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 * 239608703eSJan Kara * inode->i_rwsem (while writing or truncating, not reading or faulting) 24c1e8d7c6SMichel Lespinasse * mm->mmap_lock 25730633f0SJan Kara * mapping->invalidate_lock (in filemap_fault) 269608703eSJan Kara * page->flags PG_locked (lock_page) * (see hugetlbfs below) 2788f306b6SKirill A. Shutemov * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share) 28c8c06efaSDavidlohr Bueso * mapping->i_mmap_rwsem 29c0d0381aSMike Kravetz * hugetlb_fault_mutex (hugetlbfs specific page fault mutex) 305a505085SIngo Molnar * anon_vma->rwsem 31b8072f09SHugh Dickins * mm->page_table_lock or pte_lock 325d337b91SHugh Dickins * swap_lock (in swap_duplicate, swap_info_get) 331da177e4SLinus Torvalds * mmlist_lock (in mmput, drain_mmlist and others) 34e621900aSMatthew Wilcox (Oracle) * mapping->private_lock (in block_dirty_folio) 35e621900aSMatthew Wilcox (Oracle) * folio_lock_memcg move_lock (in block_dirty_folio) 36b93b0163SMatthew Wilcox * i_pages lock (widely used) 37e809c3feSMatthew Wilcox (Oracle) * lruvec->lru_lock (in folio_lruvec_lock_irq) 38250df6edSDave Chinner * inode->i_lock (in set_page_dirty's __mark_inode_dirty) 39f758eeabSChristoph Hellwig * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) 401da177e4SLinus Torvalds * sb_lock (within inode_lock in fs/fs-writeback.c) 41b93b0163SMatthew Wilcox * i_pages lock (widely used, in set_page_dirty, 421da177e4SLinus Torvalds * in arch-dependent flush_dcache_mmap_lock, 43f758eeabSChristoph Hellwig * within bdi.wb->list_lock in __sync_single_inode) 446a46079cSAndi Kleen * 459608703eSJan Kara * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon) 466a46079cSAndi Kleen * ->tasklist_lock 476a46079cSAndi Kleen * pte map lock 48c0d0381aSMike Kravetz * 49c0d0381aSMike Kravetz * * hugetlbfs PageHuge() pages take locks in this order: 50c0d0381aSMike Kravetz * mapping->i_mmap_rwsem 51c0d0381aSMike Kravetz * hugetlb_fault_mutex (hugetlbfs specific page fault mutex) 52c0d0381aSMike Kravetz * page->flags PG_locked (lock_page) 531da177e4SLinus Torvalds */ 541da177e4SLinus Torvalds 551da177e4SLinus Torvalds #include <linux/mm.h> 566e84f315SIngo Molnar #include <linux/sched/mm.h> 5729930025SIngo Molnar #include <linux/sched/task.h> 581da177e4SLinus Torvalds #include <linux/pagemap.h> 591da177e4SLinus Torvalds #include <linux/swap.h> 601da177e4SLinus Torvalds #include <linux/swapops.h> 611da177e4SLinus Torvalds #include <linux/slab.h> 621da177e4SLinus Torvalds #include <linux/init.h> 635ad64688SHugh Dickins #include <linux/ksm.h> 641da177e4SLinus Torvalds #include <linux/rmap.h> 651da177e4SLinus Torvalds #include <linux/rcupdate.h> 66b95f1b31SPaul Gortmaker #include <linux/export.h> 678a9f3ccdSBalbir Singh #include <linux/memcontrol.h> 68cddb8a5cSAndrea Arcangeli #include <linux/mmu_notifier.h> 6964cdd548SKOSAKI Motohiro #include <linux/migrate.h> 700fe6e20bSNaoya Horiguchi #include <linux/hugetlb.h> 71444f84fdSBen Dooks #include <linux/huge_mm.h> 72ef5d437fSJan Kara #include <linux/backing-dev.h> 7333c3fc71SVladimir Davydov #include <linux/page_idle.h> 74a5430ddaSJérôme Glisse #include <linux/memremap.h> 75bce73e48SChristian Borntraeger #include <linux/userfaultfd_k.h> 76999dad82SPeter Xu #include <linux/mm_inline.h> 771da177e4SLinus Torvalds 781da177e4SLinus Torvalds #include <asm/tlbflush.h> 791da177e4SLinus Torvalds 804cc79b33SAnshuman Khandual #define CREATE_TRACE_POINTS 8172b252aeSMel Gorman #include <trace/events/tlb.h> 824cc79b33SAnshuman Khandual #include <trace/events/migrate.h> 8372b252aeSMel Gorman 84b291f000SNick Piggin #include "internal.h" 85b291f000SNick Piggin 86fdd2e5f8SAdrian Bunk static struct kmem_cache *anon_vma_cachep; 875beb4930SRik van Riel static struct kmem_cache *anon_vma_chain_cachep; 88fdd2e5f8SAdrian Bunk 89fdd2e5f8SAdrian Bunk static inline struct anon_vma *anon_vma_alloc(void) 90fdd2e5f8SAdrian Bunk { 9101d8b20dSPeter Zijlstra struct anon_vma *anon_vma; 9201d8b20dSPeter Zijlstra 9301d8b20dSPeter Zijlstra anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); 9401d8b20dSPeter Zijlstra if (anon_vma) { 9501d8b20dSPeter Zijlstra atomic_set(&anon_vma->refcount, 1); 967a3ef208SKonstantin Khlebnikov anon_vma->degree = 1; /* Reference for first vma */ 977a3ef208SKonstantin Khlebnikov anon_vma->parent = anon_vma; 9801d8b20dSPeter Zijlstra /* 9901d8b20dSPeter Zijlstra * Initialise the anon_vma root to point to itself. If called 10001d8b20dSPeter Zijlstra * from fork, the root will be reset to the parents anon_vma. 10101d8b20dSPeter Zijlstra */ 10201d8b20dSPeter Zijlstra anon_vma->root = anon_vma; 103fdd2e5f8SAdrian Bunk } 104fdd2e5f8SAdrian Bunk 10501d8b20dSPeter Zijlstra return anon_vma; 10601d8b20dSPeter Zijlstra } 10701d8b20dSPeter Zijlstra 10801d8b20dSPeter Zijlstra static inline void anon_vma_free(struct anon_vma *anon_vma) 109fdd2e5f8SAdrian Bunk { 11001d8b20dSPeter Zijlstra VM_BUG_ON(atomic_read(&anon_vma->refcount)); 11188c22088SPeter Zijlstra 11288c22088SPeter Zijlstra /* 1132f031c6fSMatthew Wilcox (Oracle) * Synchronize against folio_lock_anon_vma_read() such that 11488c22088SPeter Zijlstra * we can safely hold the lock without the anon_vma getting 11588c22088SPeter Zijlstra * freed. 11688c22088SPeter Zijlstra * 11788c22088SPeter Zijlstra * Relies on the full mb implied by the atomic_dec_and_test() from 11888c22088SPeter Zijlstra * put_anon_vma() against the acquire barrier implied by 1192f031c6fSMatthew Wilcox (Oracle) * down_read_trylock() from folio_lock_anon_vma_read(). This orders: 12088c22088SPeter Zijlstra * 1212f031c6fSMatthew Wilcox (Oracle) * folio_lock_anon_vma_read() VS put_anon_vma() 1224fc3f1d6SIngo Molnar * down_read_trylock() atomic_dec_and_test() 12388c22088SPeter Zijlstra * LOCK MB 1244fc3f1d6SIngo Molnar * atomic_read() rwsem_is_locked() 12588c22088SPeter Zijlstra * 12688c22088SPeter Zijlstra * LOCK should suffice since the actual taking of the lock must 12788c22088SPeter Zijlstra * happen _before_ what follows. 12888c22088SPeter Zijlstra */ 1297f39dda9SHugh Dickins might_sleep(); 1305a505085SIngo Molnar if (rwsem_is_locked(&anon_vma->root->rwsem)) { 1314fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 13208b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 13388c22088SPeter Zijlstra } 13488c22088SPeter Zijlstra 135fdd2e5f8SAdrian Bunk kmem_cache_free(anon_vma_cachep, anon_vma); 136fdd2e5f8SAdrian Bunk } 1371da177e4SLinus Torvalds 138dd34739cSLinus Torvalds static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) 1395beb4930SRik van Riel { 140dd34739cSLinus Torvalds return kmem_cache_alloc(anon_vma_chain_cachep, gfp); 1415beb4930SRik van Riel } 1425beb4930SRik van Riel 143e574b5fdSNamhyung Kim static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) 1445beb4930SRik van Riel { 1455beb4930SRik van Riel kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); 1465beb4930SRik van Riel } 1475beb4930SRik van Riel 1486583a843SKautuk Consul static void anon_vma_chain_link(struct vm_area_struct *vma, 1496583a843SKautuk Consul struct anon_vma_chain *avc, 1506583a843SKautuk Consul struct anon_vma *anon_vma) 1516583a843SKautuk Consul { 1526583a843SKautuk Consul avc->vma = vma; 1536583a843SKautuk Consul avc->anon_vma = anon_vma; 1546583a843SKautuk Consul list_add(&avc->same_vma, &vma->anon_vma_chain); 155bf181b9fSMichel Lespinasse anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); 1566583a843SKautuk Consul } 1576583a843SKautuk Consul 158d9d332e0SLinus Torvalds /** 159d5a187daSVlastimil Babka * __anon_vma_prepare - attach an anon_vma to a memory region 160d9d332e0SLinus Torvalds * @vma: the memory region in question 161d9d332e0SLinus Torvalds * 162d9d332e0SLinus Torvalds * This makes sure the memory mapping described by 'vma' has 163d9d332e0SLinus Torvalds * an 'anon_vma' attached to it, so that we can associate the 164d9d332e0SLinus Torvalds * anonymous pages mapped into it with that anon_vma. 165d9d332e0SLinus Torvalds * 166d5a187daSVlastimil Babka * The common case will be that we already have one, which 167d5a187daSVlastimil Babka * is handled inline by anon_vma_prepare(). But if 16823a0790aSFigo.zhang * not we either need to find an adjacent mapping that we 169d9d332e0SLinus Torvalds * can re-use the anon_vma from (very common when the only 170d9d332e0SLinus Torvalds * reason for splitting a vma has been mprotect()), or we 171d9d332e0SLinus Torvalds * allocate a new one. 172d9d332e0SLinus Torvalds * 173d9d332e0SLinus Torvalds * Anon-vma allocations are very subtle, because we may have 1742f031c6fSMatthew Wilcox (Oracle) * optimistically looked up an anon_vma in folio_lock_anon_vma_read() 175aaf1f990SMiaohe Lin * and that may actually touch the rwsem even in the newly 176d9d332e0SLinus Torvalds * allocated vma (it depends on RCU to make sure that the 177d9d332e0SLinus Torvalds * anon_vma isn't actually destroyed). 178d9d332e0SLinus Torvalds * 179d9d332e0SLinus Torvalds * As a result, we need to do proper anon_vma locking even 180d9d332e0SLinus Torvalds * for the new allocation. At the same time, we do not want 181d9d332e0SLinus Torvalds * to do any locking for the common case of already having 182d9d332e0SLinus Torvalds * an anon_vma. 183d9d332e0SLinus Torvalds * 184c1e8d7c6SMichel Lespinasse * This must be called with the mmap_lock held for reading. 185d9d332e0SLinus Torvalds */ 186d5a187daSVlastimil Babka int __anon_vma_prepare(struct vm_area_struct *vma) 1871da177e4SLinus Torvalds { 188d5a187daSVlastimil Babka struct mm_struct *mm = vma->vm_mm; 189d5a187daSVlastimil Babka struct anon_vma *anon_vma, *allocated; 1905beb4930SRik van Riel struct anon_vma_chain *avc; 1911da177e4SLinus Torvalds 1921da177e4SLinus Torvalds might_sleep(); 1931da177e4SLinus Torvalds 194dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 1955beb4930SRik van Riel if (!avc) 1965beb4930SRik van Riel goto out_enomem; 1975beb4930SRik van Riel 1981da177e4SLinus Torvalds anon_vma = find_mergeable_anon_vma(vma); 1991da177e4SLinus Torvalds allocated = NULL; 200d9d332e0SLinus Torvalds if (!anon_vma) { 2011da177e4SLinus Torvalds anon_vma = anon_vma_alloc(); 2021da177e4SLinus Torvalds if (unlikely(!anon_vma)) 2035beb4930SRik van Riel goto out_enomem_free_avc; 2041da177e4SLinus Torvalds allocated = anon_vma; 2051da177e4SLinus Torvalds } 2061da177e4SLinus Torvalds 2074fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 2081da177e4SLinus Torvalds /* page_table_lock to protect against threads */ 2091da177e4SLinus Torvalds spin_lock(&mm->page_table_lock); 2101da177e4SLinus Torvalds if (likely(!vma->anon_vma)) { 2111da177e4SLinus Torvalds vma->anon_vma = anon_vma; 2126583a843SKautuk Consul anon_vma_chain_link(vma, avc, anon_vma); 2137a3ef208SKonstantin Khlebnikov /* vma reference or self-parent link for new root */ 2147a3ef208SKonstantin Khlebnikov anon_vma->degree++; 2151da177e4SLinus Torvalds allocated = NULL; 21631f2b0ebSOleg Nesterov avc = NULL; 2171da177e4SLinus Torvalds } 2181da177e4SLinus Torvalds spin_unlock(&mm->page_table_lock); 21908b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 22031f2b0ebSOleg Nesterov 22131f2b0ebSOleg Nesterov if (unlikely(allocated)) 22201d8b20dSPeter Zijlstra put_anon_vma(allocated); 22331f2b0ebSOleg Nesterov if (unlikely(avc)) 2245beb4930SRik van Riel anon_vma_chain_free(avc); 225d5a187daSVlastimil Babka 2261da177e4SLinus Torvalds return 0; 2275beb4930SRik van Riel 2285beb4930SRik van Riel out_enomem_free_avc: 2295beb4930SRik van Riel anon_vma_chain_free(avc); 2305beb4930SRik van Riel out_enomem: 2315beb4930SRik van Riel return -ENOMEM; 2321da177e4SLinus Torvalds } 2331da177e4SLinus Torvalds 234bb4aa396SLinus Torvalds /* 235bb4aa396SLinus Torvalds * This is a useful helper function for locking the anon_vma root as 236bb4aa396SLinus Torvalds * we traverse the vma->anon_vma_chain, looping over anon_vma's that 237bb4aa396SLinus Torvalds * have the same vma. 238bb4aa396SLinus Torvalds * 239bb4aa396SLinus Torvalds * Such anon_vma's should have the same root, so you'd expect to see 240bb4aa396SLinus Torvalds * just a single mutex_lock for the whole traversal. 241bb4aa396SLinus Torvalds */ 242bb4aa396SLinus Torvalds static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) 243bb4aa396SLinus Torvalds { 244bb4aa396SLinus Torvalds struct anon_vma *new_root = anon_vma->root; 245bb4aa396SLinus Torvalds if (new_root != root) { 246bb4aa396SLinus Torvalds if (WARN_ON_ONCE(root)) 2475a505085SIngo Molnar up_write(&root->rwsem); 248bb4aa396SLinus Torvalds root = new_root; 2495a505085SIngo Molnar down_write(&root->rwsem); 250bb4aa396SLinus Torvalds } 251bb4aa396SLinus Torvalds return root; 252bb4aa396SLinus Torvalds } 253bb4aa396SLinus Torvalds 254bb4aa396SLinus Torvalds static inline void unlock_anon_vma_root(struct anon_vma *root) 255bb4aa396SLinus Torvalds { 256bb4aa396SLinus Torvalds if (root) 2575a505085SIngo Molnar up_write(&root->rwsem); 258bb4aa396SLinus Torvalds } 259bb4aa396SLinus Torvalds 2605beb4930SRik van Riel /* 2615beb4930SRik van Riel * Attach the anon_vmas from src to dst. 2625beb4930SRik van Riel * Returns 0 on success, -ENOMEM on failure. 2637a3ef208SKonstantin Khlebnikov * 264cb152a1aSShijie Luo * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and 26547b390d2SWei Yang * anon_vma_fork(). The first three want an exact copy of src, while the last 26647b390d2SWei Yang * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent 26747b390d2SWei Yang * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call, 26847b390d2SWei Yang * we can identify this case by checking (!dst->anon_vma && src->anon_vma). 26947b390d2SWei Yang * 27047b390d2SWei Yang * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find 27147b390d2SWei Yang * and reuse existing anon_vma which has no vmas and only one child anon_vma. 27247b390d2SWei Yang * This prevents degradation of anon_vma hierarchy to endless linear chain in 27347b390d2SWei Yang * case of constantly forking task. On the other hand, an anon_vma with more 27447b390d2SWei Yang * than one child isn't reused even if there was no alive vma, thus rmap 27547b390d2SWei Yang * walker has a good chance of avoiding scanning the whole hierarchy when it 27647b390d2SWei Yang * searches where page is mapped. 2775beb4930SRik van Riel */ 2785beb4930SRik van Riel int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) 2795beb4930SRik van Riel { 2805beb4930SRik van Riel struct anon_vma_chain *avc, *pavc; 281bb4aa396SLinus Torvalds struct anon_vma *root = NULL; 2825beb4930SRik van Riel 283646d87b4SLinus Torvalds list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { 284bb4aa396SLinus Torvalds struct anon_vma *anon_vma; 285bb4aa396SLinus Torvalds 286dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); 287dd34739cSLinus Torvalds if (unlikely(!avc)) { 288dd34739cSLinus Torvalds unlock_anon_vma_root(root); 289dd34739cSLinus Torvalds root = NULL; 290dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 2915beb4930SRik van Riel if (!avc) 2925beb4930SRik van Riel goto enomem_failure; 293dd34739cSLinus Torvalds } 294bb4aa396SLinus Torvalds anon_vma = pavc->anon_vma; 295bb4aa396SLinus Torvalds root = lock_anon_vma_root(root, anon_vma); 296bb4aa396SLinus Torvalds anon_vma_chain_link(dst, avc, anon_vma); 2977a3ef208SKonstantin Khlebnikov 2987a3ef208SKonstantin Khlebnikov /* 2997a3ef208SKonstantin Khlebnikov * Reuse existing anon_vma if its degree lower than two, 3007a3ef208SKonstantin Khlebnikov * that means it has no vma and only one anon_vma child. 3017a3ef208SKonstantin Khlebnikov * 302dd062302SAdrian Huang * Do not choose parent anon_vma, otherwise first child 3037a3ef208SKonstantin Khlebnikov * will always reuse it. Root anon_vma is never reused: 3047a3ef208SKonstantin Khlebnikov * it has self-parent reference and at least one child. 3057a3ef208SKonstantin Khlebnikov */ 30647b390d2SWei Yang if (!dst->anon_vma && src->anon_vma && 30747b390d2SWei Yang anon_vma != src->anon_vma && anon_vma->degree < 2) 3087a3ef208SKonstantin Khlebnikov dst->anon_vma = anon_vma; 3095beb4930SRik van Riel } 3107a3ef208SKonstantin Khlebnikov if (dst->anon_vma) 3117a3ef208SKonstantin Khlebnikov dst->anon_vma->degree++; 312bb4aa396SLinus Torvalds unlock_anon_vma_root(root); 3135beb4930SRik van Riel return 0; 3145beb4930SRik van Riel 3155beb4930SRik van Riel enomem_failure: 3163fe89b3eSLeon Yu /* 3173fe89b3eSLeon Yu * dst->anon_vma is dropped here otherwise its degree can be incorrectly 3183fe89b3eSLeon Yu * decremented in unlink_anon_vmas(). 3193fe89b3eSLeon Yu * We can safely do this because callers of anon_vma_clone() don't care 3203fe89b3eSLeon Yu * about dst->anon_vma if anon_vma_clone() failed. 3213fe89b3eSLeon Yu */ 3223fe89b3eSLeon Yu dst->anon_vma = NULL; 3235beb4930SRik van Riel unlink_anon_vmas(dst); 3245beb4930SRik van Riel return -ENOMEM; 3251da177e4SLinus Torvalds } 3261da177e4SLinus Torvalds 3275beb4930SRik van Riel /* 3285beb4930SRik van Riel * Attach vma to its own anon_vma, as well as to the anon_vmas that 3295beb4930SRik van Riel * the corresponding VMA in the parent process is attached to. 3305beb4930SRik van Riel * Returns 0 on success, non-zero on failure. 3315beb4930SRik van Riel */ 3325beb4930SRik van Riel int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) 3331da177e4SLinus Torvalds { 3345beb4930SRik van Riel struct anon_vma_chain *avc; 3355beb4930SRik van Riel struct anon_vma *anon_vma; 336c4ea95d7SDaniel Forrest int error; 3375beb4930SRik van Riel 3385beb4930SRik van Riel /* Don't bother if the parent process has no anon_vma here. */ 3395beb4930SRik van Riel if (!pvma->anon_vma) 3405beb4930SRik van Riel return 0; 3415beb4930SRik van Riel 3427a3ef208SKonstantin Khlebnikov /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ 3437a3ef208SKonstantin Khlebnikov vma->anon_vma = NULL; 3447a3ef208SKonstantin Khlebnikov 3455beb4930SRik van Riel /* 3465beb4930SRik van Riel * First, attach the new VMA to the parent VMA's anon_vmas, 3475beb4930SRik van Riel * so rmap can find non-COWed pages in child processes. 3485beb4930SRik van Riel */ 349c4ea95d7SDaniel Forrest error = anon_vma_clone(vma, pvma); 350c4ea95d7SDaniel Forrest if (error) 351c4ea95d7SDaniel Forrest return error; 3525beb4930SRik van Riel 3537a3ef208SKonstantin Khlebnikov /* An existing anon_vma has been reused, all done then. */ 3547a3ef208SKonstantin Khlebnikov if (vma->anon_vma) 3557a3ef208SKonstantin Khlebnikov return 0; 3567a3ef208SKonstantin Khlebnikov 3575beb4930SRik van Riel /* Then add our own anon_vma. */ 3585beb4930SRik van Riel anon_vma = anon_vma_alloc(); 3595beb4930SRik van Riel if (!anon_vma) 3605beb4930SRik van Riel goto out_error; 361dd34739cSLinus Torvalds avc = anon_vma_chain_alloc(GFP_KERNEL); 3625beb4930SRik van Riel if (!avc) 3635beb4930SRik van Riel goto out_error_free_anon_vma; 3645c341ee1SRik van Riel 3655c341ee1SRik van Riel /* 366aaf1f990SMiaohe Lin * The root anon_vma's rwsem is the lock actually used when we 3675c341ee1SRik van Riel * lock any of the anon_vmas in this anon_vma tree. 3685c341ee1SRik van Riel */ 3695c341ee1SRik van Riel anon_vma->root = pvma->anon_vma->root; 3707a3ef208SKonstantin Khlebnikov anon_vma->parent = pvma->anon_vma; 37176545066SRik van Riel /* 37201d8b20dSPeter Zijlstra * With refcounts, an anon_vma can stay around longer than the 37301d8b20dSPeter Zijlstra * process it belongs to. The root anon_vma needs to be pinned until 37401d8b20dSPeter Zijlstra * this anon_vma is freed, because the lock lives in the root. 37576545066SRik van Riel */ 37676545066SRik van Riel get_anon_vma(anon_vma->root); 3775beb4930SRik van Riel /* Mark this anon_vma as the one where our new (COWed) pages go. */ 3785beb4930SRik van Riel vma->anon_vma = anon_vma; 3794fc3f1d6SIngo Molnar anon_vma_lock_write(anon_vma); 3805c341ee1SRik van Riel anon_vma_chain_link(vma, avc, anon_vma); 3817a3ef208SKonstantin Khlebnikov anon_vma->parent->degree++; 38208b52706SKonstantin Khlebnikov anon_vma_unlock_write(anon_vma); 3835beb4930SRik van Riel 3845beb4930SRik van Riel return 0; 3855beb4930SRik van Riel 3865beb4930SRik van Riel out_error_free_anon_vma: 38701d8b20dSPeter Zijlstra put_anon_vma(anon_vma); 3885beb4930SRik van Riel out_error: 3894946d54cSRik van Riel unlink_anon_vmas(vma); 3905beb4930SRik van Riel return -ENOMEM; 3915beb4930SRik van Riel } 3925beb4930SRik van Riel 3935beb4930SRik van Riel void unlink_anon_vmas(struct vm_area_struct *vma) 3945beb4930SRik van Riel { 3955beb4930SRik van Riel struct anon_vma_chain *avc, *next; 396eee2acbaSPeter Zijlstra struct anon_vma *root = NULL; 3975beb4930SRik van Riel 3985c341ee1SRik van Riel /* 3995c341ee1SRik van Riel * Unlink each anon_vma chained to the VMA. This list is ordered 4005c341ee1SRik van Riel * from newest to oldest, ensuring the root anon_vma gets freed last. 4015c341ee1SRik van Riel */ 4025beb4930SRik van Riel list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 403eee2acbaSPeter Zijlstra struct anon_vma *anon_vma = avc->anon_vma; 404eee2acbaSPeter Zijlstra 405eee2acbaSPeter Zijlstra root = lock_anon_vma_root(root, anon_vma); 406bf181b9fSMichel Lespinasse anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); 407eee2acbaSPeter Zijlstra 408eee2acbaSPeter Zijlstra /* 409eee2acbaSPeter Zijlstra * Leave empty anon_vmas on the list - we'll need 410eee2acbaSPeter Zijlstra * to free them outside the lock. 411eee2acbaSPeter Zijlstra */ 412f808c13fSDavidlohr Bueso if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) { 4137a3ef208SKonstantin Khlebnikov anon_vma->parent->degree--; 414eee2acbaSPeter Zijlstra continue; 4157a3ef208SKonstantin Khlebnikov } 416eee2acbaSPeter Zijlstra 417eee2acbaSPeter Zijlstra list_del(&avc->same_vma); 418eee2acbaSPeter Zijlstra anon_vma_chain_free(avc); 419eee2acbaSPeter Zijlstra } 420ee8ab190SLi Xinhai if (vma->anon_vma) { 4217a3ef208SKonstantin Khlebnikov vma->anon_vma->degree--; 422ee8ab190SLi Xinhai 423ee8ab190SLi Xinhai /* 424ee8ab190SLi Xinhai * vma would still be needed after unlink, and anon_vma will be prepared 425ee8ab190SLi Xinhai * when handle fault. 426ee8ab190SLi Xinhai */ 427ee8ab190SLi Xinhai vma->anon_vma = NULL; 428ee8ab190SLi Xinhai } 429eee2acbaSPeter Zijlstra unlock_anon_vma_root(root); 430eee2acbaSPeter Zijlstra 431eee2acbaSPeter Zijlstra /* 432eee2acbaSPeter Zijlstra * Iterate the list once more, it now only contains empty and unlinked 433eee2acbaSPeter Zijlstra * anon_vmas, destroy them. Could not do before due to __put_anon_vma() 4345a505085SIngo Molnar * needing to write-acquire the anon_vma->root->rwsem. 435eee2acbaSPeter Zijlstra */ 436eee2acbaSPeter Zijlstra list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 437eee2acbaSPeter Zijlstra struct anon_vma *anon_vma = avc->anon_vma; 438eee2acbaSPeter Zijlstra 439e4c5800aSKonstantin Khlebnikov VM_WARN_ON(anon_vma->degree); 440eee2acbaSPeter Zijlstra put_anon_vma(anon_vma); 441eee2acbaSPeter Zijlstra 4425beb4930SRik van Riel list_del(&avc->same_vma); 4435beb4930SRik van Riel anon_vma_chain_free(avc); 4445beb4930SRik van Riel } 4455beb4930SRik van Riel } 4465beb4930SRik van Riel 44751cc5068SAlexey Dobriyan static void anon_vma_ctor(void *data) 4481da177e4SLinus Torvalds { 4491da177e4SLinus Torvalds struct anon_vma *anon_vma = data; 4501da177e4SLinus Torvalds 4515a505085SIngo Molnar init_rwsem(&anon_vma->rwsem); 45283813267SPeter Zijlstra atomic_set(&anon_vma->refcount, 0); 453f808c13fSDavidlohr Bueso anon_vma->rb_root = RB_ROOT_CACHED; 4541da177e4SLinus Torvalds } 4551da177e4SLinus Torvalds 4561da177e4SLinus Torvalds void __init anon_vma_init(void) 4571da177e4SLinus Torvalds { 4581da177e4SLinus Torvalds anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 4595f0d5a3aSPaul E. McKenney 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, 4605d097056SVladimir Davydov anon_vma_ctor); 4615d097056SVladimir Davydov anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, 4625d097056SVladimir Davydov SLAB_PANIC|SLAB_ACCOUNT); 4631da177e4SLinus Torvalds } 4641da177e4SLinus Torvalds 4651da177e4SLinus Torvalds /* 4666111e4caSPeter Zijlstra * Getting a lock on a stable anon_vma from a page off the LRU is tricky! 4676111e4caSPeter Zijlstra * 4686111e4caSPeter Zijlstra * Since there is no serialization what so ever against page_remove_rmap() 469ad8a20cfSMiaohe Lin * the best this function can do is return a refcount increased anon_vma 470ad8a20cfSMiaohe Lin * that might have been relevant to this page. 4716111e4caSPeter Zijlstra * 4726111e4caSPeter Zijlstra * The page might have been remapped to a different anon_vma or the anon_vma 4736111e4caSPeter Zijlstra * returned may already be freed (and even reused). 4746111e4caSPeter Zijlstra * 475bc658c96SPeter Zijlstra * In case it was remapped to a different anon_vma, the new anon_vma will be a 476bc658c96SPeter Zijlstra * child of the old anon_vma, and the anon_vma lifetime rules will therefore 477bc658c96SPeter Zijlstra * ensure that any anon_vma obtained from the page will still be valid for as 478bc658c96SPeter Zijlstra * long as we observe page_mapped() [ hence all those page_mapped() tests ]. 479bc658c96SPeter Zijlstra * 4806111e4caSPeter Zijlstra * All users of this function must be very careful when walking the anon_vma 4816111e4caSPeter Zijlstra * chain and verify that the page in question is indeed mapped in it 4826111e4caSPeter Zijlstra * [ something equivalent to page_mapped_in_vma() ]. 4836111e4caSPeter Zijlstra * 484091e4299SMiles Chen * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from 485091e4299SMiles Chen * page_remove_rmap() that the anon_vma pointer from page->mapping is valid 486091e4299SMiles Chen * if there is a mapcount, we can dereference the anon_vma after observing 487091e4299SMiles Chen * those. 4881da177e4SLinus Torvalds */ 489746b18d4SPeter Zijlstra struct anon_vma *page_get_anon_vma(struct page *page) 4901da177e4SLinus Torvalds { 491746b18d4SPeter Zijlstra struct anon_vma *anon_vma = NULL; 4921da177e4SLinus Torvalds unsigned long anon_mapping; 4931da177e4SLinus Torvalds 4941da177e4SLinus Torvalds rcu_read_lock(); 4954db0c3c2SJason Low anon_mapping = (unsigned long)READ_ONCE(page->mapping); 4963ca7b3c5SHugh Dickins if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 4971da177e4SLinus Torvalds goto out; 4981da177e4SLinus Torvalds if (!page_mapped(page)) 4991da177e4SLinus Torvalds goto out; 5001da177e4SLinus Torvalds 5011da177e4SLinus Torvalds anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 502746b18d4SPeter Zijlstra if (!atomic_inc_not_zero(&anon_vma->refcount)) { 503746b18d4SPeter Zijlstra anon_vma = NULL; 504746b18d4SPeter Zijlstra goto out; 505746b18d4SPeter Zijlstra } 506f1819427SHugh Dickins 507f1819427SHugh Dickins /* 508f1819427SHugh Dickins * If this page is still mapped, then its anon_vma cannot have been 509746b18d4SPeter Zijlstra * freed. But if it has been unmapped, we have no security against the 510746b18d4SPeter Zijlstra * anon_vma structure being freed and reused (for another anon_vma: 5115f0d5a3aSPaul E. McKenney * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero() 512746b18d4SPeter Zijlstra * above cannot corrupt). 513f1819427SHugh Dickins */ 514746b18d4SPeter Zijlstra if (!page_mapped(page)) { 5157f39dda9SHugh Dickins rcu_read_unlock(); 516746b18d4SPeter Zijlstra put_anon_vma(anon_vma); 5177f39dda9SHugh Dickins return NULL; 518746b18d4SPeter Zijlstra } 5191da177e4SLinus Torvalds out: 5201da177e4SLinus Torvalds rcu_read_unlock(); 521746b18d4SPeter Zijlstra 522746b18d4SPeter Zijlstra return anon_vma; 523746b18d4SPeter Zijlstra } 524746b18d4SPeter Zijlstra 52588c22088SPeter Zijlstra /* 52688c22088SPeter Zijlstra * Similar to page_get_anon_vma() except it locks the anon_vma. 52788c22088SPeter Zijlstra * 52888c22088SPeter Zijlstra * Its a little more complex as it tries to keep the fast path to a single 52988c22088SPeter Zijlstra * atomic op -- the trylock. If we fail the trylock, we fall back to getting a 5306d4675e6SMinchan Kim * reference like with page_get_anon_vma() and then block on the mutex 5316d4675e6SMinchan Kim * on !rwc->try_lock case. 53288c22088SPeter Zijlstra */ 5336d4675e6SMinchan Kim struct anon_vma *folio_lock_anon_vma_read(struct folio *folio, 5346d4675e6SMinchan Kim struct rmap_walk_control *rwc) 535746b18d4SPeter Zijlstra { 53688c22088SPeter Zijlstra struct anon_vma *anon_vma = NULL; 537eee0f252SHugh Dickins struct anon_vma *root_anon_vma; 53888c22088SPeter Zijlstra unsigned long anon_mapping; 539746b18d4SPeter Zijlstra 54088c22088SPeter Zijlstra rcu_read_lock(); 5419595d769SMatthew Wilcox (Oracle) anon_mapping = (unsigned long)READ_ONCE(folio->mapping); 54288c22088SPeter Zijlstra if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 54388c22088SPeter Zijlstra goto out; 5449595d769SMatthew Wilcox (Oracle) if (!folio_mapped(folio)) 54588c22088SPeter Zijlstra goto out; 54688c22088SPeter Zijlstra 54788c22088SPeter Zijlstra anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 5484db0c3c2SJason Low root_anon_vma = READ_ONCE(anon_vma->root); 5494fc3f1d6SIngo Molnar if (down_read_trylock(&root_anon_vma->rwsem)) { 55088c22088SPeter Zijlstra /* 5519595d769SMatthew Wilcox (Oracle) * If the folio is still mapped, then this anon_vma is still 552eee0f252SHugh Dickins * its anon_vma, and holding the mutex ensures that it will 553bc658c96SPeter Zijlstra * not go away, see anon_vma_free(). 55488c22088SPeter Zijlstra */ 5559595d769SMatthew Wilcox (Oracle) if (!folio_mapped(folio)) { 5564fc3f1d6SIngo Molnar up_read(&root_anon_vma->rwsem); 55788c22088SPeter Zijlstra anon_vma = NULL; 55888c22088SPeter Zijlstra } 55988c22088SPeter Zijlstra goto out; 56088c22088SPeter Zijlstra } 56188c22088SPeter Zijlstra 5626d4675e6SMinchan Kim if (rwc && rwc->try_lock) { 5636d4675e6SMinchan Kim anon_vma = NULL; 5646d4675e6SMinchan Kim rwc->contended = true; 5656d4675e6SMinchan Kim goto out; 5666d4675e6SMinchan Kim } 5676d4675e6SMinchan Kim 56888c22088SPeter Zijlstra /* trylock failed, we got to sleep */ 56988c22088SPeter Zijlstra if (!atomic_inc_not_zero(&anon_vma->refcount)) { 57088c22088SPeter Zijlstra anon_vma = NULL; 57188c22088SPeter Zijlstra goto out; 57288c22088SPeter Zijlstra } 57388c22088SPeter Zijlstra 5749595d769SMatthew Wilcox (Oracle) if (!folio_mapped(folio)) { 5757f39dda9SHugh Dickins rcu_read_unlock(); 57688c22088SPeter Zijlstra put_anon_vma(anon_vma); 5777f39dda9SHugh Dickins return NULL; 57888c22088SPeter Zijlstra } 57988c22088SPeter Zijlstra 58088c22088SPeter Zijlstra /* we pinned the anon_vma, its safe to sleep */ 58188c22088SPeter Zijlstra rcu_read_unlock(); 5824fc3f1d6SIngo Molnar anon_vma_lock_read(anon_vma); 583746b18d4SPeter Zijlstra 58488c22088SPeter Zijlstra if (atomic_dec_and_test(&anon_vma->refcount)) { 58588c22088SPeter Zijlstra /* 58688c22088SPeter Zijlstra * Oops, we held the last refcount, release the lock 58788c22088SPeter Zijlstra * and bail -- can't simply use put_anon_vma() because 5884fc3f1d6SIngo Molnar * we'll deadlock on the anon_vma_lock_write() recursion. 58988c22088SPeter Zijlstra */ 5904fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 59188c22088SPeter Zijlstra __put_anon_vma(anon_vma); 59288c22088SPeter Zijlstra anon_vma = NULL; 59388c22088SPeter Zijlstra } 59488c22088SPeter Zijlstra 59588c22088SPeter Zijlstra return anon_vma; 59688c22088SPeter Zijlstra 59788c22088SPeter Zijlstra out: 59888c22088SPeter Zijlstra rcu_read_unlock(); 599746b18d4SPeter Zijlstra return anon_vma; 60034bbd704SOleg Nesterov } 60134bbd704SOleg Nesterov 6024fc3f1d6SIngo Molnar void page_unlock_anon_vma_read(struct anon_vma *anon_vma) 60334bbd704SOleg Nesterov { 6044fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 6051da177e4SLinus Torvalds } 6061da177e4SLinus Torvalds 60772b252aeSMel Gorman #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH 60872b252aeSMel Gorman /* 60972b252aeSMel Gorman * Flush TLB entries for recently unmapped pages from remote CPUs. It is 61072b252aeSMel Gorman * important if a PTE was dirty when it was unmapped that it's flushed 61172b252aeSMel Gorman * before any IO is initiated on the page to prevent lost writes. Similarly, 61272b252aeSMel Gorman * it must be flushed before freeing to prevent data leakage. 61372b252aeSMel Gorman */ 61472b252aeSMel Gorman void try_to_unmap_flush(void) 61572b252aeSMel Gorman { 61672b252aeSMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 61772b252aeSMel Gorman 61872b252aeSMel Gorman if (!tlb_ubc->flush_required) 61972b252aeSMel Gorman return; 62072b252aeSMel Gorman 621e73ad5ffSAndy Lutomirski arch_tlbbatch_flush(&tlb_ubc->arch); 62272b252aeSMel Gorman tlb_ubc->flush_required = false; 623d950c947SMel Gorman tlb_ubc->writable = false; 62472b252aeSMel Gorman } 62572b252aeSMel Gorman 626d950c947SMel Gorman /* Flush iff there are potentially writable TLB entries that can race with IO */ 627d950c947SMel Gorman void try_to_unmap_flush_dirty(void) 628d950c947SMel Gorman { 629d950c947SMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 630d950c947SMel Gorman 631d950c947SMel Gorman if (tlb_ubc->writable) 632d950c947SMel Gorman try_to_unmap_flush(); 633d950c947SMel Gorman } 634d950c947SMel Gorman 6355ee2fa2fSHuang Ying /* 6365ee2fa2fSHuang Ying * Bits 0-14 of mm->tlb_flush_batched record pending generations. 6375ee2fa2fSHuang Ying * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations. 6385ee2fa2fSHuang Ying */ 6395ee2fa2fSHuang Ying #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16 6405ee2fa2fSHuang Ying #define TLB_FLUSH_BATCH_PENDING_MASK \ 6415ee2fa2fSHuang Ying ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1) 6425ee2fa2fSHuang Ying #define TLB_FLUSH_BATCH_PENDING_LARGE \ 6435ee2fa2fSHuang Ying (TLB_FLUSH_BATCH_PENDING_MASK / 2) 6445ee2fa2fSHuang Ying 645c7ab0d2fSKirill A. Shutemov static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) 64672b252aeSMel Gorman { 64772b252aeSMel Gorman struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 6485ee2fa2fSHuang Ying int batch, nbatch; 64972b252aeSMel Gorman 650e73ad5ffSAndy Lutomirski arch_tlbbatch_add_mm(&tlb_ubc->arch, mm); 65172b252aeSMel Gorman tlb_ubc->flush_required = true; 652d950c947SMel Gorman 653d950c947SMel Gorman /* 6543ea27719SMel Gorman * Ensure compiler does not re-order the setting of tlb_flush_batched 6553ea27719SMel Gorman * before the PTE is cleared. 6563ea27719SMel Gorman */ 6573ea27719SMel Gorman barrier(); 6585ee2fa2fSHuang Ying batch = atomic_read(&mm->tlb_flush_batched); 6595ee2fa2fSHuang Ying retry: 6605ee2fa2fSHuang Ying if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) { 6615ee2fa2fSHuang Ying /* 6625ee2fa2fSHuang Ying * Prevent `pending' from catching up with `flushed' because of 6635ee2fa2fSHuang Ying * overflow. Reset `pending' and `flushed' to be 1 and 0 if 6645ee2fa2fSHuang Ying * `pending' becomes large. 6655ee2fa2fSHuang Ying */ 6665ee2fa2fSHuang Ying nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1); 6675ee2fa2fSHuang Ying if (nbatch != batch) { 6685ee2fa2fSHuang Ying batch = nbatch; 6695ee2fa2fSHuang Ying goto retry; 6705ee2fa2fSHuang Ying } 6715ee2fa2fSHuang Ying } else { 6725ee2fa2fSHuang Ying atomic_inc(&mm->tlb_flush_batched); 6735ee2fa2fSHuang Ying } 6743ea27719SMel Gorman 6753ea27719SMel Gorman /* 676d950c947SMel Gorman * If the PTE was dirty then it's best to assume it's writable. The 677d950c947SMel Gorman * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() 678d950c947SMel Gorman * before the page is queued for IO. 679d950c947SMel Gorman */ 680d950c947SMel Gorman if (writable) 681d950c947SMel Gorman tlb_ubc->writable = true; 68272b252aeSMel Gorman } 68372b252aeSMel Gorman 68472b252aeSMel Gorman /* 68572b252aeSMel Gorman * Returns true if the TLB flush should be deferred to the end of a batch of 68672b252aeSMel Gorman * unmap operations to reduce IPIs. 68772b252aeSMel Gorman */ 68872b252aeSMel Gorman static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 68972b252aeSMel Gorman { 69072b252aeSMel Gorman bool should_defer = false; 69172b252aeSMel Gorman 69272b252aeSMel Gorman if (!(flags & TTU_BATCH_FLUSH)) 69372b252aeSMel Gorman return false; 69472b252aeSMel Gorman 69572b252aeSMel Gorman /* If remote CPUs need to be flushed then defer batch the flush */ 69672b252aeSMel Gorman if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids) 69772b252aeSMel Gorman should_defer = true; 69872b252aeSMel Gorman put_cpu(); 69972b252aeSMel Gorman 70072b252aeSMel Gorman return should_defer; 70172b252aeSMel Gorman } 7023ea27719SMel Gorman 7033ea27719SMel Gorman /* 7043ea27719SMel Gorman * Reclaim unmaps pages under the PTL but do not flush the TLB prior to 7053ea27719SMel Gorman * releasing the PTL if TLB flushes are batched. It's possible for a parallel 7063ea27719SMel Gorman * operation such as mprotect or munmap to race between reclaim unmapping 7073ea27719SMel Gorman * the page and flushing the page. If this race occurs, it potentially allows 7083ea27719SMel Gorman * access to data via a stale TLB entry. Tracking all mm's that have TLB 7093ea27719SMel Gorman * batching in flight would be expensive during reclaim so instead track 7103ea27719SMel Gorman * whether TLB batching occurred in the past and if so then do a flush here 7113ea27719SMel Gorman * if required. This will cost one additional flush per reclaim cycle paid 7123ea27719SMel Gorman * by the first operation at risk such as mprotect and mumap. 7133ea27719SMel Gorman * 7143ea27719SMel Gorman * This must be called under the PTL so that an access to tlb_flush_batched 7153ea27719SMel Gorman * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise 7163ea27719SMel Gorman * via the PTL. 7173ea27719SMel Gorman */ 7183ea27719SMel Gorman void flush_tlb_batched_pending(struct mm_struct *mm) 7193ea27719SMel Gorman { 7205ee2fa2fSHuang Ying int batch = atomic_read(&mm->tlb_flush_batched); 7215ee2fa2fSHuang Ying int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK; 7225ee2fa2fSHuang Ying int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT; 7233ea27719SMel Gorman 7245ee2fa2fSHuang Ying if (pending != flushed) { 7255ee2fa2fSHuang Ying flush_tlb_mm(mm); 7263ea27719SMel Gorman /* 7275ee2fa2fSHuang Ying * If the new TLB flushing is pending during flushing, leave 7285ee2fa2fSHuang Ying * mm->tlb_flush_batched as is, to avoid losing flushing. 7293ea27719SMel Gorman */ 7305ee2fa2fSHuang Ying atomic_cmpxchg(&mm->tlb_flush_batched, batch, 7315ee2fa2fSHuang Ying pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT)); 7323ea27719SMel Gorman } 7333ea27719SMel Gorman } 73472b252aeSMel Gorman #else 735c7ab0d2fSKirill A. Shutemov static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) 73672b252aeSMel Gorman { 73772b252aeSMel Gorman } 73872b252aeSMel Gorman 73972b252aeSMel Gorman static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 74072b252aeSMel Gorman { 74172b252aeSMel Gorman return false; 74272b252aeSMel Gorman } 74372b252aeSMel Gorman #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ 74472b252aeSMel Gorman 7451da177e4SLinus Torvalds /* 746bf89c8c8SHuang Shijie * At what user virtual address is page expected in vma? 747ab941e0fSNaoya Horiguchi * Caller should check the page is actually part of the vma. 7481da177e4SLinus Torvalds */ 7491da177e4SLinus Torvalds unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 7501da177e4SLinus Torvalds { 751e05b3453SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 752e05b3453SMatthew Wilcox (Oracle) if (folio_test_anon(folio)) { 753e05b3453SMatthew Wilcox (Oracle) struct anon_vma *page__anon_vma = folio_anon_vma(folio); 7544829b906SHugh Dickins /* 7554829b906SHugh Dickins * Note: swapoff's unuse_vma() is more efficient with this 7564829b906SHugh Dickins * check, and needs it to match anon_vma when KSM is active. 7574829b906SHugh Dickins */ 7584829b906SHugh Dickins if (!vma->anon_vma || !page__anon_vma || 7594829b906SHugh Dickins vma->anon_vma->root != page__anon_vma->root) 76021d0d443SAndrea Arcangeli return -EFAULT; 76131657170SJue Wang } else if (!vma->vm_file) { 7621da177e4SLinus Torvalds return -EFAULT; 763e05b3453SMatthew Wilcox (Oracle) } else if (vma->vm_file->f_mapping != folio->mapping) { 7641da177e4SLinus Torvalds return -EFAULT; 76531657170SJue Wang } 766494334e4SHugh Dickins 767494334e4SHugh Dickins return vma_address(page, vma); 7681da177e4SLinus Torvalds } 7691da177e4SLinus Torvalds 77050722804SZach O'Keefe /* 77150722804SZach O'Keefe * Returns the actual pmd_t* where we expect 'address' to be mapped from, or 77250722804SZach O'Keefe * NULL if it doesn't exist. No guarantees / checks on what the pmd_t* 77350722804SZach O'Keefe * represents. 77450722804SZach O'Keefe */ 7756219049aSBob Liu pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) 7766219049aSBob Liu { 7776219049aSBob Liu pgd_t *pgd; 778c2febafcSKirill A. Shutemov p4d_t *p4d; 7796219049aSBob Liu pud_t *pud; 7806219049aSBob Liu pmd_t *pmd = NULL; 7816219049aSBob Liu 7826219049aSBob Liu pgd = pgd_offset(mm, address); 7836219049aSBob Liu if (!pgd_present(*pgd)) 7846219049aSBob Liu goto out; 7856219049aSBob Liu 786c2febafcSKirill A. Shutemov p4d = p4d_offset(pgd, address); 787c2febafcSKirill A. Shutemov if (!p4d_present(*p4d)) 788c2febafcSKirill A. Shutemov goto out; 789c2febafcSKirill A. Shutemov 790c2febafcSKirill A. Shutemov pud = pud_offset(p4d, address); 7916219049aSBob Liu if (!pud_present(*pud)) 7926219049aSBob Liu goto out; 7936219049aSBob Liu 7946219049aSBob Liu pmd = pmd_offset(pud, address); 7956219049aSBob Liu out: 7966219049aSBob Liu return pmd; 7976219049aSBob Liu } 7986219049aSBob Liu 799b3ac0413SMatthew Wilcox (Oracle) struct folio_referenced_arg { 8009f32624bSJoonsoo Kim int mapcount; 8019f32624bSJoonsoo Kim int referenced; 8029f32624bSJoonsoo Kim unsigned long vm_flags; 8039f32624bSJoonsoo Kim struct mem_cgroup *memcg; 8049f32624bSJoonsoo Kim }; 80581b4082dSNikita Danilov /* 806b3ac0413SMatthew Wilcox (Oracle) * arg: folio_referenced_arg will be passed 8071da177e4SLinus Torvalds */ 8082f031c6fSMatthew Wilcox (Oracle) static bool folio_referenced_one(struct folio *folio, 8092f031c6fSMatthew Wilcox (Oracle) struct vm_area_struct *vma, unsigned long address, void *arg) 8101da177e4SLinus Torvalds { 811b3ac0413SMatthew Wilcox (Oracle) struct folio_referenced_arg *pra = arg; 812b3ac0413SMatthew Wilcox (Oracle) DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); 8138749cfeaSVladimir Davydov int referenced = 0; 8142da28bfdSAndrea Arcangeli 8158eaededeSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 8168eaededeSKirill A. Shutemov address = pvmw.address; 8172da28bfdSAndrea Arcangeli 81847d4f3eeSHugh Dickins if ((vma->vm_flags & VM_LOCKED) && 819b3ac0413SMatthew Wilcox (Oracle) (!folio_test_large(folio) || !pvmw.pte)) { 82047d4f3eeSHugh Dickins /* Restore the mlock which got missed */ 821b3ac0413SMatthew Wilcox (Oracle) mlock_vma_folio(folio, vma, !pvmw.pte); 8228eaededeSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 8239f32624bSJoonsoo Kim pra->vm_flags |= VM_LOCKED; 824e4b82222SMinchan Kim return false; /* To break the loop */ 8252da28bfdSAndrea Arcangeli } 8262da28bfdSAndrea Arcangeli 8278eaededeSKirill A. Shutemov if (pvmw.pte) { 8288eaededeSKirill A. Shutemov if (ptep_clear_flush_young_notify(vma, address, 8298eaededeSKirill A. Shutemov pvmw.pte)) { 8304917e5d0SJohannes Weiner /* 8318eaededeSKirill A. Shutemov * Don't treat a reference through 8328eaededeSKirill A. Shutemov * a sequentially read mapping as such. 833b3ac0413SMatthew Wilcox (Oracle) * If the folio has been used in another mapping, 8348eaededeSKirill A. Shutemov * we will catch it; if this other mapping is 8358eaededeSKirill A. Shutemov * already gone, the unmap path will have set 836b3ac0413SMatthew Wilcox (Oracle) * the referenced flag or activated the folio. 8374917e5d0SJohannes Weiner */ 83864363aadSJoe Perches if (likely(!(vma->vm_flags & VM_SEQ_READ))) 8391da177e4SLinus Torvalds referenced++; 8404917e5d0SJohannes Weiner } 8418749cfeaSVladimir Davydov } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { 8428eaededeSKirill A. Shutemov if (pmdp_clear_flush_young_notify(vma, address, 8438eaededeSKirill A. Shutemov pvmw.pmd)) 8448749cfeaSVladimir Davydov referenced++; 8458749cfeaSVladimir Davydov } else { 846b3ac0413SMatthew Wilcox (Oracle) /* unexpected pmd-mapped folio? */ 8478749cfeaSVladimir Davydov WARN_ON_ONCE(1); 8488749cfeaSVladimir Davydov } 8498eaededeSKirill A. Shutemov 8508eaededeSKirill A. Shutemov pra->mapcount--; 8518eaededeSKirill A. Shutemov } 85271e3aac0SAndrea Arcangeli 85333c3fc71SVladimir Davydov if (referenced) 854b3ac0413SMatthew Wilcox (Oracle) folio_clear_idle(folio); 855b3ac0413SMatthew Wilcox (Oracle) if (folio_test_clear_young(folio)) 85633c3fc71SVladimir Davydov referenced++; 85733c3fc71SVladimir Davydov 8589f32624bSJoonsoo Kim if (referenced) { 8599f32624bSJoonsoo Kim pra->referenced++; 86047d4f3eeSHugh Dickins pra->vm_flags |= vma->vm_flags & ~VM_LOCKED; 8611da177e4SLinus Torvalds } 8621da177e4SLinus Torvalds 8639f32624bSJoonsoo Kim if (!pra->mapcount) 864e4b82222SMinchan Kim return false; /* To break the loop */ 8659f32624bSJoonsoo Kim 866e4b82222SMinchan Kim return true; 8679f32624bSJoonsoo Kim } 8689f32624bSJoonsoo Kim 869b3ac0413SMatthew Wilcox (Oracle) static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg) 8701da177e4SLinus Torvalds { 871b3ac0413SMatthew Wilcox (Oracle) struct folio_referenced_arg *pra = arg; 8729f32624bSJoonsoo Kim struct mem_cgroup *memcg = pra->memcg; 8731da177e4SLinus Torvalds 8749f32624bSJoonsoo Kim if (!mm_match_cgroup(vma->vm_mm, memcg)) 8759f32624bSJoonsoo Kim return true; 8761da177e4SLinus Torvalds 8779f32624bSJoonsoo Kim return false; 8781da177e4SLinus Torvalds } 8791da177e4SLinus Torvalds 8801da177e4SLinus Torvalds /** 881b3ac0413SMatthew Wilcox (Oracle) * folio_referenced() - Test if the folio was referenced. 882b3ac0413SMatthew Wilcox (Oracle) * @folio: The folio to test. 883b3ac0413SMatthew Wilcox (Oracle) * @is_locked: Caller holds lock on the folio. 88472835c86SJohannes Weiner * @memcg: target memory cgroup 885b3ac0413SMatthew Wilcox (Oracle) * @vm_flags: A combination of all the vma->vm_flags which referenced the folio. 8861da177e4SLinus Torvalds * 887b3ac0413SMatthew Wilcox (Oracle) * Quick test_and_clear_referenced for all mappings of a folio, 888b3ac0413SMatthew Wilcox (Oracle) * 8896d4675e6SMinchan Kim * Return: The number of mappings which referenced the folio. Return -1 if 8906d4675e6SMinchan Kim * the function bailed out due to rmap lock contention. 8911da177e4SLinus Torvalds */ 892b3ac0413SMatthew Wilcox (Oracle) int folio_referenced(struct folio *folio, int is_locked, 893b3ac0413SMatthew Wilcox (Oracle) struct mem_cgroup *memcg, unsigned long *vm_flags) 8941da177e4SLinus Torvalds { 8955ad64688SHugh Dickins int we_locked = 0; 896b3ac0413SMatthew Wilcox (Oracle) struct folio_referenced_arg pra = { 897b3ac0413SMatthew Wilcox (Oracle) .mapcount = folio_mapcount(folio), 8989f32624bSJoonsoo Kim .memcg = memcg, 8999f32624bSJoonsoo Kim }; 9009f32624bSJoonsoo Kim struct rmap_walk_control rwc = { 901b3ac0413SMatthew Wilcox (Oracle) .rmap_one = folio_referenced_one, 9029f32624bSJoonsoo Kim .arg = (void *)&pra, 9032f031c6fSMatthew Wilcox (Oracle) .anon_lock = folio_lock_anon_vma_read, 9046d4675e6SMinchan Kim .try_lock = true, 9059f32624bSJoonsoo Kim }; 9061da177e4SLinus Torvalds 9076fe6b7e3SWu Fengguang *vm_flags = 0; 908059d8442SHuang Shijie if (!pra.mapcount) 9099f32624bSJoonsoo Kim return 0; 9109f32624bSJoonsoo Kim 911b3ac0413SMatthew Wilcox (Oracle) if (!folio_raw_mapping(folio)) 9129f32624bSJoonsoo Kim return 0; 9139f32624bSJoonsoo Kim 914b3ac0413SMatthew Wilcox (Oracle) if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) { 915b3ac0413SMatthew Wilcox (Oracle) we_locked = folio_trylock(folio); 9169f32624bSJoonsoo Kim if (!we_locked) 9179f32624bSJoonsoo Kim return 1; 9185ad64688SHugh Dickins } 9199f32624bSJoonsoo Kim 9209f32624bSJoonsoo Kim /* 9219f32624bSJoonsoo Kim * If we are reclaiming on behalf of a cgroup, skip 9229f32624bSJoonsoo Kim * counting on behalf of references from different 9239f32624bSJoonsoo Kim * cgroups 9249f32624bSJoonsoo Kim */ 9259f32624bSJoonsoo Kim if (memcg) { 926b3ac0413SMatthew Wilcox (Oracle) rwc.invalid_vma = invalid_folio_referenced_vma; 9275ad64688SHugh Dickins } 9289f32624bSJoonsoo Kim 9292f031c6fSMatthew Wilcox (Oracle) rmap_walk(folio, &rwc); 9309f32624bSJoonsoo Kim *vm_flags = pra.vm_flags; 9319f32624bSJoonsoo Kim 9325ad64688SHugh Dickins if (we_locked) 933b3ac0413SMatthew Wilcox (Oracle) folio_unlock(folio); 9349f32624bSJoonsoo Kim 9356d4675e6SMinchan Kim return rwc.contended ? -1 : pra.referenced; 9361da177e4SLinus Torvalds } 9371da177e4SLinus Torvalds 9386a8e0596SMuchun Song static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw) 939d08b3851SPeter Zijlstra { 9406a8e0596SMuchun Song int cleaned = 0; 9416a8e0596SMuchun Song struct vm_area_struct *vma = pvmw->vma; 942ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 9436a8e0596SMuchun Song unsigned long address = pvmw->address; 944d08b3851SPeter Zijlstra 945369ea824SJérôme Glisse /* 946369ea824SJérôme Glisse * We have to assume the worse case ie pmd for invalidation. Note that 947e83c09a2SMatthew Wilcox (Oracle) * the folio can not be freed from this function. 948369ea824SJérôme Glisse */ 9497269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 9507269f999SJérôme Glisse 0, vma, vma->vm_mm, address, 9516a8e0596SMuchun Song vma_address_end(pvmw)); 952ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 953369ea824SJérôme Glisse 9546a8e0596SMuchun Song while (page_vma_mapped_walk(pvmw)) { 955f27176cfSKirill A. Shutemov int ret = 0; 956369ea824SJérôme Glisse 9576a8e0596SMuchun Song address = pvmw->address; 9586a8e0596SMuchun Song if (pvmw->pte) { 959c2fda5feSPeter Zijlstra pte_t entry; 9606a8e0596SMuchun Song pte_t *pte = pvmw->pte; 961f27176cfSKirill A. Shutemov 962f27176cfSKirill A. Shutemov if (!pte_dirty(*pte) && !pte_write(*pte)) 963f27176cfSKirill A. Shutemov continue; 964d08b3851SPeter Zijlstra 965785373b4SLinus Torvalds flush_cache_page(vma, address, pte_pfn(*pte)); 966785373b4SLinus Torvalds entry = ptep_clear_flush(vma, address, pte); 967d08b3851SPeter Zijlstra entry = pte_wrprotect(entry); 968c2fda5feSPeter Zijlstra entry = pte_mkclean(entry); 969785373b4SLinus Torvalds set_pte_at(vma->vm_mm, address, pte, entry); 970d08b3851SPeter Zijlstra ret = 1; 971f27176cfSKirill A. Shutemov } else { 972396bcc52SMatthew Wilcox (Oracle) #ifdef CONFIG_TRANSPARENT_HUGEPAGE 9736a8e0596SMuchun Song pmd_t *pmd = pvmw->pmd; 974f27176cfSKirill A. Shutemov pmd_t entry; 975d08b3851SPeter Zijlstra 976f27176cfSKirill A. Shutemov if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) 977f27176cfSKirill A. Shutemov continue; 978f27176cfSKirill A. Shutemov 9797f9c9b60SMuchun Song flush_cache_range(vma, address, 9807f9c9b60SMuchun Song address + HPAGE_PMD_SIZE); 981024eee0eSAneesh Kumar K.V entry = pmdp_invalidate(vma, address, pmd); 982f27176cfSKirill A. Shutemov entry = pmd_wrprotect(entry); 983f27176cfSKirill A. Shutemov entry = pmd_mkclean(entry); 984785373b4SLinus Torvalds set_pmd_at(vma->vm_mm, address, pmd, entry); 985f27176cfSKirill A. Shutemov ret = 1; 986f27176cfSKirill A. Shutemov #else 987e83c09a2SMatthew Wilcox (Oracle) /* unexpected pmd-mapped folio? */ 988f27176cfSKirill A. Shutemov WARN_ON_ONCE(1); 989f27176cfSKirill A. Shutemov #endif 990f27176cfSKirill A. Shutemov } 9912ec74c3eSSagi Grimberg 9920f10851eSJérôme Glisse /* 9930f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() as we are 9940f10851eSJérôme Glisse * downgrading page table protection not changing it to point 9950f10851eSJérôme Glisse * to a new page. 9960f10851eSJérôme Glisse * 997ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 9980f10851eSJérôme Glisse */ 9990f10851eSJérôme Glisse if (ret) 10006a8e0596SMuchun Song cleaned++; 10019853a407SJoonsoo Kim } 1002f27176cfSKirill A. Shutemov 1003ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1004369ea824SJérôme Glisse 10056a8e0596SMuchun Song return cleaned; 10066a8e0596SMuchun Song } 10076a8e0596SMuchun Song 10086a8e0596SMuchun Song static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma, 10096a8e0596SMuchun Song unsigned long address, void *arg) 10106a8e0596SMuchun Song { 10116a8e0596SMuchun Song DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC); 10126a8e0596SMuchun Song int *cleaned = arg; 10136a8e0596SMuchun Song 10146a8e0596SMuchun Song *cleaned += page_vma_mkclean_one(&pvmw); 10156a8e0596SMuchun Song 1016e4b82222SMinchan Kim return true; 1017d08b3851SPeter Zijlstra } 1018d08b3851SPeter Zijlstra 10199853a407SJoonsoo Kim static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) 1020d08b3851SPeter Zijlstra { 10219853a407SJoonsoo Kim if (vma->vm_flags & VM_SHARED) 1022871beb8cSFengguang Wu return false; 1023d08b3851SPeter Zijlstra 1024871beb8cSFengguang Wu return true; 1025d08b3851SPeter Zijlstra } 1026d08b3851SPeter Zijlstra 1027d9c08e22SMatthew Wilcox (Oracle) int folio_mkclean(struct folio *folio) 1028d08b3851SPeter Zijlstra { 10299853a407SJoonsoo Kim int cleaned = 0; 10309853a407SJoonsoo Kim struct address_space *mapping; 10319853a407SJoonsoo Kim struct rmap_walk_control rwc = { 10329853a407SJoonsoo Kim .arg = (void *)&cleaned, 10339853a407SJoonsoo Kim .rmap_one = page_mkclean_one, 10349853a407SJoonsoo Kim .invalid_vma = invalid_mkclean_vma, 10359853a407SJoonsoo Kim }; 1036d08b3851SPeter Zijlstra 1037d9c08e22SMatthew Wilcox (Oracle) BUG_ON(!folio_test_locked(folio)); 1038d08b3851SPeter Zijlstra 1039d9c08e22SMatthew Wilcox (Oracle) if (!folio_mapped(folio)) 10409853a407SJoonsoo Kim return 0; 1041d08b3851SPeter Zijlstra 1042d9c08e22SMatthew Wilcox (Oracle) mapping = folio_mapping(folio); 10439853a407SJoonsoo Kim if (!mapping) 10449853a407SJoonsoo Kim return 0; 10459853a407SJoonsoo Kim 10462f031c6fSMatthew Wilcox (Oracle) rmap_walk(folio, &rwc); 10479853a407SJoonsoo Kim 10489853a407SJoonsoo Kim return cleaned; 1049d08b3851SPeter Zijlstra } 1050d9c08e22SMatthew Wilcox (Oracle) EXPORT_SYMBOL_GPL(folio_mkclean); 1051d08b3851SPeter Zijlstra 10521da177e4SLinus Torvalds /** 10536a8e0596SMuchun Song * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of 10546a8e0596SMuchun Song * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff) 10556a8e0596SMuchun Song * within the @vma of shared mappings. And since clean PTEs 10566a8e0596SMuchun Song * should also be readonly, write protects them too. 10576a8e0596SMuchun Song * @pfn: start pfn. 10586a8e0596SMuchun Song * @nr_pages: number of physically contiguous pages srarting with @pfn. 10596a8e0596SMuchun Song * @pgoff: page offset that the @pfn mapped with. 10606a8e0596SMuchun Song * @vma: vma that @pfn mapped within. 10616a8e0596SMuchun Song * 10626a8e0596SMuchun Song * Returns the number of cleaned PTEs (including PMDs). 10636a8e0596SMuchun Song */ 10646a8e0596SMuchun Song int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff, 10656a8e0596SMuchun Song struct vm_area_struct *vma) 10666a8e0596SMuchun Song { 10676a8e0596SMuchun Song struct page_vma_mapped_walk pvmw = { 10686a8e0596SMuchun Song .pfn = pfn, 10696a8e0596SMuchun Song .nr_pages = nr_pages, 10706a8e0596SMuchun Song .pgoff = pgoff, 10716a8e0596SMuchun Song .vma = vma, 10726a8e0596SMuchun Song .flags = PVMW_SYNC, 10736a8e0596SMuchun Song }; 10746a8e0596SMuchun Song 10756a8e0596SMuchun Song if (invalid_mkclean_vma(vma, NULL)) 10766a8e0596SMuchun Song return 0; 10776a8e0596SMuchun Song 10786a8e0596SMuchun Song pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma); 10796a8e0596SMuchun Song VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma); 10806a8e0596SMuchun Song 10816a8e0596SMuchun Song return page_vma_mkclean_one(&pvmw); 10826a8e0596SMuchun Song } 10836a8e0596SMuchun Song 10846a8e0596SMuchun Song /** 1085c44b6743SRik van Riel * page_move_anon_rmap - move a page to our anon_vma 1086c44b6743SRik van Riel * @page: the page to move to our anon_vma 1087c44b6743SRik van Riel * @vma: the vma the page belongs to 1088c44b6743SRik van Riel * 1089c44b6743SRik van Riel * When a page belongs exclusively to one process after a COW event, 1090c44b6743SRik van Riel * that page can be moved into the anon_vma that belongs to just that 1091c44b6743SRik van Riel * process, so the rmap code will not search the parent or sibling 1092c44b6743SRik van Riel * processes. 1093c44b6743SRik van Riel */ 10945a49973dSHugh Dickins void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma) 1095c44b6743SRik van Riel { 1096c44b6743SRik van Riel struct anon_vma *anon_vma = vma->anon_vma; 10976c287605SDavid Hildenbrand struct page *subpage = page; 1098c44b6743SRik van Riel 10995a49973dSHugh Dickins page = compound_head(page); 11005a49973dSHugh Dickins 1101309381feSSasha Levin VM_BUG_ON_PAGE(!PageLocked(page), page); 110281d1b09cSSasha Levin VM_BUG_ON_VMA(!anon_vma, vma); 1103c44b6743SRik van Riel 1104c44b6743SRik van Riel anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1105414e2fb8SVladimir Davydov /* 1106414e2fb8SVladimir Davydov * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written 1107b3ac0413SMatthew Wilcox (Oracle) * simultaneously, so a concurrent reader (eg folio_referenced()'s 1108b3ac0413SMatthew Wilcox (Oracle) * folio_test_anon()) will not see one without the other. 1109414e2fb8SVladimir Davydov */ 1110414e2fb8SVladimir Davydov WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); 11116c287605SDavid Hildenbrand SetPageAnonExclusive(subpage); 1112c44b6743SRik van Riel } 1113c44b6743SRik van Riel 1114c44b6743SRik van Riel /** 111543d8eac4SRandy Dunlap * __page_set_anon_rmap - set up new anonymous rmap 1116451b9514SKirill Tkhai * @page: Page or Hugepage to add to rmap 11174e1c1975SAndi Kleen * @vma: VM area to add page to. 11184e1c1975SAndi Kleen * @address: User virtual address of the mapping 1119e8a03febSRik van Riel * @exclusive: the page is exclusively owned by the current process 11201da177e4SLinus Torvalds */ 11219617d95eSNick Piggin static void __page_set_anon_rmap(struct page *page, 1122e8a03febSRik van Riel struct vm_area_struct *vma, unsigned long address, int exclusive) 11231da177e4SLinus Torvalds { 1124e8a03febSRik van Riel struct anon_vma *anon_vma = vma->anon_vma; 11252822c1aaSNick Piggin 1126e8a03febSRik van Riel BUG_ON(!anon_vma); 1127ea90002bSLinus Torvalds 11284e1c1975SAndi Kleen if (PageAnon(page)) 11296c287605SDavid Hildenbrand goto out; 11304e1c1975SAndi Kleen 1131ea90002bSLinus Torvalds /* 1132e8a03febSRik van Riel * If the page isn't exclusively mapped into this vma, 1133e8a03febSRik van Riel * we must use the _oldest_ possible anon_vma for the 1134e8a03febSRik van Riel * page mapping! 1135ea90002bSLinus Torvalds */ 11364e1c1975SAndi Kleen if (!exclusive) 1137288468c3SAndrea Arcangeli anon_vma = anon_vma->root; 1138ea90002bSLinus Torvalds 113916f5e707SAlex Shi /* 114016f5e707SAlex Shi * page_idle does a lockless/optimistic rmap scan on page->mapping. 114116f5e707SAlex Shi * Make sure the compiler doesn't split the stores of anon_vma and 114216f5e707SAlex Shi * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code 114316f5e707SAlex Shi * could mistake the mapping for a struct address_space and crash. 114416f5e707SAlex Shi */ 11451da177e4SLinus Torvalds anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 114616f5e707SAlex Shi WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); 11474d7670e0SNick Piggin page->index = linear_page_index(vma, address); 11486c287605SDavid Hildenbrand out: 11496c287605SDavid Hildenbrand if (exclusive) 11506c287605SDavid Hildenbrand SetPageAnonExclusive(page); 11511da177e4SLinus Torvalds } 11529617d95eSNick Piggin 11539617d95eSNick Piggin /** 115443d8eac4SRandy Dunlap * __page_check_anon_rmap - sanity check anonymous rmap addition 1155c97a9e10SNick Piggin * @page: the page to add the mapping to 1156c97a9e10SNick Piggin * @vma: the vm area in which the mapping is added 1157c97a9e10SNick Piggin * @address: the user virtual address mapped 1158c97a9e10SNick Piggin */ 1159c97a9e10SNick Piggin static void __page_check_anon_rmap(struct page *page, 1160c97a9e10SNick Piggin struct vm_area_struct *vma, unsigned long address) 1161c97a9e10SNick Piggin { 1162e05b3453SMatthew Wilcox (Oracle) struct folio *folio = page_folio(page); 1163c97a9e10SNick Piggin /* 1164c97a9e10SNick Piggin * The page's anon-rmap details (mapping and index) are guaranteed to 1165c97a9e10SNick Piggin * be set up correctly at this point. 1166c97a9e10SNick Piggin * 1167c97a9e10SNick Piggin * We have exclusion against page_add_anon_rmap because the caller 116890aaca85SMiaohe Lin * always holds the page locked. 1169c97a9e10SNick Piggin * 1170c97a9e10SNick Piggin * We have exclusion against page_add_new_anon_rmap because those pages 1171c97a9e10SNick Piggin * are initially only visible via the pagetables, and the pte is locked 1172c97a9e10SNick Piggin * over the call to page_add_new_anon_rmap. 1173c97a9e10SNick Piggin */ 1174e05b3453SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root, 1175e05b3453SMatthew Wilcox (Oracle) folio); 117630c46382SYang Shi VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address), 117730c46382SYang Shi page); 1178c97a9e10SNick Piggin } 1179c97a9e10SNick Piggin 1180c97a9e10SNick Piggin /** 11819617d95eSNick Piggin * page_add_anon_rmap - add pte mapping to an anonymous page 11829617d95eSNick Piggin * @page: the page to add the mapping to 11839617d95eSNick Piggin * @vma: the vm area in which the mapping is added 11849617d95eSNick Piggin * @address: the user virtual address mapped 1185f1e2db12SDavid Hildenbrand * @flags: the rmap flags 11869617d95eSNick Piggin * 11875ad64688SHugh Dickins * The caller needs to hold the pte lock, and the page must be locked in 118880e14822SHugh Dickins * the anon_vma case: to serialize mapping,index checking after setting, 118980e14822SHugh Dickins * and to ensure that PageAnon is not being upgraded racily to PageKsm 119080e14822SHugh Dickins * (but PageKsm is never downgraded to PageAnon). 11919617d95eSNick Piggin */ 11929617d95eSNick Piggin void page_add_anon_rmap(struct page *page, 119314f9135dSDavid Hildenbrand struct vm_area_struct *vma, unsigned long address, rmap_t flags) 1194ad8c2ee8SRik van Riel { 1195d281ee61SKirill A. Shutemov bool compound = flags & RMAP_COMPOUND; 119653f9263bSKirill A. Shutemov bool first; 119753f9263bSKirill A. Shutemov 1198be5d0a74SJohannes Weiner if (unlikely(PageKsm(page))) 1199be5d0a74SJohannes Weiner lock_page_memcg(page); 1200be5d0a74SJohannes Weiner else 1201be5d0a74SJohannes Weiner VM_BUG_ON_PAGE(!PageLocked(page), page); 1202be5d0a74SJohannes Weiner 120353f9263bSKirill A. Shutemov if (compound) { 120453f9263bSKirill A. Shutemov atomic_t *mapcount; 1205e9b61f19SKirill A. Shutemov VM_BUG_ON_PAGE(!PageLocked(page), page); 120653f9263bSKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 120753f9263bSKirill A. Shutemov mapcount = compound_mapcount_ptr(page); 120853f9263bSKirill A. Shutemov first = atomic_inc_and_test(mapcount); 120953f9263bSKirill A. Shutemov } else { 121053f9263bSKirill A. Shutemov first = atomic_inc_and_test(&page->_mapcount); 121153f9263bSKirill A. Shutemov } 12126c287605SDavid Hildenbrand VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page); 12136c287605SDavid Hildenbrand VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page); 121453f9263bSKirill A. Shutemov 121553f9263bSKirill A. Shutemov if (first) { 12166c357848SMatthew Wilcox (Oracle) int nr = compound ? thp_nr_pages(page) : 1; 1217bea04b07SJianyu Zhan /* 1218bea04b07SJianyu Zhan * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1219bea04b07SJianyu Zhan * these counters are not modified in interrupt context, and 1220bea04b07SJianyu Zhan * pte lock(a spinlock) is held, which implies preemption 1221bea04b07SJianyu Zhan * disabled. 1222bea04b07SJianyu Zhan */ 122365c45377SKirill A. Shutemov if (compound) 122469473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, nr); 1225be5d0a74SJohannes Weiner __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr); 122679134171SAndrea Arcangeli } 12275ad64688SHugh Dickins 1228cea86fe2SHugh Dickins if (unlikely(PageKsm(page))) 1229be5d0a74SJohannes Weiner unlock_page_memcg(page); 123053f9263bSKirill A. Shutemov 12315dbe0af4SHugh Dickins /* address might be in next vma when migration races vma_adjust */ 1232cea86fe2SHugh Dickins else if (first) 1233d281ee61SKirill A. Shutemov __page_set_anon_rmap(page, vma, address, 123414f9135dSDavid Hildenbrand !!(flags & RMAP_EXCLUSIVE)); 123569029cd5SKAMEZAWA Hiroyuki else 1236c97a9e10SNick Piggin __page_check_anon_rmap(page, vma, address); 1237cea86fe2SHugh Dickins 1238cea86fe2SHugh Dickins mlock_vma_page(page, vma, compound); 12391da177e4SLinus Torvalds } 12401da177e4SLinus Torvalds 124143d8eac4SRandy Dunlap /** 124240f2bbf7SDavid Hildenbrand * page_add_new_anon_rmap - add mapping to a new anonymous page 12439617d95eSNick Piggin * @page: the page to add the mapping to 12449617d95eSNick Piggin * @vma: the vm area in which the mapping is added 12459617d95eSNick Piggin * @address: the user virtual address mapped 124640f2bbf7SDavid Hildenbrand * 124740f2bbf7SDavid Hildenbrand * If it's a compound page, it is accounted as a compound page. As the page 124840f2bbf7SDavid Hildenbrand * is new, it's assume to get mapped exclusively by a single process. 12499617d95eSNick Piggin * 12509617d95eSNick Piggin * Same as page_add_anon_rmap but must only be called on *new* pages. 12519617d95eSNick Piggin * This means the inc-and-test can be bypassed. 1252c97a9e10SNick Piggin * Page does not have to be locked. 12539617d95eSNick Piggin */ 12549617d95eSNick Piggin void page_add_new_anon_rmap(struct page *page, 125540f2bbf7SDavid Hildenbrand struct vm_area_struct *vma, unsigned long address) 12569617d95eSNick Piggin { 125740f2bbf7SDavid Hildenbrand const bool compound = PageCompound(page); 12586c357848SMatthew Wilcox (Oracle) int nr = compound ? thp_nr_pages(page) : 1; 1259d281ee61SKirill A. Shutemov 126081d1b09cSSasha Levin VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); 1261fa9949daSHugh Dickins __SetPageSwapBacked(page); 1262d281ee61SKirill A. Shutemov if (compound) { 1263d281ee61SKirill A. Shutemov VM_BUG_ON_PAGE(!PageTransHuge(page), page); 126453f9263bSKirill A. Shutemov /* increment count (starts at -1) */ 126553f9263bSKirill A. Shutemov atomic_set(compound_mapcount_ptr(page), 0); 126647e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 126747e29d32SJohn Hubbard 126869473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, nr); 126953f9263bSKirill A. Shutemov } else { 127053f9263bSKirill A. Shutemov /* increment count (starts at -1) */ 127153f9263bSKirill A. Shutemov atomic_set(&page->_mapcount, 0); 1272d281ee61SKirill A. Shutemov } 1273be5d0a74SJohannes Weiner __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr); 1274e8a03febSRik van Riel __page_set_anon_rmap(page, vma, address, 1); 12759617d95eSNick Piggin } 12769617d95eSNick Piggin 12771da177e4SLinus Torvalds /** 12781da177e4SLinus Torvalds * page_add_file_rmap - add pte mapping to a file page 12791da177e4SLinus Torvalds * @page: the page to add the mapping to 1280cea86fe2SHugh Dickins * @vma: the vm area in which the mapping is added 1281e8b098fcSMike Rapoport * @compound: charge the page as compound or small page 12821da177e4SLinus Torvalds * 1283b8072f09SHugh Dickins * The caller needs to hold the pte lock. 12841da177e4SLinus Torvalds */ 1285cea86fe2SHugh Dickins void page_add_file_rmap(struct page *page, 1286cea86fe2SHugh Dickins struct vm_area_struct *vma, bool compound) 12871da177e4SLinus Torvalds { 12885d543f13SHugh Dickins int i, nr = 0; 1289dd78feddSKirill A. Shutemov 1290dd78feddSKirill A. Shutemov VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page); 129162cccb8cSJohannes Weiner lock_page_memcg(page); 1292dd78feddSKirill A. Shutemov if (compound && PageTransHuge(page)) { 1293a1528e21SMuchun Song int nr_pages = thp_nr_pages(page); 1294a1528e21SMuchun Song 12955d543f13SHugh Dickins for (i = 0; i < nr_pages; i++) { 1296dd78feddSKirill A. Shutemov if (atomic_inc_and_test(&page[i]._mapcount)) 1297dd78feddSKirill A. Shutemov nr++; 1298d69b042fSBalbir Singh } 1299dd78feddSKirill A. Shutemov if (!atomic_inc_and_test(compound_mapcount_ptr(page))) 1300dd78feddSKirill A. Shutemov goto out; 1301bd55b0c2SHugh Dickins 1302bd55b0c2SHugh Dickins /* 1303bd55b0c2SHugh Dickins * It is racy to ClearPageDoubleMap in page_remove_file_rmap(); 1304bd55b0c2SHugh Dickins * but page lock is held by all page_add_file_rmap() compound 1305bd55b0c2SHugh Dickins * callers, and SetPageDoubleMap below warns if !PageLocked: 1306bd55b0c2SHugh Dickins * so here is a place that DoubleMap can be safely cleared. 1307bd55b0c2SHugh Dickins */ 1308bd55b0c2SHugh Dickins VM_WARN_ON_ONCE(!PageLocked(page)); 1309bd55b0c2SHugh Dickins if (nr == nr_pages && PageDoubleMap(page)) 1310bd55b0c2SHugh Dickins ClearPageDoubleMap(page); 1311bd55b0c2SHugh Dickins 131299cb0dbdSSong Liu if (PageSwapBacked(page)) 1313a1528e21SMuchun Song __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED, 1314a1528e21SMuchun Song nr_pages); 131599cb0dbdSSong Liu else 1316380780e7SMuchun Song __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED, 1317380780e7SMuchun Song nr_pages); 1318dd78feddSKirill A. Shutemov } else { 1319c8efc390SKirill A. Shutemov if (PageTransCompound(page) && page_mapping(page)) { 1320c8efc390SKirill A. Shutemov VM_WARN_ON_ONCE(!PageLocked(page)); 1321cea86fe2SHugh Dickins SetPageDoubleMap(compound_head(page)); 13229a73f61bSKirill A. Shutemov } 13235d543f13SHugh Dickins if (atomic_inc_and_test(&page->_mapcount)) 13245d543f13SHugh Dickins nr++; 1325dd78feddSKirill A. Shutemov } 1326dd78feddSKirill A. Shutemov out: 13275d543f13SHugh Dickins if (nr) 13285d543f13SHugh Dickins __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr); 132962cccb8cSJohannes Weiner unlock_page_memcg(page); 1330cea86fe2SHugh Dickins 1331cea86fe2SHugh Dickins mlock_vma_page(page, vma, compound); 13321da177e4SLinus Torvalds } 13331da177e4SLinus Torvalds 1334dd78feddSKirill A. Shutemov static void page_remove_file_rmap(struct page *page, bool compound) 13358186eb6aSJohannes Weiner { 13365d543f13SHugh Dickins int i, nr = 0; 1337dd78feddSKirill A. Shutemov 133857dea93aSSteve Capper VM_BUG_ON_PAGE(compound && !PageHead(page), page); 13398186eb6aSJohannes Weiner 134053f9263bSKirill A. Shutemov /* Hugepages are not counted in NR_FILE_MAPPED for now. */ 134153f9263bSKirill A. Shutemov if (unlikely(PageHuge(page))) { 134253f9263bSKirill A. Shutemov /* hugetlb pages are always mapped with pmds */ 134353f9263bSKirill A. Shutemov atomic_dec(compound_mapcount_ptr(page)); 1344be5d0a74SJohannes Weiner return; 134553f9263bSKirill A. Shutemov } 134653f9263bSKirill A. Shutemov 13478186eb6aSJohannes Weiner /* page still mapped by someone else? */ 1348dd78feddSKirill A. Shutemov if (compound && PageTransHuge(page)) { 1349a1528e21SMuchun Song int nr_pages = thp_nr_pages(page); 1350a1528e21SMuchun Song 13515d543f13SHugh Dickins for (i = 0; i < nr_pages; i++) { 1352dd78feddSKirill A. Shutemov if (atomic_add_negative(-1, &page[i]._mapcount)) 1353dd78feddSKirill A. Shutemov nr++; 1354dd78feddSKirill A. Shutemov } 1355dd78feddSKirill A. Shutemov if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 13565d543f13SHugh Dickins goto out; 135799cb0dbdSSong Liu if (PageSwapBacked(page)) 1358a1528e21SMuchun Song __mod_lruvec_page_state(page, NR_SHMEM_PMDMAPPED, 1359a1528e21SMuchun Song -nr_pages); 136099cb0dbdSSong Liu else 1361380780e7SMuchun Song __mod_lruvec_page_state(page, NR_FILE_PMDMAPPED, 1362380780e7SMuchun Song -nr_pages); 1363dd78feddSKirill A. Shutemov } else { 13645d543f13SHugh Dickins if (atomic_add_negative(-1, &page->_mapcount)) 13655d543f13SHugh Dickins nr++; 1366dd78feddSKirill A. Shutemov } 13675d543f13SHugh Dickins out: 13685d543f13SHugh Dickins if (nr) 136900f3ca2cSJohannes Weiner __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr); 13708186eb6aSJohannes Weiner } 13718186eb6aSJohannes Weiner 137253f9263bSKirill A. Shutemov static void page_remove_anon_compound_rmap(struct page *page) 137353f9263bSKirill A. Shutemov { 137453f9263bSKirill A. Shutemov int i, nr; 137553f9263bSKirill A. Shutemov 137653f9263bSKirill A. Shutemov if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 137753f9263bSKirill A. Shutemov return; 137853f9263bSKirill A. Shutemov 137953f9263bSKirill A. Shutemov /* Hugepages are not counted in NR_ANON_PAGES for now. */ 138053f9263bSKirill A. Shutemov if (unlikely(PageHuge(page))) 138153f9263bSKirill A. Shutemov return; 138253f9263bSKirill A. Shutemov 138353f9263bSKirill A. Shutemov if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) 138453f9263bSKirill A. Shutemov return; 138553f9263bSKirill A. Shutemov 138669473e5dSMuchun Song __mod_lruvec_page_state(page, NR_ANON_THPS, -thp_nr_pages(page)); 138753f9263bSKirill A. Shutemov 138853f9263bSKirill A. Shutemov if (TestClearPageDoubleMap(page)) { 138953f9263bSKirill A. Shutemov /* 139053f9263bSKirill A. Shutemov * Subpages can be mapped with PTEs too. Check how many of 1391f1fe80d4SKirill A. Shutemov * them are still mapped. 139253f9263bSKirill A. Shutemov */ 13935eaf35abSMatthew Wilcox (Oracle) for (i = 0, nr = 0; i < thp_nr_pages(page); i++) { 139453f9263bSKirill A. Shutemov if (atomic_add_negative(-1, &page[i]._mapcount)) 139553f9263bSKirill A. Shutemov nr++; 139653f9263bSKirill A. Shutemov } 1397f1fe80d4SKirill A. Shutemov 1398f1fe80d4SKirill A. Shutemov /* 1399f1fe80d4SKirill A. Shutemov * Queue the page for deferred split if at least one small 1400f1fe80d4SKirill A. Shutemov * page of the compound page is unmapped, but at least one 1401f1fe80d4SKirill A. Shutemov * small page is still mapped. 1402f1fe80d4SKirill A. Shutemov */ 14035eaf35abSMatthew Wilcox (Oracle) if (nr && nr < thp_nr_pages(page)) 1404f1fe80d4SKirill A. Shutemov deferred_split_huge_page(page); 140553f9263bSKirill A. Shutemov } else { 14065eaf35abSMatthew Wilcox (Oracle) nr = thp_nr_pages(page); 140753f9263bSKirill A. Shutemov } 140853f9263bSKirill A. Shutemov 1409f1fe80d4SKirill A. Shutemov if (nr) 1410be5d0a74SJohannes Weiner __mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr); 141153f9263bSKirill A. Shutemov } 141253f9263bSKirill A. Shutemov 14131da177e4SLinus Torvalds /** 14141da177e4SLinus Torvalds * page_remove_rmap - take down pte mapping from a page 14151da177e4SLinus Torvalds * @page: page to remove mapping from 1416cea86fe2SHugh Dickins * @vma: the vm area from which the mapping is removed 1417d281ee61SKirill A. Shutemov * @compound: uncharge the page as compound or small page 14181da177e4SLinus Torvalds * 1419b8072f09SHugh Dickins * The caller needs to hold the pte lock. 14201da177e4SLinus Torvalds */ 1421cea86fe2SHugh Dickins void page_remove_rmap(struct page *page, 1422cea86fe2SHugh Dickins struct vm_area_struct *vma, bool compound) 14231da177e4SLinus Torvalds { 1424be5d0a74SJohannes Weiner lock_page_memcg(page); 142589c06bd5SKAMEZAWA Hiroyuki 1426be5d0a74SJohannes Weiner if (!PageAnon(page)) { 1427be5d0a74SJohannes Weiner page_remove_file_rmap(page, compound); 1428be5d0a74SJohannes Weiner goto out; 1429be5d0a74SJohannes Weiner } 1430be5d0a74SJohannes Weiner 1431be5d0a74SJohannes Weiner if (compound) { 1432be5d0a74SJohannes Weiner page_remove_anon_compound_rmap(page); 1433be5d0a74SJohannes Weiner goto out; 1434be5d0a74SJohannes Weiner } 143553f9263bSKirill A. Shutemov 1436b904dcfeSKOSAKI Motohiro /* page still mapped by someone else? */ 1437b904dcfeSKOSAKI Motohiro if (!atomic_add_negative(-1, &page->_mapcount)) 1438be5d0a74SJohannes Weiner goto out; 14398186eb6aSJohannes Weiner 14401da177e4SLinus Torvalds /* 1441bea04b07SJianyu Zhan * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1442bea04b07SJianyu Zhan * these counters are not modified in interrupt context, and 1443bea04b07SJianyu Zhan * pte lock(a spinlock) is held, which implies preemption disabled. 14440fe6e20bSNaoya Horiguchi */ 1445be5d0a74SJohannes Weiner __dec_lruvec_page_state(page, NR_ANON_MAPPED); 14468186eb6aSJohannes Weiner 14479a982250SKirill A. Shutemov if (PageTransCompound(page)) 14489a982250SKirill A. Shutemov deferred_split_huge_page(compound_head(page)); 14499a982250SKirill A. Shutemov 145016f8c5b2SHugh Dickins /* 14511da177e4SLinus Torvalds * It would be tidy to reset the PageAnon mapping here, 14521da177e4SLinus Torvalds * but that might overwrite a racing page_add_anon_rmap 14531da177e4SLinus Torvalds * which increments mapcount after us but sets mapping 14542d4894b5SMel Gorman * before us: so leave the reset to free_unref_page, 14551da177e4SLinus Torvalds * and remember that it's only reliable while mapped. 14561da177e4SLinus Torvalds * Leaving it set also helps swapoff to reinstate ptes 14571da177e4SLinus Torvalds * faster for those pages still in swapcache. 14581da177e4SLinus Torvalds */ 1459be5d0a74SJohannes Weiner out: 1460be5d0a74SJohannes Weiner unlock_page_memcg(page); 1461cea86fe2SHugh Dickins 1462cea86fe2SHugh Dickins munlock_vma_page(page, vma, compound); 14631da177e4SLinus Torvalds } 14641da177e4SLinus Torvalds 14651da177e4SLinus Torvalds /* 146652629506SJoonsoo Kim * @arg: enum ttu_flags will be passed to this argument 14671da177e4SLinus Torvalds */ 14682f031c6fSMatthew Wilcox (Oracle) static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, 146952629506SJoonsoo Kim unsigned long address, void *arg) 14701da177e4SLinus Torvalds { 14711da177e4SLinus Torvalds struct mm_struct *mm = vma->vm_mm; 1472869f7ee6SMatthew Wilcox (Oracle) DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); 14731da177e4SLinus Torvalds pte_t pteval; 1474c7ab0d2fSKirill A. Shutemov struct page *subpage; 14756c287605SDavid Hildenbrand bool anon_exclusive, ret = true; 1476ac46d4f3SJérôme Glisse struct mmu_notifier_range range; 14774708f318SPalmer Dabbelt enum ttu_flags flags = (enum ttu_flags)(long)arg; 14781da177e4SLinus Torvalds 1479732ed558SHugh Dickins /* 1480732ed558SHugh Dickins * When racing against e.g. zap_pte_range() on another cpu, 1481732ed558SHugh Dickins * in between its ptep_get_and_clear_full() and page_remove_rmap(), 14821fb08ac6SYang Shi * try_to_unmap() may return before page_mapped() has become false, 1483732ed558SHugh Dickins * if page table locking is skipped: use TTU_SYNC to wait for that. 1484732ed558SHugh Dickins */ 1485732ed558SHugh Dickins if (flags & TTU_SYNC) 1486732ed558SHugh Dickins pvmw.flags = PVMW_SYNC; 1487732ed558SHugh Dickins 1488a98a2f0cSAlistair Popple if (flags & TTU_SPLIT_HUGE_PMD) 1489af28a988SMatthew Wilcox (Oracle) split_huge_pmd_address(vma, address, false, folio); 1490fec89c10SKirill A. Shutemov 1491369ea824SJérôme Glisse /* 1492017b1660SMike Kravetz * For THP, we have to assume the worse case ie pmd for invalidation. 1493017b1660SMike Kravetz * For hugetlb, it could be much worse if we need to do pud 1494017b1660SMike Kravetz * invalidation in the case of pmd sharing. 1495017b1660SMike Kravetz * 1496869f7ee6SMatthew Wilcox (Oracle) * Note that the folio can not be freed in this function as call of 1497869f7ee6SMatthew Wilcox (Oracle) * try_to_unmap() must hold a reference on the folio. 1498369ea824SJérôme Glisse */ 14992aff7a47SMatthew Wilcox (Oracle) range.end = vma_address_end(&pvmw); 15007269f999SJérôme Glisse mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1501494334e4SHugh Dickins address, range.end); 1502869f7ee6SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 1503017b1660SMike Kravetz /* 1504017b1660SMike Kravetz * If sharing is possible, start and end will be adjusted 1505017b1660SMike Kravetz * accordingly. 1506017b1660SMike Kravetz */ 1507ac46d4f3SJérôme Glisse adjust_range_if_pmd_sharing_possible(vma, &range.start, 1508ac46d4f3SJérôme Glisse &range.end); 1509017b1660SMike Kravetz } 1510ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_start(&range); 1511369ea824SJérôme Glisse 1512c7ab0d2fSKirill A. Shutemov while (page_vma_mapped_walk(&pvmw)) { 1513cea86fe2SHugh Dickins /* Unexpected PMD-mapped THP? */ 1514869f7ee6SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!pvmw.pte, folio); 1515cea86fe2SHugh Dickins 15161da177e4SLinus Torvalds /* 1517869f7ee6SMatthew Wilcox (Oracle) * If the folio is in an mlock()d vma, we must not swap it out. 15181da177e4SLinus Torvalds */ 1519efdb6720SHugh Dickins if (!(flags & TTU_IGNORE_MLOCK) && 1520efdb6720SHugh Dickins (vma->vm_flags & VM_LOCKED)) { 1521cea86fe2SHugh Dickins /* Restore the mlock which got missed */ 1522869f7ee6SMatthew Wilcox (Oracle) mlock_vma_folio(folio, vma, false); 1523c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1524efdb6720SHugh Dickins ret = false; 1525c7ab0d2fSKirill A. Shutemov break; 1526b87537d9SHugh Dickins } 1527c7ab0d2fSKirill A. Shutemov 1528869f7ee6SMatthew Wilcox (Oracle) subpage = folio_page(folio, 1529869f7ee6SMatthew Wilcox (Oracle) pte_pfn(*pvmw.pte) - folio_pfn(folio)); 1530785373b4SLinus Torvalds address = pvmw.address; 15316c287605SDavid Hildenbrand anon_exclusive = folio_test_anon(folio) && 15326c287605SDavid Hildenbrand PageAnonExclusive(subpage); 1533785373b4SLinus Torvalds 1534dfc7ab57SBaolin Wang if (folio_test_hugetlb(folio)) { 15350506c31dSBaolin Wang bool anon = folio_test_anon(folio); 15360506c31dSBaolin Wang 1537017b1660SMike Kravetz /* 1538a00a8759SBaolin Wang * The try_to_unmap() is only passed a hugetlb page 1539a00a8759SBaolin Wang * in the case where the hugetlb page is poisoned. 1540a00a8759SBaolin Wang */ 1541a00a8759SBaolin Wang VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage); 1542a00a8759SBaolin Wang /* 154354205e9cSBaolin Wang * huge_pmd_unshare may unmap an entire PMD page. 154454205e9cSBaolin Wang * There is no way of knowing exactly which PMDs may 154554205e9cSBaolin Wang * be cached for this mm, so we must flush them all. 154654205e9cSBaolin Wang * start/end were already adjusted above to cover this 154754205e9cSBaolin Wang * range. 1548017b1660SMike Kravetz */ 1549ac46d4f3SJérôme Glisse flush_cache_range(vma, range.start, range.end); 155054205e9cSBaolin Wang 1551dfc7ab57SBaolin Wang /* 1552dfc7ab57SBaolin Wang * To call huge_pmd_unshare, i_mmap_rwsem must be 1553dfc7ab57SBaolin Wang * held in write mode. Caller needs to explicitly 1554dfc7ab57SBaolin Wang * do this outside rmap routines. 1555dfc7ab57SBaolin Wang */ 15560506c31dSBaolin Wang VM_BUG_ON(!anon && !(flags & TTU_RMAP_LOCKED)); 15574ddb4d91SMike Kravetz if (!anon && huge_pmd_unshare(mm, vma, address, pvmw.pte)) { 1558ac46d4f3SJérôme Glisse flush_tlb_range(vma, range.start, range.end); 1559ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range(mm, range.start, 1560ac46d4f3SJérôme Glisse range.end); 1561017b1660SMike Kravetz 1562017b1660SMike Kravetz /* 1563017b1660SMike Kravetz * The ref count of the PMD page was dropped 1564017b1660SMike Kravetz * which is part of the way map counting 1565017b1660SMike Kravetz * is done for shared PMDs. Return 'true' 1566017b1660SMike Kravetz * here. When there is no other sharing, 1567017b1660SMike Kravetz * huge_pmd_unshare returns false and we will 1568017b1660SMike Kravetz * unmap the actual page and drop map count 1569017b1660SMike Kravetz * to zero. 1570017b1660SMike Kravetz */ 1571017b1660SMike Kravetz page_vma_mapped_walk_done(&pvmw); 1572017b1660SMike Kravetz break; 1573017b1660SMike Kravetz } 1574a00a8759SBaolin Wang pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); 157554205e9cSBaolin Wang } else { 157654205e9cSBaolin Wang flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); 1577088b8aa5SDavid Hildenbrand /* Nuke the page table entry. */ 1578088b8aa5SDavid Hildenbrand if (should_defer_flush(mm, flags)) { 157972b252aeSMel Gorman /* 1580c7ab0d2fSKirill A. Shutemov * We clear the PTE but do not flush so potentially 1581869f7ee6SMatthew Wilcox (Oracle) * a remote CPU could still be writing to the folio. 1582c7ab0d2fSKirill A. Shutemov * If the entry was previously clean then the 1583c7ab0d2fSKirill A. Shutemov * architecture must guarantee that a clear->dirty 1584c7ab0d2fSKirill A. Shutemov * transition on a cached TLB entry is written through 1585c7ab0d2fSKirill A. Shutemov * and traps if the PTE is unmapped. 158672b252aeSMel Gorman */ 1587785373b4SLinus Torvalds pteval = ptep_get_and_clear(mm, address, pvmw.pte); 158872b252aeSMel Gorman 1589c7ab0d2fSKirill A. Shutemov set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); 159072b252aeSMel Gorman } else { 1591785373b4SLinus Torvalds pteval = ptep_clear_flush(vma, address, pvmw.pte); 159272b252aeSMel Gorman } 1593a00a8759SBaolin Wang } 15941da177e4SLinus Torvalds 1595999dad82SPeter Xu /* 1596999dad82SPeter Xu * Now the pte is cleared. If this pte was uffd-wp armed, 1597999dad82SPeter Xu * we may want to replace a none pte with a marker pte if 1598999dad82SPeter Xu * it's file-backed, so we don't lose the tracking info. 1599999dad82SPeter Xu */ 1600999dad82SPeter Xu pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); 1601999dad82SPeter Xu 1602869f7ee6SMatthew Wilcox (Oracle) /* Set the dirty flag on the folio now the pte is gone. */ 16031da177e4SLinus Torvalds if (pte_dirty(pteval)) 1604869f7ee6SMatthew Wilcox (Oracle) folio_mark_dirty(folio); 16051da177e4SLinus Torvalds 1606365e9c87SHugh Dickins /* Update high watermark before we lower rss */ 1607365e9c87SHugh Dickins update_hiwater_rss(mm); 1608365e9c87SHugh Dickins 1609da358d5cSMatthew Wilcox (Oracle) if (PageHWPoison(subpage) && !(flags & TTU_IGNORE_HWPOISON)) { 16105fd27b8eSPunit Agrawal pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); 1611869f7ee6SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 1612869f7ee6SMatthew Wilcox (Oracle) hugetlb_count_sub(folio_nr_pages(folio), mm); 161318f39629SQi Zheng set_huge_pte_at(mm, address, pvmw.pte, pteval); 16145d317b2bSNaoya Horiguchi } else { 1615869f7ee6SMatthew Wilcox (Oracle) dec_mm_counter(mm, mm_counter(&folio->page)); 1616785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 16175f24ae58SNaoya Horiguchi } 1618c7ab0d2fSKirill A. Shutemov 1619bce73e48SChristian Borntraeger } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { 162045961722SKonstantin Weitz /* 162145961722SKonstantin Weitz * The guest indicated that the page content is of no 162245961722SKonstantin Weitz * interest anymore. Simply discard the pte, vmscan 162345961722SKonstantin Weitz * will take care of the rest. 1624bce73e48SChristian Borntraeger * A future reference will then fault in a new zero 1625bce73e48SChristian Borntraeger * page. When userfaultfd is active, we must not drop 1626bce73e48SChristian Borntraeger * this page though, as its main user (postcopy 1627bce73e48SChristian Borntraeger * migration) will not expect userfaults on already 1628bce73e48SChristian Borntraeger * copied pages. 162945961722SKonstantin Weitz */ 1630869f7ee6SMatthew Wilcox (Oracle) dec_mm_counter(mm, mm_counter(&folio->page)); 16310f10851eSJérôme Glisse /* We have to invalidate as we cleared the pte */ 16320f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, address, 16330f10851eSJérôme Glisse address + PAGE_SIZE); 1634869f7ee6SMatthew Wilcox (Oracle) } else if (folio_test_anon(folio)) { 1635c7ab0d2fSKirill A. Shutemov swp_entry_t entry = { .val = page_private(subpage) }; 1636179ef71cSCyrill Gorcunov pte_t swp_pte; 16371da177e4SLinus Torvalds /* 16381da177e4SLinus Torvalds * Store the swap location in the pte. 16391da177e4SLinus Torvalds * See handle_pte_fault() ... 16401da177e4SLinus Torvalds */ 1641869f7ee6SMatthew Wilcox (Oracle) if (unlikely(folio_test_swapbacked(folio) != 1642869f7ee6SMatthew Wilcox (Oracle) folio_test_swapcache(folio))) { 1643eb94a878SMinchan Kim WARN_ON_ONCE(1); 164483612a94SMinchan Kim ret = false; 1645369ea824SJérôme Glisse /* We have to invalidate as we cleared the pte */ 16460f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, address, 16470f10851eSJérôme Glisse address + PAGE_SIZE); 1648eb94a878SMinchan Kim page_vma_mapped_walk_done(&pvmw); 1649eb94a878SMinchan Kim break; 1650eb94a878SMinchan Kim } 1651854e9ed0SMinchan Kim 1652802a3a92SShaohua Li /* MADV_FREE page check */ 1653869f7ee6SMatthew Wilcox (Oracle) if (!folio_test_swapbacked(folio)) { 16546c8e2a25SMauricio Faria de Oliveira int ref_count, map_count; 16556c8e2a25SMauricio Faria de Oliveira 16566c8e2a25SMauricio Faria de Oliveira /* 16576c8e2a25SMauricio Faria de Oliveira * Synchronize with gup_pte_range(): 16586c8e2a25SMauricio Faria de Oliveira * - clear PTE; barrier; read refcount 16596c8e2a25SMauricio Faria de Oliveira * - inc refcount; barrier; read PTE 16606c8e2a25SMauricio Faria de Oliveira */ 16616c8e2a25SMauricio Faria de Oliveira smp_mb(); 16626c8e2a25SMauricio Faria de Oliveira 16636c8e2a25SMauricio Faria de Oliveira ref_count = folio_ref_count(folio); 16646c8e2a25SMauricio Faria de Oliveira map_count = folio_mapcount(folio); 16656c8e2a25SMauricio Faria de Oliveira 16666c8e2a25SMauricio Faria de Oliveira /* 16676c8e2a25SMauricio Faria de Oliveira * Order reads for page refcount and dirty flag 16686c8e2a25SMauricio Faria de Oliveira * (see comments in __remove_mapping()). 16696c8e2a25SMauricio Faria de Oliveira */ 16706c8e2a25SMauricio Faria de Oliveira smp_rmb(); 16716c8e2a25SMauricio Faria de Oliveira 16726c8e2a25SMauricio Faria de Oliveira /* 16736c8e2a25SMauricio Faria de Oliveira * The only page refs must be one from isolation 16746c8e2a25SMauricio Faria de Oliveira * plus the rmap(s) (dropped by discard:). 16756c8e2a25SMauricio Faria de Oliveira */ 16766c8e2a25SMauricio Faria de Oliveira if (ref_count == 1 + map_count && 16776c8e2a25SMauricio Faria de Oliveira !folio_test_dirty(folio)) { 16780f10851eSJérôme Glisse /* Invalidate as we cleared the pte */ 16790f10851eSJérôme Glisse mmu_notifier_invalidate_range(mm, 16800f10851eSJérôme Glisse address, address + PAGE_SIZE); 1681854e9ed0SMinchan Kim dec_mm_counter(mm, MM_ANONPAGES); 1682854e9ed0SMinchan Kim goto discard; 1683854e9ed0SMinchan Kim } 1684854e9ed0SMinchan Kim 1685802a3a92SShaohua Li /* 1686869f7ee6SMatthew Wilcox (Oracle) * If the folio was redirtied, it cannot be 1687802a3a92SShaohua Li * discarded. Remap the page to page table. 1688802a3a92SShaohua Li */ 1689785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 1690869f7ee6SMatthew Wilcox (Oracle) folio_set_swapbacked(folio); 1691e4b82222SMinchan Kim ret = false; 1692802a3a92SShaohua Li page_vma_mapped_walk_done(&pvmw); 1693802a3a92SShaohua Li break; 1694802a3a92SShaohua Li } 1695802a3a92SShaohua Li 1696570a335bSHugh Dickins if (swap_duplicate(entry) < 0) { 1697785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, pteval); 1698e4b82222SMinchan Kim ret = false; 1699c7ab0d2fSKirill A. Shutemov page_vma_mapped_walk_done(&pvmw); 1700c7ab0d2fSKirill A. Shutemov break; 1701570a335bSHugh Dickins } 1702ca827d55SKhalid Aziz if (arch_unmap_one(mm, vma, address, pteval) < 0) { 1703322842eaSDavid Hildenbrand swap_free(entry); 1704ca827d55SKhalid Aziz set_pte_at(mm, address, pvmw.pte, pteval); 1705ca827d55SKhalid Aziz ret = false; 1706ca827d55SKhalid Aziz page_vma_mapped_walk_done(&pvmw); 1707ca827d55SKhalid Aziz break; 1708ca827d55SKhalid Aziz } 1709088b8aa5SDavid Hildenbrand 1710088b8aa5SDavid Hildenbrand /* See page_try_share_anon_rmap(): clear PTE first. */ 17116c287605SDavid Hildenbrand if (anon_exclusive && 17126c287605SDavid Hildenbrand page_try_share_anon_rmap(subpage)) { 17136c287605SDavid Hildenbrand swap_free(entry); 17146c287605SDavid Hildenbrand set_pte_at(mm, address, pvmw.pte, pteval); 17156c287605SDavid Hildenbrand ret = false; 17166c287605SDavid Hildenbrand page_vma_mapped_walk_done(&pvmw); 17176c287605SDavid Hildenbrand break; 17186c287605SDavid Hildenbrand } 17196c287605SDavid Hildenbrand /* 17201493a191SDavid Hildenbrand * Note: We *don't* remember if the page was mapped 17211493a191SDavid Hildenbrand * exclusively in the swap pte if the architecture 17221493a191SDavid Hildenbrand * doesn't support __HAVE_ARCH_PTE_SWP_EXCLUSIVE. In 17231493a191SDavid Hildenbrand * that case, swapin code has to re-determine that 17241493a191SDavid Hildenbrand * manually and might detect the page as possibly 17251493a191SDavid Hildenbrand * shared, for example, if there are other references on 17261493a191SDavid Hildenbrand * the page or if the page is under writeback. We made 17271493a191SDavid Hildenbrand * sure that there are no GUP pins on the page that 17281493a191SDavid Hildenbrand * would rely on it, so for GUP pins this is fine. 17296c287605SDavid Hildenbrand */ 17301da177e4SLinus Torvalds if (list_empty(&mm->mmlist)) { 17311da177e4SLinus Torvalds spin_lock(&mmlist_lock); 1732f412ac08SHugh Dickins if (list_empty(&mm->mmlist)) 17331da177e4SLinus Torvalds list_add(&mm->mmlist, &init_mm.mmlist); 17341da177e4SLinus Torvalds spin_unlock(&mmlist_lock); 17351da177e4SLinus Torvalds } 1736d559db08SKAMEZAWA Hiroyuki dec_mm_counter(mm, MM_ANONPAGES); 1737b084d435SKAMEZAWA Hiroyuki inc_mm_counter(mm, MM_SWAPENTS); 1738179ef71cSCyrill Gorcunov swp_pte = swp_entry_to_pte(entry); 17391493a191SDavid Hildenbrand if (anon_exclusive) 17401493a191SDavid Hildenbrand swp_pte = pte_swp_mkexclusive(swp_pte); 1741179ef71cSCyrill Gorcunov if (pte_soft_dirty(pteval)) 1742179ef71cSCyrill Gorcunov swp_pte = pte_swp_mksoft_dirty(swp_pte); 1743f45ec5ffSPeter Xu if (pte_uffd_wp(pteval)) 1744f45ec5ffSPeter Xu swp_pte = pte_swp_mkuffd_wp(swp_pte); 1745785373b4SLinus Torvalds set_pte_at(mm, address, pvmw.pte, swp_pte); 17460f10851eSJérôme Glisse /* Invalidate as we cleared the pte */ 1747369ea824SJérôme Glisse mmu_notifier_invalidate_range(mm, address, 1748369ea824SJérôme Glisse address + PAGE_SIZE); 17490f10851eSJérôme Glisse } else { 17500f10851eSJérôme Glisse /* 1751869f7ee6SMatthew Wilcox (Oracle) * This is a locked file-backed folio, 1752869f7ee6SMatthew Wilcox (Oracle) * so it cannot be removed from the page 1753869f7ee6SMatthew Wilcox (Oracle) * cache and replaced by a new folio before 1754869f7ee6SMatthew Wilcox (Oracle) * mmu_notifier_invalidate_range_end, so no 1755869f7ee6SMatthew Wilcox (Oracle) * concurrent thread might update its page table 1756869f7ee6SMatthew Wilcox (Oracle) * to point at a new folio while a device is 1757869f7ee6SMatthew Wilcox (Oracle) * still using this folio. 17580f10851eSJérôme Glisse * 1759ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 17600f10851eSJérôme Glisse */ 1761869f7ee6SMatthew Wilcox (Oracle) dec_mm_counter(mm, mm_counter_file(&folio->page)); 17620f10851eSJérôme Glisse } 17630f10851eSJérôme Glisse discard: 17640f10851eSJérôme Glisse /* 17650f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() it has be 17660f10851eSJérôme Glisse * done above for all cases requiring it to happen under page 17670f10851eSJérôme Glisse * table lock before mmu_notifier_invalidate_range_end() 17680f10851eSJérôme Glisse * 1769ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 17700f10851eSJérôme Glisse */ 1771869f7ee6SMatthew Wilcox (Oracle) page_remove_rmap(subpage, vma, folio_test_hugetlb(folio)); 1772b7435507SHugh Dickins if (vma->vm_flags & VM_LOCKED) 1773adb11e78SSebastian Andrzej Siewior mlock_page_drain_local(); 1774869f7ee6SMatthew Wilcox (Oracle) folio_put(folio); 1775c7ab0d2fSKirill A. Shutemov } 1776369ea824SJérôme Glisse 1777ac46d4f3SJérôme Glisse mmu_notifier_invalidate_range_end(&range); 1778369ea824SJérôme Glisse 1779caed0f48SKOSAKI Motohiro return ret; 17801da177e4SLinus Torvalds } 17811da177e4SLinus Torvalds 178252629506SJoonsoo Kim static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) 178352629506SJoonsoo Kim { 1784222100eeSAnshuman Khandual return vma_is_temporary_stack(vma); 178552629506SJoonsoo Kim } 178652629506SJoonsoo Kim 17872f031c6fSMatthew Wilcox (Oracle) static int page_not_mapped(struct folio *folio) 178852629506SJoonsoo Kim { 17892f031c6fSMatthew Wilcox (Oracle) return !folio_mapped(folio); 17902a52bcbcSKirill A. Shutemov } 179152629506SJoonsoo Kim 17921da177e4SLinus Torvalds /** 1793869f7ee6SMatthew Wilcox (Oracle) * try_to_unmap - Try to remove all page table mappings to a folio. 1794869f7ee6SMatthew Wilcox (Oracle) * @folio: The folio to unmap. 179514fa31b8SAndi Kleen * @flags: action and flags 17961da177e4SLinus Torvalds * 17971da177e4SLinus Torvalds * Tries to remove all the page table entries which are mapping this 1798869f7ee6SMatthew Wilcox (Oracle) * folio. It is the caller's responsibility to check if the folio is 1799869f7ee6SMatthew Wilcox (Oracle) * still mapped if needed (use TTU_SYNC to prevent accounting races). 18001da177e4SLinus Torvalds * 1801869f7ee6SMatthew Wilcox (Oracle) * Context: Caller must hold the folio lock. 18021da177e4SLinus Torvalds */ 1803869f7ee6SMatthew Wilcox (Oracle) void try_to_unmap(struct folio *folio, enum ttu_flags flags) 18041da177e4SLinus Torvalds { 180552629506SJoonsoo Kim struct rmap_walk_control rwc = { 180652629506SJoonsoo Kim .rmap_one = try_to_unmap_one, 1807802a3a92SShaohua Li .arg = (void *)flags, 1808b7e188ecSMiaohe Lin .done = page_not_mapped, 18092f031c6fSMatthew Wilcox (Oracle) .anon_lock = folio_lock_anon_vma_read, 181052629506SJoonsoo Kim }; 18111da177e4SLinus Torvalds 1812a98a2f0cSAlistair Popple if (flags & TTU_RMAP_LOCKED) 18132f031c6fSMatthew Wilcox (Oracle) rmap_walk_locked(folio, &rwc); 1814a98a2f0cSAlistair Popple else 18152f031c6fSMatthew Wilcox (Oracle) rmap_walk(folio, &rwc); 1816a98a2f0cSAlistair Popple } 1817a98a2f0cSAlistair Popple 1818a98a2f0cSAlistair Popple /* 1819a98a2f0cSAlistair Popple * @arg: enum ttu_flags will be passed to this argument. 1820a98a2f0cSAlistair Popple * 1821a98a2f0cSAlistair Popple * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs 182264b586d1SHugh Dickins * containing migration entries. 1823a98a2f0cSAlistair Popple */ 18242f031c6fSMatthew Wilcox (Oracle) static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma, 1825a98a2f0cSAlistair Popple unsigned long address, void *arg) 1826a98a2f0cSAlistair Popple { 1827a98a2f0cSAlistair Popple struct mm_struct *mm = vma->vm_mm; 18284b8554c5SMatthew Wilcox (Oracle) DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); 1829a98a2f0cSAlistair Popple pte_t pteval; 1830a98a2f0cSAlistair Popple struct page *subpage; 18316c287605SDavid Hildenbrand bool anon_exclusive, ret = true; 1832a98a2f0cSAlistair Popple struct mmu_notifier_range range; 1833a98a2f0cSAlistair Popple enum ttu_flags flags = (enum ttu_flags)(long)arg; 1834a98a2f0cSAlistair Popple 1835a98a2f0cSAlistair Popple /* 1836a98a2f0cSAlistair Popple * When racing against e.g. zap_pte_range() on another cpu, 1837a98a2f0cSAlistair Popple * in between its ptep_get_and_clear_full() and page_remove_rmap(), 1838a98a2f0cSAlistair Popple * try_to_migrate() may return before page_mapped() has become false, 1839a98a2f0cSAlistair Popple * if page table locking is skipped: use TTU_SYNC to wait for that. 1840a98a2f0cSAlistair Popple */ 1841a98a2f0cSAlistair Popple if (flags & TTU_SYNC) 1842a98a2f0cSAlistair Popple pvmw.flags = PVMW_SYNC; 1843a98a2f0cSAlistair Popple 1844a98a2f0cSAlistair Popple /* 1845a98a2f0cSAlistair Popple * unmap_page() in mm/huge_memory.c is the only user of migration with 1846a98a2f0cSAlistair Popple * TTU_SPLIT_HUGE_PMD and it wants to freeze. 1847a98a2f0cSAlistair Popple */ 1848a98a2f0cSAlistair Popple if (flags & TTU_SPLIT_HUGE_PMD) 1849af28a988SMatthew Wilcox (Oracle) split_huge_pmd_address(vma, address, true, folio); 1850a98a2f0cSAlistair Popple 1851a98a2f0cSAlistair Popple /* 1852a98a2f0cSAlistair Popple * For THP, we have to assume the worse case ie pmd for invalidation. 1853a98a2f0cSAlistair Popple * For hugetlb, it could be much worse if we need to do pud 1854a98a2f0cSAlistair Popple * invalidation in the case of pmd sharing. 1855a98a2f0cSAlistair Popple * 1856a98a2f0cSAlistair Popple * Note that the page can not be free in this function as call of 1857a98a2f0cSAlistair Popple * try_to_unmap() must hold a reference on the page. 1858a98a2f0cSAlistair Popple */ 18592aff7a47SMatthew Wilcox (Oracle) range.end = vma_address_end(&pvmw); 1860a98a2f0cSAlistair Popple mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, 1861a98a2f0cSAlistair Popple address, range.end); 18624b8554c5SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 1863a98a2f0cSAlistair Popple /* 1864a98a2f0cSAlistair Popple * If sharing is possible, start and end will be adjusted 1865a98a2f0cSAlistair Popple * accordingly. 1866a98a2f0cSAlistair Popple */ 1867a98a2f0cSAlistair Popple adjust_range_if_pmd_sharing_possible(vma, &range.start, 1868a98a2f0cSAlistair Popple &range.end); 1869a98a2f0cSAlistair Popple } 1870a98a2f0cSAlistair Popple mmu_notifier_invalidate_range_start(&range); 1871a98a2f0cSAlistair Popple 1872a98a2f0cSAlistair Popple while (page_vma_mapped_walk(&pvmw)) { 1873a98a2f0cSAlistair Popple #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1874a98a2f0cSAlistair Popple /* PMD-mapped THP migration entry */ 1875a98a2f0cSAlistair Popple if (!pvmw.pte) { 18764b8554c5SMatthew Wilcox (Oracle) subpage = folio_page(folio, 18774b8554c5SMatthew Wilcox (Oracle) pmd_pfn(*pvmw.pmd) - folio_pfn(folio)); 18784b8554c5SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || 18794b8554c5SMatthew Wilcox (Oracle) !folio_test_pmd_mappable(folio), folio); 1880a98a2f0cSAlistair Popple 18817f5abe60SDavid Hildenbrand if (set_pmd_migration_entry(&pvmw, subpage)) { 18827f5abe60SDavid Hildenbrand ret = false; 18837f5abe60SDavid Hildenbrand page_vma_mapped_walk_done(&pvmw); 18847f5abe60SDavid Hildenbrand break; 18857f5abe60SDavid Hildenbrand } 1886a98a2f0cSAlistair Popple continue; 1887a98a2f0cSAlistair Popple } 1888a98a2f0cSAlistair Popple #endif 1889a98a2f0cSAlistair Popple 1890a98a2f0cSAlistair Popple /* Unexpected PMD-mapped THP? */ 18914b8554c5SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!pvmw.pte, folio); 1892a98a2f0cSAlistair Popple 18931118234eSDavid Hildenbrand if (folio_is_zone_device(folio)) { 18941118234eSDavid Hildenbrand /* 18951118234eSDavid Hildenbrand * Our PTE is a non-present device exclusive entry and 18961118234eSDavid Hildenbrand * calculating the subpage as for the common case would 18971118234eSDavid Hildenbrand * result in an invalid pointer. 18981118234eSDavid Hildenbrand * 18991118234eSDavid Hildenbrand * Since only PAGE_SIZE pages can currently be 19001118234eSDavid Hildenbrand * migrated, just set it to page. This will need to be 19011118234eSDavid Hildenbrand * changed when hugepage migrations to device private 19021118234eSDavid Hildenbrand * memory are supported. 19031118234eSDavid Hildenbrand */ 19041118234eSDavid Hildenbrand VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio); 19051118234eSDavid Hildenbrand subpage = &folio->page; 19061118234eSDavid Hildenbrand } else { 19074b8554c5SMatthew Wilcox (Oracle) subpage = folio_page(folio, 19084b8554c5SMatthew Wilcox (Oracle) pte_pfn(*pvmw.pte) - folio_pfn(folio)); 19091118234eSDavid Hildenbrand } 1910a98a2f0cSAlistair Popple address = pvmw.address; 19116c287605SDavid Hildenbrand anon_exclusive = folio_test_anon(folio) && 19126c287605SDavid Hildenbrand PageAnonExclusive(subpage); 1913a98a2f0cSAlistair Popple 1914dfc7ab57SBaolin Wang if (folio_test_hugetlb(folio)) { 19150506c31dSBaolin Wang bool anon = folio_test_anon(folio); 19160506c31dSBaolin Wang 1917a98a2f0cSAlistair Popple /* 191854205e9cSBaolin Wang * huge_pmd_unshare may unmap an entire PMD page. 191954205e9cSBaolin Wang * There is no way of knowing exactly which PMDs may 192054205e9cSBaolin Wang * be cached for this mm, so we must flush them all. 192154205e9cSBaolin Wang * start/end were already adjusted above to cover this 192254205e9cSBaolin Wang * range. 1923a98a2f0cSAlistair Popple */ 1924a98a2f0cSAlistair Popple flush_cache_range(vma, range.start, range.end); 192554205e9cSBaolin Wang 1926dfc7ab57SBaolin Wang /* 1927dfc7ab57SBaolin Wang * To call huge_pmd_unshare, i_mmap_rwsem must be 1928dfc7ab57SBaolin Wang * held in write mode. Caller needs to explicitly 1929dfc7ab57SBaolin Wang * do this outside rmap routines. 1930dfc7ab57SBaolin Wang */ 19310506c31dSBaolin Wang VM_BUG_ON(!anon && !(flags & TTU_RMAP_LOCKED)); 19324ddb4d91SMike Kravetz if (!anon && huge_pmd_unshare(mm, vma, address, pvmw.pte)) { 1933a98a2f0cSAlistair Popple flush_tlb_range(vma, range.start, range.end); 1934a98a2f0cSAlistair Popple mmu_notifier_invalidate_range(mm, range.start, 1935a98a2f0cSAlistair Popple range.end); 1936a98a2f0cSAlistair Popple 1937a98a2f0cSAlistair Popple /* 1938a98a2f0cSAlistair Popple * The ref count of the PMD page was dropped 1939a98a2f0cSAlistair Popple * which is part of the way map counting 1940a98a2f0cSAlistair Popple * is done for shared PMDs. Return 'true' 1941a98a2f0cSAlistair Popple * here. When there is no other sharing, 1942a98a2f0cSAlistair Popple * huge_pmd_unshare returns false and we will 1943a98a2f0cSAlistair Popple * unmap the actual page and drop map count 1944a98a2f0cSAlistair Popple * to zero. 1945a98a2f0cSAlistair Popple */ 1946a98a2f0cSAlistair Popple page_vma_mapped_walk_done(&pvmw); 1947a98a2f0cSAlistair Popple break; 1948a98a2f0cSAlistair Popple } 19495d4af619SBaolin Wang 19505d4af619SBaolin Wang /* Nuke the hugetlb page table entry */ 19515d4af619SBaolin Wang pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); 195254205e9cSBaolin Wang } else { 195354205e9cSBaolin Wang flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); 1954a98a2f0cSAlistair Popple /* Nuke the page table entry. */ 1955a98a2f0cSAlistair Popple pteval = ptep_clear_flush(vma, address, pvmw.pte); 19565d4af619SBaolin Wang } 1957a98a2f0cSAlistair Popple 19584b8554c5SMatthew Wilcox (Oracle) /* Set the dirty flag on the folio now the pte is gone. */ 1959a98a2f0cSAlistair Popple if (pte_dirty(pteval)) 19604b8554c5SMatthew Wilcox (Oracle) folio_mark_dirty(folio); 1961a98a2f0cSAlistair Popple 1962a98a2f0cSAlistair Popple /* Update high watermark before we lower rss */ 1963a98a2f0cSAlistair Popple update_hiwater_rss(mm); 1964a98a2f0cSAlistair Popple 1965f25cbb7aSAlex Sierra if (folio_is_device_private(folio)) { 19664b8554c5SMatthew Wilcox (Oracle) unsigned long pfn = folio_pfn(folio); 1967a98a2f0cSAlistair Popple swp_entry_t entry; 1968a98a2f0cSAlistair Popple pte_t swp_pte; 1969a98a2f0cSAlistair Popple 19706c287605SDavid Hildenbrand if (anon_exclusive) 19716c287605SDavid Hildenbrand BUG_ON(page_try_share_anon_rmap(subpage)); 19726c287605SDavid Hildenbrand 1973a98a2f0cSAlistair Popple /* 1974a98a2f0cSAlistair Popple * Store the pfn of the page in a special migration 1975a98a2f0cSAlistair Popple * pte. do_swap_page() will wait until the migration 1976a98a2f0cSAlistair Popple * pte is removed and then restart fault handling. 1977a98a2f0cSAlistair Popple */ 19783d88705cSAlistair Popple entry = pte_to_swp_entry(pteval); 19793d88705cSAlistair Popple if (is_writable_device_private_entry(entry)) 19803d88705cSAlistair Popple entry = make_writable_migration_entry(pfn); 19816c287605SDavid Hildenbrand else if (anon_exclusive) 19826c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry(pfn); 19833d88705cSAlistair Popple else 19843d88705cSAlistair Popple entry = make_readable_migration_entry(pfn); 1985a98a2f0cSAlistair Popple swp_pte = swp_entry_to_pte(entry); 1986a98a2f0cSAlistair Popple 1987a98a2f0cSAlistair Popple /* 1988a98a2f0cSAlistair Popple * pteval maps a zone device page and is therefore 1989a98a2f0cSAlistair Popple * a swap pte. 1990a98a2f0cSAlistair Popple */ 1991a98a2f0cSAlistair Popple if (pte_swp_soft_dirty(pteval)) 1992a98a2f0cSAlistair Popple swp_pte = pte_swp_mksoft_dirty(swp_pte); 1993a98a2f0cSAlistair Popple if (pte_swp_uffd_wp(pteval)) 1994a98a2f0cSAlistair Popple swp_pte = pte_swp_mkuffd_wp(swp_pte); 1995a98a2f0cSAlistair Popple set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte); 19964cc79b33SAnshuman Khandual trace_set_migration_pte(pvmw.address, pte_val(swp_pte), 19974cc79b33SAnshuman Khandual compound_order(&folio->page)); 1998a98a2f0cSAlistair Popple /* 1999a98a2f0cSAlistair Popple * No need to invalidate here it will synchronize on 2000a98a2f0cSAlistair Popple * against the special swap migration pte. 2001a98a2f0cSAlistair Popple */ 2002da358d5cSMatthew Wilcox (Oracle) } else if (PageHWPoison(subpage)) { 2003a98a2f0cSAlistair Popple pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); 20044b8554c5SMatthew Wilcox (Oracle) if (folio_test_hugetlb(folio)) { 20054b8554c5SMatthew Wilcox (Oracle) hugetlb_count_sub(folio_nr_pages(folio), mm); 200618f39629SQi Zheng set_huge_pte_at(mm, address, pvmw.pte, pteval); 2007a98a2f0cSAlistair Popple } else { 20084b8554c5SMatthew Wilcox (Oracle) dec_mm_counter(mm, mm_counter(&folio->page)); 2009a98a2f0cSAlistair Popple set_pte_at(mm, address, pvmw.pte, pteval); 2010a98a2f0cSAlistair Popple } 2011a98a2f0cSAlistair Popple 2012a98a2f0cSAlistair Popple } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { 2013a98a2f0cSAlistair Popple /* 2014a98a2f0cSAlistair Popple * The guest indicated that the page content is of no 2015a98a2f0cSAlistair Popple * interest anymore. Simply discard the pte, vmscan 2016a98a2f0cSAlistair Popple * will take care of the rest. 2017a98a2f0cSAlistair Popple * A future reference will then fault in a new zero 2018a98a2f0cSAlistair Popple * page. When userfaultfd is active, we must not drop 2019a98a2f0cSAlistair Popple * this page though, as its main user (postcopy 2020a98a2f0cSAlistair Popple * migration) will not expect userfaults on already 2021a98a2f0cSAlistair Popple * copied pages. 2022a98a2f0cSAlistair Popple */ 20234b8554c5SMatthew Wilcox (Oracle) dec_mm_counter(mm, mm_counter(&folio->page)); 2024a98a2f0cSAlistair Popple /* We have to invalidate as we cleared the pte */ 2025a98a2f0cSAlistair Popple mmu_notifier_invalidate_range(mm, address, 2026a98a2f0cSAlistair Popple address + PAGE_SIZE); 2027a98a2f0cSAlistair Popple } else { 2028a98a2f0cSAlistair Popple swp_entry_t entry; 2029a98a2f0cSAlistair Popple pte_t swp_pte; 2030a98a2f0cSAlistair Popple 2031a98a2f0cSAlistair Popple if (arch_unmap_one(mm, vma, address, pteval) < 0) { 20325d4af619SBaolin Wang if (folio_test_hugetlb(folio)) 20335d4af619SBaolin Wang set_huge_pte_at(mm, address, pvmw.pte, pteval); 20345d4af619SBaolin Wang else 2035a98a2f0cSAlistair Popple set_pte_at(mm, address, pvmw.pte, pteval); 2036a98a2f0cSAlistair Popple ret = false; 2037a98a2f0cSAlistair Popple page_vma_mapped_walk_done(&pvmw); 2038a98a2f0cSAlistair Popple break; 2039a98a2f0cSAlistair Popple } 20406c287605SDavid Hildenbrand VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) && 20416c287605SDavid Hildenbrand !anon_exclusive, subpage); 2042088b8aa5SDavid Hildenbrand 2043088b8aa5SDavid Hildenbrand /* See page_try_share_anon_rmap(): clear PTE first. */ 20446c287605SDavid Hildenbrand if (anon_exclusive && 20456c287605SDavid Hildenbrand page_try_share_anon_rmap(subpage)) { 20465d4af619SBaolin Wang if (folio_test_hugetlb(folio)) 20475d4af619SBaolin Wang set_huge_pte_at(mm, address, pvmw.pte, pteval); 20485d4af619SBaolin Wang else 20496c287605SDavid Hildenbrand set_pte_at(mm, address, pvmw.pte, pteval); 20506c287605SDavid Hildenbrand ret = false; 20516c287605SDavid Hildenbrand page_vma_mapped_walk_done(&pvmw); 20526c287605SDavid Hildenbrand break; 20536c287605SDavid Hildenbrand } 2054a98a2f0cSAlistair Popple 2055a98a2f0cSAlistair Popple /* 2056a98a2f0cSAlistair Popple * Store the pfn of the page in a special migration 2057a98a2f0cSAlistair Popple * pte. do_swap_page() will wait until the migration 2058a98a2f0cSAlistair Popple * pte is removed and then restart fault handling. 2059a98a2f0cSAlistair Popple */ 2060a98a2f0cSAlistair Popple if (pte_write(pteval)) 2061a98a2f0cSAlistair Popple entry = make_writable_migration_entry( 2062a98a2f0cSAlistair Popple page_to_pfn(subpage)); 20636c287605SDavid Hildenbrand else if (anon_exclusive) 20646c287605SDavid Hildenbrand entry = make_readable_exclusive_migration_entry( 20656c287605SDavid Hildenbrand page_to_pfn(subpage)); 2066a98a2f0cSAlistair Popple else 2067a98a2f0cSAlistair Popple entry = make_readable_migration_entry( 2068a98a2f0cSAlistair Popple page_to_pfn(subpage)); 2069*2e346877SPeter Xu if (pte_young(pteval)) 2070*2e346877SPeter Xu entry = make_migration_entry_young(entry); 2071*2e346877SPeter Xu if (pte_dirty(pteval)) 2072*2e346877SPeter Xu entry = make_migration_entry_dirty(entry); 2073a98a2f0cSAlistair Popple swp_pte = swp_entry_to_pte(entry); 2074a98a2f0cSAlistair Popple if (pte_soft_dirty(pteval)) 2075a98a2f0cSAlistair Popple swp_pte = pte_swp_mksoft_dirty(swp_pte); 2076a98a2f0cSAlistair Popple if (pte_uffd_wp(pteval)) 2077a98a2f0cSAlistair Popple swp_pte = pte_swp_mkuffd_wp(swp_pte); 20785d4af619SBaolin Wang if (folio_test_hugetlb(folio)) 207918f39629SQi Zheng set_huge_pte_at(mm, address, pvmw.pte, swp_pte); 20805d4af619SBaolin Wang else 2081a98a2f0cSAlistair Popple set_pte_at(mm, address, pvmw.pte, swp_pte); 20824cc79b33SAnshuman Khandual trace_set_migration_pte(address, pte_val(swp_pte), 20834cc79b33SAnshuman Khandual compound_order(&folio->page)); 2084a98a2f0cSAlistair Popple /* 2085a98a2f0cSAlistair Popple * No need to invalidate here it will synchronize on 2086a98a2f0cSAlistair Popple * against the special swap migration pte. 2087a98a2f0cSAlistair Popple */ 2088a98a2f0cSAlistair Popple } 2089a98a2f0cSAlistair Popple 2090a98a2f0cSAlistair Popple /* 2091a98a2f0cSAlistair Popple * No need to call mmu_notifier_invalidate_range() it has be 2092a98a2f0cSAlistair Popple * done above for all cases requiring it to happen under page 2093a98a2f0cSAlistair Popple * table lock before mmu_notifier_invalidate_range_end() 2094a98a2f0cSAlistair Popple * 2095ee65728eSMike Rapoport * See Documentation/mm/mmu_notifier.rst 2096a98a2f0cSAlistair Popple */ 20974b8554c5SMatthew Wilcox (Oracle) page_remove_rmap(subpage, vma, folio_test_hugetlb(folio)); 2098b7435507SHugh Dickins if (vma->vm_flags & VM_LOCKED) 2099adb11e78SSebastian Andrzej Siewior mlock_page_drain_local(); 21004b8554c5SMatthew Wilcox (Oracle) folio_put(folio); 2101a98a2f0cSAlistair Popple } 2102a98a2f0cSAlistair Popple 2103a98a2f0cSAlistair Popple mmu_notifier_invalidate_range_end(&range); 2104a98a2f0cSAlistair Popple 2105a98a2f0cSAlistair Popple return ret; 2106a98a2f0cSAlistair Popple } 2107a98a2f0cSAlistair Popple 2108a98a2f0cSAlistair Popple /** 2109a98a2f0cSAlistair Popple * try_to_migrate - try to replace all page table mappings with swap entries 21104b8554c5SMatthew Wilcox (Oracle) * @folio: the folio to replace page table entries for 2111a98a2f0cSAlistair Popple * @flags: action and flags 2112a98a2f0cSAlistair Popple * 21134b8554c5SMatthew Wilcox (Oracle) * Tries to remove all the page table entries which are mapping this folio and 21144b8554c5SMatthew Wilcox (Oracle) * replace them with special swap entries. Caller must hold the folio lock. 2115a98a2f0cSAlistair Popple */ 21164b8554c5SMatthew Wilcox (Oracle) void try_to_migrate(struct folio *folio, enum ttu_flags flags) 2117a98a2f0cSAlistair Popple { 2118a98a2f0cSAlistair Popple struct rmap_walk_control rwc = { 2119a98a2f0cSAlistair Popple .rmap_one = try_to_migrate_one, 2120a98a2f0cSAlistair Popple .arg = (void *)flags, 2121a98a2f0cSAlistair Popple .done = page_not_mapped, 21222f031c6fSMatthew Wilcox (Oracle) .anon_lock = folio_lock_anon_vma_read, 2123a98a2f0cSAlistair Popple }; 2124a98a2f0cSAlistair Popple 2125a98a2f0cSAlistair Popple /* 2126a98a2f0cSAlistair Popple * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and 2127a98a2f0cSAlistair Popple * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags. 2128a98a2f0cSAlistair Popple */ 2129a98a2f0cSAlistair Popple if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | 2130a98a2f0cSAlistair Popple TTU_SYNC))) 2131a98a2f0cSAlistair Popple return; 2132a98a2f0cSAlistair Popple 2133f25cbb7aSAlex Sierra if (folio_is_zone_device(folio) && 2134f25cbb7aSAlex Sierra (!folio_is_device_private(folio) && !folio_is_device_coherent(folio))) 21356c855fceSHugh Dickins return; 21366c855fceSHugh Dickins 213752629506SJoonsoo Kim /* 213852629506SJoonsoo Kim * During exec, a temporary VMA is setup and later moved. 213952629506SJoonsoo Kim * The VMA is moved under the anon_vma lock but not the 214052629506SJoonsoo Kim * page tables leading to a race where migration cannot 214152629506SJoonsoo Kim * find the migration ptes. Rather than increasing the 214252629506SJoonsoo Kim * locking requirements of exec(), migration skips 214352629506SJoonsoo Kim * temporary VMAs until after exec() completes. 214452629506SJoonsoo Kim */ 21454b8554c5SMatthew Wilcox (Oracle) if (!folio_test_ksm(folio) && folio_test_anon(folio)) 214652629506SJoonsoo Kim rwc.invalid_vma = invalid_migration_vma; 214752629506SJoonsoo Kim 21482a52bcbcSKirill A. Shutemov if (flags & TTU_RMAP_LOCKED) 21492f031c6fSMatthew Wilcox (Oracle) rmap_walk_locked(folio, &rwc); 21502a52bcbcSKirill A. Shutemov else 21512f031c6fSMatthew Wilcox (Oracle) rmap_walk(folio, &rwc); 2152b291f000SNick Piggin } 2153e9995ef9SHugh Dickins 2154b756a3b5SAlistair Popple #ifdef CONFIG_DEVICE_PRIVATE 2155b756a3b5SAlistair Popple struct make_exclusive_args { 2156b756a3b5SAlistair Popple struct mm_struct *mm; 2157b756a3b5SAlistair Popple unsigned long address; 2158b756a3b5SAlistair Popple void *owner; 2159b756a3b5SAlistair Popple bool valid; 2160b756a3b5SAlistair Popple }; 2161b756a3b5SAlistair Popple 21622f031c6fSMatthew Wilcox (Oracle) static bool page_make_device_exclusive_one(struct folio *folio, 2163b756a3b5SAlistair Popple struct vm_area_struct *vma, unsigned long address, void *priv) 2164b756a3b5SAlistair Popple { 2165b756a3b5SAlistair Popple struct mm_struct *mm = vma->vm_mm; 21660d251485SMatthew Wilcox (Oracle) DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); 2167b756a3b5SAlistair Popple struct make_exclusive_args *args = priv; 2168b756a3b5SAlistair Popple pte_t pteval; 2169b756a3b5SAlistair Popple struct page *subpage; 2170b756a3b5SAlistair Popple bool ret = true; 2171b756a3b5SAlistair Popple struct mmu_notifier_range range; 2172b756a3b5SAlistair Popple swp_entry_t entry; 2173b756a3b5SAlistair Popple pte_t swp_pte; 2174b756a3b5SAlistair Popple 2175b756a3b5SAlistair Popple mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma, 2176b756a3b5SAlistair Popple vma->vm_mm, address, min(vma->vm_end, 21770d251485SMatthew Wilcox (Oracle) address + folio_size(folio)), 21780d251485SMatthew Wilcox (Oracle) args->owner); 2179b756a3b5SAlistair Popple mmu_notifier_invalidate_range_start(&range); 2180b756a3b5SAlistair Popple 2181b756a3b5SAlistair Popple while (page_vma_mapped_walk(&pvmw)) { 2182b756a3b5SAlistair Popple /* Unexpected PMD-mapped THP? */ 21830d251485SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!pvmw.pte, folio); 2184b756a3b5SAlistair Popple 2185b756a3b5SAlistair Popple if (!pte_present(*pvmw.pte)) { 2186b756a3b5SAlistair Popple ret = false; 2187b756a3b5SAlistair Popple page_vma_mapped_walk_done(&pvmw); 2188b756a3b5SAlistair Popple break; 2189b756a3b5SAlistair Popple } 2190b756a3b5SAlistair Popple 21910d251485SMatthew Wilcox (Oracle) subpage = folio_page(folio, 21920d251485SMatthew Wilcox (Oracle) pte_pfn(*pvmw.pte) - folio_pfn(folio)); 2193b756a3b5SAlistair Popple address = pvmw.address; 2194b756a3b5SAlistair Popple 2195b756a3b5SAlistair Popple /* Nuke the page table entry. */ 2196b756a3b5SAlistair Popple flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); 2197b756a3b5SAlistair Popple pteval = ptep_clear_flush(vma, address, pvmw.pte); 2198b756a3b5SAlistair Popple 21990d251485SMatthew Wilcox (Oracle) /* Set the dirty flag on the folio now the pte is gone. */ 2200b756a3b5SAlistair Popple if (pte_dirty(pteval)) 22010d251485SMatthew Wilcox (Oracle) folio_mark_dirty(folio); 2202b756a3b5SAlistair Popple 2203b756a3b5SAlistair Popple /* 2204b756a3b5SAlistair Popple * Check that our target page is still mapped at the expected 2205b756a3b5SAlistair Popple * address. 2206b756a3b5SAlistair Popple */ 2207b756a3b5SAlistair Popple if (args->mm == mm && args->address == address && 2208b756a3b5SAlistair Popple pte_write(pteval)) 2209b756a3b5SAlistair Popple args->valid = true; 2210b756a3b5SAlistair Popple 2211b756a3b5SAlistair Popple /* 2212b756a3b5SAlistair Popple * Store the pfn of the page in a special migration 2213b756a3b5SAlistair Popple * pte. do_swap_page() will wait until the migration 2214b756a3b5SAlistair Popple * pte is removed and then restart fault handling. 2215b756a3b5SAlistair Popple */ 2216b756a3b5SAlistair Popple if (pte_write(pteval)) 2217b756a3b5SAlistair Popple entry = make_writable_device_exclusive_entry( 2218b756a3b5SAlistair Popple page_to_pfn(subpage)); 2219b756a3b5SAlistair Popple else 2220b756a3b5SAlistair Popple entry = make_readable_device_exclusive_entry( 2221b756a3b5SAlistair Popple page_to_pfn(subpage)); 2222b756a3b5SAlistair Popple swp_pte = swp_entry_to_pte(entry); 2223b756a3b5SAlistair Popple if (pte_soft_dirty(pteval)) 2224b756a3b5SAlistair Popple swp_pte = pte_swp_mksoft_dirty(swp_pte); 2225b756a3b5SAlistair Popple if (pte_uffd_wp(pteval)) 2226b756a3b5SAlistair Popple swp_pte = pte_swp_mkuffd_wp(swp_pte); 2227b756a3b5SAlistair Popple 2228b756a3b5SAlistair Popple set_pte_at(mm, address, pvmw.pte, swp_pte); 2229b756a3b5SAlistair Popple 2230b756a3b5SAlistair Popple /* 2231b756a3b5SAlistair Popple * There is a reference on the page for the swap entry which has 2232b756a3b5SAlistair Popple * been removed, so shouldn't take another. 2233b756a3b5SAlistair Popple */ 2234cea86fe2SHugh Dickins page_remove_rmap(subpage, vma, false); 2235b756a3b5SAlistair Popple } 2236b756a3b5SAlistair Popple 2237b756a3b5SAlistair Popple mmu_notifier_invalidate_range_end(&range); 2238b756a3b5SAlistair Popple 2239b756a3b5SAlistair Popple return ret; 2240b756a3b5SAlistair Popple } 2241b756a3b5SAlistair Popple 2242b756a3b5SAlistair Popple /** 22430d251485SMatthew Wilcox (Oracle) * folio_make_device_exclusive - Mark the folio exclusively owned by a device. 22440d251485SMatthew Wilcox (Oracle) * @folio: The folio to replace page table entries for. 22450d251485SMatthew Wilcox (Oracle) * @mm: The mm_struct where the folio is expected to be mapped. 22460d251485SMatthew Wilcox (Oracle) * @address: Address where the folio is expected to be mapped. 2247b756a3b5SAlistair Popple * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks 2248b756a3b5SAlistair Popple * 22490d251485SMatthew Wilcox (Oracle) * Tries to remove all the page table entries which are mapping this 22500d251485SMatthew Wilcox (Oracle) * folio and replace them with special device exclusive swap entries to 22510d251485SMatthew Wilcox (Oracle) * grant a device exclusive access to the folio. 2252b756a3b5SAlistair Popple * 22530d251485SMatthew Wilcox (Oracle) * Context: Caller must hold the folio lock. 22540d251485SMatthew Wilcox (Oracle) * Return: false if the page is still mapped, or if it could not be unmapped 2255b756a3b5SAlistair Popple * from the expected address. Otherwise returns true (success). 2256b756a3b5SAlistair Popple */ 22570d251485SMatthew Wilcox (Oracle) static bool folio_make_device_exclusive(struct folio *folio, 22580d251485SMatthew Wilcox (Oracle) struct mm_struct *mm, unsigned long address, void *owner) 2259b756a3b5SAlistair Popple { 2260b756a3b5SAlistair Popple struct make_exclusive_args args = { 2261b756a3b5SAlistair Popple .mm = mm, 2262b756a3b5SAlistair Popple .address = address, 2263b756a3b5SAlistair Popple .owner = owner, 2264b756a3b5SAlistair Popple .valid = false, 2265b756a3b5SAlistair Popple }; 2266b756a3b5SAlistair Popple struct rmap_walk_control rwc = { 2267b756a3b5SAlistair Popple .rmap_one = page_make_device_exclusive_one, 2268b756a3b5SAlistair Popple .done = page_not_mapped, 22692f031c6fSMatthew Wilcox (Oracle) .anon_lock = folio_lock_anon_vma_read, 2270b756a3b5SAlistair Popple .arg = &args, 2271b756a3b5SAlistair Popple }; 2272b756a3b5SAlistair Popple 2273b756a3b5SAlistair Popple /* 22740d251485SMatthew Wilcox (Oracle) * Restrict to anonymous folios for now to avoid potential writeback 22750d251485SMatthew Wilcox (Oracle) * issues. 2276b756a3b5SAlistair Popple */ 22770d251485SMatthew Wilcox (Oracle) if (!folio_test_anon(folio)) 2278b756a3b5SAlistair Popple return false; 2279b756a3b5SAlistair Popple 22802f031c6fSMatthew Wilcox (Oracle) rmap_walk(folio, &rwc); 2281b756a3b5SAlistair Popple 22820d251485SMatthew Wilcox (Oracle) return args.valid && !folio_mapcount(folio); 2283b756a3b5SAlistair Popple } 2284b756a3b5SAlistair Popple 2285b756a3b5SAlistair Popple /** 2286b756a3b5SAlistair Popple * make_device_exclusive_range() - Mark a range for exclusive use by a device 2287dd062302SAdrian Huang * @mm: mm_struct of associated target process 2288b756a3b5SAlistair Popple * @start: start of the region to mark for exclusive device access 2289b756a3b5SAlistair Popple * @end: end address of region 2290b756a3b5SAlistair Popple * @pages: returns the pages which were successfully marked for exclusive access 2291b756a3b5SAlistair Popple * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering 2292b756a3b5SAlistair Popple * 2293b756a3b5SAlistair Popple * Returns: number of pages found in the range by GUP. A page is marked for 2294b756a3b5SAlistair Popple * exclusive access only if the page pointer is non-NULL. 2295b756a3b5SAlistair Popple * 2296b756a3b5SAlistair Popple * This function finds ptes mapping page(s) to the given address range, locks 2297b756a3b5SAlistair Popple * them and replaces mappings with special swap entries preventing userspace CPU 2298b756a3b5SAlistair Popple * access. On fault these entries are replaced with the original mapping after 2299b756a3b5SAlistair Popple * calling MMU notifiers. 2300b756a3b5SAlistair Popple * 2301b756a3b5SAlistair Popple * A driver using this to program access from a device must use a mmu notifier 2302b756a3b5SAlistair Popple * critical section to hold a device specific lock during programming. Once 2303b756a3b5SAlistair Popple * programming is complete it should drop the page lock and reference after 2304b756a3b5SAlistair Popple * which point CPU access to the page will revoke the exclusive access. 2305b756a3b5SAlistair Popple */ 2306b756a3b5SAlistair Popple int make_device_exclusive_range(struct mm_struct *mm, unsigned long start, 2307b756a3b5SAlistair Popple unsigned long end, struct page **pages, 2308b756a3b5SAlistair Popple void *owner) 2309b756a3b5SAlistair Popple { 2310b756a3b5SAlistair Popple long npages = (end - start) >> PAGE_SHIFT; 2311b756a3b5SAlistair Popple long i; 2312b756a3b5SAlistair Popple 2313b756a3b5SAlistair Popple npages = get_user_pages_remote(mm, start, npages, 2314b756a3b5SAlistair Popple FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD, 2315b756a3b5SAlistair Popple pages, NULL, NULL); 2316b756a3b5SAlistair Popple if (npages < 0) 2317b756a3b5SAlistair Popple return npages; 2318b756a3b5SAlistair Popple 2319b756a3b5SAlistair Popple for (i = 0; i < npages; i++, start += PAGE_SIZE) { 23200d251485SMatthew Wilcox (Oracle) struct folio *folio = page_folio(pages[i]); 23210d251485SMatthew Wilcox (Oracle) if (PageTail(pages[i]) || !folio_trylock(folio)) { 23220d251485SMatthew Wilcox (Oracle) folio_put(folio); 2323b756a3b5SAlistair Popple pages[i] = NULL; 2324b756a3b5SAlistair Popple continue; 2325b756a3b5SAlistair Popple } 2326b756a3b5SAlistair Popple 23270d251485SMatthew Wilcox (Oracle) if (!folio_make_device_exclusive(folio, mm, start, owner)) { 23280d251485SMatthew Wilcox (Oracle) folio_unlock(folio); 23290d251485SMatthew Wilcox (Oracle) folio_put(folio); 2330b756a3b5SAlistair Popple pages[i] = NULL; 2331b756a3b5SAlistair Popple } 2332b756a3b5SAlistair Popple } 2333b756a3b5SAlistair Popple 2334b756a3b5SAlistair Popple return npages; 2335b756a3b5SAlistair Popple } 2336b756a3b5SAlistair Popple EXPORT_SYMBOL_GPL(make_device_exclusive_range); 2337b756a3b5SAlistair Popple #endif 2338b756a3b5SAlistair Popple 233901d8b20dSPeter Zijlstra void __put_anon_vma(struct anon_vma *anon_vma) 234076545066SRik van Riel { 234176545066SRik van Riel struct anon_vma *root = anon_vma->root; 234276545066SRik van Riel 2343624483f3SAndrey Ryabinin anon_vma_free(anon_vma); 234401d8b20dSPeter Zijlstra if (root != anon_vma && atomic_dec_and_test(&root->refcount)) 234576545066SRik van Riel anon_vma_free(root); 234676545066SRik van Riel } 234776545066SRik van Riel 23482f031c6fSMatthew Wilcox (Oracle) static struct anon_vma *rmap_walk_anon_lock(struct folio *folio, 23496d4675e6SMinchan Kim struct rmap_walk_control *rwc) 2350faecd8ddSJoonsoo Kim { 2351faecd8ddSJoonsoo Kim struct anon_vma *anon_vma; 2352faecd8ddSJoonsoo Kim 23530dd1c7bbSJoonsoo Kim if (rwc->anon_lock) 23546d4675e6SMinchan Kim return rwc->anon_lock(folio, rwc); 23550dd1c7bbSJoonsoo Kim 2356faecd8ddSJoonsoo Kim /* 23572f031c6fSMatthew Wilcox (Oracle) * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read() 2358faecd8ddSJoonsoo Kim * because that depends on page_mapped(); but not all its usages 2359c1e8d7c6SMichel Lespinasse * are holding mmap_lock. Users without mmap_lock are required to 2360faecd8ddSJoonsoo Kim * take a reference count to prevent the anon_vma disappearing 2361faecd8ddSJoonsoo Kim */ 2362e05b3453SMatthew Wilcox (Oracle) anon_vma = folio_anon_vma(folio); 2363faecd8ddSJoonsoo Kim if (!anon_vma) 2364faecd8ddSJoonsoo Kim return NULL; 2365faecd8ddSJoonsoo Kim 23666d4675e6SMinchan Kim if (anon_vma_trylock_read(anon_vma)) 23676d4675e6SMinchan Kim goto out; 23686d4675e6SMinchan Kim 23696d4675e6SMinchan Kim if (rwc->try_lock) { 23706d4675e6SMinchan Kim anon_vma = NULL; 23716d4675e6SMinchan Kim rwc->contended = true; 23726d4675e6SMinchan Kim goto out; 23736d4675e6SMinchan Kim } 23746d4675e6SMinchan Kim 2375faecd8ddSJoonsoo Kim anon_vma_lock_read(anon_vma); 23766d4675e6SMinchan Kim out: 2377faecd8ddSJoonsoo Kim return anon_vma; 2378faecd8ddSJoonsoo Kim } 2379faecd8ddSJoonsoo Kim 2380e9995ef9SHugh Dickins /* 2381e8351ac9SJoonsoo Kim * rmap_walk_anon - do something to anonymous page using the object-based 2382e8351ac9SJoonsoo Kim * rmap method 2383e8351ac9SJoonsoo Kim * @page: the page to be handled 2384e8351ac9SJoonsoo Kim * @rwc: control variable according to each walk type 2385e8351ac9SJoonsoo Kim * 2386e8351ac9SJoonsoo Kim * Find all the mappings of a page using the mapping pointer and the vma chains 2387e8351ac9SJoonsoo Kim * contained in the anon_vma struct it points to. 2388e9995ef9SHugh Dickins */ 238984fbbe21SMatthew Wilcox (Oracle) static void rmap_walk_anon(struct folio *folio, 23906d4675e6SMinchan Kim struct rmap_walk_control *rwc, bool locked) 2391e9995ef9SHugh Dickins { 2392e9995ef9SHugh Dickins struct anon_vma *anon_vma; 2393a8fa41adSKirill A. Shutemov pgoff_t pgoff_start, pgoff_end; 23945beb4930SRik van Riel struct anon_vma_chain *avc; 2395e9995ef9SHugh Dickins 2396b9773199SKirill A. Shutemov if (locked) { 2397e05b3453SMatthew Wilcox (Oracle) anon_vma = folio_anon_vma(folio); 2398b9773199SKirill A. Shutemov /* anon_vma disappear under us? */ 2399e05b3453SMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!anon_vma, folio); 2400b9773199SKirill A. Shutemov } else { 24012f031c6fSMatthew Wilcox (Oracle) anon_vma = rmap_walk_anon_lock(folio, rwc); 2402b9773199SKirill A. Shutemov } 2403e9995ef9SHugh Dickins if (!anon_vma) 24041df631aeSMinchan Kim return; 2405faecd8ddSJoonsoo Kim 24062f031c6fSMatthew Wilcox (Oracle) pgoff_start = folio_pgoff(folio); 24072f031c6fSMatthew Wilcox (Oracle) pgoff_end = pgoff_start + folio_nr_pages(folio) - 1; 2408a8fa41adSKirill A. Shutemov anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, 2409a8fa41adSKirill A. Shutemov pgoff_start, pgoff_end) { 24105beb4930SRik van Riel struct vm_area_struct *vma = avc->vma; 24112f031c6fSMatthew Wilcox (Oracle) unsigned long address = vma_address(&folio->page, vma); 24120dd1c7bbSJoonsoo Kim 2413494334e4SHugh Dickins VM_BUG_ON_VMA(address == -EFAULT, vma); 2414ad12695fSAndrea Arcangeli cond_resched(); 2415ad12695fSAndrea Arcangeli 24160dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 24170dd1c7bbSJoonsoo Kim continue; 24180dd1c7bbSJoonsoo Kim 24192f031c6fSMatthew Wilcox (Oracle) if (!rwc->rmap_one(folio, vma, address, rwc->arg)) 2420e9995ef9SHugh Dickins break; 24212f031c6fSMatthew Wilcox (Oracle) if (rwc->done && rwc->done(folio)) 24220dd1c7bbSJoonsoo Kim break; 2423e9995ef9SHugh Dickins } 2424b9773199SKirill A. Shutemov 2425b9773199SKirill A. Shutemov if (!locked) 24264fc3f1d6SIngo Molnar anon_vma_unlock_read(anon_vma); 2427e9995ef9SHugh Dickins } 2428e9995ef9SHugh Dickins 2429e8351ac9SJoonsoo Kim /* 2430e8351ac9SJoonsoo Kim * rmap_walk_file - do something to file page using the object-based rmap method 2431e8351ac9SJoonsoo Kim * @page: the page to be handled 2432e8351ac9SJoonsoo Kim * @rwc: control variable according to each walk type 2433e8351ac9SJoonsoo Kim * 2434e8351ac9SJoonsoo Kim * Find all the mappings of a page using the mapping pointer and the vma chains 2435e8351ac9SJoonsoo Kim * contained in the address_space struct it points to. 2436e8351ac9SJoonsoo Kim */ 243784fbbe21SMatthew Wilcox (Oracle) static void rmap_walk_file(struct folio *folio, 24386d4675e6SMinchan Kim struct rmap_walk_control *rwc, bool locked) 2439e9995ef9SHugh Dickins { 24402f031c6fSMatthew Wilcox (Oracle) struct address_space *mapping = folio_mapping(folio); 2441a8fa41adSKirill A. Shutemov pgoff_t pgoff_start, pgoff_end; 2442e9995ef9SHugh Dickins struct vm_area_struct *vma; 2443e9995ef9SHugh Dickins 24449f32624bSJoonsoo Kim /* 24459f32624bSJoonsoo Kim * The page lock not only makes sure that page->mapping cannot 24469f32624bSJoonsoo Kim * suddenly be NULLified by truncation, it makes sure that the 24479f32624bSJoonsoo Kim * structure at mapping cannot be freed and reused yet, 2448c8c06efaSDavidlohr Bueso * so we can safely take mapping->i_mmap_rwsem. 24499f32624bSJoonsoo Kim */ 24502f031c6fSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 24519f32624bSJoonsoo Kim 2452e9995ef9SHugh Dickins if (!mapping) 24531df631aeSMinchan Kim return; 24543dec0ba0SDavidlohr Bueso 24552f031c6fSMatthew Wilcox (Oracle) pgoff_start = folio_pgoff(folio); 24562f031c6fSMatthew Wilcox (Oracle) pgoff_end = pgoff_start + folio_nr_pages(folio) - 1; 24576d4675e6SMinchan Kim if (!locked) { 24586d4675e6SMinchan Kim if (i_mmap_trylock_read(mapping)) 24596d4675e6SMinchan Kim goto lookup; 24606d4675e6SMinchan Kim 24616d4675e6SMinchan Kim if (rwc->try_lock) { 24626d4675e6SMinchan Kim rwc->contended = true; 24636d4675e6SMinchan Kim return; 24646d4675e6SMinchan Kim } 24656d4675e6SMinchan Kim 24663dec0ba0SDavidlohr Bueso i_mmap_lock_read(mapping); 24676d4675e6SMinchan Kim } 24686d4675e6SMinchan Kim lookup: 2469a8fa41adSKirill A. Shutemov vma_interval_tree_foreach(vma, &mapping->i_mmap, 2470a8fa41adSKirill A. Shutemov pgoff_start, pgoff_end) { 24712f031c6fSMatthew Wilcox (Oracle) unsigned long address = vma_address(&folio->page, vma); 24720dd1c7bbSJoonsoo Kim 2473494334e4SHugh Dickins VM_BUG_ON_VMA(address == -EFAULT, vma); 2474ad12695fSAndrea Arcangeli cond_resched(); 2475ad12695fSAndrea Arcangeli 24760dd1c7bbSJoonsoo Kim if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 24770dd1c7bbSJoonsoo Kim continue; 24780dd1c7bbSJoonsoo Kim 24792f031c6fSMatthew Wilcox (Oracle) if (!rwc->rmap_one(folio, vma, address, rwc->arg)) 24800dd1c7bbSJoonsoo Kim goto done; 24812f031c6fSMatthew Wilcox (Oracle) if (rwc->done && rwc->done(folio)) 24820dd1c7bbSJoonsoo Kim goto done; 2483e9995ef9SHugh Dickins } 24840dd1c7bbSJoonsoo Kim 24850dd1c7bbSJoonsoo Kim done: 2486b9773199SKirill A. Shutemov if (!locked) 24873dec0ba0SDavidlohr Bueso i_mmap_unlock_read(mapping); 2488e9995ef9SHugh Dickins } 2489e9995ef9SHugh Dickins 24906d4675e6SMinchan Kim void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc) 2491e9995ef9SHugh Dickins { 24922f031c6fSMatthew Wilcox (Oracle) if (unlikely(folio_test_ksm(folio))) 24932f031c6fSMatthew Wilcox (Oracle) rmap_walk_ksm(folio, rwc); 24942f031c6fSMatthew Wilcox (Oracle) else if (folio_test_anon(folio)) 24952f031c6fSMatthew Wilcox (Oracle) rmap_walk_anon(folio, rwc, false); 2496e9995ef9SHugh Dickins else 24972f031c6fSMatthew Wilcox (Oracle) rmap_walk_file(folio, rwc, false); 2498b9773199SKirill A. Shutemov } 2499b9773199SKirill A. Shutemov 2500b9773199SKirill A. Shutemov /* Like rmap_walk, but caller holds relevant rmap lock */ 25016d4675e6SMinchan Kim void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc) 2502b9773199SKirill A. Shutemov { 2503b9773199SKirill A. Shutemov /* no ksm support for now */ 25042f031c6fSMatthew Wilcox (Oracle) VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio); 25052f031c6fSMatthew Wilcox (Oracle) if (folio_test_anon(folio)) 25062f031c6fSMatthew Wilcox (Oracle) rmap_walk_anon(folio, rwc, true); 2507b9773199SKirill A. Shutemov else 25082f031c6fSMatthew Wilcox (Oracle) rmap_walk_file(folio, rwc, true); 2509e9995ef9SHugh Dickins } 25100fe6e20bSNaoya Horiguchi 2511e3390f67SNaoya Horiguchi #ifdef CONFIG_HUGETLB_PAGE 25120fe6e20bSNaoya Horiguchi /* 2513451b9514SKirill Tkhai * The following two functions are for anonymous (private mapped) hugepages. 25140fe6e20bSNaoya Horiguchi * Unlike common anonymous pages, anonymous hugepages have no accounting code 25150fe6e20bSNaoya Horiguchi * and no lru code, because we handle hugepages differently from common pages. 251628c5209dSDavid Hildenbrand * 251728c5209dSDavid Hildenbrand * RMAP_COMPOUND is ignored. 25180fe6e20bSNaoya Horiguchi */ 251928c5209dSDavid Hildenbrand void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma, 252028c5209dSDavid Hildenbrand unsigned long address, rmap_t flags) 25210fe6e20bSNaoya Horiguchi { 25220fe6e20bSNaoya Horiguchi struct anon_vma *anon_vma = vma->anon_vma; 25230fe6e20bSNaoya Horiguchi int first; 2524a850ea30SNaoya Horiguchi 2525a850ea30SNaoya Horiguchi BUG_ON(!PageLocked(page)); 25260fe6e20bSNaoya Horiguchi BUG_ON(!anon_vma); 25275dbe0af4SHugh Dickins /* address might be in next vma when migration races vma_adjust */ 252853f9263bSKirill A. Shutemov first = atomic_inc_and_test(compound_mapcount_ptr(page)); 25296c287605SDavid Hildenbrand VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page); 25306c287605SDavid Hildenbrand VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page); 25310fe6e20bSNaoya Horiguchi if (first) 253228c5209dSDavid Hildenbrand __page_set_anon_rmap(page, vma, address, 253328c5209dSDavid Hildenbrand !!(flags & RMAP_EXCLUSIVE)); 25340fe6e20bSNaoya Horiguchi } 25350fe6e20bSNaoya Horiguchi 25360fe6e20bSNaoya Horiguchi void hugepage_add_new_anon_rmap(struct page *page, 25370fe6e20bSNaoya Horiguchi struct vm_area_struct *vma, unsigned long address) 25380fe6e20bSNaoya Horiguchi { 25390fe6e20bSNaoya Horiguchi BUG_ON(address < vma->vm_start || address >= vma->vm_end); 254053f9263bSKirill A. Shutemov atomic_set(compound_mapcount_ptr(page), 0); 254147e29d32SJohn Hubbard atomic_set(compound_pincount_ptr(page), 0); 254247e29d32SJohn Hubbard 2543451b9514SKirill Tkhai __page_set_anon_rmap(page, vma, address, 1); 25440fe6e20bSNaoya Horiguchi } 2545e3390f67SNaoya Horiguchi #endif /* CONFIG_HUGETLB_PAGE */ 2546