hugetlb.c revision 30b0a105d9f7141e4cbf72ae5511832457d89788
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/cpuset.h> 17#include <linux/mutex.h> 18#include <linux/bootmem.h> 19#include <linux/sysfs.h> 20#include <linux/slab.h> 21#include <linux/rmap.h> 22#include <linux/swap.h> 23#include <linux/swapops.h> 24#include <linux/page-isolation.h> 25 26#include <asm/page.h> 27#include <asm/pgtable.h> 28#include <asm/tlb.h> 29 30#include <linux/io.h> 31#include <linux/hugetlb.h> 32#include <linux/hugetlb_cgroup.h> 33#include <linux/node.h> 34#include "internal.h" 35 36const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 37unsigned long hugepages_treat_as_movable; 38 39int hugetlb_max_hstate __read_mostly; 40unsigned int default_hstate_idx; 41struct hstate hstates[HUGE_MAX_HSTATE]; 42 43__initdata LIST_HEAD(huge_boot_pages); 44 45/* for command line parsing */ 46static struct hstate * __initdata parsed_hstate; 47static unsigned long __initdata default_hstate_max_huge_pages; 48static unsigned long __initdata default_hstate_size; 49 50/* 51 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 52 * free_huge_pages, and surplus_huge_pages. 53 */ 54DEFINE_SPINLOCK(hugetlb_lock); 55 56static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 57{ 58 bool free = (spool->count == 0) && (spool->used_hpages == 0); 59 60 spin_unlock(&spool->lock); 61 62 /* If no pages are used, and no other handles to the subpool 63 * remain, free the subpool the subpool remain */ 64 if (free) 65 kfree(spool); 66} 67 68struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 69{ 70 struct hugepage_subpool *spool; 71 72 spool = kmalloc(sizeof(*spool), GFP_KERNEL); 73 if (!spool) 74 return NULL; 75 76 spin_lock_init(&spool->lock); 77 spool->count = 1; 78 spool->max_hpages = nr_blocks; 79 spool->used_hpages = 0; 80 81 return spool; 82} 83 84void hugepage_put_subpool(struct hugepage_subpool *spool) 85{ 86 spin_lock(&spool->lock); 87 BUG_ON(!spool->count); 88 spool->count--; 89 unlock_or_release_subpool(spool); 90} 91 92static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 93 long delta) 94{ 95 int ret = 0; 96 97 if (!spool) 98 return 0; 99 100 spin_lock(&spool->lock); 101 if ((spool->used_hpages + delta) <= spool->max_hpages) { 102 spool->used_hpages += delta; 103 } else { 104 ret = -ENOMEM; 105 } 106 spin_unlock(&spool->lock); 107 108 return ret; 109} 110 111static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 112 long delta) 113{ 114 if (!spool) 115 return; 116 117 spin_lock(&spool->lock); 118 spool->used_hpages -= delta; 119 /* If hugetlbfs_put_super couldn't free spool due to 120 * an outstanding quota reference, free it now. */ 121 unlock_or_release_subpool(spool); 122} 123 124static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 125{ 126 return HUGETLBFS_SB(inode->i_sb)->spool; 127} 128 129static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 130{ 131 return subpool_inode(file_inode(vma->vm_file)); 132} 133 134/* 135 * Region tracking -- allows tracking of reservations and instantiated pages 136 * across the pages in a mapping. 137 * 138 * The region data structures are protected by a combination of the mmap_sem 139 * and the hugetlb_instantiation_mutex. To access or modify a region the caller 140 * must either hold the mmap_sem for write, or the mmap_sem for read and 141 * the hugetlb_instantiation_mutex: 142 * 143 * down_write(&mm->mmap_sem); 144 * or 145 * down_read(&mm->mmap_sem); 146 * mutex_lock(&hugetlb_instantiation_mutex); 147 */ 148struct file_region { 149 struct list_head link; 150 long from; 151 long to; 152}; 153 154static long region_add(struct list_head *head, long f, long t) 155{ 156 struct file_region *rg, *nrg, *trg; 157 158 /* Locate the region we are either in or before. */ 159 list_for_each_entry(rg, head, link) 160 if (f <= rg->to) 161 break; 162 163 /* Round our left edge to the current segment if it encloses us. */ 164 if (f > rg->from) 165 f = rg->from; 166 167 /* Check for and consume any regions we now overlap with. */ 168 nrg = rg; 169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 170 if (&rg->link == head) 171 break; 172 if (rg->from > t) 173 break; 174 175 /* If this area reaches higher then extend our area to 176 * include it completely. If this is not the first area 177 * which we intend to reuse, free it. */ 178 if (rg->to > t) 179 t = rg->to; 180 if (rg != nrg) { 181 list_del(&rg->link); 182 kfree(rg); 183 } 184 } 185 nrg->from = f; 186 nrg->to = t; 187 return 0; 188} 189 190static long region_chg(struct list_head *head, long f, long t) 191{ 192 struct file_region *rg, *nrg; 193 long chg = 0; 194 195 /* Locate the region we are before or in. */ 196 list_for_each_entry(rg, head, link) 197 if (f <= rg->to) 198 break; 199 200 /* If we are below the current region then a new region is required. 201 * Subtle, allocate a new region at the position but make it zero 202 * size such that we can guarantee to record the reservation. */ 203 if (&rg->link == head || t < rg->from) { 204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 205 if (!nrg) 206 return -ENOMEM; 207 nrg->from = f; 208 nrg->to = f; 209 INIT_LIST_HEAD(&nrg->link); 210 list_add(&nrg->link, rg->link.prev); 211 212 return t - f; 213 } 214 215 /* Round our left edge to the current segment if it encloses us. */ 216 if (f > rg->from) 217 f = rg->from; 218 chg = t - f; 219 220 /* Check for and consume any regions we now overlap with. */ 221 list_for_each_entry(rg, rg->link.prev, link) { 222 if (&rg->link == head) 223 break; 224 if (rg->from > t) 225 return chg; 226 227 /* We overlap with this area, if it extends further than 228 * us then we must extend ourselves. Account for its 229 * existing reservation. */ 230 if (rg->to > t) { 231 chg += rg->to - t; 232 t = rg->to; 233 } 234 chg -= rg->to - rg->from; 235 } 236 return chg; 237} 238 239static long region_truncate(struct list_head *head, long end) 240{ 241 struct file_region *rg, *trg; 242 long chg = 0; 243 244 /* Locate the region we are either in or before. */ 245 list_for_each_entry(rg, head, link) 246 if (end <= rg->to) 247 break; 248 if (&rg->link == head) 249 return 0; 250 251 /* If we are in the middle of a region then adjust it. */ 252 if (end > rg->from) { 253 chg = rg->to - end; 254 rg->to = end; 255 rg = list_entry(rg->link.next, typeof(*rg), link); 256 } 257 258 /* Drop any remaining regions. */ 259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 260 if (&rg->link == head) 261 break; 262 chg += rg->to - rg->from; 263 list_del(&rg->link); 264 kfree(rg); 265 } 266 return chg; 267} 268 269static long region_count(struct list_head *head, long f, long t) 270{ 271 struct file_region *rg; 272 long chg = 0; 273 274 /* Locate each segment we overlap with, and count that overlap. */ 275 list_for_each_entry(rg, head, link) { 276 long seg_from; 277 long seg_to; 278 279 if (rg->to <= f) 280 continue; 281 if (rg->from >= t) 282 break; 283 284 seg_from = max(rg->from, f); 285 seg_to = min(rg->to, t); 286 287 chg += seg_to - seg_from; 288 } 289 290 return chg; 291} 292 293/* 294 * Convert the address within this vma to the page offset within 295 * the mapping, in pagecache page units; huge pages here. 296 */ 297static pgoff_t vma_hugecache_offset(struct hstate *h, 298 struct vm_area_struct *vma, unsigned long address) 299{ 300 return ((address - vma->vm_start) >> huge_page_shift(h)) + 301 (vma->vm_pgoff >> huge_page_order(h)); 302} 303 304pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 305 unsigned long address) 306{ 307 return vma_hugecache_offset(hstate_vma(vma), vma, address); 308} 309 310/* 311 * Return the size of the pages allocated when backing a VMA. In the majority 312 * cases this will be same size as used by the page table entries. 313 */ 314unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 315{ 316 struct hstate *hstate; 317 318 if (!is_vm_hugetlb_page(vma)) 319 return PAGE_SIZE; 320 321 hstate = hstate_vma(vma); 322 323 return 1UL << huge_page_shift(hstate); 324} 325EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 326 327/* 328 * Return the page size being used by the MMU to back a VMA. In the majority 329 * of cases, the page size used by the kernel matches the MMU size. On 330 * architectures where it differs, an architecture-specific version of this 331 * function is required. 332 */ 333#ifndef vma_mmu_pagesize 334unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 335{ 336 return vma_kernel_pagesize(vma); 337} 338#endif 339 340/* 341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom 342 * bits of the reservation map pointer, which are always clear due to 343 * alignment. 344 */ 345#define HPAGE_RESV_OWNER (1UL << 0) 346#define HPAGE_RESV_UNMAPPED (1UL << 1) 347#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 348 349/* 350 * These helpers are used to track how many pages are reserved for 351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 352 * is guaranteed to have their future faults succeed. 353 * 354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 355 * the reserve counters are updated with the hugetlb_lock held. It is safe 356 * to reset the VMA at fork() time as it is not in use yet and there is no 357 * chance of the global counters getting corrupted as a result of the values. 358 * 359 * The private mapping reservation is represented in a subtly different 360 * manner to a shared mapping. A shared mapping has a region map associated 361 * with the underlying file, this region map represents the backing file 362 * pages which have ever had a reservation assigned which this persists even 363 * after the page is instantiated. A private mapping has a region map 364 * associated with the original mmap which is attached to all VMAs which 365 * reference it, this region map represents those offsets which have consumed 366 * reservation ie. where pages have been instantiated. 367 */ 368static unsigned long get_vma_private_data(struct vm_area_struct *vma) 369{ 370 return (unsigned long)vma->vm_private_data; 371} 372 373static void set_vma_private_data(struct vm_area_struct *vma, 374 unsigned long value) 375{ 376 vma->vm_private_data = (void *)value; 377} 378 379struct resv_map { 380 struct kref refs; 381 struct list_head regions; 382}; 383 384static struct resv_map *resv_map_alloc(void) 385{ 386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 387 if (!resv_map) 388 return NULL; 389 390 kref_init(&resv_map->refs); 391 INIT_LIST_HEAD(&resv_map->regions); 392 393 return resv_map; 394} 395 396static void resv_map_release(struct kref *ref) 397{ 398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 399 400 /* Clear out any active regions before we release the map. */ 401 region_truncate(&resv_map->regions, 0); 402 kfree(resv_map); 403} 404 405static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 406{ 407 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 408 if (!(vma->vm_flags & VM_MAYSHARE)) 409 return (struct resv_map *)(get_vma_private_data(vma) & 410 ~HPAGE_RESV_MASK); 411 return NULL; 412} 413 414static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 415{ 416 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 418 419 set_vma_private_data(vma, (get_vma_private_data(vma) & 420 HPAGE_RESV_MASK) | (unsigned long)map); 421} 422 423static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 424{ 425 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 427 428 set_vma_private_data(vma, get_vma_private_data(vma) | flags); 429} 430 431static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 432{ 433 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 434 435 return (get_vma_private_data(vma) & flag) != 0; 436} 437 438/* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 439void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 440{ 441 VM_BUG_ON(!is_vm_hugetlb_page(vma)); 442 if (!(vma->vm_flags & VM_MAYSHARE)) 443 vma->vm_private_data = (void *)0; 444} 445 446/* Returns true if the VMA has associated reserve pages */ 447static int vma_has_reserves(struct vm_area_struct *vma, long chg) 448{ 449 if (vma->vm_flags & VM_NORESERVE) { 450 /* 451 * This address is already reserved by other process(chg == 0), 452 * so, we should decrement reserved count. Without decrementing, 453 * reserve count remains after releasing inode, because this 454 * allocated page will go into page cache and is regarded as 455 * coming from reserved pool in releasing step. Currently, we 456 * don't have any other solution to deal with this situation 457 * properly, so add work-around here. 458 */ 459 if (vma->vm_flags & VM_MAYSHARE && chg == 0) 460 return 1; 461 else 462 return 0; 463 } 464 465 /* Shared mappings always use reserves */ 466 if (vma->vm_flags & VM_MAYSHARE) 467 return 1; 468 469 /* 470 * Only the process that called mmap() has reserves for 471 * private mappings. 472 */ 473 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 474 return 1; 475 476 return 0; 477} 478 479static void enqueue_huge_page(struct hstate *h, struct page *page) 480{ 481 int nid = page_to_nid(page); 482 list_move(&page->lru, &h->hugepage_freelists[nid]); 483 h->free_huge_pages++; 484 h->free_huge_pages_node[nid]++; 485} 486 487static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 488{ 489 struct page *page; 490 491 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 492 if (!is_migrate_isolate_page(page)) 493 break; 494 /* 495 * if 'non-isolated free hugepage' not found on the list, 496 * the allocation fails. 497 */ 498 if (&h->hugepage_freelists[nid] == &page->lru) 499 return NULL; 500 list_move(&page->lru, &h->hugepage_activelist); 501 set_page_refcounted(page); 502 h->free_huge_pages--; 503 h->free_huge_pages_node[nid]--; 504 return page; 505} 506 507/* Movability of hugepages depends on migration support. */ 508static inline gfp_t htlb_alloc_mask(struct hstate *h) 509{ 510 if (hugepages_treat_as_movable || hugepage_migration_support(h)) 511 return GFP_HIGHUSER_MOVABLE; 512 else 513 return GFP_HIGHUSER; 514} 515 516static struct page *dequeue_huge_page_vma(struct hstate *h, 517 struct vm_area_struct *vma, 518 unsigned long address, int avoid_reserve, 519 long chg) 520{ 521 struct page *page = NULL; 522 struct mempolicy *mpol; 523 nodemask_t *nodemask; 524 struct zonelist *zonelist; 525 struct zone *zone; 526 struct zoneref *z; 527 unsigned int cpuset_mems_cookie; 528 529 /* 530 * A child process with MAP_PRIVATE mappings created by their parent 531 * have no page reserves. This check ensures that reservations are 532 * not "stolen". The child may still get SIGKILLed 533 */ 534 if (!vma_has_reserves(vma, chg) && 535 h->free_huge_pages - h->resv_huge_pages == 0) 536 goto err; 537 538 /* If reserves cannot be used, ensure enough pages are in the pool */ 539 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 540 goto err; 541 542retry_cpuset: 543 cpuset_mems_cookie = get_mems_allowed(); 544 zonelist = huge_zonelist(vma, address, 545 htlb_alloc_mask(h), &mpol, &nodemask); 546 547 for_each_zone_zonelist_nodemask(zone, z, zonelist, 548 MAX_NR_ZONES - 1, nodemask) { 549 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) { 550 page = dequeue_huge_page_node(h, zone_to_nid(zone)); 551 if (page) { 552 if (avoid_reserve) 553 break; 554 if (!vma_has_reserves(vma, chg)) 555 break; 556 557 SetPagePrivate(page); 558 h->resv_huge_pages--; 559 break; 560 } 561 } 562 } 563 564 mpol_cond_put(mpol); 565 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) 566 goto retry_cpuset; 567 return page; 568 569err: 570 return NULL; 571} 572 573static void update_and_free_page(struct hstate *h, struct page *page) 574{ 575 int i; 576 577 VM_BUG_ON(h->order >= MAX_ORDER); 578 579 h->nr_huge_pages--; 580 h->nr_huge_pages_node[page_to_nid(page)]--; 581 for (i = 0; i < pages_per_huge_page(h); i++) { 582 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 583 1 << PG_referenced | 1 << PG_dirty | 584 1 << PG_active | 1 << PG_reserved | 585 1 << PG_private | 1 << PG_writeback); 586 } 587 VM_BUG_ON(hugetlb_cgroup_from_page(page)); 588 set_compound_page_dtor(page, NULL); 589 set_page_refcounted(page); 590 arch_release_hugepage(page); 591 __free_pages(page, huge_page_order(h)); 592} 593 594struct hstate *size_to_hstate(unsigned long size) 595{ 596 struct hstate *h; 597 598 for_each_hstate(h) { 599 if (huge_page_size(h) == size) 600 return h; 601 } 602 return NULL; 603} 604 605static void free_huge_page(struct page *page) 606{ 607 /* 608 * Can't pass hstate in here because it is called from the 609 * compound page destructor. 610 */ 611 struct hstate *h = page_hstate(page); 612 int nid = page_to_nid(page); 613 struct hugepage_subpool *spool = 614 (struct hugepage_subpool *)page_private(page); 615 bool restore_reserve; 616 617 set_page_private(page, 0); 618 page->mapping = NULL; 619 BUG_ON(page_count(page)); 620 BUG_ON(page_mapcount(page)); 621 restore_reserve = PagePrivate(page); 622 ClearPagePrivate(page); 623 624 spin_lock(&hugetlb_lock); 625 hugetlb_cgroup_uncharge_page(hstate_index(h), 626 pages_per_huge_page(h), page); 627 if (restore_reserve) 628 h->resv_huge_pages++; 629 630 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 631 /* remove the page from active list */ 632 list_del(&page->lru); 633 update_and_free_page(h, page); 634 h->surplus_huge_pages--; 635 h->surplus_huge_pages_node[nid]--; 636 } else { 637 arch_clear_hugepage_flags(page); 638 enqueue_huge_page(h, page); 639 } 640 spin_unlock(&hugetlb_lock); 641 hugepage_subpool_put_pages(spool, 1); 642} 643 644static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 645{ 646 INIT_LIST_HEAD(&page->lru); 647 set_compound_page_dtor(page, free_huge_page); 648 spin_lock(&hugetlb_lock); 649 set_hugetlb_cgroup(page, NULL); 650 h->nr_huge_pages++; 651 h->nr_huge_pages_node[nid]++; 652 spin_unlock(&hugetlb_lock); 653 put_page(page); /* free it into the hugepage allocator */ 654} 655 656static void prep_compound_gigantic_page(struct page *page, unsigned long order) 657{ 658 int i; 659 int nr_pages = 1 << order; 660 struct page *p = page + 1; 661 662 /* we rely on prep_new_huge_page to set the destructor */ 663 set_compound_order(page, order); 664 __SetPageHead(page); 665 __ClearPageReserved(page); 666 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 667 __SetPageTail(p); 668 /* 669 * For gigantic hugepages allocated through bootmem at 670 * boot, it's safer to be consistent with the not-gigantic 671 * hugepages and clear the PG_reserved bit from all tail pages 672 * too. Otherwse drivers using get_user_pages() to access tail 673 * pages may get the reference counting wrong if they see 674 * PG_reserved set on a tail page (despite the head page not 675 * having PG_reserved set). Enforcing this consistency between 676 * head and tail pages allows drivers to optimize away a check 677 * on the head page when they need know if put_page() is needed 678 * after get_user_pages(). 679 */ 680 __ClearPageReserved(p); 681 set_page_count(p, 0); 682 p->first_page = page; 683 } 684} 685 686/* 687 * PageHuge() only returns true for hugetlbfs pages, but not for normal or 688 * transparent huge pages. See the PageTransHuge() documentation for more 689 * details. 690 */ 691int PageHuge(struct page *page) 692{ 693 compound_page_dtor *dtor; 694 695 if (!PageCompound(page)) 696 return 0; 697 698 page = compound_head(page); 699 dtor = get_compound_page_dtor(page); 700 701 return dtor == free_huge_page; 702} 703EXPORT_SYMBOL_GPL(PageHuge); 704 705pgoff_t __basepage_index(struct page *page) 706{ 707 struct page *page_head = compound_head(page); 708 pgoff_t index = page_index(page_head); 709 unsigned long compound_idx; 710 711 if (!PageHuge(page_head)) 712 return page_index(page); 713 714 if (compound_order(page_head) >= MAX_ORDER) 715 compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 716 else 717 compound_idx = page - page_head; 718 719 return (index << compound_order(page_head)) + compound_idx; 720} 721 722static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 723{ 724 struct page *page; 725 726 if (h->order >= MAX_ORDER) 727 return NULL; 728 729 page = alloc_pages_exact_node(nid, 730 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 731 __GFP_REPEAT|__GFP_NOWARN, 732 huge_page_order(h)); 733 if (page) { 734 if (arch_prepare_hugepage(page)) { 735 __free_pages(page, huge_page_order(h)); 736 return NULL; 737 } 738 prep_new_huge_page(h, page, nid); 739 } 740 741 return page; 742} 743 744/* 745 * common helper functions for hstate_next_node_to_{alloc|free}. 746 * We may have allocated or freed a huge page based on a different 747 * nodes_allowed previously, so h->next_node_to_{alloc|free} might 748 * be outside of *nodes_allowed. Ensure that we use an allowed 749 * node for alloc or free. 750 */ 751static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 752{ 753 nid = next_node(nid, *nodes_allowed); 754 if (nid == MAX_NUMNODES) 755 nid = first_node(*nodes_allowed); 756 VM_BUG_ON(nid >= MAX_NUMNODES); 757 758 return nid; 759} 760 761static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 762{ 763 if (!node_isset(nid, *nodes_allowed)) 764 nid = next_node_allowed(nid, nodes_allowed); 765 return nid; 766} 767 768/* 769 * returns the previously saved node ["this node"] from which to 770 * allocate a persistent huge page for the pool and advance the 771 * next node from which to allocate, handling wrap at end of node 772 * mask. 773 */ 774static int hstate_next_node_to_alloc(struct hstate *h, 775 nodemask_t *nodes_allowed) 776{ 777 int nid; 778 779 VM_BUG_ON(!nodes_allowed); 780 781 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 782 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 783 784 return nid; 785} 786 787/* 788 * helper for free_pool_huge_page() - return the previously saved 789 * node ["this node"] from which to free a huge page. Advance the 790 * next node id whether or not we find a free huge page to free so 791 * that the next attempt to free addresses the next node. 792 */ 793static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 794{ 795 int nid; 796 797 VM_BUG_ON(!nodes_allowed); 798 799 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 800 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 801 802 return nid; 803} 804 805#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 806 for (nr_nodes = nodes_weight(*mask); \ 807 nr_nodes > 0 && \ 808 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 809 nr_nodes--) 810 811#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 812 for (nr_nodes = nodes_weight(*mask); \ 813 nr_nodes > 0 && \ 814 ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 815 nr_nodes--) 816 817static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 818{ 819 struct page *page; 820 int nr_nodes, node; 821 int ret = 0; 822 823 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 824 page = alloc_fresh_huge_page_node(h, node); 825 if (page) { 826 ret = 1; 827 break; 828 } 829 } 830 831 if (ret) 832 count_vm_event(HTLB_BUDDY_PGALLOC); 833 else 834 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 835 836 return ret; 837} 838 839/* 840 * Free huge page from pool from next node to free. 841 * Attempt to keep persistent huge pages more or less 842 * balanced over allowed nodes. 843 * Called with hugetlb_lock locked. 844 */ 845static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 846 bool acct_surplus) 847{ 848 int nr_nodes, node; 849 int ret = 0; 850 851 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 852 /* 853 * If we're returning unused surplus pages, only examine 854 * nodes with surplus pages. 855 */ 856 if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 857 !list_empty(&h->hugepage_freelists[node])) { 858 struct page *page = 859 list_entry(h->hugepage_freelists[node].next, 860 struct page, lru); 861 list_del(&page->lru); 862 h->free_huge_pages--; 863 h->free_huge_pages_node[node]--; 864 if (acct_surplus) { 865 h->surplus_huge_pages--; 866 h->surplus_huge_pages_node[node]--; 867 } 868 update_and_free_page(h, page); 869 ret = 1; 870 break; 871 } 872 } 873 874 return ret; 875} 876 877/* 878 * Dissolve a given free hugepage into free buddy pages. This function does 879 * nothing for in-use (including surplus) hugepages. 880 */ 881static void dissolve_free_huge_page(struct page *page) 882{ 883 spin_lock(&hugetlb_lock); 884 if (PageHuge(page) && !page_count(page)) { 885 struct hstate *h = page_hstate(page); 886 int nid = page_to_nid(page); 887 list_del(&page->lru); 888 h->free_huge_pages--; 889 h->free_huge_pages_node[nid]--; 890 update_and_free_page(h, page); 891 } 892 spin_unlock(&hugetlb_lock); 893} 894 895/* 896 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 897 * make specified memory blocks removable from the system. 898 * Note that start_pfn should aligned with (minimum) hugepage size. 899 */ 900void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 901{ 902 unsigned int order = 8 * sizeof(void *); 903 unsigned long pfn; 904 struct hstate *h; 905 906 /* Set scan step to minimum hugepage size */ 907 for_each_hstate(h) 908 if (order > huge_page_order(h)) 909 order = huge_page_order(h); 910 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); 911 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) 912 dissolve_free_huge_page(pfn_to_page(pfn)); 913} 914 915static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 916{ 917 struct page *page; 918 unsigned int r_nid; 919 920 if (h->order >= MAX_ORDER) 921 return NULL; 922 923 /* 924 * Assume we will successfully allocate the surplus page to 925 * prevent racing processes from causing the surplus to exceed 926 * overcommit 927 * 928 * This however introduces a different race, where a process B 929 * tries to grow the static hugepage pool while alloc_pages() is 930 * called by process A. B will only examine the per-node 931 * counters in determining if surplus huge pages can be 932 * converted to normal huge pages in adjust_pool_surplus(). A 933 * won't be able to increment the per-node counter, until the 934 * lock is dropped by B, but B doesn't drop hugetlb_lock until 935 * no more huge pages can be converted from surplus to normal 936 * state (and doesn't try to convert again). Thus, we have a 937 * case where a surplus huge page exists, the pool is grown, and 938 * the surplus huge page still exists after, even though it 939 * should just have been converted to a normal huge page. This 940 * does not leak memory, though, as the hugepage will be freed 941 * once it is out of use. It also does not allow the counters to 942 * go out of whack in adjust_pool_surplus() as we don't modify 943 * the node values until we've gotten the hugepage and only the 944 * per-node value is checked there. 945 */ 946 spin_lock(&hugetlb_lock); 947 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 948 spin_unlock(&hugetlb_lock); 949 return NULL; 950 } else { 951 h->nr_huge_pages++; 952 h->surplus_huge_pages++; 953 } 954 spin_unlock(&hugetlb_lock); 955 956 if (nid == NUMA_NO_NODE) 957 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| 958 __GFP_REPEAT|__GFP_NOWARN, 959 huge_page_order(h)); 960 else 961 page = alloc_pages_exact_node(nid, 962 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 963 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 964 965 if (page && arch_prepare_hugepage(page)) { 966 __free_pages(page, huge_page_order(h)); 967 page = NULL; 968 } 969 970 spin_lock(&hugetlb_lock); 971 if (page) { 972 INIT_LIST_HEAD(&page->lru); 973 r_nid = page_to_nid(page); 974 set_compound_page_dtor(page, free_huge_page); 975 set_hugetlb_cgroup(page, NULL); 976 /* 977 * We incremented the global counters already 978 */ 979 h->nr_huge_pages_node[r_nid]++; 980 h->surplus_huge_pages_node[r_nid]++; 981 __count_vm_event(HTLB_BUDDY_PGALLOC); 982 } else { 983 h->nr_huge_pages--; 984 h->surplus_huge_pages--; 985 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 986 } 987 spin_unlock(&hugetlb_lock); 988 989 return page; 990} 991 992/* 993 * This allocation function is useful in the context where vma is irrelevant. 994 * E.g. soft-offlining uses this function because it only cares physical 995 * address of error page. 996 */ 997struct page *alloc_huge_page_node(struct hstate *h, int nid) 998{ 999 struct page *page = NULL; 1000 1001 spin_lock(&hugetlb_lock); 1002 if (h->free_huge_pages - h->resv_huge_pages > 0) 1003 page = dequeue_huge_page_node(h, nid); 1004 spin_unlock(&hugetlb_lock); 1005 1006 if (!page) 1007 page = alloc_buddy_huge_page(h, nid); 1008 1009 return page; 1010} 1011 1012/* 1013 * Increase the hugetlb pool such that it can accommodate a reservation 1014 * of size 'delta'. 1015 */ 1016static int gather_surplus_pages(struct hstate *h, int delta) 1017{ 1018 struct list_head surplus_list; 1019 struct page *page, *tmp; 1020 int ret, i; 1021 int needed, allocated; 1022 bool alloc_ok = true; 1023 1024 needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1025 if (needed <= 0) { 1026 h->resv_huge_pages += delta; 1027 return 0; 1028 } 1029 1030 allocated = 0; 1031 INIT_LIST_HEAD(&surplus_list); 1032 1033 ret = -ENOMEM; 1034retry: 1035 spin_unlock(&hugetlb_lock); 1036 for (i = 0; i < needed; i++) { 1037 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1038 if (!page) { 1039 alloc_ok = false; 1040 break; 1041 } 1042 list_add(&page->lru, &surplus_list); 1043 } 1044 allocated += i; 1045 1046 /* 1047 * After retaking hugetlb_lock, we need to recalculate 'needed' 1048 * because either resv_huge_pages or free_huge_pages may have changed. 1049 */ 1050 spin_lock(&hugetlb_lock); 1051 needed = (h->resv_huge_pages + delta) - 1052 (h->free_huge_pages + allocated); 1053 if (needed > 0) { 1054 if (alloc_ok) 1055 goto retry; 1056 /* 1057 * We were not able to allocate enough pages to 1058 * satisfy the entire reservation so we free what 1059 * we've allocated so far. 1060 */ 1061 goto free; 1062 } 1063 /* 1064 * The surplus_list now contains _at_least_ the number of extra pages 1065 * needed to accommodate the reservation. Add the appropriate number 1066 * of pages to the hugetlb pool and free the extras back to the buddy 1067 * allocator. Commit the entire reservation here to prevent another 1068 * process from stealing the pages as they are added to the pool but 1069 * before they are reserved. 1070 */ 1071 needed += allocated; 1072 h->resv_huge_pages += delta; 1073 ret = 0; 1074 1075 /* Free the needed pages to the hugetlb pool */ 1076 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1077 if ((--needed) < 0) 1078 break; 1079 /* 1080 * This page is now managed by the hugetlb allocator and has 1081 * no users -- drop the buddy allocator's reference. 1082 */ 1083 put_page_testzero(page); 1084 VM_BUG_ON(page_count(page)); 1085 enqueue_huge_page(h, page); 1086 } 1087free: 1088 spin_unlock(&hugetlb_lock); 1089 1090 /* Free unnecessary surplus pages to the buddy allocator */ 1091 list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1092 put_page(page); 1093 spin_lock(&hugetlb_lock); 1094 1095 return ret; 1096} 1097 1098/* 1099 * When releasing a hugetlb pool reservation, any surplus pages that were 1100 * allocated to satisfy the reservation must be explicitly freed if they were 1101 * never used. 1102 * Called with hugetlb_lock held. 1103 */ 1104static void return_unused_surplus_pages(struct hstate *h, 1105 unsigned long unused_resv_pages) 1106{ 1107 unsigned long nr_pages; 1108 1109 /* Uncommit the reservation */ 1110 h->resv_huge_pages -= unused_resv_pages; 1111 1112 /* Cannot return gigantic pages currently */ 1113 if (h->order >= MAX_ORDER) 1114 return; 1115 1116 nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1117 1118 /* 1119 * We want to release as many surplus pages as possible, spread 1120 * evenly across all nodes with memory. Iterate across these nodes 1121 * until we can no longer free unreserved surplus pages. This occurs 1122 * when the nodes with surplus pages have no free pages. 1123 * free_pool_huge_page() will balance the the freed pages across the 1124 * on-line nodes with memory and will handle the hstate accounting. 1125 */ 1126 while (nr_pages--) { 1127 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1128 break; 1129 } 1130} 1131 1132/* 1133 * Determine if the huge page at addr within the vma has an associated 1134 * reservation. Where it does not we will need to logically increase 1135 * reservation and actually increase subpool usage before an allocation 1136 * can occur. Where any new reservation would be required the 1137 * reservation change is prepared, but not committed. Once the page 1138 * has been allocated from the subpool and instantiated the change should 1139 * be committed via vma_commit_reservation. No action is required on 1140 * failure. 1141 */ 1142static long vma_needs_reservation(struct hstate *h, 1143 struct vm_area_struct *vma, unsigned long addr) 1144{ 1145 struct address_space *mapping = vma->vm_file->f_mapping; 1146 struct inode *inode = mapping->host; 1147 1148 if (vma->vm_flags & VM_MAYSHARE) { 1149 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1150 return region_chg(&inode->i_mapping->private_list, 1151 idx, idx + 1); 1152 1153 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1154 return 1; 1155 1156 } else { 1157 long err; 1158 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1159 struct resv_map *resv = vma_resv_map(vma); 1160 1161 err = region_chg(&resv->regions, idx, idx + 1); 1162 if (err < 0) 1163 return err; 1164 return 0; 1165 } 1166} 1167static void vma_commit_reservation(struct hstate *h, 1168 struct vm_area_struct *vma, unsigned long addr) 1169{ 1170 struct address_space *mapping = vma->vm_file->f_mapping; 1171 struct inode *inode = mapping->host; 1172 1173 if (vma->vm_flags & VM_MAYSHARE) { 1174 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1175 region_add(&inode->i_mapping->private_list, idx, idx + 1); 1176 1177 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1178 pgoff_t idx = vma_hugecache_offset(h, vma, addr); 1179 struct resv_map *resv = vma_resv_map(vma); 1180 1181 /* Mark this page used in the map. */ 1182 region_add(&resv->regions, idx, idx + 1); 1183 } 1184} 1185 1186static struct page *alloc_huge_page(struct vm_area_struct *vma, 1187 unsigned long addr, int avoid_reserve) 1188{ 1189 struct hugepage_subpool *spool = subpool_vma(vma); 1190 struct hstate *h = hstate_vma(vma); 1191 struct page *page; 1192 long chg; 1193 int ret, idx; 1194 struct hugetlb_cgroup *h_cg; 1195 1196 idx = hstate_index(h); 1197 /* 1198 * Processes that did not create the mapping will have no 1199 * reserves and will not have accounted against subpool 1200 * limit. Check that the subpool limit can be made before 1201 * satisfying the allocation MAP_NORESERVE mappings may also 1202 * need pages and subpool limit allocated allocated if no reserve 1203 * mapping overlaps. 1204 */ 1205 chg = vma_needs_reservation(h, vma, addr); 1206 if (chg < 0) 1207 return ERR_PTR(-ENOMEM); 1208 if (chg || avoid_reserve) 1209 if (hugepage_subpool_get_pages(spool, 1)) 1210 return ERR_PTR(-ENOSPC); 1211 1212 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 1213 if (ret) { 1214 if (chg || avoid_reserve) 1215 hugepage_subpool_put_pages(spool, 1); 1216 return ERR_PTR(-ENOSPC); 1217 } 1218 spin_lock(&hugetlb_lock); 1219 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 1220 if (!page) { 1221 spin_unlock(&hugetlb_lock); 1222 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1223 if (!page) { 1224 hugetlb_cgroup_uncharge_cgroup(idx, 1225 pages_per_huge_page(h), 1226 h_cg); 1227 if (chg || avoid_reserve) 1228 hugepage_subpool_put_pages(spool, 1); 1229 return ERR_PTR(-ENOSPC); 1230 } 1231 spin_lock(&hugetlb_lock); 1232 list_move(&page->lru, &h->hugepage_activelist); 1233 /* Fall through */ 1234 } 1235 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 1236 spin_unlock(&hugetlb_lock); 1237 1238 set_page_private(page, (unsigned long)spool); 1239 1240 vma_commit_reservation(h, vma, addr); 1241 return page; 1242} 1243 1244/* 1245 * alloc_huge_page()'s wrapper which simply returns the page if allocation 1246 * succeeds, otherwise NULL. This function is called from new_vma_page(), 1247 * where no ERR_VALUE is expected to be returned. 1248 */ 1249struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 1250 unsigned long addr, int avoid_reserve) 1251{ 1252 struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 1253 if (IS_ERR(page)) 1254 page = NULL; 1255 return page; 1256} 1257 1258int __weak alloc_bootmem_huge_page(struct hstate *h) 1259{ 1260 struct huge_bootmem_page *m; 1261 int nr_nodes, node; 1262 1263 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1264 void *addr; 1265 1266 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node), 1267 huge_page_size(h), huge_page_size(h), 0); 1268 1269 if (addr) { 1270 /* 1271 * Use the beginning of the huge page to store the 1272 * huge_bootmem_page struct (until gather_bootmem 1273 * puts them into the mem_map). 1274 */ 1275 m = addr; 1276 goto found; 1277 } 1278 } 1279 return 0; 1280 1281found: 1282 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1283 /* Put them into a private list first because mem_map is not up yet */ 1284 list_add(&m->list, &huge_boot_pages); 1285 m->hstate = h; 1286 return 1; 1287} 1288 1289static void prep_compound_huge_page(struct page *page, int order) 1290{ 1291 if (unlikely(order > (MAX_ORDER - 1))) 1292 prep_compound_gigantic_page(page, order); 1293 else 1294 prep_compound_page(page, order); 1295} 1296 1297/* Put bootmem huge pages into the standard lists after mem_map is up */ 1298static void __init gather_bootmem_prealloc(void) 1299{ 1300 struct huge_bootmem_page *m; 1301 1302 list_for_each_entry(m, &huge_boot_pages, list) { 1303 struct hstate *h = m->hstate; 1304 struct page *page; 1305 1306#ifdef CONFIG_HIGHMEM 1307 page = pfn_to_page(m->phys >> PAGE_SHIFT); 1308 free_bootmem_late((unsigned long)m, 1309 sizeof(struct huge_bootmem_page)); 1310#else 1311 page = virt_to_page(m); 1312#endif 1313 WARN_ON(page_count(page) != 1); 1314 prep_compound_huge_page(page, h->order); 1315 WARN_ON(PageReserved(page)); 1316 prep_new_huge_page(h, page, page_to_nid(page)); 1317 /* 1318 * If we had gigantic hugepages allocated at boot time, we need 1319 * to restore the 'stolen' pages to totalram_pages in order to 1320 * fix confusing memory reports from free(1) and another 1321 * side-effects, like CommitLimit going negative. 1322 */ 1323 if (h->order > (MAX_ORDER - 1)) 1324 adjust_managed_page_count(page, 1 << h->order); 1325 } 1326} 1327 1328static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 1329{ 1330 unsigned long i; 1331 1332 for (i = 0; i < h->max_huge_pages; ++i) { 1333 if (h->order >= MAX_ORDER) { 1334 if (!alloc_bootmem_huge_page(h)) 1335 break; 1336 } else if (!alloc_fresh_huge_page(h, 1337 &node_states[N_MEMORY])) 1338 break; 1339 } 1340 h->max_huge_pages = i; 1341} 1342 1343static void __init hugetlb_init_hstates(void) 1344{ 1345 struct hstate *h; 1346 1347 for_each_hstate(h) { 1348 /* oversize hugepages were init'ed in early boot */ 1349 if (h->order < MAX_ORDER) 1350 hugetlb_hstate_alloc_pages(h); 1351 } 1352} 1353 1354static char * __init memfmt(char *buf, unsigned long n) 1355{ 1356 if (n >= (1UL << 30)) 1357 sprintf(buf, "%lu GB", n >> 30); 1358 else if (n >= (1UL << 20)) 1359 sprintf(buf, "%lu MB", n >> 20); 1360 else 1361 sprintf(buf, "%lu KB", n >> 10); 1362 return buf; 1363} 1364 1365static void __init report_hugepages(void) 1366{ 1367 struct hstate *h; 1368 1369 for_each_hstate(h) { 1370 char buf[32]; 1371 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 1372 memfmt(buf, huge_page_size(h)), 1373 h->free_huge_pages); 1374 } 1375} 1376 1377#ifdef CONFIG_HIGHMEM 1378static void try_to_free_low(struct hstate *h, unsigned long count, 1379 nodemask_t *nodes_allowed) 1380{ 1381 int i; 1382 1383 if (h->order >= MAX_ORDER) 1384 return; 1385 1386 for_each_node_mask(i, *nodes_allowed) { 1387 struct page *page, *next; 1388 struct list_head *freel = &h->hugepage_freelists[i]; 1389 list_for_each_entry_safe(page, next, freel, lru) { 1390 if (count >= h->nr_huge_pages) 1391 return; 1392 if (PageHighMem(page)) 1393 continue; 1394 list_del(&page->lru); 1395 update_and_free_page(h, page); 1396 h->free_huge_pages--; 1397 h->free_huge_pages_node[page_to_nid(page)]--; 1398 } 1399 } 1400} 1401#else 1402static inline void try_to_free_low(struct hstate *h, unsigned long count, 1403 nodemask_t *nodes_allowed) 1404{ 1405} 1406#endif 1407 1408/* 1409 * Increment or decrement surplus_huge_pages. Keep node-specific counters 1410 * balanced by operating on them in a round-robin fashion. 1411 * Returns 1 if an adjustment was made. 1412 */ 1413static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 1414 int delta) 1415{ 1416 int nr_nodes, node; 1417 1418 VM_BUG_ON(delta != -1 && delta != 1); 1419 1420 if (delta < 0) { 1421 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1422 if (h->surplus_huge_pages_node[node]) 1423 goto found; 1424 } 1425 } else { 1426 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1427 if (h->surplus_huge_pages_node[node] < 1428 h->nr_huge_pages_node[node]) 1429 goto found; 1430 } 1431 } 1432 return 0; 1433 1434found: 1435 h->surplus_huge_pages += delta; 1436 h->surplus_huge_pages_node[node] += delta; 1437 return 1; 1438} 1439 1440#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1441static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 1442 nodemask_t *nodes_allowed) 1443{ 1444 unsigned long min_count, ret; 1445 1446 if (h->order >= MAX_ORDER) 1447 return h->max_huge_pages; 1448 1449 /* 1450 * Increase the pool size 1451 * First take pages out of surplus state. Then make up the 1452 * remaining difference by allocating fresh huge pages. 1453 * 1454 * We might race with alloc_buddy_huge_page() here and be unable 1455 * to convert a surplus huge page to a normal huge page. That is 1456 * not critical, though, it just means the overall size of the 1457 * pool might be one hugepage larger than it needs to be, but 1458 * within all the constraints specified by the sysctls. 1459 */ 1460 spin_lock(&hugetlb_lock); 1461 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1462 if (!adjust_pool_surplus(h, nodes_allowed, -1)) 1463 break; 1464 } 1465 1466 while (count > persistent_huge_pages(h)) { 1467 /* 1468 * If this allocation races such that we no longer need the 1469 * page, free_huge_page will handle it by freeing the page 1470 * and reducing the surplus. 1471 */ 1472 spin_unlock(&hugetlb_lock); 1473 ret = alloc_fresh_huge_page(h, nodes_allowed); 1474 spin_lock(&hugetlb_lock); 1475 if (!ret) 1476 goto out; 1477 1478 /* Bail for signals. Probably ctrl-c from user */ 1479 if (signal_pending(current)) 1480 goto out; 1481 } 1482 1483 /* 1484 * Decrease the pool size 1485 * First return free pages to the buddy allocator (being careful 1486 * to keep enough around to satisfy reservations). Then place 1487 * pages into surplus state as needed so the pool will shrink 1488 * to the desired size as pages become free. 1489 * 1490 * By placing pages into the surplus state independent of the 1491 * overcommit value, we are allowing the surplus pool size to 1492 * exceed overcommit. There are few sane options here. Since 1493 * alloc_buddy_huge_page() is checking the global counter, 1494 * though, we'll note that we're not allowed to exceed surplus 1495 * and won't grow the pool anywhere else. Not until one of the 1496 * sysctls are changed, or the surplus pages go out of use. 1497 */ 1498 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 1499 min_count = max(count, min_count); 1500 try_to_free_low(h, min_count, nodes_allowed); 1501 while (min_count < persistent_huge_pages(h)) { 1502 if (!free_pool_huge_page(h, nodes_allowed, 0)) 1503 break; 1504 } 1505 while (count < persistent_huge_pages(h)) { 1506 if (!adjust_pool_surplus(h, nodes_allowed, 1)) 1507 break; 1508 } 1509out: 1510 ret = persistent_huge_pages(h); 1511 spin_unlock(&hugetlb_lock); 1512 return ret; 1513} 1514 1515#define HSTATE_ATTR_RO(_name) \ 1516 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1517 1518#define HSTATE_ATTR(_name) \ 1519 static struct kobj_attribute _name##_attr = \ 1520 __ATTR(_name, 0644, _name##_show, _name##_store) 1521 1522static struct kobject *hugepages_kobj; 1523static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1524 1525static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 1526 1527static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1528{ 1529 int i; 1530 1531 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1532 if (hstate_kobjs[i] == kobj) { 1533 if (nidp) 1534 *nidp = NUMA_NO_NODE; 1535 return &hstates[i]; 1536 } 1537 1538 return kobj_to_node_hstate(kobj, nidp); 1539} 1540 1541static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1542 struct kobj_attribute *attr, char *buf) 1543{ 1544 struct hstate *h; 1545 unsigned long nr_huge_pages; 1546 int nid; 1547 1548 h = kobj_to_hstate(kobj, &nid); 1549 if (nid == NUMA_NO_NODE) 1550 nr_huge_pages = h->nr_huge_pages; 1551 else 1552 nr_huge_pages = h->nr_huge_pages_node[nid]; 1553 1554 return sprintf(buf, "%lu\n", nr_huge_pages); 1555} 1556 1557static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 1558 struct kobject *kobj, struct kobj_attribute *attr, 1559 const char *buf, size_t len) 1560{ 1561 int err; 1562 int nid; 1563 unsigned long count; 1564 struct hstate *h; 1565 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1566 1567 err = kstrtoul(buf, 10, &count); 1568 if (err) 1569 goto out; 1570 1571 h = kobj_to_hstate(kobj, &nid); 1572 if (h->order >= MAX_ORDER) { 1573 err = -EINVAL; 1574 goto out; 1575 } 1576 1577 if (nid == NUMA_NO_NODE) { 1578 /* 1579 * global hstate attribute 1580 */ 1581 if (!(obey_mempolicy && 1582 init_nodemask_of_mempolicy(nodes_allowed))) { 1583 NODEMASK_FREE(nodes_allowed); 1584 nodes_allowed = &node_states[N_MEMORY]; 1585 } 1586 } else if (nodes_allowed) { 1587 /* 1588 * per node hstate attribute: adjust count to global, 1589 * but restrict alloc/free to the specified node. 1590 */ 1591 count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 1592 init_nodemask_of_node(nodes_allowed, nid); 1593 } else 1594 nodes_allowed = &node_states[N_MEMORY]; 1595 1596 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1597 1598 if (nodes_allowed != &node_states[N_MEMORY]) 1599 NODEMASK_FREE(nodes_allowed); 1600 1601 return len; 1602out: 1603 NODEMASK_FREE(nodes_allowed); 1604 return err; 1605} 1606 1607static ssize_t nr_hugepages_show(struct kobject *kobj, 1608 struct kobj_attribute *attr, char *buf) 1609{ 1610 return nr_hugepages_show_common(kobj, attr, buf); 1611} 1612 1613static ssize_t nr_hugepages_store(struct kobject *kobj, 1614 struct kobj_attribute *attr, const char *buf, size_t len) 1615{ 1616 return nr_hugepages_store_common(false, kobj, attr, buf, len); 1617} 1618HSTATE_ATTR(nr_hugepages); 1619 1620#ifdef CONFIG_NUMA 1621 1622/* 1623 * hstate attribute for optionally mempolicy-based constraint on persistent 1624 * huge page alloc/free. 1625 */ 1626static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 1627 struct kobj_attribute *attr, char *buf) 1628{ 1629 return nr_hugepages_show_common(kobj, attr, buf); 1630} 1631 1632static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 1633 struct kobj_attribute *attr, const char *buf, size_t len) 1634{ 1635 return nr_hugepages_store_common(true, kobj, attr, buf, len); 1636} 1637HSTATE_ATTR(nr_hugepages_mempolicy); 1638#endif 1639 1640 1641static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1642 struct kobj_attribute *attr, char *buf) 1643{ 1644 struct hstate *h = kobj_to_hstate(kobj, NULL); 1645 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1646} 1647 1648static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1649 struct kobj_attribute *attr, const char *buf, size_t count) 1650{ 1651 int err; 1652 unsigned long input; 1653 struct hstate *h = kobj_to_hstate(kobj, NULL); 1654 1655 if (h->order >= MAX_ORDER) 1656 return -EINVAL; 1657 1658 err = kstrtoul(buf, 10, &input); 1659 if (err) 1660 return err; 1661 1662 spin_lock(&hugetlb_lock); 1663 h->nr_overcommit_huge_pages = input; 1664 spin_unlock(&hugetlb_lock); 1665 1666 return count; 1667} 1668HSTATE_ATTR(nr_overcommit_hugepages); 1669 1670static ssize_t free_hugepages_show(struct kobject *kobj, 1671 struct kobj_attribute *attr, char *buf) 1672{ 1673 struct hstate *h; 1674 unsigned long free_huge_pages; 1675 int nid; 1676 1677 h = kobj_to_hstate(kobj, &nid); 1678 if (nid == NUMA_NO_NODE) 1679 free_huge_pages = h->free_huge_pages; 1680 else 1681 free_huge_pages = h->free_huge_pages_node[nid]; 1682 1683 return sprintf(buf, "%lu\n", free_huge_pages); 1684} 1685HSTATE_ATTR_RO(free_hugepages); 1686 1687static ssize_t resv_hugepages_show(struct kobject *kobj, 1688 struct kobj_attribute *attr, char *buf) 1689{ 1690 struct hstate *h = kobj_to_hstate(kobj, NULL); 1691 return sprintf(buf, "%lu\n", h->resv_huge_pages); 1692} 1693HSTATE_ATTR_RO(resv_hugepages); 1694 1695static ssize_t surplus_hugepages_show(struct kobject *kobj, 1696 struct kobj_attribute *attr, char *buf) 1697{ 1698 struct hstate *h; 1699 unsigned long surplus_huge_pages; 1700 int nid; 1701 1702 h = kobj_to_hstate(kobj, &nid); 1703 if (nid == NUMA_NO_NODE) 1704 surplus_huge_pages = h->surplus_huge_pages; 1705 else 1706 surplus_huge_pages = h->surplus_huge_pages_node[nid]; 1707 1708 return sprintf(buf, "%lu\n", surplus_huge_pages); 1709} 1710HSTATE_ATTR_RO(surplus_hugepages); 1711 1712static struct attribute *hstate_attrs[] = { 1713 &nr_hugepages_attr.attr, 1714 &nr_overcommit_hugepages_attr.attr, 1715 &free_hugepages_attr.attr, 1716 &resv_hugepages_attr.attr, 1717 &surplus_hugepages_attr.attr, 1718#ifdef CONFIG_NUMA 1719 &nr_hugepages_mempolicy_attr.attr, 1720#endif 1721 NULL, 1722}; 1723 1724static struct attribute_group hstate_attr_group = { 1725 .attrs = hstate_attrs, 1726}; 1727 1728static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 1729 struct kobject **hstate_kobjs, 1730 struct attribute_group *hstate_attr_group) 1731{ 1732 int retval; 1733 int hi = hstate_index(h); 1734 1735 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 1736 if (!hstate_kobjs[hi]) 1737 return -ENOMEM; 1738 1739 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1740 if (retval) 1741 kobject_put(hstate_kobjs[hi]); 1742 1743 return retval; 1744} 1745 1746static void __init hugetlb_sysfs_init(void) 1747{ 1748 struct hstate *h; 1749 int err; 1750 1751 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1752 if (!hugepages_kobj) 1753 return; 1754 1755 for_each_hstate(h) { 1756 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 1757 hstate_kobjs, &hstate_attr_group); 1758 if (err) 1759 pr_err("Hugetlb: Unable to add hstate %s", h->name); 1760 } 1761} 1762 1763#ifdef CONFIG_NUMA 1764 1765/* 1766 * node_hstate/s - associate per node hstate attributes, via their kobjects, 1767 * with node devices in node_devices[] using a parallel array. The array 1768 * index of a node device or _hstate == node id. 1769 * This is here to avoid any static dependency of the node device driver, in 1770 * the base kernel, on the hugetlb module. 1771 */ 1772struct node_hstate { 1773 struct kobject *hugepages_kobj; 1774 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1775}; 1776struct node_hstate node_hstates[MAX_NUMNODES]; 1777 1778/* 1779 * A subset of global hstate attributes for node devices 1780 */ 1781static struct attribute *per_node_hstate_attrs[] = { 1782 &nr_hugepages_attr.attr, 1783 &free_hugepages_attr.attr, 1784 &surplus_hugepages_attr.attr, 1785 NULL, 1786}; 1787 1788static struct attribute_group per_node_hstate_attr_group = { 1789 .attrs = per_node_hstate_attrs, 1790}; 1791 1792/* 1793 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 1794 * Returns node id via non-NULL nidp. 1795 */ 1796static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 1797{ 1798 int nid; 1799 1800 for (nid = 0; nid < nr_node_ids; nid++) { 1801 struct node_hstate *nhs = &node_hstates[nid]; 1802 int i; 1803 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1804 if (nhs->hstate_kobjs[i] == kobj) { 1805 if (nidp) 1806 *nidp = nid; 1807 return &hstates[i]; 1808 } 1809 } 1810 1811 BUG(); 1812 return NULL; 1813} 1814 1815/* 1816 * Unregister hstate attributes from a single node device. 1817 * No-op if no hstate attributes attached. 1818 */ 1819static void hugetlb_unregister_node(struct node *node) 1820{ 1821 struct hstate *h; 1822 struct node_hstate *nhs = &node_hstates[node->dev.id]; 1823 1824 if (!nhs->hugepages_kobj) 1825 return; /* no hstate attributes */ 1826 1827 for_each_hstate(h) { 1828 int idx = hstate_index(h); 1829 if (nhs->hstate_kobjs[idx]) { 1830 kobject_put(nhs->hstate_kobjs[idx]); 1831 nhs->hstate_kobjs[idx] = NULL; 1832 } 1833 } 1834 1835 kobject_put(nhs->hugepages_kobj); 1836 nhs->hugepages_kobj = NULL; 1837} 1838 1839/* 1840 * hugetlb module exit: unregister hstate attributes from node devices 1841 * that have them. 1842 */ 1843static void hugetlb_unregister_all_nodes(void) 1844{ 1845 int nid; 1846 1847 /* 1848 * disable node device registrations. 1849 */ 1850 register_hugetlbfs_with_node(NULL, NULL); 1851 1852 /* 1853 * remove hstate attributes from any nodes that have them. 1854 */ 1855 for (nid = 0; nid < nr_node_ids; nid++) 1856 hugetlb_unregister_node(node_devices[nid]); 1857} 1858 1859/* 1860 * Register hstate attributes for a single node device. 1861 * No-op if attributes already registered. 1862 */ 1863static void hugetlb_register_node(struct node *node) 1864{ 1865 struct hstate *h; 1866 struct node_hstate *nhs = &node_hstates[node->dev.id]; 1867 int err; 1868 1869 if (nhs->hugepages_kobj) 1870 return; /* already allocated */ 1871 1872 nhs->hugepages_kobj = kobject_create_and_add("hugepages", 1873 &node->dev.kobj); 1874 if (!nhs->hugepages_kobj) 1875 return; 1876 1877 for_each_hstate(h) { 1878 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 1879 nhs->hstate_kobjs, 1880 &per_node_hstate_attr_group); 1881 if (err) { 1882 pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 1883 h->name, node->dev.id); 1884 hugetlb_unregister_node(node); 1885 break; 1886 } 1887 } 1888} 1889 1890/* 1891 * hugetlb init time: register hstate attributes for all registered node 1892 * devices of nodes that have memory. All on-line nodes should have 1893 * registered their associated device by this time. 1894 */ 1895static void hugetlb_register_all_nodes(void) 1896{ 1897 int nid; 1898 1899 for_each_node_state(nid, N_MEMORY) { 1900 struct node *node = node_devices[nid]; 1901 if (node->dev.id == nid) 1902 hugetlb_register_node(node); 1903 } 1904 1905 /* 1906 * Let the node device driver know we're here so it can 1907 * [un]register hstate attributes on node hotplug. 1908 */ 1909 register_hugetlbfs_with_node(hugetlb_register_node, 1910 hugetlb_unregister_node); 1911} 1912#else /* !CONFIG_NUMA */ 1913 1914static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 1915{ 1916 BUG(); 1917 if (nidp) 1918 *nidp = -1; 1919 return NULL; 1920} 1921 1922static void hugetlb_unregister_all_nodes(void) { } 1923 1924static void hugetlb_register_all_nodes(void) { } 1925 1926#endif 1927 1928static void __exit hugetlb_exit(void) 1929{ 1930 struct hstate *h; 1931 1932 hugetlb_unregister_all_nodes(); 1933 1934 for_each_hstate(h) { 1935 kobject_put(hstate_kobjs[hstate_index(h)]); 1936 } 1937 1938 kobject_put(hugepages_kobj); 1939} 1940module_exit(hugetlb_exit); 1941 1942static int __init hugetlb_init(void) 1943{ 1944 /* Some platform decide whether they support huge pages at boot 1945 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when 1946 * there is no such support 1947 */ 1948 if (HPAGE_SHIFT == 0) 1949 return 0; 1950 1951 if (!size_to_hstate(default_hstate_size)) { 1952 default_hstate_size = HPAGE_SIZE; 1953 if (!size_to_hstate(default_hstate_size)) 1954 hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1955 } 1956 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1957 if (default_hstate_max_huge_pages) 1958 default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1959 1960 hugetlb_init_hstates(); 1961 gather_bootmem_prealloc(); 1962 report_hugepages(); 1963 1964 hugetlb_sysfs_init(); 1965 hugetlb_register_all_nodes(); 1966 hugetlb_cgroup_file_init(); 1967 1968 return 0; 1969} 1970module_init(hugetlb_init); 1971 1972/* Should be called on processing a hugepagesz=... option */ 1973void __init hugetlb_add_hstate(unsigned order) 1974{ 1975 struct hstate *h; 1976 unsigned long i; 1977 1978 if (size_to_hstate(PAGE_SIZE << order)) { 1979 pr_warning("hugepagesz= specified twice, ignoring\n"); 1980 return; 1981 } 1982 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 1983 BUG_ON(order == 0); 1984 h = &hstates[hugetlb_max_hstate++]; 1985 h->order = order; 1986 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 1987 h->nr_huge_pages = 0; 1988 h->free_huge_pages = 0; 1989 for (i = 0; i < MAX_NUMNODES; ++i) 1990 INIT_LIST_HEAD(&h->hugepage_freelists[i]); 1991 INIT_LIST_HEAD(&h->hugepage_activelist); 1992 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 1993 h->next_nid_to_free = first_node(node_states[N_MEMORY]); 1994 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 1995 huge_page_size(h)/1024); 1996 1997 parsed_hstate = h; 1998} 1999 2000static int __init hugetlb_nrpages_setup(char *s) 2001{ 2002 unsigned long *mhp; 2003 static unsigned long *last_mhp; 2004 2005 /* 2006 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2007 * so this hugepages= parameter goes to the "default hstate". 2008 */ 2009 if (!hugetlb_max_hstate) 2010 mhp = &default_hstate_max_huge_pages; 2011 else 2012 mhp = &parsed_hstate->max_huge_pages; 2013 2014 if (mhp == last_mhp) { 2015 pr_warning("hugepages= specified twice without " 2016 "interleaving hugepagesz=, ignoring\n"); 2017 return 1; 2018 } 2019 2020 if (sscanf(s, "%lu", mhp) <= 0) 2021 *mhp = 0; 2022 2023 /* 2024 * Global state is always initialized later in hugetlb_init. 2025 * But we need to allocate >= MAX_ORDER hstates here early to still 2026 * use the bootmem allocator. 2027 */ 2028 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 2029 hugetlb_hstate_alloc_pages(parsed_hstate); 2030 2031 last_mhp = mhp; 2032 2033 return 1; 2034} 2035__setup("hugepages=", hugetlb_nrpages_setup); 2036 2037static int __init hugetlb_default_setup(char *s) 2038{ 2039 default_hstate_size = memparse(s, &s); 2040 return 1; 2041} 2042__setup("default_hugepagesz=", hugetlb_default_setup); 2043 2044static unsigned int cpuset_mems_nr(unsigned int *array) 2045{ 2046 int node; 2047 unsigned int nr = 0; 2048 2049 for_each_node_mask(node, cpuset_current_mems_allowed) 2050 nr += array[node]; 2051 2052 return nr; 2053} 2054 2055#ifdef CONFIG_SYSCTL 2056static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 2057 struct ctl_table *table, int write, 2058 void __user *buffer, size_t *length, loff_t *ppos) 2059{ 2060 struct hstate *h = &default_hstate; 2061 unsigned long tmp; 2062 int ret; 2063 2064 tmp = h->max_huge_pages; 2065 2066 if (write && h->order >= MAX_ORDER) 2067 return -EINVAL; 2068 2069 table->data = &tmp; 2070 table->maxlen = sizeof(unsigned long); 2071 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 2072 if (ret) 2073 goto out; 2074 2075 if (write) { 2076 NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2077 GFP_KERNEL | __GFP_NORETRY); 2078 if (!(obey_mempolicy && 2079 init_nodemask_of_mempolicy(nodes_allowed))) { 2080 NODEMASK_FREE(nodes_allowed); 2081 nodes_allowed = &node_states[N_MEMORY]; 2082 } 2083 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 2084 2085 if (nodes_allowed != &node_states[N_MEMORY]) 2086 NODEMASK_FREE(nodes_allowed); 2087 } 2088out: 2089 return ret; 2090} 2091 2092int hugetlb_sysctl_handler(struct ctl_table *table, int write, 2093 void __user *buffer, size_t *length, loff_t *ppos) 2094{ 2095 2096 return hugetlb_sysctl_handler_common(false, table, write, 2097 buffer, length, ppos); 2098} 2099 2100#ifdef CONFIG_NUMA 2101int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 2102 void __user *buffer, size_t *length, loff_t *ppos) 2103{ 2104 return hugetlb_sysctl_handler_common(true, table, write, 2105 buffer, length, ppos); 2106} 2107#endif /* CONFIG_NUMA */ 2108 2109int hugetlb_overcommit_handler(struct ctl_table *table, int write, 2110 void __user *buffer, 2111 size_t *length, loff_t *ppos) 2112{ 2113 struct hstate *h = &default_hstate; 2114 unsigned long tmp; 2115 int ret; 2116 2117 tmp = h->nr_overcommit_huge_pages; 2118 2119 if (write && h->order >= MAX_ORDER) 2120 return -EINVAL; 2121 2122 table->data = &tmp; 2123 table->maxlen = sizeof(unsigned long); 2124 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 2125 if (ret) 2126 goto out; 2127 2128 if (write) { 2129 spin_lock(&hugetlb_lock); 2130 h->nr_overcommit_huge_pages = tmp; 2131 spin_unlock(&hugetlb_lock); 2132 } 2133out: 2134 return ret; 2135} 2136 2137#endif /* CONFIG_SYSCTL */ 2138 2139void hugetlb_report_meminfo(struct seq_file *m) 2140{ 2141 struct hstate *h = &default_hstate; 2142 seq_printf(m, 2143 "HugePages_Total: %5lu\n" 2144 "HugePages_Free: %5lu\n" 2145 "HugePages_Rsvd: %5lu\n" 2146 "HugePages_Surp: %5lu\n" 2147 "Hugepagesize: %8lu kB\n", 2148 h->nr_huge_pages, 2149 h->free_huge_pages, 2150 h->resv_huge_pages, 2151 h->surplus_huge_pages, 2152 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2153} 2154 2155int hugetlb_report_node_meminfo(int nid, char *buf) 2156{ 2157 struct hstate *h = &default_hstate; 2158 return sprintf(buf, 2159 "Node %d HugePages_Total: %5u\n" 2160 "Node %d HugePages_Free: %5u\n" 2161 "Node %d HugePages_Surp: %5u\n", 2162 nid, h->nr_huge_pages_node[nid], 2163 nid, h->free_huge_pages_node[nid], 2164 nid, h->surplus_huge_pages_node[nid]); 2165} 2166 2167void hugetlb_show_meminfo(void) 2168{ 2169 struct hstate *h; 2170 int nid; 2171 2172 for_each_node_state(nid, N_MEMORY) 2173 for_each_hstate(h) 2174 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2175 nid, 2176 h->nr_huge_pages_node[nid], 2177 h->free_huge_pages_node[nid], 2178 h->surplus_huge_pages_node[nid], 2179 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2180} 2181 2182/* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 2183unsigned long hugetlb_total_pages(void) 2184{ 2185 struct hstate *h; 2186 unsigned long nr_total_pages = 0; 2187 2188 for_each_hstate(h) 2189 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2190 return nr_total_pages; 2191} 2192 2193static int hugetlb_acct_memory(struct hstate *h, long delta) 2194{ 2195 int ret = -ENOMEM; 2196 2197 spin_lock(&hugetlb_lock); 2198 /* 2199 * When cpuset is configured, it breaks the strict hugetlb page 2200 * reservation as the accounting is done on a global variable. Such 2201 * reservation is completely rubbish in the presence of cpuset because 2202 * the reservation is not checked against page availability for the 2203 * current cpuset. Application can still potentially OOM'ed by kernel 2204 * with lack of free htlb page in cpuset that the task is in. 2205 * Attempt to enforce strict accounting with cpuset is almost 2206 * impossible (or too ugly) because cpuset is too fluid that 2207 * task or memory node can be dynamically moved between cpusets. 2208 * 2209 * The change of semantics for shared hugetlb mapping with cpuset is 2210 * undesirable. However, in order to preserve some of the semantics, 2211 * we fall back to check against current free page availability as 2212 * a best attempt and hopefully to minimize the impact of changing 2213 * semantics that cpuset has. 2214 */ 2215 if (delta > 0) { 2216 if (gather_surplus_pages(h, delta) < 0) 2217 goto out; 2218 2219 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2220 return_unused_surplus_pages(h, delta); 2221 goto out; 2222 } 2223 } 2224 2225 ret = 0; 2226 if (delta < 0) 2227 return_unused_surplus_pages(h, (unsigned long) -delta); 2228 2229out: 2230 spin_unlock(&hugetlb_lock); 2231 return ret; 2232} 2233 2234static void hugetlb_vm_op_open(struct vm_area_struct *vma) 2235{ 2236 struct resv_map *resv = vma_resv_map(vma); 2237 2238 /* 2239 * This new VMA should share its siblings reservation map if present. 2240 * The VMA will only ever have a valid reservation map pointer where 2241 * it is being copied for another still existing VMA. As that VMA 2242 * has a reference to the reservation map it cannot disappear until 2243 * after this open call completes. It is therefore safe to take a 2244 * new reference here without additional locking. 2245 */ 2246 if (resv) 2247 kref_get(&resv->refs); 2248} 2249 2250static void resv_map_put(struct vm_area_struct *vma) 2251{ 2252 struct resv_map *resv = vma_resv_map(vma); 2253 2254 if (!resv) 2255 return; 2256 kref_put(&resv->refs, resv_map_release); 2257} 2258 2259static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2260{ 2261 struct hstate *h = hstate_vma(vma); 2262 struct resv_map *resv = vma_resv_map(vma); 2263 struct hugepage_subpool *spool = subpool_vma(vma); 2264 unsigned long reserve; 2265 unsigned long start; 2266 unsigned long end; 2267 2268 if (resv) { 2269 start = vma_hugecache_offset(h, vma, vma->vm_start); 2270 end = vma_hugecache_offset(h, vma, vma->vm_end); 2271 2272 reserve = (end - start) - 2273 region_count(&resv->regions, start, end); 2274 2275 resv_map_put(vma); 2276 2277 if (reserve) { 2278 hugetlb_acct_memory(h, -reserve); 2279 hugepage_subpool_put_pages(spool, reserve); 2280 } 2281 } 2282} 2283 2284/* 2285 * We cannot handle pagefaults against hugetlb pages at all. They cause 2286 * handle_mm_fault() to try to instantiate regular-sized pages in the 2287 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 2288 * this far. 2289 */ 2290static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 2291{ 2292 BUG(); 2293 return 0; 2294} 2295 2296const struct vm_operations_struct hugetlb_vm_ops = { 2297 .fault = hugetlb_vm_op_fault, 2298 .open = hugetlb_vm_op_open, 2299 .close = hugetlb_vm_op_close, 2300}; 2301 2302static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 2303 int writable) 2304{ 2305 pte_t entry; 2306 2307 if (writable) { 2308 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2309 vma->vm_page_prot))); 2310 } else { 2311 entry = huge_pte_wrprotect(mk_huge_pte(page, 2312 vma->vm_page_prot)); 2313 } 2314 entry = pte_mkyoung(entry); 2315 entry = pte_mkhuge(entry); 2316 entry = arch_make_huge_pte(entry, vma, page, writable); 2317 2318 return entry; 2319} 2320 2321static void set_huge_ptep_writable(struct vm_area_struct *vma, 2322 unsigned long address, pte_t *ptep) 2323{ 2324 pte_t entry; 2325 2326 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 2327 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 2328 update_mmu_cache(vma, address, ptep); 2329} 2330 2331 2332int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 2333 struct vm_area_struct *vma) 2334{ 2335 pte_t *src_pte, *dst_pte, entry; 2336 struct page *ptepage; 2337 unsigned long addr; 2338 int cow; 2339 struct hstate *h = hstate_vma(vma); 2340 unsigned long sz = huge_page_size(h); 2341 2342 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 2343 2344 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2345 spinlock_t *src_ptl, *dst_ptl; 2346 src_pte = huge_pte_offset(src, addr); 2347 if (!src_pte) 2348 continue; 2349 dst_pte = huge_pte_alloc(dst, addr, sz); 2350 if (!dst_pte) 2351 goto nomem; 2352 2353 /* If the pagetables are shared don't copy or take references */ 2354 if (dst_pte == src_pte) 2355 continue; 2356 2357 dst_ptl = huge_pte_lock(h, dst, dst_pte); 2358 src_ptl = huge_pte_lockptr(h, src, src_pte); 2359 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 2360 if (!huge_pte_none(huge_ptep_get(src_pte))) { 2361 if (cow) 2362 huge_ptep_set_wrprotect(src, addr, src_pte); 2363 entry = huge_ptep_get(src_pte); 2364 ptepage = pte_page(entry); 2365 get_page(ptepage); 2366 page_dup_rmap(ptepage); 2367 set_huge_pte_at(dst, addr, dst_pte, entry); 2368 } 2369 spin_unlock(src_ptl); 2370 spin_unlock(dst_ptl); 2371 } 2372 return 0; 2373 2374nomem: 2375 return -ENOMEM; 2376} 2377 2378static int is_hugetlb_entry_migration(pte_t pte) 2379{ 2380 swp_entry_t swp; 2381 2382 if (huge_pte_none(pte) || pte_present(pte)) 2383 return 0; 2384 swp = pte_to_swp_entry(pte); 2385 if (non_swap_entry(swp) && is_migration_entry(swp)) 2386 return 1; 2387 else 2388 return 0; 2389} 2390 2391static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2392{ 2393 swp_entry_t swp; 2394 2395 if (huge_pte_none(pte) || pte_present(pte)) 2396 return 0; 2397 swp = pte_to_swp_entry(pte); 2398 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2399 return 1; 2400 else 2401 return 0; 2402} 2403 2404void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 2405 unsigned long start, unsigned long end, 2406 struct page *ref_page) 2407{ 2408 int force_flush = 0; 2409 struct mm_struct *mm = vma->vm_mm; 2410 unsigned long address; 2411 pte_t *ptep; 2412 pte_t pte; 2413 spinlock_t *ptl; 2414 struct page *page; 2415 struct hstate *h = hstate_vma(vma); 2416 unsigned long sz = huge_page_size(h); 2417 const unsigned long mmun_start = start; /* For mmu_notifiers */ 2418 const unsigned long mmun_end = end; /* For mmu_notifiers */ 2419 2420 WARN_ON(!is_vm_hugetlb_page(vma)); 2421 BUG_ON(start & ~huge_page_mask(h)); 2422 BUG_ON(end & ~huge_page_mask(h)); 2423 2424 tlb_start_vma(tlb, vma); 2425 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2426again: 2427 for (address = start; address < end; address += sz) { 2428 ptep = huge_pte_offset(mm, address); 2429 if (!ptep) 2430 continue; 2431 2432 ptl = huge_pte_lock(h, mm, ptep); 2433 if (huge_pmd_unshare(mm, &address, ptep)) 2434 goto unlock; 2435 2436 pte = huge_ptep_get(ptep); 2437 if (huge_pte_none(pte)) 2438 goto unlock; 2439 2440 /* 2441 * HWPoisoned hugepage is already unmapped and dropped reference 2442 */ 2443 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 2444 huge_pte_clear(mm, address, ptep); 2445 goto unlock; 2446 } 2447 2448 page = pte_page(pte); 2449 /* 2450 * If a reference page is supplied, it is because a specific 2451 * page is being unmapped, not a range. Ensure the page we 2452 * are about to unmap is the actual page of interest. 2453 */ 2454 if (ref_page) { 2455 if (page != ref_page) 2456 goto unlock; 2457 2458 /* 2459 * Mark the VMA as having unmapped its page so that 2460 * future faults in this VMA will fail rather than 2461 * looking like data was lost 2462 */ 2463 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 2464 } 2465 2466 pte = huge_ptep_get_and_clear(mm, address, ptep); 2467 tlb_remove_tlb_entry(tlb, ptep, address); 2468 if (huge_pte_dirty(pte)) 2469 set_page_dirty(page); 2470 2471 page_remove_rmap(page); 2472 force_flush = !__tlb_remove_page(tlb, page); 2473 if (force_flush) { 2474 spin_unlock(ptl); 2475 break; 2476 } 2477 /* Bail out after unmapping reference page if supplied */ 2478 if (ref_page) { 2479 spin_unlock(ptl); 2480 break; 2481 } 2482unlock: 2483 spin_unlock(ptl); 2484 } 2485 /* 2486 * mmu_gather ran out of room to batch pages, we break out of 2487 * the PTE lock to avoid doing the potential expensive TLB invalidate 2488 * and page-free while holding it. 2489 */ 2490 if (force_flush) { 2491 force_flush = 0; 2492 tlb_flush_mmu(tlb); 2493 if (address < end && !ref_page) 2494 goto again; 2495 } 2496 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2497 tlb_end_vma(tlb, vma); 2498} 2499 2500void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2501 struct vm_area_struct *vma, unsigned long start, 2502 unsigned long end, struct page *ref_page) 2503{ 2504 __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2505 2506 /* 2507 * Clear this flag so that x86's huge_pmd_share page_table_shareable 2508 * test will fail on a vma being torn down, and not grab a page table 2509 * on its way out. We're lucky that the flag has such an appropriate 2510 * name, and can in fact be safely cleared here. We could clear it 2511 * before the __unmap_hugepage_range above, but all that's necessary 2512 * is to clear it before releasing the i_mmap_mutex. This works 2513 * because in the context this is called, the VMA is about to be 2514 * destroyed and the i_mmap_mutex is held. 2515 */ 2516 vma->vm_flags &= ~VM_MAYSHARE; 2517} 2518 2519void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 2520 unsigned long end, struct page *ref_page) 2521{ 2522 struct mm_struct *mm; 2523 struct mmu_gather tlb; 2524 2525 mm = vma->vm_mm; 2526 2527 tlb_gather_mmu(&tlb, mm, start, end); 2528 __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 2529 tlb_finish_mmu(&tlb, start, end); 2530} 2531 2532/* 2533 * This is called when the original mapper is failing to COW a MAP_PRIVATE 2534 * mappping it owns the reserve page for. The intention is to unmap the page 2535 * from other VMAs and let the children be SIGKILLed if they are faulting the 2536 * same region. 2537 */ 2538static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 2539 struct page *page, unsigned long address) 2540{ 2541 struct hstate *h = hstate_vma(vma); 2542 struct vm_area_struct *iter_vma; 2543 struct address_space *mapping; 2544 pgoff_t pgoff; 2545 2546 /* 2547 * vm_pgoff is in PAGE_SIZE units, hence the different calculation 2548 * from page cache lookup which is in HPAGE_SIZE units. 2549 */ 2550 address = address & huge_page_mask(h); 2551 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 2552 vma->vm_pgoff; 2553 mapping = file_inode(vma->vm_file)->i_mapping; 2554 2555 /* 2556 * Take the mapping lock for the duration of the table walk. As 2557 * this mapping should be shared between all the VMAs, 2558 * __unmap_hugepage_range() is called as the lock is already held 2559 */ 2560 mutex_lock(&mapping->i_mmap_mutex); 2561 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 2562 /* Do not unmap the current VMA */ 2563 if (iter_vma == vma) 2564 continue; 2565 2566 /* 2567 * Unmap the page from other VMAs without their own reserves. 2568 * They get marked to be SIGKILLed if they fault in these 2569 * areas. This is because a future no-page fault on this VMA 2570 * could insert a zeroed page instead of the data existing 2571 * from the time of fork. This would look like data corruption 2572 */ 2573 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 2574 unmap_hugepage_range(iter_vma, address, 2575 address + huge_page_size(h), page); 2576 } 2577 mutex_unlock(&mapping->i_mmap_mutex); 2578 2579 return 1; 2580} 2581 2582/* 2583 * Hugetlb_cow() should be called with page lock of the original hugepage held. 2584 * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2585 * cannot race with other handlers or page migration. 2586 * Keep the pte_same checks anyway to make transition from the mutex easier. 2587 */ 2588static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 2589 unsigned long address, pte_t *ptep, pte_t pte, 2590 struct page *pagecache_page, spinlock_t *ptl) 2591{ 2592 struct hstate *h = hstate_vma(vma); 2593 struct page *old_page, *new_page; 2594 int outside_reserve = 0; 2595 unsigned long mmun_start; /* For mmu_notifiers */ 2596 unsigned long mmun_end; /* For mmu_notifiers */ 2597 2598 old_page = pte_page(pte); 2599 2600retry_avoidcopy: 2601 /* If no-one else is actually using this page, avoid the copy 2602 * and just make the page writable */ 2603 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 2604 page_move_anon_rmap(old_page, vma, address); 2605 set_huge_ptep_writable(vma, address, ptep); 2606 return 0; 2607 } 2608 2609 /* 2610 * If the process that created a MAP_PRIVATE mapping is about to 2611 * perform a COW due to a shared page count, attempt to satisfy 2612 * the allocation without using the existing reserves. The pagecache 2613 * page is used to determine if the reserve at this address was 2614 * consumed or not. If reserves were used, a partial faulted mapping 2615 * at the time of fork() could consume its reserves on COW instead 2616 * of the full address range. 2617 */ 2618 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 2619 old_page != pagecache_page) 2620 outside_reserve = 1; 2621 2622 page_cache_get(old_page); 2623 2624 /* Drop page table lock as buddy allocator may be called */ 2625 spin_unlock(ptl); 2626 new_page = alloc_huge_page(vma, address, outside_reserve); 2627 2628 if (IS_ERR(new_page)) { 2629 long err = PTR_ERR(new_page); 2630 page_cache_release(old_page); 2631 2632 /* 2633 * If a process owning a MAP_PRIVATE mapping fails to COW, 2634 * it is due to references held by a child and an insufficient 2635 * huge page pool. To guarantee the original mappers 2636 * reliability, unmap the page from child processes. The child 2637 * may get SIGKILLed if it later faults. 2638 */ 2639 if (outside_reserve) { 2640 BUG_ON(huge_pte_none(pte)); 2641 if (unmap_ref_private(mm, vma, old_page, address)) { 2642 BUG_ON(huge_pte_none(pte)); 2643 spin_lock(ptl); 2644 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2645 if (likely(pte_same(huge_ptep_get(ptep), pte))) 2646 goto retry_avoidcopy; 2647 /* 2648 * race occurs while re-acquiring page table 2649 * lock, and our job is done. 2650 */ 2651 return 0; 2652 } 2653 WARN_ON_ONCE(1); 2654 } 2655 2656 /* Caller expects lock to be held */ 2657 spin_lock(ptl); 2658 if (err == -ENOMEM) 2659 return VM_FAULT_OOM; 2660 else 2661 return VM_FAULT_SIGBUS; 2662 } 2663 2664 /* 2665 * When the original hugepage is shared one, it does not have 2666 * anon_vma prepared. 2667 */ 2668 if (unlikely(anon_vma_prepare(vma))) { 2669 page_cache_release(new_page); 2670 page_cache_release(old_page); 2671 /* Caller expects lock to be held */ 2672 spin_lock(ptl); 2673 return VM_FAULT_OOM; 2674 } 2675 2676 copy_user_huge_page(new_page, old_page, address, vma, 2677 pages_per_huge_page(h)); 2678 __SetPageUptodate(new_page); 2679 2680 mmun_start = address & huge_page_mask(h); 2681 mmun_end = mmun_start + huge_page_size(h); 2682 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2683 /* 2684 * Retake the page table lock to check for racing updates 2685 * before the page tables are altered 2686 */ 2687 spin_lock(ptl); 2688 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2689 if (likely(pte_same(huge_ptep_get(ptep), pte))) { 2690 ClearPagePrivate(new_page); 2691 2692 /* Break COW */ 2693 huge_ptep_clear_flush(vma, address, ptep); 2694 set_huge_pte_at(mm, address, ptep, 2695 make_huge_pte(vma, new_page, 1)); 2696 page_remove_rmap(old_page); 2697 hugepage_add_new_anon_rmap(new_page, vma, address); 2698 /* Make the old page be freed below */ 2699 new_page = old_page; 2700 } 2701 spin_unlock(ptl); 2702 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2703 page_cache_release(new_page); 2704 page_cache_release(old_page); 2705 2706 /* Caller expects lock to be held */ 2707 spin_lock(ptl); 2708 return 0; 2709} 2710 2711/* Return the pagecache page at a given address within a VMA */ 2712static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2713 struct vm_area_struct *vma, unsigned long address) 2714{ 2715 struct address_space *mapping; 2716 pgoff_t idx; 2717 2718 mapping = vma->vm_file->f_mapping; 2719 idx = vma_hugecache_offset(h, vma, address); 2720 2721 return find_lock_page(mapping, idx); 2722} 2723 2724/* 2725 * Return whether there is a pagecache page to back given address within VMA. 2726 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 2727 */ 2728static bool hugetlbfs_pagecache_present(struct hstate *h, 2729 struct vm_area_struct *vma, unsigned long address) 2730{ 2731 struct address_space *mapping; 2732 pgoff_t idx; 2733 struct page *page; 2734 2735 mapping = vma->vm_file->f_mapping; 2736 idx = vma_hugecache_offset(h, vma, address); 2737 2738 page = find_get_page(mapping, idx); 2739 if (page) 2740 put_page(page); 2741 return page != NULL; 2742} 2743 2744static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2745 unsigned long address, pte_t *ptep, unsigned int flags) 2746{ 2747 struct hstate *h = hstate_vma(vma); 2748 int ret = VM_FAULT_SIGBUS; 2749 int anon_rmap = 0; 2750 pgoff_t idx; 2751 unsigned long size; 2752 struct page *page; 2753 struct address_space *mapping; 2754 pte_t new_pte; 2755 spinlock_t *ptl; 2756 2757 /* 2758 * Currently, we are forced to kill the process in the event the 2759 * original mapper has unmapped pages from the child due to a failed 2760 * COW. Warn that such a situation has occurred as it may not be obvious 2761 */ 2762 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2763 pr_warning("PID %d killed due to inadequate hugepage pool\n", 2764 current->pid); 2765 return ret; 2766 } 2767 2768 mapping = vma->vm_file->f_mapping; 2769 idx = vma_hugecache_offset(h, vma, address); 2770 2771 /* 2772 * Use page lock to guard against racing truncation 2773 * before we get page_table_lock. 2774 */ 2775retry: 2776 page = find_lock_page(mapping, idx); 2777 if (!page) { 2778 size = i_size_read(mapping->host) >> huge_page_shift(h); 2779 if (idx >= size) 2780 goto out; 2781 page = alloc_huge_page(vma, address, 0); 2782 if (IS_ERR(page)) { 2783 ret = PTR_ERR(page); 2784 if (ret == -ENOMEM) 2785 ret = VM_FAULT_OOM; 2786 else 2787 ret = VM_FAULT_SIGBUS; 2788 goto out; 2789 } 2790 clear_huge_page(page, address, pages_per_huge_page(h)); 2791 __SetPageUptodate(page); 2792 2793 if (vma->vm_flags & VM_MAYSHARE) { 2794 int err; 2795 struct inode *inode = mapping->host; 2796 2797 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 2798 if (err) { 2799 put_page(page); 2800 if (err == -EEXIST) 2801 goto retry; 2802 goto out; 2803 } 2804 ClearPagePrivate(page); 2805 2806 spin_lock(&inode->i_lock); 2807 inode->i_blocks += blocks_per_huge_page(h); 2808 spin_unlock(&inode->i_lock); 2809 } else { 2810 lock_page(page); 2811 if (unlikely(anon_vma_prepare(vma))) { 2812 ret = VM_FAULT_OOM; 2813 goto backout_unlocked; 2814 } 2815 anon_rmap = 1; 2816 } 2817 } else { 2818 /* 2819 * If memory error occurs between mmap() and fault, some process 2820 * don't have hwpoisoned swap entry for errored virtual address. 2821 * So we need to block hugepage fault by PG_hwpoison bit check. 2822 */ 2823 if (unlikely(PageHWPoison(page))) { 2824 ret = VM_FAULT_HWPOISON | 2825 VM_FAULT_SET_HINDEX(hstate_index(h)); 2826 goto backout_unlocked; 2827 } 2828 } 2829 2830 /* 2831 * If we are going to COW a private mapping later, we examine the 2832 * pending reservations for this page now. This will ensure that 2833 * any allocations necessary to record that reservation occur outside 2834 * the spinlock. 2835 */ 2836 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 2837 if (vma_needs_reservation(h, vma, address) < 0) { 2838 ret = VM_FAULT_OOM; 2839 goto backout_unlocked; 2840 } 2841 2842 ptl = huge_pte_lockptr(h, mm, ptep); 2843 spin_lock(ptl); 2844 size = i_size_read(mapping->host) >> huge_page_shift(h); 2845 if (idx >= size) 2846 goto backout; 2847 2848 ret = 0; 2849 if (!huge_pte_none(huge_ptep_get(ptep))) 2850 goto backout; 2851 2852 if (anon_rmap) { 2853 ClearPagePrivate(page); 2854 hugepage_add_new_anon_rmap(page, vma, address); 2855 } 2856 else 2857 page_dup_rmap(page); 2858 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 2859 && (vma->vm_flags & VM_SHARED))); 2860 set_huge_pte_at(mm, address, ptep, new_pte); 2861 2862 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 2863 /* Optimization, do the COW without a second fault */ 2864 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 2865 } 2866 2867 spin_unlock(ptl); 2868 unlock_page(page); 2869out: 2870 return ret; 2871 2872backout: 2873 spin_unlock(ptl); 2874backout_unlocked: 2875 unlock_page(page); 2876 put_page(page); 2877 goto out; 2878} 2879 2880int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2881 unsigned long address, unsigned int flags) 2882{ 2883 pte_t *ptep; 2884 pte_t entry; 2885 spinlock_t *ptl; 2886 int ret; 2887 struct page *page = NULL; 2888 struct page *pagecache_page = NULL; 2889 static DEFINE_MUTEX(hugetlb_instantiation_mutex); 2890 struct hstate *h = hstate_vma(vma); 2891 2892 address &= huge_page_mask(h); 2893 2894 ptep = huge_pte_offset(mm, address); 2895 if (ptep) { 2896 entry = huge_ptep_get(ptep); 2897 if (unlikely(is_hugetlb_entry_migration(entry))) { 2898 migration_entry_wait_huge(vma, mm, ptep); 2899 return 0; 2900 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2901 return VM_FAULT_HWPOISON_LARGE | 2902 VM_FAULT_SET_HINDEX(hstate_index(h)); 2903 } 2904 2905 ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 2906 if (!ptep) 2907 return VM_FAULT_OOM; 2908 2909 /* 2910 * Serialize hugepage allocation and instantiation, so that we don't 2911 * get spurious allocation failures if two CPUs race to instantiate 2912 * the same page in the page cache. 2913 */ 2914 mutex_lock(&hugetlb_instantiation_mutex); 2915 entry = huge_ptep_get(ptep); 2916 if (huge_pte_none(entry)) { 2917 ret = hugetlb_no_page(mm, vma, address, ptep, flags); 2918 goto out_mutex; 2919 } 2920 2921 ret = 0; 2922 2923 /* 2924 * If we are going to COW the mapping later, we examine the pending 2925 * reservations for this page now. This will ensure that any 2926 * allocations necessary to record that reservation occur outside the 2927 * spinlock. For private mappings, we also lookup the pagecache 2928 * page now as it is used to determine if a reservation has been 2929 * consumed. 2930 */ 2931 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 2932 if (vma_needs_reservation(h, vma, address) < 0) { 2933 ret = VM_FAULT_OOM; 2934 goto out_mutex; 2935 } 2936 2937 if (!(vma->vm_flags & VM_MAYSHARE)) 2938 pagecache_page = hugetlbfs_pagecache_page(h, 2939 vma, address); 2940 } 2941 2942 /* 2943 * hugetlb_cow() requires page locks of pte_page(entry) and 2944 * pagecache_page, so here we need take the former one 2945 * when page != pagecache_page or !pagecache_page. 2946 * Note that locking order is always pagecache_page -> page, 2947 * so no worry about deadlock. 2948 */ 2949 page = pte_page(entry); 2950 get_page(page); 2951 if (page != pagecache_page) 2952 lock_page(page); 2953 2954 ptl = huge_pte_lockptr(h, mm, ptep); 2955 spin_lock(ptl); 2956 /* Check for a racing update before calling hugetlb_cow */ 2957 if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 2958 goto out_ptl; 2959 2960 2961 if (flags & FAULT_FLAG_WRITE) { 2962 if (!huge_pte_write(entry)) { 2963 ret = hugetlb_cow(mm, vma, address, ptep, entry, 2964 pagecache_page, ptl); 2965 goto out_ptl; 2966 } 2967 entry = huge_pte_mkdirty(entry); 2968 } 2969 entry = pte_mkyoung(entry); 2970 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 2971 flags & FAULT_FLAG_WRITE)) 2972 update_mmu_cache(vma, address, ptep); 2973 2974out_ptl: 2975 spin_unlock(ptl); 2976 2977 if (pagecache_page) { 2978 unlock_page(pagecache_page); 2979 put_page(pagecache_page); 2980 } 2981 if (page != pagecache_page) 2982 unlock_page(page); 2983 put_page(page); 2984 2985out_mutex: 2986 mutex_unlock(&hugetlb_instantiation_mutex); 2987 2988 return ret; 2989} 2990 2991long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 2992 struct page **pages, struct vm_area_struct **vmas, 2993 unsigned long *position, unsigned long *nr_pages, 2994 long i, unsigned int flags) 2995{ 2996 unsigned long pfn_offset; 2997 unsigned long vaddr = *position; 2998 unsigned long remainder = *nr_pages; 2999 struct hstate *h = hstate_vma(vma); 3000 3001 while (vaddr < vma->vm_end && remainder) { 3002 pte_t *pte; 3003 spinlock_t *ptl = NULL; 3004 int absent; 3005 struct page *page; 3006 3007 /* 3008 * Some archs (sparc64, sh*) have multiple pte_ts to 3009 * each hugepage. We have to make sure we get the 3010 * first, for the page indexing below to work. 3011 * 3012 * Note that page table lock is not held when pte is null. 3013 */ 3014 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3015 if (pte) 3016 ptl = huge_pte_lock(h, mm, pte); 3017 absent = !pte || huge_pte_none(huge_ptep_get(pte)); 3018 3019 /* 3020 * When coredumping, it suits get_dump_page if we just return 3021 * an error where there's an empty slot with no huge pagecache 3022 * to back it. This way, we avoid allocating a hugepage, and 3023 * the sparse dumpfile avoids allocating disk blocks, but its 3024 * huge holes still show up with zeroes where they need to be. 3025 */ 3026 if (absent && (flags & FOLL_DUMP) && 3027 !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3028 if (pte) 3029 spin_unlock(ptl); 3030 remainder = 0; 3031 break; 3032 } 3033 3034 /* 3035 * We need call hugetlb_fault for both hugepages under migration 3036 * (in which case hugetlb_fault waits for the migration,) and 3037 * hwpoisoned hugepages (in which case we need to prevent the 3038 * caller from accessing to them.) In order to do this, we use 3039 * here is_swap_pte instead of is_hugetlb_entry_migration and 3040 * is_hugetlb_entry_hwpoisoned. This is because it simply covers 3041 * both cases, and because we can't follow correct pages 3042 * directly from any kind of swap entries. 3043 */ 3044 if (absent || is_swap_pte(huge_ptep_get(pte)) || 3045 ((flags & FOLL_WRITE) && 3046 !huge_pte_write(huge_ptep_get(pte)))) { 3047 int ret; 3048 3049 if (pte) 3050 spin_unlock(ptl); 3051 ret = hugetlb_fault(mm, vma, vaddr, 3052 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3053 if (!(ret & VM_FAULT_ERROR)) 3054 continue; 3055 3056 remainder = 0; 3057 break; 3058 } 3059 3060 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 3061 page = pte_page(huge_ptep_get(pte)); 3062same_page: 3063 if (pages) { 3064 pages[i] = mem_map_offset(page, pfn_offset); 3065 get_page(pages[i]); 3066 } 3067 3068 if (vmas) 3069 vmas[i] = vma; 3070 3071 vaddr += PAGE_SIZE; 3072 ++pfn_offset; 3073 --remainder; 3074 ++i; 3075 if (vaddr < vma->vm_end && remainder && 3076 pfn_offset < pages_per_huge_page(h)) { 3077 /* 3078 * We use pfn_offset to avoid touching the pageframes 3079 * of this compound page. 3080 */ 3081 goto same_page; 3082 } 3083 spin_unlock(ptl); 3084 } 3085 *nr_pages = remainder; 3086 *position = vaddr; 3087 3088 return i ? i : -EFAULT; 3089} 3090 3091unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 3092 unsigned long address, unsigned long end, pgprot_t newprot) 3093{ 3094 struct mm_struct *mm = vma->vm_mm; 3095 unsigned long start = address; 3096 pte_t *ptep; 3097 pte_t pte; 3098 struct hstate *h = hstate_vma(vma); 3099 unsigned long pages = 0; 3100 3101 BUG_ON(address >= end); 3102 flush_cache_range(vma, address, end); 3103 3104 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 3105 for (; address < end; address += huge_page_size(h)) { 3106 spinlock_t *ptl; 3107 ptep = huge_pte_offset(mm, address); 3108 if (!ptep) 3109 continue; 3110 ptl = huge_pte_lock(h, mm, ptep); 3111 if (huge_pmd_unshare(mm, &address, ptep)) { 3112 pages++; 3113 spin_unlock(ptl); 3114 continue; 3115 } 3116 if (!huge_pte_none(huge_ptep_get(ptep))) { 3117 pte = huge_ptep_get_and_clear(mm, address, ptep); 3118 pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3119 pte = arch_make_huge_pte(pte, vma, NULL, 0); 3120 set_huge_pte_at(mm, address, ptep, pte); 3121 pages++; 3122 } 3123 spin_unlock(ptl); 3124 } 3125 /* 3126 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3127 * may have cleared our pud entry and done put_page on the page table: 3128 * once we release i_mmap_mutex, another task can do the final put_page 3129 * and that page table be reused and filled with junk. 3130 */ 3131 flush_tlb_range(vma, start, end); 3132 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 3133 3134 return pages << h->order; 3135} 3136 3137int hugetlb_reserve_pages(struct inode *inode, 3138 long from, long to, 3139 struct vm_area_struct *vma, 3140 vm_flags_t vm_flags) 3141{ 3142 long ret, chg; 3143 struct hstate *h = hstate_inode(inode); 3144 struct hugepage_subpool *spool = subpool_inode(inode); 3145 3146 /* 3147 * Only apply hugepage reservation if asked. At fault time, an 3148 * attempt will be made for VM_NORESERVE to allocate a page 3149 * without using reserves 3150 */ 3151 if (vm_flags & VM_NORESERVE) 3152 return 0; 3153 3154 /* 3155 * Shared mappings base their reservation on the number of pages that 3156 * are already allocated on behalf of the file. Private mappings need 3157 * to reserve the full area even if read-only as mprotect() may be 3158 * called to make the mapping read-write. Assume !vma is a shm mapping 3159 */ 3160 if (!vma || vma->vm_flags & VM_MAYSHARE) 3161 chg = region_chg(&inode->i_mapping->private_list, from, to); 3162 else { 3163 struct resv_map *resv_map = resv_map_alloc(); 3164 if (!resv_map) 3165 return -ENOMEM; 3166 3167 chg = to - from; 3168 3169 set_vma_resv_map(vma, resv_map); 3170 set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 3171 } 3172 3173 if (chg < 0) { 3174 ret = chg; 3175 goto out_err; 3176 } 3177 3178 /* There must be enough pages in the subpool for the mapping */ 3179 if (hugepage_subpool_get_pages(spool, chg)) { 3180 ret = -ENOSPC; 3181 goto out_err; 3182 } 3183 3184 /* 3185 * Check enough hugepages are available for the reservation. 3186 * Hand the pages back to the subpool if there are not 3187 */ 3188 ret = hugetlb_acct_memory(h, chg); 3189 if (ret < 0) { 3190 hugepage_subpool_put_pages(spool, chg); 3191 goto out_err; 3192 } 3193 3194 /* 3195 * Account for the reservations made. Shared mappings record regions 3196 * that have reservations as they are shared by multiple VMAs. 3197 * When the last VMA disappears, the region map says how much 3198 * the reservation was and the page cache tells how much of 3199 * the reservation was consumed. Private mappings are per-VMA and 3200 * only the consumed reservations are tracked. When the VMA 3201 * disappears, the original reservation is the VMA size and the 3202 * consumed reservations are stored in the map. Hence, nothing 3203 * else has to be done for private mappings here 3204 */ 3205 if (!vma || vma->vm_flags & VM_MAYSHARE) 3206 region_add(&inode->i_mapping->private_list, from, to); 3207 return 0; 3208out_err: 3209 if (vma) 3210 resv_map_put(vma); 3211 return ret; 3212} 3213 3214void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3215{ 3216 struct hstate *h = hstate_inode(inode); 3217 long chg = region_truncate(&inode->i_mapping->private_list, offset); 3218 struct hugepage_subpool *spool = subpool_inode(inode); 3219 3220 spin_lock(&inode->i_lock); 3221 inode->i_blocks -= (blocks_per_huge_page(h) * freed); 3222 spin_unlock(&inode->i_lock); 3223 3224 hugepage_subpool_put_pages(spool, (chg - freed)); 3225 hugetlb_acct_memory(h, -(chg - freed)); 3226} 3227 3228#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 3229static unsigned long page_table_shareable(struct vm_area_struct *svma, 3230 struct vm_area_struct *vma, 3231 unsigned long addr, pgoff_t idx) 3232{ 3233 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 3234 svma->vm_start; 3235 unsigned long sbase = saddr & PUD_MASK; 3236 unsigned long s_end = sbase + PUD_SIZE; 3237 3238 /* Allow segments to share if only one is marked locked */ 3239 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 3240 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 3241 3242 /* 3243 * match the virtual addresses, permission and the alignment of the 3244 * page table page. 3245 */ 3246 if (pmd_index(addr) != pmd_index(saddr) || 3247 vm_flags != svm_flags || 3248 sbase < svma->vm_start || svma->vm_end < s_end) 3249 return 0; 3250 3251 return saddr; 3252} 3253 3254static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 3255{ 3256 unsigned long base = addr & PUD_MASK; 3257 unsigned long end = base + PUD_SIZE; 3258 3259 /* 3260 * check on proper vm_flags and page table alignment 3261 */ 3262 if (vma->vm_flags & VM_MAYSHARE && 3263 vma->vm_start <= base && end <= vma->vm_end) 3264 return 1; 3265 return 0; 3266} 3267 3268/* 3269 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 3270 * and returns the corresponding pte. While this is not necessary for the 3271 * !shared pmd case because we can allocate the pmd later as well, it makes the 3272 * code much cleaner. pmd allocation is essential for the shared case because 3273 * pud has to be populated inside the same i_mmap_mutex section - otherwise 3274 * racing tasks could either miss the sharing (see huge_pte_offset) or select a 3275 * bad pmd for sharing. 3276 */ 3277pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 3278{ 3279 struct vm_area_struct *vma = find_vma(mm, addr); 3280 struct address_space *mapping = vma->vm_file->f_mapping; 3281 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 3282 vma->vm_pgoff; 3283 struct vm_area_struct *svma; 3284 unsigned long saddr; 3285 pte_t *spte = NULL; 3286 pte_t *pte; 3287 spinlock_t *ptl; 3288 3289 if (!vma_shareable(vma, addr)) 3290 return (pte_t *)pmd_alloc(mm, pud, addr); 3291 3292 mutex_lock(&mapping->i_mmap_mutex); 3293 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 3294 if (svma == vma) 3295 continue; 3296 3297 saddr = page_table_shareable(svma, vma, addr, idx); 3298 if (saddr) { 3299 spte = huge_pte_offset(svma->vm_mm, saddr); 3300 if (spte) { 3301 get_page(virt_to_page(spte)); 3302 break; 3303 } 3304 } 3305 } 3306 3307 if (!spte) 3308 goto out; 3309 3310 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 3311 spin_lock(ptl); 3312 if (pud_none(*pud)) 3313 pud_populate(mm, pud, 3314 (pmd_t *)((unsigned long)spte & PAGE_MASK)); 3315 else 3316 put_page(virt_to_page(spte)); 3317 spin_unlock(ptl); 3318out: 3319 pte = (pte_t *)pmd_alloc(mm, pud, addr); 3320 mutex_unlock(&mapping->i_mmap_mutex); 3321 return pte; 3322} 3323 3324/* 3325 * unmap huge page backed by shared pte. 3326 * 3327 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 3328 * indicated by page_count > 1, unmap is achieved by clearing pud and 3329 * decrementing the ref count. If count == 1, the pte page is not shared. 3330 * 3331 * called with page table lock held. 3332 * 3333 * returns: 1 successfully unmapped a shared pte page 3334 * 0 the underlying pte page is not shared, or it is the last user 3335 */ 3336int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 3337{ 3338 pgd_t *pgd = pgd_offset(mm, *addr); 3339 pud_t *pud = pud_offset(pgd, *addr); 3340 3341 BUG_ON(page_count(virt_to_page(ptep)) == 0); 3342 if (page_count(virt_to_page(ptep)) == 1) 3343 return 0; 3344 3345 pud_clear(pud); 3346 put_page(virt_to_page(ptep)); 3347 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 3348 return 1; 3349} 3350#define want_pmd_share() (1) 3351#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 3352pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 3353{ 3354 return NULL; 3355} 3356#define want_pmd_share() (0) 3357#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 3358 3359#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 3360pte_t *huge_pte_alloc(struct mm_struct *mm, 3361 unsigned long addr, unsigned long sz) 3362{ 3363 pgd_t *pgd; 3364 pud_t *pud; 3365 pte_t *pte = NULL; 3366 3367 pgd = pgd_offset(mm, addr); 3368 pud = pud_alloc(mm, pgd, addr); 3369 if (pud) { 3370 if (sz == PUD_SIZE) { 3371 pte = (pte_t *)pud; 3372 } else { 3373 BUG_ON(sz != PMD_SIZE); 3374 if (want_pmd_share() && pud_none(*pud)) 3375 pte = huge_pmd_share(mm, addr, pud); 3376 else 3377 pte = (pte_t *)pmd_alloc(mm, pud, addr); 3378 } 3379 } 3380 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 3381 3382 return pte; 3383} 3384 3385pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 3386{ 3387 pgd_t *pgd; 3388 pud_t *pud; 3389 pmd_t *pmd = NULL; 3390 3391 pgd = pgd_offset(mm, addr); 3392 if (pgd_present(*pgd)) { 3393 pud = pud_offset(pgd, addr); 3394 if (pud_present(*pud)) { 3395 if (pud_huge(*pud)) 3396 return (pte_t *)pud; 3397 pmd = pmd_offset(pud, addr); 3398 } 3399 } 3400 return (pte_t *) pmd; 3401} 3402 3403struct page * 3404follow_huge_pmd(struct mm_struct *mm, unsigned long address, 3405 pmd_t *pmd, int write) 3406{ 3407 struct page *page; 3408 3409 page = pte_page(*(pte_t *)pmd); 3410 if (page) 3411 page += ((address & ~PMD_MASK) >> PAGE_SHIFT); 3412 return page; 3413} 3414 3415struct page * 3416follow_huge_pud(struct mm_struct *mm, unsigned long address, 3417 pud_t *pud, int write) 3418{ 3419 struct page *page; 3420 3421 page = pte_page(*(pte_t *)pud); 3422 if (page) 3423 page += ((address & ~PUD_MASK) >> PAGE_SHIFT); 3424 return page; 3425} 3426 3427#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 3428 3429/* Can be overriden by architectures */ 3430__attribute__((weak)) struct page * 3431follow_huge_pud(struct mm_struct *mm, unsigned long address, 3432 pud_t *pud, int write) 3433{ 3434 BUG(); 3435 return NULL; 3436} 3437 3438#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 3439 3440#ifdef CONFIG_MEMORY_FAILURE 3441 3442/* Should be called in hugetlb_lock */ 3443static int is_hugepage_on_freelist(struct page *hpage) 3444{ 3445 struct page *page; 3446 struct page *tmp; 3447 struct hstate *h = page_hstate(hpage); 3448 int nid = page_to_nid(hpage); 3449 3450 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 3451 if (page == hpage) 3452 return 1; 3453 return 0; 3454} 3455 3456/* 3457 * This function is called from memory failure code. 3458 * Assume the caller holds page lock of the head page. 3459 */ 3460int dequeue_hwpoisoned_huge_page(struct page *hpage) 3461{ 3462 struct hstate *h = page_hstate(hpage); 3463 int nid = page_to_nid(hpage); 3464 int ret = -EBUSY; 3465 3466 spin_lock(&hugetlb_lock); 3467 if (is_hugepage_on_freelist(hpage)) { 3468 /* 3469 * Hwpoisoned hugepage isn't linked to activelist or freelist, 3470 * but dangling hpage->lru can trigger list-debug warnings 3471 * (this happens when we call unpoison_memory() on it), 3472 * so let it point to itself with list_del_init(). 3473 */ 3474 list_del_init(&hpage->lru); 3475 set_page_refcounted(hpage); 3476 h->free_huge_pages--; 3477 h->free_huge_pages_node[nid]--; 3478 ret = 0; 3479 } 3480 spin_unlock(&hugetlb_lock); 3481 return ret; 3482} 3483#endif 3484 3485bool isolate_huge_page(struct page *page, struct list_head *list) 3486{ 3487 VM_BUG_ON(!PageHead(page)); 3488 if (!get_page_unless_zero(page)) 3489 return false; 3490 spin_lock(&hugetlb_lock); 3491 list_move_tail(&page->lru, list); 3492 spin_unlock(&hugetlb_lock); 3493 return true; 3494} 3495 3496void putback_active_hugepage(struct page *page) 3497{ 3498 VM_BUG_ON(!PageHead(page)); 3499 spin_lock(&hugetlb_lock); 3500 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 3501 spin_unlock(&hugetlb_lock); 3502 put_page(page); 3503} 3504 3505bool is_hugepage_active(struct page *page) 3506{ 3507 VM_BUG_ON(!PageHuge(page)); 3508 /* 3509 * This function can be called for a tail page because the caller, 3510 * scan_movable_pages, scans through a given pfn-range which typically 3511 * covers one memory block. In systems using gigantic hugepage (1GB 3512 * for x86_64,) a hugepage is larger than a memory block, and we don't 3513 * support migrating such large hugepages for now, so return false 3514 * when called for tail pages. 3515 */ 3516 if (PageTail(page)) 3517 return false; 3518 /* 3519 * Refcount of a hwpoisoned hugepages is 1, but they are not active, 3520 * so we should return false for them. 3521 */ 3522 if (unlikely(PageHWPoison(page))) 3523 return false; 3524 return page_count(page) > 0; 3525} 3526