linux/mm/rmap.c
<<
>>
Prefs
   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->mutex
  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->mutex,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
  65static struct kmem_cache *anon_vma_cachep;
  66static struct kmem_cache *anon_vma_chain_cachep;
  67
  68static 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
  85static 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() 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         * mutex_trylock() from page_lock_anon_vma(). This orders:
  97         *
  98         * page_lock_anon_vma()         VS      put_anon_vma()
  99         *   mutex_trylock()                      atomic_dec_and_test()
 100         *   LOCK                                 MB
 101         *   atomic_read()                        mutex_is_locked()
 102         *
 103         * LOCK should suffice since the actual taking of the lock must
 104         * happen _before_ what follows.
 105         */
 106        if (mutex_is_locked(&anon_vma->root->mutex)) {
 107                anon_vma_lock(anon_vma);
 108                anon_vma_unlock(anon_vma);
 109        }
 110
 111        kmem_cache_free(anon_vma_cachep, anon_vma);
 112}
 113
 114static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 115{
 116        return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 117}
 118
 119static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 120{
 121        kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 122}
 123
 124static void anon_vma_chain_link(struct vm_area_struct *vma,
 125                                struct anon_vma_chain *avc,
 126                                struct anon_vma *anon_vma)
 127{
 128        avc->vma = vma;
 129        avc->anon_vma = anon_vma;
 130        list_add(&avc->same_vma, &vma->anon_vma_chain);
 131
 132        /*
 133         * It's critical to add new vmas to the tail of the anon_vma,
 134         * see comment in huge_memory.c:__split_huge_page().
 135         */
 136        list_add_tail(&avc->same_anon_vma, &anon_vma->head);
 137}
 138
 139/**
 140 * anon_vma_prepare - attach an anon_vma to a memory region
 141 * @vma: the memory region in question
 142 *
 143 * This makes sure the memory mapping described by 'vma' has
 144 * an 'anon_vma' attached to it, so that we can associate the
 145 * anonymous pages mapped into it with that anon_vma.
 146 *
 147 * The common case will be that we already have one, but if
 148 * not we either need to find an adjacent mapping that we
 149 * can re-use the anon_vma from (very common when the only
 150 * reason for splitting a vma has been mprotect()), or we
 151 * allocate a new one.
 152 *
 153 * Anon-vma allocations are very subtle, because we may have
 154 * optimistically looked up an anon_vma in page_lock_anon_vma()
 155 * and that may actually touch the spinlock even in the newly
 156 * allocated vma (it depends on RCU to make sure that the
 157 * anon_vma isn't actually destroyed).
 158 *
 159 * As a result, we need to do proper anon_vma locking even
 160 * for the new allocation. At the same time, we do not want
 161 * to do any locking for the common case of already having
 162 * an anon_vma.
 163 *
 164 * This must be called with the mmap_sem held for reading.
 165 */
 166int anon_vma_prepare(struct vm_area_struct *vma)
 167{
 168        struct anon_vma *anon_vma = vma->anon_vma;
 169        struct anon_vma_chain *avc;
 170
 171        might_sleep();
 172        if (unlikely(!anon_vma)) {
 173                struct mm_struct *mm = vma->vm_mm;
 174                struct anon_vma *allocated;
 175
 176                avc = anon_vma_chain_alloc(GFP_KERNEL);
 177                if (!avc)
 178                        goto out_enomem;
 179
 180                anon_vma = find_mergeable_anon_vma(vma);
 181                allocated = NULL;
 182                if (!anon_vma) {
 183                        anon_vma = anon_vma_alloc();
 184                        if (unlikely(!anon_vma))
 185                                goto out_enomem_free_avc;
 186                        allocated = anon_vma;
 187                }
 188
 189                anon_vma_lock(anon_vma);
 190                /* page_table_lock to protect against threads */
 191                spin_lock(&mm->page_table_lock);
 192                if (likely(!vma->anon_vma)) {
 193                        vma->anon_vma = anon_vma;
 194                        anon_vma_chain_link(vma, avc, anon_vma);
 195                        allocated = NULL;
 196                        avc = NULL;
 197                }
 198                spin_unlock(&mm->page_table_lock);
 199                anon_vma_unlock(anon_vma);
 200
 201                if (unlikely(allocated))
 202                        put_anon_vma(allocated);
 203                if (unlikely(avc))
 204                        anon_vma_chain_free(avc);
 205        }
 206        return 0;
 207
 208 out_enomem_free_avc:
 209        anon_vma_chain_free(avc);
 210 out_enomem:
 211        return -ENOMEM;
 212}
 213
 214/*
 215 * This is a useful helper function for locking the anon_vma root as
 216 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 217 * have the same vma.
 218 *
 219 * Such anon_vma's should have the same root, so you'd expect to see
 220 * just a single mutex_lock for the whole traversal.
 221 */
 222static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 223{
 224        struct anon_vma *new_root = anon_vma->root;
 225        if (new_root != root) {
 226                if (WARN_ON_ONCE(root))
 227                        mutex_unlock(&root->mutex);
 228                root = new_root;
 229                mutex_lock(&root->mutex);
 230        }
 231        return root;
 232}
 233
 234static inline void unlock_anon_vma_root(struct anon_vma *root)
 235{
 236        if (root)
 237                mutex_unlock(&root->mutex);
 238}
 239
 240/*
 241 * Attach the anon_vmas from src to dst.
 242 * Returns 0 on success, -ENOMEM on failure.
 243 */
 244int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 245{
 246        struct anon_vma_chain *avc, *pavc;
 247        struct anon_vma *root = NULL;
 248
 249        list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 250                struct anon_vma *anon_vma;
 251
 252                avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 253                if (unlikely(!avc)) {
 254                        unlock_anon_vma_root(root);
 255                        root = NULL;
 256                        avc = anon_vma_chain_alloc(GFP_KERNEL);
 257                        if (!avc)
 258                                goto enomem_failure;
 259                }
 260                anon_vma = pavc->anon_vma;
 261                root = lock_anon_vma_root(root, anon_vma);
 262                anon_vma_chain_link(dst, avc, anon_vma);
 263        }
 264        unlock_anon_vma_root(root);
 265        return 0;
 266
 267 enomem_failure:
 268        unlink_anon_vmas(dst);
 269        return -ENOMEM;
 270}
 271
 272/*
 273 * Some rmap walk that needs to find all ptes/hugepmds without false
 274 * negatives (like migrate and split_huge_page) running concurrent
 275 * with operations that copy or move pagetables (like mremap() and
 276 * fork()) to be safe. They depend on the anon_vma "same_anon_vma"
 277 * list to be in a certain order: the dst_vma must be placed after the
 278 * src_vma in the list. This is always guaranteed by fork() but
 279 * mremap() needs to call this function to enforce it in case the
 280 * dst_vma isn't newly allocated and chained with the anon_vma_clone()
 281 * function but just an extension of a pre-existing vma through
 282 * vma_merge.
 283 *
 284 * NOTE: the same_anon_vma list can still be changed by other
 285 * processes while mremap runs because mremap doesn't hold the
 286 * anon_vma mutex to prevent modifications to the list while it
 287 * runs. All we need to enforce is that the relative order of this
 288 * process vmas isn't changing (we don't care about other vmas
 289 * order). Each vma corresponds to an anon_vma_chain structure so
 290 * there's no risk that other processes calling anon_vma_moveto_tail()
 291 * and changing the same_anon_vma list under mremap() will screw with
 292 * the relative order of this process vmas in the list, because we
 293 * they can't alter the order of any vma that belongs to this
 294 * process. And there can't be another anon_vma_moveto_tail() running
 295 * concurrently with mremap() coming from this process because we hold
 296 * the mmap_sem for the whole mremap(). fork() ordering dependency
 297 * also shouldn't be affected because fork() only cares that the
 298 * parent vmas are placed in the list before the child vmas and
 299 * anon_vma_moveto_tail() won't reorder vmas from either the fork()
 300 * parent or child.
 301 */
 302void anon_vma_moveto_tail(struct vm_area_struct *dst)
 303{
 304        struct anon_vma_chain *pavc;
 305        struct anon_vma *root = NULL;
 306
 307        list_for_each_entry_reverse(pavc, &dst->anon_vma_chain, same_vma) {
 308                struct anon_vma *anon_vma = pavc->anon_vma;
 309                VM_BUG_ON(pavc->vma != dst);
 310                root = lock_anon_vma_root(root, anon_vma);
 311                list_del(&pavc->same_anon_vma);
 312                list_add_tail(&pavc->same_anon_vma, &anon_vma->head);
 313        }
 314        unlock_anon_vma_root(root);
 315}
 316
 317/*
 318 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 319 * the corresponding VMA in the parent process is attached to.
 320 * Returns 0 on success, non-zero on failure.
 321 */
 322int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 323{
 324        struct anon_vma_chain *avc;
 325        struct anon_vma *anon_vma;
 326
 327        /* Don't bother if the parent process has no anon_vma here. */
 328        if (!pvma->anon_vma)
 329                return 0;
 330
 331        /*
 332         * First, attach the new VMA to the parent VMA's anon_vmas,
 333         * so rmap can find non-COWed pages in child processes.
 334         */
 335        if (anon_vma_clone(vma, pvma))
 336                return -ENOMEM;
 337
 338        /* Then add our own anon_vma. */
 339        anon_vma = anon_vma_alloc();
 340        if (!anon_vma)
 341                goto out_error;
 342        avc = anon_vma_chain_alloc(GFP_KERNEL);
 343        if (!avc)
 344                goto out_error_free_anon_vma;
 345
 346        /*
 347         * The root anon_vma's spinlock is the lock actually used when we
 348         * lock any of the anon_vmas in this anon_vma tree.
 349         */
 350        anon_vma->root = pvma->anon_vma->root;
 351        /*
 352         * With refcounts, an anon_vma can stay around longer than the
 353         * process it belongs to. The root anon_vma needs to be pinned until
 354         * this anon_vma is freed, because the lock lives in the root.
 355         */
 356        get_anon_vma(anon_vma->root);
 357        /* Mark this anon_vma as the one where our new (COWed) pages go. */
 358        vma->anon_vma = anon_vma;
 359        anon_vma_lock(anon_vma);
 360        anon_vma_chain_link(vma, avc, anon_vma);
 361        anon_vma_unlock(anon_vma);
 362
 363        return 0;
 364
 365 out_error_free_anon_vma:
 366        put_anon_vma(anon_vma);
 367 out_error:
 368        unlink_anon_vmas(vma);
 369        return -ENOMEM;
 370}
 371
 372void unlink_anon_vmas(struct vm_area_struct *vma)
 373{
 374        struct anon_vma_chain *avc, *next;
 375        struct anon_vma *root = NULL;
 376
 377        /*
 378         * Unlink each anon_vma chained to the VMA.  This list is ordered
 379         * from newest to oldest, ensuring the root anon_vma gets freed last.
 380         */
 381        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 382                struct anon_vma *anon_vma = avc->anon_vma;
 383
 384                root = lock_anon_vma_root(root, anon_vma);
 385                list_del(&avc->same_anon_vma);
 386
 387                /*
 388                 * Leave empty anon_vmas on the list - we'll need
 389                 * to free them outside the lock.
 390                 */
 391                if (list_empty(&anon_vma->head))
 392                        continue;
 393
 394                list_del(&avc->same_vma);
 395                anon_vma_chain_free(avc);
 396        }
 397        unlock_anon_vma_root(root);
 398
 399        /*
 400         * Iterate the list once more, it now only contains empty and unlinked
 401         * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 402         * needing to acquire the anon_vma->root->mutex.
 403         */
 404        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 405                struct anon_vma *anon_vma = avc->anon_vma;
 406
 407                put_anon_vma(anon_vma);
 408
 409                list_del(&avc->same_vma);
 410                anon_vma_chain_free(avc);
 411        }
 412}
 413
 414static void anon_vma_ctor(void *data)
 415{
 416        struct anon_vma *anon_vma = data;
 417
 418        mutex_init(&anon_vma->mutex);
 419        atomic_set(&anon_vma->refcount, 0);
 420        INIT_LIST_HEAD(&anon_vma->head);
 421}
 422
 423void __init anon_vma_init(void)
 424{
 425        anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 426                        0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
 427        anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
 428}
 429
 430/*
 431 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 432 *
 433 * Since there is no serialization what so ever against page_remove_rmap()
 434 * the best this function can do is return a locked anon_vma that might
 435 * have been relevant to this page.
 436 *
 437 * The page might have been remapped to a different anon_vma or the anon_vma
 438 * returned may already be freed (and even reused).
 439 *
 440 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 441 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 442 * ensure that any anon_vma obtained from the page will still be valid for as
 443 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 444 *
 445 * All users of this function must be very careful when walking the anon_vma
 446 * chain and verify that the page in question is indeed mapped in it
 447 * [ something equivalent to page_mapped_in_vma() ].
 448 *
 449 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 450 * that the anon_vma pointer from page->mapping is valid if there is a
 451 * mapcount, we can dereference the anon_vma after observing those.
 452 */
 453struct anon_vma *page_get_anon_vma(struct page *page)
 454{
 455        struct anon_vma *anon_vma = NULL;
 456        unsigned long anon_mapping;
 457
 458        rcu_read_lock();
 459        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 460        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 461                goto out;
 462        if (!page_mapped(page))
 463                goto out;
 464
 465        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 466        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 467                anon_vma = NULL;
 468                goto out;
 469        }
 470
 471        /*
 472         * If this page is still mapped, then its anon_vma cannot have been
 473         * freed.  But if it has been unmapped, we have no security against the
 474         * anon_vma structure being freed and reused (for another anon_vma:
 475         * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
 476         * above cannot corrupt).
 477         */
 478        if (!page_mapped(page)) {
 479                put_anon_vma(anon_vma);
 480                anon_vma = NULL;
 481        }
 482out:
 483        rcu_read_unlock();
 484
 485        return anon_vma;
 486}
 487
 488/*
 489 * Similar to page_get_anon_vma() except it locks the anon_vma.
 490 *
 491 * Its a little more complex as it tries to keep the fast path to a single
 492 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 493 * reference like with page_get_anon_vma() and then block on the mutex.
 494 */
 495struct anon_vma *page_lock_anon_vma(struct page *page)
 496{
 497        struct anon_vma *anon_vma = NULL;
 498        struct anon_vma *root_anon_vma;
 499        unsigned long anon_mapping;
 500
 501        rcu_read_lock();
 502        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 503        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 504                goto out;
 505        if (!page_mapped(page))
 506                goto out;
 507
 508        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 509        root_anon_vma = ACCESS_ONCE(anon_vma->root);
 510        if (mutex_trylock(&root_anon_vma->mutex)) {
 511                /*
 512                 * If the page is still mapped, then this anon_vma is still
 513                 * its anon_vma, and holding the mutex ensures that it will
 514                 * not go away, see anon_vma_free().
 515                 */
 516                if (!page_mapped(page)) {
 517                        mutex_unlock(&root_anon_vma->mutex);
 518                        anon_vma = NULL;
 519                }
 520                goto out;
 521        }
 522
 523        /* trylock failed, we got to sleep */
 524        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 525                anon_vma = NULL;
 526                goto out;
 527        }
 528
 529        if (!page_mapped(page)) {
 530                put_anon_vma(anon_vma);
 531                anon_vma = NULL;
 532                goto out;
 533        }
 534
 535        /* we pinned the anon_vma, its safe to sleep */
 536        rcu_read_unlock();
 537        anon_vma_lock(anon_vma);
 538
 539        if (atomic_dec_and_test(&anon_vma->refcount)) {
 540                /*
 541                 * Oops, we held the last refcount, release the lock
 542                 * and bail -- can't simply use put_anon_vma() because
 543                 * we'll deadlock on the anon_vma_lock() recursion.
 544                 */
 545                anon_vma_unlock(anon_vma);
 546                __put_anon_vma(anon_vma);
 547                anon_vma = NULL;
 548        }
 549
 550        return anon_vma;
 551
 552out:
 553        rcu_read_unlock();
 554        return anon_vma;
 555}
 556
 557void page_unlock_anon_vma(struct anon_vma *anon_vma)
 558{
 559        anon_vma_unlock(anon_vma);
 560}
 561
 562/*
 563 * At what user virtual address is page expected in @vma?
 564 * Returns virtual address or -EFAULT if page's index/offset is not
 565 * within the range mapped the @vma.
 566 */
 567inline unsigned long
 568vma_address(struct page *page, struct vm_area_struct *vma)
 569{
 570        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 571        unsigned long address;
 572
 573        if (unlikely(is_vm_hugetlb_page(vma)))
 574                pgoff = page->index << huge_page_order(page_hstate(page));
 575        address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 576        if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
 577                /* page should be within @vma mapping range */
 578                return -EFAULT;
 579        }
 580        return address;
 581}
 582
 583/*
 584 * At what user virtual address is page expected in vma?
 585 * Caller should check the page is actually part of the vma.
 586 */
 587unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 588{
 589        if (PageAnon(page)) {
 590                struct anon_vma *page__anon_vma = page_anon_vma(page);
 591                /*
 592                 * Note: swapoff's unuse_vma() is more efficient with this
 593                 * check, and needs it to match anon_vma when KSM is active.
 594                 */
 595                if (!vma->anon_vma || !page__anon_vma ||
 596                    vma->anon_vma->root != page__anon_vma->root)
 597                        return -EFAULT;
 598        } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
 599                if (!vma->vm_file ||
 600                    vma->vm_file->f_mapping != page->mapping)
 601                        return -EFAULT;
 602        } else
 603                return -EFAULT;
 604        return vma_address(page, vma);
 605}
 606
 607/*
 608 * Check that @page is mapped at @address into @mm.
 609 *
 610 * If @sync is false, page_check_address may perform a racy check to avoid
 611 * the page table lock when the pte is not present (helpful when reclaiming
 612 * highly shared pages).
 613 *
 614 * On success returns with pte mapped and locked.
 615 */
 616pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
 617                          unsigned long address, spinlock_t **ptlp, int sync)
 618{
 619        pgd_t *pgd;
 620        pud_t *pud;
 621        pmd_t *pmd;
 622        pte_t *pte;
 623        spinlock_t *ptl;
 624
 625        if (unlikely(PageHuge(page))) {
 626                pte = huge_pte_offset(mm, address);
 627                ptl = &mm->page_table_lock;
 628                goto check;
 629        }
 630
 631        pgd = pgd_offset(mm, address);
 632        if (!pgd_present(*pgd))
 633                return NULL;
 634
 635        pud = pud_offset(pgd, address);
 636        if (!pud_present(*pud))
 637                return NULL;
 638
 639        pmd = pmd_offset(pud, address);
 640        if (!pmd_present(*pmd))
 641                return NULL;
 642        if (pmd_trans_huge(*pmd))
 643                return NULL;
 644
 645        pte = pte_offset_map(pmd, address);
 646        /* Make a quick check before getting the lock */
 647        if (!sync && !pte_present(*pte)) {
 648                pte_unmap(pte);
 649                return NULL;
 650        }
 651
 652        ptl = pte_lockptr(mm, pmd);
 653check:
 654        spin_lock(ptl);
 655        if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 656                *ptlp = ptl;
 657                return pte;
 658        }
 659        pte_unmap_unlock(pte, ptl);
 660        return NULL;
 661}
 662
 663/**
 664 * page_mapped_in_vma - check whether a page is really mapped in a VMA
 665 * @page: the page to test
 666 * @vma: the VMA to test
 667 *
 668 * Returns 1 if the page is mapped into the page tables of the VMA, 0
 669 * if the page is not mapped into the page tables of this VMA.  Only
 670 * valid for normal file or anonymous VMAs.
 671 */
 672int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 673{
 674        unsigned long address;
 675        pte_t *pte;
 676        spinlock_t *ptl;
 677
 678        address = vma_address(page, vma);
 679        if (address == -EFAULT)         /* out of vma range */
 680                return 0;
 681        pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
 682        if (!pte)                       /* the page is not in this mm */
 683                return 0;
 684        pte_unmap_unlock(pte, ptl);
 685
 686        return 1;
 687}
 688
 689/*
 690 * Subfunctions of page_referenced: page_referenced_one called
 691 * repeatedly from either page_referenced_anon or page_referenced_file.
 692 */
 693int page_referenced_one(struct page *page, struct vm_area_struct *vma,
 694                        unsigned long address, unsigned int *mapcount,
 695                        unsigned long *vm_flags)
 696{
 697        struct mm_struct *mm = vma->vm_mm;
 698        int referenced = 0;
 699
 700        if (unlikely(PageTransHuge(page))) {
 701                pmd_t *pmd;
 702
 703                spin_lock(&mm->page_table_lock);
 704                /*
 705                 * rmap might return false positives; we must filter
 706                 * these out using page_check_address_pmd().
 707                 */
 708                pmd = page_check_address_pmd(page, mm, address,
 709                                             PAGE_CHECK_ADDRESS_PMD_FLAG);
 710                if (!pmd) {
 711                        spin_unlock(&mm->page_table_lock);
 712                        goto out;
 713                }
 714
 715                if (vma->vm_flags & VM_LOCKED) {
 716                        spin_unlock(&mm->page_table_lock);
 717                        *mapcount = 0;  /* break early from loop */
 718                        *vm_flags |= VM_LOCKED;
 719                        goto out;
 720                }
 721
 722                /* go ahead even if the pmd is pmd_trans_splitting() */
 723                if (pmdp_clear_flush_young_notify(vma, address, pmd))
 724                        referenced++;
 725                spin_unlock(&mm->page_table_lock);
 726        } else {
 727                pte_t *pte;
 728                spinlock_t *ptl;
 729
 730                /*
 731                 * rmap might return false positives; we must filter
 732                 * these out using page_check_address().
 733                 */
 734                pte = page_check_address(page, mm, address, &ptl, 0);
 735                if (!pte)
 736                        goto out;
 737
 738                if (vma->vm_flags & VM_LOCKED) {
 739                        pte_unmap_unlock(pte, ptl);
 740                        *mapcount = 0;  /* break early from loop */
 741                        *vm_flags |= VM_LOCKED;
 742                        goto out;
 743                }
 744
 745                if (ptep_clear_flush_young_notify(vma, address, pte)) {
 746                        /*
 747                         * Don't treat a reference through a sequentially read
 748                         * mapping as such.  If the page has been used in
 749                         * another mapping, we will catch it; if this other
 750                         * mapping is already gone, the unmap path will have
 751                         * set PG_referenced or activated the page.
 752                         */
 753                        if (likely(!VM_SequentialReadHint(vma)))
 754                                referenced++;
 755                }
 756                pte_unmap_unlock(pte, ptl);
 757        }
 758
 759        (*mapcount)--;
 760
 761        if (referenced)
 762                *vm_flags |= vma->vm_flags;
 763out:
 764        return referenced;
 765}
 766
 767static int page_referenced_anon(struct page *page,
 768                                struct mem_cgroup *memcg,
 769                                unsigned long *vm_flags)
 770{
 771        unsigned int mapcount;
 772        struct anon_vma *anon_vma;
 773        struct anon_vma_chain *avc;
 774        int referenced = 0;
 775
 776        anon_vma = page_lock_anon_vma(page);
 777        if (!anon_vma)
 778                return referenced;
 779
 780        mapcount = page_mapcount(page);
 781        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
 782                struct vm_area_struct *vma = avc->vma;
 783                unsigned long address = vma_address(page, vma);
 784                if (address == -EFAULT)
 785                        continue;
 786                /*
 787                 * If we are reclaiming on behalf of a cgroup, skip
 788                 * counting on behalf of references from different
 789                 * cgroups
 790                 */
 791                if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 792                        continue;
 793                referenced += page_referenced_one(page, vma, address,
 794                                                  &mapcount, vm_flags);
 795                if (!mapcount)
 796                        break;
 797        }
 798
 799        page_unlock_anon_vma(anon_vma);
 800        return referenced;
 801}
 802
 803/**
 804 * page_referenced_file - referenced check for object-based rmap
 805 * @page: the page we're checking references on.
 806 * @memcg: target memory control group
 807 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 808 *
 809 * For an object-based mapped page, find all the places it is mapped and
 810 * check/clear the referenced flag.  This is done by following the page->mapping
 811 * pointer, then walking the chain of vmas it holds.  It returns the number
 812 * of references it found.
 813 *
 814 * This function is only called from page_referenced for object-based pages.
 815 */
 816static int page_referenced_file(struct page *page,
 817                                struct mem_cgroup *memcg,
 818                                unsigned long *vm_flags)
 819{
 820        unsigned int mapcount;
 821        struct address_space *mapping = page->mapping;
 822        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 823        struct vm_area_struct *vma;
 824        struct prio_tree_iter iter;
 825        int referenced = 0;
 826
 827        /*
 828         * The caller's checks on page->mapping and !PageAnon have made
 829         * sure that this is a file page: the check for page->mapping
 830         * excludes the case just before it gets set on an anon page.
 831         */
 832        BUG_ON(PageAnon(page));
 833
 834        /*
 835         * The page lock not only makes sure that page->mapping cannot
 836         * suddenly be NULLified by truncation, it makes sure that the
 837         * structure at mapping cannot be freed and reused yet,
 838         * so we can safely take mapping->i_mmap_mutex.
 839         */
 840        BUG_ON(!PageLocked(page));
 841
 842        mutex_lock(&mapping->i_mmap_mutex);
 843
 844        /*
 845         * i_mmap_mutex does not stabilize mapcount at all, but mapcount
 846         * is more likely to be accurate if we note it after spinning.
 847         */
 848        mapcount = page_mapcount(page);
 849
 850        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
 851                unsigned long address = vma_address(page, vma);
 852                if (address == -EFAULT)
 853                        continue;
 854                /*
 855                 * If we are reclaiming on behalf of a cgroup, skip
 856                 * counting on behalf of references from different
 857                 * cgroups
 858                 */
 859                if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 860                        continue;
 861                referenced += page_referenced_one(page, vma, address,
 862                                                  &mapcount, vm_flags);
 863                if (!mapcount)
 864                        break;
 865        }
 866
 867        mutex_unlock(&mapping->i_mmap_mutex);
 868        return referenced;
 869}
 870
 871/**
 872 * page_referenced - test if the page was referenced
 873 * @page: the page to test
 874 * @is_locked: caller holds lock on the page
 875 * @memcg: target memory cgroup
 876 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 877 *
 878 * Quick test_and_clear_referenced for all mappings to a page,
 879 * returns the number of ptes which referenced the page.
 880 */
 881int page_referenced(struct page *page,
 882                    int is_locked,
 883                    struct mem_cgroup *memcg,
 884                    unsigned long *vm_flags)
 885{
 886        int referenced = 0;
 887        int we_locked = 0;
 888
 889        *vm_flags = 0;
 890        if (page_mapped(page) && page_rmapping(page)) {
 891                if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 892                        we_locked = trylock_page(page);
 893                        if (!we_locked) {
 894                                referenced++;
 895                                goto out;
 896                        }
 897                }
 898                if (unlikely(PageKsm(page)))
 899                        referenced += page_referenced_ksm(page, memcg,
 900                                                                vm_flags);
 901                else if (PageAnon(page))
 902                        referenced += page_referenced_anon(page, memcg,
 903                                                                vm_flags);
 904                else if (page->mapping)
 905                        referenced += page_referenced_file(page, memcg,
 906                                                                vm_flags);
 907                if (we_locked)
 908                        unlock_page(page);
 909
 910                if (page_test_and_clear_young(page_to_pfn(page)))
 911                        referenced++;
 912        }
 913out:
 914        return referenced;
 915}
 916
 917static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 918                            unsigned long address)
 919{
 920        struct mm_struct *mm = vma->vm_mm;
 921        pte_t *pte;
 922        spinlock_t *ptl;
 923        int ret = 0;
 924
 925        pte = page_check_address(page, mm, address, &ptl, 1);
 926        if (!pte)
 927                goto out;
 928
 929        if (pte_dirty(*pte) || pte_write(*pte)) {
 930                pte_t entry;
 931
 932                flush_cache_page(vma, address, pte_pfn(*pte));
 933                entry = ptep_clear_flush_notify(vma, address, pte);
 934                entry = pte_wrprotect(entry);
 935                entry = pte_mkclean(entry);
 936                set_pte_at(mm, address, pte, entry);
 937                ret = 1;
 938        }
 939
 940        pte_unmap_unlock(pte, ptl);
 941out:
 942        return ret;
 943}
 944
 945static int page_mkclean_file(struct address_space *mapping, struct page *page)
 946{
 947        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 948        struct vm_area_struct *vma;
 949        struct prio_tree_iter iter;
 950        int ret = 0;
 951
 952        BUG_ON(PageAnon(page));
 953
 954        mutex_lock(&mapping->i_mmap_mutex);
 955        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
 956                if (vma->vm_flags & VM_SHARED) {
 957                        unsigned long address = vma_address(page, vma);
 958                        if (address == -EFAULT)
 959                                continue;
 960                        ret += page_mkclean_one(page, vma, address);
 961                }
 962        }
 963        mutex_unlock(&mapping->i_mmap_mutex);
 964        return ret;
 965}
 966
 967int page_mkclean(struct page *page)
 968{
 969        int ret = 0;
 970
 971        BUG_ON(!PageLocked(page));
 972
 973        if (page_mapped(page)) {
 974                struct address_space *mapping = page_mapping(page);
 975                if (mapping)
 976                        ret = page_mkclean_file(mapping, page);
 977        }
 978
 979        return ret;
 980}
 981EXPORT_SYMBOL_GPL(page_mkclean);
 982
 983/**
 984 * page_move_anon_rmap - move a page to our anon_vma
 985 * @page:       the page to move to our anon_vma
 986 * @vma:        the vma the page belongs to
 987 * @address:    the user virtual address mapped
 988 *
 989 * When a page belongs exclusively to one process after a COW event,
 990 * that page can be moved into the anon_vma that belongs to just that
 991 * process, so the rmap code will not search the parent or sibling
 992 * processes.
 993 */
 994void page_move_anon_rmap(struct page *page,
 995        struct vm_area_struct *vma, unsigned long address)
 996{
 997        struct anon_vma *anon_vma = vma->anon_vma;
 998
 999        VM_BUG_ON(!PageLocked(page));
1000        VM_BUG_ON(!anon_vma);
1001        VM_BUG_ON(page->index != linear_page_index(vma, address));
1002
1003        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1004        page->mapping = (struct address_space *) anon_vma;
1005}
1006
1007/**
1008 * __page_set_anon_rmap - set up new anonymous rmap
1009 * @page:       Page to add to rmap     
1010 * @vma:        VM area to add page to.
1011 * @address:    User virtual address of the mapping     
1012 * @exclusive:  the page is exclusively owned by the current process
1013 */
1014static void __page_set_anon_rmap(struct page *page,
1015        struct vm_area_struct *vma, unsigned long address, int exclusive)
1016{
1017        struct anon_vma *anon_vma = vma->anon_vma;
1018
1019        BUG_ON(!anon_vma);
1020
1021        if (PageAnon(page))
1022                return;
1023
1024        /*
1025         * If the page isn't exclusively mapped into this vma,
1026         * we must use the _oldest_ possible anon_vma for the
1027         * page mapping!
1028         */
1029        if (!exclusive)
1030                anon_vma = anon_vma->root;
1031
1032        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1033        page->mapping = (struct address_space *) anon_vma;
1034        page->index = linear_page_index(vma, address);
1035}
1036
1037/**
1038 * __page_check_anon_rmap - sanity check anonymous rmap addition
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 */
1043static void __page_check_anon_rmap(struct page *page,
1044        struct vm_area_struct *vma, unsigned long address)
1045{
1046#ifdef CONFIG_DEBUG_VM
1047        /*
1048         * The page's anon-rmap details (mapping and index) are guaranteed to
1049         * be set up correctly at this point.
1050         *
1051         * We have exclusion against page_add_anon_rmap because the caller
1052         * always holds the page locked, except if called from page_dup_rmap,
1053         * in which case the page is already known to be setup.
1054         *
1055         * We have exclusion against page_add_new_anon_rmap because those pages
1056         * are initially only visible via the pagetables, and the pte is locked
1057         * over the call to page_add_new_anon_rmap.
1058         */
1059        BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1060        BUG_ON(page->index != linear_page_index(vma, address));
1061#endif
1062}
1063
1064/**
1065 * page_add_anon_rmap - add pte mapping to an anonymous page
1066 * @page:       the page to add the mapping to
1067 * @vma:        the vm area in which the mapping is added
1068 * @address:    the user virtual address mapped
1069 *
1070 * The caller needs to hold the pte lock, and the page must be locked in
1071 * the anon_vma case: to serialize mapping,index checking after setting,
1072 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1073 * (but PageKsm is never downgraded to PageAnon).
1074 */
1075void page_add_anon_rmap(struct page *page,
1076        struct vm_area_struct *vma, unsigned long address)
1077{
1078        do_page_add_anon_rmap(page, vma, address, 0);
1079}
1080
1081/*
1082 * Special version of the above for do_swap_page, which often runs
1083 * into pages that are exclusively owned by the current process.
1084 * Everybody else should continue to use page_add_anon_rmap above.
1085 */
1086void do_page_add_anon_rmap(struct page *page,
1087        struct vm_area_struct *vma, unsigned long address, int exclusive)
1088{
1089        int first = atomic_inc_and_test(&page->_mapcount);
1090        if (first) {
1091                if (!PageTransHuge(page))
1092                        __inc_zone_page_state(page, NR_ANON_PAGES);
1093                else
1094                        __inc_zone_page_state(page,
1095                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1096        }
1097        if (unlikely(PageKsm(page)))
1098                return;
1099
1100        VM_BUG_ON(!PageLocked(page));
1101        /* address might be in next vma when migration races vma_adjust */
1102        if (first)
1103                __page_set_anon_rmap(page, vma, address, exclusive);
1104        else
1105                __page_check_anon_rmap(page, vma, address);
1106}
1107
1108/**
1109 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1110 * @page:       the page to add the mapping to
1111 * @vma:        the vm area in which the mapping is added
1112 * @address:    the user virtual address mapped
1113 *
1114 * Same as page_add_anon_rmap but must only be called on *new* pages.
1115 * This means the inc-and-test can be bypassed.
1116 * Page does not have to be locked.
1117 */
1118void page_add_new_anon_rmap(struct page *page,
1119        struct vm_area_struct *vma, unsigned long address)
1120{
1121        VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1122        SetPageSwapBacked(page);
1123        atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1124        if (!PageTransHuge(page))
1125                __inc_zone_page_state(page, NR_ANON_PAGES);
1126        else
1127                __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1128        __page_set_anon_rmap(page, vma, address, 1);
1129        if (page_evictable(page, vma))
1130                lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1131        else
1132                add_page_to_unevictable_list(page);
1133}
1134
1135/**
1136 * page_add_file_rmap - add pte mapping to a file page
1137 * @page: the page to add the mapping to
1138 *
1139 * The caller needs to hold the pte lock.
1140 */
1141void page_add_file_rmap(struct page *page)
1142{
1143        bool locked;
1144        unsigned long flags;
1145
1146        mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1147        if (atomic_inc_and_test(&page->_mapcount)) {
1148                __inc_zone_page_state(page, NR_FILE_MAPPED);
1149                mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1150        }
1151        mem_cgroup_end_update_page_stat(page, &locked, &flags);
1152}
1153
1154/**
1155 * page_remove_rmap - take down pte mapping from a page
1156 * @page: page to remove mapping from
1157 *
1158 * The caller needs to hold the pte lock.
1159 */
1160void page_remove_rmap(struct page *page)
1161{
1162        struct address_space *mapping = page_mapping(page);
1163        bool anon = PageAnon(page);
1164        bool locked;
1165        unsigned long flags;
1166
1167        /*
1168         * The anon case has no mem_cgroup page_stat to update; but may
1169         * uncharge_page() below, where the lock ordering can deadlock if
1170         * we hold the lock against page_stat move: so avoid it on anon.
1171         */
1172        if (!anon)
1173                mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1174
1175        /* page still mapped by someone else? */
1176        if (!atomic_add_negative(-1, &page->_mapcount))
1177                goto out;
1178
1179        /*
1180         * Now that the last pte has gone, s390 must transfer dirty
1181         * flag from storage key to struct page.  We can usually skip
1182         * this if the page is anon, so about to be freed; but perhaps
1183         * not if it's in swapcache - there might be another pte slot
1184         * containing the swap entry, but page not yet written to swap.
1185         *
1186         * And we can skip it on file pages, so long as the filesystem
1187         * participates in dirty tracking; but need to catch shm and tmpfs
1188         * and ramfs pages which have been modified since creation by read
1189         * fault.
1190         *
1191         * Note that mapping must be decided above, before decrementing
1192         * mapcount (which luckily provides a barrier): once page is unmapped,
1193         * it could be truncated and page->mapping reset to NULL at any moment.
1194         * Note also that we are relying on page_mapping(page) to set mapping
1195         * to &swapper_space when PageSwapCache(page).
1196         */
1197        if (mapping && !mapping_cap_account_dirty(mapping) &&
1198            page_test_and_clear_dirty(page_to_pfn(page), 1))
1199                set_page_dirty(page);
1200        /*
1201         * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1202         * and not charged by memcg for now.
1203         */
1204        if (unlikely(PageHuge(page)))
1205                goto out;
1206        if (anon) {
1207                mem_cgroup_uncharge_page(page);
1208                if (!PageTransHuge(page))
1209                        __dec_zone_page_state(page, NR_ANON_PAGES);
1210                else
1211                        __dec_zone_page_state(page,
1212                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1213        } else {
1214                __dec_zone_page_state(page, NR_FILE_MAPPED);
1215                mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1216        }
1217        /*
1218         * It would be tidy to reset the PageAnon mapping here,
1219         * but that might overwrite a racing page_add_anon_rmap
1220         * which increments mapcount after us but sets mapping
1221         * before us: so leave the reset to free_hot_cold_page,
1222         * and remember that it's only reliable while mapped.
1223         * Leaving it set also helps swapoff to reinstate ptes
1224         * faster for those pages still in swapcache.
1225         */
1226out:
1227        if (!anon)
1228                mem_cgroup_end_update_page_stat(page, &locked, &flags);
1229}
1230
1231/*
1232 * Subfunctions of try_to_unmap: try_to_unmap_one called
1233 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1234 */
1235int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1236                     unsigned long address, enum ttu_flags flags)
1237{
1238        struct mm_struct *mm = vma->vm_mm;
1239        pte_t *pte;
1240        pte_t pteval;
1241        spinlock_t *ptl;
1242        int ret = SWAP_AGAIN;
1243
1244        pte = page_check_address(page, mm, address, &ptl, 0);
1245        if (!pte)
1246                goto out;
1247
1248        /*
1249         * If the page is mlock()d, we cannot swap it out.
1250         * If it's recently referenced (perhaps page_referenced
1251         * skipped over this mm) then we should reactivate it.
1252         */
1253        if (!(flags & TTU_IGNORE_MLOCK)) {
1254                if (vma->vm_flags & VM_LOCKED)
1255                        goto out_mlock;
1256
1257                if (TTU_ACTION(flags) == TTU_MUNLOCK)
1258                        goto out_unmap;
1259        }
1260        if (!(flags & TTU_IGNORE_ACCESS)) {
1261                if (ptep_clear_flush_young_notify(vma, address, pte)) {
1262                        ret = SWAP_FAIL;
1263                        goto out_unmap;
1264                }
1265        }
1266
1267        /* Nuke the page table entry. */
1268        flush_cache_page(vma, address, page_to_pfn(page));
1269        pteval = ptep_clear_flush_notify(vma, address, pte);
1270
1271        /* Move the dirty bit to the physical page now the pte is gone. */
1272        if (pte_dirty(pteval))
1273                set_page_dirty(page);
1274
1275        /* Update high watermark before we lower rss */
1276        update_hiwater_rss(mm);
1277
1278        if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1279                if (PageAnon(page))
1280                        dec_mm_counter(mm, MM_ANONPAGES);
1281                else
1282                        dec_mm_counter(mm, MM_FILEPAGES);
1283                set_pte_at(mm, address, pte,
1284                                swp_entry_to_pte(make_hwpoison_entry(page)));
1285        } else if (PageAnon(page)) {
1286                swp_entry_t entry = { .val = page_private(page) };
1287
1288                if (PageSwapCache(page)) {
1289                        /*
1290                         * Store the swap location in the pte.
1291                         * See handle_pte_fault() ...
1292                         */
1293                        if (swap_duplicate(entry) < 0) {
1294                                set_pte_at(mm, address, pte, pteval);
1295                                ret = SWAP_FAIL;
1296                                goto out_unmap;
1297                        }
1298                        if (list_empty(&mm->mmlist)) {
1299                                spin_lock(&mmlist_lock);
1300                                if (list_empty(&mm->mmlist))
1301                                        list_add(&mm->mmlist, &init_mm.mmlist);
1302                                spin_unlock(&mmlist_lock);
1303                        }
1304                        dec_mm_counter(mm, MM_ANONPAGES);
1305                        inc_mm_counter(mm, MM_SWAPENTS);
1306                } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1307                        /*
1308                         * Store the pfn of the page in a special migration
1309                         * pte. do_swap_page() will wait until the migration
1310                         * pte is removed and then restart fault handling.
1311                         */
1312                        BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1313                        entry = make_migration_entry(page, pte_write(pteval));
1314                }
1315                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1316                BUG_ON(pte_file(*pte));
1317        } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1318                   (TTU_ACTION(flags) == TTU_MIGRATION)) {
1319                /* Establish migration entry for a file page */
1320                swp_entry_t entry;
1321                entry = make_migration_entry(page, pte_write(pteval));
1322                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1323        } else
1324                dec_mm_counter(mm, MM_FILEPAGES);
1325
1326        page_remove_rmap(page);
1327        page_cache_release(page);
1328
1329out_unmap:
1330        pte_unmap_unlock(pte, ptl);
1331out:
1332        return ret;
1333
1334out_mlock:
1335        pte_unmap_unlock(pte, ptl);
1336
1337
1338        /*
1339         * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1340         * unstable result and race. Plus, We can't wait here because
1341         * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1342         * if trylock failed, the page remain in evictable lru and later
1343         * vmscan could retry to move the page to unevictable lru if the
1344         * page is actually mlocked.
1345         */
1346        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1347                if (vma->vm_flags & VM_LOCKED) {
1348                        mlock_vma_page(page);
1349                        ret = SWAP_MLOCK;
1350                }
1351                up_read(&vma->vm_mm->mmap_sem);
1352        }
1353        return ret;
1354}
1355
1356/*
1357 * objrmap doesn't work for nonlinear VMAs because the assumption that
1358 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1359 * Consequently, given a particular page and its ->index, we cannot locate the
1360 * ptes which are mapping that page without an exhaustive linear search.
1361 *
1362 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1363 * maps the file to which the target page belongs.  The ->vm_private_data field
1364 * holds the current cursor into that scan.  Successive searches will circulate
1365 * around the vma's virtual address space.
1366 *
1367 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1368 * more scanning pressure is placed against them as well.   Eventually pages
1369 * will become fully unmapped and are eligible for eviction.
1370 *
1371 * For very sparsely populated VMAs this is a little inefficient - chances are
1372 * there there won't be many ptes located within the scan cluster.  In this case
1373 * maybe we could scan further - to the end of the pte page, perhaps.
1374 *
1375 * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
1376 * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
1377 * rather than unmapping them.  If we encounter the "check_page" that vmscan is
1378 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1379 */
1380#define CLUSTER_SIZE    min(32*PAGE_SIZE, PMD_SIZE)
1381#define CLUSTER_MASK    (~(CLUSTER_SIZE - 1))
1382
1383static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1384                struct vm_area_struct *vma, struct page *check_page)
1385{
1386        struct mm_struct *mm = vma->vm_mm;
1387        pgd_t *pgd;
1388        pud_t *pud;
1389        pmd_t *pmd;
1390        pte_t *pte;
1391        pte_t pteval;
1392        spinlock_t *ptl;
1393        struct page *page;
1394        unsigned long address;
1395        unsigned long end;
1396        int ret = SWAP_AGAIN;
1397        int locked_vma = 0;
1398
1399        address = (vma->vm_start + cursor) & CLUSTER_MASK;
1400        end = address + CLUSTER_SIZE;
1401        if (address < vma->vm_start)
1402                address = vma->vm_start;
1403        if (end > vma->vm_end)
1404                end = vma->vm_end;
1405
1406        pgd = pgd_offset(mm, address);
1407        if (!pgd_present(*pgd))
1408                return ret;
1409
1410        pud = pud_offset(pgd, address);
1411        if (!pud_present(*pud))
1412                return ret;
1413
1414        pmd = pmd_offset(pud, address);
1415        if (!pmd_present(*pmd))
1416                return ret;
1417
1418        /*
1419         * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1420         * keep the sem while scanning the cluster for mlocking pages.
1421         */
1422        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1423                locked_vma = (vma->vm_flags & VM_LOCKED);
1424                if (!locked_vma)
1425                        up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1426        }
1427
1428        pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1429
1430        /* Update high watermark before we lower rss */
1431        update_hiwater_rss(mm);
1432
1433        for (; address < end; pte++, address += PAGE_SIZE) {
1434                if (!pte_present(*pte))
1435                        continue;
1436                page = vm_normal_page(vma, address, *pte);
1437                BUG_ON(!page || PageAnon(page));
1438
1439                if (locked_vma) {
1440                        mlock_vma_page(page);   /* no-op if already mlocked */
1441                        if (page == check_page)
1442                                ret = SWAP_MLOCK;
1443                        continue;       /* don't unmap */
1444                }
1445
1446                if (ptep_clear_flush_young_notify(vma, address, pte))
1447                        continue;
1448
1449                /* Nuke the page table entry. */
1450                flush_cache_page(vma, address, pte_pfn(*pte));
1451                pteval = ptep_clear_flush_notify(vma, address, pte);
1452
1453                /* If nonlinear, store the file page offset in the pte. */
1454                if (page->index != linear_page_index(vma, address))
1455                        set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1456
1457                /* Move the dirty bit to the physical page now the pte is gone. */
1458                if (pte_dirty(pteval))
1459                        set_page_dirty(page);
1460
1461                page_remove_rmap(page);
1462                page_cache_release(page);
1463                dec_mm_counter(mm, MM_FILEPAGES);
1464                (*mapcount)--;
1465        }
1466        pte_unmap_unlock(pte - 1, ptl);
1467        if (locked_vma)
1468                up_read(&vma->vm_mm->mmap_sem);
1469        return ret;
1470}
1471
1472bool is_vma_temporary_stack(struct vm_area_struct *vma)
1473{
1474        int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1475
1476        if (!maybe_stack)
1477                return false;
1478
1479        if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1480                                                VM_STACK_INCOMPLETE_SETUP)
1481                return true;
1482
1483        return false;
1484}
1485
1486/**
1487 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1488 * rmap method
1489 * @page: the page to unmap/unlock
1490 * @flags: action and flags
1491 *
1492 * Find all the mappings of a page using the mapping pointer and the vma chains
1493 * contained in the anon_vma struct it points to.
1494 *
1495 * This function is only called from try_to_unmap/try_to_munlock for
1496 * anonymous pages.
1497 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1498 * where the page was found will be held for write.  So, we won't recheck
1499 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1500 * 'LOCKED.
1501 */
1502static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1503{
1504        struct anon_vma *anon_vma;
1505        struct anon_vma_chain *avc;
1506        int ret = SWAP_AGAIN;
1507
1508        anon_vma = page_lock_anon_vma(page);
1509        if (!anon_vma)
1510                return ret;
1511
1512        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1513                struct vm_area_struct *vma = avc->vma;
1514                unsigned long address;
1515
1516                /*
1517                 * During exec, a temporary VMA is setup and later moved.
1518                 * The VMA is moved under the anon_vma lock but not the
1519                 * page tables leading to a race where migration cannot
1520                 * find the migration ptes. Rather than increasing the
1521                 * locking requirements of exec(), migration skips
1522                 * temporary VMAs until after exec() completes.
1523                 */
1524                if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1525                                is_vma_temporary_stack(vma))
1526                        continue;
1527
1528                address = vma_address(page, vma);
1529                if (address == -EFAULT)
1530                        continue;
1531                ret = try_to_unmap_one(page, vma, address, flags);
1532                if (ret != SWAP_AGAIN || !page_mapped(page))
1533                        break;
1534        }
1535
1536        page_unlock_anon_vma(anon_vma);
1537        return ret;
1538}
1539
1540/**
1541 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1542 * @page: the page to unmap/unlock
1543 * @flags: action and flags
1544 *
1545 * Find all the mappings of a page using the mapping pointer and the vma chains
1546 * contained in the address_space struct it points to.
1547 *
1548 * This function is only called from try_to_unmap/try_to_munlock for
1549 * object-based pages.
1550 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1551 * where the page was found will be held for write.  So, we won't recheck
1552 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1553 * 'LOCKED.
1554 */
1555static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1556{
1557        struct address_space *mapping = page->mapping;
1558        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1559        struct vm_area_struct *vma;
1560        struct prio_tree_iter iter;
1561        int ret = SWAP_AGAIN;
1562        unsigned long cursor;
1563        unsigned long max_nl_cursor = 0;
1564        unsigned long max_nl_size = 0;
1565        unsigned int mapcount;
1566
1567        mutex_lock(&mapping->i_mmap_mutex);
1568        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1569                unsigned long address = vma_address(page, vma);
1570                if (address == -EFAULT)
1571                        continue;
1572                ret = try_to_unmap_one(page, vma, address, flags);
1573                if (ret != SWAP_AGAIN || !page_mapped(page))
1574                        goto out;
1575        }
1576
1577        if (list_empty(&mapping->i_mmap_nonlinear))
1578                goto out;
1579
1580        /*
1581         * We don't bother to try to find the munlocked page in nonlinears.
1582         * It's costly. Instead, later, page reclaim logic may call
1583         * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1584         */
1585        if (TTU_ACTION(flags) == TTU_MUNLOCK)
1586                goto out;
1587
1588        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1589                                                shared.vm_set.list) {
1590                cursor = (unsigned long) vma->vm_private_data;
1591                if (cursor > max_nl_cursor)
1592                        max_nl_cursor = cursor;
1593                cursor = vma->vm_end - vma->vm_start;
1594                if (cursor > max_nl_size)
1595                        max_nl_size = cursor;
1596        }
1597
1598        if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1599                ret = SWAP_FAIL;
1600                goto out;
1601        }
1602
1603        /*
1604         * We don't try to search for this page in the nonlinear vmas,
1605         * and page_referenced wouldn't have found it anyway.  Instead
1606         * just walk the nonlinear vmas trying to age and unmap some.
1607         * The mapcount of the page we came in with is irrelevant,
1608         * but even so use it as a guide to how hard we should try?
1609         */
1610        mapcount = page_mapcount(page);
1611        if (!mapcount)
1612                goto out;
1613        cond_resched();
1614
1615        max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1616        if (max_nl_cursor == 0)
1617                max_nl_cursor = CLUSTER_SIZE;
1618
1619        do {
1620                list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1621                                                shared.vm_set.list) {
1622                        cursor = (unsigned long) vma->vm_private_data;
1623                        while ( cursor < max_nl_cursor &&
1624                                cursor < vma->vm_end - vma->vm_start) {
1625                                if (try_to_unmap_cluster(cursor, &mapcount,
1626                                                vma, page) == SWAP_MLOCK)
1627                                        ret = SWAP_MLOCK;
1628                                cursor += CLUSTER_SIZE;
1629                                vma->vm_private_data = (void *) cursor;
1630                                if ((int)mapcount <= 0)
1631                                        goto out;
1632                        }
1633                        vma->vm_private_data = (void *) max_nl_cursor;
1634                }
1635                cond_resched();
1636                max_nl_cursor += CLUSTER_SIZE;
1637        } while (max_nl_cursor <= max_nl_size);
1638
1639        /*
1640         * Don't loop forever (perhaps all the remaining pages are
1641         * in locked vmas).  Reset cursor on all unreserved nonlinear
1642         * vmas, now forgetting on which ones it had fallen behind.
1643         */
1644        list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1645                vma->vm_private_data = NULL;
1646out:
1647        mutex_unlock(&mapping->i_mmap_mutex);
1648        return ret;
1649}
1650
1651/**
1652 * try_to_unmap - try to remove all page table mappings to a page
1653 * @page: the page to get unmapped
1654 * @flags: action and flags
1655 *
1656 * Tries to remove all the page table entries which are mapping this
1657 * page, used in the pageout path.  Caller must hold the page lock.
1658 * Return values are:
1659 *
1660 * SWAP_SUCCESS - we succeeded in removing all mappings
1661 * SWAP_AGAIN   - we missed a mapping, try again later
1662 * SWAP_FAIL    - the page is unswappable
1663 * SWAP_MLOCK   - page is mlocked.
1664 */
1665int try_to_unmap(struct page *page, enum ttu_flags flags)
1666{
1667        int ret;
1668
1669        BUG_ON(!PageLocked(page));
1670        VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1671
1672        if (unlikely(PageKsm(page)))
1673                ret = try_to_unmap_ksm(page, flags);
1674        else if (PageAnon(page))
1675                ret = try_to_unmap_anon(page, flags);
1676        else
1677                ret = try_to_unmap_file(page, flags);
1678        if (ret != SWAP_MLOCK && !page_mapped(page))
1679                ret = SWAP_SUCCESS;
1680        return ret;
1681}
1682
1683/**
1684 * try_to_munlock - try to munlock a page
1685 * @page: the page to be munlocked
1686 *
1687 * Called from munlock code.  Checks all of the VMAs mapping the page
1688 * to make sure nobody else has this page mlocked. The page will be
1689 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1690 *
1691 * Return values are:
1692 *
1693 * SWAP_AGAIN   - no vma is holding page mlocked, or,
1694 * SWAP_AGAIN   - page mapped in mlocked vma -- couldn't acquire mmap sem
1695 * SWAP_FAIL    - page cannot be located at present
1696 * SWAP_MLOCK   - page is now mlocked.
1697 */
1698int try_to_munlock(struct page *page)
1699{
1700        VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1701
1702        if (unlikely(PageKsm(page)))
1703                return try_to_unmap_ksm(page, TTU_MUNLOCK);
1704        else if (PageAnon(page))
1705                return try_to_unmap_anon(page, TTU_MUNLOCK);
1706        else
1707                return try_to_unmap_file(page, TTU_MUNLOCK);
1708}
1709
1710void __put_anon_vma(struct anon_vma *anon_vma)
1711{
1712        struct anon_vma *root = anon_vma->root;
1713
1714        if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1715                anon_vma_free(root);
1716
1717        anon_vma_free(anon_vma);
1718}
1719
1720#ifdef CONFIG_MIGRATION
1721/*
1722 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1723 * Called by migrate.c to remove migration ptes, but might be used more later.
1724 */
1725static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1726                struct vm_area_struct *, unsigned long, void *), void *arg)
1727{
1728        struct anon_vma *anon_vma;
1729        struct anon_vma_chain *avc;
1730        int ret = SWAP_AGAIN;
1731
1732        /*
1733         * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1734         * because that depends on page_mapped(); but not all its usages
1735         * are holding mmap_sem. Users without mmap_sem are required to
1736         * take a reference count to prevent the anon_vma disappearing
1737         */
1738        anon_vma = page_anon_vma(page);
1739        if (!anon_vma)
1740                return ret;
1741        anon_vma_lock(anon_vma);
1742        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1743                struct vm_area_struct *vma = avc->vma;
1744                unsigned long address = vma_address(page, vma);
1745                if (address == -EFAULT)
1746                        continue;
1747                ret = rmap_one(page, vma, address, arg);
1748                if (ret != SWAP_AGAIN)
1749                        break;
1750        }
1751        anon_vma_unlock(anon_vma);
1752        return ret;
1753}
1754
1755static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1756                struct vm_area_struct *, unsigned long, void *), void *arg)
1757{
1758        struct address_space *mapping = page->mapping;
1759        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1760        struct vm_area_struct *vma;
1761        struct prio_tree_iter iter;
1762        int ret = SWAP_AGAIN;
1763
1764        if (!mapping)
1765                return ret;
1766        mutex_lock(&mapping->i_mmap_mutex);
1767        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1768                unsigned long address = vma_address(page, vma);
1769                if (address == -EFAULT)
1770                        continue;
1771                ret = rmap_one(page, vma, address, arg);
1772                if (ret != SWAP_AGAIN)
1773                        break;
1774        }
1775        /*
1776         * No nonlinear handling: being always shared, nonlinear vmas
1777         * never contain migration ptes.  Decide what to do about this
1778         * limitation to linear when we need rmap_walk() on nonlinear.
1779         */
1780        mutex_unlock(&mapping->i_mmap_mutex);
1781        return ret;
1782}
1783
1784int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1785                struct vm_area_struct *, unsigned long, void *), void *arg)
1786{
1787        VM_BUG_ON(!PageLocked(page));
1788
1789        if (unlikely(PageKsm(page)))
1790                return rmap_walk_ksm(page, rmap_one, arg);
1791        else if (PageAnon(page))
1792                return rmap_walk_anon(page, rmap_one, arg);
1793        else
1794                return rmap_walk_file(page, rmap_one, arg);
1795}
1796#endif /* CONFIG_MIGRATION */
1797
1798#ifdef CONFIG_HUGETLB_PAGE
1799/*
1800 * The following three functions are for anonymous (private mapped) hugepages.
1801 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1802 * and no lru code, because we handle hugepages differently from common pages.
1803 */
1804static void __hugepage_set_anon_rmap(struct page *page,
1805        struct vm_area_struct *vma, unsigned long address, int exclusive)
1806{
1807        struct anon_vma *anon_vma = vma->anon_vma;
1808
1809        BUG_ON(!anon_vma);
1810
1811        if (PageAnon(page))
1812                return;
1813        if (!exclusive)
1814                anon_vma = anon_vma->root;
1815
1816        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1817        page->mapping = (struct address_space *) anon_vma;
1818        page->index = linear_page_index(vma, address);
1819}
1820
1821void hugepage_add_anon_rmap(struct page *page,
1822                            struct vm_area_struct *vma, unsigned long address)
1823{
1824        struct anon_vma *anon_vma = vma->anon_vma;
1825        int first;
1826
1827        BUG_ON(!PageLocked(page));
1828        BUG_ON(!anon_vma);
1829        /* address might be in next vma when migration races vma_adjust */
1830        first = atomic_inc_and_test(&page->_mapcount);
1831        if (first)
1832                __hugepage_set_anon_rmap(page, vma, address, 0);
1833}
1834
1835void hugepage_add_new_anon_rmap(struct page *page,
1836                        struct vm_area_struct *vma, unsigned long address)
1837{
1838        BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1839        atomic_set(&page->_mapcount, 0);
1840        __hugepage_set_anon_rmap(page, vma, address, 1);
1841}
1842#endif /* CONFIG_HUGETLB_PAGE */
1843
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.