| 1 | /* |
| 2 | * mm/rmap.c - physical to virtual reverse mappings |
| 3 | * |
| 4 | * Copyright 2001, Rik van Riel <riel@conectiva.com.br> |
| 5 | * Released under the General Public License (GPL). |
| 6 | * |
| 7 | * Simple, low overhead reverse mapping scheme. |
| 8 | * Please try to keep this thing as modular as possible. |
| 9 | * |
| 10 | * Provides methods for unmapping each kind of mapped page: |
| 11 | * the anon methods track anonymous pages, and |
| 12 | * the file methods track pages belonging to an inode. |
| 13 | * |
| 14 | * Original design by Rik van Riel <riel@conectiva.com.br> 2001 |
| 15 | * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 |
| 16 | * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 |
| 17 | * Contributions by Hugh Dickins 2003, 2004 |
| 18 | */ |
| 19 | |
| 20 | /* |
| 21 | * Lock ordering in mm: |
| 22 | * |
| 23 | * inode->i_mutex (while writing or truncating, not reading or faulting) |
| 24 | * mm->mmap_sem |
| 25 | * page->flags PG_locked (lock_page) |
| 26 | * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share) |
| 27 | * mapping->i_mmap_rwsem |
| 28 | * anon_vma->rwsem |
| 29 | * mm->page_table_lock or pte_lock |
| 30 | * pgdat->lru_lock (in mark_page_accessed, isolate_lru_page) |
| 31 | * swap_lock (in swap_duplicate, swap_info_get) |
| 32 | * mmlist_lock (in mmput, drain_mmlist and others) |
| 33 | * mapping->private_lock (in __set_page_dirty_buffers) |
| 34 | * mem_cgroup_{begin,end}_page_stat (memcg->move_lock) |
| 35 | * i_pages lock (widely used) |
| 36 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) |
| 37 | * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) |
| 38 | * sb_lock (within inode_lock in fs/fs-writeback.c) |
| 39 | * i_pages lock (widely used, in set_page_dirty, |
| 40 | * in arch-dependent flush_dcache_mmap_lock, |
| 41 | * within bdi.wb->list_lock in __sync_single_inode) |
| 42 | * |
| 43 | * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) |
| 44 | * ->tasklist_lock |
| 45 | * pte map lock |
| 46 | */ |
| 47 | |
| 48 | #include <linux/mm.h> |
| 49 | #include <linux/sched/mm.h> |
| 50 | #include <linux/sched/task.h> |
| 51 | #include <linux/pagemap.h> |
| 52 | #include <linux/swap.h> |
| 53 | #include <linux/swapops.h> |
| 54 | #include <linux/slab.h> |
| 55 | #include <linux/init.h> |
| 56 | #include <linux/ksm.h> |
| 57 | #include <linux/rmap.h> |
| 58 | #include <linux/rcupdate.h> |
| 59 | #include <linux/export.h> |
| 60 | #include <linux/memcontrol.h> |
| 61 | #include <linux/mmu_notifier.h> |
| 62 | #include <linux/migrate.h> |
| 63 | #include <linux/hugetlb.h> |
| 64 | #include <linux/backing-dev.h> |
| 65 | #include <linux/page_idle.h> |
| 66 | #include <linux/memremap.h> |
| 67 | #include <linux/userfaultfd_k.h> |
| 68 | |
| 69 | #include <asm/tlbflush.h> |
| 70 | |
| 71 | #include <trace/events/tlb.h> |
| 72 | |
| 73 | #include "internal.h" |
| 74 | |
| 75 | static struct kmem_cache *anon_vma_cachep; |
| 76 | static struct kmem_cache *anon_vma_chain_cachep; |
| 77 | |
| 78 | static inline struct anon_vma *anon_vma_alloc(void) |
| 79 | { |
| 80 | struct anon_vma *anon_vma; |
| 81 | |
| 82 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); |
| 83 | if (anon_vma) { |
| 84 | atomic_set(&anon_vma->refcount, 1); |
| 85 | anon_vma->degree = 1; /* Reference for first vma */ |
| 86 | anon_vma->parent = anon_vma; |
| 87 | /* |
| 88 | * Initialise the anon_vma root to point to itself. If called |
| 89 | * from fork, the root will be reset to the parents anon_vma. |
| 90 | */ |
| 91 | anon_vma->root = anon_vma; |
| 92 | } |
| 93 | |
| 94 | return anon_vma; |
| 95 | } |
| 96 | |
| 97 | static inline void anon_vma_free(struct anon_vma *anon_vma) |
| 98 | { |
| 99 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); |
| 100 | |
| 101 | /* |
| 102 | * Synchronize against page_lock_anon_vma_read() such that |
| 103 | * we can safely hold the lock without the anon_vma getting |
| 104 | * freed. |
| 105 | * |
| 106 | * Relies on the full mb implied by the atomic_dec_and_test() from |
| 107 | * put_anon_vma() against the acquire barrier implied by |
| 108 | * down_read_trylock() from page_lock_anon_vma_read(). This orders: |
| 109 | * |
| 110 | * page_lock_anon_vma_read() VS put_anon_vma() |
| 111 | * down_read_trylock() atomic_dec_and_test() |
| 112 | * LOCK MB |
| 113 | * atomic_read() rwsem_is_locked() |
| 114 | * |
| 115 | * LOCK should suffice since the actual taking of the lock must |
| 116 | * happen _before_ what follows. |
| 117 | */ |
| 118 | might_sleep(); |
| 119 | if (rwsem_is_locked(&anon_vma->root->rwsem)) { |
| 120 | anon_vma_lock_write(anon_vma); |
| 121 | anon_vma_unlock_write(anon_vma); |
| 122 | } |
| 123 | |
| 124 | kmem_cache_free(anon_vma_cachep, anon_vma); |
| 125 | } |
| 126 | |
| 127 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) |
| 128 | { |
| 129 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); |
| 130 | } |
| 131 | |
| 132 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) |
| 133 | { |
| 134 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); |
| 135 | } |
| 136 | |
| 137 | static void anon_vma_chain_link(struct vm_area_struct *vma, |
| 138 | struct anon_vma_chain *avc, |
| 139 | struct anon_vma *anon_vma) |
| 140 | { |
| 141 | avc->vma = vma; |
| 142 | avc->anon_vma = anon_vma; |
| 143 | list_add(&avc->same_vma, &vma->anon_vma_chain); |
| 144 | anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); |
| 145 | } |
| 146 | |
| 147 | /** |
| 148 | * __anon_vma_prepare - attach an anon_vma to a memory region |
| 149 | * @vma: the memory region in question |
| 150 | * |
| 151 | * This makes sure the memory mapping described by 'vma' has |
| 152 | * an 'anon_vma' attached to it, so that we can associate the |
| 153 | * anonymous pages mapped into it with that anon_vma. |
| 154 | * |
| 155 | * The common case will be that we already have one, which |
| 156 | * is handled inline by anon_vma_prepare(). But if |
| 157 | * not we either need to find an adjacent mapping that we |
| 158 | * can re-use the anon_vma from (very common when the only |
| 159 | * reason for splitting a vma has been mprotect()), or we |
| 160 | * allocate a new one. |
| 161 | * |
| 162 | * Anon-vma allocations are very subtle, because we may have |
| 163 | * optimistically looked up an anon_vma in page_lock_anon_vma_read() |
| 164 | * and that may actually touch the spinlock even in the newly |
| 165 | * allocated vma (it depends on RCU to make sure that the |
| 166 | * anon_vma isn't actually destroyed). |
| 167 | * |
| 168 | * As a result, we need to do proper anon_vma locking even |
| 169 | * for the new allocation. At the same time, we do not want |
| 170 | * to do any locking for the common case of already having |
| 171 | * an anon_vma. |
| 172 | * |
| 173 | * This must be called with the mmap_sem held for reading. |
| 174 | */ |
| 175 | int __anon_vma_prepare(struct vm_area_struct *vma) |
| 176 | { |
| 177 | struct mm_struct *mm = vma->vm_mm; |
| 178 | struct anon_vma *anon_vma, *allocated; |
| 179 | struct anon_vma_chain *avc; |
| 180 | |
| 181 | might_sleep(); |
| 182 | |
| 183 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 184 | if (!avc) |
| 185 | goto out_enomem; |
| 186 | |
| 187 | anon_vma = find_mergeable_anon_vma(vma); |
| 188 | allocated = NULL; |
| 189 | if (!anon_vma) { |
| 190 | anon_vma = anon_vma_alloc(); |
| 191 | if (unlikely(!anon_vma)) |
| 192 | goto out_enomem_free_avc; |
| 193 | allocated = anon_vma; |
| 194 | } |
| 195 | |
| 196 | anon_vma_lock_write(anon_vma); |
| 197 | /* page_table_lock to protect against threads */ |
| 198 | spin_lock(&mm->page_table_lock); |
| 199 | if (likely(!vma->anon_vma)) { |
| 200 | vma->anon_vma = anon_vma; |
| 201 | anon_vma_chain_link(vma, avc, anon_vma); |
| 202 | /* vma reference or self-parent link for new root */ |
| 203 | anon_vma->degree++; |
| 204 | allocated = NULL; |
| 205 | avc = NULL; |
| 206 | } |
| 207 | spin_unlock(&mm->page_table_lock); |
| 208 | anon_vma_unlock_write(anon_vma); |
| 209 | |
| 210 | if (unlikely(allocated)) |
| 211 | put_anon_vma(allocated); |
| 212 | if (unlikely(avc)) |
| 213 | anon_vma_chain_free(avc); |
| 214 | |
| 215 | return 0; |
| 216 | |
| 217 | out_enomem_free_avc: |
| 218 | anon_vma_chain_free(avc); |
| 219 | out_enomem: |
| 220 | return -ENOMEM; |
| 221 | } |
| 222 | |
| 223 | /* |
| 224 | * This is a useful helper function for locking the anon_vma root as |
| 225 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that |
| 226 | * have the same vma. |
| 227 | * |
| 228 | * Such anon_vma's should have the same root, so you'd expect to see |
| 229 | * just a single mutex_lock for the whole traversal. |
| 230 | */ |
| 231 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) |
| 232 | { |
| 233 | struct anon_vma *new_root = anon_vma->root; |
| 234 | if (new_root != root) { |
| 235 | if (WARN_ON_ONCE(root)) |
| 236 | up_write(&root->rwsem); |
| 237 | root = new_root; |
| 238 | down_write(&root->rwsem); |
| 239 | } |
| 240 | return root; |
| 241 | } |
| 242 | |
| 243 | static inline void unlock_anon_vma_root(struct anon_vma *root) |
| 244 | { |
| 245 | if (root) |
| 246 | up_write(&root->rwsem); |
| 247 | } |
| 248 | |
| 249 | /* |
| 250 | * Attach the anon_vmas from src to dst. |
| 251 | * Returns 0 on success, -ENOMEM on failure. |
| 252 | * |
| 253 | * If dst->anon_vma is NULL this function tries to find and reuse existing |
| 254 | * anon_vma which has no vmas and only one child anon_vma. This prevents |
| 255 | * degradation of anon_vma hierarchy to endless linear chain in case of |
| 256 | * constantly forking task. On the other hand, an anon_vma with more than one |
| 257 | * child isn't reused even if there was no alive vma, thus rmap walker has a |
| 258 | * good chance of avoiding scanning the whole hierarchy when it searches where |
| 259 | * page is mapped. |
| 260 | */ |
| 261 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) |
| 262 | { |
| 263 | struct anon_vma_chain *avc, *pavc; |
| 264 | struct anon_vma *root = NULL; |
| 265 | |
| 266 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { |
| 267 | struct anon_vma *anon_vma; |
| 268 | |
| 269 | avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); |
| 270 | if (unlikely(!avc)) { |
| 271 | unlock_anon_vma_root(root); |
| 272 | root = NULL; |
| 273 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 274 | if (!avc) |
| 275 | goto enomem_failure; |
| 276 | } |
| 277 | anon_vma = pavc->anon_vma; |
| 278 | root = lock_anon_vma_root(root, anon_vma); |
| 279 | anon_vma_chain_link(dst, avc, anon_vma); |
| 280 | |
| 281 | /* |
| 282 | * Reuse existing anon_vma if its degree lower than two, |
| 283 | * that means it has no vma and only one anon_vma child. |
| 284 | * |
| 285 | * Do not chose parent anon_vma, otherwise first child |
| 286 | * will always reuse it. Root anon_vma is never reused: |
| 287 | * it has self-parent reference and at least one child. |
| 288 | */ |
| 289 | if (!dst->anon_vma && anon_vma != src->anon_vma && |
| 290 | anon_vma->degree < 2) |
| 291 | dst->anon_vma = anon_vma; |
| 292 | } |
| 293 | if (dst->anon_vma) |
| 294 | dst->anon_vma->degree++; |
| 295 | unlock_anon_vma_root(root); |
| 296 | return 0; |
| 297 | |
| 298 | enomem_failure: |
| 299 | /* |
| 300 | * dst->anon_vma is dropped here otherwise its degree can be incorrectly |
| 301 | * decremented in unlink_anon_vmas(). |
| 302 | * We can safely do this because callers of anon_vma_clone() don't care |
| 303 | * about dst->anon_vma if anon_vma_clone() failed. |
| 304 | */ |
| 305 | dst->anon_vma = NULL; |
| 306 | unlink_anon_vmas(dst); |
| 307 | return -ENOMEM; |
| 308 | } |
| 309 | |
| 310 | /* |
| 311 | * Attach vma to its own anon_vma, as well as to the anon_vmas that |
| 312 | * the corresponding VMA in the parent process is attached to. |
| 313 | * Returns 0 on success, non-zero on failure. |
| 314 | */ |
| 315 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) |
| 316 | { |
| 317 | struct anon_vma_chain *avc; |
| 318 | struct anon_vma *anon_vma; |
| 319 | int error; |
| 320 | |
| 321 | /* Don't bother if the parent process has no anon_vma here. */ |
| 322 | if (!pvma->anon_vma) |
| 323 | return 0; |
| 324 | |
| 325 | /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ |
| 326 | vma->anon_vma = NULL; |
| 327 | |
| 328 | /* |
| 329 | * First, attach the new VMA to the parent VMA's anon_vmas, |
| 330 | * so rmap can find non-COWed pages in child processes. |
| 331 | */ |
| 332 | error = anon_vma_clone(vma, pvma); |
| 333 | if (error) |
| 334 | return error; |
| 335 | |
| 336 | /* An existing anon_vma has been reused, all done then. */ |
| 337 | if (vma->anon_vma) |
| 338 | return 0; |
| 339 | |
| 340 | /* Then add our own anon_vma. */ |
| 341 | anon_vma = anon_vma_alloc(); |
| 342 | if (!anon_vma) |
| 343 | goto out_error; |
| 344 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 345 | if (!avc) |
| 346 | goto out_error_free_anon_vma; |
| 347 | |
| 348 | /* |
| 349 | * The root anon_vma's spinlock is the lock actually used when we |
| 350 | * lock any of the anon_vmas in this anon_vma tree. |
| 351 | */ |
| 352 | anon_vma->root = pvma->anon_vma->root; |
| 353 | anon_vma->parent = pvma->anon_vma; |
| 354 | /* |
| 355 | * With refcounts, an anon_vma can stay around longer than the |
| 356 | * process it belongs to. The root anon_vma needs to be pinned until |
| 357 | * this anon_vma is freed, because the lock lives in the root. |
| 358 | */ |
| 359 | get_anon_vma(anon_vma->root); |
| 360 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ |
| 361 | vma->anon_vma = anon_vma; |
| 362 | anon_vma_lock_write(anon_vma); |
| 363 | anon_vma_chain_link(vma, avc, anon_vma); |
| 364 | anon_vma->parent->degree++; |
| 365 | anon_vma_unlock_write(anon_vma); |
| 366 | |
| 367 | return 0; |
| 368 | |
| 369 | out_error_free_anon_vma: |
| 370 | put_anon_vma(anon_vma); |
| 371 | out_error: |
| 372 | unlink_anon_vmas(vma); |
| 373 | return -ENOMEM; |
| 374 | } |
| 375 | |
| 376 | void unlink_anon_vmas(struct vm_area_struct *vma) |
| 377 | { |
| 378 | struct anon_vma_chain *avc, *next; |
| 379 | struct anon_vma *root = NULL; |
| 380 | |
| 381 | /* |
| 382 | * Unlink each anon_vma chained to the VMA. This list is ordered |
| 383 | * from newest to oldest, ensuring the root anon_vma gets freed last. |
| 384 | */ |
| 385 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
| 386 | struct anon_vma *anon_vma = avc->anon_vma; |
| 387 | |
| 388 | root = lock_anon_vma_root(root, anon_vma); |
| 389 | anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); |
| 390 | |
| 391 | /* |
| 392 | * Leave empty anon_vmas on the list - we'll need |
| 393 | * to free them outside the lock. |
| 394 | */ |
| 395 | if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) { |
| 396 | anon_vma->parent->degree--; |
| 397 | continue; |
| 398 | } |
| 399 | |
| 400 | list_del(&avc->same_vma); |
| 401 | anon_vma_chain_free(avc); |
| 402 | } |
| 403 | if (vma->anon_vma) |
| 404 | vma->anon_vma->degree--; |
| 405 | unlock_anon_vma_root(root); |
| 406 | |
| 407 | /* |
| 408 | * Iterate the list once more, it now only contains empty and unlinked |
| 409 | * anon_vmas, destroy them. Could not do before due to __put_anon_vma() |
| 410 | * needing to write-acquire the anon_vma->root->rwsem. |
| 411 | */ |
| 412 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
| 413 | struct anon_vma *anon_vma = avc->anon_vma; |
| 414 | |
| 415 | VM_WARN_ON(anon_vma->degree); |
| 416 | put_anon_vma(anon_vma); |
| 417 | |
| 418 | list_del(&avc->same_vma); |
| 419 | anon_vma_chain_free(avc); |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | static void anon_vma_ctor(void *data) |
| 424 | { |
| 425 | struct anon_vma *anon_vma = data; |
| 426 | |
| 427 | init_rwsem(&anon_vma->rwsem); |
| 428 | atomic_set(&anon_vma->refcount, 0); |
| 429 | anon_vma->rb_root = RB_ROOT_CACHED; |
| 430 | } |
| 431 | |
| 432 | void __init anon_vma_init(void) |
| 433 | { |
| 434 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), |
| 435 | 0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, |
| 436 | anon_vma_ctor); |
| 437 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, |
| 438 | SLAB_PANIC|SLAB_ACCOUNT); |
| 439 | } |
| 440 | |
| 441 | /* |
| 442 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! |
| 443 | * |
| 444 | * Since there is no serialization what so ever against page_remove_rmap() |
| 445 | * the best this function can do is return a locked anon_vma that might |
| 446 | * have been relevant to this page. |
| 447 | * |
| 448 | * The page might have been remapped to a different anon_vma or the anon_vma |
| 449 | * returned may already be freed (and even reused). |
| 450 | * |
| 451 | * In case it was remapped to a different anon_vma, the new anon_vma will be a |
| 452 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore |
| 453 | * ensure that any anon_vma obtained from the page will still be valid for as |
| 454 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. |
| 455 | * |
| 456 | * All users of this function must be very careful when walking the anon_vma |
| 457 | * chain and verify that the page in question is indeed mapped in it |
| 458 | * [ something equivalent to page_mapped_in_vma() ]. |
| 459 | * |
| 460 | * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() |
| 461 | * that the anon_vma pointer from page->mapping is valid if there is a |
| 462 | * mapcount, we can dereference the anon_vma after observing those. |
| 463 | */ |
| 464 | struct anon_vma *page_get_anon_vma(struct page *page) |
| 465 | { |
| 466 | struct anon_vma *anon_vma = NULL; |
| 467 | unsigned long anon_mapping; |
| 468 | |
| 469 | rcu_read_lock(); |
| 470 | anon_mapping = (unsigned long)READ_ONCE(page->mapping); |
| 471 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
| 472 | goto out; |
| 473 | if (!page_mapped(page)) |
| 474 | goto out; |
| 475 | |
| 476 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
| 477 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
| 478 | anon_vma = NULL; |
| 479 | goto out; |
| 480 | } |
| 481 | |
| 482 | /* |
| 483 | * If this page is still mapped, then its anon_vma cannot have been |
| 484 | * freed. But if it has been unmapped, we have no security against the |
| 485 | * anon_vma structure being freed and reused (for another anon_vma: |
| 486 | * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero() |
| 487 | * above cannot corrupt). |
| 488 | */ |
| 489 | if (!page_mapped(page)) { |
| 490 | rcu_read_unlock(); |
| 491 | put_anon_vma(anon_vma); |
| 492 | return NULL; |
| 493 | } |
| 494 | out: |
| 495 | rcu_read_unlock(); |
| 496 | |
| 497 | return anon_vma; |
| 498 | } |
| 499 | |
| 500 | /* |
| 501 | * Similar to page_get_anon_vma() except it locks the anon_vma. |
| 502 | * |
| 503 | * Its a little more complex as it tries to keep the fast path to a single |
| 504 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a |
| 505 | * reference like with page_get_anon_vma() and then block on the mutex. |
| 506 | */ |
| 507 | struct anon_vma *page_lock_anon_vma_read(struct page *page) |
| 508 | { |
| 509 | struct anon_vma *anon_vma = NULL; |
| 510 | struct anon_vma *root_anon_vma; |
| 511 | unsigned long anon_mapping; |
| 512 | |
| 513 | rcu_read_lock(); |
| 514 | anon_mapping = (unsigned long)READ_ONCE(page->mapping); |
| 515 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
| 516 | goto out; |
| 517 | if (!page_mapped(page)) |
| 518 | goto out; |
| 519 | |
| 520 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
| 521 | root_anon_vma = READ_ONCE(anon_vma->root); |
| 522 | if (down_read_trylock(&root_anon_vma->rwsem)) { |
| 523 | /* |
| 524 | * If the page is still mapped, then this anon_vma is still |
| 525 | * its anon_vma, and holding the mutex ensures that it will |
| 526 | * not go away, see anon_vma_free(). |
| 527 | */ |
| 528 | if (!page_mapped(page)) { |
| 529 | up_read(&root_anon_vma->rwsem); |
| 530 | anon_vma = NULL; |
| 531 | } |
| 532 | goto out; |
| 533 | } |
| 534 | |
| 535 | /* trylock failed, we got to sleep */ |
| 536 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
| 537 | anon_vma = NULL; |
| 538 | goto out; |
| 539 | } |
| 540 | |
| 541 | if (!page_mapped(page)) { |
| 542 | rcu_read_unlock(); |
| 543 | put_anon_vma(anon_vma); |
| 544 | return NULL; |
| 545 | } |
| 546 | |
| 547 | /* we pinned the anon_vma, its safe to sleep */ |
| 548 | rcu_read_unlock(); |
| 549 | anon_vma_lock_read(anon_vma); |
| 550 | |
| 551 | if (atomic_dec_and_test(&anon_vma->refcount)) { |
| 552 | /* |
| 553 | * Oops, we held the last refcount, release the lock |
| 554 | * and bail -- can't simply use put_anon_vma() because |
| 555 | * we'll deadlock on the anon_vma_lock_write() recursion. |
| 556 | */ |
| 557 | anon_vma_unlock_read(anon_vma); |
| 558 | __put_anon_vma(anon_vma); |
| 559 | anon_vma = NULL; |
| 560 | } |
| 561 | |
| 562 | return anon_vma; |
| 563 | |
| 564 | out: |
| 565 | rcu_read_unlock(); |
| 566 | return anon_vma; |
| 567 | } |
| 568 | |
| 569 | void page_unlock_anon_vma_read(struct anon_vma *anon_vma) |
| 570 | { |
| 571 | anon_vma_unlock_read(anon_vma); |
| 572 | } |
| 573 | |
| 574 | #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH |
| 575 | /* |
| 576 | * Flush TLB entries for recently unmapped pages from remote CPUs. It is |
| 577 | * important if a PTE was dirty when it was unmapped that it's flushed |
| 578 | * before any IO is initiated on the page to prevent lost writes. Similarly, |
| 579 | * it must be flushed before freeing to prevent data leakage. |
| 580 | */ |
| 581 | void try_to_unmap_flush(void) |
| 582 | { |
| 583 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; |
| 584 | |
| 585 | if (!tlb_ubc->flush_required) |
| 586 | return; |
| 587 | |
| 588 | arch_tlbbatch_flush(&tlb_ubc->arch); |
| 589 | tlb_ubc->flush_required = false; |
| 590 | tlb_ubc->writable = false; |
| 591 | } |
| 592 | |
| 593 | /* Flush iff there are potentially writable TLB entries that can race with IO */ |
| 594 | void try_to_unmap_flush_dirty(void) |
| 595 | { |
| 596 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; |
| 597 | |
| 598 | if (tlb_ubc->writable) |
| 599 | try_to_unmap_flush(); |
| 600 | } |
| 601 | |
| 602 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) |
| 603 | { |
| 604 | struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; |
| 605 | |
| 606 | arch_tlbbatch_add_mm(&tlb_ubc->arch, mm); |
| 607 | tlb_ubc->flush_required = true; |
| 608 | |
| 609 | /* |
| 610 | * Ensure compiler does not re-order the setting of tlb_flush_batched |
| 611 | * before the PTE is cleared. |
| 612 | */ |
| 613 | barrier(); |
| 614 | mm->tlb_flush_batched = true; |
| 615 | |
| 616 | /* |
| 617 | * If the PTE was dirty then it's best to assume it's writable. The |
| 618 | * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() |
| 619 | * before the page is queued for IO. |
| 620 | */ |
| 621 | if (writable) |
| 622 | tlb_ubc->writable = true; |
| 623 | } |
| 624 | |
| 625 | /* |
| 626 | * Returns true if the TLB flush should be deferred to the end of a batch of |
| 627 | * unmap operations to reduce IPIs. |
| 628 | */ |
| 629 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) |
| 630 | { |
| 631 | bool should_defer = false; |
| 632 | |
| 633 | if (!(flags & TTU_BATCH_FLUSH)) |
| 634 | return false; |
| 635 | |
| 636 | /* If remote CPUs need to be flushed then defer batch the flush */ |
| 637 | if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids) |
| 638 | should_defer = true; |
| 639 | put_cpu(); |
| 640 | |
| 641 | return should_defer; |
| 642 | } |
| 643 | |
| 644 | /* |
| 645 | * Reclaim unmaps pages under the PTL but do not flush the TLB prior to |
| 646 | * releasing the PTL if TLB flushes are batched. It's possible for a parallel |
| 647 | * operation such as mprotect or munmap to race between reclaim unmapping |
| 648 | * the page and flushing the page. If this race occurs, it potentially allows |
| 649 | * access to data via a stale TLB entry. Tracking all mm's that have TLB |
| 650 | * batching in flight would be expensive during reclaim so instead track |
| 651 | * whether TLB batching occurred in the past and if so then do a flush here |
| 652 | * if required. This will cost one additional flush per reclaim cycle paid |
| 653 | * by the first operation at risk such as mprotect and mumap. |
| 654 | * |
| 655 | * This must be called under the PTL so that an access to tlb_flush_batched |
| 656 | * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise |
| 657 | * via the PTL. |
| 658 | */ |
| 659 | void flush_tlb_batched_pending(struct mm_struct *mm) |
| 660 | { |
| 661 | if (mm->tlb_flush_batched) { |
| 662 | flush_tlb_mm(mm); |
| 663 | |
| 664 | /* |
| 665 | * Do not allow the compiler to re-order the clearing of |
| 666 | * tlb_flush_batched before the tlb is flushed. |
| 667 | */ |
| 668 | barrier(); |
| 669 | mm->tlb_flush_batched = false; |
| 670 | } |
| 671 | } |
| 672 | #else |
| 673 | static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) |
| 674 | { |
| 675 | } |
| 676 | |
| 677 | static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) |
| 678 | { |
| 679 | return false; |
| 680 | } |
| 681 | #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ |
| 682 | |
| 683 | /* |
| 684 | * At what user virtual address is page expected in vma? |
| 685 | * Caller should check the page is actually part of the vma. |
| 686 | */ |
| 687 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) |
| 688 | { |
| 689 | unsigned long address; |
| 690 | if (PageAnon(page)) { |
| 691 | struct anon_vma *page__anon_vma = page_anon_vma(page); |
| 692 | /* |
| 693 | * Note: swapoff's unuse_vma() is more efficient with this |
| 694 | * check, and needs it to match anon_vma when KSM is active. |
| 695 | */ |
| 696 | if (!vma->anon_vma || !page__anon_vma || |
| 697 | vma->anon_vma->root != page__anon_vma->root) |
| 698 | return -EFAULT; |
| 699 | } else if (page->mapping) { |
| 700 | if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping) |
| 701 | return -EFAULT; |
| 702 | } else |
| 703 | return -EFAULT; |
| 704 | address = __vma_address(page, vma); |
| 705 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) |
| 706 | return -EFAULT; |
| 707 | return address; |
| 708 | } |
| 709 | |
| 710 | pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) |
| 711 | { |
| 712 | pgd_t *pgd; |
| 713 | p4d_t *p4d; |
| 714 | pud_t *pud; |
| 715 | pmd_t *pmd = NULL; |
| 716 | pmd_t pmde; |
| 717 | |
| 718 | pgd = pgd_offset(mm, address); |
| 719 | if (!pgd_present(*pgd)) |
| 720 | goto out; |
| 721 | |
| 722 | p4d = p4d_offset(pgd, address); |
| 723 | if (!p4d_present(*p4d)) |
| 724 | goto out; |
| 725 | |
| 726 | pud = pud_offset(p4d, address); |
| 727 | if (!pud_present(*pud)) |
| 728 | goto out; |
| 729 | |
| 730 | pmd = pmd_offset(pud, address); |
| 731 | /* |
| 732 | * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() |
| 733 | * without holding anon_vma lock for write. So when looking for a |
| 734 | * genuine pmde (in which to find pte), test present and !THP together. |
| 735 | */ |
| 736 | pmde = *pmd; |
| 737 | barrier(); |
| 738 | if (!pmd_present(pmde) || pmd_trans_huge(pmde)) |
| 739 | pmd = NULL; |
| 740 | out: |
| 741 | return pmd; |
| 742 | } |
| 743 | |
| 744 | struct page_referenced_arg { |
| 745 | int mapcount; |
| 746 | int referenced; |
| 747 | unsigned long vm_flags; |
| 748 | struct mem_cgroup *memcg; |
| 749 | }; |
| 750 | /* |
| 751 | * arg: page_referenced_arg will be passed |
| 752 | */ |
| 753 | static bool page_referenced_one(struct page *page, struct vm_area_struct *vma, |
| 754 | unsigned long address, void *arg) |
| 755 | { |
| 756 | struct page_referenced_arg *pra = arg; |
| 757 | struct page_vma_mapped_walk pvmw = { |
| 758 | .page = page, |
| 759 | .vma = vma, |
| 760 | .address = address, |
| 761 | }; |
| 762 | int referenced = 0; |
| 763 | |
| 764 | while (page_vma_mapped_walk(&pvmw)) { |
| 765 | address = pvmw.address; |
| 766 | |
| 767 | if (vma->vm_flags & VM_LOCKED) { |
| 768 | page_vma_mapped_walk_done(&pvmw); |
| 769 | pra->vm_flags |= VM_LOCKED; |
| 770 | return false; /* To break the loop */ |
| 771 | } |
| 772 | |
| 773 | if (pvmw.pte) { |
| 774 | if (ptep_clear_flush_young_notify(vma, address, |
| 775 | pvmw.pte)) { |
| 776 | /* |
| 777 | * Don't treat a reference through |
| 778 | * a sequentially read mapping as such. |
| 779 | * If the page has been used in another mapping, |
| 780 | * we will catch it; if this other mapping is |
| 781 | * already gone, the unmap path will have set |
| 782 | * PG_referenced or activated the page. |
| 783 | */ |
| 784 | if (likely(!(vma->vm_flags & VM_SEQ_READ))) |
| 785 | referenced++; |
| 786 | } |
| 787 | } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { |
| 788 | if (pmdp_clear_flush_young_notify(vma, address, |
| 789 | pvmw.pmd)) |
| 790 | referenced++; |
| 791 | } else { |
| 792 | /* unexpected pmd-mapped page? */ |
| 793 | WARN_ON_ONCE(1); |
| 794 | } |
| 795 | |
| 796 | pra->mapcount--; |
| 797 | } |
| 798 | |
| 799 | if (referenced) |
| 800 | clear_page_idle(page); |
| 801 | if (test_and_clear_page_young(page)) |
| 802 | referenced++; |
| 803 | |
| 804 | if (referenced) { |
| 805 | pra->referenced++; |
| 806 | pra->vm_flags |= vma->vm_flags; |
| 807 | } |
| 808 | |
| 809 | if (!pra->mapcount) |
| 810 | return false; /* To break the loop */ |
| 811 | |
| 812 | return true; |
| 813 | } |
| 814 | |
| 815 | static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg) |
| 816 | { |
| 817 | struct page_referenced_arg *pra = arg; |
| 818 | struct mem_cgroup *memcg = pra->memcg; |
| 819 | |
| 820 | if (!mm_match_cgroup(vma->vm_mm, memcg)) |
| 821 | return true; |
| 822 | |
| 823 | return false; |
| 824 | } |
| 825 | |
| 826 | /** |
| 827 | * page_referenced - test if the page was referenced |
| 828 | * @page: the page to test |
| 829 | * @is_locked: caller holds lock on the page |
| 830 | * @memcg: target memory cgroup |
| 831 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
| 832 | * |
| 833 | * Quick test_and_clear_referenced for all mappings to a page, |
| 834 | * returns the number of ptes which referenced the page. |
| 835 | */ |
| 836 | int page_referenced(struct page *page, |
| 837 | int is_locked, |
| 838 | struct mem_cgroup *memcg, |
| 839 | unsigned long *vm_flags) |
| 840 | { |
| 841 | int we_locked = 0; |
| 842 | struct page_referenced_arg pra = { |
| 843 | .mapcount = total_mapcount(page), |
| 844 | .memcg = memcg, |
| 845 | }; |
| 846 | struct rmap_walk_control rwc = { |
| 847 | .rmap_one = page_referenced_one, |
| 848 | .arg = (void *)&pra, |
| 849 | .anon_lock = page_lock_anon_vma_read, |
| 850 | }; |
| 851 | |
| 852 | *vm_flags = 0; |
| 853 | if (!pra.mapcount) |
| 854 | return 0; |
| 855 | |
| 856 | if (!page_rmapping(page)) |
| 857 | return 0; |
| 858 | |
| 859 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { |
| 860 | we_locked = trylock_page(page); |
| 861 | if (!we_locked) |
| 862 | return 1; |
| 863 | } |
| 864 | |
| 865 | /* |
| 866 | * If we are reclaiming on behalf of a cgroup, skip |
| 867 | * counting on behalf of references from different |
| 868 | * cgroups |
| 869 | */ |
| 870 | if (memcg) { |
| 871 | rwc.invalid_vma = invalid_page_referenced_vma; |
| 872 | } |
| 873 | |
| 874 | rmap_walk(page, &rwc); |
| 875 | *vm_flags = pra.vm_flags; |
| 876 | |
| 877 | if (we_locked) |
| 878 | unlock_page(page); |
| 879 | |
| 880 | return pra.referenced; |
| 881 | } |
| 882 | |
| 883 | static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma, |
| 884 | unsigned long address, void *arg) |
| 885 | { |
| 886 | struct page_vma_mapped_walk pvmw = { |
| 887 | .page = page, |
| 888 | .vma = vma, |
| 889 | .address = address, |
| 890 | .flags = PVMW_SYNC, |
| 891 | }; |
| 892 | struct mmu_notifier_range range; |
| 893 | int *cleaned = arg; |
| 894 | |
| 895 | /* |
| 896 | * We have to assume the worse case ie pmd for invalidation. Note that |
| 897 | * the page can not be free from this function. |
| 898 | */ |
| 899 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, |
| 900 | 0, vma, vma->vm_mm, address, |
| 901 | min(vma->vm_end, address + |
| 902 | (PAGE_SIZE << compound_order(page)))); |
| 903 | mmu_notifier_invalidate_range_start(&range); |
| 904 | |
| 905 | while (page_vma_mapped_walk(&pvmw)) { |
| 906 | unsigned long cstart; |
| 907 | int ret = 0; |
| 908 | |
| 909 | cstart = address = pvmw.address; |
| 910 | if (pvmw.pte) { |
| 911 | pte_t entry; |
| 912 | pte_t *pte = pvmw.pte; |
| 913 | |
| 914 | if (!pte_dirty(*pte) && !pte_write(*pte)) |
| 915 | continue; |
| 916 | |
| 917 | flush_cache_page(vma, address, pte_pfn(*pte)); |
| 918 | entry = ptep_clear_flush(vma, address, pte); |
| 919 | entry = pte_wrprotect(entry); |
| 920 | entry = pte_mkclean(entry); |
| 921 | set_pte_at(vma->vm_mm, address, pte, entry); |
| 922 | ret = 1; |
| 923 | } else { |
| 924 | #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE |
| 925 | pmd_t *pmd = pvmw.pmd; |
| 926 | pmd_t entry; |
| 927 | |
| 928 | if (!pmd_dirty(*pmd) && !pmd_write(*pmd)) |
| 929 | continue; |
| 930 | |
| 931 | flush_cache_page(vma, address, page_to_pfn(page)); |
| 932 | entry = pmdp_invalidate(vma, address, pmd); |
| 933 | entry = pmd_wrprotect(entry); |
| 934 | entry = pmd_mkclean(entry); |
| 935 | set_pmd_at(vma->vm_mm, address, pmd, entry); |
| 936 | cstart &= PMD_MASK; |
| 937 | ret = 1; |
| 938 | #else |
| 939 | /* unexpected pmd-mapped page? */ |
| 940 | WARN_ON_ONCE(1); |
| 941 | #endif |
| 942 | } |
| 943 | |
| 944 | /* |
| 945 | * No need to call mmu_notifier_invalidate_range() as we are |
| 946 | * downgrading page table protection not changing it to point |
| 947 | * to a new page. |
| 948 | * |
| 949 | * See Documentation/vm/mmu_notifier.rst |
| 950 | */ |
| 951 | if (ret) |
| 952 | (*cleaned)++; |
| 953 | } |
| 954 | |
| 955 | mmu_notifier_invalidate_range_end(&range); |
| 956 | |
| 957 | return true; |
| 958 | } |
| 959 | |
| 960 | static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) |
| 961 | { |
| 962 | if (vma->vm_flags & VM_SHARED) |
| 963 | return false; |
| 964 | |
| 965 | return true; |
| 966 | } |
| 967 | |
| 968 | int page_mkclean(struct page *page) |
| 969 | { |
| 970 | int cleaned = 0; |
| 971 | struct address_space *mapping; |
| 972 | struct rmap_walk_control rwc = { |
| 973 | .arg = (void *)&cleaned, |
| 974 | .rmap_one = page_mkclean_one, |
| 975 | .invalid_vma = invalid_mkclean_vma, |
| 976 | }; |
| 977 | |
| 978 | BUG_ON(!PageLocked(page)); |
| 979 | |
| 980 | if (!page_mapped(page)) |
| 981 | return 0; |
| 982 | |
| 983 | mapping = page_mapping(page); |
| 984 | if (!mapping) |
| 985 | return 0; |
| 986 | |
| 987 | rmap_walk(page, &rwc); |
| 988 | |
| 989 | return cleaned; |
| 990 | } |
| 991 | EXPORT_SYMBOL_GPL(page_mkclean); |
| 992 | |
| 993 | /** |
| 994 | * page_move_anon_rmap - move a page to our anon_vma |
| 995 | * @page: the page to move to our anon_vma |
| 996 | * @vma: the vma the page belongs to |
| 997 | * |
| 998 | * When a page belongs exclusively to one process after a COW event, |
| 999 | * that page can be moved into the anon_vma that belongs to just that |
| 1000 | * process, so the rmap code will not search the parent or sibling |
| 1001 | * processes. |
| 1002 | */ |
| 1003 | void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma) |
| 1004 | { |
| 1005 | struct anon_vma *anon_vma = vma->anon_vma; |
| 1006 | |
| 1007 | page = compound_head(page); |
| 1008 | |
| 1009 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
| 1010 | VM_BUG_ON_VMA(!anon_vma, vma); |
| 1011 | |
| 1012 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
| 1013 | /* |
| 1014 | * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written |
| 1015 | * simultaneously, so a concurrent reader (eg page_referenced()'s |
| 1016 | * PageAnon()) will not see one without the other. |
| 1017 | */ |
| 1018 | WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); |
| 1019 | } |
| 1020 | |
| 1021 | /** |
| 1022 | * __page_set_anon_rmap - set up new anonymous rmap |
| 1023 | * @page: Page or Hugepage to add to rmap |
| 1024 | * @vma: VM area to add page to. |
| 1025 | * @address: User virtual address of the mapping |
| 1026 | * @exclusive: the page is exclusively owned by the current process |
| 1027 | */ |
| 1028 | static void __page_set_anon_rmap(struct page *page, |
| 1029 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
| 1030 | { |
| 1031 | struct anon_vma *anon_vma = vma->anon_vma; |
| 1032 | |
| 1033 | BUG_ON(!anon_vma); |
| 1034 | |
| 1035 | if (PageAnon(page)) |
| 1036 | return; |
| 1037 | |
| 1038 | /* |
| 1039 | * If the page isn't exclusively mapped into this vma, |
| 1040 | * we must use the _oldest_ possible anon_vma for the |
| 1041 | * page mapping! |
| 1042 | */ |
| 1043 | if (!exclusive) |
| 1044 | anon_vma = anon_vma->root; |
| 1045 | |
| 1046 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
| 1047 | page->mapping = (struct address_space *) anon_vma; |
| 1048 | page->index = linear_page_index(vma, address); |
| 1049 | } |
| 1050 | |
| 1051 | /** |
| 1052 | * __page_check_anon_rmap - sanity check anonymous rmap addition |
| 1053 | * @page: the page to add the mapping to |
| 1054 | * @vma: the vm area in which the mapping is added |
| 1055 | * @address: the user virtual address mapped |
| 1056 | */ |
| 1057 | static void __page_check_anon_rmap(struct page *page, |
| 1058 | struct vm_area_struct *vma, unsigned long address) |
| 1059 | { |
| 1060 | #ifdef CONFIG_DEBUG_VM |
| 1061 | /* |
| 1062 | * The page's anon-rmap details (mapping and index) are guaranteed to |
| 1063 | * be set up correctly at this point. |
| 1064 | * |
| 1065 | * We have exclusion against page_add_anon_rmap because the caller |
| 1066 | * always holds the page locked, except if called from page_dup_rmap, |
| 1067 | * in which case the page is already known to be setup. |
| 1068 | * |
| 1069 | * We have exclusion against page_add_new_anon_rmap because those pages |
| 1070 | * are initially only visible via the pagetables, and the pte is locked |
| 1071 | * over the call to page_add_new_anon_rmap. |
| 1072 | */ |
| 1073 | BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); |
| 1074 | BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address)); |
| 1075 | #endif |
| 1076 | } |
| 1077 | |
| 1078 | /** |
| 1079 | * page_add_anon_rmap - add pte mapping to an anonymous page |
| 1080 | * @page: the page to add the mapping to |
| 1081 | * @vma: the vm area in which the mapping is added |
| 1082 | * @address: the user virtual address mapped |
| 1083 | * @compound: charge the page as compound or small page |
| 1084 | * |
| 1085 | * The caller needs to hold the pte lock, and the page must be locked in |
| 1086 | * the anon_vma case: to serialize mapping,index checking after setting, |
| 1087 | * and to ensure that PageAnon is not being upgraded racily to PageKsm |
| 1088 | * (but PageKsm is never downgraded to PageAnon). |
| 1089 | */ |
| 1090 | void page_add_anon_rmap(struct page *page, |
| 1091 | struct vm_area_struct *vma, unsigned long address, bool compound) |
| 1092 | { |
| 1093 | do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0); |
| 1094 | } |
| 1095 | |
| 1096 | /* |
| 1097 | * Special version of the above for do_swap_page, which often runs |
| 1098 | * into pages that are exclusively owned by the current process. |
| 1099 | * Everybody else should continue to use page_add_anon_rmap above. |
| 1100 | */ |
| 1101 | void do_page_add_anon_rmap(struct page *page, |
| 1102 | struct vm_area_struct *vma, unsigned long address, int flags) |
| 1103 | { |
| 1104 | bool compound = flags & RMAP_COMPOUND; |
| 1105 | bool first; |
| 1106 | |
| 1107 | if (compound) { |
| 1108 | atomic_t *mapcount; |
| 1109 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
| 1110 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); |
| 1111 | mapcount = compound_mapcount_ptr(page); |
| 1112 | first = atomic_inc_and_test(mapcount); |
| 1113 | } else { |
| 1114 | first = atomic_inc_and_test(&page->_mapcount); |
| 1115 | } |
| 1116 | |
| 1117 | if (first) { |
| 1118 | int nr = compound ? hpage_nr_pages(page) : 1; |
| 1119 | /* |
| 1120 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because |
| 1121 | * these counters are not modified in interrupt context, and |
| 1122 | * pte lock(a spinlock) is held, which implies preemption |
| 1123 | * disabled. |
| 1124 | */ |
| 1125 | if (compound) |
| 1126 | __inc_node_page_state(page, NR_ANON_THPS); |
| 1127 | __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); |
| 1128 | } |
| 1129 | if (unlikely(PageKsm(page))) |
| 1130 | return; |
| 1131 | |
| 1132 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
| 1133 | |
| 1134 | /* address might be in next vma when migration races vma_adjust */ |
| 1135 | if (first) |
| 1136 | __page_set_anon_rmap(page, vma, address, |
| 1137 | flags & RMAP_EXCLUSIVE); |
| 1138 | else |
| 1139 | __page_check_anon_rmap(page, vma, address); |
| 1140 | } |
| 1141 | |
| 1142 | /** |
| 1143 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page |
| 1144 | * @page: the page to add the mapping to |
| 1145 | * @vma: the vm area in which the mapping is added |
| 1146 | * @address: the user virtual address mapped |
| 1147 | * @compound: charge the page as compound or small page |
| 1148 | * |
| 1149 | * Same as page_add_anon_rmap but must only be called on *new* pages. |
| 1150 | * This means the inc-and-test can be bypassed. |
| 1151 | * Page does not have to be locked. |
| 1152 | */ |
| 1153 | void page_add_new_anon_rmap(struct page *page, |
| 1154 | struct vm_area_struct *vma, unsigned long address, bool compound) |
| 1155 | { |
| 1156 | int nr = compound ? hpage_nr_pages(page) : 1; |
| 1157 | |
| 1158 | VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); |
| 1159 | __SetPageSwapBacked(page); |
| 1160 | if (compound) { |
| 1161 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); |
| 1162 | /* increment count (starts at -1) */ |
| 1163 | atomic_set(compound_mapcount_ptr(page), 0); |
| 1164 | __inc_node_page_state(page, NR_ANON_THPS); |
| 1165 | } else { |
| 1166 | /* Anon THP always mapped first with PMD */ |
| 1167 | VM_BUG_ON_PAGE(PageTransCompound(page), page); |
| 1168 | /* increment count (starts at -1) */ |
| 1169 | atomic_set(&page->_mapcount, 0); |
| 1170 | } |
| 1171 | __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); |
| 1172 | __page_set_anon_rmap(page, vma, address, 1); |
| 1173 | } |
| 1174 | |
| 1175 | /** |
| 1176 | * page_add_file_rmap - add pte mapping to a file page |
| 1177 | * @page: the page to add the mapping to |
| 1178 | * @compound: charge the page as compound or small page |
| 1179 | * |
| 1180 | * The caller needs to hold the pte lock. |
| 1181 | */ |
| 1182 | void page_add_file_rmap(struct page *page, bool compound) |
| 1183 | { |
| 1184 | int i, nr = 1; |
| 1185 | |
| 1186 | VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page); |
| 1187 | lock_page_memcg(page); |
| 1188 | if (compound && PageTransHuge(page)) { |
| 1189 | for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { |
| 1190 | if (atomic_inc_and_test(&page[i]._mapcount)) |
| 1191 | nr++; |
| 1192 | } |
| 1193 | if (!atomic_inc_and_test(compound_mapcount_ptr(page))) |
| 1194 | goto out; |
| 1195 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); |
| 1196 | __inc_node_page_state(page, NR_SHMEM_PMDMAPPED); |
| 1197 | } else { |
| 1198 | if (PageTransCompound(page) && page_mapping(page)) { |
| 1199 | VM_WARN_ON_ONCE(!PageLocked(page)); |
| 1200 | |
| 1201 | SetPageDoubleMap(compound_head(page)); |
| 1202 | if (PageMlocked(page)) |
| 1203 | clear_page_mlock(compound_head(page)); |
| 1204 | } |
| 1205 | if (!atomic_inc_and_test(&page->_mapcount)) |
| 1206 | goto out; |
| 1207 | } |
| 1208 | __mod_lruvec_page_state(page, NR_FILE_MAPPED, nr); |
| 1209 | out: |
| 1210 | unlock_page_memcg(page); |
| 1211 | } |
| 1212 | |
| 1213 | static void page_remove_file_rmap(struct page *page, bool compound) |
| 1214 | { |
| 1215 | int i, nr = 1; |
| 1216 | |
| 1217 | VM_BUG_ON_PAGE(compound && !PageHead(page), page); |
| 1218 | lock_page_memcg(page); |
| 1219 | |
| 1220 | /* Hugepages are not counted in NR_FILE_MAPPED for now. */ |
| 1221 | if (unlikely(PageHuge(page))) { |
| 1222 | /* hugetlb pages are always mapped with pmds */ |
| 1223 | atomic_dec(compound_mapcount_ptr(page)); |
| 1224 | goto out; |
| 1225 | } |
| 1226 | |
| 1227 | /* page still mapped by someone else? */ |
| 1228 | if (compound && PageTransHuge(page)) { |
| 1229 | for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { |
| 1230 | if (atomic_add_negative(-1, &page[i]._mapcount)) |
| 1231 | nr++; |
| 1232 | } |
| 1233 | if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) |
| 1234 | goto out; |
| 1235 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); |
| 1236 | __dec_node_page_state(page, NR_SHMEM_PMDMAPPED); |
| 1237 | } else { |
| 1238 | if (!atomic_add_negative(-1, &page->_mapcount)) |
| 1239 | goto out; |
| 1240 | } |
| 1241 | |
| 1242 | /* |
| 1243 | * We use the irq-unsafe __{inc|mod}_lruvec_page_state because |
| 1244 | * these counters are not modified in interrupt context, and |
| 1245 | * pte lock(a spinlock) is held, which implies preemption disabled. |
| 1246 | */ |
| 1247 | __mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr); |
| 1248 | |
| 1249 | if (unlikely(PageMlocked(page))) |
| 1250 | clear_page_mlock(page); |
| 1251 | out: |
| 1252 | unlock_page_memcg(page); |
| 1253 | } |
| 1254 | |
| 1255 | static void page_remove_anon_compound_rmap(struct page *page) |
| 1256 | { |
| 1257 | int i, nr; |
| 1258 | |
| 1259 | if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) |
| 1260 | return; |
| 1261 | |
| 1262 | /* Hugepages are not counted in NR_ANON_PAGES for now. */ |
| 1263 | if (unlikely(PageHuge(page))) |
| 1264 | return; |
| 1265 | |
| 1266 | if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) |
| 1267 | return; |
| 1268 | |
| 1269 | __dec_node_page_state(page, NR_ANON_THPS); |
| 1270 | |
| 1271 | if (TestClearPageDoubleMap(page)) { |
| 1272 | /* |
| 1273 | * Subpages can be mapped with PTEs too. Check how many of |
| 1274 | * themi are still mapped. |
| 1275 | */ |
| 1276 | for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { |
| 1277 | if (atomic_add_negative(-1, &page[i]._mapcount)) |
| 1278 | nr++; |
| 1279 | } |
| 1280 | } else { |
| 1281 | nr = HPAGE_PMD_NR; |
| 1282 | } |
| 1283 | |
| 1284 | if (unlikely(PageMlocked(page))) |
| 1285 | clear_page_mlock(page); |
| 1286 | |
| 1287 | if (nr) { |
| 1288 | __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr); |
| 1289 | deferred_split_huge_page(page); |
| 1290 | } |
| 1291 | } |
| 1292 | |
| 1293 | /** |
| 1294 | * page_remove_rmap - take down pte mapping from a page |
| 1295 | * @page: page to remove mapping from |
| 1296 | * @compound: uncharge the page as compound or small page |
| 1297 | * |
| 1298 | * The caller needs to hold the pte lock. |
| 1299 | */ |
| 1300 | void page_remove_rmap(struct page *page, bool compound) |
| 1301 | { |
| 1302 | if (!PageAnon(page)) |
| 1303 | return page_remove_file_rmap(page, compound); |
| 1304 | |
| 1305 | if (compound) |
| 1306 | return page_remove_anon_compound_rmap(page); |
| 1307 | |
| 1308 | /* page still mapped by someone else? */ |
| 1309 | if (!atomic_add_negative(-1, &page->_mapcount)) |
| 1310 | return; |
| 1311 | |
| 1312 | /* |
| 1313 | * We use the irq-unsafe __{inc|mod}_zone_page_stat because |
| 1314 | * these counters are not modified in interrupt context, and |
| 1315 | * pte lock(a spinlock) is held, which implies preemption disabled. |
| 1316 | */ |
| 1317 | __dec_node_page_state(page, NR_ANON_MAPPED); |
| 1318 | |
| 1319 | if (unlikely(PageMlocked(page))) |
| 1320 | clear_page_mlock(page); |
| 1321 | |
| 1322 | if (PageTransCompound(page)) |
| 1323 | deferred_split_huge_page(compound_head(page)); |
| 1324 | |
| 1325 | /* |
| 1326 | * It would be tidy to reset the PageAnon mapping here, |
| 1327 | * but that might overwrite a racing page_add_anon_rmap |
| 1328 | * which increments mapcount after us but sets mapping |
| 1329 | * before us: so leave the reset to free_unref_page, |
| 1330 | * and remember that it's only reliable while mapped. |
| 1331 | * Leaving it set also helps swapoff to reinstate ptes |
| 1332 | * faster for those pages still in swapcache. |
| 1333 | */ |
| 1334 | } |
| 1335 | |
| 1336 | /* |
| 1337 | * @arg: enum ttu_flags will be passed to this argument |
| 1338 | */ |
| 1339 | static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
| 1340 | unsigned long address, void *arg) |
| 1341 | { |
| 1342 | struct mm_struct *mm = vma->vm_mm; |
| 1343 | struct page_vma_mapped_walk pvmw = { |
| 1344 | .page = page, |
| 1345 | .vma = vma, |
| 1346 | .address = address, |
| 1347 | }; |
| 1348 | pte_t pteval; |
| 1349 | struct page *subpage; |
| 1350 | bool ret = true; |
| 1351 | struct mmu_notifier_range range; |
| 1352 | enum ttu_flags flags = (enum ttu_flags)arg; |
| 1353 | |
| 1354 | /* munlock has nothing to gain from examining un-locked vmas */ |
| 1355 | if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED)) |
| 1356 | return true; |
| 1357 | |
| 1358 | if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) && |
| 1359 | is_zone_device_page(page) && !is_device_private_page(page)) |
| 1360 | return true; |
| 1361 | |
| 1362 | if (flags & TTU_SPLIT_HUGE_PMD) { |
| 1363 | split_huge_pmd_address(vma, address, |
| 1364 | flags & TTU_SPLIT_FREEZE, page); |
| 1365 | } |
| 1366 | |
| 1367 | /* |
| 1368 | * For THP, we have to assume the worse case ie pmd for invalidation. |
| 1369 | * For hugetlb, it could be much worse if we need to do pud |
| 1370 | * invalidation in the case of pmd sharing. |
| 1371 | * |
| 1372 | * Note that the page can not be free in this function as call of |
| 1373 | * try_to_unmap() must hold a reference on the page. |
| 1374 | */ |
| 1375 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
| 1376 | address, |
| 1377 | min(vma->vm_end, address + |
| 1378 | (PAGE_SIZE << compound_order(page)))); |
| 1379 | if (PageHuge(page)) { |
| 1380 | /* |
| 1381 | * If sharing is possible, start and end will be adjusted |
| 1382 | * accordingly. |
| 1383 | */ |
| 1384 | adjust_range_if_pmd_sharing_possible(vma, &range.start, |
| 1385 | &range.end); |
| 1386 | } |
| 1387 | mmu_notifier_invalidate_range_start(&range); |
| 1388 | |
| 1389 | while (page_vma_mapped_walk(&pvmw)) { |
| 1390 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
| 1391 | /* PMD-mapped THP migration entry */ |
| 1392 | if (!pvmw.pte && (flags & TTU_MIGRATION)) { |
| 1393 | VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page); |
| 1394 | |
| 1395 | set_pmd_migration_entry(&pvmw, page); |
| 1396 | continue; |
| 1397 | } |
| 1398 | #endif |
| 1399 | |
| 1400 | /* |
| 1401 | * If the page is mlock()d, we cannot swap it out. |
| 1402 | * If it's recently referenced (perhaps page_referenced |
| 1403 | * skipped over this mm) then we should reactivate it. |
| 1404 | */ |
| 1405 | if (!(flags & TTU_IGNORE_MLOCK)) { |
| 1406 | if (vma->vm_flags & VM_LOCKED) { |
| 1407 | /* PTE-mapped THP are never mlocked */ |
| 1408 | if (!PageTransCompound(page)) { |
| 1409 | /* |
| 1410 | * Holding pte lock, we do *not* need |
| 1411 | * mmap_sem here |
| 1412 | */ |
| 1413 | mlock_vma_page(page); |
| 1414 | } |
| 1415 | ret = false; |
| 1416 | page_vma_mapped_walk_done(&pvmw); |
| 1417 | break; |
| 1418 | } |
| 1419 | if (flags & TTU_MUNLOCK) |
| 1420 | continue; |
| 1421 | } |
| 1422 | |
| 1423 | /* Unexpected PMD-mapped THP? */ |
| 1424 | VM_BUG_ON_PAGE(!pvmw.pte, page); |
| 1425 | |
| 1426 | subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte); |
| 1427 | address = pvmw.address; |
| 1428 | |
| 1429 | if (PageHuge(page)) { |
| 1430 | if (huge_pmd_unshare(mm, &address, pvmw.pte)) { |
| 1431 | /* |
| 1432 | * huge_pmd_unshare unmapped an entire PMD |
| 1433 | * page. There is no way of knowing exactly |
| 1434 | * which PMDs may be cached for this mm, so |
| 1435 | * we must flush them all. start/end were |
| 1436 | * already adjusted above to cover this range. |
| 1437 | */ |
| 1438 | flush_cache_range(vma, range.start, range.end); |
| 1439 | flush_tlb_range(vma, range.start, range.end); |
| 1440 | mmu_notifier_invalidate_range(mm, range.start, |
| 1441 | range.end); |
| 1442 | |
| 1443 | /* |
| 1444 | * The ref count of the PMD page was dropped |
| 1445 | * which is part of the way map counting |
| 1446 | * is done for shared PMDs. Return 'true' |
| 1447 | * here. When there is no other sharing, |
| 1448 | * huge_pmd_unshare returns false and we will |
| 1449 | * unmap the actual page and drop map count |
| 1450 | * to zero. |
| 1451 | */ |
| 1452 | page_vma_mapped_walk_done(&pvmw); |
| 1453 | break; |
| 1454 | } |
| 1455 | } |
| 1456 | |
| 1457 | if (IS_ENABLED(CONFIG_MIGRATION) && |
| 1458 | (flags & TTU_MIGRATION) && |
| 1459 | is_zone_device_page(page)) { |
| 1460 | swp_entry_t entry; |
| 1461 | pte_t swp_pte; |
| 1462 | |
| 1463 | pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte); |
| 1464 | |
| 1465 | /* |
| 1466 | * Store the pfn of the page in a special migration |
| 1467 | * pte. do_swap_page() will wait until the migration |
| 1468 | * pte is removed and then restart fault handling. |
| 1469 | */ |
| 1470 | entry = make_migration_entry(page, 0); |
| 1471 | swp_pte = swp_entry_to_pte(entry); |
| 1472 | if (pte_soft_dirty(pteval)) |
| 1473 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
| 1474 | set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte); |
| 1475 | /* |
| 1476 | * No need to invalidate here it will synchronize on |
| 1477 | * against the special swap migration pte. |
| 1478 | */ |
| 1479 | goto discard; |
| 1480 | } |
| 1481 | |
| 1482 | if (!(flags & TTU_IGNORE_ACCESS)) { |
| 1483 | if (ptep_clear_flush_young_notify(vma, address, |
| 1484 | pvmw.pte)) { |
| 1485 | ret = false; |
| 1486 | page_vma_mapped_walk_done(&pvmw); |
| 1487 | break; |
| 1488 | } |
| 1489 | } |
| 1490 | |
| 1491 | /* Nuke the page table entry. */ |
| 1492 | flush_cache_page(vma, address, pte_pfn(*pvmw.pte)); |
| 1493 | if (should_defer_flush(mm, flags)) { |
| 1494 | /* |
| 1495 | * We clear the PTE but do not flush so potentially |
| 1496 | * a remote CPU could still be writing to the page. |
| 1497 | * If the entry was previously clean then the |
| 1498 | * architecture must guarantee that a clear->dirty |
| 1499 | * transition on a cached TLB entry is written through |
| 1500 | * and traps if the PTE is unmapped. |
| 1501 | */ |
| 1502 | pteval = ptep_get_and_clear(mm, address, pvmw.pte); |
| 1503 | |
| 1504 | set_tlb_ubc_flush_pending(mm, pte_dirty(pteval)); |
| 1505 | } else { |
| 1506 | pteval = ptep_clear_flush(vma, address, pvmw.pte); |
| 1507 | } |
| 1508 | |
| 1509 | /* Move the dirty bit to the page. Now the pte is gone. */ |
| 1510 | if (pte_dirty(pteval)) |
| 1511 | set_page_dirty(page); |
| 1512 | |
| 1513 | /* Update high watermark before we lower rss */ |
| 1514 | update_hiwater_rss(mm); |
| 1515 | |
| 1516 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { |
| 1517 | pteval = swp_entry_to_pte(make_hwpoison_entry(subpage)); |
| 1518 | if (PageHuge(page)) { |
| 1519 | int nr = 1 << compound_order(page); |
| 1520 | hugetlb_count_sub(nr, mm); |
| 1521 | set_huge_swap_pte_at(mm, address, |
| 1522 | pvmw.pte, pteval, |
| 1523 | vma_mmu_pagesize(vma)); |
| 1524 | } else { |
| 1525 | dec_mm_counter(mm, mm_counter(page)); |
| 1526 | set_pte_at(mm, address, pvmw.pte, pteval); |
| 1527 | } |
| 1528 | |
| 1529 | } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) { |
| 1530 | /* |
| 1531 | * The guest indicated that the page content is of no |
| 1532 | * interest anymore. Simply discard the pte, vmscan |
| 1533 | * will take care of the rest. |
| 1534 | * A future reference will then fault in a new zero |
| 1535 | * page. When userfaultfd is active, we must not drop |
| 1536 | * this page though, as its main user (postcopy |
| 1537 | * migration) will not expect userfaults on already |
| 1538 | * copied pages. |
| 1539 | */ |
| 1540 | dec_mm_counter(mm, mm_counter(page)); |
| 1541 | /* We have to invalidate as we cleared the pte */ |
| 1542 | mmu_notifier_invalidate_range(mm, address, |
| 1543 | address + PAGE_SIZE); |
| 1544 | } else if (IS_ENABLED(CONFIG_MIGRATION) && |
| 1545 | (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) { |
| 1546 | swp_entry_t entry; |
| 1547 | pte_t swp_pte; |
| 1548 | |
| 1549 | if (arch_unmap_one(mm, vma, address, pteval) < 0) { |
| 1550 | set_pte_at(mm, address, pvmw.pte, pteval); |
| 1551 | ret = false; |
| 1552 | page_vma_mapped_walk_done(&pvmw); |
| 1553 | break; |
| 1554 | } |
| 1555 | |
| 1556 | /* |
| 1557 | * Store the pfn of the page in a special migration |
| 1558 | * pte. do_swap_page() will wait until the migration |
| 1559 | * pte is removed and then restart fault handling. |
| 1560 | */ |
| 1561 | entry = make_migration_entry(subpage, |
| 1562 | pte_write(pteval)); |
| 1563 | swp_pte = swp_entry_to_pte(entry); |
| 1564 | if (pte_soft_dirty(pteval)) |
| 1565 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
| 1566 | set_pte_at(mm, address, pvmw.pte, swp_pte); |
| 1567 | /* |
| 1568 | * No need to invalidate here it will synchronize on |
| 1569 | * against the special swap migration pte. |
| 1570 | */ |
| 1571 | } else if (PageAnon(page)) { |
| 1572 | swp_entry_t entry = { .val = page_private(subpage) }; |
| 1573 | pte_t swp_pte; |
| 1574 | /* |
| 1575 | * Store the swap location in the pte. |
| 1576 | * See handle_pte_fault() ... |
| 1577 | */ |
| 1578 | if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) { |
| 1579 | WARN_ON_ONCE(1); |
| 1580 | ret = false; |
| 1581 | /* We have to invalidate as we cleared the pte */ |
| 1582 | mmu_notifier_invalidate_range(mm, address, |
| 1583 | address + PAGE_SIZE); |
| 1584 | page_vma_mapped_walk_done(&pvmw); |
| 1585 | break; |
| 1586 | } |
| 1587 | |
| 1588 | /* MADV_FREE page check */ |
| 1589 | if (!PageSwapBacked(page)) { |
| 1590 | if (!PageDirty(page)) { |
| 1591 | /* Invalidate as we cleared the pte */ |
| 1592 | mmu_notifier_invalidate_range(mm, |
| 1593 | address, address + PAGE_SIZE); |
| 1594 | dec_mm_counter(mm, MM_ANONPAGES); |
| 1595 | goto discard; |
| 1596 | } |
| 1597 | |
| 1598 | /* |
| 1599 | * If the page was redirtied, it cannot be |
| 1600 | * discarded. Remap the page to page table. |
| 1601 | */ |
| 1602 | set_pte_at(mm, address, pvmw.pte, pteval); |
| 1603 | SetPageSwapBacked(page); |
| 1604 | ret = false; |
| 1605 | page_vma_mapped_walk_done(&pvmw); |
| 1606 | break; |
| 1607 | } |
| 1608 | |
| 1609 | if (swap_duplicate(entry) < 0) { |
| 1610 | set_pte_at(mm, address, pvmw.pte, pteval); |
| 1611 | ret = false; |
| 1612 | page_vma_mapped_walk_done(&pvmw); |
| 1613 | break; |
| 1614 | } |
| 1615 | if (arch_unmap_one(mm, vma, address, pteval) < 0) { |
| 1616 | set_pte_at(mm, address, pvmw.pte, pteval); |
| 1617 | ret = false; |
| 1618 | page_vma_mapped_walk_done(&pvmw); |
| 1619 | break; |
| 1620 | } |
| 1621 | if (list_empty(&mm->mmlist)) { |
| 1622 | spin_lock(&mmlist_lock); |
| 1623 | if (list_empty(&mm->mmlist)) |
| 1624 | list_add(&mm->mmlist, &init_mm.mmlist); |
| 1625 | spin_unlock(&mmlist_lock); |
| 1626 | } |
| 1627 | dec_mm_counter(mm, MM_ANONPAGES); |
| 1628 | inc_mm_counter(mm, MM_SWAPENTS); |
| 1629 | swp_pte = swp_entry_to_pte(entry); |
| 1630 | if (pte_soft_dirty(pteval)) |
| 1631 | swp_pte = pte_swp_mksoft_dirty(swp_pte); |
| 1632 | set_pte_at(mm, address, pvmw.pte, swp_pte); |
| 1633 | /* Invalidate as we cleared the pte */ |
| 1634 | mmu_notifier_invalidate_range(mm, address, |
| 1635 | address + PAGE_SIZE); |
| 1636 | } else { |
| 1637 | /* |
| 1638 | * This is a locked file-backed page, thus it cannot |
| 1639 | * be removed from the page cache and replaced by a new |
| 1640 | * page before mmu_notifier_invalidate_range_end, so no |
| 1641 | * concurrent thread might update its page table to |
| 1642 | * point at new page while a device still is using this |
| 1643 | * page. |
| 1644 | * |
| 1645 | * See Documentation/vm/mmu_notifier.rst |
| 1646 | */ |
| 1647 | dec_mm_counter(mm, mm_counter_file(page)); |
| 1648 | } |
| 1649 | discard: |
| 1650 | /* |
| 1651 | * No need to call mmu_notifier_invalidate_range() it has be |
| 1652 | * done above for all cases requiring it to happen under page |
| 1653 | * table lock before mmu_notifier_invalidate_range_end() |
| 1654 | * |
| 1655 | * See Documentation/vm/mmu_notifier.rst |
| 1656 | */ |
| 1657 | page_remove_rmap(subpage, PageHuge(page)); |
| 1658 | put_page(page); |
| 1659 | } |
| 1660 | |
| 1661 | mmu_notifier_invalidate_range_end(&range); |
| 1662 | |
| 1663 | return ret; |
| 1664 | } |
| 1665 | |
| 1666 | bool is_vma_temporary_stack(struct vm_area_struct *vma) |
| 1667 | { |
| 1668 | int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); |
| 1669 | |
| 1670 | if (!maybe_stack) |
| 1671 | return false; |
| 1672 | |
| 1673 | if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == |
| 1674 | VM_STACK_INCOMPLETE_SETUP) |
| 1675 | return true; |
| 1676 | |
| 1677 | return false; |
| 1678 | } |
| 1679 | |
| 1680 | static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) |
| 1681 | { |
| 1682 | return is_vma_temporary_stack(vma); |
| 1683 | } |
| 1684 | |
| 1685 | static int page_mapcount_is_zero(struct page *page) |
| 1686 | { |
| 1687 | return !total_mapcount(page); |
| 1688 | } |
| 1689 | |
| 1690 | /** |
| 1691 | * try_to_unmap - try to remove all page table mappings to a page |
| 1692 | * @page: the page to get unmapped |
| 1693 | * @flags: action and flags |
| 1694 | * |
| 1695 | * Tries to remove all the page table entries which are mapping this |
| 1696 | * page, used in the pageout path. Caller must hold the page lock. |
| 1697 | * |
| 1698 | * If unmap is successful, return true. Otherwise, false. |
| 1699 | */ |
| 1700 | bool try_to_unmap(struct page *page, enum ttu_flags flags) |
| 1701 | { |
| 1702 | struct rmap_walk_control rwc = { |
| 1703 | .rmap_one = try_to_unmap_one, |
| 1704 | .arg = (void *)flags, |
| 1705 | .done = page_mapcount_is_zero, |
| 1706 | .anon_lock = page_lock_anon_vma_read, |
| 1707 | }; |
| 1708 | |
| 1709 | /* |
| 1710 | * During exec, a temporary VMA is setup and later moved. |
| 1711 | * The VMA is moved under the anon_vma lock but not the |
| 1712 | * page tables leading to a race where migration cannot |
| 1713 | * find the migration ptes. Rather than increasing the |
| 1714 | * locking requirements of exec(), migration skips |
| 1715 | * temporary VMAs until after exec() completes. |
| 1716 | */ |
| 1717 | if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE)) |
| 1718 | && !PageKsm(page) && PageAnon(page)) |
| 1719 | rwc.invalid_vma = invalid_migration_vma; |
| 1720 | |
| 1721 | if (flags & TTU_RMAP_LOCKED) |
| 1722 | rmap_walk_locked(page, &rwc); |
| 1723 | else |
| 1724 | rmap_walk(page, &rwc); |
| 1725 | |
| 1726 | return !page_mapcount(page) ? true : false; |
| 1727 | } |
| 1728 | |
| 1729 | static int page_not_mapped(struct page *page) |
| 1730 | { |
| 1731 | return !page_mapped(page); |
| 1732 | }; |
| 1733 | |
| 1734 | /** |
| 1735 | * try_to_munlock - try to munlock a page |
| 1736 | * @page: the page to be munlocked |
| 1737 | * |
| 1738 | * Called from munlock code. Checks all of the VMAs mapping the page |
| 1739 | * to make sure nobody else has this page mlocked. The page will be |
| 1740 | * returned with PG_mlocked cleared if no other vmas have it mlocked. |
| 1741 | */ |
| 1742 | |
| 1743 | void try_to_munlock(struct page *page) |
| 1744 | { |
| 1745 | struct rmap_walk_control rwc = { |
| 1746 | .rmap_one = try_to_unmap_one, |
| 1747 | .arg = (void *)TTU_MUNLOCK, |
| 1748 | .done = page_not_mapped, |
| 1749 | .anon_lock = page_lock_anon_vma_read, |
| 1750 | |
| 1751 | }; |
| 1752 | |
| 1753 | VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page); |
| 1754 | VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page); |
| 1755 | |
| 1756 | rmap_walk(page, &rwc); |
| 1757 | } |
| 1758 | |
| 1759 | void __put_anon_vma(struct anon_vma *anon_vma) |
| 1760 | { |
| 1761 | struct anon_vma *root = anon_vma->root; |
| 1762 | |
| 1763 | anon_vma_free(anon_vma); |
| 1764 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) |
| 1765 | anon_vma_free(root); |
| 1766 | } |
| 1767 | |
| 1768 | static struct anon_vma *rmap_walk_anon_lock(struct page *page, |
| 1769 | struct rmap_walk_control *rwc) |
| 1770 | { |
| 1771 | struct anon_vma *anon_vma; |
| 1772 | |
| 1773 | if (rwc->anon_lock) |
| 1774 | return rwc->anon_lock(page); |
| 1775 | |
| 1776 | /* |
| 1777 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() |
| 1778 | * because that depends on page_mapped(); but not all its usages |
| 1779 | * are holding mmap_sem. Users without mmap_sem are required to |
| 1780 | * take a reference count to prevent the anon_vma disappearing |
| 1781 | */ |
| 1782 | anon_vma = page_anon_vma(page); |
| 1783 | if (!anon_vma) |
| 1784 | return NULL; |
| 1785 | |
| 1786 | anon_vma_lock_read(anon_vma); |
| 1787 | return anon_vma; |
| 1788 | } |
| 1789 | |
| 1790 | /* |
| 1791 | * rmap_walk_anon - do something to anonymous page using the object-based |
| 1792 | * rmap method |
| 1793 | * @page: the page to be handled |
| 1794 | * @rwc: control variable according to each walk type |
| 1795 | * |
| 1796 | * Find all the mappings of a page using the mapping pointer and the vma chains |
| 1797 | * contained in the anon_vma struct it points to. |
| 1798 | * |
| 1799 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma |
| 1800 | * where the page was found will be held for write. So, we won't recheck |
| 1801 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be |
| 1802 | * LOCKED. |
| 1803 | */ |
| 1804 | static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc, |
| 1805 | bool locked) |
| 1806 | { |
| 1807 | struct anon_vma *anon_vma; |
| 1808 | pgoff_t pgoff_start, pgoff_end; |
| 1809 | struct anon_vma_chain *avc; |
| 1810 | |
| 1811 | if (locked) { |
| 1812 | anon_vma = page_anon_vma(page); |
| 1813 | /* anon_vma disappear under us? */ |
| 1814 | VM_BUG_ON_PAGE(!anon_vma, page); |
| 1815 | } else { |
| 1816 | anon_vma = rmap_walk_anon_lock(page, rwc); |
| 1817 | } |
| 1818 | if (!anon_vma) |
| 1819 | return; |
| 1820 | |
| 1821 | pgoff_start = page_to_pgoff(page); |
| 1822 | pgoff_end = pgoff_start + hpage_nr_pages(page) - 1; |
| 1823 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, |
| 1824 | pgoff_start, pgoff_end) { |
| 1825 | struct vm_area_struct *vma = avc->vma; |
| 1826 | unsigned long address = vma_address(page, vma); |
| 1827 | |
| 1828 | cond_resched(); |
| 1829 | |
| 1830 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
| 1831 | continue; |
| 1832 | |
| 1833 | if (!rwc->rmap_one(page, vma, address, rwc->arg)) |
| 1834 | break; |
| 1835 | if (rwc->done && rwc->done(page)) |
| 1836 | break; |
| 1837 | } |
| 1838 | |
| 1839 | if (!locked) |
| 1840 | anon_vma_unlock_read(anon_vma); |
| 1841 | } |
| 1842 | |
| 1843 | /* |
| 1844 | * rmap_walk_file - do something to file page using the object-based rmap method |
| 1845 | * @page: the page to be handled |
| 1846 | * @rwc: control variable according to each walk type |
| 1847 | * |
| 1848 | * Find all the mappings of a page using the mapping pointer and the vma chains |
| 1849 | * contained in the address_space struct it points to. |
| 1850 | * |
| 1851 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma |
| 1852 | * where the page was found will be held for write. So, we won't recheck |
| 1853 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be |
| 1854 | * LOCKED. |
| 1855 | */ |
| 1856 | static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc, |
| 1857 | bool locked) |
| 1858 | { |
| 1859 | struct address_space *mapping = page_mapping(page); |
| 1860 | pgoff_t pgoff_start, pgoff_end; |
| 1861 | struct vm_area_struct *vma; |
| 1862 | |
| 1863 | /* |
| 1864 | * The page lock not only makes sure that page->mapping cannot |
| 1865 | * suddenly be NULLified by truncation, it makes sure that the |
| 1866 | * structure at mapping cannot be freed and reused yet, |
| 1867 | * so we can safely take mapping->i_mmap_rwsem. |
| 1868 | */ |
| 1869 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
| 1870 | |
| 1871 | if (!mapping) |
| 1872 | return; |
| 1873 | |
| 1874 | pgoff_start = page_to_pgoff(page); |
| 1875 | pgoff_end = pgoff_start + hpage_nr_pages(page) - 1; |
| 1876 | if (!locked) |
| 1877 | i_mmap_lock_read(mapping); |
| 1878 | vma_interval_tree_foreach(vma, &mapping->i_mmap, |
| 1879 | pgoff_start, pgoff_end) { |
| 1880 | unsigned long address = vma_address(page, vma); |
| 1881 | |
| 1882 | cond_resched(); |
| 1883 | |
| 1884 | if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) |
| 1885 | continue; |
| 1886 | |
| 1887 | if (!rwc->rmap_one(page, vma, address, rwc->arg)) |
| 1888 | goto done; |
| 1889 | if (rwc->done && rwc->done(page)) |
| 1890 | goto done; |
| 1891 | } |
| 1892 | |
| 1893 | done: |
| 1894 | if (!locked) |
| 1895 | i_mmap_unlock_read(mapping); |
| 1896 | } |
| 1897 | |
| 1898 | void rmap_walk(struct page *page, struct rmap_walk_control *rwc) |
| 1899 | { |
| 1900 | if (unlikely(PageKsm(page))) |
| 1901 | rmap_walk_ksm(page, rwc); |
| 1902 | else if (PageAnon(page)) |
| 1903 | rmap_walk_anon(page, rwc, false); |
| 1904 | else |
| 1905 | rmap_walk_file(page, rwc, false); |
| 1906 | } |
| 1907 | |
| 1908 | /* Like rmap_walk, but caller holds relevant rmap lock */ |
| 1909 | void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc) |
| 1910 | { |
| 1911 | /* no ksm support for now */ |
| 1912 | VM_BUG_ON_PAGE(PageKsm(page), page); |
| 1913 | if (PageAnon(page)) |
| 1914 | rmap_walk_anon(page, rwc, true); |
| 1915 | else |
| 1916 | rmap_walk_file(page, rwc, true); |
| 1917 | } |
| 1918 | |
| 1919 | #ifdef CONFIG_HUGETLB_PAGE |
| 1920 | /* |
| 1921 | * The following two functions are for anonymous (private mapped) hugepages. |
| 1922 | * Unlike common anonymous pages, anonymous hugepages have no accounting code |
| 1923 | * and no lru code, because we handle hugepages differently from common pages. |
| 1924 | */ |
| 1925 | void hugepage_add_anon_rmap(struct page *page, |
| 1926 | struct vm_area_struct *vma, unsigned long address) |
| 1927 | { |
| 1928 | struct anon_vma *anon_vma = vma->anon_vma; |
| 1929 | int first; |
| 1930 | |
| 1931 | BUG_ON(!PageLocked(page)); |
| 1932 | BUG_ON(!anon_vma); |
| 1933 | /* address might be in next vma when migration races vma_adjust */ |
| 1934 | first = atomic_inc_and_test(compound_mapcount_ptr(page)); |
| 1935 | if (first) |
| 1936 | __page_set_anon_rmap(page, vma, address, 0); |
| 1937 | } |
| 1938 | |
| 1939 | void hugepage_add_new_anon_rmap(struct page *page, |
| 1940 | struct vm_area_struct *vma, unsigned long address) |
| 1941 | { |
| 1942 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
| 1943 | atomic_set(compound_mapcount_ptr(page), 0); |
| 1944 | __page_set_anon_rmap(page, vma, address, 1); |
| 1945 | } |
| 1946 | #endif /* CONFIG_HUGETLB_PAGE */ |