1 /* 1 /*
2 * mm/rmap.c - physical to virtual reverse map 2 * mm/rmap.c - physical to virtual reverse mappings
3 * 3 *
4 * Copyright 2001, Rik van Riel <riel@conectiv 4 * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5 * Released under the General Public License ( 5 * Released under the General Public License (GPL).
6 * 6 *
7 * Simple, low overhead reverse mapping scheme 7 * Simple, low overhead reverse mapping scheme.
8 * Please try to keep this thing as modular as 8 * Please try to keep this thing as modular as possible.
9 * 9 *
10 * Provides methods for unmapping each kind of 10 * Provides methods for unmapping each kind of mapped page:
11 * the anon methods track anonymous pages, and 11 * the anon methods track anonymous pages, and
12 * the file methods track pages belonging to a 12 * the file methods track pages belonging to an inode.
13 * 13 *
14 * Original design by Rik van Riel <riel@conec 14 * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15 * File methods by Dave McCracken <dmccr@us.ib 15 * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16 * Anonymous methods by Andrea Arcangeli <andr 16 * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17 * Contributions by Hugh Dickins 2003, 2004 17 * Contributions by Hugh Dickins 2003, 2004
18 */ 18 */
19 19
20 /* 20 /*
21 * Lock ordering in mm: 21 * Lock ordering in mm:
22 * 22 *
23 * inode->i_rwsem (while writing or trun 23 * inode->i_rwsem (while writing or truncating, not reading or faulting)
24 * mm->mmap_lock 24 * mm->mmap_lock
25 * mapping->invalidate_lock (in filemap_fa 25 * mapping->invalidate_lock (in filemap_fault)
26 * folio_lock !! 26 * page->flags PG_locked (lock_page)
27 * hugetlbfs_i_mmap_rwsem_key (in huge 27 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
28 * vma_start_write !! 28 * mapping->i_mmap_rwsem
29 * mapping->i_mmap_rwsem !! 29 * anon_vma->rwsem
30 * anon_vma->rwsem !! 30 * mm->page_table_lock or pte_lock
31 * mm->page_table_lock or pte_ !! 31 * swap_lock (in swap_duplicate, swap_info_get)
32 * swap_lock (in swap_duplic !! 32 * mmlist_lock (in mmput, drain_mmlist and others)
33 * mmlist_lock (in mmput, !! 33 * mapping->private_lock (in block_dirty_folio)
34 * mapping->private_lock ( !! 34 * folio_lock_memcg move_lock (in block_dirty_folio)
35 * folio_lock_memcg move !! 35 * i_pages lock (widely used)
36 * i_pages lock (widel !! 36 * lruvec->lru_lock (in folio_lruvec_lock_irq)
37 * lruvec->lru_lock !! 37 * inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38 * inode->i_lock (in set_p !! 38 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39 * bdi.wb->list_lock (in s !! 39 * sb_lock (within inode_lock in fs/fs-writeback.c)
40 * sb_lock (within inode !! 40 * i_pages lock (widely used, in set_page_dirty,
41 * i_pages lock (widely !! 41 * in arch-dependent flush_dcache_mmap_lock,
42 * in arch-dep !! 42 * within bdi.wb->list_lock in __sync_single_inode)
43 * within bdi. <<
44 * 43 *
45 * anon_vma->rwsem,mapping->i_mmap_rwsem (me 44 * anon_vma->rwsem,mapping->i_mmap_rwsem (memory_failure, collect_procs_anon)
46 * ->tasklist_lock 45 * ->tasklist_lock
47 * pte map lock 46 * pte map lock
48 * 47 *
49 * hugetlbfs PageHuge() take locks in this ord 48 * hugetlbfs PageHuge() take locks in this order:
50 * hugetlb_fault_mutex (hugetlbfs specific p 49 * hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
51 * vma_lock (hugetlb specific lock for pmd 50 * vma_lock (hugetlb specific lock for pmd_sharing)
52 * mapping->i_mmap_rwsem (also used for 51 * mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
53 * folio_lock !! 52 * page->flags PG_locked (lock_page)
54 */ 53 */
55 54
56 #include <linux/mm.h> 55 #include <linux/mm.h>
57 #include <linux/sched/mm.h> 56 #include <linux/sched/mm.h>
58 #include <linux/sched/task.h> 57 #include <linux/sched/task.h>
59 #include <linux/pagemap.h> 58 #include <linux/pagemap.h>
60 #include <linux/swap.h> 59 #include <linux/swap.h>
61 #include <linux/swapops.h> 60 #include <linux/swapops.h>
62 #include <linux/slab.h> 61 #include <linux/slab.h>
63 #include <linux/init.h> 62 #include <linux/init.h>
64 #include <linux/ksm.h> 63 #include <linux/ksm.h>
65 #include <linux/rmap.h> 64 #include <linux/rmap.h>
66 #include <linux/rcupdate.h> 65 #include <linux/rcupdate.h>
67 #include <linux/export.h> 66 #include <linux/export.h>
68 #include <linux/memcontrol.h> 67 #include <linux/memcontrol.h>
69 #include <linux/mmu_notifier.h> 68 #include <linux/mmu_notifier.h>
70 #include <linux/migrate.h> 69 #include <linux/migrate.h>
71 #include <linux/hugetlb.h> 70 #include <linux/hugetlb.h>
72 #include <linux/huge_mm.h> 71 #include <linux/huge_mm.h>
73 #include <linux/backing-dev.h> 72 #include <linux/backing-dev.h>
74 #include <linux/page_idle.h> 73 #include <linux/page_idle.h>
75 #include <linux/memremap.h> 74 #include <linux/memremap.h>
76 #include <linux/userfaultfd_k.h> 75 #include <linux/userfaultfd_k.h>
77 #include <linux/mm_inline.h> 76 #include <linux/mm_inline.h>
78 #include <linux/oom.h> <<
79 77
80 #include <asm/tlbflush.h> 78 #include <asm/tlbflush.h>
81 79
82 #define CREATE_TRACE_POINTS 80 #define CREATE_TRACE_POINTS
83 #include <trace/events/tlb.h> 81 #include <trace/events/tlb.h>
84 #include <trace/events/migrate.h> 82 #include <trace/events/migrate.h>
85 83
86 #include "internal.h" 84 #include "internal.h"
87 85
88 static struct kmem_cache *anon_vma_cachep; 86 static struct kmem_cache *anon_vma_cachep;
89 static struct kmem_cache *anon_vma_chain_cache 87 static struct kmem_cache *anon_vma_chain_cachep;
90 88
91 static inline struct anon_vma *anon_vma_alloc( 89 static inline struct anon_vma *anon_vma_alloc(void)
92 { 90 {
93 struct anon_vma *anon_vma; 91 struct anon_vma *anon_vma;
94 92
95 anon_vma = kmem_cache_alloc(anon_vma_c 93 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
96 if (anon_vma) { 94 if (anon_vma) {
97 atomic_set(&anon_vma->refcount 95 atomic_set(&anon_vma->refcount, 1);
98 anon_vma->num_children = 0; 96 anon_vma->num_children = 0;
99 anon_vma->num_active_vmas = 0; 97 anon_vma->num_active_vmas = 0;
100 anon_vma->parent = anon_vma; 98 anon_vma->parent = anon_vma;
101 /* 99 /*
102 * Initialise the anon_vma roo 100 * Initialise the anon_vma root to point to itself. If called
103 * from fork, the root will be 101 * from fork, the root will be reset to the parents anon_vma.
104 */ 102 */
105 anon_vma->root = anon_vma; 103 anon_vma->root = anon_vma;
106 } 104 }
107 105
108 return anon_vma; 106 return anon_vma;
109 } 107 }
110 108
111 static inline void anon_vma_free(struct anon_v 109 static inline void anon_vma_free(struct anon_vma *anon_vma)
112 { 110 {
113 VM_BUG_ON(atomic_read(&anon_vma->refco 111 VM_BUG_ON(atomic_read(&anon_vma->refcount));
114 112
115 /* 113 /*
116 * Synchronize against folio_lock_anon 114 * Synchronize against folio_lock_anon_vma_read() such that
117 * we can safely hold the lock without 115 * we can safely hold the lock without the anon_vma getting
118 * freed. 116 * freed.
119 * 117 *
120 * Relies on the full mb implied by th 118 * Relies on the full mb implied by the atomic_dec_and_test() from
121 * put_anon_vma() against the acquire 119 * put_anon_vma() against the acquire barrier implied by
122 * down_read_trylock() from folio_lock 120 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
123 * 121 *
124 * folio_lock_anon_vma_read() VS 122 * folio_lock_anon_vma_read() VS put_anon_vma()
125 * down_read_trylock() 123 * down_read_trylock() atomic_dec_and_test()
126 * LOCK 124 * LOCK MB
127 * atomic_read() 125 * atomic_read() rwsem_is_locked()
128 * 126 *
129 * LOCK should suffice since the actua 127 * LOCK should suffice since the actual taking of the lock must
130 * happen _before_ what follows. 128 * happen _before_ what follows.
131 */ 129 */
132 might_sleep(); 130 might_sleep();
133 if (rwsem_is_locked(&anon_vma->root->r 131 if (rwsem_is_locked(&anon_vma->root->rwsem)) {
134 anon_vma_lock_write(anon_vma); 132 anon_vma_lock_write(anon_vma);
135 anon_vma_unlock_write(anon_vma 133 anon_vma_unlock_write(anon_vma);
136 } 134 }
137 135
138 kmem_cache_free(anon_vma_cachep, anon_ 136 kmem_cache_free(anon_vma_cachep, anon_vma);
139 } 137 }
140 138
141 static inline struct anon_vma_chain *anon_vma_ 139 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
142 { 140 {
143 return kmem_cache_alloc(anon_vma_chain 141 return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
144 } 142 }
145 143
146 static void anon_vma_chain_free(struct anon_vm 144 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
147 { 145 {
148 kmem_cache_free(anon_vma_chain_cachep, 146 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
149 } 147 }
150 148
151 static void anon_vma_chain_link(struct vm_area 149 static void anon_vma_chain_link(struct vm_area_struct *vma,
152 struct anon_vm 150 struct anon_vma_chain *avc,
153 struct anon_vm 151 struct anon_vma *anon_vma)
154 { 152 {
155 avc->vma = vma; 153 avc->vma = vma;
156 avc->anon_vma = anon_vma; 154 avc->anon_vma = anon_vma;
157 list_add(&avc->same_vma, &vma->anon_vm 155 list_add(&avc->same_vma, &vma->anon_vma_chain);
158 anon_vma_interval_tree_insert(avc, &an 156 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
159 } 157 }
160 158
161 /** 159 /**
162 * __anon_vma_prepare - attach an anon_vma to 160 * __anon_vma_prepare - attach an anon_vma to a memory region
163 * @vma: the memory region in question 161 * @vma: the memory region in question
164 * 162 *
165 * This makes sure the memory mapping describe 163 * This makes sure the memory mapping described by 'vma' has
166 * an 'anon_vma' attached to it, so that we ca 164 * an 'anon_vma' attached to it, so that we can associate the
167 * anonymous pages mapped into it with that an 165 * anonymous pages mapped into it with that anon_vma.
168 * 166 *
169 * The common case will be that we already hav 167 * The common case will be that we already have one, which
170 * is handled inline by anon_vma_prepare(). Bu 168 * is handled inline by anon_vma_prepare(). But if
171 * not we either need to find an adjacent mapp 169 * not we either need to find an adjacent mapping that we
172 * can re-use the anon_vma from (very common w 170 * can re-use the anon_vma from (very common when the only
173 * reason for splitting a vma has been mprotec 171 * reason for splitting a vma has been mprotect()), or we
174 * allocate a new one. 172 * allocate a new one.
175 * 173 *
176 * Anon-vma allocations are very subtle, becau 174 * Anon-vma allocations are very subtle, because we may have
177 * optimistically looked up an anon_vma in fol 175 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
178 * and that may actually touch the rwsem even 176 * and that may actually touch the rwsem even in the newly
179 * allocated vma (it depends on RCU to make su 177 * allocated vma (it depends on RCU to make sure that the
180 * anon_vma isn't actually destroyed). 178 * anon_vma isn't actually destroyed).
181 * 179 *
182 * As a result, we need to do proper anon_vma 180 * As a result, we need to do proper anon_vma locking even
183 * for the new allocation. At the same time, w 181 * for the new allocation. At the same time, we do not want
184 * to do any locking for the common case of al 182 * to do any locking for the common case of already having
185 * an anon_vma. 183 * an anon_vma.
>> 184 *
>> 185 * This must be called with the mmap_lock held for reading.
186 */ 186 */
187 int __anon_vma_prepare(struct vm_area_struct * 187 int __anon_vma_prepare(struct vm_area_struct *vma)
188 { 188 {
189 struct mm_struct *mm = vma->vm_mm; 189 struct mm_struct *mm = vma->vm_mm;
190 struct anon_vma *anon_vma, *allocated; 190 struct anon_vma *anon_vma, *allocated;
191 struct anon_vma_chain *avc; 191 struct anon_vma_chain *avc;
192 192
193 mmap_assert_locked(mm); <<
194 might_sleep(); 193 might_sleep();
195 194
196 avc = anon_vma_chain_alloc(GFP_KERNEL) 195 avc = anon_vma_chain_alloc(GFP_KERNEL);
197 if (!avc) 196 if (!avc)
198 goto out_enomem; 197 goto out_enomem;
199 198
200 anon_vma = find_mergeable_anon_vma(vma 199 anon_vma = find_mergeable_anon_vma(vma);
201 allocated = NULL; 200 allocated = NULL;
202 if (!anon_vma) { 201 if (!anon_vma) {
203 anon_vma = anon_vma_alloc(); 202 anon_vma = anon_vma_alloc();
204 if (unlikely(!anon_vma)) 203 if (unlikely(!anon_vma))
205 goto out_enomem_free_a 204 goto out_enomem_free_avc;
206 anon_vma->num_children++; /* s 205 anon_vma->num_children++; /* self-parent link for new root */
207 allocated = anon_vma; 206 allocated = anon_vma;
208 } 207 }
209 208
210 anon_vma_lock_write(anon_vma); 209 anon_vma_lock_write(anon_vma);
211 /* page_table_lock to protect against 210 /* page_table_lock to protect against threads */
212 spin_lock(&mm->page_table_lock); 211 spin_lock(&mm->page_table_lock);
213 if (likely(!vma->anon_vma)) { 212 if (likely(!vma->anon_vma)) {
214 vma->anon_vma = anon_vma; 213 vma->anon_vma = anon_vma;
215 anon_vma_chain_link(vma, avc, 214 anon_vma_chain_link(vma, avc, anon_vma);
216 anon_vma->num_active_vmas++; 215 anon_vma->num_active_vmas++;
217 allocated = NULL; 216 allocated = NULL;
218 avc = NULL; 217 avc = NULL;
219 } 218 }
220 spin_unlock(&mm->page_table_lock); 219 spin_unlock(&mm->page_table_lock);
221 anon_vma_unlock_write(anon_vma); 220 anon_vma_unlock_write(anon_vma);
222 221
223 if (unlikely(allocated)) 222 if (unlikely(allocated))
224 put_anon_vma(allocated); 223 put_anon_vma(allocated);
225 if (unlikely(avc)) 224 if (unlikely(avc))
226 anon_vma_chain_free(avc); 225 anon_vma_chain_free(avc);
227 226
228 return 0; 227 return 0;
229 228
230 out_enomem_free_avc: 229 out_enomem_free_avc:
231 anon_vma_chain_free(avc); 230 anon_vma_chain_free(avc);
232 out_enomem: 231 out_enomem:
233 return -ENOMEM; 232 return -ENOMEM;
234 } 233 }
235 234
236 /* 235 /*
237 * This is a useful helper function for lockin 236 * This is a useful helper function for locking the anon_vma root as
238 * we traverse the vma->anon_vma_chain, loopin 237 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
239 * have the same vma. 238 * have the same vma.
240 * 239 *
241 * Such anon_vma's should have the same root, 240 * Such anon_vma's should have the same root, so you'd expect to see
242 * just a single mutex_lock for the whole trav 241 * just a single mutex_lock for the whole traversal.
243 */ 242 */
244 static inline struct anon_vma *lock_anon_vma_r 243 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
245 { 244 {
246 struct anon_vma *new_root = anon_vma-> 245 struct anon_vma *new_root = anon_vma->root;
247 if (new_root != root) { 246 if (new_root != root) {
248 if (WARN_ON_ONCE(root)) 247 if (WARN_ON_ONCE(root))
249 up_write(&root->rwsem) 248 up_write(&root->rwsem);
250 root = new_root; 249 root = new_root;
251 down_write(&root->rwsem); 250 down_write(&root->rwsem);
252 } 251 }
253 return root; 252 return root;
254 } 253 }
255 254
256 static inline void unlock_anon_vma_root(struct 255 static inline void unlock_anon_vma_root(struct anon_vma *root)
257 { 256 {
258 if (root) 257 if (root)
259 up_write(&root->rwsem); 258 up_write(&root->rwsem);
260 } 259 }
261 260
262 /* 261 /*
263 * Attach the anon_vmas from src to dst. 262 * Attach the anon_vmas from src to dst.
264 * Returns 0 on success, -ENOMEM on failure. 263 * Returns 0 on success, -ENOMEM on failure.
265 * 264 *
266 * anon_vma_clone() is called by vma_expand(), !! 265 * anon_vma_clone() is called by __vma_adjust(), __split_vma(), copy_vma() and
267 * copy_vma() and anon_vma_fork(). The first f !! 266 * anon_vma_fork(). The first three want an exact copy of src, while the last
268 * while the last one, anon_vma_fork(), may tr !! 267 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
269 * prevent endless growth of anon_vma. Since d !! 268 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
270 * call, we can identify this case by checking !! 269 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
271 * src->anon_vma). <<
272 * 270 *
273 * If (!dst->anon_vma && src->anon_vma) is tru 271 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
274 * and reuse existing anon_vma which has no vm 272 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
275 * This prevents degradation of anon_vma hiera 273 * This prevents degradation of anon_vma hierarchy to endless linear chain in
276 * case of constantly forking task. On the oth 274 * case of constantly forking task. On the other hand, an anon_vma with more
277 * than one child isn't reused even if there w 275 * than one child isn't reused even if there was no alive vma, thus rmap
278 * walker has a good chance of avoiding scanni 276 * walker has a good chance of avoiding scanning the whole hierarchy when it
279 * searches where page is mapped. 277 * searches where page is mapped.
280 */ 278 */
281 int anon_vma_clone(struct vm_area_struct *dst, 279 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
282 { 280 {
283 struct anon_vma_chain *avc, *pavc; 281 struct anon_vma_chain *avc, *pavc;
284 struct anon_vma *root = NULL; 282 struct anon_vma *root = NULL;
285 283
286 list_for_each_entry_reverse(pavc, &src 284 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
287 struct anon_vma *anon_vma; 285 struct anon_vma *anon_vma;
288 286
289 avc = anon_vma_chain_alloc(GFP 287 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
290 if (unlikely(!avc)) { 288 if (unlikely(!avc)) {
291 unlock_anon_vma_root(r 289 unlock_anon_vma_root(root);
292 root = NULL; 290 root = NULL;
293 avc = anon_vma_chain_a 291 avc = anon_vma_chain_alloc(GFP_KERNEL);
294 if (!avc) 292 if (!avc)
295 goto enomem_fa 293 goto enomem_failure;
296 } 294 }
297 anon_vma = pavc->anon_vma; 295 anon_vma = pavc->anon_vma;
298 root = lock_anon_vma_root(root 296 root = lock_anon_vma_root(root, anon_vma);
299 anon_vma_chain_link(dst, avc, 297 anon_vma_chain_link(dst, avc, anon_vma);
300 298
301 /* 299 /*
302 * Reuse existing anon_vma if 300 * Reuse existing anon_vma if it has no vma and only one
303 * anon_vma child. 301 * anon_vma child.
304 * 302 *
305 * Root anon_vma is never reus 303 * Root anon_vma is never reused:
306 * it has self-parent referenc 304 * it has self-parent reference and at least one child.
307 */ 305 */
308 if (!dst->anon_vma && src->ano 306 if (!dst->anon_vma && src->anon_vma &&
309 anon_vma->num_children < 2 307 anon_vma->num_children < 2 &&
310 anon_vma->num_active_vmas 308 anon_vma->num_active_vmas == 0)
311 dst->anon_vma = anon_v 309 dst->anon_vma = anon_vma;
312 } 310 }
313 if (dst->anon_vma) 311 if (dst->anon_vma)
314 dst->anon_vma->num_active_vmas 312 dst->anon_vma->num_active_vmas++;
315 unlock_anon_vma_root(root); 313 unlock_anon_vma_root(root);
316 return 0; 314 return 0;
317 315
318 enomem_failure: 316 enomem_failure:
319 /* 317 /*
320 * dst->anon_vma is dropped here other 318 * dst->anon_vma is dropped here otherwise its num_active_vmas can
321 * be incorrectly decremented in unlin 319 * be incorrectly decremented in unlink_anon_vmas().
322 * We can safely do this because calle 320 * We can safely do this because callers of anon_vma_clone() don't care
323 * about dst->anon_vma if anon_vma_clo 321 * about dst->anon_vma if anon_vma_clone() failed.
324 */ 322 */
325 dst->anon_vma = NULL; 323 dst->anon_vma = NULL;
326 unlink_anon_vmas(dst); 324 unlink_anon_vmas(dst);
327 return -ENOMEM; 325 return -ENOMEM;
328 } 326 }
329 327
330 /* 328 /*
331 * Attach vma to its own anon_vma, as well as 329 * Attach vma to its own anon_vma, as well as to the anon_vmas that
332 * the corresponding VMA in the parent process 330 * the corresponding VMA in the parent process is attached to.
333 * Returns 0 on success, non-zero on failure. 331 * Returns 0 on success, non-zero on failure.
334 */ 332 */
335 int anon_vma_fork(struct vm_area_struct *vma, 333 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
336 { 334 {
337 struct anon_vma_chain *avc; 335 struct anon_vma_chain *avc;
338 struct anon_vma *anon_vma; 336 struct anon_vma *anon_vma;
339 int error; 337 int error;
340 338
341 /* Don't bother if the parent process 339 /* Don't bother if the parent process has no anon_vma here. */
342 if (!pvma->anon_vma) 340 if (!pvma->anon_vma)
343 return 0; 341 return 0;
344 342
345 /* Drop inherited anon_vma, we'll reus 343 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */
346 vma->anon_vma = NULL; 344 vma->anon_vma = NULL;
347 345
348 /* 346 /*
349 * First, attach the new VMA to the pa 347 * First, attach the new VMA to the parent VMA's anon_vmas,
350 * so rmap can find non-COWed pages in 348 * so rmap can find non-COWed pages in child processes.
351 */ 349 */
352 error = anon_vma_clone(vma, pvma); 350 error = anon_vma_clone(vma, pvma);
353 if (error) 351 if (error)
354 return error; 352 return error;
355 353
356 /* An existing anon_vma has been reuse 354 /* An existing anon_vma has been reused, all done then. */
357 if (vma->anon_vma) 355 if (vma->anon_vma)
358 return 0; 356 return 0;
359 357
360 /* Then add our own anon_vma. */ 358 /* Then add our own anon_vma. */
361 anon_vma = anon_vma_alloc(); 359 anon_vma = anon_vma_alloc();
362 if (!anon_vma) 360 if (!anon_vma)
363 goto out_error; 361 goto out_error;
364 anon_vma->num_active_vmas++; 362 anon_vma->num_active_vmas++;
365 avc = anon_vma_chain_alloc(GFP_KERNEL) 363 avc = anon_vma_chain_alloc(GFP_KERNEL);
366 if (!avc) 364 if (!avc)
367 goto out_error_free_anon_vma; 365 goto out_error_free_anon_vma;
368 366
369 /* 367 /*
370 * The root anon_vma's rwsem is the lo 368 * The root anon_vma's rwsem is the lock actually used when we
371 * lock any of the anon_vmas in this a 369 * lock any of the anon_vmas in this anon_vma tree.
372 */ 370 */
373 anon_vma->root = pvma->anon_vma->root; 371 anon_vma->root = pvma->anon_vma->root;
374 anon_vma->parent = pvma->anon_vma; 372 anon_vma->parent = pvma->anon_vma;
375 /* 373 /*
376 * With refcounts, an anon_vma can sta 374 * With refcounts, an anon_vma can stay around longer than the
377 * process it belongs to. The root ano 375 * process it belongs to. The root anon_vma needs to be pinned until
378 * this anon_vma is freed, because the 376 * this anon_vma is freed, because the lock lives in the root.
379 */ 377 */
380 get_anon_vma(anon_vma->root); 378 get_anon_vma(anon_vma->root);
381 /* Mark this anon_vma as the one where 379 /* Mark this anon_vma as the one where our new (COWed) pages go. */
382 vma->anon_vma = anon_vma; 380 vma->anon_vma = anon_vma;
383 anon_vma_lock_write(anon_vma); 381 anon_vma_lock_write(anon_vma);
384 anon_vma_chain_link(vma, avc, anon_vma 382 anon_vma_chain_link(vma, avc, anon_vma);
385 anon_vma->parent->num_children++; 383 anon_vma->parent->num_children++;
386 anon_vma_unlock_write(anon_vma); 384 anon_vma_unlock_write(anon_vma);
387 385
388 return 0; 386 return 0;
389 387
390 out_error_free_anon_vma: 388 out_error_free_anon_vma:
391 put_anon_vma(anon_vma); 389 put_anon_vma(anon_vma);
392 out_error: 390 out_error:
393 unlink_anon_vmas(vma); 391 unlink_anon_vmas(vma);
394 return -ENOMEM; 392 return -ENOMEM;
395 } 393 }
396 394
397 void unlink_anon_vmas(struct vm_area_struct *v 395 void unlink_anon_vmas(struct vm_area_struct *vma)
398 { 396 {
399 struct anon_vma_chain *avc, *next; 397 struct anon_vma_chain *avc, *next;
400 struct anon_vma *root = NULL; 398 struct anon_vma *root = NULL;
401 399
402 /* 400 /*
403 * Unlink each anon_vma chained to the 401 * Unlink each anon_vma chained to the VMA. This list is ordered
404 * from newest to oldest, ensuring the 402 * from newest to oldest, ensuring the root anon_vma gets freed last.
405 */ 403 */
406 list_for_each_entry_safe(avc, next, &v 404 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
407 struct anon_vma *anon_vma = av 405 struct anon_vma *anon_vma = avc->anon_vma;
408 406
409 root = lock_anon_vma_root(root 407 root = lock_anon_vma_root(root, anon_vma);
410 anon_vma_interval_tree_remove( 408 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
411 409
412 /* 410 /*
413 * Leave empty anon_vmas on th 411 * Leave empty anon_vmas on the list - we'll need
414 * to free them outside the lo 412 * to free them outside the lock.
415 */ 413 */
416 if (RB_EMPTY_ROOT(&anon_vma->r 414 if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
417 anon_vma->parent->num_ 415 anon_vma->parent->num_children--;
418 continue; 416 continue;
419 } 417 }
420 418
421 list_del(&avc->same_vma); 419 list_del(&avc->same_vma);
422 anon_vma_chain_free(avc); 420 anon_vma_chain_free(avc);
423 } 421 }
424 if (vma->anon_vma) { 422 if (vma->anon_vma) {
425 vma->anon_vma->num_active_vmas 423 vma->anon_vma->num_active_vmas--;
426 424
427 /* 425 /*
428 * vma would still be needed a 426 * vma would still be needed after unlink, and anon_vma will be prepared
429 * when handle fault. 427 * when handle fault.
430 */ 428 */
431 vma->anon_vma = NULL; 429 vma->anon_vma = NULL;
432 } 430 }
433 unlock_anon_vma_root(root); 431 unlock_anon_vma_root(root);
434 432
435 /* 433 /*
436 * Iterate the list once more, it now 434 * Iterate the list once more, it now only contains empty and unlinked
437 * anon_vmas, destroy them. Could not 435 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
438 * needing to write-acquire the anon_v 436 * needing to write-acquire the anon_vma->root->rwsem.
439 */ 437 */
440 list_for_each_entry_safe(avc, next, &v 438 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
441 struct anon_vma *anon_vma = av 439 struct anon_vma *anon_vma = avc->anon_vma;
442 440
443 VM_WARN_ON(anon_vma->num_child 441 VM_WARN_ON(anon_vma->num_children);
444 VM_WARN_ON(anon_vma->num_activ 442 VM_WARN_ON(anon_vma->num_active_vmas);
445 put_anon_vma(anon_vma); 443 put_anon_vma(anon_vma);
446 444
447 list_del(&avc->same_vma); 445 list_del(&avc->same_vma);
448 anon_vma_chain_free(avc); 446 anon_vma_chain_free(avc);
449 } 447 }
450 } 448 }
451 449
452 static void anon_vma_ctor(void *data) 450 static void anon_vma_ctor(void *data)
453 { 451 {
454 struct anon_vma *anon_vma = data; 452 struct anon_vma *anon_vma = data;
455 453
456 init_rwsem(&anon_vma->rwsem); 454 init_rwsem(&anon_vma->rwsem);
457 atomic_set(&anon_vma->refcount, 0); 455 atomic_set(&anon_vma->refcount, 0);
458 anon_vma->rb_root = RB_ROOT_CACHED; 456 anon_vma->rb_root = RB_ROOT_CACHED;
459 } 457 }
460 458
461 void __init anon_vma_init(void) 459 void __init anon_vma_init(void)
462 { 460 {
463 anon_vma_cachep = kmem_cache_create("a 461 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
464 0, SLAB_TYPESAFE_BY_RC 462 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
465 anon_vma_ctor); 463 anon_vma_ctor);
466 anon_vma_chain_cachep = KMEM_CACHE(ano 464 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
467 SLAB_PANIC|SLAB_ACCOUN 465 SLAB_PANIC|SLAB_ACCOUNT);
468 } 466 }
469 467
470 /* 468 /*
471 * Getting a lock on a stable anon_vma from a 469 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
472 * 470 *
473 * Since there is no serialization what so eve !! 471 * Since there is no serialization what so ever against page_remove_rmap()
474 * the best this function can do is return a r 472 * the best this function can do is return a refcount increased anon_vma
475 * that might have been relevant to this page. 473 * that might have been relevant to this page.
476 * 474 *
477 * The page might have been remapped to a diff 475 * The page might have been remapped to a different anon_vma or the anon_vma
478 * returned may already be freed (and even reu 476 * returned may already be freed (and even reused).
479 * 477 *
480 * In case it was remapped to a different anon 478 * In case it was remapped to a different anon_vma, the new anon_vma will be a
481 * child of the old anon_vma, and the anon_vma 479 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
482 * ensure that any anon_vma obtained from the 480 * ensure that any anon_vma obtained from the page will still be valid for as
483 * long as we observe page_mapped() [ hence al 481 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
484 * 482 *
485 * All users of this function must be very car 483 * All users of this function must be very careful when walking the anon_vma
486 * chain and verify that the page in question 484 * chain and verify that the page in question is indeed mapped in it
487 * [ something equivalent to page_mapped_in_vm 485 * [ something equivalent to page_mapped_in_vma() ].
488 * 486 *
489 * Since anon_vma's slab is SLAB_TYPESAFE_BY_R 487 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
490 * folio_remove_rmap_*() that the anon_vma poi !! 488 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
491 * if there is a mapcount, we can dereference 489 * if there is a mapcount, we can dereference the anon_vma after observing
492 * those. 490 * those.
493 * <<
494 * NOTE: the caller should normally hold folio <<
495 * not, the caller needs to double check the a <<
496 * taking the anon_vma lock for either read or <<
497 * concurrently without folio lock protection) <<
498 * which has already covered that, and comment <<
499 */ 491 */
500 struct anon_vma *folio_get_anon_vma(struct fol 492 struct anon_vma *folio_get_anon_vma(struct folio *folio)
501 { 493 {
502 struct anon_vma *anon_vma = NULL; 494 struct anon_vma *anon_vma = NULL;
503 unsigned long anon_mapping; 495 unsigned long anon_mapping;
504 496
505 rcu_read_lock(); 497 rcu_read_lock();
506 anon_mapping = (unsigned long)READ_ONC 498 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
507 if ((anon_mapping & PAGE_MAPPING_FLAGS 499 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
508 goto out; 500 goto out;
509 if (!folio_mapped(folio)) 501 if (!folio_mapped(folio))
510 goto out; 502 goto out;
511 503
512 anon_vma = (struct anon_vma *) (anon_m 504 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
513 if (!atomic_inc_not_zero(&anon_vma->re 505 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
514 anon_vma = NULL; 506 anon_vma = NULL;
515 goto out; 507 goto out;
516 } 508 }
517 509
518 /* 510 /*
519 * If this folio is still mapped, then 511 * If this folio is still mapped, then its anon_vma cannot have been
520 * freed. But if it has been unmapped 512 * freed. But if it has been unmapped, we have no security against the
521 * anon_vma structure being freed and 513 * anon_vma structure being freed and reused (for another anon_vma:
522 * SLAB_TYPESAFE_BY_RCU guarantees tha 514 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
523 * above cannot corrupt). 515 * above cannot corrupt).
524 */ 516 */
525 if (!folio_mapped(folio)) { 517 if (!folio_mapped(folio)) {
526 rcu_read_unlock(); 518 rcu_read_unlock();
527 put_anon_vma(anon_vma); 519 put_anon_vma(anon_vma);
528 return NULL; 520 return NULL;
529 } 521 }
530 out: 522 out:
531 rcu_read_unlock(); 523 rcu_read_unlock();
532 524
533 return anon_vma; 525 return anon_vma;
534 } 526 }
535 527
536 /* 528 /*
537 * Similar to folio_get_anon_vma() except it l 529 * Similar to folio_get_anon_vma() except it locks the anon_vma.
538 * 530 *
539 * Its a little more complex as it tries to ke 531 * Its a little more complex as it tries to keep the fast path to a single
540 * atomic op -- the trylock. If we fail the tr 532 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
541 * reference like with folio_get_anon_vma() an 533 * reference like with folio_get_anon_vma() and then block on the mutex
542 * on !rwc->try_lock case. 534 * on !rwc->try_lock case.
543 */ 535 */
544 struct anon_vma *folio_lock_anon_vma_read(stru 536 struct anon_vma *folio_lock_anon_vma_read(struct folio *folio,
545 stru 537 struct rmap_walk_control *rwc)
546 { 538 {
547 struct anon_vma *anon_vma = NULL; 539 struct anon_vma *anon_vma = NULL;
548 struct anon_vma *root_anon_vma; 540 struct anon_vma *root_anon_vma;
549 unsigned long anon_mapping; 541 unsigned long anon_mapping;
550 542
551 retry: <<
552 rcu_read_lock(); 543 rcu_read_lock();
553 anon_mapping = (unsigned long)READ_ONC 544 anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
554 if ((anon_mapping & PAGE_MAPPING_FLAGS 545 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
555 goto out; 546 goto out;
556 if (!folio_mapped(folio)) 547 if (!folio_mapped(folio))
557 goto out; 548 goto out;
558 549
559 anon_vma = (struct anon_vma *) (anon_m 550 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
560 root_anon_vma = READ_ONCE(anon_vma->ro 551 root_anon_vma = READ_ONCE(anon_vma->root);
561 if (down_read_trylock(&root_anon_vma-> 552 if (down_read_trylock(&root_anon_vma->rwsem)) {
562 /* 553 /*
563 * folio_move_anon_rmap() migh <<
564 * might not hold the folio lo <<
565 */ <<
566 if (unlikely((unsigned long)RE <<
567 anon_mapping)) { <<
568 up_read(&root_anon_vma <<
569 rcu_read_unlock(); <<
570 goto retry; <<
571 } <<
572 <<
573 /* <<
574 * If the folio is still mappe 554 * If the folio is still mapped, then this anon_vma is still
575 * its anon_vma, and holding t 555 * its anon_vma, and holding the mutex ensures that it will
576 * not go away, see anon_vma_f 556 * not go away, see anon_vma_free().
577 */ 557 */
578 if (!folio_mapped(folio)) { 558 if (!folio_mapped(folio)) {
579 up_read(&root_anon_vma 559 up_read(&root_anon_vma->rwsem);
580 anon_vma = NULL; 560 anon_vma = NULL;
581 } 561 }
582 goto out; 562 goto out;
583 } 563 }
584 564
585 if (rwc && rwc->try_lock) { 565 if (rwc && rwc->try_lock) {
586 anon_vma = NULL; 566 anon_vma = NULL;
587 rwc->contended = true; 567 rwc->contended = true;
588 goto out; 568 goto out;
589 } 569 }
590 570
591 /* trylock failed, we got to sleep */ 571 /* trylock failed, we got to sleep */
592 if (!atomic_inc_not_zero(&anon_vma->re 572 if (!atomic_inc_not_zero(&anon_vma->refcount)) {
593 anon_vma = NULL; 573 anon_vma = NULL;
594 goto out; 574 goto out;
595 } 575 }
596 576
597 if (!folio_mapped(folio)) { 577 if (!folio_mapped(folio)) {
598 rcu_read_unlock(); 578 rcu_read_unlock();
599 put_anon_vma(anon_vma); 579 put_anon_vma(anon_vma);
600 return NULL; 580 return NULL;
601 } 581 }
602 582
603 /* we pinned the anon_vma, its safe to 583 /* we pinned the anon_vma, its safe to sleep */
604 rcu_read_unlock(); 584 rcu_read_unlock();
605 anon_vma_lock_read(anon_vma); 585 anon_vma_lock_read(anon_vma);
606 586
607 /* <<
608 * folio_move_anon_rmap() might have c <<
609 * not hold the folio lock here. <<
610 */ <<
611 if (unlikely((unsigned long)READ_ONCE( <<
612 anon_mapping)) { <<
613 anon_vma_unlock_read(anon_vma) <<
614 put_anon_vma(anon_vma); <<
615 anon_vma = NULL; <<
616 goto retry; <<
617 } <<
618 <<
619 if (atomic_dec_and_test(&anon_vma->ref 587 if (atomic_dec_and_test(&anon_vma->refcount)) {
620 /* 588 /*
621 * Oops, we held the last refc 589 * Oops, we held the last refcount, release the lock
622 * and bail -- can't simply us 590 * and bail -- can't simply use put_anon_vma() because
623 * we'll deadlock on the anon_ 591 * we'll deadlock on the anon_vma_lock_write() recursion.
624 */ 592 */
625 anon_vma_unlock_read(anon_vma) 593 anon_vma_unlock_read(anon_vma);
626 __put_anon_vma(anon_vma); 594 __put_anon_vma(anon_vma);
627 anon_vma = NULL; 595 anon_vma = NULL;
628 } 596 }
629 597
630 return anon_vma; 598 return anon_vma;
631 599
632 out: 600 out:
633 rcu_read_unlock(); 601 rcu_read_unlock();
634 return anon_vma; 602 return anon_vma;
635 } 603 }
636 604
637 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUS 605 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
638 /* 606 /*
639 * Flush TLB entries for recently unmapped pag 607 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
640 * important if a PTE was dirty when it was un 608 * important if a PTE was dirty when it was unmapped that it's flushed
641 * before any IO is initiated on the page to p 609 * before any IO is initiated on the page to prevent lost writes. Similarly,
642 * it must be flushed before freeing to preven 610 * it must be flushed before freeing to prevent data leakage.
643 */ 611 */
644 void try_to_unmap_flush(void) 612 void try_to_unmap_flush(void)
645 { 613 {
646 struct tlbflush_unmap_batch *tlb_ubc = 614 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
647 615
648 if (!tlb_ubc->flush_required) 616 if (!tlb_ubc->flush_required)
649 return; 617 return;
650 618
651 arch_tlbbatch_flush(&tlb_ubc->arch); 619 arch_tlbbatch_flush(&tlb_ubc->arch);
652 tlb_ubc->flush_required = false; 620 tlb_ubc->flush_required = false;
653 tlb_ubc->writable = false; 621 tlb_ubc->writable = false;
654 } 622 }
655 623
656 /* Flush iff there are potentially writable TL 624 /* Flush iff there are potentially writable TLB entries that can race with IO */
657 void try_to_unmap_flush_dirty(void) 625 void try_to_unmap_flush_dirty(void)
658 { 626 {
659 struct tlbflush_unmap_batch *tlb_ubc = 627 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
660 628
661 if (tlb_ubc->writable) 629 if (tlb_ubc->writable)
662 try_to_unmap_flush(); 630 try_to_unmap_flush();
663 } 631 }
664 632
665 /* 633 /*
666 * Bits 0-14 of mm->tlb_flush_batched record p 634 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
667 * Bits 16-30 of mm->tlb_flush_batched bit rec 635 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
668 */ 636 */
669 #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16 637 #define TLB_FLUSH_BATCH_FLUSHED_SHIFT 16
670 #define TLB_FLUSH_BATCH_PENDING_MASK 638 #define TLB_FLUSH_BATCH_PENDING_MASK \
671 ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT 639 ((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
672 #define TLB_FLUSH_BATCH_PENDING_LARGE 640 #define TLB_FLUSH_BATCH_PENDING_LARGE \
673 (TLB_FLUSH_BATCH_PENDING_MASK / 2) 641 (TLB_FLUSH_BATCH_PENDING_MASK / 2)
674 642
675 static void set_tlb_ubc_flush_pending(struct m !! 643 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
676 unsigned <<
677 { 644 {
678 struct tlbflush_unmap_batch *tlb_ubc = 645 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
679 int batch; !! 646 int batch, nbatch;
680 bool writable = pte_dirty(pteval); <<
681 <<
682 if (!pte_accessible(mm, pteval)) <<
683 return; <<
684 647
685 arch_tlbbatch_add_pending(&tlb_ubc->ar !! 648 arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
686 tlb_ubc->flush_required = true; 649 tlb_ubc->flush_required = true;
687 650
688 /* 651 /*
689 * Ensure compiler does not re-order t 652 * Ensure compiler does not re-order the setting of tlb_flush_batched
690 * before the PTE is cleared. 653 * before the PTE is cleared.
691 */ 654 */
692 barrier(); 655 barrier();
693 batch = atomic_read(&mm->tlb_flush_bat 656 batch = atomic_read(&mm->tlb_flush_batched);
694 retry: 657 retry:
695 if ((batch & TLB_FLUSH_BATCH_PENDING_M 658 if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
696 /* 659 /*
697 * Prevent `pending' from catc 660 * Prevent `pending' from catching up with `flushed' because of
698 * overflow. Reset `pending' 661 * overflow. Reset `pending' and `flushed' to be 1 and 0 if
699 * `pending' becomes large. 662 * `pending' becomes large.
700 */ 663 */
701 if (!atomic_try_cmpxchg(&mm->t !! 664 nbatch = atomic_cmpxchg(&mm->tlb_flush_batched, batch, 1);
>> 665 if (nbatch != batch) {
>> 666 batch = nbatch;
702 goto retry; 667 goto retry;
>> 668 }
703 } else { 669 } else {
704 atomic_inc(&mm->tlb_flush_batc 670 atomic_inc(&mm->tlb_flush_batched);
705 } 671 }
706 672
707 /* 673 /*
708 * If the PTE was dirty then it's best 674 * If the PTE was dirty then it's best to assume it's writable. The
709 * caller must use try_to_unmap_flush_ 675 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
710 * before the page is queued for IO. 676 * before the page is queued for IO.
711 */ 677 */
712 if (writable) 678 if (writable)
713 tlb_ubc->writable = true; 679 tlb_ubc->writable = true;
714 } 680 }
715 681
716 /* 682 /*
717 * Returns true if the TLB flush should be def 683 * Returns true if the TLB flush should be deferred to the end of a batch of
718 * unmap operations to reduce IPIs. 684 * unmap operations to reduce IPIs.
719 */ 685 */
720 static bool should_defer_flush(struct mm_struc 686 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
721 { 687 {
>> 688 bool should_defer = false;
>> 689
722 if (!(flags & TTU_BATCH_FLUSH)) 690 if (!(flags & TTU_BATCH_FLUSH))
723 return false; 691 return false;
724 692
725 return arch_tlbbatch_should_defer(mm); !! 693 /* If remote CPUs need to be flushed then defer batch the flush */
>> 694 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
>> 695 should_defer = true;
>> 696 put_cpu();
>> 697
>> 698 return should_defer;
726 } 699 }
727 700
728 /* 701 /*
729 * Reclaim unmaps pages under the PTL but do n 702 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
730 * releasing the PTL if TLB flushes are batche 703 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
731 * operation such as mprotect or munmap to rac 704 * operation such as mprotect or munmap to race between reclaim unmapping
732 * the page and flushing the page. If this rac 705 * the page and flushing the page. If this race occurs, it potentially allows
733 * access to data via a stale TLB entry. Track 706 * access to data via a stale TLB entry. Tracking all mm's that have TLB
734 * batching in flight would be expensive durin 707 * batching in flight would be expensive during reclaim so instead track
735 * whether TLB batching occurred in the past a 708 * whether TLB batching occurred in the past and if so then do a flush here
736 * if required. This will cost one additional 709 * if required. This will cost one additional flush per reclaim cycle paid
737 * by the first operation at risk such as mpro 710 * by the first operation at risk such as mprotect and mumap.
738 * 711 *
739 * This must be called under the PTL so that a 712 * This must be called under the PTL so that an access to tlb_flush_batched
740 * that is potentially a "reclaim vs mprotect/ 713 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
741 * via the PTL. 714 * via the PTL.
742 */ 715 */
743 void flush_tlb_batched_pending(struct mm_struc 716 void flush_tlb_batched_pending(struct mm_struct *mm)
744 { 717 {
745 int batch = atomic_read(&mm->tlb_flush 718 int batch = atomic_read(&mm->tlb_flush_batched);
746 int pending = batch & TLB_FLUSH_BATCH_ 719 int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
747 int flushed = batch >> TLB_FLUSH_BATCH 720 int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
748 721
749 if (pending != flushed) { 722 if (pending != flushed) {
750 arch_flush_tlb_batched_pending !! 723 flush_tlb_mm(mm);
751 /* 724 /*
752 * If the new TLB flushing is 725 * If the new TLB flushing is pending during flushing, leave
753 * mm->tlb_flush_batched as is 726 * mm->tlb_flush_batched as is, to avoid losing flushing.
754 */ 727 */
755 atomic_cmpxchg(&mm->tlb_flush_ 728 atomic_cmpxchg(&mm->tlb_flush_batched, batch,
756 pending | (pend 729 pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
757 } 730 }
758 } 731 }
759 #else 732 #else
760 static void set_tlb_ubc_flush_pending(struct m !! 733 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
761 unsigned <<
762 { 734 {
763 } 735 }
764 736
765 static bool should_defer_flush(struct mm_struc 737 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
766 { 738 {
767 return false; 739 return false;
768 } 740 }
769 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_F 741 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
770 742
771 /* 743 /*
772 * At what user virtual address is page expect 744 * At what user virtual address is page expected in vma?
773 * Caller should check the page is actually pa 745 * Caller should check the page is actually part of the vma.
774 */ 746 */
775 unsigned long page_address_in_vma(struct page 747 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
776 { 748 {
777 struct folio *folio = page_folio(page) 749 struct folio *folio = page_folio(page);
778 pgoff_t pgoff; <<
779 <<
780 if (folio_test_anon(folio)) { 750 if (folio_test_anon(folio)) {
781 struct anon_vma *page__anon_vm 751 struct anon_vma *page__anon_vma = folio_anon_vma(folio);
782 /* 752 /*
783 * Note: swapoff's unuse_vma() 753 * Note: swapoff's unuse_vma() is more efficient with this
784 * check, and needs it to matc 754 * check, and needs it to match anon_vma when KSM is active.
785 */ 755 */
786 if (!vma->anon_vma || !page__a 756 if (!vma->anon_vma || !page__anon_vma ||
787 vma->anon_vma->root != pag 757 vma->anon_vma->root != page__anon_vma->root)
788 return -EFAULT; 758 return -EFAULT;
789 } else if (!vma->vm_file) { 759 } else if (!vma->vm_file) {
790 return -EFAULT; 760 return -EFAULT;
791 } else if (vma->vm_file->f_mapping != 761 } else if (vma->vm_file->f_mapping != folio->mapping) {
792 return -EFAULT; 762 return -EFAULT;
793 } 763 }
794 764
795 /* The !page__anon_vma above handles K !! 765 return vma_address(page, vma);
796 pgoff = folio->index + folio_page_idx( <<
797 return vma_address(vma, pgoff, 1); <<
798 } 766 }
799 767
800 /* 768 /*
801 * Returns the actual pmd_t* where we expect ' 769 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
802 * NULL if it doesn't exist. No guarantees / 770 * NULL if it doesn't exist. No guarantees / checks on what the pmd_t*
803 * represents. 771 * represents.
804 */ 772 */
805 pmd_t *mm_find_pmd(struct mm_struct *mm, unsig 773 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
806 { 774 {
807 pgd_t *pgd; 775 pgd_t *pgd;
808 p4d_t *p4d; 776 p4d_t *p4d;
809 pud_t *pud; 777 pud_t *pud;
810 pmd_t *pmd = NULL; 778 pmd_t *pmd = NULL;
811 779
812 pgd = pgd_offset(mm, address); 780 pgd = pgd_offset(mm, address);
813 if (!pgd_present(*pgd)) 781 if (!pgd_present(*pgd))
814 goto out; 782 goto out;
815 783
816 p4d = p4d_offset(pgd, address); 784 p4d = p4d_offset(pgd, address);
817 if (!p4d_present(*p4d)) 785 if (!p4d_present(*p4d))
818 goto out; 786 goto out;
819 787
820 pud = pud_offset(p4d, address); 788 pud = pud_offset(p4d, address);
821 if (!pud_present(*pud)) 789 if (!pud_present(*pud))
822 goto out; 790 goto out;
823 791
824 pmd = pmd_offset(pud, address); 792 pmd = pmd_offset(pud, address);
825 out: 793 out:
826 return pmd; 794 return pmd;
827 } 795 }
828 796
829 struct folio_referenced_arg { 797 struct folio_referenced_arg {
830 int mapcount; 798 int mapcount;
831 int referenced; 799 int referenced;
832 unsigned long vm_flags; 800 unsigned long vm_flags;
833 struct mem_cgroup *memcg; 801 struct mem_cgroup *memcg;
834 }; 802 };
835 <<
836 /* 803 /*
837 * arg: folio_referenced_arg will be passed 804 * arg: folio_referenced_arg will be passed
838 */ 805 */
839 static bool folio_referenced_one(struct folio 806 static bool folio_referenced_one(struct folio *folio,
840 struct vm_area_struct *vma, un 807 struct vm_area_struct *vma, unsigned long address, void *arg)
841 { 808 {
842 struct folio_referenced_arg *pra = arg 809 struct folio_referenced_arg *pra = arg;
843 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma 810 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
844 int referenced = 0; 811 int referenced = 0;
845 unsigned long start = address, ptes = <<
846 812
847 while (page_vma_mapped_walk(&pvmw)) { 813 while (page_vma_mapped_walk(&pvmw)) {
848 address = pvmw.address; 814 address = pvmw.address;
849 815
850 if (vma->vm_flags & VM_LOCKED) !! 816 if ((vma->vm_flags & VM_LOCKED) &&
851 if (!folio_test_large( !! 817 (!folio_test_large(folio) || !pvmw.pte)) {
852 /* Restore the !! 818 /* Restore the mlock which got missed */
853 mlock_vma_foli !! 819 mlock_vma_folio(folio, vma, !pvmw.pte);
854 page_vma_mappe <<
855 pra->vm_flags <<
856 return false; <<
857 } <<
858 /* <<
859 * For large folio ful <<
860 * be handled after th <<
861 * <<
862 * For large folio cro <<
863 * expected to be pick <<
864 * should skip referen <<
865 * the range of VM_LOC <<
866 * should just count t <<
867 * the range of VM_LOC <<
868 */ <<
869 ptes++; <<
870 pra->mapcount--; <<
871 continue; <<
872 } <<
873 <<
874 /* <<
875 * Skip the non-shared swapbac <<
876 * the exiting or OOM-reaped p <<
877 * swap-out followed by an imm <<
878 */ <<
879 if ((!atomic_read(&vma->vm_mm- <<
880 check_stable_address_space <<
881 folio_test_anon(folio) && <<
882 !folio_likely_mapped_share <<
883 pra->referenced = -1; <<
884 page_vma_mapped_walk_d 820 page_vma_mapped_walk_done(&pvmw);
885 return false; !! 821 pra->vm_flags |= VM_LOCKED;
>> 822 return false; /* To break the loop */
886 } 823 }
887 824
888 if (lru_gen_enabled() && pvmw. !! 825 if (pvmw.pte) {
889 if (lru_gen_look_aroun !! 826 if (lru_gen_enabled() && pte_young(*pvmw.pte) &&
>> 827 !(vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))) {
>> 828 lru_gen_look_around(&pvmw);
890 referenced++; 829 referenced++;
891 } else if (pvmw.pte) { !! 830 }
>> 831
892 if (ptep_clear_flush_y 832 if (ptep_clear_flush_young_notify(vma, address,
893 !! 833 pvmw.pte)) {
894 referenced++; !! 834 /*
>> 835 * Don't treat a reference through
>> 836 * a sequentially read mapping as such.
>> 837 * If the folio has been used in another mapping,
>> 838 * we will catch it; if this other mapping is
>> 839 * already gone, the unmap path will have set
>> 840 * the referenced flag or activated the folio.
>> 841 */
>> 842 if (likely(!(vma->vm_flags & VM_SEQ_READ)))
>> 843 referenced++;
>> 844 }
895 } else if (IS_ENABLED(CONFIG_T 845 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
896 if (pmdp_clear_flush_y 846 if (pmdp_clear_flush_young_notify(vma, address,
897 847 pvmw.pmd))
898 referenced++; 848 referenced++;
899 } else { 849 } else {
900 /* unexpected pmd-mapp 850 /* unexpected pmd-mapped folio? */
901 WARN_ON_ONCE(1); 851 WARN_ON_ONCE(1);
902 } 852 }
903 853
904 pra->mapcount--; 854 pra->mapcount--;
905 } 855 }
906 856
907 if ((vma->vm_flags & VM_LOCKED) && <<
908 folio_test_large(folio <<
909 folio_within_vma(folio <<
910 unsigned long s_align, e_align <<
911 <<
912 s_align = ALIGN_DOWN(start, PM <<
913 e_align = ALIGN_DOWN(start + f <<
914 <<
915 /* folio doesn't cross page ta <<
916 if ((s_align == e_align) && (p <<
917 /* Restore the mlock w <<
918 mlock_vma_folio(folio, <<
919 pra->vm_flags |= VM_LO <<
920 return false; /* To br <<
921 } <<
922 } <<
923 <<
924 if (referenced) 857 if (referenced)
925 folio_clear_idle(folio); 858 folio_clear_idle(folio);
926 if (folio_test_clear_young(folio)) 859 if (folio_test_clear_young(folio))
927 referenced++; 860 referenced++;
928 861
929 if (referenced) { 862 if (referenced) {
930 pra->referenced++; 863 pra->referenced++;
931 pra->vm_flags |= vma->vm_flags 864 pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
932 } 865 }
933 866
934 if (!pra->mapcount) 867 if (!pra->mapcount)
935 return false; /* To break the 868 return false; /* To break the loop */
936 869
937 return true; 870 return true;
938 } 871 }
939 872
940 static bool invalid_folio_referenced_vma(struc 873 static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
941 { 874 {
942 struct folio_referenced_arg *pra = arg 875 struct folio_referenced_arg *pra = arg;
943 struct mem_cgroup *memcg = pra->memcg; 876 struct mem_cgroup *memcg = pra->memcg;
944 877
945 /* !! 878 if (!mm_match_cgroup(vma->vm_mm, memcg))
946 * Ignore references from this mapping <<
947 * folio has been used in another mapp <<
948 * other mapping is already gone, the <<
949 * referenced flag or activated the fo <<
950 */ <<
951 if (!vma_has_recency(vma)) <<
952 return true; <<
953 <<
954 /* <<
955 * If we are reclaiming on behalf of a <<
956 * of references from different cgroup <<
957 */ <<
958 if (memcg && !mm_match_cgroup(vma->vm_ <<
959 return true; 879 return true;
960 880
961 return false; 881 return false;
962 } 882 }
963 883
964 /** 884 /**
965 * folio_referenced() - Test if the folio was 885 * folio_referenced() - Test if the folio was referenced.
966 * @folio: The folio to test. 886 * @folio: The folio to test.
967 * @is_locked: Caller holds lock on the folio. 887 * @is_locked: Caller holds lock on the folio.
968 * @memcg: target memory cgroup 888 * @memcg: target memory cgroup
969 * @vm_flags: A combination of all the vma->vm 889 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
970 * 890 *
971 * Quick test_and_clear_referenced for all map 891 * Quick test_and_clear_referenced for all mappings of a folio,
972 * 892 *
973 * Return: The number of mappings which refere 893 * Return: The number of mappings which referenced the folio. Return -1 if
974 * the function bailed out due to rmap lock co 894 * the function bailed out due to rmap lock contention.
975 */ 895 */
976 int folio_referenced(struct folio *folio, int 896 int folio_referenced(struct folio *folio, int is_locked,
977 struct mem_cgroup *memcg, 897 struct mem_cgroup *memcg, unsigned long *vm_flags)
978 { 898 {
979 bool we_locked = false; !! 899 int we_locked = 0;
980 struct folio_referenced_arg pra = { 900 struct folio_referenced_arg pra = {
981 .mapcount = folio_mapcount(fol 901 .mapcount = folio_mapcount(folio),
982 .memcg = memcg, 902 .memcg = memcg,
983 }; 903 };
984 struct rmap_walk_control rwc = { 904 struct rmap_walk_control rwc = {
985 .rmap_one = folio_referenced_o 905 .rmap_one = folio_referenced_one,
986 .arg = (void *)&pra, 906 .arg = (void *)&pra,
987 .anon_lock = folio_lock_anon_v 907 .anon_lock = folio_lock_anon_vma_read,
988 .try_lock = true, 908 .try_lock = true,
989 .invalid_vma = invalid_folio_r <<
990 }; 909 };
991 910
992 *vm_flags = 0; 911 *vm_flags = 0;
993 if (!pra.mapcount) 912 if (!pra.mapcount)
994 return 0; 913 return 0;
995 914
996 if (!folio_raw_mapping(folio)) 915 if (!folio_raw_mapping(folio))
997 return 0; 916 return 0;
998 917
999 if (!is_locked && (!folio_test_anon(fo 918 if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
1000 we_locked = folio_trylock(fol 919 we_locked = folio_trylock(folio);
1001 if (!we_locked) 920 if (!we_locked)
1002 return 1; 921 return 1;
1003 } 922 }
1004 923
>> 924 /*
>> 925 * If we are reclaiming on behalf of a cgroup, skip
>> 926 * counting on behalf of references from different
>> 927 * cgroups
>> 928 */
>> 929 if (memcg) {
>> 930 rwc.invalid_vma = invalid_folio_referenced_vma;
>> 931 }
>> 932
1005 rmap_walk(folio, &rwc); 933 rmap_walk(folio, &rwc);
1006 *vm_flags = pra.vm_flags; 934 *vm_flags = pra.vm_flags;
1007 935
1008 if (we_locked) 936 if (we_locked)
1009 folio_unlock(folio); 937 folio_unlock(folio);
1010 938
1011 return rwc.contended ? -1 : pra.refer 939 return rwc.contended ? -1 : pra.referenced;
1012 } 940 }
1013 941
1014 static int page_vma_mkclean_one(struct page_v 942 static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
1015 { 943 {
1016 int cleaned = 0; 944 int cleaned = 0;
1017 struct vm_area_struct *vma = pvmw->vm 945 struct vm_area_struct *vma = pvmw->vma;
1018 struct mmu_notifier_range range; 946 struct mmu_notifier_range range;
1019 unsigned long address = pvmw->address 947 unsigned long address = pvmw->address;
1020 948
1021 /* 949 /*
1022 * We have to assume the worse case i 950 * We have to assume the worse case ie pmd for invalidation. Note that
1023 * the folio can not be freed from th 951 * the folio can not be freed from this function.
1024 */ 952 */
1025 mmu_notifier_range_init(&range, MMU_N !! 953 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1026 vma->vm_mm, a !! 954 0, vma, vma->vm_mm, address,
>> 955 vma_address_end(pvmw));
1027 mmu_notifier_invalidate_range_start(& 956 mmu_notifier_invalidate_range_start(&range);
1028 957
1029 while (page_vma_mapped_walk(pvmw)) { 958 while (page_vma_mapped_walk(pvmw)) {
1030 int ret = 0; 959 int ret = 0;
1031 960
1032 address = pvmw->address; 961 address = pvmw->address;
1033 if (pvmw->pte) { 962 if (pvmw->pte) {
>> 963 pte_t entry;
1034 pte_t *pte = pvmw->pt 964 pte_t *pte = pvmw->pte;
1035 pte_t entry = ptep_ge <<
1036 965
1037 if (!pte_dirty(entry) !! 966 if (!pte_dirty(*pte) && !pte_write(*pte))
1038 continue; 967 continue;
1039 968
1040 flush_cache_page(vma, !! 969 flush_cache_page(vma, address, pte_pfn(*pte));
1041 entry = ptep_clear_fl 970 entry = ptep_clear_flush(vma, address, pte);
1042 entry = pte_wrprotect 971 entry = pte_wrprotect(entry);
1043 entry = pte_mkclean(e 972 entry = pte_mkclean(entry);
1044 set_pte_at(vma->vm_mm 973 set_pte_at(vma->vm_mm, address, pte, entry);
1045 ret = 1; 974 ret = 1;
1046 } else { 975 } else {
1047 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 976 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1048 pmd_t *pmd = pvmw->pm 977 pmd_t *pmd = pvmw->pmd;
1049 pmd_t entry; 978 pmd_t entry;
1050 979
1051 if (!pmd_dirty(*pmd) 980 if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
1052 continue; 981 continue;
1053 982
1054 flush_cache_range(vma 983 flush_cache_range(vma, address,
1055 add 984 address + HPAGE_PMD_SIZE);
1056 entry = pmdp_invalida 985 entry = pmdp_invalidate(vma, address, pmd);
1057 entry = pmd_wrprotect 986 entry = pmd_wrprotect(entry);
1058 entry = pmd_mkclean(e 987 entry = pmd_mkclean(entry);
1059 set_pmd_at(vma->vm_mm 988 set_pmd_at(vma->vm_mm, address, pmd, entry);
1060 ret = 1; 989 ret = 1;
1061 #else 990 #else
1062 /* unexpected pmd-map 991 /* unexpected pmd-mapped folio? */
1063 WARN_ON_ONCE(1); 992 WARN_ON_ONCE(1);
1064 #endif 993 #endif
1065 } 994 }
1066 995
>> 996 /*
>> 997 * No need to call mmu_notifier_invalidate_range() as we are
>> 998 * downgrading page table protection not changing it to point
>> 999 * to a new page.
>> 1000 *
>> 1001 * See Documentation/mm/mmu_notifier.rst
>> 1002 */
1067 if (ret) 1003 if (ret)
1068 cleaned++; 1004 cleaned++;
1069 } 1005 }
1070 1006
1071 mmu_notifier_invalidate_range_end(&ra 1007 mmu_notifier_invalidate_range_end(&range);
1072 1008
1073 return cleaned; 1009 return cleaned;
1074 } 1010 }
1075 1011
1076 static bool page_mkclean_one(struct folio *fo 1012 static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1077 unsigned long ad 1013 unsigned long address, void *arg)
1078 { 1014 {
1079 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vm 1015 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1080 int *cleaned = arg; 1016 int *cleaned = arg;
1081 1017
1082 *cleaned += page_vma_mkclean_one(&pvm 1018 *cleaned += page_vma_mkclean_one(&pvmw);
1083 1019
1084 return true; 1020 return true;
1085 } 1021 }
1086 1022
1087 static bool invalid_mkclean_vma(struct vm_are 1023 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1088 { 1024 {
1089 if (vma->vm_flags & VM_SHARED) 1025 if (vma->vm_flags & VM_SHARED)
1090 return false; 1026 return false;
1091 1027
1092 return true; 1028 return true;
1093 } 1029 }
1094 1030
1095 int folio_mkclean(struct folio *folio) 1031 int folio_mkclean(struct folio *folio)
1096 { 1032 {
1097 int cleaned = 0; 1033 int cleaned = 0;
1098 struct address_space *mapping; 1034 struct address_space *mapping;
1099 struct rmap_walk_control rwc = { 1035 struct rmap_walk_control rwc = {
1100 .arg = (void *)&cleaned, 1036 .arg = (void *)&cleaned,
1101 .rmap_one = page_mkclean_one, 1037 .rmap_one = page_mkclean_one,
1102 .invalid_vma = invalid_mkclea 1038 .invalid_vma = invalid_mkclean_vma,
1103 }; 1039 };
1104 1040
1105 BUG_ON(!folio_test_locked(folio)); 1041 BUG_ON(!folio_test_locked(folio));
1106 1042
1107 if (!folio_mapped(folio)) 1043 if (!folio_mapped(folio))
1108 return 0; 1044 return 0;
1109 1045
1110 mapping = folio_mapping(folio); 1046 mapping = folio_mapping(folio);
1111 if (!mapping) 1047 if (!mapping)
1112 return 0; 1048 return 0;
1113 1049
1114 rmap_walk(folio, &rwc); 1050 rmap_walk(folio, &rwc);
1115 1051
1116 return cleaned; 1052 return cleaned;
1117 } 1053 }
1118 EXPORT_SYMBOL_GPL(folio_mkclean); 1054 EXPORT_SYMBOL_GPL(folio_mkclean);
1119 1055
1120 /** 1056 /**
1121 * pfn_mkclean_range - Cleans the PTEs (inclu 1057 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1122 * [@pfn, @pfn + @nr_page 1058 * [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1123 * within the @vma of sha 1059 * within the @vma of shared mappings. And since clean PTEs
1124 * should also be readonl 1060 * should also be readonly, write protects them too.
1125 * @pfn: start pfn. 1061 * @pfn: start pfn.
1126 * @nr_pages: number of physically contiguous 1062 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1127 * @pgoff: page offset that the @pfn mapped w 1063 * @pgoff: page offset that the @pfn mapped with.
1128 * @vma: vma that @pfn mapped within. 1064 * @vma: vma that @pfn mapped within.
1129 * 1065 *
1130 * Returns the number of cleaned PTEs (includ 1066 * Returns the number of cleaned PTEs (including PMDs).
1131 */ 1067 */
1132 int pfn_mkclean_range(unsigned long pfn, unsi 1068 int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1133 struct vm_area_struct * 1069 struct vm_area_struct *vma)
1134 { 1070 {
1135 struct page_vma_mapped_walk pvmw = { 1071 struct page_vma_mapped_walk pvmw = {
1136 .pfn = pfn, 1072 .pfn = pfn,
1137 .nr_pages = nr_pages, 1073 .nr_pages = nr_pages,
1138 .pgoff = pgoff, 1074 .pgoff = pgoff,
1139 .vma = vma, 1075 .vma = vma,
1140 .flags = PVMW_SYNC, 1076 .flags = PVMW_SYNC,
1141 }; 1077 };
1142 1078
1143 if (invalid_mkclean_vma(vma, NULL)) 1079 if (invalid_mkclean_vma(vma, NULL))
1144 return 0; 1080 return 0;
1145 1081
1146 pvmw.address = vma_address(vma, pgoff !! 1082 pvmw.address = vma_pgoff_address(pgoff, nr_pages, vma);
1147 VM_BUG_ON_VMA(pvmw.address == -EFAULT 1083 VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1148 1084
1149 return page_vma_mkclean_one(&pvmw); 1085 return page_vma_mkclean_one(&pvmw);
1150 } 1086 }
1151 1087
1152 static __always_inline unsigned int __folio_a !! 1088 int total_compound_mapcount(struct page *head)
1153 struct page *page, int nr_pag !! 1089 {
1154 int *nr_pmdmapped) !! 1090 int mapcount = head_compound_mapcount(head);
1155 { !! 1091 int nr_subpages;
1156 atomic_t *mapped = &folio->_nr_pages_ !! 1092 int i;
1157 const int orig_nr_pages = nr_pages; !! 1093
1158 int first = 0, nr = 0; !! 1094 /* In the common case, avoid the loop when no subpages mapped by PTE */
1159 !! 1095 if (head_subpages_mapcount(head) == 0)
1160 __folio_rmap_sanity_checks(folio, pag !! 1096 return mapcount;
1161 !! 1097 /*
1162 switch (level) { !! 1098 * Add all the PTE mappings of those subpages mapped by PTE.
1163 case RMAP_LEVEL_PTE: !! 1099 * Limit the loop, knowing that only subpages_mapcount are mapped?
1164 if (!folio_test_large(folio)) !! 1100 * Perhaps: given all the raciness, that may be a good or a bad idea.
1165 nr = atomic_inc_and_t !! 1101 */
1166 break; !! 1102 nr_subpages = thp_nr_pages(head);
1167 } !! 1103 for (i = 0; i < nr_subpages; i++)
1168 !! 1104 mapcount += atomic_read(&head[i]._mapcount);
1169 do { !! 1105
1170 first += atomic_inc_a !! 1106 /* But each of those _mapcounts was based on -1 */
1171 } while (page++, --nr_pages > !! 1107 mapcount += nr_subpages;
1172 !! 1108 return mapcount;
1173 if (first && <<
1174 atomic_add_return_relaxed <<
1175 nr = first; <<
1176 <<
1177 atomic_add(orig_nr_pages, &fo <<
1178 break; <<
1179 case RMAP_LEVEL_PMD: <<
1180 first = atomic_inc_and_test(& <<
1181 if (first) { <<
1182 nr = atomic_add_retur <<
1183 if (likely(nr < ENTIR <<
1184 *nr_pmdmapped <<
1185 nr = *nr_pmdm <<
1186 /* Raced ahea <<
1187 if (unlikely( <<
1188 nr = <<
1189 } else { <<
1190 /* Raced ahea <<
1191 nr = 0; <<
1192 } <<
1193 } <<
1194 atomic_inc(&folio->_large_map <<
1195 break; <<
1196 } <<
1197 return nr; <<
1198 } 1109 }
1199 1110
1200 /** 1111 /**
1201 * folio_move_anon_rmap - move a folio to our !! 1112 * page_move_anon_rmap - move a page to our anon_vma
1202 * @folio: The folio to move to our anon !! 1113 * @page: the page to move to our anon_vma
1203 * @vma: The vma the folio belongs to !! 1114 * @vma: the vma the page belongs to
1204 * !! 1115 *
1205 * When a folio belongs exclusively to one pr !! 1116 * When a page belongs exclusively to one process after a COW event,
1206 * that folio can be moved into the anon_vma !! 1117 * that page can be moved into the anon_vma that belongs to just that
1207 * process, so the rmap code will not search !! 1118 * process, so the rmap code will not search the parent or sibling
>> 1119 * processes.
1208 */ 1120 */
1209 void folio_move_anon_rmap(struct folio *folio !! 1121 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1210 { 1122 {
1211 void *anon_vma = vma->anon_vma; 1123 void *anon_vma = vma->anon_vma;
>> 1124 struct folio *folio = page_folio(page);
1212 1125
1213 VM_BUG_ON_FOLIO(!folio_test_locked(fo 1126 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1214 VM_BUG_ON_VMA(!anon_vma, vma); 1127 VM_BUG_ON_VMA(!anon_vma, vma);
1215 1128
1216 anon_vma += PAGE_MAPPING_ANON; 1129 anon_vma += PAGE_MAPPING_ANON;
1217 /* 1130 /*
1218 * Ensure that anon_vma and the PAGE_ 1131 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1219 * simultaneously, so a concurrent re 1132 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1220 * folio_test_anon()) will not see on 1133 * folio_test_anon()) will not see one without the other.
1221 */ 1134 */
1222 WRITE_ONCE(folio->mapping, anon_vma); 1135 WRITE_ONCE(folio->mapping, anon_vma);
>> 1136 SetPageAnonExclusive(page);
1223 } 1137 }
1224 1138
1225 /** 1139 /**
1226 * __folio_set_anon - set up a new anonymous !! 1140 * __page_set_anon_rmap - set up new anonymous rmap
1227 * @folio: The folio to set up the new a !! 1141 * @page: Page or Hugepage to add to rmap
1228 * @vma: VM area to add the folio to. !! 1142 * @vma: VM area to add page to.
1229 * @address: User virtual address of the m !! 1143 * @address: User virtual address of the mapping
1230 * @exclusive: Whether the folio is exclusiv !! 1144 * @exclusive: the page is exclusively owned by the current process
1231 */ 1145 */
1232 static void __folio_set_anon(struct folio *fo !! 1146 static void __page_set_anon_rmap(struct page *page,
1233 unsigned long ad !! 1147 struct vm_area_struct *vma, unsigned long address, int exclusive)
1234 { 1148 {
1235 struct anon_vma *anon_vma = vma->anon 1149 struct anon_vma *anon_vma = vma->anon_vma;
1236 1150
1237 BUG_ON(!anon_vma); 1151 BUG_ON(!anon_vma);
1238 1152
>> 1153 if (PageAnon(page))
>> 1154 goto out;
>> 1155
1239 /* 1156 /*
1240 * If the folio isn't exclusive to th !! 1157 * If the page isn't exclusively mapped into this vma,
1241 * possible anon_vma for the folio ma !! 1158 * we must use the _oldest_ possible anon_vma for the
>> 1159 * page mapping!
1242 */ 1160 */
1243 if (!exclusive) 1161 if (!exclusive)
1244 anon_vma = anon_vma->root; 1162 anon_vma = anon_vma->root;
1245 1163
1246 /* 1164 /*
1247 * page_idle does a lockless/optimist !! 1165 * page_idle does a lockless/optimistic rmap scan on page->mapping.
1248 * Make sure the compiler doesn't spl 1166 * Make sure the compiler doesn't split the stores of anon_vma and
1249 * the PAGE_MAPPING_ANON type identif 1167 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1250 * could mistake the mapping for a st 1168 * could mistake the mapping for a struct address_space and crash.
1251 */ 1169 */
1252 anon_vma = (void *) anon_vma + PAGE_M 1170 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1253 WRITE_ONCE(folio->mapping, (struct ad !! 1171 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1254 folio->index = linear_page_index(vma, !! 1172 page->index = linear_page_index(vma, address);
>> 1173 out:
>> 1174 if (exclusive)
>> 1175 SetPageAnonExclusive(page);
1255 } 1176 }
1256 1177
1257 /** 1178 /**
1258 * __page_check_anon_rmap - sanity check anon 1179 * __page_check_anon_rmap - sanity check anonymous rmap addition
1259 * @folio: The folio containing @page. !! 1180 * @page: the page to add the mapping to
1260 * @page: the page to check the mapping <<
1261 * @vma: the vm area in which the mapp 1181 * @vma: the vm area in which the mapping is added
1262 * @address: the user virtual address mapp 1182 * @address: the user virtual address mapped
1263 */ 1183 */
1264 static void __page_check_anon_rmap(struct fol !! 1184 static void __page_check_anon_rmap(struct page *page,
1265 struct vm_area_struct *vma, unsigned 1185 struct vm_area_struct *vma, unsigned long address)
1266 { 1186 {
>> 1187 struct folio *folio = page_folio(page);
1267 /* 1188 /*
1268 * The page's anon-rmap details (mapp 1189 * The page's anon-rmap details (mapping and index) are guaranteed to
1269 * be set up correctly at this point. 1190 * be set up correctly at this point.
1270 * 1191 *
1271 * We have exclusion against folio_ad !! 1192 * We have exclusion against page_add_anon_rmap because the caller
1272 * always holds the page locked. 1193 * always holds the page locked.
1273 * 1194 *
1274 * We have exclusion against folio_ad !! 1195 * We have exclusion against page_add_new_anon_rmap because those pages
1275 * are initially only visible via the 1196 * are initially only visible via the pagetables, and the pte is locked
1276 * over the call to folio_add_new_ano !! 1197 * over the call to page_add_new_anon_rmap.
1277 */ 1198 */
1278 VM_BUG_ON_FOLIO(folio_anon_vma(folio) 1199 VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1279 folio); 1200 folio);
1280 VM_BUG_ON_PAGE(page_to_pgoff(page) != 1201 VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1281 page); 1202 page);
1282 } 1203 }
1283 1204
1284 static void __folio_mod_stat(struct folio *fo !! 1205 /**
>> 1206 * page_add_anon_rmap - add pte mapping to an anonymous page
>> 1207 * @page: the page to add the mapping to
>> 1208 * @vma: the vm area in which the mapping is added
>> 1209 * @address: the user virtual address mapped
>> 1210 * @flags: the rmap flags
>> 1211 *
>> 1212 * The caller needs to hold the pte lock, and the page must be locked in
>> 1213 * the anon_vma case: to serialize mapping,index checking after setting,
>> 1214 * and to ensure that PageAnon is not being upgraded racily to PageKsm
>> 1215 * (but PageKsm is never downgraded to PageAnon).
>> 1216 */
>> 1217 void page_add_anon_rmap(struct page *page,
>> 1218 struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1285 { 1219 {
1286 int idx; !! 1220 atomic_t *mapped;
1287 !! 1221 int nr = 0, nr_pmdmapped = 0;
1288 if (nr) { !! 1222 bool compound = flags & RMAP_COMPOUND;
1289 idx = folio_test_anon(folio) !! 1223 bool first = true;
1290 __lruvec_stat_mod_folio(folio !! 1224
1291 } !! 1225 if (unlikely(PageKsm(page)))
1292 if (nr_pmdmapped) { !! 1226 lock_page_memcg(page);
1293 if (folio_test_anon(folio)) { !! 1227
1294 idx = NR_ANON_THPS; !! 1228 /* Is page being mapped by PTE? Is this its first map to be added? */
1295 __lruvec_stat_mod_fol !! 1229 if (likely(!compound)) {
1296 } else { !! 1230 first = atomic_inc_and_test(&page->_mapcount);
1297 /* NR_*_PMDMAPPED are !! 1231 nr = first;
1298 idx = folio_test_swap !! 1232 if (first && PageCompound(page)) {
1299 NR_SHMEM_PMDM !! 1233 mapped = subpages_mapcount_ptr(compound_head(page));
1300 __mod_node_page_state !! 1234 nr = atomic_inc_return_relaxed(mapped);
1301 !! 1235 nr = (nr < COMPOUND_MAPPED);
1302 } 1236 }
1303 } !! 1237 } else if (PageTransHuge(page)) {
1304 } !! 1238 /* That test is redundant: it's for safety or to optimize out */
1305 <<
1306 static __always_inline void __folio_add_anon_ <<
1307 struct page *page, int nr_pag <<
1308 unsigned long address, rmap_t <<
1309 { <<
1310 int i, nr, nr_pmdmapped = 0; <<
1311 <<
1312 VM_WARN_ON_FOLIO(!folio_test_anon(fol <<
1313 <<
1314 nr = __folio_add_rmap(folio, page, nr <<
1315 <<
1316 if (likely(!folio_test_ksm(folio))) <<
1317 __page_check_anon_rmap(folio, <<
1318 1239
1319 __folio_mod_stat(folio, nr, nr_pmdmap !! 1240 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1320 !! 1241 if (first) {
1321 if (flags & RMAP_EXCLUSIVE) { !! 1242 mapped = subpages_mapcount_ptr(page);
1322 switch (level) { !! 1243 nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
1323 case RMAP_LEVEL_PTE: !! 1244 if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
1324 for (i = 0; i < nr_pa !! 1245 nr_pmdmapped = thp_nr_pages(page);
1325 SetPageAnonEx !! 1246 nr = nr_pmdmapped - (nr & SUBPAGES_MAPPED);
1326 break; !! 1247 /* Raced ahead of a remove and another add? */
1327 case RMAP_LEVEL_PMD: !! 1248 if (unlikely(nr < 0))
1328 SetPageAnonExclusive( !! 1249 nr = 0;
1329 break; !! 1250 } else {
>> 1251 /* Raced ahead of a remove of COMPOUND_MAPPED */
>> 1252 nr = 0;
>> 1253 }
1330 } 1254 }
1331 } 1255 }
1332 for (i = 0; i < nr_pages; i++) { <<
1333 struct page *cur_page = page <<
1334 1256
1335 /* While PTE-mapping a THP we !! 1257 VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
1336 VM_WARN_ON_FOLIO((atomic_read !! 1258 VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
1337 (folio_test <<
1338 folio_enti <<
1339 PageAnonExcl <<
1340 } <<
1341 1259
1342 /* !! 1260 if (nr_pmdmapped)
1343 * For large folio, only mlock it if !! 1261 __mod_lruvec_page_state(page, NR_ANON_THPS, nr_pmdmapped);
1344 * not easy to check whether the larg !! 1262 if (nr)
1345 * here. Only mlock normal 4K folio a !! 1263 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1346 * large folio. !! 1264
1347 */ !! 1265 if (unlikely(PageKsm(page)))
1348 if (!folio_test_large(folio)) !! 1266 unlock_page_memcg(page);
1349 mlock_vma_folio(folio, vma); !! 1267
1350 } !! 1268 /* address might be in next vma when migration races vma_adjust */
1351 !! 1269 else if (first)
1352 /** !! 1270 __page_set_anon_rmap(page, vma, address,
1353 * folio_add_anon_rmap_ptes - add PTE mapping !! 1271 !!(flags & RMAP_EXCLUSIVE));
1354 * @folio: The folio to add the mappings !! 1272 else
1355 * @page: The first page to add !! 1273 __page_check_anon_rmap(page, vma, address);
1356 * @nr_pages: The number of pages which wil <<
1357 * @vma: The vm area in which the mapp <<
1358 * @address: The user virtual address of t <<
1359 * @flags: The rmap flags <<
1360 * <<
1361 * The page range of folio is defined by [fir <<
1362 * <<
1363 * The caller needs to hold the page table lo <<
1364 * the anon_vma case: to serialize mapping,in <<
1365 * and to ensure that an anon folio is not be <<
1366 * (but KSM folios are never downgraded). <<
1367 */ <<
1368 void folio_add_anon_rmap_ptes(struct folio *f <<
1369 int nr_pages, struct vm_area_ <<
1370 rmap_t flags) <<
1371 { <<
1372 __folio_add_anon_rmap(folio, page, nr <<
1373 RMAP_LEVEL_PTE) <<
1374 } <<
1375 1274
1376 /** !! 1275 mlock_vma_page(page, vma, compound);
1377 * folio_add_anon_rmap_pmd - add a PMD mappin <<
1378 * @folio: The folio to add the mapping <<
1379 * @page: The first page to add <<
1380 * @vma: The vm area in which the mapp <<
1381 * @address: The user virtual address of t <<
1382 * @flags: The rmap flags <<
1383 * <<
1384 * The page range of folio is defined by [fir <<
1385 * <<
1386 * The caller needs to hold the page table lo <<
1387 * the anon_vma case: to serialize mapping,in <<
1388 */ <<
1389 void folio_add_anon_rmap_pmd(struct folio *fo <<
1390 struct vm_area_struct *vma, u <<
1391 { <<
1392 #ifdef CONFIG_TRANSPARENT_HUGEPAGE <<
1393 __folio_add_anon_rmap(folio, page, HP <<
1394 RMAP_LEVEL_PMD) <<
1395 #else <<
1396 WARN_ON_ONCE(true); <<
1397 #endif <<
1398 } 1276 }
1399 1277
1400 /** 1278 /**
1401 * folio_add_new_anon_rmap - Add mapping to a !! 1279 * page_add_new_anon_rmap - add mapping to a new anonymous page
1402 * @folio: The folio to add the mapping !! 1280 * @page: the page to add the mapping to
1403 * @vma: the vm area in which the mapp 1281 * @vma: the vm area in which the mapping is added
1404 * @address: the user virtual address mapp 1282 * @address: the user virtual address mapped
1405 * @flags: The rmap flags <<
1406 * 1283 *
1407 * Like folio_add_anon_rmap_*() but must only !! 1284 * If it's a compound page, it is accounted as a compound page. As the page
1408 * This means the inc-and-test can be bypasse !! 1285 * is new, it's assume to get mapped exclusively by a single process.
1409 * The folio doesn't necessarily need to be l <<
1410 * unless two threads map it concurrently. Ho <<
1411 * locked if it's shared. <<
1412 * 1286 *
1413 * If the folio is pmd-mappable, it is accoun !! 1287 * Same as page_add_anon_rmap but must only be called on *new* pages.
>> 1288 * This means the inc-and-test can be bypassed.
>> 1289 * Page does not have to be locked.
1414 */ 1290 */
1415 void folio_add_new_anon_rmap(struct folio *fo !! 1291 void page_add_new_anon_rmap(struct page *page,
1416 unsigned long address, rmap_t !! 1292 struct vm_area_struct *vma, unsigned long address)
1417 { 1293 {
1418 const int nr = folio_nr_pages(folio); !! 1294 int nr;
1419 const bool exclusive = flags & RMAP_E <<
1420 int nr_pmdmapped = 0; <<
1421 <<
1422 VM_WARN_ON_FOLIO(folio_test_hugetlb(f <<
1423 VM_WARN_ON_FOLIO(!exclusive && !folio <<
1424 VM_BUG_ON_VMA(address < vma->vm_start <<
1425 address + (nr << PAGE <<
1426 <<
1427 /* <<
1428 * VM_DROPPABLE mappings don't swap; <<
1429 * under memory pressure. <<
1430 */ <<
1431 if (!folio_test_swapbacked(folio) && <<
1432 __folio_set_swapbacked(folio) <<
1433 __folio_set_anon(folio, vma, address, <<
1434 1295
1435 if (likely(!folio_test_large(folio))) !! 1296 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1436 /* increment count (starts at !! 1297 __SetPageSwapBacked(page);
1437 atomic_set(&folio->_mapcount, <<
1438 if (exclusive) <<
1439 SetPageAnonExclusive( <<
1440 } else if (!folio_test_pmd_mappable(f <<
1441 int i; <<
1442 <<
1443 for (i = 0; i < nr; i++) { <<
1444 struct page *page = f <<
1445 <<
1446 /* increment count (s <<
1447 atomic_set(&page->_ma <<
1448 if (exclusive) <<
1449 SetPageAnonEx <<
1450 } <<
1451 1298
>> 1299 if (likely(!PageCompound(page))) {
1452 /* increment count (starts at 1300 /* increment count (starts at -1) */
1453 atomic_set(&folio->_large_map !! 1301 atomic_set(&page->_mapcount, 0);
1454 atomic_set(&folio->_nr_pages_ !! 1302 nr = 1;
1455 } else { 1303 } else {
>> 1304 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1456 /* increment count (starts at 1305 /* increment count (starts at -1) */
1457 atomic_set(&folio->_entire_ma !! 1306 atomic_set(compound_mapcount_ptr(page), 0);
1458 /* increment count (starts at !! 1307 atomic_set(subpages_mapcount_ptr(page), COMPOUND_MAPPED);
1459 atomic_set(&folio->_large_map !! 1308 nr = thp_nr_pages(page);
1460 atomic_set(&folio->_nr_pages_ !! 1309 __mod_lruvec_page_state(page, NR_ANON_THPS, nr);
1461 if (exclusive) <<
1462 SetPageAnonExclusive( <<
1463 nr_pmdmapped = nr; <<
1464 } 1310 }
1465 1311
1466 __folio_mod_stat(folio, nr, nr_pmdmap !! 1312 __mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1467 mod_mthp_stat(folio_order(folio), MTH !! 1313 __page_set_anon_rmap(page, vma, address, 1);
1468 } 1314 }
1469 1315
1470 static __always_inline void __folio_add_file_ !! 1316 /**
1471 struct page *page, int nr_pag !! 1317 * page_add_file_rmap - add pte mapping to a file page
1472 enum rmap_level level) !! 1318 * @page: the page to add the mapping to
>> 1319 * @vma: the vm area in which the mapping is added
>> 1320 * @compound: charge the page as compound or small page
>> 1321 *
>> 1322 * The caller needs to hold the pte lock.
>> 1323 */
>> 1324 void page_add_file_rmap(struct page *page,
>> 1325 struct vm_area_struct *vma, bool compound)
1473 { 1326 {
1474 int nr, nr_pmdmapped = 0; !! 1327 atomic_t *mapped;
>> 1328 int nr = 0, nr_pmdmapped = 0;
>> 1329 bool first;
1475 1330
1476 VM_WARN_ON_FOLIO(folio_test_anon(foli !! 1331 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
>> 1332 lock_page_memcg(page);
1477 1333
1478 nr = __folio_add_rmap(folio, page, nr !! 1334 /* Is page being mapped by PTE? Is this its first map to be added? */
1479 __folio_mod_stat(folio, nr, nr_pmdmap !! 1335 if (likely(!compound)) {
>> 1336 first = atomic_inc_and_test(&page->_mapcount);
>> 1337 nr = first;
>> 1338 if (first && PageCompound(page)) {
>> 1339 mapped = subpages_mapcount_ptr(compound_head(page));
>> 1340 nr = atomic_inc_return_relaxed(mapped);
>> 1341 nr = (nr < COMPOUND_MAPPED);
>> 1342 }
>> 1343 } else if (PageTransHuge(page)) {
>> 1344 /* That test is redundant: it's for safety or to optimize out */
1480 1345
1481 /* See comments in folio_add_anon_rma !! 1346 first = atomic_inc_and_test(compound_mapcount_ptr(page));
1482 if (!folio_test_large(folio)) !! 1347 if (first) {
1483 mlock_vma_folio(folio, vma); !! 1348 mapped = subpages_mapcount_ptr(page);
1484 } !! 1349 nr = atomic_add_return_relaxed(COMPOUND_MAPPED, mapped);
>> 1350 if (likely(nr < COMPOUND_MAPPED + COMPOUND_MAPPED)) {
>> 1351 nr_pmdmapped = thp_nr_pages(page);
>> 1352 nr = nr_pmdmapped - (nr & SUBPAGES_MAPPED);
>> 1353 /* Raced ahead of a remove and another add? */
>> 1354 if (unlikely(nr < 0))
>> 1355 nr = 0;
>> 1356 } else {
>> 1357 /* Raced ahead of a remove of COMPOUND_MAPPED */
>> 1358 nr = 0;
>> 1359 }
>> 1360 }
>> 1361 }
1485 1362
1486 /** !! 1363 if (nr_pmdmapped)
1487 * folio_add_file_rmap_ptes - add PTE mapping !! 1364 __mod_lruvec_page_state(page, PageSwapBacked(page) ?
1488 * @folio: The folio to add the mappings !! 1365 NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED, nr_pmdmapped);
1489 * @page: The first page to add !! 1366 if (nr)
1490 * @nr_pages: The number of pages that will !! 1367 __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1491 * @vma: The vm area in which the mapp !! 1368 unlock_page_memcg(page);
1492 * !! 1369
1493 * The page range of the folio is defined by !! 1370 mlock_vma_page(page, vma, compound);
1494 * <<
1495 * The caller needs to hold the page table lo <<
1496 */ <<
1497 void folio_add_file_rmap_ptes(struct folio *f <<
1498 int nr_pages, struct vm_area_ <<
1499 { <<
1500 __folio_add_file_rmap(folio, page, nr <<
1501 } 1371 }
1502 1372
1503 /** 1373 /**
1504 * folio_add_file_rmap_pmd - add a PMD mappin !! 1374 * page_remove_rmap - take down pte mapping from a page
1505 * @folio: The folio to add the mapping !! 1375 * @page: page to remove mapping from
1506 * @page: The first page to add !! 1376 * @vma: the vm area from which the mapping is removed
1507 * @vma: The vm area in which the mapp !! 1377 * @compound: uncharge the page as compound or small page
1508 * !! 1378 *
1509 * The page range of the folio is defined by !! 1379 * The caller needs to hold the pte lock.
1510 * !! 1380 */
1511 * The caller needs to hold the page table lo !! 1381 void page_remove_rmap(struct page *page,
1512 */ !! 1382 struct vm_area_struct *vma, bool compound)
1513 void folio_add_file_rmap_pmd(struct folio *fo !! 1383 {
1514 struct vm_area_struct *vma) !! 1384 atomic_t *mapped;
1515 { !! 1385 int nr = 0, nr_pmdmapped = 0;
1516 #ifdef CONFIG_TRANSPARENT_HUGEPAGE !! 1386 bool last;
1517 __folio_add_file_rmap(folio, page, HP !! 1387
1518 #else !! 1388 VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1519 WARN_ON_ONCE(true); !! 1389
1520 #endif !! 1390 /* Hugetlb pages are not counted in NR_*MAPPED */
1521 } !! 1391 if (unlikely(PageHuge(page))) {
>> 1392 /* hugetlb pages are always mapped with pmds */
>> 1393 atomic_dec(compound_mapcount_ptr(page));
>> 1394 return;
>> 1395 }
1522 1396
1523 static __always_inline void __folio_remove_rm !! 1397 lock_page_memcg(page);
1524 struct page *page, int nr_pag !! 1398
1525 enum rmap_level level) !! 1399 /* Is page being unmapped by PTE? Is this its last map to be removed? */
1526 { !! 1400 if (likely(!compound)) {
1527 atomic_t *mapped = &folio->_nr_pages_ !! 1401 last = atomic_add_negative(-1, &page->_mapcount);
1528 int last = 0, nr = 0, nr_pmdmapped = !! 1402 nr = last;
1529 bool partially_mapped = false; !! 1403 if (last && PageCompound(page)) {
1530 !! 1404 mapped = subpages_mapcount_ptr(compound_head(page));
1531 __folio_rmap_sanity_checks(folio, pag !! 1405 nr = atomic_dec_return_relaxed(mapped);
1532 !! 1406 nr = (nr < COMPOUND_MAPPED);
1533 switch (level) { <<
1534 case RMAP_LEVEL_PTE: <<
1535 if (!folio_test_large(folio)) <<
1536 nr = atomic_add_negat <<
1537 break; <<
1538 } 1407 }
>> 1408 } else if (PageTransHuge(page)) {
>> 1409 /* That test is redundant: it's for safety or to optimize out */
1539 1410
1540 atomic_sub(nr_pages, &folio-> !! 1411 last = atomic_add_negative(-1, compound_mapcount_ptr(page));
1541 do { <<
1542 last += atomic_add_ne <<
1543 } while (page++, --nr_pages > <<
1544 <<
1545 if (last && <<
1546 atomic_sub_return_relaxed <<
1547 nr = last; <<
1548 <<
1549 partially_mapped = nr && atom <<
1550 break; <<
1551 case RMAP_LEVEL_PMD: <<
1552 atomic_dec(&folio->_large_map <<
1553 last = atomic_add_negative(-1 <<
1554 if (last) { 1412 if (last) {
1555 nr = atomic_sub_retur !! 1413 mapped = subpages_mapcount_ptr(page);
1556 if (likely(nr < ENTIR !! 1414 nr = atomic_sub_return_relaxed(COMPOUND_MAPPED, mapped);
1557 nr_pmdmapped !! 1415 if (likely(nr < COMPOUND_MAPPED)) {
1558 nr = nr_pmdma !! 1416 nr_pmdmapped = thp_nr_pages(page);
>> 1417 nr = nr_pmdmapped - (nr & SUBPAGES_MAPPED);
1559 /* Raced ahea 1418 /* Raced ahead of another remove and an add? */
1560 if (unlikely( 1419 if (unlikely(nr < 0))
1561 nr = 1420 nr = 0;
1562 } else { 1421 } else {
1563 /* An add of !! 1422 /* An add of COMPOUND_MAPPED raced ahead */
1564 nr = 0; 1423 nr = 0;
1565 } 1424 }
1566 } 1425 }
1567 <<
1568 partially_mapped = nr && nr < <<
1569 break; <<
1570 } 1426 }
1571 1427
1572 /* !! 1428 if (nr_pmdmapped) {
1573 * Queue anon large folio for deferre !! 1429 __mod_lruvec_page_state(page, PageAnon(page) ? NR_ANON_THPS :
1574 * the folio is unmapped and at least !! 1430 (PageSwapBacked(page) ? NR_SHMEM_PMDMAPPED :
1575 * !! 1431 NR_FILE_PMDMAPPED), -nr_pmdmapped);
1576 * Check partially_mapped first to en !! 1432 }
1577 */ !! 1433 if (nr) {
1578 if (partially_mapped && folio_test_an !! 1434 __mod_lruvec_page_state(page, PageAnon(page) ? NR_ANON_MAPPED :
1579 !folio_test_partially_mapped(foli !! 1435 NR_FILE_MAPPED, -nr);
1580 deferred_split_folio(folio, t !! 1436 /*
1581 !! 1437 * Queue anon THP for deferred split if at least one small
1582 __folio_mod_stat(folio, -nr, -nr_pmdm !! 1438 * page of the compound page is unmapped, but at least one
>> 1439 * small page is still mapped.
>> 1440 */
>> 1441 if (PageTransCompound(page) && PageAnon(page))
>> 1442 if (!compound || nr < nr_pmdmapped)
>> 1443 deferred_split_huge_page(compound_head(page));
>> 1444 }
1583 1445
1584 /* 1446 /*
1585 * It would be tidy to reset folio_te !! 1447 * It would be tidy to reset PageAnon mapping when fully unmapped,
1586 * unmapped, but that might overwrite !! 1448 * but that might overwrite a racing page_add_anon_rmap
1587 * which increments mapcount after us !! 1449 * which increments mapcount after us but sets mapping
1588 * so leave the reset to free_pages_p !! 1450 * before us: so leave the reset to free_pages_prepare,
1589 * it's only reliable while mapped. !! 1451 * and remember that it's only reliable while mapped.
1590 */ 1452 */
1591 1453
1592 munlock_vma_folio(folio, vma); !! 1454 unlock_page_memcg(page);
1593 } <<
1594 <<
1595 /** <<
1596 * folio_remove_rmap_ptes - remove PTE mappin <<
1597 * @folio: The folio to remove the mappi <<
1598 * @page: The first page to remove <<
1599 * @nr_pages: The number of pages that will <<
1600 * @vma: The vm area from which the ma <<
1601 * <<
1602 * The page range of the folio is defined by <<
1603 * <<
1604 * The caller needs to hold the page table lo <<
1605 */ <<
1606 void folio_remove_rmap_ptes(struct folio *fol <<
1607 int nr_pages, struct vm_area_ <<
1608 { <<
1609 __folio_remove_rmap(folio, page, nr_p <<
1610 } <<
1611 1455
1612 /** !! 1456 munlock_vma_page(page, vma, compound);
1613 * folio_remove_rmap_pmd - remove a PMD mappi <<
1614 * @folio: The folio to remove the mappi <<
1615 * @page: The first page to remove <<
1616 * @vma: The vm area from which the ma <<
1617 * <<
1618 * The page range of the folio is defined by <<
1619 * <<
1620 * The caller needs to hold the page table lo <<
1621 */ <<
1622 void folio_remove_rmap_pmd(struct folio *foli <<
1623 struct vm_area_struct *vma) <<
1624 { <<
1625 #ifdef CONFIG_TRANSPARENT_HUGEPAGE <<
1626 __folio_remove_rmap(folio, page, HPAG <<
1627 #else <<
1628 WARN_ON_ONCE(true); <<
1629 #endif <<
1630 } 1457 }
1631 1458
1632 /* 1459 /*
1633 * @arg: enum ttu_flags will be passed to thi 1460 * @arg: enum ttu_flags will be passed to this argument
1634 */ 1461 */
1635 static bool try_to_unmap_one(struct folio *fo 1462 static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1636 unsigned long address, v 1463 unsigned long address, void *arg)
1637 { 1464 {
1638 struct mm_struct *mm = vma->vm_mm; 1465 struct mm_struct *mm = vma->vm_mm;
1639 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vm 1466 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1640 pte_t pteval; 1467 pte_t pteval;
1641 struct page *subpage; 1468 struct page *subpage;
1642 bool anon_exclusive, ret = true; 1469 bool anon_exclusive, ret = true;
1643 struct mmu_notifier_range range; 1470 struct mmu_notifier_range range;
1644 enum ttu_flags flags = (enum ttu_flag 1471 enum ttu_flags flags = (enum ttu_flags)(long)arg;
1645 unsigned long pfn; <<
1646 unsigned long hsz = 0; <<
1647 1472
1648 /* 1473 /*
1649 * When racing against e.g. zap_pte_r 1474 * When racing against e.g. zap_pte_range() on another cpu,
1650 * in between its ptep_get_and_clear_ !! 1475 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
1651 * try_to_unmap() may return before p 1476 * try_to_unmap() may return before page_mapped() has become false,
1652 * if page table locking is skipped: 1477 * if page table locking is skipped: use TTU_SYNC to wait for that.
1653 */ 1478 */
1654 if (flags & TTU_SYNC) 1479 if (flags & TTU_SYNC)
1655 pvmw.flags = PVMW_SYNC; 1480 pvmw.flags = PVMW_SYNC;
1656 1481
>> 1482 if (flags & TTU_SPLIT_HUGE_PMD)
>> 1483 split_huge_pmd_address(vma, address, false, folio);
>> 1484
1657 /* 1485 /*
1658 * For THP, we have to assume the wor 1486 * For THP, we have to assume the worse case ie pmd for invalidation.
1659 * For hugetlb, it could be much wors 1487 * For hugetlb, it could be much worse if we need to do pud
1660 * invalidation in the case of pmd sh 1488 * invalidation in the case of pmd sharing.
1661 * 1489 *
1662 * Note that the folio can not be fre 1490 * Note that the folio can not be freed in this function as call of
1663 * try_to_unmap() must hold a referen 1491 * try_to_unmap() must hold a reference on the folio.
1664 */ 1492 */
1665 range.end = vma_address_end(&pvmw); 1493 range.end = vma_address_end(&pvmw);
1666 mmu_notifier_range_init(&range, MMU_N !! 1494 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1667 address, rang 1495 address, range.end);
1668 if (folio_test_hugetlb(folio)) { 1496 if (folio_test_hugetlb(folio)) {
1669 /* 1497 /*
1670 * If sharing is possible, st 1498 * If sharing is possible, start and end will be adjusted
1671 * accordingly. 1499 * accordingly.
1672 */ 1500 */
1673 adjust_range_if_pmd_sharing_p 1501 adjust_range_if_pmd_sharing_possible(vma, &range.start,
1674 1502 &range.end);
1675 <<
1676 /* We need the huge page size <<
1677 hsz = huge_page_size(hstate_v <<
1678 } 1503 }
1679 mmu_notifier_invalidate_range_start(& 1504 mmu_notifier_invalidate_range_start(&range);
1680 1505
1681 while (page_vma_mapped_walk(&pvmw)) { 1506 while (page_vma_mapped_walk(&pvmw)) {
>> 1507 /* Unexpected PMD-mapped THP? */
>> 1508 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
>> 1509
1682 /* 1510 /*
1683 * If the folio is in an mloc 1511 * If the folio is in an mlock()d vma, we must not swap it out.
1684 */ 1512 */
1685 if (!(flags & TTU_IGNORE_MLOC 1513 if (!(flags & TTU_IGNORE_MLOCK) &&
1686 (vma->vm_flags & VM_LOCKE 1514 (vma->vm_flags & VM_LOCKED)) {
1687 /* Restore the mlock 1515 /* Restore the mlock which got missed */
1688 if (!folio_test_large !! 1516 mlock_vma_folio(folio, vma, false);
1689 mlock_vma_fol !! 1517 page_vma_mapped_walk_done(&pvmw);
1690 goto walk_abort; !! 1518 ret = false;
1691 } !! 1519 break;
1692 <<
1693 if (!pvmw.pte) { <<
1694 if (unmap_huge_pmd_lo <<
1695 <<
1696 goto walk_don <<
1697 <<
1698 if (flags & TTU_SPLIT <<
1699 /* <<
1700 * We tempora <<
1701 * restart so <<
1702 */ <<
1703 split_huge_pm <<
1704 <<
1705 flags &= ~TTU <<
1706 page_vma_mapp <<
1707 continue; <<
1708 } <<
1709 } 1520 }
1710 1521
1711 /* Unexpected PMD-mapped THP? !! 1522 subpage = folio_page(folio,
1712 VM_BUG_ON_FOLIO(!pvmw.pte, fo !! 1523 pte_pfn(*pvmw.pte) - folio_pfn(folio));
1713 <<
1714 pfn = pte_pfn(ptep_get(pvmw.p <<
1715 subpage = folio_page(folio, p <<
1716 address = pvmw.address; 1524 address = pvmw.address;
1717 anon_exclusive = folio_test_a 1525 anon_exclusive = folio_test_anon(folio) &&
1718 PageAnonExcl 1526 PageAnonExclusive(subpage);
1719 1527
1720 if (folio_test_hugetlb(folio) 1528 if (folio_test_hugetlb(folio)) {
1721 bool anon = folio_tes 1529 bool anon = folio_test_anon(folio);
1722 1530
1723 /* 1531 /*
1724 * The try_to_unmap() 1532 * The try_to_unmap() is only passed a hugetlb page
1725 * in the case where 1533 * in the case where the hugetlb page is poisoned.
1726 */ 1534 */
1727 VM_BUG_ON_PAGE(!PageH 1535 VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1728 /* 1536 /*
1729 * huge_pmd_unshare m 1537 * huge_pmd_unshare may unmap an entire PMD page.
1730 * There is no way of 1538 * There is no way of knowing exactly which PMDs may
1731 * be cached for this 1539 * be cached for this mm, so we must flush them all.
1732 * start/end were alr 1540 * start/end were already adjusted above to cover this
1733 * range. 1541 * range.
1734 */ 1542 */
1735 flush_cache_range(vma 1543 flush_cache_range(vma, range.start, range.end);
1736 1544
1737 /* 1545 /*
1738 * To call huge_pmd_u 1546 * To call huge_pmd_unshare, i_mmap_rwsem must be
1739 * held in write mode 1547 * held in write mode. Caller needs to explicitly
1740 * do this outside rm 1548 * do this outside rmap routines.
1741 * 1549 *
1742 * We also must hold 1550 * We also must hold hugetlb vma_lock in write mode.
1743 * Lock order dictate 1551 * Lock order dictates acquiring vma_lock BEFORE
1744 * i_mmap_rwsem. We 1552 * i_mmap_rwsem. We can only try lock here and fail
1745 * if unsuccessful. 1553 * if unsuccessful.
1746 */ 1554 */
1747 if (!anon) { 1555 if (!anon) {
1748 VM_BUG_ON(!(f 1556 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1749 if (!hugetlb_ !! 1557 if (!hugetlb_vma_trylock_write(vma)) {
1750 goto !! 1558 page_vma_mapped_walk_done(&pvmw);
>> 1559 ret = false;
>> 1560 break;
>> 1561 }
1751 if (huge_pmd_ 1562 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1752 huget 1563 hugetlb_vma_unlock_write(vma);
1753 flush 1564 flush_tlb_range(vma,
1754 1565 range.start, range.end);
>> 1566 mmu_notifier_invalidate_range(mm,
>> 1567 range.start, range.end);
1755 /* 1568 /*
1756 * Th 1569 * The ref count of the PMD page was
1757 * dr 1570 * dropped which is part of the way map
1758 * co 1571 * counting is done for shared PMDs.
1759 * Re 1572 * Return 'true' here. When there is
1760 * no 1573 * no other sharing, huge_pmd_unshare
1761 * re 1574 * returns false and we will unmap the
1762 * ac 1575 * actual page and drop map count
1763 * to 1576 * to zero.
1764 */ 1577 */
1765 goto !! 1578 page_vma_mapped_walk_done(&pvmw);
>> 1579 break;
1766 } 1580 }
1767 hugetlb_vma_u 1581 hugetlb_vma_unlock_write(vma);
1768 } 1582 }
1769 pteval = huge_ptep_cl 1583 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1770 } else { 1584 } else {
1771 flush_cache_page(vma, !! 1585 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1772 /* Nuke the page tabl 1586 /* Nuke the page table entry. */
1773 if (should_defer_flus 1587 if (should_defer_flush(mm, flags)) {
1774 /* 1588 /*
1775 * We clear t 1589 * We clear the PTE but do not flush so potentially
1776 * a remote C 1590 * a remote CPU could still be writing to the folio.
1777 * If the ent 1591 * If the entry was previously clean then the
1778 * architectu 1592 * architecture must guarantee that a clear->dirty
1779 * transition 1593 * transition on a cached TLB entry is written through
1780 * and traps 1594 * and traps if the PTE is unmapped.
1781 */ 1595 */
1782 pteval = ptep 1596 pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1783 1597
1784 set_tlb_ubc_f !! 1598 set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1785 } else { 1599 } else {
1786 pteval = ptep 1600 pteval = ptep_clear_flush(vma, address, pvmw.pte);
1787 } 1601 }
1788 } 1602 }
1789 1603
1790 /* 1604 /*
1791 * Now the pte is cleared. If 1605 * Now the pte is cleared. If this pte was uffd-wp armed,
1792 * we may want to replace a n 1606 * we may want to replace a none pte with a marker pte if
1793 * it's file-backed, so we do 1607 * it's file-backed, so we don't lose the tracking info.
1794 */ 1608 */
1795 pte_install_uffd_wp_if_needed 1609 pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1796 1610
1797 /* Set the dirty flag on the 1611 /* Set the dirty flag on the folio now the pte is gone. */
1798 if (pte_dirty(pteval)) 1612 if (pte_dirty(pteval))
1799 folio_mark_dirty(foli 1613 folio_mark_dirty(folio);
1800 1614
1801 /* Update high watermark befo 1615 /* Update high watermark before we lower rss */
1802 update_hiwater_rss(mm); 1616 update_hiwater_rss(mm);
1803 1617
1804 if (PageHWPoison(subpage) && 1618 if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1805 pteval = swp_entry_to 1619 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1806 if (folio_test_hugetl 1620 if (folio_test_hugetlb(folio)) {
1807 hugetlb_count 1621 hugetlb_count_sub(folio_nr_pages(folio), mm);
1808 set_huge_pte_ !! 1622 set_huge_pte_at(mm, address, pvmw.pte, pteval);
1809 <<
1810 } else { 1623 } else {
1811 dec_mm_counte !! 1624 dec_mm_counter(mm, mm_counter(&folio->page));
1812 set_pte_at(mm 1625 set_pte_at(mm, address, pvmw.pte, pteval);
1813 } 1626 }
1814 1627
1815 } else if (pte_unused(pteval) 1628 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1816 /* 1629 /*
1817 * The guest indicate 1630 * The guest indicated that the page content is of no
1818 * interest anymore. 1631 * interest anymore. Simply discard the pte, vmscan
1819 * will take care of 1632 * will take care of the rest.
1820 * A future reference 1633 * A future reference will then fault in a new zero
1821 * page. When userfau 1634 * page. When userfaultfd is active, we must not drop
1822 * this page though, 1635 * this page though, as its main user (postcopy
1823 * migration) will no 1636 * migration) will not expect userfaults on already
1824 * copied pages. 1637 * copied pages.
1825 */ 1638 */
1826 dec_mm_counter(mm, mm !! 1639 dec_mm_counter(mm, mm_counter(&folio->page));
>> 1640 /* We have to invalidate as we cleared the pte */
>> 1641 mmu_notifier_invalidate_range(mm, address,
>> 1642 address + PAGE_SIZE);
1827 } else if (folio_test_anon(fo 1643 } else if (folio_test_anon(folio)) {
1828 swp_entry_t entry = p !! 1644 swp_entry_t entry = { .val = page_private(subpage) };
1829 pte_t swp_pte; 1645 pte_t swp_pte;
1830 /* 1646 /*
1831 * Store the swap loc 1647 * Store the swap location in the pte.
1832 * See handle_pte_fau 1648 * See handle_pte_fault() ...
1833 */ 1649 */
1834 if (unlikely(folio_te 1650 if (unlikely(folio_test_swapbacked(folio) !=
1835 folio 1651 folio_test_swapcache(folio))) {
1836 WARN_ON_ONCE( 1652 WARN_ON_ONCE(1);
1837 goto walk_abo !! 1653 ret = false;
>> 1654 /* We have to invalidate as we cleared the pte */
>> 1655 mmu_notifier_invalidate_range(mm, address,
>> 1656 address + PAGE_SIZE);
>> 1657 page_vma_mapped_walk_done(&pvmw);
>> 1658 break;
1838 } 1659 }
1839 1660
1840 /* MADV_FREE page che 1661 /* MADV_FREE page check */
1841 if (!folio_test_swapb 1662 if (!folio_test_swapbacked(folio)) {
1842 int ref_count 1663 int ref_count, map_count;
1843 1664
1844 /* 1665 /*
1845 * Synchroniz 1666 * Synchronize with gup_pte_range():
1846 * - clear PT 1667 * - clear PTE; barrier; read refcount
1847 * - inc refc 1668 * - inc refcount; barrier; read PTE
1848 */ 1669 */
1849 smp_mb(); 1670 smp_mb();
1850 1671
1851 ref_count = f 1672 ref_count = folio_ref_count(folio);
1852 map_count = f 1673 map_count = folio_mapcount(folio);
1853 1674
1854 /* 1675 /*
1855 * Order read 1676 * Order reads for page refcount and dirty flag
1856 * (see comme 1677 * (see comments in __remove_mapping()).
1857 */ 1678 */
1858 smp_rmb(); 1679 smp_rmb();
1859 1680
1860 /* 1681 /*
1861 * The only p 1682 * The only page refs must be one from isolation
1862 * plus the r 1683 * plus the rmap(s) (dropped by discard:).
1863 */ 1684 */
1864 if (ref_count 1685 if (ref_count == 1 + map_count &&
1865 (!folio_t !! 1686 !folio_test_dirty(folio)) {
1866 /* !! 1687 /* Invalidate as we cleared the pte */
1867 * Unlik !! 1688 mmu_notifier_invalidate_range(mm,
1868 * ones !! 1689 address, address + PAGE_SIZE);
1869 * been <<
1870 */ <<
1871 (vma->vm <<
1872 dec_m 1690 dec_mm_counter(mm, MM_ANONPAGES);
1873 goto 1691 goto discard;
1874 } 1692 }
1875 1693
1876 /* 1694 /*
1877 * If the fol 1695 * If the folio was redirtied, it cannot be
1878 * discarded. 1696 * discarded. Remap the page to page table.
1879 */ 1697 */
1880 set_pte_at(mm 1698 set_pte_at(mm, address, pvmw.pte, pteval);
1881 /* !! 1699 folio_set_swapbacked(folio);
1882 * Unlike MAD !! 1700 ret = false;
1883 * never get !! 1701 page_vma_mapped_walk_done(&pvmw);
1884 */ !! 1702 break;
1885 if (!(vma->vm <<
1886 folio <<
1887 goto walk_abo <<
1888 } 1703 }
1889 1704
1890 if (swap_duplicate(en 1705 if (swap_duplicate(entry) < 0) {
1891 set_pte_at(mm 1706 set_pte_at(mm, address, pvmw.pte, pteval);
1892 goto walk_abo !! 1707 ret = false;
>> 1708 page_vma_mapped_walk_done(&pvmw);
>> 1709 break;
1893 } 1710 }
1894 if (arch_unmap_one(mm 1711 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1895 swap_free(ent 1712 swap_free(entry);
1896 set_pte_at(mm 1713 set_pte_at(mm, address, pvmw.pte, pteval);
1897 goto walk_abo !! 1714 ret = false;
>> 1715 page_vma_mapped_walk_done(&pvmw);
>> 1716 break;
1898 } 1717 }
1899 1718
1900 /* See folio_try_shar !! 1719 /* See page_try_share_anon_rmap(): clear PTE first. */
1901 if (anon_exclusive && 1720 if (anon_exclusive &&
1902 folio_try_share_a !! 1721 page_try_share_anon_rmap(subpage)) {
1903 swap_free(ent 1722 swap_free(entry);
1904 set_pte_at(mm 1723 set_pte_at(mm, address, pvmw.pte, pteval);
1905 goto walk_abo !! 1724 ret = false;
>> 1725 page_vma_mapped_walk_done(&pvmw);
>> 1726 break;
1906 } 1727 }
>> 1728 /*
>> 1729 * Note: We *don't* remember if the page was mapped
>> 1730 * exclusively in the swap pte if the architecture
>> 1731 * doesn't support __HAVE_ARCH_PTE_SWP_EXCLUSIVE. In
>> 1732 * that case, swapin code has to re-determine that
>> 1733 * manually and might detect the page as possibly
>> 1734 * shared, for example, if there are other references on
>> 1735 * the page or if the page is under writeback. We made
>> 1736 * sure that there are no GUP pins on the page that
>> 1737 * would rely on it, so for GUP pins this is fine.
>> 1738 */
1907 if (list_empty(&mm->m 1739 if (list_empty(&mm->mmlist)) {
1908 spin_lock(&mm 1740 spin_lock(&mmlist_lock);
1909 if (list_empt 1741 if (list_empty(&mm->mmlist))
1910 list_ 1742 list_add(&mm->mmlist, &init_mm.mmlist);
1911 spin_unlock(& 1743 spin_unlock(&mmlist_lock);
1912 } 1744 }
1913 dec_mm_counter(mm, MM 1745 dec_mm_counter(mm, MM_ANONPAGES);
1914 inc_mm_counter(mm, MM 1746 inc_mm_counter(mm, MM_SWAPENTS);
1915 swp_pte = swp_entry_t 1747 swp_pte = swp_entry_to_pte(entry);
1916 if (anon_exclusive) 1748 if (anon_exclusive)
1917 swp_pte = pte 1749 swp_pte = pte_swp_mkexclusive(swp_pte);
1918 if (pte_soft_dirty(pt 1750 if (pte_soft_dirty(pteval))
1919 swp_pte = pte 1751 swp_pte = pte_swp_mksoft_dirty(swp_pte);
1920 if (pte_uffd_wp(pteva 1752 if (pte_uffd_wp(pteval))
1921 swp_pte = pte 1753 swp_pte = pte_swp_mkuffd_wp(swp_pte);
1922 set_pte_at(mm, addres 1754 set_pte_at(mm, address, pvmw.pte, swp_pte);
>> 1755 /* Invalidate as we cleared the pte */
>> 1756 mmu_notifier_invalidate_range(mm, address,
>> 1757 address + PAGE_SIZE);
1923 } else { 1758 } else {
1924 /* 1759 /*
1925 * This is a locked f 1760 * This is a locked file-backed folio,
1926 * so it cannot be re 1761 * so it cannot be removed from the page
1927 * cache and replaced 1762 * cache and replaced by a new folio before
1928 * mmu_notifier_inval 1763 * mmu_notifier_invalidate_range_end, so no
1929 * concurrent thread 1764 * concurrent thread might update its page table
1930 * to point at a new 1765 * to point at a new folio while a device is
1931 * still using this f 1766 * still using this folio.
1932 * 1767 *
1933 * See Documentation/ 1768 * See Documentation/mm/mmu_notifier.rst
1934 */ 1769 */
1935 dec_mm_counter(mm, mm !! 1770 dec_mm_counter(mm, mm_counter_file(&folio->page));
1936 } 1771 }
1937 discard: 1772 discard:
1938 if (unlikely(folio_test_huget !! 1773 /*
1939 hugetlb_remove_rmap(f !! 1774 * No need to call mmu_notifier_invalidate_range() it has be
1940 else !! 1775 * done above for all cases requiring it to happen under page
1941 folio_remove_rmap_pte !! 1776 * table lock before mmu_notifier_invalidate_range_end()
>> 1777 *
>> 1778 * See Documentation/mm/mmu_notifier.rst
>> 1779 */
>> 1780 page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
1942 if (vma->vm_flags & VM_LOCKED 1781 if (vma->vm_flags & VM_LOCKED)
1943 mlock_drain_local(); !! 1782 mlock_page_drain_local();
1944 folio_put(folio); 1783 folio_put(folio);
1945 continue; <<
1946 walk_abort: <<
1947 ret = false; <<
1948 walk_done: <<
1949 page_vma_mapped_walk_done(&pv <<
1950 break; <<
1951 } 1784 }
1952 1785
1953 mmu_notifier_invalidate_range_end(&ra 1786 mmu_notifier_invalidate_range_end(&range);
1954 1787
1955 return ret; 1788 return ret;
1956 } 1789 }
1957 1790
1958 static bool invalid_migration_vma(struct vm_a 1791 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1959 { 1792 {
1960 return vma_is_temporary_stack(vma); 1793 return vma_is_temporary_stack(vma);
1961 } 1794 }
1962 1795
1963 static int folio_not_mapped(struct folio *fol 1796 static int folio_not_mapped(struct folio *folio)
1964 { 1797 {
1965 return !folio_mapped(folio); 1798 return !folio_mapped(folio);
1966 } 1799 }
1967 1800
1968 /** 1801 /**
1969 * try_to_unmap - Try to remove all page tabl 1802 * try_to_unmap - Try to remove all page table mappings to a folio.
1970 * @folio: The folio to unmap. 1803 * @folio: The folio to unmap.
1971 * @flags: action and flags 1804 * @flags: action and flags
1972 * 1805 *
1973 * Tries to remove all the page table entries 1806 * Tries to remove all the page table entries which are mapping this
1974 * folio. It is the caller's responsibility 1807 * folio. It is the caller's responsibility to check if the folio is
1975 * still mapped if needed (use TTU_SYNC to pr 1808 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1976 * 1809 *
1977 * Context: Caller must hold the folio lock. 1810 * Context: Caller must hold the folio lock.
1978 */ 1811 */
1979 void try_to_unmap(struct folio *folio, enum t 1812 void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1980 { 1813 {
1981 struct rmap_walk_control rwc = { 1814 struct rmap_walk_control rwc = {
1982 .rmap_one = try_to_unmap_one, 1815 .rmap_one = try_to_unmap_one,
1983 .arg = (void *)flags, 1816 .arg = (void *)flags,
1984 .done = folio_not_mapped, 1817 .done = folio_not_mapped,
1985 .anon_lock = folio_lock_anon_ 1818 .anon_lock = folio_lock_anon_vma_read,
1986 }; 1819 };
1987 1820
1988 if (flags & TTU_RMAP_LOCKED) 1821 if (flags & TTU_RMAP_LOCKED)
1989 rmap_walk_locked(folio, &rwc) 1822 rmap_walk_locked(folio, &rwc);
1990 else 1823 else
1991 rmap_walk(folio, &rwc); 1824 rmap_walk(folio, &rwc);
1992 } 1825 }
1993 1826
1994 /* 1827 /*
1995 * @arg: enum ttu_flags will be passed to thi 1828 * @arg: enum ttu_flags will be passed to this argument.
1996 * 1829 *
1997 * If TTU_SPLIT_HUGE_PMD is specified any PMD 1830 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
1998 * containing migration entries. 1831 * containing migration entries.
1999 */ 1832 */
2000 static bool try_to_migrate_one(struct folio * 1833 static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
2001 unsigned long address, v 1834 unsigned long address, void *arg)
2002 { 1835 {
2003 struct mm_struct *mm = vma->vm_mm; 1836 struct mm_struct *mm = vma->vm_mm;
2004 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vm 1837 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2005 pte_t pteval; 1838 pte_t pteval;
2006 struct page *subpage; 1839 struct page *subpage;
2007 bool anon_exclusive, ret = true; 1840 bool anon_exclusive, ret = true;
2008 struct mmu_notifier_range range; 1841 struct mmu_notifier_range range;
2009 enum ttu_flags flags = (enum ttu_flag 1842 enum ttu_flags flags = (enum ttu_flags)(long)arg;
2010 unsigned long pfn; <<
2011 unsigned long hsz = 0; <<
2012 1843
2013 /* 1844 /*
2014 * When racing against e.g. zap_pte_r 1845 * When racing against e.g. zap_pte_range() on another cpu,
2015 * in between its ptep_get_and_clear_ !! 1846 * in between its ptep_get_and_clear_full() and page_remove_rmap(),
2016 * try_to_migrate() may return before 1847 * try_to_migrate() may return before page_mapped() has become false,
2017 * if page table locking is skipped: 1848 * if page table locking is skipped: use TTU_SYNC to wait for that.
2018 */ 1849 */
2019 if (flags & TTU_SYNC) 1850 if (flags & TTU_SYNC)
2020 pvmw.flags = PVMW_SYNC; 1851 pvmw.flags = PVMW_SYNC;
2021 1852
2022 /* 1853 /*
2023 * unmap_page() in mm/huge_memory.c i 1854 * unmap_page() in mm/huge_memory.c is the only user of migration with
2024 * TTU_SPLIT_HUGE_PMD and it wants to 1855 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
2025 */ 1856 */
2026 if (flags & TTU_SPLIT_HUGE_PMD) 1857 if (flags & TTU_SPLIT_HUGE_PMD)
2027 split_huge_pmd_address(vma, a 1858 split_huge_pmd_address(vma, address, true, folio);
2028 1859
2029 /* 1860 /*
2030 * For THP, we have to assume the wor 1861 * For THP, we have to assume the worse case ie pmd for invalidation.
2031 * For hugetlb, it could be much wors 1862 * For hugetlb, it could be much worse if we need to do pud
2032 * invalidation in the case of pmd sh 1863 * invalidation in the case of pmd sharing.
2033 * 1864 *
2034 * Note that the page can not be free 1865 * Note that the page can not be free in this function as call of
2035 * try_to_unmap() must hold a referen 1866 * try_to_unmap() must hold a reference on the page.
2036 */ 1867 */
2037 range.end = vma_address_end(&pvmw); 1868 range.end = vma_address_end(&pvmw);
2038 mmu_notifier_range_init(&range, MMU_N !! 1869 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2039 address, rang 1870 address, range.end);
2040 if (folio_test_hugetlb(folio)) { 1871 if (folio_test_hugetlb(folio)) {
2041 /* 1872 /*
2042 * If sharing is possible, st 1873 * If sharing is possible, start and end will be adjusted
2043 * accordingly. 1874 * accordingly.
2044 */ 1875 */
2045 adjust_range_if_pmd_sharing_p 1876 adjust_range_if_pmd_sharing_possible(vma, &range.start,
2046 1877 &range.end);
2047 <<
2048 /* We need the huge page size <<
2049 hsz = huge_page_size(hstate_v <<
2050 } 1878 }
2051 mmu_notifier_invalidate_range_start(& 1879 mmu_notifier_invalidate_range_start(&range);
2052 1880
2053 while (page_vma_mapped_walk(&pvmw)) { 1881 while (page_vma_mapped_walk(&pvmw)) {
2054 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION 1882 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2055 /* PMD-mapped THP migration e 1883 /* PMD-mapped THP migration entry */
2056 if (!pvmw.pte) { 1884 if (!pvmw.pte) {
2057 subpage = folio_page( 1885 subpage = folio_page(folio,
2058 pmd_pfn(*pvmw 1886 pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
2059 VM_BUG_ON_FOLIO(folio 1887 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
2060 !foli 1888 !folio_test_pmd_mappable(folio), folio);
2061 1889
2062 if (set_pmd_migration 1890 if (set_pmd_migration_entry(&pvmw, subpage)) {
2063 ret = false; 1891 ret = false;
2064 page_vma_mapp 1892 page_vma_mapped_walk_done(&pvmw);
2065 break; 1893 break;
2066 } 1894 }
2067 continue; 1895 continue;
2068 } 1896 }
2069 #endif 1897 #endif
2070 1898
2071 /* Unexpected PMD-mapped THP? 1899 /* Unexpected PMD-mapped THP? */
2072 VM_BUG_ON_FOLIO(!pvmw.pte, fo 1900 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2073 1901
2074 pfn = pte_pfn(ptep_get(pvmw.p <<
2075 <<
2076 if (folio_is_zone_device(foli 1902 if (folio_is_zone_device(folio)) {
2077 /* 1903 /*
2078 * Our PTE is a non-p 1904 * Our PTE is a non-present device exclusive entry and
2079 * calculating the su 1905 * calculating the subpage as for the common case would
2080 * result in an inval 1906 * result in an invalid pointer.
2081 * 1907 *
2082 * Since only PAGE_SI 1908 * Since only PAGE_SIZE pages can currently be
2083 * migrated, just set 1909 * migrated, just set it to page. This will need to be
2084 * changed when hugep 1910 * changed when hugepage migrations to device private
2085 * memory are support 1911 * memory are supported.
2086 */ 1912 */
2087 VM_BUG_ON_FOLIO(folio 1913 VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
2088 subpage = &folio->pag 1914 subpage = &folio->page;
2089 } else { 1915 } else {
2090 subpage = folio_page( !! 1916 subpage = folio_page(folio,
>> 1917 pte_pfn(*pvmw.pte) - folio_pfn(folio));
2091 } 1918 }
2092 address = pvmw.address; 1919 address = pvmw.address;
2093 anon_exclusive = folio_test_a 1920 anon_exclusive = folio_test_anon(folio) &&
2094 PageAnonExcl 1921 PageAnonExclusive(subpage);
2095 1922
2096 if (folio_test_hugetlb(folio) 1923 if (folio_test_hugetlb(folio)) {
2097 bool anon = folio_tes 1924 bool anon = folio_test_anon(folio);
2098 1925
2099 /* 1926 /*
2100 * huge_pmd_unshare m 1927 * huge_pmd_unshare may unmap an entire PMD page.
2101 * There is no way of 1928 * There is no way of knowing exactly which PMDs may
2102 * be cached for this 1929 * be cached for this mm, so we must flush them all.
2103 * start/end were alr 1930 * start/end were already adjusted above to cover this
2104 * range. 1931 * range.
2105 */ 1932 */
2106 flush_cache_range(vma 1933 flush_cache_range(vma, range.start, range.end);
2107 1934
2108 /* 1935 /*
2109 * To call huge_pmd_u 1936 * To call huge_pmd_unshare, i_mmap_rwsem must be
2110 * held in write mode 1937 * held in write mode. Caller needs to explicitly
2111 * do this outside rm 1938 * do this outside rmap routines.
2112 * 1939 *
2113 * We also must hold 1940 * We also must hold hugetlb vma_lock in write mode.
2114 * Lock order dictate 1941 * Lock order dictates acquiring vma_lock BEFORE
2115 * i_mmap_rwsem. We 1942 * i_mmap_rwsem. We can only try lock here and
2116 * fail if unsuccessf 1943 * fail if unsuccessful.
2117 */ 1944 */
2118 if (!anon) { 1945 if (!anon) {
2119 VM_BUG_ON(!(f 1946 VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
2120 if (!hugetlb_ 1947 if (!hugetlb_vma_trylock_write(vma)) {
2121 page_ 1948 page_vma_mapped_walk_done(&pvmw);
2122 ret = 1949 ret = false;
2123 break 1950 break;
2124 } 1951 }
2125 if (huge_pmd_ 1952 if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
2126 huget 1953 hugetlb_vma_unlock_write(vma);
2127 flush 1954 flush_tlb_range(vma,
2128 1955 range.start, range.end);
>> 1956 mmu_notifier_invalidate_range(mm,
>> 1957 range.start, range.end);
2129 1958
2130 /* 1959 /*
2131 * Th 1960 * The ref count of the PMD page was
2132 * dr 1961 * dropped which is part of the way map
2133 * co 1962 * counting is done for shared PMDs.
2134 * Re 1963 * Return 'true' here. When there is
2135 * no 1964 * no other sharing, huge_pmd_unshare
2136 * re 1965 * returns false and we will unmap the
2137 * ac 1966 * actual page and drop map count
2138 * to 1967 * to zero.
2139 */ 1968 */
2140 page_ 1969 page_vma_mapped_walk_done(&pvmw);
2141 break 1970 break;
2142 } 1971 }
2143 hugetlb_vma_u 1972 hugetlb_vma_unlock_write(vma);
2144 } 1973 }
2145 /* Nuke the hugetlb p 1974 /* Nuke the hugetlb page table entry */
2146 pteval = huge_ptep_cl 1975 pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
2147 } else { 1976 } else {
2148 flush_cache_page(vma, !! 1977 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2149 /* Nuke the page tabl 1978 /* Nuke the page table entry. */
2150 if (should_defer_flus !! 1979 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2151 /* <<
2152 * We clear t <<
2153 * a remote C <<
2154 * If the ent <<
2155 * architectu <<
2156 * transition <<
2157 * and traps <<
2158 */ <<
2159 pteval = ptep <<
2160 <<
2161 set_tlb_ubc_f <<
2162 } else { <<
2163 pteval = ptep <<
2164 } <<
2165 } 1980 }
2166 1981
2167 /* Set the dirty flag on the 1982 /* Set the dirty flag on the folio now the pte is gone. */
2168 if (pte_dirty(pteval)) 1983 if (pte_dirty(pteval))
2169 folio_mark_dirty(foli 1984 folio_mark_dirty(folio);
2170 1985
2171 /* Update high watermark befo 1986 /* Update high watermark before we lower rss */
2172 update_hiwater_rss(mm); 1987 update_hiwater_rss(mm);
2173 1988
2174 if (folio_is_device_private(f 1989 if (folio_is_device_private(folio)) {
2175 unsigned long pfn = f 1990 unsigned long pfn = folio_pfn(folio);
2176 swp_entry_t entry; 1991 swp_entry_t entry;
2177 pte_t swp_pte; 1992 pte_t swp_pte;
2178 1993
2179 if (anon_exclusive) 1994 if (anon_exclusive)
2180 WARN_ON_ONCE( !! 1995 BUG_ON(page_try_share_anon_rmap(subpage));
2181 <<
2182 1996
2183 /* 1997 /*
2184 * Store the pfn of t 1998 * Store the pfn of the page in a special migration
2185 * pte. do_swap_page( 1999 * pte. do_swap_page() will wait until the migration
2186 * pte is removed and 2000 * pte is removed and then restart fault handling.
2187 */ 2001 */
2188 entry = pte_to_swp_en 2002 entry = pte_to_swp_entry(pteval);
2189 if (is_writable_devic 2003 if (is_writable_device_private_entry(entry))
2190 entry = make_ 2004 entry = make_writable_migration_entry(pfn);
2191 else if (anon_exclusi 2005 else if (anon_exclusive)
2192 entry = make_ 2006 entry = make_readable_exclusive_migration_entry(pfn);
2193 else 2007 else
2194 entry = make_ 2008 entry = make_readable_migration_entry(pfn);
2195 swp_pte = swp_entry_t 2009 swp_pte = swp_entry_to_pte(entry);
2196 2010
2197 /* 2011 /*
2198 * pteval maps a zone 2012 * pteval maps a zone device page and is therefore
2199 * a swap pte. 2013 * a swap pte.
2200 */ 2014 */
2201 if (pte_swp_soft_dirt 2015 if (pte_swp_soft_dirty(pteval))
2202 swp_pte = pte 2016 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2203 if (pte_swp_uffd_wp(p 2017 if (pte_swp_uffd_wp(pteval))
2204 swp_pte = pte 2018 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2205 set_pte_at(mm, pvmw.a 2019 set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2206 trace_set_migration_p 2020 trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2207 !! 2021 compound_order(&folio->page));
2208 /* 2022 /*
2209 * No need to invalid 2023 * No need to invalidate here it will synchronize on
2210 * against the specia 2024 * against the special swap migration pte.
2211 */ 2025 */
2212 } else if (PageHWPoison(subpa 2026 } else if (PageHWPoison(subpage)) {
2213 pteval = swp_entry_to 2027 pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2214 if (folio_test_hugetl 2028 if (folio_test_hugetlb(folio)) {
2215 hugetlb_count 2029 hugetlb_count_sub(folio_nr_pages(folio), mm);
2216 set_huge_pte_ !! 2030 set_huge_pte_at(mm, address, pvmw.pte, pteval);
2217 <<
2218 } else { 2031 } else {
2219 dec_mm_counte !! 2032 dec_mm_counter(mm, mm_counter(&folio->page));
2220 set_pte_at(mm 2033 set_pte_at(mm, address, pvmw.pte, pteval);
2221 } 2034 }
2222 2035
2223 } else if (pte_unused(pteval) 2036 } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2224 /* 2037 /*
2225 * The guest indicate 2038 * The guest indicated that the page content is of no
2226 * interest anymore. 2039 * interest anymore. Simply discard the pte, vmscan
2227 * will take care of 2040 * will take care of the rest.
2228 * A future reference 2041 * A future reference will then fault in a new zero
2229 * page. When userfau 2042 * page. When userfaultfd is active, we must not drop
2230 * this page though, 2043 * this page though, as its main user (postcopy
2231 * migration) will no 2044 * migration) will not expect userfaults on already
2232 * copied pages. 2045 * copied pages.
2233 */ 2046 */
2234 dec_mm_counter(mm, mm !! 2047 dec_mm_counter(mm, mm_counter(&folio->page));
>> 2048 /* We have to invalidate as we cleared the pte */
>> 2049 mmu_notifier_invalidate_range(mm, address,
>> 2050 address + PAGE_SIZE);
2235 } else { 2051 } else {
2236 swp_entry_t entry; 2052 swp_entry_t entry;
2237 pte_t swp_pte; 2053 pte_t swp_pte;
2238 2054
2239 if (arch_unmap_one(mm 2055 if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2240 if (folio_tes 2056 if (folio_test_hugetlb(folio))
2241 set_h !! 2057 set_huge_pte_at(mm, address, pvmw.pte, pteval);
2242 <<
2243 else 2058 else
2244 set_p 2059 set_pte_at(mm, address, pvmw.pte, pteval);
2245 ret = false; 2060 ret = false;
2246 page_vma_mapp 2061 page_vma_mapped_walk_done(&pvmw);
2247 break; 2062 break;
2248 } 2063 }
2249 VM_BUG_ON_PAGE(pte_wr 2064 VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2250 !anon_ 2065 !anon_exclusive, subpage);
2251 2066
2252 /* See folio_try_shar !! 2067 /* See page_try_share_anon_rmap(): clear PTE first. */
2253 if (folio_test_hugetl !! 2068 if (anon_exclusive &&
2254 if (anon_excl !! 2069 page_try_share_anon_rmap(subpage)) {
2255 hugetlb_t !! 2070 if (folio_test_hugetlb(folio))
2256 set_h !! 2071 set_huge_pte_at(mm, address, pvmw.pte, pteval);
2257 !! 2072 else
2258 ret = !! 2073 set_pte_at(mm, address, pvmw.pte, pteval);
2259 page_ <<
2260 break <<
2261 } <<
2262 } else if (anon_exclu <<
2263 folio_try_ <<
2264 set_pte_at(mm <<
2265 ret = false; 2074 ret = false;
2266 page_vma_mapp 2075 page_vma_mapped_walk_done(&pvmw);
2267 break; 2076 break;
2268 } 2077 }
2269 2078
2270 /* 2079 /*
2271 * Store the pfn of t 2080 * Store the pfn of the page in a special migration
2272 * pte. do_swap_page( 2081 * pte. do_swap_page() will wait until the migration
2273 * pte is removed and 2082 * pte is removed and then restart fault handling.
2274 */ 2083 */
2275 if (pte_write(pteval) 2084 if (pte_write(pteval))
2276 entry = make_ 2085 entry = make_writable_migration_entry(
2277 2086 page_to_pfn(subpage));
2278 else if (anon_exclusi 2087 else if (anon_exclusive)
2279 entry = make_ 2088 entry = make_readable_exclusive_migration_entry(
2280 2089 page_to_pfn(subpage));
2281 else 2090 else
2282 entry = make_ 2091 entry = make_readable_migration_entry(
2283 2092 page_to_pfn(subpage));
2284 if (pte_young(pteval) 2093 if (pte_young(pteval))
2285 entry = make_ 2094 entry = make_migration_entry_young(entry);
2286 if (pte_dirty(pteval) 2095 if (pte_dirty(pteval))
2287 entry = make_ 2096 entry = make_migration_entry_dirty(entry);
2288 swp_pte = swp_entry_t 2097 swp_pte = swp_entry_to_pte(entry);
2289 if (pte_soft_dirty(pt 2098 if (pte_soft_dirty(pteval))
2290 swp_pte = pte 2099 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2291 if (pte_uffd_wp(pteva 2100 if (pte_uffd_wp(pteval))
2292 swp_pte = pte 2101 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2293 if (folio_test_hugetl 2102 if (folio_test_hugetlb(folio))
2294 set_huge_pte_ !! 2103 set_huge_pte_at(mm, address, pvmw.pte, swp_pte);
2295 <<
2296 else 2104 else
2297 set_pte_at(mm 2105 set_pte_at(mm, address, pvmw.pte, swp_pte);
2298 trace_set_migration_p 2106 trace_set_migration_pte(address, pte_val(swp_pte),
2299 !! 2107 compound_order(&folio->page));
2300 /* 2108 /*
2301 * No need to invalid 2109 * No need to invalidate here it will synchronize on
2302 * against the specia 2110 * against the special swap migration pte.
2303 */ 2111 */
2304 } 2112 }
2305 2113
2306 if (unlikely(folio_test_huget !! 2114 /*
2307 hugetlb_remove_rmap(f !! 2115 * No need to call mmu_notifier_invalidate_range() it has be
2308 else !! 2116 * done above for all cases requiring it to happen under page
2309 folio_remove_rmap_pte !! 2117 * table lock before mmu_notifier_invalidate_range_end()
>> 2118 *
>> 2119 * See Documentation/mm/mmu_notifier.rst
>> 2120 */
>> 2121 page_remove_rmap(subpage, vma, folio_test_hugetlb(folio));
2310 if (vma->vm_flags & VM_LOCKED 2122 if (vma->vm_flags & VM_LOCKED)
2311 mlock_drain_local(); !! 2123 mlock_page_drain_local();
2312 folio_put(folio); 2124 folio_put(folio);
2313 } 2125 }
2314 2126
2315 mmu_notifier_invalidate_range_end(&ra 2127 mmu_notifier_invalidate_range_end(&range);
2316 2128
2317 return ret; 2129 return ret;
2318 } 2130 }
2319 2131
2320 /** 2132 /**
2321 * try_to_migrate - try to replace all page t 2133 * try_to_migrate - try to replace all page table mappings with swap entries
2322 * @folio: the folio to replace page table en 2134 * @folio: the folio to replace page table entries for
2323 * @flags: action and flags 2135 * @flags: action and flags
2324 * 2136 *
2325 * Tries to remove all the page table entries 2137 * Tries to remove all the page table entries which are mapping this folio and
2326 * replace them with special swap entries. Ca 2138 * replace them with special swap entries. Caller must hold the folio lock.
2327 */ 2139 */
2328 void try_to_migrate(struct folio *folio, enum 2140 void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2329 { 2141 {
2330 struct rmap_walk_control rwc = { 2142 struct rmap_walk_control rwc = {
2331 .rmap_one = try_to_migrate_on 2143 .rmap_one = try_to_migrate_one,
2332 .arg = (void *)flags, 2144 .arg = (void *)flags,
2333 .done = folio_not_mapped, 2145 .done = folio_not_mapped,
2334 .anon_lock = folio_lock_anon_ 2146 .anon_lock = folio_lock_anon_vma_read,
2335 }; 2147 };
2336 2148
2337 /* 2149 /*
2338 * Migration always ignores mlock and 2150 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2339 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and !! 2151 * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
2340 */ 2152 */
2341 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_L 2153 if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2342 TTU_S !! 2154 TTU_SYNC)))
2343 return; 2155 return;
2344 2156
2345 if (folio_is_zone_device(folio) && 2157 if (folio_is_zone_device(folio) &&
2346 (!folio_is_device_private(folio) 2158 (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2347 return; 2159 return;
2348 2160
2349 /* 2161 /*
2350 * During exec, a temporary VMA is se 2162 * During exec, a temporary VMA is setup and later moved.
2351 * The VMA is moved under the anon_vm 2163 * The VMA is moved under the anon_vma lock but not the
2352 * page tables leading to a race wher 2164 * page tables leading to a race where migration cannot
2353 * find the migration ptes. Rather th 2165 * find the migration ptes. Rather than increasing the
2354 * locking requirements of exec(), mi 2166 * locking requirements of exec(), migration skips
2355 * temporary VMAs until after exec() 2167 * temporary VMAs until after exec() completes.
2356 */ 2168 */
2357 if (!folio_test_ksm(folio) && folio_t 2169 if (!folio_test_ksm(folio) && folio_test_anon(folio))
2358 rwc.invalid_vma = invalid_mig 2170 rwc.invalid_vma = invalid_migration_vma;
2359 2171
2360 if (flags & TTU_RMAP_LOCKED) 2172 if (flags & TTU_RMAP_LOCKED)
2361 rmap_walk_locked(folio, &rwc) 2173 rmap_walk_locked(folio, &rwc);
2362 else 2174 else
2363 rmap_walk(folio, &rwc); 2175 rmap_walk(folio, &rwc);
2364 } 2176 }
2365 2177
2366 #ifdef CONFIG_DEVICE_PRIVATE 2178 #ifdef CONFIG_DEVICE_PRIVATE
2367 struct make_exclusive_args { 2179 struct make_exclusive_args {
2368 struct mm_struct *mm; 2180 struct mm_struct *mm;
2369 unsigned long address; 2181 unsigned long address;
2370 void *owner; 2182 void *owner;
2371 bool valid; 2183 bool valid;
2372 }; 2184 };
2373 2185
2374 static bool page_make_device_exclusive_one(st 2186 static bool page_make_device_exclusive_one(struct folio *folio,
2375 struct vm_area_struct *vma, u 2187 struct vm_area_struct *vma, unsigned long address, void *priv)
2376 { 2188 {
2377 struct mm_struct *mm = vma->vm_mm; 2189 struct mm_struct *mm = vma->vm_mm;
2378 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vm 2190 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2379 struct make_exclusive_args *args = pr 2191 struct make_exclusive_args *args = priv;
2380 pte_t pteval; 2192 pte_t pteval;
2381 struct page *subpage; 2193 struct page *subpage;
2382 bool ret = true; 2194 bool ret = true;
2383 struct mmu_notifier_range range; 2195 struct mmu_notifier_range range;
2384 swp_entry_t entry; 2196 swp_entry_t entry;
2385 pte_t swp_pte; 2197 pte_t swp_pte;
2386 pte_t ptent; <<
2387 2198
2388 mmu_notifier_range_init_owner(&range, !! 2199 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
2389 vma->vm 2200 vma->vm_mm, address, min(vma->vm_end,
2390 address 2201 address + folio_size(folio)),
2391 args->o 2202 args->owner);
2392 mmu_notifier_invalidate_range_start(& 2203 mmu_notifier_invalidate_range_start(&range);
2393 2204
2394 while (page_vma_mapped_walk(&pvmw)) { 2205 while (page_vma_mapped_walk(&pvmw)) {
2395 /* Unexpected PMD-mapped THP? 2206 /* Unexpected PMD-mapped THP? */
2396 VM_BUG_ON_FOLIO(!pvmw.pte, fo 2207 VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2397 2208
2398 ptent = ptep_get(pvmw.pte); !! 2209 if (!pte_present(*pvmw.pte)) {
2399 if (!pte_present(ptent)) { <<
2400 ret = false; 2210 ret = false;
2401 page_vma_mapped_walk_ 2211 page_vma_mapped_walk_done(&pvmw);
2402 break; 2212 break;
2403 } 2213 }
2404 2214
2405 subpage = folio_page(folio, 2215 subpage = folio_page(folio,
2406 pte_pfn(ptent !! 2216 pte_pfn(*pvmw.pte) - folio_pfn(folio));
2407 address = pvmw.address; 2217 address = pvmw.address;
2408 2218
2409 /* Nuke the page table entry. 2219 /* Nuke the page table entry. */
2410 flush_cache_page(vma, address !! 2220 flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
2411 pteval = ptep_clear_flush(vma 2221 pteval = ptep_clear_flush(vma, address, pvmw.pte);
2412 2222
2413 /* Set the dirty flag on the 2223 /* Set the dirty flag on the folio now the pte is gone. */
2414 if (pte_dirty(pteval)) 2224 if (pte_dirty(pteval))
2415 folio_mark_dirty(foli 2225 folio_mark_dirty(folio);
2416 2226
2417 /* 2227 /*
2418 * Check that our target page 2228 * Check that our target page is still mapped at the expected
2419 * address. 2229 * address.
2420 */ 2230 */
2421 if (args->mm == mm && args->a 2231 if (args->mm == mm && args->address == address &&
2422 pte_write(pteval)) 2232 pte_write(pteval))
2423 args->valid = true; 2233 args->valid = true;
2424 2234
2425 /* 2235 /*
2426 * Store the pfn of the page 2236 * Store the pfn of the page in a special migration
2427 * pte. do_swap_page() will w 2237 * pte. do_swap_page() will wait until the migration
2428 * pte is removed and then re 2238 * pte is removed and then restart fault handling.
2429 */ 2239 */
2430 if (pte_write(pteval)) 2240 if (pte_write(pteval))
2431 entry = make_writable 2241 entry = make_writable_device_exclusive_entry(
2432 2242 page_to_pfn(subpage));
2433 else 2243 else
2434 entry = make_readable 2244 entry = make_readable_device_exclusive_entry(
2435 2245 page_to_pfn(subpage));
2436 swp_pte = swp_entry_to_pte(en 2246 swp_pte = swp_entry_to_pte(entry);
2437 if (pte_soft_dirty(pteval)) 2247 if (pte_soft_dirty(pteval))
2438 swp_pte = pte_swp_mks 2248 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2439 if (pte_uffd_wp(pteval)) 2249 if (pte_uffd_wp(pteval))
2440 swp_pte = pte_swp_mku 2250 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2441 2251
2442 set_pte_at(mm, address, pvmw. 2252 set_pte_at(mm, address, pvmw.pte, swp_pte);
2443 2253
2444 /* 2254 /*
2445 * There is a reference on th 2255 * There is a reference on the page for the swap entry which has
2446 * been removed, so shouldn't 2256 * been removed, so shouldn't take another.
2447 */ 2257 */
2448 folio_remove_rmap_pte(folio, !! 2258 page_remove_rmap(subpage, vma, false);
2449 } 2259 }
2450 2260
2451 mmu_notifier_invalidate_range_end(&ra 2261 mmu_notifier_invalidate_range_end(&range);
2452 2262
2453 return ret; 2263 return ret;
2454 } 2264 }
2455 2265
2456 /** 2266 /**
2457 * folio_make_device_exclusive - Mark the fol 2267 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2458 * @folio: The folio to replace page table en 2268 * @folio: The folio to replace page table entries for.
2459 * @mm: The mm_struct where the folio is expe 2269 * @mm: The mm_struct where the folio is expected to be mapped.
2460 * @address: Address where the folio is expec 2270 * @address: Address where the folio is expected to be mapped.
2461 * @owner: passed to MMU_NOTIFY_EXCLUSIVE ran 2271 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2462 * 2272 *
2463 * Tries to remove all the page table entries 2273 * Tries to remove all the page table entries which are mapping this
2464 * folio and replace them with special device 2274 * folio and replace them with special device exclusive swap entries to
2465 * grant a device exclusive access to the fol 2275 * grant a device exclusive access to the folio.
2466 * 2276 *
2467 * Context: Caller must hold the folio lock. 2277 * Context: Caller must hold the folio lock.
2468 * Return: false if the page is still mapped, 2278 * Return: false if the page is still mapped, or if it could not be unmapped
2469 * from the expected address. Otherwise retur 2279 * from the expected address. Otherwise returns true (success).
2470 */ 2280 */
2471 static bool folio_make_device_exclusive(struc 2281 static bool folio_make_device_exclusive(struct folio *folio,
2472 struct mm_struct *mm, unsigne 2282 struct mm_struct *mm, unsigned long address, void *owner)
2473 { 2283 {
2474 struct make_exclusive_args args = { 2284 struct make_exclusive_args args = {
2475 .mm = mm, 2285 .mm = mm,
2476 .address = address, 2286 .address = address,
2477 .owner = owner, 2287 .owner = owner,
2478 .valid = false, 2288 .valid = false,
2479 }; 2289 };
2480 struct rmap_walk_control rwc = { 2290 struct rmap_walk_control rwc = {
2481 .rmap_one = page_make_device_ 2291 .rmap_one = page_make_device_exclusive_one,
2482 .done = folio_not_mapped, 2292 .done = folio_not_mapped,
2483 .anon_lock = folio_lock_anon_ 2293 .anon_lock = folio_lock_anon_vma_read,
2484 .arg = &args, 2294 .arg = &args,
2485 }; 2295 };
2486 2296
2487 /* 2297 /*
2488 * Restrict to anonymous folios for n 2298 * Restrict to anonymous folios for now to avoid potential writeback
2489 * issues. 2299 * issues.
2490 */ 2300 */
2491 if (!folio_test_anon(folio)) 2301 if (!folio_test_anon(folio))
2492 return false; 2302 return false;
2493 2303
2494 rmap_walk(folio, &rwc); 2304 rmap_walk(folio, &rwc);
2495 2305
2496 return args.valid && !folio_mapcount( 2306 return args.valid && !folio_mapcount(folio);
2497 } 2307 }
2498 2308
2499 /** 2309 /**
2500 * make_device_exclusive_range() - Mark a ran 2310 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2501 * @mm: mm_struct of associated target proces 2311 * @mm: mm_struct of associated target process
2502 * @start: start of the region to mark for ex 2312 * @start: start of the region to mark for exclusive device access
2503 * @end: end address of region 2313 * @end: end address of region
2504 * @pages: returns the pages which were succe 2314 * @pages: returns the pages which were successfully marked for exclusive access
2505 * @owner: passed to MMU_NOTIFY_EXCLUSIVE ran 2315 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2506 * 2316 *
2507 * Returns: number of pages found in the rang 2317 * Returns: number of pages found in the range by GUP. A page is marked for
2508 * exclusive access only if the page pointer 2318 * exclusive access only if the page pointer is non-NULL.
2509 * 2319 *
2510 * This function finds ptes mapping page(s) t 2320 * This function finds ptes mapping page(s) to the given address range, locks
2511 * them and replaces mappings with special sw 2321 * them and replaces mappings with special swap entries preventing userspace CPU
2512 * access. On fault these entries are replace 2322 * access. On fault these entries are replaced with the original mapping after
2513 * calling MMU notifiers. 2323 * calling MMU notifiers.
2514 * 2324 *
2515 * A driver using this to program access from 2325 * A driver using this to program access from a device must use a mmu notifier
2516 * critical section to hold a device specific 2326 * critical section to hold a device specific lock during programming. Once
2517 * programming is complete it should drop the 2327 * programming is complete it should drop the page lock and reference after
2518 * which point CPU access to the page will re 2328 * which point CPU access to the page will revoke the exclusive access.
2519 */ 2329 */
2520 int make_device_exclusive_range(struct mm_str 2330 int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2521 unsigned long 2331 unsigned long end, struct page **pages,
2522 void *owner) 2332 void *owner)
2523 { 2333 {
2524 long npages = (end - start) >> PAGE_S 2334 long npages = (end - start) >> PAGE_SHIFT;
2525 long i; 2335 long i;
2526 2336
2527 npages = get_user_pages_remote(mm, st 2337 npages = get_user_pages_remote(mm, start, npages,
2528 FOLL_G 2338 FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2529 pages, !! 2339 pages, NULL, NULL);
2530 if (npages < 0) 2340 if (npages < 0)
2531 return npages; 2341 return npages;
2532 2342
2533 for (i = 0; i < npages; i++, start += 2343 for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2534 struct folio *folio = page_fo 2344 struct folio *folio = page_folio(pages[i]);
2535 if (PageTail(pages[i]) || !fo 2345 if (PageTail(pages[i]) || !folio_trylock(folio)) {
2536 folio_put(folio); 2346 folio_put(folio);
2537 pages[i] = NULL; 2347 pages[i] = NULL;
2538 continue; 2348 continue;
2539 } 2349 }
2540 2350
2541 if (!folio_make_device_exclus 2351 if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2542 folio_unlock(folio); 2352 folio_unlock(folio);
2543 folio_put(folio); 2353 folio_put(folio);
2544 pages[i] = NULL; 2354 pages[i] = NULL;
2545 } 2355 }
2546 } 2356 }
2547 2357
2548 return npages; 2358 return npages;
2549 } 2359 }
2550 EXPORT_SYMBOL_GPL(make_device_exclusive_range 2360 EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2551 #endif 2361 #endif
2552 2362
2553 void __put_anon_vma(struct anon_vma *anon_vma 2363 void __put_anon_vma(struct anon_vma *anon_vma)
2554 { 2364 {
2555 struct anon_vma *root = anon_vma->roo 2365 struct anon_vma *root = anon_vma->root;
2556 2366
2557 anon_vma_free(anon_vma); 2367 anon_vma_free(anon_vma);
2558 if (root != anon_vma && atomic_dec_an 2368 if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2559 anon_vma_free(root); 2369 anon_vma_free(root);
2560 } 2370 }
2561 2371
2562 static struct anon_vma *rmap_walk_anon_lock(s 2372 static struct anon_vma *rmap_walk_anon_lock(struct folio *folio,
2563 s 2373 struct rmap_walk_control *rwc)
2564 { 2374 {
2565 struct anon_vma *anon_vma; 2375 struct anon_vma *anon_vma;
2566 2376
2567 if (rwc->anon_lock) 2377 if (rwc->anon_lock)
2568 return rwc->anon_lock(folio, 2378 return rwc->anon_lock(folio, rwc);
2569 2379
2570 /* 2380 /*
2571 * Note: remove_migration_ptes() cann 2381 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2572 * because that depends on page_mappe 2382 * because that depends on page_mapped(); but not all its usages
2573 * are holding mmap_lock. Users witho 2383 * are holding mmap_lock. Users without mmap_lock are required to
2574 * take a reference count to prevent 2384 * take a reference count to prevent the anon_vma disappearing
2575 */ 2385 */
2576 anon_vma = folio_anon_vma(folio); 2386 anon_vma = folio_anon_vma(folio);
2577 if (!anon_vma) 2387 if (!anon_vma)
2578 return NULL; 2388 return NULL;
2579 2389
2580 if (anon_vma_trylock_read(anon_vma)) 2390 if (anon_vma_trylock_read(anon_vma))
2581 goto out; 2391 goto out;
2582 2392
2583 if (rwc->try_lock) { 2393 if (rwc->try_lock) {
2584 anon_vma = NULL; 2394 anon_vma = NULL;
2585 rwc->contended = true; 2395 rwc->contended = true;
2586 goto out; 2396 goto out;
2587 } 2397 }
2588 2398
2589 anon_vma_lock_read(anon_vma); 2399 anon_vma_lock_read(anon_vma);
2590 out: 2400 out:
2591 return anon_vma; 2401 return anon_vma;
2592 } 2402 }
2593 2403
2594 /* 2404 /*
2595 * rmap_walk_anon - do something to anonymous 2405 * rmap_walk_anon - do something to anonymous page using the object-based
2596 * rmap method 2406 * rmap method
2597 * @folio: the folio to be handled !! 2407 * @page: the page to be handled
2598 * @rwc: control variable according to each w 2408 * @rwc: control variable according to each walk type
2599 * @locked: caller holds relevant rmap lock <<
2600 * 2409 *
2601 * Find all the mappings of a folio using the !! 2410 * Find all the mappings of a page using the mapping pointer and the vma chains
2602 * chains contained in the anon_vma struct it !! 2411 * contained in the anon_vma struct it points to.
2603 */ 2412 */
2604 static void rmap_walk_anon(struct folio *foli 2413 static void rmap_walk_anon(struct folio *folio,
2605 struct rmap_walk_control *rwc 2414 struct rmap_walk_control *rwc, bool locked)
2606 { 2415 {
2607 struct anon_vma *anon_vma; 2416 struct anon_vma *anon_vma;
2608 pgoff_t pgoff_start, pgoff_end; 2417 pgoff_t pgoff_start, pgoff_end;
2609 struct anon_vma_chain *avc; 2418 struct anon_vma_chain *avc;
2610 2419
2611 if (locked) { 2420 if (locked) {
2612 anon_vma = folio_anon_vma(fol 2421 anon_vma = folio_anon_vma(folio);
2613 /* anon_vma disappear under u 2422 /* anon_vma disappear under us? */
2614 VM_BUG_ON_FOLIO(!anon_vma, fo 2423 VM_BUG_ON_FOLIO(!anon_vma, folio);
2615 } else { 2424 } else {
2616 anon_vma = rmap_walk_anon_loc 2425 anon_vma = rmap_walk_anon_lock(folio, rwc);
2617 } 2426 }
2618 if (!anon_vma) 2427 if (!anon_vma)
2619 return; 2428 return;
2620 2429
2621 pgoff_start = folio_pgoff(folio); 2430 pgoff_start = folio_pgoff(folio);
2622 pgoff_end = pgoff_start + folio_nr_pa 2431 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2623 anon_vma_interval_tree_foreach(avc, & 2432 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2624 pgoff_start, pgoff_en 2433 pgoff_start, pgoff_end) {
2625 struct vm_area_struct *vma = 2434 struct vm_area_struct *vma = avc->vma;
2626 unsigned long address = vma_a !! 2435 unsigned long address = vma_address(&folio->page, vma);
2627 folio_nr_page <<
2628 2436
2629 VM_BUG_ON_VMA(address == -EFA 2437 VM_BUG_ON_VMA(address == -EFAULT, vma);
2630 cond_resched(); 2438 cond_resched();
2631 2439
2632 if (rwc->invalid_vma && rwc-> 2440 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2633 continue; 2441 continue;
2634 2442
2635 if (!rwc->rmap_one(folio, vma 2443 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2636 break; 2444 break;
2637 if (rwc->done && rwc->done(fo 2445 if (rwc->done && rwc->done(folio))
2638 break; 2446 break;
2639 } 2447 }
2640 2448
2641 if (!locked) 2449 if (!locked)
2642 anon_vma_unlock_read(anon_vma 2450 anon_vma_unlock_read(anon_vma);
2643 } 2451 }
2644 2452
2645 /* 2453 /*
2646 * rmap_walk_file - do something to file page 2454 * rmap_walk_file - do something to file page using the object-based rmap method
2647 * @folio: the folio to be handled !! 2455 * @page: the page to be handled
2648 * @rwc: control variable according to each w 2456 * @rwc: control variable according to each walk type
2649 * @locked: caller holds relevant rmap lock <<
2650 * 2457 *
2651 * Find all the mappings of a folio using the !! 2458 * Find all the mappings of a page using the mapping pointer and the vma chains
2652 * contained in the address_space struct it p 2459 * contained in the address_space struct it points to.
2653 */ 2460 */
2654 static void rmap_walk_file(struct folio *foli 2461 static void rmap_walk_file(struct folio *folio,
2655 struct rmap_walk_control *rwc 2462 struct rmap_walk_control *rwc, bool locked)
2656 { 2463 {
2657 struct address_space *mapping = folio 2464 struct address_space *mapping = folio_mapping(folio);
2658 pgoff_t pgoff_start, pgoff_end; 2465 pgoff_t pgoff_start, pgoff_end;
2659 struct vm_area_struct *vma; 2466 struct vm_area_struct *vma;
2660 2467
2661 /* 2468 /*
2662 * The page lock not only makes sure 2469 * The page lock not only makes sure that page->mapping cannot
2663 * suddenly be NULLified by truncatio 2470 * suddenly be NULLified by truncation, it makes sure that the
2664 * structure at mapping cannot be fre 2471 * structure at mapping cannot be freed and reused yet,
2665 * so we can safely take mapping->i_m 2472 * so we can safely take mapping->i_mmap_rwsem.
2666 */ 2473 */
2667 VM_BUG_ON_FOLIO(!folio_test_locked(fo 2474 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2668 2475
2669 if (!mapping) 2476 if (!mapping)
2670 return; 2477 return;
2671 2478
2672 pgoff_start = folio_pgoff(folio); 2479 pgoff_start = folio_pgoff(folio);
2673 pgoff_end = pgoff_start + folio_nr_pa 2480 pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2674 if (!locked) { 2481 if (!locked) {
2675 if (i_mmap_trylock_read(mappi 2482 if (i_mmap_trylock_read(mapping))
2676 goto lookup; 2483 goto lookup;
2677 2484
2678 if (rwc->try_lock) { 2485 if (rwc->try_lock) {
2679 rwc->contended = true 2486 rwc->contended = true;
2680 return; 2487 return;
2681 } 2488 }
2682 2489
2683 i_mmap_lock_read(mapping); 2490 i_mmap_lock_read(mapping);
2684 } 2491 }
2685 lookup: 2492 lookup:
2686 vma_interval_tree_foreach(vma, &mappi 2493 vma_interval_tree_foreach(vma, &mapping->i_mmap,
2687 pgoff_start, pgoff_en 2494 pgoff_start, pgoff_end) {
2688 unsigned long address = vma_a !! 2495 unsigned long address = vma_address(&folio->page, vma);
2689 folio_nr_pages <<
2690 2496
2691 VM_BUG_ON_VMA(address == -EFA 2497 VM_BUG_ON_VMA(address == -EFAULT, vma);
2692 cond_resched(); 2498 cond_resched();
2693 2499
2694 if (rwc->invalid_vma && rwc-> 2500 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2695 continue; 2501 continue;
2696 2502
2697 if (!rwc->rmap_one(folio, vma 2503 if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2698 goto done; 2504 goto done;
2699 if (rwc->done && rwc->done(fo 2505 if (rwc->done && rwc->done(folio))
2700 goto done; 2506 goto done;
2701 } 2507 }
2702 2508
2703 done: 2509 done:
2704 if (!locked) 2510 if (!locked)
2705 i_mmap_unlock_read(mapping); 2511 i_mmap_unlock_read(mapping);
2706 } 2512 }
2707 2513
2708 void rmap_walk(struct folio *folio, struct rm 2514 void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2709 { 2515 {
2710 if (unlikely(folio_test_ksm(folio))) 2516 if (unlikely(folio_test_ksm(folio)))
2711 rmap_walk_ksm(folio, rwc); 2517 rmap_walk_ksm(folio, rwc);
2712 else if (folio_test_anon(folio)) 2518 else if (folio_test_anon(folio))
2713 rmap_walk_anon(folio, rwc, fa 2519 rmap_walk_anon(folio, rwc, false);
2714 else 2520 else
2715 rmap_walk_file(folio, rwc, fa 2521 rmap_walk_file(folio, rwc, false);
2716 } 2522 }
2717 2523
2718 /* Like rmap_walk, but caller holds relevant 2524 /* Like rmap_walk, but caller holds relevant rmap lock */
2719 void rmap_walk_locked(struct folio *folio, st 2525 void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2720 { 2526 {
2721 /* no ksm support for now */ 2527 /* no ksm support for now */
2722 VM_BUG_ON_FOLIO(folio_test_ksm(folio) 2528 VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2723 if (folio_test_anon(folio)) 2529 if (folio_test_anon(folio))
2724 rmap_walk_anon(folio, rwc, tr 2530 rmap_walk_anon(folio, rwc, true);
2725 else 2531 else
2726 rmap_walk_file(folio, rwc, tr 2532 rmap_walk_file(folio, rwc, true);
2727 } 2533 }
2728 2534
2729 #ifdef CONFIG_HUGETLB_PAGE 2535 #ifdef CONFIG_HUGETLB_PAGE
2730 /* 2536 /*
2731 * The following two functions are for anonym 2537 * The following two functions are for anonymous (private mapped) hugepages.
2732 * Unlike common anonymous pages, anonymous h 2538 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2733 * and no lru code, because we handle hugepag 2539 * and no lru code, because we handle hugepages differently from common pages.
>> 2540 *
>> 2541 * RMAP_COMPOUND is ignored.
2734 */ 2542 */
2735 void hugetlb_add_anon_rmap(struct folio *foli !! 2543 void hugepage_add_anon_rmap(struct page *page, struct vm_area_struct *vma,
2736 unsigned long address, rmap_t !! 2544 unsigned long address, rmap_t flags)
2737 { 2545 {
2738 VM_WARN_ON_FOLIO(!folio_test_hugetlb( !! 2546 struct anon_vma *anon_vma = vma->anon_vma;
2739 VM_WARN_ON_FOLIO(!folio_test_anon(fol !! 2547 int first;
2740 2548
2741 atomic_inc(&folio->_entire_mapcount); !! 2549 BUG_ON(!PageLocked(page));
2742 atomic_inc(&folio->_large_mapcount); !! 2550 BUG_ON(!anon_vma);
2743 if (flags & RMAP_EXCLUSIVE) !! 2551 /* address might be in next vma when migration races vma_adjust */
2744 SetPageAnonExclusive(&folio-> !! 2552 first = atomic_inc_and_test(compound_mapcount_ptr(page));
2745 VM_WARN_ON_FOLIO(folio_entire_mapcoun !! 2553 VM_BUG_ON_PAGE(!first && (flags & RMAP_EXCLUSIVE), page);
2746 PageAnonExclusive(&f !! 2554 VM_BUG_ON_PAGE(!first && PageAnonExclusive(page), page);
>> 2555 if (first)
>> 2556 __page_set_anon_rmap(page, vma, address,
>> 2557 !!(flags & RMAP_EXCLUSIVE));
2747 } 2558 }
2748 2559
2749 void hugetlb_add_new_anon_rmap(struct folio * !! 2560 void hugepage_add_new_anon_rmap(struct page *page,
2750 struct vm_area_struct *vma, u !! 2561 struct vm_area_struct *vma, unsigned long address)
2751 { 2562 {
2752 VM_WARN_ON_FOLIO(!folio_test_hugetlb( <<
2753 <<
2754 BUG_ON(address < vma->vm_start || add 2563 BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2755 /* increment count (starts at -1) */ 2564 /* increment count (starts at -1) */
2756 atomic_set(&folio->_entire_mapcount, !! 2565 atomic_set(compound_mapcount_ptr(page), 0);
2757 atomic_set(&folio->_large_mapcount, 0 !! 2566 ClearHPageRestoreReserve(page);
2758 folio_clear_hugetlb_restore_reserve(f !! 2567 __page_set_anon_rmap(page, vma, address, 1);
2759 __folio_set_anon(folio, vma, address, <<
2760 SetPageAnonExclusive(&folio->page); <<
2761 } 2568 }
2762 #endif /* CONFIG_HUGETLB_PAGE */ 2569 #endif /* CONFIG_HUGETLB_PAGE */
2763 2570
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.