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/module.h>
  55#include <linux/memcontrol.h>
  56#include <linux/mmu_notifier.h>
  57#include <linux/migrate.h>
  58#include <linux/hugetlb.h>
  59
  60#include <asm/tlbflush.h>
  61
  62#include "internal.h"
  63
  64static struct kmem_cache *anon_vma_cachep;
  65static struct kmem_cache *anon_vma_chain_cachep;
  66
  67static inline struct anon_vma *anon_vma_alloc(void)
  68{
  69        struct anon_vma *anon_vma;
  70
  71        anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  72        if (anon_vma) {
  73                atomic_set(&anon_vma->refcount, 1);
  74                /*
  75                 * Initialise the anon_vma root to point to itself. If called
  76                 * from fork, the root will be reset to the parents anon_vma.
  77                 */
  78                anon_vma->root = anon_vma;
  79        }
  80
  81        return anon_vma;
  82}
  83
  84static inline void anon_vma_free(struct anon_vma *anon_vma)
  85{
  86        VM_BUG_ON(atomic_read(&anon_vma->refcount));
  87
  88        /*
  89         * Synchronize against page_lock_anon_vma() such that
  90         * we can safely hold the lock without the anon_vma getting
  91         * freed.
  92         *
  93         * Relies on the full mb implied by the atomic_dec_and_test() from
  94         * put_anon_vma() against the acquire barrier implied by
  95         * mutex_trylock() from page_lock_anon_vma(). This orders:
  96         *
  97         * page_lock_anon_vma()         VS      put_anon_vma()
  98         *   mutex_trylock()                      atomic_dec_and_test()
  99         *   LOCK                                 MB
 100         *   atomic_read()                        mutex_is_locked()
 101         *
 102         * LOCK should suffice since the actual taking of the lock must
 103         * happen _before_ what follows.
 104         */
 105        if (mutex_is_locked(&anon_vma->root->mutex)) {
 106                anon_vma_lock(anon_vma);
 107                anon_vma_unlock(anon_vma);
 108        }
 109
 110        kmem_cache_free(anon_vma_cachep, anon_vma);
 111}
 112
 113static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 114{
 115        return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 116}
 117
 118static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 119{
 120        kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 121}
 122
 123/**
 124 * anon_vma_prepare - attach an anon_vma to a memory region
 125 * @vma: the memory region in question
 126 *
 127 * This makes sure the memory mapping described by 'vma' has
 128 * an 'anon_vma' attached to it, so that we can associate the
 129 * anonymous pages mapped into it with that anon_vma.
 130 *
 131 * The common case will be that we already have one, but if
 132 * not we either need to find an adjacent mapping that we
 133 * can re-use the anon_vma from (very common when the only
 134 * reason for splitting a vma has been mprotect()), or we
 135 * allocate a new one.
 136 *
 137 * Anon-vma allocations are very subtle, because we may have
 138 * optimistically looked up an anon_vma in page_lock_anon_vma()
 139 * and that may actually touch the spinlock even in the newly
 140 * allocated vma (it depends on RCU to make sure that the
 141 * anon_vma isn't actually destroyed).
 142 *
 143 * As a result, we need to do proper anon_vma locking even
 144 * for the new allocation. At the same time, we do not want
 145 * to do any locking for the common case of already having
 146 * an anon_vma.
 147 *
 148 * This must be called with the mmap_sem held for reading.
 149 */
 150int anon_vma_prepare(struct vm_area_struct *vma)
 151{
 152        struct anon_vma *anon_vma = vma->anon_vma;
 153        struct anon_vma_chain *avc;
 154
 155        might_sleep();
 156        if (unlikely(!anon_vma)) {
 157                struct mm_struct *mm = vma->vm_mm;
 158                struct anon_vma *allocated;
 159
 160                avc = anon_vma_chain_alloc(GFP_KERNEL);
 161                if (!avc)
 162                        goto out_enomem;
 163
 164                anon_vma = find_mergeable_anon_vma(vma);
 165                allocated = NULL;
 166                if (!anon_vma) {
 167                        anon_vma = anon_vma_alloc();
 168                        if (unlikely(!anon_vma))
 169                                goto out_enomem_free_avc;
 170                        allocated = anon_vma;
 171                }
 172
 173                anon_vma_lock(anon_vma);
 174                /* page_table_lock to protect against threads */
 175                spin_lock(&mm->page_table_lock);
 176                if (likely(!vma->anon_vma)) {
 177                        vma->anon_vma = anon_vma;
 178                        avc->anon_vma = anon_vma;
 179                        avc->vma = vma;
 180                        list_add(&avc->same_vma, &vma->anon_vma_chain);
 181                        list_add_tail(&avc->same_anon_vma, &anon_vma->head);
 182                        allocated = NULL;
 183                        avc = NULL;
 184                }
 185                spin_unlock(&mm->page_table_lock);
 186                anon_vma_unlock(anon_vma);
 187
 188                if (unlikely(allocated))
 189                        put_anon_vma(allocated);
 190                if (unlikely(avc))
 191                        anon_vma_chain_free(avc);
 192        }
 193        return 0;
 194
 195 out_enomem_free_avc:
 196        anon_vma_chain_free(avc);
 197 out_enomem:
 198        return -ENOMEM;
 199}
 200
 201/*
 202 * This is a useful helper function for locking the anon_vma root as
 203 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 204 * have the same vma.
 205 *
 206 * Such anon_vma's should have the same root, so you'd expect to see
 207 * just a single mutex_lock for the whole traversal.
 208 */
 209static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 210{
 211        struct anon_vma *new_root = anon_vma->root;
 212        if (new_root != root) {
 213                if (WARN_ON_ONCE(root))
 214                        mutex_unlock(&root->mutex);
 215                root = new_root;
 216                mutex_lock(&root->mutex);
 217        }
 218        return root;
 219}
 220
 221static inline void unlock_anon_vma_root(struct anon_vma *root)
 222{
 223        if (root)
 224                mutex_unlock(&root->mutex);
 225}
 226
 227static void anon_vma_chain_link(struct vm_area_struct *vma,
 228                                struct anon_vma_chain *avc,
 229                                struct anon_vma *anon_vma)
 230{
 231        avc->vma = vma;
 232        avc->anon_vma = anon_vma;
 233        list_add(&avc->same_vma, &vma->anon_vma_chain);
 234
 235        /*
 236         * It's critical to add new vmas to the tail of the anon_vma,
 237         * see comment in huge_memory.c:__split_huge_page().
 238         */
 239        list_add_tail(&avc->same_anon_vma, &anon_vma->head);
 240}
 241
 242/*
 243 * Attach the anon_vmas from src to dst.
 244 * Returns 0 on success, -ENOMEM on failure.
 245 */
 246int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 247{
 248        struct anon_vma_chain *avc, *pavc;
 249        struct anon_vma *root = NULL;
 250
 251        list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 252                struct anon_vma *anon_vma;
 253
 254                avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 255                if (unlikely(!avc)) {
 256                        unlock_anon_vma_root(root);
 257                        root = NULL;
 258                        avc = anon_vma_chain_alloc(GFP_KERNEL);
 259                        if (!avc)
 260                                goto enomem_failure;
 261                }
 262                anon_vma = pavc->anon_vma;
 263                root = lock_anon_vma_root(root, anon_vma);
 264                anon_vma_chain_link(dst, avc, anon_vma);
 265        }
 266        unlock_anon_vma_root(root);
 267        return 0;
 268
 269 enomem_failure:
 270        unlink_anon_vmas(dst);
 271        return -ENOMEM;
 272}
 273
 274/*
 275 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 276 * the corresponding VMA in the parent process is attached to.
 277 * Returns 0 on success, non-zero on failure.
 278 */
 279int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 280{
 281        struct anon_vma_chain *avc;
 282        struct anon_vma *anon_vma;
 283
 284        /* Don't bother if the parent process has no anon_vma here. */
 285        if (!pvma->anon_vma)
 286                return 0;
 287
 288        /*
 289         * First, attach the new VMA to the parent VMA's anon_vmas,
 290         * so rmap can find non-COWed pages in child processes.
 291         */
 292        if (anon_vma_clone(vma, pvma))
 293                return -ENOMEM;
 294
 295        /* Then add our own anon_vma. */
 296        anon_vma = anon_vma_alloc();
 297        if (!anon_vma)
 298                goto out_error;
 299        avc = anon_vma_chain_alloc(GFP_KERNEL);
 300        if (!avc)
 301                goto out_error_free_anon_vma;
 302
 303        /*
 304         * The root anon_vma's spinlock is the lock actually used when we
 305         * lock any of the anon_vmas in this anon_vma tree.
 306         */
 307        anon_vma->root = pvma->anon_vma->root;
 308        /*
 309         * With refcounts, an anon_vma can stay around longer than the
 310         * process it belongs to. The root anon_vma needs to be pinned until
 311         * this anon_vma is freed, because the lock lives in the root.
 312         */
 313        get_anon_vma(anon_vma->root);
 314        /* Mark this anon_vma as the one where our new (COWed) pages go. */
 315        vma->anon_vma = anon_vma;
 316        anon_vma_lock(anon_vma);
 317        anon_vma_chain_link(vma, avc, anon_vma);
 318        anon_vma_unlock(anon_vma);
 319
 320        return 0;
 321
 322 out_error_free_anon_vma:
 323        put_anon_vma(anon_vma);
 324 out_error:
 325        unlink_anon_vmas(vma);
 326        return -ENOMEM;
 327}
 328
 329void unlink_anon_vmas(struct vm_area_struct *vma)
 330{
 331        struct anon_vma_chain *avc, *next;
 332        struct anon_vma *root = NULL;
 333
 334        /*
 335         * Unlink each anon_vma chained to the VMA.  This list is ordered
 336         * from newest to oldest, ensuring the root anon_vma gets freed last.
 337         */
 338        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 339                struct anon_vma *anon_vma = avc->anon_vma;
 340
 341                root = lock_anon_vma_root(root, anon_vma);
 342                list_del(&avc->same_anon_vma);
 343
 344                /*
 345                 * Leave empty anon_vmas on the list - we'll need
 346                 * to free them outside the lock.
 347                 */
 348                if (list_empty(&anon_vma->head))
 349                        continue;
 350
 351                list_del(&avc->same_vma);
 352                anon_vma_chain_free(avc);
 353        }
 354        unlock_anon_vma_root(root);
 355
 356        /*
 357         * Iterate the list once more, it now only contains empty and unlinked
 358         * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 359         * needing to acquire the anon_vma->root->mutex.
 360         */
 361        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 362                struct anon_vma *anon_vma = avc->anon_vma;
 363
 364                put_anon_vma(anon_vma);
 365
 366                list_del(&avc->same_vma);
 367                anon_vma_chain_free(avc);
 368        }
 369}
 370
 371static void anon_vma_ctor(void *data)
 372{
 373        struct anon_vma *anon_vma = data;
 374
 375        mutex_init(&anon_vma->mutex);
 376        atomic_set(&anon_vma->refcount, 0);
 377        INIT_LIST_HEAD(&anon_vma->head);
 378}
 379
 380void __init anon_vma_init(void)
 381{
 382        anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 383                        0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
 384        anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
 385}
 386
 387/*
 388 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 389 *
 390 * Since there is no serialization what so ever against page_remove_rmap()
 391 * the best this function can do is return a locked anon_vma that might
 392 * have been relevant to this page.
 393 *
 394 * The page might have been remapped to a different anon_vma or the anon_vma
 395 * returned may already be freed (and even reused).
 396 *
 397 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 398 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 399 * ensure that any anon_vma obtained from the page will still be valid for as
 400 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 401 *
 402 * All users of this function must be very careful when walking the anon_vma
 403 * chain and verify that the page in question is indeed mapped in it
 404 * [ something equivalent to page_mapped_in_vma() ].
 405 *
 406 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 407 * that the anon_vma pointer from page->mapping is valid if there is a
 408 * mapcount, we can dereference the anon_vma after observing those.
 409 */
 410struct anon_vma *page_get_anon_vma(struct page *page)
 411{
 412        struct anon_vma *anon_vma = NULL;
 413        unsigned long anon_mapping;
 414
 415        rcu_read_lock();
 416        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 417        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 418                goto out;
 419        if (!page_mapped(page))
 420                goto out;
 421
 422        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 423        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 424                anon_vma = NULL;
 425                goto out;
 426        }
 427
 428        /*
 429         * If this page is still mapped, then its anon_vma cannot have been
 430         * freed.  But if it has been unmapped, we have no security against the
 431         * anon_vma structure being freed and reused (for another anon_vma:
 432         * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
 433         * above cannot corrupt).
 434         */
 435        if (!page_mapped(page)) {
 436                put_anon_vma(anon_vma);
 437                anon_vma = NULL;
 438        }
 439out:
 440        rcu_read_unlock();
 441
 442        return anon_vma;
 443}
 444
 445/*
 446 * Similar to page_get_anon_vma() except it locks the anon_vma.
 447 *
 448 * Its a little more complex as it tries to keep the fast path to a single
 449 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 450 * reference like with page_get_anon_vma() and then block on the mutex.
 451 */
 452struct anon_vma *page_lock_anon_vma(struct page *page)
 453{
 454        struct anon_vma *anon_vma = NULL;
 455        struct anon_vma *root_anon_vma;
 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        root_anon_vma = ACCESS_ONCE(anon_vma->root);
 467        if (mutex_trylock(&root_anon_vma->mutex)) {
 468                /*
 469                 * If the page is still mapped, then this anon_vma is still
 470                 * its anon_vma, and holding the mutex ensures that it will
 471                 * not go away, see anon_vma_free().
 472                 */
 473                if (!page_mapped(page)) {
 474                        mutex_unlock(&root_anon_vma->mutex);
 475                        anon_vma = NULL;
 476                }
 477                goto out;
 478        }
 479
 480        /* trylock failed, we got to sleep */
 481        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 482                anon_vma = NULL;
 483                goto out;
 484        }
 485
 486        if (!page_mapped(page)) {
 487                put_anon_vma(anon_vma);
 488                anon_vma = NULL;
 489                goto out;
 490        }
 491
 492        /* we pinned the anon_vma, its safe to sleep */
 493        rcu_read_unlock();
 494        anon_vma_lock(anon_vma);
 495
 496        if (atomic_dec_and_test(&anon_vma->refcount)) {
 497                /*
 498                 * Oops, we held the last refcount, release the lock
 499                 * and bail -- can't simply use put_anon_vma() because
 500                 * we'll deadlock on the anon_vma_lock() recursion.
 501                 */
 502                anon_vma_unlock(anon_vma);
 503                __put_anon_vma(anon_vma);
 504                anon_vma = NULL;
 505        }
 506
 507        return anon_vma;
 508
 509out:
 510        rcu_read_unlock();
 511        return anon_vma;
 512}
 513
 514void page_unlock_anon_vma(struct anon_vma *anon_vma)
 515{
 516        anon_vma_unlock(anon_vma);
 517}
 518
 519/*
 520 * At what user virtual address is page expected in @vma?
 521 * Returns virtual address or -EFAULT if page's index/offset is not
 522 * within the range mapped the @vma.
 523 */
 524inline unsigned long
 525vma_address(struct page *page, struct vm_area_struct *vma)
 526{
 527        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 528        unsigned long address;
 529
 530        if (unlikely(is_vm_hugetlb_page(vma)))
 531                pgoff = page->index << huge_page_order(page_hstate(page));
 532        address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 533        if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
 534                /* page should be within @vma mapping range */
 535                return -EFAULT;
 536        }
 537        return address;
 538}
 539
 540/*
 541 * At what user virtual address is page expected in vma?
 542 * Caller should check the page is actually part of the vma.
 543 */
 544unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 545{
 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        return vma_address(page, vma);
 562}
 563
 564/*
 565 * Check that @page is mapped at @address into @mm.
 566 *
 567 * If @sync is false, page_check_address may perform a racy check to avoid
 568 * the page table lock when the pte is not present (helpful when reclaiming
 569 * highly shared pages).
 570 *
 571 * On success returns with pte mapped and locked.
 572 */
 573pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
 574                          unsigned long address, spinlock_t **ptlp, int sync)
 575{
 576        pgd_t *pgd;
 577        pud_t *pud;
 578        pmd_t *pmd;
 579        pte_t *pte;
 580        spinlock_t *ptl;
 581
 582        if (unlikely(PageHuge(page))) {
 583                pte = huge_pte_offset(mm, address);
 584                ptl = &mm->page_table_lock;
 585                goto check;
 586        }
 587
 588        pgd = pgd_offset(mm, address);
 589        if (!pgd_present(*pgd))
 590                return NULL;
 591
 592        pud = pud_offset(pgd, address);
 593        if (!pud_present(*pud))
 594                return NULL;
 595
 596        pmd = pmd_offset(pud, address);
 597        if (!pmd_present(*pmd))
 598                return NULL;
 599        if (pmd_trans_huge(*pmd))
 600                return NULL;
 601
 602        pte = pte_offset_map(pmd, address);
 603        /* Make a quick check before getting the lock */
 604        if (!sync && !pte_present(*pte)) {
 605                pte_unmap(pte);
 606                return NULL;
 607        }
 608
 609        ptl = pte_lockptr(mm, pmd);
 610check:
 611        spin_lock(ptl);
 612        if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 613                *ptlp = ptl;
 614                return pte;
 615        }
 616        pte_unmap_unlock(pte, ptl);
 617        return NULL;
 618}
 619
 620/**
 621 * page_mapped_in_vma - check whether a page is really mapped in a VMA
 622 * @page: the page to test
 623 * @vma: the VMA to test
 624 *
 625 * Returns 1 if the page is mapped into the page tables of the VMA, 0
 626 * if the page is not mapped into the page tables of this VMA.  Only
 627 * valid for normal file or anonymous VMAs.
 628 */
 629int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 630{
 631        unsigned long address;
 632        pte_t *pte;
 633        spinlock_t *ptl;
 634
 635        address = vma_address(page, vma);
 636        if (address == -EFAULT)         /* out of vma range */
 637                return 0;
 638        pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
 639        if (!pte)                       /* the page is not in this mm */
 640                return 0;
 641        pte_unmap_unlock(pte, ptl);
 642
 643        return 1;
 644}
 645
 646/*
 647 * Subfunctions of page_referenced: page_referenced_one called
 648 * repeatedly from either page_referenced_anon or page_referenced_file.
 649 */
 650int page_referenced_one(struct page *page, struct vm_area_struct *vma,
 651                        unsigned long address, unsigned int *mapcount,
 652                        unsigned long *vm_flags)
 653{
 654        struct mm_struct *mm = vma->vm_mm;
 655        int referenced = 0;
 656
 657        if (unlikely(PageTransHuge(page))) {
 658                pmd_t *pmd;
 659
 660                spin_lock(&mm->page_table_lock);
 661                /*
 662                 * rmap might return false positives; we must filter
 663                 * these out using page_check_address_pmd().
 664                 */
 665                pmd = page_check_address_pmd(page, mm, address,
 666                                             PAGE_CHECK_ADDRESS_PMD_FLAG);
 667                if (!pmd) {
 668                        spin_unlock(&mm->page_table_lock);
 669                        goto out;
 670                }
 671
 672                if (vma->vm_flags & VM_LOCKED) {
 673                        spin_unlock(&mm->page_table_lock);
 674                        *mapcount = 0;  /* break early from loop */
 675                        *vm_flags |= VM_LOCKED;
 676                        goto out;
 677                }
 678
 679                /* go ahead even if the pmd is pmd_trans_splitting() */
 680                if (pmdp_clear_flush_young_notify(vma, address, pmd))
 681                        referenced++;
 682                spin_unlock(&mm->page_table_lock);
 683        } else {
 684                pte_t *pte;
 685                spinlock_t *ptl;
 686
 687                /*
 688                 * rmap might return false positives; we must filter
 689                 * these out using page_check_address().
 690                 */
 691                pte = page_check_address(page, mm, address, &ptl, 0);
 692                if (!pte)
 693                        goto out;
 694
 695                if (vma->vm_flags & VM_LOCKED) {
 696                        pte_unmap_unlock(pte, ptl);
 697                        *mapcount = 0;  /* break early from loop */
 698                        *vm_flags |= VM_LOCKED;
 699                        goto out;
 700                }
 701
 702                if (ptep_clear_flush_young_notify(vma, address, pte)) {
 703                        /*
 704                         * Don't treat a reference through a sequentially read
 705                         * mapping as such.  If the page has been used in
 706                         * another mapping, we will catch it; if this other
 707                         * mapping is already gone, the unmap path will have
 708                         * set PG_referenced or activated the page.
 709                         */
 710                        if (likely(!VM_SequentialReadHint(vma)))
 711                                referenced++;
 712                }
 713                pte_unmap_unlock(pte, ptl);
 714        }
 715
 716        /* Pretend the page is referenced if the task has the
 717           swap token and is in the middle of a page fault. */
 718        if (mm != current->mm && has_swap_token(mm) &&
 719                        rwsem_is_locked(&mm->mmap_sem))
 720                referenced++;
 721
 722        (*mapcount)--;
 723
 724        if (referenced)
 725                *vm_flags |= vma->vm_flags;
 726out:
 727        return referenced;
 728}
 729
 730static int page_referenced_anon(struct page *page,
 731                                struct mem_cgroup *mem_cont,
 732                                unsigned long *vm_flags)
 733{
 734        unsigned int mapcount;
 735        struct anon_vma *anon_vma;
 736        struct anon_vma_chain *avc;
 737        int referenced = 0;
 738
 739        anon_vma = page_lock_anon_vma(page);
 740        if (!anon_vma)
 741                return referenced;
 742
 743        mapcount = page_mapcount(page);
 744        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
 745                struct vm_area_struct *vma = avc->vma;
 746                unsigned long address = vma_address(page, vma);
 747                if (address == -EFAULT)
 748                        continue;
 749                /*
 750                 * If we are reclaiming on behalf of a cgroup, skip
 751                 * counting on behalf of references from different
 752                 * cgroups
 753                 */
 754                if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
 755                        continue;
 756                referenced += page_referenced_one(page, vma, address,
 757                                                  &mapcount, vm_flags);
 758                if (!mapcount)
 759                        break;
 760        }
 761
 762        page_unlock_anon_vma(anon_vma);
 763        return referenced;
 764}
 765
 766/**
 767 * page_referenced_file - referenced check for object-based rmap
 768 * @page: the page we're checking references on.
 769 * @mem_cont: target memory controller
 770 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 771 *
 772 * For an object-based mapped page, find all the places it is mapped and
 773 * check/clear the referenced flag.  This is done by following the page->mapping
 774 * pointer, then walking the chain of vmas it holds.  It returns the number
 775 * of references it found.
 776 *
 777 * This function is only called from page_referenced for object-based pages.
 778 */
 779static int page_referenced_file(struct page *page,
 780                                struct mem_cgroup *mem_cont,
 781                                unsigned long *vm_flags)
 782{
 783        unsigned int mapcount;
 784        struct address_space *mapping = page->mapping;
 785        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 786        struct vm_area_struct *vma;
 787        struct prio_tree_iter iter;
 788        int referenced = 0;
 789
 790        /*
 791         * The caller's checks on page->mapping and !PageAnon have made
 792         * sure that this is a file page: the check for page->mapping
 793         * excludes the case just before it gets set on an anon page.
 794         */
 795        BUG_ON(PageAnon(page));
 796
 797        /*
 798         * The page lock not only makes sure that page->mapping cannot
 799         * suddenly be NULLified by truncation, it makes sure that the
 800         * structure at mapping cannot be freed and reused yet,
 801         * so we can safely take mapping->i_mmap_mutex.
 802         */
 803        BUG_ON(!PageLocked(page));
 804
 805        mutex_lock(&mapping->i_mmap_mutex);
 806
 807        /*
 808         * i_mmap_mutex does not stabilize mapcount at all, but mapcount
 809         * is more likely to be accurate if we note it after spinning.
 810         */
 811        mapcount = page_mapcount(page);
 812
 813        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
 814                unsigned long address = vma_address(page, vma);
 815                if (address == -EFAULT)
 816                        continue;
 817                /*
 818                 * If we are reclaiming on behalf of a cgroup, skip
 819                 * counting on behalf of references from different
 820                 * cgroups
 821                 */
 822                if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
 823                        continue;
 824                referenced += page_referenced_one(page, vma, address,
 825                                                  &mapcount, vm_flags);
 826                if (!mapcount)
 827                        break;
 828        }
 829
 830        mutex_unlock(&mapping->i_mmap_mutex);
 831        return referenced;
 832}
 833
 834/**
 835 * page_referenced - test if the page was referenced
 836 * @page: the page to test
 837 * @is_locked: caller holds lock on the page
 838 * @mem_cont: target memory controller
 839 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 840 *
 841 * Quick test_and_clear_referenced for all mappings to a page,
 842 * returns the number of ptes which referenced the page.
 843 */
 844int page_referenced(struct page *page,
 845                    int is_locked,
 846                    struct mem_cgroup *mem_cont,
 847                    unsigned long *vm_flags)
 848{
 849        int referenced = 0;
 850        int we_locked = 0;
 851
 852        *vm_flags = 0;
 853        if (page_mapped(page) && page_rmapping(page)) {
 854                if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 855                        we_locked = trylock_page(page);
 856                        if (!we_locked) {
 857                                referenced++;
 858                                goto out;
 859                        }
 860                }
 861                if (unlikely(PageKsm(page)))
 862                        referenced += page_referenced_ksm(page, mem_cont,
 863                                                                vm_flags);
 864                else if (PageAnon(page))
 865                        referenced += page_referenced_anon(page, mem_cont,
 866                                                                vm_flags);
 867                else if (page->mapping)
 868                        referenced += page_referenced_file(page, mem_cont,
 869                                                                vm_flags);
 870                if (we_locked)
 871                        unlock_page(page);
 872
 873                if (page_test_and_clear_young(page_to_pfn(page)))
 874                        referenced++;
 875        }
 876out:
 877        return referenced;
 878}
 879
 880static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 881                            unsigned long address)
 882{
 883        struct mm_struct *mm = vma->vm_mm;
 884        pte_t *pte;
 885        spinlock_t *ptl;
 886        int ret = 0;
 887
 888        pte = page_check_address(page, mm, address, &ptl, 1);
 889        if (!pte)
 890                goto out;
 891
 892        if (pte_dirty(*pte) || pte_write(*pte)) {
 893                pte_t entry;
 894
 895                flush_cache_page(vma, address, pte_pfn(*pte));
 896                entry = ptep_clear_flush_notify(vma, address, pte);
 897                entry = pte_wrprotect(entry);
 898                entry = pte_mkclean(entry);
 899                set_pte_at(mm, address, pte, entry);
 900                ret = 1;
 901        }
 902
 903        pte_unmap_unlock(pte, ptl);
 904out:
 905        return ret;
 906}
 907
 908static int page_mkclean_file(struct address_space *mapping, struct page *page)
 909{
 910        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 911        struct vm_area_struct *vma;
 912        struct prio_tree_iter iter;
 913        int ret = 0;
 914
 915        BUG_ON(PageAnon(page));
 916
 917        mutex_lock(&mapping->i_mmap_mutex);
 918        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
 919                if (vma->vm_flags & VM_SHARED) {
 920                        unsigned long address = vma_address(page, vma);
 921                        if (address == -EFAULT)
 922                                continue;
 923                        ret += page_mkclean_one(page, vma, address);
 924                }
 925        }
 926        mutex_unlock(&mapping->i_mmap_mutex);
 927        return ret;
 928}
 929
 930int page_mkclean(struct page *page)
 931{
 932        int ret = 0;
 933
 934        BUG_ON(!PageLocked(page));
 935
 936        if (page_mapped(page)) {
 937                struct address_space *mapping = page_mapping(page);
 938                if (mapping) {
 939                        ret = page_mkclean_file(mapping, page);
 940                        if (page_test_and_clear_dirty(page_to_pfn(page), 1))
 941                                ret = 1;
 942                }
 943        }
 944
 945        return ret;
 946}
 947EXPORT_SYMBOL_GPL(page_mkclean);
 948
 949/**
 950 * page_move_anon_rmap - move a page to our anon_vma
 951 * @page:       the page to move to our anon_vma
 952 * @vma:        the vma the page belongs to
 953 * @address:    the user virtual address mapped
 954 *
 955 * When a page belongs exclusively to one process after a COW event,
 956 * that page can be moved into the anon_vma that belongs to just that
 957 * process, so the rmap code will not search the parent or sibling
 958 * processes.
 959 */
 960void page_move_anon_rmap(struct page *page,
 961        struct vm_area_struct *vma, unsigned long address)
 962{
 963        struct anon_vma *anon_vma = vma->anon_vma;
 964
 965        VM_BUG_ON(!PageLocked(page));
 966        VM_BUG_ON(!anon_vma);
 967        VM_BUG_ON(page->index != linear_page_index(vma, address));
 968
 969        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 970        page->mapping = (struct address_space *) anon_vma;
 971}
 972
 973/**
 974 * __page_set_anon_rmap - set up new anonymous rmap
 975 * @page:       Page to add to rmap     
 976 * @vma:        VM area to add page to.
 977 * @address:    User virtual address of the mapping     
 978 * @exclusive:  the page is exclusively owned by the current process
 979 */
 980static void __page_set_anon_rmap(struct page *page,
 981        struct vm_area_struct *vma, unsigned long address, int exclusive)
 982{
 983        struct anon_vma *anon_vma = vma->anon_vma;
 984
 985        BUG_ON(!anon_vma);
 986
 987        if (PageAnon(page))
 988                return;
 989
 990        /*
 991         * If the page isn't exclusively mapped into this vma,
 992         * we must use the _oldest_ possible anon_vma for the
 993         * page mapping!
 994         */
 995        if (!exclusive)
 996                anon_vma = anon_vma->root;
 997
 998        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 999        page->mapping = (struct address_space *) anon_vma;
1000        page->index = linear_page_index(vma, address);
1001}
1002
1003/**
1004 * __page_check_anon_rmap - sanity check anonymous rmap addition
1005 * @page:       the page to add the mapping to
1006 * @vma:        the vm area in which the mapping is added
1007 * @address:    the user virtual address mapped
1008 */
1009static void __page_check_anon_rmap(struct page *page,
1010        struct vm_area_struct *vma, unsigned long address)
1011{
1012#ifdef CONFIG_DEBUG_VM
1013        /*
1014         * The page's anon-rmap details (mapping and index) are guaranteed to
1015         * be set up correctly at this point.
1016         *
1017         * We have exclusion against page_add_anon_rmap because the caller
1018         * always holds the page locked, except if called from page_dup_rmap,
1019         * in which case the page is already known to be setup.
1020         *
1021         * We have exclusion against page_add_new_anon_rmap because those pages
1022         * are initially only visible via the pagetables, and the pte is locked
1023         * over the call to page_add_new_anon_rmap.
1024         */
1025        BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1026        BUG_ON(page->index != linear_page_index(vma, address));
1027#endif
1028}
1029
1030/**
1031 * page_add_anon_rmap - add pte mapping to an anonymous page
1032 * @page:       the page to add the mapping to
1033 * @vma:        the vm area in which the mapping is added
1034 * @address:    the user virtual address mapped
1035 *
1036 * The caller needs to hold the pte lock, and the page must be locked in
1037 * the anon_vma case: to serialize mapping,index checking after setting,
1038 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1039 * (but PageKsm is never downgraded to PageAnon).
1040 */
1041void page_add_anon_rmap(struct page *page,
1042        struct vm_area_struct *vma, unsigned long address)
1043{
1044        do_page_add_anon_rmap(page, vma, address, 0);
1045}
1046
1047/*
1048 * Special version of the above for do_swap_page, which often runs
1049 * into pages that are exclusively owned by the current process.
1050 * Everybody else should continue to use page_add_anon_rmap above.
1051 */
1052void do_page_add_anon_rmap(struct page *page,
1053        struct vm_area_struct *vma, unsigned long address, int exclusive)
1054{
1055        int first = atomic_inc_and_test(&page->_mapcount);
1056        if (first) {
1057                if (!PageTransHuge(page))
1058                        __inc_zone_page_state(page, NR_ANON_PAGES);
1059                else
1060                        __inc_zone_page_state(page,
1061                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1062        }
1063        if (unlikely(PageKsm(page)))
1064                return;
1065
1066        VM_BUG_ON(!PageLocked(page));
1067        /* address might be in next vma when migration races vma_adjust */
1068        if (first)
1069                __page_set_anon_rmap(page, vma, address, exclusive);
1070        else
1071                __page_check_anon_rmap(page, vma, address);
1072}
1073
1074/**
1075 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1076 * @page:       the page to add the mapping to
1077 * @vma:        the vm area in which the mapping is added
1078 * @address:    the user virtual address mapped
1079 *
1080 * Same as page_add_anon_rmap but must only be called on *new* pages.
1081 * This means the inc-and-test can be bypassed.
1082 * Page does not have to be locked.
1083 */
1084void page_add_new_anon_rmap(struct page *page,
1085        struct vm_area_struct *vma, unsigned long address)
1086{
1087        VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1088        SetPageSwapBacked(page);
1089        atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1090        if (!PageTransHuge(page))
1091                __inc_zone_page_state(page, NR_ANON_PAGES);
1092        else
1093                __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1094        __page_set_anon_rmap(page, vma, address, 1);
1095        if (page_evictable(page, vma))
1096                lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1097        else
1098                add_page_to_unevictable_list(page);
1099}
1100
1101/**
1102 * page_add_file_rmap - add pte mapping to a file page
1103 * @page: the page to add the mapping to
1104 *
1105 * The caller needs to hold the pte lock.
1106 */
1107void page_add_file_rmap(struct page *page)
1108{
1109        if (atomic_inc_and_test(&page->_mapcount)) {
1110                __inc_zone_page_state(page, NR_FILE_MAPPED);
1111                mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1112        }
1113}
1114
1115/**
1116 * page_remove_rmap - take down pte mapping from a page
1117 * @page: page to remove mapping from
1118 *
1119 * The caller needs to hold the pte lock.
1120 */
1121void page_remove_rmap(struct page *page)
1122{
1123        /* page still mapped by someone else? */
1124        if (!atomic_add_negative(-1, &page->_mapcount))
1125                return;
1126
1127        /*
1128         * Now that the last pte has gone, s390 must transfer dirty
1129         * flag from storage key to struct page.  We can usually skip
1130         * this if the page is anon, so about to be freed; but perhaps
1131         * not if it's in swapcache - there might be another pte slot
1132         * containing the swap entry, but page not yet written to swap.
1133         */
1134        if ((!PageAnon(page) || PageSwapCache(page)) &&
1135            page_test_and_clear_dirty(page_to_pfn(page), 1))
1136                set_page_dirty(page);
1137        /*
1138         * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1139         * and not charged by memcg for now.
1140         */
1141        if (unlikely(PageHuge(page)))
1142                return;
1143        if (PageAnon(page)) {
1144                mem_cgroup_uncharge_page(page);
1145                if (!PageTransHuge(page))
1146                        __dec_zone_page_state(page, NR_ANON_PAGES);
1147                else
1148                        __dec_zone_page_state(page,
1149                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1150        } else {
1151                __dec_zone_page_state(page, NR_FILE_MAPPED);
1152                mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1153        }
1154        /*
1155         * It would be tidy to reset the PageAnon mapping here,
1156         * but that might overwrite a racing page_add_anon_rmap
1157         * which increments mapcount after us but sets mapping
1158         * before us: so leave the reset to free_hot_cold_page,
1159         * and remember that it's only reliable while mapped.
1160         * Leaving it set also helps swapoff to reinstate ptes
1161         * faster for those pages still in swapcache.
1162         */
1163}
1164
1165/*
1166 * Subfunctions of try_to_unmap: try_to_unmap_one called
1167 * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
1168 */
1169int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1170                     unsigned long address, enum ttu_flags flags)
1171{
1172        struct mm_struct *mm = vma->vm_mm;
1173        pte_t *pte;
1174        pte_t pteval;
1175        spinlock_t *ptl;
1176        int ret = SWAP_AGAIN;
1177
1178        pte = page_check_address(page, mm, address, &ptl, 0);
1179        if (!pte)
1180                goto out;
1181
1182        /*
1183         * If the page is mlock()d, we cannot swap it out.
1184         * If it's recently referenced (perhaps page_referenced
1185         * skipped over this mm) then we should reactivate it.
1186         */
1187        if (!(flags & TTU_IGNORE_MLOCK)) {
1188                if (vma->vm_flags & VM_LOCKED)
1189                        goto out_mlock;
1190
1191                if (TTU_ACTION(flags) == TTU_MUNLOCK)
1192                        goto out_unmap;
1193        }
1194        if (!(flags & TTU_IGNORE_ACCESS)) {
1195                if (ptep_clear_flush_young_notify(vma, address, pte)) {
1196                        ret = SWAP_FAIL;
1197                        goto out_unmap;
1198                }
1199        }
1200
1201        /* Nuke the page table entry. */
1202        flush_cache_page(vma, address, page_to_pfn(page));
1203        pteval = ptep_clear_flush_notify(vma, address, pte);
1204
1205        /* Move the dirty bit to the physical page now the pte is gone. */
1206        if (pte_dirty(pteval))
1207                set_page_dirty(page);
1208
1209        /* Update high watermark before we lower rss */
1210        update_hiwater_rss(mm);
1211
1212        if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1213                if (PageAnon(page))
1214                        dec_mm_counter(mm, MM_ANONPAGES);
1215                else
1216                        dec_mm_counter(mm, MM_FILEPAGES);
1217                set_pte_at(mm, address, pte,
1218                                swp_entry_to_pte(make_hwpoison_entry(page)));
1219        } else if (PageAnon(page)) {
1220                swp_entry_t entry = { .val = page_private(page) };
1221
1222                if (PageSwapCache(page)) {
1223                        /*
1224                         * Store the swap location in the pte.
1225                         * See handle_pte_fault() ...
1226                         */
1227                        if (swap_duplicate(entry) < 0) {
1228                                set_pte_at(mm, address, pte, pteval);
1229                                ret = SWAP_FAIL;
1230                                goto out_unmap;
1231                        }
1232                        if (list_empty(&mm->mmlist)) {
1233                                spin_lock(&mmlist_lock);
1234                                if (list_empty(&mm->mmlist))
1235                                        list_add(&mm->mmlist, &init_mm.mmlist);
1236                                spin_unlock(&mmlist_lock);
1237                        }
1238                        dec_mm_counter(mm, MM_ANONPAGES);
1239                        inc_mm_counter(mm, MM_SWAPENTS);
1240                } else if (PAGE_MIGRATION) {
1241                        /*
1242                         * Store the pfn of the page in a special migration
1243                         * pte. do_swap_page() will wait until the migration
1244                         * pte is removed and then restart fault handling.
1245                         */
1246                        BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1247                        entry = make_migration_entry(page, pte_write(pteval));
1248                }
1249                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1250                BUG_ON(pte_file(*pte));
1251        } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) {
1252                /* Establish migration entry for a file page */
1253                swp_entry_t entry;
1254                entry = make_migration_entry(page, pte_write(pteval));
1255                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1256        } else
1257                dec_mm_counter(mm, MM_FILEPAGES);
1258
1259        page_remove_rmap(page);
1260        page_cache_release(page);
1261
1262out_unmap:
1263        pte_unmap_unlock(pte, ptl);
1264out:
1265        return ret;
1266
1267out_mlock:
1268        pte_unmap_unlock(pte, ptl);
1269
1270
1271        /*
1272         * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1273         * unstable result and race. Plus, We can't wait here because
1274         * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1275         * if trylock failed, the page remain in evictable lru and later
1276         * vmscan could retry to move the page to unevictable lru if the
1277         * page is actually mlocked.
1278         */
1279        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1280                if (vma->vm_flags & VM_LOCKED) {
1281                        mlock_vma_page(page);
1282                        ret = SWAP_MLOCK;
1283                }
1284                up_read(&vma->vm_mm->mmap_sem);
1285        }
1286        return ret;
1287}
1288
1289/*
1290 * objrmap doesn't work for nonlinear VMAs because the assumption that
1291 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1292 * Consequently, given a particular page and its ->index, we cannot locate the
1293 * ptes which are mapping that page without an exhaustive linear search.
1294 *
1295 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1296 * maps the file to which the target page belongs.  The ->vm_private_data field
1297 * holds the current cursor into that scan.  Successive searches will circulate
1298 * around the vma's virtual address space.
1299 *
1300 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1301 * more scanning pressure is placed against them as well.   Eventually pages
1302 * will become fully unmapped and are eligible for eviction.
1303 *
1304 * For very sparsely populated VMAs this is a little inefficient - chances are
1305 * there there won't be many ptes located within the scan cluster.  In this case
1306 * maybe we could scan further - to the end of the pte page, perhaps.
1307 *
1308 * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
1309 * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
1310 * rather than unmapping them.  If we encounter the "check_page" that vmscan is
1311 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1312 */
1313#define CLUSTER_SIZE    min(32*PAGE_SIZE, PMD_SIZE)
1314#define CLUSTER_MASK    (~(CLUSTER_SIZE - 1))
1315
1316static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1317                struct vm_area_struct *vma, struct page *check_page)
1318{
1319        struct mm_struct *mm = vma->vm_mm;
1320        pgd_t *pgd;
1321        pud_t *pud;
1322        pmd_t *pmd;
1323        pte_t *pte;
1324        pte_t pteval;
1325        spinlock_t *ptl;
1326        struct page *page;
1327        unsigned long address;
1328        unsigned long end;
1329        int ret = SWAP_AGAIN;
1330        int locked_vma = 0;
1331
1332        address = (vma->vm_start + cursor) & CLUSTER_MASK;
1333        end = address + CLUSTER_SIZE;
1334        if (address < vma->vm_start)
1335                address = vma->vm_start;
1336        if (end > vma->vm_end)
1337                end = vma->vm_end;
1338
1339        pgd = pgd_offset(mm, address);
1340        if (!pgd_present(*pgd))
1341                return ret;
1342
1343        pud = pud_offset(pgd, address);
1344        if (!pud_present(*pud))
1345                return ret;
1346
1347        pmd = pmd_offset(pud, address);
1348        if (!pmd_present(*pmd))
1349                return ret;
1350
1351        /*
1352         * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1353         * keep the sem while scanning the cluster for mlocking pages.
1354         */
1355        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1356                locked_vma = (vma->vm_flags & VM_LOCKED);
1357                if (!locked_vma)
1358                        up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1359        }
1360
1361        pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1362
1363        /* Update high watermark before we lower rss */
1364        update_hiwater_rss(mm);
1365
1366        for (; address < end; pte++, address += PAGE_SIZE) {
1367                if (!pte_present(*pte))
1368                        continue;
1369                page = vm_normal_page(vma, address, *pte);
1370                BUG_ON(!page || PageAnon(page));
1371
1372                if (locked_vma) {
1373                        mlock_vma_page(page);   /* no-op if already mlocked */
1374                        if (page == check_page)
1375                                ret = SWAP_MLOCK;
1376                        continue;       /* don't unmap */
1377                }
1378
1379                if (ptep_clear_flush_young_notify(vma, address, pte))
1380                        continue;
1381
1382                /* Nuke the page table entry. */
1383                flush_cache_page(vma, address, pte_pfn(*pte));
1384                pteval = ptep_clear_flush_notify(vma, address, pte);
1385
1386                /* If nonlinear, store the file page offset in the pte. */
1387                if (page->index != linear_page_index(vma, address))
1388                        set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1389
1390                /* Move the dirty bit to the physical page now the pte is gone. */
1391                if (pte_dirty(pteval))
1392                        set_page_dirty(page);
1393
1394                page_remove_rmap(page);
1395                page_cache_release(page);
1396                dec_mm_counter(mm, MM_FILEPAGES);
1397                (*mapcount)--;
1398        }
1399        pte_unmap_unlock(pte - 1, ptl);
1400        if (locked_vma)
1401                up_read(&vma->vm_mm->mmap_sem);
1402        return ret;
1403}
1404
1405bool is_vma_temporary_stack(struct vm_area_struct *vma)
1406{
1407        int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1408
1409        if (!maybe_stack)
1410                return false;
1411
1412        if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1413                                                VM_STACK_INCOMPLETE_SETUP)
1414                return true;
1415
1416        return false;
1417}
1418
1419/**
1420 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1421 * rmap method
1422 * @page: the page to unmap/unlock
1423 * @flags: action and flags
1424 *
1425 * Find all the mappings of a page using the mapping pointer and the vma chains
1426 * contained in the anon_vma struct it points to.
1427 *
1428 * This function is only called from try_to_unmap/try_to_munlock for
1429 * anonymous pages.
1430 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1431 * where the page was found will be held for write.  So, we won't recheck
1432 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1433 * 'LOCKED.
1434 */
1435static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1436{
1437        struct anon_vma *anon_vma;
1438        struct anon_vma_chain *avc;
1439        int ret = SWAP_AGAIN;
1440
1441        anon_vma = page_lock_anon_vma(page);
1442        if (!anon_vma)
1443                return ret;
1444
1445        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1446                struct vm_area_struct *vma = avc->vma;
1447                unsigned long address;
1448
1449                /*
1450                 * During exec, a temporary VMA is setup and later moved.
1451                 * The VMA is moved under the anon_vma lock but not the
1452                 * page tables leading to a race where migration cannot
1453                 * find the migration ptes. Rather than increasing the
1454                 * locking requirements of exec(), migration skips
1455                 * temporary VMAs until after exec() completes.
1456                 */
1457                if (PAGE_MIGRATION && (flags & TTU_MIGRATION) &&
1458                                is_vma_temporary_stack(vma))
1459                        continue;
1460
1461                address = vma_address(page, vma);
1462                if (address == -EFAULT)
1463                        continue;
1464                ret = try_to_unmap_one(page, vma, address, flags);
1465                if (ret != SWAP_AGAIN || !page_mapped(page))
1466                        break;
1467        }
1468
1469        page_unlock_anon_vma(anon_vma);
1470        return ret;
1471}
1472
1473/**
1474 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1475 * @page: the page to unmap/unlock
1476 * @flags: action and flags
1477 *
1478 * Find all the mappings of a page using the mapping pointer and the vma chains
1479 * contained in the address_space struct it points to.
1480 *
1481 * This function is only called from try_to_unmap/try_to_munlock for
1482 * object-based pages.
1483 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1484 * where the page was found will be held for write.  So, we won't recheck
1485 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1486 * 'LOCKED.
1487 */
1488static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1489{
1490        struct address_space *mapping = page->mapping;
1491        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1492        struct vm_area_struct *vma;
1493        struct prio_tree_iter iter;
1494        int ret = SWAP_AGAIN;
1495        unsigned long cursor;
1496        unsigned long max_nl_cursor = 0;
1497        unsigned long max_nl_size = 0;
1498        unsigned int mapcount;
1499
1500        mutex_lock(&mapping->i_mmap_mutex);
1501        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1502                unsigned long address = vma_address(page, vma);
1503                if (address == -EFAULT)
1504                        continue;
1505                ret = try_to_unmap_one(page, vma, address, flags);
1506                if (ret != SWAP_AGAIN || !page_mapped(page))
1507                        goto out;
1508        }
1509
1510        if (list_empty(&mapping->i_mmap_nonlinear))
1511                goto out;
1512
1513        /*
1514         * We don't bother to try to find the munlocked page in nonlinears.
1515         * It's costly. Instead, later, page reclaim logic may call
1516         * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1517         */
1518        if (TTU_ACTION(flags) == TTU_MUNLOCK)
1519                goto out;
1520
1521        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1522                                                shared.vm_set.list) {
1523                cursor = (unsigned long) vma->vm_private_data;
1524                if (cursor > max_nl_cursor)
1525                        max_nl_cursor = cursor;
1526                cursor = vma->vm_end - vma->vm_start;
1527                if (cursor > max_nl_size)
1528                        max_nl_size = cursor;
1529        }
1530
1531        if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1532                ret = SWAP_FAIL;
1533                goto out;
1534        }
1535
1536        /*
1537         * We don't try to search for this page in the nonlinear vmas,
1538         * and page_referenced wouldn't have found it anyway.  Instead
1539         * just walk the nonlinear vmas trying to age and unmap some.
1540         * The mapcount of the page we came in with is irrelevant,
1541         * but even so use it as a guide to how hard we should try?
1542         */
1543        mapcount = page_mapcount(page);
1544        if (!mapcount)
1545                goto out;
1546        cond_resched();
1547
1548        max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1549        if (max_nl_cursor == 0)
1550                max_nl_cursor = CLUSTER_SIZE;
1551
1552        do {
1553                list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1554                                                shared.vm_set.list) {
1555                        cursor = (unsigned long) vma->vm_private_data;
1556                        while ( cursor < max_nl_cursor &&
1557                                cursor < vma->vm_end - vma->vm_start) {
1558                                if (try_to_unmap_cluster(cursor, &mapcount,
1559                                                vma, page) == SWAP_MLOCK)
1560                                        ret = SWAP_MLOCK;
1561                                cursor += CLUSTER_SIZE;
1562                                vma->vm_private_data = (void *) cursor;
1563                                if ((int)mapcount <= 0)
1564                                        goto out;
1565                        }
1566                        vma->vm_private_data = (void *) max_nl_cursor;
1567                }
1568                cond_resched();
1569                max_nl_cursor += CLUSTER_SIZE;
1570        } while (max_nl_cursor <= max_nl_size);
1571
1572        /*
1573         * Don't loop forever (perhaps all the remaining pages are
1574         * in locked vmas).  Reset cursor on all unreserved nonlinear
1575         * vmas, now forgetting on which ones it had fallen behind.
1576         */
1577        list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1578                vma->vm_private_data = NULL;
1579out:
1580        mutex_unlock(&mapping->i_mmap_mutex);
1581        return ret;
1582}
1583
1584/**
1585 * try_to_unmap - try to remove all page table mappings to a page
1586 * @page: the page to get unmapped
1587 * @flags: action and flags
1588 *
1589 * Tries to remove all the page table entries which are mapping this
1590 * page, used in the pageout path.  Caller must hold the page lock.
1591 * Return values are:
1592 *
1593 * SWAP_SUCCESS - we succeeded in removing all mappings
1594 * SWAP_AGAIN   - we missed a mapping, try again later
1595 * SWAP_FAIL    - the page is unswappable
1596 * SWAP_MLOCK   - page is mlocked.
1597 */
1598int try_to_unmap(struct page *page, enum ttu_flags flags)
1599{
1600        int ret;
1601
1602        BUG_ON(!PageLocked(page));
1603        VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1604
1605        if (unlikely(PageKsm(page)))
1606                ret = try_to_unmap_ksm(page, flags);
1607        else if (PageAnon(page))
1608                ret = try_to_unmap_anon(page, flags);
1609        else
1610                ret = try_to_unmap_file(page, flags);
1611        if (ret != SWAP_MLOCK && !page_mapped(page))
1612                ret = SWAP_SUCCESS;
1613        return ret;
1614}
1615
1616/**
1617 * try_to_munlock - try to munlock a page
1618 * @page: the page to be munlocked
1619 *
1620 * Called from munlock code.  Checks all of the VMAs mapping the page
1621 * to make sure nobody else has this page mlocked. The page will be
1622 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1623 *
1624 * Return values are:
1625 *
1626 * SWAP_AGAIN   - no vma is holding page mlocked, or,
1627 * SWAP_AGAIN   - page mapped in mlocked vma -- couldn't acquire mmap sem
1628 * SWAP_FAIL    - page cannot be located at present
1629 * SWAP_MLOCK   - page is now mlocked.
1630 */
1631int try_to_munlock(struct page *page)
1632{
1633        VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1634
1635        if (unlikely(PageKsm(page)))
1636                return try_to_unmap_ksm(page, TTU_MUNLOCK);
1637        else if (PageAnon(page))
1638                return try_to_unmap_anon(page, TTU_MUNLOCK);
1639        else
1640                return try_to_unmap_file(page, TTU_MUNLOCK);
1641}
1642
1643void __put_anon_vma(struct anon_vma *anon_vma)
1644{
1645        struct anon_vma *root = anon_vma->root;
1646
1647        if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1648                anon_vma_free(root);
1649
1650        anon_vma_free(anon_vma);
1651}
1652
1653#ifdef CONFIG_MIGRATION
1654/*
1655 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1656 * Called by migrate.c to remove migration ptes, but might be used more later.
1657 */
1658static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1659                struct vm_area_struct *, unsigned long, void *), void *arg)
1660{
1661        struct anon_vma *anon_vma;
1662        struct anon_vma_chain *avc;
1663        int ret = SWAP_AGAIN;
1664
1665        /*
1666         * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1667         * because that depends on page_mapped(); but not all its usages
1668         * are holding mmap_sem. Users without mmap_sem are required to
1669         * take a reference count to prevent the anon_vma disappearing
1670         */
1671        anon_vma = page_anon_vma(page);
1672        if (!anon_vma)
1673                return ret;
1674        anon_vma_lock(anon_vma);
1675        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1676                struct vm_area_struct *vma = avc->vma;
1677                unsigned long address = vma_address(page, vma);
1678                if (address == -EFAULT)
1679                        continue;
1680                ret = rmap_one(page, vma, address, arg);
1681                if (ret != SWAP_AGAIN)
1682                        break;
1683        }
1684        anon_vma_unlock(anon_vma);
1685        return ret;
1686}
1687
1688static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1689                struct vm_area_struct *, unsigned long, void *), void *arg)
1690{
1691        struct address_space *mapping = page->mapping;
1692        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1693        struct vm_area_struct *vma;
1694        struct prio_tree_iter iter;
1695        int ret = SWAP_AGAIN;
1696
1697        if (!mapping)
1698                return ret;
1699        mutex_lock(&mapping->i_mmap_mutex);
1700        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1701                unsigned long address = vma_address(page, vma);
1702                if (address == -EFAULT)
1703                        continue;
1704                ret = rmap_one(page, vma, address, arg);
1705                if (ret != SWAP_AGAIN)
1706                        break;
1707        }
1708        /*
1709         * No nonlinear handling: being always shared, nonlinear vmas
1710         * never contain migration ptes.  Decide what to do about this
1711         * limitation to linear when we need rmap_walk() on nonlinear.
1712         */
1713        mutex_unlock(&mapping->i_mmap_mutex);
1714        return ret;
1715}
1716
1717int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1718                struct vm_area_struct *, unsigned long, void *), void *arg)
1719{
1720        VM_BUG_ON(!PageLocked(page));
1721
1722        if (unlikely(PageKsm(page)))
1723                return rmap_walk_ksm(page, rmap_one, arg);
1724        else if (PageAnon(page))
1725                return rmap_walk_anon(page, rmap_one, arg);
1726        else
1727                return rmap_walk_file(page, rmap_one, arg);
1728}
1729#endif /* CONFIG_MIGRATION */
1730
1731#ifdef CONFIG_HUGETLB_PAGE
1732/*
1733 * The following three functions are for anonymous (private mapped) hugepages.
1734 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1735 * and no lru code, because we handle hugepages differently from common pages.
1736 */
1737static void __hugepage_set_anon_rmap(struct page *page,
1738        struct vm_area_struct *vma, unsigned long address, int exclusive)
1739{
1740        struct anon_vma *anon_vma = vma->anon_vma;
1741
1742        BUG_ON(!anon_vma);
1743
1744        if (PageAnon(page))
1745                return;
1746        if (!exclusive)
1747                anon_vma = anon_vma->root;
1748
1749        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1750        page->mapping = (struct address_space *) anon_vma;
1751        page->index = linear_page_index(vma, address);
1752}
1753
1754void hugepage_add_anon_rmap(struct page *page,
1755                            struct vm_area_struct *vma, unsigned long address)
1756{
1757        struct anon_vma *anon_vma = vma->anon_vma;
1758        int first;
1759
1760        BUG_ON(!PageLocked(page));
1761        BUG_ON(!anon_vma);
1762        /* address might be in next vma when migration races vma_adjust */
1763        first = atomic_inc_and_test(&page->_mapcount);
1764        if (first)
1765                __hugepage_set_anon_rmap(page, vma, address, 0);
1766}
1767
1768void hugepage_add_new_anon_rmap(struct page *page,
1769                        struct vm_area_struct *vma, unsigned long address)
1770{
1771        BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1772        atomic_set(&page->_mapcount, 0);
1773        __hugepage_set_anon_rmap(page, vma, address, 1);
1774}
1775#endif /* CONFIG_HUGETLB_PAGE */
1776
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.