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 *       hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
  27 *         mapping->i_mmap_rwsem
  28 *           anon_vma->rwsem
  29 *             mm->page_table_lock or pte_lock
  30 *               zone->lru_lock (in mark_page_accessed, isolate_lru_page)
  31 *               swap_lock (in swap_duplicate, swap_info_get)
  32 *                 mmlist_lock (in mmput, drain_mmlist and others)
  33 *                 mapping->private_lock (in __set_page_dirty_buffers)
  34 *                   mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
  35 *                     mapping->tree_lock (widely used)
  36 *                 inode->i_lock (in set_page_dirty's __mark_inode_dirty)
  37 *                 bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
  38 *                   sb_lock (within inode_lock in fs/fs-writeback.c)
  39 *                   mapping->tree_lock (widely used, in set_page_dirty,
  40 *                             in arch-dependent flush_dcache_mmap_lock,
  41 *                             within bdi.wb->list_lock in __sync_single_inode)
  42 *
  43 * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
  44 *   ->tasklist_lock
  45 *     pte map lock
  46 */
  47
  48#include <linux/mm.h>
  49#include <linux/pagemap.h>
  50#include <linux/swap.h>
  51#include <linux/swapops.h>
  52#include <linux/slab.h>
  53#include <linux/init.h>
  54#include <linux/ksm.h>
  55#include <linux/rmap.h>
  56#include <linux/rcupdate.h>
  57#include <linux/export.h>
  58#include <linux/memcontrol.h>
  59#include <linux/mmu_notifier.h>
  60#include <linux/migrate.h>
  61#include <linux/hugetlb.h>
  62#include <linux/backing-dev.h>
  63#include <linux/page_idle.h>
  64
  65#include <asm/tlbflush.h>
  66
  67#include <trace/events/tlb.h>
  68
  69#include "internal.h"
  70
  71static struct kmem_cache *anon_vma_cachep;
  72static struct kmem_cache *anon_vma_chain_cachep;
  73
  74static inline struct anon_vma *anon_vma_alloc(void)
  75{
  76	struct anon_vma *anon_vma;
  77
  78	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  79	if (anon_vma) {
  80		atomic_set(&anon_vma->refcount, 1);
  81		anon_vma->degree = 1;	/* Reference for first vma */
  82		anon_vma->parent = anon_vma;
  83		/*
  84		 * Initialise the anon_vma root to point to itself. If called
  85		 * from fork, the root will be reset to the parents anon_vma.
  86		 */
  87		anon_vma->root = anon_vma;
  88	}
  89
  90	return anon_vma;
  91}
  92
  93static inline void anon_vma_free(struct anon_vma *anon_vma)
  94{
  95	VM_BUG_ON(atomic_read(&anon_vma->refcount));
  96
  97	/*
  98	 * Synchronize against page_lock_anon_vma_read() such that
  99	 * we can safely hold the lock without the anon_vma getting
 100	 * freed.
 101	 *
 102	 * Relies on the full mb implied by the atomic_dec_and_test() from
 103	 * put_anon_vma() against the acquire barrier implied by
 104	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
 105	 *
 106	 * page_lock_anon_vma_read()	VS	put_anon_vma()
 107	 *   down_read_trylock()		  atomic_dec_and_test()
 108	 *   LOCK				  MB
 109	 *   atomic_read()			  rwsem_is_locked()
 110	 *
 111	 * LOCK should suffice since the actual taking of the lock must
 112	 * happen _before_ what follows.
 113	 */
 114	might_sleep();
 115	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
 116		anon_vma_lock_write(anon_vma);
 117		anon_vma_unlock_write(anon_vma);
 118	}
 119
 120	kmem_cache_free(anon_vma_cachep, anon_vma);
 121}
 122
 123static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 124{
 125	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 126}
 127
 128static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 129{
 130	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 131}
 132
 133static void anon_vma_chain_link(struct vm_area_struct *vma,
 134				struct anon_vma_chain *avc,
 135				struct anon_vma *anon_vma)
 136{
 137	avc->vma = vma;
 138	avc->anon_vma = anon_vma;
 139	list_add(&avc->same_vma, &vma->anon_vma_chain);
 140	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 141}
 142
 143/**
 144 * anon_vma_prepare - attach an anon_vma to a memory region
 145 * @vma: the memory region in question
 146 *
 147 * This makes sure the memory mapping described by 'vma' has
 148 * an 'anon_vma' attached to it, so that we can associate the
 149 * anonymous pages mapped into it with that anon_vma.
 150 *
 151 * The common case will be that we already have one, but if
 152 * not we either need to find an adjacent mapping that we
 153 * can re-use the anon_vma from (very common when the only
 154 * reason for splitting a vma has been mprotect()), or we
 155 * allocate a new one.
 156 *
 157 * Anon-vma allocations are very subtle, because we may have
 158 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
 159 * and that may actually touch the spinlock even in the newly
 160 * allocated vma (it depends on RCU to make sure that the
 161 * anon_vma isn't actually destroyed).
 162 *
 163 * As a result, we need to do proper anon_vma locking even
 164 * for the new allocation. At the same time, we do not want
 165 * to do any locking for the common case of already having
 166 * an anon_vma.
 167 *
 168 * This must be called with the mmap_sem held for reading.
 169 */
 170int anon_vma_prepare(struct vm_area_struct *vma)
 171{
 172	struct anon_vma *anon_vma = vma->anon_vma;
 173	struct anon_vma_chain *avc;
 174
 175	might_sleep();
 176	if (unlikely(!anon_vma)) {
 177		struct mm_struct *mm = vma->vm_mm;
 178		struct anon_vma *allocated;
 179
 180		avc = anon_vma_chain_alloc(GFP_KERNEL);
 181		if (!avc)
 182			goto out_enomem;
 183
 184		anon_vma = find_mergeable_anon_vma(vma);
 185		allocated = NULL;
 186		if (!anon_vma) {
 187			anon_vma = anon_vma_alloc();
 188			if (unlikely(!anon_vma))
 189				goto out_enomem_free_avc;
 190			allocated = anon_vma;
 191		}
 192
 193		anon_vma_lock_write(anon_vma);
 194		/* page_table_lock to protect against threads */
 195		spin_lock(&mm->page_table_lock);
 196		if (likely(!vma->anon_vma)) {
 197			vma->anon_vma = anon_vma;
 198			anon_vma_chain_link(vma, avc, anon_vma);
 199			/* vma reference or self-parent link for new root */
 200			anon_vma->degree++;
 201			allocated = NULL;
 202			avc = NULL;
 203		}
 204		spin_unlock(&mm->page_table_lock);
 205		anon_vma_unlock_write(anon_vma);
 206
 207		if (unlikely(allocated))
 208			put_anon_vma(allocated);
 209		if (unlikely(avc))
 210			anon_vma_chain_free(avc);
 211	}
 212	return 0;
 213
 214 out_enomem_free_avc:
 215	anon_vma_chain_free(avc);
 216 out_enomem:
 217	return -ENOMEM;
 218}
 219
 220/*
 221 * This is a useful helper function for locking the anon_vma root as
 222 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 223 * have the same vma.
 224 *
 225 * Such anon_vma's should have the same root, so you'd expect to see
 226 * just a single mutex_lock for the whole traversal.
 227 */
 228static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 229{
 230	struct anon_vma *new_root = anon_vma->root;
 231	if (new_root != root) {
 232		if (WARN_ON_ONCE(root))
 233			up_write(&root->rwsem);
 234		root = new_root;
 235		down_write(&root->rwsem);
 236	}
 237	return root;
 238}
 239
 240static inline void unlock_anon_vma_root(struct anon_vma *root)
 241{
 242	if (root)
 243		up_write(&root->rwsem);
 244}
 245
 246/*
 247 * Attach the anon_vmas from src to dst.
 248 * Returns 0 on success, -ENOMEM on failure.
 249 *
 250 * If dst->anon_vma is NULL this function tries to find and reuse existing
 251 * anon_vma which has no vmas and only one child anon_vma. This prevents
 252 * degradation of anon_vma hierarchy to endless linear chain in case of
 253 * constantly forking task. On the other hand, an anon_vma with more than one
 254 * child isn't reused even if there was no alive vma, thus rmap walker has a
 255 * good chance of avoiding scanning the whole hierarchy when it searches where
 256 * page is mapped.
 257 */
 258int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 259{
 260	struct anon_vma_chain *avc, *pavc;
 261	struct anon_vma *root = NULL;
 262
 263	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 264		struct anon_vma *anon_vma;
 265
 266		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 267		if (unlikely(!avc)) {
 268			unlock_anon_vma_root(root);
 269			root = NULL;
 270			avc = anon_vma_chain_alloc(GFP_KERNEL);
 271			if (!avc)
 272				goto enomem_failure;
 273		}
 274		anon_vma = pavc->anon_vma;
 275		root = lock_anon_vma_root(root, anon_vma);
 276		anon_vma_chain_link(dst, avc, anon_vma);
 277
 278		/*
 279		 * Reuse existing anon_vma if its degree lower than two,
 280		 * that means it has no vma and only one anon_vma child.
 281		 *
 282		 * Do not chose parent anon_vma, otherwise first child
 283		 * will always reuse it. Root anon_vma is never reused:
 284		 * it has self-parent reference and at least one child.
 285		 */
 286		if (!dst->anon_vma && anon_vma != src->anon_vma &&
 287				anon_vma->degree < 2)
 288			dst->anon_vma = anon_vma;
 289	}
 290	if (dst->anon_vma)
 291		dst->anon_vma->degree++;
 292	unlock_anon_vma_root(root);
 293	return 0;
 294
 295 enomem_failure:
 296	/*
 297	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
 298	 * decremented in unlink_anon_vmas().
 299	 * We can safely do this because callers of anon_vma_clone() don't care
 300	 * about dst->anon_vma if anon_vma_clone() failed.
 301	 */
 302	dst->anon_vma = NULL;
 303	unlink_anon_vmas(dst);
 304	return -ENOMEM;
 305}
 306
 307/*
 308 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 309 * the corresponding VMA in the parent process is attached to.
 310 * Returns 0 on success, non-zero on failure.
 311 */
 312int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 313{
 314	struct anon_vma_chain *avc;
 315	struct anon_vma *anon_vma;
 316	int error;
 317
 318	/* Don't bother if the parent process has no anon_vma here. */
 319	if (!pvma->anon_vma)
 320		return 0;
 321
 322	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
 323	vma->anon_vma = NULL;
 324
 325	/*
 326	 * First, attach the new VMA to the parent VMA's anon_vmas,
 327	 * so rmap can find non-COWed pages in child processes.
 328	 */
 329	error = anon_vma_clone(vma, pvma);
 330	if (error)
 331		return error;
 332
 333	/* An existing anon_vma has been reused, all done then. */
 334	if (vma->anon_vma)
 335		return 0;
 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	anon_vma->parent = pvma->anon_vma;
 351	/*
 352	 * With refcounts, an anon_vma can stay around longer than the
 353	 * process it belongs to. The root anon_vma needs to be pinned until
 354	 * this anon_vma is freed, because the lock lives in the root.
 355	 */
 356	get_anon_vma(anon_vma->root);
 357	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 358	vma->anon_vma = anon_vma;
 359	anon_vma_lock_write(anon_vma);
 360	anon_vma_chain_link(vma, avc, anon_vma);
 361	anon_vma->parent->degree++;
 362	anon_vma_unlock_write(anon_vma);
 363
 364	return 0;
 365
 366 out_error_free_anon_vma:
 367	put_anon_vma(anon_vma);
 368 out_error:
 369	unlink_anon_vmas(vma);
 370	return -ENOMEM;
 371}
 372
 373void unlink_anon_vmas(struct vm_area_struct *vma)
 374{
 375	struct anon_vma_chain *avc, *next;
 376	struct anon_vma *root = NULL;
 377
 378	/*
 379	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 380	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 381	 */
 382	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 383		struct anon_vma *anon_vma = avc->anon_vma;
 384
 385		root = lock_anon_vma_root(root, anon_vma);
 386		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 387
 388		/*
 389		 * Leave empty anon_vmas on the list - we'll need
 390		 * to free them outside the lock.
 391		 */
 392		if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
 393			anon_vma->parent->degree--;
 394			continue;
 395		}
 396
 397		list_del(&avc->same_vma);
 398		anon_vma_chain_free(avc);
 399	}
 400	if (vma->anon_vma)
 401		vma->anon_vma->degree--;
 402	unlock_anon_vma_root(root);
 403
 404	/*
 405	 * Iterate the list once more, it now only contains empty and unlinked
 406	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 407	 * needing to write-acquire the anon_vma->root->rwsem.
 408	 */
 409	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 410		struct anon_vma *anon_vma = avc->anon_vma;
 411
 412		BUG_ON(anon_vma->degree);
 413		put_anon_vma(anon_vma);
 414
 415		list_del(&avc->same_vma);
 416		anon_vma_chain_free(avc);
 417	}
 418}
 419
 420static void anon_vma_ctor(void *data)
 421{
 422	struct anon_vma *anon_vma = data;
 423
 424	init_rwsem(&anon_vma->rwsem);
 425	atomic_set(&anon_vma->refcount, 0);
 426	anon_vma->rb_root = RB_ROOT;
 427}
 428
 429void __init anon_vma_init(void)
 430{
 431	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 432			0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
 433			anon_vma_ctor);
 434	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
 435			SLAB_PANIC|SLAB_ACCOUNT);
 436}
 437
 438/*
 439 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 440 *
 441 * Since there is no serialization what so ever against page_remove_rmap()
 442 * the best this function can do is return a locked anon_vma that might
 443 * have been relevant to this page.
 444 *
 445 * The page might have been remapped to a different anon_vma or the anon_vma
 446 * returned may already be freed (and even reused).
 447 *
 448 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 449 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 450 * ensure that any anon_vma obtained from the page will still be valid for as
 451 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 452 *
 453 * All users of this function must be very careful when walking the anon_vma
 454 * chain and verify that the page in question is indeed mapped in it
 455 * [ something equivalent to page_mapped_in_vma() ].
 456 *
 457 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
 458 * that the anon_vma pointer from page->mapping is valid if there is a
 459 * mapcount, we can dereference the anon_vma after observing those.
 460 */
 461struct anon_vma *page_get_anon_vma(struct page *page)
 462{
 463	struct anon_vma *anon_vma = NULL;
 464	unsigned long anon_mapping;
 465
 466	rcu_read_lock();
 467	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
 468	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 469		goto out;
 470	if (!page_mapped(page))
 471		goto out;
 472
 473	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 474	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 475		anon_vma = NULL;
 476		goto out;
 477	}
 478
 479	/*
 480	 * If this page is still mapped, then its anon_vma cannot have been
 481	 * freed.  But if it has been unmapped, we have no security against the
 482	 * anon_vma structure being freed and reused (for another anon_vma:
 483	 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
 484	 * above cannot corrupt).
 485	 */
 486	if (!page_mapped(page)) {
 487		rcu_read_unlock();
 488		put_anon_vma(anon_vma);
 489		return NULL;
 490	}
 491out:
 492	rcu_read_unlock();
 493
 494	return anon_vma;
 495}
 496
 497/*
 498 * Similar to page_get_anon_vma() except it locks the anon_vma.
 499 *
 500 * Its a little more complex as it tries to keep the fast path to a single
 501 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 502 * reference like with page_get_anon_vma() and then block on the mutex.
 503 */
 504struct anon_vma *page_lock_anon_vma_read(struct page *page)
 505{
 506	struct anon_vma *anon_vma = NULL;
 507	struct anon_vma *root_anon_vma;
 508	unsigned long anon_mapping;
 509
 510	rcu_read_lock();
 511	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
 512	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 513		goto out;
 514	if (!page_mapped(page))
 515		goto out;
 516
 517	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 518	root_anon_vma = READ_ONCE(anon_vma->root);
 519	if (down_read_trylock(&root_anon_vma->rwsem)) {
 520		/*
 521		 * If the page is still mapped, then this anon_vma is still
 522		 * its anon_vma, and holding the mutex ensures that it will
 523		 * not go away, see anon_vma_free().
 524		 */
 525		if (!page_mapped(page)) {
 526			up_read(&root_anon_vma->rwsem);
 527			anon_vma = NULL;
 528		}
 529		goto out;
 530	}
 531
 532	/* trylock failed, we got to sleep */
 533	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 534		anon_vma = NULL;
 535		goto out;
 536	}
 537
 538	if (!page_mapped(page)) {
 539		rcu_read_unlock();
 540		put_anon_vma(anon_vma);
 541		return NULL;
 
 542	}
 543
 544	/* we pinned the anon_vma, its safe to sleep */
 545	rcu_read_unlock();
 546	anon_vma_lock_read(anon_vma);
 547
 548	if (atomic_dec_and_test(&anon_vma->refcount)) {
 549		/*
 550		 * Oops, we held the last refcount, release the lock
 551		 * and bail -- can't simply use put_anon_vma() because
 552		 * we'll deadlock on the anon_vma_lock_write() recursion.
 553		 */
 554		anon_vma_unlock_read(anon_vma);
 555		__put_anon_vma(anon_vma);
 556		anon_vma = NULL;
 557	}
 558
 559	return anon_vma;
 560
 561out:
 562	rcu_read_unlock();
 563	return anon_vma;
 564}
 565
 566void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
 567{
 568	anon_vma_unlock_read(anon_vma);
 569}
 570
 571#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 572/*
 573 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
 574 * important if a PTE was dirty when it was unmapped that it's flushed
 575 * before any IO is initiated on the page to prevent lost writes. Similarly,
 576 * it must be flushed before freeing to prevent data leakage.
 577 */
 578void try_to_unmap_flush(void)
 
 579{
 580	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 581	int cpu;
 582
 583	if (!tlb_ubc->flush_required)
 584		return;
 585
 586	cpu = get_cpu();
 587
 588	if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
 589		count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
 590		local_flush_tlb();
 591		trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
 592	}
 593
 594	if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
 595		flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
 596	cpumask_clear(&tlb_ubc->cpumask);
 597	tlb_ubc->flush_required = false;
 598	tlb_ubc->writable = false;
 599	put_cpu();
 600}
 601
 602/* Flush iff there are potentially writable TLB entries that can race with IO */
 603void try_to_unmap_flush_dirty(void)
 604{
 605	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 606
 607	if (tlb_ubc->writable)
 608		try_to_unmap_flush();
 609}
 610
 611static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
 612		struct page *page, bool writable)
 613{
 614	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 615
 616	cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
 617	tlb_ubc->flush_required = true;
 618
 619	/*
 620	 * If the PTE was dirty then it's best to assume it's writable. The
 621	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
 622	 * before the page is queued for IO.
 623	 */
 624	if (writable)
 625		tlb_ubc->writable = true;
 626}
 627
 628/*
 629 * Returns true if the TLB flush should be deferred to the end of a batch of
 630 * unmap operations to reduce IPIs.
 631 */
 632static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 633{
 634	bool should_defer = false;
 635
 636	if (!(flags & TTU_BATCH_FLUSH))
 637		return false;
 638
 639	/* If remote CPUs need to be flushed then defer batch the flush */
 640	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
 641		should_defer = true;
 642	put_cpu();
 643
 644	return should_defer;
 645}
 646#else
 647static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
 648		struct page *page, bool writable)
 649{
 650}
 651
 652static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 653{
 654	return false;
 655}
 656#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
 657
 658/*
 659 * At what user virtual address is page expected in vma?
 660 * Caller should check the page is actually part of the vma.
 661 */
 662unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 663{
 664	unsigned long address;
 665	if (PageAnon(page)) {
 666		struct anon_vma *page__anon_vma = page_anon_vma(page);
 667		/*
 668		 * Note: swapoff's unuse_vma() is more efficient with this
 669		 * check, and needs it to match anon_vma when KSM is active.
 670		 */
 671		if (!vma->anon_vma || !page__anon_vma ||
 672		    vma->anon_vma->root != page__anon_vma->root)
 673			return -EFAULT;
 674	} else if (page->mapping) {
 675		if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
 
 676			return -EFAULT;
 677	} else
 678		return -EFAULT;
 679	address = __vma_address(page, vma);
 680	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 681		return -EFAULT;
 682	return address;
 683}
 684
 685pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 686{
 687	pgd_t *pgd;
 688	pud_t *pud;
 689	pmd_t *pmd = NULL;
 690	pmd_t pmde;
 691
 692	pgd = pgd_offset(mm, address);
 693	if (!pgd_present(*pgd))
 694		goto out;
 695
 696	pud = pud_offset(pgd, address);
 697	if (!pud_present(*pud))
 698		goto out;
 699
 700	pmd = pmd_offset(pud, address);
 701	/*
 702	 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
 703	 * without holding anon_vma lock for write.  So when looking for a
 704	 * genuine pmde (in which to find pte), test present and !THP together.
 705	 */
 706	pmde = *pmd;
 707	barrier();
 708	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
 709		pmd = NULL;
 710out:
 711	return pmd;
 712}
 713
 714/*
 715 * Check that @page is mapped at @address into @mm.
 716 *
 717 * If @sync is false, page_check_address may perform a racy check to avoid
 718 * the page table lock when the pte is not present (helpful when reclaiming
 719 * highly shared pages).
 720 *
 721 * On success returns with pte mapped and locked.
 722 */
 723pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
 724			  unsigned long address, spinlock_t **ptlp, int sync)
 725{
 726	pmd_t *pmd;
 727	pte_t *pte;
 728	spinlock_t *ptl;
 729
 730	if (unlikely(PageHuge(page))) {
 731		/* when pud is not present, pte will be NULL */
 732		pte = huge_pte_offset(mm, address);
 733		if (!pte)
 734			return NULL;
 735
 736		ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
 737		goto check;
 738	}
 739
 740	pmd = mm_find_pmd(mm, address);
 741	if (!pmd)
 742		return NULL;
 743
 
 
 
 744	pte = pte_offset_map(pmd, address);
 745	/* Make a quick check before getting the lock */
 746	if (!sync && !pte_present(*pte)) {
 747		pte_unmap(pte);
 748		return NULL;
 749	}
 750
 751	ptl = pte_lockptr(mm, pmd);
 752check:
 753	spin_lock(ptl);
 754	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 755		*ptlp = ptl;
 756		return pte;
 757	}
 758	pte_unmap_unlock(pte, ptl);
 759	return NULL;
 760}
 761
 762/**
 763 * page_mapped_in_vma - check whether a page is really mapped in a VMA
 764 * @page: the page to test
 765 * @vma: the VMA to test
 766 *
 767 * Returns 1 if the page is mapped into the page tables of the VMA, 0
 768 * if the page is not mapped into the page tables of this VMA.  Only
 769 * valid for normal file or anonymous VMAs.
 770 */
 771int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
 772{
 773	unsigned long address;
 774	pte_t *pte;
 775	spinlock_t *ptl;
 776
 777	address = __vma_address(page, vma);
 778	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 779		return 0;
 780	pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
 781	if (!pte)			/* the page is not in this mm */
 782		return 0;
 783	pte_unmap_unlock(pte, ptl);
 784
 785	return 1;
 786}
 787
 788#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 789/*
 790 * Check that @page is mapped at @address into @mm. In contrast to
 791 * page_check_address(), this function can handle transparent huge pages.
 792 *
 793 * On success returns true with pte mapped and locked. For PMD-mapped
 794 * transparent huge pages *@ptep is set to NULL.
 795 */
 796bool page_check_address_transhuge(struct page *page, struct mm_struct *mm,
 797				  unsigned long address, pmd_t **pmdp,
 798				  pte_t **ptep, spinlock_t **ptlp)
 799{
 800	pgd_t *pgd;
 801	pud_t *pud;
 802	pmd_t *pmd;
 803	pte_t *pte;
 804	spinlock_t *ptl;
 805
 806	if (unlikely(PageHuge(page))) {
 807		/* when pud is not present, pte will be NULL */
 808		pte = huge_pte_offset(mm, address);
 809		if (!pte)
 810			return false;
 811
 812		ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
 813		pmd = NULL;
 814		goto check_pte;
 815	}
 816
 817	pgd = pgd_offset(mm, address);
 818	if (!pgd_present(*pgd))
 819		return false;
 820	pud = pud_offset(pgd, address);
 821	if (!pud_present(*pud))
 822		return false;
 823	pmd = pmd_offset(pud, address);
 824
 825	if (pmd_trans_huge(*pmd)) {
 826		ptl = pmd_lock(mm, pmd);
 827		if (!pmd_present(*pmd))
 828			goto unlock_pmd;
 829		if (unlikely(!pmd_trans_huge(*pmd))) {
 830			spin_unlock(ptl);
 831			goto map_pte;
 832		}
 833
 834		if (pmd_page(*pmd) != page)
 835			goto unlock_pmd;
 836
 837		pte = NULL;
 838		goto found;
 839unlock_pmd:
 840		spin_unlock(ptl);
 841		return false;
 842	} else {
 843		pmd_t pmde = *pmd;
 844
 845		barrier();
 846		if (!pmd_present(pmde) || pmd_trans_huge(pmde))
 847			return false;
 848	}
 849map_pte:
 850	pte = pte_offset_map(pmd, address);
 851	if (!pte_present(*pte)) {
 852		pte_unmap(pte);
 853		return false;
 854	}
 855
 856	ptl = pte_lockptr(mm, pmd);
 857check_pte:
 858	spin_lock(ptl);
 859
 860	if (!pte_present(*pte)) {
 861		pte_unmap_unlock(pte, ptl);
 862		return false;
 863	}
 864
 865	/* THP can be referenced by any subpage */
 866	if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) {
 867		pte_unmap_unlock(pte, ptl);
 868		return false;
 869	}
 870found:
 871	*ptep = pte;
 872	*pmdp = pmd;
 873	*ptlp = ptl;
 874	return true;
 875}
 876#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 877
 878struct page_referenced_arg {
 879	int mapcount;
 880	int referenced;
 881	unsigned long vm_flags;
 882	struct mem_cgroup *memcg;
 883};
 884/*
 885 * arg: page_referenced_arg will be passed
 886 */
 887static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
 888			unsigned long address, void *arg)
 889{
 890	struct mm_struct *mm = vma->vm_mm;
 891	struct page_referenced_arg *pra = arg;
 892	pmd_t *pmd;
 893	pte_t *pte;
 894	spinlock_t *ptl;
 895	int referenced = 0;
 
 
 
 
 896
 897	if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl))
 898		return SWAP_AGAIN;
 
 
 
 
 
 
 899
 900	if (vma->vm_flags & VM_LOCKED) {
 901		if (pte)
 902			pte_unmap(pte);
 
 
 
 
 
 
 903		spin_unlock(ptl);
 904		pra->vm_flags |= VM_LOCKED;
 905		return SWAP_FAIL; /* To break the loop */
 906	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 907
 908	if (pte) {
 909		if (ptep_clear_flush_young_notify(vma, address, pte)) {
 910			/*
 911			 * Don't treat a reference through a sequentially read
 912			 * mapping as such.  If the page has been used in
 913			 * another mapping, we will catch it; if this other
 914			 * mapping is already gone, the unmap path will have
 915			 * set PG_referenced or activated the page.
 916			 */
 917			if (likely(!(vma->vm_flags & VM_SEQ_READ)))
 918				referenced++;
 919		}
 920		pte_unmap(pte);
 921	} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 922		if (pmdp_clear_flush_young_notify(vma, address, pmd))
 923			referenced++;
 924	} else {
 925		/* unexpected pmd-mapped page? */
 926		WARN_ON_ONCE(1);
 927	}
 928	spin_unlock(ptl);
 929
 930	if (referenced)
 931		clear_page_idle(page);
 932	if (test_and_clear_page_young(page))
 933		referenced++;
 934
 935	if (referenced) {
 936		pra->referenced++;
 937		pra->vm_flags |= vma->vm_flags;
 938	}
 939
 940	pra->mapcount--;
 941	if (!pra->mapcount)
 942		return SWAP_SUCCESS; /* To break the loop */
 943
 944	return SWAP_AGAIN;
 945}
 946
 947static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
 948{
 949	struct page_referenced_arg *pra = arg;
 950	struct mem_cgroup *memcg = pra->memcg;
 951
 952	if (!mm_match_cgroup(vma->vm_mm, memcg))
 953		return true;
 954
 955	return false;
 956}
 957
 958/**
 959 * page_referenced - test if the page was referenced
 960 * @page: the page to test
 961 * @is_locked: caller holds lock on the page
 962 * @memcg: target memory cgroup
 963 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 964 *
 965 * Quick test_and_clear_referenced for all mappings to a page,
 966 * returns the number of ptes which referenced the page.
 967 */
 968int page_referenced(struct page *page,
 969		    int is_locked,
 970		    struct mem_cgroup *memcg,
 971		    unsigned long *vm_flags)
 972{
 973	int ret;
 974	int we_locked = 0;
 975	struct page_referenced_arg pra = {
 976		.mapcount = total_mapcount(page),
 977		.memcg = memcg,
 978	};
 979	struct rmap_walk_control rwc = {
 980		.rmap_one = page_referenced_one,
 981		.arg = (void *)&pra,
 982		.anon_lock = page_lock_anon_vma_read,
 983	};
 984
 985	*vm_flags = 0;
 986	if (!page_mapped(page))
 987		return 0;
 988
 989	if (!page_rmapping(page))
 990		return 0;
 991
 992	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 993		we_locked = trylock_page(page);
 994		if (!we_locked)
 995			return 1;
 996	}
 997
 998	/*
 999	 * If we are reclaiming on behalf of a cgroup, skip
1000	 * counting on behalf of references from different
1001	 * cgroups
1002	 */
1003	if (memcg) {
1004		rwc.invalid_vma = invalid_page_referenced_vma;
1005	}
1006
1007	ret = rmap_walk(page, &rwc);
1008	*vm_flags = pra.vm_flags;
1009
1010	if (we_locked)
1011		unlock_page(page);
1012
1013	return pra.referenced;
1014}
1015
1016static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
1017			    unsigned long address, void *arg)
1018{
1019	struct mm_struct *mm = vma->vm_mm;
1020	pte_t *pte;
1021	spinlock_t *ptl;
1022	int ret = 0;
1023	int *cleaned = arg;
1024
1025	pte = page_check_address(page, mm, address, &ptl, 1);
1026	if (!pte)
1027		goto out;
1028
1029	if (pte_dirty(*pte) || pte_write(*pte)) {
1030		pte_t entry;
1031
1032		flush_cache_page(vma, address, pte_pfn(*pte));
1033		entry = ptep_clear_flush(vma, address, pte);
1034		entry = pte_wrprotect(entry);
1035		entry = pte_mkclean(entry);
1036		set_pte_at(mm, address, pte, entry);
1037		ret = 1;
1038	}
1039
1040	pte_unmap_unlock(pte, ptl);
1041
1042	if (ret) {
1043		mmu_notifier_invalidate_page(mm, address);
1044		(*cleaned)++;
1045	}
1046out:
1047	return SWAP_AGAIN;
1048}
1049
1050static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1051{
1052	if (vma->vm_flags & VM_SHARED)
1053		return false;
1054
1055	return true;
1056}
1057
1058int page_mkclean(struct page *page)
1059{
1060	int cleaned = 0;
1061	struct address_space *mapping;
1062	struct rmap_walk_control rwc = {
1063		.arg = (void *)&cleaned,
1064		.rmap_one = page_mkclean_one,
1065		.invalid_vma = invalid_mkclean_vma,
1066	};
1067
1068	BUG_ON(!PageLocked(page));
1069
1070	if (!page_mapped(page))
1071		return 0;
1072
1073	mapping = page_mapping(page);
1074	if (!mapping)
1075		return 0;
1076
1077	rmap_walk(page, &rwc);
1078
1079	return cleaned;
1080}
1081EXPORT_SYMBOL_GPL(page_mkclean);
1082
1083/**
1084 * page_move_anon_rmap - move a page to our anon_vma
1085 * @page:	the page to move to our anon_vma
1086 * @vma:	the vma the page belongs to
1087 * @address:	the user virtual address mapped
1088 *
1089 * When a page belongs exclusively to one process after a COW event,
1090 * that page can be moved into the anon_vma that belongs to just that
1091 * process, so the rmap code will not search the parent or sibling
1092 * processes.
1093 */
1094void page_move_anon_rmap(struct page *page,
1095	struct vm_area_struct *vma, unsigned long address)
1096{
1097	struct anon_vma *anon_vma = vma->anon_vma;
1098
1099	VM_BUG_ON_PAGE(!PageLocked(page), page);
1100	VM_BUG_ON_VMA(!anon_vma, vma);
1101	VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
1102
1103	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1104	/*
1105	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1106	 * simultaneously, so a concurrent reader (eg page_referenced()'s
1107	 * PageAnon()) will not see one without the other.
1108	 */
1109	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1110}
1111
1112/**
1113 * __page_set_anon_rmap - set up new anonymous rmap
1114 * @page:	Page to add to rmap	
1115 * @vma:	VM area to add page to.
1116 * @address:	User virtual address of the mapping	
1117 * @exclusive:	the page is exclusively owned by the current process
1118 */
1119static void __page_set_anon_rmap(struct page *page,
1120	struct vm_area_struct *vma, unsigned long address, int exclusive)
1121{
1122	struct anon_vma *anon_vma = vma->anon_vma;
1123
1124	BUG_ON(!anon_vma);
1125
1126	if (PageAnon(page))
1127		return;
1128
1129	/*
1130	 * If the page isn't exclusively mapped into this vma,
1131	 * we must use the _oldest_ possible anon_vma for the
1132	 * page mapping!
1133	 */
1134	if (!exclusive)
1135		anon_vma = anon_vma->root;
1136
1137	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1138	page->mapping = (struct address_space *) anon_vma;
1139	page->index = linear_page_index(vma, address);
1140}
1141
1142/**
1143 * __page_check_anon_rmap - sanity check anonymous rmap addition
1144 * @page:	the page to add the mapping to
1145 * @vma:	the vm area in which the mapping is added
1146 * @address:	the user virtual address mapped
1147 */
1148static void __page_check_anon_rmap(struct page *page,
1149	struct vm_area_struct *vma, unsigned long address)
1150{
1151#ifdef CONFIG_DEBUG_VM
1152	/*
1153	 * The page's anon-rmap details (mapping and index) are guaranteed to
1154	 * be set up correctly at this point.
1155	 *
1156	 * We have exclusion against page_add_anon_rmap because the caller
1157	 * always holds the page locked, except if called from page_dup_rmap,
1158	 * in which case the page is already known to be setup.
1159	 *
1160	 * We have exclusion against page_add_new_anon_rmap because those pages
1161	 * are initially only visible via the pagetables, and the pte is locked
1162	 * over the call to page_add_new_anon_rmap.
1163	 */
1164	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1165	BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1166#endif
1167}
1168
1169/**
1170 * page_add_anon_rmap - add pte mapping to an anonymous page
1171 * @page:	the page to add the mapping to
1172 * @vma:	the vm area in which the mapping is added
1173 * @address:	the user virtual address mapped
1174 * @compound:	charge the page as compound or small page
1175 *
1176 * The caller needs to hold the pte lock, and the page must be locked in
1177 * the anon_vma case: to serialize mapping,index checking after setting,
1178 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1179 * (but PageKsm is never downgraded to PageAnon).
1180 */
1181void page_add_anon_rmap(struct page *page,
1182	struct vm_area_struct *vma, unsigned long address, bool compound)
1183{
1184	do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1185}
1186
1187/*
1188 * Special version of the above for do_swap_page, which often runs
1189 * into pages that are exclusively owned by the current process.
1190 * Everybody else should continue to use page_add_anon_rmap above.
1191 */
1192void do_page_add_anon_rmap(struct page *page,
1193	struct vm_area_struct *vma, unsigned long address, int flags)
1194{
1195	bool compound = flags & RMAP_COMPOUND;
1196	bool first;
1197
1198	if (compound) {
1199		atomic_t *mapcount;
1200		VM_BUG_ON_PAGE(!PageLocked(page), page);
1201		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1202		mapcount = compound_mapcount_ptr(page);
1203		first = atomic_inc_and_test(mapcount);
1204	} else {
1205		first = atomic_inc_and_test(&page->_mapcount);
1206	}
1207
1208	if (first) {
1209		int nr = compound ? hpage_nr_pages(page) : 1;
1210		/*
1211		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1212		 * these counters are not modified in interrupt context, and
1213		 * pte lock(a spinlock) is held, which implies preemption
1214		 * disabled.
1215		 */
1216		if (compound) {
1217			__inc_zone_page_state(page,
1218					      NR_ANON_TRANSPARENT_HUGEPAGES);
1219		}
1220		__mod_zone_page_state(page_zone(page), NR_ANON_PAGES, nr);
1221	}
1222	if (unlikely(PageKsm(page)))
1223		return;
1224
1225	VM_BUG_ON_PAGE(!PageLocked(page), page);
1226
1227	/* address might be in next vma when migration races vma_adjust */
1228	if (first)
1229		__page_set_anon_rmap(page, vma, address,
1230				flags & RMAP_EXCLUSIVE);
1231	else
1232		__page_check_anon_rmap(page, vma, address);
1233}
1234
1235/**
1236 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1237 * @page:	the page to add the mapping to
1238 * @vma:	the vm area in which the mapping is added
1239 * @address:	the user virtual address mapped
1240 * @compound:	charge the page as compound or small page
1241 *
1242 * Same as page_add_anon_rmap but must only be called on *new* pages.
1243 * This means the inc-and-test can be bypassed.
1244 * Page does not have to be locked.
1245 */
1246void page_add_new_anon_rmap(struct page *page,
1247	struct vm_area_struct *vma, unsigned long address, bool compound)
1248{
1249	int nr = compound ? hpage_nr_pages(page) : 1;
1250
1251	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1252	SetPageSwapBacked(page);
1253	if (compound) {
1254		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1255		/* increment count (starts at -1) */
1256		atomic_set(compound_mapcount_ptr(page), 0);
1257		__inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1258	} else {
1259		/* Anon THP always mapped first with PMD */
1260		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1261		/* increment count (starts at -1) */
1262		atomic_set(&page->_mapcount, 0);
1263	}
1264	__mod_zone_page_state(page_zone(page), NR_ANON_PAGES, nr);
1265	__page_set_anon_rmap(page, vma, address, 1);
 
 
 
 
 
1266}
1267
1268/**
1269 * page_add_file_rmap - add pte mapping to a file page
1270 * @page: the page to add the mapping to
1271 *
1272 * The caller needs to hold the pte lock.
1273 */
1274void page_add_file_rmap(struct page *page)
1275{
1276	lock_page_memcg(page);
 
 
 
1277	if (atomic_inc_and_test(&page->_mapcount)) {
1278		__inc_zone_page_state(page, NR_FILE_MAPPED);
1279		mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1280	}
1281	unlock_page_memcg(page);
1282}
1283
1284static void page_remove_file_rmap(struct page *page)
1285{
1286	lock_page_memcg(page);
1287
1288	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1289	if (unlikely(PageHuge(page))) {
1290		/* hugetlb pages are always mapped with pmds */
1291		atomic_dec(compound_mapcount_ptr(page));
1292		goto out;
1293	}
1294
1295	/* page still mapped by someone else? */
1296	if (!atomic_add_negative(-1, &page->_mapcount))
1297		goto out;
1298
1299	/*
1300	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1301	 * these counters are not modified in interrupt context, and
1302	 * pte lock(a spinlock) is held, which implies preemption disabled.
1303	 */
1304	__dec_zone_page_state(page, NR_FILE_MAPPED);
1305	mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED);
1306
1307	if (unlikely(PageMlocked(page)))
1308		clear_page_mlock(page);
1309out:
1310	unlock_page_memcg(page);
1311}
1312
1313static void page_remove_anon_compound_rmap(struct page *page)
1314{
1315	int i, nr;
1316
1317	if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1318		return;
1319
1320	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1321	if (unlikely(PageHuge(page)))
1322		return;
1323
1324	if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1325		return;
1326
1327	__dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1328
1329	if (TestClearPageDoubleMap(page)) {
1330		/*
1331		 * Subpages can be mapped with PTEs too. Check how many of
1332		 * themi are still mapped.
1333		 */
1334		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1335			if (atomic_add_negative(-1, &page[i]._mapcount))
1336				nr++;
1337		}
1338	} else {
1339		nr = HPAGE_PMD_NR;
1340	}
1341
1342	if (unlikely(PageMlocked(page)))
1343		clear_page_mlock(page);
1344
1345	if (nr) {
1346		__mod_zone_page_state(page_zone(page), NR_ANON_PAGES, -nr);
1347		deferred_split_huge_page(page);
1348	}
1349}
1350
1351/**
1352 * page_remove_rmap - take down pte mapping from a page
1353 * @page:	page to remove mapping from
1354 * @compound:	uncharge the page as compound or small page
1355 *
1356 * The caller needs to hold the pte lock.
1357 */
1358void page_remove_rmap(struct page *page, bool compound)
1359{
1360	if (!PageAnon(page)) {
1361		VM_BUG_ON_PAGE(compound && !PageHuge(page), page);
1362		page_remove_file_rmap(page);
1363		return;
1364	}
1365
1366	if (compound)
1367		return page_remove_anon_compound_rmap(page);
 
 
 
 
 
1368
1369	/* page still mapped by someone else? */
1370	if (!atomic_add_negative(-1, &page->_mapcount))
1371		return;
1372
1373	/*
1374	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1375	 * these counters are not modified in interrupt context, and
1376	 * pte lock(a spinlock) is held, which implies preemption disabled.
1377	 */
1378	__dec_zone_page_state(page, NR_ANON_PAGES);
1379
 
 
 
 
 
 
 
 
 
 
 
 
1380	if (unlikely(PageMlocked(page)))
1381		clear_page_mlock(page);
1382
1383	if (PageTransCompound(page))
1384		deferred_split_huge_page(compound_head(page));
1385
1386	/*
1387	 * It would be tidy to reset the PageAnon mapping here,
1388	 * but that might overwrite a racing page_add_anon_rmap
1389	 * which increments mapcount after us but sets mapping
1390	 * before us: so leave the reset to free_hot_cold_page,
1391	 * and remember that it's only reliable while mapped.
1392	 * Leaving it set also helps swapoff to reinstate ptes
1393	 * faster for those pages still in swapcache.
1394	 */
 
 
 
 
1395}
1396
1397struct rmap_private {
1398	enum ttu_flags flags;
1399	int lazyfreed;
1400};
1401
1402/*
1403 * @arg: enum ttu_flags will be passed to this argument
1404 */
1405static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1406		     unsigned long address, void *arg)
1407{
1408	struct mm_struct *mm = vma->vm_mm;
1409	pte_t *pte;
1410	pte_t pteval;
1411	spinlock_t *ptl;
1412	int ret = SWAP_AGAIN;
1413	struct rmap_private *rp = arg;
1414	enum ttu_flags flags = rp->flags;
1415
1416	/* munlock has nothing to gain from examining un-locked vmas */
1417	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1418		goto out;
1419
1420	if (flags & TTU_SPLIT_HUGE_PMD) {
1421		split_huge_pmd_address(vma, address,
1422				flags & TTU_MIGRATION, page);
1423		/* check if we have anything to do after split */
1424		if (page_mapcount(page) == 0)
1425			goto out;
1426	}
1427
1428	pte = page_check_address(page, mm, address, &ptl, 0);
1429	if (!pte)
1430		goto out;
1431
1432	/*
1433	 * If the page is mlock()d, we cannot swap it out.
1434	 * If it's recently referenced (perhaps page_referenced
1435	 * skipped over this mm) then we should reactivate it.
1436	 */
1437	if (!(flags & TTU_IGNORE_MLOCK)) {
1438		if (vma->vm_flags & VM_LOCKED) {
1439			/* Holding pte lock, we do *not* need mmap_sem here */
1440			mlock_vma_page(page);
1441			ret = SWAP_MLOCK;
1442			goto out_unmap;
1443		}
1444		if (flags & TTU_MUNLOCK)
1445			goto out_unmap;
1446	}
1447	if (!(flags & TTU_IGNORE_ACCESS)) {
1448		if (ptep_clear_flush_young_notify(vma, address, pte)) {
1449			ret = SWAP_FAIL;
1450			goto out_unmap;
1451		}
1452  	}
1453
1454	/* Nuke the page table entry. */
1455	flush_cache_page(vma, address, page_to_pfn(page));
1456	if (should_defer_flush(mm, flags)) {
1457		/*
1458		 * We clear the PTE but do not flush so potentially a remote
1459		 * CPU could still be writing to the page. If the entry was
1460		 * previously clean then the architecture must guarantee that
1461		 * a clear->dirty transition on a cached TLB entry is written
1462		 * through and traps if the PTE is unmapped.
1463		 */
1464		pteval = ptep_get_and_clear(mm, address, pte);
1465
1466		set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1467	} else {
1468		pteval = ptep_clear_flush(vma, address, pte);
1469	}
1470
1471	/* Move the dirty bit to the physical page now the pte is gone. */
1472	if (pte_dirty(pteval))
1473		set_page_dirty(page);
1474
1475	/* Update high watermark before we lower rss */
1476	update_hiwater_rss(mm);
1477
1478	if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1479		if (PageHuge(page)) {
1480			hugetlb_count_sub(1 << compound_order(page), mm);
1481		} else {
1482			dec_mm_counter(mm, mm_counter(page));
 
1483		}
1484		set_pte_at(mm, address, pte,
1485			   swp_entry_to_pte(make_hwpoison_entry(page)));
1486	} else if (pte_unused(pteval)) {
1487		/*
1488		 * The guest indicated that the page content is of no
1489		 * interest anymore. Simply discard the pte, vmscan
1490		 * will take care of the rest.
1491		 */
1492		dec_mm_counter(mm, mm_counter(page));
1493	} else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1494		swp_entry_t entry;
1495		pte_t swp_pte;
1496		/*
1497		 * Store the pfn of the page in a special migration
1498		 * pte. do_swap_page() will wait until the migration
1499		 * pte is removed and then restart fault handling.
1500		 */
1501		entry = make_migration_entry(page, pte_write(pteval));
1502		swp_pte = swp_entry_to_pte(entry);
1503		if (pte_soft_dirty(pteval))
1504			swp_pte = pte_swp_mksoft_dirty(swp_pte);
1505		set_pte_at(mm, address, pte, swp_pte);
1506	} else if (PageAnon(page)) {
1507		swp_entry_t entry = { .val = page_private(page) };
1508		pte_t swp_pte;
1509		/*
1510		 * Store the swap location in the pte.
1511		 * See handle_pte_fault() ...
1512		 */
1513		VM_BUG_ON_PAGE(!PageSwapCache(page), page);
1514
1515		if (!PageDirty(page) && (flags & TTU_LZFREE)) {
1516			/* It's a freeable page by MADV_FREE */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1517			dec_mm_counter(mm, MM_ANONPAGES);
1518			rp->lazyfreed++;
1519			goto discard;
1520		}
1521
1522		if (swap_duplicate(entry) < 0) {
1523			set_pte_at(mm, address, pte, pteval);
1524			ret = SWAP_FAIL;
1525			goto out_unmap;
1526		}
1527		if (list_empty(&mm->mmlist)) {
1528			spin_lock(&mmlist_lock);
1529			if (list_empty(&mm->mmlist))
1530				list_add(&mm->mmlist, &init_mm.mmlist);
1531			spin_unlock(&mmlist_lock);
1532		}
1533		dec_mm_counter(mm, MM_ANONPAGES);
1534		inc_mm_counter(mm, MM_SWAPENTS);
1535		swp_pte = swp_entry_to_pte(entry);
1536		if (pte_soft_dirty(pteval))
1537			swp_pte = pte_swp_mksoft_dirty(swp_pte);
1538		set_pte_at(mm, address, pte, swp_pte);
 
 
 
 
 
 
 
1539	} else
1540		dec_mm_counter(mm, mm_counter_file(page));
1541
1542discard:
1543	page_remove_rmap(page, PageHuge(page));
1544	put_page(page);
1545
1546out_unmap:
1547	pte_unmap_unlock(pte, ptl);
1548	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
1549		mmu_notifier_invalidate_page(mm, address);
1550out:
1551	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1552}
1553
1554bool is_vma_temporary_stack(struct vm_area_struct *vma)
1555{
1556	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1557
1558	if (!maybe_stack)
1559		return false;
1560
1561	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1562						VM_STACK_INCOMPLETE_SETUP)
1563		return true;
1564
1565	return false;
1566}
1567
1568static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1569{
1570	return is_vma_temporary_stack(vma);
1571}
1572
1573static int page_mapcount_is_zero(struct page *page)
1574{
1575	return !page_mapcount(page);
1576}
1577
1578/**
1579 * try_to_unmap - try to remove all page table mappings to a page
1580 * @page: the page to get unmapped
1581 * @flags: action and flags
1582 *
1583 * Tries to remove all the page table entries which are mapping this
1584 * page, used in the pageout path.  Caller must hold the page lock.
1585 * Return values are:
1586 *
1587 * SWAP_SUCCESS	- we succeeded in removing all mappings
1588 * SWAP_AGAIN	- we missed a mapping, try again later
1589 * SWAP_FAIL	- the page is unswappable
1590 * SWAP_MLOCK	- page is mlocked.
1591 */
1592int try_to_unmap(struct page *page, enum ttu_flags flags)
1593{
1594	int ret;
1595	struct rmap_private rp = {
1596		.flags = flags,
1597		.lazyfreed = 0,
1598	};
1599
1600	struct rmap_walk_control rwc = {
1601		.rmap_one = try_to_unmap_one,
1602		.arg = &rp,
1603		.done = page_mapcount_is_zero,
 
1604		.anon_lock = page_lock_anon_vma_read,
1605	};
1606
 
 
1607	/*
1608	 * During exec, a temporary VMA is setup and later moved.
1609	 * The VMA is moved under the anon_vma lock but not the
1610	 * page tables leading to a race where migration cannot
1611	 * find the migration ptes. Rather than increasing the
1612	 * locking requirements of exec(), migration skips
1613	 * temporary VMAs until after exec() completes.
1614	 */
1615	if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1616		rwc.invalid_vma = invalid_migration_vma;
1617
1618	if (flags & TTU_RMAP_LOCKED)
1619		ret = rmap_walk_locked(page, &rwc);
1620	else
1621		ret = rmap_walk(page, &rwc);
1622
1623	if (ret != SWAP_MLOCK && !page_mapcount(page)) {
1624		ret = SWAP_SUCCESS;
1625		if (rp.lazyfreed && !PageDirty(page))
1626			ret = SWAP_LZFREE;
1627	}
1628	return ret;
1629}
1630
1631static int page_not_mapped(struct page *page)
1632{
1633	return !page_mapped(page);
1634};
1635
1636/**
1637 * try_to_munlock - try to munlock a page
1638 * @page: the page to be munlocked
1639 *
1640 * Called from munlock code.  Checks all of the VMAs mapping the page
1641 * to make sure nobody else has this page mlocked. The page will be
1642 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1643 *
1644 * Return values are:
1645 *
1646 * SWAP_AGAIN	- no vma is holding page mlocked, or,
1647 * SWAP_AGAIN	- page mapped in mlocked vma -- couldn't acquire mmap sem
1648 * SWAP_FAIL	- page cannot be located at present
1649 * SWAP_MLOCK	- page is now mlocked.
1650 */
1651int try_to_munlock(struct page *page)
1652{
1653	int ret;
1654	struct rmap_private rp = {
1655		.flags = TTU_MUNLOCK,
1656		.lazyfreed = 0,
1657	};
1658
1659	struct rmap_walk_control rwc = {
1660		.rmap_one = try_to_unmap_one,
1661		.arg = &rp,
1662		.done = page_not_mapped,
 
 
 
 
 
 
1663		.anon_lock = page_lock_anon_vma_read,
1664
1665	};
1666
1667	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1668
1669	ret = rmap_walk(page, &rwc);
1670	return ret;
1671}
1672
1673void __put_anon_vma(struct anon_vma *anon_vma)
1674{
1675	struct anon_vma *root = anon_vma->root;
1676
1677	anon_vma_free(anon_vma);
1678	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1679		anon_vma_free(root);
1680}
1681
1682static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1683					struct rmap_walk_control *rwc)
1684{
1685	struct anon_vma *anon_vma;
1686
1687	if (rwc->anon_lock)
1688		return rwc->anon_lock(page);
1689
1690	/*
1691	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1692	 * because that depends on page_mapped(); but not all its usages
1693	 * are holding mmap_sem. Users without mmap_sem are required to
1694	 * take a reference count to prevent the anon_vma disappearing
1695	 */
1696	anon_vma = page_anon_vma(page);
1697	if (!anon_vma)
1698		return NULL;
1699
1700	anon_vma_lock_read(anon_vma);
1701	return anon_vma;
1702}
1703
1704/*
1705 * rmap_walk_anon - do something to anonymous page using the object-based
1706 * rmap method
1707 * @page: the page to be handled
1708 * @rwc: control variable according to each walk type
1709 *
1710 * Find all the mappings of a page using the mapping pointer and the vma chains
1711 * contained in the anon_vma struct it points to.
1712 *
1713 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1714 * where the page was found will be held for write.  So, we won't recheck
1715 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1716 * LOCKED.
1717 */
1718static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1719		bool locked)
1720{
1721	struct anon_vma *anon_vma;
1722	pgoff_t pgoff;
1723	struct anon_vma_chain *avc;
1724	int ret = SWAP_AGAIN;
1725
1726	if (locked) {
1727		anon_vma = page_anon_vma(page);
1728		/* anon_vma disappear under us? */
1729		VM_BUG_ON_PAGE(!anon_vma, page);
1730	} else {
1731		anon_vma = rmap_walk_anon_lock(page, rwc);
1732	}
1733	if (!anon_vma)
1734		return ret;
1735
1736	pgoff = page_to_pgoff(page);
1737	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1738		struct vm_area_struct *vma = avc->vma;
1739		unsigned long address = vma_address(page, vma);
1740
1741		cond_resched();
1742
1743		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1744			continue;
1745
1746		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1747		if (ret != SWAP_AGAIN)
1748			break;
1749		if (rwc->done && rwc->done(page))
1750			break;
1751	}
1752
1753	if (!locked)
1754		anon_vma_unlock_read(anon_vma);
1755	return ret;
1756}
1757
1758/*
1759 * rmap_walk_file - do something to file page using the object-based rmap method
1760 * @page: the page to be handled
1761 * @rwc: control variable according to each walk type
1762 *
1763 * Find all the mappings of a page using the mapping pointer and the vma chains
1764 * contained in the address_space struct it points to.
1765 *
1766 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1767 * where the page was found will be held for write.  So, we won't recheck
1768 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1769 * LOCKED.
1770 */
1771static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1772		bool locked)
1773{
1774	struct address_space *mapping = page_mapping(page);
1775	pgoff_t pgoff;
1776	struct vm_area_struct *vma;
1777	int ret = SWAP_AGAIN;
1778
1779	/*
1780	 * The page lock not only makes sure that page->mapping cannot
1781	 * suddenly be NULLified by truncation, it makes sure that the
1782	 * structure at mapping cannot be freed and reused yet,
1783	 * so we can safely take mapping->i_mmap_rwsem.
1784	 */
1785	VM_BUG_ON_PAGE(!PageLocked(page), page);
1786
1787	if (!mapping)
1788		return ret;
1789
1790	pgoff = page_to_pgoff(page);
1791	if (!locked)
1792		i_mmap_lock_read(mapping);
1793	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1794		unsigned long address = vma_address(page, vma);
1795
1796		cond_resched();
1797
1798		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1799			continue;
1800
1801		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1802		if (ret != SWAP_AGAIN)
1803			goto done;
1804		if (rwc->done && rwc->done(page))
1805			goto done;
1806	}
1807
 
 
 
 
 
 
 
 
1808done:
1809	if (!locked)
1810		i_mmap_unlock_read(mapping);
1811	return ret;
1812}
1813
1814int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1815{
1816	if (unlikely(PageKsm(page)))
1817		return rmap_walk_ksm(page, rwc);
1818	else if (PageAnon(page))
1819		return rmap_walk_anon(page, rwc, false);
1820	else
1821		return rmap_walk_file(page, rwc, false);
1822}
1823
1824/* Like rmap_walk, but caller holds relevant rmap lock */
1825int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1826{
1827	/* no ksm support for now */
1828	VM_BUG_ON_PAGE(PageKsm(page), page);
1829	if (PageAnon(page))
1830		return rmap_walk_anon(page, rwc, true);
1831	else
1832		return rmap_walk_file(page, rwc, true);
1833}
1834
1835#ifdef CONFIG_HUGETLB_PAGE
1836/*
1837 * The following three functions are for anonymous (private mapped) hugepages.
1838 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1839 * and no lru code, because we handle hugepages differently from common pages.
1840 */
1841static void __hugepage_set_anon_rmap(struct page *page,
1842	struct vm_area_struct *vma, unsigned long address, int exclusive)
1843{
1844	struct anon_vma *anon_vma = vma->anon_vma;
1845
1846	BUG_ON(!anon_vma);
1847
1848	if (PageAnon(page))
1849		return;
1850	if (!exclusive)
1851		anon_vma = anon_vma->root;
1852
1853	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1854	page->mapping = (struct address_space *) anon_vma;
1855	page->index = linear_page_index(vma, address);
1856}
1857
1858void hugepage_add_anon_rmap(struct page *page,
1859			    struct vm_area_struct *vma, unsigned long address)
1860{
1861	struct anon_vma *anon_vma = vma->anon_vma;
1862	int first;
1863
1864	BUG_ON(!PageLocked(page));
1865	BUG_ON(!anon_vma);
1866	/* address might be in next vma when migration races vma_adjust */
1867	first = atomic_inc_and_test(compound_mapcount_ptr(page));
1868	if (first)
1869		__hugepage_set_anon_rmap(page, vma, address, 0);
1870}
1871
1872void hugepage_add_new_anon_rmap(struct page *page,
1873			struct vm_area_struct *vma, unsigned long address)
1874{
1875	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1876	atomic_set(compound_mapcount_ptr(page), 0);
1877	__hugepage_set_anon_rmap(page, vma, address, 1);
1878}
1879#endif /* CONFIG_HUGETLB_PAGE */