Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * mm/rmap.c - physical to virtual reverse mappings |
| 3 | * |
| 4 | * Copyright 2001, Rik van Riel <riel@conectiva.com.br> |
| 5 | * Released under the General Public License (GPL). |
| 6 | * |
| 7 | * Simple, low overhead reverse mapping scheme. |
| 8 | * Please try to keep this thing as modular as possible. |
| 9 | * |
| 10 | * Provides methods for unmapping each kind of mapped page: |
| 11 | * the anon methods track anonymous pages, and |
| 12 | * the file methods track pages belonging to an inode. |
| 13 | * |
| 14 | * Original design by Rik van Riel <riel@conectiva.com.br> 2001 |
| 15 | * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 |
| 16 | * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 |
Hugh Dickins | 98f3260 | 2009-05-21 20:33:58 +0100 | [diff] [blame] | 17 | * Contributions by Hugh Dickins 2003, 2004 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 18 | */ |
| 19 | |
| 20 | /* |
| 21 | * Lock ordering in mm: |
| 22 | * |
Jan Kara | 9608703 | 2021-04-12 15:50:21 +0200 | [diff] [blame] | 23 | * inode->i_rwsem (while writing or truncating, not reading or faulting) |
Michel Lespinasse | c1e8d7c | 2020-06-08 21:33:54 -0700 | [diff] [blame] | 24 | * mm->mmap_lock |
Jan Kara | 730633f | 2021-01-28 19:19:45 +0100 | [diff] [blame] | 25 | * mapping->invalidate_lock (in filemap_fault) |
Matthew Wilcox (Oracle) | 4dc7d37 | 2024-03-26 17:10:31 +0000 | [diff] [blame] | 26 | * folio_lock |
Mike Kravetz | 8d9bfb2 | 2022-09-14 15:18:07 -0700 | [diff] [blame] | 27 | * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below) |
Suren Baghdasaryan | 55fd6fc | 2023-02-27 09:36:14 -0800 | [diff] [blame] | 28 | * vma_start_write |
| 29 | * mapping->i_mmap_rwsem |
| 30 | * anon_vma->rwsem |
| 31 | * mm->page_table_lock or pte_lock |
| 32 | * swap_lock (in swap_duplicate, swap_info_get) |
| 33 | * mmlist_lock (in mmput, drain_mmlist and others) |
| 34 | * mapping->private_lock (in block_dirty_folio) |
| 35 | * folio_lock_memcg move_lock (in block_dirty_folio) |
| 36 | * i_pages lock (widely used) |
| 37 | * lruvec->lru_lock (in folio_lruvec_lock_irq) |
| 38 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) |
| 39 | * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) |
| 40 | * sb_lock (within inode_lock in fs/fs-writeback.c) |
| 41 | * i_pages lock (widely used, in set_page_dirty, |
| 42 | * in arch-dependent flush_dcache_mmap_lock, |
| 43 | * within bdi.wb->list_lock in __sync_single_inode) |
Andi Kleen | 6a46079 | 2009-09-16 11:50:15 +0200 | [diff] [blame] | 44 | * |
Jan Kara | 9608703 | 2021-04-12 15:50:21 +0200 | [diff] [blame] | 45 | * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon) |
Peter Zijlstra | 9b679320a | 2011-06-27 16:18:09 -0700 | [diff] [blame] | 46 | * ->tasklist_lock |
Andi Kleen | 6a46079 | 2009-09-16 11:50:15 +0200 | [diff] [blame] | 47 | * pte map lock |
Mike Kravetz | c0d0381 | 2020-04-01 21:11:05 -0700 | [diff] [blame] | 48 | * |
Mike Kravetz | 8d9bfb2 | 2022-09-14 15:18:07 -0700 | [diff] [blame] | 49 | * hugetlbfs PageHuge() take locks in this order: |
| 50 | * hugetlb_fault_mutex (hugetlbfs specific page fault mutex) |
| 51 | * vma_lock (hugetlb specific lock for pmd_sharing) |
| 52 | * mapping->i_mmap_rwsem (also used for hugetlb pmd sharing) |
Matthew Wilcox (Oracle) | 4dc7d37 | 2024-03-26 17:10:31 +0000 | [diff] [blame] | 53 | * folio_lock |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 54 | */ |
| 55 | |
| 56 | #include <linux/mm.h> |
Ingo Molnar | 6e84f31 | 2017-02-08 18:51:29 +0100 | [diff] [blame] | 57 | #include <linux/sched/mm.h> |
Ingo Molnar | 2993002 | 2017-02-08 18:51:36 +0100 | [diff] [blame] | 58 | #include <linux/sched/task.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 59 | #include <linux/pagemap.h> |
| 60 | #include <linux/swap.h> |
| 61 | #include <linux/swapops.h> |
| 62 | #include <linux/slab.h> |
| 63 | #include <linux/init.h> |
Hugh Dickins | 5ad6468 | 2009-12-14 17:59:24 -0800 | [diff] [blame] | 64 | #include <linux/ksm.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 65 | #include <linux/rmap.h> |
| 66 | #include <linux/rcupdate.h> |
Paul Gortmaker | b95f1b31 | 2011-10-16 02:01:52 -0400 | [diff] [blame] | 67 | #include <linux/export.h> |
Balbir Singh | 8a9f3cc | 2008-02-07 00:13:53 -0800 | [diff] [blame] | 68 | #include <linux/memcontrol.h> |
Andrea Arcangeli | cddb8a5 | 2008-07-28 15:46:29 -0700 | [diff] [blame] | 69 | #include <linux/mmu_notifier.h> |
KOSAKI Motohiro | 64cdd54 | 2009-01-06 14:39:16 -0800 | [diff] [blame] | 70 | #include <linux/migrate.h> |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 71 | #include <linux/hugetlb.h> |
Ben Dooks | 444f84f | 2019-10-18 20:20:17 -0700 | [diff] [blame] | 72 | #include <linux/huge_mm.h> |
Jan Kara | ef5d437 | 2012-10-25 13:37:31 -0700 | [diff] [blame] | 73 | #include <linux/backing-dev.h> |
Vladimir Davydov | 33c3fc7 | 2015-09-09 15:35:45 -0700 | [diff] [blame] | 74 | #include <linux/page_idle.h> |
Jérôme Glisse | a5430dd | 2017-09-08 16:12:17 -0700 | [diff] [blame] | 75 | #include <linux/memremap.h> |
Christian Borntraeger | bce73e4 | 2018-07-13 16:58:52 -0700 | [diff] [blame] | 76 | #include <linux/userfaultfd_k.h> |
Peter Xu | 999dad8 | 2022-05-12 20:22:53 -0700 | [diff] [blame] | 77 | #include <linux/mm_inline.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 78 | |
| 79 | #include <asm/tlbflush.h> |
| 80 | |
Anshuman Khandual | 4cc79b3 | 2022-03-24 18:10:01 -0700 | [diff] [blame] | 81 | #define CREATE_TRACE_POINTS |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 82 | #include <trace/events/tlb.h> |
Anshuman Khandual | 4cc79b3 | 2022-03-24 18:10:01 -0700 | [diff] [blame] | 83 | #include <trace/events/migrate.h> |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 84 | |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 85 | #include "internal.h" |
| 86 | |
Adrian Bunk | fdd2e5f | 2008-10-18 20:28:38 -0700 | [diff] [blame] | 87 | static struct kmem_cache *anon_vma_cachep; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 88 | static struct kmem_cache *anon_vma_chain_cachep; |
Adrian Bunk | fdd2e5f | 2008-10-18 20:28:38 -0700 | [diff] [blame] | 89 | |
| 90 | static inline struct anon_vma *anon_vma_alloc(void) |
| 91 | { |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 92 | struct anon_vma *anon_vma; |
| 93 | |
| 94 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); |
| 95 | if (anon_vma) { |
| 96 | atomic_set(&anon_vma->refcount, 1); |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 97 | anon_vma->num_children = 0; |
| 98 | anon_vma->num_active_vmas = 0; |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 99 | anon_vma->parent = anon_vma; |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 100 | /* |
| 101 | * Initialise the anon_vma root to point to itself. If called |
| 102 | * from fork, the root will be reset to the parents anon_vma. |
| 103 | */ |
| 104 | anon_vma->root = anon_vma; |
| 105 | } |
| 106 | |
| 107 | return anon_vma; |
Adrian Bunk | fdd2e5f | 2008-10-18 20:28:38 -0700 | [diff] [blame] | 108 | } |
| 109 | |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 110 | static inline void anon_vma_free(struct anon_vma *anon_vma) |
Adrian Bunk | fdd2e5f | 2008-10-18 20:28:38 -0700 | [diff] [blame] | 111 | { |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 112 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 113 | |
| 114 | /* |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 115 | * Synchronize against folio_lock_anon_vma_read() such that |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 116 | * we can safely hold the lock without the anon_vma getting |
| 117 | * freed. |
| 118 | * |
| 119 | * Relies on the full mb implied by the atomic_dec_and_test() from |
| 120 | * put_anon_vma() against the acquire barrier implied by |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 121 | * down_read_trylock() from folio_lock_anon_vma_read(). This orders: |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 122 | * |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 123 | * folio_lock_anon_vma_read() VS put_anon_vma() |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 124 | * down_read_trylock() atomic_dec_and_test() |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 125 | * LOCK MB |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 126 | * atomic_read() rwsem_is_locked() |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 127 | * |
| 128 | * LOCK should suffice since the actual taking of the lock must |
| 129 | * happen _before_ what follows. |
| 130 | */ |
Hugh Dickins | 7f39dda | 2014-06-04 16:05:33 -0700 | [diff] [blame] | 131 | might_sleep(); |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 132 | if (rwsem_is_locked(&anon_vma->root->rwsem)) { |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 133 | anon_vma_lock_write(anon_vma); |
Konstantin Khlebnikov | 08b5270 | 2013-02-22 16:34:40 -0800 | [diff] [blame] | 134 | anon_vma_unlock_write(anon_vma); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 135 | } |
| 136 | |
Adrian Bunk | fdd2e5f | 2008-10-18 20:28:38 -0700 | [diff] [blame] | 137 | kmem_cache_free(anon_vma_cachep, anon_vma); |
| 138 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 139 | |
Linus Torvalds | dd34739 | 2011-06-17 19:05:36 -0700 | [diff] [blame] | 140 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 141 | { |
Linus Torvalds | dd34739 | 2011-06-17 19:05:36 -0700 | [diff] [blame] | 142 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 143 | } |
| 144 | |
Namhyung Kim | e574b5f | 2010-10-26 14:22:02 -0700 | [diff] [blame] | 145 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 146 | { |
| 147 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); |
| 148 | } |
| 149 | |
Kautuk Consul | 6583a84 | 2012-03-21 16:34:01 -0700 | [diff] [blame] | 150 | static void anon_vma_chain_link(struct vm_area_struct *vma, |
| 151 | struct anon_vma_chain *avc, |
| 152 | struct anon_vma *anon_vma) |
| 153 | { |
| 154 | avc->vma = vma; |
| 155 | avc->anon_vma = anon_vma; |
| 156 | list_add(&avc->same_vma, &vma->anon_vma_chain); |
Michel Lespinasse | bf181b9 | 2012-10-08 16:31:39 -0700 | [diff] [blame] | 157 | anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); |
Kautuk Consul | 6583a84 | 2012-03-21 16:34:01 -0700 | [diff] [blame] | 158 | } |
| 159 | |
Linus Torvalds | d9d332e | 2008-10-19 10:32:20 -0700 | [diff] [blame] | 160 | /** |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 161 | * __anon_vma_prepare - attach an anon_vma to a memory region |
Linus Torvalds | d9d332e | 2008-10-19 10:32:20 -0700 | [diff] [blame] | 162 | * @vma: the memory region in question |
| 163 | * |
| 164 | * This makes sure the memory mapping described by 'vma' has |
| 165 | * an 'anon_vma' attached to it, so that we can associate the |
| 166 | * anonymous pages mapped into it with that anon_vma. |
| 167 | * |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 168 | * The common case will be that we already have one, which |
| 169 | * is handled inline by anon_vma_prepare(). But if |
Figo.zhang | 23a0790 | 2010-12-27 15:14:06 +0100 | [diff] [blame] | 170 | * not we either need to find an adjacent mapping that we |
Linus Torvalds | d9d332e | 2008-10-19 10:32:20 -0700 | [diff] [blame] | 171 | * can re-use the anon_vma from (very common when the only |
| 172 | * reason for splitting a vma has been mprotect()), or we |
| 173 | * allocate a new one. |
| 174 | * |
| 175 | * Anon-vma allocations are very subtle, because we may have |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 176 | * optimistically looked up an anon_vma in folio_lock_anon_vma_read() |
Miaohe Lin | aaf1f99 | 2021-02-25 17:17:53 -0800 | [diff] [blame] | 177 | * and that may actually touch the rwsem even in the newly |
Linus Torvalds | d9d332e | 2008-10-19 10:32:20 -0700 | [diff] [blame] | 178 | * allocated vma (it depends on RCU to make sure that the |
| 179 | * anon_vma isn't actually destroyed). |
| 180 | * |
| 181 | * As a result, we need to do proper anon_vma locking even |
| 182 | * for the new allocation. At the same time, we do not want |
| 183 | * to do any locking for the common case of already having |
| 184 | * an anon_vma. |
Linus Torvalds | d9d332e | 2008-10-19 10:32:20 -0700 | [diff] [blame] | 185 | */ |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 186 | int __anon_vma_prepare(struct vm_area_struct *vma) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 187 | { |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 188 | struct mm_struct *mm = vma->vm_mm; |
| 189 | struct anon_vma *anon_vma, *allocated; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 190 | struct anon_vma_chain *avc; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 191 | |
Matthew Wilcox (Oracle) | 3be5106 | 2024-04-26 15:45:00 +0100 | [diff] [blame] | 192 | mmap_assert_locked(mm); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 193 | might_sleep(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 194 | |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 195 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 196 | if (!avc) |
| 197 | goto out_enomem; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 198 | |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 199 | anon_vma = find_mergeable_anon_vma(vma); |
| 200 | allocated = NULL; |
| 201 | if (!anon_vma) { |
| 202 | anon_vma = anon_vma_alloc(); |
| 203 | if (unlikely(!anon_vma)) |
| 204 | goto out_enomem_free_avc; |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 205 | anon_vma->num_children++; /* self-parent link for new root */ |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 206 | allocated = anon_vma; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 207 | } |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 208 | |
| 209 | anon_vma_lock_write(anon_vma); |
| 210 | /* page_table_lock to protect against threads */ |
| 211 | spin_lock(&mm->page_table_lock); |
| 212 | if (likely(!vma->anon_vma)) { |
| 213 | vma->anon_vma = anon_vma; |
| 214 | anon_vma_chain_link(vma, avc, anon_vma); |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 215 | anon_vma->num_active_vmas++; |
Vlastimil Babka | d5a187d | 2016-12-12 16:44:38 -0800 | [diff] [blame] | 216 | allocated = NULL; |
| 217 | avc = NULL; |
| 218 | } |
| 219 | spin_unlock(&mm->page_table_lock); |
| 220 | anon_vma_unlock_write(anon_vma); |
| 221 | |
| 222 | if (unlikely(allocated)) |
| 223 | put_anon_vma(allocated); |
| 224 | if (unlikely(avc)) |
| 225 | anon_vma_chain_free(avc); |
| 226 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 227 | return 0; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 228 | |
| 229 | out_enomem_free_avc: |
| 230 | anon_vma_chain_free(avc); |
| 231 | out_enomem: |
| 232 | return -ENOMEM; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 233 | } |
| 234 | |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 235 | /* |
| 236 | * This is a useful helper function for locking the anon_vma root as |
| 237 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that |
| 238 | * have the same vma. |
| 239 | * |
| 240 | * Such anon_vma's should have the same root, so you'd expect to see |
| 241 | * just a single mutex_lock for the whole traversal. |
| 242 | */ |
| 243 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) |
| 244 | { |
| 245 | struct anon_vma *new_root = anon_vma->root; |
| 246 | if (new_root != root) { |
| 247 | if (WARN_ON_ONCE(root)) |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 248 | up_write(&root->rwsem); |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 249 | root = new_root; |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 250 | down_write(&root->rwsem); |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 251 | } |
| 252 | return root; |
| 253 | } |
| 254 | |
| 255 | static inline void unlock_anon_vma_root(struct anon_vma *root) |
| 256 | { |
| 257 | if (root) |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 258 | up_write(&root->rwsem); |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 259 | } |
| 260 | |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 261 | /* |
| 262 | * Attach the anon_vmas from src to dst. |
| 263 | * Returns 0 on success, -ENOMEM on failure. |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 264 | * |
Liam R. Howlett | 0503ea8 | 2023-01-20 11:26:49 -0500 | [diff] [blame] | 265 | * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(), |
| 266 | * copy_vma() and anon_vma_fork(). The first four want an exact copy of src, |
| 267 | * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to |
| 268 | * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before |
| 269 | * call, we can identify this case by checking (!dst->anon_vma && |
| 270 | * src->anon_vma). |
Wei Yang | 47b390d | 2019-11-30 17:50:56 -0800 | [diff] [blame] | 271 | * |
| 272 | * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find |
| 273 | * and reuse existing anon_vma which has no vmas and only one child anon_vma. |
| 274 | * This prevents degradation of anon_vma hierarchy to endless linear chain in |
| 275 | * case of constantly forking task. On the other hand, an anon_vma with more |
| 276 | * than one child isn't reused even if there was no alive vma, thus rmap |
| 277 | * walker has a good chance of avoiding scanning the whole hierarchy when it |
| 278 | * searches where page is mapped. |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 279 | */ |
| 280 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 281 | { |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 282 | struct anon_vma_chain *avc, *pavc; |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 283 | struct anon_vma *root = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 284 | |
Linus Torvalds | 646d87b | 2010-04-11 17:15:03 -0700 | [diff] [blame] | 285 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 286 | struct anon_vma *anon_vma; |
| 287 | |
Linus Torvalds | dd34739 | 2011-06-17 19:05:36 -0700 | [diff] [blame] | 288 | avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); |
| 289 | if (unlikely(!avc)) { |
| 290 | unlock_anon_vma_root(root); |
| 291 | root = NULL; |
| 292 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 293 | if (!avc) |
| 294 | goto enomem_failure; |
| 295 | } |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 296 | anon_vma = pavc->anon_vma; |
| 297 | root = lock_anon_vma_root(root, anon_vma); |
| 298 | anon_vma_chain_link(dst, avc, anon_vma); |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 299 | |
| 300 | /* |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 301 | * Reuse existing anon_vma if it has no vma and only one |
| 302 | * anon_vma child. |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 303 | * |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 304 | * Root anon_vma is never reused: |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 305 | * it has self-parent reference and at least one child. |
| 306 | */ |
Wei Yang | 47b390d | 2019-11-30 17:50:56 -0800 | [diff] [blame] | 307 | if (!dst->anon_vma && src->anon_vma && |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 308 | anon_vma->num_children < 2 && |
| 309 | anon_vma->num_active_vmas == 0) |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 310 | dst->anon_vma = anon_vma; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 311 | } |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 312 | if (dst->anon_vma) |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 313 | dst->anon_vma->num_active_vmas++; |
Linus Torvalds | bb4aa39 | 2011-06-16 20:44:51 -0700 | [diff] [blame] | 314 | unlock_anon_vma_root(root); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 315 | return 0; |
| 316 | |
| 317 | enomem_failure: |
Leon Yu | 3fe89b3 | 2015-03-25 15:55:11 -0700 | [diff] [blame] | 318 | /* |
Ma Wupeng | d8e454e | 2022-10-14 09:39:31 +0800 | [diff] [blame] | 319 | * dst->anon_vma is dropped here otherwise its num_active_vmas can |
| 320 | * be incorrectly decremented in unlink_anon_vmas(). |
Leon Yu | 3fe89b3 | 2015-03-25 15:55:11 -0700 | [diff] [blame] | 321 | * We can safely do this because callers of anon_vma_clone() don't care |
| 322 | * about dst->anon_vma if anon_vma_clone() failed. |
| 323 | */ |
| 324 | dst->anon_vma = NULL; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 325 | unlink_anon_vmas(dst); |
| 326 | return -ENOMEM; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 327 | } |
| 328 | |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 329 | /* |
| 330 | * Attach vma to its own anon_vma, as well as to the anon_vmas that |
| 331 | * the corresponding VMA in the parent process is attached to. |
| 332 | * Returns 0 on success, non-zero on failure. |
| 333 | */ |
| 334 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 335 | { |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 336 | struct anon_vma_chain *avc; |
| 337 | struct anon_vma *anon_vma; |
Daniel Forrest | c4ea95d | 2014-12-02 15:59:42 -0800 | [diff] [blame] | 338 | int error; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 339 | |
| 340 | /* Don't bother if the parent process has no anon_vma here. */ |
| 341 | if (!pvma->anon_vma) |
| 342 | return 0; |
| 343 | |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 344 | /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ |
| 345 | vma->anon_vma = NULL; |
| 346 | |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 347 | /* |
| 348 | * First, attach the new VMA to the parent VMA's anon_vmas, |
| 349 | * so rmap can find non-COWed pages in child processes. |
| 350 | */ |
Daniel Forrest | c4ea95d | 2014-12-02 15:59:42 -0800 | [diff] [blame] | 351 | error = anon_vma_clone(vma, pvma); |
| 352 | if (error) |
| 353 | return error; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 354 | |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 355 | /* An existing anon_vma has been reused, all done then. */ |
| 356 | if (vma->anon_vma) |
| 357 | return 0; |
| 358 | |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 359 | /* Then add our own anon_vma. */ |
| 360 | anon_vma = anon_vma_alloc(); |
| 361 | if (!anon_vma) |
| 362 | goto out_error; |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 363 | anon_vma->num_active_vmas++; |
Linus Torvalds | dd34739 | 2011-06-17 19:05:36 -0700 | [diff] [blame] | 364 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 365 | if (!avc) |
| 366 | goto out_error_free_anon_vma; |
Rik van Riel | 5c341ee1 | 2010-08-09 17:18:39 -0700 | [diff] [blame] | 367 | |
| 368 | /* |
Miaohe Lin | aaf1f99 | 2021-02-25 17:17:53 -0800 | [diff] [blame] | 369 | * The root anon_vma's rwsem is the lock actually used when we |
Rik van Riel | 5c341ee1 | 2010-08-09 17:18:39 -0700 | [diff] [blame] | 370 | * lock any of the anon_vmas in this anon_vma tree. |
| 371 | */ |
| 372 | anon_vma->root = pvma->anon_vma->root; |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 373 | anon_vma->parent = pvma->anon_vma; |
Rik van Riel | 7654506 | 2010-08-09 17:18:41 -0700 | [diff] [blame] | 374 | /* |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 375 | * With refcounts, an anon_vma can stay around longer than the |
| 376 | * process it belongs to. The root anon_vma needs to be pinned until |
| 377 | * this anon_vma is freed, because the lock lives in the root. |
Rik van Riel | 7654506 | 2010-08-09 17:18:41 -0700 | [diff] [blame] | 378 | */ |
| 379 | get_anon_vma(anon_vma->root); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 380 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ |
| 381 | vma->anon_vma = anon_vma; |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 382 | anon_vma_lock_write(anon_vma); |
Rik van Riel | 5c341ee1 | 2010-08-09 17:18:39 -0700 | [diff] [blame] | 383 | anon_vma_chain_link(vma, avc, anon_vma); |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 384 | anon_vma->parent->num_children++; |
Konstantin Khlebnikov | 08b5270 | 2013-02-22 16:34:40 -0800 | [diff] [blame] | 385 | anon_vma_unlock_write(anon_vma); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 386 | |
| 387 | return 0; |
| 388 | |
| 389 | out_error_free_anon_vma: |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 390 | put_anon_vma(anon_vma); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 391 | out_error: |
Rik van Riel | 4946d54 | 2010-04-05 12:13:33 -0400 | [diff] [blame] | 392 | unlink_anon_vmas(vma); |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 393 | return -ENOMEM; |
| 394 | } |
| 395 | |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 396 | void unlink_anon_vmas(struct vm_area_struct *vma) |
| 397 | { |
| 398 | struct anon_vma_chain *avc, *next; |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 399 | struct anon_vma *root = NULL; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 400 | |
Rik van Riel | 5c341ee1 | 2010-08-09 17:18:39 -0700 | [diff] [blame] | 401 | /* |
| 402 | * Unlink each anon_vma chained to the VMA. This list is ordered |
| 403 | * from newest to oldest, ensuring the root anon_vma gets freed last. |
| 404 | */ |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 405 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 406 | struct anon_vma *anon_vma = avc->anon_vma; |
| 407 | |
| 408 | root = lock_anon_vma_root(root, anon_vma); |
Michel Lespinasse | bf181b9 | 2012-10-08 16:31:39 -0700 | [diff] [blame] | 409 | anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 410 | |
| 411 | /* |
| 412 | * Leave empty anon_vmas on the list - we'll need |
| 413 | * to free them outside the lock. |
| 414 | */ |
Davidlohr Bueso | f808c13 | 2017-09-08 16:15:08 -0700 | [diff] [blame] | 415 | if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) { |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 416 | anon_vma->parent->num_children--; |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 417 | continue; |
Konstantin Khlebnikov | 7a3ef20 | 2015-01-08 14:32:15 -0800 | [diff] [blame] | 418 | } |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 419 | |
| 420 | list_del(&avc->same_vma); |
| 421 | anon_vma_chain_free(avc); |
| 422 | } |
Li Xinhai | ee8ab19 | 2021-02-24 12:04:49 -0800 | [diff] [blame] | 423 | if (vma->anon_vma) { |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 424 | vma->anon_vma->num_active_vmas--; |
Li Xinhai | ee8ab19 | 2021-02-24 12:04:49 -0800 | [diff] [blame] | 425 | |
| 426 | /* |
| 427 | * vma would still be needed after unlink, and anon_vma will be prepared |
| 428 | * when handle fault. |
| 429 | */ |
| 430 | vma->anon_vma = NULL; |
| 431 | } |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 432 | unlock_anon_vma_root(root); |
| 433 | |
| 434 | /* |
| 435 | * Iterate the list once more, it now only contains empty and unlinked |
| 436 | * anon_vmas, destroy them. Could not do before due to __put_anon_vma() |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 437 | * needing to write-acquire the anon_vma->root->rwsem. |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 438 | */ |
| 439 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
| 440 | struct anon_vma *anon_vma = avc->anon_vma; |
| 441 | |
Jann Horn | 2555283 | 2022-08-31 19:06:00 +0200 | [diff] [blame] | 442 | VM_WARN_ON(anon_vma->num_children); |
| 443 | VM_WARN_ON(anon_vma->num_active_vmas); |
Peter Zijlstra | eee2acb | 2011-06-17 13:54:23 +0200 | [diff] [blame] | 444 | put_anon_vma(anon_vma); |
| 445 | |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 446 | list_del(&avc->same_vma); |
| 447 | anon_vma_chain_free(avc); |
| 448 | } |
| 449 | } |
| 450 | |
Alexey Dobriyan | 51cc506 | 2008-07-25 19:45:34 -0700 | [diff] [blame] | 451 | static void anon_vma_ctor(void *data) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 452 | { |
Christoph Lameter | a35afb8 | 2007-05-16 22:10:57 -0700 | [diff] [blame] | 453 | struct anon_vma *anon_vma = data; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 454 | |
Ingo Molnar | 5a50508 | 2012-12-02 19:56:46 +0000 | [diff] [blame] | 455 | init_rwsem(&anon_vma->rwsem); |
Peter Zijlstra | 8381326 | 2011-03-22 16:32:48 -0700 | [diff] [blame] | 456 | atomic_set(&anon_vma->refcount, 0); |
Davidlohr Bueso | f808c13 | 2017-09-08 16:15:08 -0700 | [diff] [blame] | 457 | anon_vma->rb_root = RB_ROOT_CACHED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 458 | } |
| 459 | |
| 460 | void __init anon_vma_init(void) |
| 461 | { |
| 462 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), |
Paul E. McKenney | 5f0d5a3 | 2017-01-18 02:53:44 -0800 | [diff] [blame] | 463 | 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, |
Vladimir Davydov | 5d09705 | 2016-01-14 15:18:21 -0800 | [diff] [blame] | 464 | anon_vma_ctor); |
| 465 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, |
| 466 | SLAB_PANIC|SLAB_ACCOUNT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 467 | } |
| 468 | |
| 469 | /* |
Peter Zijlstra | 6111e4c | 2011-05-24 17:12:08 -0700 | [diff] [blame] | 470 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! |
| 471 | * |
David Hildenbrand | 4d8f741 | 2023-12-20 23:44:56 +0100 | [diff] [blame] | 472 | * Since there is no serialization what so ever against folio_remove_rmap_*() |
Miaohe Lin | ad8a20c | 2021-02-25 17:18:06 -0800 | [diff] [blame] | 473 | * the best this function can do is return a refcount increased anon_vma |
| 474 | * that might have been relevant to this page. |
Peter Zijlstra | 6111e4c | 2011-05-24 17:12:08 -0700 | [diff] [blame] | 475 | * |
| 476 | * The page might have been remapped to a different anon_vma or the anon_vma |
| 477 | * returned may already be freed (and even reused). |
| 478 | * |
Peter Zijlstra | bc658c9 | 2011-05-29 10:33:44 +0200 | [diff] [blame] | 479 | * In case it was remapped to a different anon_vma, the new anon_vma will be a |
| 480 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore |
| 481 | * ensure that any anon_vma obtained from the page will still be valid for as |
| 482 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. |
| 483 | * |
Peter Zijlstra | 6111e4c | 2011-05-24 17:12:08 -0700 | [diff] [blame] | 484 | * All users of this function must be very careful when walking the anon_vma |
| 485 | * chain and verify that the page in question is indeed mapped in it |
| 486 | * [ something equivalent to page_mapped_in_vma() ]. |
| 487 | * |
Miles Chen | 091e429 | 2019-11-30 17:51:23 -0800 | [diff] [blame] | 488 | * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from |
David Hildenbrand | 4d8f741 | 2023-12-20 23:44:56 +0100 | [diff] [blame] | 489 | * folio_remove_rmap_*() that the anon_vma pointer from page->mapping is valid |
Miles Chen | 091e429 | 2019-11-30 17:51:23 -0800 | [diff] [blame] | 490 | * if there is a mapcount, we can dereference the anon_vma after observing |
| 491 | * those. |
Andrea Arcangeli | adef440 | 2023-12-06 02:36:56 -0800 | [diff] [blame] | 492 | * |
| 493 | * NOTE: the caller should normally hold folio lock when calling this. If |
| 494 | * not, the caller needs to double check the anon_vma didn't change after |
| 495 | * taking the anon_vma lock for either read or write (UFFDIO_MOVE can modify it |
| 496 | * concurrently without folio lock protection). See folio_lock_anon_vma_read() |
| 497 | * which has already covered that, and comment above remap_pages(). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 498 | */ |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 499 | struct anon_vma *folio_get_anon_vma(struct folio *folio) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 500 | { |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 501 | struct anon_vma *anon_vma = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 502 | unsigned long anon_mapping; |
| 503 | |
| 504 | rcu_read_lock(); |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 505 | anon_mapping = (unsigned long)READ_ONCE(folio->mapping); |
Hugh Dickins | 3ca7b3c | 2009-12-14 17:58:57 -0800 | [diff] [blame] | 506 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 507 | goto out; |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 508 | if (!folio_mapped(folio)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 509 | goto out; |
| 510 | |
| 511 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 512 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
| 513 | anon_vma = NULL; |
| 514 | goto out; |
| 515 | } |
Hugh Dickins | f1819427 | 2010-08-25 23:12:54 -0700 | [diff] [blame] | 516 | |
| 517 | /* |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 518 | * If this folio is still mapped, then its anon_vma cannot have been |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 519 | * freed. But if it has been unmapped, we have no security against the |
| 520 | * anon_vma structure being freed and reused (for another anon_vma: |
Paul E. McKenney | 5f0d5a3 | 2017-01-18 02:53:44 -0800 | [diff] [blame] | 521 | * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero() |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 522 | * above cannot corrupt). |
Hugh Dickins | f1819427 | 2010-08-25 23:12:54 -0700 | [diff] [blame] | 523 | */ |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 524 | if (!folio_mapped(folio)) { |
Hugh Dickins | 7f39dda | 2014-06-04 16:05:33 -0700 | [diff] [blame] | 525 | rcu_read_unlock(); |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 526 | put_anon_vma(anon_vma); |
Hugh Dickins | 7f39dda | 2014-06-04 16:05:33 -0700 | [diff] [blame] | 527 | return NULL; |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 528 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 529 | out: |
| 530 | rcu_read_unlock(); |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 531 | |
| 532 | return anon_vma; |
| 533 | } |
| 534 | |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 535 | /* |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 536 | * Similar to folio_get_anon_vma() except it locks the anon_vma. |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 537 | * |
| 538 | * Its a little more complex as it tries to keep the fast path to a single |
| 539 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a |
Matthew Wilcox (Oracle) | 29eea9b | 2022-09-02 20:46:50 +0100 | [diff] [blame] | 540 | * reference like with folio_get_anon_vma() and then block on the mutex |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 541 | * on !rwc->try_lock case. |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 542 | */ |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 543 | struct anon_vma *folio_lock_anon_vma_read(struct folio *folio, |
| 544 | struct rmap_walk_control *rwc) |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 545 | { |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 546 | struct anon_vma *anon_vma = NULL; |
Hugh Dickins | eee0f25 | 2011-05-28 13:20:21 -0700 | [diff] [blame] | 547 | struct anon_vma *root_anon_vma; |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 548 | unsigned long anon_mapping; |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 549 | |
Andrea Arcangeli | 880a99b | 2023-12-06 02:36:55 -0800 | [diff] [blame] | 550 | retry: |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 551 | rcu_read_lock(); |
Matthew Wilcox (Oracle) | 9595d76 | 2022-02-01 23:33:08 -0500 | [diff] [blame] | 552 | anon_mapping = (unsigned long)READ_ONCE(folio->mapping); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 553 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
| 554 | goto out; |
Matthew Wilcox (Oracle) | 9595d76 | 2022-02-01 23:33:08 -0500 | [diff] [blame] | 555 | if (!folio_mapped(folio)) |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 556 | goto out; |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 557 | |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 558 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
Jason Low | 4db0c3c | 2015-04-15 16:14:08 -0700 | [diff] [blame] | 559 | root_anon_vma = READ_ONCE(anon_vma->root); |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 560 | if (down_read_trylock(&root_anon_vma->rwsem)) { |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 561 | /* |
Andrea Arcangeli | 880a99b | 2023-12-06 02:36:55 -0800 | [diff] [blame] | 562 | * folio_move_anon_rmap() might have changed the anon_vma as we |
| 563 | * might not hold the folio lock here. |
| 564 | */ |
| 565 | if (unlikely((unsigned long)READ_ONCE(folio->mapping) != |
| 566 | anon_mapping)) { |
| 567 | up_read(&root_anon_vma->rwsem); |
| 568 | rcu_read_unlock(); |
| 569 | goto retry; |
| 570 | } |
| 571 | |
| 572 | /* |
Matthew Wilcox (Oracle) | 9595d76 | 2022-02-01 23:33:08 -0500 | [diff] [blame] | 573 | * If the folio is still mapped, then this anon_vma is still |
Hugh Dickins | eee0f25 | 2011-05-28 13:20:21 -0700 | [diff] [blame] | 574 | * its anon_vma, and holding the mutex ensures that it will |
Peter Zijlstra | bc658c9 | 2011-05-29 10:33:44 +0200 | [diff] [blame] | 575 | * not go away, see anon_vma_free(). |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 576 | */ |
Matthew Wilcox (Oracle) | 9595d76 | 2022-02-01 23:33:08 -0500 | [diff] [blame] | 577 | if (!folio_mapped(folio)) { |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 578 | up_read(&root_anon_vma->rwsem); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 579 | anon_vma = NULL; |
| 580 | } |
| 581 | goto out; |
| 582 | } |
| 583 | |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 584 | if (rwc && rwc->try_lock) { |
| 585 | anon_vma = NULL; |
| 586 | rwc->contended = true; |
| 587 | goto out; |
| 588 | } |
| 589 | |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 590 | /* trylock failed, we got to sleep */ |
| 591 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
| 592 | anon_vma = NULL; |
| 593 | goto out; |
| 594 | } |
| 595 | |
Matthew Wilcox (Oracle) | 9595d76 | 2022-02-01 23:33:08 -0500 | [diff] [blame] | 596 | if (!folio_mapped(folio)) { |
Hugh Dickins | 7f39dda | 2014-06-04 16:05:33 -0700 | [diff] [blame] | 597 | rcu_read_unlock(); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 598 | put_anon_vma(anon_vma); |
Hugh Dickins | 7f39dda | 2014-06-04 16:05:33 -0700 | [diff] [blame] | 599 | return NULL; |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 600 | } |
| 601 | |
| 602 | /* we pinned the anon_vma, its safe to sleep */ |
| 603 | rcu_read_unlock(); |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 604 | anon_vma_lock_read(anon_vma); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 605 | |
Andrea Arcangeli | 880a99b | 2023-12-06 02:36:55 -0800 | [diff] [blame] | 606 | /* |
| 607 | * folio_move_anon_rmap() might have changed the anon_vma as we might |
| 608 | * not hold the folio lock here. |
| 609 | */ |
| 610 | if (unlikely((unsigned long)READ_ONCE(folio->mapping) != |
| 611 | anon_mapping)) { |
| 612 | anon_vma_unlock_read(anon_vma); |
| 613 | put_anon_vma(anon_vma); |
| 614 | anon_vma = NULL; |
| 615 | goto retry; |
| 616 | } |
| 617 | |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 618 | if (atomic_dec_and_test(&anon_vma->refcount)) { |
| 619 | /* |
| 620 | * Oops, we held the last refcount, release the lock |
| 621 | * and bail -- can't simply use put_anon_vma() because |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 622 | * we'll deadlock on the anon_vma_lock_write() recursion. |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 623 | */ |
Ingo Molnar | 4fc3f1d | 2012-12-02 19:56:50 +0000 | [diff] [blame] | 624 | anon_vma_unlock_read(anon_vma); |
Peter Zijlstra | 88c2208 | 2011-05-24 17:12:13 -0700 | [diff] [blame] | 625 | __put_anon_vma(anon_vma); |
| 626 | anon_vma = NULL; |
| 627 | } |
| 628 | |
| 629 | return anon_vma; |
| 630 | |
| 631 | out: |
| 632 | rcu_read_unlock(); |
Peter Zijlstra | 746b18d | 2011-05-24 17:12:10 -0700 | [diff] [blame] | 633 | return anon_vma; |
Oleg Nesterov | 34bbd70 | 2007-02-28 20:13:49 -0800 | [diff] [blame] | 634 | } |
| 635 | |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 636 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 637 | /* |
| 638 | * Flush TLB entries for recently unmapped pages from remote CPUs. It is |
| 639 | * important if a PTE was dirty when it was unmapped that it's flushed |
| 640 | * before any IO is initiated on the page to prevent lost writes. Similarly, |
| 641 | * it must be flushed before freeing to prevent data leakage. |
| 642 | */ |
| 643 | void try_to_unmap_flush(void) |
| 644 | { |
| 645 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 646 | |
| 647 | if (!tlb_ubc->flush_required) |
| 648 | return; |
| 649 | |
Andy Lutomirski | e73ad5f | 2017-05-22 15:30:03 -0700 | [diff] [blame] | 650 | arch_tlbbatch_flush(&tlb_ubc->arch); |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 651 | tlb_ubc->flush_required = false; |
Mel Gorman | d950c94 | 2015-09-04 15:47:35 -0700 | [diff] [blame] | 652 | tlb_ubc->writable = false; |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 653 | } |
| 654 | |
Mel Gorman | d950c94 | 2015-09-04 15:47:35 -0700 | [diff] [blame] | 655 | /* Flush iff there are potentially writable TLB entries that can race with IO */ |
| 656 | void try_to_unmap_flush_dirty(void) |
| 657 | { |
| 658 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; |
| 659 | |
| 660 | if (tlb_ubc->writable) |
| 661 | try_to_unmap_flush(); |
| 662 | } |
| 663 | |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 664 | /* |
| 665 | * Bits 0-14 of mm->tlb_flush_batched record pending generations. |
| 666 | * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations. |
| 667 | */ |
| 668 | #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16 |
| 669 | #define TLB_FLUSH_BATCH_PENDING_MASK \ |
| 670 | ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1) |
| 671 | #define TLB_FLUSH_BATCH_PENDING_LARGE \ |
| 672 | (TLB_FLUSH_BATCH_PENDING_MASK / 2) |
| 673 | |
Barry Song | f73419b | 2023-07-17 21:10:02 +0800 | [diff] [blame] | 674 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval, |
| 675 | unsigned long uaddr) |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 676 | { |
| 677 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; |
Uros Bizjak | bdeb918 | 2023-02-27 22:42:28 +0100 | [diff] [blame] | 678 | int batch; |
Huang Ying | 4d4b6d6 | 2023-04-24 14:54:08 +0800 | [diff] [blame] | 679 | bool writable = pte_dirty(pteval); |
| 680 | |
| 681 | if (!pte_accessible(mm, pteval)) |
| 682 | return; |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 683 | |
Barry Song | f73419b | 2023-07-17 21:10:02 +0800 | [diff] [blame] | 684 | arch_tlbbatch_add_pending(&tlb_ubc->arch, mm, uaddr); |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 685 | tlb_ubc->flush_required = true; |
Mel Gorman | d950c94 | 2015-09-04 15:47:35 -0700 | [diff] [blame] | 686 | |
| 687 | /* |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 688 | * Ensure compiler does not re-order the setting of tlb_flush_batched |
| 689 | * before the PTE is cleared. |
| 690 | */ |
| 691 | barrier(); |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 692 | batch = atomic_read(&mm->tlb_flush_batched); |
| 693 | retry: |
| 694 | if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) { |
| 695 | /* |
| 696 | * Prevent `pending' from catching up with `flushed' because of |
| 697 | * overflow. Reset `pending' and `flushed' to be 1 and 0 if |
| 698 | * `pending' becomes large. |
| 699 | */ |
Uros Bizjak | bdeb918 | 2023-02-27 22:42:28 +0100 | [diff] [blame] | 700 | if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1)) |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 701 | goto retry; |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 702 | } else { |
| 703 | atomic_inc(&mm->tlb_flush_batched); |
| 704 | } |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 705 | |
| 706 | /* |
Mel Gorman | d950c94 | 2015-09-04 15:47:35 -0700 | [diff] [blame] | 707 | * If the PTE was dirty then it's best to assume it's writable. The |
| 708 | * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() |
| 709 | * before the page is queued for IO. |
| 710 | */ |
| 711 | if (writable) |
| 712 | tlb_ubc->writable = true; |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 713 | } |
| 714 | |
| 715 | /* |
| 716 | * Returns true if the TLB flush should be deferred to the end of a batch of |
| 717 | * unmap operations to reduce IPIs. |
| 718 | */ |
| 719 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) |
| 720 | { |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 721 | if (!(flags & TTU_BATCH_FLUSH)) |
| 722 | return false; |
| 723 | |
Anshuman Khandual | 65c8d30 | 2023-07-17 21:10:01 +0800 | [diff] [blame] | 724 | return arch_tlbbatch_should_defer(mm); |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 725 | } |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 726 | |
| 727 | /* |
| 728 | * Reclaim unmaps pages under the PTL but do not flush the TLB prior to |
| 729 | * releasing the PTL if TLB flushes are batched. It's possible for a parallel |
| 730 | * operation such as mprotect or munmap to race between reclaim unmapping |
| 731 | * the page and flushing the page. If this race occurs, it potentially allows |
| 732 | * access to data via a stale TLB entry. Tracking all mm's that have TLB |
| 733 | * batching in flight would be expensive during reclaim so instead track |
| 734 | * whether TLB batching occurred in the past and if so then do a flush here |
| 735 | * if required. This will cost one additional flush per reclaim cycle paid |
| 736 | * by the first operation at risk such as mprotect and mumap. |
| 737 | * |
| 738 | * This must be called under the PTL so that an access to tlb_flush_batched |
| 739 | * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise |
| 740 | * via the PTL. |
| 741 | */ |
| 742 | void flush_tlb_batched_pending(struct mm_struct *mm) |
| 743 | { |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 744 | int batch = atomic_read(&mm->tlb_flush_batched); |
| 745 | int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK; |
| 746 | int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT; |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 747 | |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 748 | if (pending != flushed) { |
Yicong Yang | db6c1f6 | 2023-07-17 21:10:03 +0800 | [diff] [blame] | 749 | arch_flush_tlb_batched_pending(mm); |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 750 | /* |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 751 | * If the new TLB flushing is pending during flushing, leave |
| 752 | * mm->tlb_flush_batched as is, to avoid losing flushing. |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 753 | */ |
Huang Ying | 5ee2fa2 | 2022-01-14 14:09:16 -0800 | [diff] [blame] | 754 | atomic_cmpxchg(&mm->tlb_flush_batched, batch, |
| 755 | pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT)); |
Mel Gorman | 3ea2771 | 2017-08-02 13:31:52 -0700 | [diff] [blame] | 756 | } |
| 757 | } |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 758 | #else |
Barry Song | f73419b | 2023-07-17 21:10:02 +0800 | [diff] [blame] | 759 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval, |
| 760 | unsigned long uaddr) |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 761 | { |
| 762 | } |
| 763 | |
| 764 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) |
| 765 | { |
| 766 | return false; |
| 767 | } |
| 768 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ |
| 769 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 770 | /* |
Huang Shijie | bf89c8c | 2009-10-01 15:44:04 -0700 | [diff] [blame] | 771 | * At what user virtual address is page expected in vma? |
Naoya Horiguchi | ab941e0 | 2010-05-11 14:06:55 -0700 | [diff] [blame] | 772 | * Caller should check the page is actually part of the vma. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 773 | */ |
| 774 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) |
| 775 | { |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 776 | struct folio *folio = page_folio(page); |
Matthew Wilcox (Oracle) | 412ad5f | 2024-03-28 22:58:28 +0000 | [diff] [blame] | 777 | pgoff_t pgoff; |
| 778 | |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 779 | if (folio_test_anon(folio)) { |
| 780 | struct anon_vma *page__anon_vma = folio_anon_vma(folio); |
Hugh Dickins | 4829b90 | 2010-10-02 17:46:06 -0700 | [diff] [blame] | 781 | /* |
| 782 | * Note: swapoff's unuse_vma() is more efficient with this |
| 783 | * check, and needs it to match anon_vma when KSM is active. |
| 784 | */ |
| 785 | if (!vma->anon_vma || !page__anon_vma || |
| 786 | vma->anon_vma->root != page__anon_vma->root) |
Andrea Arcangeli | 21d0d44 | 2010-08-09 17:19:10 -0700 | [diff] [blame] | 787 | return -EFAULT; |
Jue Wang | 3165717 | 2021-06-15 18:24:00 -0700 | [diff] [blame] | 788 | } else if (!vma->vm_file) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 789 | return -EFAULT; |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 790 | } else if (vma->vm_file->f_mapping != folio->mapping) { |
Jue Wang | 3165717 | 2021-06-15 18:24:00 -0700 | [diff] [blame] | 791 | return -EFAULT; |
| 792 | } |
Hugh Dickins | 494334e | 2021-06-15 18:23:56 -0700 | [diff] [blame] | 793 | |
Matthew Wilcox (Oracle) | 412ad5f | 2024-03-28 22:58:28 +0000 | [diff] [blame] | 794 | /* The !page__anon_vma above handles KSM folios */ |
| 795 | pgoff = folio->index + folio_page_idx(folio, page); |
Matthew Wilcox (Oracle) | e0abfbb | 2024-03-28 22:58:29 +0000 | [diff] [blame] | 796 | return vma_address(vma, pgoff, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 797 | } |
| 798 | |
Zach O'Keefe | 5072280 | 2022-07-06 16:59:26 -0700 | [diff] [blame] | 799 | /* |
| 800 | * Returns the actual pmd_t* where we expect 'address' to be mapped from, or |
| 801 | * NULL if it doesn't exist. No guarantees / checks on what the pmd_t* |
| 802 | * represents. |
| 803 | */ |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 804 | pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) |
| 805 | { |
| 806 | pgd_t *pgd; |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 807 | p4d_t *p4d; |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 808 | pud_t *pud; |
| 809 | pmd_t *pmd = NULL; |
| 810 | |
| 811 | pgd = pgd_offset(mm, address); |
| 812 | if (!pgd_present(*pgd)) |
| 813 | goto out; |
| 814 | |
Kirill A. Shutemov | c2febaf | 2017-03-09 17:24:07 +0300 | [diff] [blame] | 815 | p4d = p4d_offset(pgd, address); |
| 816 | if (!p4d_present(*p4d)) |
| 817 | goto out; |
| 818 | |
| 819 | pud = pud_offset(p4d, address); |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 820 | if (!pud_present(*pud)) |
| 821 | goto out; |
| 822 | |
| 823 | pmd = pmd_offset(pud, address); |
Bob Liu | 6219049 | 2012-12-11 16:00:37 -0800 | [diff] [blame] | 824 | out: |
| 825 | return pmd; |
| 826 | } |
| 827 | |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 828 | struct folio_referenced_arg { |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 829 | int mapcount; |
| 830 | int referenced; |
| 831 | unsigned long vm_flags; |
| 832 | struct mem_cgroup *memcg; |
| 833 | }; |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 834 | |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 835 | /* |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 836 | * arg: folio_referenced_arg will be passed |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 837 | */ |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 838 | static bool folio_referenced_one(struct folio *folio, |
| 839 | struct vm_area_struct *vma, unsigned long address, void *arg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 840 | { |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 841 | struct folio_referenced_arg *pra = arg; |
| 842 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); |
Vladimir Davydov | 8749cfe | 2016-01-15 16:54:45 -0800 | [diff] [blame] | 843 | int referenced = 0; |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 844 | unsigned long start = address, ptes = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 845 | |
Kirill A. Shutemov | 8eaeded | 2017-02-24 14:57:48 -0800 | [diff] [blame] | 846 | while (page_vma_mapped_walk(&pvmw)) { |
| 847 | address = pvmw.address; |
Kirill A. Shutemov | b20ce5e | 2016-01-15 16:54:37 -0800 | [diff] [blame] | 848 | |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 849 | if (vma->vm_flags & VM_LOCKED) { |
| 850 | if (!folio_test_large(folio) || !pvmw.pte) { |
| 851 | /* Restore the mlock which got missed */ |
| 852 | mlock_vma_folio(folio, vma); |
| 853 | page_vma_mapped_walk_done(&pvmw); |
| 854 | pra->vm_flags |= VM_LOCKED; |
| 855 | return false; /* To break the loop */ |
| 856 | } |
| 857 | /* |
| 858 | * For large folio fully mapped to VMA, will |
| 859 | * be handled after the pvmw loop. |
| 860 | * |
| 861 | * For large folio cross VMA boundaries, it's |
| 862 | * expected to be picked by page reclaim. But |
| 863 | * should skip reference of pages which are in |
| 864 | * the range of VM_LOCKED vma. As page reclaim |
| 865 | * should just count the reference of pages out |
| 866 | * the range of VM_LOCKED vma. |
| 867 | */ |
| 868 | ptes++; |
| 869 | pra->mapcount--; |
| 870 | continue; |
Vladimir Davydov | 8749cfe | 2016-01-15 16:54:45 -0800 | [diff] [blame] | 871 | } |
Kirill A. Shutemov | 8eaeded | 2017-02-24 14:57:48 -0800 | [diff] [blame] | 872 | |
| 873 | if (pvmw.pte) { |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 874 | if (lru_gen_enabled() && |
| 875 | pte_young(ptep_get(pvmw.pte))) { |
Yu Zhao | 018ee47 | 2022-09-18 02:00:04 -0600 | [diff] [blame] | 876 | lru_gen_look_around(&pvmw); |
| 877 | referenced++; |
| 878 | } |
| 879 | |
Kirill A. Shutemov | 8eaeded | 2017-02-24 14:57:48 -0800 | [diff] [blame] | 880 | if (ptep_clear_flush_young_notify(vma, address, |
Yu Zhao | 8788f67 | 2022-12-30 14:52:51 -0700 | [diff] [blame] | 881 | pvmw.pte)) |
| 882 | referenced++; |
Kirill A. Shutemov | 8eaeded | 2017-02-24 14:57:48 -0800 | [diff] [blame] | 883 | } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { |
| 884 | if (pmdp_clear_flush_young_notify(vma, address, |
| 885 | pvmw.pmd)) |
| 886 | referenced++; |
| 887 | } else { |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 888 | /* unexpected pmd-mapped folio? */ |
Kirill A. Shutemov | 8eaeded | 2017-02-24 14:57:48 -0800 | [diff] [blame] | 889 | WARN_ON_ONCE(1); |
| 890 | } |
| 891 | |
| 892 | pra->mapcount--; |
Kirill A. Shutemov | b20ce5e | 2016-01-15 16:54:37 -0800 | [diff] [blame] | 893 | } |
Kirill A. Shutemov | b20ce5e | 2016-01-15 16:54:37 -0800 | [diff] [blame] | 894 | |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 895 | if ((vma->vm_flags & VM_LOCKED) && |
| 896 | folio_test_large(folio) && |
| 897 | folio_within_vma(folio, vma)) { |
| 898 | unsigned long s_align, e_align; |
| 899 | |
| 900 | s_align = ALIGN_DOWN(start, PMD_SIZE); |
| 901 | e_align = ALIGN_DOWN(start + folio_size(folio) - 1, PMD_SIZE); |
| 902 | |
| 903 | /* folio doesn't cross page table boundary and fully mapped */ |
| 904 | if ((s_align == e_align) && (ptes == folio_nr_pages(folio))) { |
| 905 | /* Restore the mlock which got missed */ |
| 906 | mlock_vma_folio(folio, vma); |
| 907 | pra->vm_flags |= VM_LOCKED; |
| 908 | return false; /* To break the loop */ |
| 909 | } |
| 910 | } |
| 911 | |
Vladimir Davydov | 33c3fc7 | 2015-09-09 15:35:45 -0700 | [diff] [blame] | 912 | if (referenced) |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 913 | folio_clear_idle(folio); |
| 914 | if (folio_test_clear_young(folio)) |
Vladimir Davydov | 33c3fc7 | 2015-09-09 15:35:45 -0700 | [diff] [blame] | 915 | referenced++; |
| 916 | |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 917 | if (referenced) { |
| 918 | pra->referenced++; |
Hugh Dickins | 47d4f3e | 2022-02-14 18:42:33 -0800 | [diff] [blame] | 919 | pra->vm_flags |= vma->vm_flags & ~VM_LOCKED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 920 | } |
Oleg Nesterov | 34bbd70 | 2007-02-28 20:13:49 -0800 | [diff] [blame] | 921 | |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 922 | if (!pra->mapcount) |
Minchan Kim | e4b8222 | 2017-05-03 14:54:27 -0700 | [diff] [blame] | 923 | return false; /* To break the loop */ |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 924 | |
Minchan Kim | e4b8222 | 2017-05-03 14:54:27 -0700 | [diff] [blame] | 925 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 926 | } |
| 927 | |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 928 | static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 929 | { |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 930 | struct folio_referenced_arg *pra = arg; |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 931 | struct mem_cgroup *memcg = pra->memcg; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 932 | |
Yu Zhao | 8788f67 | 2022-12-30 14:52:51 -0700 | [diff] [blame] | 933 | /* |
| 934 | * Ignore references from this mapping if it has no recency. If the |
| 935 | * folio has been used in another mapping, we will catch it; if this |
| 936 | * other mapping is already gone, the unmap path will have set the |
| 937 | * referenced flag or activated the folio in zap_pte_range(). |
| 938 | */ |
| 939 | if (!vma_has_recency(vma)) |
| 940 | return true; |
| 941 | |
| 942 | /* |
| 943 | * If we are reclaiming on behalf of a cgroup, skip counting on behalf |
| 944 | * of references from different cgroups. |
| 945 | */ |
| 946 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 947 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 948 | |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 949 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 950 | } |
| 951 | |
| 952 | /** |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 953 | * folio_referenced() - Test if the folio was referenced. |
| 954 | * @folio: The folio to test. |
| 955 | * @is_locked: Caller holds lock on the folio. |
Johannes Weiner | 72835c8 | 2012-01-12 17:18:32 -0800 | [diff] [blame] | 956 | * @memcg: target memory cgroup |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 957 | * @vm_flags: A combination of all the vma->vm_flags which referenced the folio. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 958 | * |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 959 | * Quick test_and_clear_referenced for all mappings of a folio, |
| 960 | * |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 961 | * Return: The number of mappings which referenced the folio. Return -1 if |
| 962 | * the function bailed out due to rmap lock contention. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 963 | */ |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 964 | int folio_referenced(struct folio *folio, int is_locked, |
| 965 | struct mem_cgroup *memcg, unsigned long *vm_flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 966 | { |
Hao Ge | 637a900 | 2024-04-28 09:20:48 +0800 | [diff] [blame] | 967 | bool we_locked = false; |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 968 | struct folio_referenced_arg pra = { |
| 969 | .mapcount = folio_mapcount(folio), |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 970 | .memcg = memcg, |
| 971 | }; |
| 972 | struct rmap_walk_control rwc = { |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 973 | .rmap_one = folio_referenced_one, |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 974 | .arg = (void *)&pra, |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 975 | .anon_lock = folio_lock_anon_vma_read, |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 976 | .try_lock = true, |
Yu Zhao | 8788f67 | 2022-12-30 14:52:51 -0700 | [diff] [blame] | 977 | .invalid_vma = invalid_folio_referenced_vma, |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 978 | }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 979 | |
Wu Fengguang | 6fe6b7e | 2009-06-16 15:33:05 -0700 | [diff] [blame] | 980 | *vm_flags = 0; |
Huang Shijie | 059d844 | 2019-05-13 17:21:07 -0700 | [diff] [blame] | 981 | if (!pra.mapcount) |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 982 | return 0; |
| 983 | |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 984 | if (!folio_raw_mapping(folio)) |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 985 | return 0; |
| 986 | |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 987 | if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) { |
| 988 | we_locked = folio_trylock(folio); |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 989 | if (!we_locked) |
| 990 | return 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 991 | } |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 992 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 993 | rmap_walk(folio, &rwc); |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 994 | *vm_flags = pra.vm_flags; |
| 995 | |
| 996 | if (we_locked) |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 997 | folio_unlock(folio); |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 998 | |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 999 | return rwc.contended ? -1 : pra.referenced; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1000 | } |
| 1001 | |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1002 | static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw) |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1003 | { |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1004 | int cleaned = 0; |
| 1005 | struct vm_area_struct *vma = pvmw->vma; |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1006 | struct mmu_notifier_range range; |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1007 | unsigned long address = pvmw->address; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1008 | |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1009 | /* |
| 1010 | * We have to assume the worse case ie pmd for invalidation. Note that |
Matthew Wilcox (Oracle) | e83c09a | 2022-01-20 18:20:07 -0500 | [diff] [blame] | 1011 | * the folio can not be freed from this function. |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1012 | */ |
Alistair Popple | 7d4a8be | 2023-01-10 13:57:22 +1100 | [diff] [blame] | 1013 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0, |
| 1014 | vma->vm_mm, address, vma_address_end(pvmw)); |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1015 | mmu_notifier_invalidate_range_start(&range); |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1016 | |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1017 | while (page_vma_mapped_walk(pvmw)) { |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1018 | int ret = 0; |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1019 | |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1020 | address = pvmw->address; |
| 1021 | if (pvmw->pte) { |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1022 | pte_t *pte = pvmw->pte; |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1023 | pte_t entry = ptep_get(pte); |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1024 | |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1025 | if (!pte_dirty(entry) && !pte_write(entry)) |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1026 | continue; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1027 | |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1028 | flush_cache_page(vma, address, pte_pfn(entry)); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1029 | entry = ptep_clear_flush(vma, address, pte); |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1030 | entry = pte_wrprotect(entry); |
| 1031 | entry = pte_mkclean(entry); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1032 | set_pte_at(vma->vm_mm, address, pte, entry); |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1033 | ret = 1; |
| 1034 | } else { |
Matthew Wilcox (Oracle) | 396bcc5 | 2020-04-06 20:04:35 -0700 | [diff] [blame] | 1035 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1036 | pmd_t *pmd = pvmw->pmd; |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1037 | pmd_t entry; |
| 1038 | |
| 1039 | if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) |
| 1040 | continue; |
| 1041 | |
Muchun Song | 7f9c9b6 | 2022-04-28 23:16:09 -0700 | [diff] [blame] | 1042 | flush_cache_range(vma, address, |
| 1043 | address + HPAGE_PMD_SIZE); |
Aneesh Kumar K.V | 024eee0 | 2019-05-13 17:19:11 -0700 | [diff] [blame] | 1044 | entry = pmdp_invalidate(vma, address, pmd); |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1045 | entry = pmd_wrprotect(entry); |
| 1046 | entry = pmd_mkclean(entry); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1047 | set_pmd_at(vma->vm_mm, address, pmd, entry); |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1048 | ret = 1; |
| 1049 | #else |
Matthew Wilcox (Oracle) | e83c09a | 2022-01-20 18:20:07 -0500 | [diff] [blame] | 1050 | /* unexpected pmd-mapped folio? */ |
Kirill A. Shutemov | f27176c | 2017-02-24 14:57:57 -0800 | [diff] [blame] | 1051 | WARN_ON_ONCE(1); |
| 1052 | #endif |
| 1053 | } |
| 1054 | |
Jérôme Glisse | 0f10851 | 2017-11-15 17:34:07 -0800 | [diff] [blame] | 1055 | if (ret) |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1056 | cleaned++; |
Peter Zijlstra | c2fda5f | 2006-12-22 14:25:52 +0100 | [diff] [blame] | 1057 | } |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1058 | |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1059 | mmu_notifier_invalidate_range_end(&range); |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1060 | |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1061 | return cleaned; |
| 1062 | } |
| 1063 | |
| 1064 | static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma, |
| 1065 | unsigned long address, void *arg) |
| 1066 | { |
| 1067 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC); |
| 1068 | int *cleaned = arg; |
| 1069 | |
| 1070 | *cleaned += page_vma_mkclean_one(&pvmw); |
| 1071 | |
Minchan Kim | e4b8222 | 2017-05-03 14:54:27 -0700 | [diff] [blame] | 1072 | return true; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1073 | } |
| 1074 | |
Joonsoo Kim | 9853a40 | 2014-01-21 15:49:55 -0800 | [diff] [blame] | 1075 | static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1076 | { |
Joonsoo Kim | 9853a40 | 2014-01-21 15:49:55 -0800 | [diff] [blame] | 1077 | if (vma->vm_flags & VM_SHARED) |
Fengguang Wu | 871beb8 | 2014-01-23 15:53:41 -0800 | [diff] [blame] | 1078 | return false; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1079 | |
Fengguang Wu | 871beb8 | 2014-01-23 15:53:41 -0800 | [diff] [blame] | 1080 | return true; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1081 | } |
| 1082 | |
Matthew Wilcox (Oracle) | d9c08e2 | 2021-02-28 16:02:57 -0500 | [diff] [blame] | 1083 | int folio_mkclean(struct folio *folio) |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1084 | { |
Joonsoo Kim | 9853a40 | 2014-01-21 15:49:55 -0800 | [diff] [blame] | 1085 | int cleaned = 0; |
| 1086 | struct address_space *mapping; |
| 1087 | struct rmap_walk_control rwc = { |
| 1088 | .arg = (void *)&cleaned, |
| 1089 | .rmap_one = page_mkclean_one, |
| 1090 | .invalid_vma = invalid_mkclean_vma, |
| 1091 | }; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1092 | |
Matthew Wilcox (Oracle) | d9c08e2 | 2021-02-28 16:02:57 -0500 | [diff] [blame] | 1093 | BUG_ON(!folio_test_locked(folio)); |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1094 | |
Matthew Wilcox (Oracle) | d9c08e2 | 2021-02-28 16:02:57 -0500 | [diff] [blame] | 1095 | if (!folio_mapped(folio)) |
Joonsoo Kim | 9853a40 | 2014-01-21 15:49:55 -0800 | [diff] [blame] | 1096 | return 0; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1097 | |
Matthew Wilcox (Oracle) | d9c08e2 | 2021-02-28 16:02:57 -0500 | [diff] [blame] | 1098 | mapping = folio_mapping(folio); |
Joonsoo Kim | 9853a40 | 2014-01-21 15:49:55 -0800 | [diff] [blame] | 1099 | if (!mapping) |
| 1100 | return 0; |
| 1101 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1102 | rmap_walk(folio, &rwc); |
Joonsoo Kim | 9853a40 | 2014-01-21 15:49:55 -0800 | [diff] [blame] | 1103 | |
| 1104 | return cleaned; |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1105 | } |
Matthew Wilcox (Oracle) | d9c08e2 | 2021-02-28 16:02:57 -0500 | [diff] [blame] | 1106 | EXPORT_SYMBOL_GPL(folio_mkclean); |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 1107 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1108 | /** |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1109 | * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of |
| 1110 | * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff) |
| 1111 | * within the @vma of shared mappings. And since clean PTEs |
| 1112 | * should also be readonly, write protects them too. |
| 1113 | * @pfn: start pfn. |
| 1114 | * @nr_pages: number of physically contiguous pages srarting with @pfn. |
| 1115 | * @pgoff: page offset that the @pfn mapped with. |
| 1116 | * @vma: vma that @pfn mapped within. |
| 1117 | * |
| 1118 | * Returns the number of cleaned PTEs (including PMDs). |
| 1119 | */ |
| 1120 | int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff, |
| 1121 | struct vm_area_struct *vma) |
| 1122 | { |
| 1123 | struct page_vma_mapped_walk pvmw = { |
| 1124 | .pfn = pfn, |
| 1125 | .nr_pages = nr_pages, |
| 1126 | .pgoff = pgoff, |
| 1127 | .vma = vma, |
| 1128 | .flags = PVMW_SYNC, |
| 1129 | }; |
| 1130 | |
| 1131 | if (invalid_mkclean_vma(vma, NULL)) |
| 1132 | return 0; |
| 1133 | |
Matthew Wilcox (Oracle) | e0abfbb | 2024-03-28 22:58:29 +0000 | [diff] [blame] | 1134 | pvmw.address = vma_address(vma, pgoff, nr_pages); |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1135 | VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma); |
| 1136 | |
| 1137 | return page_vma_mkclean_one(&pvmw); |
| 1138 | } |
| 1139 | |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1140 | static __always_inline unsigned int __folio_add_rmap(struct folio *folio, |
| 1141 | struct page *page, int nr_pages, enum rmap_level level, |
| 1142 | int *nr_pmdmapped) |
| 1143 | { |
| 1144 | atomic_t *mapped = &folio->_nr_pages_mapped; |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1145 | const int orig_nr_pages = nr_pages; |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1146 | int first, nr = 0; |
| 1147 | |
| 1148 | __folio_rmap_sanity_checks(folio, page, nr_pages, level); |
| 1149 | |
| 1150 | switch (level) { |
| 1151 | case RMAP_LEVEL_PTE: |
David Hildenbrand | 46d62de | 2024-04-09 21:22:46 +0200 | [diff] [blame] | 1152 | if (!folio_test_large(folio)) { |
| 1153 | nr = atomic_inc_and_test(&page->_mapcount); |
| 1154 | break; |
| 1155 | } |
| 1156 | |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1157 | do { |
| 1158 | first = atomic_inc_and_test(&page->_mapcount); |
David Hildenbrand | 46d62de | 2024-04-09 21:22:46 +0200 | [diff] [blame] | 1159 | if (first) { |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1160 | first = atomic_inc_return_relaxed(mapped); |
David Hildenbrand | 46d62de | 2024-04-09 21:22:46 +0200 | [diff] [blame] | 1161 | if (first < ENTIRELY_MAPPED) |
| 1162 | nr++; |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1163 | } |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1164 | } while (page++, --nr_pages > 0); |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1165 | atomic_add(orig_nr_pages, &folio->_large_mapcount); |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1166 | break; |
| 1167 | case RMAP_LEVEL_PMD: |
| 1168 | first = atomic_inc_and_test(&folio->_entire_mapcount); |
| 1169 | if (first) { |
David Hildenbrand | e78a13f | 2023-12-20 23:45:03 +0100 | [diff] [blame] | 1170 | nr = atomic_add_return_relaxed(ENTIRELY_MAPPED, mapped); |
| 1171 | if (likely(nr < ENTIRELY_MAPPED + ENTIRELY_MAPPED)) { |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1172 | *nr_pmdmapped = folio_nr_pages(folio); |
| 1173 | nr = *nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED); |
| 1174 | /* Raced ahead of a remove and another add? */ |
| 1175 | if (unlikely(nr < 0)) |
| 1176 | nr = 0; |
| 1177 | } else { |
David Hildenbrand | e78a13f | 2023-12-20 23:45:03 +0100 | [diff] [blame] | 1178 | /* Raced ahead of a remove of ENTIRELY_MAPPED */ |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1179 | nr = 0; |
| 1180 | } |
| 1181 | } |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1182 | atomic_inc(&folio->_large_mapcount); |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1183 | break; |
| 1184 | } |
| 1185 | return nr; |
| 1186 | } |
| 1187 | |
Muchun Song | 6a8e059 | 2022-04-28 23:16:10 -0700 | [diff] [blame] | 1188 | /** |
David Hildenbrand | 0696862 | 2023-10-02 16:29:48 +0200 | [diff] [blame] | 1189 | * folio_move_anon_rmap - move a folio to our anon_vma |
| 1190 | * @folio: The folio to move to our anon_vma |
| 1191 | * @vma: The vma the folio belongs to |
Rik van Riel | c44b674 | 2010-03-05 13:42:09 -0800 | [diff] [blame] | 1192 | * |
David Hildenbrand | 0696862 | 2023-10-02 16:29:48 +0200 | [diff] [blame] | 1193 | * When a folio belongs exclusively to one process after a COW event, |
| 1194 | * that folio can be moved into the anon_vma that belongs to just that |
| 1195 | * process, so the rmap code will not search the parent or sibling processes. |
Rik van Riel | c44b674 | 2010-03-05 13:42:09 -0800 | [diff] [blame] | 1196 | */ |
David Hildenbrand | 0696862 | 2023-10-02 16:29:48 +0200 | [diff] [blame] | 1197 | void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma) |
Rik van Riel | c44b674 | 2010-03-05 13:42:09 -0800 | [diff] [blame] | 1198 | { |
Matthew Wilcox (Oracle) | 595af4c9 | 2022-09-02 20:46:45 +0100 | [diff] [blame] | 1199 | void *anon_vma = vma->anon_vma; |
Rik van Riel | c44b674 | 2010-03-05 13:42:09 -0800 | [diff] [blame] | 1200 | |
Matthew Wilcox (Oracle) | 595af4c9 | 2022-09-02 20:46:45 +0100 | [diff] [blame] | 1201 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
Sasha Levin | 81d1b09 | 2014-10-09 15:28:10 -0700 | [diff] [blame] | 1202 | VM_BUG_ON_VMA(!anon_vma, vma); |
Rik van Riel | c44b674 | 2010-03-05 13:42:09 -0800 | [diff] [blame] | 1203 | |
Matthew Wilcox (Oracle) | 595af4c9 | 2022-09-02 20:46:45 +0100 | [diff] [blame] | 1204 | anon_vma += PAGE_MAPPING_ANON; |
Vladimir Davydov | 414e2fb | 2015-06-24 16:56:56 -0700 | [diff] [blame] | 1205 | /* |
| 1206 | * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written |
Matthew Wilcox (Oracle) | b3ac041 | 2022-01-21 11:27:31 -0500 | [diff] [blame] | 1207 | * simultaneously, so a concurrent reader (eg folio_referenced()'s |
| 1208 | * folio_test_anon()) will not see one without the other. |
Vladimir Davydov | 414e2fb | 2015-06-24 16:56:56 -0700 | [diff] [blame] | 1209 | */ |
Matthew Wilcox (Oracle) | 595af4c9 | 2022-09-02 20:46:45 +0100 | [diff] [blame] | 1210 | WRITE_ONCE(folio->mapping, anon_vma); |
Rik van Riel | c44b674 | 2010-03-05 13:42:09 -0800 | [diff] [blame] | 1211 | } |
| 1212 | |
| 1213 | /** |
David Hildenbrand | c66db8c | 2023-09-13 14:51:09 +0200 | [diff] [blame] | 1214 | * __folio_set_anon - set up a new anonymous rmap for a folio |
| 1215 | * @folio: The folio to set up the new anonymous rmap for. |
| 1216 | * @vma: VM area to add the folio to. |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1217 | * @address: User virtual address of the mapping |
David Hildenbrand | c66db8c | 2023-09-13 14:51:09 +0200 | [diff] [blame] | 1218 | * @exclusive: Whether the folio is exclusive to the process. |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1219 | */ |
David Hildenbrand | c66db8c | 2023-09-13 14:51:09 +0200 | [diff] [blame] | 1220 | static void __folio_set_anon(struct folio *folio, struct vm_area_struct *vma, |
| 1221 | unsigned long address, bool exclusive) |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1222 | { |
Rik van Riel | e8a03fe | 2010-04-14 17:59:28 -0400 | [diff] [blame] | 1223 | struct anon_vma *anon_vma = vma->anon_vma; |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1224 | |
Rik van Riel | e8a03fe | 2010-04-14 17:59:28 -0400 | [diff] [blame] | 1225 | BUG_ON(!anon_vma); |
Linus Torvalds | ea90002 | 2010-04-12 12:44:29 -0700 | [diff] [blame] | 1226 | |
| 1227 | /* |
David Hildenbrand | c66db8c | 2023-09-13 14:51:09 +0200 | [diff] [blame] | 1228 | * If the folio isn't exclusive to this vma, we must use the _oldest_ |
| 1229 | * possible anon_vma for the folio mapping! |
Linus Torvalds | ea90002 | 2010-04-12 12:44:29 -0700 | [diff] [blame] | 1230 | */ |
Andi Kleen | 4e1c197 | 2010-09-22 12:43:56 +0200 | [diff] [blame] | 1231 | if (!exclusive) |
Andrea Arcangeli | 288468c | 2010-08-09 17:19:09 -0700 | [diff] [blame] | 1232 | anon_vma = anon_vma->root; |
Linus Torvalds | ea90002 | 2010-04-12 12:44:29 -0700 | [diff] [blame] | 1233 | |
Alex Shi | 16f5e70 | 2020-12-15 12:33:42 -0800 | [diff] [blame] | 1234 | /* |
Matthew Wilcox (Oracle) | 5b4bd90 | 2023-01-16 19:29:59 +0000 | [diff] [blame] | 1235 | * page_idle does a lockless/optimistic rmap scan on folio->mapping. |
Alex Shi | 16f5e70 | 2020-12-15 12:33:42 -0800 | [diff] [blame] | 1236 | * Make sure the compiler doesn't split the stores of anon_vma and |
| 1237 | * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code |
| 1238 | * could mistake the mapping for a struct address_space and crash. |
| 1239 | */ |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1240 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
Matthew Wilcox (Oracle) | 5b4bd90 | 2023-01-16 19:29:59 +0000 | [diff] [blame] | 1241 | WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma); |
| 1242 | folio->index = linear_page_index(vma, address); |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1243 | } |
| 1244 | |
| 1245 | /** |
Randy Dunlap | 43d8eac | 2008-03-19 17:00:43 -0700 | [diff] [blame] | 1246 | * __page_check_anon_rmap - sanity check anonymous rmap addition |
Matthew Wilcox (Oracle) | dba438b | 2023-07-06 20:52:51 +0100 | [diff] [blame] | 1247 | * @folio: The folio containing @page. |
| 1248 | * @page: the page to check the mapping of |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1249 | * @vma: the vm area in which the mapping is added |
| 1250 | * @address: the user virtual address mapped |
| 1251 | */ |
Matthew Wilcox (Oracle) | dba438b | 2023-07-06 20:52:51 +0100 | [diff] [blame] | 1252 | static void __page_check_anon_rmap(struct folio *folio, struct page *page, |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1253 | struct vm_area_struct *vma, unsigned long address) |
| 1254 | { |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1255 | /* |
| 1256 | * The page's anon-rmap details (mapping and index) are guaranteed to |
| 1257 | * be set up correctly at this point. |
| 1258 | * |
David Hildenbrand | 84f0169 | 2023-12-20 23:44:45 +0100 | [diff] [blame] | 1259 | * We have exclusion against folio_add_anon_rmap_*() because the caller |
Miaohe Lin | 90aaca8 | 2021-02-25 17:17:59 -0800 | [diff] [blame] | 1260 | * always holds the page locked. |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1261 | * |
Matthew Wilcox (Oracle) | cb9089b | 2023-12-11 16:22:11 +0000 | [diff] [blame] | 1262 | * We have exclusion against folio_add_new_anon_rmap because those pages |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1263 | * are initially only visible via the pagetables, and the pte is locked |
Matthew Wilcox (Oracle) | cb9089b | 2023-12-11 16:22:11 +0000 | [diff] [blame] | 1264 | * over the call to folio_add_new_anon_rmap. |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1265 | */ |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 1266 | VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root, |
| 1267 | folio); |
Yang Shi | 30c4638 | 2019-11-30 17:51:26 -0800 | [diff] [blame] | 1268 | VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address), |
| 1269 | page); |
Nick Piggin | c97a9e1 | 2007-05-16 22:11:21 -0700 | [diff] [blame] | 1270 | } |
| 1271 | |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1272 | static void __folio_mod_stat(struct folio *folio, int nr, int nr_pmdmapped) |
| 1273 | { |
| 1274 | int idx; |
| 1275 | |
| 1276 | if (nr) { |
| 1277 | idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED; |
| 1278 | __lruvec_stat_mod_folio(folio, idx, nr); |
| 1279 | } |
| 1280 | if (nr_pmdmapped) { |
| 1281 | if (folio_test_anon(folio)) { |
| 1282 | idx = NR_ANON_THPS; |
| 1283 | __lruvec_stat_mod_folio(folio, idx, nr_pmdmapped); |
| 1284 | } else { |
| 1285 | /* NR_*_PMDMAPPED are not maintained per-memcg */ |
| 1286 | idx = folio_test_swapbacked(folio) ? |
| 1287 | NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED; |
| 1288 | __mod_node_page_state(folio_pgdat(folio), idx, |
| 1289 | nr_pmdmapped); |
| 1290 | } |
| 1291 | } |
| 1292 | } |
| 1293 | |
David Hildenbrand | 8bd5130 | 2023-12-20 23:44:38 +0100 | [diff] [blame] | 1294 | static __always_inline void __folio_add_anon_rmap(struct folio *folio, |
| 1295 | struct page *page, int nr_pages, struct vm_area_struct *vma, |
| 1296 | unsigned long address, rmap_t flags, enum rmap_level level) |
| 1297 | { |
| 1298 | int i, nr, nr_pmdmapped = 0; |
Hugh Dickins | be5ef2d | 2022-11-22 01:42:04 -0800 | [diff] [blame] | 1299 | |
Barry Song | 4c1171f | 2024-06-18 11:11:37 +1200 | [diff] [blame] | 1300 | VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); |
| 1301 | |
David Hildenbrand | 8bd5130 | 2023-12-20 23:44:38 +0100 | [diff] [blame] | 1302 | nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped); |
Hugh Dickins | 5ad6468 | 2009-12-14 17:59:24 -0800 | [diff] [blame] | 1303 | |
Barry Song | 4c1171f | 2024-06-18 11:11:37 +1200 | [diff] [blame] | 1304 | if (likely(!folio_test_ksm(folio))) |
David Hildenbrand | c5c5400 | 2023-09-13 14:51:10 +0200 | [diff] [blame] | 1305 | __page_check_anon_rmap(folio, page, vma, address); |
David Hildenbrand | 8bd5130 | 2023-12-20 23:44:38 +0100 | [diff] [blame] | 1306 | |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1307 | __folio_mod_stat(folio, nr, nr_pmdmapped); |
| 1308 | |
David Hildenbrand | 8bd5130 | 2023-12-20 23:44:38 +0100 | [diff] [blame] | 1309 | if (flags & RMAP_EXCLUSIVE) { |
| 1310 | switch (level) { |
| 1311 | case RMAP_LEVEL_PTE: |
| 1312 | for (i = 0; i < nr_pages; i++) |
| 1313 | SetPageAnonExclusive(page + i); |
| 1314 | break; |
| 1315 | case RMAP_LEVEL_PMD: |
| 1316 | SetPageAnonExclusive(page); |
| 1317 | break; |
| 1318 | } |
| 1319 | } |
| 1320 | for (i = 0; i < nr_pages; i++) { |
| 1321 | struct page *cur_page = page + i; |
| 1322 | |
| 1323 | /* While PTE-mapping a THP we have a PMD and a PTE mapping. */ |
| 1324 | VM_WARN_ON_FOLIO((atomic_read(&cur_page->_mapcount) > 0 || |
| 1325 | (folio_test_large(folio) && |
| 1326 | folio_entire_mapcount(folio) > 1)) && |
| 1327 | PageAnonExclusive(cur_page), folio); |
| 1328 | } |
Hugh Dickins | cea86fe | 2022-02-14 18:26:39 -0800 | [diff] [blame] | 1329 | |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 1330 | /* |
| 1331 | * For large folio, only mlock it if it's fully mapped to VMA. It's |
| 1332 | * not easy to check whether the large folio is fully mapped to VMA |
| 1333 | * here. Only mlock normal 4K folio and leave page reclaim to handle |
| 1334 | * large folio. |
| 1335 | */ |
| 1336 | if (!folio_test_large(folio)) |
| 1337 | mlock_vma_folio(folio, vma); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1338 | } |
| 1339 | |
Randy Dunlap | 43d8eac | 2008-03-19 17:00:43 -0700 | [diff] [blame] | 1340 | /** |
David Hildenbrand | 8bd5130 | 2023-12-20 23:44:38 +0100 | [diff] [blame] | 1341 | * folio_add_anon_rmap_ptes - add PTE mappings to a page range of an anon folio |
| 1342 | * @folio: The folio to add the mappings to |
| 1343 | * @page: The first page to add |
| 1344 | * @nr_pages: The number of pages which will be mapped |
| 1345 | * @vma: The vm area in which the mappings are added |
| 1346 | * @address: The user virtual address of the first page to map |
| 1347 | * @flags: The rmap flags |
| 1348 | * |
| 1349 | * The page range of folio is defined by [first_page, first_page + nr_pages) |
| 1350 | * |
| 1351 | * The caller needs to hold the page table lock, and the page must be locked in |
| 1352 | * the anon_vma case: to serialize mapping,index checking after setting, |
| 1353 | * and to ensure that an anon folio is not being upgraded racily to a KSM folio |
| 1354 | * (but KSM folios are never downgraded). |
| 1355 | */ |
| 1356 | void folio_add_anon_rmap_ptes(struct folio *folio, struct page *page, |
| 1357 | int nr_pages, struct vm_area_struct *vma, unsigned long address, |
| 1358 | rmap_t flags) |
| 1359 | { |
| 1360 | __folio_add_anon_rmap(folio, page, nr_pages, vma, address, flags, |
| 1361 | RMAP_LEVEL_PTE); |
| 1362 | } |
| 1363 | |
| 1364 | /** |
| 1365 | * folio_add_anon_rmap_pmd - add a PMD mapping to a page range of an anon folio |
| 1366 | * @folio: The folio to add the mapping to |
| 1367 | * @page: The first page to add |
| 1368 | * @vma: The vm area in which the mapping is added |
| 1369 | * @address: The user virtual address of the first page to map |
| 1370 | * @flags: The rmap flags |
| 1371 | * |
| 1372 | * The page range of folio is defined by [first_page, first_page + HPAGE_PMD_NR) |
| 1373 | * |
| 1374 | * The caller needs to hold the page table lock, and the page must be locked in |
| 1375 | * the anon_vma case: to serialize mapping,index checking after setting. |
| 1376 | */ |
| 1377 | void folio_add_anon_rmap_pmd(struct folio *folio, struct page *page, |
| 1378 | struct vm_area_struct *vma, unsigned long address, rmap_t flags) |
| 1379 | { |
| 1380 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 1381 | __folio_add_anon_rmap(folio, page, HPAGE_PMD_NR, vma, address, flags, |
| 1382 | RMAP_LEVEL_PMD); |
| 1383 | #else |
| 1384 | WARN_ON_ONCE(true); |
| 1385 | #endif |
| 1386 | } |
| 1387 | |
| 1388 | /** |
Matthew Wilcox (Oracle) | 4d510f3 | 2023-01-11 14:28:54 +0000 | [diff] [blame] | 1389 | * folio_add_new_anon_rmap - Add mapping to a new anonymous folio. |
| 1390 | * @folio: The folio to add the mapping to. |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1391 | * @vma: the vm area in which the mapping is added |
| 1392 | * @address: the user virtual address mapped |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1393 | * @flags: The rmap flags |
David Hildenbrand | 40f2bbf | 2022-05-09 18:20:43 -0700 | [diff] [blame] | 1394 | * |
David Hildenbrand | 84f0169 | 2023-12-20 23:44:45 +0100 | [diff] [blame] | 1395 | * Like folio_add_anon_rmap_*() but must only be called on *new* folios. |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1396 | * This means the inc-and-test can be bypassed. |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1397 | * The folio doesn't necessarily need to be locked while it's exclusive |
| 1398 | * unless two threads map it concurrently. However, the folio must be |
| 1399 | * locked if it's shared. |
Matthew Wilcox (Oracle) | 4d510f3 | 2023-01-11 14:28:54 +0000 | [diff] [blame] | 1400 | * |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1401 | * If the folio is pmd-mappable, it is accounted as a THP. |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1402 | */ |
Matthew Wilcox (Oracle) | 4d510f3 | 2023-01-11 14:28:54 +0000 | [diff] [blame] | 1403 | void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma, |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1404 | unsigned long address, rmap_t flags) |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1405 | { |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1406 | const int nr = folio_nr_pages(folio); |
| 1407 | const bool exclusive = flags & RMAP_EXCLUSIVE; |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1408 | int nr_pmdmapped = 0; |
Kirill A. Shutemov | d281ee6 | 2016-01-15 16:52:16 -0800 | [diff] [blame] | 1409 | |
David Hildenbrand | a4ea186 | 2023-12-20 23:44:30 +0100 | [diff] [blame] | 1410 | VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio); |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1411 | VM_WARN_ON_FOLIO(!exclusive && !folio_test_locked(folio), folio); |
Ryan Roberts | 372cbd4 | 2023-12-07 16:12:03 +0000 | [diff] [blame] | 1412 | VM_BUG_ON_VMA(address < vma->vm_start || |
| 1413 | address + (nr << PAGE_SHIFT) > vma->vm_end, vma); |
Barry Song | 9ae2fea | 2024-06-18 11:11:36 +1200 | [diff] [blame] | 1414 | |
Jason A. Donenfeld | 9651fce | 2022-12-08 17:55:04 +0100 | [diff] [blame] | 1415 | /* |
| 1416 | * VM_DROPPABLE mappings don't swap; instead they're just dropped when |
| 1417 | * under memory pressure. |
| 1418 | */ |
Linus Torvalds | 7a3fad3 | 2024-07-24 10:29:50 -0700 | [diff] [blame] | 1419 | if (!folio_test_swapbacked(folio) && !(vma->vm_flags & VM_DROPPABLE)) |
Barry Song | 9ae2fea | 2024-06-18 11:11:36 +1200 | [diff] [blame] | 1420 | __folio_set_swapbacked(folio); |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1421 | __folio_set_anon(folio, vma, address, exclusive); |
Hugh Dickins | d8dd5e9 | 2022-11-09 18:18:49 -0800 | [diff] [blame] | 1422 | |
Ryan Roberts | 372cbd4 | 2023-12-07 16:12:03 +0000 | [diff] [blame] | 1423 | if (likely(!folio_test_large(folio))) { |
Hugh Dickins | d8dd5e9 | 2022-11-09 18:18:49 -0800 | [diff] [blame] | 1424 | /* increment count (starts at -1) */ |
Matthew Wilcox (Oracle) | 4d510f3 | 2023-01-11 14:28:54 +0000 | [diff] [blame] | 1425 | atomic_set(&folio->_mapcount, 0); |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1426 | if (exclusive) |
| 1427 | SetPageAnonExclusive(&folio->page); |
Ryan Roberts | 372cbd4 | 2023-12-07 16:12:03 +0000 | [diff] [blame] | 1428 | } else if (!folio_test_pmd_mappable(folio)) { |
| 1429 | int i; |
| 1430 | |
| 1431 | for (i = 0; i < nr; i++) { |
| 1432 | struct page *page = folio_page(folio, i); |
| 1433 | |
| 1434 | /* increment count (starts at -1) */ |
| 1435 | atomic_set(&page->_mapcount, 0); |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1436 | if (exclusive) |
| 1437 | SetPageAnonExclusive(page); |
Ryan Roberts | 372cbd4 | 2023-12-07 16:12:03 +0000 | [diff] [blame] | 1438 | } |
| 1439 | |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1440 | /* increment count (starts at -1) */ |
| 1441 | atomic_set(&folio->_large_mapcount, nr - 1); |
Ryan Roberts | 372cbd4 | 2023-12-07 16:12:03 +0000 | [diff] [blame] | 1442 | atomic_set(&folio->_nr_pages_mapped, nr); |
Hugh Dickins | d8dd5e9 | 2022-11-09 18:18:49 -0800 | [diff] [blame] | 1443 | } else { |
Kirill A. Shutemov | 53f9263 | 2016-01-15 16:53:42 -0800 | [diff] [blame] | 1444 | /* increment count (starts at -1) */ |
Matthew Wilcox (Oracle) | 4d510f3 | 2023-01-11 14:28:54 +0000 | [diff] [blame] | 1445 | atomic_set(&folio->_entire_mapcount, 0); |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1446 | /* increment count (starts at -1) */ |
| 1447 | atomic_set(&folio->_large_mapcount, 0); |
David Hildenbrand | e78a13f | 2023-12-20 23:45:03 +0100 | [diff] [blame] | 1448 | atomic_set(&folio->_nr_pages_mapped, ENTIRELY_MAPPED); |
Barry Song | 15bde4a | 2024-06-18 11:11:35 +1200 | [diff] [blame] | 1449 | if (exclusive) |
| 1450 | SetPageAnonExclusive(&folio->page); |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1451 | nr_pmdmapped = nr; |
Kirill A. Shutemov | d281ee6 | 2016-01-15 16:52:16 -0800 | [diff] [blame] | 1452 | } |
Hugh Dickins | d8dd5e9 | 2022-11-09 18:18:49 -0800 | [diff] [blame] | 1453 | |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1454 | __folio_mod_stat(folio, nr, nr_pmdmapped); |
Nick Piggin | 9617d95 | 2006-01-06 00:11:12 -0800 | [diff] [blame] | 1455 | } |
| 1456 | |
David Hildenbrand | 68f03208 | 2023-12-20 23:44:31 +0100 | [diff] [blame] | 1457 | static __always_inline void __folio_add_file_rmap(struct folio *folio, |
| 1458 | struct page *page, int nr_pages, struct vm_area_struct *vma, |
| 1459 | enum rmap_level level) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1460 | { |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1461 | int nr, nr_pmdmapped = 0; |
Kirill A. Shutemov | dd78fed | 2016-07-26 15:25:26 -0700 | [diff] [blame] | 1462 | |
David Hildenbrand | 68f03208 | 2023-12-20 23:44:31 +0100 | [diff] [blame] | 1463 | VM_WARN_ON_FOLIO(folio_test_anon(folio), folio); |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1464 | |
David Hildenbrand | 96fd749 | 2023-12-20 23:44:37 +0100 | [diff] [blame] | 1465 | nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped); |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1466 | __folio_mod_stat(folio, nr, nr_pmdmapped); |
Hugh Dickins | cea86fe | 2022-02-14 18:26:39 -0800 | [diff] [blame] | 1467 | |
David Hildenbrand | 84f0169 | 2023-12-20 23:44:45 +0100 | [diff] [blame] | 1468 | /* See comments in folio_add_anon_rmap_*() */ |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 1469 | if (!folio_test_large(folio)) |
| 1470 | mlock_vma_folio(folio, vma); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1471 | } |
| 1472 | |
| 1473 | /** |
David Hildenbrand | 68f03208 | 2023-12-20 23:44:31 +0100 | [diff] [blame] | 1474 | * folio_add_file_rmap_ptes - add PTE mappings to a page range of a folio |
| 1475 | * @folio: The folio to add the mappings to |
| 1476 | * @page: The first page to add |
| 1477 | * @nr_pages: The number of pages that will be mapped using PTEs |
| 1478 | * @vma: The vm area in which the mappings are added |
| 1479 | * |
| 1480 | * The page range of the folio is defined by [page, page + nr_pages) |
| 1481 | * |
| 1482 | * The caller needs to hold the page table lock. |
| 1483 | */ |
| 1484 | void folio_add_file_rmap_ptes(struct folio *folio, struct page *page, |
| 1485 | int nr_pages, struct vm_area_struct *vma) |
| 1486 | { |
| 1487 | __folio_add_file_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE); |
| 1488 | } |
| 1489 | |
| 1490 | /** |
| 1491 | * folio_add_file_rmap_pmd - add a PMD mapping to a page range of a folio |
| 1492 | * @folio: The folio to add the mapping to |
| 1493 | * @page: The first page to add |
| 1494 | * @vma: The vm area in which the mapping is added |
| 1495 | * |
| 1496 | * The page range of the folio is defined by [page, page + HPAGE_PMD_NR) |
| 1497 | * |
| 1498 | * The caller needs to hold the page table lock. |
| 1499 | */ |
| 1500 | void folio_add_file_rmap_pmd(struct folio *folio, struct page *page, |
| 1501 | struct vm_area_struct *vma) |
| 1502 | { |
| 1503 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 1504 | __folio_add_file_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD); |
| 1505 | #else |
| 1506 | WARN_ON_ONCE(true); |
| 1507 | #endif |
| 1508 | } |
| 1509 | |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1510 | static __always_inline void __folio_remove_rmap(struct folio *folio, |
| 1511 | struct page *page, int nr_pages, struct vm_area_struct *vma, |
| 1512 | enum rmap_level level) |
| 1513 | { |
Matthew Wilcox (Oracle) | 62beb90 | 2023-01-11 14:28:51 +0000 | [diff] [blame] | 1514 | atomic_t *mapped = &folio->_nr_pages_mapped; |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1515 | int last, nr = 0, nr_pmdmapped = 0; |
Zi Yan | 7491f3f | 2024-05-02 09:28:51 -0400 | [diff] [blame] | 1516 | bool partially_mapped = false; |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1517 | |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1518 | __folio_rmap_sanity_checks(folio, page, nr_pages, level); |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1519 | |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1520 | switch (level) { |
| 1521 | case RMAP_LEVEL_PTE: |
David Hildenbrand | 46d62de | 2024-04-09 21:22:46 +0200 | [diff] [blame] | 1522 | if (!folio_test_large(folio)) { |
| 1523 | nr = atomic_add_negative(-1, &page->_mapcount); |
| 1524 | break; |
| 1525 | } |
| 1526 | |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1527 | atomic_sub(nr_pages, &folio->_large_mapcount); |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1528 | do { |
| 1529 | last = atomic_add_negative(-1, &page->_mapcount); |
David Hildenbrand | 46d62de | 2024-04-09 21:22:46 +0200 | [diff] [blame] | 1530 | if (last) { |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1531 | last = atomic_dec_return_relaxed(mapped); |
David Hildenbrand | 46d62de | 2024-04-09 21:22:46 +0200 | [diff] [blame] | 1532 | if (last < ENTIRELY_MAPPED) |
| 1533 | nr++; |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1534 | } |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1535 | } while (page++, --nr_pages > 0); |
Zi Yan | 7491f3f | 2024-05-02 09:28:51 -0400 | [diff] [blame] | 1536 | |
| 1537 | partially_mapped = nr && atomic_read(mapped); |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1538 | break; |
| 1539 | case RMAP_LEVEL_PMD: |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 1540 | atomic_dec(&folio->_large_mapcount); |
Matthew Wilcox (Oracle) | 62beb90 | 2023-01-11 14:28:51 +0000 | [diff] [blame] | 1541 | last = atomic_add_negative(-1, &folio->_entire_mapcount); |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1542 | if (last) { |
David Hildenbrand | e78a13f | 2023-12-20 23:45:03 +0100 | [diff] [blame] | 1543 | nr = atomic_sub_return_relaxed(ENTIRELY_MAPPED, mapped); |
| 1544 | if (likely(nr < ENTIRELY_MAPPED)) { |
Matthew Wilcox (Oracle) | 62beb90 | 2023-01-11 14:28:51 +0000 | [diff] [blame] | 1545 | nr_pmdmapped = folio_nr_pages(folio); |
Matthew Wilcox (Oracle) | eec2042 | 2023-01-11 14:28:48 +0000 | [diff] [blame] | 1546 | nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED); |
Hugh Dickins | 6287b7d | 2022-12-04 17:57:07 -0800 | [diff] [blame] | 1547 | /* Raced ahead of another remove and an add? */ |
| 1548 | if (unlikely(nr < 0)) |
| 1549 | nr = 0; |
| 1550 | } else { |
David Hildenbrand | e78a13f | 2023-12-20 23:45:03 +0100 | [diff] [blame] | 1551 | /* An add of ENTIRELY_MAPPED raced ahead */ |
Hugh Dickins | 6287b7d | 2022-12-04 17:57:07 -0800 | [diff] [blame] | 1552 | nr = 0; |
| 1553 | } |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1554 | } |
Zi Yan | 7491f3f | 2024-05-02 09:28:51 -0400 | [diff] [blame] | 1555 | |
| 1556 | partially_mapped = nr < nr_pmdmapped; |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1557 | break; |
Hugh Dickins | cb67f42 | 2022-11-02 18:51:38 -0700 | [diff] [blame] | 1558 | } |
| 1559 | |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1560 | if (nr) { |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1561 | /* |
Ryan Roberts | 7dc7c5e | 2023-12-07 16:12:02 +0000 | [diff] [blame] | 1562 | * Queue anon large folio for deferred split if at least one |
Matthew Wilcox (Oracle) | 62beb90 | 2023-01-11 14:28:51 +0000 | [diff] [blame] | 1563 | * page of the folio is unmapped and at least one page |
| 1564 | * is still mapped. |
Zi Yan | 7491f3f | 2024-05-02 09:28:51 -0400 | [diff] [blame] | 1565 | * |
| 1566 | * Check partially_mapped first to ensure it is a large folio. |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1567 | */ |
Zi Yan | 7491f3f | 2024-05-02 09:28:51 -0400 | [diff] [blame] | 1568 | if (folio_test_anon(folio) && partially_mapped && |
| 1569 | list_empty(&folio->_deferred_list)) |
| 1570 | deferred_split_folio(folio); |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1571 | } |
Yosry Ahmed | 15c0536f | 2024-05-06 21:13:33 +0000 | [diff] [blame] | 1572 | __folio_mod_stat(folio, -nr, -nr_pmdmapped); |
Kirill A. Shutemov | 9a98225 | 2016-01-15 16:54:17 -0800 | [diff] [blame] | 1573 | |
KOSAKI Motohiro | b904dcf | 2009-09-21 17:01:28 -0700 | [diff] [blame] | 1574 | /* |
Matthew Wilcox (Oracle) | 672aa27 | 2023-01-16 19:28:26 +0000 | [diff] [blame] | 1575 | * It would be tidy to reset folio_test_anon mapping when fully |
David Hildenbrand | 84f0169 | 2023-12-20 23:44:45 +0100 | [diff] [blame] | 1576 | * unmapped, but that might overwrite a racing folio_add_anon_rmap_*() |
Matthew Wilcox (Oracle) | 672aa27 | 2023-01-16 19:28:26 +0000 | [diff] [blame] | 1577 | * which increments mapcount after us but sets mapping before us: |
| 1578 | * so leave the reset to free_pages_prepare, and remember that |
| 1579 | * it's only reliable while mapped. |
KOSAKI Motohiro | b904dcf | 2009-09-21 17:01:28 -0700 | [diff] [blame] | 1580 | */ |
Hugh Dickins | 9bd3155 | 2022-11-02 18:53:45 -0700 | [diff] [blame] | 1581 | |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 1582 | munlock_vma_folio(folio, vma); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1583 | } |
| 1584 | |
David Hildenbrand | b06dc28 | 2023-12-20 23:44:47 +0100 | [diff] [blame] | 1585 | /** |
| 1586 | * folio_remove_rmap_ptes - remove PTE mappings from a page range of a folio |
| 1587 | * @folio: The folio to remove the mappings from |
| 1588 | * @page: The first page to remove |
| 1589 | * @nr_pages: The number of pages that will be removed from the mapping |
| 1590 | * @vma: The vm area from which the mappings are removed |
| 1591 | * |
| 1592 | * The page range of the folio is defined by [page, page + nr_pages) |
| 1593 | * |
| 1594 | * The caller needs to hold the page table lock. |
| 1595 | */ |
| 1596 | void folio_remove_rmap_ptes(struct folio *folio, struct page *page, |
| 1597 | int nr_pages, struct vm_area_struct *vma) |
| 1598 | { |
| 1599 | __folio_remove_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE); |
| 1600 | } |
| 1601 | |
| 1602 | /** |
| 1603 | * folio_remove_rmap_pmd - remove a PMD mapping from a page range of a folio |
| 1604 | * @folio: The folio to remove the mapping from |
| 1605 | * @page: The first page to remove |
| 1606 | * @vma: The vm area from which the mapping is removed |
| 1607 | * |
| 1608 | * The page range of the folio is defined by [page, page + HPAGE_PMD_NR) |
| 1609 | * |
| 1610 | * The caller needs to hold the page table lock. |
| 1611 | */ |
| 1612 | void folio_remove_rmap_pmd(struct folio *folio, struct page *page, |
| 1613 | struct vm_area_struct *vma) |
| 1614 | { |
| 1615 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 1616 | __folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD); |
| 1617 | #else |
| 1618 | WARN_ON_ONCE(true); |
| 1619 | #endif |
| 1620 | } |
| 1621 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1622 | /* |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1623 | * @arg: enum ttu_flags will be passed to this argument |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1624 | */ |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1625 | static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1626 | unsigned long address, void *arg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1627 | { |
| 1628 | struct mm_struct *mm = vma->vm_mm; |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1629 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1630 | pte_t pteval; |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1631 | struct page *subpage; |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 1632 | bool anon_exclusive, ret = true; |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1633 | struct mmu_notifier_range range; |
Palmer Dabbelt | 4708f31 | 2020-04-06 20:08:00 -0700 | [diff] [blame] | 1634 | enum ttu_flags flags = (enum ttu_flags)(long)arg; |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1635 | unsigned long pfn; |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 1636 | unsigned long hsz = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1637 | |
Hugh Dickins | 732ed55 | 2021-06-15 18:23:53 -0700 | [diff] [blame] | 1638 | /* |
| 1639 | * When racing against e.g. zap_pte_range() on another cpu, |
David Hildenbrand | ca1a074 | 2023-12-20 23:44:54 +0100 | [diff] [blame] | 1640 | * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(), |
Yang Shi | 1fb08ac | 2021-06-30 18:52:01 -0700 | [diff] [blame] | 1641 | * try_to_unmap() may return before page_mapped() has become false, |
Hugh Dickins | 732ed55 | 2021-06-15 18:23:53 -0700 | [diff] [blame] | 1642 | * if page table locking is skipped: use TTU_SYNC to wait for that. |
| 1643 | */ |
| 1644 | if (flags & TTU_SYNC) |
| 1645 | pvmw.flags = PVMW_SYNC; |
| 1646 | |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1647 | /* |
Mike Kravetz | 017b166 | 2018-10-05 15:51:29 -0700 | [diff] [blame] | 1648 | * For THP, we have to assume the worse case ie pmd for invalidation. |
| 1649 | * For hugetlb, it could be much worse if we need to do pud |
| 1650 | * invalidation in the case of pmd sharing. |
| 1651 | * |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1652 | * Note that the folio can not be freed in this function as call of |
| 1653 | * try_to_unmap() must hold a reference on the folio. |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1654 | */ |
Matthew Wilcox (Oracle) | 2aff7a4 | 2022-02-03 11:40:17 -0500 | [diff] [blame] | 1655 | range.end = vma_address_end(&pvmw); |
Alistair Popple | 7d4a8be | 2023-01-10 13:57:22 +1100 | [diff] [blame] | 1656 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, |
Hugh Dickins | 494334e | 2021-06-15 18:23:56 -0700 | [diff] [blame] | 1657 | address, range.end); |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1658 | if (folio_test_hugetlb(folio)) { |
Mike Kravetz | 017b166 | 2018-10-05 15:51:29 -0700 | [diff] [blame] | 1659 | /* |
| 1660 | * If sharing is possible, start and end will be adjusted |
| 1661 | * accordingly. |
| 1662 | */ |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1663 | adjust_range_if_pmd_sharing_possible(vma, &range.start, |
| 1664 | &range.end); |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 1665 | |
| 1666 | /* We need the huge page size for set_huge_pte_at() */ |
| 1667 | hsz = huge_page_size(hstate_vma(vma)); |
Mike Kravetz | 017b166 | 2018-10-05 15:51:29 -0700 | [diff] [blame] | 1668 | } |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1669 | mmu_notifier_invalidate_range_start(&range); |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1670 | |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1671 | while (page_vma_mapped_walk(&pvmw)) { |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1672 | /* |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1673 | * If the folio is in an mlock()d vma, we must not swap it out. |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1674 | */ |
Hugh Dickins | efdb672 | 2021-07-11 20:10:49 -0700 | [diff] [blame] | 1675 | if (!(flags & TTU_IGNORE_MLOCK) && |
| 1676 | (vma->vm_flags & VM_LOCKED)) { |
Hugh Dickins | cea86fe | 2022-02-14 18:26:39 -0800 | [diff] [blame] | 1677 | /* Restore the mlock which got missed */ |
Yin Fengwei | 1acbc3f | 2023-09-18 15:33:17 +0800 | [diff] [blame] | 1678 | if (!folio_test_large(folio)) |
| 1679 | mlock_vma_folio(folio, vma); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1680 | goto walk_abort; |
Hugh Dickins | b87537d9e | 2015-11-05 18:49:33 -0800 | [diff] [blame] | 1681 | } |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1682 | |
Lance Yang | 735ecdf | 2024-06-14 09:51:38 +0800 | [diff] [blame] | 1683 | if (!pvmw.pte) { |
| 1684 | if (unmap_huge_pmd_locked(vma, pvmw.address, pvmw.pmd, |
| 1685 | folio)) |
| 1686 | goto walk_done; |
| 1687 | |
| 1688 | if (flags & TTU_SPLIT_HUGE_PMD) { |
| 1689 | /* |
| 1690 | * We temporarily have to drop the PTL and |
| 1691 | * restart so we can process the PTE-mapped THP. |
| 1692 | */ |
| 1693 | split_huge_pmd_locked(vma, pvmw.address, |
| 1694 | pvmw.pmd, false, folio); |
| 1695 | flags &= ~TTU_SPLIT_HUGE_PMD; |
| 1696 | page_vma_mapped_walk_restart(&pvmw); |
| 1697 | continue; |
| 1698 | } |
Lance Yang | 29e847d | 2024-06-14 09:51:37 +0800 | [diff] [blame] | 1699 | } |
| 1700 | |
| 1701 | /* Unexpected PMD-mapped THP? */ |
| 1702 | VM_BUG_ON_FOLIO(!pvmw.pte, folio); |
| 1703 | |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1704 | pfn = pte_pfn(ptep_get(pvmw.pte)); |
| 1705 | subpage = folio_page(folio, pfn - folio_pfn(folio)); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1706 | address = pvmw.address; |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 1707 | anon_exclusive = folio_test_anon(folio) && |
| 1708 | PageAnonExclusive(subpage); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1709 | |
Baolin Wang | dfc7ab5 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 1710 | if (folio_test_hugetlb(folio)) { |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 1711 | bool anon = folio_test_anon(folio); |
| 1712 | |
Baolin Wang | 54205e9 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 1713 | /* |
Baolin Wang | a00a875 | 2022-05-13 16:48:56 -0700 | [diff] [blame] | 1714 | * The try_to_unmap() is only passed a hugetlb page |
| 1715 | * in the case where the hugetlb page is poisoned. |
| 1716 | */ |
| 1717 | VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage); |
| 1718 | /* |
Baolin Wang | 54205e9 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 1719 | * huge_pmd_unshare may unmap an entire PMD page. |
| 1720 | * There is no way of knowing exactly which PMDs may |
| 1721 | * be cached for this mm, so we must flush them all. |
| 1722 | * start/end were already adjusted above to cover this |
| 1723 | * range. |
| 1724 | */ |
| 1725 | flush_cache_range(vma, range.start, range.end); |
| 1726 | |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 1727 | /* |
| 1728 | * To call huge_pmd_unshare, i_mmap_rwsem must be |
| 1729 | * held in write mode. Caller needs to explicitly |
| 1730 | * do this outside rmap routines. |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 1731 | * |
| 1732 | * We also must hold hugetlb vma_lock in write mode. |
| 1733 | * Lock order dictates acquiring vma_lock BEFORE |
| 1734 | * i_mmap_rwsem. We can only try lock here and fail |
| 1735 | * if unsuccessful. |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 1736 | */ |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 1737 | if (!anon) { |
| 1738 | VM_BUG_ON(!(flags & TTU_RMAP_LOCKED)); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1739 | if (!hugetlb_vma_trylock_write(vma)) |
| 1740 | goto walk_abort; |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 1741 | if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) { |
| 1742 | hugetlb_vma_unlock_write(vma); |
| 1743 | flush_tlb_range(vma, |
| 1744 | range.start, range.end); |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 1745 | /* |
| 1746 | * The ref count of the PMD page was |
| 1747 | * dropped which is part of the way map |
| 1748 | * counting is done for shared PMDs. |
| 1749 | * Return 'true' here. When there is |
| 1750 | * no other sharing, huge_pmd_unshare |
| 1751 | * returns false and we will unmap the |
| 1752 | * actual page and drop map count |
| 1753 | * to zero. |
| 1754 | */ |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1755 | goto walk_done; |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 1756 | } |
| 1757 | hugetlb_vma_unlock_write(vma); |
Mike Kravetz | 017b166 | 2018-10-05 15:51:29 -0700 | [diff] [blame] | 1758 | } |
Baolin Wang | a00a875 | 2022-05-13 16:48:56 -0700 | [diff] [blame] | 1759 | pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); |
Baolin Wang | 54205e9 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 1760 | } else { |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 1761 | flush_cache_page(vma, address, pfn); |
David Hildenbrand | 088b8aa | 2022-09-01 10:35:59 +0200 | [diff] [blame] | 1762 | /* Nuke the page table entry. */ |
| 1763 | if (should_defer_flush(mm, flags)) { |
Baolin Wang | a00a875 | 2022-05-13 16:48:56 -0700 | [diff] [blame] | 1764 | /* |
| 1765 | * We clear the PTE but do not flush so potentially |
| 1766 | * a remote CPU could still be writing to the folio. |
| 1767 | * If the entry was previously clean then the |
| 1768 | * architecture must guarantee that a clear->dirty |
| 1769 | * transition on a cached TLB entry is written through |
| 1770 | * and traps if the PTE is unmapped. |
| 1771 | */ |
| 1772 | pteval = ptep_get_and_clear(mm, address, pvmw.pte); |
Mel Gorman | 72b252a | 2015-09-04 15:47:32 -0700 | [diff] [blame] | 1773 | |
Barry Song | f73419b | 2023-07-17 21:10:02 +0800 | [diff] [blame] | 1774 | set_tlb_ubc_flush_pending(mm, pteval, address); |
Baolin Wang | a00a875 | 2022-05-13 16:48:56 -0700 | [diff] [blame] | 1775 | } else { |
| 1776 | pteval = ptep_clear_flush(vma, address, pvmw.pte); |
| 1777 | } |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1778 | } |
| 1779 | |
Peter Xu | 999dad8 | 2022-05-12 20:22:53 -0700 | [diff] [blame] | 1780 | /* |
| 1781 | * Now the pte is cleared. If this pte was uffd-wp armed, |
| 1782 | * we may want to replace a none pte with a marker pte if |
| 1783 | * it's file-backed, so we don't lose the tracking info. |
| 1784 | */ |
| 1785 | pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval); |
| 1786 | |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1787 | /* Set the dirty flag on the folio now the pte is gone. */ |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1788 | if (pte_dirty(pteval)) |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1789 | folio_mark_dirty(folio); |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1790 | |
| 1791 | /* Update high watermark before we lower rss */ |
| 1792 | update_hiwater_rss(mm); |
| 1793 | |
Naoya Horiguchi | 6da6b1d | 2023-02-21 17:59:05 +0900 | [diff] [blame] | 1794 | if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) { |
Punit Agrawal | 5fd27b8 | 2017-07-06 15:39:53 -0700 | [diff] [blame] | 1795 | pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1796 | if (folio_test_hugetlb(folio)) { |
| 1797 | hugetlb_count_sub(folio_nr_pages(folio), mm); |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 1798 | set_huge_pte_at(mm, address, pvmw.pte, pteval, |
| 1799 | hsz); |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1800 | } else { |
Kefeng Wang | a23f517 | 2024-01-11 15:24:28 +0000 | [diff] [blame] | 1801 | dec_mm_counter(mm, mm_counter(folio)); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1802 | set_pte_at(mm, address, pvmw.pte, pteval); |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1803 | } |
| 1804 | |
Christian Borntraeger | bce73e4 | 2018-07-13 16:58:52 -0700 | [diff] [blame] | 1805 | } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1806 | /* |
| 1807 | * The guest indicated that the page content is of no |
| 1808 | * interest anymore. Simply discard the pte, vmscan |
| 1809 | * will take care of the rest. |
Christian Borntraeger | bce73e4 | 2018-07-13 16:58:52 -0700 | [diff] [blame] | 1810 | * A future reference will then fault in a new zero |
| 1811 | * page. When userfaultfd is active, we must not drop |
| 1812 | * this page though, as its main user (postcopy |
| 1813 | * migration) will not expect userfaults on already |
| 1814 | * copied pages. |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1815 | */ |
Kefeng Wang | a23f517 | 2024-01-11 15:24:28 +0000 | [diff] [blame] | 1816 | dec_mm_counter(mm, mm_counter(folio)); |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1817 | } else if (folio_test_anon(folio)) { |
David Hildenbrand | cfeed8f | 2023-08-21 18:08:46 +0200 | [diff] [blame] | 1818 | swp_entry_t entry = page_swap_entry(subpage); |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1819 | pte_t swp_pte; |
| 1820 | /* |
| 1821 | * Store the swap location in the pte. |
| 1822 | * See handle_pte_fault() ... |
| 1823 | */ |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1824 | if (unlikely(folio_test_swapbacked(folio) != |
| 1825 | folio_test_swapcache(folio))) { |
Minchan Kim | eb94a87 | 2017-05-03 14:52:36 -0700 | [diff] [blame] | 1826 | WARN_ON_ONCE(1); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1827 | goto walk_abort; |
Minchan Kim | eb94a87 | 2017-05-03 14:52:36 -0700 | [diff] [blame] | 1828 | } |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1829 | |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1830 | /* MADV_FREE page check */ |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1831 | if (!folio_test_swapbacked(folio)) { |
Mauricio Faria de Oliveira | 6c8e2a2 | 2022-03-24 18:14:09 -0700 | [diff] [blame] | 1832 | int ref_count, map_count; |
| 1833 | |
| 1834 | /* |
| 1835 | * Synchronize with gup_pte_range(): |
| 1836 | * - clear PTE; barrier; read refcount |
| 1837 | * - inc refcount; barrier; read PTE |
| 1838 | */ |
| 1839 | smp_mb(); |
| 1840 | |
| 1841 | ref_count = folio_ref_count(folio); |
| 1842 | map_count = folio_mapcount(folio); |
| 1843 | |
| 1844 | /* |
| 1845 | * Order reads for page refcount and dirty flag |
| 1846 | * (see comments in __remove_mapping()). |
| 1847 | */ |
| 1848 | smp_rmb(); |
| 1849 | |
| 1850 | /* |
| 1851 | * The only page refs must be one from isolation |
| 1852 | * plus the rmap(s) (dropped by discard:). |
| 1853 | */ |
| 1854 | if (ref_count == 1 + map_count && |
Jason A. Donenfeld | 9651fce | 2022-12-08 17:55:04 +0100 | [diff] [blame] | 1855 | (!folio_test_dirty(folio) || |
| 1856 | /* |
| 1857 | * Unlike MADV_FREE mappings, VM_DROPPABLE |
| 1858 | * ones can be dropped even if they've |
| 1859 | * been dirtied. |
| 1860 | */ |
| 1861 | (vma->vm_flags & VM_DROPPABLE))) { |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1862 | dec_mm_counter(mm, MM_ANONPAGES); |
| 1863 | goto discard; |
| 1864 | } |
| 1865 | |
| 1866 | /* |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1867 | * If the folio was redirtied, it cannot be |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1868 | * discarded. Remap the page to page table. |
| 1869 | */ |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1870 | set_pte_at(mm, address, pvmw.pte, pteval); |
Jason A. Donenfeld | 9651fce | 2022-12-08 17:55:04 +0100 | [diff] [blame] | 1871 | /* |
| 1872 | * Unlike MADV_FREE mappings, VM_DROPPABLE ones |
| 1873 | * never get swap backed on failure to drop. |
| 1874 | */ |
| 1875 | if (!(vma->vm_flags & VM_DROPPABLE)) |
| 1876 | folio_set_swapbacked(folio); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1877 | goto walk_abort; |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1878 | } |
| 1879 | |
| 1880 | if (swap_duplicate(entry) < 0) { |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1881 | set_pte_at(mm, address, pvmw.pte, pteval); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1882 | goto walk_abort; |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1883 | } |
Khalid Aziz | ca827d5 | 2018-02-21 10:15:44 -0700 | [diff] [blame] | 1884 | if (arch_unmap_one(mm, vma, address, pteval) < 0) { |
David Hildenbrand | 322842e | 2022-05-09 18:20:42 -0700 | [diff] [blame] | 1885 | swap_free(entry); |
Khalid Aziz | ca827d5 | 2018-02-21 10:15:44 -0700 | [diff] [blame] | 1886 | set_pte_at(mm, address, pvmw.pte, pteval); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1887 | goto walk_abort; |
Khalid Aziz | ca827d5 | 2018-02-21 10:15:44 -0700 | [diff] [blame] | 1888 | } |
David Hildenbrand | 088b8aa | 2022-09-01 10:35:59 +0200 | [diff] [blame] | 1889 | |
David Hildenbrand | e3b4b13 | 2023-12-20 23:45:02 +0100 | [diff] [blame] | 1890 | /* See folio_try_share_anon_rmap(): clear PTE first. */ |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 1891 | if (anon_exclusive && |
David Hildenbrand | e3b4b13 | 2023-12-20 23:45:02 +0100 | [diff] [blame] | 1892 | folio_try_share_anon_rmap_pte(folio, subpage)) { |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 1893 | swap_free(entry); |
| 1894 | set_pte_at(mm, address, pvmw.pte, pteval); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1895 | goto walk_abort; |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 1896 | } |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1897 | if (list_empty(&mm->mmlist)) { |
| 1898 | spin_lock(&mmlist_lock); |
| 1899 | if (list_empty(&mm->mmlist)) |
| 1900 | list_add(&mm->mmlist, &init_mm.mmlist); |
| 1901 | spin_unlock(&mmlist_lock); |
| 1902 | } |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1903 | dec_mm_counter(mm, MM_ANONPAGES); |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1904 | inc_mm_counter(mm, MM_SWAPENTS); |
| 1905 | swp_pte = swp_entry_to_pte(entry); |
David Hildenbrand | 1493a19 | 2022-05-09 18:20:45 -0700 | [diff] [blame] | 1906 | if (anon_exclusive) |
| 1907 | swp_pte = pte_swp_mkexclusive(swp_pte); |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1908 | if (pte_soft_dirty(pteval)) |
| 1909 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
Peter Xu | f45ec5f | 2020-04-06 20:06:01 -0700 | [diff] [blame] | 1910 | if (pte_uffd_wp(pteval)) |
| 1911 | swp_pte = pte_swp_mkuffd_wp(swp_pte); |
Linus Torvalds | 785373b | 2017-08-29 09:11:06 -0700 | [diff] [blame] | 1912 | set_pte_at(mm, address, pvmw.pte, swp_pte); |
Jérôme Glisse | 0f10851 | 2017-11-15 17:34:07 -0800 | [diff] [blame] | 1913 | } else { |
| 1914 | /* |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1915 | * This is a locked file-backed folio, |
| 1916 | * so it cannot be removed from the page |
| 1917 | * cache and replaced by a new folio before |
| 1918 | * mmu_notifier_invalidate_range_end, so no |
| 1919 | * concurrent thread might update its page table |
| 1920 | * to point at a new folio while a device is |
| 1921 | * still using this folio. |
Jérôme Glisse | 0f10851 | 2017-11-15 17:34:07 -0800 | [diff] [blame] | 1922 | * |
Mike Rapoport | ee65728 | 2022-06-27 09:00:26 +0300 | [diff] [blame] | 1923 | * See Documentation/mm/mmu_notifier.rst |
Jérôme Glisse | 0f10851 | 2017-11-15 17:34:07 -0800 | [diff] [blame] | 1924 | */ |
Kefeng Wang | 6b27cc6c | 2024-01-11 15:24:29 +0000 | [diff] [blame] | 1925 | dec_mm_counter(mm, mm_counter_file(folio)); |
Jérôme Glisse | 0f10851 | 2017-11-15 17:34:07 -0800 | [diff] [blame] | 1926 | } |
Minchan Kim | 854e9ed | 2016-01-15 16:54:53 -0800 | [diff] [blame] | 1927 | discard: |
David Hildenbrand | e135826 | 2023-12-20 23:44:26 +0100 | [diff] [blame] | 1928 | if (unlikely(folio_test_hugetlb(folio))) |
| 1929 | hugetlb_remove_rmap(folio); |
| 1930 | else |
David Hildenbrand | ca1a074 | 2023-12-20 23:44:54 +0100 | [diff] [blame] | 1931 | folio_remove_rmap_pte(folio, subpage, vma); |
Hugh Dickins | b743550 | 2022-02-14 18:38:47 -0800 | [diff] [blame] | 1932 | if (vma->vm_flags & VM_LOCKED) |
Lorenzo Stoakes | 96f97c4 | 2023-01-12 12:39:31 +0000 | [diff] [blame] | 1933 | mlock_drain_local(); |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1934 | folio_put(folio); |
Lance Yang | 26d21b1 | 2024-06-14 09:51:36 +0800 | [diff] [blame] | 1935 | continue; |
| 1936 | walk_abort: |
| 1937 | ret = false; |
| 1938 | walk_done: |
| 1939 | page_vma_mapped_walk_done(&pvmw); |
| 1940 | break; |
Kirill A. Shutemov | c7ab0d2 | 2017-02-24 14:58:01 -0800 | [diff] [blame] | 1941 | } |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1942 | |
Jérôme Glisse | ac46d4f | 2018-12-28 00:38:09 -0800 | [diff] [blame] | 1943 | mmu_notifier_invalidate_range_end(&range); |
Jérôme Glisse | 369ea82 | 2017-08-31 17:17:27 -0400 | [diff] [blame] | 1944 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1945 | return ret; |
| 1946 | } |
| 1947 | |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1948 | static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) |
| 1949 | { |
Anshuman Khandual | 222100e | 2020-04-01 21:07:52 -0700 | [diff] [blame] | 1950 | return vma_is_temporary_stack(vma); |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1951 | } |
| 1952 | |
Kefeng Wang | f3ad032 | 2022-09-27 14:38:26 +0800 | [diff] [blame] | 1953 | static int folio_not_mapped(struct folio *folio) |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1954 | { |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1955 | return !folio_mapped(folio); |
Kirill A. Shutemov | 2a52bcb | 2016-03-17 14:20:04 -0700 | [diff] [blame] | 1956 | } |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1957 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1958 | /** |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1959 | * try_to_unmap - Try to remove all page table mappings to a folio. |
| 1960 | * @folio: The folio to unmap. |
Andi Kleen | 14fa31b | 2009-09-16 11:50:10 +0200 | [diff] [blame] | 1961 | * @flags: action and flags |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1962 | * |
| 1963 | * Tries to remove all the page table entries which are mapping this |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1964 | * folio. It is the caller's responsibility to check if the folio is |
| 1965 | * still mapped if needed (use TTU_SYNC to prevent accounting races). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1966 | * |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1967 | * Context: Caller must hold the folio lock. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1968 | */ |
Matthew Wilcox (Oracle) | 869f7ee | 2022-02-15 09:28:49 -0500 | [diff] [blame] | 1969 | void try_to_unmap(struct folio *folio, enum ttu_flags flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1970 | { |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1971 | struct rmap_walk_control rwc = { |
| 1972 | .rmap_one = try_to_unmap_one, |
Shaohua Li | 802a3a9 | 2017-05-03 14:52:32 -0700 | [diff] [blame] | 1973 | .arg = (void *)flags, |
Kefeng Wang | f3ad032 | 2022-09-27 14:38:26 +0800 | [diff] [blame] | 1974 | .done = folio_not_mapped, |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1975 | .anon_lock = folio_lock_anon_vma_read, |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 1976 | }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1977 | |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1978 | if (flags & TTU_RMAP_LOCKED) |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1979 | rmap_walk_locked(folio, &rwc); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1980 | else |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1981 | rmap_walk(folio, &rwc); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1982 | } |
| 1983 | |
| 1984 | /* |
| 1985 | * @arg: enum ttu_flags will be passed to this argument. |
| 1986 | * |
| 1987 | * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs |
Hugh Dickins | 64b586d | 2021-07-07 13:06:17 -0700 | [diff] [blame] | 1988 | * containing migration entries. |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1989 | */ |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 1990 | static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma, |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1991 | unsigned long address, void *arg) |
| 1992 | { |
| 1993 | struct mm_struct *mm = vma->vm_mm; |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 1994 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1995 | pte_t pteval; |
| 1996 | struct page *subpage; |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 1997 | bool anon_exclusive, ret = true; |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 1998 | struct mmu_notifier_range range; |
| 1999 | enum ttu_flags flags = (enum ttu_flags)(long)arg; |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2000 | unsigned long pfn; |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 2001 | unsigned long hsz = 0; |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2002 | |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2003 | /* |
| 2004 | * When racing against e.g. zap_pte_range() on another cpu, |
David Hildenbrand | ca1a074 | 2023-12-20 23:44:54 +0100 | [diff] [blame] | 2005 | * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(), |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2006 | * try_to_migrate() may return before page_mapped() has become false, |
| 2007 | * if page table locking is skipped: use TTU_SYNC to wait for that. |
| 2008 | */ |
| 2009 | if (flags & TTU_SYNC) |
| 2010 | pvmw.flags = PVMW_SYNC; |
| 2011 | |
| 2012 | /* |
| 2013 | * unmap_page() in mm/huge_memory.c is the only user of migration with |
| 2014 | * TTU_SPLIT_HUGE_PMD and it wants to freeze. |
| 2015 | */ |
| 2016 | if (flags & TTU_SPLIT_HUGE_PMD) |
Matthew Wilcox (Oracle) | af28a98 | 2022-01-21 10:44:52 -0500 | [diff] [blame] | 2017 | split_huge_pmd_address(vma, address, true, folio); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2018 | |
| 2019 | /* |
| 2020 | * For THP, we have to assume the worse case ie pmd for invalidation. |
| 2021 | * For hugetlb, it could be much worse if we need to do pud |
| 2022 | * invalidation in the case of pmd sharing. |
| 2023 | * |
| 2024 | * Note that the page can not be free in this function as call of |
| 2025 | * try_to_unmap() must hold a reference on the page. |
| 2026 | */ |
Matthew Wilcox (Oracle) | 2aff7a4 | 2022-02-03 11:40:17 -0500 | [diff] [blame] | 2027 | range.end = vma_address_end(&pvmw); |
Alistair Popple | 7d4a8be | 2023-01-10 13:57:22 +1100 | [diff] [blame] | 2028 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2029 | address, range.end); |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2030 | if (folio_test_hugetlb(folio)) { |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2031 | /* |
| 2032 | * If sharing is possible, start and end will be adjusted |
| 2033 | * accordingly. |
| 2034 | */ |
| 2035 | adjust_range_if_pmd_sharing_possible(vma, &range.start, |
| 2036 | &range.end); |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 2037 | |
| 2038 | /* We need the huge page size for set_huge_pte_at() */ |
| 2039 | hsz = huge_page_size(hstate_vma(vma)); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2040 | } |
| 2041 | mmu_notifier_invalidate_range_start(&range); |
| 2042 | |
| 2043 | while (page_vma_mapped_walk(&pvmw)) { |
| 2044 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
| 2045 | /* PMD-mapped THP migration entry */ |
| 2046 | if (!pvmw.pte) { |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2047 | subpage = folio_page(folio, |
| 2048 | pmd_pfn(*pvmw.pmd) - folio_pfn(folio)); |
| 2049 | VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) || |
| 2050 | !folio_test_pmd_mappable(folio), folio); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2051 | |
David Hildenbrand | 7f5abe6 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2052 | if (set_pmd_migration_entry(&pvmw, subpage)) { |
| 2053 | ret = false; |
| 2054 | page_vma_mapped_walk_done(&pvmw); |
| 2055 | break; |
| 2056 | } |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2057 | continue; |
| 2058 | } |
| 2059 | #endif |
| 2060 | |
| 2061 | /* Unexpected PMD-mapped THP? */ |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2062 | VM_BUG_ON_FOLIO(!pvmw.pte, folio); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2063 | |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2064 | pfn = pte_pfn(ptep_get(pvmw.pte)); |
| 2065 | |
David Hildenbrand | 1118234 | 2022-06-23 22:53:32 +0200 | [diff] [blame] | 2066 | if (folio_is_zone_device(folio)) { |
| 2067 | /* |
| 2068 | * Our PTE is a non-present device exclusive entry and |
| 2069 | * calculating the subpage as for the common case would |
| 2070 | * result in an invalid pointer. |
| 2071 | * |
| 2072 | * Since only PAGE_SIZE pages can currently be |
| 2073 | * migrated, just set it to page. This will need to be |
| 2074 | * changed when hugepage migrations to device private |
| 2075 | * memory are supported. |
| 2076 | */ |
| 2077 | VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio); |
| 2078 | subpage = &folio->page; |
| 2079 | } else { |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2080 | subpage = folio_page(folio, pfn - folio_pfn(folio)); |
David Hildenbrand | 1118234 | 2022-06-23 22:53:32 +0200 | [diff] [blame] | 2081 | } |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2082 | address = pvmw.address; |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2083 | anon_exclusive = folio_test_anon(folio) && |
| 2084 | PageAnonExclusive(subpage); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2085 | |
Baolin Wang | dfc7ab5 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 2086 | if (folio_test_hugetlb(folio)) { |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 2087 | bool anon = folio_test_anon(folio); |
| 2088 | |
Baolin Wang | 54205e9 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 2089 | /* |
| 2090 | * huge_pmd_unshare may unmap an entire PMD page. |
| 2091 | * There is no way of knowing exactly which PMDs may |
| 2092 | * be cached for this mm, so we must flush them all. |
| 2093 | * start/end were already adjusted above to cover this |
| 2094 | * range. |
| 2095 | */ |
| 2096 | flush_cache_range(vma, range.start, range.end); |
| 2097 | |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 2098 | /* |
| 2099 | * To call huge_pmd_unshare, i_mmap_rwsem must be |
| 2100 | * held in write mode. Caller needs to explicitly |
| 2101 | * do this outside rmap routines. |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 2102 | * |
| 2103 | * We also must hold hugetlb vma_lock in write mode. |
| 2104 | * Lock order dictates acquiring vma_lock BEFORE |
| 2105 | * i_mmap_rwsem. We can only try lock here and |
| 2106 | * fail if unsuccessful. |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 2107 | */ |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 2108 | if (!anon) { |
| 2109 | VM_BUG_ON(!(flags & TTU_RMAP_LOCKED)); |
| 2110 | if (!hugetlb_vma_trylock_write(vma)) { |
| 2111 | page_vma_mapped_walk_done(&pvmw); |
| 2112 | ret = false; |
| 2113 | break; |
| 2114 | } |
| 2115 | if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) { |
| 2116 | hugetlb_vma_unlock_write(vma); |
| 2117 | flush_tlb_range(vma, |
| 2118 | range.start, range.end); |
Baolin Wang | 0506c31 | 2022-06-20 19:47:15 +0800 | [diff] [blame] | 2119 | |
Mike Kravetz | 40549ba | 2022-09-14 15:18:09 -0700 | [diff] [blame] | 2120 | /* |
| 2121 | * The ref count of the PMD page was |
| 2122 | * dropped which is part of the way map |
| 2123 | * counting is done for shared PMDs. |
| 2124 | * Return 'true' here. When there is |
| 2125 | * no other sharing, huge_pmd_unshare |
| 2126 | * returns false and we will unmap the |
| 2127 | * actual page and drop map count |
| 2128 | * to zero. |
| 2129 | */ |
| 2130 | page_vma_mapped_walk_done(&pvmw); |
| 2131 | break; |
| 2132 | } |
| 2133 | hugetlb_vma_unlock_write(vma); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2134 | } |
Baolin Wang | 5d4af61 | 2022-05-13 16:48:55 -0700 | [diff] [blame] | 2135 | /* Nuke the hugetlb page table entry */ |
| 2136 | pteval = huge_ptep_clear_flush(vma, address, pvmw.pte); |
Baolin Wang | 54205e9 | 2022-05-09 18:20:53 -0700 | [diff] [blame] | 2137 | } else { |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2138 | flush_cache_page(vma, address, pfn); |
Baolin Wang | 5d4af61 | 2022-05-13 16:48:55 -0700 | [diff] [blame] | 2139 | /* Nuke the page table entry. */ |
Huang Ying | 7e12beb | 2023-02-13 20:34:43 +0800 | [diff] [blame] | 2140 | if (should_defer_flush(mm, flags)) { |
| 2141 | /* |
| 2142 | * We clear the PTE but do not flush so potentially |
| 2143 | * a remote CPU could still be writing to the folio. |
| 2144 | * If the entry was previously clean then the |
| 2145 | * architecture must guarantee that a clear->dirty |
| 2146 | * transition on a cached TLB entry is written through |
| 2147 | * and traps if the PTE is unmapped. |
| 2148 | */ |
| 2149 | pteval = ptep_get_and_clear(mm, address, pvmw.pte); |
| 2150 | |
Barry Song | f73419b | 2023-07-17 21:10:02 +0800 | [diff] [blame] | 2151 | set_tlb_ubc_flush_pending(mm, pteval, address); |
Huang Ying | 7e12beb | 2023-02-13 20:34:43 +0800 | [diff] [blame] | 2152 | } else { |
| 2153 | pteval = ptep_clear_flush(vma, address, pvmw.pte); |
| 2154 | } |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2155 | } |
| 2156 | |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2157 | /* Set the dirty flag on the folio now the pte is gone. */ |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2158 | if (pte_dirty(pteval)) |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2159 | folio_mark_dirty(folio); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2160 | |
| 2161 | /* Update high watermark before we lower rss */ |
| 2162 | update_hiwater_rss(mm); |
| 2163 | |
Alex Sierra | f25cbb7 | 2022-07-15 10:05:10 -0500 | [diff] [blame] | 2164 | if (folio_is_device_private(folio)) { |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2165 | unsigned long pfn = folio_pfn(folio); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2166 | swp_entry_t entry; |
| 2167 | pte_t swp_pte; |
| 2168 | |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2169 | if (anon_exclusive) |
David Hildenbrand | e3b4b13 | 2023-12-20 23:45:02 +0100 | [diff] [blame] | 2170 | WARN_ON_ONCE(folio_try_share_anon_rmap_pte(folio, |
| 2171 | subpage)); |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2172 | |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2173 | /* |
| 2174 | * Store the pfn of the page in a special migration |
| 2175 | * pte. do_swap_page() will wait until the migration |
| 2176 | * pte is removed and then restart fault handling. |
| 2177 | */ |
Alistair Popple | 3d88705 | 2021-11-05 13:45:00 -0700 | [diff] [blame] | 2178 | entry = pte_to_swp_entry(pteval); |
| 2179 | if (is_writable_device_private_entry(entry)) |
| 2180 | entry = make_writable_migration_entry(pfn); |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2181 | else if (anon_exclusive) |
| 2182 | entry = make_readable_exclusive_migration_entry(pfn); |
Alistair Popple | 3d88705 | 2021-11-05 13:45:00 -0700 | [diff] [blame] | 2183 | else |
| 2184 | entry = make_readable_migration_entry(pfn); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2185 | swp_pte = swp_entry_to_pte(entry); |
| 2186 | |
| 2187 | /* |
| 2188 | * pteval maps a zone device page and is therefore |
| 2189 | * a swap pte. |
| 2190 | */ |
| 2191 | if (pte_swp_soft_dirty(pteval)) |
| 2192 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
| 2193 | if (pte_swp_uffd_wp(pteval)) |
| 2194 | swp_pte = pte_swp_mkuffd_wp(swp_pte); |
| 2195 | set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte); |
Anshuman Khandual | 4cc79b3 | 2022-03-24 18:10:01 -0700 | [diff] [blame] | 2196 | trace_set_migration_pte(pvmw.address, pte_val(swp_pte), |
Matthew Wilcox (Oracle) | 059ab7be | 2024-02-15 20:53:05 +0000 | [diff] [blame] | 2197 | folio_order(folio)); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2198 | /* |
| 2199 | * No need to invalidate here it will synchronize on |
| 2200 | * against the special swap migration pte. |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2201 | */ |
Matthew Wilcox (Oracle) | da358d5 | 2022-03-22 14:46:38 -0700 | [diff] [blame] | 2202 | } else if (PageHWPoison(subpage)) { |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2203 | pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2204 | if (folio_test_hugetlb(folio)) { |
| 2205 | hugetlb_count_sub(folio_nr_pages(folio), mm); |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 2206 | set_huge_pte_at(mm, address, pvmw.pte, pteval, |
| 2207 | hsz); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2208 | } else { |
Kefeng Wang | a23f517 | 2024-01-11 15:24:28 +0000 | [diff] [blame] | 2209 | dec_mm_counter(mm, mm_counter(folio)); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2210 | set_pte_at(mm, address, pvmw.pte, pteval); |
| 2211 | } |
| 2212 | |
| 2213 | } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { |
| 2214 | /* |
| 2215 | * The guest indicated that the page content is of no |
| 2216 | * interest anymore. Simply discard the pte, vmscan |
| 2217 | * will take care of the rest. |
| 2218 | * A future reference will then fault in a new zero |
| 2219 | * page. When userfaultfd is active, we must not drop |
| 2220 | * this page though, as its main user (postcopy |
| 2221 | * migration) will not expect userfaults on already |
| 2222 | * copied pages. |
| 2223 | */ |
Kefeng Wang | a23f517 | 2024-01-11 15:24:28 +0000 | [diff] [blame] | 2224 | dec_mm_counter(mm, mm_counter(folio)); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2225 | } else { |
| 2226 | swp_entry_t entry; |
| 2227 | pte_t swp_pte; |
| 2228 | |
| 2229 | if (arch_unmap_one(mm, vma, address, pteval) < 0) { |
Baolin Wang | 5d4af61 | 2022-05-13 16:48:55 -0700 | [diff] [blame] | 2230 | if (folio_test_hugetlb(folio)) |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 2231 | set_huge_pte_at(mm, address, pvmw.pte, |
| 2232 | pteval, hsz); |
Baolin Wang | 5d4af61 | 2022-05-13 16:48:55 -0700 | [diff] [blame] | 2233 | else |
| 2234 | set_pte_at(mm, address, pvmw.pte, pteval); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2235 | ret = false; |
| 2236 | page_vma_mapped_walk_done(&pvmw); |
| 2237 | break; |
| 2238 | } |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2239 | VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) && |
| 2240 | !anon_exclusive, subpage); |
David Hildenbrand | 088b8aa | 2022-09-01 10:35:59 +0200 | [diff] [blame] | 2241 | |
David Hildenbrand | e3b4b13 | 2023-12-20 23:45:02 +0100 | [diff] [blame] | 2242 | /* See folio_try_share_anon_rmap_pte(): clear PTE first. */ |
David Hildenbrand | 0c2ec32 | 2023-12-20 23:44:29 +0100 | [diff] [blame] | 2243 | if (folio_test_hugetlb(folio)) { |
| 2244 | if (anon_exclusive && |
| 2245 | hugetlb_try_share_anon_rmap(folio)) { |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 2246 | set_huge_pte_at(mm, address, pvmw.pte, |
| 2247 | pteval, hsz); |
David Hildenbrand | 0c2ec32 | 2023-12-20 23:44:29 +0100 | [diff] [blame] | 2248 | ret = false; |
| 2249 | page_vma_mapped_walk_done(&pvmw); |
| 2250 | break; |
| 2251 | } |
| 2252 | } else if (anon_exclusive && |
David Hildenbrand | e3b4b13 | 2023-12-20 23:45:02 +0100 | [diff] [blame] | 2253 | folio_try_share_anon_rmap_pte(folio, subpage)) { |
David Hildenbrand | 0c2ec32 | 2023-12-20 23:44:29 +0100 | [diff] [blame] | 2254 | set_pte_at(mm, address, pvmw.pte, pteval); |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2255 | ret = false; |
| 2256 | page_vma_mapped_walk_done(&pvmw); |
| 2257 | break; |
| 2258 | } |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2259 | |
| 2260 | /* |
| 2261 | * Store the pfn of the page in a special migration |
| 2262 | * pte. do_swap_page() will wait until the migration |
| 2263 | * pte is removed and then restart fault handling. |
| 2264 | */ |
| 2265 | if (pte_write(pteval)) |
| 2266 | entry = make_writable_migration_entry( |
| 2267 | page_to_pfn(subpage)); |
David Hildenbrand | 6c28760 | 2022-05-09 18:20:44 -0700 | [diff] [blame] | 2268 | else if (anon_exclusive) |
| 2269 | entry = make_readable_exclusive_migration_entry( |
| 2270 | page_to_pfn(subpage)); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2271 | else |
| 2272 | entry = make_readable_migration_entry( |
| 2273 | page_to_pfn(subpage)); |
Peter Xu | 2e34687 | 2022-08-11 12:13:29 -0400 | [diff] [blame] | 2274 | if (pte_young(pteval)) |
| 2275 | entry = make_migration_entry_young(entry); |
| 2276 | if (pte_dirty(pteval)) |
| 2277 | entry = make_migration_entry_dirty(entry); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2278 | swp_pte = swp_entry_to_pte(entry); |
| 2279 | if (pte_soft_dirty(pteval)) |
| 2280 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
| 2281 | if (pte_uffd_wp(pteval)) |
| 2282 | swp_pte = pte_swp_mkuffd_wp(swp_pte); |
Baolin Wang | 5d4af61 | 2022-05-13 16:48:55 -0700 | [diff] [blame] | 2283 | if (folio_test_hugetlb(folio)) |
Ryan Roberts | 935d4f0 | 2023-09-22 12:58:03 +0100 | [diff] [blame] | 2284 | set_huge_pte_at(mm, address, pvmw.pte, swp_pte, |
| 2285 | hsz); |
Baolin Wang | 5d4af61 | 2022-05-13 16:48:55 -0700 | [diff] [blame] | 2286 | else |
| 2287 | set_pte_at(mm, address, pvmw.pte, swp_pte); |
Anshuman Khandual | 4cc79b3 | 2022-03-24 18:10:01 -0700 | [diff] [blame] | 2288 | trace_set_migration_pte(address, pte_val(swp_pte), |
Matthew Wilcox (Oracle) | 059ab7be | 2024-02-15 20:53:05 +0000 | [diff] [blame] | 2289 | folio_order(folio)); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2290 | /* |
| 2291 | * No need to invalidate here it will synchronize on |
| 2292 | * against the special swap migration pte. |
| 2293 | */ |
| 2294 | } |
| 2295 | |
David Hildenbrand | e135826 | 2023-12-20 23:44:26 +0100 | [diff] [blame] | 2296 | if (unlikely(folio_test_hugetlb(folio))) |
| 2297 | hugetlb_remove_rmap(folio); |
| 2298 | else |
David Hildenbrand | ca1a074 | 2023-12-20 23:44:54 +0100 | [diff] [blame] | 2299 | folio_remove_rmap_pte(folio, subpage, vma); |
Hugh Dickins | b743550 | 2022-02-14 18:38:47 -0800 | [diff] [blame] | 2300 | if (vma->vm_flags & VM_LOCKED) |
Lorenzo Stoakes | 96f97c4 | 2023-01-12 12:39:31 +0000 | [diff] [blame] | 2301 | mlock_drain_local(); |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2302 | folio_put(folio); |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2303 | } |
| 2304 | |
| 2305 | mmu_notifier_invalidate_range_end(&range); |
| 2306 | |
| 2307 | return ret; |
| 2308 | } |
| 2309 | |
| 2310 | /** |
| 2311 | * try_to_migrate - try to replace all page table mappings with swap entries |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2312 | * @folio: the folio to replace page table entries for |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2313 | * @flags: action and flags |
| 2314 | * |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2315 | * Tries to remove all the page table entries which are mapping this folio and |
| 2316 | * replace them with special swap entries. Caller must hold the folio lock. |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2317 | */ |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2318 | void try_to_migrate(struct folio *folio, enum ttu_flags flags) |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2319 | { |
| 2320 | struct rmap_walk_control rwc = { |
| 2321 | .rmap_one = try_to_migrate_one, |
| 2322 | .arg = (void *)flags, |
Kefeng Wang | f3ad032 | 2022-09-27 14:38:26 +0800 | [diff] [blame] | 2323 | .done = folio_not_mapped, |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2324 | .anon_lock = folio_lock_anon_vma_read, |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2325 | }; |
| 2326 | |
| 2327 | /* |
| 2328 | * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and |
Huang Ying | 7e12beb | 2023-02-13 20:34:43 +0800 | [diff] [blame] | 2329 | * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags. |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2330 | */ |
| 2331 | if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD | |
Huang Ying | 7e12beb | 2023-02-13 20:34:43 +0800 | [diff] [blame] | 2332 | TTU_SYNC | TTU_BATCH_FLUSH))) |
Alistair Popple | a98a2f0 | 2021-06-30 18:54:16 -0700 | [diff] [blame] | 2333 | return; |
| 2334 | |
Alex Sierra | f25cbb7 | 2022-07-15 10:05:10 -0500 | [diff] [blame] | 2335 | if (folio_is_zone_device(folio) && |
| 2336 | (!folio_is_device_private(folio) && !folio_is_device_coherent(folio))) |
Hugh Dickins | 6c855fc | 2021-07-07 13:13:33 -0700 | [diff] [blame] | 2337 | return; |
| 2338 | |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 2339 | /* |
| 2340 | * During exec, a temporary VMA is setup and later moved. |
| 2341 | * The VMA is moved under the anon_vma lock but not the |
| 2342 | * page tables leading to a race where migration cannot |
| 2343 | * find the migration ptes. Rather than increasing the |
| 2344 | * locking requirements of exec(), migration skips |
| 2345 | * temporary VMAs until after exec() completes. |
| 2346 | */ |
Matthew Wilcox (Oracle) | 4b8554c | 2022-01-28 14:29:43 -0500 | [diff] [blame] | 2347 | if (!folio_test_ksm(folio) && folio_test_anon(folio)) |
Joonsoo Kim | 5262950 | 2014-01-21 15:49:50 -0800 | [diff] [blame] | 2348 | rwc.invalid_vma = invalid_migration_vma; |
| 2349 | |
Kirill A. Shutemov | 2a52bcb | 2016-03-17 14:20:04 -0700 | [diff] [blame] | 2350 | if (flags & TTU_RMAP_LOCKED) |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2351 | rmap_walk_locked(folio, &rwc); |
Kirill A. Shutemov | 2a52bcb | 2016-03-17 14:20:04 -0700 | [diff] [blame] | 2352 | else |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2353 | rmap_walk(folio, &rwc); |
Nick Piggin | b291f00 | 2008-10-18 20:26:44 -0700 | [diff] [blame] | 2354 | } |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2355 | |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2356 | #ifdef CONFIG_DEVICE_PRIVATE |
| 2357 | struct make_exclusive_args { |
| 2358 | struct mm_struct *mm; |
| 2359 | unsigned long address; |
| 2360 | void *owner; |
| 2361 | bool valid; |
| 2362 | }; |
| 2363 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2364 | static bool page_make_device_exclusive_one(struct folio *folio, |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2365 | struct vm_area_struct *vma, unsigned long address, void *priv) |
| 2366 | { |
| 2367 | struct mm_struct *mm = vma->vm_mm; |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2368 | DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2369 | struct make_exclusive_args *args = priv; |
| 2370 | pte_t pteval; |
| 2371 | struct page *subpage; |
| 2372 | bool ret = true; |
| 2373 | struct mmu_notifier_range range; |
| 2374 | swp_entry_t entry; |
| 2375 | pte_t swp_pte; |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2376 | pte_t ptent; |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2377 | |
Alistair Popple | 7d4a8be | 2023-01-10 13:57:22 +1100 | [diff] [blame] | 2378 | mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2379 | vma->vm_mm, address, min(vma->vm_end, |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2380 | address + folio_size(folio)), |
| 2381 | args->owner); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2382 | mmu_notifier_invalidate_range_start(&range); |
| 2383 | |
| 2384 | while (page_vma_mapped_walk(&pvmw)) { |
| 2385 | /* Unexpected PMD-mapped THP? */ |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2386 | VM_BUG_ON_FOLIO(!pvmw.pte, folio); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2387 | |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2388 | ptent = ptep_get(pvmw.pte); |
| 2389 | if (!pte_present(ptent)) { |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2390 | ret = false; |
| 2391 | page_vma_mapped_walk_done(&pvmw); |
| 2392 | break; |
| 2393 | } |
| 2394 | |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2395 | subpage = folio_page(folio, |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2396 | pte_pfn(ptent) - folio_pfn(folio)); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2397 | address = pvmw.address; |
| 2398 | |
| 2399 | /* Nuke the page table entry. */ |
Ryan Roberts | c33c794 | 2023-06-12 16:15:45 +0100 | [diff] [blame] | 2400 | flush_cache_page(vma, address, pte_pfn(ptent)); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2401 | pteval = ptep_clear_flush(vma, address, pvmw.pte); |
| 2402 | |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2403 | /* Set the dirty flag on the folio now the pte is gone. */ |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2404 | if (pte_dirty(pteval)) |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2405 | folio_mark_dirty(folio); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2406 | |
| 2407 | /* |
| 2408 | * Check that our target page is still mapped at the expected |
| 2409 | * address. |
| 2410 | */ |
| 2411 | if (args->mm == mm && args->address == address && |
| 2412 | pte_write(pteval)) |
| 2413 | args->valid = true; |
| 2414 | |
| 2415 | /* |
| 2416 | * Store the pfn of the page in a special migration |
| 2417 | * pte. do_swap_page() will wait until the migration |
| 2418 | * pte is removed and then restart fault handling. |
| 2419 | */ |
| 2420 | if (pte_write(pteval)) |
| 2421 | entry = make_writable_device_exclusive_entry( |
| 2422 | page_to_pfn(subpage)); |
| 2423 | else |
| 2424 | entry = make_readable_device_exclusive_entry( |
| 2425 | page_to_pfn(subpage)); |
| 2426 | swp_pte = swp_entry_to_pte(entry); |
| 2427 | if (pte_soft_dirty(pteval)) |
| 2428 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
| 2429 | if (pte_uffd_wp(pteval)) |
| 2430 | swp_pte = pte_swp_mkuffd_wp(swp_pte); |
| 2431 | |
| 2432 | set_pte_at(mm, address, pvmw.pte, swp_pte); |
| 2433 | |
| 2434 | /* |
| 2435 | * There is a reference on the page for the swap entry which has |
| 2436 | * been removed, so shouldn't take another. |
| 2437 | */ |
David Hildenbrand | ca1a074 | 2023-12-20 23:44:54 +0100 | [diff] [blame] | 2438 | folio_remove_rmap_pte(folio, subpage, vma); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2439 | } |
| 2440 | |
| 2441 | mmu_notifier_invalidate_range_end(&range); |
| 2442 | |
| 2443 | return ret; |
| 2444 | } |
| 2445 | |
| 2446 | /** |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2447 | * folio_make_device_exclusive - Mark the folio exclusively owned by a device. |
| 2448 | * @folio: The folio to replace page table entries for. |
| 2449 | * @mm: The mm_struct where the folio is expected to be mapped. |
| 2450 | * @address: Address where the folio is expected to be mapped. |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2451 | * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks |
| 2452 | * |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2453 | * Tries to remove all the page table entries which are mapping this |
| 2454 | * folio and replace them with special device exclusive swap entries to |
| 2455 | * grant a device exclusive access to the folio. |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2456 | * |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2457 | * Context: Caller must hold the folio lock. |
| 2458 | * Return: false if the page is still mapped, or if it could not be unmapped |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2459 | * from the expected address. Otherwise returns true (success). |
| 2460 | */ |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2461 | static bool folio_make_device_exclusive(struct folio *folio, |
| 2462 | struct mm_struct *mm, unsigned long address, void *owner) |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2463 | { |
| 2464 | struct make_exclusive_args args = { |
| 2465 | .mm = mm, |
| 2466 | .address = address, |
| 2467 | .owner = owner, |
| 2468 | .valid = false, |
| 2469 | }; |
| 2470 | struct rmap_walk_control rwc = { |
| 2471 | .rmap_one = page_make_device_exclusive_one, |
Kefeng Wang | f3ad032 | 2022-09-27 14:38:26 +0800 | [diff] [blame] | 2472 | .done = folio_not_mapped, |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2473 | .anon_lock = folio_lock_anon_vma_read, |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2474 | .arg = &args, |
| 2475 | }; |
| 2476 | |
| 2477 | /* |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2478 | * Restrict to anonymous folios for now to avoid potential writeback |
| 2479 | * issues. |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2480 | */ |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2481 | if (!folio_test_anon(folio)) |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2482 | return false; |
| 2483 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2484 | rmap_walk(folio, &rwc); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2485 | |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2486 | return args.valid && !folio_mapcount(folio); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2487 | } |
| 2488 | |
| 2489 | /** |
| 2490 | * make_device_exclusive_range() - Mark a range for exclusive use by a device |
Adrian Huang | dd06230 | 2022-05-09 18:20:54 -0700 | [diff] [blame] | 2491 | * @mm: mm_struct of associated target process |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2492 | * @start: start of the region to mark for exclusive device access |
| 2493 | * @end: end address of region |
| 2494 | * @pages: returns the pages which were successfully marked for exclusive access |
| 2495 | * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering |
| 2496 | * |
| 2497 | * Returns: number of pages found in the range by GUP. A page is marked for |
| 2498 | * exclusive access only if the page pointer is non-NULL. |
| 2499 | * |
| 2500 | * This function finds ptes mapping page(s) to the given address range, locks |
| 2501 | * them and replaces mappings with special swap entries preventing userspace CPU |
| 2502 | * access. On fault these entries are replaced with the original mapping after |
| 2503 | * calling MMU notifiers. |
| 2504 | * |
| 2505 | * A driver using this to program access from a device must use a mmu notifier |
| 2506 | * critical section to hold a device specific lock during programming. Once |
| 2507 | * programming is complete it should drop the page lock and reference after |
| 2508 | * which point CPU access to the page will revoke the exclusive access. |
| 2509 | */ |
| 2510 | int make_device_exclusive_range(struct mm_struct *mm, unsigned long start, |
| 2511 | unsigned long end, struct page **pages, |
| 2512 | void *owner) |
| 2513 | { |
| 2514 | long npages = (end - start) >> PAGE_SHIFT; |
| 2515 | long i; |
| 2516 | |
| 2517 | npages = get_user_pages_remote(mm, start, npages, |
| 2518 | FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD, |
Lorenzo Stoakes | ca5e863 | 2023-05-17 20:25:39 +0100 | [diff] [blame] | 2519 | pages, NULL); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2520 | if (npages < 0) |
| 2521 | return npages; |
| 2522 | |
| 2523 | for (i = 0; i < npages; i++, start += PAGE_SIZE) { |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2524 | struct folio *folio = page_folio(pages[i]); |
| 2525 | if (PageTail(pages[i]) || !folio_trylock(folio)) { |
| 2526 | folio_put(folio); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2527 | pages[i] = NULL; |
| 2528 | continue; |
| 2529 | } |
| 2530 | |
Matthew Wilcox (Oracle) | 0d25148 | 2022-01-28 16:03:42 -0500 | [diff] [blame] | 2531 | if (!folio_make_device_exclusive(folio, mm, start, owner)) { |
| 2532 | folio_unlock(folio); |
| 2533 | folio_put(folio); |
Alistair Popple | b756a3b | 2021-06-30 18:54:25 -0700 | [diff] [blame] | 2534 | pages[i] = NULL; |
| 2535 | } |
| 2536 | } |
| 2537 | |
| 2538 | return npages; |
| 2539 | } |
| 2540 | EXPORT_SYMBOL_GPL(make_device_exclusive_range); |
| 2541 | #endif |
| 2542 | |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 2543 | void __put_anon_vma(struct anon_vma *anon_vma) |
Rik van Riel | 7654506 | 2010-08-09 17:18:41 -0700 | [diff] [blame] | 2544 | { |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 2545 | struct anon_vma *root = anon_vma->root; |
Rik van Riel | 7654506 | 2010-08-09 17:18:41 -0700 | [diff] [blame] | 2546 | |
Andrey Ryabinin | 624483f | 2014-06-06 19:09:30 +0400 | [diff] [blame] | 2547 | anon_vma_free(anon_vma); |
Peter Zijlstra | 01d8b20 | 2011-03-22 16:32:49 -0700 | [diff] [blame] | 2548 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) |
| 2549 | anon_vma_free(root); |
Rik van Riel | 7654506 | 2010-08-09 17:18:41 -0700 | [diff] [blame] | 2550 | } |
Rik van Riel | 7654506 | 2010-08-09 17:18:41 -0700 | [diff] [blame] | 2551 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2552 | static struct anon_vma *rmap_walk_anon_lock(struct folio *folio, |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2553 | struct rmap_walk_control *rwc) |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2554 | { |
| 2555 | struct anon_vma *anon_vma; |
| 2556 | |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2557 | if (rwc->anon_lock) |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2558 | return rwc->anon_lock(folio, rwc); |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2559 | |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2560 | /* |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2561 | * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read() |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2562 | * because that depends on page_mapped(); but not all its usages |
Michel Lespinasse | c1e8d7c | 2020-06-08 21:33:54 -0700 | [diff] [blame] | 2563 | * are holding mmap_lock. Users without mmap_lock are required to |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2564 | * take a reference count to prevent the anon_vma disappearing |
| 2565 | */ |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 2566 | anon_vma = folio_anon_vma(folio); |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2567 | if (!anon_vma) |
| 2568 | return NULL; |
| 2569 | |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2570 | if (anon_vma_trylock_read(anon_vma)) |
| 2571 | goto out; |
| 2572 | |
| 2573 | if (rwc->try_lock) { |
| 2574 | anon_vma = NULL; |
| 2575 | rwc->contended = true; |
| 2576 | goto out; |
| 2577 | } |
| 2578 | |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2579 | anon_vma_lock_read(anon_vma); |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2580 | out: |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2581 | return anon_vma; |
| 2582 | } |
| 2583 | |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2584 | /* |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2585 | * rmap_walk_anon - do something to anonymous page using the object-based |
| 2586 | * rmap method |
Kemeng Shi | 89be82b | 2023-07-18 17:21:36 +0800 | [diff] [blame] | 2587 | * @folio: the folio to be handled |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2588 | * @rwc: control variable according to each walk type |
Kemeng Shi | 89be82b | 2023-07-18 17:21:36 +0800 | [diff] [blame] | 2589 | * @locked: caller holds relevant rmap lock |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2590 | * |
Kemeng Shi | 89be82b | 2023-07-18 17:21:36 +0800 | [diff] [blame] | 2591 | * Find all the mappings of a folio using the mapping pointer and the vma |
| 2592 | * chains contained in the anon_vma struct it points to. |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2593 | */ |
Matthew Wilcox (Oracle) | 84fbbe2 | 2022-01-29 16:16:54 -0500 | [diff] [blame] | 2594 | static void rmap_walk_anon(struct folio *folio, |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2595 | struct rmap_walk_control *rwc, bool locked) |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2596 | { |
| 2597 | struct anon_vma *anon_vma; |
Kirill A. Shutemov | a8fa41ad | 2017-02-24 14:57:54 -0800 | [diff] [blame] | 2598 | pgoff_t pgoff_start, pgoff_end; |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 2599 | struct anon_vma_chain *avc; |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2600 | |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2601 | if (locked) { |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 2602 | anon_vma = folio_anon_vma(folio); |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2603 | /* anon_vma disappear under us? */ |
Matthew Wilcox (Oracle) | e05b345 | 2022-01-29 11:52:52 -0500 | [diff] [blame] | 2604 | VM_BUG_ON_FOLIO(!anon_vma, folio); |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2605 | } else { |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2606 | anon_vma = rmap_walk_anon_lock(folio, rwc); |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2607 | } |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2608 | if (!anon_vma) |
Minchan Kim | 1df631a | 2017-05-03 14:54:23 -0700 | [diff] [blame] | 2609 | return; |
Joonsoo Kim | faecd8d | 2014-01-21 15:49:46 -0800 | [diff] [blame] | 2610 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2611 | pgoff_start = folio_pgoff(folio); |
| 2612 | pgoff_end = pgoff_start + folio_nr_pages(folio) - 1; |
Kirill A. Shutemov | a8fa41ad | 2017-02-24 14:57:54 -0800 | [diff] [blame] | 2613 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, |
| 2614 | pgoff_start, pgoff_end) { |
Rik van Riel | 5beb493 | 2010-03-05 13:42:07 -0800 | [diff] [blame] | 2615 | struct vm_area_struct *vma = avc->vma; |
Matthew Wilcox (Oracle) | e0abfbb | 2024-03-28 22:58:29 +0000 | [diff] [blame] | 2616 | unsigned long address = vma_address(vma, pgoff_start, |
| 2617 | folio_nr_pages(folio)); |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2618 | |
Hugh Dickins | 494334e | 2021-06-15 18:23:56 -0700 | [diff] [blame] | 2619 | VM_BUG_ON_VMA(address == -EFAULT, vma); |
Andrea Arcangeli | ad12695 | 2015-11-05 18:49:07 -0800 | [diff] [blame] | 2620 | cond_resched(); |
| 2621 | |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2622 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
| 2623 | continue; |
| 2624 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2625 | if (!rwc->rmap_one(folio, vma, address, rwc->arg)) |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2626 | break; |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2627 | if (rwc->done && rwc->done(folio)) |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2628 | break; |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2629 | } |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2630 | |
| 2631 | if (!locked) |
| 2632 | anon_vma_unlock_read(anon_vma); |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2633 | } |
| 2634 | |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2635 | /* |
| 2636 | * rmap_walk_file - do something to file page using the object-based rmap method |
Kemeng Shi | 89be82b | 2023-07-18 17:21:36 +0800 | [diff] [blame] | 2637 | * @folio: the folio to be handled |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2638 | * @rwc: control variable according to each walk type |
Kemeng Shi | 89be82b | 2023-07-18 17:21:36 +0800 | [diff] [blame] | 2639 | * @locked: caller holds relevant rmap lock |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2640 | * |
Kemeng Shi | 89be82b | 2023-07-18 17:21:36 +0800 | [diff] [blame] | 2641 | * Find all the mappings of a folio using the mapping pointer and the vma chains |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2642 | * contained in the address_space struct it points to. |
Joonsoo Kim | e8351ac | 2014-01-21 15:49:52 -0800 | [diff] [blame] | 2643 | */ |
Matthew Wilcox (Oracle) | 84fbbe2 | 2022-01-29 16:16:54 -0500 | [diff] [blame] | 2644 | static void rmap_walk_file(struct folio *folio, |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2645 | struct rmap_walk_control *rwc, bool locked) |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2646 | { |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2647 | struct address_space *mapping = folio_mapping(folio); |
Kirill A. Shutemov | a8fa41ad | 2017-02-24 14:57:54 -0800 | [diff] [blame] | 2648 | pgoff_t pgoff_start, pgoff_end; |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2649 | struct vm_area_struct *vma; |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2650 | |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 2651 | /* |
| 2652 | * The page lock not only makes sure that page->mapping cannot |
| 2653 | * suddenly be NULLified by truncation, it makes sure that the |
| 2654 | * structure at mapping cannot be freed and reused yet, |
Davidlohr Bueso | c8c06ef | 2014-12-12 16:54:24 -0800 | [diff] [blame] | 2655 | * so we can safely take mapping->i_mmap_rwsem. |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 2656 | */ |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2657 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
Joonsoo Kim | 9f32624 | 2014-01-21 15:49:53 -0800 | [diff] [blame] | 2658 | |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2659 | if (!mapping) |
Minchan Kim | 1df631a | 2017-05-03 14:54:23 -0700 | [diff] [blame] | 2660 | return; |
Davidlohr Bueso | 3dec0ba | 2014-12-12 16:54:27 -0800 | [diff] [blame] | 2661 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2662 | pgoff_start = folio_pgoff(folio); |
| 2663 | pgoff_end = pgoff_start + folio_nr_pages(folio) - 1; |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2664 | if (!locked) { |
| 2665 | if (i_mmap_trylock_read(mapping)) |
| 2666 | goto lookup; |
| 2667 | |
| 2668 | if (rwc->try_lock) { |
| 2669 | rwc->contended = true; |
| 2670 | return; |
| 2671 | } |
| 2672 | |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2673 | i_mmap_lock_read(mapping); |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2674 | } |
| 2675 | lookup: |
Kirill A. Shutemov | a8fa41ad | 2017-02-24 14:57:54 -0800 | [diff] [blame] | 2676 | vma_interval_tree_foreach(vma, &mapping->i_mmap, |
| 2677 | pgoff_start, pgoff_end) { |
Matthew Wilcox (Oracle) | e0abfbb | 2024-03-28 22:58:29 +0000 | [diff] [blame] | 2678 | unsigned long address = vma_address(vma, pgoff_start, |
| 2679 | folio_nr_pages(folio)); |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2680 | |
Hugh Dickins | 494334e | 2021-06-15 18:23:56 -0700 | [diff] [blame] | 2681 | VM_BUG_ON_VMA(address == -EFAULT, vma); |
Andrea Arcangeli | ad12695 | 2015-11-05 18:49:07 -0800 | [diff] [blame] | 2682 | cond_resched(); |
| 2683 | |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2684 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
| 2685 | continue; |
| 2686 | |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2687 | if (!rwc->rmap_one(folio, vma, address, rwc->arg)) |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2688 | goto done; |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2689 | if (rwc->done && rwc->done(folio)) |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2690 | goto done; |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2691 | } |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2692 | |
Joonsoo Kim | 0dd1c7b | 2014-01-21 15:49:49 -0800 | [diff] [blame] | 2693 | done: |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2694 | if (!locked) |
| 2695 | i_mmap_unlock_read(mapping); |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2696 | } |
| 2697 | |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2698 | void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc) |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2699 | { |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2700 | if (unlikely(folio_test_ksm(folio))) |
| 2701 | rmap_walk_ksm(folio, rwc); |
| 2702 | else if (folio_test_anon(folio)) |
| 2703 | rmap_walk_anon(folio, rwc, false); |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2704 | else |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2705 | rmap_walk_file(folio, rwc, false); |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2706 | } |
| 2707 | |
| 2708 | /* Like rmap_walk, but caller holds relevant rmap lock */ |
Minchan Kim | 6d4675e | 2022-05-19 14:08:54 -0700 | [diff] [blame] | 2709 | void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc) |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2710 | { |
| 2711 | /* no ksm support for now */ |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2712 | VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio); |
| 2713 | if (folio_test_anon(folio)) |
| 2714 | rmap_walk_anon(folio, rwc, true); |
Kirill A. Shutemov | b977319 | 2016-03-17 14:20:01 -0700 | [diff] [blame] | 2715 | else |
Matthew Wilcox (Oracle) | 2f031c6 | 2022-01-29 16:06:53 -0500 | [diff] [blame] | 2716 | rmap_walk_file(folio, rwc, true); |
Hugh Dickins | e9995ef | 2009-12-14 17:59:31 -0800 | [diff] [blame] | 2717 | } |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2718 | |
Naoya Horiguchi | e3390f6 | 2010-06-15 13:18:13 +0900 | [diff] [blame] | 2719 | #ifdef CONFIG_HUGETLB_PAGE |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2720 | /* |
Kirill Tkhai | 451b951 | 2018-12-28 00:39:31 -0800 | [diff] [blame] | 2721 | * The following two functions are for anonymous (private mapped) hugepages. |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2722 | * Unlike common anonymous pages, anonymous hugepages have no accounting code |
| 2723 | * and no lru code, because we handle hugepages differently from common pages. |
| 2724 | */ |
David Hildenbrand | 9d5fafd | 2023-12-20 23:44:25 +0100 | [diff] [blame] | 2725 | void hugetlb_add_anon_rmap(struct folio *folio, struct vm_area_struct *vma, |
| 2726 | unsigned long address, rmap_t flags) |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2727 | { |
David Hildenbrand | a4ea186 | 2023-12-20 23:44:30 +0100 | [diff] [blame] | 2728 | VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio); |
David Hildenbrand | c5c5400 | 2023-09-13 14:51:10 +0200 | [diff] [blame] | 2729 | VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio); |
| 2730 | |
David Hildenbrand | 132b180 | 2023-09-13 14:51:12 +0200 | [diff] [blame] | 2731 | atomic_inc(&folio->_entire_mapcount); |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 2732 | atomic_inc(&folio->_large_mapcount); |
David Hildenbrand | c66db8c | 2023-09-13 14:51:09 +0200 | [diff] [blame] | 2733 | if (flags & RMAP_EXCLUSIVE) |
David Hildenbrand | 09c5505 | 2023-09-13 14:51:13 +0200 | [diff] [blame] | 2734 | SetPageAnonExclusive(&folio->page); |
David Hildenbrand | 132b180 | 2023-09-13 14:51:12 +0200 | [diff] [blame] | 2735 | VM_WARN_ON_FOLIO(folio_entire_mapcount(folio) > 1 && |
David Hildenbrand | 09c5505 | 2023-09-13 14:51:13 +0200 | [diff] [blame] | 2736 | PageAnonExclusive(&folio->page), folio); |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2737 | } |
| 2738 | |
David Hildenbrand | 9d5fafd | 2023-12-20 23:44:25 +0100 | [diff] [blame] | 2739 | void hugetlb_add_new_anon_rmap(struct folio *folio, |
| 2740 | struct vm_area_struct *vma, unsigned long address) |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2741 | { |
David Hildenbrand | a4ea186 | 2023-12-20 23:44:30 +0100 | [diff] [blame] | 2742 | VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio); |
| 2743 | |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2744 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
Hugh Dickins | cb67f42 | 2022-11-02 18:51:38 -0700 | [diff] [blame] | 2745 | /* increment count (starts at -1) */ |
Matthew Wilcox (Oracle) | db4e5db | 2023-01-11 14:28:56 +0000 | [diff] [blame] | 2746 | atomic_set(&folio->_entire_mapcount, 0); |
David Hildenbrand | 05c5323 | 2024-04-09 21:22:47 +0200 | [diff] [blame] | 2747 | atomic_set(&folio->_large_mapcount, 0); |
Matthew Wilcox (Oracle) | db4e5db | 2023-01-11 14:28:56 +0000 | [diff] [blame] | 2748 | folio_clear_hugetlb_restore_reserve(folio); |
David Hildenbrand | c66db8c | 2023-09-13 14:51:09 +0200 | [diff] [blame] | 2749 | __folio_set_anon(folio, vma, address, true); |
| 2750 | SetPageAnonExclusive(&folio->page); |
Naoya Horiguchi | 0fe6e20 | 2010-05-28 09:29:16 +0900 | [diff] [blame] | 2751 | } |
Naoya Horiguchi | e3390f6 | 2010-06-15 13:18:13 +0900 | [diff] [blame] | 2752 | #endif /* CONFIG_HUGETLB_PAGE */ |