 * Serializes faults on the same logical page. This is used to 64 * prevent spurious OOMs when the hugepage pool is fully utilized. 65 */ 66static int num_fault_mutexes; 67static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; 68 69/* Forward declaration */ 70static int hugetlb_acct_memory(struct hstate *h, long delta); 71 72static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 73{ 74 bool free = (spool->count == 0) && (spool->used_hpages == 0); 75 76 spin_unlock(&spool->lock); 77 78 /* If no pages are used, and no other handles to the subpool 79 * remain, give up any reservations mased on minimum size and 80 * free the subpool */ 81 if (free) { 82 if (spool->min_hpages != -1) 83 hugetlb_acct_memory(spool->hstate, 84 -spool->min_hpages); 85 kfree(spool); 86 } 87} 88 89struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 90 long min_hpages) 91{ 92 struct hugepage_subpool *spool; 93 94 spool = kzalloc(sizeof(*spool), GFP_KERNEL); 95 if (!spool) 96 return NULL; 97 98 spin_lock_init(&spool->lock); 99 spool->count = 1; 100 spool->max_hpages = max_hpages; 101 spool->hstate = h; 102 spool->min_hpages = min_hpages; 103 104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 105 kfree(spool); 106 return NULL; 107 } 108 spool->rsv_hpages = min_hpages; 109 110 return spool; 111} 112 113void hugepage_put_subpool(struct hugepage_subpool *spool) 114{ 115 spin_lock(&spool->lock); 116 BUG_ON(!spool->count); 117 spool->count--; 118 unlock_or_release_subpool(spool); 119} 120 121/* 122 * Subpool accounting for allocating and reserving pages. 123 * Return -ENOMEM if there are not enough resources to satisfy the 124 * the request. Otherwise, return the number of pages by which the 125 * global pools must be adjusted (upward). The returned value may 126 * only be different than the passed value (delta) in the case where 127 * a subpool minimum size must be manitained. 128 */ 129static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 130 long delta) 131{ 132 long ret = delta; 133 134 if (!spool) 135 return ret; 136 137 spin_lock(&spool->lock); 138 139 if (spool->max_hpages != -1) { /* maximum size accounting */ 140 if ((spool->used_hpages + delta) <= spool->max_hpages) 141 spool->used_hpages += delta; 142 else { 143 ret = -ENOMEM; 144 goto unlock_ret; 145 } 146 } 147 148 if (spool->min_hpages != -1) { /* minimum size accounting */ 149 if (delta > spool->rsv_hpages) { 150 /* 151 * Asking for more reserves than those already taken on 152 * behalf of subpool. Return difference. 153 */ 154 ret = delta - spool->rsv_hpages; 155 spool->rsv_hpages = 0; 156 } else { 157 ret = 0; /* reserves already accounted for */ 158 spool->rsv_hpages -= delta; 159 } 160 } 161 162unlock_ret: 163 spin_unlock(&spool->lock); 164 return ret; 165} 166 167/* 168 * Subpool accounting for freeing and unreserving pages. 169 * Return the number of global page reservations that must be dropped. 170 * The return value may only be different than the passed value (delta) 171 * in the case where a subpool minimum size must be maintained. 172 */ 173static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 174 long delta) 175{ 176 long ret = delta; 177 178 if (!spool) 179 return delta; 180 181 spin_lock(&spool->lock); 182 183 if (spool->max_hpages != -1) /* maximum size accounting */ 184 spool->used_hpages -= delta; 185 186 if (spool->min_hpages != -1) { /* minimum size accounting */ 187 if (spool->rsv_hpages + delta <= spool->min_hpages) 188 ret = 0; 189 else 190 ret = spool->rsv_hpages + delta - spool->min_hpages; 191 192 spool->rsv_hpages += delta; 193 if (spool->rsv_hpages > spool->min_hpages) 194 spool->rsv_hpages = spool->min_hpages; 195 } 196 197 /* 198 * If hugetlbfs_put_super couldn't free spool due to an outstanding 199 * quota reference, free it now. 200 */ 201 unlock_or_release_subpool(spool); 202 203 return ret; 204} 205 206static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 207{ 208 return HUGETLBFS_SB(inode->i_sb)->spool; 209} 210 211static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 212{ 213 return subpool_inode(file_inode(vma->vm_file)); 214} 215 216/* 217 * Region tracking -- allows tracking of reservations and instantiated pages 218 * across the pages in a mapping. 219 * 220 * The region data structures are embedded into a resv_map and 221 * protected by a resv_map's lock 222 */ 223struct file_region { 224 struct list_head link; 225 long from; 226 long to; 227}; 228 229static long region_add(struct resv_map *resv, long f, long t) 230{ 231 struct list_head *head = &resv->regions; 232 struct file_region *rg, *nrg, *trg; 233 234 spin_lock(&resv->lock); 235 /* Locate the region we are either in or before. */ 236 list_for_each_entry(rg, head, link) 237 if (f <= rg->to) 238 break; 239 240 /* Round our left edge to the current segment if it encloses us. */ 241 if (f > rg->from) 242 f = rg->from; 243 244 /* Check for and consume any regions we now overlap with. */ 245 nrg = rg; 246 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 247 if (&rg->link == head) 248 break; 249 if (rg->from > t) 250 break; 251 252 /* If this area reaches higher then extend our area to 253 * include it completely. If this is not the first area 254 * which we intend to reuse, free it. */ 255 if (rg->to > t) 256 t = rg->to; 257 if (rg != nrg) { 258 list_del(&rg->link); 259 kfree(rg); 260 } 261 } 262 nrg->from = f; 263 nrg->to = t; 264 spin_unlock(&resv->lock); 265 return 0; 266} 267 268static long region_chg(struct resv_map *resv, long f, long t) 269{ 270 struct list_head *head = &resv->regions; 271 struct file_region *rg, *nrg = NULL; 272 long chg = 0; 273 274retry: 275 spin_lock(&resv->lock); 276 /* Locate the region we are before or in. */ 277 list_for_each_entry(rg, head, link) 278 if (f <= rg->to) 279 break; 280 281 /* If we are below the current region then a new region is required. 282 * Subtle, allocate a new region at the position but make it zero 283 * size such that we can guarantee to record the reservation. */ 284 if (&rg->link == head || t < rg->from) { 285 if (!nrg) { 286 spin_unlock(&resv->lock); 287 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 288 if (!nrg) 289 return -ENOMEM; 290 291 nrg->from = f; 292 nrg->to = f; 293 INIT_LIST_HEAD(&nrg->link); 294 goto retry; 295 } 296 297 list_add(&nrg->link, rg->link.prev); 298 chg = t - f; 299 goto out_nrg; 300 } 301 302 /* Round our left edge to the current segment if it encloses us. */ 303 if (f > rg->from) 304 f = rg->from; 305 chg = t - f; 306 307 /* Check for and consume any regions we now overlap with. */ 308 list_for_each_entry(rg, rg->link.prev, link) { 309 if (&rg->link == head) 310 break; 311 if (rg->from > t) 312 goto out; 313 314 /* We overlap with this area, if it extends further than 315 * us then we must extend ourselves. Account for its 316 * existing reservation. */ 317 if (rg->to > t) { 318 chg += rg->to - t; 319 t = rg->to; 320 } 321 chg -= rg->to - rg->from; 322 } 323 324out: 325 spin_unlock(&resv->lock); 326 /* We already know we raced and no longer need the new region */ 327 kfree(nrg); 328 return chg; 329out_nrg: 330 spin_unlock(&resv->lock); 331 return chg; 332} 333 334static long region_truncate(struct resv_map *resv, long end) 335{ 336 struct list_head *head = &resv->regions; 337 struct file_region *rg, *trg; 338 long chg = 0; 339 340 spin_lock(&resv->lock); 341 /* Locate the region we are either in or before. */ 342 list_for_each_entry(rg, head, link) 343 if (end <= rg->to) 344 break; 345 if (&rg->link == head) 346 goto out; 347 348 /* If we are in the middle of a region then adjust it. */ 349 if (end > rg->from) { 350 chg = rg->to - end; 351 rg->to = end; 352 rg = list_entry(rg->link.next, typeof(*rg), link); 353 } 354 355 /* Drop any remaining regions. */ 356 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 357 if (&rg->link == head) 358 break; 359 chg += rg->to - rg->from; 360 list_del(&rg->link); 361 kfree(rg); 362 } 363 364out: 365 spin_unlock(&resv->lock); 366 return chg; 367} 368 369static long region_count(struct resv_map *resv, long f, long t) 370{ 371 struct list_head *head = &resv->regions; 372 struct file_region *rg; 373 long chg = 0; 374 375 spin_lock(&resv->lock); 376 /* Locate each segment we overlap with, and count that overlap. */ 377 list_for_each_entry(rg, head, link) { 378 long seg_from; 379 long seg_to; 380 381 if (rg->to <= f) 382 continue; 383 if (rg->from >= t) 384 break; 385 386 seg_from = max(rg->from, f); 387 seg_to = min(rg->to, t); 388 389 chg += seg_to - seg_from; 390 } 391 spin_unlock(&resv->lock); 392 393 return chg; 394} 395 396/* 397 * Convert the address within this vma to the page offset within 398 * the mapping, in pagecache page units; huge pages here. 399 */ 400static pgoff_t vma_hugecache_offset(struct hstate *h, 401 struct vm_area_struct *vma, unsigned long address) 402{ 403 return ((address - vma->vm_start) >> huge_page_shift(h)) + 404 (vma->vm_pgoff >> huge_page_order(h)); 405} 406 407pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 408 unsigned long address) 409{ 410 return vma_hugecache_offset(hstate_vma(vma), vma, address); 411} 412 413/* 414 * Return the size of the pages allocated when backing a VMA. In the majority 415 * cases this will be same size as used by the page table entries. 416 */ 417unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 418{ 419 struct hstate *hstate; 420 421 if (!is_vm_hugetlb_page(vma)) 422 return PAGE_SIZE; 423 424 hstate = hstate_vma(vma); 425 426 return 1UL << huge_page_shift(hstate); 427} 428EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 429 430/* 431 * Return the page size being used by the MMU to back a VMA. In the majority 432 * of cases, the page size used by the kernel matches the MMU size. On 433 * architectures where it differs, an architecture-specific version of this 434 * function is required. 435 */ 436#ifndef vma_mmu_pagesize 437unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 438{ 439 return vma_kernel_pagesize(vma); 440} 441#endif 442 443/* 444 * Flags for MAP_PRIVATE reservations. These are stored in the bottom 445 * bits of the reservation map pointer, which are always clear due to 446 * alignment. 447 */ 448#define HPAGE_RESV_OWNER (1UL << 0) 449#define HPAGE_RESV_UNMAPPED (1UL << 1) 450#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 451 452/* 453 * These helpers are used to track how many pages are reserved for 454 * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 455 * is guaranteed to have their future faults succeed. 456 * 457 * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 458 * the reserve counters are updated with the hugetlb_lock held. It is safe 459 * to reset the VMA at fork() time as it is not in use yet and there is no 460 * chance of the global counters getting corrupted as a result of the values. 461 * 462 * The private mapping reservation is represented in a subtly different 463 * manner to a shared mapping. A shared mapping has a region map associated 464 * with the underlying file, this region map represents the backing file 465 * pages which have ever had a reservation assigned which this persists even 466 * after the page is instantiated. A private mapping has a region map 467 * associated with the original mmap which is attached to all VMAs which 468 * reference it, this region map represents those offsets which have consumed 469 * reservation ie. where pages have been instantiated. 470 */ 471static unsigned long get_vma_private_data(struct vm_area_struct *vma) 472{ 473 return (unsigned long)vma->vm_private_data; 474} 475 476static void set_vma_private_data(struct vm_area_struct *vma, 477 unsigned long value) 478{ 479 vma->vm_private_data = (void *)value; 480} 481 482struct resv_map *resv_map_alloc(void) 483{ 484 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 485 if (!resv_map) 486 return NULL; 487 488 kref_init(&resv_map->refs); 489 spin_lock_init(&resv_map->lock); 490 INIT_LIST_HEAD(&resv_map->regions); 491 492 return resv_map; 493} 494 495void resv_map_release(struct kref *ref) 496{ 497 struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 498 499 /* Clear out any active regions before we release the map. */ 500 region_truncate(resv_map, 0); 501 kfree(resv_map); 502} 503 504static inline struct resv_map *inode_resv_map(struct inode *inode) 505{ 506 return inode->i_mapping->private_data; 507} 508 509static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 510{ 511 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 512 if (vma->vm_flags & VM_MAYSHARE) { 513 struct address_space *mapping = vma->vm_file->f_mapping; 514 struct inode *inode = mapping->host; 515 516 return inode_resv_map(inode); 517 518 } else { 519 return (struct resv_map *)(get_vma_private_data(vma) & 520 ~HPAGE_RESV_MASK); 521 } 522} 523 524static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 525{ 526 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 527 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 528 529 set_vma_private_data(vma, (get_vma_private_data(vma) & 530 HPAGE_RESV_MASK) | (unsigned long)map); 531} 532 533static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 534{ 535 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 536 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 537 538 set_vma_private_data(vma, get_vma_private_data(vma) | flags); 539} 540 541static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 542{ 543 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 544 545 return (get_vma_private_data(vma) & flag) != 0; 546} 547 548/* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 549void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 550{ 551 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 552 if (!(vma->vm_flags & VM_MAYSHARE)) 553 vma->vm_private_data = (void *)0; 554} 555 556/* Returns true if the VMA has associated reserve pages */ 557static int vma_has_reserves(struct vm_area_struct *vma, long chg) 558{ 559 if (vma->vm_flags & VM_NORESERVE) { 560 /* 561 * This address is already reserved by other process(chg == 0), 562 * so, we should decrement reserved count. Without decrementing, 563 * reserve count remains after releasing inode, because this 564 * allocated page will go into page cache and is regarded as 565 * coming from reserved pool in releasing step. Currently, we 566 * don't have any other solution to deal with this situation 567 * properly, so add work-around here. 568 */ 569 if (vma->vm_flags & VM_MAYSHARE && chg == 0) 570 return 1; 571 else 572 return 0; 573 } 574 575 /* Shared mappings always use reserves */ 576 if (vma->vm_flags & VM_MAYSHARE) 577 return 1; 578 579 /* 580 * Only the process that called mmap() has reserves for 581 * private mappings. 582 */ 583 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 584 return 1; 585 586 return 0; 587} 588 589static void enqueue_huge_page(struct hstate *h, struct page *page) 590{ 591 int nid = page_to_nid(page); 592 list_move(&page->lru, &h->hugepage_freelists[nid]); 593 h->free_huge_pages++; 594 h->free_huge_pages_node[nid]++; 595} 596 597static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 598{ 599 struct page *page; 600 601 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 602 if (!is_migrate_isolate_page(page)) 603 break; 604 /* 605 * if 'non-isolated free hugepage' not found on the list, 606 * the allocation fails. 607 */ 608 if (&h->hugepage_freelists[nid] == &page->lru) 609 return NULL; 610 list_move(&page->lru, &h->hugepage_activelist); 611 set_page_refcounted(page); 612 h->free_huge_pages--; 613 h->free_huge_pages_node[nid]--; 614 return page; 615} 616 617/* Movability of hugepages depends on migration support. */ 618static inline gfp_t htlb_alloc_mask(struct hstate *h) 619{ 620 if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 621 return GFP_HIGHUSER_MOVABLE; 622 else 623 return GFP_HIGHUSER; 624} 625 626static struct page *dequeue_huge_page_vma(struct hstate *h, 627 struct vm_area_struct *vma, 628 unsigned long address, int avoid_reserve, 629 long chg) 630{ 631 struct page *page = NULL; 632 struct mempolicy *mpol; 633 nodemask_t *nodemask; 634 struct zonelist *zonelist; 635 struct zone *zone; 636 struct zoneref *z; 637 unsigned int cpuset_mems_cookie; 638 639 /* 640 * A child process with MAP_PRIVATE mappings created by their parent 641 * have no page reserves. This check ensures that reservations are 642 * not "stolen". The child may still get SIGKILLed 643 */ 644 if (!vma_has_reserves(vma, chg) && 645 h->free_huge_pages - h->resv_huge_pages == 0) 646 goto err; 647 648 /* If reserves cannot be used, ensure enough pages are in the pool */ 649 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 650 goto err; 651 652retry_cpuset: 653 cpuset_mems_cookie = read_mems_allowed_begin(); 654 zonelist = huge_zonelist(vma, address, 655 htlb_alloc_mask(h), &mpol, &nodemask); 656 657 for_each_zone_zonelist_nodemask(zone, z, zonelist, 658 MAX_NR_ZONES - 1, nodemask) { 659 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 660 page = dequeue_huge_page_node(h, zone_to_nid(zone)); 661 if (page) { 662 if (avoid_reserve) 663 break; 664 if (!vma_has_reserves(vma, chg)) 665 break; 666 667 SetPagePrivate(page); 668 h->resv_huge_pages--; 669 break; 670 } 671 } 672 } 673 674 mpol_cond_put(mpol); 675 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 676 goto retry_cpuset; 677 return page; 678 679err: 680 return NULL; 681} 682 683/* 684 * common helper functions for hstate_next_node_to_{alloc|free}. 685 * We may have allocated or freed a huge page based on a different 686 * nodes_allowed previously, so h->next_node_to_{alloc|free} might 687 * be outside of *nodes_allowed. Ensure that we use an allowed 688 * node for alloc or free. 689 */ 690static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 691{ 692 nid = next_node(nid, *nodes_allowed); 693 if (nid == MAX_NUMNODES) 694 nid = first_node(*nodes_allowed); 695 VM_BUG_ON(nid >= MAX_NUMNODES); 696 697 return nid; 698} 699 700static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 701{ 702 if (!node_isset(nid, *nodes_allowed)) 703 nid = next_node_allowed(nid, nodes_allowed); 704 return nid; 705} 706 707/* 708 * returns the previously saved node ["this node"] from which to 709 * allocate a persistent huge page for the pool and advance the 710 * next node from which to allocate, handling wrap at end of node 711 * mask. 712 */ 713static int hstate_next_node_to_alloc(struct hstate *h, 714 nodemask_t *nodes_allowed) 715{ 716 int nid; 717 718 VM_BUG_ON(!nodes_allowed); 719 720 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 721 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 722 723 return nid; 724} 725 726/* 727 * helper for free_pool_huge_page() - return the previously saved 728 * node ["this node"] from which to free a huge page. Advance the 729 * next node id whether or not we find a free huge page to free so 730 * that the next attempt to free addresses the next node. 731 */ 732static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 733{ 734 int nid; 735 736 VM_BUG_ON(!nodes_allowed); 737 738 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 739 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 740 741 return nid; 742} 743 744#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 745 for (nr_nodes = nodes_weight(*mask); \ 746 nr_nodes > 0 && \ 747 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 748 nr_nodes--) 749 750#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 751 for (nr_nodes = nodes_weight(*mask); \ 752 nr_nodes > 0 && \ 753 ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 754 nr_nodes--) 755 756#if defined(CONFIG_CMA) && defined(CONFIG_X86_64) 757static void destroy_compound_gigantic_page(struct page *page, 758 unsigned int order) 759{ 760 int i; 761 int nr_pages = 1 << order; 762 struct page *p = page + 1; 763 764 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 765 __ClearPageTail(p); 766 set_page_refcounted(p); 767 p->first_page = NULL; 768 } 769 770 set_compound_order(page, 0); 771 __ClearPageHead(page); 772} 773 774static void free_gigantic_page(struct page *page, unsigned int order) 775{ 776 free_contig_range(page_to_pfn(page), 1 << order); 777} 778 779static int __alloc_gigantic_page(unsigned long start_pfn, 780 unsigned long nr_pages) 781{ 782 unsigned long end_pfn = start_pfn + nr_pages; 783 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 784} 785 786static bool pfn_range_valid_gigantic(unsigned long start_pfn, 787 unsigned long nr_pages) 788{ 789 unsigned long i, end_pfn = start_pfn + nr_pages; 790 struct page *page; 791 792 for (i = start_pfn; i < end_pfn; i++) { 793 if (!pfn_valid(i)) 794 return false; 795 796 page = pfn_to_page(i); 797 798 if (PageReserved(page)) 799 return false; 800 801 if (page_count(page) > 0) 802 return false; 803 804 if (PageHuge(page)) 805 return false; 806 } 807 808 return true; 809} 810 811static bool zone_spans_last_pfn(const struct zone *zone, 812 unsigned long start_pfn, unsigned long nr_pages) 813{ 814 unsigned long last_pfn = start_pfn + nr_pages - 1; 815 return zone_spans_pfn(zone, last_pfn); 816} 817 818static struct page *alloc_gigantic_page(int nid, unsigned int order) 819{ 820 unsigned long nr_pages = 1 << order; 821 unsigned long ret, pfn, flags; 822 struct zone *z; 823 824 z = NODE_DATA(nid)->node_zones; 825 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 826 spin_lock_irqsave(&z->lock, flags); 827 828 pfn = ALIGN(z->zone_start_pfn, nr_pages); 829 while (zone_spans_last_pfn(z, pfn, nr_pages)) { 830 if (pfn_range_valid_gigantic(pfn, nr_pages)) { 831 /* 832 * We release the zone lock here because 833 * alloc_contig_range() will also lock the zone 834 * at some point. If there's an allocation 835 * spinning on this lock, it may win the race 836 * and cause alloc_contig_range() to fail... 837 */ 838 spin_unlock_irqrestore(&z->lock, flags); 839 ret = __alloc_gigantic_page(pfn, nr_pages); 840 if (!ret) 841 return pfn_to_page(pfn); 842 spin_lock_irqsave(&z->lock, flags); 843 } 844 pfn += nr_pages; 845 } 846 847 spin_unlock_irqrestore(&z->lock, flags); 848 } 849 850 return NULL; 851} 852 853static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 854static void prep_compound_gigantic_page(struct page *page, unsigned int order); 855 856static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 857{ 858 struct page *page; 859 860 page = alloc_gigantic_page(nid, huge_page_order(h)); 861 if (page) { 862 prep_compound_gigantic_page(page, huge_page_order(h)); 863 prep_new_huge_page(h, page, nid); 864 } 865 866 return page; 867} 868 869static int alloc_fresh_gigantic_page(struct hstate *h, 870 nodemask_t *nodes_allowed) 871{ 872 struct page *page = NULL; 873 int nr_nodes, node; 874 875 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 876 page = alloc_fresh_gigantic_page_node(h, node); 877 if (page) 878 return 1; 879 } 880 881 return 0; 882} 883 884static inline bool gigantic_page_supported(void) { return true; } 885#else 886static inline bool gigantic_page_supported(void) { return false; } 887static inline void free_gigantic_page(struct page *page, unsigned int order) { } 888static inline void destroy_compound_gigantic_page(struct page *page, 889 unsigned int order) { } 890static inline int alloc_fresh_gigantic_page(struct hstate *h, 891 nodemask_t *nodes_allowed) { return 0; } 892#endif 893 894static void update_and_free_page(struct hstate *h, struct page *page) 895{ 896 int i; 897 898 if (hstate_is_gigantic(h) && !gigantic_page_supported()) 899 return; 900 901 h->nr_huge_pages--; 902 h->nr_huge_pages_node[page_to_nid(page)]--; 903 for (i = 0; i < pages_per_huge_page(h); i++) { 904 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 905 1 << PG_referenced | 1 << PG_dirty | 906 1 << PG_active | 1 << PG_private | 907 1 << PG_writeback); 908 } 909 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 910 set_compound_page_dtor(page, NULL); 911 set_page_refcounted(page); 912 if (hstate_is_gigantic(h)) { 913 destroy_compound_gigantic_page(page, huge_page_order(h)); 914 free_gigantic_page(page, huge_page_order(h)); 915 } else { 916 arch_release_hugepage(page); 917 __free_pages(page, huge_page_order(h)); 918 } 919} 920 921struct hstate *size_to_hstate(unsigned long size) 922{ 923 struct hstate *h; 924 925 for_each_hstate(h) { 926 if (huge_page_size(h) == size) 927 return h; 928 } 929 return NULL; 930} 931 932/* 933 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked 934 * to hstate->hugepage_activelist.) 935 * 936 * This function can be called for tail pages, but never returns true for them. 937 */ 938bool page_huge_active(struct page *page) 939{ 940 VM_BUG_ON_PAGE(!PageHuge(page), page); 941 return PageHead(page) && PagePrivate(&page[1]); 942} 943 944/* never called for tail page */ 945static void set_page_huge_active(struct page *page) 946{ 947 VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 948 SetPagePrivate(&page[1]); 949} 950 951static void clear_page_huge_active(struct page *page) 952{ 953 VM_BUG_ON_PAGE(!PageHeadHuge(page), page); 954 ClearPagePrivate(&page[1]); 955} 956 957void free_huge_page(struct page *page) 958{ 959 /* 960 * Can't pass hstate in here because it is called from the 961 * compound page destructor. 962 */ 963 struct hstate *h = page_hstate(page); 964 int nid = page_to_nid(page); 965 struct hugepage_subpool *spool = 966 (struct hugepage_subpool *)page_private(page); 967 bool restore_reserve; 968 969 set_page_private(page, 0); 970 page->mapping = NULL; 971 BUG_ON(page_count(page)); 972 BUG_ON(page_mapcount(page)); 973 restore_reserve = PagePrivate(page); 974 ClearPagePrivate(page); 975 976 /* 977 * A return code of zero implies that the subpool will be under its 978 * minimum size if the reservation is not restored after page is free. 979 * Therefore, force restore_reserve operation. 980 */ 981 if (hugepage_subpool_put_pages(spool, 1) == 0) 982 restore_reserve = true; 983 984 spin_lock(&hugetlb_lock); 985 clear_page_huge_active(page); 986 hugetlb_cgroup_uncharge_page(hstate_index(h), 987 pages_per_huge_page(h), page); 988 if (restore_reserve) 989 h->resv_huge_pages++; 990 991 if (h->surplus_huge_pages_node[nid]) { 992 /* remove the page from active list */ 993 list_del(&page->lru); 994 update_and_free_page(h, page); 995 h->surplus_huge_pages--; 996 h->surplus_huge_pages_node[nid]--; 997 } else { 998 arch_clear_hugepage_flags(page); 999 enqueue_huge_page(h, page); 1000 } 1001 spin_unlock(&hugetlb_lock); 1002} 1003 1004static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 1005{ 1006 INIT_LIST_HEAD(&page->lru); 1007 set_compound_page_dtor(page, free_huge_page); 1008 spin_lock(&hugetlb_lock); 1009 set_hugetlb_cgroup(page, NULL); 1010 h->nr_huge_pages++; 1011 h->nr_huge_pages_node[nid]++; 1012 spin_unlock(&hugetlb_lock); 1013 put_page(page); /* free it into the hugepage allocator */ 1014} 1015 1016static void prep_compound_gigantic_page(struct page *page, unsigned int order) 1017{ 1018 int i; 1019 int nr_pages = 1 << order; 1020 struct page *p = page + 1; 1021 1022 /* we rely on prep_new_huge_page to set the destructor */ 1023 set_compound_order(page, order); 1024 __SetPageHead(page); 1025 __ClearPageReserved(page); 1026 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 1027 /* 1028 * For gigantic hugepages allocated through bootmem at 1029 * boot, it's safer to be consistent with the not-gigantic 1030 * hugepages and clear the PG_reserved bit from all tail pages 1031 * too. Otherwse drivers using get_user_pages() to access tail 1032 * pages may get the reference counting wrong if they see 1033 * PG_reserved set on a tail page (despite the head page not 1034 * having PG_reserved set). Enforcing this consistency between 1035 * head and tail pages allows drivers to optimize away a check 1036 * on the head page when they need know if put_page() is needed 1037 * after get_user_pages(). 1038 */ 1039 __ClearPageReserved(p); 1040 set_page_count(p, 0); 1041 p->first_page = page; 1042 /* Make sure p->first_page is always valid for PageTail() */ 1043 smp_wmb(); 1044 __SetPageTail(p); 1045 } 1046} 1047 1048/* 1049 * PageHuge() only returns true for hugetlbfs pages, but not for normal or 1050 * transparent huge pages. See the PageTransHuge() documentation for more 1051 * details. 1052 */ 1053int PageHuge(struct page *page) 1054{ 1055 if (!PageCompound(page)) 1056 return 0; 1057 1058 page = compound_head(page); 1059 return get_compound_page_dtor(page) == free_huge_page; 1060} 1061EXPORT_SYMBOL_GPL(PageHuge); 1062 1063/* 1064 * PageHeadHuge() only returns true for hugetlbfs head page, but not for 1065 * normal or transparent huge pages. 1066 */ 1067int PageHeadHuge(struct page *page_head) 1068{ 1069 if (!PageHead(page_head)) 1070 return 0; 1071 1072 return get_compound_page_dtor(page_head) == free_huge_page; 1073} 1074 1075pgoff_t __basepage_index(struct page *page) 1076{ 1077 struct page *page_head = compound_head(page); 1078 pgoff_t index = page_index(page_head); 1079 unsigned long compound_idx; 1080 1081 if (!PageHuge(page_head)) 1082 return page_index(page); 1083 1084 if (compound_order(page_head) >= MAX_ORDER) 1085 compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 1086 else 1087 compound_idx = page - page_head; 1088 1089 return (index << compound_order(page_head)) + compound_idx; 1090} 1091 1092static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 1093{ 1094 struct page *page; 1095 1096 page = alloc_pages_exact_node(nid, 1097 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1098 __GFP_REPEAT|__GFP_NOWARN, 1099 huge_page_order(h)); 1100 if (page) { 1101 if (arch_prepare_hugepage(page)) { 1102 __free_pages(page, huge_page_order(h)); 1103 return NULL; 1104 } 1105 prep_new_huge_page(h, page, nid); 1106 } 1107 1108 return page; 1109} 1110 1111static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1112{ 1113 struct page *page; 1114 int nr_nodes, node; 1115 int ret = 0; 1116 1117 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1118 page = alloc_fresh_huge_page_node(h, node); 1119 if (page) { 1120 ret = 1; 1121 break; 1122 } 1123 } 1124 1125 if (ret) 1126 count_vm_event(HTLB_BUDDY_PGALLOC); 1127 else 1128 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1129 1130 return ret; 1131} 1132 1133/* 1134 * Free huge page from pool from next node to free. 1135 * Attempt to keep persistent huge pages more or less 1136 * balanced over allowed nodes. 1137 * Called with hugetlb_lock locked. 1138 */ 1139static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1140 bool acct_surplus) 1141{ 1142 int nr_nodes, node; 1143 int ret = 0; 1144 1145 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1146 /* 1147 * If we're returning unused surplus pages, only examine 1148 * nodes with surplus pages. 1149 */ 1150 if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1151 !list_empty(&h->hugepage_freelists[node])) { 1152 struct page *page = 1153 list_entry(h->hugepage_freelists[node].next, 1154 struct page, lru); 1155 list_del(&page->lru); 1156 h->free_huge_pages--; 1157 h->free_huge_pages_node[node]--; 1158 if (acct_surplus) { 1159 h->surplus_huge_pages--; 1160 h->surplus_huge_pages_node[node]--; 1161 } 1162 update_and_free_page(h, page); 1163 ret = 1; 1164 break; 1165 } 1166 } 1167 1168 return ret; 1169} 1170 1171/* 1172 * Dissolve a given free hugepage into free buddy pages. This function does 1173 * nothing for in-use (including surplus) hugepages. 1174 */ 1175static void dissolve_free_huge_page(struct page *page) 1176{ 1177 spin_lock(&hugetlb_lock); 1178 if (PageHuge(page) && !page_count(page)) { 1179 struct hstate *h = page_hstate(page); 1180 int nid = page_to_nid(page); 1181 list_del(&page->lru); 1182 h->free_huge_pages--; 1183 h->free_huge_pages_node[nid]--; 1184 update_and_free_page(h, page); 1185 } 1186 spin_unlock(&hugetlb_lock); 1187} 1188 1189/* 1190 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1191 * make specified memory blocks removable from the system. 1192 * Note that start_pfn should aligned with (minimum) hugepage size. 1193 */ 1194void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1195{ 1196 unsigned long pfn; 1197 1198 if (!hugepages_supported()) 1199 return; 1200 1201 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); 1202 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) 1203 dissolve_free_huge_page(pfn_to_page(pfn)); 1204} 1205 1206static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 1207{ 1208 struct page *page; 1209 unsigned int r_nid; 1210 1211 if (hstate_is_gigantic(h)) 1212 return NULL; 1213 1214 /* 1215 * Assume we will successfully allocate the surplus page to 1216 * prevent racing processes from causing the surplus to exceed 1217 * overcommit 1218 * 1219 * This however introduces a different race, where a process B 1220 * tries to grow the static hugepage pool while alloc_pages() is 1221 * called by process A. B will only examine the per-node 1222 * counters in determining if surplus huge pages can be 1223 * converted to normal huge pages in adjust_pool_surplus(). A 1224 * won't be able to increment the per-node counter, until the 1225 * lock is dropped by B, but B doesn't drop hugetlb_lock until 1226 * no more huge pages can be converted from surplus to normal 1227 * state (and doesn't try to convert again). Thus, we have a 1228 * case where a surplus huge page exists, the pool is grown, and 1229 * the surplus huge page still exists after, even though it 1230 * should just have been converted to a normal huge page. This 1231 * does not leak memory, though, as the hugepage will be freed 1232 * once it is out of use. It also does not allow the counters to 1233 * go out of whack in adjust_pool_surplus() as we don't modify 1234 * the node values until we've gotten the hugepage and only the 1235 * per-node value is checked there. 1236 */ 1237 spin_lock(&hugetlb_lock); 1238 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1239 spin_unlock(&hugetlb_lock); 1240 return NULL; 1241 } else { 1242 h->nr_huge_pages++; 1243 h->surplus_huge_pages++; 1244 } 1245 spin_unlock(&hugetlb_lock); 1246 1247 if (nid == NUMA_NO_NODE) 1248 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| 1249 __GFP_REPEAT|__GFP_NOWARN, 1250 huge_page_order(h)); 1251 else 1252 page = alloc_pages_exact_node(nid, 1253 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1254 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 1255 1256 if (page && arch_prepare_hugepage(page)) { 1257 __free_pages(page, huge_page_order(h)); 1258 page = NULL; 1259 } 1260 1261 spin_lock(&hugetlb_lock); 1262 if (page) { 1263 INIT_LIST_HEAD(&page->lru); 1264 r_nid = page_to_nid(page); 1265 set_compound_page_dtor(page, free_huge_page); 1266 set_hugetlb_cgroup(page, NULL); 1267 /* 1268 * We incremented the global counters already 1269 */ 1270 h->nr_huge_pages_node[r_nid]++; 1271 h->surplus_huge_pages_node[r_nid]++; 1272 __count_vm_event(HTLB_BUDDY_PGALLOC); 1273 } else { 1274 h->nr_huge_pages--; 1275 h->surplus_huge_pages--; 1276 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1277 } 1278 spin_unlock(&hugetlb_lock); 1279 1280 return page; 1281} 1282 1283/* 1284 * This allocation function is useful in the context where vma is irrelevant. 1285 * E.g. soft-offlining uses this function because it only cares physical 1286 * address of error page. 1287 */ 1288struct page *alloc_huge_page_node(struct hstate *h, int nid) 1289{ 1290 struct page *page = NULL; 1291 1292 spin_lock(&hugetlb_lock); 1293 if (h->free_huge_pages - h->resv_huge_pages > 0) 1294 page = dequeue_huge_page_node(h, nid); 1295 spin_unlock(&hugetlb_lock); 1296 1297 if (!page) 1298 page = alloc_buddy_huge_page(h, nid); 1299 1300 return page; 1301} 1302 1303/* 1304 * Increase the hugetlb pool such that it can accommodate a reservation 1305 * of size 'delta'. 1306 */ 1307static int gather_surplus_pages(struct hstate *h, int delta) 1308{ 1309 struct list_head surplus_list; 1310 struct page *page, *tmp; 1311 int ret, i; 1312 int needed, allocated; 1313 bool alloc_ok = true; 1314 1315 needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1316 if (needed <= 0) { 1317 h->resv_huge_pages += delta; 1318 return 0; 1319 } 1320 1321 allocated = 0; 1322 INIT_LIST_HEAD(&surplus_list); 1323 1324 ret = -ENOMEM; 1325retry: 1326 spin_unlock(&hugetlb_lock); 1327 for (i = 0; i < needed; i++) { 1328 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1329 if (!page) { 1330 alloc_ok = false; 1331 break; 1332 } 1333 list_add(&page->lru, &surplus_list); 1334 } 1335 allocated += i; 1336 1337 /* 1338 * After retaking hugetlb_lock, we need to recalculate 'needed' 1339 * because either resv_huge_pages or free_huge_pages may have changed. 1340 */ 1341 spin_lock(&hugetlb_lock); 1342 needed = (h->resv_huge_pages + delta) - 1343 (h->free_huge_pages + allocated); 1344 if (needed > 0) { 1345 if (alloc_ok) 1346 goto retry; 1347 /* 1348 * We were not able to allocate enough pages to 1349 * satisfy the entire reservation so we free what 1350 * we've allocated so far. 1351 */ 1352 goto free; 1353 } 1354 /* 1355 * The surplus_list now contains _at_least_ the number of extra pages 1356 * needed to accommodate the reservation. Add the appropriate number 1357 * of pages to the hugetlb pool and free the extras back to the buddy 1358 * allocator. Commit the entire reservation here to prevent another 1359 * process from stealing the pages as they are added to the pool but 1360 * before they are reserved. 1361 */ 1362 needed += allocated; 1363 h->resv_huge_pages += delta; 1364 ret = 0; 1365 1366 /* Free the needed pages to the hugetlb pool */ 1367 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1368 if ((--needed) < 0) 1369 break; 1370 /* 1371 * This page is now managed by the hugetlb allocator and has 1372 * no users -- drop the buddy allocator's reference. 1373 */ 1374 put_page_testzero(page); 1375 VM_BUG_ON_PAGE(page_count(page), page); 1376 enqueue_huge_page(h, page); 1377 } 1378free: 1379 spin_unlock(&hugetlb_lock); 1380 1381 /* Free unnecessary surplus pages to the buddy allocator */ 1382 list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1383 put_page(page); 1384 spin_lock(&hugetlb_lock); 1385 1386 return ret; 1387} 1388 1389/* 1390 * When releasing a hugetlb pool reservation, any surplus pages that were 1391 * allocated to satisfy the reservation must be explicitly freed if they were 1392 * never used. 1393 * Called with hugetlb_lock held. 1394 */ 1395static void return_unused_surplus_pages(struct hstate *h, 1396 unsigned long unused_resv_pages) 1397{ 1398 unsigned long nr_pages; 1399 1400 /* Uncommit the reservation */ 1401 h->resv_huge_pages -= unused_resv_pages; 1402 1403 /* Cannot return gigantic pages currently */ 1404 if (hstate_is_gigantic(h)) 1405 return; 1406 1407 nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1408 1409 /* 1410 * We want to release as many surplus pages as possible, spread 1411 * evenly across all nodes with memory. Iterate across these nodes 1412 * until we can no longer free unreserved surplus pages. This occurs 1413 * when the nodes with surplus pages have no free pages. 1414 * free_pool_huge_page() will balance the the freed pages across the 1415 * on-line nodes with memory and will handle the hstate accounting. 1416 */ 1417 while (nr_pages--) { 1418 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1419 break; 1420 cond_resched_lock(&hugetlb_lock); 1421 } 1422} 1423 1424/* 1425 * Determine if the huge page at addr within the vma has an associated 1426 * reservation. Where it does not we will need to logically increase 1427 * reservation and actually increase subpool usage before an allocation 1428 * can occur. Where any new reservation would be required the 1429 * reservation change is prepared, but not committed. Once the page 1430 * has been allocated from the subpool and instantiated the change should 1431 * be committed via vma_commit_reservation. No action is required on 1432 * failure. 1433 */ 1434static long vma_needs_reservation(struct hstate *h, 1435 struct vm_area_struct *vma, unsigned long addr) 1436{ 1437 struct resv_map *resv; 1438 pgoff_t idx; 1439 long chg; 1440 1441 resv = vma_resv_map(vma); 1442 if (!resv) 1443 return 1; 1444 1445 idx = vma_hugecache_offset(h, vma, addr); 1446 chg = region_chg(resv, idx, idx + 1); 1447 1448 if (vma->vm_flags & VM_MAYSHARE) 1449 return chg; 1450 else 1451 return chg < 0 ? chg : 0; 1452} 1453static void vma_commit_reservation(struct hstate *h, 1454 struct vm_area_struct *vma, unsigned long addr) 1455{ 1456 struct resv_map *resv; 1457 pgoff_t idx; 1458 1459 resv = vma_resv_map(vma); 1460 if (!resv) 1461 return; 1462 1463 idx = vma_hugecache_offset(h, vma, addr); 1464 region_add(resv, idx, idx + 1); 1465} 1466 1467static struct page *alloc_huge_page(struct vm_area_struct *vma, 1468 unsigned long addr, int avoid_reserve) 1469{ 1470 struct hugepage_subpool *spool = subpool_vma(vma); 1471 struct hstate *h = hstate_vma(vma); 1472 struct page *page; 1473 long chg; 1474 int ret, idx; 1475 struct hugetlb_cgroup *h_cg; 1476 1477 idx = hstate_index(h); 1478 /* 1479 * Processes that did not create the mapping will have no 1480 * reserves and will not have accounted against subpool 1481 * limit. Check that the subpool limit can be made before 1482 * satisfying the allocation MAP_NORESERVE mappings may also 1483 * need pages and subpool limit allocated allocated if no reserve 1484 * mapping overlaps. 1485 */ 1486 chg = vma_needs_reservation(h, vma, addr); 1487 if (chg < 0) 1488 return ERR_PTR(-ENOMEM); 1489 if (chg || avoid_reserve) 1490 if (hugepage_subpool_get_pages(spool, 1) < 0) 1491 return ERR_PTR(-ENOSPC); 1492 1493 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 1494 if (ret) 1495 goto out_subpool_put; 1496 1497 spin_lock(&hugetlb_lock); 1498 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 1499 if (!page) { 1500 spin_unlock(&hugetlb_lock); 1501 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1502 if (!page) 1503 goto out_uncharge_cgroup; 1504 1505 spin_lock(&hugetlb_lock); 1506 list_move(&page->lru, &h->hugepage_activelist); 1507 /* Fall through */ 1508 } 1509 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 1510 spin_unlock(&hugetlb_lock); 1511 1512 set_page_private(page, (unsigned long)spool); 1513 1514 vma_commit_reservation(h, vma, addr); 1515 return page; 1516 1517out_uncharge_cgroup: 1518 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 1519out_subpool_put: 1520 if (chg || avoid_reserve) 1521 hugepage_subpool_put_pages(spool, 1); 1522 return ERR_PTR(-ENOSPC); 1523} 1524 1525/* 1526 * alloc_huge_page()'s wrapper which simply returns the page if allocation 1527 * succeeds, otherwise NULL. This function is called from new_vma_page(), 1528 * where no ERR_VALUE is expected to be returned. 1529 */ 1530struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 1531 unsigned long addr, int avoid_reserve) 1532{ 1533 struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 1534 if (IS_ERR(page)) 1535 page = NULL; 1536 return page; 1537} 1538 1539int __weak alloc_bootmem_huge_page(struct hstate *h) 1540{ 1541 struct huge_bootmem_page *m; 1542 int nr_nodes, node; 1543 1544 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1545 void *addr; 1546 1547 addr = memblock_virt_alloc_try_nid_nopanic( 1548 huge_page_size(h), huge_page_size(h), 1549 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1550 if (addr) { 1551 /* 1552 * Use the beginning of the huge page to store the 1553 * huge_bootmem_page struct (until gather_bootmem 1554 * puts them into the mem_map). 1555 */ 1556 m = addr; 1557 goto found; 1558 } 1559 } 1560 return 0; 1561 1562found: 1563 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 1564 /* Put them into a private list first because mem_map is not up yet */ 1565 list_add(&m->list, &huge_boot_pages); 1566 m->hstate = h; 1567 return 1; 1568} 1569 1570static void __init prep_compound_huge_page(struct page *page, 1571 unsigned int order) 1572{ 1573 if (unlikely(order > (MAX_ORDER - 1))) 1574 prep_compound_gigantic_page(page, order); 1575 else 1576 prep_compound_page(page, order); 1577} 1578 1579/* Put bootmem huge pages into the standard lists after mem_map is up */ 1580static void __init gather_bootmem_prealloc(void) 1581{ 1582 struct huge_bootmem_page *m; 1583 1584 list_for_each_entry(m, &huge_boot_pages, list) { 1585 struct hstate *h = m->hstate; 1586 struct page *page; 1587 1588#ifdef CONFIG_HIGHMEM 1589 page = pfn_to_page(m->phys >> PAGE_SHIFT); 1590 memblock_free_late(__pa(m), 1591 sizeof(struct huge_bootmem_page)); 1592#else 1593 page = virt_to_page(m); 1594#endif 1595 WARN_ON(page_count(page) != 1); 1596 prep_compound_huge_page(page, h->order); 1597 WARN_ON(PageReserved(page)); 1598 prep_new_huge_page(h, page, page_to_nid(page)); 1599 /* 1600 * If we had gigantic hugepages allocated at boot time, we need 1601 * to restore the 'stolen' pages to totalram_pages in order to 1602 * fix confusing memory reports from free(1) and another 1603 * side-effects, like CommitLimit going negative. 1604 */ 1605 if (hstate_is_gigantic(h)) 1606 adjust_managed_page_count(page, 1 << h->order); 1607 } 1608} 1609 1610static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 1611{ 1612 unsigned long i; 1613 1614 for (i = 0; i < h->max_huge_pages; ++i) { 1615 if (hstate_is_gigantic(h)) { 1616 if (!alloc_bootmem_huge_page(h)) 1617 break; 1618 } else if (!alloc_fresh_huge_page(h, 1619 &node_states[N_MEMORY])) 1620 break; 1621 } 1622 h->max_huge_pages = i; 1623} 1624 1625static void __init hugetlb_init_hstates(void) 1626{ 1627 struct hstate *h; 1628 1629 for_each_hstate(h) { 1630 if (minimum_order > huge_page_order(h)) 1631 minimum_order = huge_page_order(h); 1632 1633 /* oversize hugepages were init'ed in early boot */ 1634 if (!hstate_is_gigantic(h)) 1635 hugetlb_hstate_alloc_pages(h); 1636 } 1637 VM_BUG_ON(minimum_order == UINT_MAX); 1638} 1639 1640static char * __init memfmt(char *buf, unsigned long n) 1641{ 1642 if (n >= (1UL << 30)) 1643 sprintf(buf, "%lu GB", n >> 30); 1644 else if (n >= (1UL << 20)) 1645 sprintf(buf, "%lu MB", n >> 20); 1646 else 1647 sprintf(buf, "%lu KB", n >> 10); 1648 return buf; 1649} 1650 1651static void __init report_hugepages(void) 1652{ 1653 struct hstate *h; 1654 1655 for_each_hstate(h) { 1656 char buf[32]; 1657 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 1658 memfmt(buf, huge_page_size(h)), 1659 h->free_huge_pages); 1660 } 1661} 1662 1663#ifdef CONFIG_HIGHMEM 1664static void try_to_free_low(struct hstate *h, unsigned long count, 1665 nodemask_t *nodes_allowed) 1666{ 1667 int i; 1668 1669 if (hstate_is_gigantic(h)) 1670 return; 1671 1672 for_each_node_mask(i, *nodes_allowed) { 1673 struct page *page, *next; 1674 struct list_head *freel = &h->hugepage_freelists[i]; 1675 list_for_each_entry_safe(page, next, freel, lru) { 1676 if (count >= h->nr_huge_pages) 1677 return; 1678 if (PageHighMem(page)) 1679 continue; 1680 list_del(&page->lru); 1681 update_and_free_page(h, page); 1682 h->free_huge_pages--; 1683 h->free_huge_pages_node[page_to_nid(page)]--; 1684 } 1685 } 1686} 1687#else 1688static inline void try_to_free_low(struct hstate *h, unsigned long count, 1689 nodemask_t *nodes_allowed) 1690{ 1691} 1692#endif 1693 1694/* 1695 * Increment or decrement surplus_huge_pages. Keep node-specific counters 1696 * balanced by operating on them in a round-robin fashion. 1697 * Returns 1 if an adjustment was made. 1698 */ 1699static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 1700 int delta) 1701{ 1702 int nr_nodes, node; 1703 1704 VM_BUG_ON(delta != -1 && delta != 1); 1705 1706 if (delta < 0) { 1707 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1708 if (h->surplus_huge_pages_node[node]) 1709 goto found; 1710 } 1711 } else { 1712 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1713 if (h->surplus_huge_pages_node[node] < 1714 h->nr_huge_pages_node[node]) 1715 goto found; 1716 } 1717 } 1718 return 0; 1719 1720found: 1721 h->surplus_huge_pages += delta; 1722 h->surplus_huge_pages_node[node] += delta; 1723 return 1; 1724} 1725 1726#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1727static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 1728 nodemask_t *nodes_allowed) 1729{ 1730 unsigned long min_count, ret; 1731 1732 if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1733 return h->max_huge_pages; 1734 1735 /* 1736 * Increase the pool size 1737 * First take pages out of surplus state. Then make up the 1738 * remaining difference by allocating fresh huge pages. 1739 * 1740 * We might race with alloc_buddy_huge_page() here and be unable 1741 * to convert a surplus huge page to a normal huge page. That is 1742 * not critical, though, it just means the overall size of the 1743 * pool might be one hugepage larger than it needs to be, but 1744 * within all the constraints specified by the sysctls. 1745 */ 1746 spin_lock(&hugetlb_lock); 1747 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1748 if (!adjust_pool_surplus(h, nodes_allowed, -1)) 1749 break; 1750 } 1751 1752 while (count > persistent_huge_pages(h)) { 1753 /* 1754 * If this allocation races such that we no longer need the 1755 * page, free_huge_page will handle it by freeing the page 1756 * and reducing the surplus. 1757 */ 1758 spin_unlock(&hugetlb_lock); 1759 if (hstate_is_gigantic(h)) 1760 ret = alloc_fresh_gigantic_page(h, nodes_allowed); 1761 else 1762 ret = alloc_fresh_huge_page(h, nodes_allowed); 1763 spin_lock(&hugetlb_lock); 1764 if (!ret) 1765 goto out; 1766 1767 /* Bail for signals. Probably ctrl-c from user */ 1768 if (signal_pending(current)) 1769 goto out; 1770 } 1771 1772 /* 1773 * Decrease the pool size 1774 * First return free pages to the buddy allocator (being careful 1775 * to keep enough around to satisfy reservations). Then place 1776 * pages into surplus state as needed so the pool will shrink 1777 * to the desired size as pages become free. 1778 * 1779 * By placing pages into the surplus state independent of the 1780 * overcommit value, we are allowing the surplus pool size to 1781 * exceed overcommit. There are few sane options here. Since 1782 * alloc_buddy_huge_page() is checking the global counter, 1783 * though, we'll note that we're not allowed to exceed surplus 1784 * and won't grow the pool anywhere else. Not until one of the 1785 * sysctls are changed, or the surplus pages go out of use. 1786 */ 1787 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 1788 min_count = max(count, min_count); 1789 try_to_free_low(h, min_count, nodes_allowed); 1790 while (min_count < persistent_huge_pages(h)) { 1791 if (!free_pool_huge_page(h, nodes_allowed, 0)) 1792 break; 1793 cond_resched_lock(&hugetlb_lock); 1794 } 1795 while (count < persistent_huge_pages(h)) { 1796 if (!adjust_pool_surplus(h, nodes_allowed, 1)) 1797 break; 1798 } 1799out: 1800 ret = persistent_huge_pages(h); 1801 spin_unlock(&hugetlb_lock); 1802 return ret; 1803} 1804 1805#define HSTATE_ATTR_RO(_name) \ 1806 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1807 1808#define HSTATE_ATTR(_name) \ 1809 static struct kobj_attribute _name##_attr = \ 1810 __ATTR(_name, 0644, _name##_show, _name##_store) 1811 1812static struct kobject *hugepages_kobj; 1813static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1814 1815static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 1816 1817static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1818{ 1819 int i; 1820 1821 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1822 if (hstate_kobjs[i] == kobj) { 1823 if (nidp) 1824 *nidp = NUMA_NO_NODE; 1825 return &hstates[i]; 1826 } 1827 1828 return kobj_to_node_hstate(kobj, nidp); 1829} 1830 1831static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1832 struct kobj_attribute *attr, char *buf) 1833{ 1834 struct hstate *h; 1835 unsigned long nr_huge_pages; 1836 int nid; 1837 1838 h = kobj_to_hstate(kobj, &nid); 1839 if (nid == NUMA_NO_NODE) 1840 nr_huge_pages = h->nr_huge_pages; 1841 else 1842 nr_huge_pages = h->nr_huge_pages_node[nid]; 1843 1844 return sprintf(buf, "%lu\n", nr_huge_pages); 1845} 1846 1847static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 1848 struct hstate *h, int nid, 1849 unsigned long count, size_t len) 1850{ 1851 int err; 1852 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1853 1854 if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 1855 err = -EINVAL; 1856 goto out; 1857 } 1858 1859 if (nid == NUMA_NO_NODE) { 1860 /* 1861 * global hstate attribute 1862 */ 1863 if (!(obey_mempolicy && 1864 init_nodemask_of_mempolicy(nodes_allowed))) { 1865 NODEMASK_FREE(nodes_allowed); 1866 nodes_allowed = &node_states[N_MEMORY]; 1867 } 1868 } else if (nodes_allowed) { 1869 /* 1870 * per node hstate attribute: adjust count to global, 1871 * but restrict alloc/free to the specified node. 1872 */ 1873 count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 1874 init_nodemask_of_node(nodes_allowed, nid); 1875 } else 1876 nodes_allowed = &node_states[N_MEMORY]; 1877 1878 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1879 1880 if (nodes_allowed != &node_states[N_MEMORY]) 1881 NODEMASK_FREE(nodes_allowed); 1882 1883 return len; 1884out: 1885 NODEMASK_FREE(nodes_allowed); 1886 return err; 1887} 1888 1889static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 1890 struct kobject *kobj, const char *buf, 1891 size_t len) 1892{ 1893 struct hstate *h; 1894 unsigned long count; 1895 int nid; 1896 int err; 1897 1898 err = kstrtoul(buf, 10, &count); 1899 if (err) 1900 return err; 1901 1902 h = kobj_to_hstate(kobj, &nid); 1903 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 1904} 1905 1906static ssize_t nr_hugepages_show(struct kobject *kobj, 1907 struct kobj_attribute *attr, char *buf) 1908{ 1909 return nr_hugepages_show_common(kobj, attr, buf); 1910} 1911 1912static ssize_t nr_hugepages_store(struct kobject *kobj, 1913 struct kobj_attribute *attr, const char *buf, size_t len) 1914{ 1915 return nr_hugepages_store_common(false, kobj, buf, len); 1916} 1917HSTATE_ATTR(nr_hugepages); 1918 1919#ifdef CONFIG_NUMA 1920 1921/* 1922 * hstate attribute for optionally mempolicy-based constraint on persistent 1923 * huge page alloc/free. 1924 */ 1925static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 1926 struct kobj_attribute *attr, char *buf) 1927{ 1928 return nr_hugepages_show_common(kobj, attr, buf); 1929} 1930 1931static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 1932 struct kobj_attribute *attr, const char *buf, size_t len) 1933{ 1934 return nr_hugepages_store_common(true, kobj, buf, len); 1935} 1936HSTATE_ATTR(nr_hugepages_mempolicy); 1937#endif 1938 1939 1940static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1941 struct kobj_attribute *attr, char *buf) 1942{ 1943 struct hstate *h = kobj_to_hstate(kobj, NULL); 1944 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1945} 1946 1947static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1948 struct kobj_attribute *attr, const char *buf, size_t count) 1949{ 1950 int err; 1951 unsigned long input; 1952 struct hstate *h = kobj_to_hstate(kobj, NULL); 1953 1954 if (hstate_is_gigantic(h)) 1955 return -EINVAL; 1956 1957 err = kstrtoul(buf, 10, &input); 1958 if (err) 1959 return err; 1960 1961 spin_lock(&hugetlb_lock); 1962 h->nr_overcommit_huge_pages = input; 1963 spin_unlock(&hugetlb_lock); 1964 1965 return count; 1966} 1967HSTATE_ATTR(nr_overcommit_hugepages); 1968 1969static ssize_t free_hugepages_show(struct kobject *kobj, 1970 struct kobj_attribute *attr, char *buf) 1971{ 1972 struct hstate *h; 1973 unsigned long free_huge_pages; 1974 int nid; 1975 1976 h = kobj_to_hstate(kobj, &nid); 1977 if (nid == NUMA_NO_NODE) 1978 free_huge_pages = h->free_huge_pages; 1979 else 1980 free_huge_pages = h->free_huge_pages_node[nid]; 1981 1982 return sprintf(buf, "%lu\n", free_huge_pages); 1983} 1984HSTATE_ATTR_RO(free_hugepages); 1985 1986static ssize_t resv_hugepages_show(struct kobject *kobj, 1987 struct kobj_attribute *attr, char *buf) 1988{ 1989 struct hstate *h = kobj_to_hstate(kobj, NULL); 1990 return sprintf(buf, "%lu\n", h->resv_huge_pages); 1991} 1992HSTATE_ATTR_RO(resv_hugepages); 1993 1994static ssize_t surplus_hugepages_show(struct kobject *kobj, 1995 struct kobj_attribute *attr, char *buf) 1996{ 1997 struct hstate *h; 1998 unsigned long surplus_huge_pages; 1999 int nid; 2000 2001 h = kobj_to_hstate(kobj, &nid); 2002 if (nid == NUMA_NO_NODE) 2003 surplus_huge_pages = h->surplus_huge_pages; 2004 else 2005 surplus_huge_pages = h->surplus_huge_pages_node[nid]; 2006 2007 return sprintf(buf, "%lu\n", surplus_huge_pages); 2008} 2009HSTATE_ATTR_RO(surplus_hugepages); 2010 2011static struct attribute *hstate_attrs[] = { 2012 &nr_hugepages_attr.attr, 2013 &nr_overcommit_hugepages_attr.attr, 2014 &free_hugepages_attr.attr, 2015 &resv_hugepages_attr.attr, 2016 &surplus_hugepages_attr.attr, 2017#ifdef CONFIG_NUMA 2018 &nr_hugepages_mempolicy_attr.attr, 2019#endif 2020 NULL, 2021}; 2022 2023static struct attribute_group hstate_attr_group = { 2024 .attrs = hstate_attrs, 2025}; 2026 2027static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 2028 struct kobject **hstate_kobjs, 2029 struct attribute_group *hstate_attr_group) 2030{ 2031 int retval; 2032 int hi = hstate_index(h); 2033 2034 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 2035 if (!hstate_kobjs[hi]) 2036 return -ENOMEM; 2037 2038 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 2039 if (retval) 2040 kobject_put(hstate_kobjs[hi]); 2041 2042 return retval; 2043} 2044 2045static void __init hugetlb_sysfs_init(void) 2046{ 2047 struct hstate *h; 2048 int err; 2049 2050 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 2051 if (!hugepages_kobj) 2052 return; 2053 2054 for_each_hstate(h) { 2055 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 2056 hstate_kobjs, &hstate_attr_group); 2057 if (err) 2058 pr_err("Hugetlb: Unable to add hstate %s", h->name); 2059 } 2060} 2061 2062#ifdef CONFIG_NUMA 2063 2064/* 2065 * node_hstate/s - associate per node hstate attributes, via their kobjects, 2066 * with node devices in node_devices[] using a parallel array. The array 2067 * index of a node device or _hstate == node id. 2068 * This is here to avoid any static dependency of the node device driver, in 2069 * the base kernel, on the hugetlb module. 2070 */ 2071struct node_hstate { 2072 struct kobject *hugepages_kobj; 2073 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 2074}; 2075struct node_hstate node_hstates[MAX_NUMNODES]; 2076 2077/* 2078 * A subset of global hstate attributes for node devices 2079 */ 2080static struct attribute *per_node_hstate_attrs[] = { 2081 &nr_hugepages_attr.attr, 2082 &free_hugepages_attr.attr, 2083 &surplus_hugepages_attr.attr, 2084 NULL, 2085}; 2086 2087static struct attribute_group per_node_hstate_attr_group = { 2088 .attrs = per_node_hstate_attrs, 2089}; 2090 2091/* 2092 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 2093 * Returns node id via non-NULL nidp. 2094 */ 2095static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 2096{ 2097 int nid; 2098 2099 for (nid = 0; nid < nr_node_ids; nid++) { 2100 struct node_hstate *nhs = &node_hstates[nid]; 2101 int i; 2102 for (i = 0; i < HUGE_MAX_HSTATE; i++) 2103 if (nhs->hstate_kobjs[i] == kobj) { 2104 if (nidp) 2105 *nidp = nid; 2106 return &hstates[i]; 2107 } 2108 } 2109 2110 BUG(); 2111 return NULL; 2112} 2113 2114/* 2115 * Unregister hstate attributes from a single node device. 2116 * No-op if no hstate attributes attached. 2117 */ 2118static void hugetlb_unregister_node(struct node *node) 2119{ 2120 struct hstate *h; 2121 struct node_hstate *nhs = &node_hstates[node->dev.id]; 2122 2123 if (!nhs->hugepages_kobj) 2124 return; /* no hstate attributes */ 2125 2126 for_each_hstate(h) { 2127 int idx = hstate_index(h); 2128 if (nhs->hstate_kobjs[idx]) { 2129 kobject_put(nhs->hstate_kobjs[idx]); 2130 nhs->hstate_kobjs[idx] = NULL; 2131 } 2132 } 2133 2134 kobject_put(nhs->hugepages_kobj); 2135 nhs->hugepages_kobj = NULL; 2136} 2137 2138/* 2139 * hugetlb module exit: unregister hstate attributes from node devices 2140 * that have them. 2141 */ 2142static void hugetlb_unregister_all_nodes(void) 2143{ 2144 int nid; 2145 2146 /* 2147 * disable node device registrations. 2148 */ 2149 register_hugetlbfs_with_node(NULL, NULL); 2150 2151 /* 2152 * remove hstate attributes from any nodes that have them. 2153 */ 2154 for (nid = 0; nid < nr_node_ids; nid++) 2155 hugetlb_unregister_node(node_devices[nid]); 2156} 2157 2158/* 2159 * Register hstate attributes for a single node device. 2160 * No-op if attributes already registered. 2161 */ 2162static void hugetlb_register_node(struct node *node) 2163{ 2164 struct hstate *h; 2165 struct node_hstate *nhs = &node_hstates[node->dev.id]; 2166 int err; 2167 2168 if (nhs->hugepages_kobj) 2169 return; /* already allocated */ 2170 2171 nhs->hugepages_kobj = kobject_create_and_add("hugepages", 2172 &node->dev.kobj); 2173 if (!nhs->hugepages_kobj) 2174 return; 2175 2176 for_each_hstate(h) { 2177 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 2178 nhs->hstate_kobjs, 2179 &per_node_hstate_attr_group); 2180 if (err) { 2181 pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 2182 h->name, node->dev.id); 2183 hugetlb_unregister_node(node); 2184 break; 2185 } 2186 } 2187} 2188 2189/* 2190 * hugetlb init time: register hstate attributes for all registered node 2191 * devices of nodes that have memory. All on-line nodes should have 2192 * registered their associated device by this time. 2193 */ 2194static void __init hugetlb_register_all_nodes(void) 2195{ 2196 int nid; 2197 2198 for_each_node_state(nid, N_MEMORY) { 2199 struct node *node = node_devices[nid]; 2200 if (node->dev.id == nid) 2201 hugetlb_register_node(node); 2202 } 2203 2204 /* 2205 * Let the node device driver know we're here so it can 2206 * [un]register hstate attributes on node hotplug. 2207 */ 2208 register_hugetlbfs_with_node(hugetlb_register_node, 2209 hugetlb_unregister_node); 2210} 2211#else /* !CONFIG_NUMA */ 2212 2213static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 2214{ 2215 BUG(); 2216 if (nidp) 2217 *nidp = -1; 2218 return NULL; 2219} 2220 2221static void hugetlb_unregister_all_nodes(void) { } 2222 2223static void hugetlb_register_all_nodes(void) { } 2224 2225#endif 2226 2227static void __exit hugetlb_exit(void) 2228{ 2229 struct hstate *h; 2230 2231 hugetlb_unregister_all_nodes(); 2232 2233 for_each_hstate(h) { 2234 kobject_put(hstate_kobjs[hstate_index(h)]); 2235 } 2236 2237 kobject_put(hugepages_kobj); 2238 kfree(htlb_fault_mutex_table); 2239} 2240module_exit(hugetlb_exit); 2241 2242static int __init hugetlb_init(void) 2243{ 2244 int i; 2245 2246 if (!hugepages_supported()) 2247 return 0; 2248 2249 if (!size_to_hstate(default_hstate_size)) { 2250 default_hstate_size = HPAGE_SIZE; 2251 if (!size_to_hstate(default_hstate_size)) 2252 hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2253 } 2254 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2255 if (default_hstate_max_huge_pages) 2256 default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2257 2258 hugetlb_init_hstates(); 2259 gather_bootmem_prealloc(); 2260 report_hugepages(); 2261 2262 hugetlb_sysfs_init(); 2263 hugetlb_register_all_nodes(); 2264 hugetlb_cgroup_file_init(); 2265 2266#ifdef CONFIG_SMP 2267 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 2268#else 2269 num_fault_mutexes = 1; 2270#endif 2271 htlb_fault_mutex_table = 2272 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2273 BUG_ON(!htlb_fault_mutex_table); 2274 2275 for (i = 0; i < num_fault_mutexes; i++) 2276 mutex_init(&htlb_fault_mutex_table[i]); 2277 return 0; 2278} 2279module_init(hugetlb_init); 2280 2281/* Should be called on processing a hugepagesz=... option */ 2282void __init hugetlb_add_hstate(unsigned int order) 2283{ 2284 struct hstate *h; 2285 unsigned long i; 2286 2287 if (size_to_hstate(PAGE_SIZE << order)) { 2288 pr_warning("hugepagesz= specified twice, ignoring\n"); 2289 return; 2290 } 2291 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2292 BUG_ON(order == 0); 2293 h = &hstates[hugetlb_max_hstate++]; 2294 h->order = order; 2295 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 2296 h->nr_huge_pages = 0; 2297 h->free_huge_pages = 0; 2298 for (i = 0; i < MAX_NUMNODES; ++i) 2299 INIT_LIST_HEAD(&h->hugepage_freelists[i]); 2300 INIT_LIST_HEAD(&h->hugepage_activelist); 2301 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 2302 h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2303 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2304 huge_page_size(h)/1024); 2305 2306 parsed_hstate = h; 2307} 2308 2309static int __init hugetlb_nrpages_setup(char *s) 2310{ 2311 unsigned long *mhp; 2312 static unsigned long *last_mhp; 2313 2314 /* 2315 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2316 * so this hugepages= parameter goes to the "default hstate". 2317 */ 2318 if (!hugetlb_max_hstate) 2319 mhp = &default_hstate_max_huge_pages; 2320 else 2321 mhp = &parsed_hstate->max_huge_pages; 2322 2323 if (mhp == last_mhp) { 2324 pr_warning("hugepages= specified twice without " 2325 "interleaving hugepagesz=, ignoring\n"); 2326 return 1; 2327 } 2328 2329 if (sscanf(s, "%lu", mhp) <= 0) 2330 *mhp = 0; 2331 2332 /* 2333 * Global state is always initialized later in hugetlb_init. 2334 * But we need to allocate >= MAX_ORDER hstates here early to still 2335 * use the bootmem allocator. 2336 */ 2337 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 2338 hugetlb_hstate_alloc_pages(parsed_hstate); 2339 2340 last_mhp = mhp; 2341 2342 return 1; 2343} 2344__setup("hugepages=", hugetlb_nrpages_setup); 2345 2346static int __init hugetlb_default_setup(char *s) 2347{ 2348 default_hstate_size = memparse(s, &s); 2349 return 1; 2350} 2351__setup("default_hugepagesz=", hugetlb_default_setup); 2352 2353static unsigned int cpuset_mems_nr(unsigned int *array) 2354{ 2355 int node; 2356 unsigned int nr = 0; 2357 2358 for_each_node_mask(node, cpuset_current_mems_allowed) 2359 nr += array[node]; 2360 2361 return nr; 2362} 2363 2364#ifdef CONFIG_SYSCTL 2365static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 2366 struct ctl_table *table, int write, 2367 void __user *buffer, size_t *length, loff_t *ppos) 2368{ 2369 struct hstate *h = &default_hstate; 2370 unsigned long tmp = h->max_huge_pages; 2371 int ret; 2372 2373 if (!hugepages_supported()) 2374 return -ENOTSUPP; 2375 2376 table->data = &tmp; 2377 table->maxlen = sizeof(unsigned long); 2378 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 2379 if (ret) 2380 goto out; 2381 2382 if (write) 2383 ret = __nr_hugepages_store_common(obey_mempolicy, h, 2384 NUMA_NO_NODE, tmp, *length); 2385out: 2386 return ret; 2387} 2388 2389int hugetlb_sysctl_handler(struct ctl_table *table, int write, 2390 void __user *buffer, size_t *length, loff_t *ppos) 2391{ 2392 2393 return hugetlb_sysctl_handler_common(false, table, write, 2394 buffer, length, ppos); 2395} 2396 2397#ifdef CONFIG_NUMA 2398int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 2399 void __user *buffer, size_t *length, loff_t *ppos) 2400{ 2401 return hugetlb_sysctl_handler_common(true, table, write, 2402 buffer, length, ppos); 2403} 2404#endif /* CONFIG_NUMA */ 2405 2406int hugetlb_overcommit_handler(struct ctl_table *table, int write, 2407 void __user *buffer, 2408 size_t *length, loff_t *ppos) 2409{ 2410 struct hstate *h = &default_hstate; 2411 unsigned long tmp; 2412 int ret; 2413 2414 if (!hugepages_supported()) 2415 return -ENOTSUPP; 2416 2417 tmp = h->nr_overcommit_huge_pages; 2418 2419 if (write && hstate_is_gigantic(h)) 2420 return -EINVAL; 2421 2422 table->data = &tmp; 2423 table->maxlen = sizeof(unsigned long); 2424 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 2425 if (ret) 2426 goto out; 2427 2428 if (write) { 2429 spin_lock(&hugetlb_lock); 2430 h->nr_overcommit_huge_pages = tmp; 2431 spin_unlock(&hugetlb_lock); 2432 } 2433out: 2434 return ret; 2435} 2436 2437#endif /* CONFIG_SYSCTL */ 2438 2439void hugetlb_report_meminfo(struct seq_file *m) 2440{ 2441 struct hstate *h = &default_hstate; 2442 if (!hugepages_supported()) 2443 return; 2444 seq_printf(m, 2445 "HugePages_Total: %5lu\n" 2446 "HugePages_Free: %5lu\n" 2447 "HugePages_Rsvd: %5lu\n" 2448 "HugePages_Surp: %5lu\n" 2449 "Hugepagesize: %8lu kB\n", 2450 h->nr_huge_pages, 2451 h->free_huge_pages, 2452 h->resv_huge_pages, 2453 h->surplus_huge_pages, 2454 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2455} 2456 2457int hugetlb_report_node_meminfo(int nid, char *buf) 2458{ 2459 struct hstate *h = &default_hstate; 2460 if (!hugepages_supported()) 2461 return 0; 2462 return sprintf(buf, 2463 "Node %d HugePages_Total: %5u\n" 2464 "Node %d HugePages_Free: %5u\n" 2465 "Node %d HugePages_Surp: %5u\n", 2466 nid, h->nr_huge_pages_node[nid], 2467 nid, h->free_huge_pages_node[nid], 2468 nid, h->surplus_huge_pages_node[nid]); 2469} 2470 2471void hugetlb_show_meminfo(void) 2472{ 2473 struct hstate *h; 2474 int nid; 2475 2476 if (!hugepages_supported()) 2477 return; 2478 2479 for_each_node_state(nid, N_MEMORY) 2480 for_each_hstate(h) 2481 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2482 nid, 2483 h->nr_huge_pages_node[nid], 2484 h->free_huge_pages_node[nid], 2485 h->surplus_huge_pages_node[nid], 2486 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2487} 2488 2489/* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 2490unsigned long hugetlb_total_pages(void) 2491{ 2492 struct hstate *h; 2493 unsigned long nr_total_pages = 0; 2494 2495 for_each_hstate(h) 2496 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2497 return nr_total_pages; 2498} 2499 2500static int hugetlb_acct_memory(struct hstate *h, long delta) 2501{ 2502 int ret = -ENOMEM; 2503 2504 spin_lock(&hugetlb_lock); 2505 /* 2506 * When cpuset is configured, it breaks the strict hugetlb page 2507 * reservation as the accounting is done on a global variable. Such 2508 * reservation is completely rubbish in the presence of cpuset because 2509 * the reservation is not checked against page availability for the 2510 * current cpuset. Application can still potentially OOM'ed by kernel 2511 * with lack of free htlb page in cpuset that the task is in. 2512 * Attempt to enforce strict accounting with cpuset is almost 2513 * impossible (or too ugly) because cpuset is too fluid that 2514 * task or memory node can be dynamically moved between cpusets. 2515 * 2516 * The change of semantics for shared hugetlb mapping with cpuset is 2517 * undesirable. However, in order to preserve some of the semantics, 2518 * we fall back to check against current free page availability as 2519 * a best attempt and hopefully to minimize the impact of changing 2520 * semantics that cpuset has. 2521 */ 2522 if (delta > 0) { 2523 if (gather_surplus_pages(h, delta) < 0) 2524 goto out; 2525 2526 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2527 return_unused_surplus_pages(h, delta); 2528 goto out; 2529 } 2530 } 2531 2532 ret = 0; 2533 if (delta < 0) 2534 return_unused_surplus_pages(h, (unsigned long) -delta); 2535 2536out: 2537 spin_unlock(&hugetlb_lock); 2538 return ret; 2539} 2540 2541static void hugetlb_vm_op_open(struct vm_area_struct *vma) 2542{ 2543 struct resv_map *resv = vma_resv_map(vma); 2544 2545 /* 2546 * This new VMA should share its siblings reservation map if present. 2547 * The VMA will only ever have a valid reservation map pointer where 2548 * it is being copied for another still existing VMA. As that VMA 2549 * has a reference to the reservation map it cannot disappear until 2550 * after this open call completes. It is therefore safe to take a 2551 * new reference here without additional locking. 2552 */ 2553 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2554 kref_get(&resv->refs); 2555} 2556 2557static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2558{ 2559 struct hstate *h = hstate_vma(vma); 2560 struct resv_map *resv = vma_resv_map(vma); 2561 struct hugepage_subpool *spool = subpool_vma(vma); 2562 unsigned long reserve, start, end; 2563 long gbl_reserve; 2564 2565 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2566 return; 2567 2568 start = vma_hugecache_offset(h, vma, vma->vm_start); 2569 end = vma_hugecache_offset(h, vma, vma->vm_end); 2570 2571 reserve = (end - start) - region_count(resv, start, end); 2572 2573 kref_put(&resv->refs, resv_map_release); 2574 2575 if (reserve) { 2576 /* 2577 * Decrement reserve counts. The global reserve count may be 2578 * adjusted if the subpool has a minimum size. 2579 */ 2580 gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 2581 hugetlb_acct_memory(h, -gbl_reserve); 2582 } 2583} 2584 2585/* 2586 * We cannot handle pagefaults against hugetlb pages at all. They cause 2587 * handle_mm_fault() to try to instantiate regular-sized pages in the 2588 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 2589 * this far. 2590 */ 2591static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 2592{ 2593 BUG(); 2594 return 0; 2595} 2596 2597const struct vm_operations_struct hugetlb_vm_ops = { 2598 .fault = hugetlb_vm_op_fault, 2599 .open = hugetlb_vm_op_open, 2600 .close = hugetlb_vm_op_close, 2601}; 2602 2603static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 2604 int writable) 2605{ 2606 pte_t entry; 2607 2608 if (writable) { 2609 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2610 vma->vm_page_prot))); 2611 } else { 2612 entry = huge_pte_wrprotect(mk_huge_pte(page, 2613 vma->vm_page_prot)); 2614 } 2615 entry = pte_mkyoung(entry); 2616 entry = pte_mkhuge(entry); 2617 entry = arch_make_huge_pte(entry, vma, page, writable); 2618 2619 return entry; 2620} 2621 2622static void set_huge_ptep_writable(struct vm_area_struct *vma, 2623 unsigned long address, pte_t *ptep) 2624{ 2625 pte_t entry; 2626 2627 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 2628 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 2629 update_mmu_cache(vma, address, ptep); 2630} 2631 2632static int is_hugetlb_entry_migration(pte_t pte) 2633{ 2634 swp_entry_t swp; 2635 2636 if (huge_pte_none(pte) || pte_present(pte)) 2637 return 0; 2638 swp = pte_to_swp_entry(pte); 2639 if (non_swap_entry(swp) && is_migration_entry(swp)) 2640 return 1; 2641 else 2642 return 0; 2643} 2644 2645static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2646{ 2647 swp_entry_t swp; 2648 2649 if (huge_pte_none(pte) || pte_present(pte)) 2650 return 0; 2651 swp = pte_to_swp_entry(pte); 2652 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2653 return 1; 2654 else 2655 return 0; 2656} 2657 2658int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 2659 struct vm_area_struct *vma) 2660{ 2661 pte_t *src_pte, *dst_pte, entry; 2662 struct page *ptepage; 2663 unsigned long addr; 2664 int cow; 2665 struct hstate *h = hstate_vma(vma); 2666 unsigned long sz = huge_page_size(h); 2667 unsigned long mmun_start; /* For mmu_notifiers */ 2668 unsigned long mmun_end; /* For mmu_notifiers */ 2669 int ret = 0; 2670 2671 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 2672 2673 mmun_start = vma->vm_start; 2674 mmun_end = vma->vm_end; 2675 if (cow) 2676 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 2677 2678 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2679 spinlock_t *src_ptl, *dst_ptl; 2680 src_pte = huge_pte_offset(src, addr); 2681 if (!src_pte) 2682 continue; 2683 dst_pte = huge_pte_alloc(dst, addr, sz); 2684 if (!dst_pte) { 2685 ret = -ENOMEM; 2686 break; 2687 } 2688 2689 /* If the pagetables are shared don't copy or take references */ 2690 if (dst_pte == src_pte) 2691 continue; 2692 2693 dst_ptl = huge_pte_lock(h, dst, dst_pte); 2694 src_ptl = huge_pte_lockptr(h, src, src_pte); 2695 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 2696 entry = huge_ptep_get(src_pte); 2697 if (huge_pte_none(entry)) { /* skip none entry */ 2698 ; 2699 } else if (unlikely(is_hugetlb_entry_migration(entry) || 2700 is_hugetlb_entry_hwpoisoned(entry))) { 2701 swp_entry_t swp_entry = pte_to_swp_entry(entry); 2702 2703 if (is_write_migration_entry(swp_entry) && cow) { 2704 /* 2705 * COW mappings require pages in both 2706 * parent and child to be set to read. 2707 */ 2708 make_migration_entry_read(&swp_entry); 2709 entry = swp_entry_to_pte(swp_entry); 2710 set_huge_pte_at(src, addr, src_pte, entry); 2711 } 2712 set_huge_pte_at(dst, addr, dst_pte, entry); 2713 } else { 2714 if (cow) { 2715 huge_ptep_set_wrprotect(src, addr, src_pte); 2716 mmu_notifier_invalidate_range(src, mmun_start, 2717 mmun_end); 2718 } 2719 entry = huge_ptep_get(src_pte); 2720 ptepage = pte_page(entry); 2721 get_page(ptepage); 2722 page_dup_rmap(ptepage); 2723 set_huge_pte_at(dst, addr, dst_pte, entry); 2724 } 2725 spin_unlock(src_ptl); 2726 spin_unlock(dst_ptl); 2727 } 2728 2729 if (cow) 2730 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 2731 2732 return ret; 2733} 2734 2735void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 2736 unsigned long start, unsigned long end, 2737 struct page *ref_page) 2738{ 2739 int force_flush = 0; 2740 struct mm_struct *mm = vma->vm_mm; 2741 unsigned long address; 2742 pte_t *ptep; 2743 pte_t pte; 2744 spinlock_t *ptl; 2745 struct page *page; 2746 struct hstate *h = hstate_vma(vma); 2747 unsigned long sz = huge_page_size(h); 2748 const unsigned long mmun_start = start; /* For mmu_notifiers */ 2749 const unsigned long mmun_end = end; /* For mmu_notifiers */ 2750 2751 WARN_ON(!is_vm_hugetlb_page(vma)); 2752 BUG_ON(start & ~huge_page_mask(h)); 2753 BUG_ON(end & ~huge_page_mask(h)); 2754 2755 tlb_start_vma(tlb, vma); 2756 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2757 address = start; 2758again: 2759 for (; address < end; address += sz) { 2760 ptep = huge_pte_offset(mm, address); 2761 if (!ptep) 2762 continue; 2763 2764 ptl = huge_pte_lock(h, mm, ptep); 2765 if (huge_pmd_unshare(mm, &address, ptep)) 2766 goto unlock; 2767 2768 pte = huge_ptep_get(ptep); 2769 if (huge_pte_none(pte)) 2770 goto unlock; 2771 2772 /* 2773 * Migrating hugepage or HWPoisoned hugepage is already 2774 * unmapped and its refcount is dropped, so just clear pte here. 2775 */ 2776 if (unlikely(!pte_present(pte))) { 2777 huge_pte_clear(mm, address, ptep); 2778 goto unlock; 2779 } 2780 2781 page = pte_page(pte); 2782 /* 2783 * If a reference page is supplied, it is because a specific 2784 * page is being unmapped, not a range. Ensure the page we 2785 * are about to unmap is the actual page of interest. 2786 */ 2787 if (ref_page) { 2788 if (page != ref_page) 2789 goto unlock; 2790 2791 /* 2792 * Mark the VMA as having unmapped its page so that 2793 * future faults in this VMA will fail rather than 2794 * looking like data was lost 2795 */ 2796 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 2797 } 2798 2799 pte = huge_ptep_get_and_clear(mm, address, ptep); 2800 tlb_remove_tlb_entry(tlb, ptep, address); 2801 if (huge_pte_dirty(pte)) 2802 set_page_dirty(page); 2803 2804 page_remove_rmap(page); 2805 force_flush = !__tlb_remove_page(tlb, page); 2806 if (force_flush) { 2807 address += sz; 2808 spin_unlock(ptl); 2809 break; 2810 } 2811 /* Bail out after unmapping reference page if supplied */ 2812 if (ref_page) { 2813 spin_unlock(ptl); 2814 break; 2815 } 2816unlock: 2817 spin_unlock(ptl); 2818 } 2819 /* 2820 * mmu_gather ran out of room to batch pages, we break out of 2821 * the PTE lock to avoid doing the potential expensive TLB invalidate 2822 * and page-free while holding it. 2823 */ 2824 if (force_flush) { 2825 force_flush = 0; 2826 tlb_flush_mmu(tlb); 2827 if (address < end && !ref_page) 2828 goto again; 2829 } 2830 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2831 tlb_end_vma(tlb, vma); 2832} 2833 2834void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2835 struct vm_area_struct *vma, unsigned long start, 2836 unsigned long end, struct page *ref_page) 2837{ 2838 __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2839 2840 /* 2841 * Clear this flag so that x86's huge_pmd_share page_table_shareable 2842 * test will fail on a vma being torn down, and not grab a page table 2843 * on its way out. We're lucky that the flag has such an appropriate 2844 * name, and can in fact be safely cleared here. We could clear it 2845 * before the __unmap_hugepage_range above, but all that's necessary 2846 * is to clear it before releasing the i_mmap_rwsem. This works 2847 * because in the context this is called, the VMA is about to be 2848 * destroyed and the i_mmap_rwsem is held. 2849 */ 2850 vma->vm_flags &= ~VM_MAYSHARE; 2851} 2852 2853void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 2854 unsigned long end, struct page *ref_page) 2855{ 2856 struct mm_struct *mm; 2857 struct mmu_gather tlb; 2858 2859 mm = vma->vm_mm; 2860 2861 tlb_gather_mmu(&tlb, mm, start, end); 2862 __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 2863 tlb_finish_mmu(&tlb, start, end); 2864} 2865 2866/* 2867 * This is called when the original mapper is failing to COW a MAP_PRIVATE 2868 * mappping it owns the reserve page for. The intention is to unmap the page 2869 * from other VMAs and let the children be SIGKILLed if they are faulting the 2870 * same region. 2871 */ 2872static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 2873 struct page *page, unsigned long address) 2874{ 2875 struct hstate *h = hstate_vma(vma); 2876 struct vm_area_struct *iter_vma; 2877 struct address_space *mapping; 2878 pgoff_t pgoff; 2879 2880 /* 2881 * vm_pgoff is in PAGE_SIZE units, hence the different calculation 2882 * from page cache lookup which is in HPAGE_SIZE units. 2883 */ 2884 address = address & huge_page_mask(h); 2885 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 2886 vma->vm_pgoff; 2887 mapping = file_inode(vma->vm_file)->i_mapping; 2888 2889 /* 2890 * Take the mapping lock for the duration of the table walk. As 2891 * this mapping should be shared between all the VMAs, 2892 * __unmap_hugepage_range() is called as the lock is already held 2893 */ 2894 i_mmap_lock_write(mapping); 2895 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 2896 /* Do not unmap the current VMA */ 2897 if (iter_vma == vma) 2898 continue; 2899 2900 /* 2901 * Shared VMAs have their own reserves and do not affect 2902 * MAP_PRIVATE accounting but it is possible that a shared 2903 * VMA is using the same page so check and skip such VMAs. 2904 */ 2905 if (iter_vma->vm_flags & VM_MAYSHARE) 2906 continue; 2907 2908 /* 2909 * Unmap the page from other VMAs without their own reserves. 2910 * They get marked to be SIGKILLed if they fault in these 2911 * areas. This is because a future no-page fault on this VMA 2912 * could insert a zeroed page instead of the data existing 2913 * from the time of fork. This would look like data corruption 2914 */ 2915 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 2916 unmap_hugepage_range(iter_vma, address, 2917 address + huge_page_size(h), page); 2918 } 2919 i_mmap_unlock_write(mapping); 2920} 2921 2922/* 2923 * Hugetlb_cow() should be called with page lock of the original hugepage held. 2924 * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2925 * cannot race with other handlers or page migration. 2926 * Keep the pte_same checks anyway to make transition from the mutex easier. 2927 */ 2928static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 2929 unsigned long address, pte_t *ptep, pte_t pte, 2930 struct page *pagecache_page, spinlock_t *ptl) 2931{ 2932 struct hstate *h = hstate_vma(vma); 2933 struct page *old_page, *new_page; 2934 int ret = 0, outside_reserve = 0; 2935 unsigned long mmun_start; /* For mmu_notifiers */ 2936 unsigned long mmun_end; /* For mmu_notifiers */ 2937 2938 old_page = pte_page(pte); 2939 2940retry_avoidcopy: 2941 /* If no-one else is actually using this page, avoid the copy 2942 * and just make the page writable */ 2943 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 2944 page_move_anon_rmap(old_page, vma, address); 2945 set_huge_ptep_writable(vma, address, ptep); 2946 return 0; 2947 } 2948 2949 /* 2950 * If the process that created a MAP_PRIVATE mapping is about to 2951 * perform a COW due to a shared page count, attempt to satisfy 2952 * the allocation without using the existing reserves. The pagecache 2953 * page is used to determine if the reserve at this address was 2954 * consumed or not. If reserves were used, a partial faulted mapping 2955 * at the time of fork() could consume its reserves on COW instead 2956 * of the full address range. 2957 */ 2958 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 2959 old_page != pagecache_page) 2960 outside_reserve = 1; 2961 2962 page_cache_get(old_page); 2963 2964 /* 2965 * Drop page table lock as buddy allocator may be called. It will 2966 * be acquired again before returning to the caller, as expected. 2967 */ 2968 spin_unlock(ptl); 2969 new_page = alloc_huge_page(vma, address, outside_reserve); 2970 2971 if (IS_ERR(new_page)) { 2972 /* 2973 * If a process owning a MAP_PRIVATE mapping fails to COW, 2974 * it is due to references held by a child and an insufficient 2975 * huge page pool. To guarantee the original mappers 2976 * reliability, unmap the page from child processes. The child 2977 * may get SIGKILLed if it later faults. 2978 */ 2979 if (outside_reserve) { 2980 page_cache_release(old_page); 2981 BUG_ON(huge_pte_none(pte)); 2982 unmap_ref_private(mm, vma, old_page, address); 2983 BUG_ON(huge_pte_none(pte)); 2984 spin_lock(ptl); 2985 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2986 if (likely(ptep && 2987 pte_same(huge_ptep_get(ptep), pte))) 2988 goto retry_avoidcopy; 2989 /* 2990 * race occurs while re-acquiring page table 2991 * lock, and our job is done. 2992 */ 2993 return 0; 2994 } 2995 2996 ret = (PTR_ERR(new_page) == -ENOMEM) ? 2997 VM_FAULT_OOM : VM_FAULT_SIGBUS; 2998 goto out_release_old; 2999 } 3000 3001 /* 3002 * When the original hugepage is shared one, it does not have 3003 * anon_vma prepared. 3004 */ 3005 if (unlikely(anon_vma_prepare(vma))) { 3006 ret = VM_FAULT_OOM; 3007 goto out_release_all; 3008 } 3009 3010 copy_user_huge_page(new_page, old_page, address, vma, 3011 pages_per_huge_page(h)); 3012 __SetPageUptodate(new_page); 3013 set_page_huge_active(new_page); 3014 3015 mmun_start = address & huge_page_mask(h); 3016 mmun_end = mmun_start + huge_page_size(h); 3017 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 3018 3019 /* 3020 * Retake the page table lock to check for racing updates 3021 * before the page tables are altered 3022 */ 3023 spin_lock(ptl); 3024 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 3025 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 3026 ClearPagePrivate(new_page); 3027 3028 /* Break COW */ 3029 huge_ptep_clear_flush(vma, address, ptep); 3030 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 3031 set_huge_pte_at(mm, address, ptep, 3032 make_huge_pte(vma, new_page, 1)); 3033 page_remove_rmap(old_page); 3034 hugepage_add_new_anon_rmap(new_page, vma, address); 3035 /* Make the old page be freed below */ 3036 new_page = old_page; 3037 } 3038 spin_unlock(ptl); 3039 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 3040out_release_all: 3041 page_cache_release(new_page); 3042out_release_old: 3043 page_cache_release(old_page); 3044 3045 spin_lock(ptl); /* Caller expects lock to be held */ 3046 return ret; 3047} 3048 3049/* Return the pagecache page at a given address within a VMA */ 3050static struct page *hugetlbfs_pagecache_page(struct hstate *h, 3051 struct vm_area_struct *vma, unsigned long address) 3052{ 3053 struct address_space *mapping; 3054 pgoff_t idx; 3055 3056 mapping = vma->vm_file->f_mapping; 3057 idx = vma_hugecache_offset(h, vma, address); 3058 3059 return find_lock_page(mapping, idx); 3060} 3061 3062/* 3063 * Return whether there is a pagecache page to back given address within VMA. 3064 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 3065 */ 3066static bool hugetlbfs_pagecache_present(struct hstate *h, 3067 struct vm_area_struct *vma, unsigned long address) 3068{ 3069 struct address_space *mapping; 3070 pgoff_t idx; 3071 struct page *page; 3072 3073 mapping = vma->vm_file->f_mapping; 3074 idx = vma_hugecache_offset(h, vma, address); 3075 3076 page = find_get_page(mapping, idx); 3077 if (page) 3078 put_page(page); 3079 return page != NULL; 3080} 3081 3082static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 3083 struct address_space *mapping, pgoff_t idx, 3084 unsigned long address, pte_t *ptep, unsigned int flags) 3085{ 3086 struct hstate *h = hstate_vma(vma); 3087 int ret = VM_FAULT_SIGBUS; 3088 int anon_rmap = 0; 3089 unsigned long size; 3090 struct page *page; 3091 pte_t new_pte; 3092 spinlock_t *ptl; 3093 3094 /* 3095 * Currently, we are forced to kill the process in the event the 3096 * original mapper has unmapped pages from the child due to a failed 3097 * COW. Warn that such a situation has occurred as it may not be obvious 3098 */ 3099 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 3100 pr_warning("PID %d killed due to inadequate hugepage pool\n", 3101 current->pid); 3102 return ret; 3103 } 3104 3105 /* 3106 * Use page lock to guard against racing truncation 3107 * before we get page_table_lock. 3108 */ 3109retry: 3110 page = find_lock_page(mapping, idx); 3111 if (!page) { 3112 size = i_size_read(mapping->host) >> huge_page_shift(h); 3113 if (idx >= size) 3114 goto out; 3115 page = alloc_huge_page(vma, address, 0); 3116 if (IS_ERR(page)) { 3117 ret = PTR_ERR(page); 3118 if (ret == -ENOMEM) 3119 ret = VM_FAULT_OOM; 3120 else 3121 ret = VM_FAULT_SIGBUS; 3122 goto out; 3123 } 3124 clear_huge_page(page, address, pages_per_huge_page(h)); 3125 __SetPageUptodate(page); 3126 set_page_huge_active(page); 3127 3128 if (vma->vm_flags & VM_MAYSHARE) { 3129 int err; 3130 struct inode *inode = mapping->host; 3131 3132 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3133 if (err) { 3134 put_page(page); 3135 if (err == -EEXIST) 3136 goto retry; 3137 goto out; 3138 } 3139 ClearPagePrivate(page); 3140 3141 spin_lock(&inode->i_lock); 3142 inode->i_blocks += blocks_per_huge_page(h); 3143 spin_unlock(&inode->i_lock); 3144 } else { 3145 lock_page(page); 3146 if (unlikely(anon_vma_prepare(vma))) { 3147 ret = VM_FAULT_OOM; 3148 goto backout_unlocked; 3149 } 3150 anon_rmap = 1; 3151 } 3152 } else { 3153 /* 3154 * If memory error occurs between mmap() and fault, some process 3155 * don't have hwpoisoned swap entry for errored virtual address. 3156 * So we need to block hugepage fault by PG_hwpoison bit check. 3157 */ 3158 if (unlikely(PageHWPoison(page))) { 3159 ret = VM_FAULT_HWPOISON | 3160 VM_FAULT_SET_HINDEX(hstate_index(h)); 3161 goto backout_unlocked; 3162 } 3163 } 3164 3165 /* 3166 * If we are going to COW a private mapping later, we examine the 3167 * pending reservations for this page now. This will ensure that 3168 * any allocations necessary to record that reservation occur outside 3169 * the spinlock. 3170 */ 3171 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 3172 if (vma_needs_reservation(h, vma, address) < 0) { 3173 ret = VM_FAULT_OOM; 3174 goto backout_unlocked; 3175 } 3176 3177 ptl = huge_pte_lockptr(h, mm, ptep); 3178 spin_lock(ptl); 3179 size = i_size_read(mapping->host) >> huge_page_shift(h); 3180 if (idx >= size) 3181 goto backout; 3182 3183 ret = 0; 3184 if (!huge_pte_none(huge_ptep_get(ptep))) 3185 goto backout; 3186 3187 if (anon_rmap) { 3188 ClearPagePrivate(page); 3189 hugepage_add_new_anon_rmap(page, vma, address); 3190 } else 3191 page_dup_rmap(page); 3192 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 3193 && (vma->vm_flags & VM_SHARED))); 3194 set_huge_pte_at(mm, address, ptep, new_pte); 3195 3196 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3197 /* Optimization, do the COW without a second fault */ 3198 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 3199 } 3200 3201 spin_unlock(ptl); 3202 unlock_page(page); 3203out: 3204 return ret; 3205 3206backout: 3207 spin_unlock(ptl); 3208backout_unlocked: 3209 unlock_page(page); 3210 put_page(page); 3211 goto out; 3212} 3213 3214#ifdef CONFIG_SMP 3215static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 3216 struct vm_area_struct *vma, 3217 struct address_space *mapping, 3218 pgoff_t idx, unsigned long address) 3219{ 3220 unsigned long key[2]; 3221 u32 hash; 3222 3223 if (vma->vm_flags & VM_SHARED) { 3224 key[0] = (unsigned long) mapping; 3225 key[1] = idx; 3226 } else { 3227 key[0] = (unsigned long) mm; 3228 key[1] = address >> huge_page_shift(h); 3229 } 3230 3231 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 3232 3233 return hash & (num_fault_mutexes - 1); 3234} 3235#else 3236/* 3237 * For uniprocesor systems we always use a single mutex, so just 3238 * return 0 and avoid the hashing overhead. 3239 */ 3240static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 3241 struct vm_area_struct *vma, 3242 struct address_space *mapping, 3243 pgoff_t idx, unsigned long address) 3244{ 3245 return 0; 3246} 3247#endif 3248 3249int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3250 unsigned long address, unsigned int flags) 3251{ 3252 pte_t *ptep, entry; 3253 spinlock_t *ptl; 3254 int ret; 3255 u32 hash; 3256 pgoff_t idx; 3257 struct page *page = NULL; 3258 struct page *pagecache_page = NULL; 3259 struct hstate *h = hstate_vma(vma); 3260 struct address_space *mapping; 3261 int need_wait_lock = 0; 3262 3263 address &= huge_page_mask(h); 3264 3265 ptep = huge_pte_offset(mm, address); 3266 if (ptep) { 3267 entry = huge_ptep_get(ptep); 3268 if (unlikely(is_hugetlb_entry_migration(entry))) { 3269 migration_entry_wait_huge(vma, mm, ptep); 3270 return 0; 3271 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3272 return VM_FAULT_HWPOISON_LARGE | 3273 VM_FAULT_SET_HINDEX(hstate_index(h)); 3274 } 3275 3276 ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 3277 if (!ptep) 3278 return VM_FAULT_OOM; 3279 3280 mapping = vma->vm_file->f_mapping; 3281 idx = vma_hugecache_offset(h, vma, address); 3282 3283 /* 3284 * Serialize hugepage allocation and instantiation, so that we don't 3285 * get spurious allocation failures if two CPUs race to instantiate 3286 * the same page in the page cache. 3287 */ 3288 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); 3289 mutex_lock(&htlb_fault_mutex_table[hash]); 3290 3291 entry = huge_ptep_get(ptep); 3292 if (huge_pte_none(entry)) { 3293 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3294 goto out_mutex; 3295 } 3296 3297 ret = 0; 3298 3299 /* 3300 * entry could be a migration/hwpoison entry at this point, so this 3301 * check prevents the kernel from going below assuming that we have 3302 * a active hugepage in pagecache. This goto expects the 2nd page fault, 3303 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly 3304 * handle it. 3305 */ 3306 if (!pte_present(entry)) 3307 goto out_mutex; 3308 3309 /* 3310 * If we are going to COW the mapping later, we examine the pending 3311 * reservations for this page now. This will ensure that any 3312 * allocations necessary to record that reservation occur outside the 3313 * spinlock. For private mappings, we also lookup the pagecache 3314 * page now as it is used to determine if a reservation has been 3315 * consumed. 3316 */ 3317 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 3318 if (vma_needs_reservation(h, vma, address) < 0) { 3319 ret = VM_FAULT_OOM; 3320 goto out_mutex; 3321 } 3322 3323 if (!(vma->vm_flags & VM_MAYSHARE)) 3324 pagecache_page = hugetlbfs_pagecache_page(h, 3325 vma, address); 3326 } 3327 3328 ptl = huge_pte_lock(h, mm, ptep); 3329 3330 /* Check for a racing update before calling hugetlb_cow */ 3331 if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3332 goto out_ptl; 3333 3334 /* 3335 * hugetlb_cow() requires page locks of pte_page(entry) and 3336 * pagecache_page, so here we need take the former one 3337 * when page != pagecache_page or !pagecache_page. 3338 */ 3339 page = pte_page(entry); 3340 if (page != pagecache_page) 3341 if (!trylock_page(page)) { 3342 need_wait_lock = 1; 3343 goto out_ptl; 3344 } 3345 3346 get_page(page); 3347 3348 if (flags & FAULT_FLAG_WRITE) { 3349 if (!huge_pte_write(entry)) { 3350 ret = hugetlb_cow(mm, vma, address, ptep, entry, 3351 pagecache_page, ptl); 3352 goto out_put_page; 3353 } 3354 entry = huge_pte_mkdirty(entry); 3355 } 3356 entry = pte_mkyoung(entry); 3357 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3358 flags & FAULT_FLAG_WRITE)) 3359 update_mmu_cache(vma, address, ptep); 3360out_put_page: 3361 if (page != pagecache_page) 3362 unlock_page(page); 3363 put_page(page); 3364out_ptl: 3365 spin_unlock(ptl); 3366 3367 if (pagecache_page) { 3368 unlock_page(pagecache_page); 3369 put_page(pagecache_page); 3370 } 3371out_mutex: 3372 mutex_unlock(&htlb_fault_mutex_table[hash]); 3373 /* 3374 * Generally it's safe to hold refcount during waiting page lock. But 3375 * here we just wait to defer the next page fault to avoid busy loop and 3376 * the page is not used after unlocked before returning from the current 3377 * page fault. So we are safe from accessing freed page, even if we wait 3378 * here without taking refcount. 3379 */ 3380 if (need_wait_lock) 3381 wait_on_page_locked(page); 3382 return ret; 3383} 3384 3385long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 3386 struct page **pages, struct vm_area_struct **vmas, 3387 unsigned long *position, unsigned long *nr_pages, 3388 long i, unsigned int flags) 3389{ 3390 unsigned long pfn_offset; 3391 unsigned long vaddr = *position; 3392 unsigned long remainder = *nr_pages; 3393 struct hstate *h = hstate_vma(vma); 3394 3395 while (vaddr < vma->vm_end && remainder) { 3396 pte_t *pte; 3397 spinlock_t *ptl = NULL; 3398 int absent; 3399 struct page *page; 3400 3401 /* 3402 * If we have a pending SIGKILL, don't keep faulting pages and 3403 * potentially allocating memory. 3404 */ 3405 if (unlikely(fatal_signal_pending(current))) { 3406 remainder = 0; 3407 break; 3408 } 3409 3410 /* 3411 * Some archs (sparc64, sh*) have multiple pte_ts to 3412 * each hugepage. We have to make sure we get the 3413 * first, for the page indexing below to work. 3414 * 3415 * Note that page table lock is not held when pte is null. 3416 */ 3417 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3418 if (pte) 3419 ptl = huge_pte_lock(h, mm, pte); 3420 absent = !pte || huge_pte_none(huge_ptep_get(pte)); 3421 3422 /* 3423 * When coredumping, it suits get_dump_page if we just return 3424 * an error where there's an empty slot with no huge pagecache 3425 * to back it. This way, we avoid allocating a hugepage, and 3426 * the sparse dumpfile avoids allocating disk blocks, but its 3427 * huge holes still show up with zeroes where they need to be. 3428 */ 3429 if (absent && (flags & FOLL_DUMP) && 3430 !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3431 if (pte) 3432 spin_unlock(ptl); 3433 remainder = 0; 3434 break; 3435 } 3436 3437 /* 3438 * We need call hugetlb_fault for both hugepages under migration 3439 * (in which case hugetlb_fault waits for the migration,) and 3440 * hwpoisoned hugepages (in which case we need to prevent the 3441 * caller from accessing to them.) In order to do this, we use 3442 * here is_swap_pte instead of is_hugetlb_entry_migration and 3443 * is_hugetlb_entry_hwpoisoned. This is because it simply covers 3444 * both cases, and because we can't follow correct pages 3445 * directly from any kind of swap entries. 3446 */ 3447 if (absent || is_swap_pte(huge_ptep_get(pte)) || 3448 ((flags & FOLL_WRITE) && 3449 !huge_pte_write(huge_ptep_get(pte)))) { 3450 int ret; 3451 3452 if (pte) 3453 spin_unlock(ptl); 3454 ret = hugetlb_fault(mm, vma, vaddr, 3455 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3456 if (!(ret & VM_FAULT_ERROR)) 3457 continue; 3458 3459 remainder = 0; 3460 break; 3461 } 3462 3463 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 3464 page = pte_page(huge_ptep_get(pte)); 3465same_page: 3466 if (pages) { 3467 pages[i] = mem_map_offset(page, pfn_offset); 3468 get_page_foll(pages[i]); 3469 } 3470 3471 if (vmas) 3472 vmas[i] = vma; 3473 3474 vaddr += PAGE_SIZE; 3475 ++pfn_offset; 3476 --remainder; 3477 ++i; 3478 if (vaddr < vma->vm_end && remainder && 3479 pfn_offset < pages_per_huge_page(h)) { 3480 /* 3481 * We use pfn_offset to avoid touching the pageframes 3482 * of this compound page. 3483 */ 3484 goto same_page; 3485 } 3486 spin_unlock(ptl); 3487 } 3488 *nr_pages = remainder; 3489 *position = vaddr; 3490 3491 return i ? i : -EFAULT; 3492} 3493 3494unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 3495 unsigned long address, unsigned long end, pgprot_t newprot) 3496{ 3497 struct mm_struct *mm = vma->vm_mm; 3498 unsigned long start = address; 3499 pte_t *ptep; 3500 pte_t pte; 3501 struct hstate *h = hstate_vma(vma); 3502 unsigned long pages = 0; 3503 3504 BUG_ON(address >= end); 3505 flush_cache_range(vma, address, end); 3506 3507 mmu_notifier_invalidate_range_start(mm, start, end); 3508 i_mmap_lock_write(vma->vm_file->f_mapping); 3509 for (; address < end; address += huge_page_size(h)) { 3510 spinlock_t *ptl; 3511 ptep = huge_pte_offset(mm, address); 3512 if (!ptep) 3513 continue; 3514 ptl = huge_pte_lock(h, mm, ptep); 3515 if (huge_pmd_unshare(mm, &address, ptep)) { 3516 pages++; 3517 spin_unlock(ptl); 3518 continue; 3519 } 3520 pte = huge_ptep_get(ptep); 3521 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 3522 spin_unlock(ptl); 3523 continue; 3524 } 3525 if (unlikely(is_hugetlb_entry_migration(pte))) { 3526 swp_entry_t entry = pte_to_swp_entry(pte); 3527 3528 if (is_write_migration_entry(entry)) { 3529 pte_t newpte; 3530 3531 make_migration_entry_read(&entry); 3532 newpte = swp_entry_to_pte(entry); 3533 set_huge_pte_at(mm, address, ptep, newpte); 3534 pages++; 3535 } 3536 spin_unlock(ptl); 3537 continue; 3538 } 3539 if (!huge_pte_none(pte)) { 3540 pte = huge_ptep_get_and_clear(mm, address, ptep); 3541 pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3542 pte = arch_make_huge_pte(pte, vma, NULL, 0); 3543 set_huge_pte_at(mm, address, ptep, pte); 3544 pages++; 3545 } 3546 spin_unlock(ptl); 3547 } 3548 /* 3549 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 3550 * may have cleared our pud entry and done put_page on the page table: 3551 * once we release i_mmap_rwsem, another task can do the final put_page 3552 * and that page table be reused and filled with junk. 3553 */ 3554 flush_tlb_range(vma, start, end); 3555 mmu_notifier_invalidate_range(mm, start, end); 3556 i_mmap_unlock_write(vma->vm_file->f_mapping); 3557 mmu_notifier_invalidate_range_end(mm, start, end); 3558 3559 return pages << h->order; 3560} 3561 3562int hugetlb_reserve_pages(struct inode *inode, 3563 long from, long to, 3564 struct vm_area_struct *vma, 3565 vm_flags_t vm_flags) 3566{ 3567 long ret, chg; 3568 struct hstate *h = hstate_inode(inode); 3569 struct hugepage_subpool *spool = subpool_inode(inode); 3570 struct resv_map *resv_map; 3571 long gbl_reserve; 3572 3573 /* 3574 * Only apply hugepage reservation if asked. At fault time, an 3575 * attempt will be made for VM_NORESERVE to allocate a page 3576 * without using reserves 3577 */ 3578 if (vm_flags & VM_NORESERVE) 3579 return 0; 3580 3581 /* 3582 * Shared mappings base their reservation on the number of pages that 3583 * are already allocated on behalf of the file. Private mappings need 3584 * to reserve the full area even if read-only as mprotect() may be 3585 * called to make the mapping read-write. Assume !vma is a shm mapping 3586 */ 3587 if (!vma || vma->vm_flags & VM_MAYSHARE) { 3588 resv_map = inode_resv_map(inode); 3589 3590 chg = region_chg(resv_map, from, to); 3591 3592 } else { 3593 resv_map = resv_map_alloc(); 3594 if (!resv_map) 3595 return -ENOMEM; 3596 3597 chg = to - from; 3598 3599 set_vma_resv_map(vma, resv_map); 3600 set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 3601 } 3602 3603 if (chg < 0) { 3604 ret = chg; 3605 goto out_err; 3606 } 3607 3608 /* 3609 * There must be enough pages in the subpool for the mapping. If 3610 * the subpool has a minimum size, there may be some global 3611 * reservations already in place (gbl_reserve). 3612 */ 3613 gbl_reserve = hugepage_subpool_get_pages(spool, chg); 3614 if (gbl_reserve < 0) { 3615 ret = -ENOSPC; 3616 goto out_err; 3617 } 3618 3619 /* 3620 * Check enough hugepages are available for the reservation. 3621 * Hand the pages back to the subpool if there are not 3622 */ 3623 ret = hugetlb_acct_memory(h, gbl_reserve); 3624 if (ret < 0) { 3625 /* put back original number of pages, chg */ 3626 (void)hugepage_subpool_put_pages(spool, chg); 3627 goto out_err; 3628 } 3629 3630 /* 3631 * Account for the reservations made. Shared mappings record regions 3632 * that have reservations as they are shared by multiple VMAs. 3633 * When the last VMA disappears, the region map says how much 3634 * the reservation was and the page cache tells how much of 3635 * the reservation was consumed. Private mappings are per-VMA and 3636 * only the consumed reservations are tracked. When the VMA 3637 * disappears, the original reservation is the VMA size and the 3638 * consumed reservations are stored in the map. Hence, nothing 3639 * else has to be done for private mappings here 3640 */ 3641 if (!vma || vma->vm_flags & VM_MAYSHARE) 3642 region_add(resv_map, from, to); 3643 return 0; 3644out_err: 3645 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3646 kref_put(&resv_map->refs, resv_map_release); 3647 return ret; 3648} 3649 3650void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3651{ 3652 struct hstate *h = hstate_inode(inode); 3653 struct resv_map *resv_map = inode_resv_map(inode); 3654 long chg = 0; 3655 struct hugepage_subpool *spool = subpool_inode(inode); 3656 long gbl_reserve; 3657 3658 if (resv_map) 3659 chg = region_truncate(resv_map, offset); 3660 spin_lock(&inode->i_lock); 3661 inode->i_blocks -= (blocks_per_huge_page(h) * freed); 3662 spin_unlock(&inode->i_lock); 3663 3664 /* 3665 * If the subpool has a minimum size, the number of global 3666 * reservations to be released may be adjusted. 3667 */ 3668 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 3669 hugetlb_acct_memory(h, -gbl_reserve); 3670} 3671 3672#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 3673static unsigned long page_table_shareable(struct vm_area_struct *svma, 3674 struct vm_area_struct *vma, 3675 unsigned long addr, pgoff_t idx) 3676{ 3677 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 3678 svma->vm_start; 3679 unsigned long sbase = saddr & PUD_MASK; 3680 unsigned long s_end = sbase + PUD_SIZE; 3681 3682 /* Allow segments to share if only one is marked locked */ 3683 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 3684 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 3685 3686 /* 3687 * match the virtual addresses, permission and the alignment of the 3688 * page table page. 3689 */ 3690 if (pmd_index(addr) != pmd_index(saddr) || 3691 vm_flags != svm_flags || 3692 sbase < svma->vm_start || svma->vm_end < s_end) 3693 return 0; 3694 3695 return saddr; 3696} 3697 3698static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 3699{ 3700 unsigned long base = addr & PUD_MASK; 3701 unsigned long end = base + PUD_SIZE; 3702 3703 /* 3704 * check on proper vm_flags and page table alignment 3705 */ 3706 if (vma->vm_flags & VM_MAYSHARE && 3707 vma->vm_start <= base && end <= vma->vm_end) 3708 return 1; 3709 return 0; 3710} 3711 3712/* 3713 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 3714 * and returns the corresponding pte. While this is not necessary for the 3715 * !shared pmd case because we can allocate the pmd later as well, it makes the 3716 * code much cleaner. pmd allocation is essential for the shared case because 3717 * pud has to be populated inside the same i_mmap_rwsem section - otherwise 3718 * racing tasks could either miss the sharing (see huge_pte_offset) or select a 3719 * bad pmd for sharing. 3720 */ 3721pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 3722{ 3723 struct vm_area_struct *vma = find_vma(mm, addr); 3724 struct address_space *mapping = vma->vm_file->f_mapping; 3725 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 3726 vma->vm_pgoff; 3727 struct vm_area_struct *svma; 3728 unsigned long saddr; 3729 pte_t *spte = NULL; 3730 pte_t *pte; 3731 spinlock_t *ptl; 3732 3733 if (!vma_shareable(vma, addr)) 3734 return (pte_t *)pmd_alloc(mm, pud, addr); 3735 3736 i_mmap_lock_write(mapping); 3737 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 3738 if (svma == vma) 3739 continue; 3740 3741 saddr = page_table_shareable(svma, vma, addr, idx); 3742 if (saddr) { 3743 spte = huge_pte_offset(svma->vm_mm, saddr); 3744 if (spte) { 3745 mm_inc_nr_pmds(mm); 3746 get_page(virt_to_page(spte)); 3747 break; 3748 } 3749 } 3750 } 3751 3752 if (!spte) 3753 goto out; 3754 3755 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 3756 spin_lock(ptl); 3757 if (pud_none(*pud)) { 3758 pud_populate(mm, pud, 3759 (pmd_t *)((unsigned long)spte & PAGE_MASK)); 3760 } else { 3761 put_page(virt_to_page(spte)); 3762 mm_inc_nr_pmds(mm); 3763 } 3764 spin_unlock(ptl); 3765out: 3766 pte = (pte_t *)pmd_alloc(mm, pud, addr); 3767 i_mmap_unlock_write(mapping); 3768 return pte; 3769} 3770 3771/* 3772 * unmap huge page backed by shared pte. 3773 * 3774 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 3775 * indicated by page_count > 1, unmap is achieved by clearing pud and 3776 * decrementing the ref count. If count == 1, the pte page is not shared. 3777 * 3778 * called with page table lock held. 3779 * 3780 * returns: 1 successfully unmapped a shared pte page 3781 * 0 the underlying pte page is not shared, or it is the last user 3782 */ 3783int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 3784{ 3785 pgd_t *pgd = pgd_offset(mm, *addr); 3786 pud_t *pud = pud_offset(pgd, *addr); 3787 3788 BUG_ON(page_count(virt_to_page(ptep)) == 0); 3789 if (page_count(virt_to_page(ptep)) == 1) 3790 return 0; 3791 3792 pud_clear(pud); 3793 put_page(virt_to_page(ptep)); 3794 mm_dec_nr_pmds(mm); 3795 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 3796 return 1; 3797} 3798#define want_pmd_share() (1) 3799#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 3800pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 3801{ 3802 return NULL; 3803} 3804#define want_pmd_share() (0) 3805#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 3806 3807#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 3808pte_t *huge_pte_alloc(struct mm_struct *mm, 3809 unsigned long addr, unsigned long sz) 3810{ 3811 pgd_t *pgd; 3812 pud_t *pud; 3813 pte_t *pte = NULL; 3814 3815 pgd = pgd_offset(mm, addr); 3816 pud = pud_alloc(mm, pgd, addr); 3817 if (pud) { 3818 if (sz == PUD_SIZE) { 3819 pte = (pte_t *)pud; 3820 } else { 3821 BUG_ON(sz != PMD_SIZE); 3822 if (want_pmd_share() && pud_none(*pud)) 3823 pte = huge_pmd_share(mm, addr, pud); 3824 else 3825 pte = (pte_t *)pmd_alloc(mm, pud, addr); 3826 } 3827 } 3828 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 3829 3830 return pte; 3831} 3832 3833pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 3834{ 3835 pgd_t *pgd; 3836 pud_t *pud; 3837 pmd_t *pmd = NULL; 3838 3839 pgd = pgd_offset(mm, addr); 3840 if (pgd_present(*pgd)) { 3841 pud = pud_offset(pgd, addr); 3842 if (pud_present(*pud)) { 3843 if (pud_huge(*pud)) 3844 return (pte_t *)pud; 3845 pmd = pmd_offset(pud, addr); 3846 } 3847 } 3848 return (pte_t *) pmd; 3849} 3850 3851#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 3852 3853/* 3854 * These functions are overwritable if your architecture needs its own 3855 * behavior. 3856 */ 3857struct page * __weak 3858follow_huge_addr(struct mm_struct *mm, unsigned long address, 3859 int write) 3860{ 3861 return ERR_PTR(-EINVAL); 3862} 3863 3864struct page * __weak 3865follow_huge_pmd(struct mm_struct *mm, unsigned long address, 3866 pmd_t *pmd, int flags) 3867{ 3868 struct page *page = NULL; 3869 spinlock_t *ptl; 3870retry: 3871 ptl = pmd_lockptr(mm, pmd); 3872 spin_lock(ptl); 3873 /* 3874 * make sure that the address range covered by this pmd is not 3875 * unmapped from other threads. 3876 */ 3877 if (!pmd_huge(*pmd)) 3878 goto out; 3879 if (pmd_present(*pmd)) { 3880 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); 3881 if (flags & FOLL_GET) 3882 get_page(page); 3883 } else { 3884 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { 3885 spin_unlock(ptl); 3886 __migration_entry_wait(mm, (pte_t *)pmd, ptl); 3887 goto retry; 3888 } 3889 /* 3890 * hwpoisoned entry is treated as no_page_table in 3891 * follow_page_mask(). 3892 */ 3893 } 3894out: 3895 spin_unlock(ptl); 3896 return page; 3897} 3898 3899struct page * __weak 3900follow_huge_pud(struct mm_struct *mm, unsigned long address, 3901 pud_t *pud, int flags) 3902{ 3903 if (flags & FOLL_GET) 3904 return NULL; 3905 3906 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); 3907} 3908 3909#ifdef CONFIG_MEMORY_FAILURE 3910 3911/* 3912 * This function is called from memory failure code. 3913 * Assume the caller holds page lock of the head page. 3914 */ 3915int dequeue_hwpoisoned_huge_page(struct page *hpage) 3916{ 3917 struct hstate *h = page_hstate(hpage); 3918 int nid = page_to_nid(hpage); 3919 int ret = -EBUSY; 3920 3921 spin_lock(&hugetlb_lock); 3922 /* 3923 * Just checking !page_huge_active is not enough, because that could be 3924 * an isolated/hwpoisoned hugepage (which have >0 refcount). 3925 */ 3926 if (!page_huge_active(hpage) && !page_count(hpage)) { 3927 /* 3928 * Hwpoisoned hugepage isn't linked to activelist or freelist, 3929 * but dangling hpage->lru can trigger list-debug warnings 3930 * (this happens when we call unpoison_memory() on it), 3931 * so let it point to itself with list_del_init(). 3932 */ 3933 list_del_init(&hpage->lru); 3934 set_page_refcounted(hpage); 3935 h->free_huge_pages--; 3936 h->free_huge_pages_node[nid]--; 3937 ret = 0; 3938 } 3939 spin_unlock(&hugetlb_lock); 3940 return ret; 3941} 3942#endif 3943 3944bool isolate_huge_page(struct page *page, struct list_head *list) 3945{ 3946 bool ret = true; 3947 3948 VM_BUG_ON_PAGE(!PageHead(page), page); 3949 spin_lock(&hugetlb_lock); 3950 if (!page_huge_active(page) || !get_page_unless_zero(page)) { 3951 ret = false; 3952 goto unlock; 3953 } 3954 clear_page_huge_active(page); 3955 list_move_tail(&page->lru, list); 3956unlock: 3957 spin_unlock(&hugetlb_lock); 3958 return ret; 3959} 3960 3961void putback_active_hugepage(struct page *page) 3962{ 3963 VM_BUG_ON_PAGE(!PageHead(page), page); 3964 spin_lock(&hugetlb_lock); 3965 set_page_huge_active(page); 3966 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 3967 spin_unlock(&hugetlb_lock); 3968 put_page(page); 3969} 3970