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
  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
 123static void anon_vma_chain_link(struct vm_area_struct *vma,
 124                                struct anon_vma_chain *avc,
 125                                struct anon_vma *anon_vma)
 126{
 127        avc->vma = vma;
 128        avc->anon_vma = anon_vma;
 129        list_add(&avc->same_vma, &vma->anon_vma_chain);
 130
 131        /*
 132         * It's critical to add new vmas to the tail of the anon_vma,
 133         * see comment in huge_memory.c:__split_huge_page().
 134         */
 135        list_add_tail(&avc->same_anon_vma, &anon_vma->head);
 136}
 137
 138/**
 139 * anon_vma_prepare - attach an anon_vma to a memory region
 140 * @vma: the memory region in question
 141 *
 142 * This makes sure the memory mapping described by 'vma' has
 143 * an 'anon_vma' attached to it, so that we can associate the
 144 * anonymous pages mapped into it with that anon_vma.
 145 *
 146 * The common case will be that we already have one, but if
 147 * not we either need to find an adjacent mapping that we
 148 * can re-use the anon_vma from (very common when the only
 149 * reason for splitting a vma has been mprotect()), or we
 150 * allocate a new one.
 151 *
 152 * Anon-vma allocations are very subtle, because we may have
 153 * optimistically looked up an anon_vma in page_lock_anon_vma()
 154 * and that may actually touch the spinlock even in the newly
 155 * allocated vma (it depends on RCU to make sure that the
 156 * anon_vma isn't actually destroyed).
 157 *
 158 * As a result, we need to do proper anon_vma locking even
 159 * for the new allocation. At the same time, we do not want
 160 * to do any locking for the common case of already having
 161 * an anon_vma.
 162 *
 163 * This must be called with the mmap_sem held for reading.
 164 */
 165int anon_vma_prepare(struct vm_area_struct *vma)
 166{
 167        struct anon_vma *anon_vma = vma->anon_vma;
 168        struct anon_vma_chain *avc;
 169
 170        might_sleep();
 171        if (unlikely(!anon_vma)) {
 172                struct mm_struct *mm = vma->vm_mm;
 173                struct anon_vma *allocated;
 174
 175                avc = anon_vma_chain_alloc(GFP_KERNEL);
 176                if (!avc)
 177                        goto out_enomem;
 178
 179                anon_vma = find_mergeable_anon_vma(vma);
 180                allocated = NULL;
 181                if (!anon_vma) {
 182                        anon_vma = anon_vma_alloc();
 183                        if (unlikely(!anon_vma))
 184                                goto out_enomem_free_avc;
 185                        allocated = anon_vma;
 186                }
 187
 188                anon_vma_lock(anon_vma);
 189                /* page_table_lock to protect against threads */
 190                spin_lock(&mm->page_table_lock);
 191                if (likely(!vma->anon_vma)) {
 192                        vma->anon_vma = anon_vma;
 193                        anon_vma_chain_link(vma, avc, anon_vma);
 194                        allocated = NULL;
 195                        avc = NULL;
 196                }
 197                spin_unlock(&mm->page_table_lock);
 198                anon_vma_unlock(anon_vma);
 199
 200                if (unlikely(allocated))
 201                        put_anon_vma(allocated);
 202                if (unlikely(avc))
 203                        anon_vma_chain_free(avc);
 204        }
 205        return 0;
 206
 207 out_enomem_free_avc:
 208        anon_vma_chain_free(avc);
 209 out_enomem:
 210        return -ENOMEM;
 211}
 212
 213/*
 214 * This is a useful helper function for locking the anon_vma root as
 215 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 216 * have the same vma.
 217 *
 218 * Such anon_vma's should have the same root, so you'd expect to see
 219 * just a single mutex_lock for the whole traversal.
 220 */
 221static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 222{
 223        struct anon_vma *new_root = anon_vma->root;
 224        if (new_root != root) {
 225                if (WARN_ON_ONCE(root))
 226                        mutex_unlock(&root->mutex);
 227                root = new_root;
 228                mutex_lock(&root->mutex);
 229        }
 230        return root;
 231}
 232
 233static inline void unlock_anon_vma_root(struct anon_vma *root)
 234{
 235        if (root)
 236                mutex_unlock(&root->mutex);
 237}
 238
 239/*
 240 * Attach the anon_vmas from src to dst.
 241 * Returns 0 on success, -ENOMEM on failure.
 242 */
 243int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 244{
 245        struct anon_vma_chain *avc, *pavc;
 246        struct anon_vma *root = NULL;
 247
 248        list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 249                struct anon_vma *anon_vma;
 250
 251                avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 252                if (unlikely(!avc)) {
 253                        unlock_anon_vma_root(root);
 254                        root = NULL;
 255                        avc = anon_vma_chain_alloc(GFP_KERNEL);
 256                        if (!avc)
 257                                goto enomem_failure;
 258                }
 259                anon_vma = pavc->anon_vma;
 260                root = lock_anon_vma_root(root, anon_vma);
 261                anon_vma_chain_link(dst, avc, anon_vma);
 262        }
 263        unlock_anon_vma_root(root);
 264        return 0;
 265
 266 enomem_failure:
 267        unlink_anon_vmas(dst);
 268        return -ENOMEM;
 269}
 270
 271/*
 272 * Some rmap walk that needs to find all ptes/hugepmds without false
 273 * negatives (like migrate and split_huge_page) running concurrent
 274 * with operations that copy or move pagetables (like mremap() and
 275 * fork()) to be safe. They depend on the anon_vma "same_anon_vma"
 276 * list to be in a certain order: the dst_vma must be placed after the
 277 * src_vma in the list. This is always guaranteed by fork() but
 278 * mremap() needs to call this function to enforce it in case the
 279 * dst_vma isn't newly allocated and chained with the anon_vma_clone()
 280 * function but just an extension of a pre-existing vma through
 281 * vma_merge.
 282 *
 283 * NOTE: the same_anon_vma list can still be changed by other
 284 * processes while mremap runs because mremap doesn't hold the
 285 * anon_vma mutex to prevent modifications to the list while it
 286 * runs. All we need to enforce is that the relative order of this
 287 * process vmas isn't changing (we don't care about other vmas
 288 * order). Each vma corresponds to an anon_vma_chain structure so
 289 * there's no risk that other processes calling anon_vma_moveto_tail()
 290 * and changing the same_anon_vma list under mremap() will screw with
 291 * the relative order of this process vmas in the list, because we
 292 * they can't alter the order of any vma that belongs to this
 293 * process. And there can't be another anon_vma_moveto_tail() running
 294 * concurrently with mremap() coming from this process because we hold
 295 * the mmap_sem for the whole mremap(). fork() ordering dependency
 296 * also shouldn't be affected because fork() only cares that the
 297 * parent vmas are placed in the list before the child vmas and
 298 * anon_vma_moveto_tail() won't reorder vmas from either the fork()
 299 * parent or child.
 300 */
 301void anon_vma_moveto_tail(struct vm_area_struct *dst)
 302{
 303        struct anon_vma_chain *pavc;
 304        struct anon_vma *root = NULL;
 305
 306        list_for_each_entry_reverse(pavc, &dst->anon_vma_chain, same_vma) {
 307                struct anon_vma *anon_vma = pavc->anon_vma;
 308                VM_BUG_ON(pavc->vma != dst);
 309                root = lock_anon_vma_root(root, anon_vma);
 310                list_del(&pavc->same_anon_vma);
 311                list_add_tail(&pavc->same_anon_vma, &anon_vma->head);
 312        }
 313        unlock_anon_vma_root(root);
 314}
 315
 316/*
 317 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 318 * the corresponding VMA in the parent process is attached to.
 319 * Returns 0 on success, non-zero on failure.
 320 */
 321int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 322{
 323        struct anon_vma_chain *avc;
 324        struct anon_vma *anon_vma;
 325
 326        /* Don't bother if the parent process has no anon_vma here. */
 327        if (!pvma->anon_vma)
 328                return 0;
 329
 330        /*
 331         * First, attach the new VMA to the parent VMA's anon_vmas,
 332         * so rmap can find non-COWed pages in child processes.
 333         */
 334        if (anon_vma_clone(vma, pvma))
 335                return -ENOMEM;
 336
 337        /* Then add our own anon_vma. */
 338        anon_vma = anon_vma_alloc();
 339        if (!anon_vma)
 340                goto out_error;
 341        avc = anon_vma_chain_alloc(GFP_KERNEL);
 342        if (!avc)
 343                goto out_error_free_anon_vma;
 344
 345        /*
 346         * The root anon_vma's spinlock is the lock actually used when we
 347         * lock any of the anon_vmas in this anon_vma tree.
 348         */
 349        anon_vma->root = pvma->anon_vma->root;
 350        /*
 351         * With refcounts, an anon_vma can stay around longer than the
 352         * process it belongs to. The root anon_vma needs to be pinned until
 353         * this anon_vma is freed, because the lock lives in the root.
 354         */
 355        get_anon_vma(anon_vma->root);
 356        /* Mark this anon_vma as the one where our new (COWed) pages go. */
 357        vma->anon_vma = anon_vma;
 358        anon_vma_lock(anon_vma);
 359        anon_vma_chain_link(vma, avc, anon_vma);
 360        anon_vma_unlock(anon_vma);
 361
 362        return 0;
 363
 364 out_error_free_anon_vma:
 365        put_anon_vma(anon_vma);
 366 out_error:
 367        unlink_anon_vmas(vma);
 368        return -ENOMEM;
 369}
 370
 371void unlink_anon_vmas(struct vm_area_struct *vma)
 372{
 373        struct anon_vma_chain *avc, *next;
 374        struct anon_vma *root = NULL;
 375
 376        /*
 377         * Unlink each anon_vma chained to the VMA.  This list is ordered
 378         * from newest to oldest, ensuring the root anon_vma gets freed last.
 379         */
 380        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 381                struct anon_vma *anon_vma = avc->anon_vma;
 382
 383                root = lock_anon_vma_root(root, anon_vma);
 384                list_del(&avc->same_anon_vma);
 385
 386                /*
 387                 * Leave empty anon_vmas on the list - we'll need
 388                 * to free them outside the lock.
 389                 */
 390                if (list_empty(&anon_vma->head))
 391                        continue;
 392
 393                list_del(&avc->same_vma);
 394                anon_vma_chain_free(avc);
 395        }
 396        unlock_anon_vma_root(root);
 397
 398        /*
 399         * Iterate the list once more, it now only contains empty and unlinked
 400         * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 401         * needing to acquire the anon_vma->root->mutex.
 402         */
 403        list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 404                struct anon_vma *anon_vma = avc->anon_vma;
 405
 406                put_anon_vma(anon_vma);
 407
 408                list_del(&avc->same_vma);
 409                anon_vma_chain_free(avc);
 410        }
 411}
 412
 413static void anon_vma_ctor(void *data)
 414{
 415        struct anon_vma *anon_vma = data;
 416
 417        mutex_init(&anon_vma->mutex);
 418        atomic_set(&anon_vma->refcount, 0);
 419        INIT_LIST_HEAD(&anon_vma->head);
 420}
 421
 422void __init anon_vma_init(void)
 423{
 424        anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 425                        0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
 426        anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
 427}
 428
 429/*
 430 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 431 *
 432 * Since there is no serialization what so ever against page_remove_rmap()
 433 * the best this function can do is return a locked anon_vma that might
 434 * have been relevant to this page.
 435 *
 436 * The page might have been remapped to a different anon_vma or the anon_vma
 437 * returned may already be freed (and even reused).
 438 *
 439 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 440 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 441 * ensure that any anon_vma obtained from the page will still be valid for as
 442 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 443 *
 444 * All users of this function must be very careful when walking the anon_vma
 445 * chain and verify that the page in question is indeed mapped in it
 446 * [ something equivalent to page_mapped_in_vma() ].
 447 *
 448 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 449 * that the anon_vma pointer from page->mapping is valid if there is a
 450 * mapcount, we can dereference the anon_vma after observing those.
 451 */
 452struct anon_vma *page_get_anon_vma(struct page *page)
 453{
 454        struct anon_vma *anon_vma = NULL;
 455        unsigned long anon_mapping;
 456
 457        rcu_read_lock();
 458        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 459        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 460                goto out;
 461        if (!page_mapped(page))
 462                goto out;
 463
 464        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 465        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 466                anon_vma = NULL;
 467                goto out;
 468        }
 469
 470        /*
 471         * If this page is still mapped, then its anon_vma cannot have been
 472         * freed.  But if it has been unmapped, we have no security against the
 473         * anon_vma structure being freed and reused (for another anon_vma:
 474         * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
 475         * above cannot corrupt).
 476         */
 477        if (!page_mapped(page)) {
 478                put_anon_vma(anon_vma);
 479                anon_vma = NULL;
 480        }
 481out:
 482        rcu_read_unlock();
 483
 484        return anon_vma;
 485}
 486
 487/*
 488 * Similar to page_get_anon_vma() except it locks the anon_vma.
 489 *
 490 * Its a little more complex as it tries to keep the fast path to a single
 491 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 492 * reference like with page_get_anon_vma() and then block on the mutex.
 493 */
 494struct anon_vma *page_lock_anon_vma(struct page *page)
 495{
 496        struct anon_vma *anon_vma = NULL;
 497        struct anon_vma *root_anon_vma;
 498        unsigned long anon_mapping;
 499
 500        rcu_read_lock();
 501        anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
 502        if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 503                goto out;
 504        if (!page_mapped(page))
 505                goto out;
 506
 507        anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 508        root_anon_vma = ACCESS_ONCE(anon_vma->root);
 509        if (mutex_trylock(&root_anon_vma->mutex)) {
 510                /*
 511                 * If the page is still mapped, then this anon_vma is still
 512                 * its anon_vma, and holding the mutex ensures that it will
 513                 * not go away, see anon_vma_free().
 514                 */
 515                if (!page_mapped(page)) {
 516                        mutex_unlock(&root_anon_vma->mutex);
 517                        anon_vma = NULL;
 518                }
 519                goto out;
 520        }
 521
 522        /* trylock failed, we got to sleep */
 523        if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 524                anon_vma = NULL;
 525                goto out;
 526        }
 527
 528        if (!page_mapped(page)) {
 529                put_anon_vma(anon_vma);
 530                anon_vma = NULL;
 531                goto out;
 532        }
 533
 534        /* we pinned the anon_vma, its safe to sleep */
 535        rcu_read_unlock();
 536        anon_vma_lock(anon_vma);
 537
 538        if (atomic_dec_and_test(&anon_vma->refcount)) {
 539                /*
 540                 * Oops, we held the last refcount, release the lock
 541                 * and bail -- can't simply use put_anon_vma() because
 542                 * we'll deadlock on the anon_vma_lock() recursion.
 543                 */
 544                anon_vma_unlock(anon_vma);
 545                __put_anon_vma(anon_vma);
 546                anon_vma = NULL;
 547        }
 548
 549        return anon_vma;
 550
 551out:
 552        rcu_read_unlock();
 553        return anon_vma;
 554}
 555
 556void page_unlock_anon_vma(struct anon_vma *anon_vma)
 557{
 558        anon_vma_unlock(anon_vma);
 559}
 560
 561/*
 562 * At what user virtual address is page expected in @vma?
 563 * Returns virtual address or -EFAULT if page's index/offset is not
 564 * within the range mapped the @vma.
 565 */
 566inline unsigned long
 567vma_address(struct page *page, struct vm_area_struct *vma)
 568{
 569        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 570        unsigned long address;
 571
 572        if (unlikely(is_vm_hugetlb_page(vma)))
 573                pgoff = page->index << huge_page_order(page_hstate(page));
 574        address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
 575        if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
 576                /* page should be within @vma mapping range */
 577                return -EFAULT;
 578        }
 579        return address;
 580}
 581
 582/*
 583 * At what user virtual address is page expected in vma?
 584 * Caller should check the page is actually part of the vma.
 585 */
 586unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 587{
 588        if (PageAnon(page)) {
 589                struct anon_vma *page__anon_vma = page_anon_vma(page);
 590                /*
 591                 * Note: swapoff's unuse_vma() is more efficient with this
 592                 * check, and needs it to match anon_vma when KSM is active.
 593                 */
 594                if (!vma->anon_vma || !page__anon_vma ||
 595                    vma->anon_vma->root != page__anon_vma->root)
 596                        return -EFAULT;
 597        } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
 598                if (!vma->vm_file ||
 599                    vma->vm_file->f_mapping != page->mapping)
 600                        return -EFAULT;
 601        } else
 602                return -EFAULT;
 603        return vma_address(page, vma);
 604}
 605
 606/*
 607 * Check that @page is mapped at @address into @mm.
 608 *
 609 * If @sync is false, page_check_address may perform a racy check to avoid
 610 * the page table lock when the pte is not present (helpful when reclaiming
 611 * highly shared pages).
 612 *
 613 * On success returns with pte mapped and locked.
 614 */
 615pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
 616                          unsigned long address, spinlock_t **ptlp, int sync)
 617{
 618        pgd_t *pgd;
 619        pud_t *pud;
 620        pmd_t *pmd;
 621        pte_t *pte;
 622        spinlock_t *ptl;
 623
 624        if (unlikely(PageHuge(page))) {
 625                pte = huge_pte_offset(mm, address);
 626                ptl = &mm->page_table_lock;
 627                goto check;
 628        }
 629
 630        pgd = pgd_offset(mm, address);
 631        if (!pgd_present(*pgd))
 632                return NULL;
 633
 634        pud = pud_offset(pgd, address);
 635        if (!pud_present(*pud))
 636                return NULL;
 637
 638        pmd = pmd_offset(pud, address);
 639        if (!pmd_present(*pmd))
 640                return NULL;
 641        if (pmd_trans_huge(*pmd))
 642                return NULL;
 643
 644        pte = pte_offset_map(pmd, address);
 645        /* Make a quick check before getting the lock */
 646        if (!sync && !pte_present(*pte)) {
 647                pte_unmap(pte);
 648                return NULL;
 649        }
 650
 651        ptl = pte_lockptr(mm, pmd);
 652check:
 653        spin_lock(ptl);
 654        if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 655                *ptlp = ptl;
 656                return pte;
 657        }
 658        pte_unmap_unlock(pte, ptl);
 659        return NULL;
 660}
 661
 662/**
 663 * page_mapped_in_vma - check whether a page is really mapped in a VMA
 664 * @page: the page to test
 665 * @vma: the VMA to test
 666 *
 667 * Returns 1 if the page is mapped into the page tables of the VMA, 0
 668 * if the page is not mapped into the page tables of this VMA.  Only
 669 * valid for normal file or anonymous VMAs.
 670 */
 671int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 672{
 673        unsigned long address;
 674        pte_t *pte;
 675        spinlock_t *ptl;
 676
 677        address = vma_address(page, vma);
 678        if (address == -EFAULT)         /* out of vma range */
 679                return 0;
 680        pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
 681        if (!pte)                       /* the page is not in this mm */
 682                return 0;
 683        pte_unmap_unlock(pte, ptl);
 684
 685        return 1;
 686}
 687
 688/*
 689 * Subfunctions of page_referenced: page_referenced_one called
 690 * repeatedly from either page_referenced_anon or page_referenced_file.
 691 */
 692int page_referenced_one(struct page *page, struct vm_area_struct *vma,
 693                        unsigned long address, unsigned int *mapcount,
 694                        unsigned long *vm_flags)
 695{
 696        struct mm_struct *mm = vma->vm_mm;
 697        int referenced = 0;
 698
 699        if (unlikely(PageTransHuge(page))) {
 700                pmd_t *pmd;
 701
 702                spin_lock(&mm->page_table_lock);
 703                /*
 704                 * rmap might return false positives; we must filter
 705                 * these out using page_check_address_pmd().
 706                 */
 707                pmd = page_check_address_pmd(page, mm, address,
 708                                             PAGE_CHECK_ADDRESS_PMD_FLAG);
 709                if (!pmd) {
 710                        spin_unlock(&mm->page_table_lock);
 711                        goto out;
 712                }
 713
 714                if (vma->vm_flags & VM_LOCKED) {
 715                        spin_unlock(&mm->page_table_lock);
 716                        *mapcount = 0;  /* break early from loop */
 717                        *vm_flags |= VM_LOCKED;
 718                        goto out;
 719                }
 720
 721                /* go ahead even if the pmd is pmd_trans_splitting() */
 722                if (pmdp_clear_flush_young_notify(vma, address, pmd))
 723                        referenced++;
 724                spin_unlock(&mm->page_table_lock);
 725        } else {
 726                pte_t *pte;
 727                spinlock_t *ptl;
 728
 729                /*
 730                 * rmap might return false positives; we must filter
 731                 * these out using page_check_address().
 732                 */
 733                pte = page_check_address(page, mm, address, &ptl, 0);
 734                if (!pte)
 735                        goto out;
 736
 737                if (vma->vm_flags & VM_LOCKED) {
 738                        pte_unmap_unlock(pte, ptl);
 739                        *mapcount = 0;  /* break early from loop */
 740                        *vm_flags |= VM_LOCKED;
 741                        goto out;
 742                }
 743
 744                if (ptep_clear_flush_young_notify(vma, address, pte)) {
 745                        /*
 746                         * Don't treat a reference through a sequentially read
 747                         * mapping as such.  If the page has been used in
 748                         * another mapping, we will catch it; if this other
 749                         * mapping is already gone, the unmap path will have
 750                         * set PG_referenced or activated the page.
 751                         */
 752                        if (likely(!VM_SequentialReadHint(vma)))
 753                                referenced++;
 754                }
 755                pte_unmap_unlock(pte, ptl);
 756        }
 757
 758        (*mapcount)--;
 759
 760        if (referenced)
 761                *vm_flags |= vma->vm_flags;
 762out:
 763        return referenced;
 764}
 765
 766static int page_referenced_anon(struct page *page,
 767                                struct mem_cgroup *memcg,
 768                                unsigned long *vm_flags)
 769{
 770        unsigned int mapcount;
 771        struct anon_vma *anon_vma;
 772        struct anon_vma_chain *avc;
 773        int referenced = 0;
 774
 775        anon_vma = page_lock_anon_vma(page);
 776        if (!anon_vma)
 777                return referenced;
 778
 779        mapcount = page_mapcount(page);
 780        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
 781                struct vm_area_struct *vma = avc->vma;
 782                unsigned long address = vma_address(page, vma);
 783                if (address == -EFAULT)
 784                        continue;
 785                /*
 786                 * If we are reclaiming on behalf of a cgroup, skip
 787                 * counting on behalf of references from different
 788                 * cgroups
 789                 */
 790                if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 791                        continue;
 792                referenced += page_referenced_one(page, vma, address,
 793                                                  &mapcount, vm_flags);
 794                if (!mapcount)
 795                        break;
 796        }
 797
 798        page_unlock_anon_vma(anon_vma);
 799        return referenced;
 800}
 801
 802/**
 803 * page_referenced_file - referenced check for object-based rmap
 804 * @page: the page we're checking references on.
 805 * @memcg: target memory control group
 806 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 807 *
 808 * For an object-based mapped page, find all the places it is mapped and
 809 * check/clear the referenced flag.  This is done by following the page->mapping
 810 * pointer, then walking the chain of vmas it holds.  It returns the number
 811 * of references it found.
 812 *
 813 * This function is only called from page_referenced for object-based pages.
 814 */
 815static int page_referenced_file(struct page *page,
 816                                struct mem_cgroup *memcg,
 817                                unsigned long *vm_flags)
 818{
 819        unsigned int mapcount;
 820        struct address_space *mapping = page->mapping;
 821        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 822        struct vm_area_struct *vma;
 823        struct prio_tree_iter iter;
 824        int referenced = 0;
 825
 826        /*
 827         * The caller's checks on page->mapping and !PageAnon have made
 828         * sure that this is a file page: the check for page->mapping
 829         * excludes the case just before it gets set on an anon page.
 830         */
 831        BUG_ON(PageAnon(page));
 832
 833        /*
 834         * The page lock not only makes sure that page->mapping cannot
 835         * suddenly be NULLified by truncation, it makes sure that the
 836         * structure at mapping cannot be freed and reused yet,
 837         * so we can safely take mapping->i_mmap_mutex.
 838         */
 839        BUG_ON(!PageLocked(page));
 840
 841        mutex_lock(&mapping->i_mmap_mutex);
 842
 843        /*
 844         * i_mmap_mutex does not stabilize mapcount at all, but mapcount
 845         * is more likely to be accurate if we note it after spinning.
 846         */
 847        mapcount = page_mapcount(page);
 848
 849        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
 850                unsigned long address = vma_address(page, vma);
 851                if (address == -EFAULT)
 852                        continue;
 853                /*
 854                 * If we are reclaiming on behalf of a cgroup, skip
 855                 * counting on behalf of references from different
 856                 * cgroups
 857                 */
 858                if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 859                        continue;
 860                referenced += page_referenced_one(page, vma, address,
 861                                                  &mapcount, vm_flags);
 862                if (!mapcount)
 863                        break;
 864        }
 865
 866        mutex_unlock(&mapping->i_mmap_mutex);
 867        return referenced;
 868}
 869
 870/**
 871 * page_referenced - test if the page was referenced
 872 * @page: the page to test
 873 * @is_locked: caller holds lock on the page
 874 * @memcg: target memory cgroup
 875 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 876 *
 877 * Quick test_and_clear_referenced for all mappings to a page,
 878 * returns the number of ptes which referenced the page.
 879 */
 880int page_referenced(struct page *page,
 881                    int is_locked,
 882                    struct mem_cgroup *memcg,
 883                    unsigned long *vm_flags)
 884{
 885        int referenced = 0;
 886        int we_locked = 0;
 887
 888        *vm_flags = 0;
 889        if (page_mapped(page) && page_rmapping(page)) {
 890                if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 891                        we_locked = trylock_page(page);
 892                        if (!we_locked) {
 893                                referenced++;
 894                                goto out;
 895                        }
 896                }
 897                if (unlikely(PageKsm(page)))
 898                        referenced += page_referenced_ksm(page, memcg,
 899                                                                vm_flags);
 900                else if (PageAnon(page))
 901                        referenced += page_referenced_anon(page, memcg,
 902                                                                vm_flags);
 903                else if (page->mapping)
 904                        referenced += page_referenced_file(page, memcg,
 905                                                                vm_flags);
 906                if (we_locked)
 907                        unlock_page(page);
 908
 909                if (page_test_and_clear_young(page_to_pfn(page)))
 910                        referenced++;
 911        }
 912out:
 913        return referenced;
 914}
 915
 916static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 917                            unsigned long address)
 918{
 919        struct mm_struct *mm = vma->vm_mm;
 920        pte_t *pte;
 921        spinlock_t *ptl;
 922        int ret = 0;
 923
 924        pte = page_check_address(page, mm, address, &ptl, 1);
 925        if (!pte)
 926                goto out;
 927
 928        if (pte_dirty(*pte) || pte_write(*pte)) {
 929                pte_t entry;
 930
 931                flush_cache_page(vma, address, pte_pfn(*pte));
 932                entry = ptep_clear_flush_notify(vma, address, pte);
 933                entry = pte_wrprotect(entry);
 934                entry = pte_mkclean(entry);
 935                set_pte_at(mm, address, pte, entry);
 936                ret = 1;
 937        }
 938
 939        pte_unmap_unlock(pte, ptl);
 940out:
 941        return ret;
 942}
 943
 944static int page_mkclean_file(struct address_space *mapping, struct page *page)
 945{
 946        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 947        struct vm_area_struct *vma;
 948        struct prio_tree_iter iter;
 949        int ret = 0;
 950
 951        BUG_ON(PageAnon(page));
 952
 953        mutex_lock(&mapping->i_mmap_mutex);
 954        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
 955                if (vma->vm_flags & VM_SHARED) {
 956                        unsigned long address = vma_address(page, vma);
 957                        if (address == -EFAULT)
 958                                continue;
 959                        ret += page_mkclean_one(page, vma, address);
 960                }
 961        }
 962        mutex_unlock(&mapping->i_mmap_mutex);
 963        return ret;
 964}
 965
 966int page_mkclean(struct page *page)
 967{
 968        int ret = 0;
 969
 970        BUG_ON(!PageLocked(page));
 971
 972        if (page_mapped(page)) {
 973                struct address_space *mapping = page_mapping(page);
 974                if (mapping) {
 975                        ret = page_mkclean_file(mapping, page);
 976                        if (page_test_and_clear_dirty(page_to_pfn(page), 1))
 977                                ret = 1;
 978                }
 979        }
 980
 981        return ret;
 982}
 983EXPORT_SYMBOL_GPL(page_mkclean);
 984
 985/**
 986 * page_move_anon_rmap - move a page to our anon_vma
 987 * @page:       the page to move to our anon_vma
 988 * @vma:        the vma the page belongs to
 989 * @address:    the user virtual address mapped
 990 *
 991 * When a page belongs exclusively to one process after a COW event,
 992 * that page can be moved into the anon_vma that belongs to just that
 993 * process, so the rmap code will not search the parent or sibling
 994 * processes.
 995 */
 996void page_move_anon_rmap(struct page *page,
 997        struct vm_area_struct *vma, unsigned long address)
 998{
 999        struct anon_vma *anon_vma = vma->anon_vma;
1000
1001        VM_BUG_ON(!PageLocked(page));
1002        VM_BUG_ON(!anon_vma);
1003        VM_BUG_ON(page->index != linear_page_index(vma, address));
1004
1005        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1006        page->mapping = (struct address_space *) anon_vma;
1007}
1008
1009/**
1010 * __page_set_anon_rmap - set up new anonymous rmap
1011 * @page:       Page to add to rmap     
1012 * @vma:        VM area to add page to.
1013 * @address:    User virtual address of the mapping     
1014 * @exclusive:  the page is exclusively owned by the current process
1015 */
1016static void __page_set_anon_rmap(struct page *page,
1017        struct vm_area_struct *vma, unsigned long address, int exclusive)
1018{
1019        struct anon_vma *anon_vma = vma->anon_vma;
1020
1021        BUG_ON(!anon_vma);
1022
1023        if (PageAnon(page))
1024                return;
1025
1026        /*
1027         * If the page isn't exclusively mapped into this vma,
1028         * we must use the _oldest_ possible anon_vma for the
1029         * page mapping!
1030         */
1031        if (!exclusive)
1032                anon_vma = anon_vma->root;
1033
1034        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1035        page->mapping = (struct address_space *) anon_vma;
1036        page->index = linear_page_index(vma, address);
1037}
1038
1039/**
1040 * __page_check_anon_rmap - sanity check anonymous rmap addition
1041 * @page:       the page to add the mapping to
1042 * @vma:        the vm area in which the mapping is added
1043 * @address:    the user virtual address mapped
1044 */
1045static void __page_check_anon_rmap(struct page *page,
1046        struct vm_area_struct *vma, unsigned long address)
1047{
1048#ifdef CONFIG_DEBUG_VM
1049        /*
1050         * The page's anon-rmap details (mapping and index) are guaranteed to
1051         * be set up correctly at this point.
1052         *
1053         * We have exclusion against page_add_anon_rmap because the caller
1054         * always holds the page locked, except if called from page_dup_rmap,
1055         * in which case the page is already known to be setup.
1056         *
1057         * We have exclusion against page_add_new_anon_rmap because those pages
1058         * are initially only visible via the pagetables, and the pte is locked
1059         * over the call to page_add_new_anon_rmap.
1060         */
1061        BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1062        BUG_ON(page->index != linear_page_index(vma, address));
1063#endif
1064}
1065
1066/**
1067 * page_add_anon_rmap - add pte mapping to an anonymous page
1068 * @page:       the page to add the mapping to
1069 * @vma:        the vm area in which the mapping is added
1070 * @address:    the user virtual address mapped
1071 *
1072 * The caller needs to hold the pte lock, and the page must be locked in
1073 * the anon_vma case: to serialize mapping,index checking after setting,
1074 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1075 * (but PageKsm is never downgraded to PageAnon).
1076 */
1077void page_add_anon_rmap(struct page *page,
1078        struct vm_area_struct *vma, unsigned long address)
1079{
1080        do_page_add_anon_rmap(page, vma, address, 0);
1081}
1082
1083/*
1084 * Special version of the above for do_swap_page, which often runs
1085 * into pages that are exclusively owned by the current process.
1086 * Everybody else should continue to use page_add_anon_rmap above.
1087 */
1088void do_page_add_anon_rmap(struct page *page,
1089        struct vm_area_struct *vma, unsigned long address, int exclusive)
1090{
1091        int first = atomic_inc_and_test(&page->_mapcount);
1092        if (first) {
1093                if (!PageTransHuge(page))
1094                        __inc_zone_page_state(page, NR_ANON_PAGES);
1095                else
1096                        __inc_zone_page_state(page,
1097                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1098        }
1099        if (unlikely(PageKsm(page)))
1100                return;
1101
1102        VM_BUG_ON(!PageLocked(page));
1103        /* address might be in next vma when migration races vma_adjust */
1104        if (first)
1105                __page_set_anon_rmap(page, vma, address, exclusive);
1106        else
1107                __page_check_anon_rmap(page, vma, address);
1108}
1109
1110/**
1111 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1112 * @page:       the page to add the mapping to
1113 * @vma:        the vm area in which the mapping is added
1114 * @address:    the user virtual address mapped
1115 *
1116 * Same as page_add_anon_rmap but must only be called on *new* pages.
1117 * This means the inc-and-test can be bypassed.
1118 * Page does not have to be locked.
1119 */
1120void page_add_new_anon_rmap(struct page *page,
1121        struct vm_area_struct *vma, unsigned long address)
1122{
1123        VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1124        SetPageSwapBacked(page);
1125        atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1126        if (!PageTransHuge(page))
1127                __inc_zone_page_state(page, NR_ANON_PAGES);
1128        else
1129                __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1130        __page_set_anon_rmap(page, vma, address, 1);
1131        if (page_evictable(page, vma))
1132                lru_cache_add_lru(page, LRU_ACTIVE_ANON);
1133        else
1134                add_page_to_unevictable_list(page);
1135}
1136
1137/**
1138 * page_add_file_rmap - add pte mapping to a file page
1139 * @page: the page to add the mapping to
1140 *
1141 * The caller needs to hold the pte lock.
1142 */
1143void page_add_file_rmap(struct page *page)
1144{
1145        bool locked;
1146        unsigned long flags;
1147
1148        mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1149        if (atomic_inc_and_test(&page->_mapcount)) {
1150                __inc_zone_page_state(page, NR_FILE_MAPPED);
1151                mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED);
1152        }
1153        mem_cgroup_end_update_page_stat(page, &locked, &flags);
1154}
1155
1156/**
1157 * page_remove_rmap - take down pte mapping from a page
1158 * @page: page to remove mapping from
1159 *
1160 * The caller needs to hold the pte lock.
1161 */
1162void page_remove_rmap(struct page *page)
1163{
1164        bool anon = PageAnon(page);
1165        bool locked;
1166        unsigned long flags;
1167
1168        /*
1169         * The anon case has no mem_cgroup page_stat to update; but may
1170         * uncharge_page() below, where the lock ordering can deadlock if
1171         * we hold the lock against page_stat move: so avoid it on anon.
1172         */
1173        if (!anon)
1174                mem_cgroup_begin_update_page_stat(page, &locked, &flags);
1175
1176        /* page still mapped by someone else? */
1177        if (!atomic_add_negative(-1, &page->_mapcount))
1178                goto out;
1179
1180        /*
1181         * Now that the last pte has gone, s390 must transfer dirty
1182         * flag from storage key to struct page.  We can usually skip
1183         * this if the page is anon, so about to be freed; but perhaps
1184         * not if it's in swapcache - there might be another pte slot
1185         * containing the swap entry, but page not yet written to swap.
1186         */
1187        if ((!anon || PageSwapCache(page)) &&
1188            page_test_and_clear_dirty(page_to_pfn(page), 1))
1189                set_page_dirty(page);
1190        /*
1191         * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED
1192         * and not charged by memcg for now.
1193         */
1194        if (unlikely(PageHuge(page)))
1195                goto out;
1196        if (anon) {
1197                mem_cgroup_uncharge_page(page);
1198                if (!PageTransHuge(page))
1199                        __dec_zone_page_state(page, NR_ANON_PAGES);
1200                else
1201                        __dec_zone_page_state(page,
1202                                              NR_ANON_TRANSPARENT_HUGEPAGES);
1203        } else {
1204                __dec_zone_page_state(page, NR_FILE_MAPPED);
1205                mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED);
1206        }
1207        /*
1208         * It would be tidy to reset the PageAnon mapping here,
1209         * but that might overwrite a racing page_add_anon_rmap
1210         * which increments mapcount after us but sets mapping
1211         * before us: so leave the reset to free_hot_cold_page,
1212         * and remember that it's only reliable while mapped.
1213         * Leaving it set also helps swapoff to reinstate ptes
1214         * faster for those pages still in swapcache.
1215         */
1216out:
1217        if (!anon)
1218                mem_cgroup_end_update_page_stat(page, &locked, &flags);
1219}
1220
1221/*
1222 * Subfunctions of try_to_unmap: try_to_unmap_one called
1223 * repeatedly from try_to_unmap_ksm, try_to_unmap_anon or try_to_unmap_file.
1224 */
1225int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1226                     unsigned long address, enum ttu_flags flags)
1227{
1228        struct mm_struct *mm = vma->vm_mm;
1229        pte_t *pte;
1230        pte_t pteval;
1231        spinlock_t *ptl;
1232        int ret = SWAP_AGAIN;
1233
1234        pte = page_check_address(page, mm, address, &ptl, 0);
1235        if (!pte)
1236                goto out;
1237
1238        /*
1239         * If the page is mlock()d, we cannot swap it out.
1240         * If it's recently referenced (perhaps page_referenced
1241         * skipped over this mm) then we should reactivate it.
1242         */
1243        if (!(flags & TTU_IGNORE_MLOCK)) {
1244                if (vma->vm_flags & VM_LOCKED)
1245                        goto out_mlock;
1246
1247                if (TTU_ACTION(flags) == TTU_MUNLOCK)
1248                        goto out_unmap;
1249        }
1250        if (!(flags & TTU_IGNORE_ACCESS)) {
1251                if (ptep_clear_flush_young_notify(vma, address, pte)) {
1252                        ret = SWAP_FAIL;
1253                        goto out_unmap;
1254                }
1255        }
1256
1257        /* Nuke the page table entry. */
1258        flush_cache_page(vma, address, page_to_pfn(page));
1259        pteval = ptep_clear_flush_notify(vma, address, pte);
1260
1261        /* Move the dirty bit to the physical page now the pte is gone. */
1262        if (pte_dirty(pteval))
1263                set_page_dirty(page);
1264
1265        /* Update high watermark before we lower rss */
1266        update_hiwater_rss(mm);
1267
1268        if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1269                if (PageAnon(page))
1270                        dec_mm_counter(mm, MM_ANONPAGES);
1271                else
1272                        dec_mm_counter(mm, MM_FILEPAGES);
1273                set_pte_at(mm, address, pte,
1274                                swp_entry_to_pte(make_hwpoison_entry(page)));
1275        } else if (PageAnon(page)) {
1276                swp_entry_t entry = { .val = page_private(page) };
1277
1278                if (PageSwapCache(page)) {
1279                        /*
1280                         * Store the swap location in the pte.
1281                         * See handle_pte_fault() ...
1282                         */
1283                        if (swap_duplicate(entry) < 0) {
1284                                set_pte_at(mm, address, pte, pteval);
1285                                ret = SWAP_FAIL;
1286                                goto out_unmap;
1287                        }
1288                        if (list_empty(&mm->mmlist)) {
1289                                spin_lock(&mmlist_lock);
1290                                if (list_empty(&mm->mmlist))
1291                                        list_add(&mm->mmlist, &init_mm.mmlist);
1292                                spin_unlock(&mmlist_lock);
1293                        }
1294                        dec_mm_counter(mm, MM_ANONPAGES);
1295                        inc_mm_counter(mm, MM_SWAPENTS);
1296                } else if (IS_ENABLED(CONFIG_MIGRATION)) {
1297                        /*
1298                         * Store the pfn of the page in a special migration
1299                         * pte. do_swap_page() will wait until the migration
1300                         * pte is removed and then restart fault handling.
1301                         */
1302                        BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION);
1303                        entry = make_migration_entry(page, pte_write(pteval));
1304                }
1305                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1306                BUG_ON(pte_file(*pte));
1307        } else if (IS_ENABLED(CONFIG_MIGRATION) &&
1308                   (TTU_ACTION(flags) == TTU_MIGRATION)) {
1309                /* Establish migration entry for a file page */
1310                swp_entry_t entry;
1311                entry = make_migration_entry(page, pte_write(pteval));
1312                set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1313        } else
1314                dec_mm_counter(mm, MM_FILEPAGES);
1315
1316        page_remove_rmap(page);
1317        page_cache_release(page);
1318
1319out_unmap:
1320        pte_unmap_unlock(pte, ptl);
1321out:
1322        return ret;
1323
1324out_mlock:
1325        pte_unmap_unlock(pte, ptl);
1326
1327
1328        /*
1329         * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1330         * unstable result and race. Plus, We can't wait here because
1331         * we now hold anon_vma->mutex or mapping->i_mmap_mutex.
1332         * if trylock failed, the page remain in evictable lru and later
1333         * vmscan could retry to move the page to unevictable lru if the
1334         * page is actually mlocked.
1335         */
1336        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1337                if (vma->vm_flags & VM_LOCKED) {
1338                        mlock_vma_page(page);
1339                        ret = SWAP_MLOCK;
1340                }
1341                up_read(&vma->vm_mm->mmap_sem);
1342        }
1343        return ret;
1344}
1345
1346/*
1347 * objrmap doesn't work for nonlinear VMAs because the assumption that
1348 * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1349 * Consequently, given a particular page and its ->index, we cannot locate the
1350 * ptes which are mapping that page without an exhaustive linear search.
1351 *
1352 * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1353 * maps the file to which the target page belongs.  The ->vm_private_data field
1354 * holds the current cursor into that scan.  Successive searches will circulate
1355 * around the vma's virtual address space.
1356 *
1357 * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1358 * more scanning pressure is placed against them as well.   Eventually pages
1359 * will become fully unmapped and are eligible for eviction.
1360 *
1361 * For very sparsely populated VMAs this is a little inefficient - chances are
1362 * there there won't be many ptes located within the scan cluster.  In this case
1363 * maybe we could scan further - to the end of the pte page, perhaps.
1364 *
1365 * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
1366 * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
1367 * rather than unmapping them.  If we encounter the "check_page" that vmscan is
1368 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1369 */
1370#define CLUSTER_SIZE    min(32*PAGE_SIZE, PMD_SIZE)
1371#define CLUSTER_MASK    (~(CLUSTER_SIZE - 1))
1372
1373static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1374                struct vm_area_struct *vma, struct page *check_page)
1375{
1376        struct mm_struct *mm = vma->vm_mm;
1377        pgd_t *pgd;
1378        pud_t *pud;
1379        pmd_t *pmd;
1380        pte_t *pte;
1381        pte_t pteval;
1382        spinlock_t *ptl;
1383        struct page *page;
1384        unsigned long address;
1385        unsigned long end;
1386        int ret = SWAP_AGAIN;
1387        int locked_vma = 0;
1388
1389        address = (vma->vm_start + cursor) & CLUSTER_MASK;
1390        end = address + CLUSTER_SIZE;
1391        if (address < vma->vm_start)
1392                address = vma->vm_start;
1393        if (end > vma->vm_end)
1394                end = vma->vm_end;
1395
1396        pgd = pgd_offset(mm, address);
1397        if (!pgd_present(*pgd))
1398                return ret;
1399
1400        pud = pud_offset(pgd, address);
1401        if (!pud_present(*pud))
1402                return ret;
1403
1404        pmd = pmd_offset(pud, address);
1405        if (!pmd_present(*pmd))
1406                return ret;
1407
1408        /*
1409         * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1410         * keep the sem while scanning the cluster for mlocking pages.
1411         */
1412        if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1413                locked_vma = (vma->vm_flags & VM_LOCKED);
1414                if (!locked_vma)
1415                        up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1416        }
1417
1418        pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1419
1420        /* Update high watermark before we lower rss */
1421        update_hiwater_rss(mm);
1422
1423        for (; address < end; pte++, address += PAGE_SIZE) {
1424                if (!pte_present(*pte))
1425                        continue;
1426                page = vm_normal_page(vma, address, *pte);
1427                BUG_ON(!page || PageAnon(page));
1428
1429                if (locked_vma) {
1430                        mlock_vma_page(page);   /* no-op if already mlocked */
1431                        if (page == check_page)
1432                                ret = SWAP_MLOCK;
1433                        continue;       /* don't unmap */
1434                }
1435
1436                if (ptep_clear_flush_young_notify(vma, address, pte))
1437                        continue;
1438
1439                /* Nuke the page table entry. */
1440                flush_cache_page(vma, address, pte_pfn(*pte));
1441                pteval = ptep_clear_flush_notify(vma, address, pte);
1442
1443                /* If nonlinear, store the file page offset in the pte. */
1444                if (page->index != linear_page_index(vma, address))
1445                        set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
1446
1447                /* Move the dirty bit to the physical page now the pte is gone. */
1448                if (pte_dirty(pteval))
1449                        set_page_dirty(page);
1450
1451                page_remove_rmap(page);
1452                page_cache_release(page);
1453                dec_mm_counter(mm, MM_FILEPAGES);
1454                (*mapcount)--;
1455        }
1456        pte_unmap_unlock(pte - 1, ptl);
1457        if (locked_vma)
1458                up_read(&vma->vm_mm->mmap_sem);
1459        return ret;
1460}
1461
1462bool is_vma_temporary_stack(struct vm_area_struct *vma)
1463{
1464        int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1465
1466        if (!maybe_stack)
1467                return false;
1468
1469        if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1470                                                VM_STACK_INCOMPLETE_SETUP)
1471                return true;
1472
1473        return false;
1474}
1475
1476/**
1477 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1478 * rmap method
1479 * @page: the page to unmap/unlock
1480 * @flags: action and flags
1481 *
1482 * Find all the mappings of a page using the mapping pointer and the vma chains
1483 * contained in the anon_vma struct it points to.
1484 *
1485 * This function is only called from try_to_unmap/try_to_munlock for
1486 * anonymous pages.
1487 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1488 * where the page was found will be held for write.  So, we won't recheck
1489 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1490 * 'LOCKED.
1491 */
1492static int try_to_unmap_anon(struct page *page, enum ttu_flags flags)
1493{
1494        struct anon_vma *anon_vma;
1495        struct anon_vma_chain *avc;
1496        int ret = SWAP_AGAIN;
1497
1498        anon_vma = page_lock_anon_vma(page);
1499        if (!anon_vma)
1500                return ret;
1501
1502        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1503                struct vm_area_struct *vma = avc->vma;
1504                unsigned long address;
1505
1506                /*
1507                 * During exec, a temporary VMA is setup and later moved.
1508                 * The VMA is moved under the anon_vma lock but not the
1509                 * page tables leading to a race where migration cannot
1510                 * find the migration ptes. Rather than increasing the
1511                 * locking requirements of exec(), migration skips
1512                 * temporary VMAs until after exec() completes.
1513                 */
1514                if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1515                                is_vma_temporary_stack(vma))
1516                        continue;
1517
1518                address = vma_address(page, vma);
1519                if (address == -EFAULT)
1520                        continue;
1521                ret = try_to_unmap_one(page, vma, address, flags);
1522                if (ret != SWAP_AGAIN || !page_mapped(page))
1523                        break;
1524        }
1525
1526        page_unlock_anon_vma(anon_vma);
1527        return ret;
1528}
1529
1530/**
1531 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1532 * @page: the page to unmap/unlock
1533 * @flags: action and flags
1534 *
1535 * Find all the mappings of a page using the mapping pointer and the vma chains
1536 * contained in the address_space struct it points to.
1537 *
1538 * This function is only called from try_to_unmap/try_to_munlock for
1539 * object-based pages.
1540 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1541 * where the page was found will be held for write.  So, we won't recheck
1542 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1543 * 'LOCKED.
1544 */
1545static int try_to_unmap_file(struct page *page, enum ttu_flags flags)
1546{
1547        struct address_space *mapping = page->mapping;
1548        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1549        struct vm_area_struct *vma;
1550        struct prio_tree_iter iter;
1551        int ret = SWAP_AGAIN;
1552        unsigned long cursor;
1553        unsigned long max_nl_cursor = 0;
1554        unsigned long max_nl_size = 0;
1555        unsigned int mapcount;
1556
1557        mutex_lock(&mapping->i_mmap_mutex);
1558        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1559                unsigned long address = vma_address(page, vma);
1560                if (address == -EFAULT)
1561                        continue;
1562                ret = try_to_unmap_one(page, vma, address, flags);
1563                if (ret != SWAP_AGAIN || !page_mapped(page))
1564                        goto out;
1565        }
1566
1567        if (list_empty(&mapping->i_mmap_nonlinear))
1568                goto out;
1569
1570        /*
1571         * We don't bother to try to find the munlocked page in nonlinears.
1572         * It's costly. Instead, later, page reclaim logic may call
1573         * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily.
1574         */
1575        if (TTU_ACTION(flags) == TTU_MUNLOCK)
1576                goto out;
1577
1578        list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1579                                                shared.vm_set.list) {
1580                cursor = (unsigned long) vma->vm_private_data;
1581                if (cursor > max_nl_cursor)
1582                        max_nl_cursor = cursor;
1583                cursor = vma->vm_end - vma->vm_start;
1584                if (cursor > max_nl_size)
1585                        max_nl_size = cursor;
1586        }
1587
1588        if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */
1589                ret = SWAP_FAIL;
1590                goto out;
1591        }
1592
1593        /*
1594         * We don't try to search for this page in the nonlinear vmas,
1595         * and page_referenced wouldn't have found it anyway.  Instead
1596         * just walk the nonlinear vmas trying to age and unmap some.
1597         * The mapcount of the page we came in with is irrelevant,
1598         * but even so use it as a guide to how hard we should try?
1599         */
1600        mapcount = page_mapcount(page);
1601        if (!mapcount)
1602                goto out;
1603        cond_resched();
1604
1605        max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1606        if (max_nl_cursor == 0)
1607                max_nl_cursor = CLUSTER_SIZE;
1608
1609        do {
1610                list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1611                                                shared.vm_set.list) {
1612                        cursor = (unsigned long) vma->vm_private_data;
1613                        while ( cursor < max_nl_cursor &&
1614                                cursor < vma->vm_end - vma->vm_start) {
1615                                if (try_to_unmap_cluster(cursor, &mapcount,
1616                                                vma, page) == SWAP_MLOCK)
1617                                        ret = SWAP_MLOCK;
1618                                cursor += CLUSTER_SIZE;
1619                                vma->vm_private_data = (void *) cursor;
1620                                if ((int)mapcount <= 0)
1621                                        goto out;
1622                        }
1623                        vma->vm_private_data = (void *) max_nl_cursor;
1624                }
1625                cond_resched();
1626                max_nl_cursor += CLUSTER_SIZE;
1627        } while (max_nl_cursor <= max_nl_size);
1628
1629        /*
1630         * Don't loop forever (perhaps all the remaining pages are
1631         * in locked vmas).  Reset cursor on all unreserved nonlinear
1632         * vmas, now forgetting on which ones it had fallen behind.
1633         */
1634        list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1635                vma->vm_private_data = NULL;
1636out:
1637        mutex_unlock(&mapping->i_mmap_mutex);
1638        return ret;
1639}
1640
1641/**
1642 * try_to_unmap - try to remove all page table mappings to a page
1643 * @page: the page to get unmapped
1644 * @flags: action and flags
1645 *
1646 * Tries to remove all the page table entries which are mapping this
1647 * page, used in the pageout path.  Caller must hold the page lock.
1648 * Return values are:
1649 *
1650 * SWAP_SUCCESS - we succeeded in removing all mappings
1651 * SWAP_AGAIN   - we missed a mapping, try again later
1652 * SWAP_FAIL    - the page is unswappable
1653 * SWAP_MLOCK   - page is mlocked.
1654 */
1655int try_to_unmap(struct page *page, enum ttu_flags flags)
1656{
1657        int ret;
1658
1659        BUG_ON(!PageLocked(page));
1660        VM_BUG_ON(!PageHuge(page) && PageTransHuge(page));
1661
1662        if (unlikely(PageKsm(page)))
1663                ret = try_to_unmap_ksm(page, flags);
1664        else if (PageAnon(page))
1665                ret = try_to_unmap_anon(page, flags);
1666        else
1667                ret = try_to_unmap_file(page, flags);
1668        if (ret != SWAP_MLOCK && !page_mapped(page))
1669                ret = SWAP_SUCCESS;
1670        return ret;
1671}
1672
1673/**
1674 * try_to_munlock - try to munlock a page
1675 * @page: the page to be munlocked
1676 *
1677 * Called from munlock code.  Checks all of the VMAs mapping the page
1678 * to make sure nobody else has this page mlocked. The page will be
1679 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1680 *
1681 * Return values are:
1682 *
1683 * SWAP_AGAIN   - no vma is holding page mlocked, or,
1684 * SWAP_AGAIN   - page mapped in mlocked vma -- couldn't acquire mmap sem
1685 * SWAP_FAIL    - page cannot be located at present
1686 * SWAP_MLOCK   - page is now mlocked.
1687 */
1688int try_to_munlock(struct page *page)
1689{
1690        VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1691
1692        if (unlikely(PageKsm(page)))
1693                return try_to_unmap_ksm(page, TTU_MUNLOCK);
1694        else if (PageAnon(page))
1695                return try_to_unmap_anon(page, TTU_MUNLOCK);
1696        else
1697                return try_to_unmap_file(page, TTU_MUNLOCK);
1698}
1699
1700void __put_anon_vma(struct anon_vma *anon_vma)
1701{
1702        struct anon_vma *root = anon_vma->root;
1703
1704        if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1705                anon_vma_free(root);
1706
1707        anon_vma_free(anon_vma);
1708}
1709
1710#ifdef CONFIG_MIGRATION
1711/*
1712 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file():
1713 * Called by migrate.c to remove migration ptes, but might be used more later.
1714 */
1715static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *,
1716                struct vm_area_struct *, unsigned long, void *), void *arg)
1717{
1718        struct anon_vma *anon_vma;
1719        struct anon_vma_chain *avc;
1720        int ret = SWAP_AGAIN;
1721
1722        /*
1723         * Note: remove_migration_ptes() cannot use page_lock_anon_vma()
1724         * because that depends on page_mapped(); but not all its usages
1725         * are holding mmap_sem. Users without mmap_sem are required to
1726         * take a reference count to prevent the anon_vma disappearing
1727         */
1728        anon_vma = page_anon_vma(page);
1729        if (!anon_vma)
1730                return ret;
1731        anon_vma_lock(anon_vma);
1732        list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
1733                struct vm_area_struct *vma = avc->vma;
1734                unsigned long address = vma_address(page, vma);
1735                if (address == -EFAULT)
1736                        continue;
1737                ret = rmap_one(page, vma, address, arg);
1738                if (ret != SWAP_AGAIN)
1739                        break;
1740        }
1741        anon_vma_unlock(anon_vma);
1742        return ret;
1743}
1744
1745static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *,
1746                struct vm_area_struct *, unsigned long, void *), void *arg)
1747{
1748        struct address_space *mapping = page->mapping;
1749        pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1750        struct vm_area_struct *vma;
1751        struct prio_tree_iter iter;
1752        int ret = SWAP_AGAIN;
1753
1754        if (!mapping)
1755                return ret;
1756        mutex_lock(&mapping->i_mmap_mutex);
1757        vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1758                unsigned long address = vma_address(page, vma);
1759                if (address == -EFAULT)
1760                        continue;
1761                ret = rmap_one(page, vma, address, arg);
1762                if (ret != SWAP_AGAIN)
1763                        break;
1764        }
1765        /*
1766         * No nonlinear handling: being always shared, nonlinear vmas
1767         * never contain migration ptes.  Decide what to do about this
1768         * limitation to linear when we need rmap_walk() on nonlinear.
1769         */
1770        mutex_unlock(&mapping->i_mmap_mutex);
1771        return ret;
1772}
1773
1774int rmap_walk(struct page *page, int (*rmap_one)(struct page *,
1775                struct vm_area_struct *, unsigned long, void *), void *arg)
1776{
1777        VM_BUG_ON(!PageLocked(page));
1778
1779        if (unlikely(PageKsm(page)))
1780                return rmap_walk_ksm(page, rmap_one, arg);
1781        else if (PageAnon(page))
1782                return rmap_walk_anon(page, rmap_one, arg);
1783        else
1784                return rmap_walk_file(page, rmap_one, arg);
1785}
1786#endif /* CONFIG_MIGRATION */
1787
1788#ifdef CONFIG_HUGETLB_PAGE
1789/*
1790 * The following three functions are for anonymous (private mapped) hugepages.
1791 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1792 * and no lru code, because we handle hugepages differently from common pages.
1793 */
1794static void __hugepage_set_anon_rmap(struct page *page,
1795        struct vm_area_struct *vma, unsigned long address, int exclusive)
1796{
1797        struct anon_vma *anon_vma = vma->anon_vma;
1798
1799        BUG_ON(!anon_vma);
1800
1801        if (PageAnon(page))
1802                return;
1803        if (!exclusive)
1804                anon_vma = anon_vma->root;
1805
1806        anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1807        page->mapping = (struct address_space *) anon_vma;
1808        page->index = linear_page_index(vma, address);
1809}
1810
1811void hugepage_add_anon_rmap(struct page *page,
1812                            struct vm_area_struct *vma, unsigned long address)
1813{
1814        struct anon_vma *anon_vma = vma->anon_vma;
1815        int first;
1816
1817        BUG_ON(!PageLocked(page));
1818        BUG_ON(!anon_vma);
1819        /* address might be in next vma when migration races vma_adjust */
1820        first = atomic_inc_and_test(&page->_mapcount);
1821        if (first)
1822                __hugepage_set_anon_rmap(page, vma, address, 0);
1823}
1824
1825void hugepage_add_new_anon_rmap(struct page *page,
1826                        struct vm_area_struct *vma, unsigned long address)
1827{
1828        BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1829        atomic_set(&page->_mapcount, 0);
1830        __hugepage_set_anon_rmap(page, vma, address, 1);
1831}
1832#endif /* CONFIG_HUGETLB_PAGE */
1833
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.