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        anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 132}
 133
 134/**
 135 * anon_vma_prepare - attach an anon_vma to a memory region
 136 * @vma: the memory region in question
 137 *
 138 * This makes sure the memory mapping described by 'vma' has
 139 * an 'anon_vma' attached to it, so that we can associate the
 140 * anonymous pages mapped into it with that anon_vma.
 141 *
 142 * The common case will be that we already have one, but if
 143 * not we either need to find an adjacent mapping that we
 144 * can re-use the anon_vma from (very common when the only
 145 * reason for splitting a vma has been mprotect()), or we
 146 * allocate a new one.
 147 *
 148 * Anon-vma allocations are very subtle, because we may have
 149 * optimistically looked up an anon_vma in page_lock_anon_vma()
 150 * and that may actually touch the spinlock even in the newly
 151 * allocated vma (it depends on RCU to make sure that the
 152 * anon_vma isn't actually destroyed).
 153 *
 154 * As a result, we need to do proper anon_vma locking even
 155 * for the new allocation. At the same time, we do not want
 156 * to do any locking for the common case of already having
 157 * an anon_vma.
 158 *
 159 * This must be called with the mmap_sem held for reading.
 160 */
 161int anon_vma_prepare(struct vm_area_struct *vma)
 162{
 163        struct anon_vma *anon_vma = vma->anon_vma;
 164        struct anon_vma_chain *avc;
 165
 166        might_sleep();
 167        if (unlikely(!anon_vma)) {
 168                struct mm_struct *mm = vma->vm_mm;
 169                struct anon_vma *allocated;
 170
 171                avc = anon_vma_chain_alloc(GFP_KERNEL);
 172                if (!avc)
 173                        goto out_enomem;
 174
 175                anon_vma = find_mergeable_anon_vma(vma);
 176                allocated = NULL;
 177                if (!anon_vma) {
 178                        anon_vma = anon_vma_alloc();
 179                        if (unlikely(!anon_vma))
 180                                goto out_enomem_free_avc;
 181                        allocated = anon_vma;
 182                }
 183
 184                anon_vma_lock(anon_vma);
 185                /* page_table_lock to protect against threads */
 186                spin_lock(&mm->page_table_lock);
 187                if (likely(!vma->anon_vma)) {
 188                        vma->anon_vma = anon_vma;
 189                        anon_vma_chain_link(vma, avc, anon_vma);
 190                        allocated = NULL;
 191                        avc = NULL;
 192                }
 193                spin_unlock(&mm->page_table_lock);
 194                anon_vma_unlock(anon_vma);
 195
 196                if (unlikely(allocated))
 197                        put_anon_vma(allocated);
 198                if (unlikely(avc))
 199                        anon_vma_chain_free(avc);
 200        }
 201        return 0;
 202
 203 out_enomem_free_avc:
 204        anon_vma_chain_free(avc);
 205 out_enomem:
 206        return -ENOMEM;
 207}
 208
 209/*
 210 * This is a useful helper function for locking the anon_vma root as
 211 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 212 * have the same vma.
 213 *
 214 * Such anon_vma's should have the same root, so you'd expect to see
 215 * just a single mutex_lock for the whole traversal.
 216 */
 217static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 218{
 219        struct anon_vma *new_root = anon_vma->root;
 220        if (new_root != root) {
 221                if (WARN_ON_ONCE(root))
 222                        mutex_unlock(&root->mutex);
 223                root = new_root;
 224                mutex_lock(&root->mutex);
 225        }
 226        return root;
 227}
 228
 229static inline void unlock_anon_vma_root(struct anon_vma *root)
 230{
 231        if (root)
 232                mutex_unlock(&root->mutex);
 233}
 234
 235/*
 236 * Attach the anon_vmas from src to dst.
 237 * Returns 0 on success, -ENOMEM on failure.
 238 */
 239int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 240{
 241        struct anon_vma_chain *avc, *pavc;
 242        struct anon_vma *root = NULL;
 243
 244        list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 245                struct anon_vma *anon_vma;
 246
 247                avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 248                if (unlikely(!avc)) {
 249                        unlock_anon_vma_root(root);
 250                        root = NULL;
 251                        avc = anon_vma_chain_alloc(GFP_KERNEL);
 252                        if (!avc)
 253                                goto enomem_failure;
 254                }
 255                anon_vma = pavc->anon_vma;
 256                root = lock_anon_vma_root(root, anon_vma);
 257                anon_vma_chain_link(dst, avc, anon_vma);
 258        }
 259        unlock_anon_vma_root(root);
 260        return 0;
 261
 262 enomem_failure:
 263        unlink_anon_vmas(dst);
 264        return -ENOMEM;
 265}
 266
 267/*
 268 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 269 * the corresponding VMA in the parent process is attached to.
 270 * Returns 0 on success, non-zero on failure.
 271 */
 272int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 273{
 274        struct anon_vma_chain *avc;
 275        struct anon_vma *anon_vma;
 276
 277        /* Don't bother if the parent process has no anon_vma here. */
 278        if (!pvma->anon_vma)
 279                return 0;
 280
 281        /*
 282         * First, attach the new VMA to the parent VMA's anon_vmas,
 283         * so rmap can find non-COWed pages in child processes.
 284         */
 285        if (anon_vma_clone(vma, pvma))
 286                return -ENOMEM;
 287
 288        /* Then add our own anon_vma. */
 289        anon_vma = anon_vma_alloc();
 290        if (!anon_vma)
 291                goto out_error;
 292        avc = anon_vma_chain_alloc(GFP_KERNEL);
 293        if (!avc)
 294                goto out_error_free_anon_vma;
 295
 296        /*
 297         * The root anon_vma's spinlock is the lock actually used when we
 298         * lock any of the anon_vmas in this anon_vma tree.
 299         */
 300        anon_vma->root = pvma->anon_vma->root;
 301        /*
 302         * With refcounts, an anon_vma can stay around longer than the
 303         * process it belongs to. The root anon_vma needs to be pinned until
 304         * this anon_vma is freed, because the lock lives in the root.
 305         */
 306        get_anon_vma(anon_vma->root);
 307        /* Mark this anon_vma as the one where our new (COWed) pages go. */
 308        vma->anon_vma = anon_vma;
 309        anon_vma_lock(anon_vma);
 310        anon_vma_chain_link(vma, avc, anon_vma);
 311        anon_vma_unlock(anon_vma);
 312
 313        return 0;
 314
 315 out_error_free_anon_vma:
 316        put_anon_vma(anon_vma);
 317 out_error:
 318        unlink_anon_vmas(vma);
 319        return -ENOMEM;
 320}
 321
 322void unlink_anon_vmas(struct vm_area_struct *vma)
 323{
 324        struct anon_vma_chain *avc, *next;
 325        struct anon_vma *root = NULL;
 326
 327        /*
 328         * Unlink each anon_vma chained to the VMA.  This list is ordered
 329         * from newest to oldest, ensuring the root anon_vma gets freed last.
 330         */
 331        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 332                struct anon_vma *anon_vma = avc->anon_vma;
 333
 334                root = lock_anon_vma_root(root, anon_vma);
 335                anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 336
 337                /*
 338                 * Leave empty anon_vmas on the list - we'll need
 339                 * to free them outside the lock.
 340                 */
 341                if (RB_EMPTY_ROOT(&anon_vma->rb_root))
 342                        continue;
 343
 344                list_del(&avc->same_vma);
 345                anon_vma_chain_free(avc);
 346        }
 347        unlock_anon_vma_root(root);
 348
 349        /*
 350         * Iterate the list once more, it now only contains empty and unlinked
 351         * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 352         * needing to acquire the anon_vma->root->mutex.
 353         */
 354        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 355                struct anon_vma *anon_vma = avc->anon_vma;
 356
 357                put_anon_vma(anon_vma);
 358
 359                list_del(&avc->same_vma);
 360                anon_vma_chain_free(avc);
 361        }
 362}
 363
 364static void anon_vma_ctor(void *data)
 365{
 366        struct anon_vma *anon_vma = data;
 367
 368        mutex_init(&anon_vma->mutex);
 369        atomic_set(&anon_vma->refcount, 0);
 370        anon_vma->rb_root = RB_ROOT;
 371}
 372
 373void __init anon_vma_init(void)
 374{
 375        anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 376                        0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
 377        anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
 378}
 379
 380/*
 381 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 382 *
 383 * Since there is no serialization what so ever against page_remove_rmap()
 384 * the best this function can do is return a locked anon_vma that might
 385 * have been relevant to this page.
 386 *
 387 * The page might have been remapped to a different anon_vma or the anon_vma
 388 * returned may already be freed (and even reused).
 389 *
 390 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 391 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 392 * ensure that any anon_vma obtained from the page will still be valid for as
 393 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 394 *
 395 * All users of this function must be very careful when walking the anon_vma
 396 * chain and verify that the page in question is indeed mapped in it
 397 * [ something equivalent to page_mapped_in_vma() ].
 398 *
 399 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 400 * that the anon_vma pointer from page->mapping is valid if there is a
 401 * mapcount, we can dereference the anon_vma after observing those.
 402 */
 403struct anon_vma *page_get_anon_vma(struct page *page)
 404{
 405        struct anon_vma *anon_vma = NULL;
 406        unsigned long anon_mapping;
 407
 408        rcu_read_lock();
 409        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 410        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 411                goto out;
 412        if (!page_mapped(page))
 413                goto out;
 414
 415        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 416        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 417                anon_vma = NULL;
 418                goto out;
 419        }
 420
 421        /*
 422         * If this page is still mapped, then its anon_vma cannot have been
 423         * freed.  But if it has been unmapped, we have no security against the
 424         * anon_vma structure being freed and reused (for another anon_vma:
 425         * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
 426         * above cannot corrupt).
 427         */
 428        if (!page_mapped(page)) {
 429                put_anon_vma(anon_vma);
 430                anon_vma = NULL;
 431        }
 432out:
 433        rcu_read_unlock();
 434
 435        return anon_vma;
 436}
 437
 438/*
 439 * Similar to page_get_anon_vma() except it locks the anon_vma.
 440 *
 441 * Its a little more complex as it tries to keep the fast path to a single
 442 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 443 * reference like with page_get_anon_vma() and then block on the mutex.
 444 */
 445struct anon_vma *page_lock_anon_vma(struct page *page)
 446{
 447        struct anon_vma *anon_vma = NULL;
 448        struct anon_vma *root_anon_vma;
 449        unsigned long anon_mapping;
 450
 451        rcu_read_lock();
 452        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 453        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 454                goto out;
 455        if (!page_mapped(page))
 456                goto out;
 457
 458        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 459        root_anon_vma = ACCESS_ONCE(anon_vma->root);
 460        if (mutex_trylock(&root_anon_vma->mutex)) {
 461                /*
 462                 * If the page is still mapped, then this anon_vma is still
 463                 * its anon_vma, and holding the mutex ensures that it will
 464                 * not go away, see anon_vma_free().
 465                 */
 466                if (!page_mapped(page)) {
 467                        mutex_unlock(&root_anon_vma->mutex);
 468                        anon_vma = NULL;
 469                }
 470                goto out;
 471        }
 472
 473        /* trylock failed, we got to sleep */
 474        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 475                anon_vma = NULL;
 476                goto out;
 477        }
 478
 479        if (!page_mapped(page)) {
 480                put_anon_vma(anon_vma);
 481                anon_vma = NULL;
 482                goto out;
 483        }
 484
 485        /* we pinned the anon_vma, its safe to sleep */
 486        rcu_read_unlock();
 487        anon_vma_lock(anon_vma);
 488
 489        if (atomic_dec_and_test(&anon_vma->refcount)) {
 490                /*
 491                 * Oops, we held the last refcount, release the lock
 492                 * and bail -- can't simply use put_anon_vma() because
 493                 * we'll deadlock on the anon_vma_lock() recursion.
 494                 */
 495                anon_vma_unlock(anon_vma);
 496                __put_anon_vma(anon_vma);
 497                anon_vma = NULL;
 498        }
 499
 500        return anon_vma;
 501
 502out:
 503        rcu_read_unlock();
 504        return anon_vma;
 505}
 506
 507void page_unlock_anon_vma(struct anon_vma *anon_vma)
 508{
 509        anon_vma_unlock(anon_vma);
 510}
 511
 512/*
 513 * At what user virtual address is page expected in @vma?
 514 */
 515static inline unsigned long
 516__vma_address(struct page *page, struct vm_area_struct *vma)
 517{
 518        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 519
 520        if (unlikely(is_vm_hugetlb_page(vma)))
 521                pgoff = page->index << huge_page_order(page_hstate(page));
 522
 523        return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 524}
 525
 526inline unsigned long
 527vma_address(struct page *page, struct vm_area_struct *vma)
 528{
 529        unsigned long address = __vma_address(page, vma);
 530
 531        /* page should be within @vma mapping range */
 532        VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
 533
 534        return address;
 535}
 536
 537/*
 538 * At what user virtual address is page expected in vma?
 539 * Caller should check the page is actually part of the vma.
 540 */
 541unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 542{
 543        unsigned long address;
 544        if (PageAnon(page)) {
 545                struct anon_vma *page__anon_vma = page_anon_vma(page);
 546                /*
 547                 * Note: swapoff's unuse_vma() is more efficient with this
 548                 * check, and needs it to match anon_vma when KSM is active.
 549                 */
 550                if (!vma->anon_vma || !page__anon_vma ||
 551                    vma->anon_vma->root != page__anon_vma->root)
 552                        return -EFAULT;
 553        } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
 554                if (!vma->vm_file ||
 555                    vma->vm_file->f_mapping != page->mapping)
 556                        return -EFAULT;
 557        } else
 558                return -EFAULT;
 559        address = __vma_address(page, vma);
 560        if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 561                return -EFAULT;
 562        return address;
 563}
 564
 565/*
 566 * Check that @page is mapped at @address into @mm.
 567 *
 568 * If @sync is false, page_check_address may perform a racy check to avoid
 569 * the page table lock when the pte is not present (helpful when reclaiming
 570 * highly shared pages).
 571 *
 572 * On success returns with pte mapped and locked.
 573 */
 574pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
 575                          unsigned long address, spinlock_t **ptlp, int sync)
 576{
 577        pgd_t *pgd;
 578        pud_t *pud;
 579        pmd_t *pmd;
 580        pte_t *pte;
 581        spinlock_t *ptl;
 582
 583        if (unlikely(PageHuge(page))) {
 584                pte = huge_pte_offset(mm, address);
 585                ptl = &mm->page_table_lock;
 586                goto check;
 587        }
 588
 589        pgd = pgd_offset(mm, address);
 590        if (!pgd_present(*pgd))
 591                return NULL;
 592
 593        pud = pud_offset(pgd, address);
 594        if (!pud_present(*pud))
 595                return NULL;
 596
 597        pmd = pmd_offset(pud, address);
 598        if (!pmd_present(*pmd))
 599                return NULL;
 600        if (pmd_trans_huge(*pmd))
 601                return NULL;
 602
 603        pte = pte_offset_map(pmd, address);
 604        /* Make a quick check before getting the lock */
 605        if (!sync && !pte_present(*pte)) {
 606                pte_unmap(pte);
 607                return NULL;
 608        }
 609
 610        ptl = pte_lockptr(mm, pmd);
 611check:
 612        spin_lock(ptl);
 613        if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 614                *ptlp = ptl;
 615                return pte;
 616        }
 617        pte_unmap_unlock(pte, ptl);
 618        return NULL;
 619}
 620
 621/**
 622 * page_mapped_in_vma - check whether a page is really mapped in a VMA
 623 * @page: the page to test
 624 * @vma: the VMA to test
 625 *
 626 * Returns 1 if the page is mapped into the page tables of the VMA, 0
 627 * if the page is not mapped into the page tables of this VMA.  Only
 628 * valid for normal file or anonymous VMAs.
 629 */
 630int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 631{
 632        unsigned long address;
 633        pte_t *pte;
 634        spinlock_t *ptl;
 635
 636        address = __vma_address(page, vma);
 637        if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 638                return 0;
 639        pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
 640        if (!pte)                       /* the page is not in this mm */
 641                return 0;
 642        pte_unmap_unlock(pte, ptl);
 643
 644        return 1;
 645}
 646
 647/*
 648 * Subfunctions of page_referenced: page_referenced_one called
 649 * repeatedly from either page_referenced_anon or page_referenced_file.
 650 */
 651int page_referenced_one(struct page *page, struct vm_area_struct *vma,
 652                        unsigned long address, unsigned int *mapcount,
 653                        unsigned long *vm_flags)
 654{
 655        struct mm_struct *mm = vma->vm_mm;
 656        int referenced = 0;
 657
 658        if (unlikely(PageTransHuge(page))) {
 659                pmd_t *pmd;
 660
 661                spin_lock(&mm->page_table_lock);
 662                /*
 663                 * rmap might return false positives; we must filter
 664                 * these out using page_check_address_pmd().
 665                 */
 666                pmd = page_check_address_pmd(page, mm, address,
 667                                             PAGE_CHECK_ADDRESS_PMD_FLAG);
 668                if (!pmd) {
 669                        spin_unlock(&mm->page_table_lock);
 670                        goto out;
 671                }
 672
 673                if (vma->vm_flags & VM_LOCKED) {
 674                        spin_unlock(&mm->page_table_lock);
 675                        *mapcount = 0;  /* break early from loop */
 676                        *vm_flags |= VM_LOCKED;
 677                        goto out;
 678                }
 679
 680                /* go ahead even if the pmd is pmd_trans_splitting() */
 681                if (pmdp_clear_flush_young_notify(vma, address, pmd))
 682                        referenced++;
 683                spin_unlock(&mm->page_table_lock);
 684        } else {
 685                pte_t *pte;
 686                spinlock_t *ptl;
 687
 688                /*
 689                 * rmap might return false positives; we must filter
 690                 * these out using page_check_address().
 691                 */
 692                pte = page_check_address(page, mm, address, &ptl, 0);
 693                if (!pte)
 694                        goto out;
 695
 696                if (vma->vm_flags & VM_LOCKED) {
 697                        pte_unmap_unlock(pte, ptl);
 698                        *mapcount = 0;  /* break early from loop */
 699                        *vm_flags |= VM_LOCKED;
 700                        goto out;
 701                }
 702
 703                if (ptep_clear_flush_young_notify(vma, address, pte)) {
 704                        /*
 705                         * Don't treat a reference through a sequentially read
 706                         * mapping as such.  If the page has been used in
 707                         * another mapping, we will catch it; if this other
 708                         * mapping is already gone, the unmap path will have
 709                         * set PG_referenced or activated the page.
 710                         */
 711                        if (likely(!VM_SequentialReadHint(vma)))
 712                                referenced++;
 713                }
 714                pte_unmap_unlock(pte, ptl);
 715        }
 716
 717        (*mapcount)--;
 718
 719        if (referenced)
 720                *vm_flags |= vma->vm_flags;
 721out:
 722        return referenced;
 723}
 724
 725static int page_referenced_anon(struct page *page,
 726                                struct mem_cgroup *memcg,
 727                                unsigned long *vm_flags)
 728{
 729        unsigned int mapcount;
 730        struct anon_vma *anon_vma;
 731        pgoff_t pgoff;
 732        struct anon_vma_chain *avc;
 733        int referenced = 0;
 734
 735        anon_vma = page_lock_anon_vma(page);
 736        if (!anon_vma)
 737                return referenced;
 738
 739        mapcount = page_mapcount(page);
 740        pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 741        anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
 742                struct vm_area_struct *vma = avc->vma;
 743                unsigned long address = vma_address(page, vma);
 744                /*
 745                 * If we are reclaiming on behalf of a cgroup, skip
 746                 * counting on behalf of references from different
 747                 * cgroups
 748                 */
 749                if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 750                        continue;
 751                referenced += page_referenced_one(page, vma, address,
 752                                                  &mapcount, vm_flags);
 753                if (!mapcount)
 754                        break;
 755        }
 756
 757        page_unlock_anon_vma(anon_vma);
 758        return referenced;
 759}
 760
 761/**
 762 * page_referenced_file - referenced check for object-based rmap
 763 * @page: the page we're checking references on.
 764 * @memcg: target memory control group
 765 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 766 *
 767 * For an object-based mapped page, find all the places it is mapped and
 768 * check/clear the referenced flag.  This is done by following the page->mapping
 769 * pointer, then walking the chain of vmas it holds.  It returns the number
 770 * of references it found.
 771 *
 772 * This function is only called from page_referenced for object-based pages.
 773 */
 774static int page_referenced_file(struct page *page,
 775                                struct mem_cgroup *memcg,
 776                                unsigned long *vm_flags)
 777{
 778        unsigned int mapcount;
 779        struct address_space *mapping = page->mapping;
 780        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 781        struct vm_area_struct *vma;
 782        int referenced = 0;
 783
 784        /*
 785         * The caller's checks on page->mapping and !PageAnon have made
 786         * sure that this is a file page: the check for page->mapping
 787         * excludes the case just before it gets set on an anon page.
 788         */
 789        BUG_ON(PageAnon(page));
 790
 791        /*
 792         * The page lock not only makes sure that page->mapping cannot
 793         * suddenly be NULLified by truncation, it makes sure that the
 794         * structure at mapping cannot be freed and reused yet,
 795         * so we can safely take mapping->i_mmap_mutex.
 796         */
 797        BUG_ON(!PageLocked(page));
 798
 799        mutex_lock(&mapping->i_mmap_mutex);
 800
 801        /*
 802         * i_mmap_mutex does not stabilize mapcount at all, but mapcount
 803         * is more likely to be accurate if we note it after spinning.
 804         */
 805        mapcount = page_mapcount(page);
 806
 807        vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 808                unsigned long address = vma_address(page, vma);
 809                /*
 810                 * If we are reclaiming on behalf of a cgroup, skip
 811                 * counting on behalf of references from different
 812                 * cgroups
 813                 */
 814                if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 815                        continue;
 816                referenced += page_referenced_one(page, vma, address,
 817                                                  &mapcount, vm_flags);
 818                if (!mapcount)
 819                        break;
 820        }
 821
 822        mutex_unlock(&mapping->i_mmap_mutex);
 823        return referenced;
 824}
 825
 826/**
 827 * page_referenced - test if the page was referenced
 828 * @page: the page to test
 829 * @is_locked: caller holds lock on the page
 830 * @memcg: target memory cgroup
 831 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 832 *
 833 * Quick test_and_clear_referenced for all mappings to a page,
 834 * returns the number of ptes which referenced the page.
 835 */
 836int page_referenced(struct page *page,
 837                    int is_locked,
 838                    struct mem_cgroup *memcg,
 839                    unsigned long *vm_flags)
 840{
 841        int referenced = 0;
 842        int we_locked = 0;
 843
 844        *vm_flags = 0;
 845        if (page_mapped(page) && page_rmapping(page)) {
 846                if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 847                        we_locked = trylock_page(page);
 848                        if (!we_locked) {
 849                                referenced++;
 850                                goto out;
 851                        }
 852                }
 853                if (unlikely(PageKsm(page)))
 854                        referenced += page_referenced_ksm(page, memcg,
 855                                                                vm_flags);
 856                else if (PageAnon(page))
 857                        referenced += page_referenced_anon(page, memcg,
 858                                                                vm_flags);
 859                else if (page->mapping)
 860                        referenced += page_referenced_file(page, memcg,
 861                                                                vm_flags);
 862                if (we_locked)
 863                        unlock_page(page);
 864
 865                if (page_test_and_clear_young(page_to_pfn(page)))
 866                        referenced++;
 867        }
 868out:
 869        return referenced;
 870}
 871
 872static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 873                            unsigned long address)
 874{
 875        struct mm_struct *mm = vma->vm_mm;
 876        pte_t *pte;
 877        spinlock_t *ptl;
 878        int ret = 0;
 879
 880        pte = page_check_address(page, mm, address, &ptl, 1);
 881        if (!pte)
 882                goto out;
 883
 884        if (pte_dirty(*pte) || pte_write(*pte)) {
 885                pte_t entry;
 886
 887                flush_cache_page(vma, address, pte_pfn(*pte));
 888                entry = ptep_clear_flush(vma, address, pte);
 889                entry = pte_wrprotect(entry);
 890                entry = pte_mkclean(entry);
 891                set_pte_at(mm, address, pte, entry);
 892                ret = 1;
 893        }
 894
 895        pte_unmap_unlock(pte, ptl);
 896
 897        if (ret)
 898                mmu_notifier_invalidate_page(mm, address);
 899out:
 900        return ret;
 901}
 902
 903static int page_mkclean_file(struct address_space *mapping, struct page *page)
 904{
 905        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 906        struct vm_area_struct *vma;
 907        int ret = 0;
 908
 909        BUG_ON(PageAnon(page));
 910
 911        mutex_lock(&mapping->i_mmap_mutex);
 912        vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 913                if (vma->vm_flags & VM_SHARED) {
 914                        unsigned long address = vma_address(page, vma);
 915                        ret += page_mkclean_one(page, vma, address);
 916                }
 917        }
 918        mutex_unlock(&mapping->i_mmap_mutex);
 919        return ret;
 920}
 921
 922int page_mkclean(struct page *page)
 923{
 924        int ret = 0;
 925
 926        BUG_ON(!PageLocked(page));
 927
 928        if (page_mapped(page)) {
 929                struct address_space *mapping = page_mapping(page);
 930                if (mapping)
 931                        ret = page_mkclean_file(mapping, page);
 932        }
 933
 934        return ret;
 935}
 936EXPORT_SYMBOL_GPL(page_mkclean);
 937
 938/**
 939 * page_move_anon_rmap - move a page to our anon_vma
 940 * @page:       the page to move to our anon_vma
 941 * @vma:        the vma the page belongs to
 942 * @address:    the user virtual address mapped
 943 *
 944 * When a page belongs exclusively to one process after a COW event,
 945 * that page can be moved into the anon_vma that belongs to just that
 946 * process, so the rmap code will not search the parent or sibling
 947 * processes.
 948 */
 949void page_move_anon_rmap(struct page *page,
 950        struct vm_area_struct *vma, unsigned long address)
 951{
 952        struct anon_vma *anon_vma = vma->anon_vma;
 953
 954        VM_BUG_ON(!PageLocked(page));
 955        VM_BUG_ON(!anon_vma);
 956        VM_BUG_ON(page->index != linear_page_index(vma, address));
 957
 958        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 959        page->mapping = (struct address_space *) anon_vma;
 960}
 961
 962/**
 963 * __page_set_anon_rmap - set up new anonymous rmap
 964 * @page:       Page to add to rmap     
 965 * @vma:        VM area to add page to.
 966 * @address:    User virtual address of the mapping     
 967 * @exclusive:  the page is exclusively owned by the current process
 968 */
 969static void __page_set_anon_rmap(struct page *page,
 970        struct vm_area_struct *vma, unsigned long address, int exclusive)
 971{
 972        struct anon_vma *anon_vma = vma->anon_vma;
 973
 974        BUG_ON(!anon_vma);
 975
 976        if (PageAnon(page))
 977                return;
 978
 979        /*
 980         * If the page isn't exclusively mapped into this vma,
 981         * we must use the _oldest_ possible anon_vma for the
 982         * page mapping!
 983         */
 984        if (!exclusive)
 985                anon_vma = anon_vma->root;
 986
 987        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
 988        page->mapping = (struct address_space *) anon_vma;
 989        page->index = linear_page_index(vma, address);
 990}
 991
 992/**
 993 * __page_check_anon_rmap - sanity check anonymous rmap addition
 994 * @page:       the page to add the mapping to
 995 * @vma:        the vm area in which the mapping is added
 996 * @address:    the user virtual address mapped
 997 */
 998static void __page_check_anon_rmap(struct page *page,
 999        struct vm_area_struct *vma, unsigned long address)
1000{
1001#ifdef CONFIG_DEBUG_VM
1002        /*
1003         * The page's anon-rmap details (mapping and index) are guaranteed to
1004         * be set up correctly at this point.
1005         *
1006         * We have exclusion against page_add_anon_rmap because the caller
1007         * always holds the page locked, except if called from page_dup_rmap,
1008         * in which case the page is already known to be setup.
1009         *
1010         * We have exclusion against page_add_new_anon_rmap because those pages
1011         * are initially only visible via the pagetables, and the pte is locked
1012         * over the call to page_add_new_anon_rmap.
1013         */
1014        BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1015        BUG_ON(page->index != linear_page_index(vma, address));
1016#endif
1017}
1018
1019/**
1020 * page_add_anon_rmap - add pte mapping to an anonymous page
1021 * @page:       the page to add the mapping to
1022 * @vma:        the vm area in which the mapping is added
1023 * @address:    the user virtual address mapped
1024 *
1025 * The caller needs to hold the pte lock, and the page must be locked in
1026 * the anon_vma case: to serialize mapping,index checking after setting,
1027 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1028 * (but PageKsm is never downgraded to PageAnon).
1029 */
1030void page_add_anon_rmap(struct page *page,
1031        struct vm_area_struct *vma, unsigned long address)
1032{
1033        do_page_add_anon_rmap(page, vma, address, 0);
1034}
1035
1036/*
1037 * Special version of the above for do_swap_page, which often runs
1038 * into pages that are exclusively owned by the current process.
1039 * Everybody else should continue to use page_add_anon_rmap above.
1040 */
1041void do_page_add_anon_rmap(struct page *page,
1042        struct vm_area_struct *vma, unsigned long address, int exclusive)
1043{
1044        int first = atomic_inc_and_test(&page->_mapcount);
1045        if (first) {
1046                if (!PageTransHuge(page))
1047                        __inc_zone_page_state(page, NR_ANON_PAGES);
1048                else
1049                        __inc_zone_page_state(page,
1050                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1051        }
1052        if (unlikely(PageKsm(page)))
1053                return;
1054
1055        VM_BUG_ON(!PageLocked(page));
1056        /* address might be in next vma when migration races vma_adjust */
1057        if (first)
1058                __page_set_anon_rmap(page, vma, address, exclusive);
1059        else
1060                __page_check_anon_rmap(page, vma, address);
1061}
1062
1063/**
1064 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1065 * @page:       the page to add the mapping to
1066 * @vma:        the vm area in which the mapping is added
1067 * @address:    the user virtual address mapped
1068 *
1069 * Same as page_add_anon_rmap but must only be called on *new* pages.
1070 * This means the inc-and-test can be bypassed.
1071 * Page does not have to be locked.
1072 */
1073void page_add_new_anon_rmap(struct page *page,
1074        struct vm_area_struct *vma, unsigned long address)
1075{
1076        VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1077        SetPageSwapBacked(page);
1078        atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1079        if (!PageTransHuge(page))
1080                __inc_zone_page_state(page, NR_ANON_PAGES);
1081        else
1082                __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1083        __page_set_anon_rmap(page, vma, address, 1);
1084        if (!mlocked_vma_newpage(vma, page))
1085                lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1086        else
1087                add_page_to_unevictable_list(page);
1088}
1089
1090/**
1091 * page_add_file_rmap - add pte mapping to a file page
1092 * @page: the page to add the mapping to
1093 *
1094 * The caller needs to hold the pte lock.
1095 */
1096void page_add_file_rmap(struct page *page)
1097{
1098        bool locked;
1099        unsigned long flags;
1100
1101        mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1102        if (atomic_inc_and_test(&page->_mapcount)) {
1103                __inc_zone_page_state(page, NR_FILE_MAPPED);
1104                mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1105        }
1106        mem_cgroup_end_update_page_stat(page, &locked, &flags);
1107}
1108
1109/**
1110 * page_remove_rmap - take down pte mapping from a page
1111 * @page: page to remove mapping from
1112 *
1113 * The caller needs to hold the pte lock.
1114 */
1115void page_remove_rmap(struct page *page)
1116{
1117        struct address_space *mapping = page_mapping(page);
1118        bool anon = PageAnon(page);
1119        bool locked;
1120        unsigned long flags;
1121
1122        /*
1123         * The anon case has no mem_cgroup page_stat to update; but may
1124         * uncharge_page() below, where the lock ordering can deadlock if
1125         * we hold the lock against page_stat move: so avoid it on anon.
1126         */
1127        if (!anon)
1128                mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1129
1130        /* page still mapped by someone else? */
1131        if (!atomic_add_negative(-1, &page->_mapcount))
1132                goto out;
1133
1134        /*
1135         * Now that the last pte has gone, s390 must transfer dirty
1136         * flag from storage key to struct page.  We can usually skip
1137         * this if the page is anon, so about to be freed; but perhaps
1138         * not if it's in swapcache - there might be another pte slot
1139         * containing the swap entry, but page not yet written to swap.
1140         *
1141         * And we can skip it on file pages, so long as the filesystem
1142         * participates in dirty tracking; but need to catch shm and tmpfs
1143         * and ramfs pages which have been modified since creation by read
1144         * fault.
1145         *
1146         * Note that mapping must be decided above, before decrementing
1147         * mapcount (which luckily provides a barrier): once page is unmapped,
1148         * it could be truncated and page->mapping reset to NULL at any moment.
1149         * Note also that we are relying on page_mapping(page) to set mapping
1150         * to &swapper_space when PageSwapCache(page).
1151         */
1152        if (mapping && !mapping_cap_account_dirty(mapping) &&
1153            page_test_and_clear_dirty(page_to_pfn(page), 1))
1154                set_page_dirty(page);
1155        /*
1156         * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1157         * and not charged by memcg for now.
1158         */
1159        if (unlikely(PageHuge(page)))
1160                goto out;
1161        if (anon) {
1162                mem_cgroup_uncharge_page(page);
1163                if (!PageTransHuge(page))
1164                        __dec_zone_page_state(page, NR_ANON_PAGES);
1165                else
1166                        __dec_zone_page_state(page,
1167                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1168        } else {
1169                __dec_zone_page_state(page, NR_FILE_MAPPED);
1170                mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1171                mem_cgroup_end_update_page_stat(page, &locked, &flags);
1172        }
1173        if (unlikely(PageMlocked(page)))
1174                clear_page_mlock(page);
1175        /*
1176         * It would be tidy to reset the PageAnon mapping here,
1177         * but that might overwrite a racing page_add_anon_rmap
1178         * which increments mapcount after us but sets mapping
1179         * before us: so leave the reset to free_hot_cold_page,
1180         * and remember that it's only reliable while mapped.
1181         * Leaving it set also helps swapoff to reinstate ptes
1182         * faster for those pages still in swapcache.
1183         */
1184        return;
1185out:
1186        if (!anon)
1187                mem_cgroup_end_update_page_stat(page, &locked, &flags);
1188}
1189
1190/*
1191 * Subfunctions of try_to_unmap: try_to_unmap_one called
1192 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1193 */
1194int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1195                     unsigned long address, enum ttu_flags flags)
1196{
1197        struct mm_struct *mm = vma->vm_mm;
1198        pte_t *pte;
1199        pte_t pteval;
1200        spinlock_t *ptl;
1201        int ret = SWAP_AGAIN;
1202
1203        pte = page_check_address(page, mm, address, &ptl, 0);
1204        if (!pte)
1205                goto out;
1206
1207        /*
1208         * If the page is mlock()d, we cannot swap it out.
1209         * If it's recently referenced (perhaps page_referenced
1210         * skipped over this mm) then we should reactivate it.
1211         */
1212        if (!(flags & TTU_IGNORE_MLOCK)) {
1213                if (vma->vm_flags & VM_LOCKED)
1214                        goto out_mlock;
1215
1216                if (TTU_ACTION(flags) == TTU_MUNLOCK)
1217                        goto out_unmap;
1218        }
1219        if (!(flags & TTU_IGNORE_ACCESS)) {
1220                if (ptep_clear_flush_young_notify(vma, address, pte)) {
1221                        ret = SWAP_FAIL;
1222                        goto out_unmap;
1223                }
1224        }
1225
1226        /* Nuke the page table entry. */
1227        flush_cache_page(vma, address, page_to_pfn(page));
1228        pteval = ptep_clear_flush(vma, address, pte);
1229
1230        /* Move the dirty bit to the physical page now the pte is gone. */
1231        if (pte_dirty(pteval))
1232                set_page_dirty(page);
1233
1234        /* Update high watermark before we lower rss */
1235        update_hiwater_rss(mm);
1236
1237        if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1238                if (PageAnon(page))
1239                        dec_mm_counter(mm, MM_ANONPAGES);
1240                else
1241                        dec_mm_counter(mm, MM_FILEPAGES);
1242                set_pte_at(mm, address, pte,
1243                                swp_entry_to_pte(make_hwpoison_entry(page)));
1244        } else if (PageAnon(page)) {
1245                swp_entry_t entry = { .val = page_private(page) };
1246
1247                if (PageSwapCache(page)) {
1248                        /*
1249                         * Store the swap location in the pte.
1250                         * See handle_pte_fault() ...
1251                         */
1252                        if (swap_duplicate(entry) < 0) {
1253                                set_pte_at(mm, address, pte, pteval);
1254                                ret = SWAP_FAIL;
1255                                goto out_unmap;
1256                        }
1257                        if (list_empty(&mm->mmlist)) {
1258                                spin_lock(&mmlist_lock);
1259                                if (list_empty(&mm->mmlist))
1260                                        list_add(&mm->mmlist, &init_mm.mmlist);
1261                                spin_unlock(&mmlist_lock);
1262                        }
1263                        dec_mm_counter(mm, MM_ANONPAGES);
1264                        inc_mm_counter(mm, MM_SWAPENTS);
1265                } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1266                        /*
1267                         * Store the pfn of the page in a special migration
1268                         * pte. do_swap_page() will wait until the migration
1269                         * pte is removed and then restart fault handling.
1270                         */
1271                        BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1272                        entry = make_migration_entry(page, pte_write(pteval));
1273                }
1274                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1275                BUG_ON(pte_file(*pte));
1276        } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1277                   (TTU_ACTION(flags) == TTU_MIGRATION)) {
1278                /* Establish migration entry for a file page */
1279                swp_entry_t entry;
1280                entry = make_migration_entry(page, pte_write(pteval));
1281                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1282        } else
1283                dec_mm_counter(mm, MM_FILEPAGES);
1284
1285        page_remove_rmap(page);
1286        page_cache_release(page);
1287
1288out_unmap:
1289        pte_unmap_unlock(pte, ptl);
1290        if (ret != SWAP_FAIL)
1291                mmu_notifier_invalidate_page(mm, address);
1292out:
1293        return ret;
1294
1295out_mlock:
1296        pte_unmap_unlock(pte, ptl);
1297
1298
1299        /*
1300         * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1301         * unstable result and race. Plus, We can't wait here because
1302         * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1303         * if trylock failed, the page remain in evictable lru and later
1304         * vmscan could retry to move the page to unevictable lru if the
1305         * page is actually mlocked.
1306         */
1307        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1308                if (vma->vm_flags & VM_LOCKED) {
1309                        mlock_vma_page(page);
1310                        ret = SWAP_MLOCK;
1311                }
1312                up_read(&vma->vm_mm->mmap_sem);
1313        }
1314        return ret;
1315}
1316
1317/*
1318 * objrmap doesn't work for nonlinear VMAs because the assumption that
1319 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1320 * Consequently, given a particular page and its ->index, we cannot locate the
1321 * ptes which are mapping that page without an exhaustive linear search.
1322 *
1323 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1324 * maps the file to which the target page belongs.  The ->vm_private_data field
1325 * holds the current cursor into that scan.  Successive searches will circulate
1326 * around the vma's virtual address space.
1327 *
1328 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1329 * more scanning pressure is placed against them as well.   Eventually pages
1330 * will become fully unmapped and are eligible for eviction.
1331 *
1332 * For very sparsely populated VMAs this is a little inefficient - chances are
1333 * there there won't be many ptes located within the scan cluster.  In this case
1334 * maybe we could scan further - to the end of the pte page, perhaps.
1335 *
1336 * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
1337 * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
1338 * rather than unmapping them.  If we encounter the "check_page" that vmscan is
1339 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1340 */
1341#define CLUSTER_SIZE    min(32*PAGE_SIZE, PMD_SIZE)
1342#define CLUSTER_MASK    (~(CLUSTER_SIZE - 1))
1343
1344static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1345                struct vm_area_struct *vma, struct page *check_page)
1346{
1347        struct mm_struct *mm = vma->vm_mm;
1348        pgd_t *pgd;
1349        pud_t *pud;
1350        pmd_t *pmd;
1351        pte_t *pte;
1352        pte_t pteval;
1353        spinlock_t *ptl;
1354        struct page *page;
1355        unsigned long address;
1356        unsigned long mmun_start;       /* For mmu_notifiers */
1357        unsigned long mmun_end;         /* For mmu_notifiers */
1358        unsigned long end;
1359        int ret = SWAP_AGAIN;
1360        int locked_vma = 0;
1361
1362        address = (vma->vm_start + cursor) & CLUSTER_MASK;
1363        end = address + CLUSTER_SIZE;
1364        if (address < vma->vm_start)
1365                address = vma->vm_start;
1366        if (end > vma->vm_end)
1367                end = vma->vm_end;
1368
1369        pgd = pgd_offset(mm, address);
1370        if (!pgd_present(*pgd))
1371                return ret;
1372
1373        pud = pud_offset(pgd, address);
1374        if (!pud_present(*pud))
1375                return ret;
1376
1377        pmd = pmd_offset(pud, address);
1378        if (!pmd_present(*pmd))
1379                return ret;
1380
1381        mmun_start = address;
1382        mmun_end   = end;
1383        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1384
1385        /*
1386         * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1387         * keep the sem while scanning the cluster for mlocking pages.
1388         */
1389        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1390                locked_vma = (vma->vm_flags & VM_LOCKED);
1391                if (!locked_vma)
1392                        up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1393        }
1394
1395        pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1396
1397        /* Update high watermark before we lower rss */
1398        update_hiwater_rss(mm);
1399
1400        for (; address < end; pte++, address += PAGE_SIZE) {
1401                if (!pte_present(*pte))
1402                        continue;
1403                page = vm_normal_page(vma, address, *pte);
1404                BUG_ON(!page || PageAnon(page));
1405
1406                if (locked_vma) {
1407                        mlock_vma_page(page);   /* no-op if already mlocked */
1408                        if (page == check_page)
1409                                ret = SWAP_MLOCK;
1410                        continue;       /* don't unmap */
1411                }
1412
1413                if (ptep_clear_flush_young_notify(vma, address, pte))
1414                        continue;
1415
1416                /* Nuke the page table entry. */
1417                flush_cache_page(vma, address, pte_pfn(*pte));
1418                pteval = ptep_clear_flush(vma, address, pte);
1419
1420                /* If nonlinear, store the file page offset in the pte. */
1421                if (page->index != linear_page_index(vma, address))
1422                        set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1423
1424                /* Move the dirty bit to the physical page now the pte is gone. */
1425                if (pte_dirty(pteval))
1426                        set_page_dirty(page);
1427
1428                page_remove_rmap(page);
1429                page_cache_release(page);
1430                dec_mm_counter(mm, MM_FILEPAGES);
1431                (*mapcount)--;
1432        }
1433        pte_unmap_unlock(pte - 1, ptl);
1434        mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1435        if (locked_vma)
1436                up_read(&vma->vm_mm->mmap_sem);
1437        return ret;
1438}
1439
1440bool is_vma_temporary_stack(struct vm_area_struct *vma)
1441{
1442        int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1443
1444        if (!maybe_stack)
1445                return false;
1446
1447        if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1448                                                VM_STACK_INCOMPLETE_SETUP)
1449                return true;
1450
1451        return false;
1452}
1453
1454/**
1455 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1456 * rmap method
1457 * @page: the page to unmap/unlock
1458 * @flags: action and flags
1459 *
1460 * Find all the mappings of a page using the mapping pointer and the vma chains
1461 * contained in the anon_vma struct it points to.
1462 *
1463 * This function is only called from try_to_unmap/try_to_munlock for
1464 * anonymous pages.
1465 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1466 * where the page was found will be held for write.  So, we won't recheck
1467 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1468 * 'LOCKED.
1469 */
1470static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1471{
1472        struct anon_vma *anon_vma;
1473        pgoff_t pgoff;
1474        struct anon_vma_chain *avc;
1475        int ret = SWAP_AGAIN;
1476
1477        anon_vma = page_lock_anon_vma(page);
1478        if (!anon_vma)
1479                return ret;
1480
1481        pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1482        anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1483                struct vm_area_struct *vma = avc->vma;
1484                unsigned long address;
1485
1486                /*
1487                 * During exec, a temporary VMA is setup and later moved.
1488                 * The VMA is moved under the anon_vma lock but not the
1489                 * page tables leading to a race where migration cannot
1490                 * find the migration ptes. Rather than increasing the
1491                 * locking requirements of exec(), migration skips
1492                 * temporary VMAs until after exec() completes.
1493                 */
1494                if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1495                                is_vma_temporary_stack(vma))
1496                        continue;
1497
1498                address = vma_address(page, vma);
1499                ret = try_to_unmap_one(page, vma, address, flags);
1500                if (ret != SWAP_AGAIN || !page_mapped(page))
1501                        break;
1502        }
1503
1504        page_unlock_anon_vma(anon_vma);
1505        return ret;
1506}
1507
1508/**
1509 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1510 * @page: the page to unmap/unlock
1511 * @flags: action and flags
1512 *
1513 * Find all the mappings of a page using the mapping pointer and the vma chains
1514 * contained in the address_space struct it points to.
1515 *
1516 * This function is only called from try_to_unmap/try_to_munlock for
1517 * object-based pages.
1518 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1519 * where the page was found will be held for write.  So, we won't recheck
1520 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1521 * 'LOCKED.
1522 */
1523static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1524{
1525        struct address_space *mapping = page->mapping;
1526        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1527        struct vm_area_struct *vma;
1528        int ret = SWAP_AGAIN;
1529        unsigned long cursor;
1530        unsigned long max_nl_cursor = 0;
1531        unsigned long max_nl_size = 0;
1532        unsigned int mapcount;
1533
1534        mutex_lock(&mapping->i_mmap_mutex);
1535        vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1536                unsigned long address = vma_address(page, vma);
1537                ret = try_to_unmap_one(page, vma, address, flags);
1538                if (ret != SWAP_AGAIN || !page_mapped(page))
1539                        goto out;
1540        }
1541
1542        if (list_empty(&mapping->i_mmap_nonlinear))
1543                goto out;
1544
1545        /*
1546         * We don't bother to try to find the munlocked page in nonlinears.
1547         * It's costly. Instead, later, page reclaim logic may call
1548         * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1549         */
1550        if (TTU_ACTION(flags) == TTU_MUNLOCK)
1551                goto out;
1552
1553        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1554                                                        shared.nonlinear) {
1555                cursor = (unsigned long) vma->vm_private_data;
1556                if (cursor > max_nl_cursor)
1557                        max_nl_cursor = cursor;
1558                cursor = vma->vm_end - vma->vm_start;
1559                if (cursor > max_nl_size)
1560                        max_nl_size = cursor;
1561        }
1562
1563        if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1564                ret = SWAP_FAIL;
1565                goto out;
1566        }
1567
1568        /*
1569         * We don't try to search for this page in the nonlinear vmas,
1570         * and page_referenced wouldn't have found it anyway.  Instead
1571         * just walk the nonlinear vmas trying to age and unmap some.
1572         * The mapcount of the page we came in with is irrelevant,
1573         * but even so use it as a guide to how hard we should try?
1574         */
1575        mapcount = page_mapcount(page);
1576        if (!mapcount)
1577                goto out;
1578        cond_resched();
1579
1580        max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1581        if (max_nl_cursor == 0)
1582                max_nl_cursor = CLUSTER_SIZE;
1583
1584        do {
1585                list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1586                                                        shared.nonlinear) {
1587                        cursor = (unsigned long) vma->vm_private_data;
1588                        while ( cursor < max_nl_cursor &&
1589                                cursor < vma->vm_end - vma->vm_start) {
1590                                if (try_to_unmap_cluster(cursor, &mapcount,
1591                                                vma, page) == SWAP_MLOCK)
1592                                        ret = SWAP_MLOCK;
1593                                cursor += CLUSTER_SIZE;
1594                                vma->vm_private_data = (void *) cursor;
1595                                if ((int)mapcount <= 0)
1596                                        goto out;
1597                        }
1598                        vma->vm_private_data = (void *) max_nl_cursor;
1599                }
1600                cond_resched();
1601                max_nl_cursor += CLUSTER_SIZE;
1602        } while (max_nl_cursor <= max_nl_size);
1603
1604        /*
1605         * Don't loop forever (perhaps all the remaining pages are
1606         * in locked vmas).  Reset cursor on all unreserved nonlinear
1607         * vmas, now forgetting on which ones it had fallen behind.
1608         */
1609        list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
1610                vma->vm_private_data = NULL;
1611out:
1612        mutex_unlock(&mapping->i_mmap_mutex);
1613        return ret;
1614}
1615
1616/**
1617 * try_to_unmap - try to remove all page table mappings to a page
1618 * @page: the page to get unmapped
1619 * @flags: action and flags
1620 *
1621 * Tries to remove all the page table entries which are mapping this
1622 * page, used in the pageout path.  Caller must hold the page lock.
1623 * Return values are:
1624 *
1625 * SWAP_SUCCESS - we succeeded in removing all mappings
1626 * SWAP_AGAIN   - we missed a mapping, try again later
1627 * SWAP_FAIL    - the page is unswappable
1628 * SWAP_MLOCK   - page is mlocked.
1629 */
1630int try_to_unmap(struct page *page, enum ttu_flags flags)
1631{
1632        int ret;
1633
1634        BUG_ON(!PageLocked(page));
1635        VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1636
1637        if (unlikely(PageKsm(page)))
1638                ret = try_to_unmap_ksm(page, flags);
1639        else if (PageAnon(page))
1640                ret = try_to_unmap_anon(page, flags);
1641        else
1642                ret = try_to_unmap_file(page, flags);
1643        if (ret != SWAP_MLOCK && !page_mapped(page))
1644                ret = SWAP_SUCCESS;
1645        return ret;
1646}
1647
1648/**
1649 * try_to_munlock - try to munlock a page
1650 * @page: the page to be munlocked
1651 *
1652 * Called from munlock code.  Checks all of the VMAs mapping the page
1653 * to make sure nobody else has this page mlocked. The page will be
1654 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1655 *
1656 * Return values are:
1657 *
1658 * SWAP_AGAIN   - no vma is holding page mlocked, or,
1659 * SWAP_AGAIN   - page mapped in mlocked vma -- couldn't acquire mmap sem
1660 * SWAP_FAIL    - page cannot be located at present
1661 * SWAP_MLOCK   - page is now mlocked.
1662 */
1663int try_to_munlock(struct page *page)
1664{
1665        VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1666
1667        if (unlikely(PageKsm(page)))
1668                return try_to_unmap_ksm(page, TTU_MUNLOCK);
1669        else if (PageAnon(page))
1670                return try_to_unmap_anon(page, TTU_MUNLOCK);
1671        else
1672                return try_to_unmap_file(page, TTU_MUNLOCK);
1673}
1674
1675void __put_anon_vma(struct anon_vma *anon_vma)
1676{
1677        struct anon_vma *root = anon_vma->root;
1678
1679        if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1680                anon_vma_free(root);
1681
1682        anon_vma_free(anon_vma);
1683}
1684
1685#ifdef CONFIG_MIGRATION
1686/*
1687 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1688 * Called by migrate.c to remove migration ptes, but might be used more later.
1689 */
1690static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1691                struct vm_area_struct *, unsigned long, void *), void *arg)
1692{
1693        struct anon_vma *anon_vma;
1694        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1695        struct anon_vma_chain *avc;
1696        int ret = SWAP_AGAIN;
1697
1698        /*
1699         * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1700         * because that depends on page_mapped(); but not all its usages
1701         * are holding mmap_sem. Users without mmap_sem are required to
1702         * take a reference count to prevent the anon_vma disappearing
1703         */
1704        anon_vma = page_anon_vma(page);
1705        if (!anon_vma)
1706                return ret;
1707        anon_vma_lock(anon_vma);
1708        anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1709                struct vm_area_struct *vma = avc->vma;
1710                unsigned long address = vma_address(page, vma);
1711                ret = rmap_one(page, vma, address, arg);
1712                if (ret != SWAP_AGAIN)
1713                        break;
1714        }
1715        anon_vma_unlock(anon_vma);
1716        return ret;
1717}
1718
1719static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1720                struct vm_area_struct *, unsigned long, void *), void *arg)
1721{
1722        struct address_space *mapping = page->mapping;
1723        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1724        struct vm_area_struct *vma;
1725        int ret = SWAP_AGAIN;
1726
1727        if (!mapping)
1728                return ret;
1729        mutex_lock(&mapping->i_mmap_mutex);
1730        vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1731                unsigned long address = vma_address(page, vma);
1732                ret = rmap_one(page, vma, address, arg);
1733                if (ret != SWAP_AGAIN)
1734                        break;
1735        }
1736        /*
1737         * No nonlinear handling: being always shared, nonlinear vmas
1738         * never contain migration ptes.  Decide what to do about this
1739         * limitation to linear when we need rmap_walk() on nonlinear.
1740         */
1741        mutex_unlock(&mapping->i_mmap_mutex);
1742        return ret;
1743}
1744
1745int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1746                struct vm_area_struct *, unsigned long, void *), void *arg)
1747{
1748        VM_BUG_ON(!PageLocked(page));
1749
1750        if (unlikely(PageKsm(page)))
1751                return rmap_walk_ksm(page, rmap_one, arg);
1752        else if (PageAnon(page))
1753                return rmap_walk_anon(page, rmap_one, arg);
1754        else
1755                return rmap_walk_file(page, rmap_one, arg);
1756}
1757#endif /* CONFIG_MIGRATION */
1758
1759#ifdef CONFIG_HUGETLB_PAGE
1760/*
1761 * The following three functions are for anonymous (private mapped) hugepages.
1762 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1763 * and no lru code, because we handle hugepages differently from common pages.
1764 */
1765static void __hugepage_set_anon_rmap(struct page *page,
1766        struct vm_area_struct *vma, unsigned long address, int exclusive)
1767{
1768        struct anon_vma *anon_vma = vma->anon_vma;
1769
1770        BUG_ON(!anon_vma);
1771
1772        if (PageAnon(page))
1773                return;
1774        if (!exclusive)
1775                anon_vma = anon_vma->root;
1776
1777        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1778        page->mapping = (struct address_space *) anon_vma;
1779        page->index = linear_page_index(vma, address);
1780}
1781
1782void hugepage_add_anon_rmap(struct page *page,
1783                            struct vm_area_struct *vma, unsigned long address)
1784{
1785        struct anon_vma *anon_vma = vma->anon_vma;
1786        int first;
1787
1788        BUG_ON(!PageLocked(page));
1789        BUG_ON(!anon_vma);
1790        /* address might be in next vma when migration races vma_adjust */
1791        first = atomic_inc_and_test(&page->_mapcount);
1792        if (first)
1793                __hugepage_set_anon_rmap(page, vma, address, 0);
1794}
1795
1796void hugepage_add_new_anon_rmap(struct page *page,
1797                        struct vm_area_struct *vma, unsigned long address)
1798{
1799        BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1800        atomic_set(&page->_mapcount, 0);
1801        __hugepage_set_anon_rmap(page, vma, address, 1);
1802}
1803#endif /* CONFIG_HUGETLB_PAGE */
1804
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.