11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * Generic hugetlb support. 36d49e352SNadia Yvette Chambers * (C) Nadia Yvette Chambers, April 2004 41da177e4SLinus Torvalds */ 51da177e4SLinus Torvalds #include <linux/list.h> 61da177e4SLinus Torvalds #include <linux/init.h> 71da177e4SLinus Torvalds #include <linux/module.h> 81da177e4SLinus Torvalds #include <linux/mm.h> 9e1759c21SAlexey Dobriyan #include <linux/seq_file.h> 101da177e4SLinus Torvalds #include <linux/sysctl.h> 111da177e4SLinus Torvalds #include <linux/highmem.h> 12cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 131da177e4SLinus Torvalds #include <linux/nodemask.h> 1463551ae0SDavid Gibson #include <linux/pagemap.h> 155da7ca86SChristoph Lameter #include <linux/mempolicy.h> 163b32123dSGideon Israel Dsouza #include <linux/compiler.h> 17aea47ff3SChristoph Lameter #include <linux/cpuset.h> 183935baa9SDavid Gibson #include <linux/mutex.h> 19aa888a74SAndi Kleen #include <linux/bootmem.h> 20a3437870SNishanth Aravamudan #include <linux/sysfs.h> 215a0e3ad6STejun Heo #include <linux/slab.h> 220fe6e20bSNaoya Horiguchi #include <linux/rmap.h> 23fd6a03edSNaoya Horiguchi #include <linux/swap.h> 24fd6a03edSNaoya Horiguchi #include <linux/swapops.h> 25c8721bbbSNaoya Horiguchi #include <linux/page-isolation.h> 268382d914SDavidlohr Bueso #include <linux/jhash.h> 27d6606683SLinus Torvalds 2863551ae0SDavid Gibson #include <asm/page.h> 2963551ae0SDavid Gibson #include <asm/pgtable.h> 3024669e58SAneesh Kumar K.V #include <asm/tlb.h> 3163551ae0SDavid Gibson 3224669e58SAneesh Kumar K.V #include <linux/io.h> 3363551ae0SDavid Gibson #include <linux/hugetlb.h> 349dd540e2SAneesh Kumar K.V #include <linux/hugetlb_cgroup.h> 359a305230SLee Schermerhorn #include <linux/node.h> 367835e98bSNick Piggin #include "internal.h" 371da177e4SLinus Torvalds 381da177e4SLinus Torvalds const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; 39396faf03SMel Gorman unsigned long hugepages_treat_as_movable; 40a5516438SAndi Kleen 41c3f38a38SAneesh Kumar K.V int hugetlb_max_hstate __read_mostly; 42e5ff2159SAndi Kleen unsigned int default_hstate_idx; 43e5ff2159SAndi Kleen struct hstate hstates[HUGE_MAX_HSTATE]; 44e5ff2159SAndi Kleen 4553ba51d2SJon Tollefson __initdata LIST_HEAD(huge_boot_pages); 4653ba51d2SJon Tollefson 47e5ff2159SAndi Kleen /* for command line parsing */ 48e5ff2159SAndi Kleen static struct hstate * __initdata parsed_hstate; 49e5ff2159SAndi Kleen static unsigned long __initdata default_hstate_max_huge_pages; 50e11bfbfcSNick Piggin static unsigned long __initdata default_hstate_size; 51e5ff2159SAndi Kleen 523935baa9SDavid Gibson /* 5331caf665SNaoya Horiguchi * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 5431caf665SNaoya Horiguchi * free_huge_pages, and surplus_huge_pages. 553935baa9SDavid Gibson */ 56c3f38a38SAneesh Kumar K.V DEFINE_SPINLOCK(hugetlb_lock); 570bd0f9fbSEric Paris 588382d914SDavidlohr Bueso /* 598382d914SDavidlohr Bueso * Serializes faults on the same logical page. This is used to 608382d914SDavidlohr Bueso * prevent spurious OOMs when the hugepage pool is fully utilized. 618382d914SDavidlohr Bueso */ 628382d914SDavidlohr Bueso static int num_fault_mutexes; 638382d914SDavidlohr Bueso static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; 648382d914SDavidlohr Bueso 6590481622SDavid Gibson static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) 6690481622SDavid Gibson { 6790481622SDavid Gibson bool free = (spool->count == 0) && (spool->used_hpages == 0); 6890481622SDavid Gibson 6990481622SDavid Gibson spin_unlock(&spool->lock); 7090481622SDavid Gibson 7190481622SDavid Gibson /* If no pages are used, and no other handles to the subpool 7290481622SDavid Gibson * remain, free the subpool the subpool remain */ 7390481622SDavid Gibson if (free) 7490481622SDavid Gibson kfree(spool); 7590481622SDavid Gibson } 7690481622SDavid Gibson 7790481622SDavid Gibson struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) 7890481622SDavid Gibson { 7990481622SDavid Gibson struct hugepage_subpool *spool; 8090481622SDavid Gibson 8190481622SDavid Gibson spool = kmalloc(sizeof(*spool), GFP_KERNEL); 8290481622SDavid Gibson if (!spool) 8390481622SDavid Gibson return NULL; 8490481622SDavid Gibson 8590481622SDavid Gibson spin_lock_init(&spool->lock); 8690481622SDavid Gibson spool->count = 1; 8790481622SDavid Gibson spool->max_hpages = nr_blocks; 8890481622SDavid Gibson spool->used_hpages = 0; 8990481622SDavid Gibson 9090481622SDavid Gibson return spool; 9190481622SDavid Gibson } 9290481622SDavid Gibson 9390481622SDavid Gibson void hugepage_put_subpool(struct hugepage_subpool *spool) 9490481622SDavid Gibson { 9590481622SDavid Gibson spin_lock(&spool->lock); 9690481622SDavid Gibson BUG_ON(!spool->count); 9790481622SDavid Gibson spool->count--; 9890481622SDavid Gibson unlock_or_release_subpool(spool); 9990481622SDavid Gibson } 10090481622SDavid Gibson 10190481622SDavid Gibson static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, 10290481622SDavid Gibson long delta) 10390481622SDavid Gibson { 10490481622SDavid Gibson int ret = 0; 10590481622SDavid Gibson 10690481622SDavid Gibson if (!spool) 10790481622SDavid Gibson return 0; 10890481622SDavid Gibson 10990481622SDavid Gibson spin_lock(&spool->lock); 11090481622SDavid Gibson if ((spool->used_hpages + delta) <= spool->max_hpages) { 11190481622SDavid Gibson spool->used_hpages += delta; 11290481622SDavid Gibson } else { 11390481622SDavid Gibson ret = -ENOMEM; 11490481622SDavid Gibson } 11590481622SDavid Gibson spin_unlock(&spool->lock); 11690481622SDavid Gibson 11790481622SDavid Gibson return ret; 11890481622SDavid Gibson } 11990481622SDavid Gibson 12090481622SDavid Gibson static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, 12190481622SDavid Gibson long delta) 12290481622SDavid Gibson { 12390481622SDavid Gibson if (!spool) 12490481622SDavid Gibson return; 12590481622SDavid Gibson 12690481622SDavid Gibson spin_lock(&spool->lock); 12790481622SDavid Gibson spool->used_hpages -= delta; 12890481622SDavid Gibson /* If hugetlbfs_put_super couldn't free spool due to 12990481622SDavid Gibson * an outstanding quota reference, free it now. */ 13090481622SDavid Gibson unlock_or_release_subpool(spool); 13190481622SDavid Gibson } 13290481622SDavid Gibson 13390481622SDavid Gibson static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 13490481622SDavid Gibson { 13590481622SDavid Gibson return HUGETLBFS_SB(inode->i_sb)->spool; 13690481622SDavid Gibson } 13790481622SDavid Gibson 13890481622SDavid Gibson static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 13990481622SDavid Gibson { 140496ad9aaSAl Viro return subpool_inode(file_inode(vma->vm_file)); 14190481622SDavid Gibson } 14290481622SDavid Gibson 143e7c4b0bfSAndy Whitcroft /* 14496822904SAndy Whitcroft * Region tracking -- allows tracking of reservations and instantiated pages 14596822904SAndy Whitcroft * across the pages in a mapping. 14684afd99bSAndy Whitcroft * 1477b24d861SDavidlohr Bueso * The region data structures are embedded into a resv_map and 1487b24d861SDavidlohr Bueso * protected by a resv_map's lock 14996822904SAndy Whitcroft */ 15096822904SAndy Whitcroft struct file_region { 15196822904SAndy Whitcroft struct list_head link; 15296822904SAndy Whitcroft long from; 15396822904SAndy Whitcroft long to; 15496822904SAndy Whitcroft }; 15596822904SAndy Whitcroft 1561406ec9bSJoonsoo Kim static long region_add(struct resv_map *resv, long f, long t) 15796822904SAndy Whitcroft { 1581406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 15996822904SAndy Whitcroft struct file_region *rg, *nrg, *trg; 16096822904SAndy Whitcroft 1617b24d861SDavidlohr Bueso spin_lock(&resv->lock); 16296822904SAndy Whitcroft /* Locate the region we are either in or before. */ 16396822904SAndy Whitcroft list_for_each_entry(rg, head, link) 16496822904SAndy Whitcroft if (f <= rg->to) 16596822904SAndy Whitcroft break; 16696822904SAndy Whitcroft 16796822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 16896822904SAndy Whitcroft if (f > rg->from) 16996822904SAndy Whitcroft f = rg->from; 17096822904SAndy Whitcroft 17196822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 17296822904SAndy Whitcroft nrg = rg; 17396822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 17496822904SAndy Whitcroft if (&rg->link == head) 17596822904SAndy Whitcroft break; 17696822904SAndy Whitcroft if (rg->from > t) 17796822904SAndy Whitcroft break; 17896822904SAndy Whitcroft 17996822904SAndy Whitcroft /* If this area reaches higher then extend our area to 18096822904SAndy Whitcroft * include it completely. If this is not the first area 18196822904SAndy Whitcroft * which we intend to reuse, free it. */ 18296822904SAndy Whitcroft if (rg->to > t) 18396822904SAndy Whitcroft t = rg->to; 18496822904SAndy Whitcroft if (rg != nrg) { 18596822904SAndy Whitcroft list_del(&rg->link); 18696822904SAndy Whitcroft kfree(rg); 18796822904SAndy Whitcroft } 18896822904SAndy Whitcroft } 18996822904SAndy Whitcroft nrg->from = f; 19096822904SAndy Whitcroft nrg->to = t; 1917b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 19296822904SAndy Whitcroft return 0; 19396822904SAndy Whitcroft } 19496822904SAndy Whitcroft 1951406ec9bSJoonsoo Kim static long region_chg(struct resv_map *resv, long f, long t) 19696822904SAndy Whitcroft { 1971406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 1987b24d861SDavidlohr Bueso struct file_region *rg, *nrg = NULL; 19996822904SAndy Whitcroft long chg = 0; 20096822904SAndy Whitcroft 2017b24d861SDavidlohr Bueso retry: 2027b24d861SDavidlohr Bueso spin_lock(&resv->lock); 20396822904SAndy Whitcroft /* Locate the region we are before or in. */ 20496822904SAndy Whitcroft list_for_each_entry(rg, head, link) 20596822904SAndy Whitcroft if (f <= rg->to) 20696822904SAndy Whitcroft break; 20796822904SAndy Whitcroft 20896822904SAndy Whitcroft /* If we are below the current region then a new region is required. 20996822904SAndy Whitcroft * Subtle, allocate a new region at the position but make it zero 21096822904SAndy Whitcroft * size such that we can guarantee to record the reservation. */ 21196822904SAndy Whitcroft if (&rg->link == head || t < rg->from) { 2127b24d861SDavidlohr Bueso if (!nrg) { 2137b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 21496822904SAndy Whitcroft nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 21596822904SAndy Whitcroft if (!nrg) 21696822904SAndy Whitcroft return -ENOMEM; 2177b24d861SDavidlohr Bueso 21896822904SAndy Whitcroft nrg->from = f; 21996822904SAndy Whitcroft nrg->to = f; 22096822904SAndy Whitcroft INIT_LIST_HEAD(&nrg->link); 2217b24d861SDavidlohr Bueso goto retry; 2227b24d861SDavidlohr Bueso } 22396822904SAndy Whitcroft 2247b24d861SDavidlohr Bueso list_add(&nrg->link, rg->link.prev); 2257b24d861SDavidlohr Bueso chg = t - f; 2267b24d861SDavidlohr Bueso goto out_nrg; 22796822904SAndy Whitcroft } 22896822904SAndy Whitcroft 22996822904SAndy Whitcroft /* Round our left edge to the current segment if it encloses us. */ 23096822904SAndy Whitcroft if (f > rg->from) 23196822904SAndy Whitcroft f = rg->from; 23296822904SAndy Whitcroft chg = t - f; 23396822904SAndy Whitcroft 23496822904SAndy Whitcroft /* Check for and consume any regions we now overlap with. */ 23596822904SAndy Whitcroft list_for_each_entry(rg, rg->link.prev, link) { 23696822904SAndy Whitcroft if (&rg->link == head) 23796822904SAndy Whitcroft break; 23896822904SAndy Whitcroft if (rg->from > t) 2397b24d861SDavidlohr Bueso goto out; 24096822904SAndy Whitcroft 24125985edcSLucas De Marchi /* We overlap with this area, if it extends further than 24296822904SAndy Whitcroft * us then we must extend ourselves. Account for its 24396822904SAndy Whitcroft * existing reservation. */ 24496822904SAndy Whitcroft if (rg->to > t) { 24596822904SAndy Whitcroft chg += rg->to - t; 24696822904SAndy Whitcroft t = rg->to; 24796822904SAndy Whitcroft } 24896822904SAndy Whitcroft chg -= rg->to - rg->from; 24996822904SAndy Whitcroft } 2507b24d861SDavidlohr Bueso 2517b24d861SDavidlohr Bueso out: 2527b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 2537b24d861SDavidlohr Bueso /* We already know we raced and no longer need the new region */ 2547b24d861SDavidlohr Bueso kfree(nrg); 2557b24d861SDavidlohr Bueso return chg; 2567b24d861SDavidlohr Bueso out_nrg: 2577b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 25896822904SAndy Whitcroft return chg; 25996822904SAndy Whitcroft } 26096822904SAndy Whitcroft 2611406ec9bSJoonsoo Kim static long region_truncate(struct resv_map *resv, long end) 26296822904SAndy Whitcroft { 2631406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 26496822904SAndy Whitcroft struct file_region *rg, *trg; 26596822904SAndy Whitcroft long chg = 0; 26696822904SAndy Whitcroft 2677b24d861SDavidlohr Bueso spin_lock(&resv->lock); 26896822904SAndy Whitcroft /* Locate the region we are either in or before. */ 26996822904SAndy Whitcroft list_for_each_entry(rg, head, link) 27096822904SAndy Whitcroft if (end <= rg->to) 27196822904SAndy Whitcroft break; 27296822904SAndy Whitcroft if (&rg->link == head) 2737b24d861SDavidlohr Bueso goto out; 27496822904SAndy Whitcroft 27596822904SAndy Whitcroft /* If we are in the middle of a region then adjust it. */ 27696822904SAndy Whitcroft if (end > rg->from) { 27796822904SAndy Whitcroft chg = rg->to - end; 27896822904SAndy Whitcroft rg->to = end; 27996822904SAndy Whitcroft rg = list_entry(rg->link.next, typeof(*rg), link); 28096822904SAndy Whitcroft } 28196822904SAndy Whitcroft 28296822904SAndy Whitcroft /* Drop any remaining regions. */ 28396822904SAndy Whitcroft list_for_each_entry_safe(rg, trg, rg->link.prev, link) { 28496822904SAndy Whitcroft if (&rg->link == head) 28596822904SAndy Whitcroft break; 28696822904SAndy Whitcroft chg += rg->to - rg->from; 28796822904SAndy Whitcroft list_del(&rg->link); 28896822904SAndy Whitcroft kfree(rg); 28996822904SAndy Whitcroft } 2907b24d861SDavidlohr Bueso 2917b24d861SDavidlohr Bueso out: 2927b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 29396822904SAndy Whitcroft return chg; 29496822904SAndy Whitcroft } 29596822904SAndy Whitcroft 2961406ec9bSJoonsoo Kim static long region_count(struct resv_map *resv, long f, long t) 29784afd99bSAndy Whitcroft { 2981406ec9bSJoonsoo Kim struct list_head *head = &resv->regions; 29984afd99bSAndy Whitcroft struct file_region *rg; 30084afd99bSAndy Whitcroft long chg = 0; 30184afd99bSAndy Whitcroft 3027b24d861SDavidlohr Bueso spin_lock(&resv->lock); 30384afd99bSAndy Whitcroft /* Locate each segment we overlap with, and count that overlap. */ 30484afd99bSAndy Whitcroft list_for_each_entry(rg, head, link) { 305f2135a4aSWang Sheng-Hui long seg_from; 306f2135a4aSWang Sheng-Hui long seg_to; 30784afd99bSAndy Whitcroft 30884afd99bSAndy Whitcroft if (rg->to <= f) 30984afd99bSAndy Whitcroft continue; 31084afd99bSAndy Whitcroft if (rg->from >= t) 31184afd99bSAndy Whitcroft break; 31284afd99bSAndy Whitcroft 31384afd99bSAndy Whitcroft seg_from = max(rg->from, f); 31484afd99bSAndy Whitcroft seg_to = min(rg->to, t); 31584afd99bSAndy Whitcroft 31684afd99bSAndy Whitcroft chg += seg_to - seg_from; 31784afd99bSAndy Whitcroft } 3187b24d861SDavidlohr Bueso spin_unlock(&resv->lock); 31984afd99bSAndy Whitcroft 32084afd99bSAndy Whitcroft return chg; 32184afd99bSAndy Whitcroft } 32284afd99bSAndy Whitcroft 32396822904SAndy Whitcroft /* 324e7c4b0bfSAndy Whitcroft * Convert the address within this vma to the page offset within 325e7c4b0bfSAndy Whitcroft * the mapping, in pagecache page units; huge pages here. 326e7c4b0bfSAndy Whitcroft */ 327a5516438SAndi Kleen static pgoff_t vma_hugecache_offset(struct hstate *h, 328a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 329e7c4b0bfSAndy Whitcroft { 330a5516438SAndi Kleen return ((address - vma->vm_start) >> huge_page_shift(h)) + 331a5516438SAndi Kleen (vma->vm_pgoff >> huge_page_order(h)); 332e7c4b0bfSAndy Whitcroft } 333e7c4b0bfSAndy Whitcroft 3340fe6e20bSNaoya Horiguchi pgoff_t linear_hugepage_index(struct vm_area_struct *vma, 3350fe6e20bSNaoya Horiguchi unsigned long address) 3360fe6e20bSNaoya Horiguchi { 3370fe6e20bSNaoya Horiguchi return vma_hugecache_offset(hstate_vma(vma), vma, address); 3380fe6e20bSNaoya Horiguchi } 3390fe6e20bSNaoya Horiguchi 34084afd99bSAndy Whitcroft /* 34108fba699SMel Gorman * Return the size of the pages allocated when backing a VMA. In the majority 34208fba699SMel Gorman * cases this will be same size as used by the page table entries. 34308fba699SMel Gorman */ 34408fba699SMel Gorman unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 34508fba699SMel Gorman { 34608fba699SMel Gorman struct hstate *hstate; 34708fba699SMel Gorman 34808fba699SMel Gorman if (!is_vm_hugetlb_page(vma)) 34908fba699SMel Gorman return PAGE_SIZE; 35008fba699SMel Gorman 35108fba699SMel Gorman hstate = hstate_vma(vma); 35208fba699SMel Gorman 3532415cf12SWanpeng Li return 1UL << huge_page_shift(hstate); 35408fba699SMel Gorman } 355f340ca0fSJoerg Roedel EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 35608fba699SMel Gorman 35708fba699SMel Gorman /* 3583340289dSMel Gorman * Return the page size being used by the MMU to back a VMA. In the majority 3593340289dSMel Gorman * of cases, the page size used by the kernel matches the MMU size. On 3603340289dSMel Gorman * architectures where it differs, an architecture-specific version of this 3613340289dSMel Gorman * function is required. 3623340289dSMel Gorman */ 3633340289dSMel Gorman #ifndef vma_mmu_pagesize 3643340289dSMel Gorman unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 3653340289dSMel Gorman { 3663340289dSMel Gorman return vma_kernel_pagesize(vma); 3673340289dSMel Gorman } 3683340289dSMel Gorman #endif 3693340289dSMel Gorman 3703340289dSMel Gorman /* 37184afd99bSAndy Whitcroft * Flags for MAP_PRIVATE reservations. These are stored in the bottom 37284afd99bSAndy Whitcroft * bits of the reservation map pointer, which are always clear due to 37384afd99bSAndy Whitcroft * alignment. 37484afd99bSAndy Whitcroft */ 37584afd99bSAndy Whitcroft #define HPAGE_RESV_OWNER (1UL << 0) 37684afd99bSAndy Whitcroft #define HPAGE_RESV_UNMAPPED (1UL << 1) 37704f2cbe3SMel Gorman #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 37884afd99bSAndy Whitcroft 379a1e78772SMel Gorman /* 380a1e78772SMel Gorman * These helpers are used to track how many pages are reserved for 381a1e78772SMel Gorman * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 382a1e78772SMel Gorman * is guaranteed to have their future faults succeed. 383a1e78772SMel Gorman * 384a1e78772SMel Gorman * With the exception of reset_vma_resv_huge_pages() which is called at fork(), 385a1e78772SMel Gorman * the reserve counters are updated with the hugetlb_lock held. It is safe 386a1e78772SMel Gorman * to reset the VMA at fork() time as it is not in use yet and there is no 387a1e78772SMel Gorman * chance of the global counters getting corrupted as a result of the values. 38884afd99bSAndy Whitcroft * 38984afd99bSAndy Whitcroft * The private mapping reservation is represented in a subtly different 39084afd99bSAndy Whitcroft * manner to a shared mapping. A shared mapping has a region map associated 39184afd99bSAndy Whitcroft * with the underlying file, this region map represents the backing file 39284afd99bSAndy Whitcroft * pages which have ever had a reservation assigned which this persists even 39384afd99bSAndy Whitcroft * after the page is instantiated. A private mapping has a region map 39484afd99bSAndy Whitcroft * associated with the original mmap which is attached to all VMAs which 39584afd99bSAndy Whitcroft * reference it, this region map represents those offsets which have consumed 39684afd99bSAndy Whitcroft * reservation ie. where pages have been instantiated. 397a1e78772SMel Gorman */ 398e7c4b0bfSAndy Whitcroft static unsigned long get_vma_private_data(struct vm_area_struct *vma) 399e7c4b0bfSAndy Whitcroft { 400e7c4b0bfSAndy Whitcroft return (unsigned long)vma->vm_private_data; 401e7c4b0bfSAndy Whitcroft } 402e7c4b0bfSAndy Whitcroft 403e7c4b0bfSAndy Whitcroft static void set_vma_private_data(struct vm_area_struct *vma, 404e7c4b0bfSAndy Whitcroft unsigned long value) 405e7c4b0bfSAndy Whitcroft { 406e7c4b0bfSAndy Whitcroft vma->vm_private_data = (void *)value; 407e7c4b0bfSAndy Whitcroft } 408e7c4b0bfSAndy Whitcroft 4099119a41eSJoonsoo Kim struct resv_map *resv_map_alloc(void) 41084afd99bSAndy Whitcroft { 41184afd99bSAndy Whitcroft struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 41284afd99bSAndy Whitcroft if (!resv_map) 41384afd99bSAndy Whitcroft return NULL; 41484afd99bSAndy Whitcroft 41584afd99bSAndy Whitcroft kref_init(&resv_map->refs); 4167b24d861SDavidlohr Bueso spin_lock_init(&resv_map->lock); 41784afd99bSAndy Whitcroft INIT_LIST_HEAD(&resv_map->regions); 41884afd99bSAndy Whitcroft 41984afd99bSAndy Whitcroft return resv_map; 42084afd99bSAndy Whitcroft } 42184afd99bSAndy Whitcroft 4229119a41eSJoonsoo Kim void resv_map_release(struct kref *ref) 42384afd99bSAndy Whitcroft { 42484afd99bSAndy Whitcroft struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 42584afd99bSAndy Whitcroft 42684afd99bSAndy Whitcroft /* Clear out any active regions before we release the map. */ 4271406ec9bSJoonsoo Kim region_truncate(resv_map, 0); 42884afd99bSAndy Whitcroft kfree(resv_map); 42984afd99bSAndy Whitcroft } 43084afd99bSAndy Whitcroft 4314e35f483SJoonsoo Kim static inline struct resv_map *inode_resv_map(struct inode *inode) 4324e35f483SJoonsoo Kim { 4334e35f483SJoonsoo Kim return inode->i_mapping->private_data; 4344e35f483SJoonsoo Kim } 4354e35f483SJoonsoo Kim 43684afd99bSAndy Whitcroft static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 437a1e78772SMel Gorman { 438a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 4394e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) { 4404e35f483SJoonsoo Kim struct address_space *mapping = vma->vm_file->f_mapping; 4414e35f483SJoonsoo Kim struct inode *inode = mapping->host; 4424e35f483SJoonsoo Kim 4434e35f483SJoonsoo Kim return inode_resv_map(inode); 4444e35f483SJoonsoo Kim 4454e35f483SJoonsoo Kim } else { 44684afd99bSAndy Whitcroft return (struct resv_map *)(get_vma_private_data(vma) & 44784afd99bSAndy Whitcroft ~HPAGE_RESV_MASK); 4484e35f483SJoonsoo Kim } 449a1e78772SMel Gorman } 450a1e78772SMel Gorman 45184afd99bSAndy Whitcroft static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 452a1e78772SMel Gorman { 453a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 454f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 455a1e78772SMel Gorman 45684afd99bSAndy Whitcroft set_vma_private_data(vma, (get_vma_private_data(vma) & 45784afd99bSAndy Whitcroft HPAGE_RESV_MASK) | (unsigned long)map); 45804f2cbe3SMel Gorman } 45904f2cbe3SMel Gorman 46004f2cbe3SMel Gorman static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 46104f2cbe3SMel Gorman { 46204f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 463f83a275dSMel Gorman VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); 464e7c4b0bfSAndy Whitcroft 465e7c4b0bfSAndy Whitcroft set_vma_private_data(vma, get_vma_private_data(vma) | flags); 46604f2cbe3SMel Gorman } 46704f2cbe3SMel Gorman 46804f2cbe3SMel Gorman static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 46904f2cbe3SMel Gorman { 47004f2cbe3SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 471e7c4b0bfSAndy Whitcroft 472e7c4b0bfSAndy Whitcroft return (get_vma_private_data(vma) & flag) != 0; 473a1e78772SMel Gorman } 474a1e78772SMel Gorman 47504f2cbe3SMel Gorman /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ 476a1e78772SMel Gorman void reset_vma_resv_huge_pages(struct vm_area_struct *vma) 477a1e78772SMel Gorman { 478a1e78772SMel Gorman VM_BUG_ON(!is_vm_hugetlb_page(vma)); 479f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 480a1e78772SMel Gorman vma->vm_private_data = (void *)0; 481a1e78772SMel Gorman } 482a1e78772SMel Gorman 483a1e78772SMel Gorman /* Returns true if the VMA has associated reserve pages */ 484af0ed73eSJoonsoo Kim static int vma_has_reserves(struct vm_area_struct *vma, long chg) 485a1e78772SMel Gorman { 486af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_NORESERVE) { 487af0ed73eSJoonsoo Kim /* 488af0ed73eSJoonsoo Kim * This address is already reserved by other process(chg == 0), 489af0ed73eSJoonsoo Kim * so, we should decrement reserved count. Without decrementing, 490af0ed73eSJoonsoo Kim * reserve count remains after releasing inode, because this 491af0ed73eSJoonsoo Kim * allocated page will go into page cache and is regarded as 492af0ed73eSJoonsoo Kim * coming from reserved pool in releasing step. Currently, we 493af0ed73eSJoonsoo Kim * don't have any other solution to deal with this situation 494af0ed73eSJoonsoo Kim * properly, so add work-around here. 495af0ed73eSJoonsoo Kim */ 496af0ed73eSJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE && chg == 0) 497af0ed73eSJoonsoo Kim return 1; 498af0ed73eSJoonsoo Kim else 49972231b03SJoonsoo Kim return 0; 500af0ed73eSJoonsoo Kim } 501a63884e9SJoonsoo Kim 502a63884e9SJoonsoo Kim /* Shared mappings always use reserves */ 503f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) 504a1e78772SMel Gorman return 1; 505a63884e9SJoonsoo Kim 506a63884e9SJoonsoo Kim /* 507a63884e9SJoonsoo Kim * Only the process that called mmap() has reserves for 508a63884e9SJoonsoo Kim * private mappings. 509a63884e9SJoonsoo Kim */ 5107f09ca51SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5117f09ca51SMel Gorman return 1; 512a63884e9SJoonsoo Kim 5137f09ca51SMel Gorman return 0; 514a1e78772SMel Gorman } 515a1e78772SMel Gorman 516a5516438SAndi Kleen static void enqueue_huge_page(struct hstate *h, struct page *page) 5171da177e4SLinus Torvalds { 5181da177e4SLinus Torvalds int nid = page_to_nid(page); 5190edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_freelists[nid]); 520a5516438SAndi Kleen h->free_huge_pages++; 521a5516438SAndi Kleen h->free_huge_pages_node[nid]++; 5221da177e4SLinus Torvalds } 5231da177e4SLinus Torvalds 524bf50bab2SNaoya Horiguchi static struct page *dequeue_huge_page_node(struct hstate *h, int nid) 525bf50bab2SNaoya Horiguchi { 526bf50bab2SNaoya Horiguchi struct page *page; 527bf50bab2SNaoya Horiguchi 528c8721bbbSNaoya Horiguchi list_for_each_entry(page, &h->hugepage_freelists[nid], lru) 529c8721bbbSNaoya Horiguchi if (!is_migrate_isolate_page(page)) 530c8721bbbSNaoya Horiguchi break; 531c8721bbbSNaoya Horiguchi /* 532c8721bbbSNaoya Horiguchi * if 'non-isolated free hugepage' not found on the list, 533c8721bbbSNaoya Horiguchi * the allocation fails. 534c8721bbbSNaoya Horiguchi */ 535c8721bbbSNaoya Horiguchi if (&h->hugepage_freelists[nid] == &page->lru) 536bf50bab2SNaoya Horiguchi return NULL; 5370edaecfaSAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 538a9869b83SNaoya Horiguchi set_page_refcounted(page); 539bf50bab2SNaoya Horiguchi h->free_huge_pages--; 540bf50bab2SNaoya Horiguchi h->free_huge_pages_node[nid]--; 541bf50bab2SNaoya Horiguchi return page; 542bf50bab2SNaoya Horiguchi } 543bf50bab2SNaoya Horiguchi 54486cdb465SNaoya Horiguchi /* Movability of hugepages depends on migration support. */ 54586cdb465SNaoya Horiguchi static inline gfp_t htlb_alloc_mask(struct hstate *h) 54686cdb465SNaoya Horiguchi { 547100873d7SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_supported(h)) 54886cdb465SNaoya Horiguchi return GFP_HIGHUSER_MOVABLE; 54986cdb465SNaoya Horiguchi else 55086cdb465SNaoya Horiguchi return GFP_HIGHUSER; 55186cdb465SNaoya Horiguchi } 55286cdb465SNaoya Horiguchi 553a5516438SAndi Kleen static struct page *dequeue_huge_page_vma(struct hstate *h, 554a5516438SAndi Kleen struct vm_area_struct *vma, 555af0ed73eSJoonsoo Kim unsigned long address, int avoid_reserve, 556af0ed73eSJoonsoo Kim long chg) 5571da177e4SLinus Torvalds { 558b1c12cbcSKonstantin Khlebnikov struct page *page = NULL; 559480eccf9SLee Schermerhorn struct mempolicy *mpol; 56019770b32SMel Gorman nodemask_t *nodemask; 561c0ff7453SMiao Xie struct zonelist *zonelist; 562dd1a239fSMel Gorman struct zone *zone; 563dd1a239fSMel Gorman struct zoneref *z; 564cc9a6c87SMel Gorman unsigned int cpuset_mems_cookie; 5651da177e4SLinus Torvalds 566a1e78772SMel Gorman /* 567a1e78772SMel Gorman * A child process with MAP_PRIVATE mappings created by their parent 568a1e78772SMel Gorman * have no page reserves. This check ensures that reservations are 569a1e78772SMel Gorman * not "stolen". The child may still get SIGKILLed 570a1e78772SMel Gorman */ 571af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg) && 572a5516438SAndi Kleen h->free_huge_pages - h->resv_huge_pages == 0) 573c0ff7453SMiao Xie goto err; 574a1e78772SMel Gorman 57504f2cbe3SMel Gorman /* If reserves cannot be used, ensure enough pages are in the pool */ 576a5516438SAndi Kleen if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) 5776eab04a8SJustin P. Mattock goto err; 57804f2cbe3SMel Gorman 5799966c4bbSJoonsoo Kim retry_cpuset: 580d26914d1SMel Gorman cpuset_mems_cookie = read_mems_allowed_begin(); 5819966c4bbSJoonsoo Kim zonelist = huge_zonelist(vma, address, 58286cdb465SNaoya Horiguchi htlb_alloc_mask(h), &mpol, &nodemask); 5839966c4bbSJoonsoo Kim 58419770b32SMel Gorman for_each_zone_zonelist_nodemask(zone, z, zonelist, 58519770b32SMel Gorman MAX_NR_ZONES - 1, nodemask) { 58686cdb465SNaoya Horiguchi if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) { 587bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, zone_to_nid(zone)); 588bf50bab2SNaoya Horiguchi if (page) { 589af0ed73eSJoonsoo Kim if (avoid_reserve) 590af0ed73eSJoonsoo Kim break; 591af0ed73eSJoonsoo Kim if (!vma_has_reserves(vma, chg)) 592af0ed73eSJoonsoo Kim break; 593af0ed73eSJoonsoo Kim 59407443a85SJoonsoo Kim SetPagePrivate(page); 595a63884e9SJoonsoo Kim h->resv_huge_pages--; 5965ab3ee7bSKen Chen break; 5971da177e4SLinus Torvalds } 5983abf7afdSAndrew Morton } 599bf50bab2SNaoya Horiguchi } 600cc9a6c87SMel Gorman 601cc9a6c87SMel Gorman mpol_cond_put(mpol); 602d26914d1SMel Gorman if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 603cc9a6c87SMel Gorman goto retry_cpuset; 604cc9a6c87SMel Gorman return page; 605cc9a6c87SMel Gorman 606c0ff7453SMiao Xie err: 607cc9a6c87SMel Gorman return NULL; 6081da177e4SLinus Torvalds } 6091da177e4SLinus Torvalds 6101cac6f2cSLuiz Capitulino /* 6111cac6f2cSLuiz Capitulino * common helper functions for hstate_next_node_to_{alloc|free}. 6121cac6f2cSLuiz Capitulino * We may have allocated or freed a huge page based on a different 6131cac6f2cSLuiz Capitulino * nodes_allowed previously, so h->next_node_to_{alloc|free} might 6141cac6f2cSLuiz Capitulino * be outside of *nodes_allowed. Ensure that we use an allowed 6151cac6f2cSLuiz Capitulino * node for alloc or free. 6161cac6f2cSLuiz Capitulino */ 6171cac6f2cSLuiz Capitulino static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 6181cac6f2cSLuiz Capitulino { 6191cac6f2cSLuiz Capitulino nid = next_node(nid, *nodes_allowed); 6201cac6f2cSLuiz Capitulino if (nid == MAX_NUMNODES) 6211cac6f2cSLuiz Capitulino nid = first_node(*nodes_allowed); 6221cac6f2cSLuiz Capitulino VM_BUG_ON(nid >= MAX_NUMNODES); 6231cac6f2cSLuiz Capitulino 6241cac6f2cSLuiz Capitulino return nid; 6251cac6f2cSLuiz Capitulino } 6261cac6f2cSLuiz Capitulino 6271cac6f2cSLuiz Capitulino static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 6281cac6f2cSLuiz Capitulino { 6291cac6f2cSLuiz Capitulino if (!node_isset(nid, *nodes_allowed)) 6301cac6f2cSLuiz Capitulino nid = next_node_allowed(nid, nodes_allowed); 6311cac6f2cSLuiz Capitulino return nid; 6321cac6f2cSLuiz Capitulino } 6331cac6f2cSLuiz Capitulino 6341cac6f2cSLuiz Capitulino /* 6351cac6f2cSLuiz Capitulino * returns the previously saved node ["this node"] from which to 6361cac6f2cSLuiz Capitulino * allocate a persistent huge page for the pool and advance the 6371cac6f2cSLuiz Capitulino * next node from which to allocate, handling wrap at end of node 6381cac6f2cSLuiz Capitulino * mask. 6391cac6f2cSLuiz Capitulino */ 6401cac6f2cSLuiz Capitulino static int hstate_next_node_to_alloc(struct hstate *h, 6411cac6f2cSLuiz Capitulino nodemask_t *nodes_allowed) 6421cac6f2cSLuiz Capitulino { 6431cac6f2cSLuiz Capitulino int nid; 6441cac6f2cSLuiz Capitulino 6451cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 6461cac6f2cSLuiz Capitulino 6471cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 6481cac6f2cSLuiz Capitulino h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 6491cac6f2cSLuiz Capitulino 6501cac6f2cSLuiz Capitulino return nid; 6511cac6f2cSLuiz Capitulino } 6521cac6f2cSLuiz Capitulino 6531cac6f2cSLuiz Capitulino /* 6541cac6f2cSLuiz Capitulino * helper for free_pool_huge_page() - return the previously saved 6551cac6f2cSLuiz Capitulino * node ["this node"] from which to free a huge page. Advance the 6561cac6f2cSLuiz Capitulino * next node id whether or not we find a free huge page to free so 6571cac6f2cSLuiz Capitulino * that the next attempt to free addresses the next node. 6581cac6f2cSLuiz Capitulino */ 6591cac6f2cSLuiz Capitulino static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 6601cac6f2cSLuiz Capitulino { 6611cac6f2cSLuiz Capitulino int nid; 6621cac6f2cSLuiz Capitulino 6631cac6f2cSLuiz Capitulino VM_BUG_ON(!nodes_allowed); 6641cac6f2cSLuiz Capitulino 6651cac6f2cSLuiz Capitulino nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 6661cac6f2cSLuiz Capitulino h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 6671cac6f2cSLuiz Capitulino 6681cac6f2cSLuiz Capitulino return nid; 6691cac6f2cSLuiz Capitulino } 6701cac6f2cSLuiz Capitulino 6711cac6f2cSLuiz Capitulino #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 6721cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 6731cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 6741cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 6751cac6f2cSLuiz Capitulino nr_nodes--) 6761cac6f2cSLuiz Capitulino 6771cac6f2cSLuiz Capitulino #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 6781cac6f2cSLuiz Capitulino for (nr_nodes = nodes_weight(*mask); \ 6791cac6f2cSLuiz Capitulino nr_nodes > 0 && \ 6801cac6f2cSLuiz Capitulino ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 6811cac6f2cSLuiz Capitulino nr_nodes--) 6821cac6f2cSLuiz Capitulino 683944d9fecSLuiz Capitulino #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) 684944d9fecSLuiz Capitulino static void destroy_compound_gigantic_page(struct page *page, 685944d9fecSLuiz Capitulino unsigned long order) 686944d9fecSLuiz Capitulino { 687944d9fecSLuiz Capitulino int i; 688944d9fecSLuiz Capitulino int nr_pages = 1 << order; 689944d9fecSLuiz Capitulino struct page *p = page + 1; 690944d9fecSLuiz Capitulino 691944d9fecSLuiz Capitulino for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 692944d9fecSLuiz Capitulino __ClearPageTail(p); 693944d9fecSLuiz Capitulino set_page_refcounted(p); 694944d9fecSLuiz Capitulino p->first_page = NULL; 695944d9fecSLuiz Capitulino } 696944d9fecSLuiz Capitulino 697944d9fecSLuiz Capitulino set_compound_order(page, 0); 698944d9fecSLuiz Capitulino __ClearPageHead(page); 699944d9fecSLuiz Capitulino } 700944d9fecSLuiz Capitulino 701944d9fecSLuiz Capitulino static void free_gigantic_page(struct page *page, unsigned order) 702944d9fecSLuiz Capitulino { 703944d9fecSLuiz Capitulino free_contig_range(page_to_pfn(page), 1 << order); 704944d9fecSLuiz Capitulino } 705944d9fecSLuiz Capitulino 706944d9fecSLuiz Capitulino static int __alloc_gigantic_page(unsigned long start_pfn, 707944d9fecSLuiz Capitulino unsigned long nr_pages) 708944d9fecSLuiz Capitulino { 709944d9fecSLuiz Capitulino unsigned long end_pfn = start_pfn + nr_pages; 710944d9fecSLuiz Capitulino return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 711944d9fecSLuiz Capitulino } 712944d9fecSLuiz Capitulino 713944d9fecSLuiz Capitulino static bool pfn_range_valid_gigantic(unsigned long start_pfn, 714944d9fecSLuiz Capitulino unsigned long nr_pages) 715944d9fecSLuiz Capitulino { 716944d9fecSLuiz Capitulino unsigned long i, end_pfn = start_pfn + nr_pages; 717944d9fecSLuiz Capitulino struct page *page; 718944d9fecSLuiz Capitulino 719944d9fecSLuiz Capitulino for (i = start_pfn; i < end_pfn; i++) { 720944d9fecSLuiz Capitulino if (!pfn_valid(i)) 721944d9fecSLuiz Capitulino return false; 722944d9fecSLuiz Capitulino 723944d9fecSLuiz Capitulino page = pfn_to_page(i); 724944d9fecSLuiz Capitulino 725944d9fecSLuiz Capitulino if (PageReserved(page)) 726944d9fecSLuiz Capitulino return false; 727944d9fecSLuiz Capitulino 728944d9fecSLuiz Capitulino if (page_count(page) > 0) 729944d9fecSLuiz Capitulino return false; 730944d9fecSLuiz Capitulino 731944d9fecSLuiz Capitulino if (PageHuge(page)) 732944d9fecSLuiz Capitulino return false; 733944d9fecSLuiz Capitulino } 734944d9fecSLuiz Capitulino 735944d9fecSLuiz Capitulino return true; 736944d9fecSLuiz Capitulino } 737944d9fecSLuiz Capitulino 738944d9fecSLuiz Capitulino static bool zone_spans_last_pfn(const struct zone *zone, 739944d9fecSLuiz Capitulino unsigned long start_pfn, unsigned long nr_pages) 740944d9fecSLuiz Capitulino { 741944d9fecSLuiz Capitulino unsigned long last_pfn = start_pfn + nr_pages - 1; 742944d9fecSLuiz Capitulino return zone_spans_pfn(zone, last_pfn); 743944d9fecSLuiz Capitulino } 744944d9fecSLuiz Capitulino 745944d9fecSLuiz Capitulino static struct page *alloc_gigantic_page(int nid, unsigned order) 746944d9fecSLuiz Capitulino { 747944d9fecSLuiz Capitulino unsigned long nr_pages = 1 << order; 748944d9fecSLuiz Capitulino unsigned long ret, pfn, flags; 749944d9fecSLuiz Capitulino struct zone *z; 750944d9fecSLuiz Capitulino 751944d9fecSLuiz Capitulino z = NODE_DATA(nid)->node_zones; 752944d9fecSLuiz Capitulino for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { 753944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 754944d9fecSLuiz Capitulino 755944d9fecSLuiz Capitulino pfn = ALIGN(z->zone_start_pfn, nr_pages); 756944d9fecSLuiz Capitulino while (zone_spans_last_pfn(z, pfn, nr_pages)) { 757944d9fecSLuiz Capitulino if (pfn_range_valid_gigantic(pfn, nr_pages)) { 758944d9fecSLuiz Capitulino /* 759944d9fecSLuiz Capitulino * We release the zone lock here because 760944d9fecSLuiz Capitulino * alloc_contig_range() will also lock the zone 761944d9fecSLuiz Capitulino * at some point. If there's an allocation 762944d9fecSLuiz Capitulino * spinning on this lock, it may win the race 763944d9fecSLuiz Capitulino * and cause alloc_contig_range() to fail... 764944d9fecSLuiz Capitulino */ 765944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 766944d9fecSLuiz Capitulino ret = __alloc_gigantic_page(pfn, nr_pages); 767944d9fecSLuiz Capitulino if (!ret) 768944d9fecSLuiz Capitulino return pfn_to_page(pfn); 769944d9fecSLuiz Capitulino spin_lock_irqsave(&z->lock, flags); 770944d9fecSLuiz Capitulino } 771944d9fecSLuiz Capitulino pfn += nr_pages; 772944d9fecSLuiz Capitulino } 773944d9fecSLuiz Capitulino 774944d9fecSLuiz Capitulino spin_unlock_irqrestore(&z->lock, flags); 775944d9fecSLuiz Capitulino } 776944d9fecSLuiz Capitulino 777944d9fecSLuiz Capitulino return NULL; 778944d9fecSLuiz Capitulino } 779944d9fecSLuiz Capitulino 780944d9fecSLuiz Capitulino static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); 781944d9fecSLuiz Capitulino static void prep_compound_gigantic_page(struct page *page, unsigned long order); 782944d9fecSLuiz Capitulino 783944d9fecSLuiz Capitulino static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) 784944d9fecSLuiz Capitulino { 785944d9fecSLuiz Capitulino struct page *page; 786944d9fecSLuiz Capitulino 787944d9fecSLuiz Capitulino page = alloc_gigantic_page(nid, huge_page_order(h)); 788944d9fecSLuiz Capitulino if (page) { 789944d9fecSLuiz Capitulino prep_compound_gigantic_page(page, huge_page_order(h)); 790944d9fecSLuiz Capitulino prep_new_huge_page(h, page, nid); 791944d9fecSLuiz Capitulino } 792944d9fecSLuiz Capitulino 793944d9fecSLuiz Capitulino return page; 794944d9fecSLuiz Capitulino } 795944d9fecSLuiz Capitulino 796944d9fecSLuiz Capitulino static int alloc_fresh_gigantic_page(struct hstate *h, 797944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) 798944d9fecSLuiz Capitulino { 799944d9fecSLuiz Capitulino struct page *page = NULL; 800944d9fecSLuiz Capitulino int nr_nodes, node; 801944d9fecSLuiz Capitulino 802944d9fecSLuiz Capitulino for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 803944d9fecSLuiz Capitulino page = alloc_fresh_gigantic_page_node(h, node); 804944d9fecSLuiz Capitulino if (page) 805944d9fecSLuiz Capitulino return 1; 806944d9fecSLuiz Capitulino } 807944d9fecSLuiz Capitulino 808944d9fecSLuiz Capitulino return 0; 809944d9fecSLuiz Capitulino } 810944d9fecSLuiz Capitulino 811944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return true; } 812944d9fecSLuiz Capitulino #else 813944d9fecSLuiz Capitulino static inline bool gigantic_page_supported(void) { return false; } 814944d9fecSLuiz Capitulino static inline void free_gigantic_page(struct page *page, unsigned order) { } 815944d9fecSLuiz Capitulino static inline void destroy_compound_gigantic_page(struct page *page, 816944d9fecSLuiz Capitulino unsigned long order) { } 817944d9fecSLuiz Capitulino static inline int alloc_fresh_gigantic_page(struct hstate *h, 818944d9fecSLuiz Capitulino nodemask_t *nodes_allowed) { return 0; } 819944d9fecSLuiz Capitulino #endif 820944d9fecSLuiz Capitulino 821a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 8226af2acb6SAdam Litke { 8236af2acb6SAdam Litke int i; 824a5516438SAndi Kleen 825944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 826944d9fecSLuiz Capitulino return; 82718229df5SAndy Whitcroft 828a5516438SAndi Kleen h->nr_huge_pages--; 829a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 830a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 83132f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 83232f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 833a7407a27SLuiz Capitulino 1 << PG_active | 1 << PG_private | 834a7407a27SLuiz Capitulino 1 << PG_writeback); 8356af2acb6SAdam Litke } 836309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 8376af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 8386af2acb6SAdam Litke set_page_refcounted(page); 839944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) { 840944d9fecSLuiz Capitulino destroy_compound_gigantic_page(page, huge_page_order(h)); 841944d9fecSLuiz Capitulino free_gigantic_page(page, huge_page_order(h)); 842944d9fecSLuiz Capitulino } else { 8437f2e9525SGerald Schaefer arch_release_hugepage(page); 844a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 8456af2acb6SAdam Litke } 846944d9fecSLuiz Capitulino } 8476af2acb6SAdam Litke 848e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 849e5ff2159SAndi Kleen { 850e5ff2159SAndi Kleen struct hstate *h; 851e5ff2159SAndi Kleen 852e5ff2159SAndi Kleen for_each_hstate(h) { 853e5ff2159SAndi Kleen if (huge_page_size(h) == size) 854e5ff2159SAndi Kleen return h; 855e5ff2159SAndi Kleen } 856e5ff2159SAndi Kleen return NULL; 857e5ff2159SAndi Kleen } 858e5ff2159SAndi Kleen 8598f1d26d0SAtsushi Kumagai void free_huge_page(struct page *page) 86027a85ef1SDavid Gibson { 861a5516438SAndi Kleen /* 862a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 863a5516438SAndi Kleen * compound page destructor. 864a5516438SAndi Kleen */ 865e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 8667893d1d5SAdam Litke int nid = page_to_nid(page); 86790481622SDavid Gibson struct hugepage_subpool *spool = 86890481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 86907443a85SJoonsoo Kim bool restore_reserve; 87027a85ef1SDavid Gibson 871e5df70abSAndy Whitcroft set_page_private(page, 0); 87223be7468SMel Gorman page->mapping = NULL; 8737893d1d5SAdam Litke BUG_ON(page_count(page)); 8740fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 87507443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 87616c794b4SJoonsoo Kim ClearPagePrivate(page); 87727a85ef1SDavid Gibson 87827a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 8796d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 8806d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 88107443a85SJoonsoo Kim if (restore_reserve) 88207443a85SJoonsoo Kim h->resv_huge_pages++; 88307443a85SJoonsoo Kim 884944d9fecSLuiz Capitulino if (h->surplus_huge_pages_node[nid]) { 8850edaecfaSAneesh Kumar K.V /* remove the page from active list */ 8860edaecfaSAneesh Kumar K.V list_del(&page->lru); 887a5516438SAndi Kleen update_and_free_page(h, page); 888a5516438SAndi Kleen h->surplus_huge_pages--; 889a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 8907893d1d5SAdam Litke } else { 8915d3a551cSWill Deacon arch_clear_hugepage_flags(page); 892a5516438SAndi Kleen enqueue_huge_page(h, page); 8937893d1d5SAdam Litke } 89427a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 89590481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 89627a85ef1SDavid Gibson } 89727a85ef1SDavid Gibson 898a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 899b7ba30c6SAndi Kleen { 9000edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 901b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 902b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 9039dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 904a5516438SAndi Kleen h->nr_huge_pages++; 905a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 906b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 907b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 908b7ba30c6SAndi Kleen } 909b7ba30c6SAndi Kleen 9102906dd52SLuiz Capitulino static void prep_compound_gigantic_page(struct page *page, unsigned long order) 91120a0307cSWu Fengguang { 91220a0307cSWu Fengguang int i; 91320a0307cSWu Fengguang int nr_pages = 1 << order; 91420a0307cSWu Fengguang struct page *p = page + 1; 91520a0307cSWu Fengguang 91620a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 91720a0307cSWu Fengguang set_compound_order(page, order); 91820a0307cSWu Fengguang __SetPageHead(page); 919ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 92020a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 92120a0307cSWu Fengguang __SetPageTail(p); 922ef5a22beSAndrea Arcangeli /* 923ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 924ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 925ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 926ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 927ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 928ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 929ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 930ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 931ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 932ef5a22beSAndrea Arcangeli * after get_user_pages(). 933ef5a22beSAndrea Arcangeli */ 934ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 93558a84aa9SYouquan Song set_page_count(p, 0); 93620a0307cSWu Fengguang p->first_page = page; 93720a0307cSWu Fengguang } 93820a0307cSWu Fengguang } 93920a0307cSWu Fengguang 9407795912cSAndrew Morton /* 9417795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 9427795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 9437795912cSAndrew Morton * details. 9447795912cSAndrew Morton */ 94520a0307cSWu Fengguang int PageHuge(struct page *page) 94620a0307cSWu Fengguang { 94720a0307cSWu Fengguang if (!PageCompound(page)) 94820a0307cSWu Fengguang return 0; 94920a0307cSWu Fengguang 95020a0307cSWu Fengguang page = compound_head(page); 951758f66a2SAndrew Morton return get_compound_page_dtor(page) == free_huge_page; 95220a0307cSWu Fengguang } 95343131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 95443131e14SNaoya Horiguchi 95527c73ae7SAndrea Arcangeli /* 95627c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 95727c73ae7SAndrea Arcangeli * normal or transparent huge pages. 95827c73ae7SAndrea Arcangeli */ 95927c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 96027c73ae7SAndrea Arcangeli { 96127c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 96227c73ae7SAndrea Arcangeli return 0; 96327c73ae7SAndrea Arcangeli 964758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 96527c73ae7SAndrea Arcangeli } 96627c73ae7SAndrea Arcangeli 96713d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 96813d60f4bSZhang Yi { 96913d60f4bSZhang Yi struct page *page_head = compound_head(page); 97013d60f4bSZhang Yi pgoff_t index = page_index(page_head); 97113d60f4bSZhang Yi unsigned long compound_idx; 97213d60f4bSZhang Yi 97313d60f4bSZhang Yi if (!PageHuge(page_head)) 97413d60f4bSZhang Yi return page_index(page); 97513d60f4bSZhang Yi 97613d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 97713d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 97813d60f4bSZhang Yi else 97913d60f4bSZhang Yi compound_idx = page - page_head; 98013d60f4bSZhang Yi 98113d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 98213d60f4bSZhang Yi } 98313d60f4bSZhang Yi 984a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 9851da177e4SLinus Torvalds { 9861da177e4SLinus Torvalds struct page *page; 987f96efd58SJoe Jin 9886484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 98986cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 990551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 991a5516438SAndi Kleen huge_page_order(h)); 9921da177e4SLinus Torvalds if (page) { 9937f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 994caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 9957b8ee84dSHarvey Harrison return NULL; 9967f2e9525SGerald Schaefer } 997a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 9981da177e4SLinus Torvalds } 99963b4613cSNishanth Aravamudan 100063b4613cSNishanth Aravamudan return page; 100163b4613cSNishanth Aravamudan } 100263b4613cSNishanth Aravamudan 1003b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 1004b2261026SJoonsoo Kim { 1005b2261026SJoonsoo Kim struct page *page; 1006b2261026SJoonsoo Kim int nr_nodes, node; 1007b2261026SJoonsoo Kim int ret = 0; 1008b2261026SJoonsoo Kim 1009b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1010b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 1011b2261026SJoonsoo Kim if (page) { 1012b2261026SJoonsoo Kim ret = 1; 1013b2261026SJoonsoo Kim break; 1014b2261026SJoonsoo Kim } 1015b2261026SJoonsoo Kim } 1016b2261026SJoonsoo Kim 1017b2261026SJoonsoo Kim if (ret) 1018b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 1019b2261026SJoonsoo Kim else 1020b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1021b2261026SJoonsoo Kim 1022b2261026SJoonsoo Kim return ret; 1023b2261026SJoonsoo Kim } 1024b2261026SJoonsoo Kim 1025e8c5c824SLee Schermerhorn /* 1026e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 1027e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 1028e8c5c824SLee Schermerhorn * balanced over allowed nodes. 1029e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 1030e8c5c824SLee Schermerhorn */ 10316ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 10326ae11b27SLee Schermerhorn bool acct_surplus) 1033e8c5c824SLee Schermerhorn { 1034b2261026SJoonsoo Kim int nr_nodes, node; 1035e8c5c824SLee Schermerhorn int ret = 0; 1036e8c5c824SLee Schermerhorn 1037b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1038685f3457SLee Schermerhorn /* 1039685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 1040685f3457SLee Schermerhorn * nodes with surplus pages. 1041685f3457SLee Schermerhorn */ 1042b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 1043b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 1044e8c5c824SLee Schermerhorn struct page *page = 1045b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 1046e8c5c824SLee Schermerhorn struct page, lru); 1047e8c5c824SLee Schermerhorn list_del(&page->lru); 1048e8c5c824SLee Schermerhorn h->free_huge_pages--; 1049b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 1050685f3457SLee Schermerhorn if (acct_surplus) { 1051685f3457SLee Schermerhorn h->surplus_huge_pages--; 1052b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 1053685f3457SLee Schermerhorn } 1054e8c5c824SLee Schermerhorn update_and_free_page(h, page); 1055e8c5c824SLee Schermerhorn ret = 1; 10569a76db09SLee Schermerhorn break; 1057e8c5c824SLee Schermerhorn } 1058b2261026SJoonsoo Kim } 1059e8c5c824SLee Schermerhorn 1060e8c5c824SLee Schermerhorn return ret; 1061e8c5c824SLee Schermerhorn } 1062e8c5c824SLee Schermerhorn 1063c8721bbbSNaoya Horiguchi /* 1064c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 1065c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 1066c8721bbbSNaoya Horiguchi */ 1067c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 1068c8721bbbSNaoya Horiguchi { 1069c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 1070c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 1071c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 1072c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 1073c8721bbbSNaoya Horiguchi list_del(&page->lru); 1074c8721bbbSNaoya Horiguchi h->free_huge_pages--; 1075c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 1076c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 1077c8721bbbSNaoya Horiguchi } 1078c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 1079c8721bbbSNaoya Horiguchi } 1080c8721bbbSNaoya Horiguchi 1081c8721bbbSNaoya Horiguchi /* 1082c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 1083c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 1084c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 1085c8721bbbSNaoya Horiguchi */ 1086c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 1087c8721bbbSNaoya Horiguchi { 1088c8721bbbSNaoya Horiguchi unsigned int order = 8 * sizeof(void *); 1089c8721bbbSNaoya Horiguchi unsigned long pfn; 1090c8721bbbSNaoya Horiguchi struct hstate *h; 1091c8721bbbSNaoya Horiguchi 1092c8721bbbSNaoya Horiguchi /* Set scan step to minimum hugepage size */ 1093c8721bbbSNaoya Horiguchi for_each_hstate(h) 1094c8721bbbSNaoya Horiguchi if (order > huge_page_order(h)) 1095c8721bbbSNaoya Horiguchi order = huge_page_order(h); 1096c8721bbbSNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); 1097c8721bbbSNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) 1098c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 1099c8721bbbSNaoya Horiguchi } 1100c8721bbbSNaoya Horiguchi 1101bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 11027893d1d5SAdam Litke { 11037893d1d5SAdam Litke struct page *page; 1104bf50bab2SNaoya Horiguchi unsigned int r_nid; 11057893d1d5SAdam Litke 1106bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1107aa888a74SAndi Kleen return NULL; 1108aa888a74SAndi Kleen 1109d1c3fb1fSNishanth Aravamudan /* 1110d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 1111d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 1112d1c3fb1fSNishanth Aravamudan * overcommit 1113d1c3fb1fSNishanth Aravamudan * 1114d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 1115d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 1116d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 1117d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 1118d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 1119d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 1120d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 1121d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 1122d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 1123d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 1124d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 1125d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 1126d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 1127d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 1128d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 1129d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 1130d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 1131d1c3fb1fSNishanth Aravamudan */ 1132d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 1133a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 1134d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1135d1c3fb1fSNishanth Aravamudan return NULL; 1136d1c3fb1fSNishanth Aravamudan } else { 1137a5516438SAndi Kleen h->nr_huge_pages++; 1138a5516438SAndi Kleen h->surplus_huge_pages++; 1139d1c3fb1fSNishanth Aravamudan } 1140d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1141d1c3fb1fSNishanth Aravamudan 1142bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 114386cdb465SNaoya Horiguchi page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| 1144551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1145a5516438SAndi Kleen huge_page_order(h)); 1146bf50bab2SNaoya Horiguchi else 1147bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 114886cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1149bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 1150d1c3fb1fSNishanth Aravamudan 1151caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 1152caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 1153ea5768c7SHillf Danton page = NULL; 1154caff3a2cSGerald Schaefer } 1155caff3a2cSGerald Schaefer 11567893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1157d1c3fb1fSNishanth Aravamudan if (page) { 11580edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1159bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1160d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 11619dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1162d1c3fb1fSNishanth Aravamudan /* 1163d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1164d1c3fb1fSNishanth Aravamudan */ 1165bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1166bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 11673b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1168d1c3fb1fSNishanth Aravamudan } else { 1169a5516438SAndi Kleen h->nr_huge_pages--; 1170a5516438SAndi Kleen h->surplus_huge_pages--; 11713b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 11727893d1d5SAdam Litke } 1173d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 11747893d1d5SAdam Litke 11757893d1d5SAdam Litke return page; 11767893d1d5SAdam Litke } 11777893d1d5SAdam Litke 1178e4e574b7SAdam Litke /* 1179bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1180bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1181bf50bab2SNaoya Horiguchi * address of error page. 1182bf50bab2SNaoya Horiguchi */ 1183bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1184bf50bab2SNaoya Horiguchi { 11854ef91848SJoonsoo Kim struct page *page = NULL; 1186bf50bab2SNaoya Horiguchi 1187bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 11884ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1189bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1190bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1191bf50bab2SNaoya Horiguchi 119294ae8ba7SAneesh Kumar K.V if (!page) 1193bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 1194bf50bab2SNaoya Horiguchi 1195bf50bab2SNaoya Horiguchi return page; 1196bf50bab2SNaoya Horiguchi } 1197bf50bab2SNaoya Horiguchi 1198bf50bab2SNaoya Horiguchi /* 119925985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1200e4e574b7SAdam Litke * of size 'delta'. 1201e4e574b7SAdam Litke */ 1202a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1203e4e574b7SAdam Litke { 1204e4e574b7SAdam Litke struct list_head surplus_list; 1205e4e574b7SAdam Litke struct page *page, *tmp; 1206e4e574b7SAdam Litke int ret, i; 1207e4e574b7SAdam Litke int needed, allocated; 120828073b02SHillf Danton bool alloc_ok = true; 1209e4e574b7SAdam Litke 1210a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1211ac09b3a1SAdam Litke if (needed <= 0) { 1212a5516438SAndi Kleen h->resv_huge_pages += delta; 1213e4e574b7SAdam Litke return 0; 1214ac09b3a1SAdam Litke } 1215e4e574b7SAdam Litke 1216e4e574b7SAdam Litke allocated = 0; 1217e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1218e4e574b7SAdam Litke 1219e4e574b7SAdam Litke ret = -ENOMEM; 1220e4e574b7SAdam Litke retry: 1221e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1222e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1223bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 122428073b02SHillf Danton if (!page) { 122528073b02SHillf Danton alloc_ok = false; 122628073b02SHillf Danton break; 122728073b02SHillf Danton } 1228e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1229e4e574b7SAdam Litke } 123028073b02SHillf Danton allocated += i; 1231e4e574b7SAdam Litke 1232e4e574b7SAdam Litke /* 1233e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1234e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1235e4e574b7SAdam Litke */ 1236e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1237a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1238a5516438SAndi Kleen (h->free_huge_pages + allocated); 123928073b02SHillf Danton if (needed > 0) { 124028073b02SHillf Danton if (alloc_ok) 1241e4e574b7SAdam Litke goto retry; 124228073b02SHillf Danton /* 124328073b02SHillf Danton * We were not able to allocate enough pages to 124428073b02SHillf Danton * satisfy the entire reservation so we free what 124528073b02SHillf Danton * we've allocated so far. 124628073b02SHillf Danton */ 124728073b02SHillf Danton goto free; 124828073b02SHillf Danton } 1249e4e574b7SAdam Litke /* 1250e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 125125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1252e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1253ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1254ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1255ac09b3a1SAdam Litke * before they are reserved. 1256e4e574b7SAdam Litke */ 1257e4e574b7SAdam Litke needed += allocated; 1258a5516438SAndi Kleen h->resv_huge_pages += delta; 1259e4e574b7SAdam Litke ret = 0; 1260a9869b83SNaoya Horiguchi 126119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 126219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 126319fc3f0aSAdam Litke if ((--needed) < 0) 126419fc3f0aSAdam Litke break; 1265a9869b83SNaoya Horiguchi /* 1266a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1267a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1268a9869b83SNaoya Horiguchi */ 1269a9869b83SNaoya Horiguchi put_page_testzero(page); 1270309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1271a5516438SAndi Kleen enqueue_huge_page(h, page); 127219fc3f0aSAdam Litke } 127328073b02SHillf Danton free: 1274b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 127519fc3f0aSAdam Litke 127619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1277c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1278a9869b83SNaoya Horiguchi put_page(page); 127919fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1280e4e574b7SAdam Litke 1281e4e574b7SAdam Litke return ret; 1282e4e574b7SAdam Litke } 1283e4e574b7SAdam Litke 1284e4e574b7SAdam Litke /* 1285e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1286e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1287e4e574b7SAdam Litke * never used. 1288685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1289e4e574b7SAdam Litke */ 1290a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1291a5516438SAndi Kleen unsigned long unused_resv_pages) 1292e4e574b7SAdam Litke { 1293e4e574b7SAdam Litke unsigned long nr_pages; 1294e4e574b7SAdam Litke 1295ac09b3a1SAdam Litke /* Uncommit the reservation */ 1296a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1297ac09b3a1SAdam Litke 1298aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1299bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1300aa888a74SAndi Kleen return; 1301aa888a74SAndi Kleen 1302a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1303e4e574b7SAdam Litke 1304685f3457SLee Schermerhorn /* 1305685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 13069b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 13079b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 13089b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 13099b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 13109b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1311685f3457SLee Schermerhorn */ 1312685f3457SLee Schermerhorn while (nr_pages--) { 13138cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1314685f3457SLee Schermerhorn break; 13157848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1316e4e574b7SAdam Litke } 1317e4e574b7SAdam Litke } 1318e4e574b7SAdam Litke 1319c37f9fb1SAndy Whitcroft /* 1320c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1321c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 132290481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 132390481622SDavid Gibson * can occur. Where any new reservation would be required the 132490481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 132590481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 132690481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 132790481622SDavid Gibson * failure. 1328c37f9fb1SAndy Whitcroft */ 1329e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1330a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1331c37f9fb1SAndy Whitcroft { 13324e35f483SJoonsoo Kim struct resv_map *resv; 13334e35f483SJoonsoo Kim pgoff_t idx; 13344e35f483SJoonsoo Kim long chg; 1335c37f9fb1SAndy Whitcroft 13364e35f483SJoonsoo Kim resv = vma_resv_map(vma); 13374e35f483SJoonsoo Kim if (!resv) 1338c37f9fb1SAndy Whitcroft return 1; 1339c37f9fb1SAndy Whitcroft 13404e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 13414e35f483SJoonsoo Kim chg = region_chg(resv, idx, idx + 1); 134284afd99bSAndy Whitcroft 13434e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 13444e35f483SJoonsoo Kim return chg; 13454e35f483SJoonsoo Kim else 13464e35f483SJoonsoo Kim return chg < 0 ? chg : 0; 134784afd99bSAndy Whitcroft } 1348a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1349a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1350c37f9fb1SAndy Whitcroft { 13514e35f483SJoonsoo Kim struct resv_map *resv; 13524e35f483SJoonsoo Kim pgoff_t idx; 1353c37f9fb1SAndy Whitcroft 13544e35f483SJoonsoo Kim resv = vma_resv_map(vma); 13554e35f483SJoonsoo Kim if (!resv) 13564e35f483SJoonsoo Kim return; 13579119a41eSJoonsoo Kim 13584e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 13591406ec9bSJoonsoo Kim region_add(resv, idx, idx + 1); 1360c37f9fb1SAndy Whitcroft } 1361c37f9fb1SAndy Whitcroft 1362348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 136304f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1364348ea204SAdam Litke { 136590481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1366a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1367348ea204SAdam Litke struct page *page; 1368e2f17d94SRoel Kluin long chg; 13696d76dcf4SAneesh Kumar K.V int ret, idx; 13706d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 13712fc39cecSAdam Litke 13726d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1373a1e78772SMel Gorman /* 137490481622SDavid Gibson * Processes that did not create the mapping will have no 137590481622SDavid Gibson * reserves and will not have accounted against subpool 137690481622SDavid Gibson * limit. Check that the subpool limit can be made before 137790481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 137890481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 137990481622SDavid Gibson * mapping overlaps. 1380a1e78772SMel Gorman */ 1381a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1382c37f9fb1SAndy Whitcroft if (chg < 0) 138376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 13848bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 13858bb3f12eSJoonsoo Kim if (hugepage_subpool_get_pages(spool, 1)) 138676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 138790d8b7e6SAdam Litke 13886d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 13898f34af6fSJianyu Zhan if (ret) 13908f34af6fSJianyu Zhan goto out_subpool_put; 13918f34af6fSJianyu Zhan 1392a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1393af0ed73eSJoonsoo Kim page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 139481a6fcaeSJoonsoo Kim if (!page) { 139594ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1396bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 13978f34af6fSJianyu Zhan if (!page) 13988f34af6fSJianyu Zhan goto out_uncharge_cgroup; 13998f34af6fSJianyu Zhan 140079dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 140179dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 140281a6fcaeSJoonsoo Kim /* Fall through */ 1403a1e78772SMel Gorman } 140481a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 140581a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1406a1e78772SMel Gorman 140790481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1408a1e78772SMel Gorman 1409a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 14107893d1d5SAdam Litke return page; 14118f34af6fSJianyu Zhan 14128f34af6fSJianyu Zhan out_uncharge_cgroup: 14138f34af6fSJianyu Zhan hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 14148f34af6fSJianyu Zhan out_subpool_put: 14158f34af6fSJianyu Zhan if (chg || avoid_reserve) 14168f34af6fSJianyu Zhan hugepage_subpool_put_pages(spool, 1); 14178f34af6fSJianyu Zhan return ERR_PTR(-ENOSPC); 1418b45b5bd6SDavid Gibson } 1419b45b5bd6SDavid Gibson 142074060e4dSNaoya Horiguchi /* 142174060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 142274060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 142374060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 142474060e4dSNaoya Horiguchi */ 142574060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 142674060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 142774060e4dSNaoya Horiguchi { 142874060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 142974060e4dSNaoya Horiguchi if (IS_ERR(page)) 143074060e4dSNaoya Horiguchi page = NULL; 143174060e4dSNaoya Horiguchi return page; 143274060e4dSNaoya Horiguchi } 143374060e4dSNaoya Horiguchi 143491f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1435aa888a74SAndi Kleen { 1436aa888a74SAndi Kleen struct huge_bootmem_page *m; 1437b2261026SJoonsoo Kim int nr_nodes, node; 1438aa888a74SAndi Kleen 1439b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1440aa888a74SAndi Kleen void *addr; 1441aa888a74SAndi Kleen 14428b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 14438b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 14448b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1445aa888a74SAndi Kleen if (addr) { 1446aa888a74SAndi Kleen /* 1447aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1448aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1449aa888a74SAndi Kleen * puts them into the mem_map). 1450aa888a74SAndi Kleen */ 1451aa888a74SAndi Kleen m = addr; 1452aa888a74SAndi Kleen goto found; 1453aa888a74SAndi Kleen } 1454aa888a74SAndi Kleen } 1455aa888a74SAndi Kleen return 0; 1456aa888a74SAndi Kleen 1457aa888a74SAndi Kleen found: 1458aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1459aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1460aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1461aa888a74SAndi Kleen m->hstate = h; 1462aa888a74SAndi Kleen return 1; 1463aa888a74SAndi Kleen } 1464aa888a74SAndi Kleen 1465f412c97aSDavid Rientjes static void __init prep_compound_huge_page(struct page *page, int order) 146618229df5SAndy Whitcroft { 146718229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 146818229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 146918229df5SAndy Whitcroft else 147018229df5SAndy Whitcroft prep_compound_page(page, order); 147118229df5SAndy Whitcroft } 147218229df5SAndy Whitcroft 1473aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1474aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1475aa888a74SAndi Kleen { 1476aa888a74SAndi Kleen struct huge_bootmem_page *m; 1477aa888a74SAndi Kleen 1478aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1479aa888a74SAndi Kleen struct hstate *h = m->hstate; 1480ee8f248dSBecky Bruce struct page *page; 1481ee8f248dSBecky Bruce 1482ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1483ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 14848b89a116SGrygorii Strashko memblock_free_late(__pa(m), 1485ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1486ee8f248dSBecky Bruce #else 1487ee8f248dSBecky Bruce page = virt_to_page(m); 1488ee8f248dSBecky Bruce #endif 1489aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 149018229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1491ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 1492aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1493b0320c7bSRafael Aquini /* 1494b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1495b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1496b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1497b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1498b0320c7bSRafael Aquini */ 1499bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 15003dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 1501aa888a74SAndi Kleen } 1502aa888a74SAndi Kleen } 1503aa888a74SAndi Kleen 15048faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 15051da177e4SLinus Torvalds { 15061da177e4SLinus Torvalds unsigned long i; 15071da177e4SLinus Torvalds 1508e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1509bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) { 1510aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1511aa888a74SAndi Kleen break; 15129b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 15138cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 15141da177e4SLinus Torvalds break; 15151da177e4SLinus Torvalds } 15168faa8b07SAndi Kleen h->max_huge_pages = i; 1517e5ff2159SAndi Kleen } 1518e5ff2159SAndi Kleen 1519e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1520e5ff2159SAndi Kleen { 1521e5ff2159SAndi Kleen struct hstate *h; 1522e5ff2159SAndi Kleen 1523e5ff2159SAndi Kleen for_each_hstate(h) { 15248faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 1525bae7f4aeSLuiz Capitulino if (!hstate_is_gigantic(h)) 15268faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1527e5ff2159SAndi Kleen } 1528e5ff2159SAndi Kleen } 1529e5ff2159SAndi Kleen 15304abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 15314abd32dbSAndi Kleen { 15324abd32dbSAndi Kleen if (n >= (1UL << 30)) 15334abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 15344abd32dbSAndi Kleen else if (n >= (1UL << 20)) 15354abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 15364abd32dbSAndi Kleen else 15374abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 15384abd32dbSAndi Kleen return buf; 15394abd32dbSAndi Kleen } 15404abd32dbSAndi Kleen 1541e5ff2159SAndi Kleen static void __init report_hugepages(void) 1542e5ff2159SAndi Kleen { 1543e5ff2159SAndi Kleen struct hstate *h; 1544e5ff2159SAndi Kleen 1545e5ff2159SAndi Kleen for_each_hstate(h) { 15464abd32dbSAndi Kleen char buf[32]; 1547ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 15484abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 15494abd32dbSAndi Kleen h->free_huge_pages); 1550e5ff2159SAndi Kleen } 1551e5ff2159SAndi Kleen } 1552e5ff2159SAndi Kleen 15531da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 15546ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 15556ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 15561da177e4SLinus Torvalds { 15574415cc8dSChristoph Lameter int i; 15584415cc8dSChristoph Lameter 1559bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1560aa888a74SAndi Kleen return; 1561aa888a74SAndi Kleen 15626ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 15631da177e4SLinus Torvalds struct page *page, *next; 1564a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1565a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1566a5516438SAndi Kleen if (count >= h->nr_huge_pages) 15676b0c880dSAdam Litke return; 15681da177e4SLinus Torvalds if (PageHighMem(page)) 15691da177e4SLinus Torvalds continue; 15701da177e4SLinus Torvalds list_del(&page->lru); 1571e5ff2159SAndi Kleen update_and_free_page(h, page); 1572a5516438SAndi Kleen h->free_huge_pages--; 1573a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 15741da177e4SLinus Torvalds } 15751da177e4SLinus Torvalds } 15761da177e4SLinus Torvalds } 15771da177e4SLinus Torvalds #else 15786ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 15796ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 15801da177e4SLinus Torvalds { 15811da177e4SLinus Torvalds } 15821da177e4SLinus Torvalds #endif 15831da177e4SLinus Torvalds 158420a0307cSWu Fengguang /* 158520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 158620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 158720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 158820a0307cSWu Fengguang */ 15896ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 15906ae11b27SLee Schermerhorn int delta) 159120a0307cSWu Fengguang { 1592b2261026SJoonsoo Kim int nr_nodes, node; 159320a0307cSWu Fengguang 159420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 159520a0307cSWu Fengguang 1596e8c5c824SLee Schermerhorn if (delta < 0) { 1597b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1598b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 1599b2261026SJoonsoo Kim goto found; 1600b2261026SJoonsoo Kim } 1601b2261026SJoonsoo Kim } else { 1602b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1603b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 1604b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 1605b2261026SJoonsoo Kim goto found; 1606e8c5c824SLee Schermerhorn } 16079a76db09SLee Schermerhorn } 1608b2261026SJoonsoo Kim return 0; 160920a0307cSWu Fengguang 1610b2261026SJoonsoo Kim found: 161120a0307cSWu Fengguang h->surplus_huge_pages += delta; 1612b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 1613b2261026SJoonsoo Kim return 1; 161420a0307cSWu Fengguang } 161520a0307cSWu Fengguang 1616a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 16176ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 16186ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 16191da177e4SLinus Torvalds { 16207893d1d5SAdam Litke unsigned long min_count, ret; 16211da177e4SLinus Torvalds 1622944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) 1623aa888a74SAndi Kleen return h->max_huge_pages; 1624aa888a74SAndi Kleen 16257893d1d5SAdam Litke /* 16267893d1d5SAdam Litke * Increase the pool size 16277893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 16287893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1629d1c3fb1fSNishanth Aravamudan * 1630d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1631d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1632d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1633d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1634d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 16357893d1d5SAdam Litke */ 16361da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1637a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 16386ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 16397893d1d5SAdam Litke break; 16407893d1d5SAdam Litke } 16417893d1d5SAdam Litke 1642a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 16437893d1d5SAdam Litke /* 16447893d1d5SAdam Litke * If this allocation races such that we no longer need the 16457893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 16467893d1d5SAdam Litke * and reducing the surplus. 16477893d1d5SAdam Litke */ 16487893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 1649944d9fecSLuiz Capitulino if (hstate_is_gigantic(h)) 1650944d9fecSLuiz Capitulino ret = alloc_fresh_gigantic_page(h, nodes_allowed); 1651944d9fecSLuiz Capitulino else 16526ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 16537893d1d5SAdam Litke spin_lock(&hugetlb_lock); 16547893d1d5SAdam Litke if (!ret) 16557893d1d5SAdam Litke goto out; 16567893d1d5SAdam Litke 1657536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1658536240f2SMel Gorman if (signal_pending(current)) 1659536240f2SMel Gorman goto out; 16607893d1d5SAdam Litke } 16617893d1d5SAdam Litke 16627893d1d5SAdam Litke /* 16637893d1d5SAdam Litke * Decrease the pool size 16647893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 16657893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 16667893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 16677893d1d5SAdam Litke * to the desired size as pages become free. 1668d1c3fb1fSNishanth Aravamudan * 1669d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1670d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1671d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1672d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1673d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1674d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1675d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 16767893d1d5SAdam Litke */ 1677a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 16786b0c880dSAdam Litke min_count = max(count, min_count); 16796ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1680a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 16816ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 16821da177e4SLinus Torvalds break; 168355f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 16841da177e4SLinus Torvalds } 1685a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 16866ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 16877893d1d5SAdam Litke break; 16887893d1d5SAdam Litke } 16897893d1d5SAdam Litke out: 1690a5516438SAndi Kleen ret = persistent_huge_pages(h); 16911da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 16927893d1d5SAdam Litke return ret; 16931da177e4SLinus Torvalds } 16941da177e4SLinus Torvalds 1695a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1696a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1697a3437870SNishanth Aravamudan 1698a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1699a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1700a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1701a3437870SNishanth Aravamudan 1702a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1703a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1704a3437870SNishanth Aravamudan 17059a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 17069a305230SLee Schermerhorn 17079a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1708a3437870SNishanth Aravamudan { 1709a3437870SNishanth Aravamudan int i; 17109a305230SLee Schermerhorn 1711a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 17129a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 17139a305230SLee Schermerhorn if (nidp) 17149a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1715a3437870SNishanth Aravamudan return &hstates[i]; 17169a305230SLee Schermerhorn } 17179a305230SLee Schermerhorn 17189a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1719a3437870SNishanth Aravamudan } 1720a3437870SNishanth Aravamudan 172106808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1722a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1723a3437870SNishanth Aravamudan { 17249a305230SLee Schermerhorn struct hstate *h; 17259a305230SLee Schermerhorn unsigned long nr_huge_pages; 17269a305230SLee Schermerhorn int nid; 17279a305230SLee Schermerhorn 17289a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 17299a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 17309a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 17319a305230SLee Schermerhorn else 17329a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 17339a305230SLee Schermerhorn 17349a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1735a3437870SNishanth Aravamudan } 1736adbe8726SEric B Munson 1737238d3c13SDavid Rientjes static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 1738238d3c13SDavid Rientjes struct hstate *h, int nid, 1739238d3c13SDavid Rientjes unsigned long count, size_t len) 1740a3437870SNishanth Aravamudan { 1741a3437870SNishanth Aravamudan int err; 1742bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1743a3437870SNishanth Aravamudan 1744944d9fecSLuiz Capitulino if (hstate_is_gigantic(h) && !gigantic_page_supported()) { 1745adbe8726SEric B Munson err = -EINVAL; 1746adbe8726SEric B Munson goto out; 1747adbe8726SEric B Munson } 1748adbe8726SEric B Munson 17499a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 17509a305230SLee Schermerhorn /* 17519a305230SLee Schermerhorn * global hstate attribute 17529a305230SLee Schermerhorn */ 17539a305230SLee Schermerhorn if (!(obey_mempolicy && 17549a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 175506808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 17568cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 175706808b08SLee Schermerhorn } 17589a305230SLee Schermerhorn } else if (nodes_allowed) { 17599a305230SLee Schermerhorn /* 17609a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 17619a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 17629a305230SLee Schermerhorn */ 17639a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 17649a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 17659a305230SLee Schermerhorn } else 17668cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 17679a305230SLee Schermerhorn 176806808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1769a3437870SNishanth Aravamudan 17708cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 177106808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 177206808b08SLee Schermerhorn 177306808b08SLee Schermerhorn return len; 1774adbe8726SEric B Munson out: 1775adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1776adbe8726SEric B Munson return err; 177706808b08SLee Schermerhorn } 177806808b08SLee Schermerhorn 1779238d3c13SDavid Rientjes static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 1780238d3c13SDavid Rientjes struct kobject *kobj, const char *buf, 1781238d3c13SDavid Rientjes size_t len) 1782238d3c13SDavid Rientjes { 1783238d3c13SDavid Rientjes struct hstate *h; 1784238d3c13SDavid Rientjes unsigned long count; 1785238d3c13SDavid Rientjes int nid; 1786238d3c13SDavid Rientjes int err; 1787238d3c13SDavid Rientjes 1788238d3c13SDavid Rientjes err = kstrtoul(buf, 10, &count); 1789238d3c13SDavid Rientjes if (err) 1790238d3c13SDavid Rientjes return err; 1791238d3c13SDavid Rientjes 1792238d3c13SDavid Rientjes h = kobj_to_hstate(kobj, &nid); 1793238d3c13SDavid Rientjes return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 1794238d3c13SDavid Rientjes } 1795238d3c13SDavid Rientjes 179606808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 179706808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 179806808b08SLee Schermerhorn { 179906808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 180006808b08SLee Schermerhorn } 180106808b08SLee Schermerhorn 180206808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 180306808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 180406808b08SLee Schermerhorn { 1805238d3c13SDavid Rientjes return nr_hugepages_store_common(false, kobj, buf, len); 1806a3437870SNishanth Aravamudan } 1807a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1808a3437870SNishanth Aravamudan 180906808b08SLee Schermerhorn #ifdef CONFIG_NUMA 181006808b08SLee Schermerhorn 181106808b08SLee Schermerhorn /* 181206808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 181306808b08SLee Schermerhorn * huge page alloc/free. 181406808b08SLee Schermerhorn */ 181506808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 181606808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 181706808b08SLee Schermerhorn { 181806808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 181906808b08SLee Schermerhorn } 182006808b08SLee Schermerhorn 182106808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 182206808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 182306808b08SLee Schermerhorn { 1824238d3c13SDavid Rientjes return nr_hugepages_store_common(true, kobj, buf, len); 182506808b08SLee Schermerhorn } 182606808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 182706808b08SLee Schermerhorn #endif 182806808b08SLee Schermerhorn 182906808b08SLee Schermerhorn 1830a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1831a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1832a3437870SNishanth Aravamudan { 18339a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1834a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1835a3437870SNishanth Aravamudan } 1836adbe8726SEric B Munson 1837a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1838a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1839a3437870SNishanth Aravamudan { 1840a3437870SNishanth Aravamudan int err; 1841a3437870SNishanth Aravamudan unsigned long input; 18429a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1843a3437870SNishanth Aravamudan 1844bae7f4aeSLuiz Capitulino if (hstate_is_gigantic(h)) 1845adbe8726SEric B Munson return -EINVAL; 1846adbe8726SEric B Munson 18473dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 1848a3437870SNishanth Aravamudan if (err) 184973ae31e5SEric B Munson return err; 1850a3437870SNishanth Aravamudan 1851a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1852a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1853a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1854a3437870SNishanth Aravamudan 1855a3437870SNishanth Aravamudan return count; 1856a3437870SNishanth Aravamudan } 1857a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1858a3437870SNishanth Aravamudan 1859a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1860a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1861a3437870SNishanth Aravamudan { 18629a305230SLee Schermerhorn struct hstate *h; 18639a305230SLee Schermerhorn unsigned long free_huge_pages; 18649a305230SLee Schermerhorn int nid; 18659a305230SLee Schermerhorn 18669a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 18679a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 18689a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 18699a305230SLee Schermerhorn else 18709a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 18719a305230SLee Schermerhorn 18729a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1873a3437870SNishanth Aravamudan } 1874a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1875a3437870SNishanth Aravamudan 1876a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1877a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1878a3437870SNishanth Aravamudan { 18799a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1880a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1881a3437870SNishanth Aravamudan } 1882a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1883a3437870SNishanth Aravamudan 1884a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1885a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1886a3437870SNishanth Aravamudan { 18879a305230SLee Schermerhorn struct hstate *h; 18889a305230SLee Schermerhorn unsigned long surplus_huge_pages; 18899a305230SLee Schermerhorn int nid; 18909a305230SLee Schermerhorn 18919a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 18929a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 18939a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 18949a305230SLee Schermerhorn else 18959a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 18969a305230SLee Schermerhorn 18979a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1898a3437870SNishanth Aravamudan } 1899a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1900a3437870SNishanth Aravamudan 1901a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1902a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1903a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1904a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1905a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1906a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 190706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 190806808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 190906808b08SLee Schermerhorn #endif 1910a3437870SNishanth Aravamudan NULL, 1911a3437870SNishanth Aravamudan }; 1912a3437870SNishanth Aravamudan 1913a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1914a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1915a3437870SNishanth Aravamudan }; 1916a3437870SNishanth Aravamudan 1917094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 19189a305230SLee Schermerhorn struct kobject **hstate_kobjs, 19199a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1920a3437870SNishanth Aravamudan { 1921a3437870SNishanth Aravamudan int retval; 1922972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1923a3437870SNishanth Aravamudan 19249a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 19259a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1926a3437870SNishanth Aravamudan return -ENOMEM; 1927a3437870SNishanth Aravamudan 19289a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1929a3437870SNishanth Aravamudan if (retval) 19309a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1931a3437870SNishanth Aravamudan 1932a3437870SNishanth Aravamudan return retval; 1933a3437870SNishanth Aravamudan } 1934a3437870SNishanth Aravamudan 1935a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1936a3437870SNishanth Aravamudan { 1937a3437870SNishanth Aravamudan struct hstate *h; 1938a3437870SNishanth Aravamudan int err; 1939a3437870SNishanth Aravamudan 1940a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1941a3437870SNishanth Aravamudan if (!hugepages_kobj) 1942a3437870SNishanth Aravamudan return; 1943a3437870SNishanth Aravamudan 1944a3437870SNishanth Aravamudan for_each_hstate(h) { 19459a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 19469a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1947a3437870SNishanth Aravamudan if (err) 1948ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 1949a3437870SNishanth Aravamudan } 1950a3437870SNishanth Aravamudan } 1951a3437870SNishanth Aravamudan 19529a305230SLee Schermerhorn #ifdef CONFIG_NUMA 19539a305230SLee Schermerhorn 19549a305230SLee Schermerhorn /* 19559a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 195610fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 195710fbcf4cSKay Sievers * index of a node device or _hstate == node id. 195810fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 19599a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 19609a305230SLee Schermerhorn */ 19619a305230SLee Schermerhorn struct node_hstate { 19629a305230SLee Schermerhorn struct kobject *hugepages_kobj; 19639a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 19649a305230SLee Schermerhorn }; 19659a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 19669a305230SLee Schermerhorn 19679a305230SLee Schermerhorn /* 196810fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 19699a305230SLee Schermerhorn */ 19709a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 19719a305230SLee Schermerhorn &nr_hugepages_attr.attr, 19729a305230SLee Schermerhorn &free_hugepages_attr.attr, 19739a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 19749a305230SLee Schermerhorn NULL, 19759a305230SLee Schermerhorn }; 19769a305230SLee Schermerhorn 19779a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 19789a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 19799a305230SLee Schermerhorn }; 19809a305230SLee Schermerhorn 19819a305230SLee Schermerhorn /* 198210fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 19839a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 19849a305230SLee Schermerhorn */ 19859a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 19869a305230SLee Schermerhorn { 19879a305230SLee Schermerhorn int nid; 19889a305230SLee Schermerhorn 19899a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 19909a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 19919a305230SLee Schermerhorn int i; 19929a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 19939a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 19949a305230SLee Schermerhorn if (nidp) 19959a305230SLee Schermerhorn *nidp = nid; 19969a305230SLee Schermerhorn return &hstates[i]; 19979a305230SLee Schermerhorn } 19989a305230SLee Schermerhorn } 19999a305230SLee Schermerhorn 20009a305230SLee Schermerhorn BUG(); 20019a305230SLee Schermerhorn return NULL; 20029a305230SLee Schermerhorn } 20039a305230SLee Schermerhorn 20049a305230SLee Schermerhorn /* 200510fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 20069a305230SLee Schermerhorn * No-op if no hstate attributes attached. 20079a305230SLee Schermerhorn */ 20083cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 20099a305230SLee Schermerhorn { 20109a305230SLee Schermerhorn struct hstate *h; 201110fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 20129a305230SLee Schermerhorn 20139a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 20149b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 20159a305230SLee Schermerhorn 2016972dc4deSAneesh Kumar K.V for_each_hstate(h) { 2017972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 2018972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 2019972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 2020972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 2021972dc4deSAneesh Kumar K.V } 20229a305230SLee Schermerhorn } 20239a305230SLee Schermerhorn 20249a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 20259a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 20269a305230SLee Schermerhorn } 20279a305230SLee Schermerhorn 20289a305230SLee Schermerhorn /* 202910fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 20309a305230SLee Schermerhorn * that have them. 20319a305230SLee Schermerhorn */ 20329a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 20339a305230SLee Schermerhorn { 20349a305230SLee Schermerhorn int nid; 20359a305230SLee Schermerhorn 20369a305230SLee Schermerhorn /* 203710fbcf4cSKay Sievers * disable node device registrations. 20389a305230SLee Schermerhorn */ 20399a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 20409a305230SLee Schermerhorn 20419a305230SLee Schermerhorn /* 20429a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 20439a305230SLee Schermerhorn */ 20449a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 20458732794bSWen Congyang hugetlb_unregister_node(node_devices[nid]); 20469a305230SLee Schermerhorn } 20479a305230SLee Schermerhorn 20489a305230SLee Schermerhorn /* 204910fbcf4cSKay Sievers * Register hstate attributes for a single node device. 20509a305230SLee Schermerhorn * No-op if attributes already registered. 20519a305230SLee Schermerhorn */ 20523cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 20539a305230SLee Schermerhorn { 20549a305230SLee Schermerhorn struct hstate *h; 205510fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 20569a305230SLee Schermerhorn int err; 20579a305230SLee Schermerhorn 20589a305230SLee Schermerhorn if (nhs->hugepages_kobj) 20599a305230SLee Schermerhorn return; /* already allocated */ 20609a305230SLee Schermerhorn 20619a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 206210fbcf4cSKay Sievers &node->dev.kobj); 20639a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 20649a305230SLee Schermerhorn return; 20659a305230SLee Schermerhorn 20669a305230SLee Schermerhorn for_each_hstate(h) { 20679a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 20689a305230SLee Schermerhorn nhs->hstate_kobjs, 20699a305230SLee Schermerhorn &per_node_hstate_attr_group); 20709a305230SLee Schermerhorn if (err) { 2071ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 207210fbcf4cSKay Sievers h->name, node->dev.id); 20739a305230SLee Schermerhorn hugetlb_unregister_node(node); 20749a305230SLee Schermerhorn break; 20759a305230SLee Schermerhorn } 20769a305230SLee Schermerhorn } 20779a305230SLee Schermerhorn } 20789a305230SLee Schermerhorn 20799a305230SLee Schermerhorn /* 20809b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 208110fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 208210fbcf4cSKay Sievers * registered their associated device by this time. 20839a305230SLee Schermerhorn */ 20849a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 20859a305230SLee Schermerhorn { 20869a305230SLee Schermerhorn int nid; 20879a305230SLee Schermerhorn 20888cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 20898732794bSWen Congyang struct node *node = node_devices[nid]; 209010fbcf4cSKay Sievers if (node->dev.id == nid) 20919a305230SLee Schermerhorn hugetlb_register_node(node); 20929a305230SLee Schermerhorn } 20939a305230SLee Schermerhorn 20949a305230SLee Schermerhorn /* 209510fbcf4cSKay Sievers * Let the node device driver know we're here so it can 20969a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 20979a305230SLee Schermerhorn */ 20989a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 20999a305230SLee Schermerhorn hugetlb_unregister_node); 21009a305230SLee Schermerhorn } 21019a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 21029a305230SLee Schermerhorn 21039a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 21049a305230SLee Schermerhorn { 21059a305230SLee Schermerhorn BUG(); 21069a305230SLee Schermerhorn if (nidp) 21079a305230SLee Schermerhorn *nidp = -1; 21089a305230SLee Schermerhorn return NULL; 21099a305230SLee Schermerhorn } 21109a305230SLee Schermerhorn 21119a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 21129a305230SLee Schermerhorn 21139a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 21149a305230SLee Schermerhorn 21159a305230SLee Schermerhorn #endif 21169a305230SLee Schermerhorn 2117a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 2118a3437870SNishanth Aravamudan { 2119a3437870SNishanth Aravamudan struct hstate *h; 2120a3437870SNishanth Aravamudan 21219a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 21229a305230SLee Schermerhorn 2123a3437870SNishanth Aravamudan for_each_hstate(h) { 2124972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 2125a3437870SNishanth Aravamudan } 2126a3437870SNishanth Aravamudan 2127a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 21288382d914SDavidlohr Bueso kfree(htlb_fault_mutex_table); 2129a3437870SNishanth Aravamudan } 2130a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 2131a3437870SNishanth Aravamudan 2132a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 2133a3437870SNishanth Aravamudan { 21348382d914SDavidlohr Bueso int i; 21358382d914SDavidlohr Bueso 2136457c1b27SNishanth Aravamudan if (!hugepages_supported()) 21370ef89d25SBenjamin Herrenschmidt return 0; 2138a3437870SNishanth Aravamudan 2139e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 2140e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 2141e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 2142a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 2143a3437870SNishanth Aravamudan } 2144972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 2145e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 2146e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 2147a3437870SNishanth Aravamudan 2148a3437870SNishanth Aravamudan hugetlb_init_hstates(); 2149aa888a74SAndi Kleen gather_bootmem_prealloc(); 2150a3437870SNishanth Aravamudan report_hugepages(); 2151a3437870SNishanth Aravamudan 2152a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 21539a305230SLee Schermerhorn hugetlb_register_all_nodes(); 21547179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 21559a305230SLee Schermerhorn 21568382d914SDavidlohr Bueso #ifdef CONFIG_SMP 21578382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 21588382d914SDavidlohr Bueso #else 21598382d914SDavidlohr Bueso num_fault_mutexes = 1; 21608382d914SDavidlohr Bueso #endif 21618382d914SDavidlohr Bueso htlb_fault_mutex_table = 21628382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 21638382d914SDavidlohr Bueso BUG_ON(!htlb_fault_mutex_table); 21648382d914SDavidlohr Bueso 21658382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 21668382d914SDavidlohr Bueso mutex_init(&htlb_fault_mutex_table[i]); 2167a3437870SNishanth Aravamudan return 0; 2168a3437870SNishanth Aravamudan } 2169a3437870SNishanth Aravamudan module_init(hugetlb_init); 2170a3437870SNishanth Aravamudan 2171a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 2172a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 2173a3437870SNishanth Aravamudan { 2174a3437870SNishanth Aravamudan struct hstate *h; 21758faa8b07SAndi Kleen unsigned long i; 21768faa8b07SAndi Kleen 2177a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2178ffb22af5SAndrew Morton pr_warning("hugepagesz= specified twice, ignoring\n"); 2179a3437870SNishanth Aravamudan return; 2180a3437870SNishanth Aravamudan } 218147d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2182a3437870SNishanth Aravamudan BUG_ON(order == 0); 218347d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2184a3437870SNishanth Aravamudan h->order = order; 2185a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 21868faa8b07SAndi Kleen h->nr_huge_pages = 0; 21878faa8b07SAndi Kleen h->free_huge_pages = 0; 21888faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 21898faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 21900edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 21918cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 21928cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2193a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2194a3437870SNishanth Aravamudan huge_page_size(h)/1024); 21958faa8b07SAndi Kleen 2196a3437870SNishanth Aravamudan parsed_hstate = h; 2197a3437870SNishanth Aravamudan } 2198a3437870SNishanth Aravamudan 2199e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2200a3437870SNishanth Aravamudan { 2201a3437870SNishanth Aravamudan unsigned long *mhp; 22028faa8b07SAndi Kleen static unsigned long *last_mhp; 2203a3437870SNishanth Aravamudan 2204a3437870SNishanth Aravamudan /* 220547d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2206a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2207a3437870SNishanth Aravamudan */ 220847d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 2209a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2210a3437870SNishanth Aravamudan else 2211a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2212a3437870SNishanth Aravamudan 22138faa8b07SAndi Kleen if (mhp == last_mhp) { 2214ffb22af5SAndrew Morton pr_warning("hugepages= specified twice without " 22158faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 22168faa8b07SAndi Kleen return 1; 22178faa8b07SAndi Kleen } 22188faa8b07SAndi Kleen 2219a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2220a3437870SNishanth Aravamudan *mhp = 0; 2221a3437870SNishanth Aravamudan 22228faa8b07SAndi Kleen /* 22238faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 22248faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 22258faa8b07SAndi Kleen * use the bootmem allocator. 22268faa8b07SAndi Kleen */ 222747d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 22288faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 22298faa8b07SAndi Kleen 22308faa8b07SAndi Kleen last_mhp = mhp; 22318faa8b07SAndi Kleen 2232a3437870SNishanth Aravamudan return 1; 2233a3437870SNishanth Aravamudan } 2234e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2235e11bfbfcSNick Piggin 2236e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2237e11bfbfcSNick Piggin { 2238e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2239e11bfbfcSNick Piggin return 1; 2240e11bfbfcSNick Piggin } 2241e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2242a3437870SNishanth Aravamudan 22438a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 22448a213460SNishanth Aravamudan { 22458a213460SNishanth Aravamudan int node; 22468a213460SNishanth Aravamudan unsigned int nr = 0; 22478a213460SNishanth Aravamudan 22488a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 22498a213460SNishanth Aravamudan nr += array[node]; 22508a213460SNishanth Aravamudan 22518a213460SNishanth Aravamudan return nr; 22528a213460SNishanth Aravamudan } 22538a213460SNishanth Aravamudan 22548a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 225506808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 225606808b08SLee Schermerhorn struct ctl_table *table, int write, 225706808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 22581da177e4SLinus Torvalds { 2259e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2260238d3c13SDavid Rientjes unsigned long tmp = h->max_huge_pages; 226108d4a246SMichal Hocko int ret; 2262e5ff2159SAndi Kleen 2263457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2264457c1b27SNishanth Aravamudan return -ENOTSUPP; 2265457c1b27SNishanth Aravamudan 2266e5ff2159SAndi Kleen table->data = &tmp; 2267e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 226808d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 226908d4a246SMichal Hocko if (ret) 227008d4a246SMichal Hocko goto out; 2271e5ff2159SAndi Kleen 2272238d3c13SDavid Rientjes if (write) 2273238d3c13SDavid Rientjes ret = __nr_hugepages_store_common(obey_mempolicy, h, 2274238d3c13SDavid Rientjes NUMA_NO_NODE, tmp, *length); 227508d4a246SMichal Hocko out: 227608d4a246SMichal Hocko return ret; 22771da177e4SLinus Torvalds } 2278396faf03SMel Gorman 227906808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 228006808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 228106808b08SLee Schermerhorn { 228206808b08SLee Schermerhorn 228306808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 228406808b08SLee Schermerhorn buffer, length, ppos); 228506808b08SLee Schermerhorn } 228606808b08SLee Schermerhorn 228706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 228806808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 228906808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 229006808b08SLee Schermerhorn { 229106808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 229206808b08SLee Schermerhorn buffer, length, ppos); 229306808b08SLee Schermerhorn } 229406808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 229506808b08SLee Schermerhorn 2296a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 22978d65af78SAlexey Dobriyan void __user *buffer, 2298a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2299a3d0c6aaSNishanth Aravamudan { 2300a5516438SAndi Kleen struct hstate *h = &default_hstate; 2301e5ff2159SAndi Kleen unsigned long tmp; 230208d4a246SMichal Hocko int ret; 2303e5ff2159SAndi Kleen 2304457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2305457c1b27SNishanth Aravamudan return -ENOTSUPP; 2306457c1b27SNishanth Aravamudan 2307e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2308e5ff2159SAndi Kleen 2309bae7f4aeSLuiz Capitulino if (write && hstate_is_gigantic(h)) 2310adbe8726SEric B Munson return -EINVAL; 2311adbe8726SEric B Munson 2312e5ff2159SAndi Kleen table->data = &tmp; 2313e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 231408d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 231508d4a246SMichal Hocko if (ret) 231608d4a246SMichal Hocko goto out; 2317e5ff2159SAndi Kleen 2318e5ff2159SAndi Kleen if (write) { 2319064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2320e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2321a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2322e5ff2159SAndi Kleen } 232308d4a246SMichal Hocko out: 232408d4a246SMichal Hocko return ret; 2325a3d0c6aaSNishanth Aravamudan } 2326a3d0c6aaSNishanth Aravamudan 23271da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 23281da177e4SLinus Torvalds 2329e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 23301da177e4SLinus Torvalds { 2331a5516438SAndi Kleen struct hstate *h = &default_hstate; 2332457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2333457c1b27SNishanth Aravamudan return; 2334e1759c21SAlexey Dobriyan seq_printf(m, 23351da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 23361da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2337b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 23387893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 23394f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2340a5516438SAndi Kleen h->nr_huge_pages, 2341a5516438SAndi Kleen h->free_huge_pages, 2342a5516438SAndi Kleen h->resv_huge_pages, 2343a5516438SAndi Kleen h->surplus_huge_pages, 2344a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 23451da177e4SLinus Torvalds } 23461da177e4SLinus Torvalds 23471da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 23481da177e4SLinus Torvalds { 2349a5516438SAndi Kleen struct hstate *h = &default_hstate; 2350457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2351457c1b27SNishanth Aravamudan return 0; 23521da177e4SLinus Torvalds return sprintf(buf, 23531da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2354a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2355a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2356a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2357a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2358a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 23591da177e4SLinus Torvalds } 23601da177e4SLinus Torvalds 2361949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2362949f7ec5SDavid Rientjes { 2363949f7ec5SDavid Rientjes struct hstate *h; 2364949f7ec5SDavid Rientjes int nid; 2365949f7ec5SDavid Rientjes 2366457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2367457c1b27SNishanth Aravamudan return; 2368457c1b27SNishanth Aravamudan 2369949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2370949f7ec5SDavid Rientjes for_each_hstate(h) 2371949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2372949f7ec5SDavid Rientjes nid, 2373949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2374949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2375949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2376949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2377949f7ec5SDavid Rientjes } 2378949f7ec5SDavid Rientjes 23791da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 23801da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 23811da177e4SLinus Torvalds { 2382d0028588SWanpeng Li struct hstate *h; 2383d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2384d0028588SWanpeng Li 2385d0028588SWanpeng Li for_each_hstate(h) 2386d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2387d0028588SWanpeng Li return nr_total_pages; 23881da177e4SLinus Torvalds } 23891da177e4SLinus Torvalds 2390a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2391fc1b8a73SMel Gorman { 2392fc1b8a73SMel Gorman int ret = -ENOMEM; 2393fc1b8a73SMel Gorman 2394fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2395fc1b8a73SMel Gorman /* 2396fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2397fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2398fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2399fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2400fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2401fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2402fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2403fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2404fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2405fc1b8a73SMel Gorman * 2406fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2407fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2408fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2409fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2410fc1b8a73SMel Gorman * semantics that cpuset has. 2411fc1b8a73SMel Gorman */ 2412fc1b8a73SMel Gorman if (delta > 0) { 2413a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2414fc1b8a73SMel Gorman goto out; 2415fc1b8a73SMel Gorman 2416a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2417a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2418fc1b8a73SMel Gorman goto out; 2419fc1b8a73SMel Gorman } 2420fc1b8a73SMel Gorman } 2421fc1b8a73SMel Gorman 2422fc1b8a73SMel Gorman ret = 0; 2423fc1b8a73SMel Gorman if (delta < 0) 2424a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2425fc1b8a73SMel Gorman 2426fc1b8a73SMel Gorman out: 2427fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2428fc1b8a73SMel Gorman return ret; 2429fc1b8a73SMel Gorman } 2430fc1b8a73SMel Gorman 243184afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 243284afd99bSAndy Whitcroft { 2433f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 243484afd99bSAndy Whitcroft 243584afd99bSAndy Whitcroft /* 243684afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 243784afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 243884afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 243925985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 244084afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 244184afd99bSAndy Whitcroft * new reference here without additional locking. 244284afd99bSAndy Whitcroft */ 24434e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2444f522c3acSJoonsoo Kim kref_get(&resv->refs); 244584afd99bSAndy Whitcroft } 244684afd99bSAndy Whitcroft 2447a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2448a1e78772SMel Gorman { 2449a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2450f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 245190481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 24524e35f483SJoonsoo Kim unsigned long reserve, start, end; 245384afd99bSAndy Whitcroft 24544e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 24554e35f483SJoonsoo Kim return; 24564e35f483SJoonsoo Kim 2457a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2458a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 245984afd99bSAndy Whitcroft 24604e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 246184afd99bSAndy Whitcroft 2462f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 246384afd99bSAndy Whitcroft 24647251ff78SAdam Litke if (reserve) { 2465a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 246690481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 24677251ff78SAdam Litke } 2468a1e78772SMel Gorman } 2469a1e78772SMel Gorman 24701da177e4SLinus Torvalds /* 24711da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 24721da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 24731da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 24741da177e4SLinus Torvalds * this far. 24751da177e4SLinus Torvalds */ 2476d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 24771da177e4SLinus Torvalds { 24781da177e4SLinus Torvalds BUG(); 2479d0217ac0SNick Piggin return 0; 24801da177e4SLinus Torvalds } 24811da177e4SLinus Torvalds 2482f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2483d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 248484afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2485a1e78772SMel Gorman .close = hugetlb_vm_op_close, 24861da177e4SLinus Torvalds }; 24871da177e4SLinus Torvalds 24881e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 24891e8f889bSDavid Gibson int writable) 249063551ae0SDavid Gibson { 249163551ae0SDavid Gibson pte_t entry; 249263551ae0SDavid Gibson 24931e8f889bSDavid Gibson if (writable) { 2494106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2495106c992aSGerald Schaefer vma->vm_page_prot))); 249663551ae0SDavid Gibson } else { 2497106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 2498106c992aSGerald Schaefer vma->vm_page_prot)); 249963551ae0SDavid Gibson } 250063551ae0SDavid Gibson entry = pte_mkyoung(entry); 250163551ae0SDavid Gibson entry = pte_mkhuge(entry); 2502d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 250363551ae0SDavid Gibson 250463551ae0SDavid Gibson return entry; 250563551ae0SDavid Gibson } 250663551ae0SDavid Gibson 25071e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 25081e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 25091e8f889bSDavid Gibson { 25101e8f889bSDavid Gibson pte_t entry; 25111e8f889bSDavid Gibson 2512106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 251332f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 25144b3073e1SRussell King update_mmu_cache(vma, address, ptep); 25151e8f889bSDavid Gibson } 25161e8f889bSDavid Gibson 25174a705fefSNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 25184a705fefSNaoya Horiguchi { 25194a705fefSNaoya Horiguchi swp_entry_t swp; 25204a705fefSNaoya Horiguchi 25214a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 25224a705fefSNaoya Horiguchi return 0; 25234a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 25244a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_migration_entry(swp)) 25254a705fefSNaoya Horiguchi return 1; 25264a705fefSNaoya Horiguchi else 25274a705fefSNaoya Horiguchi return 0; 25284a705fefSNaoya Horiguchi } 25294a705fefSNaoya Horiguchi 25304a705fefSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 25314a705fefSNaoya Horiguchi { 25324a705fefSNaoya Horiguchi swp_entry_t swp; 25334a705fefSNaoya Horiguchi 25344a705fefSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 25354a705fefSNaoya Horiguchi return 0; 25364a705fefSNaoya Horiguchi swp = pte_to_swp_entry(pte); 25374a705fefSNaoya Horiguchi if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 25384a705fefSNaoya Horiguchi return 1; 25394a705fefSNaoya Horiguchi else 25404a705fefSNaoya Horiguchi return 0; 25414a705fefSNaoya Horiguchi } 25421e8f889bSDavid Gibson 254363551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 254463551ae0SDavid Gibson struct vm_area_struct *vma) 254563551ae0SDavid Gibson { 254663551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 254763551ae0SDavid Gibson struct page *ptepage; 25481c59827dSHugh Dickins unsigned long addr; 25491e8f889bSDavid Gibson int cow; 2550a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2551a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2552e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 2553e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 2554e8569dd2SAndreas Sandberg int ret = 0; 25551e8f889bSDavid Gibson 25561e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 255763551ae0SDavid Gibson 2558e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 2559e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 2560e8569dd2SAndreas Sandberg if (cow) 2561e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 2562e8569dd2SAndreas Sandberg 2563a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2564cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 2565c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2566c74df32cSHugh Dickins if (!src_pte) 2567c74df32cSHugh Dickins continue; 2568a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 2569e8569dd2SAndreas Sandberg if (!dst_pte) { 2570e8569dd2SAndreas Sandberg ret = -ENOMEM; 2571e8569dd2SAndreas Sandberg break; 2572e8569dd2SAndreas Sandberg } 2573c5c99429SLarry Woodman 2574c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2575c5c99429SLarry Woodman if (dst_pte == src_pte) 2576c5c99429SLarry Woodman continue; 2577c5c99429SLarry Woodman 2578cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 2579cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 2580cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 25814a705fefSNaoya Horiguchi entry = huge_ptep_get(src_pte); 25824a705fefSNaoya Horiguchi if (huge_pte_none(entry)) { /* skip none entry */ 25834a705fefSNaoya Horiguchi ; 25844a705fefSNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_migration(entry) || 25854a705fefSNaoya Horiguchi is_hugetlb_entry_hwpoisoned(entry))) { 25864a705fefSNaoya Horiguchi swp_entry_t swp_entry = pte_to_swp_entry(entry); 25874a705fefSNaoya Horiguchi 25884a705fefSNaoya Horiguchi if (is_write_migration_entry(swp_entry) && cow) { 25894a705fefSNaoya Horiguchi /* 25904a705fefSNaoya Horiguchi * COW mappings require pages in both 25914a705fefSNaoya Horiguchi * parent and child to be set to read. 25924a705fefSNaoya Horiguchi */ 25934a705fefSNaoya Horiguchi make_migration_entry_read(&swp_entry); 25944a705fefSNaoya Horiguchi entry = swp_entry_to_pte(swp_entry); 25954a705fefSNaoya Horiguchi set_huge_pte_at(src, addr, src_pte, entry); 25964a705fefSNaoya Horiguchi } 25974a705fefSNaoya Horiguchi set_huge_pte_at(dst, addr, dst_pte, entry); 25984a705fefSNaoya Horiguchi } else { 25991e8f889bSDavid Gibson if (cow) 26007f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 26010253d634SNaoya Horiguchi entry = huge_ptep_get(src_pte); 260263551ae0SDavid Gibson ptepage = pte_page(entry); 260363551ae0SDavid Gibson get_page(ptepage); 26040fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 260563551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 26061c59827dSHugh Dickins } 2607cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 2608cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 260963551ae0SDavid Gibson } 261063551ae0SDavid Gibson 2611e8569dd2SAndreas Sandberg if (cow) 2612e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 2613e8569dd2SAndreas Sandberg 2614e8569dd2SAndreas Sandberg return ret; 261563551ae0SDavid Gibson } 261663551ae0SDavid Gibson 261724669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 261824669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 261924669e58SAneesh Kumar K.V struct page *ref_page) 262063551ae0SDavid Gibson { 262124669e58SAneesh Kumar K.V int force_flush = 0; 262263551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 262363551ae0SDavid Gibson unsigned long address; 2624c7546f8fSDavid Gibson pte_t *ptep; 262563551ae0SDavid Gibson pte_t pte; 2626cb900f41SKirill A. Shutemov spinlock_t *ptl; 262763551ae0SDavid Gibson struct page *page; 2628a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2629a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 26302ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 26312ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 2632a5516438SAndi Kleen 263363551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2634a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2635a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 263663551ae0SDavid Gibson 263724669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 26382ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 263924669e58SAneesh Kumar K.V again: 2640a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2641c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2642c7546f8fSDavid Gibson if (!ptep) 2643c7546f8fSDavid Gibson continue; 2644c7546f8fSDavid Gibson 2645cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 264639dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 2647cb900f41SKirill A. Shutemov goto unlock; 264839dde65cSChen, Kenneth W 26496629326bSHillf Danton pte = huge_ptep_get(ptep); 26506629326bSHillf Danton if (huge_pte_none(pte)) 2651cb900f41SKirill A. Shutemov goto unlock; 26526629326bSHillf Danton 26536629326bSHillf Danton /* 26546629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 26556629326bSHillf Danton */ 26568c4894c6SNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 2657106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 2658cb900f41SKirill A. Shutemov goto unlock; 26598c4894c6SNaoya Horiguchi } 26606629326bSHillf Danton 26616629326bSHillf Danton page = pte_page(pte); 266204f2cbe3SMel Gorman /* 266304f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 266404f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 266504f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 266604f2cbe3SMel Gorman */ 266704f2cbe3SMel Gorman if (ref_page) { 266804f2cbe3SMel Gorman if (page != ref_page) 2669cb900f41SKirill A. Shutemov goto unlock; 267004f2cbe3SMel Gorman 267104f2cbe3SMel Gorman /* 267204f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 267304f2cbe3SMel Gorman * future faults in this VMA will fail rather than 267404f2cbe3SMel Gorman * looking like data was lost 267504f2cbe3SMel Gorman */ 267604f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 267704f2cbe3SMel Gorman } 267804f2cbe3SMel Gorman 2679c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 268024669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 2681106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 26826649a386SKen Chen set_page_dirty(page); 26839e81130bSHillf Danton 268424669e58SAneesh Kumar K.V page_remove_rmap(page); 268524669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 2686cb900f41SKirill A. Shutemov if (force_flush) { 2687cb900f41SKirill A. Shutemov spin_unlock(ptl); 26889e81130bSHillf Danton break; 268963551ae0SDavid Gibson } 2690cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 2691cb900f41SKirill A. Shutemov if (ref_page) { 2692cb900f41SKirill A. Shutemov spin_unlock(ptl); 2693cb900f41SKirill A. Shutemov break; 2694cb900f41SKirill A. Shutemov } 2695cb900f41SKirill A. Shutemov unlock: 2696cb900f41SKirill A. Shutemov spin_unlock(ptl); 2697cb900f41SKirill A. Shutemov } 269824669e58SAneesh Kumar K.V /* 269924669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 270024669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 270124669e58SAneesh Kumar K.V * and page-free while holding it. 270224669e58SAneesh Kumar K.V */ 270324669e58SAneesh Kumar K.V if (force_flush) { 270424669e58SAneesh Kumar K.V force_flush = 0; 270524669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 270624669e58SAneesh Kumar K.V if (address < end && !ref_page) 270724669e58SAneesh Kumar K.V goto again; 2708fe1668aeSChen, Kenneth W } 27092ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 271024669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 27111da177e4SLinus Torvalds } 271263551ae0SDavid Gibson 2713d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2714d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2715d833352aSMel Gorman unsigned long end, struct page *ref_page) 2716d833352aSMel Gorman { 2717d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2718d833352aSMel Gorman 2719d833352aSMel Gorman /* 2720d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2721d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2722d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2723d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2724d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2725d833352aSMel Gorman * is to clear it before releasing the i_mmap_mutex. This works 2726d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2727d833352aSMel Gorman * destroyed and the i_mmap_mutex is held. 2728d833352aSMel Gorman */ 2729d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2730d833352aSMel Gorman } 2731d833352aSMel Gorman 2732502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 273304f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2734502717f4SChen, Kenneth W { 273524669e58SAneesh Kumar K.V struct mm_struct *mm; 273624669e58SAneesh Kumar K.V struct mmu_gather tlb; 273724669e58SAneesh Kumar K.V 273824669e58SAneesh Kumar K.V mm = vma->vm_mm; 273924669e58SAneesh Kumar K.V 27402b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 274124669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 274224669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2743502717f4SChen, Kenneth W } 2744502717f4SChen, Kenneth W 274504f2cbe3SMel Gorman /* 274604f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 274704f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 274804f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 274904f2cbe3SMel Gorman * same region. 275004f2cbe3SMel Gorman */ 27512f4612afSDavidlohr Bueso static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 27522a4b3dedSHarvey Harrison struct page *page, unsigned long address) 275304f2cbe3SMel Gorman { 27547526674dSAdam Litke struct hstate *h = hstate_vma(vma); 275504f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 275604f2cbe3SMel Gorman struct address_space *mapping; 275704f2cbe3SMel Gorman pgoff_t pgoff; 275804f2cbe3SMel Gorman 275904f2cbe3SMel Gorman /* 276004f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 276104f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 276204f2cbe3SMel Gorman */ 27637526674dSAdam Litke address = address & huge_page_mask(h); 276436e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 276536e4f20aSMichal Hocko vma->vm_pgoff; 2766496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 276704f2cbe3SMel Gorman 27684eb2b1dcSMel Gorman /* 27694eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 27704eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 27714eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 27724eb2b1dcSMel Gorman */ 27733d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 27746b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 277504f2cbe3SMel Gorman /* Do not unmap the current VMA */ 277604f2cbe3SMel Gorman if (iter_vma == vma) 277704f2cbe3SMel Gorman continue; 277804f2cbe3SMel Gorman 277904f2cbe3SMel Gorman /* 278004f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 278104f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 278204f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 278304f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 278404f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 278504f2cbe3SMel Gorman */ 278604f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 278724669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 278824669e58SAneesh Kumar K.V address + huge_page_size(h), page); 278904f2cbe3SMel Gorman } 27903d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 279104f2cbe3SMel Gorman } 279204f2cbe3SMel Gorman 27930fe6e20bSNaoya Horiguchi /* 27940fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2795ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2796ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2797ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 27980fe6e20bSNaoya Horiguchi */ 27991e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 280004f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 2801cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 28021e8f889bSDavid Gibson { 2803a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 28041e8f889bSDavid Gibson struct page *old_page, *new_page; 2805ad4404a2SDavidlohr Bueso int ret = 0, outside_reserve = 0; 28062ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 28072ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 28081e8f889bSDavid Gibson 28091e8f889bSDavid Gibson old_page = pte_page(pte); 28101e8f889bSDavid Gibson 281104f2cbe3SMel Gorman retry_avoidcopy: 28121e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 28131e8f889bSDavid Gibson * and just make the page writable */ 281437a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 28150fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 28161e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 281783c54070SNick Piggin return 0; 28181e8f889bSDavid Gibson } 28191e8f889bSDavid Gibson 282004f2cbe3SMel Gorman /* 282104f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 282204f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 282304f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 282404f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 282504f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 282604f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 282704f2cbe3SMel Gorman * of the full address range. 282804f2cbe3SMel Gorman */ 28295944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 283004f2cbe3SMel Gorman old_page != pagecache_page) 283104f2cbe3SMel Gorman outside_reserve = 1; 283204f2cbe3SMel Gorman 28331e8f889bSDavid Gibson page_cache_get(old_page); 2834b76c8cfbSLarry Woodman 2835ad4404a2SDavidlohr Bueso /* 2836ad4404a2SDavidlohr Bueso * Drop page table lock as buddy allocator may be called. It will 2837ad4404a2SDavidlohr Bueso * be acquired again before returning to the caller, as expected. 2838ad4404a2SDavidlohr Bueso */ 2839cb900f41SKirill A. Shutemov spin_unlock(ptl); 284004f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 28411e8f889bSDavid Gibson 28422fc39cecSAdam Litke if (IS_ERR(new_page)) { 284304f2cbe3SMel Gorman /* 284404f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 284504f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 284604f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 284704f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 284804f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 284904f2cbe3SMel Gorman */ 285004f2cbe3SMel Gorman if (outside_reserve) { 2851ad4404a2SDavidlohr Bueso page_cache_release(old_page); 285204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 28532f4612afSDavidlohr Bueso unmap_ref_private(mm, vma, old_page, address); 285404f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2855cb900f41SKirill A. Shutemov spin_lock(ptl); 2856a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2857a9af0c5dSNaoya Horiguchi if (likely(ptep && 2858a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 285904f2cbe3SMel Gorman goto retry_avoidcopy; 2860a734bcc8SHillf Danton /* 2861cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 2862cb900f41SKirill A. Shutemov * lock, and our job is done. 2863a734bcc8SHillf Danton */ 2864a734bcc8SHillf Danton return 0; 286504f2cbe3SMel Gorman } 286604f2cbe3SMel Gorman 2867ad4404a2SDavidlohr Bueso ret = (PTR_ERR(new_page) == -ENOMEM) ? 2868ad4404a2SDavidlohr Bueso VM_FAULT_OOM : VM_FAULT_SIGBUS; 2869ad4404a2SDavidlohr Bueso goto out_release_old; 28701e8f889bSDavid Gibson } 28711e8f889bSDavid Gibson 28720fe6e20bSNaoya Horiguchi /* 28730fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 28740fe6e20bSNaoya Horiguchi * anon_vma prepared. 28750fe6e20bSNaoya Horiguchi */ 287644e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2877ad4404a2SDavidlohr Bueso ret = VM_FAULT_OOM; 2878ad4404a2SDavidlohr Bueso goto out_release_all; 287944e2aa93SDean Nelson } 28800fe6e20bSNaoya Horiguchi 288147ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 288247ad8475SAndrea Arcangeli pages_per_huge_page(h)); 28830ed361deSNick Piggin __SetPageUptodate(new_page); 28841e8f889bSDavid Gibson 28852ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 28862ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 28872ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2888ad4404a2SDavidlohr Bueso 2889b76c8cfbSLarry Woodman /* 2890cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 2891b76c8cfbSLarry Woodman * before the page tables are altered 2892b76c8cfbSLarry Woodman */ 2893cb900f41SKirill A. Shutemov spin_lock(ptl); 2894a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2895a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 289607443a85SJoonsoo Kim ClearPagePrivate(new_page); 289707443a85SJoonsoo Kim 28981e8f889bSDavid Gibson /* Break COW */ 28998fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 29001e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 29011e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 29020fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2903cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 29041e8f889bSDavid Gibson /* Make the old page be freed below */ 29051e8f889bSDavid Gibson new_page = old_page; 29061e8f889bSDavid Gibson } 2907cb900f41SKirill A. Shutemov spin_unlock(ptl); 29082ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 2909ad4404a2SDavidlohr Bueso out_release_all: 29101e8f889bSDavid Gibson page_cache_release(new_page); 2911ad4404a2SDavidlohr Bueso out_release_old: 29121e8f889bSDavid Gibson page_cache_release(old_page); 29138312034fSJoonsoo Kim 2914ad4404a2SDavidlohr Bueso spin_lock(ptl); /* Caller expects lock to be held */ 2915ad4404a2SDavidlohr Bueso return ret; 29161e8f889bSDavid Gibson } 29171e8f889bSDavid Gibson 291804f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2919a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2920a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 292104f2cbe3SMel Gorman { 292204f2cbe3SMel Gorman struct address_space *mapping; 2923e7c4b0bfSAndy Whitcroft pgoff_t idx; 292404f2cbe3SMel Gorman 292504f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2926a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 292704f2cbe3SMel Gorman 292804f2cbe3SMel Gorman return find_lock_page(mapping, idx); 292904f2cbe3SMel Gorman } 293004f2cbe3SMel Gorman 29313ae77f43SHugh Dickins /* 29323ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 29333ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 29343ae77f43SHugh Dickins */ 29353ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 29362a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 29372a15efc9SHugh Dickins { 29382a15efc9SHugh Dickins struct address_space *mapping; 29392a15efc9SHugh Dickins pgoff_t idx; 29402a15efc9SHugh Dickins struct page *page; 29412a15efc9SHugh Dickins 29422a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 29432a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 29442a15efc9SHugh Dickins 29452a15efc9SHugh Dickins page = find_get_page(mapping, idx); 29462a15efc9SHugh Dickins if (page) 29472a15efc9SHugh Dickins put_page(page); 29482a15efc9SHugh Dickins return page != NULL; 29492a15efc9SHugh Dickins } 29502a15efc9SHugh Dickins 2951a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 29528382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 2953788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2954ac9b9c66SHugh Dickins { 2955a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2956ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2957409eb8c2SHillf Danton int anon_rmap = 0; 29584c887265SAdam Litke unsigned long size; 29594c887265SAdam Litke struct page *page; 29601e8f889bSDavid Gibson pte_t new_pte; 2961cb900f41SKirill A. Shutemov spinlock_t *ptl; 29624c887265SAdam Litke 296304f2cbe3SMel Gorman /* 296404f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 296504f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 296625985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 296704f2cbe3SMel Gorman */ 296804f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2969ffb22af5SAndrew Morton pr_warning("PID %d killed due to inadequate hugepage pool\n", 297004f2cbe3SMel Gorman current->pid); 297104f2cbe3SMel Gorman return ret; 297204f2cbe3SMel Gorman } 297304f2cbe3SMel Gorman 29744c887265SAdam Litke /* 29754c887265SAdam Litke * Use page lock to guard against racing truncation 29764c887265SAdam Litke * before we get page_table_lock. 29774c887265SAdam Litke */ 29786bda666aSChristoph Lameter retry: 29796bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 29806bda666aSChristoph Lameter if (!page) { 2981a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2982ebed4bfcSHugh Dickins if (idx >= size) 2983ebed4bfcSHugh Dickins goto out; 298404f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 29852fc39cecSAdam Litke if (IS_ERR(page)) { 298676dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 298776dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 298876dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 298976dcee75SAneesh Kumar K.V else 299076dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 29916bda666aSChristoph Lameter goto out; 29926bda666aSChristoph Lameter } 299347ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 29940ed361deSNick Piggin __SetPageUptodate(page); 2995ac9b9c66SHugh Dickins 2996f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 29976bda666aSChristoph Lameter int err; 299845c682a6SKen Chen struct inode *inode = mapping->host; 29996bda666aSChristoph Lameter 30006bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 30016bda666aSChristoph Lameter if (err) { 30026bda666aSChristoph Lameter put_page(page); 30036bda666aSChristoph Lameter if (err == -EEXIST) 30046bda666aSChristoph Lameter goto retry; 30056bda666aSChristoph Lameter goto out; 30066bda666aSChristoph Lameter } 300707443a85SJoonsoo Kim ClearPagePrivate(page); 300845c682a6SKen Chen 300945c682a6SKen Chen spin_lock(&inode->i_lock); 3010a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 301145c682a6SKen Chen spin_unlock(&inode->i_lock); 301223be7468SMel Gorman } else { 30136bda666aSChristoph Lameter lock_page(page); 30140fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 30150fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 30160fe6e20bSNaoya Horiguchi goto backout_unlocked; 301723be7468SMel Gorman } 3018409eb8c2SHillf Danton anon_rmap = 1; 30190fe6e20bSNaoya Horiguchi } 30200fe6e20bSNaoya Horiguchi } else { 302157303d80SAndy Whitcroft /* 3022998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 3023998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 3024998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 3025fd6a03edSNaoya Horiguchi */ 3026fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 3027aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 3028972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3029fd6a03edSNaoya Horiguchi goto backout_unlocked; 30306bda666aSChristoph Lameter } 3031998b4382SNaoya Horiguchi } 30321e8f889bSDavid Gibson 303357303d80SAndy Whitcroft /* 303457303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 303557303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 303657303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 303757303d80SAndy Whitcroft * the spinlock. 303857303d80SAndy Whitcroft */ 3039788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 30402b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 30412b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 30422b26736cSAndy Whitcroft goto backout_unlocked; 30432b26736cSAndy Whitcroft } 304457303d80SAndy Whitcroft 3045cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3046cb900f41SKirill A. Shutemov spin_lock(ptl); 3047a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 30484c887265SAdam Litke if (idx >= size) 30494c887265SAdam Litke goto backout; 30504c887265SAdam Litke 305183c54070SNick Piggin ret = 0; 30527f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 30534c887265SAdam Litke goto backout; 30544c887265SAdam Litke 305507443a85SJoonsoo Kim if (anon_rmap) { 305607443a85SJoonsoo Kim ClearPagePrivate(page); 3057409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 3058ac714904SChoi Gi-yong } else 3059409eb8c2SHillf Danton page_dup_rmap(page); 30601e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 30611e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 30621e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 30631e8f889bSDavid Gibson 3064788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 30651e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 3066cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 30671e8f889bSDavid Gibson } 30681e8f889bSDavid Gibson 3069cb900f41SKirill A. Shutemov spin_unlock(ptl); 30704c887265SAdam Litke unlock_page(page); 30714c887265SAdam Litke out: 3072ac9b9c66SHugh Dickins return ret; 30734c887265SAdam Litke 30744c887265SAdam Litke backout: 3075cb900f41SKirill A. Shutemov spin_unlock(ptl); 30762b26736cSAndy Whitcroft backout_unlocked: 30774c887265SAdam Litke unlock_page(page); 30784c887265SAdam Litke put_page(page); 30794c887265SAdam Litke goto out; 3080ac9b9c66SHugh Dickins } 3081ac9b9c66SHugh Dickins 30828382d914SDavidlohr Bueso #ifdef CONFIG_SMP 30838382d914SDavidlohr Bueso static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 30848382d914SDavidlohr Bueso struct vm_area_struct *vma, 30858382d914SDavidlohr Bueso struct address_space *mapping, 30868382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 30878382d914SDavidlohr Bueso { 30888382d914SDavidlohr Bueso unsigned long key[2]; 30898382d914SDavidlohr Bueso u32 hash; 30908382d914SDavidlohr Bueso 30918382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 30928382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 30938382d914SDavidlohr Bueso key[1] = idx; 30948382d914SDavidlohr Bueso } else { 30958382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 30968382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 30978382d914SDavidlohr Bueso } 30988382d914SDavidlohr Bueso 30998382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 31008382d914SDavidlohr Bueso 31018382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 31028382d914SDavidlohr Bueso } 31038382d914SDavidlohr Bueso #else 31048382d914SDavidlohr Bueso /* 31058382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 31068382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 31078382d914SDavidlohr Bueso */ 31088382d914SDavidlohr Bueso static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 31098382d914SDavidlohr Bueso struct vm_area_struct *vma, 31108382d914SDavidlohr Bueso struct address_space *mapping, 31118382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 31128382d914SDavidlohr Bueso { 31138382d914SDavidlohr Bueso return 0; 31148382d914SDavidlohr Bueso } 31158382d914SDavidlohr Bueso #endif 31168382d914SDavidlohr Bueso 311786e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 3118788c7df4SHugh Dickins unsigned long address, unsigned int flags) 311986e5216fSAdam Litke { 31208382d914SDavidlohr Bueso pte_t *ptep, entry; 3121cb900f41SKirill A. Shutemov spinlock_t *ptl; 31221e8f889bSDavid Gibson int ret; 31238382d914SDavidlohr Bueso u32 hash; 31248382d914SDavidlohr Bueso pgoff_t idx; 31250fe6e20bSNaoya Horiguchi struct page *page = NULL; 312657303d80SAndy Whitcroft struct page *pagecache_page = NULL; 3127a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 31288382d914SDavidlohr Bueso struct address_space *mapping; 312986e5216fSAdam Litke 31301e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 31311e16a539SKAMEZAWA Hiroyuki 3132fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 3133fd6a03edSNaoya Horiguchi if (ptep) { 3134fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 3135290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 3136cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 3137290408d4SNaoya Horiguchi return 0; 3138290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 3139aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 3140972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 3141fd6a03edSNaoya Horiguchi } 3142fd6a03edSNaoya Horiguchi 3143a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 314486e5216fSAdam Litke if (!ptep) 314586e5216fSAdam Litke return VM_FAULT_OOM; 314686e5216fSAdam Litke 31478382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 31488382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 31498382d914SDavidlohr Bueso 31503935baa9SDavid Gibson /* 31513935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 31523935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 31533935baa9SDavid Gibson * the same page in the page cache. 31543935baa9SDavid Gibson */ 31558382d914SDavidlohr Bueso hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); 31568382d914SDavidlohr Bueso mutex_lock(&htlb_fault_mutex_table[hash]); 31578382d914SDavidlohr Bueso 31587f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 31597f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 31608382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3161b4d1d99fSDavid Gibson goto out_mutex; 31623935baa9SDavid Gibson } 316386e5216fSAdam Litke 316483c54070SNick Piggin ret = 0; 31651e8f889bSDavid Gibson 316657303d80SAndy Whitcroft /* 316757303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 316857303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 316957303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 317057303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 317157303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 317257303d80SAndy Whitcroft * consumed. 317357303d80SAndy Whitcroft */ 3174106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 31752b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 31762b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3177b4d1d99fSDavid Gibson goto out_mutex; 31782b26736cSAndy Whitcroft } 317957303d80SAndy Whitcroft 3180f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 318157303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 318257303d80SAndy Whitcroft vma, address); 318357303d80SAndy Whitcroft } 318457303d80SAndy Whitcroft 318556c9cfb1SNaoya Horiguchi /* 318656c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 318756c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 318856c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 318956c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 319056c9cfb1SNaoya Horiguchi * so no worry about deadlock. 319156c9cfb1SNaoya Horiguchi */ 31920fe6e20bSNaoya Horiguchi page = pte_page(entry); 319366aebce7SChris Metcalf get_page(page); 319456c9cfb1SNaoya Horiguchi if (page != pagecache_page) 31950fe6e20bSNaoya Horiguchi lock_page(page); 31960fe6e20bSNaoya Horiguchi 3197cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3198cb900f41SKirill A. Shutemov spin_lock(ptl); 31991e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3200b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3201cb900f41SKirill A. Shutemov goto out_ptl; 3202b4d1d99fSDavid Gibson 3203b4d1d99fSDavid Gibson 3204788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3205106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 320657303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3207cb900f41SKirill A. Shutemov pagecache_page, ptl); 3208cb900f41SKirill A. Shutemov goto out_ptl; 3209b4d1d99fSDavid Gibson } 3210106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3211b4d1d99fSDavid Gibson } 3212b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3213788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3214788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 32154b3073e1SRussell King update_mmu_cache(vma, address, ptep); 3216b4d1d99fSDavid Gibson 3217cb900f41SKirill A. Shutemov out_ptl: 3218cb900f41SKirill A. Shutemov spin_unlock(ptl); 321957303d80SAndy Whitcroft 322057303d80SAndy Whitcroft if (pagecache_page) { 322157303d80SAndy Whitcroft unlock_page(pagecache_page); 322257303d80SAndy Whitcroft put_page(pagecache_page); 322357303d80SAndy Whitcroft } 32241f64d69cSDean Nelson if (page != pagecache_page) 322556c9cfb1SNaoya Horiguchi unlock_page(page); 322666aebce7SChris Metcalf put_page(page); 322757303d80SAndy Whitcroft 3228b4d1d99fSDavid Gibson out_mutex: 32298382d914SDavidlohr Bueso mutex_unlock(&htlb_fault_mutex_table[hash]); 32301e8f889bSDavid Gibson return ret; 323186e5216fSAdam Litke } 323286e5216fSAdam Litke 323328a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 323463551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 323528a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 323628a35716SMichel Lespinasse long i, unsigned int flags) 323763551ae0SDavid Gibson { 3238d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3239d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 324028a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3241a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 324263551ae0SDavid Gibson 324363551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 324463551ae0SDavid Gibson pte_t *pte; 3245cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 32462a15efc9SHugh Dickins int absent; 324763551ae0SDavid Gibson struct page *page; 324863551ae0SDavid Gibson 32494c887265SAdam Litke /* 32504c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 32512a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 32524c887265SAdam Litke * first, for the page indexing below to work. 3253cb900f41SKirill A. Shutemov * 3254cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 32554c887265SAdam Litke */ 3256a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3257cb900f41SKirill A. Shutemov if (pte) 3258cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 32592a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 326063551ae0SDavid Gibson 32612a15efc9SHugh Dickins /* 32622a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 32633ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 32643ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 32653ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 32663ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 32672a15efc9SHugh Dickins */ 32683ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 32693ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3270cb900f41SKirill A. Shutemov if (pte) 3271cb900f41SKirill A. Shutemov spin_unlock(ptl); 32722a15efc9SHugh Dickins remainder = 0; 32732a15efc9SHugh Dickins break; 32742a15efc9SHugh Dickins } 32752a15efc9SHugh Dickins 32769cc3a5bdSNaoya Horiguchi /* 32779cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 32789cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 32799cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 32809cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 32819cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 32829cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 32839cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 32849cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 32859cc3a5bdSNaoya Horiguchi */ 32869cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3287106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3288106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 32894c887265SAdam Litke int ret; 32904c887265SAdam Litke 3291cb900f41SKirill A. Shutemov if (pte) 3292cb900f41SKirill A. Shutemov spin_unlock(ptl); 32932a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 32942a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3295a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 32964c887265SAdam Litke continue; 32974c887265SAdam Litke 32981c59827dSHugh Dickins remainder = 0; 32991c59827dSHugh Dickins break; 33001c59827dSHugh Dickins } 330163551ae0SDavid Gibson 3302a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 33037f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3304d5d4b0aaSChen, Kenneth W same_page: 3305d6692183SChen, Kenneth W if (pages) { 330669d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3307a0368d4eSAndrea Arcangeli get_page_foll(pages[i]); 3308d6692183SChen, Kenneth W } 330963551ae0SDavid Gibson 331063551ae0SDavid Gibson if (vmas) 331163551ae0SDavid Gibson vmas[i] = vma; 331263551ae0SDavid Gibson 331363551ae0SDavid Gibson vaddr += PAGE_SIZE; 3314d5d4b0aaSChen, Kenneth W ++pfn_offset; 331563551ae0SDavid Gibson --remainder; 331663551ae0SDavid Gibson ++i; 3317d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3318a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3319d5d4b0aaSChen, Kenneth W /* 3320d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3321d5d4b0aaSChen, Kenneth W * of this compound page. 3322d5d4b0aaSChen, Kenneth W */ 3323d5d4b0aaSChen, Kenneth W goto same_page; 3324d5d4b0aaSChen, Kenneth W } 3325cb900f41SKirill A. Shutemov spin_unlock(ptl); 332663551ae0SDavid Gibson } 332728a35716SMichel Lespinasse *nr_pages = remainder; 332863551ae0SDavid Gibson *position = vaddr; 332963551ae0SDavid Gibson 33302a15efc9SHugh Dickins return i ? i : -EFAULT; 333163551ae0SDavid Gibson } 33328f860591SZhang, Yanmin 33337da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 33348f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 33358f860591SZhang, Yanmin { 33368f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 33378f860591SZhang, Yanmin unsigned long start = address; 33388f860591SZhang, Yanmin pte_t *ptep; 33398f860591SZhang, Yanmin pte_t pte; 3340a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 33417da4d641SPeter Zijlstra unsigned long pages = 0; 33428f860591SZhang, Yanmin 33438f860591SZhang, Yanmin BUG_ON(address >= end); 33448f860591SZhang, Yanmin flush_cache_range(vma, address, end); 33458f860591SZhang, Yanmin 3346a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 33473d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 3348a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3349cb900f41SKirill A. Shutemov spinlock_t *ptl; 33508f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 33518f860591SZhang, Yanmin if (!ptep) 33528f860591SZhang, Yanmin continue; 3353cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 33547da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 33557da4d641SPeter Zijlstra pages++; 3356cb900f41SKirill A. Shutemov spin_unlock(ptl); 335739dde65cSChen, Kenneth W continue; 33587da4d641SPeter Zijlstra } 33597f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 33608f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3361106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3362be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 33638f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 33647da4d641SPeter Zijlstra pages++; 33658f860591SZhang, Yanmin } 3366cb900f41SKirill A. Shutemov spin_unlock(ptl); 33678f860591SZhang, Yanmin } 3368d833352aSMel Gorman /* 3369d833352aSMel Gorman * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3370d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3371d833352aSMel Gorman * once we release i_mmap_mutex, another task can do the final put_page 3372d833352aSMel Gorman * and that page table be reused and filled with junk. 3373d833352aSMel Gorman */ 33748f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 3375d833352aSMel Gorman mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 3376a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 33777da4d641SPeter Zijlstra 33787da4d641SPeter Zijlstra return pages << h->order; 33798f860591SZhang, Yanmin } 33808f860591SZhang, Yanmin 3381a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3382a1e78772SMel Gorman long from, long to, 33835a6fe125SMel Gorman struct vm_area_struct *vma, 3384ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3385e4e574b7SAdam Litke { 338617c9d12eSMel Gorman long ret, chg; 3387a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 338890481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 33899119a41eSJoonsoo Kim struct resv_map *resv_map; 3390e4e574b7SAdam Litke 3391a1e78772SMel Gorman /* 339217c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 339317c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 339490481622SDavid Gibson * without using reserves 339517c9d12eSMel Gorman */ 3396ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 339717c9d12eSMel Gorman return 0; 339817c9d12eSMel Gorman 339917c9d12eSMel Gorman /* 3400a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3401a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3402a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3403a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3404a1e78772SMel Gorman */ 34059119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 34064e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 34079119a41eSJoonsoo Kim 34081406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 34099119a41eSJoonsoo Kim 34109119a41eSJoonsoo Kim } else { 34119119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 34125a6fe125SMel Gorman if (!resv_map) 34135a6fe125SMel Gorman return -ENOMEM; 34145a6fe125SMel Gorman 341517c9d12eSMel Gorman chg = to - from; 341617c9d12eSMel Gorman 34175a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 34185a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 34195a6fe125SMel Gorman } 34205a6fe125SMel Gorman 3421c50ac050SDave Hansen if (chg < 0) { 3422c50ac050SDave Hansen ret = chg; 3423c50ac050SDave Hansen goto out_err; 3424c50ac050SDave Hansen } 342517c9d12eSMel Gorman 342690481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3427c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3428c50ac050SDave Hansen ret = -ENOSPC; 3429c50ac050SDave Hansen goto out_err; 3430c50ac050SDave Hansen } 343117c9d12eSMel Gorman 343217c9d12eSMel Gorman /* 343317c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 343490481622SDavid Gibson * Hand the pages back to the subpool if there are not 343517c9d12eSMel Gorman */ 343617c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 343717c9d12eSMel Gorman if (ret < 0) { 343890481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3439c50ac050SDave Hansen goto out_err; 344017c9d12eSMel Gorman } 344117c9d12eSMel Gorman 344217c9d12eSMel Gorman /* 344317c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 344417c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 344517c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 344617c9d12eSMel Gorman * the reservation was and the page cache tells how much of 344717c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 344817c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 344917c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 345017c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 345117c9d12eSMel Gorman * else has to be done for private mappings here 345217c9d12eSMel Gorman */ 3453f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 34541406ec9bSJoonsoo Kim region_add(resv_map, from, to); 3455a43a8c39SChen, Kenneth W return 0; 3456c50ac050SDave Hansen out_err: 3457f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3458f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 3459c50ac050SDave Hansen return ret; 3460a43a8c39SChen, Kenneth W } 3461a43a8c39SChen, Kenneth W 3462a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3463a43a8c39SChen, Kenneth W { 3464a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 34654e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 34669119a41eSJoonsoo Kim long chg = 0; 346790481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 346845c682a6SKen Chen 34699119a41eSJoonsoo Kim if (resv_map) 34701406ec9bSJoonsoo Kim chg = region_truncate(resv_map, offset); 347145c682a6SKen Chen spin_lock(&inode->i_lock); 3472e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 347345c682a6SKen Chen spin_unlock(&inode->i_lock); 347445c682a6SKen Chen 347590481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3476a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3477a43a8c39SChen, Kenneth W } 347893f70f90SNaoya Horiguchi 34793212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 34803212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 34813212b535SSteve Capper struct vm_area_struct *vma, 34823212b535SSteve Capper unsigned long addr, pgoff_t idx) 34833212b535SSteve Capper { 34843212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 34853212b535SSteve Capper svma->vm_start; 34863212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 34873212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 34883212b535SSteve Capper 34893212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 34903212b535SSteve Capper unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 34913212b535SSteve Capper unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 34923212b535SSteve Capper 34933212b535SSteve Capper /* 34943212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 34953212b535SSteve Capper * page table page. 34963212b535SSteve Capper */ 34973212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 34983212b535SSteve Capper vm_flags != svm_flags || 34993212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 35003212b535SSteve Capper return 0; 35013212b535SSteve Capper 35023212b535SSteve Capper return saddr; 35033212b535SSteve Capper } 35043212b535SSteve Capper 35053212b535SSteve Capper static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 35063212b535SSteve Capper { 35073212b535SSteve Capper unsigned long base = addr & PUD_MASK; 35083212b535SSteve Capper unsigned long end = base + PUD_SIZE; 35093212b535SSteve Capper 35103212b535SSteve Capper /* 35113212b535SSteve Capper * check on proper vm_flags and page table alignment 35123212b535SSteve Capper */ 35133212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 35143212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 35153212b535SSteve Capper return 1; 35163212b535SSteve Capper return 0; 35173212b535SSteve Capper } 35183212b535SSteve Capper 35193212b535SSteve Capper /* 35203212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 35213212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 35223212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 35233212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 35243212b535SSteve Capper * pud has to be populated inside the same i_mmap_mutex section - otherwise 35253212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 35263212b535SSteve Capper * bad pmd for sharing. 35273212b535SSteve Capper */ 35283212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 35293212b535SSteve Capper { 35303212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 35313212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 35323212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 35333212b535SSteve Capper vma->vm_pgoff; 35343212b535SSteve Capper struct vm_area_struct *svma; 35353212b535SSteve Capper unsigned long saddr; 35363212b535SSteve Capper pte_t *spte = NULL; 35373212b535SSteve Capper pte_t *pte; 3538cb900f41SKirill A. Shutemov spinlock_t *ptl; 35393212b535SSteve Capper 35403212b535SSteve Capper if (!vma_shareable(vma, addr)) 35413212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 35423212b535SSteve Capper 35433212b535SSteve Capper mutex_lock(&mapping->i_mmap_mutex); 35443212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 35453212b535SSteve Capper if (svma == vma) 35463212b535SSteve Capper continue; 35473212b535SSteve Capper 35483212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 35493212b535SSteve Capper if (saddr) { 35503212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 35513212b535SSteve Capper if (spte) { 35523212b535SSteve Capper get_page(virt_to_page(spte)); 35533212b535SSteve Capper break; 35543212b535SSteve Capper } 35553212b535SSteve Capper } 35563212b535SSteve Capper } 35573212b535SSteve Capper 35583212b535SSteve Capper if (!spte) 35593212b535SSteve Capper goto out; 35603212b535SSteve Capper 3561cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 3562cb900f41SKirill A. Shutemov spin_lock(ptl); 35633212b535SSteve Capper if (pud_none(*pud)) 35643212b535SSteve Capper pud_populate(mm, pud, 35653212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 35663212b535SSteve Capper else 35673212b535SSteve Capper put_page(virt_to_page(spte)); 3568cb900f41SKirill A. Shutemov spin_unlock(ptl); 35693212b535SSteve Capper out: 35703212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 35713212b535SSteve Capper mutex_unlock(&mapping->i_mmap_mutex); 35723212b535SSteve Capper return pte; 35733212b535SSteve Capper } 35743212b535SSteve Capper 35753212b535SSteve Capper /* 35763212b535SSteve Capper * unmap huge page backed by shared pte. 35773212b535SSteve Capper * 35783212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 35793212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 35803212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 35813212b535SSteve Capper * 3582cb900f41SKirill A. Shutemov * called with page table lock held. 35833212b535SSteve Capper * 35843212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 35853212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 35863212b535SSteve Capper */ 35873212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 35883212b535SSteve Capper { 35893212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 35903212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 35913212b535SSteve Capper 35923212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 35933212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 35943212b535SSteve Capper return 0; 35953212b535SSteve Capper 35963212b535SSteve Capper pud_clear(pud); 35973212b535SSteve Capper put_page(virt_to_page(ptep)); 35983212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 35993212b535SSteve Capper return 1; 36003212b535SSteve Capper } 36019e5fc74cSSteve Capper #define want_pmd_share() (1) 36029e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 36039e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 36049e5fc74cSSteve Capper { 36059e5fc74cSSteve Capper return NULL; 36069e5fc74cSSteve Capper } 36079e5fc74cSSteve Capper #define want_pmd_share() (0) 36083212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 36093212b535SSteve Capper 36109e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 36119e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 36129e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 36139e5fc74cSSteve Capper { 36149e5fc74cSSteve Capper pgd_t *pgd; 36159e5fc74cSSteve Capper pud_t *pud; 36169e5fc74cSSteve Capper pte_t *pte = NULL; 36179e5fc74cSSteve Capper 36189e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 36199e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 36209e5fc74cSSteve Capper if (pud) { 36219e5fc74cSSteve Capper if (sz == PUD_SIZE) { 36229e5fc74cSSteve Capper pte = (pte_t *)pud; 36239e5fc74cSSteve Capper } else { 36249e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 36259e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 36269e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 36279e5fc74cSSteve Capper else 36289e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 36299e5fc74cSSteve Capper } 36309e5fc74cSSteve Capper } 36319e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 36329e5fc74cSSteve Capper 36339e5fc74cSSteve Capper return pte; 36349e5fc74cSSteve Capper } 36359e5fc74cSSteve Capper 36369e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 36379e5fc74cSSteve Capper { 36389e5fc74cSSteve Capper pgd_t *pgd; 36399e5fc74cSSteve Capper pud_t *pud; 36409e5fc74cSSteve Capper pmd_t *pmd = NULL; 36419e5fc74cSSteve Capper 36429e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 36439e5fc74cSSteve Capper if (pgd_present(*pgd)) { 36449e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 36459e5fc74cSSteve Capper if (pud_present(*pud)) { 36469e5fc74cSSteve Capper if (pud_huge(*pud)) 36479e5fc74cSSteve Capper return (pte_t *)pud; 36489e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 36499e5fc74cSSteve Capper } 36509e5fc74cSSteve Capper } 36519e5fc74cSSteve Capper return (pte_t *) pmd; 36529e5fc74cSSteve Capper } 36539e5fc74cSSteve Capper 36549e5fc74cSSteve Capper struct page * 36559e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 36569e5fc74cSSteve Capper pmd_t *pmd, int write) 36579e5fc74cSSteve Capper { 36589e5fc74cSSteve Capper struct page *page; 36599e5fc74cSSteve Capper 36609e5fc74cSSteve Capper page = pte_page(*(pte_t *)pmd); 36619e5fc74cSSteve Capper if (page) 36629e5fc74cSSteve Capper page += ((address & ~PMD_MASK) >> PAGE_SHIFT); 36639e5fc74cSSteve Capper return page; 36649e5fc74cSSteve Capper } 36659e5fc74cSSteve Capper 36669e5fc74cSSteve Capper struct page * 36679e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 36689e5fc74cSSteve Capper pud_t *pud, int write) 36699e5fc74cSSteve Capper { 36709e5fc74cSSteve Capper struct page *page; 36719e5fc74cSSteve Capper 36729e5fc74cSSteve Capper page = pte_page(*(pte_t *)pud); 36739e5fc74cSSteve Capper if (page) 36749e5fc74cSSteve Capper page += ((address & ~PUD_MASK) >> PAGE_SHIFT); 36759e5fc74cSSteve Capper return page; 36769e5fc74cSSteve Capper } 36779e5fc74cSSteve Capper 36789e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 36799e5fc74cSSteve Capper 36809e5fc74cSSteve Capper /* Can be overriden by architectures */ 36813b32123dSGideon Israel Dsouza struct page * __weak 36829e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 36839e5fc74cSSteve Capper pud_t *pud, int write) 36849e5fc74cSSteve Capper { 36859e5fc74cSSteve Capper BUG(); 36869e5fc74cSSteve Capper return NULL; 36879e5fc74cSSteve Capper } 36889e5fc74cSSteve Capper 36899e5fc74cSSteve Capper #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 36909e5fc74cSSteve Capper 3691d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3692d5bd9106SAndi Kleen 36936de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 36946de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 36956de2b1aaSNaoya Horiguchi { 36966de2b1aaSNaoya Horiguchi struct page *page; 36976de2b1aaSNaoya Horiguchi struct page *tmp; 36986de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 36996de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 37006de2b1aaSNaoya Horiguchi 37016de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 37026de2b1aaSNaoya Horiguchi if (page == hpage) 37036de2b1aaSNaoya Horiguchi return 1; 37046de2b1aaSNaoya Horiguchi return 0; 37056de2b1aaSNaoya Horiguchi } 37066de2b1aaSNaoya Horiguchi 370793f70f90SNaoya Horiguchi /* 370893f70f90SNaoya Horiguchi * This function is called from memory failure code. 370993f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 371093f70f90SNaoya Horiguchi */ 37116de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 371293f70f90SNaoya Horiguchi { 371393f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 371493f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 37156de2b1aaSNaoya Horiguchi int ret = -EBUSY; 371693f70f90SNaoya Horiguchi 371793f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 37186de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 371956f2fb14SNaoya Horiguchi /* 372056f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 372156f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 372256f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 372356f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 372456f2fb14SNaoya Horiguchi */ 372556f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 37268c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 372793f70f90SNaoya Horiguchi h->free_huge_pages--; 372893f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 37296de2b1aaSNaoya Horiguchi ret = 0; 373093f70f90SNaoya Horiguchi } 37316de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 37326de2b1aaSNaoya Horiguchi return ret; 37336de2b1aaSNaoya Horiguchi } 37346de2b1aaSNaoya Horiguchi #endif 373531caf665SNaoya Horiguchi 373631caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 373731caf665SNaoya Horiguchi { 3738309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 373931caf665SNaoya Horiguchi if (!get_page_unless_zero(page)) 374031caf665SNaoya Horiguchi return false; 374131caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 374231caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 374331caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 374431caf665SNaoya Horiguchi return true; 374531caf665SNaoya Horiguchi } 374631caf665SNaoya Horiguchi 374731caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 374831caf665SNaoya Horiguchi { 3749309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 375031caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 375131caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 375231caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 375331caf665SNaoya Horiguchi put_page(page); 375431caf665SNaoya Horiguchi } 3755c8721bbbSNaoya Horiguchi 3756c8721bbbSNaoya Horiguchi bool is_hugepage_active(struct page *page) 3757c8721bbbSNaoya Horiguchi { 3758309381feSSasha Levin VM_BUG_ON_PAGE(!PageHuge(page), page); 3759c8721bbbSNaoya Horiguchi /* 3760c8721bbbSNaoya Horiguchi * This function can be called for a tail page because the caller, 3761c8721bbbSNaoya Horiguchi * scan_movable_pages, scans through a given pfn-range which typically 3762c8721bbbSNaoya Horiguchi * covers one memory block. In systems using gigantic hugepage (1GB 3763c8721bbbSNaoya Horiguchi * for x86_64,) a hugepage is larger than a memory block, and we don't 3764c8721bbbSNaoya Horiguchi * support migrating such large hugepages for now, so return false 3765c8721bbbSNaoya Horiguchi * when called for tail pages. 3766c8721bbbSNaoya Horiguchi */ 3767c8721bbbSNaoya Horiguchi if (PageTail(page)) 3768c8721bbbSNaoya Horiguchi return false; 3769c8721bbbSNaoya Horiguchi /* 3770c8721bbbSNaoya Horiguchi * Refcount of a hwpoisoned hugepages is 1, but they are not active, 3771c8721bbbSNaoya Horiguchi * so we should return false for them. 3772c8721bbbSNaoya Horiguchi */ 3773c8721bbbSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 3774c8721bbbSNaoya Horiguchi return false; 3775c8721bbbSNaoya Horiguchi return page_count(page) > 0; 3776c8721bbbSNaoya Horiguchi } 3777