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_lock
  25 *     page->flags PG_locked (lock_page)   * (see huegtlbfs below)
  26 *       hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
  27 *         mapping->i_mmap_rwsem
  28 *           hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
  29 *           anon_vma->rwsem
  30 *             mm->page_table_lock or pte_lock
  31 *               pgdat->lru_lock (in mark_page_accessed, isolate_lru_page)
  32 *               swap_lock (in swap_duplicate, swap_info_get)
  33 *                 mmlist_lock (in mmput, drain_mmlist and others)
  34 *                 mapping->private_lock (in __set_page_dirty_buffers)
  35 *                   mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
  36 *                     i_pages lock (widely used)
  37 *                 inode->i_lock (in set_page_dirty's __mark_inode_dirty)
  38 *                 bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
  39 *                   sb_lock (within inode_lock in fs/fs-writeback.c)
  40 *                   i_pages lock (widely used, in set_page_dirty,
  41 *                             in arch-dependent flush_dcache_mmap_lock,
  42 *                             within bdi.wb->list_lock in __sync_single_inode)
 
  43 *
  44 * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
  45 *   ->tasklist_lock
  46 *     pte map lock
  47 *
  48 * * hugetlbfs PageHuge() pages take locks in this order:
  49 *         mapping->i_mmap_rwsem
  50 *           hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
  51 *             page->flags PG_locked (lock_page)
 
  52 */
  53
  54#include <linux/mm.h>
  55#include <linux/sched/mm.h>
  56#include <linux/sched/task.h>
  57#include <linux/pagemap.h>
  58#include <linux/swap.h>
  59#include <linux/swapops.h>
  60#include <linux/slab.h>
  61#include <linux/init.h>
  62#include <linux/ksm.h>
  63#include <linux/rmap.h>
  64#include <linux/rcupdate.h>
  65#include <linux/export.h>
  66#include <linux/memcontrol.h>
  67#include <linux/mmu_notifier.h>
  68#include <linux/migrate.h>
  69#include <linux/hugetlb.h>
  70#include <linux/huge_mm.h>
  71#include <linux/backing-dev.h>
  72#include <linux/page_idle.h>
  73#include <linux/memremap.h>
  74#include <linux/userfaultfd_k.h>
 
  75
  76#include <asm/tlbflush.h>
  77
 
  78#include <trace/events/tlb.h>
 
  79
  80#include "internal.h"
  81
  82static struct kmem_cache *anon_vma_cachep;
  83static struct kmem_cache *anon_vma_chain_cachep;
  84
  85static inline struct anon_vma *anon_vma_alloc(void)
  86{
  87	struct anon_vma *anon_vma;
  88
  89	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  90	if (anon_vma) {
  91		atomic_set(&anon_vma->refcount, 1);
  92		anon_vma->degree = 1;	/* Reference for first vma */
 
  93		anon_vma->parent = anon_vma;
  94		/*
  95		 * Initialise the anon_vma root to point to itself. If called
  96		 * from fork, the root will be reset to the parents anon_vma.
  97		 */
  98		anon_vma->root = anon_vma;
  99	}
 100
 101	return anon_vma;
 102}
 103
 104static inline void anon_vma_free(struct anon_vma *anon_vma)
 105{
 106	VM_BUG_ON(atomic_read(&anon_vma->refcount));
 107
 108	/*
 109	 * Synchronize against page_lock_anon_vma_read() such that
 110	 * we can safely hold the lock without the anon_vma getting
 111	 * freed.
 112	 *
 113	 * Relies on the full mb implied by the atomic_dec_and_test() from
 114	 * put_anon_vma() against the acquire barrier implied by
 115	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
 116	 *
 117	 * page_lock_anon_vma_read()	VS	put_anon_vma()
 118	 *   down_read_trylock()		  atomic_dec_and_test()
 119	 *   LOCK				  MB
 120	 *   atomic_read()			  rwsem_is_locked()
 121	 *
 122	 * LOCK should suffice since the actual taking of the lock must
 123	 * happen _before_ what follows.
 124	 */
 125	might_sleep();
 126	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
 127		anon_vma_lock_write(anon_vma);
 128		anon_vma_unlock_write(anon_vma);
 129	}
 130
 131	kmem_cache_free(anon_vma_cachep, anon_vma);
 132}
 133
 134static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 135{
 136	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 137}
 138
 139static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 140{
 141	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 142}
 143
 144static void anon_vma_chain_link(struct vm_area_struct *vma,
 145				struct anon_vma_chain *avc,
 146				struct anon_vma *anon_vma)
 147{
 148	avc->vma = vma;
 149	avc->anon_vma = anon_vma;
 150	list_add(&avc->same_vma, &vma->anon_vma_chain);
 151	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 152}
 153
 154/**
 155 * __anon_vma_prepare - attach an anon_vma to a memory region
 156 * @vma: the memory region in question
 157 *
 158 * This makes sure the memory mapping described by 'vma' has
 159 * an 'anon_vma' attached to it, so that we can associate the
 160 * anonymous pages mapped into it with that anon_vma.
 161 *
 162 * The common case will be that we already have one, which
 163 * is handled inline by anon_vma_prepare(). But if
 164 * not we either need to find an adjacent mapping that we
 165 * can re-use the anon_vma from (very common when the only
 166 * reason for splitting a vma has been mprotect()), or we
 167 * allocate a new one.
 168 *
 169 * Anon-vma allocations are very subtle, because we may have
 170 * optimistically looked up an anon_vma in page_lock_anon_vma_read()
 171 * and that may actually touch the spinlock even in the newly
 172 * allocated vma (it depends on RCU to make sure that the
 173 * anon_vma isn't actually destroyed).
 174 *
 175 * As a result, we need to do proper anon_vma locking even
 176 * for the new allocation. At the same time, we do not want
 177 * to do any locking for the common case of already having
 178 * an anon_vma.
 179 *
 180 * This must be called with the mmap_lock held for reading.
 181 */
 182int __anon_vma_prepare(struct vm_area_struct *vma)
 183{
 184	struct mm_struct *mm = vma->vm_mm;
 185	struct anon_vma *anon_vma, *allocated;
 186	struct anon_vma_chain *avc;
 187
 188	might_sleep();
 189
 190	avc = anon_vma_chain_alloc(GFP_KERNEL);
 191	if (!avc)
 192		goto out_enomem;
 193
 194	anon_vma = find_mergeable_anon_vma(vma);
 195	allocated = NULL;
 196	if (!anon_vma) {
 197		anon_vma = anon_vma_alloc();
 198		if (unlikely(!anon_vma))
 199			goto out_enomem_free_avc;
 
 200		allocated = anon_vma;
 201	}
 202
 203	anon_vma_lock_write(anon_vma);
 204	/* page_table_lock to protect against threads */
 205	spin_lock(&mm->page_table_lock);
 206	if (likely(!vma->anon_vma)) {
 207		vma->anon_vma = anon_vma;
 208		anon_vma_chain_link(vma, avc, anon_vma);
 209		/* vma reference or self-parent link for new root */
 210		anon_vma->degree++;
 211		allocated = NULL;
 212		avc = NULL;
 213	}
 214	spin_unlock(&mm->page_table_lock);
 215	anon_vma_unlock_write(anon_vma);
 216
 217	if (unlikely(allocated))
 218		put_anon_vma(allocated);
 219	if (unlikely(avc))
 220		anon_vma_chain_free(avc);
 221
 222	return 0;
 223
 224 out_enomem_free_avc:
 225	anon_vma_chain_free(avc);
 226 out_enomem:
 227	return -ENOMEM;
 228}
 229
 230/*
 231 * This is a useful helper function for locking the anon_vma root as
 232 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 233 * have the same vma.
 234 *
 235 * Such anon_vma's should have the same root, so you'd expect to see
 236 * just a single mutex_lock for the whole traversal.
 237 */
 238static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 239{
 240	struct anon_vma *new_root = anon_vma->root;
 241	if (new_root != root) {
 242		if (WARN_ON_ONCE(root))
 243			up_write(&root->rwsem);
 244		root = new_root;
 245		down_write(&root->rwsem);
 246	}
 247	return root;
 248}
 249
 250static inline void unlock_anon_vma_root(struct anon_vma *root)
 251{
 252	if (root)
 253		up_write(&root->rwsem);
 254}
 255
 256/*
 257 * Attach the anon_vmas from src to dst.
 258 * Returns 0 on success, -ENOMEM on failure.
 259 *
 260 * anon_vma_clone() is called by __vma_split(), __split_vma(), copy_vma() and
 261 * anon_vma_fork(). The first three want an exact copy of src, while the last
 262 * one, anon_vma_fork(), may try to reuse an existing anon_vma to prevent
 263 * endless growth of anon_vma. Since dst->anon_vma is set to NULL before call,
 264 * we can identify this case by checking (!dst->anon_vma && src->anon_vma).
 
 265 *
 266 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
 267 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
 268 * This prevents degradation of anon_vma hierarchy to endless linear chain in
 269 * case of constantly forking task. On the other hand, an anon_vma with more
 270 * than one child isn't reused even if there was no alive vma, thus rmap
 271 * walker has a good chance of avoiding scanning the whole hierarchy when it
 272 * searches where page is mapped.
 273 */
 274int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 275{
 276	struct anon_vma_chain *avc, *pavc;
 277	struct anon_vma *root = NULL;
 278
 279	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 280		struct anon_vma *anon_vma;
 281
 282		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 283		if (unlikely(!avc)) {
 284			unlock_anon_vma_root(root);
 285			root = NULL;
 286			avc = anon_vma_chain_alloc(GFP_KERNEL);
 287			if (!avc)
 288				goto enomem_failure;
 289		}
 290		anon_vma = pavc->anon_vma;
 291		root = lock_anon_vma_root(root, anon_vma);
 292		anon_vma_chain_link(dst, avc, anon_vma);
 293
 294		/*
 295		 * Reuse existing anon_vma if its degree lower than two,
 296		 * that means it has no vma and only one anon_vma child.
 297		 *
 298		 * Do not chose parent anon_vma, otherwise first child
 299		 * will always reuse it. Root anon_vma is never reused:
 300		 * it has self-parent reference and at least one child.
 301		 */
 302		if (!dst->anon_vma && src->anon_vma &&
 303		    anon_vma != src->anon_vma && anon_vma->degree < 2)
 
 304			dst->anon_vma = anon_vma;
 305	}
 306	if (dst->anon_vma)
 307		dst->anon_vma->degree++;
 308	unlock_anon_vma_root(root);
 309	return 0;
 310
 311 enomem_failure:
 312	/*
 313	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
 314	 * decremented in unlink_anon_vmas().
 315	 * We can safely do this because callers of anon_vma_clone() don't care
 316	 * about dst->anon_vma if anon_vma_clone() failed.
 317	 */
 318	dst->anon_vma = NULL;
 319	unlink_anon_vmas(dst);
 320	return -ENOMEM;
 321}
 322
 323/*
 324 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 325 * the corresponding VMA in the parent process is attached to.
 326 * Returns 0 on success, non-zero on failure.
 327 */
 328int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 329{
 330	struct anon_vma_chain *avc;
 331	struct anon_vma *anon_vma;
 332	int error;
 333
 334	/* Don't bother if the parent process has no anon_vma here. */
 335	if (!pvma->anon_vma)
 336		return 0;
 337
 338	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
 339	vma->anon_vma = NULL;
 340
 341	/*
 342	 * First, attach the new VMA to the parent VMA's anon_vmas,
 343	 * so rmap can find non-COWed pages in child processes.
 344	 */
 345	error = anon_vma_clone(vma, pvma);
 346	if (error)
 347		return error;
 348
 349	/* An existing anon_vma has been reused, all done then. */
 350	if (vma->anon_vma)
 351		return 0;
 352
 353	/* Then add our own anon_vma. */
 354	anon_vma = anon_vma_alloc();
 355	if (!anon_vma)
 356		goto out_error;
 
 357	avc = anon_vma_chain_alloc(GFP_KERNEL);
 358	if (!avc)
 359		goto out_error_free_anon_vma;
 360
 361	/*
 362	 * The root anon_vma's spinlock is the lock actually used when we
 363	 * lock any of the anon_vmas in this anon_vma tree.
 364	 */
 365	anon_vma->root = pvma->anon_vma->root;
 366	anon_vma->parent = pvma->anon_vma;
 367	/*
 368	 * With refcounts, an anon_vma can stay around longer than the
 369	 * process it belongs to. The root anon_vma needs to be pinned until
 370	 * this anon_vma is freed, because the lock lives in the root.
 371	 */
 372	get_anon_vma(anon_vma->root);
 373	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 374	vma->anon_vma = anon_vma;
 375	anon_vma_lock_write(anon_vma);
 376	anon_vma_chain_link(vma, avc, anon_vma);
 377	anon_vma->parent->degree++;
 378	anon_vma_unlock_write(anon_vma);
 379
 380	return 0;
 381
 382 out_error_free_anon_vma:
 383	put_anon_vma(anon_vma);
 384 out_error:
 385	unlink_anon_vmas(vma);
 386	return -ENOMEM;
 387}
 388
 389void unlink_anon_vmas(struct vm_area_struct *vma)
 390{
 391	struct anon_vma_chain *avc, *next;
 392	struct anon_vma *root = NULL;
 393
 394	/*
 395	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 396	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 397	 */
 398	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 399		struct anon_vma *anon_vma = avc->anon_vma;
 400
 401		root = lock_anon_vma_root(root, anon_vma);
 402		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 403
 404		/*
 405		 * Leave empty anon_vmas on the list - we'll need
 406		 * to free them outside the lock.
 407		 */
 408		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
 409			anon_vma->parent->degree--;
 410			continue;
 411		}
 412
 413		list_del(&avc->same_vma);
 414		anon_vma_chain_free(avc);
 415	}
 416	if (vma->anon_vma)
 417		vma->anon_vma->degree--;
 
 
 
 
 
 
 
 418	unlock_anon_vma_root(root);
 419
 420	/*
 421	 * Iterate the list once more, it now only contains empty and unlinked
 422	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 423	 * needing to write-acquire the anon_vma->root->rwsem.
 424	 */
 425	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 426		struct anon_vma *anon_vma = avc->anon_vma;
 427
 428		VM_WARN_ON(anon_vma->degree);
 
 429		put_anon_vma(anon_vma);
 430
 431		list_del(&avc->same_vma);
 432		anon_vma_chain_free(avc);
 433	}
 434}
 435
 436static void anon_vma_ctor(void *data)
 437{
 438	struct anon_vma *anon_vma = data;
 439
 440	init_rwsem(&anon_vma->rwsem);
 441	atomic_set(&anon_vma->refcount, 0);
 442	anon_vma->rb_root = RB_ROOT_CACHED;
 443}
 444
 445void __init anon_vma_init(void)
 446{
 447	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 448			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
 449			anon_vma_ctor);
 450	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
 451			SLAB_PANIC|SLAB_ACCOUNT);
 452}
 453
 454/*
 455 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 456 *
 457 * Since there is no serialization what so ever against page_remove_rmap()
 458 * the best this function can do is return a locked anon_vma that might
 459 * have been relevant to this page.
 460 *
 461 * The page might have been remapped to a different anon_vma or the anon_vma
 462 * returned may already be freed (and even reused).
 463 *
 464 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 465 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 466 * ensure that any anon_vma obtained from the page will still be valid for as
 467 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 468 *
 469 * All users of this function must be very careful when walking the anon_vma
 470 * chain and verify that the page in question is indeed mapped in it
 471 * [ something equivalent to page_mapped_in_vma() ].
 472 *
 473 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
 474 * page_remove_rmap() that the anon_vma pointer from page->mapping is valid
 475 * if there is a mapcount, we can dereference the anon_vma after observing
 476 * those.
 
 
 
 
 
 
 477 */
 478struct anon_vma *page_get_anon_vma(struct page *page)
 479{
 480	struct anon_vma *anon_vma = NULL;
 481	unsigned long anon_mapping;
 482
 483	rcu_read_lock();
 484	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
 485	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 486		goto out;
 487	if (!page_mapped(page))
 488		goto out;
 489
 490	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 491	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 492		anon_vma = NULL;
 493		goto out;
 494	}
 495
 496	/*
 497	 * If this page is still mapped, then its anon_vma cannot have been
 498	 * freed.  But if it has been unmapped, we have no security against the
 499	 * anon_vma structure being freed and reused (for another anon_vma:
 500	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
 501	 * above cannot corrupt).
 502	 */
 503	if (!page_mapped(page)) {
 504		rcu_read_unlock();
 505		put_anon_vma(anon_vma);
 506		return NULL;
 507	}
 508out:
 509	rcu_read_unlock();
 510
 511	return anon_vma;
 512}
 513
 514/*
 515 * Similar to page_get_anon_vma() except it locks the anon_vma.
 516 *
 517 * Its a little more complex as it tries to keep the fast path to a single
 518 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 519 * reference like with page_get_anon_vma() and then block on the mutex.
 
 520 */
 521struct anon_vma *page_lock_anon_vma_read(struct page *page)
 
 522{
 523	struct anon_vma *anon_vma = NULL;
 524	struct anon_vma *root_anon_vma;
 525	unsigned long anon_mapping;
 526
 
 527	rcu_read_lock();
 528	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
 529	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 530		goto out;
 531	if (!page_mapped(page))
 532		goto out;
 533
 534	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 535	root_anon_vma = READ_ONCE(anon_vma->root);
 536	if (down_read_trylock(&root_anon_vma->rwsem)) {
 537		/*
 538		 * If the page is still mapped, then this anon_vma is still
 
 
 
 
 
 
 
 
 
 
 
 539		 * its anon_vma, and holding the mutex ensures that it will
 540		 * not go away, see anon_vma_free().
 541		 */
 542		if (!page_mapped(page)) {
 543			up_read(&root_anon_vma->rwsem);
 544			anon_vma = NULL;
 545		}
 546		goto out;
 547	}
 548
 
 
 
 
 
 
 549	/* trylock failed, we got to sleep */
 550	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 551		anon_vma = NULL;
 552		goto out;
 553	}
 554
 555	if (!page_mapped(page)) {
 556		rcu_read_unlock();
 557		put_anon_vma(anon_vma);
 558		return NULL;
 559	}
 560
 561	/* we pinned the anon_vma, its safe to sleep */
 562	rcu_read_unlock();
 563	anon_vma_lock_read(anon_vma);
 564
 
 
 
 
 
 
 
 
 
 
 
 
 565	if (atomic_dec_and_test(&anon_vma->refcount)) {
 566		/*
 567		 * Oops, we held the last refcount, release the lock
 568		 * and bail -- can't simply use put_anon_vma() because
 569		 * we'll deadlock on the anon_vma_lock_write() recursion.
 570		 */
 571		anon_vma_unlock_read(anon_vma);
 572		__put_anon_vma(anon_vma);
 573		anon_vma = NULL;
 574	}
 575
 576	return anon_vma;
 577
 578out:
 579	rcu_read_unlock();
 580	return anon_vma;
 581}
 582
 583void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
 584{
 585	anon_vma_unlock_read(anon_vma);
 586}
 587
 588#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 589/*
 590 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
 591 * important if a PTE was dirty when it was unmapped that it's flushed
 592 * before any IO is initiated on the page to prevent lost writes. Similarly,
 593 * it must be flushed before freeing to prevent data leakage.
 594 */
 595void try_to_unmap_flush(void)
 596{
 597	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 598
 599	if (!tlb_ubc->flush_required)
 600		return;
 601
 602	arch_tlbbatch_flush(&tlb_ubc->arch);
 603	tlb_ubc->flush_required = false;
 604	tlb_ubc->writable = false;
 605}
 606
 607/* Flush iff there are potentially writable TLB entries that can race with IO */
 608void try_to_unmap_flush_dirty(void)
 609{
 610	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 611
 612	if (tlb_ubc->writable)
 613		try_to_unmap_flush();
 614}
 615
 616static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 
 
 
 
 
 
 
 
 
 
 
 617{
 618	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 
 
 
 
 
 619
 620	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
 621	tlb_ubc->flush_required = true;
 622
 623	/*
 624	 * Ensure compiler does not re-order the setting of tlb_flush_batched
 625	 * before the PTE is cleared.
 626	 */
 627	barrier();
 628	mm->tlb_flush_batched = true;
 
 
 
 
 
 
 
 
 
 
 
 
 629
 630	/*
 631	 * If the PTE was dirty then it's best to assume it's writable. The
 632	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
 633	 * before the page is queued for IO.
 634	 */
 635	if (writable)
 636		tlb_ubc->writable = true;
 637}
 638
 639/*
 640 * Returns true if the TLB flush should be deferred to the end of a batch of
 641 * unmap operations to reduce IPIs.
 642 */
 643static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 644{
 645	bool should_defer = false;
 646
 647	if (!(flags & TTU_BATCH_FLUSH))
 648		return false;
 649
 650	/* If remote CPUs need to be flushed then defer batch the flush */
 651	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
 652		should_defer = true;
 653	put_cpu();
 654
 655	return should_defer;
 656}
 657
 658/*
 659 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
 660 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
 661 * operation such as mprotect or munmap to race between reclaim unmapping
 662 * the page and flushing the page. If this race occurs, it potentially allows
 663 * access to data via a stale TLB entry. Tracking all mm's that have TLB
 664 * batching in flight would be expensive during reclaim so instead track
 665 * whether TLB batching occurred in the past and if so then do a flush here
 666 * if required. This will cost one additional flush per reclaim cycle paid
 667 * by the first operation at risk such as mprotect and mumap.
 668 *
 669 * This must be called under the PTL so that an access to tlb_flush_batched
 670 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
 671 * via the PTL.
 672 */
 673void flush_tlb_batched_pending(struct mm_struct *mm)
 674{
 675	if (data_race(mm->tlb_flush_batched)) {
 676		flush_tlb_mm(mm);
 
 677
 
 
 678		/*
 679		 * Do not allow the compiler to re-order the clearing of
 680		 * tlb_flush_batched before the tlb is flushed.
 681		 */
 682		barrier();
 683		mm->tlb_flush_batched = false;
 684	}
 685}
 686#else
 687static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 
 688{
 689}
 690
 691static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 692{
 693	return false;
 694}
 695#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
 696
 697/*
 698 * At what user virtual address is page expected in vma?
 699 * Caller should check the page is actually part of the vma.
 700 */
 701unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
 702{
 703	unsigned long address;
 704	if (PageAnon(page)) {
 705		struct anon_vma *page__anon_vma = page_anon_vma(page);
 706		/*
 707		 * Note: swapoff's unuse_vma() is more efficient with this
 708		 * check, and needs it to match anon_vma when KSM is active.
 709		 */
 710		if (!vma->anon_vma || !page__anon_vma ||
 711		    vma->anon_vma->root != page__anon_vma->root)
 712			return -EFAULT;
 713	} else if (page->mapping) {
 714		if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
 715			return -EFAULT;
 716	} else
 717		return -EFAULT;
 718	address = __vma_address(page, vma);
 719	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
 720		return -EFAULT;
 721	return address;
 
 
 722}
 723
 
 
 
 
 
 724pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 725{
 726	pgd_t *pgd;
 727	p4d_t *p4d;
 728	pud_t *pud;
 729	pmd_t *pmd = NULL;
 730	pmd_t pmde;
 731
 732	pgd = pgd_offset(mm, address);
 733	if (!pgd_present(*pgd))
 734		goto out;
 735
 736	p4d = p4d_offset(pgd, address);
 737	if (!p4d_present(*p4d))
 738		goto out;
 739
 740	pud = pud_offset(p4d, address);
 741	if (!pud_present(*pud))
 742		goto out;
 743
 744	pmd = pmd_offset(pud, address);
 745	/*
 746	 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
 747	 * without holding anon_vma lock for write.  So when looking for a
 748	 * genuine pmde (in which to find pte), test present and !THP together.
 749	 */
 750	pmde = *pmd;
 751	barrier();
 752	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
 753		pmd = NULL;
 754out:
 755	return pmd;
 756}
 757
 758struct page_referenced_arg {
 759	int mapcount;
 760	int referenced;
 761	unsigned long vm_flags;
 762	struct mem_cgroup *memcg;
 763};
 
 764/*
 765 * arg: page_referenced_arg will be passed
 766 */
 767static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
 768			unsigned long address, void *arg)
 769{
 770	struct page_referenced_arg *pra = arg;
 771	struct page_vma_mapped_walk pvmw = {
 772		.page = page,
 773		.vma = vma,
 774		.address = address,
 775	};
 776	int referenced = 0;
 
 777
 778	while (page_vma_mapped_walk(&pvmw)) {
 779		address = pvmw.address;
 780
 781		if (vma->vm_flags & VM_LOCKED) {
 782			page_vma_mapped_walk_done(&pvmw);
 783			pra->vm_flags |= VM_LOCKED;
 784			return false; /* To break the loop */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 785		}
 786
 787		if (pvmw.pte) {
 788			if (ptep_clear_flush_young_notify(vma, address,
 789						pvmw.pte)) {
 790				/*
 791				 * Don't treat a reference through
 792				 * a sequentially read mapping as such.
 793				 * If the page has been used in another mapping,
 794				 * we will catch it; if this other mapping is
 795				 * already gone, the unmap path will have set
 796				 * PG_referenced or activated the page.
 797				 */
 798				if (likely(!(vma->vm_flags & VM_SEQ_READ)))
 799					referenced++;
 800			}
 
 
 
 
 801		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 802			if (pmdp_clear_flush_young_notify(vma, address,
 803						pvmw.pmd))
 804				referenced++;
 805		} else {
 806			/* unexpected pmd-mapped page? */
 807			WARN_ON_ONCE(1);
 808		}
 809
 810		pra->mapcount--;
 811	}
 812
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 813	if (referenced)
 814		clear_page_idle(page);
 815	if (test_and_clear_page_young(page))
 816		referenced++;
 817
 818	if (referenced) {
 819		pra->referenced++;
 820		pra->vm_flags |= vma->vm_flags;
 821	}
 822
 823	if (!pra->mapcount)
 824		return false; /* To break the loop */
 825
 826	return true;
 827}
 828
 829static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
 830{
 831	struct page_referenced_arg *pra = arg;
 832	struct mem_cgroup *memcg = pra->memcg;
 833
 834	if (!mm_match_cgroup(vma->vm_mm, memcg))
 
 
 
 
 
 
 
 
 
 
 
 
 
 835		return true;
 836
 837	return false;
 838}
 839
 840/**
 841 * page_referenced - test if the page was referenced
 842 * @page: the page to test
 843 * @is_locked: caller holds lock on the page
 844 * @memcg: target memory cgroup
 845 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
 
 
 846 *
 847 * Quick test_and_clear_referenced for all mappings to a page,
 848 * returns the number of ptes which referenced the page.
 849 */
 850int page_referenced(struct page *page,
 851		    int is_locked,
 852		    struct mem_cgroup *memcg,
 853		    unsigned long *vm_flags)
 854{
 855	int we_locked = 0;
 856	struct page_referenced_arg pra = {
 857		.mapcount = total_mapcount(page),
 858		.memcg = memcg,
 859	};
 860	struct rmap_walk_control rwc = {
 861		.rmap_one = page_referenced_one,
 862		.arg = (void *)&pra,
 863		.anon_lock = page_lock_anon_vma_read,
 
 
 864	};
 865
 866	*vm_flags = 0;
 867	if (!pra.mapcount)
 868		return 0;
 869
 870	if (!page_rmapping(page))
 871		return 0;
 872
 873	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
 874		we_locked = trylock_page(page);
 875		if (!we_locked)
 876			return 1;
 877	}
 878
 879	/*
 880	 * If we are reclaiming on behalf of a cgroup, skip
 881	 * counting on behalf of references from different
 882	 * cgroups
 883	 */
 884	if (memcg) {
 885		rwc.invalid_vma = invalid_page_referenced_vma;
 886	}
 887
 888	rmap_walk(page, &rwc);
 889	*vm_flags = pra.vm_flags;
 890
 891	if (we_locked)
 892		unlock_page(page);
 893
 894	return pra.referenced;
 895}
 896
 897static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
 898			    unsigned long address, void *arg)
 899{
 900	struct page_vma_mapped_walk pvmw = {
 901		.page = page,
 902		.vma = vma,
 903		.address = address,
 904		.flags = PVMW_SYNC,
 905	};
 906	struct mmu_notifier_range range;
 907	int *cleaned = arg;
 908
 909	/*
 910	 * We have to assume the worse case ie pmd for invalidation. Note that
 911	 * the page can not be free from this function.
 912	 */
 913	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
 914				0, vma, vma->vm_mm, address,
 915				min(vma->vm_end, address + page_size(page)));
 916	mmu_notifier_invalidate_range_start(&range);
 917
 918	while (page_vma_mapped_walk(&pvmw)) {
 919		int ret = 0;
 920
 921		address = pvmw.address;
 922		if (pvmw.pte) {
 923			pte_t entry;
 924			pte_t *pte = pvmw.pte;
 925
 926			if (!pte_dirty(*pte) && !pte_write(*pte))
 927				continue;
 928
 929			flush_cache_page(vma, address, pte_pfn(*pte));
 930			entry = ptep_clear_flush(vma, address, pte);
 931			entry = pte_wrprotect(entry);
 932			entry = pte_mkclean(entry);
 933			set_pte_at(vma->vm_mm, address, pte, entry);
 934			ret = 1;
 935		} else {
 936#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 937			pmd_t *pmd = pvmw.pmd;
 938			pmd_t entry;
 939
 940			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
 941				continue;
 942
 943			flush_cache_page(vma, address, page_to_pfn(page));
 
 944			entry = pmdp_invalidate(vma, address, pmd);
 945			entry = pmd_wrprotect(entry);
 946			entry = pmd_mkclean(entry);
 947			set_pmd_at(vma->vm_mm, address, pmd, entry);
 948			ret = 1;
 949#else
 950			/* unexpected pmd-mapped page? */
 951			WARN_ON_ONCE(1);
 952#endif
 953		}
 954
 955		/*
 956		 * No need to call mmu_notifier_invalidate_range() as we are
 957		 * downgrading page table protection not changing it to point
 958		 * to a new page.
 959		 *
 960		 * See Documentation/vm/mmu_notifier.rst
 961		 */
 962		if (ret)
 963			(*cleaned)++;
 964	}
 965
 966	mmu_notifier_invalidate_range_end(&range);
 967
 
 
 
 
 
 
 
 
 
 
 
 968	return true;
 969}
 970
 971static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
 972{
 973	if (vma->vm_flags & VM_SHARED)
 974		return false;
 975
 976	return true;
 977}
 978
 979int page_mkclean(struct page *page)
 980{
 981	int cleaned = 0;
 982	struct address_space *mapping;
 983	struct rmap_walk_control rwc = {
 984		.arg = (void *)&cleaned,
 985		.rmap_one = page_mkclean_one,
 986		.invalid_vma = invalid_mkclean_vma,
 987	};
 988
 989	BUG_ON(!PageLocked(page));
 990
 991	if (!page_mapped(page))
 992		return 0;
 993
 994	mapping = page_mapping(page);
 995	if (!mapping)
 996		return 0;
 997
 998	rmap_walk(page, &rwc);
 999
1000	return cleaned;
1001}
1002EXPORT_SYMBOL_GPL(page_mkclean);
1003
1004/**
1005 * page_move_anon_rmap - move a page to our anon_vma
1006 * @page:	the page to move to our anon_vma
1007 * @vma:	the vma the page belongs to
1008 *
1009 * When a page belongs exclusively to one process after a COW event,
1010 * that page can be moved into the anon_vma that belongs to just that
1011 * process, so the rmap code will not search the parent or sibling
1012 * processes.
 
 
1013 */
1014void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
 
1015{
1016	struct anon_vma *anon_vma = vma->anon_vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1017
1018	page = compound_head(page);
 
 
 
 
 
 
 
 
 
 
 
1019
1020	VM_BUG_ON_PAGE(!PageLocked(page), page);
1021	VM_BUG_ON_VMA(!anon_vma, vma);
1022
1023	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1024	/*
1025	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1026	 * simultaneously, so a concurrent reader (eg page_referenced()'s
1027	 * PageAnon()) will not see one without the other.
1028	 */
1029	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1030}
1031
1032/**
1033 * __page_set_anon_rmap - set up new anonymous rmap
1034 * @page:	Page or Hugepage to add to rmap
1035 * @vma:	VM area to add page to.
1036 * @address:	User virtual address of the mapping	
1037 * @exclusive:	the page is exclusively owned by the current process
1038 */
1039static void __page_set_anon_rmap(struct page *page,
1040	struct vm_area_struct *vma, unsigned long address, int exclusive)
1041{
1042	struct anon_vma *anon_vma = vma->anon_vma;
1043
1044	BUG_ON(!anon_vma);
1045
1046	if (PageAnon(page))
1047		return;
1048
1049	/*
1050	 * If the page isn't exclusively mapped into this vma,
1051	 * we must use the _oldest_ possible anon_vma for the
1052	 * page mapping!
1053	 */
1054	if (!exclusive)
1055		anon_vma = anon_vma->root;
1056
 
 
 
 
 
 
1057	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1058	page->mapping = (struct address_space *) anon_vma;
1059	page->index = linear_page_index(vma, address);
1060}
1061
1062/**
1063 * __page_check_anon_rmap - sanity check anonymous rmap addition
1064 * @page:	the page to add the mapping to
 
1065 * @vma:	the vm area in which the mapping is added
1066 * @address:	the user virtual address mapped
1067 */
1068static void __page_check_anon_rmap(struct page *page,
1069	struct vm_area_struct *vma, unsigned long address)
1070{
1071	/*
1072	 * The page's anon-rmap details (mapping and index) are guaranteed to
1073	 * be set up correctly at this point.
1074	 *
1075	 * We have exclusion against page_add_anon_rmap because the caller
1076	 * always holds the page locked, except if called from page_dup_rmap,
1077	 * in which case the page is already known to be setup.
1078	 *
1079	 * We have exclusion against page_add_new_anon_rmap because those pages
1080	 * are initially only visible via the pagetables, and the pte is locked
1081	 * over the call to page_add_new_anon_rmap.
1082	 */
1083	VM_BUG_ON_PAGE(page_anon_vma(page)->root != vma->anon_vma->root, page);
 
1084	VM_BUG_ON_PAGE(page_to_pgoff(page) != linear_page_index(vma, address),
1085		       page);
1086}
1087
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1088/**
1089 * page_add_anon_rmap - add pte mapping to an anonymous page
1090 * @page:	the page to add the mapping to
1091 * @vma:	the vm area in which the mapping is added
1092 * @address:	the user virtual address mapped
1093 * @compound:	charge the page as compound or small page
 
 
 
 
1094 *
1095 * The caller needs to hold the pte lock, and the page must be locked in
1096 * the anon_vma case: to serialize mapping,index checking after setting,
1097 * and to ensure that PageAnon is not being upgraded racily to PageKsm
1098 * (but PageKsm is never downgraded to PageAnon).
1099 */
1100void page_add_anon_rmap(struct page *page,
1101	struct vm_area_struct *vma, unsigned long address, bool compound)
 
1102{
1103	do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
 
1104}
1105
1106/*
1107 * Special version of the above for do_swap_page, which often runs
1108 * into pages that are exclusively owned by the current process.
1109 * Everybody else should continue to use page_add_anon_rmap above.
 
 
 
 
 
 
 
 
1110 */
1111void do_page_add_anon_rmap(struct page *page,
1112	struct vm_area_struct *vma, unsigned long address, int flags)
1113{
1114	bool compound = flags & RMAP_COMPOUND;
1115	bool first;
1116
1117	if (unlikely(PageKsm(page)))
1118		lock_page_memcg(page);
1119	else
1120		VM_BUG_ON_PAGE(!PageLocked(page), page);
1121
1122	if (compound) {
1123		atomic_t *mapcount;
1124		VM_BUG_ON_PAGE(!PageLocked(page), page);
1125		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1126		mapcount = compound_mapcount_ptr(page);
1127		first = atomic_inc_and_test(mapcount);
1128	} else {
1129		first = atomic_inc_and_test(&page->_mapcount);
1130	}
1131
1132	if (first) {
1133		int nr = compound ? thp_nr_pages(page) : 1;
1134		/*
1135		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1136		 * these counters are not modified in interrupt context, and
1137		 * pte lock(a spinlock) is held, which implies preemption
1138		 * disabled.
1139		 */
1140		if (compound)
1141			__inc_lruvec_page_state(page, NR_ANON_THPS);
1142		__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1143	}
1144
1145	if (unlikely(PageKsm(page))) {
1146		unlock_page_memcg(page);
1147		return;
1148	}
1149
1150	/* address might be in next vma when migration races vma_adjust */
1151	if (first)
1152		__page_set_anon_rmap(page, vma, address,
1153				flags & RMAP_EXCLUSIVE);
1154	else
1155		__page_check_anon_rmap(page, vma, address);
1156}
1157
1158/**
1159 * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1160 * @page:	the page to add the mapping to
1161 * @vma:	the vm area in which the mapping is added
1162 * @address:	the user virtual address mapped
1163 * @compound:	charge the page as compound or small page
1164 *
1165 * Same as page_add_anon_rmap but must only be called on *new* pages.
1166 * This means the inc-and-test can be bypassed.
1167 * Page does not have to be locked.
 
 
 
1168 */
1169void page_add_new_anon_rmap(struct page *page,
1170	struct vm_area_struct *vma, unsigned long address, bool compound)
1171{
1172	int nr = compound ? thp_nr_pages(page) : 1;
 
 
 
 
 
 
1173
1174	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1175	__SetPageSwapBacked(page);
1176	if (compound) {
1177		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1178		/* increment count (starts at -1) */
1179		atomic_set(compound_mapcount_ptr(page), 0);
1180		if (hpage_pincount_available(page))
1181			atomic_set(compound_pincount_ptr(page), 0);
 
 
 
 
 
 
 
 
 
1182
1183		__inc_lruvec_page_state(page, NR_ANON_THPS);
1184	} else {
1185		/* Anon THP always mapped first with PMD */
1186		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1187		/* increment count (starts at -1) */
1188		atomic_set(&page->_mapcount, 0);
 
 
 
1189	}
1190	__mod_lruvec_page_state(page, NR_ANON_MAPPED, nr);
1191	__page_set_anon_rmap(page, vma, address, 1);
1192}
1193
1194/**
1195 * page_add_file_rmap - add pte mapping to a file page
1196 * @page: the page to add the mapping to
1197 * @compound: charge the page as compound or small page
1198 *
1199 * The caller needs to hold the pte lock.
1200 */
1201void page_add_file_rmap(struct page *page, bool compound)
1202{
1203	int i, nr = 1;
1204
1205	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1206	lock_page_memcg(page);
1207	if (compound && PageTransHuge(page)) {
1208		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1209			if (atomic_inc_and_test(&page[i]._mapcount))
1210				nr++;
1211		}
1212		if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1213			goto out;
1214		if (PageSwapBacked(page))
1215			__inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1216		else
1217			__inc_node_page_state(page, NR_FILE_PMDMAPPED);
1218	} else {
1219		if (PageTransCompound(page) && page_mapping(page)) {
1220			VM_WARN_ON_ONCE(!PageLocked(page));
1221
1222			SetPageDoubleMap(compound_head(page));
1223			if (PageMlocked(page))
1224				clear_page_mlock(compound_head(page));
1225		}
1226		if (!atomic_inc_and_test(&page->_mapcount))
1227			goto out;
1228	}
1229	__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1230out:
1231	unlock_page_memcg(page);
1232}
1233
1234static void page_remove_file_rmap(struct page *page, bool compound)
 
 
1235{
1236	int i, nr = 1;
1237
1238	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1239
1240	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1241	if (unlikely(PageHuge(page))) {
1242		/* hugetlb pages are always mapped with pmds */
1243		atomic_dec(compound_mapcount_ptr(page));
1244		return;
1245	}
1246
1247	/* page still mapped by someone else? */
1248	if (compound && PageTransHuge(page)) {
1249		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1250			if (atomic_add_negative(-1, &page[i]._mapcount))
1251				nr++;
1252		}
1253		if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1254			return;
1255		if (PageSwapBacked(page))
1256			__dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1257		else
1258			__dec_node_page_state(page, NR_FILE_PMDMAPPED);
1259	} else {
1260		if (!atomic_add_negative(-1, &page->_mapcount))
1261			return;
1262	}
1263
1264	/*
1265	 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1266	 * these counters are not modified in interrupt context, and
1267	 * pte lock(a spinlock) is held, which implies preemption disabled.
1268	 */
1269	__mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1270
1271	if (unlikely(PageMlocked(page)))
1272		clear_page_mlock(page);
 
1273}
1274
1275static void page_remove_anon_compound_rmap(struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
1276{
1277	int i, nr;
1278
1279	if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1280		return;
1281
1282	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1283	if (unlikely(PageHuge(page)))
1284		return;
1285
1286	if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1287		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1288
1289	__dec_lruvec_page_state(page, NR_ANON_THPS);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1290
1291	if (TestClearPageDoubleMap(page)) {
1292		/*
1293		 * Subpages can be mapped with PTEs too. Check how many of
1294		 * them are still mapped.
1295		 */
1296		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1297			if (atomic_add_negative(-1, &page[i]._mapcount))
1298				nr++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1299		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1300
1301		/*
1302		 * Queue the page for deferred split if at least one small
1303		 * page of the compound page is unmapped, but at least one
1304		 * small page is still mapped.
1305		 */
1306		if (nr && nr < HPAGE_PMD_NR)
1307			deferred_split_huge_page(page);
1308	} else {
1309		nr = HPAGE_PMD_NR;
1310	}
1311
1312	if (unlikely(PageMlocked(page)))
1313		clear_page_mlock(page);
 
 
 
 
 
1314
1315	if (nr)
1316		__mod_lruvec_page_state(page, NR_ANON_MAPPED, -nr);
1317}
1318
1319/**
1320 * page_remove_rmap - take down pte mapping from a page
1321 * @page:	page to remove mapping from
1322 * @compound:	uncharge the page as compound or small page
 
 
1323 *
1324 * The caller needs to hold the pte lock.
 
 
1325 */
1326void page_remove_rmap(struct page *page, bool compound)
 
1327{
1328	lock_page_memcg(page);
1329
1330	if (!PageAnon(page)) {
1331		page_remove_file_rmap(page, compound);
1332		goto out;
1333	}
1334
1335	if (compound) {
1336		page_remove_anon_compound_rmap(page);
1337		goto out;
1338	}
1339
1340	/* page still mapped by someone else? */
1341	if (!atomic_add_negative(-1, &page->_mapcount))
1342		goto out;
1343
1344	/*
1345	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1346	 * these counters are not modified in interrupt context, and
1347	 * pte lock(a spinlock) is held, which implies preemption disabled.
1348	 */
1349	__dec_lruvec_page_state(page, NR_ANON_MAPPED);
1350
1351	if (unlikely(PageMlocked(page)))
1352		clear_page_mlock(page);
1353
1354	if (PageTransCompound(page))
1355		deferred_split_huge_page(compound_head(page));
1356
1357	/*
1358	 * It would be tidy to reset the PageAnon mapping here,
1359	 * but that might overwrite a racing page_add_anon_rmap
1360	 * which increments mapcount after us but sets mapping
1361	 * before us: so leave the reset to free_unref_page,
1362	 * and remember that it's only reliable while mapped.
1363	 * Leaving it set also helps swapoff to reinstate ptes
1364	 * faster for those pages still in swapcache.
1365	 */
1366out:
1367	unlock_page_memcg(page);
1368}
1369
1370/*
1371 * @arg: enum ttu_flags will be passed to this argument
1372 */
1373static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1374		     unsigned long address, void *arg)
1375{
1376	struct mm_struct *mm = vma->vm_mm;
1377	struct page_vma_mapped_walk pvmw = {
1378		.page = page,
1379		.vma = vma,
1380		.address = address,
1381	};
1382	pte_t pteval;
1383	struct page *subpage;
1384	bool ret = true;
1385	struct mmu_notifier_range range;
1386	enum ttu_flags flags = (enum ttu_flags)(long)arg;
 
 
1387
1388	/* munlock has nothing to gain from examining un-locked vmas */
1389	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1390		return true;
1391
1392	if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1393	    is_zone_device_page(page) && !is_device_private_page(page))
1394		return true;
 
1395
1396	if (flags & TTU_SPLIT_HUGE_PMD) {
1397		split_huge_pmd_address(vma, address,
1398				flags & TTU_SPLIT_FREEZE, page);
1399	}
1400
1401	/*
1402	 * For THP, we have to assume the worse case ie pmd for invalidation.
1403	 * For hugetlb, it could be much worse if we need to do pud
1404	 * invalidation in the case of pmd sharing.
1405	 *
1406	 * Note that the page can not be free in this function as call of
1407	 * try_to_unmap() must hold a reference on the page.
1408	 */
1409	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1410				address,
1411				min(vma->vm_end, address + page_size(page)));
1412	if (PageHuge(page)) {
1413		/*
1414		 * If sharing is possible, start and end will be adjusted
1415		 * accordingly.
1416		 *
1417		 * If called for a huge page, caller must hold i_mmap_rwsem
1418		 * in write mode as it is possible to call huge_pmd_unshare.
1419		 */
1420		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1421						     &range.end);
 
 
 
1422	}
1423	mmu_notifier_invalidate_range_start(&range);
1424
1425	while (page_vma_mapped_walk(&pvmw)) {
1426#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1427		/* PMD-mapped THP migration entry */
1428		if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1429			VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1430
1431			set_pmd_migration_entry(&pvmw, page);
1432			continue;
1433		}
1434#endif
1435
1436		/*
1437		 * If the page is mlock()d, we cannot swap it out.
1438		 * If it's recently referenced (perhaps page_referenced
1439		 * skipped over this mm) then we should reactivate it.
1440		 */
1441		if (!(flags & TTU_IGNORE_MLOCK)) {
1442			if (vma->vm_flags & VM_LOCKED) {
1443				/* PTE-mapped THP are never mlocked */
1444				if (!PageTransCompound(page)) {
1445					/*
1446					 * Holding pte lock, we do *not* need
1447					 * mmap_lock here
1448					 */
1449					mlock_vma_page(page);
1450				}
1451				ret = false;
1452				page_vma_mapped_walk_done(&pvmw);
1453				break;
1454			}
1455			if (flags & TTU_MUNLOCK)
1456				continue;
1457		}
1458
1459		/* Unexpected PMD-mapped THP? */
1460		VM_BUG_ON_PAGE(!pvmw.pte, page);
1461
1462		subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1463		address = pvmw.address;
 
 
1464
1465		if (PageHuge(page)) {
1466			/*
1467			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1468			 * held in write mode.  Caller needs to explicitly
1469			 * do this outside rmap routines.
1470			 */
1471			VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1472			if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
1473				/*
1474				 * huge_pmd_unshare unmapped an entire PMD
1475				 * page.  There is no way of knowing exactly
1476				 * which PMDs may be cached for this mm, so
1477				 * we must flush them all.  start/end were
1478				 * already adjusted above to cover this range.
1479				 */
1480				flush_cache_range(vma, range.start, range.end);
1481				flush_tlb_range(vma, range.start, range.end);
1482				mmu_notifier_invalidate_range(mm, range.start,
1483							      range.end);
1484
1485				/*
1486				 * The ref count of the PMD page was dropped
1487				 * which is part of the way map counting
1488				 * is done for shared PMDs.  Return 'true'
1489				 * here.  When there is no other sharing,
1490				 * huge_pmd_unshare returns false and we will
1491				 * unmap the actual page and drop map count
1492				 * to zero.
1493				 */
1494				page_vma_mapped_walk_done(&pvmw);
1495				break;
1496			}
1497		}
1498
1499		if (IS_ENABLED(CONFIG_MIGRATION) &&
1500		    (flags & TTU_MIGRATION) &&
1501		    is_zone_device_page(page)) {
1502			swp_entry_t entry;
1503			pte_t swp_pte;
1504
1505			pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1506
1507			/*
1508			 * Store the pfn of the page in a special migration
1509			 * pte. do_swap_page() will wait until the migration
1510			 * pte is removed and then restart fault handling.
1511			 */
1512			entry = make_migration_entry(page, 0);
1513			swp_pte = swp_entry_to_pte(entry);
1514
1515			/*
1516			 * pteval maps a zone device page and is therefore
1517			 * a swap pte.
 
 
 
1518			 */
1519			if (pte_swp_soft_dirty(pteval))
1520				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1521			if (pte_swp_uffd_wp(pteval))
1522				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1523			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1524			/*
1525			 * No need to invalidate here it will synchronize on
1526			 * against the special swap migration pte.
 
1527			 *
1528			 * The assignment to subpage above was computed from a
1529			 * swap PTE which results in an invalid pointer.
1530			 * Since only PAGE_SIZE pages can currently be
1531			 * migrated, just set it to page. This will need to be
1532			 * changed when hugepage migrations to device private
1533			 * memory are supported.
1534			 */
1535			subpage = page;
1536			goto discard;
1537		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1538
1539		if (!(flags & TTU_IGNORE_ACCESS)) {
1540			if (ptep_clear_flush_young_notify(vma, address,
1541						pvmw.pte)) {
1542				ret = false;
1543				page_vma_mapped_walk_done(&pvmw);
1544				break;
1545			}
1546		}
1547
1548		/* Nuke the page table entry. */
1549		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1550		if (should_defer_flush(mm, flags)) {
1551			/*
1552			 * We clear the PTE but do not flush so potentially
1553			 * a remote CPU could still be writing to the page.
1554			 * If the entry was previously clean then the
1555			 * architecture must guarantee that a clear->dirty
1556			 * transition on a cached TLB entry is written through
1557			 * and traps if the PTE is unmapped.
1558			 */
1559			pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1560
1561			set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1562		} else {
1563			pteval = ptep_clear_flush(vma, address, pvmw.pte);
1564		}
1565
1566		/* Move the dirty bit to the page. Now the pte is gone. */
1567		if (pte_dirty(pteval))
1568			set_page_dirty(page);
1569
1570		/* Update high watermark before we lower rss */
1571		update_hiwater_rss(mm);
1572
1573		if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1574			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1575			if (PageHuge(page)) {
1576				hugetlb_count_sub(compound_nr(page), mm);
1577				set_huge_swap_pte_at(mm, address,
1578						     pvmw.pte, pteval,
1579						     vma_mmu_pagesize(vma));
1580			} else {
1581				dec_mm_counter(mm, mm_counter(page));
1582				set_pte_at(mm, address, pvmw.pte, pteval);
1583			}
1584
1585		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1586			/*
1587			 * The guest indicated that the page content is of no
1588			 * interest anymore. Simply discard the pte, vmscan
1589			 * will take care of the rest.
1590			 * A future reference will then fault in a new zero
1591			 * page. When userfaultfd is active, we must not drop
1592			 * this page though, as its main user (postcopy
1593			 * migration) will not expect userfaults on already
1594			 * copied pages.
1595			 */
1596			dec_mm_counter(mm, mm_counter(page));
1597			/* We have to invalidate as we cleared the pte */
1598			mmu_notifier_invalidate_range(mm, address,
1599						      address + PAGE_SIZE);
1600		} else if (IS_ENABLED(CONFIG_MIGRATION) &&
1601				(flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1602			swp_entry_t entry;
1603			pte_t swp_pte;
1604
1605			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1606				set_pte_at(mm, address, pvmw.pte, pteval);
1607				ret = false;
1608				page_vma_mapped_walk_done(&pvmw);
1609				break;
1610			}
1611
1612			/*
1613			 * Store the pfn of the page in a special migration
1614			 * pte. do_swap_page() will wait until the migration
1615			 * pte is removed and then restart fault handling.
1616			 */
1617			entry = make_migration_entry(subpage,
1618					pte_write(pteval));
1619			swp_pte = swp_entry_to_pte(entry);
1620			if (pte_soft_dirty(pteval))
1621				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1622			if (pte_uffd_wp(pteval))
1623				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1624			set_pte_at(mm, address, pvmw.pte, swp_pte);
1625			/*
1626			 * No need to invalidate here it will synchronize on
1627			 * against the special swap migration pte.
1628			 */
1629		} else if (PageAnon(page)) {
1630			swp_entry_t entry = { .val = page_private(subpage) };
1631			pte_t swp_pte;
1632			/*
1633			 * Store the swap location in the pte.
1634			 * See handle_pte_fault() ...
1635			 */
1636			if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
 
1637				WARN_ON_ONCE(1);
1638				ret = false;
1639				/* We have to invalidate as we cleared the pte */
1640				mmu_notifier_invalidate_range(mm, address,
1641							address + PAGE_SIZE);
1642				page_vma_mapped_walk_done(&pvmw);
1643				break;
1644			}
1645
1646			/* MADV_FREE page check */
1647			if (!PageSwapBacked(page)) {
1648				if (!PageDirty(page)) {
1649					/* Invalidate as we cleared the pte */
1650					mmu_notifier_invalidate_range(mm,
1651						address, address + PAGE_SIZE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1652					dec_mm_counter(mm, MM_ANONPAGES);
1653					goto discard;
1654				}
1655
1656				/*
1657				 * If the page was redirtied, it cannot be
1658				 * discarded. Remap the page to page table.
1659				 */
1660				set_pte_at(mm, address, pvmw.pte, pteval);
1661				SetPageSwapBacked(page);
1662				ret = false;
1663				page_vma_mapped_walk_done(&pvmw);
1664				break;
1665			}
1666
1667			if (swap_duplicate(entry) < 0) {
1668				set_pte_at(mm, address, pvmw.pte, pteval);
1669				ret = false;
1670				page_vma_mapped_walk_done(&pvmw);
1671				break;
1672			}
1673			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
 
 
 
 
 
 
 
 
 
 
 
1674				set_pte_at(mm, address, pvmw.pte, pteval);
1675				ret = false;
1676				page_vma_mapped_walk_done(&pvmw);
1677				break;
1678			}
1679			if (list_empty(&mm->mmlist)) {
1680				spin_lock(&mmlist_lock);
1681				if (list_empty(&mm->mmlist))
1682					list_add(&mm->mmlist, &init_mm.mmlist);
1683				spin_unlock(&mmlist_lock);
1684			}
1685			dec_mm_counter(mm, MM_ANONPAGES);
1686			inc_mm_counter(mm, MM_SWAPENTS);
1687			swp_pte = swp_entry_to_pte(entry);
 
 
1688			if (pte_soft_dirty(pteval))
1689				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1690			if (pte_uffd_wp(pteval))
1691				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1692			set_pte_at(mm, address, pvmw.pte, swp_pte);
1693			/* Invalidate as we cleared the pte */
1694			mmu_notifier_invalidate_range(mm, address,
1695						      address + PAGE_SIZE);
1696		} else {
1697			/*
1698			 * This is a locked file-backed page, thus it cannot
1699			 * be removed from the page cache and replaced by a new
1700			 * page before mmu_notifier_invalidate_range_end, so no
1701			 * concurrent thread might update its page table to
1702			 * point at new page while a device still is using this
1703			 * page.
 
1704			 *
1705			 * See Documentation/vm/mmu_notifier.rst
1706			 */
1707			dec_mm_counter(mm, mm_counter_file(page));
1708		}
1709discard:
1710		/*
1711		 * No need to call mmu_notifier_invalidate_range() it has be
1712		 * done above for all cases requiring it to happen under page
1713		 * table lock before mmu_notifier_invalidate_range_end()
1714		 *
1715		 * See Documentation/vm/mmu_notifier.rst
1716		 */
1717		page_remove_rmap(subpage, PageHuge(page));
1718		put_page(page);
1719	}
1720
1721	mmu_notifier_invalidate_range_end(&range);
1722
1723	return ret;
1724}
1725
1726static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1727{
1728	return vma_is_temporary_stack(vma);
1729}
1730
1731static int page_mapcount_is_zero(struct page *page)
1732{
1733	return !total_mapcount(page);
1734}
1735
1736/**
1737 * try_to_unmap - try to remove all page table mappings to a page
1738 * @page: the page to get unmapped
1739 * @flags: action and flags
1740 *
1741 * Tries to remove all the page table entries which are mapping this
1742 * page, used in the pageout path.  Caller must hold the page lock.
 
1743 *
1744 * If unmap is successful, return true. Otherwise, false.
1745 */
1746bool try_to_unmap(struct page *page, enum ttu_flags flags)
1747{
1748	struct rmap_walk_control rwc = {
1749		.rmap_one = try_to_unmap_one,
1750		.arg = (void *)flags,
1751		.done = page_mapcount_is_zero,
1752		.anon_lock = page_lock_anon_vma_read,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1753	};
1754
1755	/*
 
 
 
 
 
 
 
 
 
 
 
 
1756	 * During exec, a temporary VMA is setup and later moved.
1757	 * The VMA is moved under the anon_vma lock but not the
1758	 * page tables leading to a race where migration cannot
1759	 * find the migration ptes. Rather than increasing the
1760	 * locking requirements of exec(), migration skips
1761	 * temporary VMAs until after exec() completes.
1762	 */
1763	if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1764	    && !PageKsm(page) && PageAnon(page))
1765		rwc.invalid_vma = invalid_migration_vma;
1766
1767	if (flags & TTU_RMAP_LOCKED)
1768		rmap_walk_locked(page, &rwc);
1769	else
1770		rmap_walk(page, &rwc);
1771
1772	return !page_mapcount(page) ? true : false;
1773}
1774
1775static int page_not_mapped(struct page *page)
1776{
1777	return !page_mapped(page);
 
 
 
1778};
1779
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1780/**
1781 * try_to_munlock - try to munlock a page
1782 * @page: the page to be munlocked
 
 
 
 
 
 
 
1783 *
1784 * Called from munlock code.  Checks all of the VMAs mapping the page
1785 * to make sure nobody else has this page mlocked. The page will be
1786 * returned with PG_mlocked cleared if no other vmas have it mlocked.
1787 */
1788
1789void try_to_munlock(struct page *page)
1790{
 
 
 
 
 
 
1791	struct rmap_walk_control rwc = {
1792		.rmap_one = try_to_unmap_one,
1793		.arg = (void *)TTU_MUNLOCK,
1794		.done = page_not_mapped,
1795		.anon_lock = page_lock_anon_vma_read,
1796
1797	};
1798
1799	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1800	VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1801
1802	rmap_walk(page, &rwc);
1803}
 
 
1804
1805void __put_anon_vma(struct anon_vma *anon_vma)
1806{
1807	struct anon_vma *root = anon_vma->root;
1808
1809	anon_vma_free(anon_vma);
1810	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1811		anon_vma_free(root);
1812}
1813
1814static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1815					struct rmap_walk_control *rwc)
1816{
1817	struct anon_vma *anon_vma;
1818
1819	if (rwc->anon_lock)
1820		return rwc->anon_lock(page);
1821
1822	/*
1823	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1824	 * because that depends on page_mapped(); but not all its usages
1825	 * are holding mmap_lock. Users without mmap_lock are required to
1826	 * take a reference count to prevent the anon_vma disappearing
1827	 */
1828	anon_vma = page_anon_vma(page);
1829	if (!anon_vma)
1830		return NULL;
1831
 
 
 
 
 
 
 
 
 
1832	anon_vma_lock_read(anon_vma);
 
1833	return anon_vma;
1834}
1835
1836/*
1837 * rmap_walk_anon - do something to anonymous page using the object-based
1838 * rmap method
1839 * @page: the page to be handled
1840 * @rwc: control variable according to each walk type
 
1841 *
1842 * Find all the mappings of a page using the mapping pointer and the vma chains
1843 * contained in the anon_vma struct it points to.
1844 *
1845 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1846 * where the page was found will be held for write.  So, we won't recheck
1847 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1848 * LOCKED.
1849 */
1850static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1851		bool locked)
1852{
1853	struct anon_vma *anon_vma;
1854	pgoff_t pgoff_start, pgoff_end;
1855	struct anon_vma_chain *avc;
1856
1857	if (locked) {
1858		anon_vma = page_anon_vma(page);
1859		/* anon_vma disappear under us? */
1860		VM_BUG_ON_PAGE(!anon_vma, page);
1861	} else {
1862		anon_vma = rmap_walk_anon_lock(page, rwc);
1863	}
1864	if (!anon_vma)
1865		return;
1866
1867	pgoff_start = page_to_pgoff(page);
1868	pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
1869	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1870			pgoff_start, pgoff_end) {
1871		struct vm_area_struct *vma = avc->vma;
1872		unsigned long address = vma_address(page, vma);
1873
 
1874		cond_resched();
1875
1876		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1877			continue;
1878
1879		if (!rwc->rmap_one(page, vma, address, rwc->arg))
1880			break;
1881		if (rwc->done && rwc->done(page))
1882			break;
1883	}
1884
1885	if (!locked)
1886		anon_vma_unlock_read(anon_vma);
1887}
1888
1889/*
1890 * rmap_walk_file - do something to file page using the object-based rmap method
1891 * @page: the page to be handled
1892 * @rwc: control variable according to each walk type
 
1893 *
1894 * Find all the mappings of a page using the mapping pointer and the vma chains
1895 * contained in the address_space struct it points to.
1896 *
1897 * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
1898 * where the page was found will be held for write.  So, we won't recheck
1899 * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1900 * LOCKED.
1901 */
1902static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1903		bool locked)
1904{
1905	struct address_space *mapping = page_mapping(page);
1906	pgoff_t pgoff_start, pgoff_end;
1907	struct vm_area_struct *vma;
1908
1909	/*
1910	 * The page lock not only makes sure that page->mapping cannot
1911	 * suddenly be NULLified by truncation, it makes sure that the
1912	 * structure at mapping cannot be freed and reused yet,
1913	 * so we can safely take mapping->i_mmap_rwsem.
1914	 */
1915	VM_BUG_ON_PAGE(!PageLocked(page), page);
1916
1917	if (!mapping)
1918		return;
1919
1920	pgoff_start = page_to_pgoff(page);
1921	pgoff_end = pgoff_start + thp_nr_pages(page) - 1;
1922	if (!locked)
 
 
 
 
 
 
 
 
1923		i_mmap_lock_read(mapping);
 
 
1924	vma_interval_tree_foreach(vma, &mapping->i_mmap,
1925			pgoff_start, pgoff_end) {
1926		unsigned long address = vma_address(page, vma);
1927
 
1928		cond_resched();
1929
1930		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1931			continue;
1932
1933		if (!rwc->rmap_one(page, vma, address, rwc->arg))
1934			goto done;
1935		if (rwc->done && rwc->done(page))
1936			goto done;
1937	}
1938
1939done:
1940	if (!locked)
1941		i_mmap_unlock_read(mapping);
1942}
1943
1944void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1945{
1946	if (unlikely(PageKsm(page)))
1947		rmap_walk_ksm(page, rwc);
1948	else if (PageAnon(page))
1949		rmap_walk_anon(page, rwc, false);
1950	else
1951		rmap_walk_file(page, rwc, false);
1952}
1953
1954/* Like rmap_walk, but caller holds relevant rmap lock */
1955void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1956{
1957	/* no ksm support for now */
1958	VM_BUG_ON_PAGE(PageKsm(page), page);
1959	if (PageAnon(page))
1960		rmap_walk_anon(page, rwc, true);
1961	else
1962		rmap_walk_file(page, rwc, true);
1963}
1964
1965#ifdef CONFIG_HUGETLB_PAGE
1966/*
1967 * The following two functions are for anonymous (private mapped) hugepages.
1968 * Unlike common anonymous pages, anonymous hugepages have no accounting code
1969 * and no lru code, because we handle hugepages differently from common pages.
1970 */
1971void hugepage_add_anon_rmap(struct page *page,
1972			    struct vm_area_struct *vma, unsigned long address)
1973{
1974	struct anon_vma *anon_vma = vma->anon_vma;
1975	int first;
1976
1977	BUG_ON(!PageLocked(page));
1978	BUG_ON(!anon_vma);
1979	/* address might be in next vma when migration races vma_adjust */
1980	first = atomic_inc_and_test(compound_mapcount_ptr(page));
1981	if (first)
1982		__page_set_anon_rmap(page, vma, address, 0);
1983}
1984
1985void hugepage_add_new_anon_rmap(struct page *page,
1986			struct vm_area_struct *vma, unsigned long address)
1987{
1988	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1989	atomic_set(compound_mapcount_ptr(page), 0);
1990	if (hpage_pincount_available(page))
1991		atomic_set(compound_pincount_ptr(page), 0);
1992
1993	__page_set_anon_rmap(page, vma, address, 1);
 
 
 
 
 
1994}
1995#endif /* CONFIG_HUGETLB_PAGE */