11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * mm/rmap.c - physical to virtual reverse mappings 31da177e4SLinus Torvalds * 41da177e4SLinus Torvalds * Copyright 2001, Rik van Riel <riel@conectiva.com.br> 51da177e4SLinus Torvalds * Released under the General Public License (GPL). 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Simple, low overhead reverse mapping scheme. 81da177e4SLinus Torvalds * Please try to keep this thing as modular as possible. 91da177e4SLinus Torvalds * 101da177e4SLinus Torvalds * Provides methods for unmapping each kind of mapped page: 111da177e4SLinus Torvalds * the anon methods track anonymous pages, and 121da177e4SLinus Torvalds * the file methods track pages belonging to an inode. 131da177e4SLinus Torvalds * 141da177e4SLinus Torvalds * Original design by Rik van Riel <riel@conectiva.com.br> 2001 151da177e4SLinus Torvalds * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 161da177e4SLinus Torvalds * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 1798f32602SHugh Dickins * Contributions by Hugh Dickins 2003, 2004 181da177e4SLinus Torvalds */ 191da177e4SLinus Torvalds 201da177e4SLinus Torvalds /* 211da177e4SLinus Torvalds * Lock ordering in mm: 221da177e4SLinus Torvalds * 231b1dcc1bSJes Sorensen * inode->i_mutex (while writing or truncating, not reading or faulting) 24c1e8d7c6SMichel Lespinasse * mm->mmap_lock 25c0d0381aSMike Kravetz * page->flags PG_locked (lock_page) * (see huegtlbfs below) 2688f306b6SKirill A. Shutemov * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share) 27c8c06efaSDavidlohr Bueso * mapping->i_mmap_rwsem 28c0d0381aSMike Kravetz * hugetlb_fault_mutex (hugetlbfs specific page fault mutex) 295a505085SIngo Molnar * anon_vma->rwsem 30b8072f09SHugh Dickins * mm->page_table_lock or pte_lock 315d337b91SHugh Dickins * swap_lock (in swap_duplicate, swap_info_get) 321da177e4SLinus Torvalds * mmlist_lock (in mmput, drain_mmlist and others) 331da177e4SLinus Torvalds * mapping->private_lock (in __set_page_dirty_buffers) 3415b44736SHugh Dickins * lock_page_memcg move_lock (in __set_page_dirty_buffers) 35b93b0163SMatthew Wilcox * i_pages lock (widely used) 3615b44736SHugh Dickins * lruvec->lru_lock (in lock_page_lruvec_irq) 37250df6edSDave Chinner * inode->i_lock (in set_page_dirty's __mark_inode_dirty) 38f758eeabSChristoph Hellwig * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) 391da177e4SLinus Torvalds * sb_lock (within inode_lock in fs/fs-writeback.c) 40b93b0163SMatthew Wilcox * i_pages lock (widely used, in set_page_dirty, 411da177e4SLinus Torvalds * in arch-dependent flush_dcache_mmap_lock, 42f758eeabSChristoph Hellwig * within bdi.wb->list_lock in __sync_single_inode) 436a46079cSAndi Kleen * 445a505085SIngo Molnar * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) 456a46079cSAndi Kleen * ->tasklist_lock 466a46079cSAndi Kleen * pte map lock 47c0d0381aSMike Kravetz * 48c0d0381aSMike Kravetz * * hugetlbfs PageHuge() pages take locks in this order: 49c0d0381aSMike Kravetz * mapping->i_mmap_rwsem 50c0d0381aSMike Kravetz * hugetlb_fault_mutex (hugetlbfs specific page fault mutex) 51c0d0381aSMike Kravetz * page->flags PG_locked (lock_page) 521da177e4SLinus Torvalds */ 531da177e4SLinus Torvalds 541da177e4SLinus Torvalds #include <linux/mm.h> 556e84f315SIngo Molnar #include <linux/sched/mm.h> 5629930025SIngo Molnar #include <linux/sched/task.h> 571da177e4SLinus Torvalds #include <linux/pagemap.h> 581da177e4SLinus Torvalds #include <linux/swap.h> 591da177e4SLinus Torvalds #include <linux/swapops.h> 601da177e4SLinus Torvalds #include <linux/slab.h> 611da177e4SLinus Torvalds #include <linux/init.h> 625ad64688SHugh Dickins #include <linux/ksm.h> 631da177e4SLinus Torvalds #include <linux/rmap.h> 641da177e4SLinus Torvalds #include <linux/rcupdate.h> 65b95f1b31SPaul Gortmaker #include <linux/export.h> 668a9f3ccdSBalbir Singh #include <linux/memcontrol.h> 67cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 6864cdd548SKOSAKI Motohiro #include <linux/migrate.h> 690fe6e20bSNaoya Horiguchi #include <linux/hugetlb.h> 70444f84fdSBen Dooks #include <linux/huge_mm.h> 71ef5d437fSJan Kara #include <linux/backing-dev.h> 7233c3fc71SVladimir Davydov #include <linux/page_idle.h> 73a5430ddaSJérôme Glisse #include <linux/memremap.h> 74bce73e48SChristian Borntraeger #include <linux/userfaultfd_k.h> 751da177e4SLinus Torvalds 761da177e4SLinus Torvalds #include <asm/tlbflush.h> 771da177e4SLinus Torvalds 7872b252aeSMel Gorman #include <trace/events/tlb.h> 7972b252aeSMel Gorman 80b291f000SNick Piggin #include "internal.h" 81b291f000SNick Piggin 82fdd2e5f8SAdrian Bunk static struct kmem_cache *anon_vma_cachep; 835beb4930SRik van Riel static struct kmem_cache *anon_vma_chain_cachep; 84fdd2e5f8SAdrian Bunk 85fdd2e5f8SAdrian Bunk static inline struct anon_vma *anon_vma_alloc(void) 86fdd2e5f8SAdrian Bunk { 8701d8b20dSPeter Zijlstra struct anon_vma *anon_vma; 8801d8b20dSPeter Zijlstra 8901d8b20dSPeter Zijlstra anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); 9001d8b20dSPeter Zijlstra if (anon_vma) { 9101d8b20dSPeter Zijlstra atomic_set(&anon_vma->refcount, 1); 927a3ef208SKonstantin Khlebnikov anon_vma->degree = 1; /* Reference for first vma */ 937a3ef208SKonstantin Khlebnikov anon_vma->parent = anon_vma; 9401d8b20dSPeter Zijlstra /* 9501d8b20dSPeter Zijlstra * Initialise the anon_vma root to point to itself. If called 9601d8b20dSPeter Zijlstra * from fork, the root will be reset to the parents anon_vma. 9701d8b20dSPeter Zijlstra */ 9801d8b20dSPeter Zijlstra anon_vma->root = anon_vma; 99fdd2e5f8SAdrian Bunk } 100fdd2e5f8SAdrian Bunk 10101d8b20dSPeter Zijlstra return anon_vma; 10201d8b20dSPeter Zijlstra } 10301d8b20dSPeter Zijlstra 10401d8b20dSPeter Zijlstra static inline void anon_vma_free(struct anon_vma *anon_vma) 105fdd2e5f8SAdrian Bunk { 10601d8b20dSPeter Zijlstra VM_BUG_ON(atomic_read(&anon_vma->refcount)); 10788c22088SPeter Zijlstra 10888c22088SPeter Zijlstra /* 1094fc3f1d6SIngo Molnar * Synchronize against page_lock_anon_vma_read() such that 11088c22088SPeter Zijlstra * we can safely hold the lock without the anon_vma getting 11188c22088SPeter Zijlstra * freed. 11288c22088SPeter Zijlstra * 11388c22088SPeter Zijlstra * Relies on the full mb implied by the atomic_dec_and_test() from 11488c22088SPeter Zijlstra * put_anon_vma() against the acquire barrier implied by 1154fc3f1d6SIngo Molnar * down_read_trylock() from page_lock_anon_vma_read(). This orders: 11688c22088SPeter Zijlstra * 1174fc3f1d6SIngo Molnar * page_lock_anon_vma_read() VS put_anon_vma() 1184fc3f1d6SIngo Molnar * down_read_trylock() atomic_dec_and_test() 11988c22088SPeter Zijlstra * LOCK MB 1204fc3f1d6SIngo Molnar * atomic_read() rwsem_is_locked() 12188c22088SPeter Zijlstra * 12288c22088SPeter Zijlstra * LOCK should suffice since the actual taking of the lock must 12388c22088SPeter Zijlstra * happen _before_ what follows. 12488c22088SPeter Zijlstra */ 1257f39dda9SHugh Dickins might_sleep(); 1265a505085SIngo Molnar if (rwsem_is_locked(&anon_vma->root->rwsem)) { 1274fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 12808b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 12988c22088SPeter Zijlstra } 13088c22088SPeter Zijlstra 131fdd2e5f8SAdrian Bunk kmem_cache_free(anon_vma_cachep, anon_vma); 132fdd2e5f8SAdrian Bunk } 1331da177e4SLinus Torvalds 134dd34739cSLinus Torvalds static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) 1355beb4930SRik van Riel { 136dd34739cSLinus Torvalds return kmem_cache_alloc(anon_vma_chain_cachep, gfp); 1375beb4930SRik van Riel } 1385beb4930SRik van Riel 139e574b5fdSNamhyung Kim static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) 1405beb4930SRik van Riel { 1415beb4930SRik van Riel kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); 1425beb4930SRik van Riel } 1435beb4930SRik van Riel 1446583a843SKautuk Consul static void anon_vma_chain_link(struct vm_area_struct *vma, 1456583a843SKautuk Consul struct anon_vma_chain *avc, 1466583a843SKautuk Consul struct anon_vma *anon_vma) 1476583a843SKautuk Consul { 1486583a843SKautuk Consul avc->vma = vma; 1496583a843SKautuk Consul avc->anon_vma = anon_vma; 1506583a843SKautuk Consul list_add(&avc->same_vma, &vma->anon_vma_chain); 151bf181b9fSMichel Lespinasse anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); 1526583a843SKautuk Consul } 1536583a843SKautuk Consul 154d9d332e0SLinus Torvalds /** 155d5a187daSVlastimil Babka * __anon_vma_prepare - attach an anon_vma to a memory region 156d9d332e0SLinus Torvalds * @vma: the memory region in question 157d9d332e0SLinus Torvalds * 158d9d332e0SLinus Torvalds * This makes sure the memory mapping described by 'vma' has 159d9d332e0SLinus Torvalds * an 'anon_vma' attached to it, so that we can associate the 160d9d332e0SLinus Torvalds * anonymous pages mapped into it with that anon_vma. 161d9d332e0SLinus Torvalds * 162d5a187daSVlastimil Babka * The common case will be that we already have one, which 163d5a187daSVlastimil Babka * is handled inline by anon_vma_prepare(). But if 16423a0790aSFigo.zhang * not we either need to find an adjacent mapping that we 165d9d332e0SLinus Torvalds * can re-use the anon_vma from (very common when the only 166d9d332e0SLinus Torvalds * reason for splitting a vma has been mprotect()), or we 167d9d332e0SLinus Torvalds * allocate a new one. 168d9d332e0SLinus Torvalds * 169d9d332e0SLinus Torvalds * Anon-vma allocations are very subtle, because we may have 1704fc3f1d6SIngo Molnar * optimistically looked up an anon_vma in page_lock_anon_vma_read() 171aaf1f990SMiaohe Lin * and that may actually touch the rwsem even in the newly 172d9d332e0SLinus Torvalds * allocated vma (it depends on RCU to make sure that the 173d9d332e0SLinus Torvalds * anon_vma isn't actually destroyed). 174d9d332e0SLinus Torvalds * 175d9d332e0SLinus Torvalds * As a result, we need to do proper anon_vma locking even 176d9d332e0SLinus Torvalds * for the new allocation. At the same time, we do not want 177d9d332e0SLinus Torvalds * to do any locking for the common case of already having 178d9d332e0SLinus Torvalds * an anon_vma. 179d9d332e0SLinus Torvalds * 180c1e8d7c6SMichel Lespinasse * This must be called with the mmap_lock held for reading. 181d9d332e0SLinus Torvalds */ 182d5a187daSVlastimil Babka int __anon_vma_prepare(struct vm_area_struct *vma) 1831da177e4SLinus Torvalds { 184d5a187daSVlastimil Babka struct mm_struct *mm = vma->vm_mm; 185d5a187daSVlastimil Babka struct anon_vma *anon_vma, *allocated; 1865beb4930SRik van Riel struct anon_vma_chain *avc; 1871da177e4SLinus Torvalds 1881da177e4SLinus Torvalds might_sleep(); 1891da177e4SLinus Torvalds 190dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 1915beb4930SRik van Riel if (!avc) 1925beb4930SRik van Riel goto out_enomem; 1935beb4930SRik van Riel 1941da177e4SLinus Torvalds anon_vma = find_mergeable_anon_vma(vma); 1951da177e4SLinus Torvalds allocated = NULL; 196d9d332e0SLinus Torvalds if (!anon_vma) { 1971da177e4SLinus Torvalds anon_vma = anon_vma_alloc(); 1981da177e4SLinus Torvalds if (unlikely(!anon_vma)) 1995beb4930SRik van Riel goto out_enomem_free_avc; 2001da177e4SLinus Torvalds allocated = anon_vma; 2011da177e4SLinus Torvalds } 2021da177e4SLinus Torvalds 2034fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 2041da177e4SLinus Torvalds /* page_table_lock to protect against threads */ 2051da177e4SLinus Torvalds spin_lock(&mm->page_table_lock); 2061da177e4SLinus Torvalds if (likely(!vma->anon_vma)) { 2071da177e4SLinus Torvalds vma->anon_vma = anon_vma; 2086583a843SKautuk Consul anon_vma_chain_link(vma, avc, anon_vma); 2097a3ef208SKonstantin Khlebnikov /* vma reference or self-parent link for new root */ 2107a3ef208SKonstantin Khlebnikov anon_vma->degree++; 2111da177e4SLinus Torvalds allocated = NULL; 21231f2b0ebSOleg Nesterov avc = NULL; 2131da177e4SLinus Torvalds } 2141da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 21508b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 21631f2b0ebSOleg Nesterov 21731f2b0ebSOleg Nesterov if (unlikely(allocated)) 21801d8b20dSPeter Zijlstra put_anon_vma(allocated); 21931f2b0ebSOleg Nesterov if (unlikely(avc)) 2205beb4930SRik van Riel anon_vma_chain_free(avc); 221d5a187daSVlastimil Babka 2221da177e4SLinus Torvalds return 0; 2235beb4930SRik van Riel 2245beb4930SRik van Riel out_enomem_free_avc: 2255beb4930SRik van Riel anon_vma_chain_free(avc); 2265beb4930SRik van Riel out_enomem: 2275beb4930SRik van Riel return -ENOMEM; 2281da177e4SLinus Torvalds } 2291da177e4SLinus Torvalds 230bb4aa396SLinus Torvalds /* 231bb4aa396SLinus Torvalds * This is a useful helper function for locking the anon_vma root as 232bb4aa396SLinus Torvalds * we traverse the vma->anon_vma_chain, looping over anon_vma's that 233bb4aa396SLinus Torvalds * have the same vma. 234bb4aa396SLinus Torvalds * 235bb4aa396SLinus Torvalds * Such anon_vma's should have the same root, so you'd expect to see 236bb4aa396SLinus Torvalds * just a single mutex_lock for the whole traversal. 237bb4aa396SLinus Torvalds */ 238bb4aa396SLinus Torvalds static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) 239bb4aa396SLinus Torvalds { 240bb4aa396SLinus Torvalds struct anon_vma *new_root = anon_vma->root; 241bb4aa396SLinus Torvalds if (new_root != root) { 242bb4aa396SLinus Torvalds if (WARN_ON_ONCE(root)) 2435a505085SIngo Molnar up_write(&root->rwsem); 244bb4aa396SLinus Torvalds root = new_root; 2455a505085SIngo Molnar down_write(&root->rwsem); 246bb4aa396SLinus Torvalds } 247bb4aa396SLinus Torvalds return root; 248bb4aa396SLinus Torvalds } 249bb4aa396SLinus Torvalds 250bb4aa396SLinus Torvalds static inline void unlock_anon_vma_root(struct anon_vma *root) 251bb4aa396SLinus Torvalds { 252bb4aa396SLinus Torvalds if (root) 2535a505085SIngo Molnar up_write(&root->rwsem); 254bb4aa396SLinus Torvalds } 255bb4aa396SLinus Torvalds 2565beb4930SRik van Riel /* 2575beb4930SRik van Riel * Attach the anon_vmas from src to dst. 2585beb4930SRik van Riel * Returns 0 on success, -ENOMEM on failure. 2597a3ef208SKonstantin Khlebnikov * 26047b390d2SWei Yang * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and 26147b390d2SWei Yang * anon_vma_fork(). The first three want an exact copy of src, while the last 26247b390d2SWei Yang * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent 26347b390d2SWei Yang * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call, 26447b390d2SWei Yang * we can identify this case by checking (!dst->anon_vma && src->anon_vma). 26547b390d2SWei Yang * 26647b390d2SWei Yang * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find 26747b390d2SWei Yang * and reuse existing anon_vma which has no vmas and only one child anon_vma. 26847b390d2SWei Yang * This prevents degradation of anon_vma hierarchy to endless linear chain in 26947b390d2SWei Yang * case of constantly forking task. On the other hand, an anon_vma with more 27047b390d2SWei Yang * than one child isn't reused even if there was no alive vma, thus rmap 27147b390d2SWei Yang * walker has a good chance of avoiding scanning the whole hierarchy when it 27247b390d2SWei Yang * searches where page is mapped. 2735beb4930SRik van Riel */ 2745beb4930SRik van Riel int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) 2755beb4930SRik van Riel { 2765beb4930SRik van Riel struct anon_vma_chain *avc, *pavc; 277bb4aa396SLinus Torvalds struct anon_vma *root = NULL; 2785beb4930SRik van Riel 279646d87b4SLinus Torvalds list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { 280bb4aa396SLinus Torvalds struct anon_vma *anon_vma; 281bb4aa396SLinus Torvalds 282dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); 283dd34739cSLinus Torvalds if (unlikely(!avc)) { 284dd34739cSLinus Torvalds unlock_anon_vma_root(root); 285dd34739cSLinus Torvalds root = NULL; 286dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 2875beb4930SRik van Riel if (!avc) 2885beb4930SRik van Riel goto enomem_failure; 289dd34739cSLinus Torvalds } 290bb4aa396SLinus Torvalds anon_vma = pavc->anon_vma; 291bb4aa396SLinus Torvalds root = lock_anon_vma_root(root, anon_vma); 292bb4aa396SLinus Torvalds anon_vma_chain_link(dst, avc, anon_vma); 2937a3ef208SKonstantin Khlebnikov 2947a3ef208SKonstantin Khlebnikov /* 2957a3ef208SKonstantin Khlebnikov * Reuse existing anon_vma if its degree lower than two, 2967a3ef208SKonstantin Khlebnikov * that means it has no vma and only one anon_vma child. 2977a3ef208SKonstantin Khlebnikov * 2987a3ef208SKonstantin Khlebnikov * Do not chose parent anon_vma, otherwise first child 2997a3ef208SKonstantin Khlebnikov * will always reuse it. Root anon_vma is never reused: 3007a3ef208SKonstantin Khlebnikov * it has self-parent reference and at least one child. 3017a3ef208SKonstantin Khlebnikov */ 30247b390d2SWei Yang if (!dst->anon_vma && src->anon_vma && 30347b390d2SWei Yang anon_vma != src->anon_vma && anon_vma->degree < 2) 3047a3ef208SKonstantin Khlebnikov dst->anon_vma = anon_vma; 3055beb4930SRik van Riel } 3067a3ef208SKonstantin Khlebnikov if (dst->anon_vma) 3077a3ef208SKonstantin Khlebnikov dst->anon_vma->degree++; 308bb4aa396SLinus Torvalds unlock_anon_vma_root(root); 3095beb4930SRik van Riel return 0; 3105beb4930SRik van Riel 3115beb4930SRik van Riel enomem_failure: 3123fe89b3eSLeon Yu /* 3133fe89b3eSLeon Yu * dst->anon_vma is dropped here otherwise its degree can be incorrectly 3143fe89b3eSLeon Yu * decremented in unlink_anon_vmas(). 3153fe89b3eSLeon Yu * We can safely do this because callers of anon_vma_clone() don't care 3163fe89b3eSLeon Yu * about dst->anon_vma if anon_vma_clone() failed. 3173fe89b3eSLeon Yu */ 3183fe89b3eSLeon Yu dst->anon_vma = NULL; 3195beb4930SRik van Riel unlink_anon_vmas(dst); 3205beb4930SRik van Riel return -ENOMEM; 3211da177e4SLinus Torvalds } 3221da177e4SLinus Torvalds 3235beb4930SRik van Riel /* 3245beb4930SRik van Riel * Attach vma to its own anon_vma, as well as to the anon_vmas that 3255beb4930SRik van Riel * the corresponding VMA in the parent process is attached to. 3265beb4930SRik van Riel * Returns 0 on success, non-zero on failure. 3275beb4930SRik van Riel */ 3285beb4930SRik van Riel int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) 3291da177e4SLinus Torvalds { 3305beb4930SRik van Riel struct anon_vma_chain *avc; 3315beb4930SRik van Riel struct anon_vma *anon_vma; 332c4ea95d7SDaniel Forrest int error; 3335beb4930SRik van Riel 3345beb4930SRik van Riel /* Don't bother if the parent process has no anon_vma here. */ 3355beb4930SRik van Riel if (!pvma->anon_vma) 3365beb4930SRik van Riel return 0; 3375beb4930SRik van Riel 3387a3ef208SKonstantin Khlebnikov /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ 3397a3ef208SKonstantin Khlebnikov vma->anon_vma = NULL; 3407a3ef208SKonstantin Khlebnikov 3415beb4930SRik van Riel /* 3425beb4930SRik van Riel * First, attach the new VMA to the parent VMA's anon_vmas, 3435beb4930SRik van Riel * so rmap can find non-COWed pages in child processes. 3445beb4930SRik van Riel */ 345c4ea95d7SDaniel Forrest error = anon_vma_clone(vma, pvma); 346c4ea95d7SDaniel Forrest if (error) 347c4ea95d7SDaniel Forrest return error; 3485beb4930SRik van Riel 3497a3ef208SKonstantin Khlebnikov /* An existing anon_vma has been reused, all done then. */ 3507a3ef208SKonstantin Khlebnikov if (vma->anon_vma) 3517a3ef208SKonstantin Khlebnikov return 0; 3527a3ef208SKonstantin Khlebnikov 3535beb4930SRik van Riel /* Then add our own anon_vma. */ 3545beb4930SRik van Riel anon_vma = anon_vma_alloc(); 3555beb4930SRik van Riel if (!anon_vma) 3565beb4930SRik van Riel goto out_error; 357dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 3585beb4930SRik van Riel if (!avc) 3595beb4930SRik van Riel goto out_error_free_anon_vma; 3605c341ee1SRik van Riel 3615c341ee1SRik van Riel /* 362aaf1f990SMiaohe Lin * The root anon_vma's rwsem is the lock actually used when we 3635c341ee1SRik van Riel * lock any of the anon_vmas in this anon_vma tree. 3645c341ee1SRik van Riel */ 3655c341ee1SRik van Riel anon_vma->root = pvma->anon_vma->root; 3667a3ef208SKonstantin Khlebnikov anon_vma->parent = pvma->anon_vma; 36776545066SRik van Riel /* 36801d8b20dSPeter Zijlstra * With refcounts, an anon_vma can stay around longer than the 36901d8b20dSPeter Zijlstra * process it belongs to. The root anon_vma needs to be pinned until 37001d8b20dSPeter Zijlstra * this anon_vma is freed, because the lock lives in the root. 37176545066SRik van Riel */ 37276545066SRik van Riel get_anon_vma(anon_vma->root); 3735beb4930SRik van Riel /* Mark this anon_vma as the one where our new (COWed) pages go. */ 3745beb4930SRik van Riel vma->anon_vma = anon_vma; 3754fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 3765c341ee1SRik van Riel anon_vma_chain_link(vma, avc, anon_vma); 3777a3ef208SKonstantin Khlebnikov anon_vma->parent->degree++; 37808b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 3795beb4930SRik van Riel 3805beb4930SRik van Riel return 0; 3815beb4930SRik van Riel 3825beb4930SRik van Riel out_error_free_anon_vma: 38301d8b20dSPeter Zijlstra put_anon_vma(anon_vma); 3845beb4930SRik van Riel out_error: 3854946d54cSRik van Riel unlink_anon_vmas(vma); 3865beb4930SRik van Riel return -ENOMEM; 3875beb4930SRik van Riel } 3885beb4930SRik van Riel 3895beb4930SRik van Riel void unlink_anon_vmas(struct vm_area_struct *vma) 3905beb4930SRik van Riel { 3915beb4930SRik van Riel struct anon_vma_chain *avc, *next; 392eee2acbaSPeter Zijlstra struct anon_vma *root = NULL; 3935beb4930SRik van Riel 3945c341ee1SRik van Riel /* 3955c341ee1SRik van Riel * Unlink each anon_vma chained to the VMA. This list is ordered 3965c341ee1SRik van Riel * from newest to oldest, ensuring the root anon_vma gets freed last. 3975c341ee1SRik van Riel */ 3985beb4930SRik van Riel list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 399eee2acbaSPeter Zijlstra struct anon_vma *anon_vma = avc->anon_vma; 400eee2acbaSPeter Zijlstra 401eee2acbaSPeter Zijlstra root = lock_anon_vma_root(root, anon_vma); 402bf181b9fSMichel Lespinasse anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); 403eee2acbaSPeter Zijlstra 404eee2acbaSPeter Zijlstra /* 405eee2acbaSPeter Zijlstra * Leave empty anon_vmas on the list - we'll need 406eee2acbaSPeter Zijlstra * to free them outside the lock. 407eee2acbaSPeter Zijlstra */ 408f808c13fSDavidlohr Bueso if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) { 4097a3ef208SKonstantin Khlebnikov anon_vma->parent->degree--; 410eee2acbaSPeter Zijlstra continue; 4117a3ef208SKonstantin Khlebnikov } 412eee2acbaSPeter Zijlstra 413eee2acbaSPeter Zijlstra list_del(&avc->same_vma); 414eee2acbaSPeter Zijlstra anon_vma_chain_free(avc); 415eee2acbaSPeter Zijlstra } 416ee8ab190SLi Xinhai if (vma->anon_vma) { 4177a3ef208SKonstantin Khlebnikov vma->anon_vma->degree--; 418ee8ab190SLi Xinhai 419ee8ab190SLi Xinhai /* 420ee8ab190SLi Xinhai * vma would still be needed after unlink, and anon_vma will be prepared 421ee8ab190SLi Xinhai * when handle fault. 422ee8ab190SLi Xinhai */ 423ee8ab190SLi Xinhai vma->anon_vma = NULL; 424ee8ab190SLi Xinhai } 425eee2acbaSPeter Zijlstra unlock_anon_vma_root(root); 426eee2acbaSPeter Zijlstra 427eee2acbaSPeter Zijlstra /* 428eee2acbaSPeter Zijlstra * Iterate the list once more, it now only contains empty and unlinked 429eee2acbaSPeter Zijlstra * anon_vmas, destroy them. Could not do before due to __put_anon_vma() 4305a505085SIngo Molnar * needing to write-acquire the anon_vma->root->rwsem. 431eee2acbaSPeter Zijlstra */ 432eee2acbaSPeter Zijlstra list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 433eee2acbaSPeter Zijlstra struct anon_vma *anon_vma = avc->anon_vma; 434eee2acbaSPeter Zijlstra 435e4c5800aSKonstantin Khlebnikov VM_WARN_ON(anon_vma->degree); 436eee2acbaSPeter Zijlstra put_anon_vma(anon_vma); 437eee2acbaSPeter Zijlstra 4385beb4930SRik van Riel list_del(&avc->same_vma); 4395beb4930SRik van Riel anon_vma_chain_free(avc); 4405beb4930SRik van Riel } 4415beb4930SRik van Riel } 4425beb4930SRik van Riel 44351cc5068SAlexey Dobriyan static void anon_vma_ctor(void *data) 4441da177e4SLinus Torvalds { 4451da177e4SLinus Torvalds struct anon_vma *anon_vma = data; 4461da177e4SLinus Torvalds 4475a505085SIngo Molnar init_rwsem(&anon_vma->rwsem); 44883813267SPeter Zijlstra atomic_set(&anon_vma->refcount, 0); 449f808c13fSDavidlohr Bueso anon_vma->rb_root = RB_ROOT_CACHED; 4501da177e4SLinus Torvalds } 4511da177e4SLinus Torvalds 4521da177e4SLinus Torvalds void __init anon_vma_init(void) 4531da177e4SLinus Torvalds { 4541da177e4SLinus Torvalds anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 4555f0d5a3aSPaul E. McKenney 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, 4565d097056SVladimir Davydov anon_vma_ctor); 4575d097056SVladimir Davydov anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, 4585d097056SVladimir Davydov SLAB_PANIC|SLAB_ACCOUNT); 4591da177e4SLinus Torvalds } 4601da177e4SLinus Torvalds 4611da177e4SLinus Torvalds /* 4626111e4caSPeter Zijlstra * Getting a lock on a stable anon_vma from a page off the LRU is tricky! 4636111e4caSPeter Zijlstra * 4646111e4caSPeter Zijlstra * Since there is no serialization what so ever against page_remove_rmap() 4656111e4caSPeter Zijlstra * the best this function can do is return a locked anon_vma that might 4666111e4caSPeter Zijlstra * have been relevant to this page. 4676111e4caSPeter Zijlstra * 4686111e4caSPeter Zijlstra * The page might have been remapped to a different anon_vma or the anon_vma 4696111e4caSPeter Zijlstra * returned may already be freed (and even reused). 4706111e4caSPeter Zijlstra * 471bc658c96SPeter Zijlstra * In case it was remapped to a different anon_vma, the new anon_vma will be a 472bc658c96SPeter Zijlstra * child of the old anon_vma, and the anon_vma lifetime rules will therefore 473bc658c96SPeter Zijlstra * ensure that any anon_vma obtained from the page will still be valid for as 474bc658c96SPeter Zijlstra * long as we observe page_mapped() [ hence all those page_mapped() tests ]. 475bc658c96SPeter Zijlstra * 4766111e4caSPeter Zijlstra * All users of this function must be very careful when walking the anon_vma 4776111e4caSPeter Zijlstra * chain and verify that the page in question is indeed mapped in it 4786111e4caSPeter Zijlstra * [ something equivalent to page_mapped_in_vma() ]. 4796111e4caSPeter Zijlstra * 480091e4299SMiles Chen * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from 481091e4299SMiles Chen * page_remove_rmap() that the anon_vma pointer from page->mapping is valid 482091e4299SMiles Chen * if there is a mapcount, we can dereference the anon_vma after observing 483091e4299SMiles Chen * those. 4841da177e4SLinus Torvalds */ 485746b18d4SPeter Zijlstra struct anon_vma *page_get_anon_vma(struct page *page) 4861da177e4SLinus Torvalds { 487746b18d4SPeter Zijlstra struct anon_vma *anon_vma = NULL; 4881da177e4SLinus Torvalds unsigned long anon_mapping; 4891da177e4SLinus Torvalds 4901da177e4SLinus Torvalds rcu_read_lock(); 4914db0c3c2SJason Low anon_mapping = (unsigned long)READ_ONCE(page->mapping); 4923ca7b3c5SHugh Dickins if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 4931da177e4SLinus Torvalds goto out; 4941da177e4SLinus Torvalds if (!page_mapped(page)) 4951da177e4SLinus Torvalds goto out; 4961da177e4SLinus Torvalds 4971da177e4SLinus Torvalds anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 498746b18d4SPeter Zijlstra if (!atomic_inc_not_zero(&anon_vma->refcount)) { 499746b18d4SPeter Zijlstra anon_vma = NULL; 500746b18d4SPeter Zijlstra goto out; 501746b18d4SPeter Zijlstra } 502f1819427SHugh Dickins 503f1819427SHugh Dickins /* 504f1819427SHugh Dickins * If this page is still mapped, then its anon_vma cannot have been 505746b18d4SPeter Zijlstra * freed. But if it has been unmapped, we have no security against the 506746b18d4SPeter Zijlstra * anon_vma structure being freed and reused (for another anon_vma: 5075f0d5a3aSPaul E. McKenney * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero() 508746b18d4SPeter Zijlstra * above cannot corrupt). 509f1819427SHugh Dickins */ 510746b18d4SPeter Zijlstra if (!page_mapped(page)) { 5117f39dda9SHugh Dickins rcu_read_unlock(); 512746b18d4SPeter Zijlstra put_anon_vma(anon_vma); 5137f39dda9SHugh Dickins return NULL; 514746b18d4SPeter Zijlstra } 5151da177e4SLinus Torvalds out: 5161da177e4SLinus Torvalds rcu_read_unlock(); 517746b18d4SPeter Zijlstra 518746b18d4SPeter Zijlstra return anon_vma; 519746b18d4SPeter Zijlstra } 520746b18d4SPeter Zijlstra 52188c22088SPeter Zijlstra /* 52288c22088SPeter Zijlstra * Similar to page_get_anon_vma() except it locks the anon_vma. 52388c22088SPeter Zijlstra * 52488c22088SPeter Zijlstra * Its a little more complex as it tries to keep the fast path to a single 52588c22088SPeter Zijlstra * atomic op -- the trylock. If we fail the trylock, we fall back to getting a 52688c22088SPeter Zijlstra * reference like with page_get_anon_vma() and then block on the mutex. 52788c22088SPeter Zijlstra */ 5284fc3f1d6SIngo Molnar struct anon_vma *page_lock_anon_vma_read(struct page *page) 529746b18d4SPeter Zijlstra { 53088c22088SPeter Zijlstra struct anon_vma *anon_vma = NULL; 531eee0f252SHugh Dickins struct anon_vma *root_anon_vma; 53288c22088SPeter Zijlstra unsigned long anon_mapping; 533746b18d4SPeter Zijlstra 53488c22088SPeter Zijlstra rcu_read_lock(); 5354db0c3c2SJason Low anon_mapping = (unsigned long)READ_ONCE(page->mapping); 53688c22088SPeter Zijlstra if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 53788c22088SPeter Zijlstra goto out; 53888c22088SPeter Zijlstra if (!page_mapped(page)) 53988c22088SPeter Zijlstra goto out; 54088c22088SPeter Zijlstra 54188c22088SPeter Zijlstra anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 5424db0c3c2SJason Low root_anon_vma = READ_ONCE(anon_vma->root); 5434fc3f1d6SIngo Molnar if (down_read_trylock(&root_anon_vma->rwsem)) { 54488c22088SPeter Zijlstra /* 545eee0f252SHugh Dickins * If the page is still mapped, then this anon_vma is still 546eee0f252SHugh Dickins * its anon_vma, and holding the mutex ensures that it will 547bc658c96SPeter Zijlstra * not go away, see anon_vma_free(). 54888c22088SPeter Zijlstra */ 549eee0f252SHugh Dickins if (!page_mapped(page)) { 5504fc3f1d6SIngo Molnar up_read(&root_anon_vma->rwsem); 55188c22088SPeter Zijlstra anon_vma = NULL; 55288c22088SPeter Zijlstra } 55388c22088SPeter Zijlstra goto out; 55488c22088SPeter Zijlstra } 55588c22088SPeter Zijlstra 55688c22088SPeter Zijlstra /* trylock failed, we got to sleep */ 55788c22088SPeter Zijlstra if (!atomic_inc_not_zero(&anon_vma->refcount)) { 55888c22088SPeter Zijlstra anon_vma = NULL; 55988c22088SPeter Zijlstra goto out; 56088c22088SPeter Zijlstra } 56188c22088SPeter Zijlstra 56288c22088SPeter Zijlstra if (!page_mapped(page)) { 5637f39dda9SHugh Dickins rcu_read_unlock(); 56488c22088SPeter Zijlstra put_anon_vma(anon_vma); 5657f39dda9SHugh Dickins return NULL; 56688c22088SPeter Zijlstra } 56788c22088SPeter Zijlstra 56888c22088SPeter Zijlstra /* we pinned the anon_vma, its safe to sleep */ 56988c22088SPeter Zijlstra rcu_read_unlock(); 5704fc3f1d6SIngo Molnar anon_vma_lock_read(anon_vma); 571746b18d4SPeter Zijlstra 57288c22088SPeter Zijlstra if (atomic_dec_and_test(&anon_vma->refcount)) { 57388c22088SPeter Zijlstra /* 57488c22088SPeter Zijlstra * Oops, we held the last refcount, release the lock 57588c22088SPeter Zijlstra * and bail -- can't simply use put_anon_vma() because 5764fc3f1d6SIngo Molnar * we'll deadlock on the anon_vma_lock_write() recursion. 57788c22088SPeter Zijlstra */ 5784fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 57988c22088SPeter Zijlstra __put_anon_vma(anon_vma); 58088c22088SPeter Zijlstra anon_vma = NULL; 58188c22088SPeter Zijlstra } 58288c22088SPeter Zijlstra 58388c22088SPeter Zijlstra return anon_vma; 58488c22088SPeter Zijlstra 58588c22088SPeter Zijlstra out: 58688c22088SPeter Zijlstra rcu_read_unlock(); 587746b18d4SPeter Zijlstra return anon_vma; 58834bbd704SOleg Nesterov } 58934bbd704SOleg Nesterov 5904fc3f1d6SIngo Molnar void page_unlock_anon_vma_read(struct anon_vma *anon_vma) 59134bbd704SOleg Nesterov { 5924fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 5931da177e4SLinus Torvalds } 5941da177e4SLinus Torvalds 59572b252aeSMel Gorman #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH 59672b252aeSMel Gorman /* 59772b252aeSMel Gorman * Flush TLB entries for recently unmapped pages from remote CPUs. It is 59872b252aeSMel Gorman * important if a PTE was dirty when it was unmapped that it's flushed 59972b252aeSMel Gorman * before any IO is initiated on the page to prevent lost writes. Similarly, 60072b252aeSMel Gorman * it must be flushed before freeing to prevent data leakage. 60172b252aeSMel Gorman */ 60272b252aeSMel Gorman void try_to_unmap_flush(void) 60372b252aeSMel Gorman { 60472b252aeSMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 60572b252aeSMel Gorman 60672b252aeSMel Gorman if (!tlb_ubc->flush_required) 60772b252aeSMel Gorman return; 60872b252aeSMel Gorman 609e73ad5ffSAndy Lutomirski arch_tlbbatch_flush(&tlb_ubc->arch); 61072b252aeSMel Gorman tlb_ubc->flush_required = false; 611d950c947SMel Gorman tlb_ubc->writable = false; 61272b252aeSMel Gorman } 61372b252aeSMel Gorman 614d950c947SMel Gorman /* Flush iff there are potentially writable TLB entries that can race with IO */ 615d950c947SMel Gorman void try_to_unmap_flush_dirty(void) 616d950c947SMel Gorman { 617d950c947SMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 618d950c947SMel Gorman 619d950c947SMel Gorman if (tlb_ubc->writable) 620d950c947SMel Gorman try_to_unmap_flush(); 621d950c947SMel Gorman } 622d950c947SMel Gorman 623c7ab0d2fSKirill A. Shutemov static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) 62472b252aeSMel Gorman { 62572b252aeSMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 62672b252aeSMel Gorman 627e73ad5ffSAndy Lutomirski arch_tlbbatch_add_mm(&tlb_ubc->arch, mm); 62872b252aeSMel Gorman tlb_ubc->flush_required = true; 629d950c947SMel Gorman 630d950c947SMel Gorman /* 6313ea27719SMel Gorman * Ensure compiler does not re-order the setting of tlb_flush_batched 6323ea27719SMel Gorman * before the PTE is cleared. 6333ea27719SMel Gorman */ 6343ea27719SMel Gorman barrier(); 6353ea27719SMel Gorman mm->tlb_flush_batched = true; 6363ea27719SMel Gorman 6373ea27719SMel Gorman /* 638d950c947SMel Gorman * If the PTE was dirty then it's best to assume it's writable. The 639d950c947SMel Gorman * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() 640d950c947SMel Gorman * before the page is queued for IO. 641d950c947SMel Gorman */ 642d950c947SMel Gorman if (writable) 643d950c947SMel Gorman tlb_ubc->writable = true; 64472b252aeSMel Gorman } 64572b252aeSMel Gorman 64672b252aeSMel Gorman /* 64772b252aeSMel Gorman * Returns true if the TLB flush should be deferred to the end of a batch of 64872b252aeSMel Gorman * unmap operations to reduce IPIs. 64972b252aeSMel Gorman */ 65072b252aeSMel Gorman static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 65172b252aeSMel Gorman { 65272b252aeSMel Gorman bool should_defer = false; 65372b252aeSMel Gorman 65472b252aeSMel Gorman if (!(flags & TTU_BATCH_FLUSH)) 65572b252aeSMel Gorman return false; 65672b252aeSMel Gorman 65772b252aeSMel Gorman /* If remote CPUs need to be flushed then defer batch the flush */ 65872b252aeSMel Gorman if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids) 65972b252aeSMel Gorman should_defer = true; 66072b252aeSMel Gorman put_cpu(); 66172b252aeSMel Gorman 66272b252aeSMel Gorman return should_defer; 66372b252aeSMel Gorman } 6643ea27719SMel Gorman 6653ea27719SMel Gorman /* 6663ea27719SMel Gorman * Reclaim unmaps pages under the PTL but do not flush the TLB prior to 6673ea27719SMel Gorman * releasing the PTL if TLB flushes are batched. It's possible for a parallel 6683ea27719SMel Gorman * operation such as mprotect or munmap to race between reclaim unmapping 6693ea27719SMel Gorman * the page and flushing the page. If this race occurs, it potentially allows 6703ea27719SMel Gorman * access to data via a stale TLB entry. Tracking all mm's that have TLB 6713ea27719SMel Gorman * batching in flight would be expensive during reclaim so instead track 6723ea27719SMel Gorman * whether TLB batching occurred in the past and if so then do a flush here 6733ea27719SMel Gorman * if required. This will cost one additional flush per reclaim cycle paid 6743ea27719SMel Gorman * by the first operation at risk such as mprotect and mumap. 6753ea27719SMel Gorman * 6763ea27719SMel Gorman * This must be called under the PTL so that an access to tlb_flush_batched 6773ea27719SMel Gorman * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise 6783ea27719SMel Gorman * via the PTL. 6793ea27719SMel Gorman */ 6803ea27719SMel Gorman void flush_tlb_batched_pending(struct mm_struct *mm) 6813ea27719SMel Gorman { 6829c1177b6SQian Cai if (data_race(mm->tlb_flush_batched)) { 6833ea27719SMel Gorman flush_tlb_mm(mm); 6843ea27719SMel Gorman 6853ea27719SMel Gorman /* 6863ea27719SMel Gorman * Do not allow the compiler to re-order the clearing of 6873ea27719SMel Gorman * tlb_flush_batched before the tlb is flushed. 6883ea27719SMel Gorman */ 6893ea27719SMel Gorman barrier(); 6903ea27719SMel Gorman mm->tlb_flush_batched = false; 6913ea27719SMel Gorman } 6923ea27719SMel Gorman } 69372b252aeSMel Gorman #else 694c7ab0d2fSKirill A. Shutemov static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) 69572b252aeSMel Gorman { 69672b252aeSMel Gorman } 69772b252aeSMel Gorman 69872b252aeSMel Gorman static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 69972b252aeSMel Gorman { 70072b252aeSMel Gorman return false; 70172b252aeSMel Gorman } 70272b252aeSMel Gorman #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ 70372b252aeSMel Gorman 7041da177e4SLinus Torvalds /* 705bf89c8c8SHuang Shijie * At what user virtual address is page expected in vma? 706ab941e0fSNaoya Horiguchi * Caller should check the page is actually part of the vma. 7071da177e4SLinus Torvalds */ 7081da177e4SLinus Torvalds unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 7091da177e4SLinus Torvalds { 71086c2ad19SMichel Lespinasse unsigned long address; 71121d0d443SAndrea Arcangeli if (PageAnon(page)) { 7124829b906SHugh Dickins struct anon_vma *page__anon_vma = page_anon_vma(page); 7134829b906SHugh Dickins /* 7144829b906SHugh Dickins * Note: swapoff's unuse_vma() is more efficient with this 7154829b906SHugh Dickins * check, and needs it to match anon_vma when KSM is active. 7164829b906SHugh Dickins */ 7174829b906SHugh Dickins if (!vma->anon_vma || !page__anon_vma || 7184829b906SHugh Dickins vma->anon_vma->root != page__anon_vma->root) 71921d0d443SAndrea Arcangeli return -EFAULT; 72027ba0644SKirill A. Shutemov } else if (page->mapping) { 72127ba0644SKirill A. Shutemov if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping) 7221da177e4SLinus Torvalds return -EFAULT; 7231da177e4SLinus Torvalds } else 7241da177e4SLinus Torvalds return -EFAULT; 72586c2ad19SMichel Lespinasse address = __vma_address(page, vma); 72686c2ad19SMichel Lespinasse if (unlikely(address < vma->vm_start || address >= vma->vm_end)) 72786c2ad19SMichel Lespinasse return -EFAULT; 72886c2ad19SMichel Lespinasse return address; 7291da177e4SLinus Torvalds } 7301da177e4SLinus Torvalds 7316219049aSBob Liu pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) 7326219049aSBob Liu { 7336219049aSBob Liu pgd_t *pgd; 734c2febafcSKirill A. Shutemov p4d_t *p4d; 7356219049aSBob Liu pud_t *pud; 7366219049aSBob Liu pmd_t *pmd = NULL; 737f72e7dcdSHugh Dickins pmd_t pmde; 7386219049aSBob Liu 7396219049aSBob Liu pgd = pgd_offset(mm, address); 7406219049aSBob Liu if (!pgd_present(*pgd)) 7416219049aSBob Liu goto out; 7426219049aSBob Liu 743c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, address); 744c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 745c2febafcSKirill A. Shutemov goto out; 746c2febafcSKirill A. Shutemov 747c2febafcSKirill A. Shutemov pud = pud_offset(p4d, address); 7486219049aSBob Liu if (!pud_present(*pud)) 7496219049aSBob Liu goto out; 7506219049aSBob Liu 7516219049aSBob Liu pmd = pmd_offset(pud, address); 752f72e7dcdSHugh Dickins /* 7538809aa2dSAneesh Kumar K.V * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() 754f72e7dcdSHugh Dickins * without holding anon_vma lock for write. So when looking for a 755f72e7dcdSHugh Dickins * genuine pmde (in which to find pte), test present and !THP together. 756f72e7dcdSHugh Dickins */ 757e37c6982SChristian Borntraeger pmde = *pmd; 758e37c6982SChristian Borntraeger barrier(); 759f72e7dcdSHugh Dickins if (!pmd_present(pmde) || pmd_trans_huge(pmde)) 7606219049aSBob Liu pmd = NULL; 7616219049aSBob Liu out: 7626219049aSBob Liu return pmd; 7636219049aSBob Liu } 7646219049aSBob Liu 7659f32624bSJoonsoo Kim struct page_referenced_arg { 7669f32624bSJoonsoo Kim int mapcount; 7679f32624bSJoonsoo Kim int referenced; 7689f32624bSJoonsoo Kim unsigned long vm_flags; 7699f32624bSJoonsoo Kim struct mem_cgroup *memcg; 7709f32624bSJoonsoo Kim }; 77181b4082dSNikita Danilov /* 7729f32624bSJoonsoo Kim * arg: page_referenced_arg will be passed 7731da177e4SLinus Torvalds */ 774e4b82222SMinchan Kim static bool page_referenced_one(struct page *page, struct vm_area_struct *vma, 7759f32624bSJoonsoo Kim unsigned long address, void *arg) 7761da177e4SLinus Torvalds { 7779f32624bSJoonsoo Kim struct page_referenced_arg *pra = arg; 7788eaededeSKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 7798eaededeSKirill A. Shutemov .page = page, 7808eaededeSKirill A. Shutemov .vma = vma, 7818eaededeSKirill A. Shutemov .address = address, 7828eaededeSKirill A. Shutemov }; 7838749cfeaSVladimir Davydov int referenced = 0; 7842da28bfdSAndrea Arcangeli 7858eaededeSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 7868eaededeSKirill A. Shutemov address = pvmw.address; 7872da28bfdSAndrea Arcangeli 788b20ce5e0SKirill A. Shutemov if (vma->vm_flags & VM_LOCKED) { 7898eaededeSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 7909f32624bSJoonsoo Kim pra->vm_flags |= VM_LOCKED; 791e4b82222SMinchan Kim return false; /* To break the loop */ 7922da28bfdSAndrea Arcangeli } 7932da28bfdSAndrea Arcangeli 7948eaededeSKirill A. Shutemov if (pvmw.pte) { 7958eaededeSKirill A. Shutemov if (ptep_clear_flush_young_notify(vma, address, 7968eaededeSKirill A. Shutemov pvmw.pte)) { 7974917e5d0SJohannes Weiner /* 7988eaededeSKirill A. Shutemov * Don't treat a reference through 7998eaededeSKirill A. Shutemov * a sequentially read mapping as such. 8008eaededeSKirill A. Shutemov * If the page has been used in another mapping, 8018eaededeSKirill A. Shutemov * we will catch it; if this other mapping is 8028eaededeSKirill A. Shutemov * already gone, the unmap path will have set 8038eaededeSKirill A. Shutemov * PG_referenced or activated the page. 8044917e5d0SJohannes Weiner */ 80564363aadSJoe Perches if (likely(!(vma->vm_flags & VM_SEQ_READ))) 8061da177e4SLinus Torvalds referenced++; 8074917e5d0SJohannes Weiner } 8088749cfeaSVladimir Davydov } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { 8098eaededeSKirill A. Shutemov if (pmdp_clear_flush_young_notify(vma, address, 8108eaededeSKirill A. Shutemov pvmw.pmd)) 8118749cfeaSVladimir Davydov referenced++; 8128749cfeaSVladimir Davydov } else { 8138749cfeaSVladimir Davydov /* unexpected pmd-mapped page? */ 8148749cfeaSVladimir Davydov WARN_ON_ONCE(1); 8158749cfeaSVladimir Davydov } 8168eaededeSKirill A. Shutemov 8178eaededeSKirill A. Shutemov pra->mapcount--; 8188eaededeSKirill A. Shutemov } 81971e3aac0SAndrea Arcangeli 82033c3fc71SVladimir Davydov if (referenced) 82133c3fc71SVladimir Davydov clear_page_idle(page); 82233c3fc71SVladimir Davydov if (test_and_clear_page_young(page)) 82333c3fc71SVladimir Davydov referenced++; 82433c3fc71SVladimir Davydov 8259f32624bSJoonsoo Kim if (referenced) { 8269f32624bSJoonsoo Kim pra->referenced++; 8279f32624bSJoonsoo Kim pra->vm_flags |= vma->vm_flags; 8281da177e4SLinus Torvalds } 8291da177e4SLinus Torvalds 8309f32624bSJoonsoo Kim if (!pra->mapcount) 831e4b82222SMinchan Kim return false; /* To break the loop */ 8329f32624bSJoonsoo Kim 833e4b82222SMinchan Kim return true; 8349f32624bSJoonsoo Kim } 8359f32624bSJoonsoo Kim 8369f32624bSJoonsoo Kim static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg) 8371da177e4SLinus Torvalds { 8389f32624bSJoonsoo Kim struct page_referenced_arg *pra = arg; 8399f32624bSJoonsoo Kim struct mem_cgroup *memcg = pra->memcg; 8401da177e4SLinus Torvalds 8419f32624bSJoonsoo Kim if (!mm_match_cgroup(vma->vm_mm, memcg)) 8429f32624bSJoonsoo Kim return true; 8431da177e4SLinus Torvalds 8449f32624bSJoonsoo Kim return false; 8451da177e4SLinus Torvalds } 8461da177e4SLinus Torvalds 8471da177e4SLinus Torvalds /** 8481da177e4SLinus Torvalds * page_referenced - test if the page was referenced 8491da177e4SLinus Torvalds * @page: the page to test 8501da177e4SLinus Torvalds * @is_locked: caller holds lock on the page 85172835c86SJohannes Weiner * @memcg: target memory cgroup 8526fe6b7e3SWu Fengguang * @vm_flags: collect encountered vma->vm_flags who actually referenced the page 8531da177e4SLinus Torvalds * 8541da177e4SLinus Torvalds * Quick test_and_clear_referenced for all mappings to a page, 8551da177e4SLinus Torvalds * returns the number of ptes which referenced the page. 8561da177e4SLinus Torvalds */ 8576fe6b7e3SWu Fengguang int page_referenced(struct page *page, 8586fe6b7e3SWu Fengguang int is_locked, 85972835c86SJohannes Weiner struct mem_cgroup *memcg, 8606fe6b7e3SWu Fengguang unsigned long *vm_flags) 8611da177e4SLinus Torvalds { 8625ad64688SHugh Dickins int we_locked = 0; 8639f32624bSJoonsoo Kim struct page_referenced_arg pra = { 864b20ce5e0SKirill A. Shutemov .mapcount = total_mapcount(page), 8659f32624bSJoonsoo Kim .memcg = memcg, 8669f32624bSJoonsoo Kim }; 8679f32624bSJoonsoo Kim struct rmap_walk_control rwc = { 8689f32624bSJoonsoo Kim .rmap_one = page_referenced_one, 8699f32624bSJoonsoo Kim .arg = (void *)&pra, 8709f32624bSJoonsoo Kim .anon_lock = page_lock_anon_vma_read, 8719f32624bSJoonsoo Kim }; 8721da177e4SLinus Torvalds 8736fe6b7e3SWu Fengguang *vm_flags = 0; 874059d8442SHuang Shijie if (!pra.mapcount) 8759f32624bSJoonsoo Kim return 0; 8769f32624bSJoonsoo Kim 8779f32624bSJoonsoo Kim if (!page_rmapping(page)) 8789f32624bSJoonsoo Kim return 0; 8799f32624bSJoonsoo Kim 8805ad64688SHugh Dickins if (!is_locked && (!PageAnon(page) || PageKsm(page))) { 8815ad64688SHugh Dickins we_locked = trylock_page(page); 8829f32624bSJoonsoo Kim if (!we_locked) 8839f32624bSJoonsoo Kim return 1; 8845ad64688SHugh Dickins } 8859f32624bSJoonsoo Kim 8869f32624bSJoonsoo Kim /* 8879f32624bSJoonsoo Kim * If we are reclaiming on behalf of a cgroup, skip 8889f32624bSJoonsoo Kim * counting on behalf of references from different 8899f32624bSJoonsoo Kim * cgroups 8909f32624bSJoonsoo Kim */ 8919f32624bSJoonsoo Kim if (memcg) { 8929f32624bSJoonsoo Kim rwc.invalid_vma = invalid_page_referenced_vma; 8935ad64688SHugh Dickins } 8949f32624bSJoonsoo Kim 895c24f386cSMinchan Kim rmap_walk(page, &rwc); 8969f32624bSJoonsoo Kim *vm_flags = pra.vm_flags; 8979f32624bSJoonsoo Kim 8985ad64688SHugh Dickins if (we_locked) 8991da177e4SLinus Torvalds unlock_page(page); 9009f32624bSJoonsoo Kim 9019f32624bSJoonsoo Kim return pra.referenced; 9021da177e4SLinus Torvalds } 9031da177e4SLinus Torvalds 904e4b82222SMinchan Kim static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma, 9059853a407SJoonsoo Kim unsigned long address, void *arg) 906d08b3851SPeter Zijlstra { 907f27176cfSKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 908f27176cfSKirill A. Shutemov .page = page, 909f27176cfSKirill A. Shutemov .vma = vma, 910f27176cfSKirill A. Shutemov .address = address, 911f27176cfSKirill A. Shutemov .flags = PVMW_SYNC, 912f27176cfSKirill A. Shutemov }; 913ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 9149853a407SJoonsoo Kim int *cleaned = arg; 915d08b3851SPeter Zijlstra 916369ea824SJérôme Glisse /* 917369ea824SJérôme Glisse * We have to assume the worse case ie pmd for invalidation. Note that 918369ea824SJérôme Glisse * the page can not be free from this function. 919369ea824SJérôme Glisse */ 9207269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 9217269f999SJérôme Glisse 0, vma, vma->vm_mm, address, 922a50b854eSMatthew Wilcox (Oracle) min(vma->vm_end, address + page_size(page))); 923ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 924369ea824SJérôme Glisse 925f27176cfSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 926f27176cfSKirill A. Shutemov int ret = 0; 927369ea824SJérôme Glisse 9281f18b296SYueHaibing address = pvmw.address; 929f27176cfSKirill A. Shutemov if (pvmw.pte) { 930c2fda5feSPeter Zijlstra pte_t entry; 931f27176cfSKirill A. Shutemov pte_t *pte = pvmw.pte; 932f27176cfSKirill A. Shutemov 933f27176cfSKirill A. Shutemov if (!pte_dirty(*pte) && !pte_write(*pte)) 934f27176cfSKirill A. Shutemov continue; 935d08b3851SPeter Zijlstra 936785373b4SLinus Torvalds flush_cache_page(vma, address, pte_pfn(*pte)); 937785373b4SLinus Torvalds entry = ptep_clear_flush(vma, address, pte); 938d08b3851SPeter Zijlstra entry = pte_wrprotect(entry); 939c2fda5feSPeter Zijlstra entry = pte_mkclean(entry); 940785373b4SLinus Torvalds set_pte_at(vma->vm_mm, address, pte, entry); 941d08b3851SPeter Zijlstra ret = 1; 942f27176cfSKirill A. Shutemov } else { 943396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_TRANSPARENT_HUGEPAGE 944f27176cfSKirill A. Shutemov pmd_t *pmd = pvmw.pmd; 945f27176cfSKirill A. Shutemov pmd_t entry; 946d08b3851SPeter Zijlstra 947f27176cfSKirill A. Shutemov if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) 948f27176cfSKirill A. Shutemov continue; 949f27176cfSKirill A. Shutemov 950785373b4SLinus Torvalds flush_cache_page(vma, address, page_to_pfn(page)); 951024eee0eSAneesh Kumar K.V entry = pmdp_invalidate(vma, address, pmd); 952f27176cfSKirill A. Shutemov entry = pmd_wrprotect(entry); 953f27176cfSKirill A. Shutemov entry = pmd_mkclean(entry); 954785373b4SLinus Torvalds set_pmd_at(vma->vm_mm, address, pmd, entry); 955f27176cfSKirill A. Shutemov ret = 1; 956f27176cfSKirill A. Shutemov #else 957f27176cfSKirill A. Shutemov /* unexpected pmd-mapped page? */ 958f27176cfSKirill A. Shutemov WARN_ON_ONCE(1); 959f27176cfSKirill A. Shutemov #endif 960f27176cfSKirill A. Shutemov } 9612ec74c3eSSagi Grimberg 9620f10851eSJérôme Glisse /* 9630f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() as we are 9640f10851eSJérôme Glisse * downgrading page table protection not changing it to point 9650f10851eSJérôme Glisse * to a new page. 9660f10851eSJérôme Glisse * 967ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 9680f10851eSJérôme Glisse */ 9690f10851eSJérôme Glisse if (ret) 9709853a407SJoonsoo Kim (*cleaned)++; 9719853a407SJoonsoo Kim } 972f27176cfSKirill A. Shutemov 973ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 974369ea824SJérôme Glisse 975e4b82222SMinchan Kim return true; 976d08b3851SPeter Zijlstra } 977d08b3851SPeter Zijlstra 9789853a407SJoonsoo Kim static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) 979d08b3851SPeter Zijlstra { 9809853a407SJoonsoo Kim if (vma->vm_flags & VM_SHARED) 981871beb8cSFengguang Wu return false; 982d08b3851SPeter Zijlstra 983871beb8cSFengguang Wu return true; 984d08b3851SPeter Zijlstra } 985d08b3851SPeter Zijlstra 986d08b3851SPeter Zijlstra int page_mkclean(struct page *page) 987d08b3851SPeter Zijlstra { 9889853a407SJoonsoo Kim int cleaned = 0; 9899853a407SJoonsoo Kim struct address_space *mapping; 9909853a407SJoonsoo Kim struct rmap_walk_control rwc = { 9919853a407SJoonsoo Kim .arg = (void *)&cleaned, 9929853a407SJoonsoo Kim .rmap_one = page_mkclean_one, 9939853a407SJoonsoo Kim .invalid_vma = invalid_mkclean_vma, 9949853a407SJoonsoo Kim }; 995d08b3851SPeter Zijlstra 996d08b3851SPeter Zijlstra BUG_ON(!PageLocked(page)); 997d08b3851SPeter Zijlstra 9989853a407SJoonsoo Kim if (!page_mapped(page)) 9999853a407SJoonsoo Kim return 0; 1000d08b3851SPeter Zijlstra 10019853a407SJoonsoo Kim mapping = page_mapping(page); 10029853a407SJoonsoo Kim if (!mapping) 10039853a407SJoonsoo Kim return 0; 10049853a407SJoonsoo Kim 10059853a407SJoonsoo Kim rmap_walk(page, &rwc); 10069853a407SJoonsoo Kim 10079853a407SJoonsoo Kim return cleaned; 1008d08b3851SPeter Zijlstra } 100960b59beaSJaya Kumar EXPORT_SYMBOL_GPL(page_mkclean); 1010d08b3851SPeter Zijlstra 10111da177e4SLinus Torvalds /** 1012c44b6743SRik van Riel * page_move_anon_rmap - move a page to our anon_vma 1013c44b6743SRik van Riel * @page: the page to move to our anon_vma 1014c44b6743SRik van Riel * @vma: the vma the page belongs to 1015c44b6743SRik van Riel * 1016c44b6743SRik van Riel * When a page belongs exclusively to one process after a COW event, 1017c44b6743SRik van Riel * that page can be moved into the anon_vma that belongs to just that 1018c44b6743SRik van Riel * process, so the rmap code will not search the parent or sibling 1019c44b6743SRik van Riel * processes. 1020c44b6743SRik van Riel */ 10215a49973dSHugh Dickins void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma) 1022c44b6743SRik van Riel { 1023c44b6743SRik van Riel struct anon_vma *anon_vma = vma->anon_vma; 1024c44b6743SRik van Riel 10255a49973dSHugh Dickins page = compound_head(page); 10265a49973dSHugh Dickins 1027309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 102881d1b09cSSasha Levin VM_BUG_ON_VMA(!anon_vma, vma); 1029c44b6743SRik van Riel 1030c44b6743SRik van Riel anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1031414e2fb8SVladimir Davydov /* 1032414e2fb8SVladimir Davydov * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written 1033414e2fb8SVladimir Davydov * simultaneously, so a concurrent reader (eg page_referenced()'s 1034414e2fb8SVladimir Davydov * PageAnon()) will not see one without the other. 1035414e2fb8SVladimir Davydov */ 1036414e2fb8SVladimir Davydov WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); 1037c44b6743SRik van Riel } 1038c44b6743SRik van Riel 1039c44b6743SRik van Riel /** 104043d8eac4SRandy Dunlap * __page_set_anon_rmap - set up new anonymous rmap 1041451b9514SKirill Tkhai * @page: Page or Hugepage to add to rmap 10424e1c1975SAndi Kleen * @vma: VM area to add page to. 10434e1c1975SAndi Kleen * @address: User virtual address of the mapping 1044e8a03febSRik van Riel * @exclusive: the page is exclusively owned by the current process 10451da177e4SLinus Torvalds */ 10469617d95eSNick Piggin static void __page_set_anon_rmap(struct page *page, 1047e8a03febSRik van Riel struct vm_area_struct *vma, unsigned long address, int exclusive) 10481da177e4SLinus Torvalds { 1049e8a03febSRik van Riel struct anon_vma *anon_vma = vma->anon_vma; 10502822c1aaSNick Piggin 1051e8a03febSRik van Riel BUG_ON(!anon_vma); 1052ea90002bSLinus Torvalds 10534e1c1975SAndi Kleen if (PageAnon(page)) 10544e1c1975SAndi Kleen return; 10554e1c1975SAndi Kleen 1056ea90002bSLinus Torvalds /* 1057e8a03febSRik van Riel * If the page isn't exclusively mapped into this vma, 1058e8a03febSRik van Riel * we must use the _oldest_ possible anon_vma for the 1059e8a03febSRik van Riel * page mapping! 1060ea90002bSLinus Torvalds */ 10614e1c1975SAndi Kleen if (!exclusive) 1062288468c3SAndrea Arcangeli anon_vma = anon_vma->root; 1063ea90002bSLinus Torvalds 106416f5e707SAlex Shi /* 106516f5e707SAlex Shi * page_idle does a lockless/optimistic rmap scan on page->mapping. 106616f5e707SAlex Shi * Make sure the compiler doesn't split the stores of anon_vma and 106716f5e707SAlex Shi * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code 106816f5e707SAlex Shi * could mistake the mapping for a struct address_space and crash. 106916f5e707SAlex Shi */ 10701da177e4SLinus Torvalds anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 107116f5e707SAlex Shi WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); 10724d7670e0SNick Piggin page->index = linear_page_index(vma, address); 10731da177e4SLinus Torvalds } 10749617d95eSNick Piggin 10759617d95eSNick Piggin /** 107643d8eac4SRandy Dunlap * __page_check_anon_rmap - sanity check anonymous rmap addition 1077c97a9e10SNick Piggin * @page: the page to add the mapping to 1078c97a9e10SNick Piggin * @vma: the vm area in which the mapping is added 1079c97a9e10SNick Piggin * @address: the user virtual address mapped 1080c97a9e10SNick Piggin */ 1081c97a9e10SNick Piggin static void __page_check_anon_rmap(struct page *page, 1082c97a9e10SNick Piggin struct vm_area_struct *vma, unsigned long address) 1083c97a9e10SNick Piggin { 1084c97a9e10SNick Piggin /* 1085c97a9e10SNick Piggin * The page's anon-rmap details (mapping and index) are guaranteed to 1086c97a9e10SNick Piggin * be set up correctly at this point. 1087c97a9e10SNick Piggin * 1088c97a9e10SNick Piggin * We have exclusion against page_add_anon_rmap because the caller 1089*90aaca85SMiaohe Lin * always holds the page locked. 1090c97a9e10SNick Piggin * 1091c97a9e10SNick Piggin * We have exclusion against page_add_new_anon_rmap because those pages 1092c97a9e10SNick Piggin * are initially only visible via the pagetables, and the pte is locked 1093c97a9e10SNick Piggin * over the call to page_add_new_anon_rmap. 1094c97a9e10SNick Piggin */ 109530c46382SYang Shi VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page); 109630c46382SYang Shi VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address), 109730c46382SYang Shi page); 1098c97a9e10SNick Piggin } 1099c97a9e10SNick Piggin 1100c97a9e10SNick Piggin /** 11019617d95eSNick Piggin * page_add_anon_rmap - add pte mapping to an anonymous page 11029617d95eSNick Piggin * @page: the page to add the mapping to 11039617d95eSNick Piggin * @vma: the vm area in which the mapping is added 11049617d95eSNick Piggin * @address: the user virtual address mapped 1105d281ee61SKirill A. Shutemov * @compound: charge the page as compound or small page 11069617d95eSNick Piggin * 11075ad64688SHugh Dickins * The caller needs to hold the pte lock, and the page must be locked in 110880e14822SHugh Dickins * the anon_vma case: to serialize mapping,index checking after setting, 110980e14822SHugh Dickins * and to ensure that PageAnon is not being upgraded racily to PageKsm 111080e14822SHugh Dickins * (but PageKsm is never downgraded to PageAnon). 11119617d95eSNick Piggin */ 11129617d95eSNick Piggin void page_add_anon_rmap(struct page *page, 1113d281ee61SKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, bool compound) 11149617d95eSNick Piggin { 1115d281ee61SKirill A. Shutemov do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0); 1116ad8c2ee8SRik van Riel } 1117ad8c2ee8SRik van Riel 1118ad8c2ee8SRik van Riel /* 1119ad8c2ee8SRik van Riel * Special version of the above for do_swap_page, which often runs 1120ad8c2ee8SRik van Riel * into pages that are exclusively owned by the current process. 1121ad8c2ee8SRik van Riel * Everybody else should continue to use page_add_anon_rmap above. 1122ad8c2ee8SRik van Riel */ 1123ad8c2ee8SRik van Riel void do_page_add_anon_rmap(struct page *page, 1124d281ee61SKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, int flags) 1125ad8c2ee8SRik van Riel { 1126d281ee61SKirill A. Shutemov bool compound = flags & RMAP_COMPOUND; 112753f9263bSKirill A. Shutemov bool first; 112853f9263bSKirill A. Shutemov 1129be5d0a74SJohannes Weiner if (unlikely(PageKsm(page))) 1130be5d0a74SJohannes Weiner lock_page_memcg(page); 1131be5d0a74SJohannes Weiner else 1132be5d0a74SJohannes Weiner VM_BUG_ON_PAGE(!PageLocked(page), page); 1133be5d0a74SJohannes Weiner 113453f9263bSKirill A. Shutemov if (compound) { 113553f9263bSKirill A. Shutemov atomic_t *mapcount; 1136e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 113753f9263bSKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 113853f9263bSKirill A. Shutemov mapcount = compound_mapcount_ptr(page); 113953f9263bSKirill A. Shutemov first = atomic_inc_and_test(mapcount); 114053f9263bSKirill A. Shutemov } else { 114153f9263bSKirill A. Shutemov first = atomic_inc_and_test(&page->_mapcount); 114253f9263bSKirill A. Shutemov } 114353f9263bSKirill A. Shutemov 114453f9263bSKirill A. Shutemov if (first) { 11456c357848SMatthew Wilcox (Oracle) int nr = compound ? thp_nr_pages(page) : 1; 1146bea04b07SJianyu Zhan /* 1147bea04b07SJianyu Zhan * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1148bea04b07SJianyu Zhan * these counters are not modified in interrupt context, and 1149bea04b07SJianyu Zhan * pte lock(a spinlock) is held, which implies preemption 1150bea04b07SJianyu Zhan * disabled. 1151bea04b07SJianyu Zhan */ 115265c45377SKirill A. Shutemov if (compound) 115369473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, nr); 1154be5d0a74SJohannes Weiner __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr); 115579134171SAndrea Arcangeli } 11565ad64688SHugh Dickins 1157be5d0a74SJohannes Weiner if (unlikely(PageKsm(page))) { 1158be5d0a74SJohannes Weiner unlock_page_memcg(page); 1159be5d0a74SJohannes Weiner return; 1160be5d0a74SJohannes Weiner } 116153f9263bSKirill A. Shutemov 11625dbe0af4SHugh Dickins /* address might be in next vma when migration races vma_adjust */ 11635ad64688SHugh Dickins if (first) 1164d281ee61SKirill A. Shutemov __page_set_anon_rmap(page, vma, address, 1165d281ee61SKirill A. Shutemov flags & RMAP_EXCLUSIVE); 116669029cd5SKAMEZAWA Hiroyuki else 1167c97a9e10SNick Piggin __page_check_anon_rmap(page, vma, address); 11681da177e4SLinus Torvalds } 11691da177e4SLinus Torvalds 117043d8eac4SRandy Dunlap /** 11719617d95eSNick Piggin * page_add_new_anon_rmap - add pte mapping to a new anonymous page 11729617d95eSNick Piggin * @page: the page to add the mapping to 11739617d95eSNick Piggin * @vma: the vm area in which the mapping is added 11749617d95eSNick Piggin * @address: the user virtual address mapped 1175d281ee61SKirill A. Shutemov * @compound: charge the page as compound or small page 11769617d95eSNick Piggin * 11779617d95eSNick Piggin * Same as page_add_anon_rmap but must only be called on *new* pages. 11789617d95eSNick Piggin * This means the inc-and-test can be bypassed. 1179c97a9e10SNick Piggin * Page does not have to be locked. 11809617d95eSNick Piggin */ 11819617d95eSNick Piggin void page_add_new_anon_rmap(struct page *page, 1182d281ee61SKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, bool compound) 11839617d95eSNick Piggin { 11846c357848SMatthew Wilcox (Oracle) int nr = compound ? thp_nr_pages(page) : 1; 1185d281ee61SKirill A. Shutemov 118681d1b09cSSasha Levin VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); 1187fa9949daSHugh Dickins __SetPageSwapBacked(page); 1188d281ee61SKirill A. Shutemov if (compound) { 1189d281ee61SKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 119053f9263bSKirill A. Shutemov /* increment count (starts at -1) */ 119153f9263bSKirill A. Shutemov atomic_set(compound_mapcount_ptr(page), 0); 119247e29d32SJohn Hubbard if (hpage_pincount_available(page)) 119347e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 119447e29d32SJohn Hubbard 119569473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, nr); 119653f9263bSKirill A. Shutemov } else { 119753f9263bSKirill A. Shutemov /* Anon THP always mapped first with PMD */ 119853f9263bSKirill A. Shutemov VM_BUG_ON_PAGE(PageTransCompound(page), page); 119953f9263bSKirill A. Shutemov /* increment count (starts at -1) */ 120053f9263bSKirill A. Shutemov atomic_set(&page->_mapcount, 0); 1201d281ee61SKirill A. Shutemov } 1202be5d0a74SJohannes Weiner __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr); 1203e8a03febSRik van Riel __page_set_anon_rmap(page, vma, address, 1); 12049617d95eSNick Piggin } 12059617d95eSNick Piggin 12061da177e4SLinus Torvalds /** 12071da177e4SLinus Torvalds * page_add_file_rmap - add pte mapping to a file page 12081da177e4SLinus Torvalds * @page: the page to add the mapping to 1209e8b098fcSMike Rapoport * @compound: charge the page as compound or small page 12101da177e4SLinus Torvalds * 1211b8072f09SHugh Dickins * The caller needs to hold the pte lock. 12121da177e4SLinus Torvalds */ 1213dd78feddSKirill A. Shutemov void page_add_file_rmap(struct page *page, bool compound) 12141da177e4SLinus Torvalds { 1215dd78feddSKirill A. Shutemov int i, nr = 1; 1216dd78feddSKirill A. Shutemov 1217dd78feddSKirill A. Shutemov VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page); 121862cccb8cSJohannes Weiner lock_page_memcg(page); 1219dd78feddSKirill A. Shutemov if (compound && PageTransHuge(page)) { 1220a1528e21SMuchun Song int nr_pages = thp_nr_pages(page); 1221a1528e21SMuchun Song 1222a1528e21SMuchun Song for (i = 0, nr = 0; i < nr_pages; i++) { 1223dd78feddSKirill A. Shutemov if (atomic_inc_and_test(&page[i]._mapcount)) 1224dd78feddSKirill A. Shutemov nr++; 1225d69b042fSBalbir Singh } 1226dd78feddSKirill A. Shutemov if (!atomic_inc_and_test(compound_mapcount_ptr(page))) 1227dd78feddSKirill A. Shutemov goto out; 122899cb0dbdSSong Liu if (PageSwapBacked(page)) 1229a1528e21SMuchun Song __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED, 1230a1528e21SMuchun Song nr_pages); 123199cb0dbdSSong Liu else 1232380780e7SMuchun Song __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED, 1233380780e7SMuchun Song nr_pages); 1234dd78feddSKirill A. Shutemov } else { 1235c8efc390SKirill A. Shutemov if (PageTransCompound(page) && page_mapping(page)) { 1236c8efc390SKirill A. Shutemov VM_WARN_ON_ONCE(!PageLocked(page)); 1237c8efc390SKirill A. Shutemov 12389a73f61bSKirill A. Shutemov SetPageDoubleMap(compound_head(page)); 12399a73f61bSKirill A. Shutemov if (PageMlocked(page)) 12409a73f61bSKirill A. Shutemov clear_page_mlock(compound_head(page)); 12419a73f61bSKirill A. Shutemov } 1242dd78feddSKirill A. Shutemov if (!atomic_inc_and_test(&page->_mapcount)) 1243dd78feddSKirill A. Shutemov goto out; 1244dd78feddSKirill A. Shutemov } 124500f3ca2cSJohannes Weiner __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr); 1246dd78feddSKirill A. Shutemov out: 124762cccb8cSJohannes Weiner unlock_page_memcg(page); 12481da177e4SLinus Torvalds } 12491da177e4SLinus Torvalds 1250dd78feddSKirill A. Shutemov static void page_remove_file_rmap(struct page *page, bool compound) 12518186eb6aSJohannes Weiner { 1252dd78feddSKirill A. Shutemov int i, nr = 1; 1253dd78feddSKirill A. Shutemov 125457dea93aSSteve Capper VM_BUG_ON_PAGE(compound && !PageHead(page), page); 12558186eb6aSJohannes Weiner 125653f9263bSKirill A. Shutemov /* Hugepages are not counted in NR_FILE_MAPPED for now. */ 125753f9263bSKirill A. Shutemov if (unlikely(PageHuge(page))) { 125853f9263bSKirill A. Shutemov /* hugetlb pages are always mapped with pmds */ 125953f9263bSKirill A. Shutemov atomic_dec(compound_mapcount_ptr(page)); 1260be5d0a74SJohannes Weiner return; 126153f9263bSKirill A. Shutemov } 126253f9263bSKirill A. Shutemov 12638186eb6aSJohannes Weiner /* page still mapped by someone else? */ 1264dd78feddSKirill A. Shutemov if (compound && PageTransHuge(page)) { 1265a1528e21SMuchun Song int nr_pages = thp_nr_pages(page); 1266a1528e21SMuchun Song 1267a1528e21SMuchun Song for (i = 0, nr = 0; i < nr_pages; i++) { 1268dd78feddSKirill A. Shutemov if (atomic_add_negative(-1, &page[i]._mapcount)) 1269dd78feddSKirill A. Shutemov nr++; 1270dd78feddSKirill A. Shutemov } 1271dd78feddSKirill A. Shutemov if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 1272be5d0a74SJohannes Weiner return; 127399cb0dbdSSong Liu if (PageSwapBacked(page)) 1274a1528e21SMuchun Song __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED, 1275a1528e21SMuchun Song -nr_pages); 127699cb0dbdSSong Liu else 1277380780e7SMuchun Song __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED, 1278380780e7SMuchun Song -nr_pages); 1279dd78feddSKirill A. Shutemov } else { 12808186eb6aSJohannes Weiner if (!atomic_add_negative(-1, &page->_mapcount)) 1281be5d0a74SJohannes Weiner return; 1282dd78feddSKirill A. Shutemov } 12838186eb6aSJohannes Weiner 12848186eb6aSJohannes Weiner /* 128500f3ca2cSJohannes Weiner * We use the irq-unsafe __{inc|mod}_lruvec_page_state because 12868186eb6aSJohannes Weiner * these counters are not modified in interrupt context, and 12878186eb6aSJohannes Weiner * pte lock(a spinlock) is held, which implies preemption disabled. 12888186eb6aSJohannes Weiner */ 128900f3ca2cSJohannes Weiner __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr); 12908186eb6aSJohannes Weiner 12918186eb6aSJohannes Weiner if (unlikely(PageMlocked(page))) 12928186eb6aSJohannes Weiner clear_page_mlock(page); 12938186eb6aSJohannes Weiner } 12948186eb6aSJohannes Weiner 129553f9263bSKirill A. Shutemov static void page_remove_anon_compound_rmap(struct page *page) 129653f9263bSKirill A. Shutemov { 129753f9263bSKirill A. Shutemov int i, nr; 129853f9263bSKirill A. Shutemov 129953f9263bSKirill A. Shutemov if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 130053f9263bSKirill A. Shutemov return; 130153f9263bSKirill A. Shutemov 130253f9263bSKirill A. Shutemov /* Hugepages are not counted in NR_ANON_PAGES for now. */ 130353f9263bSKirill A. Shutemov if (unlikely(PageHuge(page))) 130453f9263bSKirill A. Shutemov return; 130553f9263bSKirill A. Shutemov 130653f9263bSKirill A. Shutemov if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) 130753f9263bSKirill A. Shutemov return; 130853f9263bSKirill A. Shutemov 130969473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, -thp_nr_pages(page)); 131053f9263bSKirill A. Shutemov 131153f9263bSKirill A. Shutemov if (TestClearPageDoubleMap(page)) { 131253f9263bSKirill A. Shutemov /* 131353f9263bSKirill A. Shutemov * Subpages can be mapped with PTEs too. Check how many of 1314f1fe80d4SKirill A. Shutemov * them are still mapped. 131553f9263bSKirill A. Shutemov */ 13165eaf35abSMatthew Wilcox (Oracle) for (i = 0, nr = 0; i < thp_nr_pages(page); i++) { 131753f9263bSKirill A. Shutemov if (atomic_add_negative(-1, &page[i]._mapcount)) 131853f9263bSKirill A. Shutemov nr++; 131953f9263bSKirill A. Shutemov } 1320f1fe80d4SKirill A. Shutemov 1321f1fe80d4SKirill A. Shutemov /* 1322f1fe80d4SKirill A. Shutemov * Queue the page for deferred split if at least one small 1323f1fe80d4SKirill A. Shutemov * page of the compound page is unmapped, but at least one 1324f1fe80d4SKirill A. Shutemov * small page is still mapped. 1325f1fe80d4SKirill A. Shutemov */ 13265eaf35abSMatthew Wilcox (Oracle) if (nr && nr < thp_nr_pages(page)) 1327f1fe80d4SKirill A. Shutemov deferred_split_huge_page(page); 132853f9263bSKirill A. Shutemov } else { 13295eaf35abSMatthew Wilcox (Oracle) nr = thp_nr_pages(page); 133053f9263bSKirill A. Shutemov } 133153f9263bSKirill A. Shutemov 1332e90309c9SKirill A. Shutemov if (unlikely(PageMlocked(page))) 1333e90309c9SKirill A. Shutemov clear_page_mlock(page); 1334e90309c9SKirill A. Shutemov 1335f1fe80d4SKirill A. Shutemov if (nr) 1336be5d0a74SJohannes Weiner __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr); 133753f9263bSKirill A. Shutemov } 133853f9263bSKirill A. Shutemov 13391da177e4SLinus Torvalds /** 13401da177e4SLinus Torvalds * page_remove_rmap - take down pte mapping from a page 13411da177e4SLinus Torvalds * @page: page to remove mapping from 1342d281ee61SKirill A. Shutemov * @compound: uncharge the page as compound or small page 13431da177e4SLinus Torvalds * 1344b8072f09SHugh Dickins * The caller needs to hold the pte lock. 13451da177e4SLinus Torvalds */ 1346d281ee61SKirill A. Shutemov void page_remove_rmap(struct page *page, bool compound) 13471da177e4SLinus Torvalds { 1348be5d0a74SJohannes Weiner lock_page_memcg(page); 134989c06bd5SKAMEZAWA Hiroyuki 1350be5d0a74SJohannes Weiner if (!PageAnon(page)) { 1351be5d0a74SJohannes Weiner page_remove_file_rmap(page, compound); 1352be5d0a74SJohannes Weiner goto out; 1353be5d0a74SJohannes Weiner } 1354be5d0a74SJohannes Weiner 1355be5d0a74SJohannes Weiner if (compound) { 1356be5d0a74SJohannes Weiner page_remove_anon_compound_rmap(page); 1357be5d0a74SJohannes Weiner goto out; 1358be5d0a74SJohannes Weiner } 135953f9263bSKirill A. Shutemov 1360b904dcfeSKOSAKI Motohiro /* page still mapped by someone else? */ 1361b904dcfeSKOSAKI Motohiro if (!atomic_add_negative(-1, &page->_mapcount)) 1362be5d0a74SJohannes Weiner goto out; 13638186eb6aSJohannes Weiner 13641da177e4SLinus Torvalds /* 1365bea04b07SJianyu Zhan * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1366bea04b07SJianyu Zhan * these counters are not modified in interrupt context, and 1367bea04b07SJianyu Zhan * pte lock(a spinlock) is held, which implies preemption disabled. 13680fe6e20bSNaoya Horiguchi */ 1369be5d0a74SJohannes Weiner __dec_lruvec_page_state(page, NR_ANON_MAPPED); 13708186eb6aSJohannes Weiner 1371e6c509f8SHugh Dickins if (unlikely(PageMlocked(page))) 1372e6c509f8SHugh Dickins clear_page_mlock(page); 13738186eb6aSJohannes Weiner 13749a982250SKirill A. Shutemov if (PageTransCompound(page)) 13759a982250SKirill A. Shutemov deferred_split_huge_page(compound_head(page)); 13769a982250SKirill A. Shutemov 137716f8c5b2SHugh Dickins /* 13781da177e4SLinus Torvalds * It would be tidy to reset the PageAnon mapping here, 13791da177e4SLinus Torvalds * but that might overwrite a racing page_add_anon_rmap 13801da177e4SLinus Torvalds * which increments mapcount after us but sets mapping 13812d4894b5SMel Gorman * before us: so leave the reset to free_unref_page, 13821da177e4SLinus Torvalds * and remember that it's only reliable while mapped. 13831da177e4SLinus Torvalds * Leaving it set also helps swapoff to reinstate ptes 13841da177e4SLinus Torvalds * faster for those pages still in swapcache. 13851da177e4SLinus Torvalds */ 1386be5d0a74SJohannes Weiner out: 1387be5d0a74SJohannes Weiner unlock_page_memcg(page); 13881da177e4SLinus Torvalds } 13891da177e4SLinus Torvalds 13901da177e4SLinus Torvalds /* 139152629506SJoonsoo Kim * @arg: enum ttu_flags will be passed to this argument 13921da177e4SLinus Torvalds */ 1393e4b82222SMinchan Kim static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma, 139452629506SJoonsoo Kim unsigned long address, void *arg) 13951da177e4SLinus Torvalds { 13961da177e4SLinus Torvalds struct mm_struct *mm = vma->vm_mm; 1397c7ab0d2fSKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 1398c7ab0d2fSKirill A. Shutemov .page = page, 1399c7ab0d2fSKirill A. Shutemov .vma = vma, 1400c7ab0d2fSKirill A. Shutemov .address = address, 1401c7ab0d2fSKirill A. Shutemov }; 14021da177e4SLinus Torvalds pte_t pteval; 1403c7ab0d2fSKirill A. Shutemov struct page *subpage; 1404785373b4SLinus Torvalds bool ret = true; 1405ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 14064708f318SPalmer Dabbelt enum ttu_flags flags = (enum ttu_flags)(long)arg; 14071da177e4SLinus Torvalds 1408b87537d9SHugh Dickins /* munlock has nothing to gain from examining un-locked vmas */ 1409b87537d9SHugh Dickins if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED)) 1410e4b82222SMinchan Kim return true; 1411b87537d9SHugh Dickins 1412a5430ddaSJérôme Glisse if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) && 1413a5430ddaSJérôme Glisse is_zone_device_page(page) && !is_device_private_page(page)) 1414a5430ddaSJérôme Glisse return true; 1415a5430ddaSJérôme Glisse 1416fec89c10SKirill A. Shutemov if (flags & TTU_SPLIT_HUGE_PMD) { 1417fec89c10SKirill A. Shutemov split_huge_pmd_address(vma, address, 1418b5ff8161SNaoya Horiguchi flags & TTU_SPLIT_FREEZE, page); 1419fec89c10SKirill A. Shutemov } 1420fec89c10SKirill A. Shutemov 1421369ea824SJérôme Glisse /* 1422017b1660SMike Kravetz * For THP, we have to assume the worse case ie pmd for invalidation. 1423017b1660SMike Kravetz * For hugetlb, it could be much worse if we need to do pud 1424017b1660SMike Kravetz * invalidation in the case of pmd sharing. 1425017b1660SMike Kravetz * 1426017b1660SMike Kravetz * Note that the page can not be free in this function as call of 1427017b1660SMike Kravetz * try_to_unmap() must hold a reference on the page. 1428369ea824SJérôme Glisse */ 14297269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 14306f4f13e8SJérôme Glisse address, 1431a50b854eSMatthew Wilcox (Oracle) min(vma->vm_end, address + page_size(page))); 1432017b1660SMike Kravetz if (PageHuge(page)) { 1433017b1660SMike Kravetz /* 1434017b1660SMike Kravetz * If sharing is possible, start and end will be adjusted 1435017b1660SMike Kravetz * accordingly. 1436017b1660SMike Kravetz */ 1437ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, 1438ac46d4f3SJérôme Glisse &range.end); 1439017b1660SMike Kravetz } 1440ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1441369ea824SJérôme Glisse 1442c7ab0d2fSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 1443616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1444616b8371SZi Yan /* PMD-mapped THP migration entry */ 1445616b8371SZi Yan if (!pvmw.pte && (flags & TTU_MIGRATION)) { 1446616b8371SZi Yan VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); 1447616b8371SZi Yan 1448616b8371SZi Yan set_pmd_migration_entry(&pvmw, page); 1449616b8371SZi Yan continue; 1450616b8371SZi Yan } 1451616b8371SZi Yan #endif 1452616b8371SZi Yan 14531da177e4SLinus Torvalds /* 14541da177e4SLinus Torvalds * If the page is mlock()d, we cannot swap it out. 14551da177e4SLinus Torvalds * If it's recently referenced (perhaps page_referenced 14561da177e4SLinus Torvalds * skipped over this mm) then we should reactivate it. 14571da177e4SLinus Torvalds */ 145814fa31b8SAndi Kleen if (!(flags & TTU_IGNORE_MLOCK)) { 1459b87537d9SHugh Dickins if (vma->vm_flags & VM_LOCKED) { 14609a73f61bSKirill A. Shutemov /* PTE-mapped THP are never mlocked */ 14619a73f61bSKirill A. Shutemov if (!PageTransCompound(page)) { 14629a73f61bSKirill A. Shutemov /* 14639a73f61bSKirill A. Shutemov * Holding pte lock, we do *not* need 1464c1e8d7c6SMichel Lespinasse * mmap_lock here 14659a73f61bSKirill A. Shutemov */ 1466b87537d9SHugh Dickins mlock_vma_page(page); 14679a73f61bSKirill A. Shutemov } 1468e4b82222SMinchan Kim ret = false; 1469c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1470c7ab0d2fSKirill A. Shutemov break; 1471b87537d9SHugh Dickins } 1472daa5ba76SKonstantin Khlebnikov if (flags & TTU_MUNLOCK) 1473c7ab0d2fSKirill A. Shutemov continue; 147414fa31b8SAndi Kleen } 1475c7ab0d2fSKirill A. Shutemov 14768346242aSKirill A. Shutemov /* Unexpected PMD-mapped THP? */ 14778346242aSKirill A. Shutemov VM_BUG_ON_PAGE(!pvmw.pte, page); 14788346242aSKirill A. Shutemov 14798346242aSKirill A. Shutemov subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte); 1480785373b4SLinus Torvalds address = pvmw.address; 1481785373b4SLinus Torvalds 1482336bf30eSMike Kravetz if (PageHuge(page) && !PageAnon(page)) { 1483c0d0381aSMike Kravetz /* 1484c0d0381aSMike Kravetz * To call huge_pmd_unshare, i_mmap_rwsem must be 1485c0d0381aSMike Kravetz * held in write mode. Caller needs to explicitly 1486c0d0381aSMike Kravetz * do this outside rmap routines. 1487c0d0381aSMike Kravetz */ 1488c0d0381aSMike Kravetz VM_BUG_ON(!(flags & TTU_RMAP_LOCKED)); 148934ae204fSMike Kravetz if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) { 1490017b1660SMike Kravetz /* 1491017b1660SMike Kravetz * huge_pmd_unshare unmapped an entire PMD 1492017b1660SMike Kravetz * page. There is no way of knowing exactly 1493017b1660SMike Kravetz * which PMDs may be cached for this mm, so 1494017b1660SMike Kravetz * we must flush them all. start/end were 1495017b1660SMike Kravetz * already adjusted above to cover this range. 1496017b1660SMike Kravetz */ 1497ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 1498ac46d4f3SJérôme Glisse flush_tlb_range(vma, range.start, range.end); 1499ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, 1500ac46d4f3SJérôme Glisse range.end); 1501017b1660SMike Kravetz 1502017b1660SMike Kravetz /* 1503017b1660SMike Kravetz * The ref count of the PMD page was dropped 1504017b1660SMike Kravetz * which is part of the way map counting 1505017b1660SMike Kravetz * is done for shared PMDs. Return 'true' 1506017b1660SMike Kravetz * here. When there is no other sharing, 1507017b1660SMike Kravetz * huge_pmd_unshare returns false and we will 1508017b1660SMike Kravetz * unmap the actual page and drop map count 1509017b1660SMike Kravetz * to zero. 1510017b1660SMike Kravetz */ 1511017b1660SMike Kravetz page_vma_mapped_walk_done(&pvmw); 1512017b1660SMike Kravetz break; 1513017b1660SMike Kravetz } 1514017b1660SMike Kravetz } 15158346242aSKirill A. Shutemov 1516a5430ddaSJérôme Glisse if (IS_ENABLED(CONFIG_MIGRATION) && 1517a5430ddaSJérôme Glisse (flags & TTU_MIGRATION) && 1518a5430ddaSJérôme Glisse is_zone_device_page(page)) { 1519a5430ddaSJérôme Glisse swp_entry_t entry; 1520a5430ddaSJérôme Glisse pte_t swp_pte; 1521a5430ddaSJérôme Glisse 1522a5430ddaSJérôme Glisse pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte); 1523a5430ddaSJérôme Glisse 1524a5430ddaSJérôme Glisse /* 1525a5430ddaSJérôme Glisse * Store the pfn of the page in a special migration 1526a5430ddaSJérôme Glisse * pte. do_swap_page() will wait until the migration 1527a5430ddaSJérôme Glisse * pte is removed and then restart fault handling. 1528a5430ddaSJérôme Glisse */ 1529a5430ddaSJérôme Glisse entry = make_migration_entry(page, 0); 1530a5430ddaSJérôme Glisse swp_pte = swp_entry_to_pte(entry); 1531ad7df764SAlistair Popple 1532ad7df764SAlistair Popple /* 1533ad7df764SAlistair Popple * pteval maps a zone device page and is therefore 1534ad7df764SAlistair Popple * a swap pte. 1535ad7df764SAlistair Popple */ 1536ad7df764SAlistair Popple if (pte_swp_soft_dirty(pteval)) 1537a5430ddaSJérôme Glisse swp_pte = pte_swp_mksoft_dirty(swp_pte); 1538ad7df764SAlistair Popple if (pte_swp_uffd_wp(pteval)) 1539f45ec5ffSPeter Xu swp_pte = pte_swp_mkuffd_wp(swp_pte); 1540a5430ddaSJérôme Glisse set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte); 15410f10851eSJérôme Glisse /* 15420f10851eSJérôme Glisse * No need to invalidate here it will synchronize on 15430f10851eSJérôme Glisse * against the special swap migration pte. 15441de13ee5SRalph Campbell * 15451de13ee5SRalph Campbell * The assignment to subpage above was computed from a 15461de13ee5SRalph Campbell * swap PTE which results in an invalid pointer. 15471de13ee5SRalph Campbell * Since only PAGE_SIZE pages can currently be 15481de13ee5SRalph Campbell * migrated, just set it to page. This will need to be 15491de13ee5SRalph Campbell * changed when hugepage migrations to device private 15501de13ee5SRalph Campbell * memory are supported. 15510f10851eSJérôme Glisse */ 15521de13ee5SRalph Campbell subpage = page; 1553a5430ddaSJérôme Glisse goto discard; 1554a5430ddaSJérôme Glisse } 1555a5430ddaSJérôme Glisse 15561da177e4SLinus Torvalds /* Nuke the page table entry. */ 1557785373b4SLinus Torvalds flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); 155872b252aeSMel Gorman if (should_defer_flush(mm, flags)) { 155972b252aeSMel Gorman /* 1560c7ab0d2fSKirill A. Shutemov * We clear the PTE but do not flush so potentially 1561c7ab0d2fSKirill A. Shutemov * a remote CPU could still be writing to the page. 1562c7ab0d2fSKirill A. Shutemov * If the entry was previously clean then the 1563c7ab0d2fSKirill A. Shutemov * architecture must guarantee that a clear->dirty 1564c7ab0d2fSKirill A. Shutemov * transition on a cached TLB entry is written through 1565c7ab0d2fSKirill A. Shutemov * and traps if the PTE is unmapped. 156672b252aeSMel Gorman */ 1567785373b4SLinus Torvalds pteval = ptep_get_and_clear(mm, address, pvmw.pte); 156872b252aeSMel Gorman 1569c7ab0d2fSKirill A. Shutemov set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); 157072b252aeSMel Gorman } else { 1571785373b4SLinus Torvalds pteval = ptep_clear_flush(vma, address, pvmw.pte); 157272b252aeSMel Gorman } 15731da177e4SLinus Torvalds 1574c7ab0d2fSKirill A. Shutemov /* Move the dirty bit to the page. Now the pte is gone. */ 15751da177e4SLinus Torvalds if (pte_dirty(pteval)) 15761da177e4SLinus Torvalds set_page_dirty(page); 15771da177e4SLinus Torvalds 1578365e9c87SHugh Dickins /* Update high watermark before we lower rss */ 1579365e9c87SHugh Dickins update_hiwater_rss(mm); 1580365e9c87SHugh Dickins 1581888b9f7cSAndi Kleen if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { 15825fd27b8eSPunit Agrawal pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); 15835d317b2bSNaoya Horiguchi if (PageHuge(page)) { 1584d8c6546bSMatthew Wilcox (Oracle) hugetlb_count_sub(compound_nr(page), mm); 1585785373b4SLinus Torvalds set_huge_swap_pte_at(mm, address, 15865fd27b8eSPunit Agrawal pvmw.pte, pteval, 15875fd27b8eSPunit Agrawal vma_mmu_pagesize(vma)); 15885d317b2bSNaoya Horiguchi } else { 1589eca56ff9SJerome Marchand dec_mm_counter(mm, mm_counter(page)); 1590785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 15915f24ae58SNaoya Horiguchi } 1592c7ab0d2fSKirill A. Shutemov 1593bce73e48SChristian Borntraeger } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { 159445961722SKonstantin Weitz /* 159545961722SKonstantin Weitz * The guest indicated that the page content is of no 159645961722SKonstantin Weitz * interest anymore. Simply discard the pte, vmscan 159745961722SKonstantin Weitz * will take care of the rest. 1598bce73e48SChristian Borntraeger * A future reference will then fault in a new zero 1599bce73e48SChristian Borntraeger * page. When userfaultfd is active, we must not drop 1600bce73e48SChristian Borntraeger * this page though, as its main user (postcopy 1601bce73e48SChristian Borntraeger * migration) will not expect userfaults on already 1602bce73e48SChristian Borntraeger * copied pages. 160345961722SKonstantin Weitz */ 1604eca56ff9SJerome Marchand dec_mm_counter(mm, mm_counter(page)); 16050f10851eSJérôme Glisse /* We have to invalidate as we cleared the pte */ 16060f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, address, 16070f10851eSJérôme Glisse address + PAGE_SIZE); 1608c7ab0d2fSKirill A. Shutemov } else if (IS_ENABLED(CONFIG_MIGRATION) && 1609b5ff8161SNaoya Horiguchi (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) { 1610470f119fSHugh Dickins swp_entry_t entry; 1611470f119fSHugh Dickins pte_t swp_pte; 1612ca827d55SKhalid Aziz 1613ca827d55SKhalid Aziz if (arch_unmap_one(mm, vma, address, pteval) < 0) { 1614ca827d55SKhalid Aziz set_pte_at(mm, address, pvmw.pte, pteval); 1615ca827d55SKhalid Aziz ret = false; 1616ca827d55SKhalid Aziz page_vma_mapped_walk_done(&pvmw); 1617ca827d55SKhalid Aziz break; 1618ca827d55SKhalid Aziz } 1619ca827d55SKhalid Aziz 1620470f119fSHugh Dickins /* 1621470f119fSHugh Dickins * Store the pfn of the page in a special migration 1622470f119fSHugh Dickins * pte. do_swap_page() will wait until the migration 1623470f119fSHugh Dickins * pte is removed and then restart fault handling. 1624470f119fSHugh Dickins */ 1625c7ab0d2fSKirill A. Shutemov entry = make_migration_entry(subpage, 1626c7ab0d2fSKirill A. Shutemov pte_write(pteval)); 1627470f119fSHugh Dickins swp_pte = swp_entry_to_pte(entry); 1628470f119fSHugh Dickins if (pte_soft_dirty(pteval)) 1629470f119fSHugh Dickins swp_pte = pte_swp_mksoft_dirty(swp_pte); 1630f45ec5ffSPeter Xu if (pte_uffd_wp(pteval)) 1631f45ec5ffSPeter Xu swp_pte = pte_swp_mkuffd_wp(swp_pte); 1632785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, swp_pte); 16330f10851eSJérôme Glisse /* 16340f10851eSJérôme Glisse * No need to invalidate here it will synchronize on 16350f10851eSJérôme Glisse * against the special swap migration pte. 16360f10851eSJérôme Glisse */ 1637888b9f7cSAndi Kleen } else if (PageAnon(page)) { 1638c7ab0d2fSKirill A. Shutemov swp_entry_t entry = { .val = page_private(subpage) }; 1639179ef71cSCyrill Gorcunov pte_t swp_pte; 16401da177e4SLinus Torvalds /* 16411da177e4SLinus Torvalds * Store the swap location in the pte. 16421da177e4SLinus Torvalds * See handle_pte_fault() ... 16431da177e4SLinus Torvalds */ 1644eb94a878SMinchan Kim if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) { 1645eb94a878SMinchan Kim WARN_ON_ONCE(1); 164683612a94SMinchan Kim ret = false; 1647369ea824SJérôme Glisse /* We have to invalidate as we cleared the pte */ 16480f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, address, 16490f10851eSJérôme Glisse address + PAGE_SIZE); 1650eb94a878SMinchan Kim page_vma_mapped_walk_done(&pvmw); 1651eb94a878SMinchan Kim break; 1652eb94a878SMinchan Kim } 1653854e9ed0SMinchan Kim 1654802a3a92SShaohua Li /* MADV_FREE page check */ 1655802a3a92SShaohua Li if (!PageSwapBacked(page)) { 1656a128ca71SShaohua Li if (!PageDirty(page)) { 16570f10851eSJérôme Glisse /* Invalidate as we cleared the pte */ 16580f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, 16590f10851eSJérôme Glisse address, address + PAGE_SIZE); 1660854e9ed0SMinchan Kim dec_mm_counter(mm, MM_ANONPAGES); 1661854e9ed0SMinchan Kim goto discard; 1662854e9ed0SMinchan Kim } 1663854e9ed0SMinchan Kim 1664802a3a92SShaohua Li /* 1665802a3a92SShaohua Li * If the page was redirtied, it cannot be 1666802a3a92SShaohua Li * discarded. Remap the page to page table. 1667802a3a92SShaohua Li */ 1668785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 166918863d3aSMinchan Kim SetPageSwapBacked(page); 1670e4b82222SMinchan Kim ret = false; 1671802a3a92SShaohua Li page_vma_mapped_walk_done(&pvmw); 1672802a3a92SShaohua Li break; 1673802a3a92SShaohua Li } 1674802a3a92SShaohua Li 1675570a335bSHugh Dickins if (swap_duplicate(entry) < 0) { 1676785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 1677e4b82222SMinchan Kim ret = false; 1678c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1679c7ab0d2fSKirill A. Shutemov break; 1680570a335bSHugh Dickins } 1681ca827d55SKhalid Aziz if (arch_unmap_one(mm, vma, address, pteval) < 0) { 1682ca827d55SKhalid Aziz set_pte_at(mm, address, pvmw.pte, pteval); 1683ca827d55SKhalid Aziz ret = false; 1684ca827d55SKhalid Aziz page_vma_mapped_walk_done(&pvmw); 1685ca827d55SKhalid Aziz break; 1686ca827d55SKhalid Aziz } 16871da177e4SLinus Torvalds if (list_empty(&mm->mmlist)) { 16881da177e4SLinus Torvalds spin_lock(&mmlist_lock); 1689f412ac08SHugh Dickins if (list_empty(&mm->mmlist)) 16901da177e4SLinus Torvalds list_add(&mm->mmlist, &init_mm.mmlist); 16911da177e4SLinus Torvalds spin_unlock(&mmlist_lock); 16921da177e4SLinus Torvalds } 1693d559db08SKAMEZAWA Hiroyuki dec_mm_counter(mm, MM_ANONPAGES); 1694b084d435SKAMEZAWA Hiroyuki inc_mm_counter(mm, MM_SWAPENTS); 1695179ef71cSCyrill Gorcunov swp_pte = swp_entry_to_pte(entry); 1696179ef71cSCyrill Gorcunov if (pte_soft_dirty(pteval)) 1697179ef71cSCyrill Gorcunov swp_pte = pte_swp_mksoft_dirty(swp_pte); 1698f45ec5ffSPeter Xu if (pte_uffd_wp(pteval)) 1699f45ec5ffSPeter Xu swp_pte = pte_swp_mkuffd_wp(swp_pte); 1700785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, swp_pte); 17010f10851eSJérôme Glisse /* Invalidate as we cleared the pte */ 1702369ea824SJérôme Glisse mmu_notifier_invalidate_range(mm, address, 1703369ea824SJérôme Glisse address + PAGE_SIZE); 17040f10851eSJérôme Glisse } else { 17050f10851eSJérôme Glisse /* 1706906f9cdfSHugh Dickins * This is a locked file-backed page, thus it cannot 1707906f9cdfSHugh Dickins * be removed from the page cache and replaced by a new 1708906f9cdfSHugh Dickins * page before mmu_notifier_invalidate_range_end, so no 17090f10851eSJérôme Glisse * concurrent thread might update its page table to 17100f10851eSJérôme Glisse * point at new page while a device still is using this 17110f10851eSJérôme Glisse * page. 17120f10851eSJérôme Glisse * 1713ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 17140f10851eSJérôme Glisse */ 17150f10851eSJérôme Glisse dec_mm_counter(mm, mm_counter_file(page)); 17160f10851eSJérôme Glisse } 17170f10851eSJérôme Glisse discard: 17180f10851eSJérôme Glisse /* 17190f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() it has be 17200f10851eSJérôme Glisse * done above for all cases requiring it to happen under page 17210f10851eSJérôme Glisse * table lock before mmu_notifier_invalidate_range_end() 17220f10851eSJérôme Glisse * 1723ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 17240f10851eSJérôme Glisse */ 17250f10851eSJérôme Glisse page_remove_rmap(subpage, PageHuge(page)); 17260f10851eSJérôme Glisse put_page(page); 1727c7ab0d2fSKirill A. Shutemov } 1728369ea824SJérôme Glisse 1729ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1730369ea824SJérôme Glisse 1731caed0f48SKOSAKI Motohiro return ret; 17321da177e4SLinus Torvalds } 17331da177e4SLinus Torvalds 173452629506SJoonsoo Kim static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) 173552629506SJoonsoo Kim { 1736222100eeSAnshuman Khandual return vma_is_temporary_stack(vma); 173752629506SJoonsoo Kim } 173852629506SJoonsoo Kim 17392a52bcbcSKirill A. Shutemov static int page_mapcount_is_zero(struct page *page) 174052629506SJoonsoo Kim { 1741c7ab0d2fSKirill A. Shutemov return !total_mapcount(page); 17422a52bcbcSKirill A. Shutemov } 174352629506SJoonsoo Kim 17441da177e4SLinus Torvalds /** 17451da177e4SLinus Torvalds * try_to_unmap - try to remove all page table mappings to a page 17461da177e4SLinus Torvalds * @page: the page to get unmapped 174714fa31b8SAndi Kleen * @flags: action and flags 17481da177e4SLinus Torvalds * 17491da177e4SLinus Torvalds * Tries to remove all the page table entries which are mapping this 17501da177e4SLinus Torvalds * page, used in the pageout path. Caller must hold the page lock. 17511da177e4SLinus Torvalds * 1752666e5a40SMinchan Kim * If unmap is successful, return true. Otherwise, false. 17531da177e4SLinus Torvalds */ 1754666e5a40SMinchan Kim bool try_to_unmap(struct page *page, enum ttu_flags flags) 17551da177e4SLinus Torvalds { 175652629506SJoonsoo Kim struct rmap_walk_control rwc = { 175752629506SJoonsoo Kim .rmap_one = try_to_unmap_one, 1758802a3a92SShaohua Li .arg = (void *)flags, 17592a52bcbcSKirill A. Shutemov .done = page_mapcount_is_zero, 176052629506SJoonsoo Kim .anon_lock = page_lock_anon_vma_read, 176152629506SJoonsoo Kim }; 17621da177e4SLinus Torvalds 176352629506SJoonsoo Kim /* 176452629506SJoonsoo Kim * During exec, a temporary VMA is setup and later moved. 176552629506SJoonsoo Kim * The VMA is moved under the anon_vma lock but not the 176652629506SJoonsoo Kim * page tables leading to a race where migration cannot 176752629506SJoonsoo Kim * find the migration ptes. Rather than increasing the 176852629506SJoonsoo Kim * locking requirements of exec(), migration skips 176952629506SJoonsoo Kim * temporary VMAs until after exec() completes. 177052629506SJoonsoo Kim */ 1771b5ff8161SNaoya Horiguchi if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE)) 1772b5ff8161SNaoya Horiguchi && !PageKsm(page) && PageAnon(page)) 177352629506SJoonsoo Kim rwc.invalid_vma = invalid_migration_vma; 177452629506SJoonsoo Kim 17752a52bcbcSKirill A. Shutemov if (flags & TTU_RMAP_LOCKED) 177633fc80e2SMinchan Kim rmap_walk_locked(page, &rwc); 17772a52bcbcSKirill A. Shutemov else 177833fc80e2SMinchan Kim rmap_walk(page, &rwc); 177952629506SJoonsoo Kim 1780666e5a40SMinchan Kim return !page_mapcount(page) ? true : false; 17811da177e4SLinus Torvalds } 178281b4082dSNikita Danilov 17832a52bcbcSKirill A. Shutemov static int page_not_mapped(struct page *page) 17842a52bcbcSKirill A. Shutemov { 17852a52bcbcSKirill A. Shutemov return !page_mapped(page); 1786e0af87ffSMiaohe Lin } 17872a52bcbcSKirill A. Shutemov 1788b291f000SNick Piggin /** 1789b291f000SNick Piggin * try_to_munlock - try to munlock a page 1790b291f000SNick Piggin * @page: the page to be munlocked 1791b291f000SNick Piggin * 1792b291f000SNick Piggin * Called from munlock code. Checks all of the VMAs mapping the page 1793b291f000SNick Piggin * to make sure nobody else has this page mlocked. The page will be 1794b291f000SNick Piggin * returned with PG_mlocked cleared if no other vmas have it mlocked. 1795b291f000SNick Piggin */ 1796854e9ed0SMinchan Kim 1797192d7232SMinchan Kim void try_to_munlock(struct page *page) 1798192d7232SMinchan Kim { 1799e8351ac9SJoonsoo Kim struct rmap_walk_control rwc = { 1800e8351ac9SJoonsoo Kim .rmap_one = try_to_unmap_one, 1801802a3a92SShaohua Li .arg = (void *)TTU_MUNLOCK, 1802e8351ac9SJoonsoo Kim .done = page_not_mapped, 1803e8351ac9SJoonsoo Kim .anon_lock = page_lock_anon_vma_read, 1804e8351ac9SJoonsoo Kim 1805e8351ac9SJoonsoo Kim }; 1806e8351ac9SJoonsoo Kim 1807309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page); 1808192d7232SMinchan Kim VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page); 1809b291f000SNick Piggin 1810192d7232SMinchan Kim rmap_walk(page, &rwc); 1811b291f000SNick Piggin } 1812e9995ef9SHugh Dickins 181301d8b20dSPeter Zijlstra void __put_anon_vma(struct anon_vma *anon_vma) 181476545066SRik van Riel { 181576545066SRik van Riel struct anon_vma *root = anon_vma->root; 181676545066SRik van Riel 1817624483f3SAndrey Ryabinin anon_vma_free(anon_vma); 181801d8b20dSPeter Zijlstra if (root != anon_vma && atomic_dec_and_test(&root->refcount)) 181976545066SRik van Riel anon_vma_free(root); 182076545066SRik van Riel } 182176545066SRik van Riel 18220dd1c7bbSJoonsoo Kim static struct anon_vma *rmap_walk_anon_lock(struct page *page, 18230dd1c7bbSJoonsoo Kim struct rmap_walk_control *rwc) 1824faecd8ddSJoonsoo Kim { 1825faecd8ddSJoonsoo Kim struct anon_vma *anon_vma; 1826faecd8ddSJoonsoo Kim 18270dd1c7bbSJoonsoo Kim if (rwc->anon_lock) 18280dd1c7bbSJoonsoo Kim return rwc->anon_lock(page); 18290dd1c7bbSJoonsoo Kim 1830faecd8ddSJoonsoo Kim /* 1831faecd8ddSJoonsoo Kim * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() 1832faecd8ddSJoonsoo Kim * because that depends on page_mapped(); but not all its usages 1833c1e8d7c6SMichel Lespinasse * are holding mmap_lock. Users without mmap_lock are required to 1834faecd8ddSJoonsoo Kim * take a reference count to prevent the anon_vma disappearing 1835faecd8ddSJoonsoo Kim */ 1836faecd8ddSJoonsoo Kim anon_vma = page_anon_vma(page); 1837faecd8ddSJoonsoo Kim if (!anon_vma) 1838faecd8ddSJoonsoo Kim return NULL; 1839faecd8ddSJoonsoo Kim 1840faecd8ddSJoonsoo Kim anon_vma_lock_read(anon_vma); 1841faecd8ddSJoonsoo Kim return anon_vma; 1842faecd8ddSJoonsoo Kim } 1843faecd8ddSJoonsoo Kim 1844e9995ef9SHugh Dickins /* 1845e8351ac9SJoonsoo Kim * rmap_walk_anon - do something to anonymous page using the object-based 1846e8351ac9SJoonsoo Kim * rmap method 1847e8351ac9SJoonsoo Kim * @page: the page to be handled 1848e8351ac9SJoonsoo Kim * @rwc: control variable according to each walk type 1849e8351ac9SJoonsoo Kim * 1850e8351ac9SJoonsoo Kim * Find all the mappings of a page using the mapping pointer and the vma chains 1851e8351ac9SJoonsoo Kim * contained in the anon_vma struct it points to. 1852e8351ac9SJoonsoo Kim * 1853c1e8d7c6SMichel Lespinasse * When called from try_to_munlock(), the mmap_lock of the mm containing the vma 1854e8351ac9SJoonsoo Kim * where the page was found will be held for write. So, we won't recheck 1855e8351ac9SJoonsoo Kim * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1856e8351ac9SJoonsoo Kim * LOCKED. 1857e9995ef9SHugh Dickins */ 18581df631aeSMinchan Kim static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc, 1859b9773199SKirill A. Shutemov bool locked) 1860e9995ef9SHugh Dickins { 1861e9995ef9SHugh Dickins struct anon_vma *anon_vma; 1862a8fa41adSKirill A. Shutemov pgoff_t pgoff_start, pgoff_end; 18635beb4930SRik van Riel struct anon_vma_chain *avc; 1864e9995ef9SHugh Dickins 1865b9773199SKirill A. Shutemov if (locked) { 1866b9773199SKirill A. Shutemov anon_vma = page_anon_vma(page); 1867b9773199SKirill A. Shutemov /* anon_vma disappear under us? */ 1868b9773199SKirill A. Shutemov VM_BUG_ON_PAGE(!anon_vma, page); 1869b9773199SKirill A. Shutemov } else { 18700dd1c7bbSJoonsoo Kim anon_vma = rmap_walk_anon_lock(page, rwc); 1871b9773199SKirill A. Shutemov } 1872e9995ef9SHugh Dickins if (!anon_vma) 18731df631aeSMinchan Kim return; 1874faecd8ddSJoonsoo Kim 1875a8fa41adSKirill A. Shutemov pgoff_start = page_to_pgoff(page); 18766c357848SMatthew Wilcox (Oracle) pgoff_end = pgoff_start + thp_nr_pages(page) - 1; 1877a8fa41adSKirill A. Shutemov anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 1878a8fa41adSKirill A. Shutemov pgoff_start, pgoff_end) { 18795beb4930SRik van Riel struct vm_area_struct *vma = avc->vma; 1880e9995ef9SHugh Dickins unsigned long address = vma_address(page, vma); 18810dd1c7bbSJoonsoo Kim 1882ad12695fSAndrea Arcangeli cond_resched(); 1883ad12695fSAndrea Arcangeli 18840dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 18850dd1c7bbSJoonsoo Kim continue; 18860dd1c7bbSJoonsoo Kim 1887e4b82222SMinchan Kim if (!rwc->rmap_one(page, vma, address, rwc->arg)) 1888e9995ef9SHugh Dickins break; 18890dd1c7bbSJoonsoo Kim if (rwc->done && rwc->done(page)) 18900dd1c7bbSJoonsoo Kim break; 1891e9995ef9SHugh Dickins } 1892b9773199SKirill A. Shutemov 1893b9773199SKirill A. Shutemov if (!locked) 18944fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 1895e9995ef9SHugh Dickins } 1896e9995ef9SHugh Dickins 1897e8351ac9SJoonsoo Kim /* 1898e8351ac9SJoonsoo Kim * rmap_walk_file - do something to file page using the object-based rmap method 1899e8351ac9SJoonsoo Kim * @page: the page to be handled 1900e8351ac9SJoonsoo Kim * @rwc: control variable according to each walk type 1901e8351ac9SJoonsoo Kim * 1902e8351ac9SJoonsoo Kim * Find all the mappings of a page using the mapping pointer and the vma chains 1903e8351ac9SJoonsoo Kim * contained in the address_space struct it points to. 1904e8351ac9SJoonsoo Kim * 1905c1e8d7c6SMichel Lespinasse * When called from try_to_munlock(), the mmap_lock of the mm containing the vma 1906e8351ac9SJoonsoo Kim * where the page was found will be held for write. So, we won't recheck 1907e8351ac9SJoonsoo Kim * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1908e8351ac9SJoonsoo Kim * LOCKED. 1909e8351ac9SJoonsoo Kim */ 19101df631aeSMinchan Kim static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc, 1911b9773199SKirill A. Shutemov bool locked) 1912e9995ef9SHugh Dickins { 1913b9773199SKirill A. Shutemov struct address_space *mapping = page_mapping(page); 1914a8fa41adSKirill A. Shutemov pgoff_t pgoff_start, pgoff_end; 1915e9995ef9SHugh Dickins struct vm_area_struct *vma; 1916e9995ef9SHugh Dickins 19179f32624bSJoonsoo Kim /* 19189f32624bSJoonsoo Kim * The page lock not only makes sure that page->mapping cannot 19199f32624bSJoonsoo Kim * suddenly be NULLified by truncation, it makes sure that the 19209f32624bSJoonsoo Kim * structure at mapping cannot be freed and reused yet, 1921c8c06efaSDavidlohr Bueso * so we can safely take mapping->i_mmap_rwsem. 19229f32624bSJoonsoo Kim */ 192381d1b09cSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 19249f32624bSJoonsoo Kim 1925e9995ef9SHugh Dickins if (!mapping) 19261df631aeSMinchan Kim return; 19273dec0ba0SDavidlohr Bueso 1928a8fa41adSKirill A. Shutemov pgoff_start = page_to_pgoff(page); 19296c357848SMatthew Wilcox (Oracle) pgoff_end = pgoff_start + thp_nr_pages(page) - 1; 1930b9773199SKirill A. Shutemov if (!locked) 19313dec0ba0SDavidlohr Bueso i_mmap_lock_read(mapping); 1932a8fa41adSKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, 1933a8fa41adSKirill A. Shutemov pgoff_start, pgoff_end) { 1934e9995ef9SHugh Dickins unsigned long address = vma_address(page, vma); 19350dd1c7bbSJoonsoo Kim 1936ad12695fSAndrea Arcangeli cond_resched(); 1937ad12695fSAndrea Arcangeli 19380dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 19390dd1c7bbSJoonsoo Kim continue; 19400dd1c7bbSJoonsoo Kim 1941e4b82222SMinchan Kim if (!rwc->rmap_one(page, vma, address, rwc->arg)) 19420dd1c7bbSJoonsoo Kim goto done; 19430dd1c7bbSJoonsoo Kim if (rwc->done && rwc->done(page)) 19440dd1c7bbSJoonsoo Kim goto done; 1945e9995ef9SHugh Dickins } 19460dd1c7bbSJoonsoo Kim 19470dd1c7bbSJoonsoo Kim done: 1948b9773199SKirill A. Shutemov if (!locked) 19493dec0ba0SDavidlohr Bueso i_mmap_unlock_read(mapping); 1950e9995ef9SHugh Dickins } 1951e9995ef9SHugh Dickins 19521df631aeSMinchan Kim void rmap_walk(struct page *page, struct rmap_walk_control *rwc) 1953e9995ef9SHugh Dickins { 1954e9995ef9SHugh Dickins if (unlikely(PageKsm(page))) 19551df631aeSMinchan Kim rmap_walk_ksm(page, rwc); 1956e9995ef9SHugh Dickins else if (PageAnon(page)) 19571df631aeSMinchan Kim rmap_walk_anon(page, rwc, false); 1958e9995ef9SHugh Dickins else 19591df631aeSMinchan Kim rmap_walk_file(page, rwc, false); 1960b9773199SKirill A. Shutemov } 1961b9773199SKirill A. Shutemov 1962b9773199SKirill A. Shutemov /* Like rmap_walk, but caller holds relevant rmap lock */ 19631df631aeSMinchan Kim void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc) 1964b9773199SKirill A. Shutemov { 1965b9773199SKirill A. Shutemov /* no ksm support for now */ 1966b9773199SKirill A. Shutemov VM_BUG_ON_PAGE(PageKsm(page), page); 1967b9773199SKirill A. Shutemov if (PageAnon(page)) 19681df631aeSMinchan Kim rmap_walk_anon(page, rwc, true); 1969b9773199SKirill A. Shutemov else 19701df631aeSMinchan Kim rmap_walk_file(page, rwc, true); 1971e9995ef9SHugh Dickins } 19720fe6e20bSNaoya Horiguchi 1973e3390f67SNaoya Horiguchi #ifdef CONFIG_HUGETLB_PAGE 19740fe6e20bSNaoya Horiguchi /* 1975451b9514SKirill Tkhai * The following two functions are for anonymous (private mapped) hugepages. 19760fe6e20bSNaoya Horiguchi * Unlike common anonymous pages, anonymous hugepages have no accounting code 19770fe6e20bSNaoya Horiguchi * and no lru code, because we handle hugepages differently from common pages. 19780fe6e20bSNaoya Horiguchi */ 19790fe6e20bSNaoya Horiguchi void hugepage_add_anon_rmap(struct page *page, 19800fe6e20bSNaoya Horiguchi struct vm_area_struct *vma, unsigned long address) 19810fe6e20bSNaoya Horiguchi { 19820fe6e20bSNaoya Horiguchi struct anon_vma *anon_vma = vma->anon_vma; 19830fe6e20bSNaoya Horiguchi int first; 1984a850ea30SNaoya Horiguchi 1985a850ea30SNaoya Horiguchi BUG_ON(!PageLocked(page)); 19860fe6e20bSNaoya Horiguchi BUG_ON(!anon_vma); 19875dbe0af4SHugh Dickins /* address might be in next vma when migration races vma_adjust */ 198853f9263bSKirill A. Shutemov first = atomic_inc_and_test(compound_mapcount_ptr(page)); 19890fe6e20bSNaoya Horiguchi if (first) 1990451b9514SKirill Tkhai __page_set_anon_rmap(page, vma, address, 0); 19910fe6e20bSNaoya Horiguchi } 19920fe6e20bSNaoya Horiguchi 19930fe6e20bSNaoya Horiguchi void hugepage_add_new_anon_rmap(struct page *page, 19940fe6e20bSNaoya Horiguchi struct vm_area_struct *vma, unsigned long address) 19950fe6e20bSNaoya Horiguchi { 19960fe6e20bSNaoya Horiguchi BUG_ON(address < vma->vm_start || address >= vma->vm_end); 199753f9263bSKirill A. Shutemov atomic_set(compound_mapcount_ptr(page), 0); 199847e29d32SJohn Hubbard if (hpage_pincount_available(page)) 199947e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 200047e29d32SJohn Hubbard 2001451b9514SKirill Tkhai __page_set_anon_rmap(page, vma, address, 1); 20020fe6e20bSNaoya Horiguchi } 2003e3390f67SNaoya Horiguchi #endif /* CONFIG_HUGETLB_PAGE */ 2004