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) 241da177e4SLinus Torvalds * mm->mmap_sem 251da177e4SLinus Torvalds * page->flags PG_locked (lock_page) 2688f306b6SKirill A. Shutemov * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share) 27c8c06efaSDavidlohr Bueso * mapping->i_mmap_rwsem 285a505085SIngo Molnar * anon_vma->rwsem 29b8072f09SHugh Dickins * mm->page_table_lock or pte_lock 30f4b7e272SAndrey Ryabinin * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page) 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) 34c4843a75SGreg Thelen * mem_cgroup_{begin,end}_page_stat (memcg->move_lock) 35b93b0163SMatthew Wilcox * i_pages lock (widely used) 36250df6edSDave Chinner * inode->i_lock (in set_page_dirty's __mark_inode_dirty) 37f758eeabSChristoph Hellwig * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) 381da177e4SLinus Torvalds * sb_lock (within inode_lock in fs/fs-writeback.c) 39b93b0163SMatthew Wilcox * i_pages lock (widely used, in set_page_dirty, 401da177e4SLinus Torvalds * in arch-dependent flush_dcache_mmap_lock, 41f758eeabSChristoph Hellwig * within bdi.wb->list_lock in __sync_single_inode) 426a46079cSAndi Kleen * 435a505085SIngo Molnar * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) 446a46079cSAndi Kleen * ->tasklist_lock 456a46079cSAndi Kleen * pte map lock 461da177e4SLinus Torvalds */ 471da177e4SLinus Torvalds 481da177e4SLinus Torvalds #include <linux/mm.h> 496e84f315SIngo Molnar #include <linux/sched/mm.h> 5029930025SIngo Molnar #include <linux/sched/task.h> 511da177e4SLinus Torvalds #include <linux/pagemap.h> 521da177e4SLinus Torvalds #include <linux/swap.h> 531da177e4SLinus Torvalds #include <linux/swapops.h> 541da177e4SLinus Torvalds #include <linux/slab.h> 551da177e4SLinus Torvalds #include <linux/init.h> 565ad64688SHugh Dickins #include <linux/ksm.h> 571da177e4SLinus Torvalds #include <linux/rmap.h> 581da177e4SLinus Torvalds #include <linux/rcupdate.h> 59b95f1b31SPaul Gortmaker #include <linux/export.h> 608a9f3ccdSBalbir Singh #include <linux/memcontrol.h> 61cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 6264cdd548SKOSAKI Motohiro #include <linux/migrate.h> 630fe6e20bSNaoya Horiguchi #include <linux/hugetlb.h> 64444f84fdSBen Dooks #include <linux/huge_mm.h> 65ef5d437fSJan Kara #include <linux/backing-dev.h> 6633c3fc71SVladimir Davydov #include <linux/page_idle.h> 67a5430ddaSJérôme Glisse #include <linux/memremap.h> 68bce73e48SChristian Borntraeger #include <linux/userfaultfd_k.h> 691da177e4SLinus Torvalds 701da177e4SLinus Torvalds #include <asm/tlbflush.h> 711da177e4SLinus Torvalds 7272b252aeSMel Gorman #include <trace/events/tlb.h> 7372b252aeSMel Gorman 74b291f000SNick Piggin #include "internal.h" 75b291f000SNick Piggin 76fdd2e5f8SAdrian Bunk static struct kmem_cache *anon_vma_cachep; 775beb4930SRik van Riel static struct kmem_cache *anon_vma_chain_cachep; 78fdd2e5f8SAdrian Bunk 79fdd2e5f8SAdrian Bunk static inline struct anon_vma *anon_vma_alloc(void) 80fdd2e5f8SAdrian Bunk { 8101d8b20dSPeter Zijlstra struct anon_vma *anon_vma; 8201d8b20dSPeter Zijlstra 8301d8b20dSPeter Zijlstra anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); 8401d8b20dSPeter Zijlstra if (anon_vma) { 8501d8b20dSPeter Zijlstra atomic_set(&anon_vma->refcount, 1); 867a3ef208SKonstantin Khlebnikov anon_vma->degree = 1; /* Reference for first vma */ 877a3ef208SKonstantin Khlebnikov anon_vma->parent = anon_vma; 8801d8b20dSPeter Zijlstra /* 8901d8b20dSPeter Zijlstra * Initialise the anon_vma root to point to itself. If called 9001d8b20dSPeter Zijlstra * from fork, the root will be reset to the parents anon_vma. 9101d8b20dSPeter Zijlstra */ 9201d8b20dSPeter Zijlstra anon_vma->root = anon_vma; 93fdd2e5f8SAdrian Bunk } 94fdd2e5f8SAdrian Bunk 9501d8b20dSPeter Zijlstra return anon_vma; 9601d8b20dSPeter Zijlstra } 9701d8b20dSPeter Zijlstra 9801d8b20dSPeter Zijlstra static inline void anon_vma_free(struct anon_vma *anon_vma) 99fdd2e5f8SAdrian Bunk { 10001d8b20dSPeter Zijlstra VM_BUG_ON(atomic_read(&anon_vma->refcount)); 10188c22088SPeter Zijlstra 10288c22088SPeter Zijlstra /* 1034fc3f1d6SIngo Molnar * Synchronize against page_lock_anon_vma_read() such that 10488c22088SPeter Zijlstra * we can safely hold the lock without the anon_vma getting 10588c22088SPeter Zijlstra * freed. 10688c22088SPeter Zijlstra * 10788c22088SPeter Zijlstra * Relies on the full mb implied by the atomic_dec_and_test() from 10888c22088SPeter Zijlstra * put_anon_vma() against the acquire barrier implied by 1094fc3f1d6SIngo Molnar * down_read_trylock() from page_lock_anon_vma_read(). This orders: 11088c22088SPeter Zijlstra * 1114fc3f1d6SIngo Molnar * page_lock_anon_vma_read() VS put_anon_vma() 1124fc3f1d6SIngo Molnar * down_read_trylock() atomic_dec_and_test() 11388c22088SPeter Zijlstra * LOCK MB 1144fc3f1d6SIngo Molnar * atomic_read() rwsem_is_locked() 11588c22088SPeter Zijlstra * 11688c22088SPeter Zijlstra * LOCK should suffice since the actual taking of the lock must 11788c22088SPeter Zijlstra * happen _before_ what follows. 11888c22088SPeter Zijlstra */ 1197f39dda9SHugh Dickins might_sleep(); 1205a505085SIngo Molnar if (rwsem_is_locked(&anon_vma->root->rwsem)) { 1214fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 12208b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 12388c22088SPeter Zijlstra } 12488c22088SPeter Zijlstra 125fdd2e5f8SAdrian Bunk kmem_cache_free(anon_vma_cachep, anon_vma); 126fdd2e5f8SAdrian Bunk } 1271da177e4SLinus Torvalds 128dd34739cSLinus Torvalds static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) 1295beb4930SRik van Riel { 130dd34739cSLinus Torvalds return kmem_cache_alloc(anon_vma_chain_cachep, gfp); 1315beb4930SRik van Riel } 1325beb4930SRik van Riel 133e574b5fdSNamhyung Kim static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) 1345beb4930SRik van Riel { 1355beb4930SRik van Riel kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); 1365beb4930SRik van Riel } 1375beb4930SRik van Riel 1386583a843SKautuk Consul static void anon_vma_chain_link(struct vm_area_struct *vma, 1396583a843SKautuk Consul struct anon_vma_chain *avc, 1406583a843SKautuk Consul struct anon_vma *anon_vma) 1416583a843SKautuk Consul { 1426583a843SKautuk Consul avc->vma = vma; 1436583a843SKautuk Consul avc->anon_vma = anon_vma; 1446583a843SKautuk Consul list_add(&avc->same_vma, &vma->anon_vma_chain); 145bf181b9fSMichel Lespinasse anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); 1466583a843SKautuk Consul } 1476583a843SKautuk Consul 148d9d332e0SLinus Torvalds /** 149d5a187daSVlastimil Babka * __anon_vma_prepare - attach an anon_vma to a memory region 150d9d332e0SLinus Torvalds * @vma: the memory region in question 151d9d332e0SLinus Torvalds * 152d9d332e0SLinus Torvalds * This makes sure the memory mapping described by 'vma' has 153d9d332e0SLinus Torvalds * an 'anon_vma' attached to it, so that we can associate the 154d9d332e0SLinus Torvalds * anonymous pages mapped into it with that anon_vma. 155d9d332e0SLinus Torvalds * 156d5a187daSVlastimil Babka * The common case will be that we already have one, which 157d5a187daSVlastimil Babka * is handled inline by anon_vma_prepare(). But if 15823a0790aSFigo.zhang * not we either need to find an adjacent mapping that we 159d9d332e0SLinus Torvalds * can re-use the anon_vma from (very common when the only 160d9d332e0SLinus Torvalds * reason for splitting a vma has been mprotect()), or we 161d9d332e0SLinus Torvalds * allocate a new one. 162d9d332e0SLinus Torvalds * 163d9d332e0SLinus Torvalds * Anon-vma allocations are very subtle, because we may have 1644fc3f1d6SIngo Molnar * optimistically looked up an anon_vma in page_lock_anon_vma_read() 165d9d332e0SLinus Torvalds * and that may actually touch the spinlock even in the newly 166d9d332e0SLinus Torvalds * allocated vma (it depends on RCU to make sure that the 167d9d332e0SLinus Torvalds * anon_vma isn't actually destroyed). 168d9d332e0SLinus Torvalds * 169d9d332e0SLinus Torvalds * As a result, we need to do proper anon_vma locking even 170d9d332e0SLinus Torvalds * for the new allocation. At the same time, we do not want 171d9d332e0SLinus Torvalds * to do any locking for the common case of already having 172d9d332e0SLinus Torvalds * an anon_vma. 173d9d332e0SLinus Torvalds * 174d9d332e0SLinus Torvalds * This must be called with the mmap_sem held for reading. 175d9d332e0SLinus Torvalds */ 176d5a187daSVlastimil Babka int __anon_vma_prepare(struct vm_area_struct *vma) 1771da177e4SLinus Torvalds { 178d5a187daSVlastimil Babka struct mm_struct *mm = vma->vm_mm; 179d5a187daSVlastimil Babka struct anon_vma *anon_vma, *allocated; 1805beb4930SRik van Riel struct anon_vma_chain *avc; 1811da177e4SLinus Torvalds 1821da177e4SLinus Torvalds might_sleep(); 1831da177e4SLinus Torvalds 184dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 1855beb4930SRik van Riel if (!avc) 1865beb4930SRik van Riel goto out_enomem; 1875beb4930SRik van Riel 1881da177e4SLinus Torvalds anon_vma = find_mergeable_anon_vma(vma); 1891da177e4SLinus Torvalds allocated = NULL; 190d9d332e0SLinus Torvalds if (!anon_vma) { 1911da177e4SLinus Torvalds anon_vma = anon_vma_alloc(); 1921da177e4SLinus Torvalds if (unlikely(!anon_vma)) 1935beb4930SRik van Riel goto out_enomem_free_avc; 1941da177e4SLinus Torvalds allocated = anon_vma; 1951da177e4SLinus Torvalds } 1961da177e4SLinus Torvalds 1974fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 1981da177e4SLinus Torvalds /* page_table_lock to protect against threads */ 1991da177e4SLinus Torvalds spin_lock(&mm->page_table_lock); 2001da177e4SLinus Torvalds if (likely(!vma->anon_vma)) { 2011da177e4SLinus Torvalds vma->anon_vma = anon_vma; 2026583a843SKautuk Consul anon_vma_chain_link(vma, avc, anon_vma); 2037a3ef208SKonstantin Khlebnikov /* vma reference or self-parent link for new root */ 2047a3ef208SKonstantin Khlebnikov anon_vma->degree++; 2051da177e4SLinus Torvalds allocated = NULL; 20631f2b0ebSOleg Nesterov avc = NULL; 2071da177e4SLinus Torvalds } 2081da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 20908b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 21031f2b0ebSOleg Nesterov 21131f2b0ebSOleg Nesterov if (unlikely(allocated)) 21201d8b20dSPeter Zijlstra put_anon_vma(allocated); 21331f2b0ebSOleg Nesterov if (unlikely(avc)) 2145beb4930SRik van Riel anon_vma_chain_free(avc); 215d5a187daSVlastimil Babka 2161da177e4SLinus Torvalds return 0; 2175beb4930SRik van Riel 2185beb4930SRik van Riel out_enomem_free_avc: 2195beb4930SRik van Riel anon_vma_chain_free(avc); 2205beb4930SRik van Riel out_enomem: 2215beb4930SRik van Riel return -ENOMEM; 2221da177e4SLinus Torvalds } 2231da177e4SLinus Torvalds 224bb4aa396SLinus Torvalds /* 225bb4aa396SLinus Torvalds * This is a useful helper function for locking the anon_vma root as 226bb4aa396SLinus Torvalds * we traverse the vma->anon_vma_chain, looping over anon_vma's that 227bb4aa396SLinus Torvalds * have the same vma. 228bb4aa396SLinus Torvalds * 229bb4aa396SLinus Torvalds * Such anon_vma's should have the same root, so you'd expect to see 230bb4aa396SLinus Torvalds * just a single mutex_lock for the whole traversal. 231bb4aa396SLinus Torvalds */ 232bb4aa396SLinus Torvalds static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) 233bb4aa396SLinus Torvalds { 234bb4aa396SLinus Torvalds struct anon_vma *new_root = anon_vma->root; 235bb4aa396SLinus Torvalds if (new_root != root) { 236bb4aa396SLinus Torvalds if (WARN_ON_ONCE(root)) 2375a505085SIngo Molnar up_write(&root->rwsem); 238bb4aa396SLinus Torvalds root = new_root; 2395a505085SIngo Molnar down_write(&root->rwsem); 240bb4aa396SLinus Torvalds } 241bb4aa396SLinus Torvalds return root; 242bb4aa396SLinus Torvalds } 243bb4aa396SLinus Torvalds 244bb4aa396SLinus Torvalds static inline void unlock_anon_vma_root(struct anon_vma *root) 245bb4aa396SLinus Torvalds { 246bb4aa396SLinus Torvalds if (root) 2475a505085SIngo Molnar up_write(&root->rwsem); 248bb4aa396SLinus Torvalds } 249bb4aa396SLinus Torvalds 2505beb4930SRik van Riel /* 2515beb4930SRik van Riel * Attach the anon_vmas from src to dst. 2525beb4930SRik van Riel * Returns 0 on success, -ENOMEM on failure. 2537a3ef208SKonstantin Khlebnikov * 254*47b390d2SWei Yang * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and 255*47b390d2SWei Yang * anon_vma_fork(). The first three want an exact copy of src, while the last 256*47b390d2SWei Yang * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent 257*47b390d2SWei Yang * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call, 258*47b390d2SWei Yang * we can identify this case by checking (!dst->anon_vma && src->anon_vma). 259*47b390d2SWei Yang * 260*47b390d2SWei Yang * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find 261*47b390d2SWei Yang * and reuse existing anon_vma which has no vmas and only one child anon_vma. 262*47b390d2SWei Yang * This prevents degradation of anon_vma hierarchy to endless linear chain in 263*47b390d2SWei Yang * case of constantly forking task. On the other hand, an anon_vma with more 264*47b390d2SWei Yang * than one child isn't reused even if there was no alive vma, thus rmap 265*47b390d2SWei Yang * walker has a good chance of avoiding scanning the whole hierarchy when it 266*47b390d2SWei Yang * searches where page is mapped. 2675beb4930SRik van Riel */ 2685beb4930SRik van Riel int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) 2695beb4930SRik van Riel { 2705beb4930SRik van Riel struct anon_vma_chain *avc, *pavc; 271bb4aa396SLinus Torvalds struct anon_vma *root = NULL; 2725beb4930SRik van Riel 273646d87b4SLinus Torvalds list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { 274bb4aa396SLinus Torvalds struct anon_vma *anon_vma; 275bb4aa396SLinus Torvalds 276dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); 277dd34739cSLinus Torvalds if (unlikely(!avc)) { 278dd34739cSLinus Torvalds unlock_anon_vma_root(root); 279dd34739cSLinus Torvalds root = NULL; 280dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 2815beb4930SRik van Riel if (!avc) 2825beb4930SRik van Riel goto enomem_failure; 283dd34739cSLinus Torvalds } 284bb4aa396SLinus Torvalds anon_vma = pavc->anon_vma; 285bb4aa396SLinus Torvalds root = lock_anon_vma_root(root, anon_vma); 286bb4aa396SLinus Torvalds anon_vma_chain_link(dst, avc, anon_vma); 2877a3ef208SKonstantin Khlebnikov 2887a3ef208SKonstantin Khlebnikov /* 2897a3ef208SKonstantin Khlebnikov * Reuse existing anon_vma if its degree lower than two, 2907a3ef208SKonstantin Khlebnikov * that means it has no vma and only one anon_vma child. 2917a3ef208SKonstantin Khlebnikov * 2927a3ef208SKonstantin Khlebnikov * Do not chose parent anon_vma, otherwise first child 2937a3ef208SKonstantin Khlebnikov * will always reuse it. Root anon_vma is never reused: 2947a3ef208SKonstantin Khlebnikov * it has self-parent reference and at least one child. 2957a3ef208SKonstantin Khlebnikov */ 296*47b390d2SWei Yang if (!dst->anon_vma && src->anon_vma && 297*47b390d2SWei Yang anon_vma != src->anon_vma && anon_vma->degree < 2) 2987a3ef208SKonstantin Khlebnikov dst->anon_vma = anon_vma; 2995beb4930SRik van Riel } 3007a3ef208SKonstantin Khlebnikov if (dst->anon_vma) 3017a3ef208SKonstantin Khlebnikov dst->anon_vma->degree++; 302bb4aa396SLinus Torvalds unlock_anon_vma_root(root); 3035beb4930SRik van Riel return 0; 3045beb4930SRik van Riel 3055beb4930SRik van Riel enomem_failure: 3063fe89b3eSLeon Yu /* 3073fe89b3eSLeon Yu * dst->anon_vma is dropped here otherwise its degree can be incorrectly 3083fe89b3eSLeon Yu * decremented in unlink_anon_vmas(). 3093fe89b3eSLeon Yu * We can safely do this because callers of anon_vma_clone() don't care 3103fe89b3eSLeon Yu * about dst->anon_vma if anon_vma_clone() failed. 3113fe89b3eSLeon Yu */ 3123fe89b3eSLeon Yu dst->anon_vma = NULL; 3135beb4930SRik van Riel unlink_anon_vmas(dst); 3145beb4930SRik van Riel return -ENOMEM; 3151da177e4SLinus Torvalds } 3161da177e4SLinus Torvalds 3175beb4930SRik van Riel /* 3185beb4930SRik van Riel * Attach vma to its own anon_vma, as well as to the anon_vmas that 3195beb4930SRik van Riel * the corresponding VMA in the parent process is attached to. 3205beb4930SRik van Riel * Returns 0 on success, non-zero on failure. 3215beb4930SRik van Riel */ 3225beb4930SRik van Riel int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) 3231da177e4SLinus Torvalds { 3245beb4930SRik van Riel struct anon_vma_chain *avc; 3255beb4930SRik van Riel struct anon_vma *anon_vma; 326c4ea95d7SDaniel Forrest int error; 3275beb4930SRik van Riel 3285beb4930SRik van Riel /* Don't bother if the parent process has no anon_vma here. */ 3295beb4930SRik van Riel if (!pvma->anon_vma) 3305beb4930SRik van Riel return 0; 3315beb4930SRik van Riel 3327a3ef208SKonstantin Khlebnikov /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ 3337a3ef208SKonstantin Khlebnikov vma->anon_vma = NULL; 3347a3ef208SKonstantin Khlebnikov 3355beb4930SRik van Riel /* 3365beb4930SRik van Riel * First, attach the new VMA to the parent VMA's anon_vmas, 3375beb4930SRik van Riel * so rmap can find non-COWed pages in child processes. 3385beb4930SRik van Riel */ 339c4ea95d7SDaniel Forrest error = anon_vma_clone(vma, pvma); 340c4ea95d7SDaniel Forrest if (error) 341c4ea95d7SDaniel Forrest return error; 3425beb4930SRik van Riel 3437a3ef208SKonstantin Khlebnikov /* An existing anon_vma has been reused, all done then. */ 3447a3ef208SKonstantin Khlebnikov if (vma->anon_vma) 3457a3ef208SKonstantin Khlebnikov return 0; 3467a3ef208SKonstantin Khlebnikov 3475beb4930SRik van Riel /* Then add our own anon_vma. */ 3485beb4930SRik van Riel anon_vma = anon_vma_alloc(); 3495beb4930SRik van Riel if (!anon_vma) 3505beb4930SRik van Riel goto out_error; 351dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 3525beb4930SRik van Riel if (!avc) 3535beb4930SRik van Riel goto out_error_free_anon_vma; 3545c341ee1SRik van Riel 3555c341ee1SRik van Riel /* 3565c341ee1SRik van Riel * The root anon_vma's spinlock is the lock actually used when we 3575c341ee1SRik van Riel * lock any of the anon_vmas in this anon_vma tree. 3585c341ee1SRik van Riel */ 3595c341ee1SRik van Riel anon_vma->root = pvma->anon_vma->root; 3607a3ef208SKonstantin Khlebnikov anon_vma->parent = pvma->anon_vma; 36176545066SRik van Riel /* 36201d8b20dSPeter Zijlstra * With refcounts, an anon_vma can stay around longer than the 36301d8b20dSPeter Zijlstra * process it belongs to. The root anon_vma needs to be pinned until 36401d8b20dSPeter Zijlstra * this anon_vma is freed, because the lock lives in the root. 36576545066SRik van Riel */ 36676545066SRik van Riel get_anon_vma(anon_vma->root); 3675beb4930SRik van Riel /* Mark this anon_vma as the one where our new (COWed) pages go. */ 3685beb4930SRik van Riel vma->anon_vma = anon_vma; 3694fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 3705c341ee1SRik van Riel anon_vma_chain_link(vma, avc, anon_vma); 3717a3ef208SKonstantin Khlebnikov anon_vma->parent->degree++; 37208b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 3735beb4930SRik van Riel 3745beb4930SRik van Riel return 0; 3755beb4930SRik van Riel 3765beb4930SRik van Riel out_error_free_anon_vma: 37701d8b20dSPeter Zijlstra put_anon_vma(anon_vma); 3785beb4930SRik van Riel out_error: 3794946d54cSRik van Riel unlink_anon_vmas(vma); 3805beb4930SRik van Riel return -ENOMEM; 3815beb4930SRik van Riel } 3825beb4930SRik van Riel 3835beb4930SRik van Riel void unlink_anon_vmas(struct vm_area_struct *vma) 3845beb4930SRik van Riel { 3855beb4930SRik van Riel struct anon_vma_chain *avc, *next; 386eee2acbaSPeter Zijlstra struct anon_vma *root = NULL; 3875beb4930SRik van Riel 3885c341ee1SRik van Riel /* 3895c341ee1SRik van Riel * Unlink each anon_vma chained to the VMA. This list is ordered 3905c341ee1SRik van Riel * from newest to oldest, ensuring the root anon_vma gets freed last. 3915c341ee1SRik van Riel */ 3925beb4930SRik van Riel list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 393eee2acbaSPeter Zijlstra struct anon_vma *anon_vma = avc->anon_vma; 394eee2acbaSPeter Zijlstra 395eee2acbaSPeter Zijlstra root = lock_anon_vma_root(root, anon_vma); 396bf181b9fSMichel Lespinasse anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); 397eee2acbaSPeter Zijlstra 398eee2acbaSPeter Zijlstra /* 399eee2acbaSPeter Zijlstra * Leave empty anon_vmas on the list - we'll need 400eee2acbaSPeter Zijlstra * to free them outside the lock. 401eee2acbaSPeter Zijlstra */ 402f808c13fSDavidlohr Bueso if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) { 4037a3ef208SKonstantin Khlebnikov anon_vma->parent->degree--; 404eee2acbaSPeter Zijlstra continue; 4057a3ef208SKonstantin Khlebnikov } 406eee2acbaSPeter Zijlstra 407eee2acbaSPeter Zijlstra list_del(&avc->same_vma); 408eee2acbaSPeter Zijlstra anon_vma_chain_free(avc); 409eee2acbaSPeter Zijlstra } 4107a3ef208SKonstantin Khlebnikov if (vma->anon_vma) 4117a3ef208SKonstantin Khlebnikov vma->anon_vma->degree--; 412eee2acbaSPeter Zijlstra unlock_anon_vma_root(root); 413eee2acbaSPeter Zijlstra 414eee2acbaSPeter Zijlstra /* 415eee2acbaSPeter Zijlstra * Iterate the list once more, it now only contains empty and unlinked 416eee2acbaSPeter Zijlstra * anon_vmas, destroy them. Could not do before due to __put_anon_vma() 4175a505085SIngo Molnar * needing to write-acquire the anon_vma->root->rwsem. 418eee2acbaSPeter Zijlstra */ 419eee2acbaSPeter Zijlstra list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 420eee2acbaSPeter Zijlstra struct anon_vma *anon_vma = avc->anon_vma; 421eee2acbaSPeter Zijlstra 422e4c5800aSKonstantin Khlebnikov VM_WARN_ON(anon_vma->degree); 423eee2acbaSPeter Zijlstra put_anon_vma(anon_vma); 424eee2acbaSPeter Zijlstra 4255beb4930SRik van Riel list_del(&avc->same_vma); 4265beb4930SRik van Riel anon_vma_chain_free(avc); 4275beb4930SRik van Riel } 4285beb4930SRik van Riel } 4295beb4930SRik van Riel 43051cc5068SAlexey Dobriyan static void anon_vma_ctor(void *data) 4311da177e4SLinus Torvalds { 4321da177e4SLinus Torvalds struct anon_vma *anon_vma = data; 4331da177e4SLinus Torvalds 4345a505085SIngo Molnar init_rwsem(&anon_vma->rwsem); 43583813267SPeter Zijlstra atomic_set(&anon_vma->refcount, 0); 436f808c13fSDavidlohr Bueso anon_vma->rb_root = RB_ROOT_CACHED; 4371da177e4SLinus Torvalds } 4381da177e4SLinus Torvalds 4391da177e4SLinus Torvalds void __init anon_vma_init(void) 4401da177e4SLinus Torvalds { 4411da177e4SLinus Torvalds anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 4425f0d5a3aSPaul E. McKenney 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, 4435d097056SVladimir Davydov anon_vma_ctor); 4445d097056SVladimir Davydov anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, 4455d097056SVladimir Davydov SLAB_PANIC|SLAB_ACCOUNT); 4461da177e4SLinus Torvalds } 4471da177e4SLinus Torvalds 4481da177e4SLinus Torvalds /* 4496111e4caSPeter Zijlstra * Getting a lock on a stable anon_vma from a page off the LRU is tricky! 4506111e4caSPeter Zijlstra * 4516111e4caSPeter Zijlstra * Since there is no serialization what so ever against page_remove_rmap() 4526111e4caSPeter Zijlstra * the best this function can do is return a locked anon_vma that might 4536111e4caSPeter Zijlstra * have been relevant to this page. 4546111e4caSPeter Zijlstra * 4556111e4caSPeter Zijlstra * The page might have been remapped to a different anon_vma or the anon_vma 4566111e4caSPeter Zijlstra * returned may already be freed (and even reused). 4576111e4caSPeter Zijlstra * 458bc658c96SPeter Zijlstra * In case it was remapped to a different anon_vma, the new anon_vma will be a 459bc658c96SPeter Zijlstra * child of the old anon_vma, and the anon_vma lifetime rules will therefore 460bc658c96SPeter Zijlstra * ensure that any anon_vma obtained from the page will still be valid for as 461bc658c96SPeter Zijlstra * long as we observe page_mapped() [ hence all those page_mapped() tests ]. 462bc658c96SPeter Zijlstra * 4636111e4caSPeter Zijlstra * All users of this function must be very careful when walking the anon_vma 4646111e4caSPeter Zijlstra * chain and verify that the page in question is indeed mapped in it 4656111e4caSPeter Zijlstra * [ something equivalent to page_mapped_in_vma() ]. 4666111e4caSPeter Zijlstra * 4676111e4caSPeter Zijlstra * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() 4686111e4caSPeter Zijlstra * that the anon_vma pointer from page->mapping is valid if there is a 4696111e4caSPeter Zijlstra * mapcount, we can dereference the anon_vma after observing those. 4701da177e4SLinus Torvalds */ 471746b18d4SPeter Zijlstra struct anon_vma *page_get_anon_vma(struct page *page) 4721da177e4SLinus Torvalds { 473746b18d4SPeter Zijlstra struct anon_vma *anon_vma = NULL; 4741da177e4SLinus Torvalds unsigned long anon_mapping; 4751da177e4SLinus Torvalds 4761da177e4SLinus Torvalds rcu_read_lock(); 4774db0c3c2SJason Low anon_mapping = (unsigned long)READ_ONCE(page->mapping); 4783ca7b3c5SHugh Dickins if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 4791da177e4SLinus Torvalds goto out; 4801da177e4SLinus Torvalds if (!page_mapped(page)) 4811da177e4SLinus Torvalds goto out; 4821da177e4SLinus Torvalds 4831da177e4SLinus Torvalds anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 484746b18d4SPeter Zijlstra if (!atomic_inc_not_zero(&anon_vma->refcount)) { 485746b18d4SPeter Zijlstra anon_vma = NULL; 486746b18d4SPeter Zijlstra goto out; 487746b18d4SPeter Zijlstra } 488f1819427SHugh Dickins 489f1819427SHugh Dickins /* 490f1819427SHugh Dickins * If this page is still mapped, then its anon_vma cannot have been 491746b18d4SPeter Zijlstra * freed. But if it has been unmapped, we have no security against the 492746b18d4SPeter Zijlstra * anon_vma structure being freed and reused (for another anon_vma: 4935f0d5a3aSPaul E. McKenney * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero() 494746b18d4SPeter Zijlstra * above cannot corrupt). 495f1819427SHugh Dickins */ 496746b18d4SPeter Zijlstra if (!page_mapped(page)) { 4977f39dda9SHugh Dickins rcu_read_unlock(); 498746b18d4SPeter Zijlstra put_anon_vma(anon_vma); 4997f39dda9SHugh Dickins return NULL; 500746b18d4SPeter Zijlstra } 5011da177e4SLinus Torvalds out: 5021da177e4SLinus Torvalds rcu_read_unlock(); 503746b18d4SPeter Zijlstra 504746b18d4SPeter Zijlstra return anon_vma; 505746b18d4SPeter Zijlstra } 506746b18d4SPeter Zijlstra 50788c22088SPeter Zijlstra /* 50888c22088SPeter Zijlstra * Similar to page_get_anon_vma() except it locks the anon_vma. 50988c22088SPeter Zijlstra * 51088c22088SPeter Zijlstra * Its a little more complex as it tries to keep the fast path to a single 51188c22088SPeter Zijlstra * atomic op -- the trylock. If we fail the trylock, we fall back to getting a 51288c22088SPeter Zijlstra * reference like with page_get_anon_vma() and then block on the mutex. 51388c22088SPeter Zijlstra */ 5144fc3f1d6SIngo Molnar struct anon_vma *page_lock_anon_vma_read(struct page *page) 515746b18d4SPeter Zijlstra { 51688c22088SPeter Zijlstra struct anon_vma *anon_vma = NULL; 517eee0f252SHugh Dickins struct anon_vma *root_anon_vma; 51888c22088SPeter Zijlstra unsigned long anon_mapping; 519746b18d4SPeter Zijlstra 52088c22088SPeter Zijlstra rcu_read_lock(); 5214db0c3c2SJason Low anon_mapping = (unsigned long)READ_ONCE(page->mapping); 52288c22088SPeter Zijlstra if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 52388c22088SPeter Zijlstra goto out; 52488c22088SPeter Zijlstra if (!page_mapped(page)) 52588c22088SPeter Zijlstra goto out; 52688c22088SPeter Zijlstra 52788c22088SPeter Zijlstra anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 5284db0c3c2SJason Low root_anon_vma = READ_ONCE(anon_vma->root); 5294fc3f1d6SIngo Molnar if (down_read_trylock(&root_anon_vma->rwsem)) { 53088c22088SPeter Zijlstra /* 531eee0f252SHugh Dickins * If the page is still mapped, then this anon_vma is still 532eee0f252SHugh Dickins * its anon_vma, and holding the mutex ensures that it will 533bc658c96SPeter Zijlstra * not go away, see anon_vma_free(). 53488c22088SPeter Zijlstra */ 535eee0f252SHugh Dickins if (!page_mapped(page)) { 5364fc3f1d6SIngo Molnar up_read(&root_anon_vma->rwsem); 53788c22088SPeter Zijlstra anon_vma = NULL; 53888c22088SPeter Zijlstra } 53988c22088SPeter Zijlstra goto out; 54088c22088SPeter Zijlstra } 54188c22088SPeter Zijlstra 54288c22088SPeter Zijlstra /* trylock failed, we got to sleep */ 54388c22088SPeter Zijlstra if (!atomic_inc_not_zero(&anon_vma->refcount)) { 54488c22088SPeter Zijlstra anon_vma = NULL; 54588c22088SPeter Zijlstra goto out; 54688c22088SPeter Zijlstra } 54788c22088SPeter Zijlstra 54888c22088SPeter Zijlstra if (!page_mapped(page)) { 5497f39dda9SHugh Dickins rcu_read_unlock(); 55088c22088SPeter Zijlstra put_anon_vma(anon_vma); 5517f39dda9SHugh Dickins return NULL; 55288c22088SPeter Zijlstra } 55388c22088SPeter Zijlstra 55488c22088SPeter Zijlstra /* we pinned the anon_vma, its safe to sleep */ 55588c22088SPeter Zijlstra rcu_read_unlock(); 5564fc3f1d6SIngo Molnar anon_vma_lock_read(anon_vma); 557746b18d4SPeter Zijlstra 55888c22088SPeter Zijlstra if (atomic_dec_and_test(&anon_vma->refcount)) { 55988c22088SPeter Zijlstra /* 56088c22088SPeter Zijlstra * Oops, we held the last refcount, release the lock 56188c22088SPeter Zijlstra * and bail -- can't simply use put_anon_vma() because 5624fc3f1d6SIngo Molnar * we'll deadlock on the anon_vma_lock_write() recursion. 56388c22088SPeter Zijlstra */ 5644fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 56588c22088SPeter Zijlstra __put_anon_vma(anon_vma); 56688c22088SPeter Zijlstra anon_vma = NULL; 56788c22088SPeter Zijlstra } 56888c22088SPeter Zijlstra 56988c22088SPeter Zijlstra return anon_vma; 57088c22088SPeter Zijlstra 57188c22088SPeter Zijlstra out: 57288c22088SPeter Zijlstra rcu_read_unlock(); 573746b18d4SPeter Zijlstra return anon_vma; 57434bbd704SOleg Nesterov } 57534bbd704SOleg Nesterov 5764fc3f1d6SIngo Molnar void page_unlock_anon_vma_read(struct anon_vma *anon_vma) 57734bbd704SOleg Nesterov { 5784fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 5791da177e4SLinus Torvalds } 5801da177e4SLinus Torvalds 58172b252aeSMel Gorman #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH 58272b252aeSMel Gorman /* 58372b252aeSMel Gorman * Flush TLB entries for recently unmapped pages from remote CPUs. It is 58472b252aeSMel Gorman * important if a PTE was dirty when it was unmapped that it's flushed 58572b252aeSMel Gorman * before any IO is initiated on the page to prevent lost writes. Similarly, 58672b252aeSMel Gorman * it must be flushed before freeing to prevent data leakage. 58772b252aeSMel Gorman */ 58872b252aeSMel Gorman void try_to_unmap_flush(void) 58972b252aeSMel Gorman { 59072b252aeSMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 59172b252aeSMel Gorman 59272b252aeSMel Gorman if (!tlb_ubc->flush_required) 59372b252aeSMel Gorman return; 59472b252aeSMel Gorman 595e73ad5ffSAndy Lutomirski arch_tlbbatch_flush(&tlb_ubc->arch); 59672b252aeSMel Gorman tlb_ubc->flush_required = false; 597d950c947SMel Gorman tlb_ubc->writable = false; 59872b252aeSMel Gorman } 59972b252aeSMel Gorman 600d950c947SMel Gorman /* Flush iff there are potentially writable TLB entries that can race with IO */ 601d950c947SMel Gorman void try_to_unmap_flush_dirty(void) 602d950c947SMel Gorman { 603d950c947SMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 604d950c947SMel Gorman 605d950c947SMel Gorman if (tlb_ubc->writable) 606d950c947SMel Gorman try_to_unmap_flush(); 607d950c947SMel Gorman } 608d950c947SMel Gorman 609c7ab0d2fSKirill A. Shutemov static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) 61072b252aeSMel Gorman { 61172b252aeSMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 61272b252aeSMel Gorman 613e73ad5ffSAndy Lutomirski arch_tlbbatch_add_mm(&tlb_ubc->arch, mm); 61472b252aeSMel Gorman tlb_ubc->flush_required = true; 615d950c947SMel Gorman 616d950c947SMel Gorman /* 6173ea27719SMel Gorman * Ensure compiler does not re-order the setting of tlb_flush_batched 6183ea27719SMel Gorman * before the PTE is cleared. 6193ea27719SMel Gorman */ 6203ea27719SMel Gorman barrier(); 6213ea27719SMel Gorman mm->tlb_flush_batched = true; 6223ea27719SMel Gorman 6233ea27719SMel Gorman /* 624d950c947SMel Gorman * If the PTE was dirty then it's best to assume it's writable. The 625d950c947SMel Gorman * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() 626d950c947SMel Gorman * before the page is queued for IO. 627d950c947SMel Gorman */ 628d950c947SMel Gorman if (writable) 629d950c947SMel Gorman tlb_ubc->writable = true; 63072b252aeSMel Gorman } 63172b252aeSMel Gorman 63272b252aeSMel Gorman /* 63372b252aeSMel Gorman * Returns true if the TLB flush should be deferred to the end of a batch of 63472b252aeSMel Gorman * unmap operations to reduce IPIs. 63572b252aeSMel Gorman */ 63672b252aeSMel Gorman static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 63772b252aeSMel Gorman { 63872b252aeSMel Gorman bool should_defer = false; 63972b252aeSMel Gorman 64072b252aeSMel Gorman if (!(flags & TTU_BATCH_FLUSH)) 64172b252aeSMel Gorman return false; 64272b252aeSMel Gorman 64372b252aeSMel Gorman /* If remote CPUs need to be flushed then defer batch the flush */ 64472b252aeSMel Gorman if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids) 64572b252aeSMel Gorman should_defer = true; 64672b252aeSMel Gorman put_cpu(); 64772b252aeSMel Gorman 64872b252aeSMel Gorman return should_defer; 64972b252aeSMel Gorman } 6503ea27719SMel Gorman 6513ea27719SMel Gorman /* 6523ea27719SMel Gorman * Reclaim unmaps pages under the PTL but do not flush the TLB prior to 6533ea27719SMel Gorman * releasing the PTL if TLB flushes are batched. It's possible for a parallel 6543ea27719SMel Gorman * operation such as mprotect or munmap to race between reclaim unmapping 6553ea27719SMel Gorman * the page and flushing the page. If this race occurs, it potentially allows 6563ea27719SMel Gorman * access to data via a stale TLB entry. Tracking all mm's that have TLB 6573ea27719SMel Gorman * batching in flight would be expensive during reclaim so instead track 6583ea27719SMel Gorman * whether TLB batching occurred in the past and if so then do a flush here 6593ea27719SMel Gorman * if required. This will cost one additional flush per reclaim cycle paid 6603ea27719SMel Gorman * by the first operation at risk such as mprotect and mumap. 6613ea27719SMel Gorman * 6623ea27719SMel Gorman * This must be called under the PTL so that an access to tlb_flush_batched 6633ea27719SMel Gorman * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise 6643ea27719SMel Gorman * via the PTL. 6653ea27719SMel Gorman */ 6663ea27719SMel Gorman void flush_tlb_batched_pending(struct mm_struct *mm) 6673ea27719SMel Gorman { 6683ea27719SMel Gorman if (mm->tlb_flush_batched) { 6693ea27719SMel Gorman flush_tlb_mm(mm); 6703ea27719SMel Gorman 6713ea27719SMel Gorman /* 6723ea27719SMel Gorman * Do not allow the compiler to re-order the clearing of 6733ea27719SMel Gorman * tlb_flush_batched before the tlb is flushed. 6743ea27719SMel Gorman */ 6753ea27719SMel Gorman barrier(); 6763ea27719SMel Gorman mm->tlb_flush_batched = false; 6773ea27719SMel Gorman } 6783ea27719SMel Gorman } 67972b252aeSMel Gorman #else 680c7ab0d2fSKirill A. Shutemov static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) 68172b252aeSMel Gorman { 68272b252aeSMel Gorman } 68372b252aeSMel Gorman 68472b252aeSMel Gorman static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 68572b252aeSMel Gorman { 68672b252aeSMel Gorman return false; 68772b252aeSMel Gorman } 68872b252aeSMel Gorman #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ 68972b252aeSMel Gorman 6901da177e4SLinus Torvalds /* 691bf89c8c8SHuang Shijie * At what user virtual address is page expected in vma? 692ab941e0fSNaoya Horiguchi * Caller should check the page is actually part of the vma. 6931da177e4SLinus Torvalds */ 6941da177e4SLinus Torvalds unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 6951da177e4SLinus Torvalds { 69686c2ad19SMichel Lespinasse unsigned long address; 69721d0d443SAndrea Arcangeli if (PageAnon(page)) { 6984829b906SHugh Dickins struct anon_vma *page__anon_vma = page_anon_vma(page); 6994829b906SHugh Dickins /* 7004829b906SHugh Dickins * Note: swapoff's unuse_vma() is more efficient with this 7014829b906SHugh Dickins * check, and needs it to match anon_vma when KSM is active. 7024829b906SHugh Dickins */ 7034829b906SHugh Dickins if (!vma->anon_vma || !page__anon_vma || 7044829b906SHugh Dickins vma->anon_vma->root != page__anon_vma->root) 70521d0d443SAndrea Arcangeli return -EFAULT; 70627ba0644SKirill A. Shutemov } else if (page->mapping) { 70727ba0644SKirill A. Shutemov if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping) 7081da177e4SLinus Torvalds return -EFAULT; 7091da177e4SLinus Torvalds } else 7101da177e4SLinus Torvalds return -EFAULT; 71186c2ad19SMichel Lespinasse address = __vma_address(page, vma); 71286c2ad19SMichel Lespinasse if (unlikely(address < vma->vm_start || address >= vma->vm_end)) 71386c2ad19SMichel Lespinasse return -EFAULT; 71486c2ad19SMichel Lespinasse return address; 7151da177e4SLinus Torvalds } 7161da177e4SLinus Torvalds 7176219049aSBob Liu pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) 7186219049aSBob Liu { 7196219049aSBob Liu pgd_t *pgd; 720c2febafcSKirill A. Shutemov p4d_t *p4d; 7216219049aSBob Liu pud_t *pud; 7226219049aSBob Liu pmd_t *pmd = NULL; 723f72e7dcdSHugh Dickins pmd_t pmde; 7246219049aSBob Liu 7256219049aSBob Liu pgd = pgd_offset(mm, address); 7266219049aSBob Liu if (!pgd_present(*pgd)) 7276219049aSBob Liu goto out; 7286219049aSBob Liu 729c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, address); 730c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 731c2febafcSKirill A. Shutemov goto out; 732c2febafcSKirill A. Shutemov 733c2febafcSKirill A. Shutemov pud = pud_offset(p4d, address); 7346219049aSBob Liu if (!pud_present(*pud)) 7356219049aSBob Liu goto out; 7366219049aSBob Liu 7376219049aSBob Liu pmd = pmd_offset(pud, address); 738f72e7dcdSHugh Dickins /* 7398809aa2dSAneesh Kumar K.V * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() 740f72e7dcdSHugh Dickins * without holding anon_vma lock for write. So when looking for a 741f72e7dcdSHugh Dickins * genuine pmde (in which to find pte), test present and !THP together. 742f72e7dcdSHugh Dickins */ 743e37c6982SChristian Borntraeger pmde = *pmd; 744e37c6982SChristian Borntraeger barrier(); 745f72e7dcdSHugh Dickins if (!pmd_present(pmde) || pmd_trans_huge(pmde)) 7466219049aSBob Liu pmd = NULL; 7476219049aSBob Liu out: 7486219049aSBob Liu return pmd; 7496219049aSBob Liu } 7506219049aSBob Liu 7519f32624bSJoonsoo Kim struct page_referenced_arg { 7529f32624bSJoonsoo Kim int mapcount; 7539f32624bSJoonsoo Kim int referenced; 7549f32624bSJoonsoo Kim unsigned long vm_flags; 7559f32624bSJoonsoo Kim struct mem_cgroup *memcg; 7569f32624bSJoonsoo Kim }; 75781b4082dSNikita Danilov /* 7589f32624bSJoonsoo Kim * arg: page_referenced_arg will be passed 7591da177e4SLinus Torvalds */ 760e4b82222SMinchan Kim static bool page_referenced_one(struct page *page, struct vm_area_struct *vma, 7619f32624bSJoonsoo Kim unsigned long address, void *arg) 7621da177e4SLinus Torvalds { 7639f32624bSJoonsoo Kim struct page_referenced_arg *pra = arg; 7648eaededeSKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 7658eaededeSKirill A. Shutemov .page = page, 7668eaededeSKirill A. Shutemov .vma = vma, 7678eaededeSKirill A. Shutemov .address = address, 7688eaededeSKirill A. Shutemov }; 7698749cfeaSVladimir Davydov int referenced = 0; 7702da28bfdSAndrea Arcangeli 7718eaededeSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 7728eaededeSKirill A. Shutemov address = pvmw.address; 7732da28bfdSAndrea Arcangeli 774b20ce5e0SKirill A. Shutemov if (vma->vm_flags & VM_LOCKED) { 7758eaededeSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 7769f32624bSJoonsoo Kim pra->vm_flags |= VM_LOCKED; 777e4b82222SMinchan Kim return false; /* To break the loop */ 7782da28bfdSAndrea Arcangeli } 7792da28bfdSAndrea Arcangeli 7808eaededeSKirill A. Shutemov if (pvmw.pte) { 7818eaededeSKirill A. Shutemov if (ptep_clear_flush_young_notify(vma, address, 7828eaededeSKirill A. Shutemov pvmw.pte)) { 7834917e5d0SJohannes Weiner /* 7848eaededeSKirill A. Shutemov * Don't treat a reference through 7858eaededeSKirill A. Shutemov * a sequentially read mapping as such. 7868eaededeSKirill A. Shutemov * If the page has been used in another mapping, 7878eaededeSKirill A. Shutemov * we will catch it; if this other mapping is 7888eaededeSKirill A. Shutemov * already gone, the unmap path will have set 7898eaededeSKirill A. Shutemov * PG_referenced or activated the page. 7904917e5d0SJohannes Weiner */ 79164363aadSJoe Perches if (likely(!(vma->vm_flags & VM_SEQ_READ))) 7921da177e4SLinus Torvalds referenced++; 7934917e5d0SJohannes Weiner } 7948749cfeaSVladimir Davydov } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { 7958eaededeSKirill A. Shutemov if (pmdp_clear_flush_young_notify(vma, address, 7968eaededeSKirill A. Shutemov pvmw.pmd)) 7978749cfeaSVladimir Davydov referenced++; 7988749cfeaSVladimir Davydov } else { 7998749cfeaSVladimir Davydov /* unexpected pmd-mapped page? */ 8008749cfeaSVladimir Davydov WARN_ON_ONCE(1); 8018749cfeaSVladimir Davydov } 8028eaededeSKirill A. Shutemov 8038eaededeSKirill A. Shutemov pra->mapcount--; 8048eaededeSKirill A. Shutemov } 80571e3aac0SAndrea Arcangeli 80633c3fc71SVladimir Davydov if (referenced) 80733c3fc71SVladimir Davydov clear_page_idle(page); 80833c3fc71SVladimir Davydov if (test_and_clear_page_young(page)) 80933c3fc71SVladimir Davydov referenced++; 81033c3fc71SVladimir Davydov 8119f32624bSJoonsoo Kim if (referenced) { 8129f32624bSJoonsoo Kim pra->referenced++; 8139f32624bSJoonsoo Kim pra->vm_flags |= vma->vm_flags; 8141da177e4SLinus Torvalds } 8151da177e4SLinus Torvalds 8169f32624bSJoonsoo Kim if (!pra->mapcount) 817e4b82222SMinchan Kim return false; /* To break the loop */ 8189f32624bSJoonsoo Kim 819e4b82222SMinchan Kim return true; 8209f32624bSJoonsoo Kim } 8219f32624bSJoonsoo Kim 8229f32624bSJoonsoo Kim static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg) 8231da177e4SLinus Torvalds { 8249f32624bSJoonsoo Kim struct page_referenced_arg *pra = arg; 8259f32624bSJoonsoo Kim struct mem_cgroup *memcg = pra->memcg; 8261da177e4SLinus Torvalds 8279f32624bSJoonsoo Kim if (!mm_match_cgroup(vma->vm_mm, memcg)) 8289f32624bSJoonsoo Kim return true; 8291da177e4SLinus Torvalds 8309f32624bSJoonsoo Kim return false; 8311da177e4SLinus Torvalds } 8321da177e4SLinus Torvalds 8331da177e4SLinus Torvalds /** 8341da177e4SLinus Torvalds * page_referenced - test if the page was referenced 8351da177e4SLinus Torvalds * @page: the page to test 8361da177e4SLinus Torvalds * @is_locked: caller holds lock on the page 83772835c86SJohannes Weiner * @memcg: target memory cgroup 8386fe6b7e3SWu Fengguang * @vm_flags: collect encountered vma->vm_flags who actually referenced the page 8391da177e4SLinus Torvalds * 8401da177e4SLinus Torvalds * Quick test_and_clear_referenced for all mappings to a page, 8411da177e4SLinus Torvalds * returns the number of ptes which referenced the page. 8421da177e4SLinus Torvalds */ 8436fe6b7e3SWu Fengguang int page_referenced(struct page *page, 8446fe6b7e3SWu Fengguang int is_locked, 84572835c86SJohannes Weiner struct mem_cgroup *memcg, 8466fe6b7e3SWu Fengguang unsigned long *vm_flags) 8471da177e4SLinus Torvalds { 8485ad64688SHugh Dickins int we_locked = 0; 8499f32624bSJoonsoo Kim struct page_referenced_arg pra = { 850b20ce5e0SKirill A. Shutemov .mapcount = total_mapcount(page), 8519f32624bSJoonsoo Kim .memcg = memcg, 8529f32624bSJoonsoo Kim }; 8539f32624bSJoonsoo Kim struct rmap_walk_control rwc = { 8549f32624bSJoonsoo Kim .rmap_one = page_referenced_one, 8559f32624bSJoonsoo Kim .arg = (void *)&pra, 8569f32624bSJoonsoo Kim .anon_lock = page_lock_anon_vma_read, 8579f32624bSJoonsoo Kim }; 8581da177e4SLinus Torvalds 8596fe6b7e3SWu Fengguang *vm_flags = 0; 860059d8442SHuang Shijie if (!pra.mapcount) 8619f32624bSJoonsoo Kim return 0; 8629f32624bSJoonsoo Kim 8639f32624bSJoonsoo Kim if (!page_rmapping(page)) 8649f32624bSJoonsoo Kim return 0; 8659f32624bSJoonsoo Kim 8665ad64688SHugh Dickins if (!is_locked && (!PageAnon(page) || PageKsm(page))) { 8675ad64688SHugh Dickins we_locked = trylock_page(page); 8689f32624bSJoonsoo Kim if (!we_locked) 8699f32624bSJoonsoo Kim return 1; 8705ad64688SHugh Dickins } 8719f32624bSJoonsoo Kim 8729f32624bSJoonsoo Kim /* 8739f32624bSJoonsoo Kim * If we are reclaiming on behalf of a cgroup, skip 8749f32624bSJoonsoo Kim * counting on behalf of references from different 8759f32624bSJoonsoo Kim * cgroups 8769f32624bSJoonsoo Kim */ 8779f32624bSJoonsoo Kim if (memcg) { 8789f32624bSJoonsoo Kim rwc.invalid_vma = invalid_page_referenced_vma; 8795ad64688SHugh Dickins } 8809f32624bSJoonsoo Kim 881c24f386cSMinchan Kim rmap_walk(page, &rwc); 8829f32624bSJoonsoo Kim *vm_flags = pra.vm_flags; 8839f32624bSJoonsoo Kim 8845ad64688SHugh Dickins if (we_locked) 8851da177e4SLinus Torvalds unlock_page(page); 8869f32624bSJoonsoo Kim 8879f32624bSJoonsoo Kim return pra.referenced; 8881da177e4SLinus Torvalds } 8891da177e4SLinus Torvalds 890e4b82222SMinchan Kim static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma, 8919853a407SJoonsoo Kim unsigned long address, void *arg) 892d08b3851SPeter Zijlstra { 893f27176cfSKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 894f27176cfSKirill A. Shutemov .page = page, 895f27176cfSKirill A. Shutemov .vma = vma, 896f27176cfSKirill A. Shutemov .address = address, 897f27176cfSKirill A. Shutemov .flags = PVMW_SYNC, 898f27176cfSKirill A. Shutemov }; 899ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 9009853a407SJoonsoo Kim int *cleaned = arg; 901d08b3851SPeter Zijlstra 902369ea824SJérôme Glisse /* 903369ea824SJérôme Glisse * We have to assume the worse case ie pmd for invalidation. Note that 904369ea824SJérôme Glisse * the page can not be free from this function. 905369ea824SJérôme Glisse */ 9067269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 9077269f999SJérôme Glisse 0, vma, vma->vm_mm, address, 908a50b854eSMatthew Wilcox (Oracle) min(vma->vm_end, address + page_size(page))); 909ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 910369ea824SJérôme Glisse 911f27176cfSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 912f27176cfSKirill A. Shutemov int ret = 0; 913369ea824SJérôme Glisse 9141f18b296SYueHaibing address = pvmw.address; 915f27176cfSKirill A. Shutemov if (pvmw.pte) { 916c2fda5feSPeter Zijlstra pte_t entry; 917f27176cfSKirill A. Shutemov pte_t *pte = pvmw.pte; 918f27176cfSKirill A. Shutemov 919f27176cfSKirill A. Shutemov if (!pte_dirty(*pte) && !pte_write(*pte)) 920f27176cfSKirill A. Shutemov continue; 921d08b3851SPeter Zijlstra 922785373b4SLinus Torvalds flush_cache_page(vma, address, pte_pfn(*pte)); 923785373b4SLinus Torvalds entry = ptep_clear_flush(vma, address, pte); 924d08b3851SPeter Zijlstra entry = pte_wrprotect(entry); 925c2fda5feSPeter Zijlstra entry = pte_mkclean(entry); 926785373b4SLinus Torvalds set_pte_at(vma->vm_mm, address, pte, entry); 927d08b3851SPeter Zijlstra ret = 1; 928f27176cfSKirill A. Shutemov } else { 929f27176cfSKirill A. Shutemov #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE 930f27176cfSKirill A. Shutemov pmd_t *pmd = pvmw.pmd; 931f27176cfSKirill A. Shutemov pmd_t entry; 932d08b3851SPeter Zijlstra 933f27176cfSKirill A. Shutemov if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) 934f27176cfSKirill A. Shutemov continue; 935f27176cfSKirill A. Shutemov 936785373b4SLinus Torvalds flush_cache_page(vma, address, page_to_pfn(page)); 937024eee0eSAneesh Kumar K.V entry = pmdp_invalidate(vma, address, pmd); 938f27176cfSKirill A. Shutemov entry = pmd_wrprotect(entry); 939f27176cfSKirill A. Shutemov entry = pmd_mkclean(entry); 940785373b4SLinus Torvalds set_pmd_at(vma->vm_mm, address, pmd, entry); 941f27176cfSKirill A. Shutemov ret = 1; 942f27176cfSKirill A. Shutemov #else 943f27176cfSKirill A. Shutemov /* unexpected pmd-mapped page? */ 944f27176cfSKirill A. Shutemov WARN_ON_ONCE(1); 945f27176cfSKirill A. Shutemov #endif 946f27176cfSKirill A. Shutemov } 9472ec74c3eSSagi Grimberg 9480f10851eSJérôme Glisse /* 9490f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() as we are 9500f10851eSJérôme Glisse * downgrading page table protection not changing it to point 9510f10851eSJérôme Glisse * to a new page. 9520f10851eSJérôme Glisse * 953ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 9540f10851eSJérôme Glisse */ 9550f10851eSJérôme Glisse if (ret) 9569853a407SJoonsoo Kim (*cleaned)++; 9579853a407SJoonsoo Kim } 958f27176cfSKirill A. Shutemov 959ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 960369ea824SJérôme Glisse 961e4b82222SMinchan Kim return true; 962d08b3851SPeter Zijlstra } 963d08b3851SPeter Zijlstra 9649853a407SJoonsoo Kim static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) 965d08b3851SPeter Zijlstra { 9669853a407SJoonsoo Kim if (vma->vm_flags & VM_SHARED) 967871beb8cSFengguang Wu return false; 968d08b3851SPeter Zijlstra 969871beb8cSFengguang Wu return true; 970d08b3851SPeter Zijlstra } 971d08b3851SPeter Zijlstra 972d08b3851SPeter Zijlstra int page_mkclean(struct page *page) 973d08b3851SPeter Zijlstra { 9749853a407SJoonsoo Kim int cleaned = 0; 9759853a407SJoonsoo Kim struct address_space *mapping; 9769853a407SJoonsoo Kim struct rmap_walk_control rwc = { 9779853a407SJoonsoo Kim .arg = (void *)&cleaned, 9789853a407SJoonsoo Kim .rmap_one = page_mkclean_one, 9799853a407SJoonsoo Kim .invalid_vma = invalid_mkclean_vma, 9809853a407SJoonsoo Kim }; 981d08b3851SPeter Zijlstra 982d08b3851SPeter Zijlstra BUG_ON(!PageLocked(page)); 983d08b3851SPeter Zijlstra 9849853a407SJoonsoo Kim if (!page_mapped(page)) 9859853a407SJoonsoo Kim return 0; 986d08b3851SPeter Zijlstra 9879853a407SJoonsoo Kim mapping = page_mapping(page); 9889853a407SJoonsoo Kim if (!mapping) 9899853a407SJoonsoo Kim return 0; 9909853a407SJoonsoo Kim 9919853a407SJoonsoo Kim rmap_walk(page, &rwc); 9929853a407SJoonsoo Kim 9939853a407SJoonsoo Kim return cleaned; 994d08b3851SPeter Zijlstra } 99560b59beaSJaya Kumar EXPORT_SYMBOL_GPL(page_mkclean); 996d08b3851SPeter Zijlstra 9971da177e4SLinus Torvalds /** 998c44b6743SRik van Riel * page_move_anon_rmap - move a page to our anon_vma 999c44b6743SRik van Riel * @page: the page to move to our anon_vma 1000c44b6743SRik van Riel * @vma: the vma the page belongs to 1001c44b6743SRik van Riel * 1002c44b6743SRik van Riel * When a page belongs exclusively to one process after a COW event, 1003c44b6743SRik van Riel * that page can be moved into the anon_vma that belongs to just that 1004c44b6743SRik van Riel * process, so the rmap code will not search the parent or sibling 1005c44b6743SRik van Riel * processes. 1006c44b6743SRik van Riel */ 10075a49973dSHugh Dickins void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma) 1008c44b6743SRik van Riel { 1009c44b6743SRik van Riel struct anon_vma *anon_vma = vma->anon_vma; 1010c44b6743SRik van Riel 10115a49973dSHugh Dickins page = compound_head(page); 10125a49973dSHugh Dickins 1013309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 101481d1b09cSSasha Levin VM_BUG_ON_VMA(!anon_vma, vma); 1015c44b6743SRik van Riel 1016c44b6743SRik van Riel anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1017414e2fb8SVladimir Davydov /* 1018414e2fb8SVladimir Davydov * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written 1019414e2fb8SVladimir Davydov * simultaneously, so a concurrent reader (eg page_referenced()'s 1020414e2fb8SVladimir Davydov * PageAnon()) will not see one without the other. 1021414e2fb8SVladimir Davydov */ 1022414e2fb8SVladimir Davydov WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); 1023c44b6743SRik van Riel } 1024c44b6743SRik van Riel 1025c44b6743SRik van Riel /** 102643d8eac4SRandy Dunlap * __page_set_anon_rmap - set up new anonymous rmap 1027451b9514SKirill Tkhai * @page: Page or Hugepage to add to rmap 10284e1c1975SAndi Kleen * @vma: VM area to add page to. 10294e1c1975SAndi Kleen * @address: User virtual address of the mapping 1030e8a03febSRik van Riel * @exclusive: the page is exclusively owned by the current process 10311da177e4SLinus Torvalds */ 10329617d95eSNick Piggin static void __page_set_anon_rmap(struct page *page, 1033e8a03febSRik van Riel struct vm_area_struct *vma, unsigned long address, int exclusive) 10341da177e4SLinus Torvalds { 1035e8a03febSRik van Riel struct anon_vma *anon_vma = vma->anon_vma; 10362822c1aaSNick Piggin 1037e8a03febSRik van Riel BUG_ON(!anon_vma); 1038ea90002bSLinus Torvalds 10394e1c1975SAndi Kleen if (PageAnon(page)) 10404e1c1975SAndi Kleen return; 10414e1c1975SAndi Kleen 1042ea90002bSLinus Torvalds /* 1043e8a03febSRik van Riel * If the page isn't exclusively mapped into this vma, 1044e8a03febSRik van Riel * we must use the _oldest_ possible anon_vma for the 1045e8a03febSRik van Riel * page mapping! 1046ea90002bSLinus Torvalds */ 10474e1c1975SAndi Kleen if (!exclusive) 1048288468c3SAndrea Arcangeli anon_vma = anon_vma->root; 1049ea90002bSLinus Torvalds 10501da177e4SLinus Torvalds anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 10512822c1aaSNick Piggin page->mapping = (struct address_space *) anon_vma; 10524d7670e0SNick Piggin page->index = linear_page_index(vma, address); 10531da177e4SLinus Torvalds } 10549617d95eSNick Piggin 10559617d95eSNick Piggin /** 105643d8eac4SRandy Dunlap * __page_check_anon_rmap - sanity check anonymous rmap addition 1057c97a9e10SNick Piggin * @page: the page to add the mapping to 1058c97a9e10SNick Piggin * @vma: the vm area in which the mapping is added 1059c97a9e10SNick Piggin * @address: the user virtual address mapped 1060c97a9e10SNick Piggin */ 1061c97a9e10SNick Piggin static void __page_check_anon_rmap(struct page *page, 1062c97a9e10SNick Piggin struct vm_area_struct *vma, unsigned long address) 1063c97a9e10SNick Piggin { 1064c97a9e10SNick Piggin #ifdef CONFIG_DEBUG_VM 1065c97a9e10SNick Piggin /* 1066c97a9e10SNick Piggin * The page's anon-rmap details (mapping and index) are guaranteed to 1067c97a9e10SNick Piggin * be set up correctly at this point. 1068c97a9e10SNick Piggin * 1069c97a9e10SNick Piggin * We have exclusion against page_add_anon_rmap because the caller 1070c97a9e10SNick Piggin * always holds the page locked, except if called from page_dup_rmap, 1071c97a9e10SNick Piggin * in which case the page is already known to be setup. 1072c97a9e10SNick Piggin * 1073c97a9e10SNick Piggin * We have exclusion against page_add_new_anon_rmap because those pages 1074c97a9e10SNick Piggin * are initially only visible via the pagetables, and the pte is locked 1075c97a9e10SNick Piggin * over the call to page_add_new_anon_rmap. 1076c97a9e10SNick Piggin */ 107744ab57a0SAndrea Arcangeli BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); 107853f9263bSKirill A. Shutemov BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address)); 1079c97a9e10SNick Piggin #endif 1080c97a9e10SNick Piggin } 1081c97a9e10SNick Piggin 1082c97a9e10SNick Piggin /** 10839617d95eSNick Piggin * page_add_anon_rmap - add pte mapping to an anonymous page 10849617d95eSNick Piggin * @page: the page to add the mapping to 10859617d95eSNick Piggin * @vma: the vm area in which the mapping is added 10869617d95eSNick Piggin * @address: the user virtual address mapped 1087d281ee61SKirill A. Shutemov * @compound: charge the page as compound or small page 10889617d95eSNick Piggin * 10895ad64688SHugh Dickins * The caller needs to hold the pte lock, and the page must be locked in 109080e14822SHugh Dickins * the anon_vma case: to serialize mapping,index checking after setting, 109180e14822SHugh Dickins * and to ensure that PageAnon is not being upgraded racily to PageKsm 109280e14822SHugh Dickins * (but PageKsm is never downgraded to PageAnon). 10939617d95eSNick Piggin */ 10949617d95eSNick Piggin void page_add_anon_rmap(struct page *page, 1095d281ee61SKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, bool compound) 10969617d95eSNick Piggin { 1097d281ee61SKirill A. Shutemov do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0); 1098ad8c2ee8SRik van Riel } 1099ad8c2ee8SRik van Riel 1100ad8c2ee8SRik van Riel /* 1101ad8c2ee8SRik van Riel * Special version of the above for do_swap_page, which often runs 1102ad8c2ee8SRik van Riel * into pages that are exclusively owned by the current process. 1103ad8c2ee8SRik van Riel * Everybody else should continue to use page_add_anon_rmap above. 1104ad8c2ee8SRik van Riel */ 1105ad8c2ee8SRik van Riel void do_page_add_anon_rmap(struct page *page, 1106d281ee61SKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, int flags) 1107ad8c2ee8SRik van Riel { 1108d281ee61SKirill A. Shutemov bool compound = flags & RMAP_COMPOUND; 110953f9263bSKirill A. Shutemov bool first; 111053f9263bSKirill A. Shutemov 111153f9263bSKirill A. Shutemov if (compound) { 111253f9263bSKirill A. Shutemov atomic_t *mapcount; 1113e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 111453f9263bSKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 111553f9263bSKirill A. Shutemov mapcount = compound_mapcount_ptr(page); 111653f9263bSKirill A. Shutemov first = atomic_inc_and_test(mapcount); 111753f9263bSKirill A. Shutemov } else { 111853f9263bSKirill A. Shutemov first = atomic_inc_and_test(&page->_mapcount); 111953f9263bSKirill A. Shutemov } 112053f9263bSKirill A. Shutemov 112153f9263bSKirill A. Shutemov if (first) { 1122d281ee61SKirill A. Shutemov int nr = compound ? hpage_nr_pages(page) : 1; 1123bea04b07SJianyu Zhan /* 1124bea04b07SJianyu Zhan * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1125bea04b07SJianyu Zhan * these counters are not modified in interrupt context, and 1126bea04b07SJianyu Zhan * pte lock(a spinlock) is held, which implies preemption 1127bea04b07SJianyu Zhan * disabled. 1128bea04b07SJianyu Zhan */ 112965c45377SKirill A. Shutemov if (compound) 113011fb9989SMel Gorman __inc_node_page_state(page, NR_ANON_THPS); 11314b9d0fabSMel Gorman __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); 113279134171SAndrea Arcangeli } 11335ad64688SHugh Dickins if (unlikely(PageKsm(page))) 11345ad64688SHugh Dickins return; 11355ad64688SHugh Dickins 1136309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 113753f9263bSKirill A. Shutemov 11385dbe0af4SHugh Dickins /* address might be in next vma when migration races vma_adjust */ 11395ad64688SHugh Dickins if (first) 1140d281ee61SKirill A. Shutemov __page_set_anon_rmap(page, vma, address, 1141d281ee61SKirill A. Shutemov flags & RMAP_EXCLUSIVE); 114269029cd5SKAMEZAWA Hiroyuki else 1143c97a9e10SNick Piggin __page_check_anon_rmap(page, vma, address); 11441da177e4SLinus Torvalds } 11451da177e4SLinus Torvalds 114643d8eac4SRandy Dunlap /** 11479617d95eSNick Piggin * page_add_new_anon_rmap - add pte mapping to a new anonymous page 11489617d95eSNick Piggin * @page: the page to add the mapping to 11499617d95eSNick Piggin * @vma: the vm area in which the mapping is added 11509617d95eSNick Piggin * @address: the user virtual address mapped 1151d281ee61SKirill A. Shutemov * @compound: charge the page as compound or small page 11529617d95eSNick Piggin * 11539617d95eSNick Piggin * Same as page_add_anon_rmap but must only be called on *new* pages. 11549617d95eSNick Piggin * This means the inc-and-test can be bypassed. 1155c97a9e10SNick Piggin * Page does not have to be locked. 11569617d95eSNick Piggin */ 11579617d95eSNick Piggin void page_add_new_anon_rmap(struct page *page, 1158d281ee61SKirill A. Shutemov struct vm_area_struct *vma, unsigned long address, bool compound) 11599617d95eSNick Piggin { 1160d281ee61SKirill A. Shutemov int nr = compound ? hpage_nr_pages(page) : 1; 1161d281ee61SKirill A. Shutemov 116281d1b09cSSasha Levin VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); 1163fa9949daSHugh Dickins __SetPageSwapBacked(page); 1164d281ee61SKirill A. Shutemov if (compound) { 1165d281ee61SKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 116653f9263bSKirill A. Shutemov /* increment count (starts at -1) */ 116753f9263bSKirill A. Shutemov atomic_set(compound_mapcount_ptr(page), 0); 116811fb9989SMel Gorman __inc_node_page_state(page, NR_ANON_THPS); 116953f9263bSKirill A. Shutemov } else { 117053f9263bSKirill A. Shutemov /* Anon THP always mapped first with PMD */ 117153f9263bSKirill A. Shutemov VM_BUG_ON_PAGE(PageTransCompound(page), page); 117253f9263bSKirill A. Shutemov /* increment count (starts at -1) */ 117353f9263bSKirill A. Shutemov atomic_set(&page->_mapcount, 0); 1174d281ee61SKirill A. Shutemov } 11754b9d0fabSMel Gorman __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); 1176e8a03febSRik van Riel __page_set_anon_rmap(page, vma, address, 1); 11779617d95eSNick Piggin } 11789617d95eSNick Piggin 11791da177e4SLinus Torvalds /** 11801da177e4SLinus Torvalds * page_add_file_rmap - add pte mapping to a file page 11811da177e4SLinus Torvalds * @page: the page to add the mapping to 1182e8b098fcSMike Rapoport * @compound: charge the page as compound or small page 11831da177e4SLinus Torvalds * 1184b8072f09SHugh Dickins * The caller needs to hold the pte lock. 11851da177e4SLinus Torvalds */ 1186dd78feddSKirill A. Shutemov void page_add_file_rmap(struct page *page, bool compound) 11871da177e4SLinus Torvalds { 1188dd78feddSKirill A. Shutemov int i, nr = 1; 1189dd78feddSKirill A. Shutemov 1190dd78feddSKirill A. Shutemov VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page); 119162cccb8cSJohannes Weiner lock_page_memcg(page); 1192dd78feddSKirill A. Shutemov if (compound && PageTransHuge(page)) { 1193dd78feddSKirill A. Shutemov for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { 1194dd78feddSKirill A. Shutemov if (atomic_inc_and_test(&page[i]._mapcount)) 1195dd78feddSKirill A. Shutemov nr++; 1196d69b042fSBalbir Singh } 1197dd78feddSKirill A. Shutemov if (!atomic_inc_and_test(compound_mapcount_ptr(page))) 1198dd78feddSKirill A. Shutemov goto out; 119999cb0dbdSSong Liu if (PageSwapBacked(page)) 120011fb9989SMel Gorman __inc_node_page_state(page, NR_SHMEM_PMDMAPPED); 120199cb0dbdSSong Liu else 120299cb0dbdSSong Liu __inc_node_page_state(page, NR_FILE_PMDMAPPED); 1203dd78feddSKirill A. Shutemov } else { 1204c8efc390SKirill A. Shutemov if (PageTransCompound(page) && page_mapping(page)) { 1205c8efc390SKirill A. Shutemov VM_WARN_ON_ONCE(!PageLocked(page)); 1206c8efc390SKirill A. Shutemov 12079a73f61bSKirill A. Shutemov SetPageDoubleMap(compound_head(page)); 12089a73f61bSKirill A. Shutemov if (PageMlocked(page)) 12099a73f61bSKirill A. Shutemov clear_page_mlock(compound_head(page)); 12109a73f61bSKirill A. Shutemov } 1211dd78feddSKirill A. Shutemov if (!atomic_inc_and_test(&page->_mapcount)) 1212dd78feddSKirill A. Shutemov goto out; 1213dd78feddSKirill A. Shutemov } 121400f3ca2cSJohannes Weiner __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr); 1215dd78feddSKirill A. Shutemov out: 121662cccb8cSJohannes Weiner unlock_page_memcg(page); 12171da177e4SLinus Torvalds } 12181da177e4SLinus Torvalds 1219dd78feddSKirill A. Shutemov static void page_remove_file_rmap(struct page *page, bool compound) 12208186eb6aSJohannes Weiner { 1221dd78feddSKirill A. Shutemov int i, nr = 1; 1222dd78feddSKirill A. Shutemov 122357dea93aSSteve Capper VM_BUG_ON_PAGE(compound && !PageHead(page), page); 122462cccb8cSJohannes Weiner lock_page_memcg(page); 12258186eb6aSJohannes Weiner 122653f9263bSKirill A. Shutemov /* Hugepages are not counted in NR_FILE_MAPPED for now. */ 122753f9263bSKirill A. Shutemov if (unlikely(PageHuge(page))) { 122853f9263bSKirill A. Shutemov /* hugetlb pages are always mapped with pmds */ 122953f9263bSKirill A. Shutemov atomic_dec(compound_mapcount_ptr(page)); 123053f9263bSKirill A. Shutemov goto out; 123153f9263bSKirill A. Shutemov } 123253f9263bSKirill A. Shutemov 12338186eb6aSJohannes Weiner /* page still mapped by someone else? */ 1234dd78feddSKirill A. Shutemov if (compound && PageTransHuge(page)) { 1235dd78feddSKirill A. Shutemov for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { 1236dd78feddSKirill A. Shutemov if (atomic_add_negative(-1, &page[i]._mapcount)) 1237dd78feddSKirill A. Shutemov nr++; 1238dd78feddSKirill A. Shutemov } 1239dd78feddSKirill A. Shutemov if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 1240dd78feddSKirill A. Shutemov goto out; 124199cb0dbdSSong Liu if (PageSwapBacked(page)) 124211fb9989SMel Gorman __dec_node_page_state(page, NR_SHMEM_PMDMAPPED); 124399cb0dbdSSong Liu else 124499cb0dbdSSong Liu __dec_node_page_state(page, NR_FILE_PMDMAPPED); 1245dd78feddSKirill A. Shutemov } else { 12468186eb6aSJohannes Weiner if (!atomic_add_negative(-1, &page->_mapcount)) 12478186eb6aSJohannes Weiner goto out; 1248dd78feddSKirill A. Shutemov } 12498186eb6aSJohannes Weiner 12508186eb6aSJohannes Weiner /* 125100f3ca2cSJohannes Weiner * We use the irq-unsafe __{inc|mod}_lruvec_page_state because 12528186eb6aSJohannes Weiner * these counters are not modified in interrupt context, and 12538186eb6aSJohannes Weiner * pte lock(a spinlock) is held, which implies preemption disabled. 12548186eb6aSJohannes Weiner */ 125500f3ca2cSJohannes Weiner __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr); 12568186eb6aSJohannes Weiner 12578186eb6aSJohannes Weiner if (unlikely(PageMlocked(page))) 12588186eb6aSJohannes Weiner clear_page_mlock(page); 12598186eb6aSJohannes Weiner out: 126062cccb8cSJohannes Weiner unlock_page_memcg(page); 12618186eb6aSJohannes Weiner } 12628186eb6aSJohannes Weiner 126353f9263bSKirill A. Shutemov static void page_remove_anon_compound_rmap(struct page *page) 126453f9263bSKirill A. Shutemov { 126553f9263bSKirill A. Shutemov int i, nr; 126653f9263bSKirill A. Shutemov 126753f9263bSKirill A. Shutemov if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 126853f9263bSKirill A. Shutemov return; 126953f9263bSKirill A. Shutemov 127053f9263bSKirill A. Shutemov /* Hugepages are not counted in NR_ANON_PAGES for now. */ 127153f9263bSKirill A. Shutemov if (unlikely(PageHuge(page))) 127253f9263bSKirill A. Shutemov return; 127353f9263bSKirill A. Shutemov 127453f9263bSKirill A. Shutemov if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) 127553f9263bSKirill A. Shutemov return; 127653f9263bSKirill A. Shutemov 127711fb9989SMel Gorman __dec_node_page_state(page, NR_ANON_THPS); 127853f9263bSKirill A. Shutemov 127953f9263bSKirill A. Shutemov if (TestClearPageDoubleMap(page)) { 128053f9263bSKirill A. Shutemov /* 128153f9263bSKirill A. Shutemov * Subpages can be mapped with PTEs too. Check how many of 128253f9263bSKirill A. Shutemov * themi are still mapped. 128353f9263bSKirill A. Shutemov */ 128453f9263bSKirill A. Shutemov for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { 128553f9263bSKirill A. Shutemov if (atomic_add_negative(-1, &page[i]._mapcount)) 128653f9263bSKirill A. Shutemov nr++; 128753f9263bSKirill A. Shutemov } 128853f9263bSKirill A. Shutemov } else { 128953f9263bSKirill A. Shutemov nr = HPAGE_PMD_NR; 129053f9263bSKirill A. Shutemov } 129153f9263bSKirill A. Shutemov 1292e90309c9SKirill A. Shutemov if (unlikely(PageMlocked(page))) 1293e90309c9SKirill A. Shutemov clear_page_mlock(page); 1294e90309c9SKirill A. Shutemov 12959a982250SKirill A. Shutemov if (nr) { 12964b9d0fabSMel Gorman __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr); 12979a982250SKirill A. Shutemov deferred_split_huge_page(page); 12989a982250SKirill A. Shutemov } 129953f9263bSKirill A. Shutemov } 130053f9263bSKirill A. Shutemov 13011da177e4SLinus Torvalds /** 13021da177e4SLinus Torvalds * page_remove_rmap - take down pte mapping from a page 13031da177e4SLinus Torvalds * @page: page to remove mapping from 1304d281ee61SKirill A. Shutemov * @compound: uncharge the page as compound or small page 13051da177e4SLinus Torvalds * 1306b8072f09SHugh Dickins * The caller needs to hold the pte lock. 13071da177e4SLinus Torvalds */ 1308d281ee61SKirill A. Shutemov void page_remove_rmap(struct page *page, bool compound) 13091da177e4SLinus Torvalds { 1310dd78feddSKirill A. Shutemov if (!PageAnon(page)) 1311dd78feddSKirill A. Shutemov return page_remove_file_rmap(page, compound); 131289c06bd5SKAMEZAWA Hiroyuki 131353f9263bSKirill A. Shutemov if (compound) 131453f9263bSKirill A. Shutemov return page_remove_anon_compound_rmap(page); 131553f9263bSKirill A. Shutemov 1316b904dcfeSKOSAKI Motohiro /* page still mapped by someone else? */ 1317b904dcfeSKOSAKI Motohiro if (!atomic_add_negative(-1, &page->_mapcount)) 13188186eb6aSJohannes Weiner return; 13198186eb6aSJohannes Weiner 13201da177e4SLinus Torvalds /* 1321bea04b07SJianyu Zhan * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1322bea04b07SJianyu Zhan * these counters are not modified in interrupt context, and 1323bea04b07SJianyu Zhan * pte lock(a spinlock) is held, which implies preemption disabled. 13240fe6e20bSNaoya Horiguchi */ 13254b9d0fabSMel Gorman __dec_node_page_state(page, NR_ANON_MAPPED); 13268186eb6aSJohannes Weiner 1327e6c509f8SHugh Dickins if (unlikely(PageMlocked(page))) 1328e6c509f8SHugh Dickins clear_page_mlock(page); 13298186eb6aSJohannes Weiner 13309a982250SKirill A. Shutemov if (PageTransCompound(page)) 13319a982250SKirill A. Shutemov deferred_split_huge_page(compound_head(page)); 13329a982250SKirill A. Shutemov 133316f8c5b2SHugh Dickins /* 13341da177e4SLinus Torvalds * It would be tidy to reset the PageAnon mapping here, 13351da177e4SLinus Torvalds * but that might overwrite a racing page_add_anon_rmap 13361da177e4SLinus Torvalds * which increments mapcount after us but sets mapping 13372d4894b5SMel Gorman * before us: so leave the reset to free_unref_page, 13381da177e4SLinus Torvalds * and remember that it's only reliable while mapped. 13391da177e4SLinus Torvalds * Leaving it set also helps swapoff to reinstate ptes 13401da177e4SLinus Torvalds * faster for those pages still in swapcache. 13411da177e4SLinus Torvalds */ 13421da177e4SLinus Torvalds } 13431da177e4SLinus Torvalds 13441da177e4SLinus Torvalds /* 134552629506SJoonsoo Kim * @arg: enum ttu_flags will be passed to this argument 13461da177e4SLinus Torvalds */ 1347e4b82222SMinchan Kim static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma, 134852629506SJoonsoo Kim unsigned long address, void *arg) 13491da177e4SLinus Torvalds { 13501da177e4SLinus Torvalds struct mm_struct *mm = vma->vm_mm; 1351c7ab0d2fSKirill A. Shutemov struct page_vma_mapped_walk pvmw = { 1352c7ab0d2fSKirill A. Shutemov .page = page, 1353c7ab0d2fSKirill A. Shutemov .vma = vma, 1354c7ab0d2fSKirill A. Shutemov .address = address, 1355c7ab0d2fSKirill A. Shutemov }; 13561da177e4SLinus Torvalds pte_t pteval; 1357c7ab0d2fSKirill A. Shutemov struct page *subpage; 1358785373b4SLinus Torvalds bool ret = true; 1359ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 1360802a3a92SShaohua Li enum ttu_flags flags = (enum ttu_flags)arg; 13611da177e4SLinus Torvalds 1362b87537d9SHugh Dickins /* munlock has nothing to gain from examining un-locked vmas */ 1363b87537d9SHugh Dickins if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED)) 1364e4b82222SMinchan Kim return true; 1365b87537d9SHugh Dickins 1366a5430ddaSJérôme Glisse if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) && 1367a5430ddaSJérôme Glisse is_zone_device_page(page) && !is_device_private_page(page)) 1368a5430ddaSJérôme Glisse return true; 1369a5430ddaSJérôme Glisse 1370fec89c10SKirill A. Shutemov if (flags & TTU_SPLIT_HUGE_PMD) { 1371fec89c10SKirill A. Shutemov split_huge_pmd_address(vma, address, 1372b5ff8161SNaoya Horiguchi flags & TTU_SPLIT_FREEZE, page); 1373fec89c10SKirill A. Shutemov } 1374fec89c10SKirill A. Shutemov 1375369ea824SJérôme Glisse /* 1376017b1660SMike Kravetz * For THP, we have to assume the worse case ie pmd for invalidation. 1377017b1660SMike Kravetz * For hugetlb, it could be much worse if we need to do pud 1378017b1660SMike Kravetz * invalidation in the case of pmd sharing. 1379017b1660SMike Kravetz * 1380017b1660SMike Kravetz * Note that the page can not be free in this function as call of 1381017b1660SMike Kravetz * try_to_unmap() must hold a reference on the page. 1382369ea824SJérôme Glisse */ 13837269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 13846f4f13e8SJérôme Glisse address, 1385a50b854eSMatthew Wilcox (Oracle) min(vma->vm_end, address + page_size(page))); 1386017b1660SMike Kravetz if (PageHuge(page)) { 1387017b1660SMike Kravetz /* 1388017b1660SMike Kravetz * If sharing is possible, start and end will be adjusted 1389017b1660SMike Kravetz * accordingly. 1390017b1660SMike Kravetz */ 1391ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, 1392ac46d4f3SJérôme Glisse &range.end); 1393017b1660SMike Kravetz } 1394ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1395369ea824SJérôme Glisse 1396c7ab0d2fSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 1397616b8371SZi Yan #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1398616b8371SZi Yan /* PMD-mapped THP migration entry */ 1399616b8371SZi Yan if (!pvmw.pte && (flags & TTU_MIGRATION)) { 1400616b8371SZi Yan VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); 1401616b8371SZi Yan 1402616b8371SZi Yan set_pmd_migration_entry(&pvmw, page); 1403616b8371SZi Yan continue; 1404616b8371SZi Yan } 1405616b8371SZi Yan #endif 1406616b8371SZi Yan 14071da177e4SLinus Torvalds /* 14081da177e4SLinus Torvalds * If the page is mlock()d, we cannot swap it out. 14091da177e4SLinus Torvalds * If it's recently referenced (perhaps page_referenced 14101da177e4SLinus Torvalds * skipped over this mm) then we should reactivate it. 14111da177e4SLinus Torvalds */ 141214fa31b8SAndi Kleen if (!(flags & TTU_IGNORE_MLOCK)) { 1413b87537d9SHugh Dickins if (vma->vm_flags & VM_LOCKED) { 14149a73f61bSKirill A. Shutemov /* PTE-mapped THP are never mlocked */ 14159a73f61bSKirill A. Shutemov if (!PageTransCompound(page)) { 14169a73f61bSKirill A. Shutemov /* 14179a73f61bSKirill A. Shutemov * Holding pte lock, we do *not* need 14189a73f61bSKirill A. Shutemov * mmap_sem here 14199a73f61bSKirill A. Shutemov */ 1420b87537d9SHugh Dickins mlock_vma_page(page); 14219a73f61bSKirill A. Shutemov } 1422e4b82222SMinchan Kim ret = false; 1423c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1424c7ab0d2fSKirill A. Shutemov break; 1425b87537d9SHugh Dickins } 1426daa5ba76SKonstantin Khlebnikov if (flags & TTU_MUNLOCK) 1427c7ab0d2fSKirill A. Shutemov continue; 142814fa31b8SAndi Kleen } 1429c7ab0d2fSKirill A. Shutemov 14308346242aSKirill A. Shutemov /* Unexpected PMD-mapped THP? */ 14318346242aSKirill A. Shutemov VM_BUG_ON_PAGE(!pvmw.pte, page); 14328346242aSKirill A. Shutemov 14338346242aSKirill A. Shutemov subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte); 1434785373b4SLinus Torvalds address = pvmw.address; 1435785373b4SLinus Torvalds 1436017b1660SMike Kravetz if (PageHuge(page)) { 1437017b1660SMike Kravetz if (huge_pmd_unshare(mm, &address, pvmw.pte)) { 1438017b1660SMike Kravetz /* 1439017b1660SMike Kravetz * huge_pmd_unshare unmapped an entire PMD 1440017b1660SMike Kravetz * page. There is no way of knowing exactly 1441017b1660SMike Kravetz * which PMDs may be cached for this mm, so 1442017b1660SMike Kravetz * we must flush them all. start/end were 1443017b1660SMike Kravetz * already adjusted above to cover this range. 1444017b1660SMike Kravetz */ 1445ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 1446ac46d4f3SJérôme Glisse flush_tlb_range(vma, range.start, range.end); 1447ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, 1448ac46d4f3SJérôme Glisse range.end); 1449017b1660SMike Kravetz 1450017b1660SMike Kravetz /* 1451017b1660SMike Kravetz * The ref count of the PMD page was dropped 1452017b1660SMike Kravetz * which is part of the way map counting 1453017b1660SMike Kravetz * is done for shared PMDs. Return 'true' 1454017b1660SMike Kravetz * here. When there is no other sharing, 1455017b1660SMike Kravetz * huge_pmd_unshare returns false and we will 1456017b1660SMike Kravetz * unmap the actual page and drop map count 1457017b1660SMike Kravetz * to zero. 1458017b1660SMike Kravetz */ 1459017b1660SMike Kravetz page_vma_mapped_walk_done(&pvmw); 1460017b1660SMike Kravetz break; 1461017b1660SMike Kravetz } 1462017b1660SMike Kravetz } 14638346242aSKirill A. Shutemov 1464a5430ddaSJérôme Glisse if (IS_ENABLED(CONFIG_MIGRATION) && 1465a5430ddaSJérôme Glisse (flags & TTU_MIGRATION) && 1466a5430ddaSJérôme Glisse is_zone_device_page(page)) { 1467a5430ddaSJérôme Glisse swp_entry_t entry; 1468a5430ddaSJérôme Glisse pte_t swp_pte; 1469a5430ddaSJérôme Glisse 1470a5430ddaSJérôme Glisse pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte); 1471a5430ddaSJérôme Glisse 1472a5430ddaSJérôme Glisse /* 1473a5430ddaSJérôme Glisse * Store the pfn of the page in a special migration 1474a5430ddaSJérôme Glisse * pte. do_swap_page() will wait until the migration 1475a5430ddaSJérôme Glisse * pte is removed and then restart fault handling. 1476a5430ddaSJérôme Glisse */ 1477a5430ddaSJérôme Glisse entry = make_migration_entry(page, 0); 1478a5430ddaSJérôme Glisse swp_pte = swp_entry_to_pte(entry); 1479a5430ddaSJérôme Glisse if (pte_soft_dirty(pteval)) 1480a5430ddaSJérôme Glisse swp_pte = pte_swp_mksoft_dirty(swp_pte); 1481a5430ddaSJérôme Glisse set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte); 14820f10851eSJérôme Glisse /* 14830f10851eSJérôme Glisse * No need to invalidate here it will synchronize on 14840f10851eSJérôme Glisse * against the special swap migration pte. 14851de13ee5SRalph Campbell * 14861de13ee5SRalph Campbell * The assignment to subpage above was computed from a 14871de13ee5SRalph Campbell * swap PTE which results in an invalid pointer. 14881de13ee5SRalph Campbell * Since only PAGE_SIZE pages can currently be 14891de13ee5SRalph Campbell * migrated, just set it to page. This will need to be 14901de13ee5SRalph Campbell * changed when hugepage migrations to device private 14911de13ee5SRalph Campbell * memory are supported. 14920f10851eSJérôme Glisse */ 14931de13ee5SRalph Campbell subpage = page; 1494a5430ddaSJérôme Glisse goto discard; 1495a5430ddaSJérôme Glisse } 1496a5430ddaSJérôme Glisse 149714fa31b8SAndi Kleen if (!(flags & TTU_IGNORE_ACCESS)) { 1498785373b4SLinus Torvalds if (ptep_clear_flush_young_notify(vma, address, 1499c7ab0d2fSKirill A. Shutemov pvmw.pte)) { 1500e4b82222SMinchan Kim ret = false; 1501c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1502c7ab0d2fSKirill A. Shutemov break; 15031da177e4SLinus Torvalds } 1504b291f000SNick Piggin } 15051da177e4SLinus Torvalds 15061da177e4SLinus Torvalds /* Nuke the page table entry. */ 1507785373b4SLinus Torvalds flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); 150872b252aeSMel Gorman if (should_defer_flush(mm, flags)) { 150972b252aeSMel Gorman /* 1510c7ab0d2fSKirill A. Shutemov * We clear the PTE but do not flush so potentially 1511c7ab0d2fSKirill A. Shutemov * a remote CPU could still be writing to the page. 1512c7ab0d2fSKirill A. Shutemov * If the entry was previously clean then the 1513c7ab0d2fSKirill A. Shutemov * architecture must guarantee that a clear->dirty 1514c7ab0d2fSKirill A. Shutemov * transition on a cached TLB entry is written through 1515c7ab0d2fSKirill A. Shutemov * and traps if the PTE is unmapped. 151672b252aeSMel Gorman */ 1517785373b4SLinus Torvalds pteval = ptep_get_and_clear(mm, address, pvmw.pte); 151872b252aeSMel Gorman 1519c7ab0d2fSKirill A. Shutemov set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); 152072b252aeSMel Gorman } else { 1521785373b4SLinus Torvalds pteval = ptep_clear_flush(vma, address, pvmw.pte); 152272b252aeSMel Gorman } 15231da177e4SLinus Torvalds 1524c7ab0d2fSKirill A. Shutemov /* Move the dirty bit to the page. Now the pte is gone. */ 15251da177e4SLinus Torvalds if (pte_dirty(pteval)) 15261da177e4SLinus Torvalds set_page_dirty(page); 15271da177e4SLinus Torvalds 1528365e9c87SHugh Dickins /* Update high watermark before we lower rss */ 1529365e9c87SHugh Dickins update_hiwater_rss(mm); 1530365e9c87SHugh Dickins 1531888b9f7cSAndi Kleen if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { 15325fd27b8eSPunit Agrawal pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); 15335d317b2bSNaoya Horiguchi if (PageHuge(page)) { 1534d8c6546bSMatthew Wilcox (Oracle) hugetlb_count_sub(compound_nr(page), mm); 1535785373b4SLinus Torvalds set_huge_swap_pte_at(mm, address, 15365fd27b8eSPunit Agrawal pvmw.pte, pteval, 15375fd27b8eSPunit Agrawal vma_mmu_pagesize(vma)); 15385d317b2bSNaoya Horiguchi } else { 1539eca56ff9SJerome Marchand dec_mm_counter(mm, mm_counter(page)); 1540785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 15415f24ae58SNaoya Horiguchi } 1542c7ab0d2fSKirill A. Shutemov 1543bce73e48SChristian Borntraeger } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { 154445961722SKonstantin Weitz /* 154545961722SKonstantin Weitz * The guest indicated that the page content is of no 154645961722SKonstantin Weitz * interest anymore. Simply discard the pte, vmscan 154745961722SKonstantin Weitz * will take care of the rest. 1548bce73e48SChristian Borntraeger * A future reference will then fault in a new zero 1549bce73e48SChristian Borntraeger * page. When userfaultfd is active, we must not drop 1550bce73e48SChristian Borntraeger * this page though, as its main user (postcopy 1551bce73e48SChristian Borntraeger * migration) will not expect userfaults on already 1552bce73e48SChristian Borntraeger * copied pages. 155345961722SKonstantin Weitz */ 1554eca56ff9SJerome Marchand dec_mm_counter(mm, mm_counter(page)); 15550f10851eSJérôme Glisse /* We have to invalidate as we cleared the pte */ 15560f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, address, 15570f10851eSJérôme Glisse address + PAGE_SIZE); 1558c7ab0d2fSKirill A. Shutemov } else if (IS_ENABLED(CONFIG_MIGRATION) && 1559b5ff8161SNaoya Horiguchi (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) { 1560470f119fSHugh Dickins swp_entry_t entry; 1561470f119fSHugh Dickins pte_t swp_pte; 1562ca827d55SKhalid Aziz 1563ca827d55SKhalid Aziz if (arch_unmap_one(mm, vma, address, pteval) < 0) { 1564ca827d55SKhalid Aziz set_pte_at(mm, address, pvmw.pte, pteval); 1565ca827d55SKhalid Aziz ret = false; 1566ca827d55SKhalid Aziz page_vma_mapped_walk_done(&pvmw); 1567ca827d55SKhalid Aziz break; 1568ca827d55SKhalid Aziz } 1569ca827d55SKhalid Aziz 1570470f119fSHugh Dickins /* 1571470f119fSHugh Dickins * Store the pfn of the page in a special migration 1572470f119fSHugh Dickins * pte. do_swap_page() will wait until the migration 1573470f119fSHugh Dickins * pte is removed and then restart fault handling. 1574470f119fSHugh Dickins */ 1575c7ab0d2fSKirill A. Shutemov entry = make_migration_entry(subpage, 1576c7ab0d2fSKirill A. Shutemov pte_write(pteval)); 1577470f119fSHugh Dickins swp_pte = swp_entry_to_pte(entry); 1578470f119fSHugh Dickins if (pte_soft_dirty(pteval)) 1579470f119fSHugh Dickins swp_pte = pte_swp_mksoft_dirty(swp_pte); 1580785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, swp_pte); 15810f10851eSJérôme Glisse /* 15820f10851eSJérôme Glisse * No need to invalidate here it will synchronize on 15830f10851eSJérôme Glisse * against the special swap migration pte. 15840f10851eSJérôme Glisse */ 1585888b9f7cSAndi Kleen } else if (PageAnon(page)) { 1586c7ab0d2fSKirill A. Shutemov swp_entry_t entry = { .val = page_private(subpage) }; 1587179ef71cSCyrill Gorcunov pte_t swp_pte; 15881da177e4SLinus Torvalds /* 15891da177e4SLinus Torvalds * Store the swap location in the pte. 15901da177e4SLinus Torvalds * See handle_pte_fault() ... 15911da177e4SLinus Torvalds */ 1592eb94a878SMinchan Kim if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) { 1593eb94a878SMinchan Kim WARN_ON_ONCE(1); 159483612a94SMinchan Kim ret = false; 1595369ea824SJérôme Glisse /* We have to invalidate as we cleared the pte */ 15960f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, address, 15970f10851eSJérôme Glisse address + PAGE_SIZE); 1598eb94a878SMinchan Kim page_vma_mapped_walk_done(&pvmw); 1599eb94a878SMinchan Kim break; 1600eb94a878SMinchan Kim } 1601854e9ed0SMinchan Kim 1602802a3a92SShaohua Li /* MADV_FREE page check */ 1603802a3a92SShaohua Li if (!PageSwapBacked(page)) { 1604a128ca71SShaohua Li if (!PageDirty(page)) { 16050f10851eSJérôme Glisse /* Invalidate as we cleared the pte */ 16060f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, 16070f10851eSJérôme Glisse address, address + PAGE_SIZE); 1608854e9ed0SMinchan Kim dec_mm_counter(mm, MM_ANONPAGES); 1609854e9ed0SMinchan Kim goto discard; 1610854e9ed0SMinchan Kim } 1611854e9ed0SMinchan Kim 1612802a3a92SShaohua Li /* 1613802a3a92SShaohua Li * If the page was redirtied, it cannot be 1614802a3a92SShaohua Li * discarded. Remap the page to page table. 1615802a3a92SShaohua Li */ 1616785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 161718863d3aSMinchan Kim SetPageSwapBacked(page); 1618e4b82222SMinchan Kim ret = false; 1619802a3a92SShaohua Li page_vma_mapped_walk_done(&pvmw); 1620802a3a92SShaohua Li break; 1621802a3a92SShaohua Li } 1622802a3a92SShaohua Li 1623570a335bSHugh Dickins if (swap_duplicate(entry) < 0) { 1624785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 1625e4b82222SMinchan Kim ret = false; 1626c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1627c7ab0d2fSKirill A. Shutemov break; 1628570a335bSHugh Dickins } 1629ca827d55SKhalid Aziz if (arch_unmap_one(mm, vma, address, pteval) < 0) { 1630ca827d55SKhalid Aziz set_pte_at(mm, address, pvmw.pte, pteval); 1631ca827d55SKhalid Aziz ret = false; 1632ca827d55SKhalid Aziz page_vma_mapped_walk_done(&pvmw); 1633ca827d55SKhalid Aziz break; 1634ca827d55SKhalid Aziz } 16351da177e4SLinus Torvalds if (list_empty(&mm->mmlist)) { 16361da177e4SLinus Torvalds spin_lock(&mmlist_lock); 1637f412ac08SHugh Dickins if (list_empty(&mm->mmlist)) 16381da177e4SLinus Torvalds list_add(&mm->mmlist, &init_mm.mmlist); 16391da177e4SLinus Torvalds spin_unlock(&mmlist_lock); 16401da177e4SLinus Torvalds } 1641d559db08SKAMEZAWA Hiroyuki dec_mm_counter(mm, MM_ANONPAGES); 1642b084d435SKAMEZAWA Hiroyuki inc_mm_counter(mm, MM_SWAPENTS); 1643179ef71cSCyrill Gorcunov swp_pte = swp_entry_to_pte(entry); 1644179ef71cSCyrill Gorcunov if (pte_soft_dirty(pteval)) 1645179ef71cSCyrill Gorcunov swp_pte = pte_swp_mksoft_dirty(swp_pte); 1646785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, swp_pte); 16470f10851eSJérôme Glisse /* Invalidate as we cleared the pte */ 1648369ea824SJérôme Glisse mmu_notifier_invalidate_range(mm, address, 1649369ea824SJérôme Glisse address + PAGE_SIZE); 16500f10851eSJérôme Glisse } else { 16510f10851eSJérôme Glisse /* 1652906f9cdfSHugh Dickins * This is a locked file-backed page, thus it cannot 1653906f9cdfSHugh Dickins * be removed from the page cache and replaced by a new 1654906f9cdfSHugh Dickins * page before mmu_notifier_invalidate_range_end, so no 16550f10851eSJérôme Glisse * concurrent thread might update its page table to 16560f10851eSJérôme Glisse * point at new page while a device still is using this 16570f10851eSJérôme Glisse * page. 16580f10851eSJérôme Glisse * 1659ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 16600f10851eSJérôme Glisse */ 16610f10851eSJérôme Glisse dec_mm_counter(mm, mm_counter_file(page)); 16620f10851eSJérôme Glisse } 16630f10851eSJérôme Glisse discard: 16640f10851eSJérôme Glisse /* 16650f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() it has be 16660f10851eSJérôme Glisse * done above for all cases requiring it to happen under page 16670f10851eSJérôme Glisse * table lock before mmu_notifier_invalidate_range_end() 16680f10851eSJérôme Glisse * 1669ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 16700f10851eSJérôme Glisse */ 16710f10851eSJérôme Glisse page_remove_rmap(subpage, PageHuge(page)); 16720f10851eSJérôme Glisse put_page(page); 1673c7ab0d2fSKirill A. Shutemov } 1674369ea824SJérôme Glisse 1675ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1676369ea824SJérôme Glisse 1677caed0f48SKOSAKI Motohiro return ret; 16781da177e4SLinus Torvalds } 16791da177e4SLinus Torvalds 168071e3aac0SAndrea Arcangeli bool is_vma_temporary_stack(struct vm_area_struct *vma) 1681a8bef8ffSMel Gorman { 1682a8bef8ffSMel Gorman int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); 1683a8bef8ffSMel Gorman 1684a8bef8ffSMel Gorman if (!maybe_stack) 1685a8bef8ffSMel Gorman return false; 1686a8bef8ffSMel Gorman 1687a8bef8ffSMel Gorman if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == 1688a8bef8ffSMel Gorman VM_STACK_INCOMPLETE_SETUP) 1689a8bef8ffSMel Gorman return true; 1690a8bef8ffSMel Gorman 1691a8bef8ffSMel Gorman return false; 1692a8bef8ffSMel Gorman } 1693a8bef8ffSMel Gorman 169452629506SJoonsoo Kim static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) 169552629506SJoonsoo Kim { 169652629506SJoonsoo Kim return is_vma_temporary_stack(vma); 169752629506SJoonsoo Kim } 169852629506SJoonsoo Kim 16992a52bcbcSKirill A. Shutemov static int page_mapcount_is_zero(struct page *page) 170052629506SJoonsoo Kim { 1701c7ab0d2fSKirill A. Shutemov return !total_mapcount(page); 17022a52bcbcSKirill A. Shutemov } 170352629506SJoonsoo Kim 17041da177e4SLinus Torvalds /** 17051da177e4SLinus Torvalds * try_to_unmap - try to remove all page table mappings to a page 17061da177e4SLinus Torvalds * @page: the page to get unmapped 170714fa31b8SAndi Kleen * @flags: action and flags 17081da177e4SLinus Torvalds * 17091da177e4SLinus Torvalds * Tries to remove all the page table entries which are mapping this 17101da177e4SLinus Torvalds * page, used in the pageout path. Caller must hold the page lock. 17111da177e4SLinus Torvalds * 1712666e5a40SMinchan Kim * If unmap is successful, return true. Otherwise, false. 17131da177e4SLinus Torvalds */ 1714666e5a40SMinchan Kim bool try_to_unmap(struct page *page, enum ttu_flags flags) 17151da177e4SLinus Torvalds { 171652629506SJoonsoo Kim struct rmap_walk_control rwc = { 171752629506SJoonsoo Kim .rmap_one = try_to_unmap_one, 1718802a3a92SShaohua Li .arg = (void *)flags, 17192a52bcbcSKirill A. Shutemov .done = page_mapcount_is_zero, 172052629506SJoonsoo Kim .anon_lock = page_lock_anon_vma_read, 172152629506SJoonsoo Kim }; 17221da177e4SLinus Torvalds 172352629506SJoonsoo Kim /* 172452629506SJoonsoo Kim * During exec, a temporary VMA is setup and later moved. 172552629506SJoonsoo Kim * The VMA is moved under the anon_vma lock but not the 172652629506SJoonsoo Kim * page tables leading to a race where migration cannot 172752629506SJoonsoo Kim * find the migration ptes. Rather than increasing the 172852629506SJoonsoo Kim * locking requirements of exec(), migration skips 172952629506SJoonsoo Kim * temporary VMAs until after exec() completes. 173052629506SJoonsoo Kim */ 1731b5ff8161SNaoya Horiguchi if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE)) 1732b5ff8161SNaoya Horiguchi && !PageKsm(page) && PageAnon(page)) 173352629506SJoonsoo Kim rwc.invalid_vma = invalid_migration_vma; 173452629506SJoonsoo Kim 17352a52bcbcSKirill A. Shutemov if (flags & TTU_RMAP_LOCKED) 173633fc80e2SMinchan Kim rmap_walk_locked(page, &rwc); 17372a52bcbcSKirill A. Shutemov else 173833fc80e2SMinchan Kim rmap_walk(page, &rwc); 173952629506SJoonsoo Kim 1740666e5a40SMinchan Kim return !page_mapcount(page) ? true : false; 17411da177e4SLinus Torvalds } 174281b4082dSNikita Danilov 17432a52bcbcSKirill A. Shutemov static int page_not_mapped(struct page *page) 17442a52bcbcSKirill A. Shutemov { 17452a52bcbcSKirill A. Shutemov return !page_mapped(page); 17462a52bcbcSKirill A. Shutemov }; 17472a52bcbcSKirill A. Shutemov 1748b291f000SNick Piggin /** 1749b291f000SNick Piggin * try_to_munlock - try to munlock a page 1750b291f000SNick Piggin * @page: the page to be munlocked 1751b291f000SNick Piggin * 1752b291f000SNick Piggin * Called from munlock code. Checks all of the VMAs mapping the page 1753b291f000SNick Piggin * to make sure nobody else has this page mlocked. The page will be 1754b291f000SNick Piggin * returned with PG_mlocked cleared if no other vmas have it mlocked. 1755b291f000SNick Piggin */ 1756854e9ed0SMinchan Kim 1757192d7232SMinchan Kim void try_to_munlock(struct page *page) 1758192d7232SMinchan Kim { 1759e8351ac9SJoonsoo Kim struct rmap_walk_control rwc = { 1760e8351ac9SJoonsoo Kim .rmap_one = try_to_unmap_one, 1761802a3a92SShaohua Li .arg = (void *)TTU_MUNLOCK, 1762e8351ac9SJoonsoo Kim .done = page_not_mapped, 1763e8351ac9SJoonsoo Kim .anon_lock = page_lock_anon_vma_read, 1764e8351ac9SJoonsoo Kim 1765e8351ac9SJoonsoo Kim }; 1766e8351ac9SJoonsoo Kim 1767309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page); 1768192d7232SMinchan Kim VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page); 1769b291f000SNick Piggin 1770192d7232SMinchan Kim rmap_walk(page, &rwc); 1771b291f000SNick Piggin } 1772e9995ef9SHugh Dickins 177301d8b20dSPeter Zijlstra void __put_anon_vma(struct anon_vma *anon_vma) 177476545066SRik van Riel { 177576545066SRik van Riel struct anon_vma *root = anon_vma->root; 177676545066SRik van Riel 1777624483f3SAndrey Ryabinin anon_vma_free(anon_vma); 177801d8b20dSPeter Zijlstra if (root != anon_vma && atomic_dec_and_test(&root->refcount)) 177976545066SRik van Riel anon_vma_free(root); 178076545066SRik van Riel } 178176545066SRik van Riel 17820dd1c7bbSJoonsoo Kim static struct anon_vma *rmap_walk_anon_lock(struct page *page, 17830dd1c7bbSJoonsoo Kim struct rmap_walk_control *rwc) 1784faecd8ddSJoonsoo Kim { 1785faecd8ddSJoonsoo Kim struct anon_vma *anon_vma; 1786faecd8ddSJoonsoo Kim 17870dd1c7bbSJoonsoo Kim if (rwc->anon_lock) 17880dd1c7bbSJoonsoo Kim return rwc->anon_lock(page); 17890dd1c7bbSJoonsoo Kim 1790faecd8ddSJoonsoo Kim /* 1791faecd8ddSJoonsoo Kim * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() 1792faecd8ddSJoonsoo Kim * because that depends on page_mapped(); but not all its usages 1793faecd8ddSJoonsoo Kim * are holding mmap_sem. Users without mmap_sem are required to 1794faecd8ddSJoonsoo Kim * take a reference count to prevent the anon_vma disappearing 1795faecd8ddSJoonsoo Kim */ 1796faecd8ddSJoonsoo Kim anon_vma = page_anon_vma(page); 1797faecd8ddSJoonsoo Kim if (!anon_vma) 1798faecd8ddSJoonsoo Kim return NULL; 1799faecd8ddSJoonsoo Kim 1800faecd8ddSJoonsoo Kim anon_vma_lock_read(anon_vma); 1801faecd8ddSJoonsoo Kim return anon_vma; 1802faecd8ddSJoonsoo Kim } 1803faecd8ddSJoonsoo Kim 1804e9995ef9SHugh Dickins /* 1805e8351ac9SJoonsoo Kim * rmap_walk_anon - do something to anonymous page using the object-based 1806e8351ac9SJoonsoo Kim * rmap method 1807e8351ac9SJoonsoo Kim * @page: the page to be handled 1808e8351ac9SJoonsoo Kim * @rwc: control variable according to each walk type 1809e8351ac9SJoonsoo Kim * 1810e8351ac9SJoonsoo Kim * Find all the mappings of a page using the mapping pointer and the vma chains 1811e8351ac9SJoonsoo Kim * contained in the anon_vma struct it points to. 1812e8351ac9SJoonsoo Kim * 1813e8351ac9SJoonsoo Kim * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1814e8351ac9SJoonsoo Kim * where the page was found will be held for write. So, we won't recheck 1815e8351ac9SJoonsoo Kim * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1816e8351ac9SJoonsoo Kim * LOCKED. 1817e9995ef9SHugh Dickins */ 18181df631aeSMinchan Kim static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc, 1819b9773199SKirill A. Shutemov bool locked) 1820e9995ef9SHugh Dickins { 1821e9995ef9SHugh Dickins struct anon_vma *anon_vma; 1822a8fa41adSKirill A. Shutemov pgoff_t pgoff_start, pgoff_end; 18235beb4930SRik van Riel struct anon_vma_chain *avc; 1824e9995ef9SHugh Dickins 1825b9773199SKirill A. Shutemov if (locked) { 1826b9773199SKirill A. Shutemov anon_vma = page_anon_vma(page); 1827b9773199SKirill A. Shutemov /* anon_vma disappear under us? */ 1828b9773199SKirill A. Shutemov VM_BUG_ON_PAGE(!anon_vma, page); 1829b9773199SKirill A. Shutemov } else { 18300dd1c7bbSJoonsoo Kim anon_vma = rmap_walk_anon_lock(page, rwc); 1831b9773199SKirill A. Shutemov } 1832e9995ef9SHugh Dickins if (!anon_vma) 18331df631aeSMinchan Kim return; 1834faecd8ddSJoonsoo Kim 1835a8fa41adSKirill A. Shutemov pgoff_start = page_to_pgoff(page); 1836a8fa41adSKirill A. Shutemov pgoff_end = pgoff_start + hpage_nr_pages(page) - 1; 1837a8fa41adSKirill A. Shutemov anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 1838a8fa41adSKirill A. Shutemov pgoff_start, pgoff_end) { 18395beb4930SRik van Riel struct vm_area_struct *vma = avc->vma; 1840e9995ef9SHugh Dickins unsigned long address = vma_address(page, vma); 18410dd1c7bbSJoonsoo Kim 1842ad12695fSAndrea Arcangeli cond_resched(); 1843ad12695fSAndrea Arcangeli 18440dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 18450dd1c7bbSJoonsoo Kim continue; 18460dd1c7bbSJoonsoo Kim 1847e4b82222SMinchan Kim if (!rwc->rmap_one(page, vma, address, rwc->arg)) 1848e9995ef9SHugh Dickins break; 18490dd1c7bbSJoonsoo Kim if (rwc->done && rwc->done(page)) 18500dd1c7bbSJoonsoo Kim break; 1851e9995ef9SHugh Dickins } 1852b9773199SKirill A. Shutemov 1853b9773199SKirill A. Shutemov if (!locked) 18544fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 1855e9995ef9SHugh Dickins } 1856e9995ef9SHugh Dickins 1857e8351ac9SJoonsoo Kim /* 1858e8351ac9SJoonsoo Kim * rmap_walk_file - do something to file page using the object-based rmap method 1859e8351ac9SJoonsoo Kim * @page: the page to be handled 1860e8351ac9SJoonsoo Kim * @rwc: control variable according to each walk type 1861e8351ac9SJoonsoo Kim * 1862e8351ac9SJoonsoo Kim * Find all the mappings of a page using the mapping pointer and the vma chains 1863e8351ac9SJoonsoo Kim * contained in the address_space struct it points to. 1864e8351ac9SJoonsoo Kim * 1865e8351ac9SJoonsoo Kim * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1866e8351ac9SJoonsoo Kim * where the page was found will be held for write. So, we won't recheck 1867e8351ac9SJoonsoo Kim * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1868e8351ac9SJoonsoo Kim * LOCKED. 1869e8351ac9SJoonsoo Kim */ 18701df631aeSMinchan Kim static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc, 1871b9773199SKirill A. Shutemov bool locked) 1872e9995ef9SHugh Dickins { 1873b9773199SKirill A. Shutemov struct address_space *mapping = page_mapping(page); 1874a8fa41adSKirill A. Shutemov pgoff_t pgoff_start, pgoff_end; 1875e9995ef9SHugh Dickins struct vm_area_struct *vma; 1876e9995ef9SHugh Dickins 18779f32624bSJoonsoo Kim /* 18789f32624bSJoonsoo Kim * The page lock not only makes sure that page->mapping cannot 18799f32624bSJoonsoo Kim * suddenly be NULLified by truncation, it makes sure that the 18809f32624bSJoonsoo Kim * structure at mapping cannot be freed and reused yet, 1881c8c06efaSDavidlohr Bueso * so we can safely take mapping->i_mmap_rwsem. 18829f32624bSJoonsoo Kim */ 188381d1b09cSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 18849f32624bSJoonsoo Kim 1885e9995ef9SHugh Dickins if (!mapping) 18861df631aeSMinchan Kim return; 18873dec0ba0SDavidlohr Bueso 1888a8fa41adSKirill A. Shutemov pgoff_start = page_to_pgoff(page); 1889a8fa41adSKirill A. Shutemov pgoff_end = pgoff_start + hpage_nr_pages(page) - 1; 1890b9773199SKirill A. Shutemov if (!locked) 18913dec0ba0SDavidlohr Bueso i_mmap_lock_read(mapping); 1892a8fa41adSKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, 1893a8fa41adSKirill A. Shutemov pgoff_start, pgoff_end) { 1894e9995ef9SHugh Dickins unsigned long address = vma_address(page, vma); 18950dd1c7bbSJoonsoo Kim 1896ad12695fSAndrea Arcangeli cond_resched(); 1897ad12695fSAndrea Arcangeli 18980dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 18990dd1c7bbSJoonsoo Kim continue; 19000dd1c7bbSJoonsoo Kim 1901e4b82222SMinchan Kim if (!rwc->rmap_one(page, vma, address, rwc->arg)) 19020dd1c7bbSJoonsoo Kim goto done; 19030dd1c7bbSJoonsoo Kim if (rwc->done && rwc->done(page)) 19040dd1c7bbSJoonsoo Kim goto done; 1905e9995ef9SHugh Dickins } 19060dd1c7bbSJoonsoo Kim 19070dd1c7bbSJoonsoo Kim done: 1908b9773199SKirill A. Shutemov if (!locked) 19093dec0ba0SDavidlohr Bueso i_mmap_unlock_read(mapping); 1910e9995ef9SHugh Dickins } 1911e9995ef9SHugh Dickins 19121df631aeSMinchan Kim void rmap_walk(struct page *page, struct rmap_walk_control *rwc) 1913e9995ef9SHugh Dickins { 1914e9995ef9SHugh Dickins if (unlikely(PageKsm(page))) 19151df631aeSMinchan Kim rmap_walk_ksm(page, rwc); 1916e9995ef9SHugh Dickins else if (PageAnon(page)) 19171df631aeSMinchan Kim rmap_walk_anon(page, rwc, false); 1918e9995ef9SHugh Dickins else 19191df631aeSMinchan Kim rmap_walk_file(page, rwc, false); 1920b9773199SKirill A. Shutemov } 1921b9773199SKirill A. Shutemov 1922b9773199SKirill A. Shutemov /* Like rmap_walk, but caller holds relevant rmap lock */ 19231df631aeSMinchan Kim void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc) 1924b9773199SKirill A. Shutemov { 1925b9773199SKirill A. Shutemov /* no ksm support for now */ 1926b9773199SKirill A. Shutemov VM_BUG_ON_PAGE(PageKsm(page), page); 1927b9773199SKirill A. Shutemov if (PageAnon(page)) 19281df631aeSMinchan Kim rmap_walk_anon(page, rwc, true); 1929b9773199SKirill A. Shutemov else 19301df631aeSMinchan Kim rmap_walk_file(page, rwc, true); 1931e9995ef9SHugh Dickins } 19320fe6e20bSNaoya Horiguchi 1933e3390f67SNaoya Horiguchi #ifdef CONFIG_HUGETLB_PAGE 19340fe6e20bSNaoya Horiguchi /* 1935451b9514SKirill Tkhai * The following two functions are for anonymous (private mapped) hugepages. 19360fe6e20bSNaoya Horiguchi * Unlike common anonymous pages, anonymous hugepages have no accounting code 19370fe6e20bSNaoya Horiguchi * and no lru code, because we handle hugepages differently from common pages. 19380fe6e20bSNaoya Horiguchi */ 19390fe6e20bSNaoya Horiguchi void hugepage_add_anon_rmap(struct page *page, 19400fe6e20bSNaoya Horiguchi struct vm_area_struct *vma, unsigned long address) 19410fe6e20bSNaoya Horiguchi { 19420fe6e20bSNaoya Horiguchi struct anon_vma *anon_vma = vma->anon_vma; 19430fe6e20bSNaoya Horiguchi int first; 1944a850ea30SNaoya Horiguchi 1945a850ea30SNaoya Horiguchi BUG_ON(!PageLocked(page)); 19460fe6e20bSNaoya Horiguchi BUG_ON(!anon_vma); 19475dbe0af4SHugh Dickins /* address might be in next vma when migration races vma_adjust */ 194853f9263bSKirill A. Shutemov first = atomic_inc_and_test(compound_mapcount_ptr(page)); 19490fe6e20bSNaoya Horiguchi if (first) 1950451b9514SKirill Tkhai __page_set_anon_rmap(page, vma, address, 0); 19510fe6e20bSNaoya Horiguchi } 19520fe6e20bSNaoya Horiguchi 19530fe6e20bSNaoya Horiguchi void hugepage_add_new_anon_rmap(struct page *page, 19540fe6e20bSNaoya Horiguchi struct vm_area_struct *vma, unsigned long address) 19550fe6e20bSNaoya Horiguchi { 19560fe6e20bSNaoya Horiguchi BUG_ON(address < vma->vm_start || address >= vma->vm_end); 195753f9263bSKirill A. Shutemov atomic_set(compound_mapcount_ptr(page), 0); 1958451b9514SKirill Tkhai __page_set_anon_rmap(page, vma, address, 1); 19590fe6e20bSNaoya Horiguchi } 1960e3390f67SNaoya Horiguchi #endif /* CONFIG_HUGETLB_PAGE */ 1961