1 /* 2 * Generic hugetlb support. 3 * (C) Nadia Yvette Chambers, April 2004 4 */ 5 #include <linux/list.h> 6 #include <linux/init.h> 7 #include <linux/module.h> 8 #include <linux/mm.h> 9 #include <linux/seq_file.h> 10 #include <linux/sysctl.h> 11 #include <linux/highmem.h> 12 #include <linux/mmu_notifier.h> 13 #include <linux/nodemask.h> 14 #include <linux/pagemap.h> 15 #include <linux/mempolicy.h> 16 #include <linux/compiler.h> 17 #include <linux/cpuset.h> 18 #include <linux/mutex.h> 19 #include <linux/bootmem.h> 20 #include <linux/sysfs.h> 21 #include <linux/slab.h> 22 #include <linux/rmap.h> 23 #include <linux/swap.h> 24 #include <linux/swapops.h> 25 #include <linux/page-isolation.h> 26 #include <linux/jhash.h> 27 28 #include <asm/page.h> 29 #include <asm/pgtable.h> 30 #include <asm/tlb.h> 31 32 #include <linux/io.h> 33 #include <linux/hugetlb.h> 34 #include <linux/hugetlb_cgroup.h> 35 #include <linux/node.h> 36 #include "internal.h" 37 38 unsigned long hugepages_treat_as_movable; 39 40 int hugetlb_max_hstate __read_mostly; 41 unsigned int default_hstate_idx; 42 struct hstate hstates[HUGE_MAX_HSTATE]; 43 44 __initdata LIST_HEAD(huge_boot_pages); 45 46 /* for command line parsing */ 47 static struct hstate * __initdata parsed_hstate; 48 static unsigned long __initdata default_hstate_max_huge_pages; 49 static unsigned long __initdata default_hstate_size; 50 51 /* 52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 53 * free_huge_pages, and surplus_huge_pages. 54 */ 55 DEFINE_SPINLOCK(hugetlb_lock); 56 57 /* 58 * Serializes faults on the same logical page. This is used to 59 * prevent spurious OOMs when the hugepage pool is fully utilized. 60 */ 61 static int num_fault_mutexes; 62 static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; 63 64 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 65 { 66 bool free = (spool->count == 0) && (spool->used_hpages == 0); 67 68 spin_unlock(&spool->lock); 69 70 /* If no pages are used, and no other handles to the subpool 71 * remain, free the subpool the subpool remain */ 72 if (free) 73 kfree(spool); 74 } 75 76 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 77 { 78 struct hugepage_subpool *spool; 79 80 spool = kmalloc(sizeof(*spool), GFP_KERNEL); 81 if (!spool) 82 return NULL; 83 84 spin_lock_init(&spool->lock); 85 spool->count = 1; 86 spool->max_hpages = nr_blocks; 87 spool->used_hpages = 0; 88 89 return spool; 90 } 91 92 void hugepage_put_subpool(struct hugepage_subpool *spool) 93 { 94 spin_lock(&spool->lock); 95 BUG_ON(!spool->count); 96 spool->count--; 97 unlock_or_release_subpool(spool); 98 } 99 100 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 101 long delta) 102 { 103 int ret = 0; 104 105 if (!spool) 106 return 0; 107 108 spin_lock(&spool->lock); 109 if ((spool->used_hpages + delta) <= spool->max_hpages) { 110 spool->used_hpages += delta; 111 } else { 112 ret = -ENOMEM; 113 } 114 spin_unlock(&spool->lock); 115 116 return ret; 117 } 118 119 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 120 long delta) 121 { 122 if (!spool) 123 return; 124 125 spin_lock(&spool->lock); 126 spool->used_hpages -= delta; 127 /* If hugetlbfs_put_super couldn't free spool due to 128 * an outstanding quota reference, free it now. */ 129 unlock_or_release_subpool(spool); 130 } 131 132 static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 133 { 134 return HUGETLBFS_SB(inode->i_sb)->spool; 135 } 136 137 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 138 { 139 return subpool_inode(file_inode(vma->vm_file)); 140 } 141 142 /* 143 * Region tracking -- allows tracking of reservations and instantiated pages 144 * across the pages in a mapping. 145 * 146 * The region data structures are embedded into a resv_map and 147 * protected by a resv_map's lock 148 */ 149 struct file_region { 150 struct list_head link; 151 long from; 152 long to; 153 }; 154 155 static long region_add(struct resv_map *resv, long f, long t) 156 { 157 struct list_head *head = &resv->regions; 158 struct file_region *rg, *nrg, *trg; 159 160 spin_lock(&resv->lock); 161 /* Locate the region we are either in or before. */ 162 list_for_each_entry(rg, head, link) 163 if (f <= rg->to) 164 break; 165 166 /* Round our left edge to the current segment if it encloses us. */ 167 if (f > rg->from) 168 f = rg->from; 169 170 /* Check for and consume any regions we now overlap with. */ 171 nrg = rg; 172 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 173 if (&rg->link == head) 174 break; 175 if (rg->from > t) 176 break; 177 178 /* If this area reaches higher then extend our area to 179 * include it completely. If this is not the first area 180 * which we intend to reuse, free it. */ 181 if (rg->to > t) 182 t = rg->to; 183 if (rg != nrg) { 184 list_del(&rg->link); 185 kfree(rg); 186 } 187 } 188 nrg->from = f; 189 nrg->to = t; 190 spin_unlock(&resv->lock); 191 return 0; 192 } 193 194 static long region_chg(struct resv_map *resv, long f, long t) 195 { 196 struct list_head *head = &resv->regions; 197 struct file_region *rg, *nrg = NULL; 198 long chg = 0; 199 200 retry: 201 spin_lock(&resv->lock); 202 /* Locate the region we are before or in. */ 203 list_for_each_entry(rg, head, link) 204 if (f <= rg->to) 205 break; 206 207 /* If we are below the current region then a new region is required. 208 * Subtle, allocate a new region at the position but make it zero 209 * size such that we can guarantee to record the reservation. */ 210 if (&rg->link == head || t < rg->from) { 211 if (!nrg) { 212 spin_unlock(&resv->lock); 213 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 214 if (!nrg) 215 return -ENOMEM; 216 217 nrg->from = f; 218 nrg->to = f; 219 INIT_LIST_HEAD(&nrg->link); 220 goto retry; 221 } 222 223 list_add(&nrg->link, rg->link.prev); 224 chg = t - f; 225 goto out_nrg; 226 } 227 228 /* Round our left edge to the current segment if it encloses us. */ 229 if (f > rg->from) 230 f = rg->from; 231 chg = t - f; 232 233 /* Check for and consume any regions we now overlap with. */ 234 list_for_each_entry(rg, rg->link.prev, link) { 235 if (&rg->link == head) 236 break; 237 if (rg->from > t) 238 goto out; 239 240 /* We overlap with this area, if it extends further than 241 * us then we must extend ourselves. Account for its 242 * existing reservation. */ 243 if (rg->to > t) { 244 chg += rg->to - t; 245 t = rg->to; 246 } 247 chg -= rg->to - rg->from; 248 } 249 250 out: 251 spin_unlock(&resv->lock); 252 /* We already know we raced and no longer need the new region */ 253 kfree(nrg); 254 return chg; 255 out_nrg: 256 spin_unlock(&resv->lock); 257 return chg; 258 } 259 260 static long region_truncate(struct resv_map *resv, long end) 261 { 262 struct list_head *head = &resv->regions; 263 struct file_region *rg, *trg; 264 long chg = 0; 265 266 spin_lock(&resv->lock); 267 /* Locate the region we are either in or before. */ 268 list_for_each_entry(rg, head, link) 269 if (end <= rg->to) 270 break; 271 if (&rg->link == head) 272 goto out; 273 274 /* If we are in the middle of a region then adjust it. */ 275 if (end > rg->from) { 276 chg = rg->to - end; 277 rg->to = end; 278 rg = list_entry(rg->link.next, typeof(*rg), link); 279 } 280 281 /* Drop any remaining regions. */ 282 list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 283 if (&rg->link == head) 284 break; 285 chg += rg->to - rg->from; 286 list_del(&rg->link); 287 kfree(rg); 288 } 289 290 out: 291 spin_unlock(&resv->lock); 292 return chg; 293 } 294 295 static long region_count(struct resv_map *resv, long f, long t) 296 { 297 struct list_head *head = &resv->regions; 298 struct file_region *rg; 299 long chg = 0; 300 301 spin_lock(&resv->lock); 302 /* Locate each segment we overlap with, and count that overlap. */ 303 list_for_each_entry(rg, head, link) { 304 long seg_from; 305 long seg_to; 306 307 if (rg->to <= f) 308 continue; 309 if (rg->from >= t) 310 break; 311 312 seg_from = max(rg->from, f); 313 seg_to = min(rg->to, t); 314 315 chg += seg_to - seg_from; 316 } 317 spin_unlock(&resv->lock); 318 319 return chg; 320 } 321 322 /* 323 * Convert the address within this vma to the page offset within 324 * the mapping, in pagecache page units; huge pages here. 325 */ 326 static pgoff_t vma_hugecache_offset(struct hstate *h, 327 struct vm_area_struct *vma, unsigned long address) 328 { 329 return ((address - vma->vm_start) >> huge_page_shift(h)) + 330 (vma->vm_pgoff >> huge_page_order(h)); 331 } 332 333 pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 334 unsigned long address) 335 { 336 return vma_hugecache_offset(hstate_vma(vma), vma, address); 337 } 338 339 /* 340 * Return the size of the pages allocated when backing a VMA. In the majority 341 * cases this will be same size as used by the page table entries. 342 */ 343 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 344 { 345 struct hstate *hstate; 346 347 if (!is_vm_hugetlb_page(vma)) 348 return PAGE_SIZE; 349 350 hstate = hstate_vma(vma); 351 352 return 1UL << huge_page_shift(hstate); 353 } 354 EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 355 356 /* 357 * Return the page size being used by the MMU to back a VMA. In the majority 358 * of cases, the page size used by the kernel matches the MMU size. On 359 * architectures where it differs, an architecture-specific version of this 360 * function is required. 361 */ 362 #ifndef vma_mmu_pagesize 363 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 364 { 365 return vma_kernel_pagesize(vma); 366 } 367 #endif 368 369 /* 370 * Flags for MAP_PRIVATE reservations. These are stored in the bottom 371 * bits of the reservation map pointer, which are always clear due to 372 * alignment. 373 */ 374 #define HPAGE_RESV_OWNER (1UL << 0) 375 #define HPAGE_RESV_UNMAPPED (1UL << 1) 376 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 377 378 /* 379 * These helpers are used to track how many pages are reserved for 380 * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 381 * is guaranteed to have their future faults succeed. 382 * 383 * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 384 * the reserve counters are updated with the hugetlb_lock held. It is safe 385 * to reset the VMA at fork() time as it is not in use yet and there is no 386 * chance of the global counters getting corrupted as a result of the values. 387 * 388 * The private mapping reservation is represented in a subtly different 389 * manner to a shared mapping. A shared mapping has a region map associated 390 * with the underlying file, this region map represents the backing file 391 * pages which have ever had a reservation assigned which this persists even 392 * after the page is instantiated. A private mapping has a region map 393 * associated with the original mmap which is attached to all VMAs which 394 * reference it, this region map represents those offsets which have consumed 395 * reservation ie. where pages have been instantiated. 396 */ 397 static unsigned long get_vma_private_data(struct vm_area_struct *vma) 398 { 399 return (unsigned long)vma->vm_private_data; 400 } 401 402 static void set_vma_private_data(struct vm_area_struct *vma, 403 unsigned long value) 404 { 405 vma->vm_private_data = (void *)value; 406 } 407 408 struct resv_map *resv_map_alloc(void) 409 { 410 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 411 if (!resv_map) 412 return NULL; 413 414 kref_init(&resv_map->refs); 415 spin_lock_init(&resv_map->lock); 416 INIT_LIST_HEAD(&resv_map->regions); 417 418 return resv_map; 419 } 420 421 void resv_map_release(struct kref *ref) 422 { 423 struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 424 425 /* Clear out any active regions before we release the map. */ 426 region_truncate(resv_map, 0); 427 kfree(resv_map); 428 } 429 430 static inline struct resv_map *inode_resv_map(struct inode *inode) 431 { 432 return inode->i_mapping->private_data; 433 } 434 435 static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 436 { 437 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 438 if (vma->vm_flags & VM_MAYSHARE) { 439 struct address_space *mapping = vma->vm_file->f_mapping; 440 struct inode *inode = mapping->host; 441 442 return inode_resv_map(inode); 443 444 } else { 445 return (struct resv_map *)(get_vma_private_data(vma) & 446 ~HPAGE_RESV_MASK); 447 } 448 } 449 450 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 451 { 452 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 453 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 454 455 set_vma_private_data(vma, (get_vma_private_data(vma) & 456 HPAGE_RESV_MASK) | (unsigned long)map); 457 } 458 459 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 460 { 461 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 462 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 463 464 set_vma_private_data(vma, get_vma_private_data(vma) | flags); 465 } 466 467 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 468 { 469 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 470 471 return (get_vma_private_data(vma) & flag) != 0; 472 } 473 474 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 475 void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 476 { 477 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 478 if (!(vma->vm_flags & VM_MAYSHARE)) 479 vma->vm_private_data = (void *)0; 480 } 481 482 /* Returns true if the VMA has associated reserve pages */ 483 static int vma_has_reserves(struct vm_area_struct *vma, long chg) 484 { 485 if (vma->vm_flags & VM_NORESERVE) { 486 /* 487 * This address is already reserved by other process(chg == 0), 488 * so, we should decrement reserved count. Without decrementing, 489 * reserve count remains after releasing inode, because this 490 * allocated page will go into page cache and is regarded as 491 * coming from reserved pool in releasing step. Currently, we 492 * don't have any other solution to deal with this situation 493 * properly, so add work-around here. 494 */ 495 if (vma->vm_flags & VM_MAYSHARE && chg == 0) 496 return 1; 497 else 498 return 0; 499 } 500 501 /* Shared mappings always use reserves */ 502 if (vma->vm_flags & VM_MAYSHARE) 503 return 1; 504 505 /* 506 * Only the process that called mmap() has reserves for 507 * private mappings. 508 */ 509 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 510 return 1; 511 512 return 0; 513 } 514 515 static void enqueue_huge_page(struct hstate *h, struct page *page) 516 { 517 int nid = page_to_nid(page); 518 list_move(&page->lru, &h->hugepage_freelists[nid]); 519 h->free_huge_pages++; 520 h->free_huge_pages_node[nid]++; 521 } 522 523 static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 524 { 525 struct page *page; 526 527 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 528 if (!is_migrate_isolate_page(page)) 529 break; 530 /* 531 * if 'non-isolated free hugepage' not found on the list, 532 * the allocation fails. 533 */ 534 if (&h->hugepage_freelists[nid] == &page->lru) 535 return NULL; 536 list_move(&page->lru, &h->hugepage_activelist); 537 set_page_refcounted(page); 538 h->free_huge_pages--; 539 h->free_huge_pages_node[nid]--; 540 return page; 541 } 542 543 /* Movability of hugepages depends on migration support. */ 544 static inline gfp_t htlb_alloc_mask(struct hstate *h) 545 { 546 if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 547 return GFP_HIGHUSER_MOVABLE; 548 else 549 return GFP_HIGHUSER; 550 } 551 552 static struct page *dequeue_huge_page_vma(struct hstate *h, 553 struct vm_area_struct *vma, 554 unsigned long address, int avoid_reserve, 555 long chg) 556 { 557 struct page *page = NULL; 558 struct mempolicy *mpol; 559 nodemask_t *nodemask; 560 struct zonelist *zonelist; 561 struct zone *zone; 562 struct zoneref *z; 563 unsigned int cpuset_mems_cookie; 564 565 /* 566 * A child process with MAP_PRIVATE mappings created by their parent 567 * have no page reserves. This check ensures that reservations are 568 * not "stolen". The child may still get SIGKILLed 569 */ 570 if (!vma_has_reserves(vma, chg) && 571 h->free_huge_pages - h->resv_huge_pages == 0) 572 goto err; 573 574 /* If reserves cannot be used, ensure enough pages are in the pool */ 575 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 576 goto err; 577 578 retry_cpuset: 579 cpuset_mems_cookie = read_mems_allowed_begin(); 580 zonelist = huge_zonelist(vma, address, 581 htlb_alloc_mask(h), &mpol, &nodemask); 582 583 for_each_zone_zonelist_nodemask(zone, z, zonelist, 584 MAX_NR_ZONES - 1, nodemask) { 585 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { 586 page = dequeue_huge_page_node(h, zone_to_nid(zone)); 587 if (page) { 588 if (avoid_reserve) 589 break; 590 if (!vma_has_reserves(vma, chg)) 591 break; 592 593 SetPagePrivate(page); 594 h->resv_huge_pages--; 595 break; 596 } 597 } 598 } 599 600 mpol_cond_put(mpol); 601 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 602 goto retry_cpuset; 603 return page; 604 605 err: 606 return NULL; 607 } 608 609 /* 610 * common helper functions for hstate_next_node_to_{alloc|free}. 611 * We may have allocated or freed a huge page based on a different 612 * nodes_allowed previously, so h->next_node_to_{alloc|free} might 613 * be outside of *nodes_allowed. Ensure that we use an allowed 614 * node for alloc or free. 615 */ 616 static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 617 { 618 nid = next_node(nid, *nodes_allowed); 619 if (nid == MAX_NUMNODES) 620 nid = first_node(*nodes_allowed); 621 VM_BUG_ON(nid >= MAX_NUMNODES); 622 623 return nid; 624 } 625 626 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 627 { 628 if (!node_isset(nid, *nodes_allowed)) 629 nid = next_node_allowed(nid, nodes_allowed); 630 return nid; 631 } 632 633 /* 634 * returns the previously saved node ["this node"] from which to 635 * allocate a persistent huge page for the pool and advance the 636 * next node from which to allocate, handling wrap at end of node 637 * mask. 638 */ 639 static int hstate_next_node_to_alloc(struct hstate *h, 640 nodemask_t *nodes_allowed) 641 { 642 int nid; 643 644 VM_BUG_ON(!nodes_allowed); 645 646 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 647 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 648 649 return nid; 650 } 651 652 /* 653 * helper for free_pool_huge_page() - return the previously saved 654 * node ["this node"] from which to free a huge page. Advance the 655 * next node id whether or not we find a free huge page to free so 656 * that the next attempt to free addresses the next node. 657 */ 658 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 659 { 660 int nid; 661 662 VM_BUG_ON(!nodes_allowed); 663 664 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 665 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 666 667 return nid; 668 } 669 670 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 671 for (nr_nodes = nodes_weight(*mask); \ 672 nr_nodes > 0 && \ 673 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 674 nr_nodes--) 675 676 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 677 for (nr_nodes = nodes_weight(*mask); \ 678 nr_nodes > 0 && \ 679 ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 680 nr_nodes--) 681 682 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) 683 static void destroy_compound_gigantic_page(struct page *page, 684 unsigned long order) 685 { 686 int i; 687 int nr_pages = 1 << order; 688 struct page *p = page + 1; 689 690 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 691 __ClearPageTail(p); 692 set_page_refcounted(p); 693 p->first_page = NULL; 694 } 695 696 set_compound_order(page, 0); 697 __ClearPageHead(page); 698 } 699 700 static void free_gigantic_page(struct page *page, unsigned order) 701 { 702 free_contig_range(page_to_pfn(page), 1 << order); 703 } 704 705 static int __alloc_gigantic_page(unsigned long start_pfn, 706 unsigned long nr_pages) 707 { 708 unsigned long end_pfn = start_pfn + nr_pages; 709 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 710 } 711 712 static bool pfn_range_valid_gigantic(unsigned long start_pfn, 713 unsigned long nr_pages) 714 { 715 unsigned long i, end_pfn = start_pfn + nr_pages; 716 struct page *page; 717 718 for (i = start_pfn; i < end_pfn; i++) { 719 if (!pfn_valid(i)) 720 return false; 721 722 page = pfn_to_page(i); 723 724 if (PageReserved(page)) 725 return false; 726 727 if (page_count(page) > 0) 728 return false; 729 730 if (PageHuge(page)) 731 return false; 732 } 733 734 return true; 735 } 736 737 static bool zone_spans_last_pfn(const struct zone *zone, 738 unsigned long start_pfn, unsigned long nr_pages) 739 { 740 unsigned long last_pfn = start_pfn + nr_pages - 1; 741 return zone_spans_pfn(zone, last_pfn); 742 } 743 744 static struct page *alloc_gigantic_page(int nid, unsigned order) 745 { 746 unsigned long nr_pages = 1 << order; 747 unsigned long ret, pfn, flags; 748 struct zone *z; 749 750 z = NODE_DATA(nid)->node_zones; 751 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 752 spin_lock_irqsave(&z->lock, flags); 753 754 pfn = ALIGN(z->zone_start_pfn, nr_pages); 755 while (zone_spans_last_pfn(z, pfn, nr_pages)) { 756 if (pfn_range_valid_gigantic(pfn, nr_pages)) { 757 /* 758 * We release the zone lock here because 759 * alloc_contig_range() will also lock the zone 760 * at some point. If there's an allocation 761 * spinning on this lock, it may win the race 762 * and cause alloc_contig_range() to fail... 763 */ 764 spin_unlock_irqrestore(&z->lock, flags); 765 ret = __alloc_gigantic_page(pfn, nr_pages); 766 if (!ret) 767 return pfn_to_page(pfn); 768 spin_lock_irqsave(&z->lock, flags); 769 } 770 pfn += nr_pages; 771 } 772 773 spin_unlock_irqrestore(&z->lock, flags); 774 } 775 776 return NULL; 777 } 778 779 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 780 static void prep_compound_gigantic_page(struct page *page, unsigned long order); 781 782 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 783 { 784 struct page *page; 785 786 page = alloc_gigantic_page(nid, huge_page_order(h)); 787 if (page) { 788 prep_compound_gigantic_page(page, huge_page_order(h)); 789 prep_new_huge_page(h, page, nid); 790 } 791 792 return page; 793 } 794 795 static int alloc_fresh_gigantic_page(struct hstate *h, 796 nodemask_t *nodes_allowed) 797 { 798 struct page *page = NULL; 799 int nr_nodes, node; 800 801 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 802 page = alloc_fresh_gigantic_page_node(h, node); 803 if (page) 804 return 1; 805 } 806 807 return 0; 808 } 809 810 static inline bool gigantic_page_supported(void) { return true; } 811 #else 812 static inline bool gigantic_page_supported(void) { return false; } 813 static inline void free_gigantic_page(struct page *page, unsigned order) { } 814 static inline void destroy_compound_gigantic_page(struct page *page, 815 unsigned long order) { } 816 static inline int alloc_fresh_gigantic_page(struct hstate *h, 817 nodemask_t *nodes_allowed) { return 0; } 818 #endif 819 820 static void update_and_free_page(struct hstate *h, struct page *page) 821 { 822 int i; 823 824 if (hstate_is_gigantic(h) && !gigantic_page_supported()) 825 return; 826 827 h->nr_huge_pages--; 828 h->nr_huge_pages_node[page_to_nid(page)]--; 829 for (i = 0; i < pages_per_huge_page(h); i++) { 830 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 831 1 << PG_referenced | 1 << PG_dirty | 832 1 << PG_active | 1 << PG_private | 833 1 << PG_writeback); 834 } 835 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 836 set_compound_page_dtor(page, NULL); 837 set_page_refcounted(page); 838 if (hstate_is_gigantic(h)) { 839 destroy_compound_gigantic_page(page, huge_page_order(h)); 840 free_gigantic_page(page, huge_page_order(h)); 841 } else { 842 arch_release_hugepage(page); 843 __free_pages(page, huge_page_order(h)); 844 } 845 } 846 847 struct hstate *size_to_hstate(unsigned long size) 848 { 849 struct hstate *h; 850 851 for_each_hstate(h) { 852 if (huge_page_size(h) == size) 853 return h; 854 } 855 return NULL; 856 } 857 858 void free_huge_page(struct page *page) 859 { 860 /* 861 * Can't pass hstate in here because it is called from the 862 * compound page destructor. 863 */ 864 struct hstate *h = page_hstate(page); 865 int nid = page_to_nid(page); 866 struct hugepage_subpool *spool = 867 (struct hugepage_subpool *)page_private(page); 868 bool restore_reserve; 869 870 set_page_private(page, 0); 871 page->mapping = NULL; 872 BUG_ON(page_count(page)); 873 BUG_ON(page_mapcount(page)); 874 restore_reserve = PagePrivate(page); 875 ClearPagePrivate(page); 876 877 spin_lock(&hugetlb_lock); 878 hugetlb_cgroup_uncharge_page(hstate_index(h), 879 pages_per_huge_page(h), page); 880 if (restore_reserve) 881 h->resv_huge_pages++; 882 883 if (h->surplus_huge_pages_node[nid]) { 884 /* remove the page from active list */ 885 list_del(&page->lru); 886 update_and_free_page(h, page); 887 h->surplus_huge_pages--; 888 h->surplus_huge_pages_node[nid]--; 889 } else { 890 arch_clear_hugepage_flags(page); 891 enqueue_huge_page(h, page); 892 } 893 spin_unlock(&hugetlb_lock); 894 hugepage_subpool_put_pages(spool, 1); 895 } 896 897 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 898 { 899 INIT_LIST_HEAD(&page->lru); 900 set_compound_page_dtor(page, free_huge_page); 901 spin_lock(&hugetlb_lock); 902 set_hugetlb_cgroup(page, NULL); 903 h->nr_huge_pages++; 904 h->nr_huge_pages_node[nid]++; 905 spin_unlock(&hugetlb_lock); 906 put_page(page); /* free it into the hugepage allocator */ 907 } 908 909 static void prep_compound_gigantic_page(struct page *page, unsigned long order) 910 { 911 int i; 912 int nr_pages = 1 << order; 913 struct page *p = page + 1; 914 915 /* we rely on prep_new_huge_page to set the destructor */ 916 set_compound_order(page, order); 917 __SetPageHead(page); 918 __ClearPageReserved(page); 919 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 920 __SetPageTail(p); 921 /* 922 * For gigantic hugepages allocated through bootmem at 923 * boot, it's safer to be consistent with the not-gigantic 924 * hugepages and clear the PG_reserved bit from all tail pages 925 * too. Otherwse drivers using get_user_pages() to access tail 926 * pages may get the reference counting wrong if they see 927 * PG_reserved set on a tail page (despite the head page not 928 * having PG_reserved set). Enforcing this consistency between 929 * head and tail pages allows drivers to optimize away a check 930 * on the head page when they need know if put_page() is needed 931 * after get_user_pages(). 932 */ 933 __ClearPageReserved(p); 934 set_page_count(p, 0); 935 p->first_page = page; 936 } 937 } 938 939 /* 940 * PageHuge() only returns true for hugetlbfs pages, but not for normal or 941 * transparent huge pages. See the PageTransHuge() documentation for more 942 * details. 943 */ 944 int PageHuge(struct page *page) 945 { 946 if (!PageCompound(page)) 947 return 0; 948 949 page = compound_head(page); 950 return get_compound_page_dtor(page) == free_huge_page; 951 } 952 EXPORT_SYMBOL_GPL(PageHuge); 953 954 /* 955 * PageHeadHuge() only returns true for hugetlbfs head page, but not for 956 * normal or transparent huge pages. 957 */ 958 int PageHeadHuge(struct page *page_head) 959 { 960 if (!PageHead(page_head)) 961 return 0; 962 963 return get_compound_page_dtor(page_head) == free_huge_page; 964 } 965 966 pgoff_t __basepage_index(struct page *page) 967 { 968 struct page *page_head = compound_head(page); 969 pgoff_t index = page_index(page_head); 970 unsigned long compound_idx; 971 972 if (!PageHuge(page_head)) 973 return page_index(page); 974 975 if (compound_order(page_head) >= MAX_ORDER) 976 compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 977 else 978 compound_idx = page - page_head; 979 980 return (index << compound_order(page_head)) + compound_idx; 981 } 982 983 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 984 { 985 struct page *page; 986 987 page = alloc_pages_exact_node(nid, 988 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 989 __GFP_REPEAT|__GFP_NOWARN, 990 huge_page_order(h)); 991 if (page) { 992 if (arch_prepare_hugepage(page)) { 993 __free_pages(page, huge_page_order(h)); 994 return NULL; 995 } 996 prep_new_huge_page(h, page, nid); 997 } 998 999 return page; 1000 } 1001 1002 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1003 { 1004 struct page *page; 1005 int nr_nodes, node; 1006 int ret = 0; 1007 1008 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1009 page = alloc_fresh_huge_page_node(h, node); 1010 if (page) { 1011 ret = 1; 1012 break; 1013 } 1014 } 1015 1016 if (ret) 1017 count_vm_event(HTLB_BUDDY_PGALLOC); 1018 else 1019 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1020 1021 return ret; 1022 } 1023 1024 /* 1025 * Free huge page from pool from next node to free. 1026 * Attempt to keep persistent huge pages more or less 1027 * balanced over allowed nodes. 1028 * Called with hugetlb_lock locked. 1029 */ 1030 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 1031 bool acct_surplus) 1032 { 1033 int nr_nodes, node; 1034 int ret = 0; 1035 1036 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1037 /* 1038 * If we're returning unused surplus pages, only examine 1039 * nodes with surplus pages. 1040 */ 1041 if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1042 !list_empty(&h->hugepage_freelists[node])) { 1043 struct page *page = 1044 list_entry(h->hugepage_freelists[node].next, 1045 struct page, lru); 1046 list_del(&page->lru); 1047 h->free_huge_pages--; 1048 h->free_huge_pages_node[node]--; 1049 if (acct_surplus) { 1050 h->surplus_huge_pages--; 1051 h->surplus_huge_pages_node[node]--; 1052 } 1053 update_and_free_page(h, page); 1054 ret = 1; 1055 break; 1056 } 1057 } 1058 1059 return ret; 1060 } 1061 1062 /* 1063 * Dissolve a given free hugepage into free buddy pages. This function does 1064 * nothing for in-use (including surplus) hugepages. 1065 */ 1066 static void dissolve_free_huge_page(struct page *page) 1067 { 1068 spin_lock(&hugetlb_lock); 1069 if (PageHuge(page) && !page_count(page)) { 1070 struct hstate *h = page_hstate(page); 1071 int nid = page_to_nid(page); 1072 list_del(&page->lru); 1073 h->free_huge_pages--; 1074 h->free_huge_pages_node[nid]--; 1075 update_and_free_page(h, page); 1076 } 1077 spin_unlock(&hugetlb_lock); 1078 } 1079 1080 /* 1081 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1082 * make specified memory blocks removable from the system. 1083 * Note that start_pfn should aligned with (minimum) hugepage size. 1084 */ 1085 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1086 { 1087 unsigned int order = 8 * sizeof(void *); 1088 unsigned long pfn; 1089 struct hstate *h; 1090 1091 if (!hugepages_supported()) 1092 return; 1093 1094 /* Set scan step to minimum hugepage size */ 1095 for_each_hstate(h) 1096 if (order > huge_page_order(h)) 1097 order = huge_page_order(h); 1098 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); 1099 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) 1100 dissolve_free_huge_page(pfn_to_page(pfn)); 1101 } 1102 1103 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 1104 { 1105 struct page *page; 1106 unsigned int r_nid; 1107 1108 if (hstate_is_gigantic(h)) 1109 return NULL; 1110 1111 /* 1112 * Assume we will successfully allocate the surplus page to 1113 * prevent racing processes from causing the surplus to exceed 1114 * overcommit 1115 * 1116 * This however introduces a different race, where a process B 1117 * tries to grow the static hugepage pool while alloc_pages() is 1118 * called by process A. B will only examine the per-node 1119 * counters in determining if surplus huge pages can be 1120 * converted to normal huge pages in adjust_pool_surplus(). A 1121 * won't be able to increment the per-node counter, until the 1122 * lock is dropped by B, but B doesn't drop hugetlb_lock until 1123 * no more huge pages can be converted from surplus to normal 1124 * state (and doesn't try to convert again). Thus, we have a 1125 * case where a surplus huge page exists, the pool is grown, and 1126 * the surplus huge page still exists after, even though it 1127 * should just have been converted to a normal huge page. This 1128 * does not leak memory, though, as the hugepage will be freed 1129 * once it is out of use. It also does not allow the counters to 1130 * go out of whack in adjust_pool_surplus() as we don't modify 1131 * the node values until we've gotten the hugepage and only the 1132 * per-node value is checked there. 1133 */ 1134 spin_lock(&hugetlb_lock); 1135 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1136 spin_unlock(&hugetlb_lock); 1137 return NULL; 1138 } else { 1139 h->nr_huge_pages++; 1140 h->surplus_huge_pages++; 1141 } 1142 spin_unlock(&hugetlb_lock); 1143 1144 if (nid == NUMA_NO_NODE) 1145 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| 1146 __GFP_REPEAT|__GFP_NOWARN, 1147 huge_page_order(h)); 1148 else 1149 page = alloc_pages_exact_node(nid, 1150 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1151 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 1152 1153 if (page && arch_prepare_hugepage(page)) { 1154 __free_pages(page, huge_page_order(h)); 1155 page = NULL; 1156 } 1157 1158 spin_lock(&hugetlb_lock); 1159 if (page) { 1160 INIT_LIST_HEAD(&page->lru); 1161 r_nid = page_to_nid(page); 1162 set_compound_page_dtor(page, free_huge_page); 1163 set_hugetlb_cgroup(page, NULL); 1164 /* 1165 * We incremented the global counters already 1166 */ 1167 h->nr_huge_pages_node[r_nid]++; 1168 h->surplus_huge_pages_node[r_nid]++; 1169 __count_vm_event(HTLB_BUDDY_PGALLOC); 1170 } else { 1171 h->nr_huge_pages--; 1172 h->surplus_huge_pages--; 1173 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1174 } 1175 spin_unlock(&hugetlb_lock); 1176 1177 return page; 1178 } 1179 1180 /* 1181 * This allocation function is useful in the context where vma is irrelevant. 1182 * E.g. soft-offlining uses this function because it only cares physical 1183 * address of error page. 1184 */ 1185 struct page *alloc_huge_page_node(struct hstate *h, int nid) 1186 { 1187 struct page *page = NULL; 1188 1189 spin_lock(&hugetlb_lock); 1190 if (h->free_huge_pages - h->resv_huge_pages > 0) 1191 page = dequeue_huge_page_node(h, nid); 1192 spin_unlock(&hugetlb_lock); 1193 1194 if (!page) 1195 page = alloc_buddy_huge_page(h, nid); 1196 1197 return page; 1198 } 1199 1200 /* 1201 * Increase the hugetlb pool such that it can accommodate a reservation 1202 * of size 'delta'. 1203 */ 1204 static int gather_surplus_pages(struct hstate *h, int delta) 1205 { 1206 struct list_head surplus_list; 1207 struct page *page, *tmp; 1208 int ret, i; 1209 int needed, allocated; 1210 bool alloc_ok = true; 1211 1212 needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1213 if (needed <= 0) { 1214 h->resv_huge_pages += delta; 1215 return 0; 1216 } 1217 1218 allocated = 0; 1219 INIT_LIST_HEAD(&surplus_list); 1220 1221 ret = -ENOMEM; 1222 retry: 1223 spin_unlock(&hugetlb_lock); 1224 for (i = 0; i < needed; i++) { 1225 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1226 if (!page) { 1227 alloc_ok = false; 1228 break; 1229 } 1230 list_add(&page->lru, &surplus_list); 1231 } 1232 allocated += i; 1233 1234 /* 1235 * After retaking hugetlb_lock, we need to recalculate 'needed' 1236 * because either resv_huge_pages or free_huge_pages may have changed. 1237 */ 1238 spin_lock(&hugetlb_lock); 1239 needed = (h->resv_huge_pages + delta) - 1240 (h->free_huge_pages + allocated); 1241 if (needed > 0) { 1242 if (alloc_ok) 1243 goto retry; 1244 /* 1245 * We were not able to allocate enough pages to 1246 * satisfy the entire reservation so we free what 1247 * we've allocated so far. 1248 */ 1249 goto free; 1250 } 1251 /* 1252 * The surplus_list now contains _at_least_ the number of extra pages 1253 * needed to accommodate the reservation. Add the appropriate number 1254 * of pages to the hugetlb pool and free the extras back to the buddy 1255 * allocator. Commit the entire reservation here to prevent another 1256 * process from stealing the pages as they are added to the pool but 1257 * before they are reserved. 1258 */ 1259 needed += allocated; 1260 h->resv_huge_pages += delta; 1261 ret = 0; 1262 1263 /* Free the needed pages to the hugetlb pool */ 1264 list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 1265 if ((--needed) < 0) 1266 break; 1267 /* 1268 * This page is now managed by the hugetlb allocator and has 1269 * no users -- drop the buddy allocator's reference. 1270 */ 1271 put_page_testzero(page); 1272 VM_BUG_ON_PAGE(page_count(page), page); 1273 enqueue_huge_page(h, page); 1274 } 1275 free: 1276 spin_unlock(&hugetlb_lock); 1277 1278 /* Free unnecessary surplus pages to the buddy allocator */ 1279 list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1280 put_page(page); 1281 spin_lock(&hugetlb_lock); 1282 1283 return ret; 1284 } 1285 1286 /* 1287 * When releasing a hugetlb pool reservation, any surplus pages that were 1288 * allocated to satisfy the reservation must be explicitly freed if they were 1289 * never used. 1290 * Called with hugetlb_lock held. 1291 */ 1292 static void return_unused_surplus_pages(struct hstate *h, 1293 unsigned long unused_resv_pages) 1294 { 1295 unsigned long nr_pages; 1296 1297 /* Uncommit the reservation */ 1298 h->resv_huge_pages -= unused_resv_pages; 1299 1300 /* Cannot return gigantic pages currently */ 1301 if (hstate_is_gigantic(h)) 1302 return; 1303 1304 nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1305 1306 /* 1307 * We want to release as many surplus pages as possible, spread 1308 * evenly across all nodes with memory. Iterate across these nodes 1309 * until we can no longer free unreserved surplus pages. This occurs 1310 * when the nodes with surplus pages have no free pages. 1311 * free_pool_huge_page() will balance the the freed pages across the 1312 * on-line nodes with memory and will handle the hstate accounting. 1313 */ 1314 while (nr_pages--) { 1315 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1316 break; 1317 cond_resched_lock(&hugetlb_lock); 1318 } 1319 } 1320 1321 /* 1322 * Determine if the huge page at addr within the vma has an associated 1323 * reservation. Where it does not we will need to logically increase 1324 * reservation and actually increase subpool usage before an allocation 1325 * can occur. Where any new reservation would be required the 1326 * reservation change is prepared, but not committed. Once the page 1327 * has been allocated from the subpool and instantiated the change should 1328 * be committed via vma_commit_reservation. No action is required on 1329 * failure. 1330 */ 1331 static long vma_needs_reservation(struct hstate *h, 1332 struct vm_area_struct *vma, unsigned long addr) 1333 { 1334 struct resv_map *resv; 1335 pgoff_t idx; 1336 long chg; 1337 1338 resv = vma_resv_map(vma); 1339 if (!resv) 1340 return 1; 1341 1342 idx = vma_hugecache_offset(h, vma, addr); 1343 chg = region_chg(resv, idx, idx + 1); 1344 1345 if (vma->vm_flags & VM_MAYSHARE) 1346 return chg; 1347 else 1348 return chg < 0 ? chg : 0; 1349 } 1350 static void vma_commit_reservation(struct hstate *h, 1351 struct vm_area_struct *vma, unsigned long addr) 1352 { 1353 struct resv_map *resv; 1354 pgoff_t idx; 1355 1356 resv = vma_resv_map(vma); 1357 if (!resv) 1358 return; 1359 1360 idx = vma_hugecache_offset(h, vma, addr); 1361 region_add(resv, idx, idx + 1); 1362 } 1363 1364 static struct page *alloc_huge_page(struct vm_area_struct *vma, 1365 unsigned long addr, int avoid_reserve) 1366 { 1367 struct hugepage_subpool *spool = subpool_vma(vma); 1368 struct hstate *h = hstate_vma(vma); 1369 struct page *page; 1370 long chg; 1371 int ret, idx; 1372 struct hugetlb_cgroup *h_cg; 1373 1374 idx = hstate_index(h); 1375 /* 1376 * Processes that did not create the mapping will have no 1377 * reserves and will not have accounted against subpool 1378 * limit. Check that the subpool limit can be made before 1379 * satisfying the allocation MAP_NORESERVE mappings may also 1380 * need pages and subpool limit allocated allocated if no reserve 1381 * mapping overlaps. 1382 */ 1383 chg = vma_needs_reservation(h, vma, addr); 1384 if (chg < 0) 1385 return ERR_PTR(-ENOMEM); 1386 if (chg || avoid_reserve) 1387 if (hugepage_subpool_get_pages(spool, 1)) 1388 return ERR_PTR(-ENOSPC); 1389 1390 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 1391 if (ret) 1392 goto out_subpool_put; 1393 1394 spin_lock(&hugetlb_lock); 1395 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 1396 if (!page) { 1397 spin_unlock(&hugetlb_lock); 1398 page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1399 if (!page) 1400 goto out_uncharge_cgroup; 1401 1402 spin_lock(&hugetlb_lock); 1403 list_move(&page->lru, &h->hugepage_activelist); 1404 /* Fall through */ 1405 } 1406 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 1407 spin_unlock(&hugetlb_lock); 1408 1409 set_page_private(page, (unsigned long)spool); 1410 1411 vma_commit_reservation(h, vma, addr); 1412 return page; 1413 1414 out_uncharge_cgroup: 1415 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 1416 out_subpool_put: 1417 if (chg || avoid_reserve) 1418 hugepage_subpool_put_pages(spool, 1); 1419 return ERR_PTR(-ENOSPC); 1420 } 1421 1422 /* 1423 * alloc_huge_page()'s wrapper which simply returns the page if allocation 1424 * succeeds, otherwise NULL. This function is called from new_vma_page(), 1425 * where no ERR_VALUE is expected to be returned. 1426 */ 1427 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 1428 unsigned long addr, int avoid_reserve) 1429 { 1430 struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 1431 if (IS_ERR(page)) 1432 page = NULL; 1433 return page; 1434 } 1435 1436 int __weak alloc_bootmem_huge_page(struct hstate *h) 1437 { 1438 struct huge_bootmem_page *m; 1439 int nr_nodes, node; 1440 1441 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1442 void *addr; 1443 1444 addr = memblock_virt_alloc_try_nid_nopanic( 1445 huge_page_size(h), huge_page_size(h), 1446 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1447 if (addr) { 1448 /* 1449 * Use the beginning of the huge page to store the 1450 * huge_bootmem_page struct (until gather_bootmem 1451 * puts them into the mem_map). 1452 */ 1453 m = addr; 1454 goto found; 1455 } 1456 } 1457 return 0; 1458 1459 found: 1460 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); 1461 /* Put them into a private list first because mem_map is not up yet */ 1462 list_add(&m->list, &huge_boot_pages); 1463 m->hstate = h; 1464 return 1; 1465 } 1466 1467 static void __init prep_compound_huge_page(struct page *page, int order) 1468 { 1469 if (unlikely(order > (MAX_ORDER - 1))) 1470 prep_compound_gigantic_page(page, order); 1471 else 1472 prep_compound_page(page, order); 1473 } 1474 1475 /* Put bootmem huge pages into the standard lists after mem_map is up */ 1476 static void __init gather_bootmem_prealloc(void) 1477 { 1478 struct huge_bootmem_page *m; 1479 1480 list_for_each_entry(m, &huge_boot_pages, list) { 1481 struct hstate *h = m->hstate; 1482 struct page *page; 1483 1484 #ifdef CONFIG_HIGHMEM 1485 page = pfn_to_page(m->phys >> PAGE_SHIFT); 1486 memblock_free_late(__pa(m), 1487 sizeof(struct huge_bootmem_page)); 1488 #else 1489 page = virt_to_page(m); 1490 #endif 1491 WARN_ON(page_count(page) != 1); 1492 prep_compound_huge_page(page, h->order); 1493 WARN_ON(PageReserved(page)); 1494 prep_new_huge_page(h, page, page_to_nid(page)); 1495 /* 1496 * If we had gigantic hugepages allocated at boot time, we need 1497 * to restore the 'stolen' pages to totalram_pages in order to 1498 * fix confusing memory reports from free(1) and another 1499 * side-effects, like CommitLimit going negative. 1500 */ 1501 if (hstate_is_gigantic(h)) 1502 adjust_managed_page_count(page, 1 << h->order); 1503 } 1504 } 1505 1506 static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 1507 { 1508 unsigned long i; 1509 1510 for (i = 0; i < h->max_huge_pages; ++i) { 1511 if (hstate_is_gigantic(h)) { 1512 if (!alloc_bootmem_huge_page(h)) 1513 break; 1514 } else if (!alloc_fresh_huge_page(h, 1515 &node_states[N_MEMORY])) 1516 break; 1517 } 1518 h->max_huge_pages = i; 1519 } 1520 1521 static void __init hugetlb_init_hstates(void) 1522 { 1523 struct hstate *h; 1524 1525 for_each_hstate(h) { 1526 /* oversize hugepages were init'ed in early boot */ 1527 if (!hstate_is_gigantic(h)) 1528 hugetlb_hstate_alloc_pages(h); 1529 } 1530 } 1531 1532 static char * __init memfmt(char *buf, unsigned long n) 1533 { 1534 if (n >= (1UL << 30)) 1535 sprintf(buf, "%lu GB", n >> 30); 1536 else if (n >= (1UL << 20)) 1537 sprintf(buf, "%lu MB", n >> 20); 1538 else 1539 sprintf(buf, "%lu KB", n >> 10); 1540 return buf; 1541 } 1542 1543 static void __init report_hugepages(void) 1544 { 1545 struct hstate *h; 1546 1547 for_each_hstate(h) { 1548 char buf[32]; 1549 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 1550 memfmt(buf, huge_page_size(h)), 1551 h->free_huge_pages); 1552 } 1553 } 1554 1555 #ifdef CONFIG_HIGHMEM 1556 static void try_to_free_low(struct hstate *h, unsigned long count, 1557 nodemask_t *nodes_allowed) 1558 { 1559 int i; 1560 1561 if (hstate_is_gigantic(h)) 1562 return; 1563 1564 for_each_node_mask(i, *nodes_allowed) { 1565 struct page *page, *next; 1566 struct list_head *freel = &h->hugepage_freelists[i]; 1567 list_for_each_entry_safe(page, next, freel, lru) { 1568 if (count >= h->nr_huge_pages) 1569 return; 1570 if (PageHighMem(page)) 1571 continue; 1572 list_del(&page->lru); 1573 update_and_free_page(h, page); 1574 h->free_huge_pages--; 1575 h->free_huge_pages_node[page_to_nid(page)]--; 1576 } 1577 } 1578 } 1579 #else 1580 static inline void try_to_free_low(struct hstate *h, unsigned long count, 1581 nodemask_t *nodes_allowed) 1582 { 1583 } 1584 #endif 1585 1586 /* 1587 * Increment or decrement surplus_huge_pages. Keep node-specific counters 1588 * balanced by operating on them in a round-robin fashion. 1589 * Returns 1 if an adjustment was made. 1590 */ 1591 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 1592 int delta) 1593 { 1594 int nr_nodes, node; 1595 1596 VM_BUG_ON(delta != -1 && delta != 1); 1597 1598 if (delta < 0) { 1599 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1600 if (h->surplus_huge_pages_node[node]) 1601 goto found; 1602 } 1603 } else { 1604 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1605 if (h->surplus_huge_pages_node[node] < 1606 h->nr_huge_pages_node[node]) 1607 goto found; 1608 } 1609 } 1610 return 0; 1611 1612 found: 1613 h->surplus_huge_pages += delta; 1614 h->surplus_huge_pages_node[node] += delta; 1615 return 1; 1616 } 1617 1618 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 1619 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 1620 nodemask_t *nodes_allowed) 1621 { 1622 unsigned long min_count, ret; 1623 1624 if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1625 return h->max_huge_pages; 1626 1627 /* 1628 * Increase the pool size 1629 * First take pages out of surplus state. Then make up the 1630 * remaining difference by allocating fresh huge pages. 1631 * 1632 * We might race with alloc_buddy_huge_page() here and be unable 1633 * to convert a surplus huge page to a normal huge page. That is 1634 * not critical, though, it just means the overall size of the 1635 * pool might be one hugepage larger than it needs to be, but 1636 * within all the constraints specified by the sysctls. 1637 */ 1638 spin_lock(&hugetlb_lock); 1639 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 1640 if (!adjust_pool_surplus(h, nodes_allowed, -1)) 1641 break; 1642 } 1643 1644 while (count > persistent_huge_pages(h)) { 1645 /* 1646 * If this allocation races such that we no longer need the 1647 * page, free_huge_page will handle it by freeing the page 1648 * and reducing the surplus. 1649 */ 1650 spin_unlock(&hugetlb_lock); 1651 if (hstate_is_gigantic(h)) 1652 ret = alloc_fresh_gigantic_page(h, nodes_allowed); 1653 else 1654 ret = alloc_fresh_huge_page(h, nodes_allowed); 1655 spin_lock(&hugetlb_lock); 1656 if (!ret) 1657 goto out; 1658 1659 /* Bail for signals. Probably ctrl-c from user */ 1660 if (signal_pending(current)) 1661 goto out; 1662 } 1663 1664 /* 1665 * Decrease the pool size 1666 * First return free pages to the buddy allocator (being careful 1667 * to keep enough around to satisfy reservations). Then place 1668 * pages into surplus state as needed so the pool will shrink 1669 * to the desired size as pages become free. 1670 * 1671 * By placing pages into the surplus state independent of the 1672 * overcommit value, we are allowing the surplus pool size to 1673 * exceed overcommit. There are few sane options here. Since 1674 * alloc_buddy_huge_page() is checking the global counter, 1675 * though, we'll note that we're not allowed to exceed surplus 1676 * and won't grow the pool anywhere else. Not until one of the 1677 * sysctls are changed, or the surplus pages go out of use. 1678 */ 1679 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 1680 min_count = max(count, min_count); 1681 try_to_free_low(h, min_count, nodes_allowed); 1682 while (min_count < persistent_huge_pages(h)) { 1683 if (!free_pool_huge_page(h, nodes_allowed, 0)) 1684 break; 1685 cond_resched_lock(&hugetlb_lock); 1686 } 1687 while (count < persistent_huge_pages(h)) { 1688 if (!adjust_pool_surplus(h, nodes_allowed, 1)) 1689 break; 1690 } 1691 out: 1692 ret = persistent_huge_pages(h); 1693 spin_unlock(&hugetlb_lock); 1694 return ret; 1695 } 1696 1697 #define HSTATE_ATTR_RO(_name) \ 1698 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1699 1700 #define HSTATE_ATTR(_name) \ 1701 static struct kobj_attribute _name##_attr = \ 1702 __ATTR(_name, 0644, _name##_show, _name##_store) 1703 1704 static struct kobject *hugepages_kobj; 1705 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1706 1707 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 1708 1709 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1710 { 1711 int i; 1712 1713 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1714 if (hstate_kobjs[i] == kobj) { 1715 if (nidp) 1716 *nidp = NUMA_NO_NODE; 1717 return &hstates[i]; 1718 } 1719 1720 return kobj_to_node_hstate(kobj, nidp); 1721 } 1722 1723 static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1724 struct kobj_attribute *attr, char *buf) 1725 { 1726 struct hstate *h; 1727 unsigned long nr_huge_pages; 1728 int nid; 1729 1730 h = kobj_to_hstate(kobj, &nid); 1731 if (nid == NUMA_NO_NODE) 1732 nr_huge_pages = h->nr_huge_pages; 1733 else 1734 nr_huge_pages = h->nr_huge_pages_node[nid]; 1735 1736 return sprintf(buf, "%lu\n", nr_huge_pages); 1737 } 1738 1739 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 1740 struct hstate *h, int nid, 1741 unsigned long count, size_t len) 1742 { 1743 int err; 1744 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1745 1746 if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 1747 err = -EINVAL; 1748 goto out; 1749 } 1750 1751 if (nid == NUMA_NO_NODE) { 1752 /* 1753 * global hstate attribute 1754 */ 1755 if (!(obey_mempolicy && 1756 init_nodemask_of_mempolicy(nodes_allowed))) { 1757 NODEMASK_FREE(nodes_allowed); 1758 nodes_allowed = &node_states[N_MEMORY]; 1759 } 1760 } else if (nodes_allowed) { 1761 /* 1762 * per node hstate attribute: adjust count to global, 1763 * but restrict alloc/free to the specified node. 1764 */ 1765 count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 1766 init_nodemask_of_node(nodes_allowed, nid); 1767 } else 1768 nodes_allowed = &node_states[N_MEMORY]; 1769 1770 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1771 1772 if (nodes_allowed != &node_states[N_MEMORY]) 1773 NODEMASK_FREE(nodes_allowed); 1774 1775 return len; 1776 out: 1777 NODEMASK_FREE(nodes_allowed); 1778 return err; 1779 } 1780 1781 static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 1782 struct kobject *kobj, const char *buf, 1783 size_t len) 1784 { 1785 struct hstate *h; 1786 unsigned long count; 1787 int nid; 1788 int err; 1789 1790 err = kstrtoul(buf, 10, &count); 1791 if (err) 1792 return err; 1793 1794 h = kobj_to_hstate(kobj, &nid); 1795 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 1796 } 1797 1798 static ssize_t nr_hugepages_show(struct kobject *kobj, 1799 struct kobj_attribute *attr, char *buf) 1800 { 1801 return nr_hugepages_show_common(kobj, attr, buf); 1802 } 1803 1804 static ssize_t nr_hugepages_store(struct kobject *kobj, 1805 struct kobj_attribute *attr, const char *buf, size_t len) 1806 { 1807 return nr_hugepages_store_common(false, kobj, buf, len); 1808 } 1809 HSTATE_ATTR(nr_hugepages); 1810 1811 #ifdef CONFIG_NUMA 1812 1813 /* 1814 * hstate attribute for optionally mempolicy-based constraint on persistent 1815 * huge page alloc/free. 1816 */ 1817 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 1818 struct kobj_attribute *attr, char *buf) 1819 { 1820 return nr_hugepages_show_common(kobj, attr, buf); 1821 } 1822 1823 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 1824 struct kobj_attribute *attr, const char *buf, size_t len) 1825 { 1826 return nr_hugepages_store_common(true, kobj, buf, len); 1827 } 1828 HSTATE_ATTR(nr_hugepages_mempolicy); 1829 #endif 1830 1831 1832 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1833 struct kobj_attribute *attr, char *buf) 1834 { 1835 struct hstate *h = kobj_to_hstate(kobj, NULL); 1836 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1837 } 1838 1839 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1840 struct kobj_attribute *attr, const char *buf, size_t count) 1841 { 1842 int err; 1843 unsigned long input; 1844 struct hstate *h = kobj_to_hstate(kobj, NULL); 1845 1846 if (hstate_is_gigantic(h)) 1847 return -EINVAL; 1848 1849 err = kstrtoul(buf, 10, &input); 1850 if (err) 1851 return err; 1852 1853 spin_lock(&hugetlb_lock); 1854 h->nr_overcommit_huge_pages = input; 1855 spin_unlock(&hugetlb_lock); 1856 1857 return count; 1858 } 1859 HSTATE_ATTR(nr_overcommit_hugepages); 1860 1861 static ssize_t free_hugepages_show(struct kobject *kobj, 1862 struct kobj_attribute *attr, char *buf) 1863 { 1864 struct hstate *h; 1865 unsigned long free_huge_pages; 1866 int nid; 1867 1868 h = kobj_to_hstate(kobj, &nid); 1869 if (nid == NUMA_NO_NODE) 1870 free_huge_pages = h->free_huge_pages; 1871 else 1872 free_huge_pages = h->free_huge_pages_node[nid]; 1873 1874 return sprintf(buf, "%lu\n", free_huge_pages); 1875 } 1876 HSTATE_ATTR_RO(free_hugepages); 1877 1878 static ssize_t resv_hugepages_show(struct kobject *kobj, 1879 struct kobj_attribute *attr, char *buf) 1880 { 1881 struct hstate *h = kobj_to_hstate(kobj, NULL); 1882 return sprintf(buf, "%lu\n", h->resv_huge_pages); 1883 } 1884 HSTATE_ATTR_RO(resv_hugepages); 1885 1886 static ssize_t surplus_hugepages_show(struct kobject *kobj, 1887 struct kobj_attribute *attr, char *buf) 1888 { 1889 struct hstate *h; 1890 unsigned long surplus_huge_pages; 1891 int nid; 1892 1893 h = kobj_to_hstate(kobj, &nid); 1894 if (nid == NUMA_NO_NODE) 1895 surplus_huge_pages = h->surplus_huge_pages; 1896 else 1897 surplus_huge_pages = h->surplus_huge_pages_node[nid]; 1898 1899 return sprintf(buf, "%lu\n", surplus_huge_pages); 1900 } 1901 HSTATE_ATTR_RO(surplus_hugepages); 1902 1903 static struct attribute *hstate_attrs[] = { 1904 &nr_hugepages_attr.attr, 1905 &nr_overcommit_hugepages_attr.attr, 1906 &free_hugepages_attr.attr, 1907 &resv_hugepages_attr.attr, 1908 &surplus_hugepages_attr.attr, 1909 #ifdef CONFIG_NUMA 1910 &nr_hugepages_mempolicy_attr.attr, 1911 #endif 1912 NULL, 1913 }; 1914 1915 static struct attribute_group hstate_attr_group = { 1916 .attrs = hstate_attrs, 1917 }; 1918 1919 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 1920 struct kobject **hstate_kobjs, 1921 struct attribute_group *hstate_attr_group) 1922 { 1923 int retval; 1924 int hi = hstate_index(h); 1925 1926 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 1927 if (!hstate_kobjs[hi]) 1928 return -ENOMEM; 1929 1930 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1931 if (retval) 1932 kobject_put(hstate_kobjs[hi]); 1933 1934 return retval; 1935 } 1936 1937 static void __init hugetlb_sysfs_init(void) 1938 { 1939 struct hstate *h; 1940 int err; 1941 1942 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1943 if (!hugepages_kobj) 1944 return; 1945 1946 for_each_hstate(h) { 1947 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 1948 hstate_kobjs, &hstate_attr_group); 1949 if (err) 1950 pr_err("Hugetlb: Unable to add hstate %s", h->name); 1951 } 1952 } 1953 1954 #ifdef CONFIG_NUMA 1955 1956 /* 1957 * node_hstate/s - associate per node hstate attributes, via their kobjects, 1958 * with node devices in node_devices[] using a parallel array. The array 1959 * index of a node device or _hstate == node id. 1960 * This is here to avoid any static dependency of the node device driver, in 1961 * the base kernel, on the hugetlb module. 1962 */ 1963 struct node_hstate { 1964 struct kobject *hugepages_kobj; 1965 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1966 }; 1967 struct node_hstate node_hstates[MAX_NUMNODES]; 1968 1969 /* 1970 * A subset of global hstate attributes for node devices 1971 */ 1972 static struct attribute *per_node_hstate_attrs[] = { 1973 &nr_hugepages_attr.attr, 1974 &free_hugepages_attr.attr, 1975 &surplus_hugepages_attr.attr, 1976 NULL, 1977 }; 1978 1979 static struct attribute_group per_node_hstate_attr_group = { 1980 .attrs = per_node_hstate_attrs, 1981 }; 1982 1983 /* 1984 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 1985 * Returns node id via non-NULL nidp. 1986 */ 1987 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 1988 { 1989 int nid; 1990 1991 for (nid = 0; nid < nr_node_ids; nid++) { 1992 struct node_hstate *nhs = &node_hstates[nid]; 1993 int i; 1994 for (i = 0; i < HUGE_MAX_HSTATE; i++) 1995 if (nhs->hstate_kobjs[i] == kobj) { 1996 if (nidp) 1997 *nidp = nid; 1998 return &hstates[i]; 1999 } 2000 } 2001 2002 BUG(); 2003 return NULL; 2004 } 2005 2006 /* 2007 * Unregister hstate attributes from a single node device. 2008 * No-op if no hstate attributes attached. 2009 */ 2010 static void hugetlb_unregister_node(struct node *node) 2011 { 2012 struct hstate *h; 2013 struct node_hstate *nhs = &node_hstates[node->dev.id]; 2014 2015 if (!nhs->hugepages_kobj) 2016 return; /* no hstate attributes */ 2017 2018 for_each_hstate(h) { 2019 int idx = hstate_index(h); 2020 if (nhs->hstate_kobjs[idx]) { 2021 kobject_put(nhs->hstate_kobjs[idx]); 2022 nhs->hstate_kobjs[idx] = NULL; 2023 } 2024 } 2025 2026 kobject_put(nhs->hugepages_kobj); 2027 nhs->hugepages_kobj = NULL; 2028 } 2029 2030 /* 2031 * hugetlb module exit: unregister hstate attributes from node devices 2032 * that have them. 2033 */ 2034 static void hugetlb_unregister_all_nodes(void) 2035 { 2036 int nid; 2037 2038 /* 2039 * disable node device registrations. 2040 */ 2041 register_hugetlbfs_with_node(NULL, NULL); 2042 2043 /* 2044 * remove hstate attributes from any nodes that have them. 2045 */ 2046 for (nid = 0; nid < nr_node_ids; nid++) 2047 hugetlb_unregister_node(node_devices[nid]); 2048 } 2049 2050 /* 2051 * Register hstate attributes for a single node device. 2052 * No-op if attributes already registered. 2053 */ 2054 static void hugetlb_register_node(struct node *node) 2055 { 2056 struct hstate *h; 2057 struct node_hstate *nhs = &node_hstates[node->dev.id]; 2058 int err; 2059 2060 if (nhs->hugepages_kobj) 2061 return; /* already allocated */ 2062 2063 nhs->hugepages_kobj = kobject_create_and_add("hugepages", 2064 &node->dev.kobj); 2065 if (!nhs->hugepages_kobj) 2066 return; 2067 2068 for_each_hstate(h) { 2069 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 2070 nhs->hstate_kobjs, 2071 &per_node_hstate_attr_group); 2072 if (err) { 2073 pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 2074 h->name, node->dev.id); 2075 hugetlb_unregister_node(node); 2076 break; 2077 } 2078 } 2079 } 2080 2081 /* 2082 * hugetlb init time: register hstate attributes for all registered node 2083 * devices of nodes that have memory. All on-line nodes should have 2084 * registered their associated device by this time. 2085 */ 2086 static void __init hugetlb_register_all_nodes(void) 2087 { 2088 int nid; 2089 2090 for_each_node_state(nid, N_MEMORY) { 2091 struct node *node = node_devices[nid]; 2092 if (node->dev.id == nid) 2093 hugetlb_register_node(node); 2094 } 2095 2096 /* 2097 * Let the node device driver know we're here so it can 2098 * [un]register hstate attributes on node hotplug. 2099 */ 2100 register_hugetlbfs_with_node(hugetlb_register_node, 2101 hugetlb_unregister_node); 2102 } 2103 #else /* !CONFIG_NUMA */ 2104 2105 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 2106 { 2107 BUG(); 2108 if (nidp) 2109 *nidp = -1; 2110 return NULL; 2111 } 2112 2113 static void hugetlb_unregister_all_nodes(void) { } 2114 2115 static void hugetlb_register_all_nodes(void) { } 2116 2117 #endif 2118 2119 static void __exit hugetlb_exit(void) 2120 { 2121 struct hstate *h; 2122 2123 hugetlb_unregister_all_nodes(); 2124 2125 for_each_hstate(h) { 2126 kobject_put(hstate_kobjs[hstate_index(h)]); 2127 } 2128 2129 kobject_put(hugepages_kobj); 2130 kfree(htlb_fault_mutex_table); 2131 } 2132 module_exit(hugetlb_exit); 2133 2134 static int __init hugetlb_init(void) 2135 { 2136 int i; 2137 2138 if (!hugepages_supported()) 2139 return 0; 2140 2141 if (!size_to_hstate(default_hstate_size)) { 2142 default_hstate_size = HPAGE_SIZE; 2143 if (!size_to_hstate(default_hstate_size)) 2144 hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2145 } 2146 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2147 if (default_hstate_max_huge_pages) 2148 default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2149 2150 hugetlb_init_hstates(); 2151 gather_bootmem_prealloc(); 2152 report_hugepages(); 2153 2154 hugetlb_sysfs_init(); 2155 hugetlb_register_all_nodes(); 2156 hugetlb_cgroup_file_init(); 2157 2158 #ifdef CONFIG_SMP 2159 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 2160 #else 2161 num_fault_mutexes = 1; 2162 #endif 2163 htlb_fault_mutex_table = 2164 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 2165 BUG_ON(!htlb_fault_mutex_table); 2166 2167 for (i = 0; i < num_fault_mutexes; i++) 2168 mutex_init(&htlb_fault_mutex_table[i]); 2169 return 0; 2170 } 2171 module_init(hugetlb_init); 2172 2173 /* Should be called on processing a hugepagesz=... option */ 2174 void __init hugetlb_add_hstate(unsigned order) 2175 { 2176 struct hstate *h; 2177 unsigned long i; 2178 2179 if (size_to_hstate(PAGE_SIZE << order)) { 2180 pr_warning("hugepagesz= specified twice, ignoring\n"); 2181 return; 2182 } 2183 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2184 BUG_ON(order == 0); 2185 h = &hstates[hugetlb_max_hstate++]; 2186 h->order = order; 2187 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 2188 h->nr_huge_pages = 0; 2189 h->free_huge_pages = 0; 2190 for (i = 0; i < MAX_NUMNODES; ++i) 2191 INIT_LIST_HEAD(&h->hugepage_freelists[i]); 2192 INIT_LIST_HEAD(&h->hugepage_activelist); 2193 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 2194 h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2195 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2196 huge_page_size(h)/1024); 2197 2198 parsed_hstate = h; 2199 } 2200 2201 static int __init hugetlb_nrpages_setup(char *s) 2202 { 2203 unsigned long *mhp; 2204 static unsigned long *last_mhp; 2205 2206 /* 2207 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2208 * so this hugepages= parameter goes to the "default hstate". 2209 */ 2210 if (!hugetlb_max_hstate) 2211 mhp = &default_hstate_max_huge_pages; 2212 else 2213 mhp = &parsed_hstate->max_huge_pages; 2214 2215 if (mhp == last_mhp) { 2216 pr_warning("hugepages= specified twice without " 2217 "interleaving hugepagesz=, ignoring\n"); 2218 return 1; 2219 } 2220 2221 if (sscanf(s, "%lu", mhp) <= 0) 2222 *mhp = 0; 2223 2224 /* 2225 * Global state is always initialized later in hugetlb_init. 2226 * But we need to allocate >= MAX_ORDER hstates here early to still 2227 * use the bootmem allocator. 2228 */ 2229 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 2230 hugetlb_hstate_alloc_pages(parsed_hstate); 2231 2232 last_mhp = mhp; 2233 2234 return 1; 2235 } 2236 __setup("hugepages=", hugetlb_nrpages_setup); 2237 2238 static int __init hugetlb_default_setup(char *s) 2239 { 2240 default_hstate_size = memparse(s, &s); 2241 return 1; 2242 } 2243 __setup("default_hugepagesz=", hugetlb_default_setup); 2244 2245 static unsigned int cpuset_mems_nr(unsigned int *array) 2246 { 2247 int node; 2248 unsigned int nr = 0; 2249 2250 for_each_node_mask(node, cpuset_current_mems_allowed) 2251 nr += array[node]; 2252 2253 return nr; 2254 } 2255 2256 #ifdef CONFIG_SYSCTL 2257 static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 2258 struct ctl_table *table, int write, 2259 void __user *buffer, size_t *length, loff_t *ppos) 2260 { 2261 struct hstate *h = &default_hstate; 2262 unsigned long tmp = h->max_huge_pages; 2263 int ret; 2264 2265 if (!hugepages_supported()) 2266 return -ENOTSUPP; 2267 2268 table->data = &tmp; 2269 table->maxlen = sizeof(unsigned long); 2270 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 2271 if (ret) 2272 goto out; 2273 2274 if (write) 2275 ret = __nr_hugepages_store_common(obey_mempolicy, h, 2276 NUMA_NO_NODE, tmp, *length); 2277 out: 2278 return ret; 2279 } 2280 2281 int hugetlb_sysctl_handler(struct ctl_table *table, int write, 2282 void __user *buffer, size_t *length, loff_t *ppos) 2283 { 2284 2285 return hugetlb_sysctl_handler_common(false, table, write, 2286 buffer, length, ppos); 2287 } 2288 2289 #ifdef CONFIG_NUMA 2290 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 2291 void __user *buffer, size_t *length, loff_t *ppos) 2292 { 2293 return hugetlb_sysctl_handler_common(true, table, write, 2294 buffer, length, ppos); 2295 } 2296 #endif /* CONFIG_NUMA */ 2297 2298 int hugetlb_overcommit_handler(struct ctl_table *table, int write, 2299 void __user *buffer, 2300 size_t *length, loff_t *ppos) 2301 { 2302 struct hstate *h = &default_hstate; 2303 unsigned long tmp; 2304 int ret; 2305 2306 if (!hugepages_supported()) 2307 return -ENOTSUPP; 2308 2309 tmp = h->nr_overcommit_huge_pages; 2310 2311 if (write && hstate_is_gigantic(h)) 2312 return -EINVAL; 2313 2314 table->data = &tmp; 2315 table->maxlen = sizeof(unsigned long); 2316 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 2317 if (ret) 2318 goto out; 2319 2320 if (write) { 2321 spin_lock(&hugetlb_lock); 2322 h->nr_overcommit_huge_pages = tmp; 2323 spin_unlock(&hugetlb_lock); 2324 } 2325 out: 2326 return ret; 2327 } 2328 2329 #endif /* CONFIG_SYSCTL */ 2330 2331 void hugetlb_report_meminfo(struct seq_file *m) 2332 { 2333 struct hstate *h = &default_hstate; 2334 if (!hugepages_supported()) 2335 return; 2336 seq_printf(m, 2337 "HugePages_Total: %5lu\n" 2338 "HugePages_Free: %5lu\n" 2339 "HugePages_Rsvd: %5lu\n" 2340 "HugePages_Surp: %5lu\n" 2341 "Hugepagesize: %8lu kB\n", 2342 h->nr_huge_pages, 2343 h->free_huge_pages, 2344 h->resv_huge_pages, 2345 h->surplus_huge_pages, 2346 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2347 } 2348 2349 int hugetlb_report_node_meminfo(int nid, char *buf) 2350 { 2351 struct hstate *h = &default_hstate; 2352 if (!hugepages_supported()) 2353 return 0; 2354 return sprintf(buf, 2355 "Node %d HugePages_Total: %5u\n" 2356 "Node %d HugePages_Free: %5u\n" 2357 "Node %d HugePages_Surp: %5u\n", 2358 nid, h->nr_huge_pages_node[nid], 2359 nid, h->free_huge_pages_node[nid], 2360 nid, h->surplus_huge_pages_node[nid]); 2361 } 2362 2363 void hugetlb_show_meminfo(void) 2364 { 2365 struct hstate *h; 2366 int nid; 2367 2368 if (!hugepages_supported()) 2369 return; 2370 2371 for_each_node_state(nid, N_MEMORY) 2372 for_each_hstate(h) 2373 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2374 nid, 2375 h->nr_huge_pages_node[nid], 2376 h->free_huge_pages_node[nid], 2377 h->surplus_huge_pages_node[nid], 2378 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2379 } 2380 2381 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 2382 unsigned long hugetlb_total_pages(void) 2383 { 2384 struct hstate *h; 2385 unsigned long nr_total_pages = 0; 2386 2387 for_each_hstate(h) 2388 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2389 return nr_total_pages; 2390 } 2391 2392 static int hugetlb_acct_memory(struct hstate *h, long delta) 2393 { 2394 int ret = -ENOMEM; 2395 2396 spin_lock(&hugetlb_lock); 2397 /* 2398 * When cpuset is configured, it breaks the strict hugetlb page 2399 * reservation as the accounting is done on a global variable. Such 2400 * reservation is completely rubbish in the presence of cpuset because 2401 * the reservation is not checked against page availability for the 2402 * current cpuset. Application can still potentially OOM'ed by kernel 2403 * with lack of free htlb page in cpuset that the task is in. 2404 * Attempt to enforce strict accounting with cpuset is almost 2405 * impossible (or too ugly) because cpuset is too fluid that 2406 * task or memory node can be dynamically moved between cpusets. 2407 * 2408 * The change of semantics for shared hugetlb mapping with cpuset is 2409 * undesirable. However, in order to preserve some of the semantics, 2410 * we fall back to check against current free page availability as 2411 * a best attempt and hopefully to minimize the impact of changing 2412 * semantics that cpuset has. 2413 */ 2414 if (delta > 0) { 2415 if (gather_surplus_pages(h, delta) < 0) 2416 goto out; 2417 2418 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2419 return_unused_surplus_pages(h, delta); 2420 goto out; 2421 } 2422 } 2423 2424 ret = 0; 2425 if (delta < 0) 2426 return_unused_surplus_pages(h, (unsigned long) -delta); 2427 2428 out: 2429 spin_unlock(&hugetlb_lock); 2430 return ret; 2431 } 2432 2433 static void hugetlb_vm_op_open(struct vm_area_struct *vma) 2434 { 2435 struct resv_map *resv = vma_resv_map(vma); 2436 2437 /* 2438 * This new VMA should share its siblings reservation map if present. 2439 * The VMA will only ever have a valid reservation map pointer where 2440 * it is being copied for another still existing VMA. As that VMA 2441 * has a reference to the reservation map it cannot disappear until 2442 * after this open call completes. It is therefore safe to take a 2443 * new reference here without additional locking. 2444 */ 2445 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2446 kref_get(&resv->refs); 2447 } 2448 2449 static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2450 { 2451 struct hstate *h = hstate_vma(vma); 2452 struct resv_map *resv = vma_resv_map(vma); 2453 struct hugepage_subpool *spool = subpool_vma(vma); 2454 unsigned long reserve, start, end; 2455 2456 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2457 return; 2458 2459 start = vma_hugecache_offset(h, vma, vma->vm_start); 2460 end = vma_hugecache_offset(h, vma, vma->vm_end); 2461 2462 reserve = (end - start) - region_count(resv, start, end); 2463 2464 kref_put(&resv->refs, resv_map_release); 2465 2466 if (reserve) { 2467 hugetlb_acct_memory(h, -reserve); 2468 hugepage_subpool_put_pages(spool, reserve); 2469 } 2470 } 2471 2472 /* 2473 * We cannot handle pagefaults against hugetlb pages at all. They cause 2474 * handle_mm_fault() to try to instantiate regular-sized pages in the 2475 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 2476 * this far. 2477 */ 2478 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 2479 { 2480 BUG(); 2481 return 0; 2482 } 2483 2484 const struct vm_operations_struct hugetlb_vm_ops = { 2485 .fault = hugetlb_vm_op_fault, 2486 .open = hugetlb_vm_op_open, 2487 .close = hugetlb_vm_op_close, 2488 }; 2489 2490 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 2491 int writable) 2492 { 2493 pte_t entry; 2494 2495 if (writable) { 2496 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2497 vma->vm_page_prot))); 2498 } else { 2499 entry = huge_pte_wrprotect(mk_huge_pte(page, 2500 vma->vm_page_prot)); 2501 } 2502 entry = pte_mkyoung(entry); 2503 entry = pte_mkhuge(entry); 2504 entry = arch_make_huge_pte(entry, vma, page, writable); 2505 2506 return entry; 2507 } 2508 2509 static void set_huge_ptep_writable(struct vm_area_struct *vma, 2510 unsigned long address, pte_t *ptep) 2511 { 2512 pte_t entry; 2513 2514 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 2515 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 2516 update_mmu_cache(vma, address, ptep); 2517 } 2518 2519 static int is_hugetlb_entry_migration(pte_t pte) 2520 { 2521 swp_entry_t swp; 2522 2523 if (huge_pte_none(pte) || pte_present(pte)) 2524 return 0; 2525 swp = pte_to_swp_entry(pte); 2526 if (non_swap_entry(swp) && is_migration_entry(swp)) 2527 return 1; 2528 else 2529 return 0; 2530 } 2531 2532 static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2533 { 2534 swp_entry_t swp; 2535 2536 if (huge_pte_none(pte) || pte_present(pte)) 2537 return 0; 2538 swp = pte_to_swp_entry(pte); 2539 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2540 return 1; 2541 else 2542 return 0; 2543 } 2544 2545 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 2546 struct vm_area_struct *vma) 2547 { 2548 pte_t *src_pte, *dst_pte, entry; 2549 struct page *ptepage; 2550 unsigned long addr; 2551 int cow; 2552 struct hstate *h = hstate_vma(vma); 2553 unsigned long sz = huge_page_size(h); 2554 unsigned long mmun_start; /* For mmu_notifiers */ 2555 unsigned long mmun_end; /* For mmu_notifiers */ 2556 int ret = 0; 2557 2558 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 2559 2560 mmun_start = vma->vm_start; 2561 mmun_end = vma->vm_end; 2562 if (cow) 2563 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 2564 2565 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2566 spinlock_t *src_ptl, *dst_ptl; 2567 src_pte = huge_pte_offset(src, addr); 2568 if (!src_pte) 2569 continue; 2570 dst_pte = huge_pte_alloc(dst, addr, sz); 2571 if (!dst_pte) { 2572 ret = -ENOMEM; 2573 break; 2574 } 2575 2576 /* If the pagetables are shared don't copy or take references */ 2577 if (dst_pte == src_pte) 2578 continue; 2579 2580 dst_ptl = huge_pte_lock(h, dst, dst_pte); 2581 src_ptl = huge_pte_lockptr(h, src, src_pte); 2582 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 2583 entry = huge_ptep_get(src_pte); 2584 if (huge_pte_none(entry)) { /* skip none entry */ 2585 ; 2586 } else if (unlikely(is_hugetlb_entry_migration(entry) || 2587 is_hugetlb_entry_hwpoisoned(entry))) { 2588 swp_entry_t swp_entry = pte_to_swp_entry(entry); 2589 2590 if (is_write_migration_entry(swp_entry) && cow) { 2591 /* 2592 * COW mappings require pages in both 2593 * parent and child to be set to read. 2594 */ 2595 make_migration_entry_read(&swp_entry); 2596 entry = swp_entry_to_pte(swp_entry); 2597 set_huge_pte_at(src, addr, src_pte, entry); 2598 } 2599 set_huge_pte_at(dst, addr, dst_pte, entry); 2600 } else { 2601 if (cow) { 2602 huge_ptep_set_wrprotect(src, addr, src_pte); 2603 mmu_notifier_invalidate_range(src, mmun_start, 2604 mmun_end); 2605 } 2606 entry = huge_ptep_get(src_pte); 2607 ptepage = pte_page(entry); 2608 get_page(ptepage); 2609 page_dup_rmap(ptepage); 2610 set_huge_pte_at(dst, addr, dst_pte, entry); 2611 } 2612 spin_unlock(src_ptl); 2613 spin_unlock(dst_ptl); 2614 } 2615 2616 if (cow) 2617 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 2618 2619 return ret; 2620 } 2621 2622 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 2623 unsigned long start, unsigned long end, 2624 struct page *ref_page) 2625 { 2626 int force_flush = 0; 2627 struct mm_struct *mm = vma->vm_mm; 2628 unsigned long address; 2629 pte_t *ptep; 2630 pte_t pte; 2631 spinlock_t *ptl; 2632 struct page *page; 2633 struct hstate *h = hstate_vma(vma); 2634 unsigned long sz = huge_page_size(h); 2635 const unsigned long mmun_start = start; /* For mmu_notifiers */ 2636 const unsigned long mmun_end = end; /* For mmu_notifiers */ 2637 2638 WARN_ON(!is_vm_hugetlb_page(vma)); 2639 BUG_ON(start & ~huge_page_mask(h)); 2640 BUG_ON(end & ~huge_page_mask(h)); 2641 2642 tlb_start_vma(tlb, vma); 2643 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2644 address = start; 2645 again: 2646 for (; address < end; address += sz) { 2647 ptep = huge_pte_offset(mm, address); 2648 if (!ptep) 2649 continue; 2650 2651 ptl = huge_pte_lock(h, mm, ptep); 2652 if (huge_pmd_unshare(mm, &address, ptep)) 2653 goto unlock; 2654 2655 pte = huge_ptep_get(ptep); 2656 if (huge_pte_none(pte)) 2657 goto unlock; 2658 2659 /* 2660 * HWPoisoned hugepage is already unmapped and dropped reference 2661 */ 2662 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 2663 huge_pte_clear(mm, address, ptep); 2664 goto unlock; 2665 } 2666 2667 page = pte_page(pte); 2668 /* 2669 * If a reference page is supplied, it is because a specific 2670 * page is being unmapped, not a range. Ensure the page we 2671 * are about to unmap is the actual page of interest. 2672 */ 2673 if (ref_page) { 2674 if (page != ref_page) 2675 goto unlock; 2676 2677 /* 2678 * Mark the VMA as having unmapped its page so that 2679 * future faults in this VMA will fail rather than 2680 * looking like data was lost 2681 */ 2682 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 2683 } 2684 2685 pte = huge_ptep_get_and_clear(mm, address, ptep); 2686 tlb_remove_tlb_entry(tlb, ptep, address); 2687 if (huge_pte_dirty(pte)) 2688 set_page_dirty(page); 2689 2690 page_remove_rmap(page); 2691 force_flush = !__tlb_remove_page(tlb, page); 2692 if (force_flush) { 2693 address += sz; 2694 spin_unlock(ptl); 2695 break; 2696 } 2697 /* Bail out after unmapping reference page if supplied */ 2698 if (ref_page) { 2699 spin_unlock(ptl); 2700 break; 2701 } 2702 unlock: 2703 spin_unlock(ptl); 2704 } 2705 /* 2706 * mmu_gather ran out of room to batch pages, we break out of 2707 * the PTE lock to avoid doing the potential expensive TLB invalidate 2708 * and page-free while holding it. 2709 */ 2710 if (force_flush) { 2711 force_flush = 0; 2712 tlb_flush_mmu(tlb); 2713 if (address < end && !ref_page) 2714 goto again; 2715 } 2716 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2717 tlb_end_vma(tlb, vma); 2718 } 2719 2720 void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2721 struct vm_area_struct *vma, unsigned long start, 2722 unsigned long end, struct page *ref_page) 2723 { 2724 __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2725 2726 /* 2727 * Clear this flag so that x86's huge_pmd_share page_table_shareable 2728 * test will fail on a vma being torn down, and not grab a page table 2729 * on its way out. We're lucky that the flag has such an appropriate 2730 * name, and can in fact be safely cleared here. We could clear it 2731 * before the __unmap_hugepage_range above, but all that's necessary 2732 * is to clear it before releasing the i_mmap_rwsem. This works 2733 * because in the context this is called, the VMA is about to be 2734 * destroyed and the i_mmap_rwsem is held. 2735 */ 2736 vma->vm_flags &= ~VM_MAYSHARE; 2737 } 2738 2739 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 2740 unsigned long end, struct page *ref_page) 2741 { 2742 struct mm_struct *mm; 2743 struct mmu_gather tlb; 2744 2745 mm = vma->vm_mm; 2746 2747 tlb_gather_mmu(&tlb, mm, start, end); 2748 __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 2749 tlb_finish_mmu(&tlb, start, end); 2750 } 2751 2752 /* 2753 * This is called when the original mapper is failing to COW a MAP_PRIVATE 2754 * mappping it owns the reserve page for. The intention is to unmap the page 2755 * from other VMAs and let the children be SIGKILLed if they are faulting the 2756 * same region. 2757 */ 2758 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 2759 struct page *page, unsigned long address) 2760 { 2761 struct hstate *h = hstate_vma(vma); 2762 struct vm_area_struct *iter_vma; 2763 struct address_space *mapping; 2764 pgoff_t pgoff; 2765 2766 /* 2767 * vm_pgoff is in PAGE_SIZE units, hence the different calculation 2768 * from page cache lookup which is in HPAGE_SIZE units. 2769 */ 2770 address = address & huge_page_mask(h); 2771 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 2772 vma->vm_pgoff; 2773 mapping = file_inode(vma->vm_file)->i_mapping; 2774 2775 /* 2776 * Take the mapping lock for the duration of the table walk. As 2777 * this mapping should be shared between all the VMAs, 2778 * __unmap_hugepage_range() is called as the lock is already held 2779 */ 2780 i_mmap_lock_write(mapping); 2781 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 2782 /* Do not unmap the current VMA */ 2783 if (iter_vma == vma) 2784 continue; 2785 2786 /* 2787 * Unmap the page from other VMAs without their own reserves. 2788 * They get marked to be SIGKILLed if they fault in these 2789 * areas. This is because a future no-page fault on this VMA 2790 * could insert a zeroed page instead of the data existing 2791 * from the time of fork. This would look like data corruption 2792 */ 2793 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 2794 unmap_hugepage_range(iter_vma, address, 2795 address + huge_page_size(h), page); 2796 } 2797 i_mmap_unlock_write(mapping); 2798 } 2799 2800 /* 2801 * Hugetlb_cow() should be called with page lock of the original hugepage held. 2802 * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2803 * cannot race with other handlers or page migration. 2804 * Keep the pte_same checks anyway to make transition from the mutex easier. 2805 */ 2806 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 2807 unsigned long address, pte_t *ptep, pte_t pte, 2808 struct page *pagecache_page, spinlock_t *ptl) 2809 { 2810 struct hstate *h = hstate_vma(vma); 2811 struct page *old_page, *new_page; 2812 int ret = 0, outside_reserve = 0; 2813 unsigned long mmun_start; /* For mmu_notifiers */ 2814 unsigned long mmun_end; /* For mmu_notifiers */ 2815 2816 old_page = pte_page(pte); 2817 2818 retry_avoidcopy: 2819 /* If no-one else is actually using this page, avoid the copy 2820 * and just make the page writable */ 2821 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 2822 page_move_anon_rmap(old_page, vma, address); 2823 set_huge_ptep_writable(vma, address, ptep); 2824 return 0; 2825 } 2826 2827 /* 2828 * If the process that created a MAP_PRIVATE mapping is about to 2829 * perform a COW due to a shared page count, attempt to satisfy 2830 * the allocation without using the existing reserves. The pagecache 2831 * page is used to determine if the reserve at this address was 2832 * consumed or not. If reserves were used, a partial faulted mapping 2833 * at the time of fork() could consume its reserves on COW instead 2834 * of the full address range. 2835 */ 2836 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 2837 old_page != pagecache_page) 2838 outside_reserve = 1; 2839 2840 page_cache_get(old_page); 2841 2842 /* 2843 * Drop page table lock as buddy allocator may be called. It will 2844 * be acquired again before returning to the caller, as expected. 2845 */ 2846 spin_unlock(ptl); 2847 new_page = alloc_huge_page(vma, address, outside_reserve); 2848 2849 if (IS_ERR(new_page)) { 2850 /* 2851 * If a process owning a MAP_PRIVATE mapping fails to COW, 2852 * it is due to references held by a child and an insufficient 2853 * huge page pool. To guarantee the original mappers 2854 * reliability, unmap the page from child processes. The child 2855 * may get SIGKILLed if it later faults. 2856 */ 2857 if (outside_reserve) { 2858 page_cache_release(old_page); 2859 BUG_ON(huge_pte_none(pte)); 2860 unmap_ref_private(mm, vma, old_page, address); 2861 BUG_ON(huge_pte_none(pte)); 2862 spin_lock(ptl); 2863 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2864 if (likely(ptep && 2865 pte_same(huge_ptep_get(ptep), pte))) 2866 goto retry_avoidcopy; 2867 /* 2868 * race occurs while re-acquiring page table 2869 * lock, and our job is done. 2870 */ 2871 return 0; 2872 } 2873 2874 ret = (PTR_ERR(new_page) == -ENOMEM) ? 2875 VM_FAULT_OOM : VM_FAULT_SIGBUS; 2876 goto out_release_old; 2877 } 2878 2879 /* 2880 * When the original hugepage is shared one, it does not have 2881 * anon_vma prepared. 2882 */ 2883 if (unlikely(anon_vma_prepare(vma))) { 2884 ret = VM_FAULT_OOM; 2885 goto out_release_all; 2886 } 2887 2888 copy_user_huge_page(new_page, old_page, address, vma, 2889 pages_per_huge_page(h)); 2890 __SetPageUptodate(new_page); 2891 2892 mmun_start = address & huge_page_mask(h); 2893 mmun_end = mmun_start + huge_page_size(h); 2894 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2895 2896 /* 2897 * Retake the page table lock to check for racing updates 2898 * before the page tables are altered 2899 */ 2900 spin_lock(ptl); 2901 ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2902 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 2903 ClearPagePrivate(new_page); 2904 2905 /* Break COW */ 2906 huge_ptep_clear_flush(vma, address, ptep); 2907 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); 2908 set_huge_pte_at(mm, address, ptep, 2909 make_huge_pte(vma, new_page, 1)); 2910 page_remove_rmap(old_page); 2911 hugepage_add_new_anon_rmap(new_page, vma, address); 2912 /* Make the old page be freed below */ 2913 new_page = old_page; 2914 } 2915 spin_unlock(ptl); 2916 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2917 out_release_all: 2918 page_cache_release(new_page); 2919 out_release_old: 2920 page_cache_release(old_page); 2921 2922 spin_lock(ptl); /* Caller expects lock to be held */ 2923 return ret; 2924 } 2925 2926 /* Return the pagecache page at a given address within a VMA */ 2927 static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2928 struct vm_area_struct *vma, unsigned long address) 2929 { 2930 struct address_space *mapping; 2931 pgoff_t idx; 2932 2933 mapping = vma->vm_file->f_mapping; 2934 idx = vma_hugecache_offset(h, vma, address); 2935 2936 return find_lock_page(mapping, idx); 2937 } 2938 2939 /* 2940 * Return whether there is a pagecache page to back given address within VMA. 2941 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 2942 */ 2943 static bool hugetlbfs_pagecache_present(struct hstate *h, 2944 struct vm_area_struct *vma, unsigned long address) 2945 { 2946 struct address_space *mapping; 2947 pgoff_t idx; 2948 struct page *page; 2949 2950 mapping = vma->vm_file->f_mapping; 2951 idx = vma_hugecache_offset(h, vma, address); 2952 2953 page = find_get_page(mapping, idx); 2954 if (page) 2955 put_page(page); 2956 return page != NULL; 2957 } 2958 2959 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 2960 struct address_space *mapping, pgoff_t idx, 2961 unsigned long address, pte_t *ptep, unsigned int flags) 2962 { 2963 struct hstate *h = hstate_vma(vma); 2964 int ret = VM_FAULT_SIGBUS; 2965 int anon_rmap = 0; 2966 unsigned long size; 2967 struct page *page; 2968 pte_t new_pte; 2969 spinlock_t *ptl; 2970 2971 /* 2972 * Currently, we are forced to kill the process in the event the 2973 * original mapper has unmapped pages from the child due to a failed 2974 * COW. Warn that such a situation has occurred as it may not be obvious 2975 */ 2976 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2977 pr_warning("PID %d killed due to inadequate hugepage pool\n", 2978 current->pid); 2979 return ret; 2980 } 2981 2982 /* 2983 * Use page lock to guard against racing truncation 2984 * before we get page_table_lock. 2985 */ 2986 retry: 2987 page = find_lock_page(mapping, idx); 2988 if (!page) { 2989 size = i_size_read(mapping->host) >> huge_page_shift(h); 2990 if (idx >= size) 2991 goto out; 2992 page = alloc_huge_page(vma, address, 0); 2993 if (IS_ERR(page)) { 2994 ret = PTR_ERR(page); 2995 if (ret == -ENOMEM) 2996 ret = VM_FAULT_OOM; 2997 else 2998 ret = VM_FAULT_SIGBUS; 2999 goto out; 3000 } 3001 clear_huge_page(page, address, pages_per_huge_page(h)); 3002 __SetPageUptodate(page); 3003 3004 if (vma->vm_flags & VM_MAYSHARE) { 3005 int err; 3006 struct inode *inode = mapping->host; 3007 3008 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 3009 if (err) { 3010 put_page(page); 3011 if (err == -EEXIST) 3012 goto retry; 3013 goto out; 3014 } 3015 ClearPagePrivate(page); 3016 3017 spin_lock(&inode->i_lock); 3018 inode->i_blocks += blocks_per_huge_page(h); 3019 spin_unlock(&inode->i_lock); 3020 } else { 3021 lock_page(page); 3022 if (unlikely(anon_vma_prepare(vma))) { 3023 ret = VM_FAULT_OOM; 3024 goto backout_unlocked; 3025 } 3026 anon_rmap = 1; 3027 } 3028 } else { 3029 /* 3030 * If memory error occurs between mmap() and fault, some process 3031 * don't have hwpoisoned swap entry for errored virtual address. 3032 * So we need to block hugepage fault by PG_hwpoison bit check. 3033 */ 3034 if (unlikely(PageHWPoison(page))) { 3035 ret = VM_FAULT_HWPOISON | 3036 VM_FAULT_SET_HINDEX(hstate_index(h)); 3037 goto backout_unlocked; 3038 } 3039 } 3040 3041 /* 3042 * If we are going to COW a private mapping later, we examine the 3043 * pending reservations for this page now. This will ensure that 3044 * any allocations necessary to record that reservation occur outside 3045 * the spinlock. 3046 */ 3047 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 3048 if (vma_needs_reservation(h, vma, address) < 0) { 3049 ret = VM_FAULT_OOM; 3050 goto backout_unlocked; 3051 } 3052 3053 ptl = huge_pte_lockptr(h, mm, ptep); 3054 spin_lock(ptl); 3055 size = i_size_read(mapping->host) >> huge_page_shift(h); 3056 if (idx >= size) 3057 goto backout; 3058 3059 ret = 0; 3060 if (!huge_pte_none(huge_ptep_get(ptep))) 3061 goto backout; 3062 3063 if (anon_rmap) { 3064 ClearPagePrivate(page); 3065 hugepage_add_new_anon_rmap(page, vma, address); 3066 } else 3067 page_dup_rmap(page); 3068 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 3069 && (vma->vm_flags & VM_SHARED))); 3070 set_huge_pte_at(mm, address, ptep, new_pte); 3071 3072 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 3073 /* Optimization, do the COW without a second fault */ 3074 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 3075 } 3076 3077 spin_unlock(ptl); 3078 unlock_page(page); 3079 out: 3080 return ret; 3081 3082 backout: 3083 spin_unlock(ptl); 3084 backout_unlocked: 3085 unlock_page(page); 3086 put_page(page); 3087 goto out; 3088 } 3089 3090 #ifdef CONFIG_SMP 3091 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 3092 struct vm_area_struct *vma, 3093 struct address_space *mapping, 3094 pgoff_t idx, unsigned long address) 3095 { 3096 unsigned long key[2]; 3097 u32 hash; 3098 3099 if (vma->vm_flags & VM_SHARED) { 3100 key[0] = (unsigned long) mapping; 3101 key[1] = idx; 3102 } else { 3103 key[0] = (unsigned long) mm; 3104 key[1] = address >> huge_page_shift(h); 3105 } 3106 3107 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 3108 3109 return hash & (num_fault_mutexes - 1); 3110 } 3111 #else 3112 /* 3113 * For uniprocesor systems we always use a single mutex, so just 3114 * return 0 and avoid the hashing overhead. 3115 */ 3116 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 3117 struct vm_area_struct *vma, 3118 struct address_space *mapping, 3119 pgoff_t idx, unsigned long address) 3120 { 3121 return 0; 3122 } 3123 #endif 3124 3125 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3126 unsigned long address, unsigned int flags) 3127 { 3128 pte_t *ptep, entry; 3129 spinlock_t *ptl; 3130 int ret; 3131 u32 hash; 3132 pgoff_t idx; 3133 struct page *page = NULL; 3134 struct page *pagecache_page = NULL; 3135 struct hstate *h = hstate_vma(vma); 3136 struct address_space *mapping; 3137 3138 address &= huge_page_mask(h); 3139 3140 ptep = huge_pte_offset(mm, address); 3141 if (ptep) { 3142 entry = huge_ptep_get(ptep); 3143 if (unlikely(is_hugetlb_entry_migration(entry))) { 3144 migration_entry_wait_huge(vma, mm, ptep); 3145 return 0; 3146 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3147 return VM_FAULT_HWPOISON_LARGE | 3148 VM_FAULT_SET_HINDEX(hstate_index(h)); 3149 } 3150 3151 ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 3152 if (!ptep) 3153 return VM_FAULT_OOM; 3154 3155 mapping = vma->vm_file->f_mapping; 3156 idx = vma_hugecache_offset(h, vma, address); 3157 3158 /* 3159 * Serialize hugepage allocation and instantiation, so that we don't 3160 * get spurious allocation failures if two CPUs race to instantiate 3161 * the same page in the page cache. 3162 */ 3163 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); 3164 mutex_lock(&htlb_fault_mutex_table[hash]); 3165 3166 entry = huge_ptep_get(ptep); 3167 if (huge_pte_none(entry)) { 3168 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3169 goto out_mutex; 3170 } 3171 3172 ret = 0; 3173 3174 /* 3175 * If we are going to COW the mapping later, we examine the pending 3176 * reservations for this page now. This will ensure that any 3177 * allocations necessary to record that reservation occur outside the 3178 * spinlock. For private mappings, we also lookup the pagecache 3179 * page now as it is used to determine if a reservation has been 3180 * consumed. 3181 */ 3182 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 3183 if (vma_needs_reservation(h, vma, address) < 0) { 3184 ret = VM_FAULT_OOM; 3185 goto out_mutex; 3186 } 3187 3188 if (!(vma->vm_flags & VM_MAYSHARE)) 3189 pagecache_page = hugetlbfs_pagecache_page(h, 3190 vma, address); 3191 } 3192 3193 /* 3194 * hugetlb_cow() requires page locks of pte_page(entry) and 3195 * pagecache_page, so here we need take the former one 3196 * when page != pagecache_page or !pagecache_page. 3197 * Note that locking order is always pagecache_page -> page, 3198 * so no worry about deadlock. 3199 */ 3200 page = pte_page(entry); 3201 get_page(page); 3202 if (page != pagecache_page) 3203 lock_page(page); 3204 3205 ptl = huge_pte_lockptr(h, mm, ptep); 3206 spin_lock(ptl); 3207 /* Check for a racing update before calling hugetlb_cow */ 3208 if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3209 goto out_ptl; 3210 3211 3212 if (flags & FAULT_FLAG_WRITE) { 3213 if (!huge_pte_write(entry)) { 3214 ret = hugetlb_cow(mm, vma, address, ptep, entry, 3215 pagecache_page, ptl); 3216 goto out_ptl; 3217 } 3218 entry = huge_pte_mkdirty(entry); 3219 } 3220 entry = pte_mkyoung(entry); 3221 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3222 flags & FAULT_FLAG_WRITE)) 3223 update_mmu_cache(vma, address, ptep); 3224 3225 out_ptl: 3226 spin_unlock(ptl); 3227 3228 if (pagecache_page) { 3229 unlock_page(pagecache_page); 3230 put_page(pagecache_page); 3231 } 3232 if (page != pagecache_page) 3233 unlock_page(page); 3234 put_page(page); 3235 3236 out_mutex: 3237 mutex_unlock(&htlb_fault_mutex_table[hash]); 3238 return ret; 3239 } 3240 3241 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 3242 struct page **pages, struct vm_area_struct **vmas, 3243 unsigned long *position, unsigned long *nr_pages, 3244 long i, unsigned int flags) 3245 { 3246 unsigned long pfn_offset; 3247 unsigned long vaddr = *position; 3248 unsigned long remainder = *nr_pages; 3249 struct hstate *h = hstate_vma(vma); 3250 3251 while (vaddr < vma->vm_end && remainder) { 3252 pte_t *pte; 3253 spinlock_t *ptl = NULL; 3254 int absent; 3255 struct page *page; 3256 3257 /* 3258 * Some archs (sparc64, sh*) have multiple pte_ts to 3259 * each hugepage. We have to make sure we get the 3260 * first, for the page indexing below to work. 3261 * 3262 * Note that page table lock is not held when pte is null. 3263 */ 3264 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3265 if (pte) 3266 ptl = huge_pte_lock(h, mm, pte); 3267 absent = !pte || huge_pte_none(huge_ptep_get(pte)); 3268 3269 /* 3270 * When coredumping, it suits get_dump_page if we just return 3271 * an error where there's an empty slot with no huge pagecache 3272 * to back it. This way, we avoid allocating a hugepage, and 3273 * the sparse dumpfile avoids allocating disk blocks, but its 3274 * huge holes still show up with zeroes where they need to be. 3275 */ 3276 if (absent && (flags & FOLL_DUMP) && 3277 !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3278 if (pte) 3279 spin_unlock(ptl); 3280 remainder = 0; 3281 break; 3282 } 3283 3284 /* 3285 * We need call hugetlb_fault for both hugepages under migration 3286 * (in which case hugetlb_fault waits for the migration,) and 3287 * hwpoisoned hugepages (in which case we need to prevent the 3288 * caller from accessing to them.) In order to do this, we use 3289 * here is_swap_pte instead of is_hugetlb_entry_migration and 3290 * is_hugetlb_entry_hwpoisoned. This is because it simply covers 3291 * both cases, and because we can't follow correct pages 3292 * directly from any kind of swap entries. 3293 */ 3294 if (absent || is_swap_pte(huge_ptep_get(pte)) || 3295 ((flags & FOLL_WRITE) && 3296 !huge_pte_write(huge_ptep_get(pte)))) { 3297 int ret; 3298 3299 if (pte) 3300 spin_unlock(ptl); 3301 ret = hugetlb_fault(mm, vma, vaddr, 3302 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3303 if (!(ret & VM_FAULT_ERROR)) 3304 continue; 3305 3306 remainder = 0; 3307 break; 3308 } 3309 3310 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 3311 page = pte_page(huge_ptep_get(pte)); 3312 same_page: 3313 if (pages) { 3314 pages[i] = mem_map_offset(page, pfn_offset); 3315 get_page_foll(pages[i]); 3316 } 3317 3318 if (vmas) 3319 vmas[i] = vma; 3320 3321 vaddr += PAGE_SIZE; 3322 ++pfn_offset; 3323 --remainder; 3324 ++i; 3325 if (vaddr < vma->vm_end && remainder && 3326 pfn_offset < pages_per_huge_page(h)) { 3327 /* 3328 * We use pfn_offset to avoid touching the pageframes 3329 * of this compound page. 3330 */ 3331 goto same_page; 3332 } 3333 spin_unlock(ptl); 3334 } 3335 *nr_pages = remainder; 3336 *position = vaddr; 3337 3338 return i ? i : -EFAULT; 3339 } 3340 3341 unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 3342 unsigned long address, unsigned long end, pgprot_t newprot) 3343 { 3344 struct mm_struct *mm = vma->vm_mm; 3345 unsigned long start = address; 3346 pte_t *ptep; 3347 pte_t pte; 3348 struct hstate *h = hstate_vma(vma); 3349 unsigned long pages = 0; 3350 3351 BUG_ON(address >= end); 3352 flush_cache_range(vma, address, end); 3353 3354 mmu_notifier_invalidate_range_start(mm, start, end); 3355 i_mmap_lock_write(vma->vm_file->f_mapping); 3356 for (; address < end; address += huge_page_size(h)) { 3357 spinlock_t *ptl; 3358 ptep = huge_pte_offset(mm, address); 3359 if (!ptep) 3360 continue; 3361 ptl = huge_pte_lock(h, mm, ptep); 3362 if (huge_pmd_unshare(mm, &address, ptep)) { 3363 pages++; 3364 spin_unlock(ptl); 3365 continue; 3366 } 3367 if (!huge_pte_none(huge_ptep_get(ptep))) { 3368 pte = huge_ptep_get_and_clear(mm, address, ptep); 3369 pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3370 pte = arch_make_huge_pte(pte, vma, NULL, 0); 3371 set_huge_pte_at(mm, address, ptep, pte); 3372 pages++; 3373 } 3374 spin_unlock(ptl); 3375 } 3376 /* 3377 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 3378 * may have cleared our pud entry and done put_page on the page table: 3379 * once we release i_mmap_rwsem, another task can do the final put_page 3380 * and that page table be reused and filled with junk. 3381 */ 3382 flush_tlb_range(vma, start, end); 3383 mmu_notifier_invalidate_range(mm, start, end); 3384 i_mmap_unlock_write(vma->vm_file->f_mapping); 3385 mmu_notifier_invalidate_range_end(mm, start, end); 3386 3387 return pages << h->order; 3388 } 3389 3390 int hugetlb_reserve_pages(struct inode *inode, 3391 long from, long to, 3392 struct vm_area_struct *vma, 3393 vm_flags_t vm_flags) 3394 { 3395 long ret, chg; 3396 struct hstate *h = hstate_inode(inode); 3397 struct hugepage_subpool *spool = subpool_inode(inode); 3398 struct resv_map *resv_map; 3399 3400 /* 3401 * Only apply hugepage reservation if asked. At fault time, an 3402 * attempt will be made for VM_NORESERVE to allocate a page 3403 * without using reserves 3404 */ 3405 if (vm_flags & VM_NORESERVE) 3406 return 0; 3407 3408 /* 3409 * Shared mappings base their reservation on the number of pages that 3410 * are already allocated on behalf of the file. Private mappings need 3411 * to reserve the full area even if read-only as mprotect() may be 3412 * called to make the mapping read-write. Assume !vma is a shm mapping 3413 */ 3414 if (!vma || vma->vm_flags & VM_MAYSHARE) { 3415 resv_map = inode_resv_map(inode); 3416 3417 chg = region_chg(resv_map, from, to); 3418 3419 } else { 3420 resv_map = resv_map_alloc(); 3421 if (!resv_map) 3422 return -ENOMEM; 3423 3424 chg = to - from; 3425 3426 set_vma_resv_map(vma, resv_map); 3427 set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 3428 } 3429 3430 if (chg < 0) { 3431 ret = chg; 3432 goto out_err; 3433 } 3434 3435 /* There must be enough pages in the subpool for the mapping */ 3436 if (hugepage_subpool_get_pages(spool, chg)) { 3437 ret = -ENOSPC; 3438 goto out_err; 3439 } 3440 3441 /* 3442 * Check enough hugepages are available for the reservation. 3443 * Hand the pages back to the subpool if there are not 3444 */ 3445 ret = hugetlb_acct_memory(h, chg); 3446 if (ret < 0) { 3447 hugepage_subpool_put_pages(spool, chg); 3448 goto out_err; 3449 } 3450 3451 /* 3452 * Account for the reservations made. Shared mappings record regions 3453 * that have reservations as they are shared by multiple VMAs. 3454 * When the last VMA disappears, the region map says how much 3455 * the reservation was and the page cache tells how much of 3456 * the reservation was consumed. Private mappings are per-VMA and 3457 * only the consumed reservations are tracked. When the VMA 3458 * disappears, the original reservation is the VMA size and the 3459 * consumed reservations are stored in the map. Hence, nothing 3460 * else has to be done for private mappings here 3461 */ 3462 if (!vma || vma->vm_flags & VM_MAYSHARE) 3463 region_add(resv_map, from, to); 3464 return 0; 3465 out_err: 3466 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3467 kref_put(&resv_map->refs, resv_map_release); 3468 return ret; 3469 } 3470 3471 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3472 { 3473 struct hstate *h = hstate_inode(inode); 3474 struct resv_map *resv_map = inode_resv_map(inode); 3475 long chg = 0; 3476 struct hugepage_subpool *spool = subpool_inode(inode); 3477 3478 if (resv_map) 3479 chg = region_truncate(resv_map, offset); 3480 spin_lock(&inode->i_lock); 3481 inode->i_blocks -= (blocks_per_huge_page(h) * freed); 3482 spin_unlock(&inode->i_lock); 3483 3484 hugepage_subpool_put_pages(spool, (chg - freed)); 3485 hugetlb_acct_memory(h, -(chg - freed)); 3486 } 3487 3488 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 3489 static unsigned long page_table_shareable(struct vm_area_struct *svma, 3490 struct vm_area_struct *vma, 3491 unsigned long addr, pgoff_t idx) 3492 { 3493 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 3494 svma->vm_start; 3495 unsigned long sbase = saddr & PUD_MASK; 3496 unsigned long s_end = sbase + PUD_SIZE; 3497 3498 /* Allow segments to share if only one is marked locked */ 3499 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 3500 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 3501 3502 /* 3503 * match the virtual addresses, permission and the alignment of the 3504 * page table page. 3505 */ 3506 if (pmd_index(addr) != pmd_index(saddr) || 3507 vm_flags != svm_flags || 3508 sbase < svma->vm_start || svma->vm_end < s_end) 3509 return 0; 3510 3511 return saddr; 3512 } 3513 3514 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 3515 { 3516 unsigned long base = addr & PUD_MASK; 3517 unsigned long end = base + PUD_SIZE; 3518 3519 /* 3520 * check on proper vm_flags and page table alignment 3521 */ 3522 if (vma->vm_flags & VM_MAYSHARE && 3523 vma->vm_start <= base && end <= vma->vm_end) 3524 return 1; 3525 return 0; 3526 } 3527 3528 /* 3529 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 3530 * and returns the corresponding pte. While this is not necessary for the 3531 * !shared pmd case because we can allocate the pmd later as well, it makes the 3532 * code much cleaner. pmd allocation is essential for the shared case because 3533 * pud has to be populated inside the same i_mmap_rwsem section - otherwise 3534 * racing tasks could either miss the sharing (see huge_pte_offset) or select a 3535 * bad pmd for sharing. 3536 */ 3537 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 3538 { 3539 struct vm_area_struct *vma = find_vma(mm, addr); 3540 struct address_space *mapping = vma->vm_file->f_mapping; 3541 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 3542 vma->vm_pgoff; 3543 struct vm_area_struct *svma; 3544 unsigned long saddr; 3545 pte_t *spte = NULL; 3546 pte_t *pte; 3547 spinlock_t *ptl; 3548 3549 if (!vma_shareable(vma, addr)) 3550 return (pte_t *)pmd_alloc(mm, pud, addr); 3551 3552 i_mmap_lock_write(mapping); 3553 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 3554 if (svma == vma) 3555 continue; 3556 3557 saddr = page_table_shareable(svma, vma, addr, idx); 3558 if (saddr) { 3559 spte = huge_pte_offset(svma->vm_mm, saddr); 3560 if (spte) { 3561 get_page(virt_to_page(spte)); 3562 break; 3563 } 3564 } 3565 } 3566 3567 if (!spte) 3568 goto out; 3569 3570 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 3571 spin_lock(ptl); 3572 if (pud_none(*pud)) 3573 pud_populate(mm, pud, 3574 (pmd_t *)((unsigned long)spte & PAGE_MASK)); 3575 else 3576 put_page(virt_to_page(spte)); 3577 spin_unlock(ptl); 3578 out: 3579 pte = (pte_t *)pmd_alloc(mm, pud, addr); 3580 i_mmap_unlock_write(mapping); 3581 return pte; 3582 } 3583 3584 /* 3585 * unmap huge page backed by shared pte. 3586 * 3587 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 3588 * indicated by page_count > 1, unmap is achieved by clearing pud and 3589 * decrementing the ref count. If count == 1, the pte page is not shared. 3590 * 3591 * called with page table lock held. 3592 * 3593 * returns: 1 successfully unmapped a shared pte page 3594 * 0 the underlying pte page is not shared, or it is the last user 3595 */ 3596 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 3597 { 3598 pgd_t *pgd = pgd_offset(mm, *addr); 3599 pud_t *pud = pud_offset(pgd, *addr); 3600 3601 BUG_ON(page_count(virt_to_page(ptep)) == 0); 3602 if (page_count(virt_to_page(ptep)) == 1) 3603 return 0; 3604 3605 pud_clear(pud); 3606 put_page(virt_to_page(ptep)); 3607 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 3608 return 1; 3609 } 3610 #define want_pmd_share() (1) 3611 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 3612 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 3613 { 3614 return NULL; 3615 } 3616 #define want_pmd_share() (0) 3617 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 3618 3619 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 3620 pte_t *huge_pte_alloc(struct mm_struct *mm, 3621 unsigned long addr, unsigned long sz) 3622 { 3623 pgd_t *pgd; 3624 pud_t *pud; 3625 pte_t *pte = NULL; 3626 3627 pgd = pgd_offset(mm, addr); 3628 pud = pud_alloc(mm, pgd, addr); 3629 if (pud) { 3630 if (sz == PUD_SIZE) { 3631 pte = (pte_t *)pud; 3632 } else { 3633 BUG_ON(sz != PMD_SIZE); 3634 if (want_pmd_share() && pud_none(*pud)) 3635 pte = huge_pmd_share(mm, addr, pud); 3636 else 3637 pte = (pte_t *)pmd_alloc(mm, pud, addr); 3638 } 3639 } 3640 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 3641 3642 return pte; 3643 } 3644 3645 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 3646 { 3647 pgd_t *pgd; 3648 pud_t *pud; 3649 pmd_t *pmd = NULL; 3650 3651 pgd = pgd_offset(mm, addr); 3652 if (pgd_present(*pgd)) { 3653 pud = pud_offset(pgd, addr); 3654 if (pud_present(*pud)) { 3655 if (pud_huge(*pud)) 3656 return (pte_t *)pud; 3657 pmd = pmd_offset(pud, addr); 3658 } 3659 } 3660 return (pte_t *) pmd; 3661 } 3662 3663 struct page * 3664 follow_huge_pmd(struct mm_struct *mm, unsigned long address, 3665 pmd_t *pmd, int write) 3666 { 3667 struct page *page; 3668 3669 page = pte_page(*(pte_t *)pmd); 3670 if (page) 3671 page += ((address & ~PMD_MASK) >> PAGE_SHIFT); 3672 return page; 3673 } 3674 3675 struct page * 3676 follow_huge_pud(struct mm_struct *mm, unsigned long address, 3677 pud_t *pud, int write) 3678 { 3679 struct page *page; 3680 3681 page = pte_page(*(pte_t *)pud); 3682 if (page) 3683 page += ((address & ~PUD_MASK) >> PAGE_SHIFT); 3684 return page; 3685 } 3686 3687 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 3688 3689 /* Can be overriden by architectures */ 3690 struct page * __weak 3691 follow_huge_pud(struct mm_struct *mm, unsigned long address, 3692 pud_t *pud, int write) 3693 { 3694 BUG(); 3695 return NULL; 3696 } 3697 3698 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 3699 3700 #ifdef CONFIG_MEMORY_FAILURE 3701 3702 /* Should be called in hugetlb_lock */ 3703 static int is_hugepage_on_freelist(struct page *hpage) 3704 { 3705 struct page *page; 3706 struct page *tmp; 3707 struct hstate *h = page_hstate(hpage); 3708 int nid = page_to_nid(hpage); 3709 3710 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 3711 if (page == hpage) 3712 return 1; 3713 return 0; 3714 } 3715 3716 /* 3717 * This function is called from memory failure code. 3718 * Assume the caller holds page lock of the head page. 3719 */ 3720 int dequeue_hwpoisoned_huge_page(struct page *hpage) 3721 { 3722 struct hstate *h = page_hstate(hpage); 3723 int nid = page_to_nid(hpage); 3724 int ret = -EBUSY; 3725 3726 spin_lock(&hugetlb_lock); 3727 if (is_hugepage_on_freelist(hpage)) { 3728 /* 3729 * Hwpoisoned hugepage isn't linked to activelist or freelist, 3730 * but dangling hpage->lru can trigger list-debug warnings 3731 * (this happens when we call unpoison_memory() on it), 3732 * so let it point to itself with list_del_init(). 3733 */ 3734 list_del_init(&hpage->lru); 3735 set_page_refcounted(hpage); 3736 h->free_huge_pages--; 3737 h->free_huge_pages_node[nid]--; 3738 ret = 0; 3739 } 3740 spin_unlock(&hugetlb_lock); 3741 return ret; 3742 } 3743 #endif 3744 3745 bool isolate_huge_page(struct page *page, struct list_head *list) 3746 { 3747 VM_BUG_ON_PAGE(!PageHead(page), page); 3748 if (!get_page_unless_zero(page)) 3749 return false; 3750 spin_lock(&hugetlb_lock); 3751 list_move_tail(&page->lru, list); 3752 spin_unlock(&hugetlb_lock); 3753 return true; 3754 } 3755 3756 void putback_active_hugepage(struct page *page) 3757 { 3758 VM_BUG_ON_PAGE(!PageHead(page), page); 3759 spin_lock(&hugetlb_lock); 3760 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 3761 spin_unlock(&hugetlb_lock); 3762 put_page(page); 3763 } 3764 3765 bool is_hugepage_active(struct page *page) 3766 { 3767 VM_BUG_ON_PAGE(!PageHuge(page), page); 3768 /* 3769 * This function can be called for a tail page because the caller, 3770 * scan_movable_pages, scans through a given pfn-range which typically 3771 * covers one memory block. In systems using gigantic hugepage (1GB 3772 * for x86_64,) a hugepage is larger than a memory block, and we don't 3773 * support migrating such large hugepages for now, so return false 3774 * when called for tail pages. 3775 */ 3776 if (PageTail(page)) 3777 return false; 3778 /* 3779 * Refcount of a hwpoisoned hugepages is 1, but they are not active, 3780 * so we should return false for them. 3781 */ 3782 if (unlikely(PageHWPoison(page))) 3783 return false; 3784 return page_count(page) > 0; 3785 } 3786