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 { 54786cdb465SNaoya Horiguchi if (hugepages_treat_as_movable || hugepage_migration_support(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 610a5516438SAndi Kleen static void update_and_free_page(struct hstate *h, struct page *page) 6116af2acb6SAdam Litke { 6126af2acb6SAdam Litke int i; 613a5516438SAndi Kleen 61418229df5SAndy Whitcroft VM_BUG_ON(h->order >= MAX_ORDER); 61518229df5SAndy Whitcroft 616a5516438SAndi Kleen h->nr_huge_pages--; 617a5516438SAndi Kleen h->nr_huge_pages_node[page_to_nid(page)]--; 618a5516438SAndi Kleen for (i = 0; i < pages_per_huge_page(h); i++) { 61932f84528SChris Forbes page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 62032f84528SChris Forbes 1 << PG_referenced | 1 << PG_dirty | 62132f84528SChris Forbes 1 << PG_active | 1 << PG_reserved | 6226af2acb6SAdam Litke 1 << PG_private | 1 << PG_writeback); 6236af2acb6SAdam Litke } 624309381feSSasha Levin VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); 6256af2acb6SAdam Litke set_compound_page_dtor(page, NULL); 6266af2acb6SAdam Litke set_page_refcounted(page); 6277f2e9525SGerald Schaefer arch_release_hugepage(page); 628a5516438SAndi Kleen __free_pages(page, huge_page_order(h)); 6296af2acb6SAdam Litke } 6306af2acb6SAdam Litke 631e5ff2159SAndi Kleen struct hstate *size_to_hstate(unsigned long size) 632e5ff2159SAndi Kleen { 633e5ff2159SAndi Kleen struct hstate *h; 634e5ff2159SAndi Kleen 635e5ff2159SAndi Kleen for_each_hstate(h) { 636e5ff2159SAndi Kleen if (huge_page_size(h) == size) 637e5ff2159SAndi Kleen return h; 638e5ff2159SAndi Kleen } 639e5ff2159SAndi Kleen return NULL; 640e5ff2159SAndi Kleen } 641e5ff2159SAndi Kleen 64227a85ef1SDavid Gibson static void free_huge_page(struct page *page) 64327a85ef1SDavid Gibson { 644a5516438SAndi Kleen /* 645a5516438SAndi Kleen * Can't pass hstate in here because it is called from the 646a5516438SAndi Kleen * compound page destructor. 647a5516438SAndi Kleen */ 648e5ff2159SAndi Kleen struct hstate *h = page_hstate(page); 6497893d1d5SAdam Litke int nid = page_to_nid(page); 65090481622SDavid Gibson struct hugepage_subpool *spool = 65190481622SDavid Gibson (struct hugepage_subpool *)page_private(page); 65207443a85SJoonsoo Kim bool restore_reserve; 65327a85ef1SDavid Gibson 654e5df70abSAndy Whitcroft set_page_private(page, 0); 65523be7468SMel Gorman page->mapping = NULL; 6567893d1d5SAdam Litke BUG_ON(page_count(page)); 6570fe6e20bSNaoya Horiguchi BUG_ON(page_mapcount(page)); 65807443a85SJoonsoo Kim restore_reserve = PagePrivate(page); 65916c794b4SJoonsoo Kim ClearPagePrivate(page); 66027a85ef1SDavid Gibson 66127a85ef1SDavid Gibson spin_lock(&hugetlb_lock); 6626d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_page(hstate_index(h), 6636d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), page); 66407443a85SJoonsoo Kim if (restore_reserve) 66507443a85SJoonsoo Kim h->resv_huge_pages++; 66607443a85SJoonsoo Kim 667aa888a74SAndi Kleen if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { 6680edaecfaSAneesh Kumar K.V /* remove the page from active list */ 6690edaecfaSAneesh Kumar K.V list_del(&page->lru); 670a5516438SAndi Kleen update_and_free_page(h, page); 671a5516438SAndi Kleen h->surplus_huge_pages--; 672a5516438SAndi Kleen h->surplus_huge_pages_node[nid]--; 6737893d1d5SAdam Litke } else { 6745d3a551cSWill Deacon arch_clear_hugepage_flags(page); 675a5516438SAndi Kleen enqueue_huge_page(h, page); 6767893d1d5SAdam Litke } 67727a85ef1SDavid Gibson spin_unlock(&hugetlb_lock); 67890481622SDavid Gibson hugepage_subpool_put_pages(spool, 1); 67927a85ef1SDavid Gibson } 68027a85ef1SDavid Gibson 681a5516438SAndi Kleen static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) 682b7ba30c6SAndi Kleen { 6830edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 684b7ba30c6SAndi Kleen set_compound_page_dtor(page, free_huge_page); 685b7ba30c6SAndi Kleen spin_lock(&hugetlb_lock); 6869dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 687a5516438SAndi Kleen h->nr_huge_pages++; 688a5516438SAndi Kleen h->nr_huge_pages_node[nid]++; 689b7ba30c6SAndi Kleen spin_unlock(&hugetlb_lock); 690b7ba30c6SAndi Kleen put_page(page); /* free it into the hugepage allocator */ 691b7ba30c6SAndi Kleen } 692b7ba30c6SAndi Kleen 693f412c97aSDavid Rientjes static void __init prep_compound_gigantic_page(struct page *page, 694f412c97aSDavid Rientjes unsigned long order) 69520a0307cSWu Fengguang { 69620a0307cSWu Fengguang int i; 69720a0307cSWu Fengguang int nr_pages = 1 << order; 69820a0307cSWu Fengguang struct page *p = page + 1; 69920a0307cSWu Fengguang 70020a0307cSWu Fengguang /* we rely on prep_new_huge_page to set the destructor */ 70120a0307cSWu Fengguang set_compound_order(page, order); 70220a0307cSWu Fengguang __SetPageHead(page); 703ef5a22beSAndrea Arcangeli __ClearPageReserved(page); 70420a0307cSWu Fengguang for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { 70520a0307cSWu Fengguang __SetPageTail(p); 706ef5a22beSAndrea Arcangeli /* 707ef5a22beSAndrea Arcangeli * For gigantic hugepages allocated through bootmem at 708ef5a22beSAndrea Arcangeli * boot, it's safer to be consistent with the not-gigantic 709ef5a22beSAndrea Arcangeli * hugepages and clear the PG_reserved bit from all tail pages 710ef5a22beSAndrea Arcangeli * too. Otherwse drivers using get_user_pages() to access tail 711ef5a22beSAndrea Arcangeli * pages may get the reference counting wrong if they see 712ef5a22beSAndrea Arcangeli * PG_reserved set on a tail page (despite the head page not 713ef5a22beSAndrea Arcangeli * having PG_reserved set). Enforcing this consistency between 714ef5a22beSAndrea Arcangeli * head and tail pages allows drivers to optimize away a check 715ef5a22beSAndrea Arcangeli * on the head page when they need know if put_page() is needed 716ef5a22beSAndrea Arcangeli * after get_user_pages(). 717ef5a22beSAndrea Arcangeli */ 718ef5a22beSAndrea Arcangeli __ClearPageReserved(p); 71958a84aa9SYouquan Song set_page_count(p, 0); 72020a0307cSWu Fengguang p->first_page = page; 72120a0307cSWu Fengguang } 72220a0307cSWu Fengguang } 72320a0307cSWu Fengguang 7247795912cSAndrew Morton /* 7257795912cSAndrew Morton * PageHuge() only returns true for hugetlbfs pages, but not for normal or 7267795912cSAndrew Morton * transparent huge pages. See the PageTransHuge() documentation for more 7277795912cSAndrew Morton * details. 7287795912cSAndrew Morton */ 72920a0307cSWu Fengguang int PageHuge(struct page *page) 73020a0307cSWu Fengguang { 73120a0307cSWu Fengguang if (!PageCompound(page)) 73220a0307cSWu Fengguang return 0; 73320a0307cSWu Fengguang 73420a0307cSWu Fengguang page = compound_head(page); 735758f66a2SAndrew Morton return get_compound_page_dtor(page) == free_huge_page; 73620a0307cSWu Fengguang } 73743131e14SNaoya Horiguchi EXPORT_SYMBOL_GPL(PageHuge); 73843131e14SNaoya Horiguchi 73927c73ae7SAndrea Arcangeli /* 74027c73ae7SAndrea Arcangeli * PageHeadHuge() only returns true for hugetlbfs head page, but not for 74127c73ae7SAndrea Arcangeli * normal or transparent huge pages. 74227c73ae7SAndrea Arcangeli */ 74327c73ae7SAndrea Arcangeli int PageHeadHuge(struct page *page_head) 74427c73ae7SAndrea Arcangeli { 74527c73ae7SAndrea Arcangeli if (!PageHead(page_head)) 74627c73ae7SAndrea Arcangeli return 0; 74727c73ae7SAndrea Arcangeli 748758f66a2SAndrew Morton return get_compound_page_dtor(page_head) == free_huge_page; 74927c73ae7SAndrea Arcangeli } 75027c73ae7SAndrea Arcangeli 75113d60f4bSZhang Yi pgoff_t __basepage_index(struct page *page) 75213d60f4bSZhang Yi { 75313d60f4bSZhang Yi struct page *page_head = compound_head(page); 75413d60f4bSZhang Yi pgoff_t index = page_index(page_head); 75513d60f4bSZhang Yi unsigned long compound_idx; 75613d60f4bSZhang Yi 75713d60f4bSZhang Yi if (!PageHuge(page_head)) 75813d60f4bSZhang Yi return page_index(page); 75913d60f4bSZhang Yi 76013d60f4bSZhang Yi if (compound_order(page_head) >= MAX_ORDER) 76113d60f4bSZhang Yi compound_idx = page_to_pfn(page) - page_to_pfn(page_head); 76213d60f4bSZhang Yi else 76313d60f4bSZhang Yi compound_idx = page - page_head; 76413d60f4bSZhang Yi 76513d60f4bSZhang Yi return (index << compound_order(page_head)) + compound_idx; 76613d60f4bSZhang Yi } 76713d60f4bSZhang Yi 768a5516438SAndi Kleen static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) 7691da177e4SLinus Torvalds { 7701da177e4SLinus Torvalds struct page *page; 771f96efd58SJoe Jin 772aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 773aa888a74SAndi Kleen return NULL; 774aa888a74SAndi Kleen 7756484eb3eSMel Gorman page = alloc_pages_exact_node(nid, 77686cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 777551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 778a5516438SAndi Kleen huge_page_order(h)); 7791da177e4SLinus Torvalds if (page) { 7807f2e9525SGerald Schaefer if (arch_prepare_hugepage(page)) { 781caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 7827b8ee84dSHarvey Harrison return NULL; 7837f2e9525SGerald Schaefer } 784a5516438SAndi Kleen prep_new_huge_page(h, page, nid); 7851da177e4SLinus Torvalds } 78663b4613cSNishanth Aravamudan 78763b4613cSNishanth Aravamudan return page; 78863b4613cSNishanth Aravamudan } 78963b4613cSNishanth Aravamudan 7905ced66c9SAndi Kleen /* 7916ae11b27SLee Schermerhorn * common helper functions for hstate_next_node_to_{alloc|free}. 7926ae11b27SLee Schermerhorn * We may have allocated or freed a huge page based on a different 7936ae11b27SLee Schermerhorn * nodes_allowed previously, so h->next_node_to_{alloc|free} might 7946ae11b27SLee Schermerhorn * be outside of *nodes_allowed. Ensure that we use an allowed 7956ae11b27SLee Schermerhorn * node for alloc or free. 7969a76db09SLee Schermerhorn */ 7976ae11b27SLee Schermerhorn static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 7989a76db09SLee Schermerhorn { 7996ae11b27SLee Schermerhorn nid = next_node(nid, *nodes_allowed); 8009a76db09SLee Schermerhorn if (nid == MAX_NUMNODES) 8016ae11b27SLee Schermerhorn nid = first_node(*nodes_allowed); 8029a76db09SLee Schermerhorn VM_BUG_ON(nid >= MAX_NUMNODES); 8039a76db09SLee Schermerhorn 8049a76db09SLee Schermerhorn return nid; 8059a76db09SLee Schermerhorn } 8069a76db09SLee Schermerhorn 8076ae11b27SLee Schermerhorn static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 8085ced66c9SAndi Kleen { 8096ae11b27SLee Schermerhorn if (!node_isset(nid, *nodes_allowed)) 8106ae11b27SLee Schermerhorn nid = next_node_allowed(nid, nodes_allowed); 8119a76db09SLee Schermerhorn return nid; 8125ced66c9SAndi Kleen } 8135ced66c9SAndi Kleen 8146ae11b27SLee Schermerhorn /* 8156ae11b27SLee Schermerhorn * returns the previously saved node ["this node"] from which to 8166ae11b27SLee Schermerhorn * allocate a persistent huge page for the pool and advance the 8176ae11b27SLee Schermerhorn * next node from which to allocate, handling wrap at end of node 8186ae11b27SLee Schermerhorn * mask. 8196ae11b27SLee Schermerhorn */ 8206ae11b27SLee Schermerhorn static int hstate_next_node_to_alloc(struct hstate *h, 8216ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 8226ae11b27SLee Schermerhorn { 8236ae11b27SLee Schermerhorn int nid; 8246ae11b27SLee Schermerhorn 8256ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 8266ae11b27SLee Schermerhorn 8276ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); 8286ae11b27SLee Schermerhorn h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); 8296ae11b27SLee Schermerhorn 8306ae11b27SLee Schermerhorn return nid; 8316ae11b27SLee Schermerhorn } 8326ae11b27SLee Schermerhorn 833e8c5c824SLee Schermerhorn /* 8346ae11b27SLee Schermerhorn * helper for free_pool_huge_page() - return the previously saved 8356ae11b27SLee Schermerhorn * node ["this node"] from which to free a huge page. Advance the 8366ae11b27SLee Schermerhorn * next node id whether or not we find a free huge page to free so 8376ae11b27SLee Schermerhorn * that the next attempt to free addresses the next node. 838e8c5c824SLee Schermerhorn */ 8396ae11b27SLee Schermerhorn static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 840e8c5c824SLee Schermerhorn { 8416ae11b27SLee Schermerhorn int nid; 8429a76db09SLee Schermerhorn 8436ae11b27SLee Schermerhorn VM_BUG_ON(!nodes_allowed); 8446ae11b27SLee Schermerhorn 8456ae11b27SLee Schermerhorn nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 8466ae11b27SLee Schermerhorn h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 8476ae11b27SLee Schermerhorn 8489a76db09SLee Schermerhorn return nid; 849e8c5c824SLee Schermerhorn } 850e8c5c824SLee Schermerhorn 851b2261026SJoonsoo Kim #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ 852b2261026SJoonsoo Kim for (nr_nodes = nodes_weight(*mask); \ 853b2261026SJoonsoo Kim nr_nodes > 0 && \ 854b2261026SJoonsoo Kim ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ 855b2261026SJoonsoo Kim nr_nodes--) 856b2261026SJoonsoo Kim 857b2261026SJoonsoo Kim #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 858b2261026SJoonsoo Kim for (nr_nodes = nodes_weight(*mask); \ 859b2261026SJoonsoo Kim nr_nodes > 0 && \ 860b2261026SJoonsoo Kim ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 861b2261026SJoonsoo Kim nr_nodes--) 862b2261026SJoonsoo Kim 863b2261026SJoonsoo Kim static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) 864b2261026SJoonsoo Kim { 865b2261026SJoonsoo Kim struct page *page; 866b2261026SJoonsoo Kim int nr_nodes, node; 867b2261026SJoonsoo Kim int ret = 0; 868b2261026SJoonsoo Kim 869b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 870b2261026SJoonsoo Kim page = alloc_fresh_huge_page_node(h, node); 871b2261026SJoonsoo Kim if (page) { 872b2261026SJoonsoo Kim ret = 1; 873b2261026SJoonsoo Kim break; 874b2261026SJoonsoo Kim } 875b2261026SJoonsoo Kim } 876b2261026SJoonsoo Kim 877b2261026SJoonsoo Kim if (ret) 878b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC); 879b2261026SJoonsoo Kim else 880b2261026SJoonsoo Kim count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 881b2261026SJoonsoo Kim 882b2261026SJoonsoo Kim return ret; 883b2261026SJoonsoo Kim } 884b2261026SJoonsoo Kim 885e8c5c824SLee Schermerhorn /* 886e8c5c824SLee Schermerhorn * Free huge page from pool from next node to free. 887e8c5c824SLee Schermerhorn * Attempt to keep persistent huge pages more or less 888e8c5c824SLee Schermerhorn * balanced over allowed nodes. 889e8c5c824SLee Schermerhorn * Called with hugetlb_lock locked. 890e8c5c824SLee Schermerhorn */ 8916ae11b27SLee Schermerhorn static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, 8926ae11b27SLee Schermerhorn bool acct_surplus) 893e8c5c824SLee Schermerhorn { 894b2261026SJoonsoo Kim int nr_nodes, node; 895e8c5c824SLee Schermerhorn int ret = 0; 896e8c5c824SLee Schermerhorn 897b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 898685f3457SLee Schermerhorn /* 899685f3457SLee Schermerhorn * If we're returning unused surplus pages, only examine 900685f3457SLee Schermerhorn * nodes with surplus pages. 901685f3457SLee Schermerhorn */ 902b2261026SJoonsoo Kim if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 903b2261026SJoonsoo Kim !list_empty(&h->hugepage_freelists[node])) { 904e8c5c824SLee Schermerhorn struct page *page = 905b2261026SJoonsoo Kim list_entry(h->hugepage_freelists[node].next, 906e8c5c824SLee Schermerhorn struct page, lru); 907e8c5c824SLee Schermerhorn list_del(&page->lru); 908e8c5c824SLee Schermerhorn h->free_huge_pages--; 909b2261026SJoonsoo Kim h->free_huge_pages_node[node]--; 910685f3457SLee Schermerhorn if (acct_surplus) { 911685f3457SLee Schermerhorn h->surplus_huge_pages--; 912b2261026SJoonsoo Kim h->surplus_huge_pages_node[node]--; 913685f3457SLee Schermerhorn } 914e8c5c824SLee Schermerhorn update_and_free_page(h, page); 915e8c5c824SLee Schermerhorn ret = 1; 9169a76db09SLee Schermerhorn break; 917e8c5c824SLee Schermerhorn } 918b2261026SJoonsoo Kim } 919e8c5c824SLee Schermerhorn 920e8c5c824SLee Schermerhorn return ret; 921e8c5c824SLee Schermerhorn } 922e8c5c824SLee Schermerhorn 923c8721bbbSNaoya Horiguchi /* 924c8721bbbSNaoya Horiguchi * Dissolve a given free hugepage into free buddy pages. This function does 925c8721bbbSNaoya Horiguchi * nothing for in-use (including surplus) hugepages. 926c8721bbbSNaoya Horiguchi */ 927c8721bbbSNaoya Horiguchi static void dissolve_free_huge_page(struct page *page) 928c8721bbbSNaoya Horiguchi { 929c8721bbbSNaoya Horiguchi spin_lock(&hugetlb_lock); 930c8721bbbSNaoya Horiguchi if (PageHuge(page) && !page_count(page)) { 931c8721bbbSNaoya Horiguchi struct hstate *h = page_hstate(page); 932c8721bbbSNaoya Horiguchi int nid = page_to_nid(page); 933c8721bbbSNaoya Horiguchi list_del(&page->lru); 934c8721bbbSNaoya Horiguchi h->free_huge_pages--; 935c8721bbbSNaoya Horiguchi h->free_huge_pages_node[nid]--; 936c8721bbbSNaoya Horiguchi update_and_free_page(h, page); 937c8721bbbSNaoya Horiguchi } 938c8721bbbSNaoya Horiguchi spin_unlock(&hugetlb_lock); 939c8721bbbSNaoya Horiguchi } 940c8721bbbSNaoya Horiguchi 941c8721bbbSNaoya Horiguchi /* 942c8721bbbSNaoya Horiguchi * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 943c8721bbbSNaoya Horiguchi * make specified memory blocks removable from the system. 944c8721bbbSNaoya Horiguchi * Note that start_pfn should aligned with (minimum) hugepage size. 945c8721bbbSNaoya Horiguchi */ 946c8721bbbSNaoya Horiguchi void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) 947c8721bbbSNaoya Horiguchi { 948c8721bbbSNaoya Horiguchi unsigned int order = 8 * sizeof(void *); 949c8721bbbSNaoya Horiguchi unsigned long pfn; 950c8721bbbSNaoya Horiguchi struct hstate *h; 951c8721bbbSNaoya Horiguchi 952c8721bbbSNaoya Horiguchi /* Set scan step to minimum hugepage size */ 953c8721bbbSNaoya Horiguchi for_each_hstate(h) 954c8721bbbSNaoya Horiguchi if (order > huge_page_order(h)) 955c8721bbbSNaoya Horiguchi order = huge_page_order(h); 956c8721bbbSNaoya Horiguchi VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); 957c8721bbbSNaoya Horiguchi for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) 958c8721bbbSNaoya Horiguchi dissolve_free_huge_page(pfn_to_page(pfn)); 959c8721bbbSNaoya Horiguchi } 960c8721bbbSNaoya Horiguchi 961bf50bab2SNaoya Horiguchi static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) 9627893d1d5SAdam Litke { 9637893d1d5SAdam Litke struct page *page; 964bf50bab2SNaoya Horiguchi unsigned int r_nid; 9657893d1d5SAdam Litke 966aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 967aa888a74SAndi Kleen return NULL; 968aa888a74SAndi Kleen 969d1c3fb1fSNishanth Aravamudan /* 970d1c3fb1fSNishanth Aravamudan * Assume we will successfully allocate the surplus page to 971d1c3fb1fSNishanth Aravamudan * prevent racing processes from causing the surplus to exceed 972d1c3fb1fSNishanth Aravamudan * overcommit 973d1c3fb1fSNishanth Aravamudan * 974d1c3fb1fSNishanth Aravamudan * This however introduces a different race, where a process B 975d1c3fb1fSNishanth Aravamudan * tries to grow the static hugepage pool while alloc_pages() is 976d1c3fb1fSNishanth Aravamudan * called by process A. B will only examine the per-node 977d1c3fb1fSNishanth Aravamudan * counters in determining if surplus huge pages can be 978d1c3fb1fSNishanth Aravamudan * converted to normal huge pages in adjust_pool_surplus(). A 979d1c3fb1fSNishanth Aravamudan * won't be able to increment the per-node counter, until the 980d1c3fb1fSNishanth Aravamudan * lock is dropped by B, but B doesn't drop hugetlb_lock until 981d1c3fb1fSNishanth Aravamudan * no more huge pages can be converted from surplus to normal 982d1c3fb1fSNishanth Aravamudan * state (and doesn't try to convert again). Thus, we have a 983d1c3fb1fSNishanth Aravamudan * case where a surplus huge page exists, the pool is grown, and 984d1c3fb1fSNishanth Aravamudan * the surplus huge page still exists after, even though it 985d1c3fb1fSNishanth Aravamudan * should just have been converted to a normal huge page. This 986d1c3fb1fSNishanth Aravamudan * does not leak memory, though, as the hugepage will be freed 987d1c3fb1fSNishanth Aravamudan * once it is out of use. It also does not allow the counters to 988d1c3fb1fSNishanth Aravamudan * go out of whack in adjust_pool_surplus() as we don't modify 989d1c3fb1fSNishanth Aravamudan * the node values until we've gotten the hugepage and only the 990d1c3fb1fSNishanth Aravamudan * per-node value is checked there. 991d1c3fb1fSNishanth Aravamudan */ 992d1c3fb1fSNishanth Aravamudan spin_lock(&hugetlb_lock); 993a5516438SAndi Kleen if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 994d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 995d1c3fb1fSNishanth Aravamudan return NULL; 996d1c3fb1fSNishanth Aravamudan } else { 997a5516438SAndi Kleen h->nr_huge_pages++; 998a5516438SAndi Kleen h->surplus_huge_pages++; 999d1c3fb1fSNishanth Aravamudan } 1000d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 1001d1c3fb1fSNishanth Aravamudan 1002bf50bab2SNaoya Horiguchi if (nid == NUMA_NO_NODE) 100386cdb465SNaoya Horiguchi page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| 1004551883aeSNishanth Aravamudan __GFP_REPEAT|__GFP_NOWARN, 1005a5516438SAndi Kleen huge_page_order(h)); 1006bf50bab2SNaoya Horiguchi else 1007bf50bab2SNaoya Horiguchi page = alloc_pages_exact_node(nid, 100886cdb465SNaoya Horiguchi htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| 1009bf50bab2SNaoya Horiguchi __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); 1010d1c3fb1fSNishanth Aravamudan 1011caff3a2cSGerald Schaefer if (page && arch_prepare_hugepage(page)) { 1012caff3a2cSGerald Schaefer __free_pages(page, huge_page_order(h)); 1013ea5768c7SHillf Danton page = NULL; 1014caff3a2cSGerald Schaefer } 1015caff3a2cSGerald Schaefer 10167893d1d5SAdam Litke spin_lock(&hugetlb_lock); 1017d1c3fb1fSNishanth Aravamudan if (page) { 10180edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&page->lru); 1019bf50bab2SNaoya Horiguchi r_nid = page_to_nid(page); 1020d1c3fb1fSNishanth Aravamudan set_compound_page_dtor(page, free_huge_page); 10219dd540e2SAneesh Kumar K.V set_hugetlb_cgroup(page, NULL); 1022d1c3fb1fSNishanth Aravamudan /* 1023d1c3fb1fSNishanth Aravamudan * We incremented the global counters already 1024d1c3fb1fSNishanth Aravamudan */ 1025bf50bab2SNaoya Horiguchi h->nr_huge_pages_node[r_nid]++; 1026bf50bab2SNaoya Horiguchi h->surplus_huge_pages_node[r_nid]++; 10273b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC); 1028d1c3fb1fSNishanth Aravamudan } else { 1029a5516438SAndi Kleen h->nr_huge_pages--; 1030a5516438SAndi Kleen h->surplus_huge_pages--; 10313b116300SAdam Litke __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 10327893d1d5SAdam Litke } 1033d1c3fb1fSNishanth Aravamudan spin_unlock(&hugetlb_lock); 10347893d1d5SAdam Litke 10357893d1d5SAdam Litke return page; 10367893d1d5SAdam Litke } 10377893d1d5SAdam Litke 1038e4e574b7SAdam Litke /* 1039bf50bab2SNaoya Horiguchi * This allocation function is useful in the context where vma is irrelevant. 1040bf50bab2SNaoya Horiguchi * E.g. soft-offlining uses this function because it only cares physical 1041bf50bab2SNaoya Horiguchi * address of error page. 1042bf50bab2SNaoya Horiguchi */ 1043bf50bab2SNaoya Horiguchi struct page *alloc_huge_page_node(struct hstate *h, int nid) 1044bf50bab2SNaoya Horiguchi { 10454ef91848SJoonsoo Kim struct page *page = NULL; 1046bf50bab2SNaoya Horiguchi 1047bf50bab2SNaoya Horiguchi spin_lock(&hugetlb_lock); 10484ef91848SJoonsoo Kim if (h->free_huge_pages - h->resv_huge_pages > 0) 1049bf50bab2SNaoya Horiguchi page = dequeue_huge_page_node(h, nid); 1050bf50bab2SNaoya Horiguchi spin_unlock(&hugetlb_lock); 1051bf50bab2SNaoya Horiguchi 105294ae8ba7SAneesh Kumar K.V if (!page) 1053bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, nid); 1054bf50bab2SNaoya Horiguchi 1055bf50bab2SNaoya Horiguchi return page; 1056bf50bab2SNaoya Horiguchi } 1057bf50bab2SNaoya Horiguchi 1058bf50bab2SNaoya Horiguchi /* 105925985edcSLucas De Marchi * Increase the hugetlb pool such that it can accommodate a reservation 1060e4e574b7SAdam Litke * of size 'delta'. 1061e4e574b7SAdam Litke */ 1062a5516438SAndi Kleen static int gather_surplus_pages(struct hstate *h, int delta) 1063e4e574b7SAdam Litke { 1064e4e574b7SAdam Litke struct list_head surplus_list; 1065e4e574b7SAdam Litke struct page *page, *tmp; 1066e4e574b7SAdam Litke int ret, i; 1067e4e574b7SAdam Litke int needed, allocated; 106828073b02SHillf Danton bool alloc_ok = true; 1069e4e574b7SAdam Litke 1070a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 1071ac09b3a1SAdam Litke if (needed <= 0) { 1072a5516438SAndi Kleen h->resv_huge_pages += delta; 1073e4e574b7SAdam Litke return 0; 1074ac09b3a1SAdam Litke } 1075e4e574b7SAdam Litke 1076e4e574b7SAdam Litke allocated = 0; 1077e4e574b7SAdam Litke INIT_LIST_HEAD(&surplus_list); 1078e4e574b7SAdam Litke 1079e4e574b7SAdam Litke ret = -ENOMEM; 1080e4e574b7SAdam Litke retry: 1081e4e574b7SAdam Litke spin_unlock(&hugetlb_lock); 1082e4e574b7SAdam Litke for (i = 0; i < needed; i++) { 1083bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 108428073b02SHillf Danton if (!page) { 108528073b02SHillf Danton alloc_ok = false; 108628073b02SHillf Danton break; 108728073b02SHillf Danton } 1088e4e574b7SAdam Litke list_add(&page->lru, &surplus_list); 1089e4e574b7SAdam Litke } 109028073b02SHillf Danton allocated += i; 1091e4e574b7SAdam Litke 1092e4e574b7SAdam Litke /* 1093e4e574b7SAdam Litke * After retaking hugetlb_lock, we need to recalculate 'needed' 1094e4e574b7SAdam Litke * because either resv_huge_pages or free_huge_pages may have changed. 1095e4e574b7SAdam Litke */ 1096e4e574b7SAdam Litke spin_lock(&hugetlb_lock); 1097a5516438SAndi Kleen needed = (h->resv_huge_pages + delta) - 1098a5516438SAndi Kleen (h->free_huge_pages + allocated); 109928073b02SHillf Danton if (needed > 0) { 110028073b02SHillf Danton if (alloc_ok) 1101e4e574b7SAdam Litke goto retry; 110228073b02SHillf Danton /* 110328073b02SHillf Danton * We were not able to allocate enough pages to 110428073b02SHillf Danton * satisfy the entire reservation so we free what 110528073b02SHillf Danton * we've allocated so far. 110628073b02SHillf Danton */ 110728073b02SHillf Danton goto free; 110828073b02SHillf Danton } 1109e4e574b7SAdam Litke /* 1110e4e574b7SAdam Litke * The surplus_list now contains _at_least_ the number of extra pages 111125985edcSLucas De Marchi * needed to accommodate the reservation. Add the appropriate number 1112e4e574b7SAdam Litke * of pages to the hugetlb pool and free the extras back to the buddy 1113ac09b3a1SAdam Litke * allocator. Commit the entire reservation here to prevent another 1114ac09b3a1SAdam Litke * process from stealing the pages as they are added to the pool but 1115ac09b3a1SAdam Litke * before they are reserved. 1116e4e574b7SAdam Litke */ 1117e4e574b7SAdam Litke needed += allocated; 1118a5516438SAndi Kleen h->resv_huge_pages += delta; 1119e4e574b7SAdam Litke ret = 0; 1120a9869b83SNaoya Horiguchi 112119fc3f0aSAdam Litke /* Free the needed pages to the hugetlb pool */ 112219fc3f0aSAdam Litke list_for_each_entry_safe(page, tmp, &surplus_list, lru) { 112319fc3f0aSAdam Litke if ((--needed) < 0) 112419fc3f0aSAdam Litke break; 1125a9869b83SNaoya Horiguchi /* 1126a9869b83SNaoya Horiguchi * This page is now managed by the hugetlb allocator and has 1127a9869b83SNaoya Horiguchi * no users -- drop the buddy allocator's reference. 1128a9869b83SNaoya Horiguchi */ 1129a9869b83SNaoya Horiguchi put_page_testzero(page); 1130309381feSSasha Levin VM_BUG_ON_PAGE(page_count(page), page); 1131a5516438SAndi Kleen enqueue_huge_page(h, page); 113219fc3f0aSAdam Litke } 113328073b02SHillf Danton free: 1134b0365c8dSHillf Danton spin_unlock(&hugetlb_lock); 113519fc3f0aSAdam Litke 113619fc3f0aSAdam Litke /* Free unnecessary surplus pages to the buddy allocator */ 1137c0d934baSJoonsoo Kim list_for_each_entry_safe(page, tmp, &surplus_list, lru) 1138a9869b83SNaoya Horiguchi put_page(page); 113919fc3f0aSAdam Litke spin_lock(&hugetlb_lock); 1140e4e574b7SAdam Litke 1141e4e574b7SAdam Litke return ret; 1142e4e574b7SAdam Litke } 1143e4e574b7SAdam Litke 1144e4e574b7SAdam Litke /* 1145e4e574b7SAdam Litke * When releasing a hugetlb pool reservation, any surplus pages that were 1146e4e574b7SAdam Litke * allocated to satisfy the reservation must be explicitly freed if they were 1147e4e574b7SAdam Litke * never used. 1148685f3457SLee Schermerhorn * Called with hugetlb_lock held. 1149e4e574b7SAdam Litke */ 1150a5516438SAndi Kleen static void return_unused_surplus_pages(struct hstate *h, 1151a5516438SAndi Kleen unsigned long unused_resv_pages) 1152e4e574b7SAdam Litke { 1153e4e574b7SAdam Litke unsigned long nr_pages; 1154e4e574b7SAdam Litke 1155ac09b3a1SAdam Litke /* Uncommit the reservation */ 1156a5516438SAndi Kleen h->resv_huge_pages -= unused_resv_pages; 1157ac09b3a1SAdam Litke 1158aa888a74SAndi Kleen /* Cannot return gigantic pages currently */ 1159aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1160aa888a74SAndi Kleen return; 1161aa888a74SAndi Kleen 1162a5516438SAndi Kleen nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 1163e4e574b7SAdam Litke 1164685f3457SLee Schermerhorn /* 1165685f3457SLee Schermerhorn * We want to release as many surplus pages as possible, spread 11669b5e5d0fSLee Schermerhorn * evenly across all nodes with memory. Iterate across these nodes 11679b5e5d0fSLee Schermerhorn * until we can no longer free unreserved surplus pages. This occurs 11689b5e5d0fSLee Schermerhorn * when the nodes with surplus pages have no free pages. 11699b5e5d0fSLee Schermerhorn * free_pool_huge_page() will balance the the freed pages across the 11709b5e5d0fSLee Schermerhorn * on-line nodes with memory and will handle the hstate accounting. 1171685f3457SLee Schermerhorn */ 1172685f3457SLee Schermerhorn while (nr_pages--) { 11738cebfcd0SLai Jiangshan if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) 1174685f3457SLee Schermerhorn break; 11757848a4bfSMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 1176e4e574b7SAdam Litke } 1177e4e574b7SAdam Litke } 1178e4e574b7SAdam Litke 1179c37f9fb1SAndy Whitcroft /* 1180c37f9fb1SAndy Whitcroft * Determine if the huge page at addr within the vma has an associated 1181c37f9fb1SAndy Whitcroft * reservation. Where it does not we will need to logically increase 118290481622SDavid Gibson * reservation and actually increase subpool usage before an allocation 118390481622SDavid Gibson * can occur. Where any new reservation would be required the 118490481622SDavid Gibson * reservation change is prepared, but not committed. Once the page 118590481622SDavid Gibson * has been allocated from the subpool and instantiated the change should 118690481622SDavid Gibson * be committed via vma_commit_reservation. No action is required on 118790481622SDavid Gibson * failure. 1188c37f9fb1SAndy Whitcroft */ 1189e2f17d94SRoel Kluin static long vma_needs_reservation(struct hstate *h, 1190a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1191c37f9fb1SAndy Whitcroft { 11924e35f483SJoonsoo Kim struct resv_map *resv; 11934e35f483SJoonsoo Kim pgoff_t idx; 11944e35f483SJoonsoo Kim long chg; 1195c37f9fb1SAndy Whitcroft 11964e35f483SJoonsoo Kim resv = vma_resv_map(vma); 11974e35f483SJoonsoo Kim if (!resv) 1198c37f9fb1SAndy Whitcroft return 1; 1199c37f9fb1SAndy Whitcroft 12004e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 12014e35f483SJoonsoo Kim chg = region_chg(resv, idx, idx + 1); 120284afd99bSAndy Whitcroft 12034e35f483SJoonsoo Kim if (vma->vm_flags & VM_MAYSHARE) 12044e35f483SJoonsoo Kim return chg; 12054e35f483SJoonsoo Kim else 12064e35f483SJoonsoo Kim return chg < 0 ? chg : 0; 120784afd99bSAndy Whitcroft } 1208a5516438SAndi Kleen static void vma_commit_reservation(struct hstate *h, 1209a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long addr) 1210c37f9fb1SAndy Whitcroft { 12114e35f483SJoonsoo Kim struct resv_map *resv; 12124e35f483SJoonsoo Kim pgoff_t idx; 1213c37f9fb1SAndy Whitcroft 12144e35f483SJoonsoo Kim resv = vma_resv_map(vma); 12154e35f483SJoonsoo Kim if (!resv) 12164e35f483SJoonsoo Kim return; 12179119a41eSJoonsoo Kim 12184e35f483SJoonsoo Kim idx = vma_hugecache_offset(h, vma, addr); 12191406ec9bSJoonsoo Kim region_add(resv, idx, idx + 1); 1220c37f9fb1SAndy Whitcroft } 1221c37f9fb1SAndy Whitcroft 1222348ea204SAdam Litke static struct page *alloc_huge_page(struct vm_area_struct *vma, 122304f2cbe3SMel Gorman unsigned long addr, int avoid_reserve) 1224348ea204SAdam Litke { 122590481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 1226a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 1227348ea204SAdam Litke struct page *page; 1228e2f17d94SRoel Kluin long chg; 12296d76dcf4SAneesh Kumar K.V int ret, idx; 12306d76dcf4SAneesh Kumar K.V struct hugetlb_cgroup *h_cg; 12312fc39cecSAdam Litke 12326d76dcf4SAneesh Kumar K.V idx = hstate_index(h); 1233a1e78772SMel Gorman /* 123490481622SDavid Gibson * Processes that did not create the mapping will have no 123590481622SDavid Gibson * reserves and will not have accounted against subpool 123690481622SDavid Gibson * limit. Check that the subpool limit can be made before 123790481622SDavid Gibson * satisfying the allocation MAP_NORESERVE mappings may also 123890481622SDavid Gibson * need pages and subpool limit allocated allocated if no reserve 123990481622SDavid Gibson * mapping overlaps. 1240a1e78772SMel Gorman */ 1241a5516438SAndi Kleen chg = vma_needs_reservation(h, vma, addr); 1242c37f9fb1SAndy Whitcroft if (chg < 0) 124376dcee75SAneesh Kumar K.V return ERR_PTR(-ENOMEM); 12448bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 12458bb3f12eSJoonsoo Kim if (hugepage_subpool_get_pages(spool, 1)) 124676dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 124790d8b7e6SAdam Litke 12486d76dcf4SAneesh Kumar K.V ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 12496d76dcf4SAneesh Kumar K.V if (ret) { 12508bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 12518bb3f12eSJoonsoo Kim hugepage_subpool_put_pages(spool, 1); 12526d76dcf4SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 12536d76dcf4SAneesh Kumar K.V } 1254a1e78772SMel Gorman spin_lock(&hugetlb_lock); 1255af0ed73eSJoonsoo Kim page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); 125681a6fcaeSJoonsoo Kim if (!page) { 125794ae8ba7SAneesh Kumar K.V spin_unlock(&hugetlb_lock); 1258bf50bab2SNaoya Horiguchi page = alloc_buddy_huge_page(h, NUMA_NO_NODE); 1259a1e78772SMel Gorman if (!page) { 12606d76dcf4SAneesh Kumar K.V hugetlb_cgroup_uncharge_cgroup(idx, 12616d76dcf4SAneesh Kumar K.V pages_per_huge_page(h), 12626d76dcf4SAneesh Kumar K.V h_cg); 12638bb3f12eSJoonsoo Kim if (chg || avoid_reserve) 12648bb3f12eSJoonsoo Kim hugepage_subpool_put_pages(spool, 1); 126576dcee75SAneesh Kumar K.V return ERR_PTR(-ENOSPC); 1266a1e78772SMel Gorman } 126779dbb236SAneesh Kumar K.V spin_lock(&hugetlb_lock); 126879dbb236SAneesh Kumar K.V list_move(&page->lru, &h->hugepage_activelist); 126981a6fcaeSJoonsoo Kim /* Fall through */ 1270a1e78772SMel Gorman } 127181a6fcaeSJoonsoo Kim hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); 127281a6fcaeSJoonsoo Kim spin_unlock(&hugetlb_lock); 1273a1e78772SMel Gorman 127490481622SDavid Gibson set_page_private(page, (unsigned long)spool); 1275a1e78772SMel Gorman 1276a5516438SAndi Kleen vma_commit_reservation(h, vma, addr); 12777893d1d5SAdam Litke return page; 1278b45b5bd6SDavid Gibson } 1279b45b5bd6SDavid Gibson 128074060e4dSNaoya Horiguchi /* 128174060e4dSNaoya Horiguchi * alloc_huge_page()'s wrapper which simply returns the page if allocation 128274060e4dSNaoya Horiguchi * succeeds, otherwise NULL. This function is called from new_vma_page(), 128374060e4dSNaoya Horiguchi * where no ERR_VALUE is expected to be returned. 128474060e4dSNaoya Horiguchi */ 128574060e4dSNaoya Horiguchi struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, 128674060e4dSNaoya Horiguchi unsigned long addr, int avoid_reserve) 128774060e4dSNaoya Horiguchi { 128874060e4dSNaoya Horiguchi struct page *page = alloc_huge_page(vma, addr, avoid_reserve); 128974060e4dSNaoya Horiguchi if (IS_ERR(page)) 129074060e4dSNaoya Horiguchi page = NULL; 129174060e4dSNaoya Horiguchi return page; 129274060e4dSNaoya Horiguchi } 129374060e4dSNaoya Horiguchi 129491f47662SCyrill Gorcunov int __weak alloc_bootmem_huge_page(struct hstate *h) 1295aa888a74SAndi Kleen { 1296aa888a74SAndi Kleen struct huge_bootmem_page *m; 1297b2261026SJoonsoo Kim int nr_nodes, node; 1298aa888a74SAndi Kleen 1299b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { 1300aa888a74SAndi Kleen void *addr; 1301aa888a74SAndi Kleen 13028b89a116SGrygorii Strashko addr = memblock_virt_alloc_try_nid_nopanic( 13038b89a116SGrygorii Strashko huge_page_size(h), huge_page_size(h), 13048b89a116SGrygorii Strashko 0, BOOTMEM_ALLOC_ACCESSIBLE, node); 1305aa888a74SAndi Kleen if (addr) { 1306aa888a74SAndi Kleen /* 1307aa888a74SAndi Kleen * Use the beginning of the huge page to store the 1308aa888a74SAndi Kleen * huge_bootmem_page struct (until gather_bootmem 1309aa888a74SAndi Kleen * puts them into the mem_map). 1310aa888a74SAndi Kleen */ 1311aa888a74SAndi Kleen m = addr; 1312aa888a74SAndi Kleen goto found; 1313aa888a74SAndi Kleen } 1314aa888a74SAndi Kleen } 1315aa888a74SAndi Kleen return 0; 1316aa888a74SAndi Kleen 1317aa888a74SAndi Kleen found: 1318aa888a74SAndi Kleen BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); 1319aa888a74SAndi Kleen /* Put them into a private list first because mem_map is not up yet */ 1320aa888a74SAndi Kleen list_add(&m->list, &huge_boot_pages); 1321aa888a74SAndi Kleen m->hstate = h; 1322aa888a74SAndi Kleen return 1; 1323aa888a74SAndi Kleen } 1324aa888a74SAndi Kleen 1325f412c97aSDavid Rientjes static void __init prep_compound_huge_page(struct page *page, int order) 132618229df5SAndy Whitcroft { 132718229df5SAndy Whitcroft if (unlikely(order > (MAX_ORDER - 1))) 132818229df5SAndy Whitcroft prep_compound_gigantic_page(page, order); 132918229df5SAndy Whitcroft else 133018229df5SAndy Whitcroft prep_compound_page(page, order); 133118229df5SAndy Whitcroft } 133218229df5SAndy Whitcroft 1333aa888a74SAndi Kleen /* Put bootmem huge pages into the standard lists after mem_map is up */ 1334aa888a74SAndi Kleen static void __init gather_bootmem_prealloc(void) 1335aa888a74SAndi Kleen { 1336aa888a74SAndi Kleen struct huge_bootmem_page *m; 1337aa888a74SAndi Kleen 1338aa888a74SAndi Kleen list_for_each_entry(m, &huge_boot_pages, list) { 1339aa888a74SAndi Kleen struct hstate *h = m->hstate; 1340ee8f248dSBecky Bruce struct page *page; 1341ee8f248dSBecky Bruce 1342ee8f248dSBecky Bruce #ifdef CONFIG_HIGHMEM 1343ee8f248dSBecky Bruce page = pfn_to_page(m->phys >> PAGE_SHIFT); 13448b89a116SGrygorii Strashko memblock_free_late(__pa(m), 1345ee8f248dSBecky Bruce sizeof(struct huge_bootmem_page)); 1346ee8f248dSBecky Bruce #else 1347ee8f248dSBecky Bruce page = virt_to_page(m); 1348ee8f248dSBecky Bruce #endif 1349aa888a74SAndi Kleen WARN_ON(page_count(page) != 1); 135018229df5SAndy Whitcroft prep_compound_huge_page(page, h->order); 1351ef5a22beSAndrea Arcangeli WARN_ON(PageReserved(page)); 1352aa888a74SAndi Kleen prep_new_huge_page(h, page, page_to_nid(page)); 1353b0320c7bSRafael Aquini /* 1354b0320c7bSRafael Aquini * If we had gigantic hugepages allocated at boot time, we need 1355b0320c7bSRafael Aquini * to restore the 'stolen' pages to totalram_pages in order to 1356b0320c7bSRafael Aquini * fix confusing memory reports from free(1) and another 1357b0320c7bSRafael Aquini * side-effects, like CommitLimit going negative. 1358b0320c7bSRafael Aquini */ 1359b0320c7bSRafael Aquini if (h->order > (MAX_ORDER - 1)) 13603dcc0571SJiang Liu adjust_managed_page_count(page, 1 << h->order); 1361aa888a74SAndi Kleen } 1362aa888a74SAndi Kleen } 1363aa888a74SAndi Kleen 13648faa8b07SAndi Kleen static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 13651da177e4SLinus Torvalds { 13661da177e4SLinus Torvalds unsigned long i; 13671da177e4SLinus Torvalds 1368e5ff2159SAndi Kleen for (i = 0; i < h->max_huge_pages; ++i) { 1369aa888a74SAndi Kleen if (h->order >= MAX_ORDER) { 1370aa888a74SAndi Kleen if (!alloc_bootmem_huge_page(h)) 1371aa888a74SAndi Kleen break; 13729b5e5d0fSLee Schermerhorn } else if (!alloc_fresh_huge_page(h, 13738cebfcd0SLai Jiangshan &node_states[N_MEMORY])) 13741da177e4SLinus Torvalds break; 13751da177e4SLinus Torvalds } 13768faa8b07SAndi Kleen h->max_huge_pages = i; 1377e5ff2159SAndi Kleen } 1378e5ff2159SAndi Kleen 1379e5ff2159SAndi Kleen static void __init hugetlb_init_hstates(void) 1380e5ff2159SAndi Kleen { 1381e5ff2159SAndi Kleen struct hstate *h; 1382e5ff2159SAndi Kleen 1383e5ff2159SAndi Kleen for_each_hstate(h) { 13848faa8b07SAndi Kleen /* oversize hugepages were init'ed in early boot */ 13858faa8b07SAndi Kleen if (h->order < MAX_ORDER) 13868faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(h); 1387e5ff2159SAndi Kleen } 1388e5ff2159SAndi Kleen } 1389e5ff2159SAndi Kleen 13904abd32dbSAndi Kleen static char * __init memfmt(char *buf, unsigned long n) 13914abd32dbSAndi Kleen { 13924abd32dbSAndi Kleen if (n >= (1UL << 30)) 13934abd32dbSAndi Kleen sprintf(buf, "%lu GB", n >> 30); 13944abd32dbSAndi Kleen else if (n >= (1UL << 20)) 13954abd32dbSAndi Kleen sprintf(buf, "%lu MB", n >> 20); 13964abd32dbSAndi Kleen else 13974abd32dbSAndi Kleen sprintf(buf, "%lu KB", n >> 10); 13984abd32dbSAndi Kleen return buf; 13994abd32dbSAndi Kleen } 14004abd32dbSAndi Kleen 1401e5ff2159SAndi Kleen static void __init report_hugepages(void) 1402e5ff2159SAndi Kleen { 1403e5ff2159SAndi Kleen struct hstate *h; 1404e5ff2159SAndi Kleen 1405e5ff2159SAndi Kleen for_each_hstate(h) { 14064abd32dbSAndi Kleen char buf[32]; 1407ffb22af5SAndrew Morton pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", 14084abd32dbSAndi Kleen memfmt(buf, huge_page_size(h)), 14094abd32dbSAndi Kleen h->free_huge_pages); 1410e5ff2159SAndi Kleen } 1411e5ff2159SAndi Kleen } 1412e5ff2159SAndi Kleen 14131da177e4SLinus Torvalds #ifdef CONFIG_HIGHMEM 14146ae11b27SLee Schermerhorn static void try_to_free_low(struct hstate *h, unsigned long count, 14156ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 14161da177e4SLinus Torvalds { 14174415cc8dSChristoph Lameter int i; 14184415cc8dSChristoph Lameter 1419aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1420aa888a74SAndi Kleen return; 1421aa888a74SAndi Kleen 14226ae11b27SLee Schermerhorn for_each_node_mask(i, *nodes_allowed) { 14231da177e4SLinus Torvalds struct page *page, *next; 1424a5516438SAndi Kleen struct list_head *freel = &h->hugepage_freelists[i]; 1425a5516438SAndi Kleen list_for_each_entry_safe(page, next, freel, lru) { 1426a5516438SAndi Kleen if (count >= h->nr_huge_pages) 14276b0c880dSAdam Litke return; 14281da177e4SLinus Torvalds if (PageHighMem(page)) 14291da177e4SLinus Torvalds continue; 14301da177e4SLinus Torvalds list_del(&page->lru); 1431e5ff2159SAndi Kleen update_and_free_page(h, page); 1432a5516438SAndi Kleen h->free_huge_pages--; 1433a5516438SAndi Kleen h->free_huge_pages_node[page_to_nid(page)]--; 14341da177e4SLinus Torvalds } 14351da177e4SLinus Torvalds } 14361da177e4SLinus Torvalds } 14371da177e4SLinus Torvalds #else 14386ae11b27SLee Schermerhorn static inline void try_to_free_low(struct hstate *h, unsigned long count, 14396ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 14401da177e4SLinus Torvalds { 14411da177e4SLinus Torvalds } 14421da177e4SLinus Torvalds #endif 14431da177e4SLinus Torvalds 144420a0307cSWu Fengguang /* 144520a0307cSWu Fengguang * Increment or decrement surplus_huge_pages. Keep node-specific counters 144620a0307cSWu Fengguang * balanced by operating on them in a round-robin fashion. 144720a0307cSWu Fengguang * Returns 1 if an adjustment was made. 144820a0307cSWu Fengguang */ 14496ae11b27SLee Schermerhorn static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 14506ae11b27SLee Schermerhorn int delta) 145120a0307cSWu Fengguang { 1452b2261026SJoonsoo Kim int nr_nodes, node; 145320a0307cSWu Fengguang 145420a0307cSWu Fengguang VM_BUG_ON(delta != -1 && delta != 1); 145520a0307cSWu Fengguang 1456e8c5c824SLee Schermerhorn if (delta < 0) { 1457b2261026SJoonsoo Kim for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { 1458b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node]) 1459b2261026SJoonsoo Kim goto found; 1460b2261026SJoonsoo Kim } 1461b2261026SJoonsoo Kim } else { 1462b2261026SJoonsoo Kim for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 1463b2261026SJoonsoo Kim if (h->surplus_huge_pages_node[node] < 1464b2261026SJoonsoo Kim h->nr_huge_pages_node[node]) 1465b2261026SJoonsoo Kim goto found; 1466e8c5c824SLee Schermerhorn } 14679a76db09SLee Schermerhorn } 1468b2261026SJoonsoo Kim return 0; 146920a0307cSWu Fengguang 1470b2261026SJoonsoo Kim found: 147120a0307cSWu Fengguang h->surplus_huge_pages += delta; 1472b2261026SJoonsoo Kim h->surplus_huge_pages_node[node] += delta; 1473b2261026SJoonsoo Kim return 1; 147420a0307cSWu Fengguang } 147520a0307cSWu Fengguang 1476a5516438SAndi Kleen #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 14776ae11b27SLee Schermerhorn static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, 14786ae11b27SLee Schermerhorn nodemask_t *nodes_allowed) 14791da177e4SLinus Torvalds { 14807893d1d5SAdam Litke unsigned long min_count, ret; 14811da177e4SLinus Torvalds 1482aa888a74SAndi Kleen if (h->order >= MAX_ORDER) 1483aa888a74SAndi Kleen return h->max_huge_pages; 1484aa888a74SAndi Kleen 14857893d1d5SAdam Litke /* 14867893d1d5SAdam Litke * Increase the pool size 14877893d1d5SAdam Litke * First take pages out of surplus state. Then make up the 14887893d1d5SAdam Litke * remaining difference by allocating fresh huge pages. 1489d1c3fb1fSNishanth Aravamudan * 1490d1c3fb1fSNishanth Aravamudan * We might race with alloc_buddy_huge_page() here and be unable 1491d1c3fb1fSNishanth Aravamudan * to convert a surplus huge page to a normal huge page. That is 1492d1c3fb1fSNishanth Aravamudan * not critical, though, it just means the overall size of the 1493d1c3fb1fSNishanth Aravamudan * pool might be one hugepage larger than it needs to be, but 1494d1c3fb1fSNishanth Aravamudan * within all the constraints specified by the sysctls. 14957893d1d5SAdam Litke */ 14961da177e4SLinus Torvalds spin_lock(&hugetlb_lock); 1497a5516438SAndi Kleen while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 14986ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, -1)) 14997893d1d5SAdam Litke break; 15007893d1d5SAdam Litke } 15017893d1d5SAdam Litke 1502a5516438SAndi Kleen while (count > persistent_huge_pages(h)) { 15037893d1d5SAdam Litke /* 15047893d1d5SAdam Litke * If this allocation races such that we no longer need the 15057893d1d5SAdam Litke * page, free_huge_page will handle it by freeing the page 15067893d1d5SAdam Litke * and reducing the surplus. 15077893d1d5SAdam Litke */ 15087893d1d5SAdam Litke spin_unlock(&hugetlb_lock); 15096ae11b27SLee Schermerhorn ret = alloc_fresh_huge_page(h, nodes_allowed); 15107893d1d5SAdam Litke spin_lock(&hugetlb_lock); 15117893d1d5SAdam Litke if (!ret) 15127893d1d5SAdam Litke goto out; 15137893d1d5SAdam Litke 1514536240f2SMel Gorman /* Bail for signals. Probably ctrl-c from user */ 1515536240f2SMel Gorman if (signal_pending(current)) 1516536240f2SMel Gorman goto out; 15177893d1d5SAdam Litke } 15187893d1d5SAdam Litke 15197893d1d5SAdam Litke /* 15207893d1d5SAdam Litke * Decrease the pool size 15217893d1d5SAdam Litke * First return free pages to the buddy allocator (being careful 15227893d1d5SAdam Litke * to keep enough around to satisfy reservations). Then place 15237893d1d5SAdam Litke * pages into surplus state as needed so the pool will shrink 15247893d1d5SAdam Litke * to the desired size as pages become free. 1525d1c3fb1fSNishanth Aravamudan * 1526d1c3fb1fSNishanth Aravamudan * By placing pages into the surplus state independent of the 1527d1c3fb1fSNishanth Aravamudan * overcommit value, we are allowing the surplus pool size to 1528d1c3fb1fSNishanth Aravamudan * exceed overcommit. There are few sane options here. Since 1529d1c3fb1fSNishanth Aravamudan * alloc_buddy_huge_page() is checking the global counter, 1530d1c3fb1fSNishanth Aravamudan * though, we'll note that we're not allowed to exceed surplus 1531d1c3fb1fSNishanth Aravamudan * and won't grow the pool anywhere else. Not until one of the 1532d1c3fb1fSNishanth Aravamudan * sysctls are changed, or the surplus pages go out of use. 15337893d1d5SAdam Litke */ 1534a5516438SAndi Kleen min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 15356b0c880dSAdam Litke min_count = max(count, min_count); 15366ae11b27SLee Schermerhorn try_to_free_low(h, min_count, nodes_allowed); 1537a5516438SAndi Kleen while (min_count < persistent_huge_pages(h)) { 15386ae11b27SLee Schermerhorn if (!free_pool_huge_page(h, nodes_allowed, 0)) 15391da177e4SLinus Torvalds break; 154055f67141SMizuma, Masayoshi cond_resched_lock(&hugetlb_lock); 15411da177e4SLinus Torvalds } 1542a5516438SAndi Kleen while (count < persistent_huge_pages(h)) { 15436ae11b27SLee Schermerhorn if (!adjust_pool_surplus(h, nodes_allowed, 1)) 15447893d1d5SAdam Litke break; 15457893d1d5SAdam Litke } 15467893d1d5SAdam Litke out: 1547a5516438SAndi Kleen ret = persistent_huge_pages(h); 15481da177e4SLinus Torvalds spin_unlock(&hugetlb_lock); 15497893d1d5SAdam Litke return ret; 15501da177e4SLinus Torvalds } 15511da177e4SLinus Torvalds 1552a3437870SNishanth Aravamudan #define HSTATE_ATTR_RO(_name) \ 1553a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 1554a3437870SNishanth Aravamudan 1555a3437870SNishanth Aravamudan #define HSTATE_ATTR(_name) \ 1556a3437870SNishanth Aravamudan static struct kobj_attribute _name##_attr = \ 1557a3437870SNishanth Aravamudan __ATTR(_name, 0644, _name##_show, _name##_store) 1558a3437870SNishanth Aravamudan 1559a3437870SNishanth Aravamudan static struct kobject *hugepages_kobj; 1560a3437870SNishanth Aravamudan static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 1561a3437870SNishanth Aravamudan 15629a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 15639a305230SLee Schermerhorn 15649a305230SLee Schermerhorn static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 1565a3437870SNishanth Aravamudan { 1566a3437870SNishanth Aravamudan int i; 15679a305230SLee Schermerhorn 1568a3437870SNishanth Aravamudan for (i = 0; i < HUGE_MAX_HSTATE; i++) 15699a305230SLee Schermerhorn if (hstate_kobjs[i] == kobj) { 15709a305230SLee Schermerhorn if (nidp) 15719a305230SLee Schermerhorn *nidp = NUMA_NO_NODE; 1572a3437870SNishanth Aravamudan return &hstates[i]; 15739a305230SLee Schermerhorn } 15749a305230SLee Schermerhorn 15759a305230SLee Schermerhorn return kobj_to_node_hstate(kobj, nidp); 1576a3437870SNishanth Aravamudan } 1577a3437870SNishanth Aravamudan 157806808b08SLee Schermerhorn static ssize_t nr_hugepages_show_common(struct kobject *kobj, 1579a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1580a3437870SNishanth Aravamudan { 15819a305230SLee Schermerhorn struct hstate *h; 15829a305230SLee Schermerhorn unsigned long nr_huge_pages; 15839a305230SLee Schermerhorn int nid; 15849a305230SLee Schermerhorn 15859a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 15869a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 15879a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages; 15889a305230SLee Schermerhorn else 15899a305230SLee Schermerhorn nr_huge_pages = h->nr_huge_pages_node[nid]; 15909a305230SLee Schermerhorn 15919a305230SLee Schermerhorn return sprintf(buf, "%lu\n", nr_huge_pages); 1592a3437870SNishanth Aravamudan } 1593adbe8726SEric B Munson 159406808b08SLee Schermerhorn static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 159506808b08SLee Schermerhorn struct kobject *kobj, struct kobj_attribute *attr, 159606808b08SLee Schermerhorn const char *buf, size_t len) 1597a3437870SNishanth Aravamudan { 1598a3437870SNishanth Aravamudan int err; 15999a305230SLee Schermerhorn int nid; 160006808b08SLee Schermerhorn unsigned long count; 16019a305230SLee Schermerhorn struct hstate *h; 1602bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); 1603a3437870SNishanth Aravamudan 16043dbb95f7SJingoo Han err = kstrtoul(buf, 10, &count); 160573ae31e5SEric B Munson if (err) 1606adbe8726SEric B Munson goto out; 1607a3437870SNishanth Aravamudan 16089a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 1609adbe8726SEric B Munson if (h->order >= MAX_ORDER) { 1610adbe8726SEric B Munson err = -EINVAL; 1611adbe8726SEric B Munson goto out; 1612adbe8726SEric B Munson } 1613adbe8726SEric B Munson 16149a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) { 16159a305230SLee Schermerhorn /* 16169a305230SLee Schermerhorn * global hstate attribute 16179a305230SLee Schermerhorn */ 16189a305230SLee Schermerhorn if (!(obey_mempolicy && 16199a305230SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 162006808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 16218cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 162206808b08SLee Schermerhorn } 16239a305230SLee Schermerhorn } else if (nodes_allowed) { 16249a305230SLee Schermerhorn /* 16259a305230SLee Schermerhorn * per node hstate attribute: adjust count to global, 16269a305230SLee Schermerhorn * but restrict alloc/free to the specified node. 16279a305230SLee Schermerhorn */ 16289a305230SLee Schermerhorn count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; 16299a305230SLee Schermerhorn init_nodemask_of_node(nodes_allowed, nid); 16309a305230SLee Schermerhorn } else 16318cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 16329a305230SLee Schermerhorn 163306808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); 1634a3437870SNishanth Aravamudan 16358cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 163606808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 163706808b08SLee Schermerhorn 163806808b08SLee Schermerhorn return len; 1639adbe8726SEric B Munson out: 1640adbe8726SEric B Munson NODEMASK_FREE(nodes_allowed); 1641adbe8726SEric B Munson return err; 164206808b08SLee Schermerhorn } 164306808b08SLee Schermerhorn 164406808b08SLee Schermerhorn static ssize_t nr_hugepages_show(struct kobject *kobj, 164506808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 164606808b08SLee Schermerhorn { 164706808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 164806808b08SLee Schermerhorn } 164906808b08SLee Schermerhorn 165006808b08SLee Schermerhorn static ssize_t nr_hugepages_store(struct kobject *kobj, 165106808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 165206808b08SLee Schermerhorn { 165306808b08SLee Schermerhorn return nr_hugepages_store_common(false, kobj, attr, buf, len); 1654a3437870SNishanth Aravamudan } 1655a3437870SNishanth Aravamudan HSTATE_ATTR(nr_hugepages); 1656a3437870SNishanth Aravamudan 165706808b08SLee Schermerhorn #ifdef CONFIG_NUMA 165806808b08SLee Schermerhorn 165906808b08SLee Schermerhorn /* 166006808b08SLee Schermerhorn * hstate attribute for optionally mempolicy-based constraint on persistent 166106808b08SLee Schermerhorn * huge page alloc/free. 166206808b08SLee Schermerhorn */ 166306808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 166406808b08SLee Schermerhorn struct kobj_attribute *attr, char *buf) 166506808b08SLee Schermerhorn { 166606808b08SLee Schermerhorn return nr_hugepages_show_common(kobj, attr, buf); 166706808b08SLee Schermerhorn } 166806808b08SLee Schermerhorn 166906808b08SLee Schermerhorn static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 167006808b08SLee Schermerhorn struct kobj_attribute *attr, const char *buf, size_t len) 167106808b08SLee Schermerhorn { 167206808b08SLee Schermerhorn return nr_hugepages_store_common(true, kobj, attr, buf, len); 167306808b08SLee Schermerhorn } 167406808b08SLee Schermerhorn HSTATE_ATTR(nr_hugepages_mempolicy); 167506808b08SLee Schermerhorn #endif 167606808b08SLee Schermerhorn 167706808b08SLee Schermerhorn 1678a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 1679a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1680a3437870SNishanth Aravamudan { 16819a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1682a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); 1683a3437870SNishanth Aravamudan } 1684adbe8726SEric B Munson 1685a3437870SNishanth Aravamudan static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 1686a3437870SNishanth Aravamudan struct kobj_attribute *attr, const char *buf, size_t count) 1687a3437870SNishanth Aravamudan { 1688a3437870SNishanth Aravamudan int err; 1689a3437870SNishanth Aravamudan unsigned long input; 16909a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1691a3437870SNishanth Aravamudan 1692adbe8726SEric B Munson if (h->order >= MAX_ORDER) 1693adbe8726SEric B Munson return -EINVAL; 1694adbe8726SEric B Munson 16953dbb95f7SJingoo Han err = kstrtoul(buf, 10, &input); 1696a3437870SNishanth Aravamudan if (err) 169773ae31e5SEric B Munson return err; 1698a3437870SNishanth Aravamudan 1699a3437870SNishanth Aravamudan spin_lock(&hugetlb_lock); 1700a3437870SNishanth Aravamudan h->nr_overcommit_huge_pages = input; 1701a3437870SNishanth Aravamudan spin_unlock(&hugetlb_lock); 1702a3437870SNishanth Aravamudan 1703a3437870SNishanth Aravamudan return count; 1704a3437870SNishanth Aravamudan } 1705a3437870SNishanth Aravamudan HSTATE_ATTR(nr_overcommit_hugepages); 1706a3437870SNishanth Aravamudan 1707a3437870SNishanth Aravamudan static ssize_t free_hugepages_show(struct kobject *kobj, 1708a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1709a3437870SNishanth Aravamudan { 17109a305230SLee Schermerhorn struct hstate *h; 17119a305230SLee Schermerhorn unsigned long free_huge_pages; 17129a305230SLee Schermerhorn int nid; 17139a305230SLee Schermerhorn 17149a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 17159a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 17169a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages; 17179a305230SLee Schermerhorn else 17189a305230SLee Schermerhorn free_huge_pages = h->free_huge_pages_node[nid]; 17199a305230SLee Schermerhorn 17209a305230SLee Schermerhorn return sprintf(buf, "%lu\n", free_huge_pages); 1721a3437870SNishanth Aravamudan } 1722a3437870SNishanth Aravamudan HSTATE_ATTR_RO(free_hugepages); 1723a3437870SNishanth Aravamudan 1724a3437870SNishanth Aravamudan static ssize_t resv_hugepages_show(struct kobject *kobj, 1725a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1726a3437870SNishanth Aravamudan { 17279a305230SLee Schermerhorn struct hstate *h = kobj_to_hstate(kobj, NULL); 1728a3437870SNishanth Aravamudan return sprintf(buf, "%lu\n", h->resv_huge_pages); 1729a3437870SNishanth Aravamudan } 1730a3437870SNishanth Aravamudan HSTATE_ATTR_RO(resv_hugepages); 1731a3437870SNishanth Aravamudan 1732a3437870SNishanth Aravamudan static ssize_t surplus_hugepages_show(struct kobject *kobj, 1733a3437870SNishanth Aravamudan struct kobj_attribute *attr, char *buf) 1734a3437870SNishanth Aravamudan { 17359a305230SLee Schermerhorn struct hstate *h; 17369a305230SLee Schermerhorn unsigned long surplus_huge_pages; 17379a305230SLee Schermerhorn int nid; 17389a305230SLee Schermerhorn 17399a305230SLee Schermerhorn h = kobj_to_hstate(kobj, &nid); 17409a305230SLee Schermerhorn if (nid == NUMA_NO_NODE) 17419a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages; 17429a305230SLee Schermerhorn else 17439a305230SLee Schermerhorn surplus_huge_pages = h->surplus_huge_pages_node[nid]; 17449a305230SLee Schermerhorn 17459a305230SLee Schermerhorn return sprintf(buf, "%lu\n", surplus_huge_pages); 1746a3437870SNishanth Aravamudan } 1747a3437870SNishanth Aravamudan HSTATE_ATTR_RO(surplus_hugepages); 1748a3437870SNishanth Aravamudan 1749a3437870SNishanth Aravamudan static struct attribute *hstate_attrs[] = { 1750a3437870SNishanth Aravamudan &nr_hugepages_attr.attr, 1751a3437870SNishanth Aravamudan &nr_overcommit_hugepages_attr.attr, 1752a3437870SNishanth Aravamudan &free_hugepages_attr.attr, 1753a3437870SNishanth Aravamudan &resv_hugepages_attr.attr, 1754a3437870SNishanth Aravamudan &surplus_hugepages_attr.attr, 175506808b08SLee Schermerhorn #ifdef CONFIG_NUMA 175606808b08SLee Schermerhorn &nr_hugepages_mempolicy_attr.attr, 175706808b08SLee Schermerhorn #endif 1758a3437870SNishanth Aravamudan NULL, 1759a3437870SNishanth Aravamudan }; 1760a3437870SNishanth Aravamudan 1761a3437870SNishanth Aravamudan static struct attribute_group hstate_attr_group = { 1762a3437870SNishanth Aravamudan .attrs = hstate_attrs, 1763a3437870SNishanth Aravamudan }; 1764a3437870SNishanth Aravamudan 1765094e9539SJeff Mahoney static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 17669a305230SLee Schermerhorn struct kobject **hstate_kobjs, 17679a305230SLee Schermerhorn struct attribute_group *hstate_attr_group) 1768a3437870SNishanth Aravamudan { 1769a3437870SNishanth Aravamudan int retval; 1770972dc4deSAneesh Kumar K.V int hi = hstate_index(h); 1771a3437870SNishanth Aravamudan 17729a305230SLee Schermerhorn hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 17739a305230SLee Schermerhorn if (!hstate_kobjs[hi]) 1774a3437870SNishanth Aravamudan return -ENOMEM; 1775a3437870SNishanth Aravamudan 17769a305230SLee Schermerhorn retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 1777a3437870SNishanth Aravamudan if (retval) 17789a305230SLee Schermerhorn kobject_put(hstate_kobjs[hi]); 1779a3437870SNishanth Aravamudan 1780a3437870SNishanth Aravamudan return retval; 1781a3437870SNishanth Aravamudan } 1782a3437870SNishanth Aravamudan 1783a3437870SNishanth Aravamudan static void __init hugetlb_sysfs_init(void) 1784a3437870SNishanth Aravamudan { 1785a3437870SNishanth Aravamudan struct hstate *h; 1786a3437870SNishanth Aravamudan int err; 1787a3437870SNishanth Aravamudan 1788a3437870SNishanth Aravamudan hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 1789a3437870SNishanth Aravamudan if (!hugepages_kobj) 1790a3437870SNishanth Aravamudan return; 1791a3437870SNishanth Aravamudan 1792a3437870SNishanth Aravamudan for_each_hstate(h) { 17939a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 17949a305230SLee Schermerhorn hstate_kobjs, &hstate_attr_group); 1795a3437870SNishanth Aravamudan if (err) 1796ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s", h->name); 1797a3437870SNishanth Aravamudan } 1798a3437870SNishanth Aravamudan } 1799a3437870SNishanth Aravamudan 18009a305230SLee Schermerhorn #ifdef CONFIG_NUMA 18019a305230SLee Schermerhorn 18029a305230SLee Schermerhorn /* 18039a305230SLee Schermerhorn * node_hstate/s - associate per node hstate attributes, via their kobjects, 180410fbcf4cSKay Sievers * with node devices in node_devices[] using a parallel array. The array 180510fbcf4cSKay Sievers * index of a node device or _hstate == node id. 180610fbcf4cSKay Sievers * This is here to avoid any static dependency of the node device driver, in 18079a305230SLee Schermerhorn * the base kernel, on the hugetlb module. 18089a305230SLee Schermerhorn */ 18099a305230SLee Schermerhorn struct node_hstate { 18109a305230SLee Schermerhorn struct kobject *hugepages_kobj; 18119a305230SLee Schermerhorn struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 18129a305230SLee Schermerhorn }; 18139a305230SLee Schermerhorn struct node_hstate node_hstates[MAX_NUMNODES]; 18149a305230SLee Schermerhorn 18159a305230SLee Schermerhorn /* 181610fbcf4cSKay Sievers * A subset of global hstate attributes for node devices 18179a305230SLee Schermerhorn */ 18189a305230SLee Schermerhorn static struct attribute *per_node_hstate_attrs[] = { 18199a305230SLee Schermerhorn &nr_hugepages_attr.attr, 18209a305230SLee Schermerhorn &free_hugepages_attr.attr, 18219a305230SLee Schermerhorn &surplus_hugepages_attr.attr, 18229a305230SLee Schermerhorn NULL, 18239a305230SLee Schermerhorn }; 18249a305230SLee Schermerhorn 18259a305230SLee Schermerhorn static struct attribute_group per_node_hstate_attr_group = { 18269a305230SLee Schermerhorn .attrs = per_node_hstate_attrs, 18279a305230SLee Schermerhorn }; 18289a305230SLee Schermerhorn 18299a305230SLee Schermerhorn /* 183010fbcf4cSKay Sievers * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 18319a305230SLee Schermerhorn * Returns node id via non-NULL nidp. 18329a305230SLee Schermerhorn */ 18339a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 18349a305230SLee Schermerhorn { 18359a305230SLee Schermerhorn int nid; 18369a305230SLee Schermerhorn 18379a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) { 18389a305230SLee Schermerhorn struct node_hstate *nhs = &node_hstates[nid]; 18399a305230SLee Schermerhorn int i; 18409a305230SLee Schermerhorn for (i = 0; i < HUGE_MAX_HSTATE; i++) 18419a305230SLee Schermerhorn if (nhs->hstate_kobjs[i] == kobj) { 18429a305230SLee Schermerhorn if (nidp) 18439a305230SLee Schermerhorn *nidp = nid; 18449a305230SLee Schermerhorn return &hstates[i]; 18459a305230SLee Schermerhorn } 18469a305230SLee Schermerhorn } 18479a305230SLee Schermerhorn 18489a305230SLee Schermerhorn BUG(); 18499a305230SLee Schermerhorn return NULL; 18509a305230SLee Schermerhorn } 18519a305230SLee Schermerhorn 18529a305230SLee Schermerhorn /* 185310fbcf4cSKay Sievers * Unregister hstate attributes from a single node device. 18549a305230SLee Schermerhorn * No-op if no hstate attributes attached. 18559a305230SLee Schermerhorn */ 18563cd8b44fSClaudiu Ghioc static void hugetlb_unregister_node(struct node *node) 18579a305230SLee Schermerhorn { 18589a305230SLee Schermerhorn struct hstate *h; 185910fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 18609a305230SLee Schermerhorn 18619a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 18629b5e5d0fSLee Schermerhorn return; /* no hstate attributes */ 18639a305230SLee Schermerhorn 1864972dc4deSAneesh Kumar K.V for_each_hstate(h) { 1865972dc4deSAneesh Kumar K.V int idx = hstate_index(h); 1866972dc4deSAneesh Kumar K.V if (nhs->hstate_kobjs[idx]) { 1867972dc4deSAneesh Kumar K.V kobject_put(nhs->hstate_kobjs[idx]); 1868972dc4deSAneesh Kumar K.V nhs->hstate_kobjs[idx] = NULL; 1869972dc4deSAneesh Kumar K.V } 18709a305230SLee Schermerhorn } 18719a305230SLee Schermerhorn 18729a305230SLee Schermerhorn kobject_put(nhs->hugepages_kobj); 18739a305230SLee Schermerhorn nhs->hugepages_kobj = NULL; 18749a305230SLee Schermerhorn } 18759a305230SLee Schermerhorn 18769a305230SLee Schermerhorn /* 187710fbcf4cSKay Sievers * hugetlb module exit: unregister hstate attributes from node devices 18789a305230SLee Schermerhorn * that have them. 18799a305230SLee Schermerhorn */ 18809a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) 18819a305230SLee Schermerhorn { 18829a305230SLee Schermerhorn int nid; 18839a305230SLee Schermerhorn 18849a305230SLee Schermerhorn /* 188510fbcf4cSKay Sievers * disable node device registrations. 18869a305230SLee Schermerhorn */ 18879a305230SLee Schermerhorn register_hugetlbfs_with_node(NULL, NULL); 18889a305230SLee Schermerhorn 18899a305230SLee Schermerhorn /* 18909a305230SLee Schermerhorn * remove hstate attributes from any nodes that have them. 18919a305230SLee Schermerhorn */ 18929a305230SLee Schermerhorn for (nid = 0; nid < nr_node_ids; nid++) 18938732794bSWen Congyang hugetlb_unregister_node(node_devices[nid]); 18949a305230SLee Schermerhorn } 18959a305230SLee Schermerhorn 18969a305230SLee Schermerhorn /* 189710fbcf4cSKay Sievers * Register hstate attributes for a single node device. 18989a305230SLee Schermerhorn * No-op if attributes already registered. 18999a305230SLee Schermerhorn */ 19003cd8b44fSClaudiu Ghioc static void hugetlb_register_node(struct node *node) 19019a305230SLee Schermerhorn { 19029a305230SLee Schermerhorn struct hstate *h; 190310fbcf4cSKay Sievers struct node_hstate *nhs = &node_hstates[node->dev.id]; 19049a305230SLee Schermerhorn int err; 19059a305230SLee Schermerhorn 19069a305230SLee Schermerhorn if (nhs->hugepages_kobj) 19079a305230SLee Schermerhorn return; /* already allocated */ 19089a305230SLee Schermerhorn 19099a305230SLee Schermerhorn nhs->hugepages_kobj = kobject_create_and_add("hugepages", 191010fbcf4cSKay Sievers &node->dev.kobj); 19119a305230SLee Schermerhorn if (!nhs->hugepages_kobj) 19129a305230SLee Schermerhorn return; 19139a305230SLee Schermerhorn 19149a305230SLee Schermerhorn for_each_hstate(h) { 19159a305230SLee Schermerhorn err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 19169a305230SLee Schermerhorn nhs->hstate_kobjs, 19179a305230SLee Schermerhorn &per_node_hstate_attr_group); 19189a305230SLee Schermerhorn if (err) { 1919ffb22af5SAndrew Morton pr_err("Hugetlb: Unable to add hstate %s for node %d\n", 192010fbcf4cSKay Sievers h->name, node->dev.id); 19219a305230SLee Schermerhorn hugetlb_unregister_node(node); 19229a305230SLee Schermerhorn break; 19239a305230SLee Schermerhorn } 19249a305230SLee Schermerhorn } 19259a305230SLee Schermerhorn } 19269a305230SLee Schermerhorn 19279a305230SLee Schermerhorn /* 19289b5e5d0fSLee Schermerhorn * hugetlb init time: register hstate attributes for all registered node 192910fbcf4cSKay Sievers * devices of nodes that have memory. All on-line nodes should have 193010fbcf4cSKay Sievers * registered their associated device by this time. 19319a305230SLee Schermerhorn */ 19329a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) 19339a305230SLee Schermerhorn { 19349a305230SLee Schermerhorn int nid; 19359a305230SLee Schermerhorn 19368cebfcd0SLai Jiangshan for_each_node_state(nid, N_MEMORY) { 19378732794bSWen Congyang struct node *node = node_devices[nid]; 193810fbcf4cSKay Sievers if (node->dev.id == nid) 19399a305230SLee Schermerhorn hugetlb_register_node(node); 19409a305230SLee Schermerhorn } 19419a305230SLee Schermerhorn 19429a305230SLee Schermerhorn /* 194310fbcf4cSKay Sievers * Let the node device driver know we're here so it can 19449a305230SLee Schermerhorn * [un]register hstate attributes on node hotplug. 19459a305230SLee Schermerhorn */ 19469a305230SLee Schermerhorn register_hugetlbfs_with_node(hugetlb_register_node, 19479a305230SLee Schermerhorn hugetlb_unregister_node); 19489a305230SLee Schermerhorn } 19499a305230SLee Schermerhorn #else /* !CONFIG_NUMA */ 19509a305230SLee Schermerhorn 19519a305230SLee Schermerhorn static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 19529a305230SLee Schermerhorn { 19539a305230SLee Schermerhorn BUG(); 19549a305230SLee Schermerhorn if (nidp) 19559a305230SLee Schermerhorn *nidp = -1; 19569a305230SLee Schermerhorn return NULL; 19579a305230SLee Schermerhorn } 19589a305230SLee Schermerhorn 19599a305230SLee Schermerhorn static void hugetlb_unregister_all_nodes(void) { } 19609a305230SLee Schermerhorn 19619a305230SLee Schermerhorn static void hugetlb_register_all_nodes(void) { } 19629a305230SLee Schermerhorn 19639a305230SLee Schermerhorn #endif 19649a305230SLee Schermerhorn 1965a3437870SNishanth Aravamudan static void __exit hugetlb_exit(void) 1966a3437870SNishanth Aravamudan { 1967a3437870SNishanth Aravamudan struct hstate *h; 1968a3437870SNishanth Aravamudan 19699a305230SLee Schermerhorn hugetlb_unregister_all_nodes(); 19709a305230SLee Schermerhorn 1971a3437870SNishanth Aravamudan for_each_hstate(h) { 1972972dc4deSAneesh Kumar K.V kobject_put(hstate_kobjs[hstate_index(h)]); 1973a3437870SNishanth Aravamudan } 1974a3437870SNishanth Aravamudan 1975a3437870SNishanth Aravamudan kobject_put(hugepages_kobj); 19768382d914SDavidlohr Bueso kfree(htlb_fault_mutex_table); 1977a3437870SNishanth Aravamudan } 1978a3437870SNishanth Aravamudan module_exit(hugetlb_exit); 1979a3437870SNishanth Aravamudan 1980a3437870SNishanth Aravamudan static int __init hugetlb_init(void) 1981a3437870SNishanth Aravamudan { 19828382d914SDavidlohr Bueso int i; 19838382d914SDavidlohr Bueso 1984457c1b27SNishanth Aravamudan if (!hugepages_supported()) 19850ef89d25SBenjamin Herrenschmidt return 0; 1986a3437870SNishanth Aravamudan 1987e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) { 1988e11bfbfcSNick Piggin default_hstate_size = HPAGE_SIZE; 1989e11bfbfcSNick Piggin if (!size_to_hstate(default_hstate_size)) 1990a3437870SNishanth Aravamudan hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 1991a3437870SNishanth Aravamudan } 1992972dc4deSAneesh Kumar K.V default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); 1993e11bfbfcSNick Piggin if (default_hstate_max_huge_pages) 1994e11bfbfcSNick Piggin default_hstate.max_huge_pages = default_hstate_max_huge_pages; 1995a3437870SNishanth Aravamudan 1996a3437870SNishanth Aravamudan hugetlb_init_hstates(); 1997aa888a74SAndi Kleen gather_bootmem_prealloc(); 1998a3437870SNishanth Aravamudan report_hugepages(); 1999a3437870SNishanth Aravamudan 2000a3437870SNishanth Aravamudan hugetlb_sysfs_init(); 20019a305230SLee Schermerhorn hugetlb_register_all_nodes(); 20027179e7bfSJianguo Wu hugetlb_cgroup_file_init(); 20039a305230SLee Schermerhorn 20048382d914SDavidlohr Bueso #ifdef CONFIG_SMP 20058382d914SDavidlohr Bueso num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 20068382d914SDavidlohr Bueso #else 20078382d914SDavidlohr Bueso num_fault_mutexes = 1; 20088382d914SDavidlohr Bueso #endif 20098382d914SDavidlohr Bueso htlb_fault_mutex_table = 20108382d914SDavidlohr Bueso kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); 20118382d914SDavidlohr Bueso BUG_ON(!htlb_fault_mutex_table); 20128382d914SDavidlohr Bueso 20138382d914SDavidlohr Bueso for (i = 0; i < num_fault_mutexes; i++) 20148382d914SDavidlohr Bueso mutex_init(&htlb_fault_mutex_table[i]); 2015a3437870SNishanth Aravamudan return 0; 2016a3437870SNishanth Aravamudan } 2017a3437870SNishanth Aravamudan module_init(hugetlb_init); 2018a3437870SNishanth Aravamudan 2019a3437870SNishanth Aravamudan /* Should be called on processing a hugepagesz=... option */ 2020a3437870SNishanth Aravamudan void __init hugetlb_add_hstate(unsigned order) 2021a3437870SNishanth Aravamudan { 2022a3437870SNishanth Aravamudan struct hstate *h; 20238faa8b07SAndi Kleen unsigned long i; 20248faa8b07SAndi Kleen 2025a3437870SNishanth Aravamudan if (size_to_hstate(PAGE_SIZE << order)) { 2026ffb22af5SAndrew Morton pr_warning("hugepagesz= specified twice, ignoring\n"); 2027a3437870SNishanth Aravamudan return; 2028a3437870SNishanth Aravamudan } 202947d38344SAneesh Kumar K.V BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 2030a3437870SNishanth Aravamudan BUG_ON(order == 0); 203147d38344SAneesh Kumar K.V h = &hstates[hugetlb_max_hstate++]; 2032a3437870SNishanth Aravamudan h->order = order; 2033a3437870SNishanth Aravamudan h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); 20348faa8b07SAndi Kleen h->nr_huge_pages = 0; 20358faa8b07SAndi Kleen h->free_huge_pages = 0; 20368faa8b07SAndi Kleen for (i = 0; i < MAX_NUMNODES; ++i) 20378faa8b07SAndi Kleen INIT_LIST_HEAD(&h->hugepage_freelists[i]); 20380edaecfaSAneesh Kumar K.V INIT_LIST_HEAD(&h->hugepage_activelist); 20398cebfcd0SLai Jiangshan h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); 20408cebfcd0SLai Jiangshan h->next_nid_to_free = first_node(node_states[N_MEMORY]); 2041a3437870SNishanth Aravamudan snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 2042a3437870SNishanth Aravamudan huge_page_size(h)/1024); 20438faa8b07SAndi Kleen 2044a3437870SNishanth Aravamudan parsed_hstate = h; 2045a3437870SNishanth Aravamudan } 2046a3437870SNishanth Aravamudan 2047e11bfbfcSNick Piggin static int __init hugetlb_nrpages_setup(char *s) 2048a3437870SNishanth Aravamudan { 2049a3437870SNishanth Aravamudan unsigned long *mhp; 20508faa8b07SAndi Kleen static unsigned long *last_mhp; 2051a3437870SNishanth Aravamudan 2052a3437870SNishanth Aravamudan /* 205347d38344SAneesh Kumar K.V * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, 2054a3437870SNishanth Aravamudan * so this hugepages= parameter goes to the "default hstate". 2055a3437870SNishanth Aravamudan */ 205647d38344SAneesh Kumar K.V if (!hugetlb_max_hstate) 2057a3437870SNishanth Aravamudan mhp = &default_hstate_max_huge_pages; 2058a3437870SNishanth Aravamudan else 2059a3437870SNishanth Aravamudan mhp = &parsed_hstate->max_huge_pages; 2060a3437870SNishanth Aravamudan 20618faa8b07SAndi Kleen if (mhp == last_mhp) { 2062ffb22af5SAndrew Morton pr_warning("hugepages= specified twice without " 20638faa8b07SAndi Kleen "interleaving hugepagesz=, ignoring\n"); 20648faa8b07SAndi Kleen return 1; 20658faa8b07SAndi Kleen } 20668faa8b07SAndi Kleen 2067a3437870SNishanth Aravamudan if (sscanf(s, "%lu", mhp) <= 0) 2068a3437870SNishanth Aravamudan *mhp = 0; 2069a3437870SNishanth Aravamudan 20708faa8b07SAndi Kleen /* 20718faa8b07SAndi Kleen * Global state is always initialized later in hugetlb_init. 20728faa8b07SAndi Kleen * But we need to allocate >= MAX_ORDER hstates here early to still 20738faa8b07SAndi Kleen * use the bootmem allocator. 20748faa8b07SAndi Kleen */ 207547d38344SAneesh Kumar K.V if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) 20768faa8b07SAndi Kleen hugetlb_hstate_alloc_pages(parsed_hstate); 20778faa8b07SAndi Kleen 20788faa8b07SAndi Kleen last_mhp = mhp; 20798faa8b07SAndi Kleen 2080a3437870SNishanth Aravamudan return 1; 2081a3437870SNishanth Aravamudan } 2082e11bfbfcSNick Piggin __setup("hugepages=", hugetlb_nrpages_setup); 2083e11bfbfcSNick Piggin 2084e11bfbfcSNick Piggin static int __init hugetlb_default_setup(char *s) 2085e11bfbfcSNick Piggin { 2086e11bfbfcSNick Piggin default_hstate_size = memparse(s, &s); 2087e11bfbfcSNick Piggin return 1; 2088e11bfbfcSNick Piggin } 2089e11bfbfcSNick Piggin __setup("default_hugepagesz=", hugetlb_default_setup); 2090a3437870SNishanth Aravamudan 20918a213460SNishanth Aravamudan static unsigned int cpuset_mems_nr(unsigned int *array) 20928a213460SNishanth Aravamudan { 20938a213460SNishanth Aravamudan int node; 20948a213460SNishanth Aravamudan unsigned int nr = 0; 20958a213460SNishanth Aravamudan 20968a213460SNishanth Aravamudan for_each_node_mask(node, cpuset_current_mems_allowed) 20978a213460SNishanth Aravamudan nr += array[node]; 20988a213460SNishanth Aravamudan 20998a213460SNishanth Aravamudan return nr; 21008a213460SNishanth Aravamudan } 21018a213460SNishanth Aravamudan 21028a213460SNishanth Aravamudan #ifdef CONFIG_SYSCTL 210306808b08SLee Schermerhorn static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 210406808b08SLee Schermerhorn struct ctl_table *table, int write, 210506808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 21061da177e4SLinus Torvalds { 2107e5ff2159SAndi Kleen struct hstate *h = &default_hstate; 2108e5ff2159SAndi Kleen unsigned long tmp; 210908d4a246SMichal Hocko int ret; 2110e5ff2159SAndi Kleen 2111457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2112457c1b27SNishanth Aravamudan return -ENOTSUPP; 2113457c1b27SNishanth Aravamudan 2114e5ff2159SAndi Kleen tmp = h->max_huge_pages; 2115e5ff2159SAndi Kleen 2116adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2117adbe8726SEric B Munson return -EINVAL; 2118adbe8726SEric B Munson 2119e5ff2159SAndi Kleen table->data = &tmp; 2120e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 212108d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 212208d4a246SMichal Hocko if (ret) 212308d4a246SMichal Hocko goto out; 2124e5ff2159SAndi Kleen 212506808b08SLee Schermerhorn if (write) { 2126bad44b5bSDavid Rientjes NODEMASK_ALLOC(nodemask_t, nodes_allowed, 2127bad44b5bSDavid Rientjes GFP_KERNEL | __GFP_NORETRY); 212806808b08SLee Schermerhorn if (!(obey_mempolicy && 212906808b08SLee Schermerhorn init_nodemask_of_mempolicy(nodes_allowed))) { 213006808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 21318cebfcd0SLai Jiangshan nodes_allowed = &node_states[N_MEMORY]; 213206808b08SLee Schermerhorn } 213306808b08SLee Schermerhorn h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); 213406808b08SLee Schermerhorn 21358cebfcd0SLai Jiangshan if (nodes_allowed != &node_states[N_MEMORY]) 213606808b08SLee Schermerhorn NODEMASK_FREE(nodes_allowed); 213706808b08SLee Schermerhorn } 213808d4a246SMichal Hocko out: 213908d4a246SMichal Hocko return ret; 21401da177e4SLinus Torvalds } 2141396faf03SMel Gorman 214206808b08SLee Schermerhorn int hugetlb_sysctl_handler(struct ctl_table *table, int write, 214306808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 214406808b08SLee Schermerhorn { 214506808b08SLee Schermerhorn 214606808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(false, table, write, 214706808b08SLee Schermerhorn buffer, length, ppos); 214806808b08SLee Schermerhorn } 214906808b08SLee Schermerhorn 215006808b08SLee Schermerhorn #ifdef CONFIG_NUMA 215106808b08SLee Schermerhorn int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, 215206808b08SLee Schermerhorn void __user *buffer, size_t *length, loff_t *ppos) 215306808b08SLee Schermerhorn { 215406808b08SLee Schermerhorn return hugetlb_sysctl_handler_common(true, table, write, 215506808b08SLee Schermerhorn buffer, length, ppos); 215606808b08SLee Schermerhorn } 215706808b08SLee Schermerhorn #endif /* CONFIG_NUMA */ 215806808b08SLee Schermerhorn 2159a3d0c6aaSNishanth Aravamudan int hugetlb_overcommit_handler(struct ctl_table *table, int write, 21608d65af78SAlexey Dobriyan void __user *buffer, 2161a3d0c6aaSNishanth Aravamudan size_t *length, loff_t *ppos) 2162a3d0c6aaSNishanth Aravamudan { 2163a5516438SAndi Kleen struct hstate *h = &default_hstate; 2164e5ff2159SAndi Kleen unsigned long tmp; 216508d4a246SMichal Hocko int ret; 2166e5ff2159SAndi Kleen 2167457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2168457c1b27SNishanth Aravamudan return -ENOTSUPP; 2169457c1b27SNishanth Aravamudan 2170e5ff2159SAndi Kleen tmp = h->nr_overcommit_huge_pages; 2171e5ff2159SAndi Kleen 2172adbe8726SEric B Munson if (write && h->order >= MAX_ORDER) 2173adbe8726SEric B Munson return -EINVAL; 2174adbe8726SEric B Munson 2175e5ff2159SAndi Kleen table->data = &tmp; 2176e5ff2159SAndi Kleen table->maxlen = sizeof(unsigned long); 217708d4a246SMichal Hocko ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); 217808d4a246SMichal Hocko if (ret) 217908d4a246SMichal Hocko goto out; 2180e5ff2159SAndi Kleen 2181e5ff2159SAndi Kleen if (write) { 2182064d9efeSNishanth Aravamudan spin_lock(&hugetlb_lock); 2183e5ff2159SAndi Kleen h->nr_overcommit_huge_pages = tmp; 2184a3d0c6aaSNishanth Aravamudan spin_unlock(&hugetlb_lock); 2185e5ff2159SAndi Kleen } 218608d4a246SMichal Hocko out: 218708d4a246SMichal Hocko return ret; 2188a3d0c6aaSNishanth Aravamudan } 2189a3d0c6aaSNishanth Aravamudan 21901da177e4SLinus Torvalds #endif /* CONFIG_SYSCTL */ 21911da177e4SLinus Torvalds 2192e1759c21SAlexey Dobriyan void hugetlb_report_meminfo(struct seq_file *m) 21931da177e4SLinus Torvalds { 2194a5516438SAndi Kleen struct hstate *h = &default_hstate; 2195457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2196457c1b27SNishanth Aravamudan return; 2197e1759c21SAlexey Dobriyan seq_printf(m, 21981da177e4SLinus Torvalds "HugePages_Total: %5lu\n" 21991da177e4SLinus Torvalds "HugePages_Free: %5lu\n" 2200b45b5bd6SDavid Gibson "HugePages_Rsvd: %5lu\n" 22017893d1d5SAdam Litke "HugePages_Surp: %5lu\n" 22024f98a2feSRik van Riel "Hugepagesize: %8lu kB\n", 2203a5516438SAndi Kleen h->nr_huge_pages, 2204a5516438SAndi Kleen h->free_huge_pages, 2205a5516438SAndi Kleen h->resv_huge_pages, 2206a5516438SAndi Kleen h->surplus_huge_pages, 2207a5516438SAndi Kleen 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 22081da177e4SLinus Torvalds } 22091da177e4SLinus Torvalds 22101da177e4SLinus Torvalds int hugetlb_report_node_meminfo(int nid, char *buf) 22111da177e4SLinus Torvalds { 2212a5516438SAndi Kleen struct hstate *h = &default_hstate; 2213457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2214457c1b27SNishanth Aravamudan return 0; 22151da177e4SLinus Torvalds return sprintf(buf, 22161da177e4SLinus Torvalds "Node %d HugePages_Total: %5u\n" 2217a1de0919SNishanth Aravamudan "Node %d HugePages_Free: %5u\n" 2218a1de0919SNishanth Aravamudan "Node %d HugePages_Surp: %5u\n", 2219a5516438SAndi Kleen nid, h->nr_huge_pages_node[nid], 2220a5516438SAndi Kleen nid, h->free_huge_pages_node[nid], 2221a5516438SAndi Kleen nid, h->surplus_huge_pages_node[nid]); 22221da177e4SLinus Torvalds } 22231da177e4SLinus Torvalds 2224949f7ec5SDavid Rientjes void hugetlb_show_meminfo(void) 2225949f7ec5SDavid Rientjes { 2226949f7ec5SDavid Rientjes struct hstate *h; 2227949f7ec5SDavid Rientjes int nid; 2228949f7ec5SDavid Rientjes 2229457c1b27SNishanth Aravamudan if (!hugepages_supported()) 2230457c1b27SNishanth Aravamudan return; 2231457c1b27SNishanth Aravamudan 2232949f7ec5SDavid Rientjes for_each_node_state(nid, N_MEMORY) 2233949f7ec5SDavid Rientjes for_each_hstate(h) 2234949f7ec5SDavid Rientjes pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 2235949f7ec5SDavid Rientjes nid, 2236949f7ec5SDavid Rientjes h->nr_huge_pages_node[nid], 2237949f7ec5SDavid Rientjes h->free_huge_pages_node[nid], 2238949f7ec5SDavid Rientjes h->surplus_huge_pages_node[nid], 2239949f7ec5SDavid Rientjes 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); 2240949f7ec5SDavid Rientjes } 2241949f7ec5SDavid Rientjes 22421da177e4SLinus Torvalds /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 22431da177e4SLinus Torvalds unsigned long hugetlb_total_pages(void) 22441da177e4SLinus Torvalds { 2245d0028588SWanpeng Li struct hstate *h; 2246d0028588SWanpeng Li unsigned long nr_total_pages = 0; 2247d0028588SWanpeng Li 2248d0028588SWanpeng Li for_each_hstate(h) 2249d0028588SWanpeng Li nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 2250d0028588SWanpeng Li return nr_total_pages; 22511da177e4SLinus Torvalds } 22521da177e4SLinus Torvalds 2253a5516438SAndi Kleen static int hugetlb_acct_memory(struct hstate *h, long delta) 2254fc1b8a73SMel Gorman { 2255fc1b8a73SMel Gorman int ret = -ENOMEM; 2256fc1b8a73SMel Gorman 2257fc1b8a73SMel Gorman spin_lock(&hugetlb_lock); 2258fc1b8a73SMel Gorman /* 2259fc1b8a73SMel Gorman * When cpuset is configured, it breaks the strict hugetlb page 2260fc1b8a73SMel Gorman * reservation as the accounting is done on a global variable. Such 2261fc1b8a73SMel Gorman * reservation is completely rubbish in the presence of cpuset because 2262fc1b8a73SMel Gorman * the reservation is not checked against page availability for the 2263fc1b8a73SMel Gorman * current cpuset. Application can still potentially OOM'ed by kernel 2264fc1b8a73SMel Gorman * with lack of free htlb page in cpuset that the task is in. 2265fc1b8a73SMel Gorman * Attempt to enforce strict accounting with cpuset is almost 2266fc1b8a73SMel Gorman * impossible (or too ugly) because cpuset is too fluid that 2267fc1b8a73SMel Gorman * task or memory node can be dynamically moved between cpusets. 2268fc1b8a73SMel Gorman * 2269fc1b8a73SMel Gorman * The change of semantics for shared hugetlb mapping with cpuset is 2270fc1b8a73SMel Gorman * undesirable. However, in order to preserve some of the semantics, 2271fc1b8a73SMel Gorman * we fall back to check against current free page availability as 2272fc1b8a73SMel Gorman * a best attempt and hopefully to minimize the impact of changing 2273fc1b8a73SMel Gorman * semantics that cpuset has. 2274fc1b8a73SMel Gorman */ 2275fc1b8a73SMel Gorman if (delta > 0) { 2276a5516438SAndi Kleen if (gather_surplus_pages(h, delta) < 0) 2277fc1b8a73SMel Gorman goto out; 2278fc1b8a73SMel Gorman 2279a5516438SAndi Kleen if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { 2280a5516438SAndi Kleen return_unused_surplus_pages(h, delta); 2281fc1b8a73SMel Gorman goto out; 2282fc1b8a73SMel Gorman } 2283fc1b8a73SMel Gorman } 2284fc1b8a73SMel Gorman 2285fc1b8a73SMel Gorman ret = 0; 2286fc1b8a73SMel Gorman if (delta < 0) 2287a5516438SAndi Kleen return_unused_surplus_pages(h, (unsigned long) -delta); 2288fc1b8a73SMel Gorman 2289fc1b8a73SMel Gorman out: 2290fc1b8a73SMel Gorman spin_unlock(&hugetlb_lock); 2291fc1b8a73SMel Gorman return ret; 2292fc1b8a73SMel Gorman } 2293fc1b8a73SMel Gorman 229484afd99bSAndy Whitcroft static void hugetlb_vm_op_open(struct vm_area_struct *vma) 229584afd99bSAndy Whitcroft { 2296f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 229784afd99bSAndy Whitcroft 229884afd99bSAndy Whitcroft /* 229984afd99bSAndy Whitcroft * This new VMA should share its siblings reservation map if present. 230084afd99bSAndy Whitcroft * The VMA will only ever have a valid reservation map pointer where 230184afd99bSAndy Whitcroft * it is being copied for another still existing VMA. As that VMA 230225985edcSLucas De Marchi * has a reference to the reservation map it cannot disappear until 230384afd99bSAndy Whitcroft * after this open call completes. It is therefore safe to take a 230484afd99bSAndy Whitcroft * new reference here without additional locking. 230584afd99bSAndy Whitcroft */ 23064e35f483SJoonsoo Kim if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 2307f522c3acSJoonsoo Kim kref_get(&resv->refs); 230884afd99bSAndy Whitcroft } 230984afd99bSAndy Whitcroft 2310a1e78772SMel Gorman static void hugetlb_vm_op_close(struct vm_area_struct *vma) 2311a1e78772SMel Gorman { 2312a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2313f522c3acSJoonsoo Kim struct resv_map *resv = vma_resv_map(vma); 231490481622SDavid Gibson struct hugepage_subpool *spool = subpool_vma(vma); 23154e35f483SJoonsoo Kim unsigned long reserve, start, end; 231684afd99bSAndy Whitcroft 23174e35f483SJoonsoo Kim if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 23184e35f483SJoonsoo Kim return; 23194e35f483SJoonsoo Kim 2320a5516438SAndi Kleen start = vma_hugecache_offset(h, vma, vma->vm_start); 2321a5516438SAndi Kleen end = vma_hugecache_offset(h, vma, vma->vm_end); 232284afd99bSAndy Whitcroft 23234e35f483SJoonsoo Kim reserve = (end - start) - region_count(resv, start, end); 232484afd99bSAndy Whitcroft 2325f031dd27SJoonsoo Kim kref_put(&resv->refs, resv_map_release); 232684afd99bSAndy Whitcroft 23277251ff78SAdam Litke if (reserve) { 2328a5516438SAndi Kleen hugetlb_acct_memory(h, -reserve); 232990481622SDavid Gibson hugepage_subpool_put_pages(spool, reserve); 23307251ff78SAdam Litke } 2331a1e78772SMel Gorman } 2332a1e78772SMel Gorman 23331da177e4SLinus Torvalds /* 23341da177e4SLinus Torvalds * We cannot handle pagefaults against hugetlb pages at all. They cause 23351da177e4SLinus Torvalds * handle_mm_fault() to try to instantiate regular-sized pages in the 23361da177e4SLinus Torvalds * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get 23371da177e4SLinus Torvalds * this far. 23381da177e4SLinus Torvalds */ 2339d0217ac0SNick Piggin static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 23401da177e4SLinus Torvalds { 23411da177e4SLinus Torvalds BUG(); 2342d0217ac0SNick Piggin return 0; 23431da177e4SLinus Torvalds } 23441da177e4SLinus Torvalds 2345f0f37e2fSAlexey Dobriyan const struct vm_operations_struct hugetlb_vm_ops = { 2346d0217ac0SNick Piggin .fault = hugetlb_vm_op_fault, 234784afd99bSAndy Whitcroft .open = hugetlb_vm_op_open, 2348a1e78772SMel Gorman .close = hugetlb_vm_op_close, 23491da177e4SLinus Torvalds }; 23501da177e4SLinus Torvalds 23511e8f889bSDavid Gibson static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 23521e8f889bSDavid Gibson int writable) 235363551ae0SDavid Gibson { 235463551ae0SDavid Gibson pte_t entry; 235563551ae0SDavid Gibson 23561e8f889bSDavid Gibson if (writable) { 2357106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 2358106c992aSGerald Schaefer vma->vm_page_prot))); 235963551ae0SDavid Gibson } else { 2360106c992aSGerald Schaefer entry = huge_pte_wrprotect(mk_huge_pte(page, 2361106c992aSGerald Schaefer vma->vm_page_prot)); 236263551ae0SDavid Gibson } 236363551ae0SDavid Gibson entry = pte_mkyoung(entry); 236463551ae0SDavid Gibson entry = pte_mkhuge(entry); 2365d9ed9faaSChris Metcalf entry = arch_make_huge_pte(entry, vma, page, writable); 236663551ae0SDavid Gibson 236763551ae0SDavid Gibson return entry; 236863551ae0SDavid Gibson } 236963551ae0SDavid Gibson 23701e8f889bSDavid Gibson static void set_huge_ptep_writable(struct vm_area_struct *vma, 23711e8f889bSDavid Gibson unsigned long address, pte_t *ptep) 23721e8f889bSDavid Gibson { 23731e8f889bSDavid Gibson pte_t entry; 23741e8f889bSDavid Gibson 2375106c992aSGerald Schaefer entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); 237632f84528SChris Forbes if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 23774b3073e1SRussell King update_mmu_cache(vma, address, ptep); 23781e8f889bSDavid Gibson } 23791e8f889bSDavid Gibson 23801e8f889bSDavid Gibson 238163551ae0SDavid Gibson int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 238263551ae0SDavid Gibson struct vm_area_struct *vma) 238363551ae0SDavid Gibson { 238463551ae0SDavid Gibson pte_t *src_pte, *dst_pte, entry; 238563551ae0SDavid Gibson struct page *ptepage; 23861c59827dSHugh Dickins unsigned long addr; 23871e8f889bSDavid Gibson int cow; 2388a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2389a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 2390e8569dd2SAndreas Sandberg unsigned long mmun_start; /* For mmu_notifiers */ 2391e8569dd2SAndreas Sandberg unsigned long mmun_end; /* For mmu_notifiers */ 2392e8569dd2SAndreas Sandberg int ret = 0; 23931e8f889bSDavid Gibson 23941e8f889bSDavid Gibson cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; 239563551ae0SDavid Gibson 2396e8569dd2SAndreas Sandberg mmun_start = vma->vm_start; 2397e8569dd2SAndreas Sandberg mmun_end = vma->vm_end; 2398e8569dd2SAndreas Sandberg if (cow) 2399e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); 2400e8569dd2SAndreas Sandberg 2401a5516438SAndi Kleen for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { 2402cb900f41SKirill A. Shutemov spinlock_t *src_ptl, *dst_ptl; 2403c74df32cSHugh Dickins src_pte = huge_pte_offset(src, addr); 2404c74df32cSHugh Dickins if (!src_pte) 2405c74df32cSHugh Dickins continue; 2406a5516438SAndi Kleen dst_pte = huge_pte_alloc(dst, addr, sz); 2407e8569dd2SAndreas Sandberg if (!dst_pte) { 2408e8569dd2SAndreas Sandberg ret = -ENOMEM; 2409e8569dd2SAndreas Sandberg break; 2410e8569dd2SAndreas Sandberg } 2411c5c99429SLarry Woodman 2412c5c99429SLarry Woodman /* If the pagetables are shared don't copy or take references */ 2413c5c99429SLarry Woodman if (dst_pte == src_pte) 2414c5c99429SLarry Woodman continue; 2415c5c99429SLarry Woodman 2416cb900f41SKirill A. Shutemov dst_ptl = huge_pte_lock(h, dst, dst_pte); 2417cb900f41SKirill A. Shutemov src_ptl = huge_pte_lockptr(h, src, src_pte); 2418cb900f41SKirill A. Shutemov spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 24197f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(src_pte))) { 24201e8f889bSDavid Gibson if (cow) 24217f2e9525SGerald Schaefer huge_ptep_set_wrprotect(src, addr, src_pte); 24227f2e9525SGerald Schaefer entry = huge_ptep_get(src_pte); 242363551ae0SDavid Gibson ptepage = pte_page(entry); 242463551ae0SDavid Gibson get_page(ptepage); 24250fe6e20bSNaoya Horiguchi page_dup_rmap(ptepage); 242663551ae0SDavid Gibson set_huge_pte_at(dst, addr, dst_pte, entry); 24271c59827dSHugh Dickins } 2428cb900f41SKirill A. Shutemov spin_unlock(src_ptl); 2429cb900f41SKirill A. Shutemov spin_unlock(dst_ptl); 243063551ae0SDavid Gibson } 243163551ae0SDavid Gibson 2432e8569dd2SAndreas Sandberg if (cow) 2433e8569dd2SAndreas Sandberg mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); 2434e8569dd2SAndreas Sandberg 2435e8569dd2SAndreas Sandberg return ret; 243663551ae0SDavid Gibson } 243763551ae0SDavid Gibson 2438290408d4SNaoya Horiguchi static int is_hugetlb_entry_migration(pte_t pte) 2439290408d4SNaoya Horiguchi { 2440290408d4SNaoya Horiguchi swp_entry_t swp; 2441290408d4SNaoya Horiguchi 2442290408d4SNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2443290408d4SNaoya Horiguchi return 0; 2444290408d4SNaoya Horiguchi swp = pte_to_swp_entry(pte); 244532f84528SChris Forbes if (non_swap_entry(swp) && is_migration_entry(swp)) 2446290408d4SNaoya Horiguchi return 1; 244732f84528SChris Forbes else 2448290408d4SNaoya Horiguchi return 0; 2449290408d4SNaoya Horiguchi } 2450290408d4SNaoya Horiguchi 2451fd6a03edSNaoya Horiguchi static int is_hugetlb_entry_hwpoisoned(pte_t pte) 2452fd6a03edSNaoya Horiguchi { 2453fd6a03edSNaoya Horiguchi swp_entry_t swp; 2454fd6a03edSNaoya Horiguchi 2455fd6a03edSNaoya Horiguchi if (huge_pte_none(pte) || pte_present(pte)) 2456fd6a03edSNaoya Horiguchi return 0; 2457fd6a03edSNaoya Horiguchi swp = pte_to_swp_entry(pte); 245832f84528SChris Forbes if (non_swap_entry(swp) && is_hwpoison_entry(swp)) 2459fd6a03edSNaoya Horiguchi return 1; 246032f84528SChris Forbes else 2461fd6a03edSNaoya Horiguchi return 0; 2462fd6a03edSNaoya Horiguchi } 2463fd6a03edSNaoya Horiguchi 246424669e58SAneesh Kumar K.V void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 246524669e58SAneesh Kumar K.V unsigned long start, unsigned long end, 246624669e58SAneesh Kumar K.V struct page *ref_page) 246763551ae0SDavid Gibson { 246824669e58SAneesh Kumar K.V int force_flush = 0; 246963551ae0SDavid Gibson struct mm_struct *mm = vma->vm_mm; 247063551ae0SDavid Gibson unsigned long address; 2471c7546f8fSDavid Gibson pte_t *ptep; 247263551ae0SDavid Gibson pte_t pte; 2473cb900f41SKirill A. Shutemov spinlock_t *ptl; 247463551ae0SDavid Gibson struct page *page; 2475a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2476a5516438SAndi Kleen unsigned long sz = huge_page_size(h); 24772ec74c3eSSagi Grimberg const unsigned long mmun_start = start; /* For mmu_notifiers */ 24782ec74c3eSSagi Grimberg const unsigned long mmun_end = end; /* For mmu_notifiers */ 2479a5516438SAndi Kleen 248063551ae0SDavid Gibson WARN_ON(!is_vm_hugetlb_page(vma)); 2481a5516438SAndi Kleen BUG_ON(start & ~huge_page_mask(h)); 2482a5516438SAndi Kleen BUG_ON(end & ~huge_page_mask(h)); 248363551ae0SDavid Gibson 248424669e58SAneesh Kumar K.V tlb_start_vma(tlb, vma); 24852ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 248624669e58SAneesh Kumar K.V again: 2487a5516438SAndi Kleen for (address = start; address < end; address += sz) { 2488c7546f8fSDavid Gibson ptep = huge_pte_offset(mm, address); 2489c7546f8fSDavid Gibson if (!ptep) 2490c7546f8fSDavid Gibson continue; 2491c7546f8fSDavid Gibson 2492cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 249339dde65cSChen, Kenneth W if (huge_pmd_unshare(mm, &address, ptep)) 2494cb900f41SKirill A. Shutemov goto unlock; 249539dde65cSChen, Kenneth W 24966629326bSHillf Danton pte = huge_ptep_get(ptep); 24976629326bSHillf Danton if (huge_pte_none(pte)) 2498cb900f41SKirill A. Shutemov goto unlock; 24996629326bSHillf Danton 25006629326bSHillf Danton /* 25016629326bSHillf Danton * HWPoisoned hugepage is already unmapped and dropped reference 25026629326bSHillf Danton */ 25038c4894c6SNaoya Horiguchi if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 2504106c992aSGerald Schaefer huge_pte_clear(mm, address, ptep); 2505cb900f41SKirill A. Shutemov goto unlock; 25068c4894c6SNaoya Horiguchi } 25076629326bSHillf Danton 25086629326bSHillf Danton page = pte_page(pte); 250904f2cbe3SMel Gorman /* 251004f2cbe3SMel Gorman * If a reference page is supplied, it is because a specific 251104f2cbe3SMel Gorman * page is being unmapped, not a range. Ensure the page we 251204f2cbe3SMel Gorman * are about to unmap is the actual page of interest. 251304f2cbe3SMel Gorman */ 251404f2cbe3SMel Gorman if (ref_page) { 251504f2cbe3SMel Gorman if (page != ref_page) 2516cb900f41SKirill A. Shutemov goto unlock; 251704f2cbe3SMel Gorman 251804f2cbe3SMel Gorman /* 251904f2cbe3SMel Gorman * Mark the VMA as having unmapped its page so that 252004f2cbe3SMel Gorman * future faults in this VMA will fail rather than 252104f2cbe3SMel Gorman * looking like data was lost 252204f2cbe3SMel Gorman */ 252304f2cbe3SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 252404f2cbe3SMel Gorman } 252504f2cbe3SMel Gorman 2526c7546f8fSDavid Gibson pte = huge_ptep_get_and_clear(mm, address, ptep); 252724669e58SAneesh Kumar K.V tlb_remove_tlb_entry(tlb, ptep, address); 2528106c992aSGerald Schaefer if (huge_pte_dirty(pte)) 25296649a386SKen Chen set_page_dirty(page); 25309e81130bSHillf Danton 253124669e58SAneesh Kumar K.V page_remove_rmap(page); 253224669e58SAneesh Kumar K.V force_flush = !__tlb_remove_page(tlb, page); 2533cb900f41SKirill A. Shutemov if (force_flush) { 2534cb900f41SKirill A. Shutemov spin_unlock(ptl); 25359e81130bSHillf Danton break; 253663551ae0SDavid Gibson } 2537cb900f41SKirill A. Shutemov /* Bail out after unmapping reference page if supplied */ 2538cb900f41SKirill A. Shutemov if (ref_page) { 2539cb900f41SKirill A. Shutemov spin_unlock(ptl); 2540cb900f41SKirill A. Shutemov break; 2541cb900f41SKirill A. Shutemov } 2542cb900f41SKirill A. Shutemov unlock: 2543cb900f41SKirill A. Shutemov spin_unlock(ptl); 2544cb900f41SKirill A. Shutemov } 254524669e58SAneesh Kumar K.V /* 254624669e58SAneesh Kumar K.V * mmu_gather ran out of room to batch pages, we break out of 254724669e58SAneesh Kumar K.V * the PTE lock to avoid doing the potential expensive TLB invalidate 254824669e58SAneesh Kumar K.V * and page-free while holding it. 254924669e58SAneesh Kumar K.V */ 255024669e58SAneesh Kumar K.V if (force_flush) { 255124669e58SAneesh Kumar K.V force_flush = 0; 255224669e58SAneesh Kumar K.V tlb_flush_mmu(tlb); 255324669e58SAneesh Kumar K.V if (address < end && !ref_page) 255424669e58SAneesh Kumar K.V goto again; 2555fe1668aeSChen, Kenneth W } 25562ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 255724669e58SAneesh Kumar K.V tlb_end_vma(tlb, vma); 25581da177e4SLinus Torvalds } 255963551ae0SDavid Gibson 2560d833352aSMel Gorman void __unmap_hugepage_range_final(struct mmu_gather *tlb, 2561d833352aSMel Gorman struct vm_area_struct *vma, unsigned long start, 2562d833352aSMel Gorman unsigned long end, struct page *ref_page) 2563d833352aSMel Gorman { 2564d833352aSMel Gorman __unmap_hugepage_range(tlb, vma, start, end, ref_page); 2565d833352aSMel Gorman 2566d833352aSMel Gorman /* 2567d833352aSMel Gorman * Clear this flag so that x86's huge_pmd_share page_table_shareable 2568d833352aSMel Gorman * test will fail on a vma being torn down, and not grab a page table 2569d833352aSMel Gorman * on its way out. We're lucky that the flag has such an appropriate 2570d833352aSMel Gorman * name, and can in fact be safely cleared here. We could clear it 2571d833352aSMel Gorman * before the __unmap_hugepage_range above, but all that's necessary 2572d833352aSMel Gorman * is to clear it before releasing the i_mmap_mutex. This works 2573d833352aSMel Gorman * because in the context this is called, the VMA is about to be 2574d833352aSMel Gorman * destroyed and the i_mmap_mutex is held. 2575d833352aSMel Gorman */ 2576d833352aSMel Gorman vma->vm_flags &= ~VM_MAYSHARE; 2577d833352aSMel Gorman } 2578d833352aSMel Gorman 2579502717f4SChen, Kenneth W void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 258004f2cbe3SMel Gorman unsigned long end, struct page *ref_page) 2581502717f4SChen, Kenneth W { 258224669e58SAneesh Kumar K.V struct mm_struct *mm; 258324669e58SAneesh Kumar K.V struct mmu_gather tlb; 258424669e58SAneesh Kumar K.V 258524669e58SAneesh Kumar K.V mm = vma->vm_mm; 258624669e58SAneesh Kumar K.V 25872b047252SLinus Torvalds tlb_gather_mmu(&tlb, mm, start, end); 258824669e58SAneesh Kumar K.V __unmap_hugepage_range(&tlb, vma, start, end, ref_page); 258924669e58SAneesh Kumar K.V tlb_finish_mmu(&tlb, start, end); 2590502717f4SChen, Kenneth W } 2591502717f4SChen, Kenneth W 259204f2cbe3SMel Gorman /* 259304f2cbe3SMel Gorman * This is called when the original mapper is failing to COW a MAP_PRIVATE 259404f2cbe3SMel Gorman * mappping it owns the reserve page for. The intention is to unmap the page 259504f2cbe3SMel Gorman * from other VMAs and let the children be SIGKILLed if they are faulting the 259604f2cbe3SMel Gorman * same region. 259704f2cbe3SMel Gorman */ 25982a4b3dedSHarvey Harrison static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 25992a4b3dedSHarvey Harrison struct page *page, unsigned long address) 260004f2cbe3SMel Gorman { 26017526674dSAdam Litke struct hstate *h = hstate_vma(vma); 260204f2cbe3SMel Gorman struct vm_area_struct *iter_vma; 260304f2cbe3SMel Gorman struct address_space *mapping; 260404f2cbe3SMel Gorman pgoff_t pgoff; 260504f2cbe3SMel Gorman 260604f2cbe3SMel Gorman /* 260704f2cbe3SMel Gorman * vm_pgoff is in PAGE_SIZE units, hence the different calculation 260804f2cbe3SMel Gorman * from page cache lookup which is in HPAGE_SIZE units. 260904f2cbe3SMel Gorman */ 26107526674dSAdam Litke address = address & huge_page_mask(h); 261136e4f20aSMichal Hocko pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 261236e4f20aSMichal Hocko vma->vm_pgoff; 2613496ad9aaSAl Viro mapping = file_inode(vma->vm_file)->i_mapping; 261404f2cbe3SMel Gorman 26154eb2b1dcSMel Gorman /* 26164eb2b1dcSMel Gorman * Take the mapping lock for the duration of the table walk. As 26174eb2b1dcSMel Gorman * this mapping should be shared between all the VMAs, 26184eb2b1dcSMel Gorman * __unmap_hugepage_range() is called as the lock is already held 26194eb2b1dcSMel Gorman */ 26203d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 26216b2dbba8SMichel Lespinasse vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 262204f2cbe3SMel Gorman /* Do not unmap the current VMA */ 262304f2cbe3SMel Gorman if (iter_vma == vma) 262404f2cbe3SMel Gorman continue; 262504f2cbe3SMel Gorman 262604f2cbe3SMel Gorman /* 262704f2cbe3SMel Gorman * Unmap the page from other VMAs without their own reserves. 262804f2cbe3SMel Gorman * They get marked to be SIGKILLed if they fault in these 262904f2cbe3SMel Gorman * areas. This is because a future no-page fault on this VMA 263004f2cbe3SMel Gorman * could insert a zeroed page instead of the data existing 263104f2cbe3SMel Gorman * from the time of fork. This would look like data corruption 263204f2cbe3SMel Gorman */ 263304f2cbe3SMel Gorman if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 263424669e58SAneesh Kumar K.V unmap_hugepage_range(iter_vma, address, 263524669e58SAneesh Kumar K.V address + huge_page_size(h), page); 263604f2cbe3SMel Gorman } 26373d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 263804f2cbe3SMel Gorman 263904f2cbe3SMel Gorman return 1; 264004f2cbe3SMel Gorman } 264104f2cbe3SMel Gorman 26420fe6e20bSNaoya Horiguchi /* 26430fe6e20bSNaoya Horiguchi * Hugetlb_cow() should be called with page lock of the original hugepage held. 2644ef009b25SMichal Hocko * Called with hugetlb_instantiation_mutex held and pte_page locked so we 2645ef009b25SMichal Hocko * cannot race with other handlers or page migration. 2646ef009b25SMichal Hocko * Keep the pte_same checks anyway to make transition from the mutex easier. 26470fe6e20bSNaoya Horiguchi */ 26481e8f889bSDavid Gibson static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, 264904f2cbe3SMel Gorman unsigned long address, pte_t *ptep, pte_t pte, 2650cb900f41SKirill A. Shutemov struct page *pagecache_page, spinlock_t *ptl) 26511e8f889bSDavid Gibson { 2652a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 26531e8f889bSDavid Gibson struct page *old_page, *new_page; 265404f2cbe3SMel Gorman int outside_reserve = 0; 26552ec74c3eSSagi Grimberg unsigned long mmun_start; /* For mmu_notifiers */ 26562ec74c3eSSagi Grimberg unsigned long mmun_end; /* For mmu_notifiers */ 26571e8f889bSDavid Gibson 26581e8f889bSDavid Gibson old_page = pte_page(pte); 26591e8f889bSDavid Gibson 266004f2cbe3SMel Gorman retry_avoidcopy: 26611e8f889bSDavid Gibson /* If no-one else is actually using this page, avoid the copy 26621e8f889bSDavid Gibson * and just make the page writable */ 266337a2140dSJoonsoo Kim if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { 26640fe6e20bSNaoya Horiguchi page_move_anon_rmap(old_page, vma, address); 26651e8f889bSDavid Gibson set_huge_ptep_writable(vma, address, ptep); 266683c54070SNick Piggin return 0; 26671e8f889bSDavid Gibson } 26681e8f889bSDavid Gibson 266904f2cbe3SMel Gorman /* 267004f2cbe3SMel Gorman * If the process that created a MAP_PRIVATE mapping is about to 267104f2cbe3SMel Gorman * perform a COW due to a shared page count, attempt to satisfy 267204f2cbe3SMel Gorman * the allocation without using the existing reserves. The pagecache 267304f2cbe3SMel Gorman * page is used to determine if the reserve at this address was 267404f2cbe3SMel Gorman * consumed or not. If reserves were used, a partial faulted mapping 267504f2cbe3SMel Gorman * at the time of fork() could consume its reserves on COW instead 267604f2cbe3SMel Gorman * of the full address range. 267704f2cbe3SMel Gorman */ 26785944d011SJoonsoo Kim if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 267904f2cbe3SMel Gorman old_page != pagecache_page) 268004f2cbe3SMel Gorman outside_reserve = 1; 268104f2cbe3SMel Gorman 26821e8f889bSDavid Gibson page_cache_get(old_page); 2683b76c8cfbSLarry Woodman 2684cb900f41SKirill A. Shutemov /* Drop page table lock as buddy allocator may be called */ 2685cb900f41SKirill A. Shutemov spin_unlock(ptl); 268604f2cbe3SMel Gorman new_page = alloc_huge_page(vma, address, outside_reserve); 26871e8f889bSDavid Gibson 26882fc39cecSAdam Litke if (IS_ERR(new_page)) { 268976dcee75SAneesh Kumar K.V long err = PTR_ERR(new_page); 26901e8f889bSDavid Gibson page_cache_release(old_page); 269104f2cbe3SMel Gorman 269204f2cbe3SMel Gorman /* 269304f2cbe3SMel Gorman * If a process owning a MAP_PRIVATE mapping fails to COW, 269404f2cbe3SMel Gorman * it is due to references held by a child and an insufficient 269504f2cbe3SMel Gorman * huge page pool. To guarantee the original mappers 269604f2cbe3SMel Gorman * reliability, unmap the page from child processes. The child 269704f2cbe3SMel Gorman * may get SIGKILLed if it later faults. 269804f2cbe3SMel Gorman */ 269904f2cbe3SMel Gorman if (outside_reserve) { 270004f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 270104f2cbe3SMel Gorman if (unmap_ref_private(mm, vma, old_page, address)) { 270204f2cbe3SMel Gorman BUG_ON(huge_pte_none(pte)); 2703cb900f41SKirill A. Shutemov spin_lock(ptl); 2704a734bcc8SHillf Danton ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2705a9af0c5dSNaoya Horiguchi if (likely(ptep && 2706a9af0c5dSNaoya Horiguchi pte_same(huge_ptep_get(ptep), pte))) 270704f2cbe3SMel Gorman goto retry_avoidcopy; 2708a734bcc8SHillf Danton /* 2709cb900f41SKirill A. Shutemov * race occurs while re-acquiring page table 2710cb900f41SKirill A. Shutemov * lock, and our job is done. 2711a734bcc8SHillf Danton */ 2712a734bcc8SHillf Danton return 0; 271304f2cbe3SMel Gorman } 271404f2cbe3SMel Gorman WARN_ON_ONCE(1); 271504f2cbe3SMel Gorman } 271604f2cbe3SMel Gorman 2717b76c8cfbSLarry Woodman /* Caller expects lock to be held */ 2718cb900f41SKirill A. Shutemov spin_lock(ptl); 271976dcee75SAneesh Kumar K.V if (err == -ENOMEM) 272076dcee75SAneesh Kumar K.V return VM_FAULT_OOM; 272176dcee75SAneesh Kumar K.V else 272276dcee75SAneesh Kumar K.V return VM_FAULT_SIGBUS; 27231e8f889bSDavid Gibson } 27241e8f889bSDavid Gibson 27250fe6e20bSNaoya Horiguchi /* 27260fe6e20bSNaoya Horiguchi * When the original hugepage is shared one, it does not have 27270fe6e20bSNaoya Horiguchi * anon_vma prepared. 27280fe6e20bSNaoya Horiguchi */ 272944e2aa93SDean Nelson if (unlikely(anon_vma_prepare(vma))) { 2730ea4039a3SHillf Danton page_cache_release(new_page); 2731ea4039a3SHillf Danton page_cache_release(old_page); 273244e2aa93SDean Nelson /* Caller expects lock to be held */ 2733cb900f41SKirill A. Shutemov spin_lock(ptl); 27340fe6e20bSNaoya Horiguchi return VM_FAULT_OOM; 273544e2aa93SDean Nelson } 27360fe6e20bSNaoya Horiguchi 273747ad8475SAndrea Arcangeli copy_user_huge_page(new_page, old_page, address, vma, 273847ad8475SAndrea Arcangeli pages_per_huge_page(h)); 27390ed361deSNick Piggin __SetPageUptodate(new_page); 27401e8f889bSDavid Gibson 27412ec74c3eSSagi Grimberg mmun_start = address & huge_page_mask(h); 27422ec74c3eSSagi Grimberg mmun_end = mmun_start + huge_page_size(h); 27432ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); 2744b76c8cfbSLarry Woodman /* 2745cb900f41SKirill A. Shutemov * Retake the page table lock to check for racing updates 2746b76c8cfbSLarry Woodman * before the page tables are altered 2747b76c8cfbSLarry Woodman */ 2748cb900f41SKirill A. Shutemov spin_lock(ptl); 2749a5516438SAndi Kleen ptep = huge_pte_offset(mm, address & huge_page_mask(h)); 2750a9af0c5dSNaoya Horiguchi if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { 275107443a85SJoonsoo Kim ClearPagePrivate(new_page); 275207443a85SJoonsoo Kim 27531e8f889bSDavid Gibson /* Break COW */ 27548fe627ecSGerald Schaefer huge_ptep_clear_flush(vma, address, ptep); 27551e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, 27561e8f889bSDavid Gibson make_huge_pte(vma, new_page, 1)); 27570fe6e20bSNaoya Horiguchi page_remove_rmap(old_page); 2758cd67f0d2SNaoya Horiguchi hugepage_add_new_anon_rmap(new_page, vma, address); 27591e8f889bSDavid Gibson /* Make the old page be freed below */ 27601e8f889bSDavid Gibson new_page = old_page; 27611e8f889bSDavid Gibson } 2762cb900f41SKirill A. Shutemov spin_unlock(ptl); 27632ec74c3eSSagi Grimberg mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); 27641e8f889bSDavid Gibson page_cache_release(new_page); 27651e8f889bSDavid Gibson page_cache_release(old_page); 27668312034fSJoonsoo Kim 27678312034fSJoonsoo Kim /* Caller expects lock to be held */ 2768cb900f41SKirill A. Shutemov spin_lock(ptl); 276983c54070SNick Piggin return 0; 27701e8f889bSDavid Gibson } 27711e8f889bSDavid Gibson 277204f2cbe3SMel Gorman /* Return the pagecache page at a given address within a VMA */ 2773a5516438SAndi Kleen static struct page *hugetlbfs_pagecache_page(struct hstate *h, 2774a5516438SAndi Kleen struct vm_area_struct *vma, unsigned long address) 277504f2cbe3SMel Gorman { 277604f2cbe3SMel Gorman struct address_space *mapping; 2777e7c4b0bfSAndy Whitcroft pgoff_t idx; 277804f2cbe3SMel Gorman 277904f2cbe3SMel Gorman mapping = vma->vm_file->f_mapping; 2780a5516438SAndi Kleen idx = vma_hugecache_offset(h, vma, address); 278104f2cbe3SMel Gorman 278204f2cbe3SMel Gorman return find_lock_page(mapping, idx); 278304f2cbe3SMel Gorman } 278404f2cbe3SMel Gorman 27853ae77f43SHugh Dickins /* 27863ae77f43SHugh Dickins * Return whether there is a pagecache page to back given address within VMA. 27873ae77f43SHugh Dickins * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. 27883ae77f43SHugh Dickins */ 27893ae77f43SHugh Dickins static bool hugetlbfs_pagecache_present(struct hstate *h, 27902a15efc9SHugh Dickins struct vm_area_struct *vma, unsigned long address) 27912a15efc9SHugh Dickins { 27922a15efc9SHugh Dickins struct address_space *mapping; 27932a15efc9SHugh Dickins pgoff_t idx; 27942a15efc9SHugh Dickins struct page *page; 27952a15efc9SHugh Dickins 27962a15efc9SHugh Dickins mapping = vma->vm_file->f_mapping; 27972a15efc9SHugh Dickins idx = vma_hugecache_offset(h, vma, address); 27982a15efc9SHugh Dickins 27992a15efc9SHugh Dickins page = find_get_page(mapping, idx); 28002a15efc9SHugh Dickins if (page) 28012a15efc9SHugh Dickins put_page(page); 28022a15efc9SHugh Dickins return page != NULL; 28032a15efc9SHugh Dickins } 28042a15efc9SHugh Dickins 2805a1ed3ddaSRobert P. J. Day static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, 28068382d914SDavidlohr Bueso struct address_space *mapping, pgoff_t idx, 2807788c7df4SHugh Dickins unsigned long address, pte_t *ptep, unsigned int flags) 2808ac9b9c66SHugh Dickins { 2809a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 2810ac9b9c66SHugh Dickins int ret = VM_FAULT_SIGBUS; 2811409eb8c2SHillf Danton int anon_rmap = 0; 28124c887265SAdam Litke unsigned long size; 28134c887265SAdam Litke struct page *page; 28141e8f889bSDavid Gibson pte_t new_pte; 2815cb900f41SKirill A. Shutemov spinlock_t *ptl; 28164c887265SAdam Litke 281704f2cbe3SMel Gorman /* 281804f2cbe3SMel Gorman * Currently, we are forced to kill the process in the event the 281904f2cbe3SMel Gorman * original mapper has unmapped pages from the child due to a failed 282025985edcSLucas De Marchi * COW. Warn that such a situation has occurred as it may not be obvious 282104f2cbe3SMel Gorman */ 282204f2cbe3SMel Gorman if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 2823ffb22af5SAndrew Morton pr_warning("PID %d killed due to inadequate hugepage pool\n", 282404f2cbe3SMel Gorman current->pid); 282504f2cbe3SMel Gorman return ret; 282604f2cbe3SMel Gorman } 282704f2cbe3SMel Gorman 28284c887265SAdam Litke /* 28294c887265SAdam Litke * Use page lock to guard against racing truncation 28304c887265SAdam Litke * before we get page_table_lock. 28314c887265SAdam Litke */ 28326bda666aSChristoph Lameter retry: 28336bda666aSChristoph Lameter page = find_lock_page(mapping, idx); 28346bda666aSChristoph Lameter if (!page) { 2835a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 2836ebed4bfcSHugh Dickins if (idx >= size) 2837ebed4bfcSHugh Dickins goto out; 283804f2cbe3SMel Gorman page = alloc_huge_page(vma, address, 0); 28392fc39cecSAdam Litke if (IS_ERR(page)) { 284076dcee75SAneesh Kumar K.V ret = PTR_ERR(page); 284176dcee75SAneesh Kumar K.V if (ret == -ENOMEM) 284276dcee75SAneesh Kumar K.V ret = VM_FAULT_OOM; 284376dcee75SAneesh Kumar K.V else 284476dcee75SAneesh Kumar K.V ret = VM_FAULT_SIGBUS; 28456bda666aSChristoph Lameter goto out; 28466bda666aSChristoph Lameter } 284747ad8475SAndrea Arcangeli clear_huge_page(page, address, pages_per_huge_page(h)); 28480ed361deSNick Piggin __SetPageUptodate(page); 2849ac9b9c66SHugh Dickins 2850f83a275dSMel Gorman if (vma->vm_flags & VM_MAYSHARE) { 28516bda666aSChristoph Lameter int err; 285245c682a6SKen Chen struct inode *inode = mapping->host; 28536bda666aSChristoph Lameter 28546bda666aSChristoph Lameter err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); 28556bda666aSChristoph Lameter if (err) { 28566bda666aSChristoph Lameter put_page(page); 28576bda666aSChristoph Lameter if (err == -EEXIST) 28586bda666aSChristoph Lameter goto retry; 28596bda666aSChristoph Lameter goto out; 28606bda666aSChristoph Lameter } 286107443a85SJoonsoo Kim ClearPagePrivate(page); 286245c682a6SKen Chen 286345c682a6SKen Chen spin_lock(&inode->i_lock); 2864a5516438SAndi Kleen inode->i_blocks += blocks_per_huge_page(h); 286545c682a6SKen Chen spin_unlock(&inode->i_lock); 286623be7468SMel Gorman } else { 28676bda666aSChristoph Lameter lock_page(page); 28680fe6e20bSNaoya Horiguchi if (unlikely(anon_vma_prepare(vma))) { 28690fe6e20bSNaoya Horiguchi ret = VM_FAULT_OOM; 28700fe6e20bSNaoya Horiguchi goto backout_unlocked; 287123be7468SMel Gorman } 2872409eb8c2SHillf Danton anon_rmap = 1; 28730fe6e20bSNaoya Horiguchi } 28740fe6e20bSNaoya Horiguchi } else { 287557303d80SAndy Whitcroft /* 2876998b4382SNaoya Horiguchi * If memory error occurs between mmap() and fault, some process 2877998b4382SNaoya Horiguchi * don't have hwpoisoned swap entry for errored virtual address. 2878998b4382SNaoya Horiguchi * So we need to block hugepage fault by PG_hwpoison bit check. 2879fd6a03edSNaoya Horiguchi */ 2880fd6a03edSNaoya Horiguchi if (unlikely(PageHWPoison(page))) { 2881aa50d3a7SAndi Kleen ret = VM_FAULT_HWPOISON | 2882972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2883fd6a03edSNaoya Horiguchi goto backout_unlocked; 28846bda666aSChristoph Lameter } 2885998b4382SNaoya Horiguchi } 28861e8f889bSDavid Gibson 288757303d80SAndy Whitcroft /* 288857303d80SAndy Whitcroft * If we are going to COW a private mapping later, we examine the 288957303d80SAndy Whitcroft * pending reservations for this page now. This will ensure that 289057303d80SAndy Whitcroft * any allocations necessary to record that reservation occur outside 289157303d80SAndy Whitcroft * the spinlock. 289257303d80SAndy Whitcroft */ 2893788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) 28942b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 28952b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 28962b26736cSAndy Whitcroft goto backout_unlocked; 28972b26736cSAndy Whitcroft } 289857303d80SAndy Whitcroft 2899cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 2900cb900f41SKirill A. Shutemov spin_lock(ptl); 2901a5516438SAndi Kleen size = i_size_read(mapping->host) >> huge_page_shift(h); 29024c887265SAdam Litke if (idx >= size) 29034c887265SAdam Litke goto backout; 29044c887265SAdam Litke 290583c54070SNick Piggin ret = 0; 29067f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) 29074c887265SAdam Litke goto backout; 29084c887265SAdam Litke 290907443a85SJoonsoo Kim if (anon_rmap) { 291007443a85SJoonsoo Kim ClearPagePrivate(page); 2911409eb8c2SHillf Danton hugepage_add_new_anon_rmap(page, vma, address); 2912ac714904SChoi Gi-yong } else 2913409eb8c2SHillf Danton page_dup_rmap(page); 29141e8f889bSDavid Gibson new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) 29151e8f889bSDavid Gibson && (vma->vm_flags & VM_SHARED))); 29161e8f889bSDavid Gibson set_huge_pte_at(mm, address, ptep, new_pte); 29171e8f889bSDavid Gibson 2918788c7df4SHugh Dickins if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 29191e8f889bSDavid Gibson /* Optimization, do the COW without a second fault */ 2920cb900f41SKirill A. Shutemov ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); 29211e8f889bSDavid Gibson } 29221e8f889bSDavid Gibson 2923cb900f41SKirill A. Shutemov spin_unlock(ptl); 29244c887265SAdam Litke unlock_page(page); 29254c887265SAdam Litke out: 2926ac9b9c66SHugh Dickins return ret; 29274c887265SAdam Litke 29284c887265SAdam Litke backout: 2929cb900f41SKirill A. Shutemov spin_unlock(ptl); 29302b26736cSAndy Whitcroft backout_unlocked: 29314c887265SAdam Litke unlock_page(page); 29324c887265SAdam Litke put_page(page); 29334c887265SAdam Litke goto out; 2934ac9b9c66SHugh Dickins } 2935ac9b9c66SHugh Dickins 29368382d914SDavidlohr Bueso #ifdef CONFIG_SMP 29378382d914SDavidlohr Bueso static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 29388382d914SDavidlohr Bueso struct vm_area_struct *vma, 29398382d914SDavidlohr Bueso struct address_space *mapping, 29408382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 29418382d914SDavidlohr Bueso { 29428382d914SDavidlohr Bueso unsigned long key[2]; 29438382d914SDavidlohr Bueso u32 hash; 29448382d914SDavidlohr Bueso 29458382d914SDavidlohr Bueso if (vma->vm_flags & VM_SHARED) { 29468382d914SDavidlohr Bueso key[0] = (unsigned long) mapping; 29478382d914SDavidlohr Bueso key[1] = idx; 29488382d914SDavidlohr Bueso } else { 29498382d914SDavidlohr Bueso key[0] = (unsigned long) mm; 29508382d914SDavidlohr Bueso key[1] = address >> huge_page_shift(h); 29518382d914SDavidlohr Bueso } 29528382d914SDavidlohr Bueso 29538382d914SDavidlohr Bueso hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); 29548382d914SDavidlohr Bueso 29558382d914SDavidlohr Bueso return hash & (num_fault_mutexes - 1); 29568382d914SDavidlohr Bueso } 29578382d914SDavidlohr Bueso #else 29588382d914SDavidlohr Bueso /* 29598382d914SDavidlohr Bueso * For uniprocesor systems we always use a single mutex, so just 29608382d914SDavidlohr Bueso * return 0 and avoid the hashing overhead. 29618382d914SDavidlohr Bueso */ 29628382d914SDavidlohr Bueso static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, 29638382d914SDavidlohr Bueso struct vm_area_struct *vma, 29648382d914SDavidlohr Bueso struct address_space *mapping, 29658382d914SDavidlohr Bueso pgoff_t idx, unsigned long address) 29668382d914SDavidlohr Bueso { 29678382d914SDavidlohr Bueso return 0; 29688382d914SDavidlohr Bueso } 29698382d914SDavidlohr Bueso #endif 29708382d914SDavidlohr Bueso 297186e5216fSAdam Litke int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 2972788c7df4SHugh Dickins unsigned long address, unsigned int flags) 297386e5216fSAdam Litke { 29748382d914SDavidlohr Bueso pte_t *ptep, entry; 2975cb900f41SKirill A. Shutemov spinlock_t *ptl; 29761e8f889bSDavid Gibson int ret; 29778382d914SDavidlohr Bueso u32 hash; 29788382d914SDavidlohr Bueso pgoff_t idx; 29790fe6e20bSNaoya Horiguchi struct page *page = NULL; 298057303d80SAndy Whitcroft struct page *pagecache_page = NULL; 2981a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 29828382d914SDavidlohr Bueso struct address_space *mapping; 298386e5216fSAdam Litke 29841e16a539SKAMEZAWA Hiroyuki address &= huge_page_mask(h); 29851e16a539SKAMEZAWA Hiroyuki 2986fd6a03edSNaoya Horiguchi ptep = huge_pte_offset(mm, address); 2987fd6a03edSNaoya Horiguchi if (ptep) { 2988fd6a03edSNaoya Horiguchi entry = huge_ptep_get(ptep); 2989290408d4SNaoya Horiguchi if (unlikely(is_hugetlb_entry_migration(entry))) { 2990cb900f41SKirill A. Shutemov migration_entry_wait_huge(vma, mm, ptep); 2991290408d4SNaoya Horiguchi return 0; 2992290408d4SNaoya Horiguchi } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) 2993aa50d3a7SAndi Kleen return VM_FAULT_HWPOISON_LARGE | 2994972dc4deSAneesh Kumar K.V VM_FAULT_SET_HINDEX(hstate_index(h)); 2995fd6a03edSNaoya Horiguchi } 2996fd6a03edSNaoya Horiguchi 2997a5516438SAndi Kleen ptep = huge_pte_alloc(mm, address, huge_page_size(h)); 299886e5216fSAdam Litke if (!ptep) 299986e5216fSAdam Litke return VM_FAULT_OOM; 300086e5216fSAdam Litke 30018382d914SDavidlohr Bueso mapping = vma->vm_file->f_mapping; 30028382d914SDavidlohr Bueso idx = vma_hugecache_offset(h, vma, address); 30038382d914SDavidlohr Bueso 30043935baa9SDavid Gibson /* 30053935baa9SDavid Gibson * Serialize hugepage allocation and instantiation, so that we don't 30063935baa9SDavid Gibson * get spurious allocation failures if two CPUs race to instantiate 30073935baa9SDavid Gibson * the same page in the page cache. 30083935baa9SDavid Gibson */ 30098382d914SDavidlohr Bueso hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); 30108382d914SDavidlohr Bueso mutex_lock(&htlb_fault_mutex_table[hash]); 30118382d914SDavidlohr Bueso 30127f2e9525SGerald Schaefer entry = huge_ptep_get(ptep); 30137f2e9525SGerald Schaefer if (huge_pte_none(entry)) { 30148382d914SDavidlohr Bueso ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); 3015b4d1d99fSDavid Gibson goto out_mutex; 30163935baa9SDavid Gibson } 301786e5216fSAdam Litke 301883c54070SNick Piggin ret = 0; 30191e8f889bSDavid Gibson 302057303d80SAndy Whitcroft /* 302157303d80SAndy Whitcroft * If we are going to COW the mapping later, we examine the pending 302257303d80SAndy Whitcroft * reservations for this page now. This will ensure that any 302357303d80SAndy Whitcroft * allocations necessary to record that reservation occur outside the 302457303d80SAndy Whitcroft * spinlock. For private mappings, we also lookup the pagecache 302557303d80SAndy Whitcroft * page now as it is used to determine if a reservation has been 302657303d80SAndy Whitcroft * consumed. 302757303d80SAndy Whitcroft */ 3028106c992aSGerald Schaefer if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { 30292b26736cSAndy Whitcroft if (vma_needs_reservation(h, vma, address) < 0) { 30302b26736cSAndy Whitcroft ret = VM_FAULT_OOM; 3031b4d1d99fSDavid Gibson goto out_mutex; 30322b26736cSAndy Whitcroft } 303357303d80SAndy Whitcroft 3034f83a275dSMel Gorman if (!(vma->vm_flags & VM_MAYSHARE)) 303557303d80SAndy Whitcroft pagecache_page = hugetlbfs_pagecache_page(h, 303657303d80SAndy Whitcroft vma, address); 303757303d80SAndy Whitcroft } 303857303d80SAndy Whitcroft 303956c9cfb1SNaoya Horiguchi /* 304056c9cfb1SNaoya Horiguchi * hugetlb_cow() requires page locks of pte_page(entry) and 304156c9cfb1SNaoya Horiguchi * pagecache_page, so here we need take the former one 304256c9cfb1SNaoya Horiguchi * when page != pagecache_page or !pagecache_page. 304356c9cfb1SNaoya Horiguchi * Note that locking order is always pagecache_page -> page, 304456c9cfb1SNaoya Horiguchi * so no worry about deadlock. 304556c9cfb1SNaoya Horiguchi */ 30460fe6e20bSNaoya Horiguchi page = pte_page(entry); 304766aebce7SChris Metcalf get_page(page); 304856c9cfb1SNaoya Horiguchi if (page != pagecache_page) 30490fe6e20bSNaoya Horiguchi lock_page(page); 30500fe6e20bSNaoya Horiguchi 3051cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(h, mm, ptep); 3052cb900f41SKirill A. Shutemov spin_lock(ptl); 30531e8f889bSDavid Gibson /* Check for a racing update before calling hugetlb_cow */ 3054b4d1d99fSDavid Gibson if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) 3055cb900f41SKirill A. Shutemov goto out_ptl; 3056b4d1d99fSDavid Gibson 3057b4d1d99fSDavid Gibson 3058788c7df4SHugh Dickins if (flags & FAULT_FLAG_WRITE) { 3059106c992aSGerald Schaefer if (!huge_pte_write(entry)) { 306057303d80SAndy Whitcroft ret = hugetlb_cow(mm, vma, address, ptep, entry, 3061cb900f41SKirill A. Shutemov pagecache_page, ptl); 3062cb900f41SKirill A. Shutemov goto out_ptl; 3063b4d1d99fSDavid Gibson } 3064106c992aSGerald Schaefer entry = huge_pte_mkdirty(entry); 3065b4d1d99fSDavid Gibson } 3066b4d1d99fSDavid Gibson entry = pte_mkyoung(entry); 3067788c7df4SHugh Dickins if (huge_ptep_set_access_flags(vma, address, ptep, entry, 3068788c7df4SHugh Dickins flags & FAULT_FLAG_WRITE)) 30694b3073e1SRussell King update_mmu_cache(vma, address, ptep); 3070b4d1d99fSDavid Gibson 3071cb900f41SKirill A. Shutemov out_ptl: 3072cb900f41SKirill A. Shutemov spin_unlock(ptl); 307357303d80SAndy Whitcroft 307457303d80SAndy Whitcroft if (pagecache_page) { 307557303d80SAndy Whitcroft unlock_page(pagecache_page); 307657303d80SAndy Whitcroft put_page(pagecache_page); 307757303d80SAndy Whitcroft } 30781f64d69cSDean Nelson if (page != pagecache_page) 307956c9cfb1SNaoya Horiguchi unlock_page(page); 308066aebce7SChris Metcalf put_page(page); 308157303d80SAndy Whitcroft 3082b4d1d99fSDavid Gibson out_mutex: 30838382d914SDavidlohr Bueso mutex_unlock(&htlb_fault_mutex_table[hash]); 30841e8f889bSDavid Gibson return ret; 308586e5216fSAdam Litke } 308686e5216fSAdam Litke 308728a35716SMichel Lespinasse long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, 308863551ae0SDavid Gibson struct page **pages, struct vm_area_struct **vmas, 308928a35716SMichel Lespinasse unsigned long *position, unsigned long *nr_pages, 309028a35716SMichel Lespinasse long i, unsigned int flags) 309163551ae0SDavid Gibson { 3092d5d4b0aaSChen, Kenneth W unsigned long pfn_offset; 3093d5d4b0aaSChen, Kenneth W unsigned long vaddr = *position; 309428a35716SMichel Lespinasse unsigned long remainder = *nr_pages; 3095a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 309663551ae0SDavid Gibson 309763551ae0SDavid Gibson while (vaddr < vma->vm_end && remainder) { 309863551ae0SDavid Gibson pte_t *pte; 3099cb900f41SKirill A. Shutemov spinlock_t *ptl = NULL; 31002a15efc9SHugh Dickins int absent; 310163551ae0SDavid Gibson struct page *page; 310263551ae0SDavid Gibson 31034c887265SAdam Litke /* 31044c887265SAdam Litke * Some archs (sparc64, sh*) have multiple pte_ts to 31052a15efc9SHugh Dickins * each hugepage. We have to make sure we get the 31064c887265SAdam Litke * first, for the page indexing below to work. 3107cb900f41SKirill A. Shutemov * 3108cb900f41SKirill A. Shutemov * Note that page table lock is not held when pte is null. 31094c887265SAdam Litke */ 3110a5516438SAndi Kleen pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); 3111cb900f41SKirill A. Shutemov if (pte) 3112cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, pte); 31132a15efc9SHugh Dickins absent = !pte || huge_pte_none(huge_ptep_get(pte)); 311463551ae0SDavid Gibson 31152a15efc9SHugh Dickins /* 31162a15efc9SHugh Dickins * When coredumping, it suits get_dump_page if we just return 31173ae77f43SHugh Dickins * an error where there's an empty slot with no huge pagecache 31183ae77f43SHugh Dickins * to back it. This way, we avoid allocating a hugepage, and 31193ae77f43SHugh Dickins * the sparse dumpfile avoids allocating disk blocks, but its 31203ae77f43SHugh Dickins * huge holes still show up with zeroes where they need to be. 31212a15efc9SHugh Dickins */ 31223ae77f43SHugh Dickins if (absent && (flags & FOLL_DUMP) && 31233ae77f43SHugh Dickins !hugetlbfs_pagecache_present(h, vma, vaddr)) { 3124cb900f41SKirill A. Shutemov if (pte) 3125cb900f41SKirill A. Shutemov spin_unlock(ptl); 31262a15efc9SHugh Dickins remainder = 0; 31272a15efc9SHugh Dickins break; 31282a15efc9SHugh Dickins } 31292a15efc9SHugh Dickins 31309cc3a5bdSNaoya Horiguchi /* 31319cc3a5bdSNaoya Horiguchi * We need call hugetlb_fault for both hugepages under migration 31329cc3a5bdSNaoya Horiguchi * (in which case hugetlb_fault waits for the migration,) and 31339cc3a5bdSNaoya Horiguchi * hwpoisoned hugepages (in which case we need to prevent the 31349cc3a5bdSNaoya Horiguchi * caller from accessing to them.) In order to do this, we use 31359cc3a5bdSNaoya Horiguchi * here is_swap_pte instead of is_hugetlb_entry_migration and 31369cc3a5bdSNaoya Horiguchi * is_hugetlb_entry_hwpoisoned. This is because it simply covers 31379cc3a5bdSNaoya Horiguchi * both cases, and because we can't follow correct pages 31389cc3a5bdSNaoya Horiguchi * directly from any kind of swap entries. 31399cc3a5bdSNaoya Horiguchi */ 31409cc3a5bdSNaoya Horiguchi if (absent || is_swap_pte(huge_ptep_get(pte)) || 3141106c992aSGerald Schaefer ((flags & FOLL_WRITE) && 3142106c992aSGerald Schaefer !huge_pte_write(huge_ptep_get(pte)))) { 31434c887265SAdam Litke int ret; 31444c887265SAdam Litke 3145cb900f41SKirill A. Shutemov if (pte) 3146cb900f41SKirill A. Shutemov spin_unlock(ptl); 31472a15efc9SHugh Dickins ret = hugetlb_fault(mm, vma, vaddr, 31482a15efc9SHugh Dickins (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); 3149a89182c7SAdam Litke if (!(ret & VM_FAULT_ERROR)) 31504c887265SAdam Litke continue; 31514c887265SAdam Litke 31521c59827dSHugh Dickins remainder = 0; 31531c59827dSHugh Dickins break; 31541c59827dSHugh Dickins } 315563551ae0SDavid Gibson 3156a5516438SAndi Kleen pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; 31577f2e9525SGerald Schaefer page = pte_page(huge_ptep_get(pte)); 3158d5d4b0aaSChen, Kenneth W same_page: 3159d6692183SChen, Kenneth W if (pages) { 316069d177c2SAndy Whitcroft pages[i] = mem_map_offset(page, pfn_offset); 3161a0368d4eSAndrea Arcangeli get_page_foll(pages[i]); 3162d6692183SChen, Kenneth W } 316363551ae0SDavid Gibson 316463551ae0SDavid Gibson if (vmas) 316563551ae0SDavid Gibson vmas[i] = vma; 316663551ae0SDavid Gibson 316763551ae0SDavid Gibson vaddr += PAGE_SIZE; 3168d5d4b0aaSChen, Kenneth W ++pfn_offset; 316963551ae0SDavid Gibson --remainder; 317063551ae0SDavid Gibson ++i; 3171d5d4b0aaSChen, Kenneth W if (vaddr < vma->vm_end && remainder && 3172a5516438SAndi Kleen pfn_offset < pages_per_huge_page(h)) { 3173d5d4b0aaSChen, Kenneth W /* 3174d5d4b0aaSChen, Kenneth W * We use pfn_offset to avoid touching the pageframes 3175d5d4b0aaSChen, Kenneth W * of this compound page. 3176d5d4b0aaSChen, Kenneth W */ 3177d5d4b0aaSChen, Kenneth W goto same_page; 3178d5d4b0aaSChen, Kenneth W } 3179cb900f41SKirill A. Shutemov spin_unlock(ptl); 318063551ae0SDavid Gibson } 318128a35716SMichel Lespinasse *nr_pages = remainder; 318263551ae0SDavid Gibson *position = vaddr; 318363551ae0SDavid Gibson 31842a15efc9SHugh Dickins return i ? i : -EFAULT; 318563551ae0SDavid Gibson } 31868f860591SZhang, Yanmin 31877da4d641SPeter Zijlstra unsigned long hugetlb_change_protection(struct vm_area_struct *vma, 31888f860591SZhang, Yanmin unsigned long address, unsigned long end, pgprot_t newprot) 31898f860591SZhang, Yanmin { 31908f860591SZhang, Yanmin struct mm_struct *mm = vma->vm_mm; 31918f860591SZhang, Yanmin unsigned long start = address; 31928f860591SZhang, Yanmin pte_t *ptep; 31938f860591SZhang, Yanmin pte_t pte; 3194a5516438SAndi Kleen struct hstate *h = hstate_vma(vma); 31957da4d641SPeter Zijlstra unsigned long pages = 0; 31968f860591SZhang, Yanmin 31978f860591SZhang, Yanmin BUG_ON(address >= end); 31988f860591SZhang, Yanmin flush_cache_range(vma, address, end); 31998f860591SZhang, Yanmin 3200a5338093SRik van Riel mmu_notifier_invalidate_range_start(mm, start, end); 32013d48ae45SPeter Zijlstra mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); 3202a5516438SAndi Kleen for (; address < end; address += huge_page_size(h)) { 3203cb900f41SKirill A. Shutemov spinlock_t *ptl; 32048f860591SZhang, Yanmin ptep = huge_pte_offset(mm, address); 32058f860591SZhang, Yanmin if (!ptep) 32068f860591SZhang, Yanmin continue; 3207cb900f41SKirill A. Shutemov ptl = huge_pte_lock(h, mm, ptep); 32087da4d641SPeter Zijlstra if (huge_pmd_unshare(mm, &address, ptep)) { 32097da4d641SPeter Zijlstra pages++; 3210cb900f41SKirill A. Shutemov spin_unlock(ptl); 321139dde65cSChen, Kenneth W continue; 32127da4d641SPeter Zijlstra } 32137f2e9525SGerald Schaefer if (!huge_pte_none(huge_ptep_get(ptep))) { 32148f860591SZhang, Yanmin pte = huge_ptep_get_and_clear(mm, address, ptep); 3215106c992aSGerald Schaefer pte = pte_mkhuge(huge_pte_modify(pte, newprot)); 3216be7517d6STony Lu pte = arch_make_huge_pte(pte, vma, NULL, 0); 32178f860591SZhang, Yanmin set_huge_pte_at(mm, address, ptep, pte); 32187da4d641SPeter Zijlstra pages++; 32198f860591SZhang, Yanmin } 3220cb900f41SKirill A. Shutemov spin_unlock(ptl); 32218f860591SZhang, Yanmin } 3222d833352aSMel Gorman /* 3223d833352aSMel Gorman * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare 3224d833352aSMel Gorman * may have cleared our pud entry and done put_page on the page table: 3225d833352aSMel Gorman * once we release i_mmap_mutex, another task can do the final put_page 3226d833352aSMel Gorman * and that page table be reused and filled with junk. 3227d833352aSMel Gorman */ 32288f860591SZhang, Yanmin flush_tlb_range(vma, start, end); 3229d833352aSMel Gorman mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); 3230a5338093SRik van Riel mmu_notifier_invalidate_range_end(mm, start, end); 32317da4d641SPeter Zijlstra 32327da4d641SPeter Zijlstra return pages << h->order; 32338f860591SZhang, Yanmin } 32348f860591SZhang, Yanmin 3235a1e78772SMel Gorman int hugetlb_reserve_pages(struct inode *inode, 3236a1e78772SMel Gorman long from, long to, 32375a6fe125SMel Gorman struct vm_area_struct *vma, 3238ca16d140SKOSAKI Motohiro vm_flags_t vm_flags) 3239e4e574b7SAdam Litke { 324017c9d12eSMel Gorman long ret, chg; 3241a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 324290481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 32439119a41eSJoonsoo Kim struct resv_map *resv_map; 3244e4e574b7SAdam Litke 3245a1e78772SMel Gorman /* 324617c9d12eSMel Gorman * Only apply hugepage reservation if asked. At fault time, an 324717c9d12eSMel Gorman * attempt will be made for VM_NORESERVE to allocate a page 324890481622SDavid Gibson * without using reserves 324917c9d12eSMel Gorman */ 3250ca16d140SKOSAKI Motohiro if (vm_flags & VM_NORESERVE) 325117c9d12eSMel Gorman return 0; 325217c9d12eSMel Gorman 325317c9d12eSMel Gorman /* 3254a1e78772SMel Gorman * Shared mappings base their reservation on the number of pages that 3255a1e78772SMel Gorman * are already allocated on behalf of the file. Private mappings need 3256a1e78772SMel Gorman * to reserve the full area even if read-only as mprotect() may be 3257a1e78772SMel Gorman * called to make the mapping read-write. Assume !vma is a shm mapping 3258a1e78772SMel Gorman */ 32599119a41eSJoonsoo Kim if (!vma || vma->vm_flags & VM_MAYSHARE) { 32604e35f483SJoonsoo Kim resv_map = inode_resv_map(inode); 32619119a41eSJoonsoo Kim 32621406ec9bSJoonsoo Kim chg = region_chg(resv_map, from, to); 32639119a41eSJoonsoo Kim 32649119a41eSJoonsoo Kim } else { 32659119a41eSJoonsoo Kim resv_map = resv_map_alloc(); 32665a6fe125SMel Gorman if (!resv_map) 32675a6fe125SMel Gorman return -ENOMEM; 32685a6fe125SMel Gorman 326917c9d12eSMel Gorman chg = to - from; 327017c9d12eSMel Gorman 32715a6fe125SMel Gorman set_vma_resv_map(vma, resv_map); 32725a6fe125SMel Gorman set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 32735a6fe125SMel Gorman } 32745a6fe125SMel Gorman 3275c50ac050SDave Hansen if (chg < 0) { 3276c50ac050SDave Hansen ret = chg; 3277c50ac050SDave Hansen goto out_err; 3278c50ac050SDave Hansen } 327917c9d12eSMel Gorman 328090481622SDavid Gibson /* There must be enough pages in the subpool for the mapping */ 3281c50ac050SDave Hansen if (hugepage_subpool_get_pages(spool, chg)) { 3282c50ac050SDave Hansen ret = -ENOSPC; 3283c50ac050SDave Hansen goto out_err; 3284c50ac050SDave Hansen } 328517c9d12eSMel Gorman 328617c9d12eSMel Gorman /* 328717c9d12eSMel Gorman * Check enough hugepages are available for the reservation. 328890481622SDavid Gibson * Hand the pages back to the subpool if there are not 328917c9d12eSMel Gorman */ 329017c9d12eSMel Gorman ret = hugetlb_acct_memory(h, chg); 329117c9d12eSMel Gorman if (ret < 0) { 329290481622SDavid Gibson hugepage_subpool_put_pages(spool, chg); 3293c50ac050SDave Hansen goto out_err; 329417c9d12eSMel Gorman } 329517c9d12eSMel Gorman 329617c9d12eSMel Gorman /* 329717c9d12eSMel Gorman * Account for the reservations made. Shared mappings record regions 329817c9d12eSMel Gorman * that have reservations as they are shared by multiple VMAs. 329917c9d12eSMel Gorman * When the last VMA disappears, the region map says how much 330017c9d12eSMel Gorman * the reservation was and the page cache tells how much of 330117c9d12eSMel Gorman * the reservation was consumed. Private mappings are per-VMA and 330217c9d12eSMel Gorman * only the consumed reservations are tracked. When the VMA 330317c9d12eSMel Gorman * disappears, the original reservation is the VMA size and the 330417c9d12eSMel Gorman * consumed reservations are stored in the map. Hence, nothing 330517c9d12eSMel Gorman * else has to be done for private mappings here 330617c9d12eSMel Gorman */ 3307f83a275dSMel Gorman if (!vma || vma->vm_flags & VM_MAYSHARE) 33081406ec9bSJoonsoo Kim region_add(resv_map, from, to); 3309a43a8c39SChen, Kenneth W return 0; 3310c50ac050SDave Hansen out_err: 3311f031dd27SJoonsoo Kim if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 3312f031dd27SJoonsoo Kim kref_put(&resv_map->refs, resv_map_release); 3313c50ac050SDave Hansen return ret; 3314a43a8c39SChen, Kenneth W } 3315a43a8c39SChen, Kenneth W 3316a43a8c39SChen, Kenneth W void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) 3317a43a8c39SChen, Kenneth W { 3318a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 33194e35f483SJoonsoo Kim struct resv_map *resv_map = inode_resv_map(inode); 33209119a41eSJoonsoo Kim long chg = 0; 332190481622SDavid Gibson struct hugepage_subpool *spool = subpool_inode(inode); 332245c682a6SKen Chen 33239119a41eSJoonsoo Kim if (resv_map) 33241406ec9bSJoonsoo Kim chg = region_truncate(resv_map, offset); 332545c682a6SKen Chen spin_lock(&inode->i_lock); 3326e4c6f8beSEric Sandeen inode->i_blocks -= (blocks_per_huge_page(h) * freed); 332745c682a6SKen Chen spin_unlock(&inode->i_lock); 332845c682a6SKen Chen 332990481622SDavid Gibson hugepage_subpool_put_pages(spool, (chg - freed)); 3330a5516438SAndi Kleen hugetlb_acct_memory(h, -(chg - freed)); 3331a43a8c39SChen, Kenneth W } 333293f70f90SNaoya Horiguchi 33333212b535SSteve Capper #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE 33343212b535SSteve Capper static unsigned long page_table_shareable(struct vm_area_struct *svma, 33353212b535SSteve Capper struct vm_area_struct *vma, 33363212b535SSteve Capper unsigned long addr, pgoff_t idx) 33373212b535SSteve Capper { 33383212b535SSteve Capper unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 33393212b535SSteve Capper svma->vm_start; 33403212b535SSteve Capper unsigned long sbase = saddr & PUD_MASK; 33413212b535SSteve Capper unsigned long s_end = sbase + PUD_SIZE; 33423212b535SSteve Capper 33433212b535SSteve Capper /* Allow segments to share if only one is marked locked */ 33443212b535SSteve Capper unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; 33453212b535SSteve Capper unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; 33463212b535SSteve Capper 33473212b535SSteve Capper /* 33483212b535SSteve Capper * match the virtual addresses, permission and the alignment of the 33493212b535SSteve Capper * page table page. 33503212b535SSteve Capper */ 33513212b535SSteve Capper if (pmd_index(addr) != pmd_index(saddr) || 33523212b535SSteve Capper vm_flags != svm_flags || 33533212b535SSteve Capper sbase < svma->vm_start || svma->vm_end < s_end) 33543212b535SSteve Capper return 0; 33553212b535SSteve Capper 33563212b535SSteve Capper return saddr; 33573212b535SSteve Capper } 33583212b535SSteve Capper 33593212b535SSteve Capper static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) 33603212b535SSteve Capper { 33613212b535SSteve Capper unsigned long base = addr & PUD_MASK; 33623212b535SSteve Capper unsigned long end = base + PUD_SIZE; 33633212b535SSteve Capper 33643212b535SSteve Capper /* 33653212b535SSteve Capper * check on proper vm_flags and page table alignment 33663212b535SSteve Capper */ 33673212b535SSteve Capper if (vma->vm_flags & VM_MAYSHARE && 33683212b535SSteve Capper vma->vm_start <= base && end <= vma->vm_end) 33693212b535SSteve Capper return 1; 33703212b535SSteve Capper return 0; 33713212b535SSteve Capper } 33723212b535SSteve Capper 33733212b535SSteve Capper /* 33743212b535SSteve Capper * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 33753212b535SSteve Capper * and returns the corresponding pte. While this is not necessary for the 33763212b535SSteve Capper * !shared pmd case because we can allocate the pmd later as well, it makes the 33773212b535SSteve Capper * code much cleaner. pmd allocation is essential for the shared case because 33783212b535SSteve Capper * pud has to be populated inside the same i_mmap_mutex section - otherwise 33793212b535SSteve Capper * racing tasks could either miss the sharing (see huge_pte_offset) or select a 33803212b535SSteve Capper * bad pmd for sharing. 33813212b535SSteve Capper */ 33823212b535SSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 33833212b535SSteve Capper { 33843212b535SSteve Capper struct vm_area_struct *vma = find_vma(mm, addr); 33853212b535SSteve Capper struct address_space *mapping = vma->vm_file->f_mapping; 33863212b535SSteve Capper pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 33873212b535SSteve Capper vma->vm_pgoff; 33883212b535SSteve Capper struct vm_area_struct *svma; 33893212b535SSteve Capper unsigned long saddr; 33903212b535SSteve Capper pte_t *spte = NULL; 33913212b535SSteve Capper pte_t *pte; 3392cb900f41SKirill A. Shutemov spinlock_t *ptl; 33933212b535SSteve Capper 33943212b535SSteve Capper if (!vma_shareable(vma, addr)) 33953212b535SSteve Capper return (pte_t *)pmd_alloc(mm, pud, addr); 33963212b535SSteve Capper 33973212b535SSteve Capper mutex_lock(&mapping->i_mmap_mutex); 33983212b535SSteve Capper vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 33993212b535SSteve Capper if (svma == vma) 34003212b535SSteve Capper continue; 34013212b535SSteve Capper 34023212b535SSteve Capper saddr = page_table_shareable(svma, vma, addr, idx); 34033212b535SSteve Capper if (saddr) { 34043212b535SSteve Capper spte = huge_pte_offset(svma->vm_mm, saddr); 34053212b535SSteve Capper if (spte) { 34063212b535SSteve Capper get_page(virt_to_page(spte)); 34073212b535SSteve Capper break; 34083212b535SSteve Capper } 34093212b535SSteve Capper } 34103212b535SSteve Capper } 34113212b535SSteve Capper 34123212b535SSteve Capper if (!spte) 34133212b535SSteve Capper goto out; 34143212b535SSteve Capper 3415cb900f41SKirill A. Shutemov ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); 3416cb900f41SKirill A. Shutemov spin_lock(ptl); 34173212b535SSteve Capper if (pud_none(*pud)) 34183212b535SSteve Capper pud_populate(mm, pud, 34193212b535SSteve Capper (pmd_t *)((unsigned long)spte & PAGE_MASK)); 34203212b535SSteve Capper else 34213212b535SSteve Capper put_page(virt_to_page(spte)); 3422cb900f41SKirill A. Shutemov spin_unlock(ptl); 34233212b535SSteve Capper out: 34243212b535SSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 34253212b535SSteve Capper mutex_unlock(&mapping->i_mmap_mutex); 34263212b535SSteve Capper return pte; 34273212b535SSteve Capper } 34283212b535SSteve Capper 34293212b535SSteve Capper /* 34303212b535SSteve Capper * unmap huge page backed by shared pte. 34313212b535SSteve Capper * 34323212b535SSteve Capper * Hugetlb pte page is ref counted at the time of mapping. If pte is shared 34333212b535SSteve Capper * indicated by page_count > 1, unmap is achieved by clearing pud and 34343212b535SSteve Capper * decrementing the ref count. If count == 1, the pte page is not shared. 34353212b535SSteve Capper * 3436cb900f41SKirill A. Shutemov * called with page table lock held. 34373212b535SSteve Capper * 34383212b535SSteve Capper * returns: 1 successfully unmapped a shared pte page 34393212b535SSteve Capper * 0 the underlying pte page is not shared, or it is the last user 34403212b535SSteve Capper */ 34413212b535SSteve Capper int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) 34423212b535SSteve Capper { 34433212b535SSteve Capper pgd_t *pgd = pgd_offset(mm, *addr); 34443212b535SSteve Capper pud_t *pud = pud_offset(pgd, *addr); 34453212b535SSteve Capper 34463212b535SSteve Capper BUG_ON(page_count(virt_to_page(ptep)) == 0); 34473212b535SSteve Capper if (page_count(virt_to_page(ptep)) == 1) 34483212b535SSteve Capper return 0; 34493212b535SSteve Capper 34503212b535SSteve Capper pud_clear(pud); 34513212b535SSteve Capper put_page(virt_to_page(ptep)); 34523212b535SSteve Capper *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; 34533212b535SSteve Capper return 1; 34543212b535SSteve Capper } 34559e5fc74cSSteve Capper #define want_pmd_share() (1) 34569e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 34579e5fc74cSSteve Capper pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) 34589e5fc74cSSteve Capper { 34599e5fc74cSSteve Capper return NULL; 34609e5fc74cSSteve Capper } 34619e5fc74cSSteve Capper #define want_pmd_share() (0) 34623212b535SSteve Capper #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ 34633212b535SSteve Capper 34649e5fc74cSSteve Capper #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 34659e5fc74cSSteve Capper pte_t *huge_pte_alloc(struct mm_struct *mm, 34669e5fc74cSSteve Capper unsigned long addr, unsigned long sz) 34679e5fc74cSSteve Capper { 34689e5fc74cSSteve Capper pgd_t *pgd; 34699e5fc74cSSteve Capper pud_t *pud; 34709e5fc74cSSteve Capper pte_t *pte = NULL; 34719e5fc74cSSteve Capper 34729e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 34739e5fc74cSSteve Capper pud = pud_alloc(mm, pgd, addr); 34749e5fc74cSSteve Capper if (pud) { 34759e5fc74cSSteve Capper if (sz == PUD_SIZE) { 34769e5fc74cSSteve Capper pte = (pte_t *)pud; 34779e5fc74cSSteve Capper } else { 34789e5fc74cSSteve Capper BUG_ON(sz != PMD_SIZE); 34799e5fc74cSSteve Capper if (want_pmd_share() && pud_none(*pud)) 34809e5fc74cSSteve Capper pte = huge_pmd_share(mm, addr, pud); 34819e5fc74cSSteve Capper else 34829e5fc74cSSteve Capper pte = (pte_t *)pmd_alloc(mm, pud, addr); 34839e5fc74cSSteve Capper } 34849e5fc74cSSteve Capper } 34859e5fc74cSSteve Capper BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); 34869e5fc74cSSteve Capper 34879e5fc74cSSteve Capper return pte; 34889e5fc74cSSteve Capper } 34899e5fc74cSSteve Capper 34909e5fc74cSSteve Capper pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 34919e5fc74cSSteve Capper { 34929e5fc74cSSteve Capper pgd_t *pgd; 34939e5fc74cSSteve Capper pud_t *pud; 34949e5fc74cSSteve Capper pmd_t *pmd = NULL; 34959e5fc74cSSteve Capper 34969e5fc74cSSteve Capper pgd = pgd_offset(mm, addr); 34979e5fc74cSSteve Capper if (pgd_present(*pgd)) { 34989e5fc74cSSteve Capper pud = pud_offset(pgd, addr); 34999e5fc74cSSteve Capper if (pud_present(*pud)) { 35009e5fc74cSSteve Capper if (pud_huge(*pud)) 35019e5fc74cSSteve Capper return (pte_t *)pud; 35029e5fc74cSSteve Capper pmd = pmd_offset(pud, addr); 35039e5fc74cSSteve Capper } 35049e5fc74cSSteve Capper } 35059e5fc74cSSteve Capper return (pte_t *) pmd; 35069e5fc74cSSteve Capper } 35079e5fc74cSSteve Capper 35089e5fc74cSSteve Capper struct page * 35099e5fc74cSSteve Capper follow_huge_pmd(struct mm_struct *mm, unsigned long address, 35109e5fc74cSSteve Capper pmd_t *pmd, int write) 35119e5fc74cSSteve Capper { 35129e5fc74cSSteve Capper struct page *page; 35139e5fc74cSSteve Capper 35149e5fc74cSSteve Capper page = pte_page(*(pte_t *)pmd); 35159e5fc74cSSteve Capper if (page) 35169e5fc74cSSteve Capper page += ((address & ~PMD_MASK) >> PAGE_SHIFT); 35179e5fc74cSSteve Capper return page; 35189e5fc74cSSteve Capper } 35199e5fc74cSSteve Capper 35209e5fc74cSSteve Capper struct page * 35219e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 35229e5fc74cSSteve Capper pud_t *pud, int write) 35239e5fc74cSSteve Capper { 35249e5fc74cSSteve Capper struct page *page; 35259e5fc74cSSteve Capper 35269e5fc74cSSteve Capper page = pte_page(*(pte_t *)pud); 35279e5fc74cSSteve Capper if (page) 35289e5fc74cSSteve Capper page += ((address & ~PUD_MASK) >> PAGE_SHIFT); 35299e5fc74cSSteve Capper return page; 35309e5fc74cSSteve Capper } 35319e5fc74cSSteve Capper 35329e5fc74cSSteve Capper #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 35339e5fc74cSSteve Capper 35349e5fc74cSSteve Capper /* Can be overriden by architectures */ 35353b32123dSGideon Israel Dsouza struct page * __weak 35369e5fc74cSSteve Capper follow_huge_pud(struct mm_struct *mm, unsigned long address, 35379e5fc74cSSteve Capper pud_t *pud, int write) 35389e5fc74cSSteve Capper { 35399e5fc74cSSteve Capper BUG(); 35409e5fc74cSSteve Capper return NULL; 35419e5fc74cSSteve Capper } 35429e5fc74cSSteve Capper 35439e5fc74cSSteve Capper #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 35449e5fc74cSSteve Capper 3545d5bd9106SAndi Kleen #ifdef CONFIG_MEMORY_FAILURE 3546d5bd9106SAndi Kleen 35476de2b1aaSNaoya Horiguchi /* Should be called in hugetlb_lock */ 35486de2b1aaSNaoya Horiguchi static int is_hugepage_on_freelist(struct page *hpage) 35496de2b1aaSNaoya Horiguchi { 35506de2b1aaSNaoya Horiguchi struct page *page; 35516de2b1aaSNaoya Horiguchi struct page *tmp; 35526de2b1aaSNaoya Horiguchi struct hstate *h = page_hstate(hpage); 35536de2b1aaSNaoya Horiguchi int nid = page_to_nid(hpage); 35546de2b1aaSNaoya Horiguchi 35556de2b1aaSNaoya Horiguchi list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) 35566de2b1aaSNaoya Horiguchi if (page == hpage) 35576de2b1aaSNaoya Horiguchi return 1; 35586de2b1aaSNaoya Horiguchi return 0; 35596de2b1aaSNaoya Horiguchi } 35606de2b1aaSNaoya Horiguchi 356193f70f90SNaoya Horiguchi /* 356293f70f90SNaoya Horiguchi * This function is called from memory failure code. 356393f70f90SNaoya Horiguchi * Assume the caller holds page lock of the head page. 356493f70f90SNaoya Horiguchi */ 35656de2b1aaSNaoya Horiguchi int dequeue_hwpoisoned_huge_page(struct page *hpage) 356693f70f90SNaoya Horiguchi { 356793f70f90SNaoya Horiguchi struct hstate *h = page_hstate(hpage); 356893f70f90SNaoya Horiguchi int nid = page_to_nid(hpage); 35696de2b1aaSNaoya Horiguchi int ret = -EBUSY; 357093f70f90SNaoya Horiguchi 357193f70f90SNaoya Horiguchi spin_lock(&hugetlb_lock); 35726de2b1aaSNaoya Horiguchi if (is_hugepage_on_freelist(hpage)) { 357356f2fb14SNaoya Horiguchi /* 357456f2fb14SNaoya Horiguchi * Hwpoisoned hugepage isn't linked to activelist or freelist, 357556f2fb14SNaoya Horiguchi * but dangling hpage->lru can trigger list-debug warnings 357656f2fb14SNaoya Horiguchi * (this happens when we call unpoison_memory() on it), 357756f2fb14SNaoya Horiguchi * so let it point to itself with list_del_init(). 357856f2fb14SNaoya Horiguchi */ 357956f2fb14SNaoya Horiguchi list_del_init(&hpage->lru); 35808c6c2ecbSNaoya Horiguchi set_page_refcounted(hpage); 358193f70f90SNaoya Horiguchi h->free_huge_pages--; 358293f70f90SNaoya Horiguchi h->free_huge_pages_node[nid]--; 35836de2b1aaSNaoya Horiguchi ret = 0; 358493f70f90SNaoya Horiguchi } 35856de2b1aaSNaoya Horiguchi spin_unlock(&hugetlb_lock); 35866de2b1aaSNaoya Horiguchi return ret; 35876de2b1aaSNaoya Horiguchi } 35886de2b1aaSNaoya Horiguchi #endif 358931caf665SNaoya Horiguchi 359031caf665SNaoya Horiguchi bool isolate_huge_page(struct page *page, struct list_head *list) 359131caf665SNaoya Horiguchi { 3592309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 359331caf665SNaoya Horiguchi if (!get_page_unless_zero(page)) 359431caf665SNaoya Horiguchi return false; 359531caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 359631caf665SNaoya Horiguchi list_move_tail(&page->lru, list); 359731caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 359831caf665SNaoya Horiguchi return true; 359931caf665SNaoya Horiguchi } 360031caf665SNaoya Horiguchi 360131caf665SNaoya Horiguchi void putback_active_hugepage(struct page *page) 360231caf665SNaoya Horiguchi { 3603309381feSSasha Levin VM_BUG_ON_PAGE(!PageHead(page), page); 360431caf665SNaoya Horiguchi spin_lock(&hugetlb_lock); 360531caf665SNaoya Horiguchi list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); 360631caf665SNaoya Horiguchi spin_unlock(&hugetlb_lock); 360731caf665SNaoya Horiguchi put_page(page); 360831caf665SNaoya Horiguchi } 3609c8721bbbSNaoya Horiguchi 3610c8721bbbSNaoya Horiguchi bool is_hugepage_active(struct page *page) 3611c8721bbbSNaoya Horiguchi { 3612309381feSSasha Levin VM_BUG_ON_PAGE(!PageHuge(page), page); 3613c8721bbbSNaoya Horiguchi /* 3614c8721bbbSNaoya Horiguchi * This function can be called for a tail page because the caller, 3615c8721bbbSNaoya Horiguchi * scan_movable_pages, scans through a given pfn-range which typically 3616c8721bbbSNaoya Horiguchi * covers one memory block. In systems using gigantic hugepage (1GB 3617c8721bbbSNaoya Horiguchi * for x86_64,) a hugepage is larger than a memory block, and we don't 3618c8721bbbSNaoya Horiguchi * support migrating such large hugepages for now, so return false 3619c8721bbbSNaoya Horiguchi * when called for tail pages. 3620c8721bbbSNaoya Horiguchi */ 3621c8721bbbSNaoya Horiguchi if (PageTail(page)) 3622c8721bbbSNaoya Horiguchi return false; 3623c8721bbbSNaoya Horiguchi /* 3624c8721bbbSNaoya Horiguchi * Refcount of a hwpoisoned hugepages is 1, but they are not active, 3625c8721bbbSNaoya Horiguchi * so we should return false for them. 3626c8721bbbSNaoya Horiguchi */ 3627c8721bbbSNaoya Horiguchi if (unlikely(PageHWPoison(page))) 3628c8721bbbSNaoya Horiguchi return false; 3629c8721bbbSNaoya Horiguchi return page_count(page) > 0; 3630c8721bbbSNaoya Horiguchi } 3631