| 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 | * mapping->i_mmap_mutex |
| 27 | * anon_vma->rwsem |
| 28 | * mm->page_table_lock or pte_lock |
| 29 | * zone->lru_lock (in mark_page_accessed, isolate_lru_page) |
| 30 | * swap_lock (in swap_duplicate, swap_info_get) |
| 31 | * mmlist_lock (in mmput, drain_mmlist and others) |
| 32 | * mapping->private_lock (in __set_page_dirty_buffers) |
| 33 | * inode->i_lock (in set_page_dirty's __mark_inode_dirty) |
| 34 | * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) |
| 35 | * sb_lock (within inode_lock in fs/fs-writeback.c) |
| 36 | * mapping->tree_lock (widely used, in set_page_dirty, |
| 37 | * in arch-dependent flush_dcache_mmap_lock, |
| 38 | * within bdi.wb->list_lock in __sync_single_inode) |
| 39 | * |
| 40 | * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) |
| 41 | * ->tasklist_lock |
| 42 | * pte map lock |
| 43 | */ |
| 44 | |
| 45 | #include <linux/mm.h> |
| 46 | #include <linux/pagemap.h> |
| 47 | #include <linux/swap.h> |
| 48 | #include <linux/swapops.h> |
| 49 | #include <linux/slab.h> |
| 50 | #include <linux/init.h> |
| 51 | #include <linux/ksm.h> |
| 52 | #include <linux/rmap.h> |
| 53 | #include <linux/rcupdate.h> |
| 54 | #include <linux/export.h> |
| 55 | #include <linux/memcontrol.h> |
| 56 | #include <linux/mmu_notifier.h> |
| 57 | #include <linux/migrate.h> |
| 58 | #include <linux/hugetlb.h> |
| 59 | #include <linux/backing-dev.h> |
| 60 | |
| 61 | #include <asm/tlbflush.h> |
| 62 | |
| 63 | #include "internal.h" |
| 64 | |
| 65 | static struct kmem_cache *anon_vma_cachep; |
| 66 | static struct kmem_cache *anon_vma_chain_cachep; |
| 67 | |
| 68 | static inline struct anon_vma *anon_vma_alloc(void) |
| 69 | { |
| 70 | struct anon_vma *anon_vma; |
| 71 | |
| 72 | anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); |
| 73 | if (anon_vma) { |
| 74 | atomic_set(&anon_vma->refcount, 1); |
| 75 | /* |
| 76 | * Initialise the anon_vma root to point to itself. If called |
| 77 | * from fork, the root will be reset to the parents anon_vma. |
| 78 | */ |
| 79 | anon_vma->root = anon_vma; |
| 80 | } |
| 81 | |
| 82 | return anon_vma; |
| 83 | } |
| 84 | |
| 85 | static inline void anon_vma_free(struct anon_vma *anon_vma) |
| 86 | { |
| 87 | VM_BUG_ON(atomic_read(&anon_vma->refcount)); |
| 88 | |
| 89 | /* |
| 90 | * Synchronize against page_lock_anon_vma_read() such that |
| 91 | * we can safely hold the lock without the anon_vma getting |
| 92 | * freed. |
| 93 | * |
| 94 | * Relies on the full mb implied by the atomic_dec_and_test() from |
| 95 | * put_anon_vma() against the acquire barrier implied by |
| 96 | * down_read_trylock() from page_lock_anon_vma_read(). This orders: |
| 97 | * |
| 98 | * page_lock_anon_vma_read() VS put_anon_vma() |
| 99 | * down_read_trylock() atomic_dec_and_test() |
| 100 | * LOCK MB |
| 101 | * atomic_read() rwsem_is_locked() |
| 102 | * |
| 103 | * LOCK should suffice since the actual taking of the lock must |
| 104 | * happen _before_ what follows. |
| 105 | */ |
| 106 | might_sleep(); |
| 107 | if (rwsem_is_locked(&anon_vma->root->rwsem)) { |
| 108 | anon_vma_lock_write(anon_vma); |
| 109 | anon_vma_unlock_write(anon_vma); |
| 110 | } |
| 111 | |
| 112 | kmem_cache_free(anon_vma_cachep, anon_vma); |
| 113 | } |
| 114 | |
| 115 | static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) |
| 116 | { |
| 117 | return kmem_cache_alloc(anon_vma_chain_cachep, gfp); |
| 118 | } |
| 119 | |
| 120 | static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) |
| 121 | { |
| 122 | kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); |
| 123 | } |
| 124 | |
| 125 | static void anon_vma_chain_link(struct vm_area_struct *vma, |
| 126 | struct anon_vma_chain *avc, |
| 127 | struct anon_vma *anon_vma) |
| 128 | { |
| 129 | avc->vma = vma; |
| 130 | avc->anon_vma = anon_vma; |
| 131 | list_add(&avc->same_vma, &vma->anon_vma_chain); |
| 132 | anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); |
| 133 | } |
| 134 | |
| 135 | /** |
| 136 | * anon_vma_prepare - attach an anon_vma to a memory region |
| 137 | * @vma: the memory region in question |
| 138 | * |
| 139 | * This makes sure the memory mapping described by 'vma' has |
| 140 | * an 'anon_vma' attached to it, so that we can associate the |
| 141 | * anonymous pages mapped into it with that anon_vma. |
| 142 | * |
| 143 | * The common case will be that we already have one, but if |
| 144 | * not we either need to find an adjacent mapping that we |
| 145 | * can re-use the anon_vma from (very common when the only |
| 146 | * reason for splitting a vma has been mprotect()), or we |
| 147 | * allocate a new one. |
| 148 | * |
| 149 | * Anon-vma allocations are very subtle, because we may have |
| 150 | * optimistically looked up an anon_vma in page_lock_anon_vma_read() |
| 151 | * and that may actually touch the spinlock even in the newly |
| 152 | * allocated vma (it depends on RCU to make sure that the |
| 153 | * anon_vma isn't actually destroyed). |
| 154 | * |
| 155 | * As a result, we need to do proper anon_vma locking even |
| 156 | * for the new allocation. At the same time, we do not want |
| 157 | * to do any locking for the common case of already having |
| 158 | * an anon_vma. |
| 159 | * |
| 160 | * This must be called with the mmap_sem held for reading. |
| 161 | */ |
| 162 | int anon_vma_prepare(struct vm_area_struct *vma) |
| 163 | { |
| 164 | struct anon_vma *anon_vma = vma->anon_vma; |
| 165 | struct anon_vma_chain *avc; |
| 166 | |
| 167 | might_sleep(); |
| 168 | if (unlikely(!anon_vma)) { |
| 169 | struct mm_struct *mm = vma->vm_mm; |
| 170 | struct anon_vma *allocated; |
| 171 | |
| 172 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 173 | if (!avc) |
| 174 | goto out_enomem; |
| 175 | |
| 176 | anon_vma = find_mergeable_anon_vma(vma); |
| 177 | allocated = NULL; |
| 178 | if (!anon_vma) { |
| 179 | anon_vma = anon_vma_alloc(); |
| 180 | if (unlikely(!anon_vma)) |
| 181 | goto out_enomem_free_avc; |
| 182 | allocated = anon_vma; |
| 183 | } |
| 184 | |
| 185 | anon_vma_lock_write(anon_vma); |
| 186 | /* page_table_lock to protect against threads */ |
| 187 | spin_lock(&mm->page_table_lock); |
| 188 | if (likely(!vma->anon_vma)) { |
| 189 | vma->anon_vma = anon_vma; |
| 190 | anon_vma_chain_link(vma, avc, anon_vma); |
| 191 | allocated = NULL; |
| 192 | avc = NULL; |
| 193 | } |
| 194 | spin_unlock(&mm->page_table_lock); |
| 195 | anon_vma_unlock_write(anon_vma); |
| 196 | |
| 197 | if (unlikely(allocated)) |
| 198 | put_anon_vma(allocated); |
| 199 | if (unlikely(avc)) |
| 200 | anon_vma_chain_free(avc); |
| 201 | } |
| 202 | return 0; |
| 203 | |
| 204 | out_enomem_free_avc: |
| 205 | anon_vma_chain_free(avc); |
| 206 | out_enomem: |
| 207 | return -ENOMEM; |
| 208 | } |
| 209 | |
| 210 | /* |
| 211 | * This is a useful helper function for locking the anon_vma root as |
| 212 | * we traverse the vma->anon_vma_chain, looping over anon_vma's that |
| 213 | * have the same vma. |
| 214 | * |
| 215 | * Such anon_vma's should have the same root, so you'd expect to see |
| 216 | * just a single mutex_lock for the whole traversal. |
| 217 | */ |
| 218 | static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) |
| 219 | { |
| 220 | struct anon_vma *new_root = anon_vma->root; |
| 221 | if (new_root != root) { |
| 222 | if (WARN_ON_ONCE(root)) |
| 223 | up_write(&root->rwsem); |
| 224 | root = new_root; |
| 225 | down_write(&root->rwsem); |
| 226 | } |
| 227 | return root; |
| 228 | } |
| 229 | |
| 230 | static inline void unlock_anon_vma_root(struct anon_vma *root) |
| 231 | { |
| 232 | if (root) |
| 233 | up_write(&root->rwsem); |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * Attach the anon_vmas from src to dst. |
| 238 | * Returns 0 on success, -ENOMEM on failure. |
| 239 | */ |
| 240 | int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) |
| 241 | { |
| 242 | struct anon_vma_chain *avc, *pavc; |
| 243 | struct anon_vma *root = NULL; |
| 244 | |
| 245 | list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { |
| 246 | struct anon_vma *anon_vma; |
| 247 | |
| 248 | avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); |
| 249 | if (unlikely(!avc)) { |
| 250 | unlock_anon_vma_root(root); |
| 251 | root = NULL; |
| 252 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 253 | if (!avc) |
| 254 | goto enomem_failure; |
| 255 | } |
| 256 | anon_vma = pavc->anon_vma; |
| 257 | root = lock_anon_vma_root(root, anon_vma); |
| 258 | anon_vma_chain_link(dst, avc, anon_vma); |
| 259 | } |
| 260 | unlock_anon_vma_root(root); |
| 261 | return 0; |
| 262 | |
| 263 | enomem_failure: |
| 264 | unlink_anon_vmas(dst); |
| 265 | return -ENOMEM; |
| 266 | } |
| 267 | |
| 268 | /* |
| 269 | * Attach vma to its own anon_vma, as well as to the anon_vmas that |
| 270 | * the corresponding VMA in the parent process is attached to. |
| 271 | * Returns 0 on success, non-zero on failure. |
| 272 | */ |
| 273 | int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) |
| 274 | { |
| 275 | struct anon_vma_chain *avc; |
| 276 | struct anon_vma *anon_vma; |
| 277 | |
| 278 | /* Don't bother if the parent process has no anon_vma here. */ |
| 279 | if (!pvma->anon_vma) |
| 280 | return 0; |
| 281 | |
| 282 | /* |
| 283 | * First, attach the new VMA to the parent VMA's anon_vmas, |
| 284 | * so rmap can find non-COWed pages in child processes. |
| 285 | */ |
| 286 | if (anon_vma_clone(vma, pvma)) |
| 287 | return -ENOMEM; |
| 288 | |
| 289 | /* Then add our own anon_vma. */ |
| 290 | anon_vma = anon_vma_alloc(); |
| 291 | if (!anon_vma) |
| 292 | goto out_error; |
| 293 | avc = anon_vma_chain_alloc(GFP_KERNEL); |
| 294 | if (!avc) |
| 295 | goto out_error_free_anon_vma; |
| 296 | |
| 297 | /* |
| 298 | * The root anon_vma's spinlock is the lock actually used when we |
| 299 | * lock any of the anon_vmas in this anon_vma tree. |
| 300 | */ |
| 301 | anon_vma->root = pvma->anon_vma->root; |
| 302 | /* |
| 303 | * With refcounts, an anon_vma can stay around longer than the |
| 304 | * process it belongs to. The root anon_vma needs to be pinned until |
| 305 | * this anon_vma is freed, because the lock lives in the root. |
| 306 | */ |
| 307 | get_anon_vma(anon_vma->root); |
| 308 | /* Mark this anon_vma as the one where our new (COWed) pages go. */ |
| 309 | vma->anon_vma = anon_vma; |
| 310 | anon_vma_lock_write(anon_vma); |
| 311 | anon_vma_chain_link(vma, avc, anon_vma); |
| 312 | anon_vma_unlock_write(anon_vma); |
| 313 | |
| 314 | return 0; |
| 315 | |
| 316 | out_error_free_anon_vma: |
| 317 | put_anon_vma(anon_vma); |
| 318 | out_error: |
| 319 | unlink_anon_vmas(vma); |
| 320 | return -ENOMEM; |
| 321 | } |
| 322 | |
| 323 | void unlink_anon_vmas(struct vm_area_struct *vma) |
| 324 | { |
| 325 | struct anon_vma_chain *avc, *next; |
| 326 | struct anon_vma *root = NULL; |
| 327 | |
| 328 | /* |
| 329 | * Unlink each anon_vma chained to the VMA. This list is ordered |
| 330 | * from newest to oldest, ensuring the root anon_vma gets freed last. |
| 331 | */ |
| 332 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
| 333 | struct anon_vma *anon_vma = avc->anon_vma; |
| 334 | |
| 335 | root = lock_anon_vma_root(root, anon_vma); |
| 336 | anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); |
| 337 | |
| 338 | /* |
| 339 | * Leave empty anon_vmas on the list - we'll need |
| 340 | * to free them outside the lock. |
| 341 | */ |
| 342 | if (RB_EMPTY_ROOT(&anon_vma->rb_root)) |
| 343 | continue; |
| 344 | |
| 345 | list_del(&avc->same_vma); |
| 346 | anon_vma_chain_free(avc); |
| 347 | } |
| 348 | unlock_anon_vma_root(root); |
| 349 | |
| 350 | /* |
| 351 | * Iterate the list once more, it now only contains empty and unlinked |
| 352 | * anon_vmas, destroy them. Could not do before due to __put_anon_vma() |
| 353 | * needing to write-acquire the anon_vma->root->rwsem. |
| 354 | */ |
| 355 | list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { |
| 356 | struct anon_vma *anon_vma = avc->anon_vma; |
| 357 | |
| 358 | put_anon_vma(anon_vma); |
| 359 | |
| 360 | list_del(&avc->same_vma); |
| 361 | anon_vma_chain_free(avc); |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | static void anon_vma_ctor(void *data) |
| 366 | { |
| 367 | struct anon_vma *anon_vma = data; |
| 368 | |
| 369 | init_rwsem(&anon_vma->rwsem); |
| 370 | atomic_set(&anon_vma->refcount, 0); |
| 371 | anon_vma->rb_root = RB_ROOT; |
| 372 | } |
| 373 | |
| 374 | void __init anon_vma_init(void) |
| 375 | { |
| 376 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), |
| 377 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); |
| 378 | anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); |
| 379 | } |
| 380 | |
| 381 | /* |
| 382 | * Getting a lock on a stable anon_vma from a page off the LRU is tricky! |
| 383 | * |
| 384 | * Since there is no serialization what so ever against page_remove_rmap() |
| 385 | * the best this function can do is return a locked anon_vma that might |
| 386 | * have been relevant to this page. |
| 387 | * |
| 388 | * The page might have been remapped to a different anon_vma or the anon_vma |
| 389 | * returned may already be freed (and even reused). |
| 390 | * |
| 391 | * In case it was remapped to a different anon_vma, the new anon_vma will be a |
| 392 | * child of the old anon_vma, and the anon_vma lifetime rules will therefore |
| 393 | * ensure that any anon_vma obtained from the page will still be valid for as |
| 394 | * long as we observe page_mapped() [ hence all those page_mapped() tests ]. |
| 395 | * |
| 396 | * All users of this function must be very careful when walking the anon_vma |
| 397 | * chain and verify that the page in question is indeed mapped in it |
| 398 | * [ something equivalent to page_mapped_in_vma() ]. |
| 399 | * |
| 400 | * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() |
| 401 | * that the anon_vma pointer from page->mapping is valid if there is a |
| 402 | * mapcount, we can dereference the anon_vma after observing those. |
| 403 | */ |
| 404 | struct anon_vma *page_get_anon_vma(struct page *page) |
| 405 | { |
| 406 | struct anon_vma *anon_vma = NULL; |
| 407 | unsigned long anon_mapping; |
| 408 | |
| 409 | rcu_read_lock(); |
| 410 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); |
| 411 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
| 412 | goto out; |
| 413 | if (!page_mapped(page)) |
| 414 | goto out; |
| 415 | |
| 416 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
| 417 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
| 418 | anon_vma = NULL; |
| 419 | goto out; |
| 420 | } |
| 421 | |
| 422 | /* |
| 423 | * If this page is still mapped, then its anon_vma cannot have been |
| 424 | * freed. But if it has been unmapped, we have no security against the |
| 425 | * anon_vma structure being freed and reused (for another anon_vma: |
| 426 | * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero() |
| 427 | * above cannot corrupt). |
| 428 | */ |
| 429 | if (!page_mapped(page)) { |
| 430 | rcu_read_unlock(); |
| 431 | put_anon_vma(anon_vma); |
| 432 | return NULL; |
| 433 | } |
| 434 | out: |
| 435 | rcu_read_unlock(); |
| 436 | |
| 437 | return anon_vma; |
| 438 | } |
| 439 | |
| 440 | /* |
| 441 | * Similar to page_get_anon_vma() except it locks the anon_vma. |
| 442 | * |
| 443 | * Its a little more complex as it tries to keep the fast path to a single |
| 444 | * atomic op -- the trylock. If we fail the trylock, we fall back to getting a |
| 445 | * reference like with page_get_anon_vma() and then block on the mutex. |
| 446 | */ |
| 447 | struct anon_vma *page_lock_anon_vma_read(struct page *page) |
| 448 | { |
| 449 | struct anon_vma *anon_vma = NULL; |
| 450 | struct anon_vma *root_anon_vma; |
| 451 | unsigned long anon_mapping; |
| 452 | |
| 453 | rcu_read_lock(); |
| 454 | anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); |
| 455 | if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) |
| 456 | goto out; |
| 457 | if (!page_mapped(page)) |
| 458 | goto out; |
| 459 | |
| 460 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
| 461 | root_anon_vma = ACCESS_ONCE(anon_vma->root); |
| 462 | if (down_read_trylock(&root_anon_vma->rwsem)) { |
| 463 | /* |
| 464 | * If the page is still mapped, then this anon_vma is still |
| 465 | * its anon_vma, and holding the mutex ensures that it will |
| 466 | * not go away, see anon_vma_free(). |
| 467 | */ |
| 468 | if (!page_mapped(page)) { |
| 469 | up_read(&root_anon_vma->rwsem); |
| 470 | anon_vma = NULL; |
| 471 | } |
| 472 | goto out; |
| 473 | } |
| 474 | |
| 475 | /* trylock failed, we got to sleep */ |
| 476 | if (!atomic_inc_not_zero(&anon_vma->refcount)) { |
| 477 | anon_vma = NULL; |
| 478 | goto out; |
| 479 | } |
| 480 | |
| 481 | if (!page_mapped(page)) { |
| 482 | rcu_read_unlock(); |
| 483 | put_anon_vma(anon_vma); |
| 484 | return NULL; |
| 485 | } |
| 486 | |
| 487 | /* we pinned the anon_vma, its safe to sleep */ |
| 488 | rcu_read_unlock(); |
| 489 | anon_vma_lock_read(anon_vma); |
| 490 | |
| 491 | if (atomic_dec_and_test(&anon_vma->refcount)) { |
| 492 | /* |
| 493 | * Oops, we held the last refcount, release the lock |
| 494 | * and bail -- can't simply use put_anon_vma() because |
| 495 | * we'll deadlock on the anon_vma_lock_write() recursion. |
| 496 | */ |
| 497 | anon_vma_unlock_read(anon_vma); |
| 498 | __put_anon_vma(anon_vma); |
| 499 | anon_vma = NULL; |
| 500 | } |
| 501 | |
| 502 | return anon_vma; |
| 503 | |
| 504 | out: |
| 505 | rcu_read_unlock(); |
| 506 | return anon_vma; |
| 507 | } |
| 508 | |
| 509 | void page_unlock_anon_vma_read(struct anon_vma *anon_vma) |
| 510 | { |
| 511 | anon_vma_unlock_read(anon_vma); |
| 512 | } |
| 513 | |
| 514 | /* |
| 515 | * At what user virtual address is page expected in @vma? |
| 516 | */ |
| 517 | static inline unsigned long |
| 518 | __vma_address(struct page *page, struct vm_area_struct *vma) |
| 519 | { |
| 520 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 521 | |
| 522 | if (unlikely(is_vm_hugetlb_page(vma))) |
| 523 | pgoff = page->index << huge_page_order(page_hstate(page)); |
| 524 | |
| 525 | return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); |
| 526 | } |
| 527 | |
| 528 | inline unsigned long |
| 529 | vma_address(struct page *page, struct vm_area_struct *vma) |
| 530 | { |
| 531 | unsigned long address = __vma_address(page, vma); |
| 532 | |
| 533 | /* page should be within @vma mapping range */ |
| 534 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
| 535 | |
| 536 | return address; |
| 537 | } |
| 538 | |
| 539 | /* |
| 540 | * At what user virtual address is page expected in vma? |
| 541 | * Caller should check the page is actually part of the vma. |
| 542 | */ |
| 543 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) |
| 544 | { |
| 545 | unsigned long address; |
| 546 | if (PageAnon(page)) { |
| 547 | struct anon_vma *page__anon_vma = page_anon_vma(page); |
| 548 | /* |
| 549 | * Note: swapoff's unuse_vma() is more efficient with this |
| 550 | * check, and needs it to match anon_vma when KSM is active. |
| 551 | */ |
| 552 | if (!vma->anon_vma || !page__anon_vma || |
| 553 | vma->anon_vma->root != page__anon_vma->root) |
| 554 | return -EFAULT; |
| 555 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { |
| 556 | if (!vma->vm_file || |
| 557 | vma->vm_file->f_mapping != page->mapping) |
| 558 | return -EFAULT; |
| 559 | } else |
| 560 | return -EFAULT; |
| 561 | address = __vma_address(page, vma); |
| 562 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) |
| 563 | return -EFAULT; |
| 564 | return address; |
| 565 | } |
| 566 | |
| 567 | pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) |
| 568 | { |
| 569 | pgd_t *pgd; |
| 570 | pud_t *pud; |
| 571 | pmd_t *pmd = NULL; |
| 572 | |
| 573 | pgd = pgd_offset(mm, address); |
| 574 | if (!pgd_present(*pgd)) |
| 575 | goto out; |
| 576 | |
| 577 | pud = pud_offset(pgd, address); |
| 578 | if (!pud_present(*pud)) |
| 579 | goto out; |
| 580 | |
| 581 | pmd = pmd_offset(pud, address); |
| 582 | if (!pmd_present(*pmd)) |
| 583 | pmd = NULL; |
| 584 | out: |
| 585 | return pmd; |
| 586 | } |
| 587 | |
| 588 | /* |
| 589 | * Check that @page is mapped at @address into @mm. |
| 590 | * |
| 591 | * If @sync is false, page_check_address may perform a racy check to avoid |
| 592 | * the page table lock when the pte is not present (helpful when reclaiming |
| 593 | * highly shared pages). |
| 594 | * |
| 595 | * On success returns with pte mapped and locked. |
| 596 | */ |
| 597 | pte_t *__page_check_address(struct page *page, struct mm_struct *mm, |
| 598 | unsigned long address, spinlock_t **ptlp, int sync) |
| 599 | { |
| 600 | pmd_t *pmd; |
| 601 | pte_t *pte; |
| 602 | spinlock_t *ptl; |
| 603 | |
| 604 | if (unlikely(PageHuge(page))) { |
| 605 | /* when pud is not present, pte will be NULL */ |
| 606 | pte = huge_pte_offset(mm, address); |
| 607 | if (!pte) |
| 608 | return NULL; |
| 609 | |
| 610 | ptl = &mm->page_table_lock; |
| 611 | goto check; |
| 612 | } |
| 613 | |
| 614 | pmd = mm_find_pmd(mm, address); |
| 615 | if (!pmd) |
| 616 | return NULL; |
| 617 | |
| 618 | if (pmd_trans_huge(*pmd)) |
| 619 | return NULL; |
| 620 | |
| 621 | pte = pte_offset_map(pmd, address); |
| 622 | /* Make a quick check before getting the lock */ |
| 623 | if (!sync && !pte_present(*pte)) { |
| 624 | pte_unmap(pte); |
| 625 | return NULL; |
| 626 | } |
| 627 | |
| 628 | ptl = pte_lockptr(mm, pmd); |
| 629 | check: |
| 630 | spin_lock(ptl); |
| 631 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { |
| 632 | *ptlp = ptl; |
| 633 | return pte; |
| 634 | } |
| 635 | pte_unmap_unlock(pte, ptl); |
| 636 | return NULL; |
| 637 | } |
| 638 | |
| 639 | /** |
| 640 | * page_mapped_in_vma - check whether a page is really mapped in a VMA |
| 641 | * @page: the page to test |
| 642 | * @vma: the VMA to test |
| 643 | * |
| 644 | * Returns 1 if the page is mapped into the page tables of the VMA, 0 |
| 645 | * if the page is not mapped into the page tables of this VMA. Only |
| 646 | * valid for normal file or anonymous VMAs. |
| 647 | */ |
| 648 | int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) |
| 649 | { |
| 650 | unsigned long address; |
| 651 | pte_t *pte; |
| 652 | spinlock_t *ptl; |
| 653 | |
| 654 | address = __vma_address(page, vma); |
| 655 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) |
| 656 | return 0; |
| 657 | pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); |
| 658 | if (!pte) /* the page is not in this mm */ |
| 659 | return 0; |
| 660 | pte_unmap_unlock(pte, ptl); |
| 661 | |
| 662 | return 1; |
| 663 | } |
| 664 | |
| 665 | /* |
| 666 | * Subfunctions of page_referenced: page_referenced_one called |
| 667 | * repeatedly from either page_referenced_anon or page_referenced_file. |
| 668 | */ |
| 669 | int page_referenced_one(struct page *page, struct vm_area_struct *vma, |
| 670 | unsigned long address, unsigned int *mapcount, |
| 671 | unsigned long *vm_flags) |
| 672 | { |
| 673 | struct mm_struct *mm = vma->vm_mm; |
| 674 | int referenced = 0; |
| 675 | |
| 676 | if (unlikely(PageTransHuge(page))) { |
| 677 | pmd_t *pmd; |
| 678 | |
| 679 | spin_lock(&mm->page_table_lock); |
| 680 | /* |
| 681 | * rmap might return false positives; we must filter |
| 682 | * these out using page_check_address_pmd(). |
| 683 | */ |
| 684 | pmd = page_check_address_pmd(page, mm, address, |
| 685 | PAGE_CHECK_ADDRESS_PMD_FLAG); |
| 686 | if (!pmd) { |
| 687 | spin_unlock(&mm->page_table_lock); |
| 688 | goto out; |
| 689 | } |
| 690 | |
| 691 | if (vma->vm_flags & VM_LOCKED) { |
| 692 | spin_unlock(&mm->page_table_lock); |
| 693 | *mapcount = 0; /* break early from loop */ |
| 694 | *vm_flags |= VM_LOCKED; |
| 695 | goto out; |
| 696 | } |
| 697 | |
| 698 | /* go ahead even if the pmd is pmd_trans_splitting() */ |
| 699 | if (pmdp_clear_flush_young_notify(vma, address, pmd)) |
| 700 | referenced++; |
| 701 | spin_unlock(&mm->page_table_lock); |
| 702 | } else { |
| 703 | pte_t *pte; |
| 704 | spinlock_t *ptl; |
| 705 | |
| 706 | /* |
| 707 | * rmap might return false positives; we must filter |
| 708 | * these out using page_check_address(). |
| 709 | */ |
| 710 | pte = page_check_address(page, mm, address, &ptl, 0); |
| 711 | if (!pte) |
| 712 | goto out; |
| 713 | |
| 714 | if (vma->vm_flags & VM_LOCKED) { |
| 715 | pte_unmap_unlock(pte, ptl); |
| 716 | *mapcount = 0; /* break early from loop */ |
| 717 | *vm_flags |= VM_LOCKED; |
| 718 | goto out; |
| 719 | } |
| 720 | |
| 721 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
| 722 | /* |
| 723 | * Don't treat a reference through a sequentially read |
| 724 | * mapping as such. If the page has been used in |
| 725 | * another mapping, we will catch it; if this other |
| 726 | * mapping is already gone, the unmap path will have |
| 727 | * set PG_referenced or activated the page. |
| 728 | */ |
| 729 | if (likely(!VM_SequentialReadHint(vma))) |
| 730 | referenced++; |
| 731 | } |
| 732 | pte_unmap_unlock(pte, ptl); |
| 733 | } |
| 734 | |
| 735 | (*mapcount)--; |
| 736 | |
| 737 | if (referenced) |
| 738 | *vm_flags |= vma->vm_flags; |
| 739 | out: |
| 740 | return referenced; |
| 741 | } |
| 742 | |
| 743 | static int page_referenced_anon(struct page *page, |
| 744 | struct mem_cgroup *memcg, |
| 745 | unsigned long *vm_flags) |
| 746 | { |
| 747 | unsigned int mapcount; |
| 748 | struct anon_vma *anon_vma; |
| 749 | pgoff_t pgoff; |
| 750 | struct anon_vma_chain *avc; |
| 751 | int referenced = 0; |
| 752 | |
| 753 | anon_vma = page_lock_anon_vma_read(page); |
| 754 | if (!anon_vma) |
| 755 | return referenced; |
| 756 | |
| 757 | mapcount = page_mapcount(page); |
| 758 | pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 759 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
| 760 | struct vm_area_struct *vma = avc->vma; |
| 761 | unsigned long address = vma_address(page, vma); |
| 762 | /* |
| 763 | * If we are reclaiming on behalf of a cgroup, skip |
| 764 | * counting on behalf of references from different |
| 765 | * cgroups |
| 766 | */ |
| 767 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) |
| 768 | continue; |
| 769 | referenced += page_referenced_one(page, vma, address, |
| 770 | &mapcount, vm_flags); |
| 771 | if (!mapcount) |
| 772 | break; |
| 773 | } |
| 774 | |
| 775 | page_unlock_anon_vma_read(anon_vma); |
| 776 | return referenced; |
| 777 | } |
| 778 | |
| 779 | /** |
| 780 | * page_referenced_file - referenced check for object-based rmap |
| 781 | * @page: the page we're checking references on. |
| 782 | * @memcg: target memory control group |
| 783 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
| 784 | * |
| 785 | * For an object-based mapped page, find all the places it is mapped and |
| 786 | * check/clear the referenced flag. This is done by following the page->mapping |
| 787 | * pointer, then walking the chain of vmas it holds. It returns the number |
| 788 | * of references it found. |
| 789 | * |
| 790 | * This function is only called from page_referenced for object-based pages. |
| 791 | */ |
| 792 | static int page_referenced_file(struct page *page, |
| 793 | struct mem_cgroup *memcg, |
| 794 | unsigned long *vm_flags) |
| 795 | { |
| 796 | unsigned int mapcount; |
| 797 | struct address_space *mapping = page->mapping; |
| 798 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 799 | struct vm_area_struct *vma; |
| 800 | int referenced = 0; |
| 801 | |
| 802 | /* |
| 803 | * The caller's checks on page->mapping and !PageAnon have made |
| 804 | * sure that this is a file page: the check for page->mapping |
| 805 | * excludes the case just before it gets set on an anon page. |
| 806 | */ |
| 807 | BUG_ON(PageAnon(page)); |
| 808 | |
| 809 | /* |
| 810 | * The page lock not only makes sure that page->mapping cannot |
| 811 | * suddenly be NULLified by truncation, it makes sure that the |
| 812 | * structure at mapping cannot be freed and reused yet, |
| 813 | * so we can safely take mapping->i_mmap_mutex. |
| 814 | */ |
| 815 | BUG_ON(!PageLocked(page)); |
| 816 | |
| 817 | mutex_lock(&mapping->i_mmap_mutex); |
| 818 | |
| 819 | /* |
| 820 | * i_mmap_mutex does not stabilize mapcount at all, but mapcount |
| 821 | * is more likely to be accurate if we note it after spinning. |
| 822 | */ |
| 823 | mapcount = page_mapcount(page); |
| 824 | |
| 825 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { |
| 826 | unsigned long address = vma_address(page, vma); |
| 827 | /* |
| 828 | * If we are reclaiming on behalf of a cgroup, skip |
| 829 | * counting on behalf of references from different |
| 830 | * cgroups |
| 831 | */ |
| 832 | if (memcg && !mm_match_cgroup(vma->vm_mm, memcg)) |
| 833 | continue; |
| 834 | referenced += page_referenced_one(page, vma, address, |
| 835 | &mapcount, vm_flags); |
| 836 | if (!mapcount) |
| 837 | break; |
| 838 | } |
| 839 | |
| 840 | mutex_unlock(&mapping->i_mmap_mutex); |
| 841 | return referenced; |
| 842 | } |
| 843 | |
| 844 | /** |
| 845 | * page_referenced - test if the page was referenced |
| 846 | * @page: the page to test |
| 847 | * @is_locked: caller holds lock on the page |
| 848 | * @memcg: target memory cgroup |
| 849 | * @vm_flags: collect encountered vma->vm_flags who actually referenced the page |
| 850 | * |
| 851 | * Quick test_and_clear_referenced for all mappings to a page, |
| 852 | * returns the number of ptes which referenced the page. |
| 853 | */ |
| 854 | int page_referenced(struct page *page, |
| 855 | int is_locked, |
| 856 | struct mem_cgroup *memcg, |
| 857 | unsigned long *vm_flags) |
| 858 | { |
| 859 | int referenced = 0; |
| 860 | int we_locked = 0; |
| 861 | |
| 862 | *vm_flags = 0; |
| 863 | if (page_mapped(page) && page_rmapping(page)) { |
| 864 | if (!is_locked && (!PageAnon(page) || PageKsm(page))) { |
| 865 | we_locked = trylock_page(page); |
| 866 | if (!we_locked) { |
| 867 | referenced++; |
| 868 | goto out; |
| 869 | } |
| 870 | } |
| 871 | if (unlikely(PageKsm(page))) |
| 872 | referenced += page_referenced_ksm(page, memcg, |
| 873 | vm_flags); |
| 874 | else if (PageAnon(page)) |
| 875 | referenced += page_referenced_anon(page, memcg, |
| 876 | vm_flags); |
| 877 | else if (page->mapping) |
| 878 | referenced += page_referenced_file(page, memcg, |
| 879 | vm_flags); |
| 880 | if (we_locked) |
| 881 | unlock_page(page); |
| 882 | |
| 883 | if (page_test_and_clear_young(page_to_pfn(page))) |
| 884 | referenced++; |
| 885 | } |
| 886 | out: |
| 887 | return referenced; |
| 888 | } |
| 889 | |
| 890 | static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, |
| 891 | unsigned long address) |
| 892 | { |
| 893 | struct mm_struct *mm = vma->vm_mm; |
| 894 | pte_t *pte; |
| 895 | spinlock_t *ptl; |
| 896 | int ret = 0; |
| 897 | |
| 898 | pte = page_check_address(page, mm, address, &ptl, 1); |
| 899 | if (!pte) |
| 900 | goto out; |
| 901 | |
| 902 | if (pte_dirty(*pte) || pte_write(*pte)) { |
| 903 | pte_t entry; |
| 904 | |
| 905 | flush_cache_page(vma, address, pte_pfn(*pte)); |
| 906 | entry = ptep_clear_flush(vma, address, pte); |
| 907 | entry = pte_wrprotect(entry); |
| 908 | entry = pte_mkclean(entry); |
| 909 | set_pte_at(mm, address, pte, entry); |
| 910 | ret = 1; |
| 911 | } |
| 912 | |
| 913 | pte_unmap_unlock(pte, ptl); |
| 914 | |
| 915 | if (ret) |
| 916 | mmu_notifier_invalidate_page(mm, address); |
| 917 | out: |
| 918 | return ret; |
| 919 | } |
| 920 | |
| 921 | static int page_mkclean_file(struct address_space *mapping, struct page *page) |
| 922 | { |
| 923 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 924 | struct vm_area_struct *vma; |
| 925 | int ret = 0; |
| 926 | |
| 927 | BUG_ON(PageAnon(page)); |
| 928 | |
| 929 | mutex_lock(&mapping->i_mmap_mutex); |
| 930 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { |
| 931 | if (vma->vm_flags & VM_SHARED) { |
| 932 | unsigned long address = vma_address(page, vma); |
| 933 | ret += page_mkclean_one(page, vma, address); |
| 934 | } |
| 935 | } |
| 936 | mutex_unlock(&mapping->i_mmap_mutex); |
| 937 | return ret; |
| 938 | } |
| 939 | |
| 940 | int page_mkclean(struct page *page) |
| 941 | { |
| 942 | int ret = 0; |
| 943 | |
| 944 | BUG_ON(!PageLocked(page)); |
| 945 | |
| 946 | if (page_mapped(page)) { |
| 947 | struct address_space *mapping = page_mapping(page); |
| 948 | if (mapping) |
| 949 | ret = page_mkclean_file(mapping, page); |
| 950 | } |
| 951 | |
| 952 | return ret; |
| 953 | } |
| 954 | EXPORT_SYMBOL_GPL(page_mkclean); |
| 955 | |
| 956 | /** |
| 957 | * page_move_anon_rmap - move a page to our anon_vma |
| 958 | * @page: the page to move to our anon_vma |
| 959 | * @vma: the vma the page belongs to |
| 960 | * @address: the user virtual address mapped |
| 961 | * |
| 962 | * When a page belongs exclusively to one process after a COW event, |
| 963 | * that page can be moved into the anon_vma that belongs to just that |
| 964 | * process, so the rmap code will not search the parent or sibling |
| 965 | * processes. |
| 966 | */ |
| 967 | void page_move_anon_rmap(struct page *page, |
| 968 | struct vm_area_struct *vma, unsigned long address) |
| 969 | { |
| 970 | struct anon_vma *anon_vma = vma->anon_vma; |
| 971 | |
| 972 | VM_BUG_ON(!PageLocked(page)); |
| 973 | VM_BUG_ON(!anon_vma); |
| 974 | VM_BUG_ON(page->index != linear_page_index(vma, address)); |
| 975 | |
| 976 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
| 977 | page->mapping = (struct address_space *) anon_vma; |
| 978 | } |
| 979 | |
| 980 | /** |
| 981 | * __page_set_anon_rmap - set up new anonymous rmap |
| 982 | * @page: Page to add to rmap |
| 983 | * @vma: VM area to add page to. |
| 984 | * @address: User virtual address of the mapping |
| 985 | * @exclusive: the page is exclusively owned by the current process |
| 986 | */ |
| 987 | static void __page_set_anon_rmap(struct page *page, |
| 988 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
| 989 | { |
| 990 | struct anon_vma *anon_vma = vma->anon_vma; |
| 991 | |
| 992 | BUG_ON(!anon_vma); |
| 993 | |
| 994 | if (PageAnon(page)) |
| 995 | return; |
| 996 | |
| 997 | /* |
| 998 | * If the page isn't exclusively mapped into this vma, |
| 999 | * we must use the _oldest_ possible anon_vma for the |
| 1000 | * page mapping! |
| 1001 | */ |
| 1002 | if (!exclusive) |
| 1003 | anon_vma = anon_vma->root; |
| 1004 | |
| 1005 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
| 1006 | page->mapping = (struct address_space *) anon_vma; |
| 1007 | page->index = linear_page_index(vma, address); |
| 1008 | } |
| 1009 | |
| 1010 | /** |
| 1011 | * __page_check_anon_rmap - sanity check anonymous rmap addition |
| 1012 | * @page: the page to add the mapping to |
| 1013 | * @vma: the vm area in which the mapping is added |
| 1014 | * @address: the user virtual address mapped |
| 1015 | */ |
| 1016 | static void __page_check_anon_rmap(struct page *page, |
| 1017 | struct vm_area_struct *vma, unsigned long address) |
| 1018 | { |
| 1019 | #ifdef CONFIG_DEBUG_VM |
| 1020 | /* |
| 1021 | * The page's anon-rmap details (mapping and index) are guaranteed to |
| 1022 | * be set up correctly at this point. |
| 1023 | * |
| 1024 | * We have exclusion against page_add_anon_rmap because the caller |
| 1025 | * always holds the page locked, except if called from page_dup_rmap, |
| 1026 | * in which case the page is already known to be setup. |
| 1027 | * |
| 1028 | * We have exclusion against page_add_new_anon_rmap because those pages |
| 1029 | * are initially only visible via the pagetables, and the pte is locked |
| 1030 | * over the call to page_add_new_anon_rmap. |
| 1031 | */ |
| 1032 | BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); |
| 1033 | BUG_ON(page->index != linear_page_index(vma, address)); |
| 1034 | #endif |
| 1035 | } |
| 1036 | |
| 1037 | /** |
| 1038 | * page_add_anon_rmap - add pte mapping to an anonymous page |
| 1039 | * @page: the page to add the mapping to |
| 1040 | * @vma: the vm area in which the mapping is added |
| 1041 | * @address: the user virtual address mapped |
| 1042 | * |
| 1043 | * The caller needs to hold the pte lock, and the page must be locked in |
| 1044 | * the anon_vma case: to serialize mapping,index checking after setting, |
| 1045 | * and to ensure that PageAnon is not being upgraded racily to PageKsm |
| 1046 | * (but PageKsm is never downgraded to PageAnon). |
| 1047 | */ |
| 1048 | void page_add_anon_rmap(struct page *page, |
| 1049 | struct vm_area_struct *vma, unsigned long address) |
| 1050 | { |
| 1051 | do_page_add_anon_rmap(page, vma, address, 0); |
| 1052 | } |
| 1053 | |
| 1054 | /* |
| 1055 | * Special version of the above for do_swap_page, which often runs |
| 1056 | * into pages that are exclusively owned by the current process. |
| 1057 | * Everybody else should continue to use page_add_anon_rmap above. |
| 1058 | */ |
| 1059 | void do_page_add_anon_rmap(struct page *page, |
| 1060 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
| 1061 | { |
| 1062 | int first = atomic_inc_and_test(&page->_mapcount); |
| 1063 | if (first) { |
| 1064 | if (!PageTransHuge(page)) |
| 1065 | __inc_zone_page_state(page, NR_ANON_PAGES); |
| 1066 | else |
| 1067 | __inc_zone_page_state(page, |
| 1068 | NR_ANON_TRANSPARENT_HUGEPAGES); |
| 1069 | } |
| 1070 | if (unlikely(PageKsm(page))) |
| 1071 | return; |
| 1072 | |
| 1073 | VM_BUG_ON(!PageLocked(page)); |
| 1074 | /* address might be in next vma when migration races vma_adjust */ |
| 1075 | if (first) |
| 1076 | __page_set_anon_rmap(page, vma, address, exclusive); |
| 1077 | else |
| 1078 | __page_check_anon_rmap(page, vma, address); |
| 1079 | } |
| 1080 | |
| 1081 | /** |
| 1082 | * page_add_new_anon_rmap - add pte mapping to a new anonymous page |
| 1083 | * @page: the page to add the mapping to |
| 1084 | * @vma: the vm area in which the mapping is added |
| 1085 | * @address: the user virtual address mapped |
| 1086 | * |
| 1087 | * Same as page_add_anon_rmap but must only be called on *new* pages. |
| 1088 | * This means the inc-and-test can be bypassed. |
| 1089 | * Page does not have to be locked. |
| 1090 | */ |
| 1091 | void page_add_new_anon_rmap(struct page *page, |
| 1092 | struct vm_area_struct *vma, unsigned long address) |
| 1093 | { |
| 1094 | VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
| 1095 | SetPageSwapBacked(page); |
| 1096 | atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ |
| 1097 | if (!PageTransHuge(page)) |
| 1098 | __inc_zone_page_state(page, NR_ANON_PAGES); |
| 1099 | else |
| 1100 | __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); |
| 1101 | __page_set_anon_rmap(page, vma, address, 1); |
| 1102 | if (!mlocked_vma_newpage(vma, page)) |
| 1103 | lru_cache_add_lru(page, LRU_ACTIVE_ANON); |
| 1104 | else |
| 1105 | add_page_to_unevictable_list(page); |
| 1106 | } |
| 1107 | |
| 1108 | /** |
| 1109 | * page_add_file_rmap - add pte mapping to a file page |
| 1110 | * @page: the page to add the mapping to |
| 1111 | * |
| 1112 | * The caller needs to hold the pte lock. |
| 1113 | */ |
| 1114 | void page_add_file_rmap(struct page *page) |
| 1115 | { |
| 1116 | bool locked; |
| 1117 | unsigned long flags; |
| 1118 | |
| 1119 | mem_cgroup_begin_update_page_stat(page, &locked, &flags); |
| 1120 | if (atomic_inc_and_test(&page->_mapcount)) { |
| 1121 | __inc_zone_page_state(page, NR_FILE_MAPPED); |
| 1122 | mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED); |
| 1123 | } |
| 1124 | mem_cgroup_end_update_page_stat(page, &locked, &flags); |
| 1125 | } |
| 1126 | |
| 1127 | /** |
| 1128 | * page_remove_rmap - take down pte mapping from a page |
| 1129 | * @page: page to remove mapping from |
| 1130 | * |
| 1131 | * The caller needs to hold the pte lock. |
| 1132 | */ |
| 1133 | void page_remove_rmap(struct page *page) |
| 1134 | { |
| 1135 | bool anon = PageAnon(page); |
| 1136 | bool locked; |
| 1137 | unsigned long flags; |
| 1138 | |
| 1139 | /* |
| 1140 | * The anon case has no mem_cgroup page_stat to update; but may |
| 1141 | * uncharge_page() below, where the lock ordering can deadlock if |
| 1142 | * we hold the lock against page_stat move: so avoid it on anon. |
| 1143 | */ |
| 1144 | if (!anon) |
| 1145 | mem_cgroup_begin_update_page_stat(page, &locked, &flags); |
| 1146 | |
| 1147 | /* page still mapped by someone else? */ |
| 1148 | if (!atomic_add_negative(-1, &page->_mapcount)) |
| 1149 | goto out; |
| 1150 | |
| 1151 | /* |
| 1152 | * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED |
| 1153 | * and not charged by memcg for now. |
| 1154 | */ |
| 1155 | if (unlikely(PageHuge(page))) |
| 1156 | goto out; |
| 1157 | if (anon) { |
| 1158 | mem_cgroup_uncharge_page(page); |
| 1159 | if (!PageTransHuge(page)) |
| 1160 | __dec_zone_page_state(page, NR_ANON_PAGES); |
| 1161 | else |
| 1162 | __dec_zone_page_state(page, |
| 1163 | NR_ANON_TRANSPARENT_HUGEPAGES); |
| 1164 | } else { |
| 1165 | __dec_zone_page_state(page, NR_FILE_MAPPED); |
| 1166 | mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED); |
| 1167 | mem_cgroup_end_update_page_stat(page, &locked, &flags); |
| 1168 | } |
| 1169 | if (unlikely(PageMlocked(page))) |
| 1170 | clear_page_mlock(page); |
| 1171 | /* |
| 1172 | * It would be tidy to reset the PageAnon mapping here, |
| 1173 | * but that might overwrite a racing page_add_anon_rmap |
| 1174 | * which increments mapcount after us but sets mapping |
| 1175 | * before us: so leave the reset to free_hot_cold_page, |
| 1176 | * and remember that it's only reliable while mapped. |
| 1177 | * Leaving it set also helps swapoff to reinstate ptes |
| 1178 | * faster for those pages still in swapcache. |
| 1179 | */ |
| 1180 | return; |
| 1181 | out: |
| 1182 | if (!anon) |
| 1183 | mem_cgroup_end_update_page_stat(page, &locked, &flags); |
| 1184 | } |
| 1185 | |
| 1186 | /* |
| 1187 | * Subfunctions of try_to_unmap: try_to_unmap_one called |
| 1188 | * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file. |
| 1189 | */ |
| 1190 | int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, |
| 1191 | unsigned long address, enum ttu_flags flags) |
| 1192 | { |
| 1193 | struct mm_struct *mm = vma->vm_mm; |
| 1194 | pte_t *pte; |
| 1195 | pte_t pteval; |
| 1196 | spinlock_t *ptl; |
| 1197 | int ret = SWAP_AGAIN; |
| 1198 | |
| 1199 | pte = page_check_address(page, mm, address, &ptl, 0); |
| 1200 | if (!pte) |
| 1201 | goto out; |
| 1202 | |
| 1203 | /* |
| 1204 | * If the page is mlock()d, we cannot swap it out. |
| 1205 | * If it's recently referenced (perhaps page_referenced |
| 1206 | * skipped over this mm) then we should reactivate it. |
| 1207 | */ |
| 1208 | if (!(flags & TTU_IGNORE_MLOCK)) { |
| 1209 | if (vma->vm_flags & VM_LOCKED) |
| 1210 | goto out_mlock; |
| 1211 | |
| 1212 | if (TTU_ACTION(flags) == TTU_MUNLOCK) |
| 1213 | goto out_unmap; |
| 1214 | } |
| 1215 | if (!(flags & TTU_IGNORE_ACCESS)) { |
| 1216 | if (ptep_clear_flush_young_notify(vma, address, pte)) { |
| 1217 | ret = SWAP_FAIL; |
| 1218 | goto out_unmap; |
| 1219 | } |
| 1220 | } |
| 1221 | |
| 1222 | /* Nuke the page table entry. */ |
| 1223 | flush_cache_page(vma, address, page_to_pfn(page)); |
| 1224 | pteval = ptep_clear_flush(vma, address, pte); |
| 1225 | |
| 1226 | /* Move the dirty bit to the physical page now the pte is gone. */ |
| 1227 | if (pte_dirty(pteval)) |
| 1228 | set_page_dirty(page); |
| 1229 | |
| 1230 | /* Update high watermark before we lower rss */ |
| 1231 | update_hiwater_rss(mm); |
| 1232 | |
| 1233 | if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { |
| 1234 | if (!PageHuge(page)) { |
| 1235 | if (PageAnon(page)) |
| 1236 | dec_mm_counter(mm, MM_ANONPAGES); |
| 1237 | else |
| 1238 | dec_mm_counter(mm, MM_FILEPAGES); |
| 1239 | } |
| 1240 | set_pte_at(mm, address, pte, |
| 1241 | swp_entry_to_pte(make_hwpoison_entry(page))); |
| 1242 | } else if (PageAnon(page)) { |
| 1243 | swp_entry_t entry = { .val = page_private(page) }; |
| 1244 | |
| 1245 | if (PageSwapCache(page)) { |
| 1246 | /* |
| 1247 | * Store the swap location in the pte. |
| 1248 | * See handle_pte_fault() ... |
| 1249 | */ |
| 1250 | if (swap_duplicate(entry) < 0) { |
| 1251 | set_pte_at(mm, address, pte, pteval); |
| 1252 | ret = SWAP_FAIL; |
| 1253 | goto out_unmap; |
| 1254 | } |
| 1255 | if (list_empty(&mm->mmlist)) { |
| 1256 | spin_lock(&mmlist_lock); |
| 1257 | if (list_empty(&mm->mmlist)) |
| 1258 | list_add(&mm->mmlist, &init_mm.mmlist); |
| 1259 | spin_unlock(&mmlist_lock); |
| 1260 | } |
| 1261 | dec_mm_counter(mm, MM_ANONPAGES); |
| 1262 | inc_mm_counter(mm, MM_SWAPENTS); |
| 1263 | } else if (IS_ENABLED(CONFIG_MIGRATION)) { |
| 1264 | /* |
| 1265 | * Store the pfn of the page in a special migration |
| 1266 | * pte. do_swap_page() will wait until the migration |
| 1267 | * pte is removed and then restart fault handling. |
| 1268 | */ |
| 1269 | BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION); |
| 1270 | entry = make_migration_entry(page, pte_write(pteval)); |
| 1271 | } |
| 1272 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); |
| 1273 | BUG_ON(pte_file(*pte)); |
| 1274 | } else if (IS_ENABLED(CONFIG_MIGRATION) && |
| 1275 | (TTU_ACTION(flags) == TTU_MIGRATION)) { |
| 1276 | /* Establish migration entry for a file page */ |
| 1277 | swp_entry_t entry; |
| 1278 | entry = make_migration_entry(page, pte_write(pteval)); |
| 1279 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); |
| 1280 | } else |
| 1281 | dec_mm_counter(mm, MM_FILEPAGES); |
| 1282 | |
| 1283 | page_remove_rmap(page); |
| 1284 | page_cache_release(page); |
| 1285 | |
| 1286 | out_unmap: |
| 1287 | pte_unmap_unlock(pte, ptl); |
| 1288 | if (ret != SWAP_FAIL) |
| 1289 | mmu_notifier_invalidate_page(mm, address); |
| 1290 | out: |
| 1291 | return ret; |
| 1292 | |
| 1293 | out_mlock: |
| 1294 | pte_unmap_unlock(pte, ptl); |
| 1295 | |
| 1296 | |
| 1297 | /* |
| 1298 | * We need mmap_sem locking, Otherwise VM_LOCKED check makes |
| 1299 | * unstable result and race. Plus, We can't wait here because |
| 1300 | * we now hold anon_vma->rwsem or mapping->i_mmap_mutex. |
| 1301 | * if trylock failed, the page remain in evictable lru and later |
| 1302 | * vmscan could retry to move the page to unevictable lru if the |
| 1303 | * page is actually mlocked. |
| 1304 | */ |
| 1305 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { |
| 1306 | if (vma->vm_flags & VM_LOCKED) { |
| 1307 | mlock_vma_page(page); |
| 1308 | ret = SWAP_MLOCK; |
| 1309 | } |
| 1310 | up_read(&vma->vm_mm->mmap_sem); |
| 1311 | } |
| 1312 | return ret; |
| 1313 | } |
| 1314 | |
| 1315 | /* |
| 1316 | * objrmap doesn't work for nonlinear VMAs because the assumption that |
| 1317 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. |
| 1318 | * Consequently, given a particular page and its ->index, we cannot locate the |
| 1319 | * ptes which are mapping that page without an exhaustive linear search. |
| 1320 | * |
| 1321 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which |
| 1322 | * maps the file to which the target page belongs. The ->vm_private_data field |
| 1323 | * holds the current cursor into that scan. Successive searches will circulate |
| 1324 | * around the vma's virtual address space. |
| 1325 | * |
| 1326 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, |
| 1327 | * more scanning pressure is placed against them as well. Eventually pages |
| 1328 | * will become fully unmapped and are eligible for eviction. |
| 1329 | * |
| 1330 | * For very sparsely populated VMAs this is a little inefficient - chances are |
| 1331 | * there there won't be many ptes located within the scan cluster. In this case |
| 1332 | * maybe we could scan further - to the end of the pte page, perhaps. |
| 1333 | * |
| 1334 | * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can |
| 1335 | * acquire it without blocking. If vma locked, mlock the pages in the cluster, |
| 1336 | * rather than unmapping them. If we encounter the "check_page" that vmscan is |
| 1337 | * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN. |
| 1338 | */ |
| 1339 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) |
| 1340 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) |
| 1341 | |
| 1342 | static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, |
| 1343 | struct vm_area_struct *vma, struct page *check_page) |
| 1344 | { |
| 1345 | struct mm_struct *mm = vma->vm_mm; |
| 1346 | pmd_t *pmd; |
| 1347 | pte_t *pte; |
| 1348 | pte_t pteval; |
| 1349 | spinlock_t *ptl; |
| 1350 | struct page *page; |
| 1351 | unsigned long address; |
| 1352 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 1353 | unsigned long mmun_end; /* For mmu_notifiers */ |
| 1354 | unsigned long end; |
| 1355 | int ret = SWAP_AGAIN; |
| 1356 | int locked_vma = 0; |
| 1357 | |
| 1358 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
| 1359 | end = address + CLUSTER_SIZE; |
| 1360 | if (address < vma->vm_start) |
| 1361 | address = vma->vm_start; |
| 1362 | if (end > vma->vm_end) |
| 1363 | end = vma->vm_end; |
| 1364 | |
| 1365 | pmd = mm_find_pmd(mm, address); |
| 1366 | if (!pmd) |
| 1367 | return ret; |
| 1368 | |
| 1369 | mmun_start = address; |
| 1370 | mmun_end = end; |
| 1371 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
| 1372 | |
| 1373 | /* |
| 1374 | * If we can acquire the mmap_sem for read, and vma is VM_LOCKED, |
| 1375 | * keep the sem while scanning the cluster for mlocking pages. |
| 1376 | */ |
| 1377 | if (down_read_trylock(&vma->vm_mm->mmap_sem)) { |
| 1378 | locked_vma = (vma->vm_flags & VM_LOCKED); |
| 1379 | if (!locked_vma) |
| 1380 | up_read(&vma->vm_mm->mmap_sem); /* don't need it */ |
| 1381 | } |
| 1382 | |
| 1383 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
| 1384 | |
| 1385 | /* Update high watermark before we lower rss */ |
| 1386 | update_hiwater_rss(mm); |
| 1387 | |
| 1388 | for (; address < end; pte++, address += PAGE_SIZE) { |
| 1389 | if (!pte_present(*pte)) |
| 1390 | continue; |
| 1391 | page = vm_normal_page(vma, address, *pte); |
| 1392 | BUG_ON(!page || PageAnon(page)); |
| 1393 | |
| 1394 | if (locked_vma) { |
| 1395 | if (page == check_page) { |
| 1396 | /* we know we have check_page locked */ |
| 1397 | mlock_vma_page(page); |
| 1398 | ret = SWAP_MLOCK; |
| 1399 | } else if (trylock_page(page)) { |
| 1400 | /* |
| 1401 | * If we can lock the page, perform mlock. |
| 1402 | * Otherwise leave the page alone, it will be |
| 1403 | * eventually encountered again later. |
| 1404 | */ |
| 1405 | mlock_vma_page(page); |
| 1406 | unlock_page(page); |
| 1407 | } |
| 1408 | continue; /* don't unmap */ |
| 1409 | } |
| 1410 | |
| 1411 | if (ptep_clear_flush_young_notify(vma, address, pte)) |
| 1412 | continue; |
| 1413 | |
| 1414 | /* Nuke the page table entry. */ |
| 1415 | flush_cache_page(vma, address, pte_pfn(*pte)); |
| 1416 | pteval = ptep_clear_flush(vma, address, pte); |
| 1417 | |
| 1418 | /* If nonlinear, store the file page offset in the pte. */ |
| 1419 | if (page->index != linear_page_index(vma, address)) |
| 1420 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); |
| 1421 | |
| 1422 | /* Move the dirty bit to the physical page now the pte is gone. */ |
| 1423 | if (pte_dirty(pteval)) |
| 1424 | set_page_dirty(page); |
| 1425 | |
| 1426 | page_remove_rmap(page); |
| 1427 | page_cache_release(page); |
| 1428 | dec_mm_counter(mm, MM_FILEPAGES); |
| 1429 | (*mapcount)--; |
| 1430 | } |
| 1431 | pte_unmap_unlock(pte - 1, ptl); |
| 1432 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
| 1433 | if (locked_vma) |
| 1434 | up_read(&vma->vm_mm->mmap_sem); |
| 1435 | return ret; |
| 1436 | } |
| 1437 | |
| 1438 | bool is_vma_temporary_stack(struct vm_area_struct *vma) |
| 1439 | { |
| 1440 | int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); |
| 1441 | |
| 1442 | if (!maybe_stack) |
| 1443 | return false; |
| 1444 | |
| 1445 | if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == |
| 1446 | VM_STACK_INCOMPLETE_SETUP) |
| 1447 | return true; |
| 1448 | |
| 1449 | return false; |
| 1450 | } |
| 1451 | |
| 1452 | /** |
| 1453 | * try_to_unmap_anon - unmap or unlock anonymous page using the object-based |
| 1454 | * rmap method |
| 1455 | * @page: the page to unmap/unlock |
| 1456 | * @flags: action and flags |
| 1457 | * |
| 1458 | * Find all the mappings of a page using the mapping pointer and the vma chains |
| 1459 | * contained in the anon_vma struct it points to. |
| 1460 | * |
| 1461 | * This function is only called from try_to_unmap/try_to_munlock for |
| 1462 | * anonymous pages. |
| 1463 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma |
| 1464 | * where the page was found will be held for write. So, we won't recheck |
| 1465 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be |
| 1466 | * 'LOCKED. |
| 1467 | */ |
| 1468 | static int try_to_unmap_anon(struct page *page, enum ttu_flags flags) |
| 1469 | { |
| 1470 | struct anon_vma *anon_vma; |
| 1471 | pgoff_t pgoff; |
| 1472 | struct anon_vma_chain *avc; |
| 1473 | int ret = SWAP_AGAIN; |
| 1474 | |
| 1475 | anon_vma = page_lock_anon_vma_read(page); |
| 1476 | if (!anon_vma) |
| 1477 | return ret; |
| 1478 | |
| 1479 | pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 1480 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
| 1481 | struct vm_area_struct *vma = avc->vma; |
| 1482 | unsigned long address; |
| 1483 | |
| 1484 | /* |
| 1485 | * During exec, a temporary VMA is setup and later moved. |
| 1486 | * The VMA is moved under the anon_vma lock but not the |
| 1487 | * page tables leading to a race where migration cannot |
| 1488 | * find the migration ptes. Rather than increasing the |
| 1489 | * locking requirements of exec(), migration skips |
| 1490 | * temporary VMAs until after exec() completes. |
| 1491 | */ |
| 1492 | if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) && |
| 1493 | is_vma_temporary_stack(vma)) |
| 1494 | continue; |
| 1495 | |
| 1496 | address = vma_address(page, vma); |
| 1497 | ret = try_to_unmap_one(page, vma, address, flags); |
| 1498 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
| 1499 | break; |
| 1500 | } |
| 1501 | |
| 1502 | page_unlock_anon_vma_read(anon_vma); |
| 1503 | return ret; |
| 1504 | } |
| 1505 | |
| 1506 | /** |
| 1507 | * try_to_unmap_file - unmap/unlock file page using the object-based rmap method |
| 1508 | * @page: the page to unmap/unlock |
| 1509 | * @flags: action and flags |
| 1510 | * |
| 1511 | * Find all the mappings of a page using the mapping pointer and the vma chains |
| 1512 | * contained in the address_space struct it points to. |
| 1513 | * |
| 1514 | * This function is only called from try_to_unmap/try_to_munlock for |
| 1515 | * object-based pages. |
| 1516 | * When called from try_to_munlock(), the mmap_sem of the mm containing the vma |
| 1517 | * where the page was found will be held for write. So, we won't recheck |
| 1518 | * vm_flags for that VMA. That should be OK, because that vma shouldn't be |
| 1519 | * 'LOCKED. |
| 1520 | */ |
| 1521 | static int try_to_unmap_file(struct page *page, enum ttu_flags flags) |
| 1522 | { |
| 1523 | struct address_space *mapping = page->mapping; |
| 1524 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 1525 | struct vm_area_struct *vma; |
| 1526 | int ret = SWAP_AGAIN; |
| 1527 | unsigned long cursor; |
| 1528 | unsigned long max_nl_cursor = 0; |
| 1529 | unsigned long max_nl_size = 0; |
| 1530 | unsigned int mapcount; |
| 1531 | |
| 1532 | if (PageHuge(page)) |
| 1533 | pgoff = page->index << compound_order(page); |
| 1534 | |
| 1535 | mutex_lock(&mapping->i_mmap_mutex); |
| 1536 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { |
| 1537 | unsigned long address = vma_address(page, vma); |
| 1538 | ret = try_to_unmap_one(page, vma, address, flags); |
| 1539 | if (ret != SWAP_AGAIN || !page_mapped(page)) |
| 1540 | goto out; |
| 1541 | } |
| 1542 | |
| 1543 | if (list_empty(&mapping->i_mmap_nonlinear)) |
| 1544 | goto out; |
| 1545 | |
| 1546 | /* |
| 1547 | * We don't bother to try to find the munlocked page in nonlinears. |
| 1548 | * It's costly. Instead, later, page reclaim logic may call |
| 1549 | * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily. |
| 1550 | */ |
| 1551 | if (TTU_ACTION(flags) == TTU_MUNLOCK) |
| 1552 | goto out; |
| 1553 | |
| 1554 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
| 1555 | shared.nonlinear) { |
| 1556 | cursor = (unsigned long) vma->vm_private_data; |
| 1557 | if (cursor > max_nl_cursor) |
| 1558 | max_nl_cursor = cursor; |
| 1559 | cursor = vma->vm_end - vma->vm_start; |
| 1560 | if (cursor > max_nl_size) |
| 1561 | max_nl_size = cursor; |
| 1562 | } |
| 1563 | |
| 1564 | if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ |
| 1565 | ret = SWAP_FAIL; |
| 1566 | goto out; |
| 1567 | } |
| 1568 | |
| 1569 | /* |
| 1570 | * We don't try to search for this page in the nonlinear vmas, |
| 1571 | * and page_referenced wouldn't have found it anyway. Instead |
| 1572 | * just walk the nonlinear vmas trying to age and unmap some. |
| 1573 | * The mapcount of the page we came in with is irrelevant, |
| 1574 | * but even so use it as a guide to how hard we should try? |
| 1575 | */ |
| 1576 | mapcount = page_mapcount(page); |
| 1577 | if (!mapcount) |
| 1578 | goto out; |
| 1579 | cond_resched(); |
| 1580 | |
| 1581 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; |
| 1582 | if (max_nl_cursor == 0) |
| 1583 | max_nl_cursor = CLUSTER_SIZE; |
| 1584 | |
| 1585 | do { |
| 1586 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
| 1587 | shared.nonlinear) { |
| 1588 | cursor = (unsigned long) vma->vm_private_data; |
| 1589 | while ( cursor < max_nl_cursor && |
| 1590 | cursor < vma->vm_end - vma->vm_start) { |
| 1591 | if (try_to_unmap_cluster(cursor, &mapcount, |
| 1592 | vma, page) == SWAP_MLOCK) |
| 1593 | ret = SWAP_MLOCK; |
| 1594 | cursor += CLUSTER_SIZE; |
| 1595 | vma->vm_private_data = (void *) cursor; |
| 1596 | if ((int)mapcount <= 0) |
| 1597 | goto out; |
| 1598 | } |
| 1599 | vma->vm_private_data = (void *) max_nl_cursor; |
| 1600 | } |
| 1601 | cond_resched(); |
| 1602 | max_nl_cursor += CLUSTER_SIZE; |
| 1603 | } while (max_nl_cursor <= max_nl_size); |
| 1604 | |
| 1605 | /* |
| 1606 | * Don't loop forever (perhaps all the remaining pages are |
| 1607 | * in locked vmas). Reset cursor on all unreserved nonlinear |
| 1608 | * vmas, now forgetting on which ones it had fallen behind. |
| 1609 | */ |
| 1610 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear) |
| 1611 | vma->vm_private_data = NULL; |
| 1612 | out: |
| 1613 | mutex_unlock(&mapping->i_mmap_mutex); |
| 1614 | return ret; |
| 1615 | } |
| 1616 | |
| 1617 | /** |
| 1618 | * try_to_unmap - try to remove all page table mappings to a page |
| 1619 | * @page: the page to get unmapped |
| 1620 | * @flags: action and flags |
| 1621 | * |
| 1622 | * Tries to remove all the page table entries which are mapping this |
| 1623 | * page, used in the pageout path. Caller must hold the page lock. |
| 1624 | * Return values are: |
| 1625 | * |
| 1626 | * SWAP_SUCCESS - we succeeded in removing all mappings |
| 1627 | * SWAP_AGAIN - we missed a mapping, try again later |
| 1628 | * SWAP_FAIL - the page is unswappable |
| 1629 | * SWAP_MLOCK - page is mlocked. |
| 1630 | */ |
| 1631 | int try_to_unmap(struct page *page, enum ttu_flags flags) |
| 1632 | { |
| 1633 | int ret; |
| 1634 | |
| 1635 | BUG_ON(!PageLocked(page)); |
| 1636 | VM_BUG_ON(!PageHuge(page) && PageTransHuge(page)); |
| 1637 | |
| 1638 | if (unlikely(PageKsm(page))) |
| 1639 | ret = try_to_unmap_ksm(page, flags); |
| 1640 | else if (PageAnon(page)) |
| 1641 | ret = try_to_unmap_anon(page, flags); |
| 1642 | else |
| 1643 | ret = try_to_unmap_file(page, flags); |
| 1644 | if (ret != SWAP_MLOCK && !page_mapped(page)) |
| 1645 | ret = SWAP_SUCCESS; |
| 1646 | return ret; |
| 1647 | } |
| 1648 | |
| 1649 | /** |
| 1650 | * try_to_munlock - try to munlock a page |
| 1651 | * @page: the page to be munlocked |
| 1652 | * |
| 1653 | * Called from munlock code. Checks all of the VMAs mapping the page |
| 1654 | * to make sure nobody else has this page mlocked. The page will be |
| 1655 | * returned with PG_mlocked cleared if no other vmas have it mlocked. |
| 1656 | * |
| 1657 | * Return values are: |
| 1658 | * |
| 1659 | * SWAP_AGAIN - no vma is holding page mlocked, or, |
| 1660 | * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem |
| 1661 | * SWAP_FAIL - page cannot be located at present |
| 1662 | * SWAP_MLOCK - page is now mlocked. |
| 1663 | */ |
| 1664 | int try_to_munlock(struct page *page) |
| 1665 | { |
| 1666 | VM_BUG_ON(!PageLocked(page) || PageLRU(page)); |
| 1667 | |
| 1668 | if (unlikely(PageKsm(page))) |
| 1669 | return try_to_unmap_ksm(page, TTU_MUNLOCK); |
| 1670 | else if (PageAnon(page)) |
| 1671 | return try_to_unmap_anon(page, TTU_MUNLOCK); |
| 1672 | else |
| 1673 | return try_to_unmap_file(page, TTU_MUNLOCK); |
| 1674 | } |
| 1675 | |
| 1676 | void __put_anon_vma(struct anon_vma *anon_vma) |
| 1677 | { |
| 1678 | struct anon_vma *root = anon_vma->root; |
| 1679 | |
| 1680 | anon_vma_free(anon_vma); |
| 1681 | if (root != anon_vma && atomic_dec_and_test(&root->refcount)) |
| 1682 | anon_vma_free(root); |
| 1683 | } |
| 1684 | |
| 1685 | #ifdef CONFIG_MIGRATION |
| 1686 | /* |
| 1687 | * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file(): |
| 1688 | * Called by migrate.c to remove migration ptes, but might be used more later. |
| 1689 | */ |
| 1690 | static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *, |
| 1691 | struct vm_area_struct *, unsigned long, void *), void *arg) |
| 1692 | { |
| 1693 | struct anon_vma *anon_vma; |
| 1694 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 1695 | struct anon_vma_chain *avc; |
| 1696 | int ret = SWAP_AGAIN; |
| 1697 | |
| 1698 | /* |
| 1699 | * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() |
| 1700 | * because that depends on page_mapped(); but not all its usages |
| 1701 | * are holding mmap_sem. Users without mmap_sem are required to |
| 1702 | * take a reference count to prevent the anon_vma disappearing |
| 1703 | */ |
| 1704 | anon_vma = page_anon_vma(page); |
| 1705 | if (!anon_vma) |
| 1706 | return ret; |
| 1707 | anon_vma_lock_read(anon_vma); |
| 1708 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
| 1709 | struct vm_area_struct *vma = avc->vma; |
| 1710 | unsigned long address = vma_address(page, vma); |
| 1711 | ret = rmap_one(page, vma, address, arg); |
| 1712 | if (ret != SWAP_AGAIN) |
| 1713 | break; |
| 1714 | } |
| 1715 | anon_vma_unlock_read(anon_vma); |
| 1716 | return ret; |
| 1717 | } |
| 1718 | |
| 1719 | static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *, |
| 1720 | struct vm_area_struct *, unsigned long, void *), void *arg) |
| 1721 | { |
| 1722 | struct address_space *mapping = page->mapping; |
| 1723 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 1724 | struct vm_area_struct *vma; |
| 1725 | int ret = SWAP_AGAIN; |
| 1726 | |
| 1727 | if (!mapping) |
| 1728 | return ret; |
| 1729 | mutex_lock(&mapping->i_mmap_mutex); |
| 1730 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { |
| 1731 | unsigned long address = vma_address(page, vma); |
| 1732 | ret = rmap_one(page, vma, address, arg); |
| 1733 | if (ret != SWAP_AGAIN) |
| 1734 | break; |
| 1735 | } |
| 1736 | /* |
| 1737 | * No nonlinear handling: being always shared, nonlinear vmas |
| 1738 | * never contain migration ptes. Decide what to do about this |
| 1739 | * limitation to linear when we need rmap_walk() on nonlinear. |
| 1740 | */ |
| 1741 | mutex_unlock(&mapping->i_mmap_mutex); |
| 1742 | return ret; |
| 1743 | } |
| 1744 | |
| 1745 | int rmap_walk(struct page *page, int (*rmap_one)(struct page *, |
| 1746 | struct vm_area_struct *, unsigned long, void *), void *arg) |
| 1747 | { |
| 1748 | VM_BUG_ON(!PageLocked(page)); |
| 1749 | |
| 1750 | if (unlikely(PageKsm(page))) |
| 1751 | return rmap_walk_ksm(page, rmap_one, arg); |
| 1752 | else if (PageAnon(page)) |
| 1753 | return rmap_walk_anon(page, rmap_one, arg); |
| 1754 | else |
| 1755 | return rmap_walk_file(page, rmap_one, arg); |
| 1756 | } |
| 1757 | #endif /* CONFIG_MIGRATION */ |
| 1758 | |
| 1759 | #ifdef CONFIG_HUGETLB_PAGE |
| 1760 | /* |
| 1761 | * The following three functions are for anonymous (private mapped) hugepages. |
| 1762 | * Unlike common anonymous pages, anonymous hugepages have no accounting code |
| 1763 | * and no lru code, because we handle hugepages differently from common pages. |
| 1764 | */ |
| 1765 | static void __hugepage_set_anon_rmap(struct page *page, |
| 1766 | struct vm_area_struct *vma, unsigned long address, int exclusive) |
| 1767 | { |
| 1768 | struct anon_vma *anon_vma = vma->anon_vma; |
| 1769 | |
| 1770 | BUG_ON(!anon_vma); |
| 1771 | |
| 1772 | if (PageAnon(page)) |
| 1773 | return; |
| 1774 | if (!exclusive) |
| 1775 | anon_vma = anon_vma->root; |
| 1776 | |
| 1777 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
| 1778 | page->mapping = (struct address_space *) anon_vma; |
| 1779 | page->index = linear_page_index(vma, address); |
| 1780 | } |
| 1781 | |
| 1782 | void hugepage_add_anon_rmap(struct page *page, |
| 1783 | struct vm_area_struct *vma, unsigned long address) |
| 1784 | { |
| 1785 | struct anon_vma *anon_vma = vma->anon_vma; |
| 1786 | int first; |
| 1787 | |
| 1788 | BUG_ON(!PageLocked(page)); |
| 1789 | BUG_ON(!anon_vma); |
| 1790 | /* address might be in next vma when migration races vma_adjust */ |
| 1791 | first = atomic_inc_and_test(&page->_mapcount); |
| 1792 | if (first) |
| 1793 | __hugepage_set_anon_rmap(page, vma, address, 0); |
| 1794 | } |
| 1795 | |
| 1796 | void hugepage_add_new_anon_rmap(struct page *page, |
| 1797 | struct vm_area_struct *vma, unsigned long address) |
| 1798 | { |
| 1799 | BUG_ON(address < vma->vm_start || address >= vma->vm_end); |
| 1800 | atomic_set(&page->_mapcount, 0); |
| 1801 | __hugepage_set_anon_rmap(page, vma, address, 1); |
| 1802 | } |
| 1803 | #endif /* CONFIG_HUGETLB_PAGE */ |