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_rwsem	(while writing or truncating, not reading or faulting)
  24 *   mm->mmap_lock
  25 *     mapping->invalidate_lock (in filemap_fault)
  26 *       folio_lock
  27 *         hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share, see hugetlbfs below)
  28 *           vma_start_write
  29 *             mapping->i_mmap_rwsem
  30 *               anon_vma->rwsem
  31 *                 mm->page_table_lock or pte_lock
  32 *                   swap_lock (in swap_duplicate, swap_info_get)
  33 *                     mmlist_lock (in mmput, drain_mmlist and others)
  34 *                     mapping->private_lock (in block_dirty_folio)
 
  35 *                         i_pages lock (widely used)
  36 *                           lruvec->lru_lock (in folio_lruvec_lock_irq)
  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_mmap_rwsem   (memory_failure, collect_procs_anon)
  45 *   ->tasklist_lock
  46 *     pte map lock
  47 *
  48 * hugetlbfs PageHuge() take locks in this order:
  49 *   hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
  50 *     vma_lock (hugetlb specific lock for pmd_sharing)
  51 *       mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)
  52 *         folio_lock
  53 */
  54
  55#include <linux/mm.h>
  56#include <linux/sched/mm.h>
  57#include <linux/sched/task.h>
  58#include <linux/pagemap.h>
  59#include <linux/swap.h>
  60#include <linux/swapops.h>
  61#include <linux/slab.h>
  62#include <linux/init.h>
  63#include <linux/ksm.h>
  64#include <linux/rmap.h>
  65#include <linux/rcupdate.h>
  66#include <linux/export.h>
  67#include <linux/memcontrol.h>
  68#include <linux/mmu_notifier.h>
  69#include <linux/migrate.h>
  70#include <linux/hugetlb.h>
  71#include <linux/huge_mm.h>
  72#include <linux/backing-dev.h>
  73#include <linux/page_idle.h>
  74#include <linux/memremap.h>
  75#include <linux/userfaultfd_k.h>
  76#include <linux/mm_inline.h>
  77#include <linux/oom.h>
  78
  79#include <asm/tlbflush.h>
  80
  81#define CREATE_TRACE_POINTS
  82#include <trace/events/tlb.h>
  83#include <trace/events/migrate.h>
  84
  85#include "internal.h"
  86
  87static struct kmem_cache *anon_vma_cachep;
  88static struct kmem_cache *anon_vma_chain_cachep;
  89
  90static inline struct anon_vma *anon_vma_alloc(void)
  91{
  92	struct anon_vma *anon_vma;
  93
  94	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
  95	if (anon_vma) {
  96		atomic_set(&anon_vma->refcount, 1);
  97		anon_vma->num_children = 0;
  98		anon_vma->num_active_vmas = 0;
  99		anon_vma->parent = anon_vma;
 100		/*
 101		 * Initialise the anon_vma root to point to itself. If called
 102		 * from fork, the root will be reset to the parents anon_vma.
 103		 */
 104		anon_vma->root = anon_vma;
 105	}
 106
 107	return anon_vma;
 108}
 109
 110static inline void anon_vma_free(struct anon_vma *anon_vma)
 111{
 112	VM_BUG_ON(atomic_read(&anon_vma->refcount));
 113
 114	/*
 115	 * Synchronize against folio_lock_anon_vma_read() such that
 116	 * we can safely hold the lock without the anon_vma getting
 117	 * freed.
 118	 *
 119	 * Relies on the full mb implied by the atomic_dec_and_test() from
 120	 * put_anon_vma() against the acquire barrier implied by
 121	 * down_read_trylock() from folio_lock_anon_vma_read(). This orders:
 122	 *
 123	 * folio_lock_anon_vma_read()	VS	put_anon_vma()
 124	 *   down_read_trylock()		  atomic_dec_and_test()
 125	 *   LOCK				  MB
 126	 *   atomic_read()			  rwsem_is_locked()
 127	 *
 128	 * LOCK should suffice since the actual taking of the lock must
 129	 * happen _before_ what follows.
 130	 */
 131	might_sleep();
 132	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
 133		anon_vma_lock_write(anon_vma);
 134		anon_vma_unlock_write(anon_vma);
 135	}
 136
 137	kmem_cache_free(anon_vma_cachep, anon_vma);
 138}
 139
 140static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
 141{
 142	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
 143}
 144
 145static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
 146{
 147	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
 148}
 149
 150static void anon_vma_chain_link(struct vm_area_struct *vma,
 151				struct anon_vma_chain *avc,
 152				struct anon_vma *anon_vma)
 153{
 154	avc->vma = vma;
 155	avc->anon_vma = anon_vma;
 156	list_add(&avc->same_vma, &vma->anon_vma_chain);
 157	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
 158}
 159
 160/**
 161 * __anon_vma_prepare - attach an anon_vma to a memory region
 162 * @vma: the memory region in question
 163 *
 164 * This makes sure the memory mapping described by 'vma' has
 165 * an 'anon_vma' attached to it, so that we can associate the
 166 * anonymous pages mapped into it with that anon_vma.
 167 *
 168 * The common case will be that we already have one, which
 169 * is handled inline by anon_vma_prepare(). But if
 170 * not we either need to find an adjacent mapping that we
 171 * can re-use the anon_vma from (very common when the only
 172 * reason for splitting a vma has been mprotect()), or we
 173 * allocate a new one.
 174 *
 175 * Anon-vma allocations are very subtle, because we may have
 176 * optimistically looked up an anon_vma in folio_lock_anon_vma_read()
 177 * and that may actually touch the rwsem even in the newly
 178 * allocated vma (it depends on RCU to make sure that the
 179 * anon_vma isn't actually destroyed).
 180 *
 181 * As a result, we need to do proper anon_vma locking even
 182 * for the new allocation. At the same time, we do not want
 183 * to do any locking for the common case of already having
 184 * an anon_vma.
 
 
 185 */
 186int __anon_vma_prepare(struct vm_area_struct *vma)
 187{
 188	struct mm_struct *mm = vma->vm_mm;
 189	struct anon_vma *anon_vma, *allocated;
 190	struct anon_vma_chain *avc;
 191
 192	mmap_assert_locked(mm);
 193	might_sleep();
 194
 195	avc = anon_vma_chain_alloc(GFP_KERNEL);
 196	if (!avc)
 197		goto out_enomem;
 198
 199	anon_vma = find_mergeable_anon_vma(vma);
 200	allocated = NULL;
 201	if (!anon_vma) {
 202		anon_vma = anon_vma_alloc();
 203		if (unlikely(!anon_vma))
 204			goto out_enomem_free_avc;
 205		anon_vma->num_children++; /* self-parent link for new root */
 206		allocated = anon_vma;
 207	}
 208
 209	anon_vma_lock_write(anon_vma);
 210	/* page_table_lock to protect against threads */
 211	spin_lock(&mm->page_table_lock);
 212	if (likely(!vma->anon_vma)) {
 213		vma->anon_vma = anon_vma;
 214		anon_vma_chain_link(vma, avc, anon_vma);
 215		anon_vma->num_active_vmas++;
 216		allocated = NULL;
 217		avc = NULL;
 218	}
 219	spin_unlock(&mm->page_table_lock);
 220	anon_vma_unlock_write(anon_vma);
 221
 222	if (unlikely(allocated))
 223		put_anon_vma(allocated);
 224	if (unlikely(avc))
 225		anon_vma_chain_free(avc);
 226
 227	return 0;
 228
 229 out_enomem_free_avc:
 230	anon_vma_chain_free(avc);
 231 out_enomem:
 232	return -ENOMEM;
 233}
 234
 235/*
 236 * This is a useful helper function for locking the anon_vma root as
 237 * we traverse the vma->anon_vma_chain, looping over anon_vma's that
 238 * have the same vma.
 239 *
 240 * Such anon_vma's should have the same root, so you'd expect to see
 241 * just a single mutex_lock for the whole traversal.
 242 */
 243static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
 244{
 245	struct anon_vma *new_root = anon_vma->root;
 246	if (new_root != root) {
 247		if (WARN_ON_ONCE(root))
 248			up_write(&root->rwsem);
 249		root = new_root;
 250		down_write(&root->rwsem);
 251	}
 252	return root;
 253}
 254
 255static inline void unlock_anon_vma_root(struct anon_vma *root)
 256{
 257	if (root)
 258		up_write(&root->rwsem);
 259}
 260
 261/*
 262 * Attach the anon_vmas from src to dst.
 263 * Returns 0 on success, -ENOMEM on failure.
 264 *
 265 * anon_vma_clone() is called by vma_expand(), vma_merge(), __split_vma(),
 266 * copy_vma() and anon_vma_fork(). The first four want an exact copy of src,
 267 * while the last one, anon_vma_fork(), may try to reuse an existing anon_vma to
 268 * prevent endless growth of anon_vma. Since dst->anon_vma is set to NULL before
 269 * call, we can identify this case by checking (!dst->anon_vma &&
 270 * src->anon_vma).
 271 *
 272 * If (!dst->anon_vma && src->anon_vma) is true, this function tries to find
 273 * and reuse existing anon_vma which has no vmas and only one child anon_vma.
 274 * This prevents degradation of anon_vma hierarchy to endless linear chain in
 275 * case of constantly forking task. On the other hand, an anon_vma with more
 276 * than one child isn't reused even if there was no alive vma, thus rmap
 277 * walker has a good chance of avoiding scanning the whole hierarchy when it
 278 * searches where page is mapped.
 279 */
 280int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
 281{
 282	struct anon_vma_chain *avc, *pavc;
 283	struct anon_vma *root = NULL;
 284
 285	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
 286		struct anon_vma *anon_vma;
 287
 288		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
 289		if (unlikely(!avc)) {
 290			unlock_anon_vma_root(root);
 291			root = NULL;
 292			avc = anon_vma_chain_alloc(GFP_KERNEL);
 293			if (!avc)
 294				goto enomem_failure;
 295		}
 296		anon_vma = pavc->anon_vma;
 297		root = lock_anon_vma_root(root, anon_vma);
 298		anon_vma_chain_link(dst, avc, anon_vma);
 299
 300		/*
 301		 * Reuse existing anon_vma if it has no vma and only one
 302		 * anon_vma child.
 303		 *
 304		 * Root anon_vma is never reused:
 305		 * it has self-parent reference and at least one child.
 306		 */
 307		if (!dst->anon_vma && src->anon_vma &&
 308		    anon_vma->num_children < 2 &&
 309		    anon_vma->num_active_vmas == 0)
 310			dst->anon_vma = anon_vma;
 311	}
 312	if (dst->anon_vma)
 313		dst->anon_vma->num_active_vmas++;
 314	unlock_anon_vma_root(root);
 315	return 0;
 316
 317 enomem_failure:
 318	/*
 319	 * dst->anon_vma is dropped here otherwise its num_active_vmas can
 320	 * be incorrectly decremented in unlink_anon_vmas().
 321	 * We can safely do this because callers of anon_vma_clone() don't care
 322	 * about dst->anon_vma if anon_vma_clone() failed.
 323	 */
 324	dst->anon_vma = NULL;
 325	unlink_anon_vmas(dst);
 326	return -ENOMEM;
 327}
 328
 329/*
 330 * Attach vma to its own anon_vma, as well as to the anon_vmas that
 331 * the corresponding VMA in the parent process is attached to.
 332 * Returns 0 on success, non-zero on failure.
 333 */
 334int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
 335{
 336	struct anon_vma_chain *avc;
 337	struct anon_vma *anon_vma;
 338	int error;
 339
 340	/* Don't bother if the parent process has no anon_vma here. */
 341	if (!pvma->anon_vma)
 342		return 0;
 343
 344	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
 345	vma->anon_vma = NULL;
 346
 347	/*
 348	 * First, attach the new VMA to the parent VMA's anon_vmas,
 349	 * so rmap can find non-COWed pages in child processes.
 350	 */
 351	error = anon_vma_clone(vma, pvma);
 352	if (error)
 353		return error;
 354
 355	/* An existing anon_vma has been reused, all done then. */
 356	if (vma->anon_vma)
 357		return 0;
 358
 359	/* Then add our own anon_vma. */
 360	anon_vma = anon_vma_alloc();
 361	if (!anon_vma)
 362		goto out_error;
 363	anon_vma->num_active_vmas++;
 364	avc = anon_vma_chain_alloc(GFP_KERNEL);
 365	if (!avc)
 366		goto out_error_free_anon_vma;
 367
 368	/*
 369	 * The root anon_vma's rwsem is the lock actually used when we
 370	 * lock any of the anon_vmas in this anon_vma tree.
 371	 */
 372	anon_vma->root = pvma->anon_vma->root;
 373	anon_vma->parent = pvma->anon_vma;
 374	/*
 375	 * With refcounts, an anon_vma can stay around longer than the
 376	 * process it belongs to. The root anon_vma needs to be pinned until
 377	 * this anon_vma is freed, because the lock lives in the root.
 378	 */
 379	get_anon_vma(anon_vma->root);
 380	/* Mark this anon_vma as the one where our new (COWed) pages go. */
 381	vma->anon_vma = anon_vma;
 382	anon_vma_lock_write(anon_vma);
 383	anon_vma_chain_link(vma, avc, anon_vma);
 384	anon_vma->parent->num_children++;
 385	anon_vma_unlock_write(anon_vma);
 386
 387	return 0;
 388
 389 out_error_free_anon_vma:
 390	put_anon_vma(anon_vma);
 391 out_error:
 392	unlink_anon_vmas(vma);
 393	return -ENOMEM;
 394}
 395
 396void unlink_anon_vmas(struct vm_area_struct *vma)
 397{
 398	struct anon_vma_chain *avc, *next;
 399	struct anon_vma *root = NULL;
 400
 401	/*
 402	 * Unlink each anon_vma chained to the VMA.  This list is ordered
 403	 * from newest to oldest, ensuring the root anon_vma gets freed last.
 404	 */
 405	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 406		struct anon_vma *anon_vma = avc->anon_vma;
 407
 408		root = lock_anon_vma_root(root, anon_vma);
 409		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
 410
 411		/*
 412		 * Leave empty anon_vmas on the list - we'll need
 413		 * to free them outside the lock.
 414		 */
 415		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
 416			anon_vma->parent->num_children--;
 417			continue;
 418		}
 419
 420		list_del(&avc->same_vma);
 421		anon_vma_chain_free(avc);
 422	}
 423	if (vma->anon_vma) {
 424		vma->anon_vma->num_active_vmas--;
 425
 426		/*
 427		 * vma would still be needed after unlink, and anon_vma will be prepared
 428		 * when handle fault.
 429		 */
 430		vma->anon_vma = NULL;
 431	}
 432	unlock_anon_vma_root(root);
 433
 434	/*
 435	 * Iterate the list once more, it now only contains empty and unlinked
 436	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
 437	 * needing to write-acquire the anon_vma->root->rwsem.
 438	 */
 439	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
 440		struct anon_vma *anon_vma = avc->anon_vma;
 441
 442		VM_WARN_ON(anon_vma->num_children);
 443		VM_WARN_ON(anon_vma->num_active_vmas);
 444		put_anon_vma(anon_vma);
 445
 446		list_del(&avc->same_vma);
 447		anon_vma_chain_free(avc);
 448	}
 449}
 450
 451static void anon_vma_ctor(void *data)
 452{
 453	struct anon_vma *anon_vma = data;
 454
 455	init_rwsem(&anon_vma->rwsem);
 456	atomic_set(&anon_vma->refcount, 0);
 457	anon_vma->rb_root = RB_ROOT_CACHED;
 458}
 459
 460void __init anon_vma_init(void)
 461{
 462	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
 463			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
 464			anon_vma_ctor);
 465	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
 466			SLAB_PANIC|SLAB_ACCOUNT);
 467}
 468
 469/*
 470 * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
 471 *
 472 * Since there is no serialization what so ever against folio_remove_rmap_*()
 473 * the best this function can do is return a refcount increased anon_vma
 474 * that might have been relevant to this page.
 475 *
 476 * The page might have been remapped to a different anon_vma or the anon_vma
 477 * returned may already be freed (and even reused).
 478 *
 479 * In case it was remapped to a different anon_vma, the new anon_vma will be a
 480 * child of the old anon_vma, and the anon_vma lifetime rules will therefore
 481 * ensure that any anon_vma obtained from the page will still be valid for as
 482 * long as we observe page_mapped() [ hence all those page_mapped() tests ].
 483 *
 484 * All users of this function must be very careful when walking the anon_vma
 485 * chain and verify that the page in question is indeed mapped in it
 486 * [ something equivalent to page_mapped_in_vma() ].
 487 *
 488 * Since anon_vma's slab is SLAB_TYPESAFE_BY_RCU and we know from
 489 * folio_remove_rmap_*() that the anon_vma pointer from page->mapping is valid
 490 * if there is a mapcount, we can dereference the anon_vma after observing
 491 * those.
 492 *
 493 * NOTE: the caller should normally hold folio lock when calling this.  If
 494 * not, the caller needs to double check the anon_vma didn't change after
 495 * taking the anon_vma lock for either read or write (UFFDIO_MOVE can modify it
 496 * concurrently without folio lock protection). See folio_lock_anon_vma_read()
 497 * which has already covered that, and comment above remap_pages().
 498 */
 499struct anon_vma *folio_get_anon_vma(const struct folio *folio)
 500{
 501	struct anon_vma *anon_vma = NULL;
 502	unsigned long anon_mapping;
 503
 504	rcu_read_lock();
 505	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
 506	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 507		goto out;
 508	if (!folio_mapped(folio))
 509		goto out;
 510
 511	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 512	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 513		anon_vma = NULL;
 514		goto out;
 515	}
 516
 517	/*
 518	 * If this folio is still mapped, then its anon_vma cannot have been
 519	 * freed.  But if it has been unmapped, we have no security against the
 520	 * anon_vma structure being freed and reused (for another anon_vma:
 521	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
 522	 * above cannot corrupt).
 523	 */
 524	if (!folio_mapped(folio)) {
 525		rcu_read_unlock();
 526		put_anon_vma(anon_vma);
 527		return NULL;
 528	}
 529out:
 530	rcu_read_unlock();
 531
 532	return anon_vma;
 533}
 534
 535/*
 536 * Similar to folio_get_anon_vma() except it locks the anon_vma.
 537 *
 538 * Its a little more complex as it tries to keep the fast path to a single
 539 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
 540 * reference like with folio_get_anon_vma() and then block on the mutex
 541 * on !rwc->try_lock case.
 542 */
 543struct anon_vma *folio_lock_anon_vma_read(const struct folio *folio,
 544					  struct rmap_walk_control *rwc)
 545{
 546	struct anon_vma *anon_vma = NULL;
 547	struct anon_vma *root_anon_vma;
 548	unsigned long anon_mapping;
 549
 550retry:
 551	rcu_read_lock();
 552	anon_mapping = (unsigned long)READ_ONCE(folio->mapping);
 553	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
 554		goto out;
 555	if (!folio_mapped(folio))
 556		goto out;
 557
 558	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
 559	root_anon_vma = READ_ONCE(anon_vma->root);
 560	if (down_read_trylock(&root_anon_vma->rwsem)) {
 561		/*
 562		 * folio_move_anon_rmap() might have changed the anon_vma as we
 563		 * might not hold the folio lock here.
 564		 */
 565		if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
 566			     anon_mapping)) {
 567			up_read(&root_anon_vma->rwsem);
 568			rcu_read_unlock();
 569			goto retry;
 570		}
 571
 572		/*
 573		 * If the folio is still mapped, then this anon_vma is still
 574		 * its anon_vma, and holding the mutex ensures that it will
 575		 * not go away, see anon_vma_free().
 576		 */
 577		if (!folio_mapped(folio)) {
 578			up_read(&root_anon_vma->rwsem);
 579			anon_vma = NULL;
 580		}
 581		goto out;
 582	}
 583
 584	if (rwc && rwc->try_lock) {
 585		anon_vma = NULL;
 586		rwc->contended = true;
 587		goto out;
 588	}
 589
 590	/* trylock failed, we got to sleep */
 591	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
 592		anon_vma = NULL;
 593		goto out;
 594	}
 595
 596	if (!folio_mapped(folio)) {
 597		rcu_read_unlock();
 598		put_anon_vma(anon_vma);
 599		return NULL;
 600	}
 601
 602	/* we pinned the anon_vma, its safe to sleep */
 603	rcu_read_unlock();
 604	anon_vma_lock_read(anon_vma);
 605
 606	/*
 607	 * folio_move_anon_rmap() might have changed the anon_vma as we might
 608	 * not hold the folio lock here.
 609	 */
 610	if (unlikely((unsigned long)READ_ONCE(folio->mapping) !=
 611		     anon_mapping)) {
 612		anon_vma_unlock_read(anon_vma);
 613		put_anon_vma(anon_vma);
 614		anon_vma = NULL;
 615		goto retry;
 616	}
 617
 618	if (atomic_dec_and_test(&anon_vma->refcount)) {
 619		/*
 620		 * Oops, we held the last refcount, release the lock
 621		 * and bail -- can't simply use put_anon_vma() because
 622		 * we'll deadlock on the anon_vma_lock_write() recursion.
 623		 */
 624		anon_vma_unlock_read(anon_vma);
 625		__put_anon_vma(anon_vma);
 626		anon_vma = NULL;
 627	}
 628
 629	return anon_vma;
 630
 631out:
 632	rcu_read_unlock();
 633	return anon_vma;
 634}
 635
 636#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 637/*
 638 * Flush TLB entries for recently unmapped pages from remote CPUs. It is
 639 * important if a PTE was dirty when it was unmapped that it's flushed
 640 * before any IO is initiated on the page to prevent lost writes. Similarly,
 641 * it must be flushed before freeing to prevent data leakage.
 642 */
 643void try_to_unmap_flush(void)
 644{
 645	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 646
 647	if (!tlb_ubc->flush_required)
 648		return;
 649
 650	arch_tlbbatch_flush(&tlb_ubc->arch);
 651	tlb_ubc->flush_required = false;
 652	tlb_ubc->writable = false;
 653}
 654
 655/* Flush iff there are potentially writable TLB entries that can race with IO */
 656void try_to_unmap_flush_dirty(void)
 657{
 658	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 659
 660	if (tlb_ubc->writable)
 661		try_to_unmap_flush();
 662}
 663
 664/*
 665 * Bits 0-14 of mm->tlb_flush_batched record pending generations.
 666 * Bits 16-30 of mm->tlb_flush_batched bit record flushed generations.
 667 */
 668#define TLB_FLUSH_BATCH_FLUSHED_SHIFT	16
 669#define TLB_FLUSH_BATCH_PENDING_MASK			\
 670	((1 << (TLB_FLUSH_BATCH_FLUSHED_SHIFT - 1)) - 1)
 671#define TLB_FLUSH_BATCH_PENDING_LARGE			\
 672	(TLB_FLUSH_BATCH_PENDING_MASK / 2)
 673
 674static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
 675				      unsigned long uaddr)
 676{
 677	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
 678	int batch;
 679	bool writable = pte_dirty(pteval);
 680
 681	if (!pte_accessible(mm, pteval))
 682		return;
 683
 684	arch_tlbbatch_add_pending(&tlb_ubc->arch, mm, uaddr);
 685	tlb_ubc->flush_required = true;
 686
 687	/*
 688	 * Ensure compiler does not re-order the setting of tlb_flush_batched
 689	 * before the PTE is cleared.
 690	 */
 691	barrier();
 692	batch = atomic_read(&mm->tlb_flush_batched);
 693retry:
 694	if ((batch & TLB_FLUSH_BATCH_PENDING_MASK) > TLB_FLUSH_BATCH_PENDING_LARGE) {
 695		/*
 696		 * Prevent `pending' from catching up with `flushed' because of
 697		 * overflow.  Reset `pending' and `flushed' to be 1 and 0 if
 698		 * `pending' becomes large.
 699		 */
 700		if (!atomic_try_cmpxchg(&mm->tlb_flush_batched, &batch, 1))
 701			goto retry;
 702	} else {
 703		atomic_inc(&mm->tlb_flush_batched);
 704	}
 705
 706	/*
 707	 * If the PTE was dirty then it's best to assume it's writable. The
 708	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
 709	 * before the page is queued for IO.
 710	 */
 711	if (writable)
 712		tlb_ubc->writable = true;
 713}
 714
 715/*
 716 * Returns true if the TLB flush should be deferred to the end of a batch of
 717 * unmap operations to reduce IPIs.
 718 */
 719static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 720{
 721	if (!(flags & TTU_BATCH_FLUSH))
 722		return false;
 723
 724	return arch_tlbbatch_should_defer(mm);
 725}
 726
 727/*
 728 * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
 729 * releasing the PTL if TLB flushes are batched. It's possible for a parallel
 730 * operation such as mprotect or munmap to race between reclaim unmapping
 731 * the page and flushing the page. If this race occurs, it potentially allows
 732 * access to data via a stale TLB entry. Tracking all mm's that have TLB
 733 * batching in flight would be expensive during reclaim so instead track
 734 * whether TLB batching occurred in the past and if so then do a flush here
 735 * if required. This will cost one additional flush per reclaim cycle paid
 736 * by the first operation at risk such as mprotect and mumap.
 737 *
 738 * This must be called under the PTL so that an access to tlb_flush_batched
 739 * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
 740 * via the PTL.
 741 */
 742void flush_tlb_batched_pending(struct mm_struct *mm)
 743{
 744	int batch = atomic_read(&mm->tlb_flush_batched);
 745	int pending = batch & TLB_FLUSH_BATCH_PENDING_MASK;
 746	int flushed = batch >> TLB_FLUSH_BATCH_FLUSHED_SHIFT;
 747
 748	if (pending != flushed) {
 749		arch_flush_tlb_batched_pending(mm);
 750		/*
 751		 * If the new TLB flushing is pending during flushing, leave
 752		 * mm->tlb_flush_batched as is, to avoid losing flushing.
 753		 */
 754		atomic_cmpxchg(&mm->tlb_flush_batched, batch,
 755			       pending | (pending << TLB_FLUSH_BATCH_FLUSHED_SHIFT));
 756	}
 757}
 758#else
 759static void set_tlb_ubc_flush_pending(struct mm_struct *mm, pte_t pteval,
 760				      unsigned long uaddr)
 761{
 762}
 763
 764static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 765{
 766	return false;
 767}
 768#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
 769
 770/**
 771 * page_address_in_vma - The virtual address of a page in this VMA.
 772 * @folio: The folio containing the page.
 773 * @page: The page within the folio.
 774 * @vma: The VMA we need to know the address in.
 775 *
 776 * Calculates the user virtual address of this page in the specified VMA.
 777 * It is the caller's responsibililty to check the page is actually
 778 * within the VMA.  There may not currently be a PTE pointing at this
 779 * page, but if a page fault occurs at this address, this is the page
 780 * which will be accessed.
 781 *
 782 * Context: Caller should hold a reference to the folio.  Caller should
 783 * hold a lock (eg the i_mmap_lock or the mmap_lock) which keeps the
 784 * VMA from being altered.
 785 *
 786 * Return: The virtual address corresponding to this page in the VMA.
 787 */
 788unsigned long page_address_in_vma(const struct folio *folio,
 789		const struct page *page, const struct vm_area_struct *vma)
 790{
 
 791	if (folio_test_anon(folio)) {
 792		struct anon_vma *page__anon_vma = folio_anon_vma(folio);
 793		/*
 794		 * Note: swapoff's unuse_vma() is more efficient with this
 795		 * check, and needs it to match anon_vma when KSM is active.
 796		 */
 797		if (!vma->anon_vma || !page__anon_vma ||
 798		    vma->anon_vma->root != page__anon_vma->root)
 799			return -EFAULT;
 800	} else if (!vma->vm_file) {
 801		return -EFAULT;
 802	} else if (vma->vm_file->f_mapping != folio->mapping) {
 803		return -EFAULT;
 804	}
 805
 806	/* KSM folios don't reach here because of the !page__anon_vma check */
 807	return vma_address(vma, page_pgoff(folio, page), 1);
 808}
 809
 810/*
 811 * Returns the actual pmd_t* where we expect 'address' to be mapped from, or
 812 * NULL if it doesn't exist.  No guarantees / checks on what the pmd_t*
 813 * represents.
 814 */
 815pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
 816{
 817	pgd_t *pgd;
 818	p4d_t *p4d;
 819	pud_t *pud;
 820	pmd_t *pmd = NULL;
 821
 822	pgd = pgd_offset(mm, address);
 823	if (!pgd_present(*pgd))
 824		goto out;
 825
 826	p4d = p4d_offset(pgd, address);
 827	if (!p4d_present(*p4d))
 828		goto out;
 829
 830	pud = pud_offset(p4d, address);
 831	if (!pud_present(*pud))
 832		goto out;
 833
 834	pmd = pmd_offset(pud, address);
 835out:
 836	return pmd;
 837}
 838
 839struct folio_referenced_arg {
 840	int mapcount;
 841	int referenced;
 842	unsigned long vm_flags;
 843	struct mem_cgroup *memcg;
 844};
 845
 846/*
 847 * arg: folio_referenced_arg will be passed
 848 */
 849static bool folio_referenced_one(struct folio *folio,
 850		struct vm_area_struct *vma, unsigned long address, void *arg)
 851{
 852	struct folio_referenced_arg *pra = arg;
 853	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
 854	int referenced = 0;
 855	unsigned long start = address, ptes = 0;
 856
 857	while (page_vma_mapped_walk(&pvmw)) {
 858		address = pvmw.address;
 859
 860		if (vma->vm_flags & VM_LOCKED) {
 861			if (!folio_test_large(folio) || !pvmw.pte) {
 862				/* Restore the mlock which got missed */
 863				mlock_vma_folio(folio, vma);
 864				page_vma_mapped_walk_done(&pvmw);
 865				pra->vm_flags |= VM_LOCKED;
 866				return false; /* To break the loop */
 867			}
 868			/*
 869			 * For large folio fully mapped to VMA, will
 870			 * be handled after the pvmw loop.
 871			 *
 872			 * For large folio cross VMA boundaries, it's
 873			 * expected to be picked  by page reclaim. But
 874			 * should skip reference of pages which are in
 875			 * the range of VM_LOCKED vma. As page reclaim
 876			 * should just count the reference of pages out
 877			 * the range of VM_LOCKED vma.
 878			 */
 879			ptes++;
 880			pra->mapcount--;
 881			continue;
 882		}
 883
 884		/*
 885		 * Skip the non-shared swapbacked folio mapped solely by
 886		 * the exiting or OOM-reaped process. This avoids redundant
 887		 * swap-out followed by an immediate unmap.
 888		 */
 889		if ((!atomic_read(&vma->vm_mm->mm_users) ||
 890		    check_stable_address_space(vma->vm_mm)) &&
 891		    folio_test_anon(folio) && folio_test_swapbacked(folio) &&
 892		    !folio_likely_mapped_shared(folio)) {
 893			pra->referenced = -1;
 894			page_vma_mapped_walk_done(&pvmw);
 895			return false;
 896		}
 897
 898		if (lru_gen_enabled() && pvmw.pte) {
 899			if (lru_gen_look_around(&pvmw))
 900				referenced++;
 901		} else if (pvmw.pte) {
 
 902			if (ptep_clear_flush_young_notify(vma, address,
 903						pvmw.pte))
 904				referenced++;
 905		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 906			if (pmdp_clear_flush_young_notify(vma, address,
 907						pvmw.pmd))
 908				referenced++;
 909		} else {
 910			/* unexpected pmd-mapped folio? */
 911			WARN_ON_ONCE(1);
 912		}
 913
 914		pra->mapcount--;
 915	}
 916
 917	if ((vma->vm_flags & VM_LOCKED) &&
 918			folio_test_large(folio) &&
 919			folio_within_vma(folio, vma)) {
 920		unsigned long s_align, e_align;
 921
 922		s_align = ALIGN_DOWN(start, PMD_SIZE);
 923		e_align = ALIGN_DOWN(start + folio_size(folio) - 1, PMD_SIZE);
 924
 925		/* folio doesn't cross page table boundary and fully mapped */
 926		if ((s_align == e_align) && (ptes == folio_nr_pages(folio))) {
 927			/* Restore the mlock which got missed */
 928			mlock_vma_folio(folio, vma);
 929			pra->vm_flags |= VM_LOCKED;
 930			return false; /* To break the loop */
 931		}
 932	}
 933
 934	if (referenced)
 935		folio_clear_idle(folio);
 936	if (folio_test_clear_young(folio))
 937		referenced++;
 938
 939	if (referenced) {
 940		pra->referenced++;
 941		pra->vm_flags |= vma->vm_flags & ~VM_LOCKED;
 942	}
 943
 944	if (!pra->mapcount)
 945		return false; /* To break the loop */
 946
 947	return true;
 948}
 949
 950static bool invalid_folio_referenced_vma(struct vm_area_struct *vma, void *arg)
 951{
 952	struct folio_referenced_arg *pra = arg;
 953	struct mem_cgroup *memcg = pra->memcg;
 954
 955	/*
 956	 * Ignore references from this mapping if it has no recency. If the
 957	 * folio has been used in another mapping, we will catch it; if this
 958	 * other mapping is already gone, the unmap path will have set the
 959	 * referenced flag or activated the folio in zap_pte_range().
 960	 */
 961	if (!vma_has_recency(vma))
 962		return true;
 963
 964	/*
 965	 * If we are reclaiming on behalf of a cgroup, skip counting on behalf
 966	 * of references from different cgroups.
 967	 */
 968	if (memcg && !mm_match_cgroup(vma->vm_mm, memcg))
 969		return true;
 970
 971	return false;
 972}
 973
 974/**
 975 * folio_referenced() - Test if the folio was referenced.
 976 * @folio: The folio to test.
 977 * @is_locked: Caller holds lock on the folio.
 978 * @memcg: target memory cgroup
 979 * @vm_flags: A combination of all the vma->vm_flags which referenced the folio.
 980 *
 981 * Quick test_and_clear_referenced for all mappings of a folio,
 982 *
 983 * Return: The number of mappings which referenced the folio. Return -1 if
 984 * the function bailed out due to rmap lock contention.
 985 */
 986int folio_referenced(struct folio *folio, int is_locked,
 987		     struct mem_cgroup *memcg, unsigned long *vm_flags)
 988{
 989	bool we_locked = false;
 990	struct folio_referenced_arg pra = {
 991		.mapcount = folio_mapcount(folio),
 992		.memcg = memcg,
 993	};
 994	struct rmap_walk_control rwc = {
 995		.rmap_one = folio_referenced_one,
 996		.arg = (void *)&pra,
 997		.anon_lock = folio_lock_anon_vma_read,
 998		.try_lock = true,
 999		.invalid_vma = invalid_folio_referenced_vma,
1000	};
1001
1002	*vm_flags = 0;
1003	if (!pra.mapcount)
1004		return 0;
1005
1006	if (!folio_raw_mapping(folio))
1007		return 0;
1008
1009	if (!is_locked && (!folio_test_anon(folio) || folio_test_ksm(folio))) {
1010		we_locked = folio_trylock(folio);
1011		if (!we_locked)
1012			return 1;
1013	}
1014
1015	rmap_walk(folio, &rwc);
1016	*vm_flags = pra.vm_flags;
1017
1018	if (we_locked)
1019		folio_unlock(folio);
1020
1021	return rwc.contended ? -1 : pra.referenced;
1022}
1023
1024static int page_vma_mkclean_one(struct page_vma_mapped_walk *pvmw)
1025{
1026	int cleaned = 0;
1027	struct vm_area_struct *vma = pvmw->vma;
1028	struct mmu_notifier_range range;
1029	unsigned long address = pvmw->address;
1030
1031	/*
1032	 * We have to assume the worse case ie pmd for invalidation. Note that
1033	 * the folio can not be freed from this function.
1034	 */
1035	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE, 0,
1036				vma->vm_mm, address, vma_address_end(pvmw));
1037	mmu_notifier_invalidate_range_start(&range);
1038
1039	while (page_vma_mapped_walk(pvmw)) {
1040		int ret = 0;
1041
1042		address = pvmw->address;
1043		if (pvmw->pte) {
1044			pte_t *pte = pvmw->pte;
1045			pte_t entry = ptep_get(pte);
1046
1047			if (!pte_dirty(entry) && !pte_write(entry))
1048				continue;
1049
1050			flush_cache_page(vma, address, pte_pfn(entry));
1051			entry = ptep_clear_flush(vma, address, pte);
1052			entry = pte_wrprotect(entry);
1053			entry = pte_mkclean(entry);
1054			set_pte_at(vma->vm_mm, address, pte, entry);
1055			ret = 1;
1056		} else {
1057#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1058			pmd_t *pmd = pvmw->pmd;
1059			pmd_t entry;
1060
1061			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
1062				continue;
1063
1064			flush_cache_range(vma, address,
1065					  address + HPAGE_PMD_SIZE);
1066			entry = pmdp_invalidate(vma, address, pmd);
1067			entry = pmd_wrprotect(entry);
1068			entry = pmd_mkclean(entry);
1069			set_pmd_at(vma->vm_mm, address, pmd, entry);
1070			ret = 1;
1071#else
1072			/* unexpected pmd-mapped folio? */
1073			WARN_ON_ONCE(1);
1074#endif
1075		}
1076
1077		if (ret)
1078			cleaned++;
1079	}
1080
1081	mmu_notifier_invalidate_range_end(&range);
1082
1083	return cleaned;
1084}
1085
1086static bool page_mkclean_one(struct folio *folio, struct vm_area_struct *vma,
1087			     unsigned long address, void *arg)
1088{
1089	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
1090	int *cleaned = arg;
1091
1092	*cleaned += page_vma_mkclean_one(&pvmw);
1093
1094	return true;
1095}
1096
1097static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1098{
1099	if (vma->vm_flags & VM_SHARED)
1100		return false;
1101
1102	return true;
1103}
1104
1105int folio_mkclean(struct folio *folio)
1106{
1107	int cleaned = 0;
1108	struct address_space *mapping;
1109	struct rmap_walk_control rwc = {
1110		.arg = (void *)&cleaned,
1111		.rmap_one = page_mkclean_one,
1112		.invalid_vma = invalid_mkclean_vma,
1113	};
1114
1115	BUG_ON(!folio_test_locked(folio));
1116
1117	if (!folio_mapped(folio))
1118		return 0;
1119
1120	mapping = folio_mapping(folio);
1121	if (!mapping)
1122		return 0;
1123
1124	rmap_walk(folio, &rwc);
1125
1126	return cleaned;
1127}
1128EXPORT_SYMBOL_GPL(folio_mkclean);
1129
1130/**
1131 * pfn_mkclean_range - Cleans the PTEs (including PMDs) mapped with range of
1132 *                     [@pfn, @pfn + @nr_pages) at the specific offset (@pgoff)
1133 *                     within the @vma of shared mappings. And since clean PTEs
1134 *                     should also be readonly, write protects them too.
1135 * @pfn: start pfn.
1136 * @nr_pages: number of physically contiguous pages srarting with @pfn.
1137 * @pgoff: page offset that the @pfn mapped with.
1138 * @vma: vma that @pfn mapped within.
1139 *
1140 * Returns the number of cleaned PTEs (including PMDs).
1141 */
1142int pfn_mkclean_range(unsigned long pfn, unsigned long nr_pages, pgoff_t pgoff,
1143		      struct vm_area_struct *vma)
1144{
1145	struct page_vma_mapped_walk pvmw = {
1146		.pfn		= pfn,
1147		.nr_pages	= nr_pages,
1148		.pgoff		= pgoff,
1149		.vma		= vma,
1150		.flags		= PVMW_SYNC,
1151	};
1152
1153	if (invalid_mkclean_vma(vma, NULL))
1154		return 0;
1155
1156	pvmw.address = vma_address(vma, pgoff, nr_pages);
1157	VM_BUG_ON_VMA(pvmw.address == -EFAULT, vma);
1158
1159	return page_vma_mkclean_one(&pvmw);
1160}
1161
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1162static __always_inline unsigned int __folio_add_rmap(struct folio *folio,
1163		struct page *page, int nr_pages, enum rmap_level level,
1164		int *nr_pmdmapped)
1165{
1166	atomic_t *mapped = &folio->_nr_pages_mapped;
1167	const int orig_nr_pages = nr_pages;
1168	int first = 0, nr = 0;
1169
1170	__folio_rmap_sanity_checks(folio, page, nr_pages, level);
1171
1172	switch (level) {
1173	case RMAP_LEVEL_PTE:
1174		if (!folio_test_large(folio)) {
1175			nr = atomic_inc_and_test(&folio->_mapcount);
1176			break;
1177		}
1178
1179		do {
1180			first += atomic_inc_and_test(&page->_mapcount);
1181		} while (page++, --nr_pages > 0);
1182
1183		if (first &&
1184		    atomic_add_return_relaxed(first, mapped) < ENTIRELY_MAPPED)
1185			nr = first;
1186
1187		atomic_add(orig_nr_pages, &folio->_large_mapcount);
 
 
1188		break;
1189	case RMAP_LEVEL_PMD:
1190		first = atomic_inc_and_test(&folio->_entire_mapcount);
1191		if (first) {
1192			nr = atomic_add_return_relaxed(ENTIRELY_MAPPED, mapped);
1193			if (likely(nr < ENTIRELY_MAPPED + ENTIRELY_MAPPED)) {
1194				*nr_pmdmapped = folio_nr_pages(folio);
1195				nr = *nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1196				/* Raced ahead of a remove and another add? */
1197				if (unlikely(nr < 0))
1198					nr = 0;
1199			} else {
1200				/* Raced ahead of a remove of ENTIRELY_MAPPED */
1201				nr = 0;
1202			}
1203		}
1204		atomic_inc(&folio->_large_mapcount);
1205		break;
1206	}
1207	return nr;
1208}
1209
1210/**
1211 * folio_move_anon_rmap - move a folio to our anon_vma
1212 * @folio:	The folio to move to our anon_vma
1213 * @vma:	The vma the folio belongs to
1214 *
1215 * When a folio belongs exclusively to one process after a COW event,
1216 * that folio can be moved into the anon_vma that belongs to just that
1217 * process, so the rmap code will not search the parent or sibling processes.
1218 */
1219void folio_move_anon_rmap(struct folio *folio, struct vm_area_struct *vma)
1220{
1221	void *anon_vma = vma->anon_vma;
1222
1223	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1224	VM_BUG_ON_VMA(!anon_vma, vma);
1225
1226	anon_vma += PAGE_MAPPING_ANON;
1227	/*
1228	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1229	 * simultaneously, so a concurrent reader (eg folio_referenced()'s
1230	 * folio_test_anon()) will not see one without the other.
1231	 */
1232	WRITE_ONCE(folio->mapping, anon_vma);
1233}
1234
1235/**
1236 * __folio_set_anon - set up a new anonymous rmap for a folio
1237 * @folio:	The folio to set up the new anonymous rmap for.
1238 * @vma:	VM area to add the folio to.
1239 * @address:	User virtual address of the mapping
1240 * @exclusive:	Whether the folio is exclusive to the process.
1241 */
1242static void __folio_set_anon(struct folio *folio, struct vm_area_struct *vma,
1243			     unsigned long address, bool exclusive)
1244{
1245	struct anon_vma *anon_vma = vma->anon_vma;
1246
1247	BUG_ON(!anon_vma);
1248
1249	/*
1250	 * If the folio isn't exclusive to this vma, we must use the _oldest_
1251	 * possible anon_vma for the folio mapping!
1252	 */
1253	if (!exclusive)
1254		anon_vma = anon_vma->root;
1255
1256	/*
1257	 * page_idle does a lockless/optimistic rmap scan on folio->mapping.
1258	 * Make sure the compiler doesn't split the stores of anon_vma and
1259	 * the PAGE_MAPPING_ANON type identifier, otherwise the rmap code
1260	 * could mistake the mapping for a struct address_space and crash.
1261	 */
1262	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1263	WRITE_ONCE(folio->mapping, (struct address_space *) anon_vma);
1264	folio->index = linear_page_index(vma, address);
1265}
1266
1267/**
1268 * __page_check_anon_rmap - sanity check anonymous rmap addition
1269 * @folio:	The folio containing @page.
1270 * @page:	the page to check the mapping of
1271 * @vma:	the vm area in which the mapping is added
1272 * @address:	the user virtual address mapped
1273 */
1274static void __page_check_anon_rmap(const struct folio *folio,
1275		const struct page *page, struct vm_area_struct *vma,
1276		unsigned long address)
1277{
1278	/*
1279	 * The page's anon-rmap details (mapping and index) are guaranteed to
1280	 * be set up correctly at this point.
1281	 *
1282	 * We have exclusion against folio_add_anon_rmap_*() because the caller
1283	 * always holds the page locked.
1284	 *
1285	 * We have exclusion against folio_add_new_anon_rmap because those pages
1286	 * are initially only visible via the pagetables, and the pte is locked
1287	 * over the call to folio_add_new_anon_rmap.
1288	 */
1289	VM_BUG_ON_FOLIO(folio_anon_vma(folio)->root != vma->anon_vma->root,
1290			folio);
1291	VM_BUG_ON_PAGE(page_pgoff(folio, page) != linear_page_index(vma, address),
1292		       page);
1293}
1294
1295static void __folio_mod_stat(struct folio *folio, int nr, int nr_pmdmapped)
1296{
1297	int idx;
1298
1299	if (nr) {
1300		idx = folio_test_anon(folio) ? NR_ANON_MAPPED : NR_FILE_MAPPED;
1301		__lruvec_stat_mod_folio(folio, idx, nr);
1302	}
1303	if (nr_pmdmapped) {
1304		if (folio_test_anon(folio)) {
1305			idx = NR_ANON_THPS;
1306			__lruvec_stat_mod_folio(folio, idx, nr_pmdmapped);
1307		} else {
1308			/* NR_*_PMDMAPPED are not maintained per-memcg */
1309			idx = folio_test_swapbacked(folio) ?
1310				NR_SHMEM_PMDMAPPED : NR_FILE_PMDMAPPED;
1311			__mod_node_page_state(folio_pgdat(folio), idx,
1312					      nr_pmdmapped);
1313		}
1314	}
1315}
1316
1317static __always_inline void __folio_add_anon_rmap(struct folio *folio,
1318		struct page *page, int nr_pages, struct vm_area_struct *vma,
1319		unsigned long address, rmap_t flags, enum rmap_level level)
1320{
1321	int i, nr, nr_pmdmapped = 0;
1322
1323	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
1324
1325	nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
 
 
 
 
1326
1327	if (likely(!folio_test_ksm(folio)))
 
 
 
 
 
 
 
 
 
 
 
 
1328		__page_check_anon_rmap(folio, page, vma, address);
1329
1330	__folio_mod_stat(folio, nr, nr_pmdmapped);
1331
1332	if (flags & RMAP_EXCLUSIVE) {
1333		switch (level) {
1334		case RMAP_LEVEL_PTE:
1335			for (i = 0; i < nr_pages; i++)
1336				SetPageAnonExclusive(page + i);
1337			break;
1338		case RMAP_LEVEL_PMD:
1339			SetPageAnonExclusive(page);
1340			break;
1341		}
1342	}
1343	for (i = 0; i < nr_pages; i++) {
1344		struct page *cur_page = page + i;
1345
1346		/* While PTE-mapping a THP we have a PMD and a PTE mapping. */
1347		VM_WARN_ON_FOLIO((atomic_read(&cur_page->_mapcount) > 0 ||
1348				  (folio_test_large(folio) &&
1349				   folio_entire_mapcount(folio) > 1)) &&
1350				 PageAnonExclusive(cur_page), folio);
1351	}
1352
1353	/*
1354	 * For large folio, only mlock it if it's fully mapped to VMA. It's
1355	 * not easy to check whether the large folio is fully mapped to VMA
1356	 * here. Only mlock normal 4K folio and leave page reclaim to handle
1357	 * large folio.
1358	 */
1359	if (!folio_test_large(folio))
1360		mlock_vma_folio(folio, vma);
1361}
1362
1363/**
1364 * folio_add_anon_rmap_ptes - add PTE mappings to a page range of an anon folio
1365 * @folio:	The folio to add the mappings to
1366 * @page:	The first page to add
1367 * @nr_pages:	The number of pages which will be mapped
1368 * @vma:	The vm area in which the mappings are added
1369 * @address:	The user virtual address of the first page to map
1370 * @flags:	The rmap flags
1371 *
1372 * The page range of folio is defined by [first_page, first_page + nr_pages)
1373 *
1374 * The caller needs to hold the page table lock, and the page must be locked in
1375 * the anon_vma case: to serialize mapping,index checking after setting,
1376 * and to ensure that an anon folio is not being upgraded racily to a KSM folio
1377 * (but KSM folios are never downgraded).
1378 */
1379void folio_add_anon_rmap_ptes(struct folio *folio, struct page *page,
1380		int nr_pages, struct vm_area_struct *vma, unsigned long address,
1381		rmap_t flags)
1382{
1383	__folio_add_anon_rmap(folio, page, nr_pages, vma, address, flags,
1384			      RMAP_LEVEL_PTE);
1385}
1386
1387/**
1388 * folio_add_anon_rmap_pmd - add a PMD mapping to a page range of an anon folio
1389 * @folio:	The folio to add the mapping to
1390 * @page:	The first page to add
1391 * @vma:	The vm area in which the mapping is added
1392 * @address:	The user virtual address of the first page to map
1393 * @flags:	The rmap flags
1394 *
1395 * The page range of folio is defined by [first_page, first_page + HPAGE_PMD_NR)
1396 *
1397 * The caller needs to hold the page table lock, and the page must be locked in
1398 * the anon_vma case: to serialize mapping,index checking after setting.
1399 */
1400void folio_add_anon_rmap_pmd(struct folio *folio, struct page *page,
1401		struct vm_area_struct *vma, unsigned long address, rmap_t flags)
1402{
1403#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1404	__folio_add_anon_rmap(folio, page, HPAGE_PMD_NR, vma, address, flags,
1405			      RMAP_LEVEL_PMD);
1406#else
1407	WARN_ON_ONCE(true);
1408#endif
1409}
1410
1411/**
1412 * folio_add_new_anon_rmap - Add mapping to a new anonymous folio.
1413 * @folio:	The folio to add the mapping to.
1414 * @vma:	the vm area in which the mapping is added
1415 * @address:	the user virtual address mapped
1416 * @flags:	The rmap flags
1417 *
1418 * Like folio_add_anon_rmap_*() but must only be called on *new* folios.
1419 * This means the inc-and-test can be bypassed.
1420 * The folio doesn't necessarily need to be locked while it's exclusive
1421 * unless two threads map it concurrently. However, the folio must be
1422 * locked if it's shared.
1423 *
1424 * If the folio is pmd-mappable, it is accounted as a THP.
 
1425 */
1426void folio_add_new_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
1427		unsigned long address, rmap_t flags)
1428{
1429	const int nr = folio_nr_pages(folio);
1430	const bool exclusive = flags & RMAP_EXCLUSIVE;
1431	int nr_pmdmapped = 0;
1432
1433	VM_WARN_ON_FOLIO(folio_test_hugetlb(folio), folio);
1434	VM_WARN_ON_FOLIO(!exclusive && !folio_test_locked(folio), folio);
1435	VM_BUG_ON_VMA(address < vma->vm_start ||
1436			address + (nr << PAGE_SHIFT) > vma->vm_end, vma);
1437
1438	/*
1439	 * VM_DROPPABLE mappings don't swap; instead they're just dropped when
1440	 * under memory pressure.
1441	 */
1442	if (!folio_test_swapbacked(folio) && !(vma->vm_flags & VM_DROPPABLE))
1443		__folio_set_swapbacked(folio);
1444	__folio_set_anon(folio, vma, address, exclusive);
1445
1446	if (likely(!folio_test_large(folio))) {
1447		/* increment count (starts at -1) */
1448		atomic_set(&folio->_mapcount, 0);
1449		if (exclusive)
1450			SetPageAnonExclusive(&folio->page);
1451	} else if (!folio_test_pmd_mappable(folio)) {
1452		int i;
1453
1454		for (i = 0; i < nr; i++) {
1455			struct page *page = folio_page(folio, i);
1456
1457			/* increment count (starts at -1) */
1458			atomic_set(&page->_mapcount, 0);
1459			if (exclusive)
1460				SetPageAnonExclusive(page);
1461		}
1462
1463		/* increment count (starts at -1) */
1464		atomic_set(&folio->_large_mapcount, nr - 1);
1465		atomic_set(&folio->_nr_pages_mapped, nr);
1466	} else {
1467		/* increment count (starts at -1) */
1468		atomic_set(&folio->_entire_mapcount, 0);
1469		/* increment count (starts at -1) */
1470		atomic_set(&folio->_large_mapcount, 0);
1471		atomic_set(&folio->_nr_pages_mapped, ENTIRELY_MAPPED);
1472		if (exclusive)
1473			SetPageAnonExclusive(&folio->page);
1474		nr_pmdmapped = nr;
1475	}
1476
1477	__folio_mod_stat(folio, nr, nr_pmdmapped);
1478	mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON, 1);
1479}
1480
1481static __always_inline void __folio_add_file_rmap(struct folio *folio,
1482		struct page *page, int nr_pages, struct vm_area_struct *vma,
1483		enum rmap_level level)
1484{
1485	int nr, nr_pmdmapped = 0;
1486
1487	VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
1488
1489	nr = __folio_add_rmap(folio, page, nr_pages, level, &nr_pmdmapped);
1490	__folio_mod_stat(folio, nr, nr_pmdmapped);
 
 
 
 
1491
1492	/* See comments in folio_add_anon_rmap_*() */
1493	if (!folio_test_large(folio))
1494		mlock_vma_folio(folio, vma);
1495}
1496
1497/**
1498 * folio_add_file_rmap_ptes - add PTE mappings to a page range of a folio
1499 * @folio:	The folio to add the mappings to
1500 * @page:	The first page to add
1501 * @nr_pages:	The number of pages that will be mapped using PTEs
1502 * @vma:	The vm area in which the mappings are added
1503 *
1504 * The page range of the folio is defined by [page, page + nr_pages)
1505 *
1506 * The caller needs to hold the page table lock.
1507 */
1508void folio_add_file_rmap_ptes(struct folio *folio, struct page *page,
1509		int nr_pages, struct vm_area_struct *vma)
1510{
1511	__folio_add_file_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1512}
1513
1514/**
1515 * folio_add_file_rmap_pmd - add a PMD mapping to a page range of a folio
1516 * @folio:	The folio to add the mapping to
1517 * @page:	The first page to add
1518 * @vma:	The vm area in which the mapping is added
1519 *
1520 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1521 *
1522 * The caller needs to hold the page table lock.
1523 */
1524void folio_add_file_rmap_pmd(struct folio *folio, struct page *page,
1525		struct vm_area_struct *vma)
1526{
1527#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1528	__folio_add_file_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1529#else
1530	WARN_ON_ONCE(true);
1531#endif
1532}
1533
1534static __always_inline void __folio_remove_rmap(struct folio *folio,
1535		struct page *page, int nr_pages, struct vm_area_struct *vma,
1536		enum rmap_level level)
1537{
1538	atomic_t *mapped = &folio->_nr_pages_mapped;
1539	int last = 0, nr = 0, nr_pmdmapped = 0;
1540	bool partially_mapped = false;
1541
1542	__folio_rmap_sanity_checks(folio, page, nr_pages, level);
1543
1544	switch (level) {
1545	case RMAP_LEVEL_PTE:
1546		if (!folio_test_large(folio)) {
1547			nr = atomic_add_negative(-1, &folio->_mapcount);
1548			break;
1549		}
1550
1551		atomic_sub(nr_pages, &folio->_large_mapcount);
1552		do {
1553			last += atomic_add_negative(-1, &page->_mapcount);
1554		} while (page++, --nr_pages > 0);
1555
1556		if (last &&
1557		    atomic_sub_return_relaxed(last, mapped) < ENTIRELY_MAPPED)
1558			nr = last;
1559
1560		partially_mapped = nr && atomic_read(mapped);
 
 
1561		break;
1562	case RMAP_LEVEL_PMD:
1563		atomic_dec(&folio->_large_mapcount);
1564		last = atomic_add_negative(-1, &folio->_entire_mapcount);
1565		if (last) {
1566			nr = atomic_sub_return_relaxed(ENTIRELY_MAPPED, mapped);
1567			if (likely(nr < ENTIRELY_MAPPED)) {
1568				nr_pmdmapped = folio_nr_pages(folio);
1569				nr = nr_pmdmapped - (nr & FOLIO_PAGES_MAPPED);
1570				/* Raced ahead of another remove and an add? */
1571				if (unlikely(nr < 0))
1572					nr = 0;
1573			} else {
1574				/* An add of ENTIRELY_MAPPED raced ahead */
1575				nr = 0;
1576			}
1577		}
1578
1579		partially_mapped = nr && nr < nr_pmdmapped;
1580		break;
1581	}
1582
1583	/*
1584	 * Queue anon large folio for deferred split if at least one page of
1585	 * the folio is unmapped and at least one page is still mapped.
1586	 *
1587	 * Check partially_mapped first to ensure it is a large folio.
1588	 */
1589	if (partially_mapped && folio_test_anon(folio) &&
1590	    !folio_test_partially_mapped(folio))
1591		deferred_split_folio(folio, true);
 
 
 
1592
1593	__folio_mod_stat(folio, -nr, -nr_pmdmapped);
 
 
 
 
 
 
 
 
1594
1595	/*
1596	 * It would be tidy to reset folio_test_anon mapping when fully
1597	 * unmapped, but that might overwrite a racing folio_add_anon_rmap_*()
1598	 * which increments mapcount after us but sets mapping before us:
1599	 * so leave the reset to free_pages_prepare, and remember that
1600	 * it's only reliable while mapped.
1601	 */
1602
1603	munlock_vma_folio(folio, vma);
1604}
1605
1606/**
1607 * folio_remove_rmap_ptes - remove PTE mappings from a page range of a folio
1608 * @folio:	The folio to remove the mappings from
1609 * @page:	The first page to remove
1610 * @nr_pages:	The number of pages that will be removed from the mapping
1611 * @vma:	The vm area from which the mappings are removed
1612 *
1613 * The page range of the folio is defined by [page, page + nr_pages)
1614 *
1615 * The caller needs to hold the page table lock.
1616 */
1617void folio_remove_rmap_ptes(struct folio *folio, struct page *page,
1618		int nr_pages, struct vm_area_struct *vma)
1619{
1620	__folio_remove_rmap(folio, page, nr_pages, vma, RMAP_LEVEL_PTE);
1621}
1622
1623/**
1624 * folio_remove_rmap_pmd - remove a PMD mapping from a page range of a folio
1625 * @folio:	The folio to remove the mapping from
1626 * @page:	The first page to remove
1627 * @vma:	The vm area from which the mapping is removed
1628 *
1629 * The page range of the folio is defined by [page, page + HPAGE_PMD_NR)
1630 *
1631 * The caller needs to hold the page table lock.
1632 */
1633void folio_remove_rmap_pmd(struct folio *folio, struct page *page,
1634		struct vm_area_struct *vma)
1635{
1636#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1637	__folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, RMAP_LEVEL_PMD);
1638#else
1639	WARN_ON_ONCE(true);
1640#endif
1641}
1642
1643/*
1644 * @arg: enum ttu_flags will be passed to this argument
1645 */
1646static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
1647		     unsigned long address, void *arg)
1648{
1649	struct mm_struct *mm = vma->vm_mm;
1650	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
1651	pte_t pteval;
1652	struct page *subpage;
1653	bool anon_exclusive, ret = true;
1654	struct mmu_notifier_range range;
1655	enum ttu_flags flags = (enum ttu_flags)(long)arg;
1656	unsigned long pfn;
1657	unsigned long hsz = 0;
1658
1659	/*
1660	 * When racing against e.g. zap_pte_range() on another cpu,
1661	 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
1662	 * try_to_unmap() may return before page_mapped() has become false,
1663	 * if page table locking is skipped: use TTU_SYNC to wait for that.
1664	 */
1665	if (flags & TTU_SYNC)
1666		pvmw.flags = PVMW_SYNC;
1667
 
 
 
1668	/*
1669	 * For THP, we have to assume the worse case ie pmd for invalidation.
1670	 * For hugetlb, it could be much worse if we need to do pud
1671	 * invalidation in the case of pmd sharing.
1672	 *
1673	 * Note that the folio can not be freed in this function as call of
1674	 * try_to_unmap() must hold a reference on the folio.
1675	 */
1676	range.end = vma_address_end(&pvmw);
1677	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
1678				address, range.end);
1679	if (folio_test_hugetlb(folio)) {
1680		/*
1681		 * If sharing is possible, start and end will be adjusted
1682		 * accordingly.
1683		 */
1684		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1685						     &range.end);
1686
1687		/* We need the huge page size for set_huge_pte_at() */
1688		hsz = huge_page_size(hstate_vma(vma));
1689	}
1690	mmu_notifier_invalidate_range_start(&range);
1691
1692	while (page_vma_mapped_walk(&pvmw)) {
 
 
 
1693		/*
1694		 * If the folio is in an mlock()d vma, we must not swap it out.
1695		 */
1696		if (!(flags & TTU_IGNORE_MLOCK) &&
1697		    (vma->vm_flags & VM_LOCKED)) {
1698			/* Restore the mlock which got missed */
1699			if (!folio_test_large(folio))
1700				mlock_vma_folio(folio, vma);
1701			goto walk_abort;
1702		}
1703
1704		if (!pvmw.pte) {
1705			if (unmap_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
1706						  folio))
1707				goto walk_done;
1708
1709			if (flags & TTU_SPLIT_HUGE_PMD) {
1710				/*
1711				 * We temporarily have to drop the PTL and
1712				 * restart so we can process the PTE-mapped THP.
1713				 */
1714				split_huge_pmd_locked(vma, pvmw.address,
1715						      pvmw.pmd, false, folio);
1716				flags &= ~TTU_SPLIT_HUGE_PMD;
1717				page_vma_mapped_walk_restart(&pvmw);
1718				continue;
1719			}
1720		}
1721
1722		/* Unexpected PMD-mapped THP? */
1723		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
1724
1725		pfn = pte_pfn(ptep_get(pvmw.pte));
1726		subpage = folio_page(folio, pfn - folio_pfn(folio));
1727		address = pvmw.address;
1728		anon_exclusive = folio_test_anon(folio) &&
1729				 PageAnonExclusive(subpage);
1730
1731		if (folio_test_hugetlb(folio)) {
1732			bool anon = folio_test_anon(folio);
1733
1734			/*
1735			 * The try_to_unmap() is only passed a hugetlb page
1736			 * in the case where the hugetlb page is poisoned.
1737			 */
1738			VM_BUG_ON_PAGE(!PageHWPoison(subpage), subpage);
1739			/*
1740			 * huge_pmd_unshare may unmap an entire PMD page.
1741			 * There is no way of knowing exactly which PMDs may
1742			 * be cached for this mm, so we must flush them all.
1743			 * start/end were already adjusted above to cover this
1744			 * range.
1745			 */
1746			flush_cache_range(vma, range.start, range.end);
1747
1748			/*
1749			 * To call huge_pmd_unshare, i_mmap_rwsem must be
1750			 * held in write mode.  Caller needs to explicitly
1751			 * do this outside rmap routines.
1752			 *
1753			 * We also must hold hugetlb vma_lock in write mode.
1754			 * Lock order dictates acquiring vma_lock BEFORE
1755			 * i_mmap_rwsem.  We can only try lock here and fail
1756			 * if unsuccessful.
1757			 */
1758			if (!anon) {
1759				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
1760				if (!hugetlb_vma_trylock_write(vma))
1761					goto walk_abort;
 
 
 
1762				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
1763					hugetlb_vma_unlock_write(vma);
1764					flush_tlb_range(vma,
1765						range.start, range.end);
1766					/*
1767					 * The ref count of the PMD page was
1768					 * dropped which is part of the way map
1769					 * counting is done for shared PMDs.
1770					 * Return 'true' here.  When there is
1771					 * no other sharing, huge_pmd_unshare
1772					 * returns false and we will unmap the
1773					 * actual page and drop map count
1774					 * to zero.
1775					 */
1776					goto walk_done;
 
1777				}
1778				hugetlb_vma_unlock_write(vma);
1779			}
1780			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
1781		} else {
1782			flush_cache_page(vma, address, pfn);
1783			/* Nuke the page table entry. */
1784			if (should_defer_flush(mm, flags)) {
1785				/*
1786				 * We clear the PTE but do not flush so potentially
1787				 * a remote CPU could still be writing to the folio.
1788				 * If the entry was previously clean then the
1789				 * architecture must guarantee that a clear->dirty
1790				 * transition on a cached TLB entry is written through
1791				 * and traps if the PTE is unmapped.
1792				 */
1793				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1794
1795				set_tlb_ubc_flush_pending(mm, pteval, address);
1796			} else {
1797				pteval = ptep_clear_flush(vma, address, pvmw.pte);
1798			}
1799		}
1800
1801		/*
1802		 * Now the pte is cleared. If this pte was uffd-wp armed,
1803		 * we may want to replace a none pte with a marker pte if
1804		 * it's file-backed, so we don't lose the tracking info.
1805		 */
1806		pte_install_uffd_wp_if_needed(vma, address, pvmw.pte, pteval);
1807
1808		/* Set the dirty flag on the folio now the pte is gone. */
1809		if (pte_dirty(pteval))
1810			folio_mark_dirty(folio);
1811
1812		/* Update high watermark before we lower rss */
1813		update_hiwater_rss(mm);
1814
1815		if (PageHWPoison(subpage) && (flags & TTU_HWPOISON)) {
1816			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1817			if (folio_test_hugetlb(folio)) {
1818				hugetlb_count_sub(folio_nr_pages(folio), mm);
1819				set_huge_pte_at(mm, address, pvmw.pte, pteval,
1820						hsz);
1821			} else {
1822				dec_mm_counter(mm, mm_counter(folio));
1823				set_pte_at(mm, address, pvmw.pte, pteval);
1824			}
1825
1826		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1827			/*
1828			 * The guest indicated that the page content is of no
1829			 * interest anymore. Simply discard the pte, vmscan
1830			 * will take care of the rest.
1831			 * A future reference will then fault in a new zero
1832			 * page. When userfaultfd is active, we must not drop
1833			 * this page though, as its main user (postcopy
1834			 * migration) will not expect userfaults on already
1835			 * copied pages.
1836			 */
1837			dec_mm_counter(mm, mm_counter(folio));
1838		} else if (folio_test_anon(folio)) {
1839			swp_entry_t entry = page_swap_entry(subpage);
1840			pte_t swp_pte;
1841			/*
1842			 * Store the swap location in the pte.
1843			 * See handle_pte_fault() ...
1844			 */
1845			if (unlikely(folio_test_swapbacked(folio) !=
1846					folio_test_swapcache(folio))) {
1847				WARN_ON_ONCE(1);
1848				goto walk_abort;
 
 
1849			}
1850
1851			/* MADV_FREE page check */
1852			if (!folio_test_swapbacked(folio)) {
1853				int ref_count, map_count;
1854
1855				/*
1856				 * Synchronize with gup_pte_range():
1857				 * - clear PTE; barrier; read refcount
1858				 * - inc refcount; barrier; read PTE
1859				 */
1860				smp_mb();
1861
1862				ref_count = folio_ref_count(folio);
1863				map_count = folio_mapcount(folio);
1864
1865				/*
1866				 * Order reads for page refcount and dirty flag
1867				 * (see comments in __remove_mapping()).
1868				 */
1869				smp_rmb();
1870
1871				/*
1872				 * The only page refs must be one from isolation
1873				 * plus the rmap(s) (dropped by discard:).
1874				 */
1875				if (ref_count == 1 + map_count &&
1876				    (!folio_test_dirty(folio) ||
1877				     /*
1878				      * Unlike MADV_FREE mappings, VM_DROPPABLE
1879				      * ones can be dropped even if they've
1880				      * been dirtied.
1881				      */
1882				     (vma->vm_flags & VM_DROPPABLE))) {
1883					dec_mm_counter(mm, MM_ANONPAGES);
1884					goto discard;
1885				}
1886
1887				/*
1888				 * If the folio was redirtied, it cannot be
1889				 * discarded. Remap the page to page table.
1890				 */
1891				set_pte_at(mm, address, pvmw.pte, pteval);
1892				/*
1893				 * Unlike MADV_FREE mappings, VM_DROPPABLE ones
1894				 * never get swap backed on failure to drop.
1895				 */
1896				if (!(vma->vm_flags & VM_DROPPABLE))
1897					folio_set_swapbacked(folio);
1898				goto walk_abort;
1899			}
1900
1901			if (swap_duplicate(entry) < 0) {
1902				set_pte_at(mm, address, pvmw.pte, pteval);
1903				goto walk_abort;
 
 
1904			}
1905			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1906				swap_free(entry);
1907				set_pte_at(mm, address, pvmw.pte, pteval);
1908				goto walk_abort;
 
 
1909			}
1910
1911			/* See folio_try_share_anon_rmap(): clear PTE first. */
1912			if (anon_exclusive &&
1913			    folio_try_share_anon_rmap_pte(folio, subpage)) {
1914				swap_free(entry);
1915				set_pte_at(mm, address, pvmw.pte, pteval);
1916				goto walk_abort;
 
 
1917			}
1918			if (list_empty(&mm->mmlist)) {
1919				spin_lock(&mmlist_lock);
1920				if (list_empty(&mm->mmlist))
1921					list_add(&mm->mmlist, &init_mm.mmlist);
1922				spin_unlock(&mmlist_lock);
1923			}
1924			dec_mm_counter(mm, MM_ANONPAGES);
1925			inc_mm_counter(mm, MM_SWAPENTS);
1926			swp_pte = swp_entry_to_pte(entry);
1927			if (anon_exclusive)
1928				swp_pte = pte_swp_mkexclusive(swp_pte);
1929			if (pte_soft_dirty(pteval))
1930				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1931			if (pte_uffd_wp(pteval))
1932				swp_pte = pte_swp_mkuffd_wp(swp_pte);
1933			set_pte_at(mm, address, pvmw.pte, swp_pte);
1934		} else {
1935			/*
1936			 * This is a locked file-backed folio,
1937			 * so it cannot be removed from the page
1938			 * cache and replaced by a new folio before
1939			 * mmu_notifier_invalidate_range_end, so no
1940			 * concurrent thread might update its page table
1941			 * to point at a new folio while a device is
1942			 * still using this folio.
1943			 *
1944			 * See Documentation/mm/mmu_notifier.rst
1945			 */
1946			dec_mm_counter(mm, mm_counter_file(folio));
1947		}
1948discard:
1949		if (unlikely(folio_test_hugetlb(folio)))
1950			hugetlb_remove_rmap(folio);
1951		else
1952			folio_remove_rmap_pte(folio, subpage, vma);
1953		if (vma->vm_flags & VM_LOCKED)
1954			mlock_drain_local();
1955		folio_put(folio);
1956		continue;
1957walk_abort:
1958		ret = false;
1959walk_done:
1960		page_vma_mapped_walk_done(&pvmw);
1961		break;
1962	}
1963
1964	mmu_notifier_invalidate_range_end(&range);
1965
1966	return ret;
1967}
1968
1969static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1970{
1971	return vma_is_temporary_stack(vma);
1972}
1973
1974static int folio_not_mapped(struct folio *folio)
1975{
1976	return !folio_mapped(folio);
1977}
1978
1979/**
1980 * try_to_unmap - Try to remove all page table mappings to a folio.
1981 * @folio: The folio to unmap.
1982 * @flags: action and flags
1983 *
1984 * Tries to remove all the page table entries which are mapping this
1985 * folio.  It is the caller's responsibility to check if the folio is
1986 * still mapped if needed (use TTU_SYNC to prevent accounting races).
1987 *
1988 * Context: Caller must hold the folio lock.
1989 */
1990void try_to_unmap(struct folio *folio, enum ttu_flags flags)
1991{
1992	struct rmap_walk_control rwc = {
1993		.rmap_one = try_to_unmap_one,
1994		.arg = (void *)flags,
1995		.done = folio_not_mapped,
1996		.anon_lock = folio_lock_anon_vma_read,
1997	};
1998
1999	if (flags & TTU_RMAP_LOCKED)
2000		rmap_walk_locked(folio, &rwc);
2001	else
2002		rmap_walk(folio, &rwc);
2003}
2004
2005/*
2006 * @arg: enum ttu_flags will be passed to this argument.
2007 *
2008 * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
2009 * containing migration entries.
2010 */
2011static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
2012		     unsigned long address, void *arg)
2013{
2014	struct mm_struct *mm = vma->vm_mm;
2015	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2016	pte_t pteval;
2017	struct page *subpage;
2018	bool anon_exclusive, ret = true;
2019	struct mmu_notifier_range range;
2020	enum ttu_flags flags = (enum ttu_flags)(long)arg;
2021	unsigned long pfn;
2022	unsigned long hsz = 0;
2023
2024	/*
2025	 * When racing against e.g. zap_pte_range() on another cpu,
2026	 * in between its ptep_get_and_clear_full() and folio_remove_rmap_*(),
2027	 * try_to_migrate() may return before page_mapped() has become false,
2028	 * if page table locking is skipped: use TTU_SYNC to wait for that.
2029	 */
2030	if (flags & TTU_SYNC)
2031		pvmw.flags = PVMW_SYNC;
2032
2033	/*
2034	 * unmap_page() in mm/huge_memory.c is the only user of migration with
2035	 * TTU_SPLIT_HUGE_PMD and it wants to freeze.
2036	 */
2037	if (flags & TTU_SPLIT_HUGE_PMD)
2038		split_huge_pmd_address(vma, address, true, folio);
2039
2040	/*
2041	 * For THP, we have to assume the worse case ie pmd for invalidation.
2042	 * For hugetlb, it could be much worse if we need to do pud
2043	 * invalidation in the case of pmd sharing.
2044	 *
2045	 * Note that the page can not be free in this function as call of
2046	 * try_to_unmap() must hold a reference on the page.
2047	 */
2048	range.end = vma_address_end(&pvmw);
2049	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2050				address, range.end);
2051	if (folio_test_hugetlb(folio)) {
2052		/*
2053		 * If sharing is possible, start and end will be adjusted
2054		 * accordingly.
2055		 */
2056		adjust_range_if_pmd_sharing_possible(vma, &range.start,
2057						     &range.end);
2058
2059		/* We need the huge page size for set_huge_pte_at() */
2060		hsz = huge_page_size(hstate_vma(vma));
2061	}
2062	mmu_notifier_invalidate_range_start(&range);
2063
2064	while (page_vma_mapped_walk(&pvmw)) {
2065#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2066		/* PMD-mapped THP migration entry */
2067		if (!pvmw.pte) {
2068			subpage = folio_page(folio,
2069				pmd_pfn(*pvmw.pmd) - folio_pfn(folio));
2070			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
2071					!folio_test_pmd_mappable(folio), folio);
2072
2073			if (set_pmd_migration_entry(&pvmw, subpage)) {
2074				ret = false;
2075				page_vma_mapped_walk_done(&pvmw);
2076				break;
2077			}
2078			continue;
2079		}
2080#endif
2081
2082		/* Unexpected PMD-mapped THP? */
2083		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2084
2085		pfn = pte_pfn(ptep_get(pvmw.pte));
2086
2087		if (folio_is_zone_device(folio)) {
2088			/*
2089			 * Our PTE is a non-present device exclusive entry and
2090			 * calculating the subpage as for the common case would
2091			 * result in an invalid pointer.
2092			 *
2093			 * Since only PAGE_SIZE pages can currently be
2094			 * migrated, just set it to page. This will need to be
2095			 * changed when hugepage migrations to device private
2096			 * memory are supported.
2097			 */
2098			VM_BUG_ON_FOLIO(folio_nr_pages(folio) > 1, folio);
2099			subpage = &folio->page;
2100		} else {
2101			subpage = folio_page(folio, pfn - folio_pfn(folio));
2102		}
2103		address = pvmw.address;
2104		anon_exclusive = folio_test_anon(folio) &&
2105				 PageAnonExclusive(subpage);
2106
2107		if (folio_test_hugetlb(folio)) {
2108			bool anon = folio_test_anon(folio);
2109
2110			/*
2111			 * huge_pmd_unshare may unmap an entire PMD page.
2112			 * There is no way of knowing exactly which PMDs may
2113			 * be cached for this mm, so we must flush them all.
2114			 * start/end were already adjusted above to cover this
2115			 * range.
2116			 */
2117			flush_cache_range(vma, range.start, range.end);
2118
2119			/*
2120			 * To call huge_pmd_unshare, i_mmap_rwsem must be
2121			 * held in write mode.  Caller needs to explicitly
2122			 * do this outside rmap routines.
2123			 *
2124			 * We also must hold hugetlb vma_lock in write mode.
2125			 * Lock order dictates acquiring vma_lock BEFORE
2126			 * i_mmap_rwsem.  We can only try lock here and
2127			 * fail if unsuccessful.
2128			 */
2129			if (!anon) {
2130				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
2131				if (!hugetlb_vma_trylock_write(vma)) {
2132					page_vma_mapped_walk_done(&pvmw);
2133					ret = false;
2134					break;
2135				}
2136				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
2137					hugetlb_vma_unlock_write(vma);
2138					flush_tlb_range(vma,
2139						range.start, range.end);
2140
2141					/*
2142					 * The ref count of the PMD page was
2143					 * dropped which is part of the way map
2144					 * counting is done for shared PMDs.
2145					 * Return 'true' here.  When there is
2146					 * no other sharing, huge_pmd_unshare
2147					 * returns false and we will unmap the
2148					 * actual page and drop map count
2149					 * to zero.
2150					 */
2151					page_vma_mapped_walk_done(&pvmw);
2152					break;
2153				}
2154				hugetlb_vma_unlock_write(vma);
2155			}
2156			/* Nuke the hugetlb page table entry */
2157			pteval = huge_ptep_clear_flush(vma, address, pvmw.pte);
2158		} else {
2159			flush_cache_page(vma, address, pfn);
2160			/* Nuke the page table entry. */
2161			if (should_defer_flush(mm, flags)) {
2162				/*
2163				 * We clear the PTE but do not flush so potentially
2164				 * a remote CPU could still be writing to the folio.
2165				 * If the entry was previously clean then the
2166				 * architecture must guarantee that a clear->dirty
2167				 * transition on a cached TLB entry is written through
2168				 * and traps if the PTE is unmapped.
2169				 */
2170				pteval = ptep_get_and_clear(mm, address, pvmw.pte);
2171
2172				set_tlb_ubc_flush_pending(mm, pteval, address);
2173			} else {
2174				pteval = ptep_clear_flush(vma, address, pvmw.pte);
2175			}
2176		}
2177
2178		/* Set the dirty flag on the folio now the pte is gone. */
2179		if (pte_dirty(pteval))
2180			folio_mark_dirty(folio);
2181
2182		/* Update high watermark before we lower rss */
2183		update_hiwater_rss(mm);
2184
2185		if (folio_is_device_private(folio)) {
2186			unsigned long pfn = folio_pfn(folio);
2187			swp_entry_t entry;
2188			pte_t swp_pte;
2189
2190			if (anon_exclusive)
2191				WARN_ON_ONCE(folio_try_share_anon_rmap_pte(folio,
2192									   subpage));
2193
2194			/*
2195			 * Store the pfn of the page in a special migration
2196			 * pte. do_swap_page() will wait until the migration
2197			 * pte is removed and then restart fault handling.
2198			 */
2199			entry = pte_to_swp_entry(pteval);
2200			if (is_writable_device_private_entry(entry))
2201				entry = make_writable_migration_entry(pfn);
2202			else if (anon_exclusive)
2203				entry = make_readable_exclusive_migration_entry(pfn);
2204			else
2205				entry = make_readable_migration_entry(pfn);
2206			swp_pte = swp_entry_to_pte(entry);
2207
2208			/*
2209			 * pteval maps a zone device page and is therefore
2210			 * a swap pte.
2211			 */
2212			if (pte_swp_soft_dirty(pteval))
2213				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2214			if (pte_swp_uffd_wp(pteval))
2215				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2216			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
2217			trace_set_migration_pte(pvmw.address, pte_val(swp_pte),
2218						folio_order(folio));
2219			/*
2220			 * No need to invalidate here it will synchronize on
2221			 * against the special swap migration pte.
2222			 */
2223		} else if (PageHWPoison(subpage)) {
2224			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
2225			if (folio_test_hugetlb(folio)) {
2226				hugetlb_count_sub(folio_nr_pages(folio), mm);
2227				set_huge_pte_at(mm, address, pvmw.pte, pteval,
2228						hsz);
2229			} else {
2230				dec_mm_counter(mm, mm_counter(folio));
2231				set_pte_at(mm, address, pvmw.pte, pteval);
2232			}
2233
2234		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
2235			/*
2236			 * The guest indicated that the page content is of no
2237			 * interest anymore. Simply discard the pte, vmscan
2238			 * will take care of the rest.
2239			 * A future reference will then fault in a new zero
2240			 * page. When userfaultfd is active, we must not drop
2241			 * this page though, as its main user (postcopy
2242			 * migration) will not expect userfaults on already
2243			 * copied pages.
2244			 */
2245			dec_mm_counter(mm, mm_counter(folio));
2246		} else {
2247			swp_entry_t entry;
2248			pte_t swp_pte;
2249
2250			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
2251				if (folio_test_hugetlb(folio))
2252					set_huge_pte_at(mm, address, pvmw.pte,
2253							pteval, hsz);
2254				else
2255					set_pte_at(mm, address, pvmw.pte, pteval);
2256				ret = false;
2257				page_vma_mapped_walk_done(&pvmw);
2258				break;
2259			}
2260			VM_BUG_ON_PAGE(pte_write(pteval) && folio_test_anon(folio) &&
2261				       !anon_exclusive, subpage);
2262
2263			/* See folio_try_share_anon_rmap_pte(): clear PTE first. */
2264			if (folio_test_hugetlb(folio)) {
2265				if (anon_exclusive &&
2266				    hugetlb_try_share_anon_rmap(folio)) {
2267					set_huge_pte_at(mm, address, pvmw.pte,
2268							pteval, hsz);
2269					ret = false;
2270					page_vma_mapped_walk_done(&pvmw);
2271					break;
2272				}
2273			} else if (anon_exclusive &&
2274				   folio_try_share_anon_rmap_pte(folio, subpage)) {
2275				set_pte_at(mm, address, pvmw.pte, pteval);
2276				ret = false;
2277				page_vma_mapped_walk_done(&pvmw);
2278				break;
2279			}
2280
2281			/*
2282			 * Store the pfn of the page in a special migration
2283			 * pte. do_swap_page() will wait until the migration
2284			 * pte is removed and then restart fault handling.
2285			 */
2286			if (pte_write(pteval))
2287				entry = make_writable_migration_entry(
2288							page_to_pfn(subpage));
2289			else if (anon_exclusive)
2290				entry = make_readable_exclusive_migration_entry(
2291							page_to_pfn(subpage));
2292			else
2293				entry = make_readable_migration_entry(
2294							page_to_pfn(subpage));
2295			if (pte_young(pteval))
2296				entry = make_migration_entry_young(entry);
2297			if (pte_dirty(pteval))
2298				entry = make_migration_entry_dirty(entry);
2299			swp_pte = swp_entry_to_pte(entry);
2300			if (pte_soft_dirty(pteval))
2301				swp_pte = pte_swp_mksoft_dirty(swp_pte);
2302			if (pte_uffd_wp(pteval))
2303				swp_pte = pte_swp_mkuffd_wp(swp_pte);
2304			if (folio_test_hugetlb(folio))
2305				set_huge_pte_at(mm, address, pvmw.pte, swp_pte,
2306						hsz);
2307			else
2308				set_pte_at(mm, address, pvmw.pte, swp_pte);
2309			trace_set_migration_pte(address, pte_val(swp_pte),
2310						folio_order(folio));
2311			/*
2312			 * No need to invalidate here it will synchronize on
2313			 * against the special swap migration pte.
2314			 */
2315		}
2316
2317		if (unlikely(folio_test_hugetlb(folio)))
2318			hugetlb_remove_rmap(folio);
2319		else
2320			folio_remove_rmap_pte(folio, subpage, vma);
2321		if (vma->vm_flags & VM_LOCKED)
2322			mlock_drain_local();
2323		folio_put(folio);
2324	}
2325
2326	mmu_notifier_invalidate_range_end(&range);
2327
2328	return ret;
2329}
2330
2331/**
2332 * try_to_migrate - try to replace all page table mappings with swap entries
2333 * @folio: the folio to replace page table entries for
2334 * @flags: action and flags
2335 *
2336 * Tries to remove all the page table entries which are mapping this folio and
2337 * replace them with special swap entries. Caller must hold the folio lock.
2338 */
2339void try_to_migrate(struct folio *folio, enum ttu_flags flags)
2340{
2341	struct rmap_walk_control rwc = {
2342		.rmap_one = try_to_migrate_one,
2343		.arg = (void *)flags,
2344		.done = folio_not_mapped,
2345		.anon_lock = folio_lock_anon_vma_read,
2346	};
2347
2348	/*
2349	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
2350	 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
2351	 */
2352	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2353					TTU_SYNC | TTU_BATCH_FLUSH)))
2354		return;
2355
2356	if (folio_is_zone_device(folio) &&
2357	    (!folio_is_device_private(folio) && !folio_is_device_coherent(folio)))
2358		return;
2359
2360	/*
2361	 * During exec, a temporary VMA is setup and later moved.
2362	 * The VMA is moved under the anon_vma lock but not the
2363	 * page tables leading to a race where migration cannot
2364	 * find the migration ptes. Rather than increasing the
2365	 * locking requirements of exec(), migration skips
2366	 * temporary VMAs until after exec() completes.
2367	 */
2368	if (!folio_test_ksm(folio) && folio_test_anon(folio))
2369		rwc.invalid_vma = invalid_migration_vma;
2370
2371	if (flags & TTU_RMAP_LOCKED)
2372		rmap_walk_locked(folio, &rwc);
2373	else
2374		rmap_walk(folio, &rwc);
2375}
2376
2377#ifdef CONFIG_DEVICE_PRIVATE
2378struct make_exclusive_args {
2379	struct mm_struct *mm;
2380	unsigned long address;
2381	void *owner;
2382	bool valid;
2383};
2384
2385static bool page_make_device_exclusive_one(struct folio *folio,
2386		struct vm_area_struct *vma, unsigned long address, void *priv)
2387{
2388	struct mm_struct *mm = vma->vm_mm;
2389	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, 0);
2390	struct make_exclusive_args *args = priv;
2391	pte_t pteval;
2392	struct page *subpage;
2393	bool ret = true;
2394	struct mmu_notifier_range range;
2395	swp_entry_t entry;
2396	pte_t swp_pte;
2397	pte_t ptent;
2398
2399	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0,
2400				      vma->vm_mm, address, min(vma->vm_end,
2401				      address + folio_size(folio)),
2402				      args->owner);
2403	mmu_notifier_invalidate_range_start(&range);
2404
2405	while (page_vma_mapped_walk(&pvmw)) {
2406		/* Unexpected PMD-mapped THP? */
2407		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
2408
2409		ptent = ptep_get(pvmw.pte);
2410		if (!pte_present(ptent)) {
2411			ret = false;
2412			page_vma_mapped_walk_done(&pvmw);
2413			break;
2414		}
2415
2416		subpage = folio_page(folio,
2417				pte_pfn(ptent) - folio_pfn(folio));
2418		address = pvmw.address;
2419
2420		/* Nuke the page table entry. */
2421		flush_cache_page(vma, address, pte_pfn(ptent));
2422		pteval = ptep_clear_flush(vma, address, pvmw.pte);
2423
2424		/* Set the dirty flag on the folio now the pte is gone. */
2425		if (pte_dirty(pteval))
2426			folio_mark_dirty(folio);
2427
2428		/*
2429		 * Check that our target page is still mapped at the expected
2430		 * address.
2431		 */
2432		if (args->mm == mm && args->address == address &&
2433		    pte_write(pteval))
2434			args->valid = true;
2435
2436		/*
2437		 * Store the pfn of the page in a special migration
2438		 * pte. do_swap_page() will wait until the migration
2439		 * pte is removed and then restart fault handling.
2440		 */
2441		if (pte_write(pteval))
2442			entry = make_writable_device_exclusive_entry(
2443							page_to_pfn(subpage));
2444		else
2445			entry = make_readable_device_exclusive_entry(
2446							page_to_pfn(subpage));
2447		swp_pte = swp_entry_to_pte(entry);
2448		if (pte_soft_dirty(pteval))
2449			swp_pte = pte_swp_mksoft_dirty(swp_pte);
2450		if (pte_uffd_wp(pteval))
2451			swp_pte = pte_swp_mkuffd_wp(swp_pte);
2452
2453		set_pte_at(mm, address, pvmw.pte, swp_pte);
2454
2455		/*
2456		 * There is a reference on the page for the swap entry which has
2457		 * been removed, so shouldn't take another.
2458		 */
2459		folio_remove_rmap_pte(folio, subpage, vma);
2460	}
2461
2462	mmu_notifier_invalidate_range_end(&range);
2463
2464	return ret;
2465}
2466
2467/**
2468 * folio_make_device_exclusive - Mark the folio exclusively owned by a device.
2469 * @folio: The folio to replace page table entries for.
2470 * @mm: The mm_struct where the folio is expected to be mapped.
2471 * @address: Address where the folio is expected to be mapped.
2472 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
2473 *
2474 * Tries to remove all the page table entries which are mapping this
2475 * folio and replace them with special device exclusive swap entries to
2476 * grant a device exclusive access to the folio.
2477 *
2478 * Context: Caller must hold the folio lock.
2479 * Return: false if the page is still mapped, or if it could not be unmapped
2480 * from the expected address. Otherwise returns true (success).
2481 */
2482static bool folio_make_device_exclusive(struct folio *folio,
2483		struct mm_struct *mm, unsigned long address, void *owner)
2484{
2485	struct make_exclusive_args args = {
2486		.mm = mm,
2487		.address = address,
2488		.owner = owner,
2489		.valid = false,
2490	};
2491	struct rmap_walk_control rwc = {
2492		.rmap_one = page_make_device_exclusive_one,
2493		.done = folio_not_mapped,
2494		.anon_lock = folio_lock_anon_vma_read,
2495		.arg = &args,
2496	};
2497
2498	/*
2499	 * Restrict to anonymous folios for now to avoid potential writeback
2500	 * issues.
2501	 */
2502	if (!folio_test_anon(folio))
2503		return false;
2504
2505	rmap_walk(folio, &rwc);
2506
2507	return args.valid && !folio_mapcount(folio);
2508}
2509
2510/**
2511 * make_device_exclusive_range() - Mark a range for exclusive use by a device
2512 * @mm: mm_struct of associated target process
2513 * @start: start of the region to mark for exclusive device access
2514 * @end: end address of region
2515 * @pages: returns the pages which were successfully marked for exclusive access
2516 * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
2517 *
2518 * Returns: number of pages found in the range by GUP. A page is marked for
2519 * exclusive access only if the page pointer is non-NULL.
2520 *
2521 * This function finds ptes mapping page(s) to the given address range, locks
2522 * them and replaces mappings with special swap entries preventing userspace CPU
2523 * access. On fault these entries are replaced with the original mapping after
2524 * calling MMU notifiers.
2525 *
2526 * A driver using this to program access from a device must use a mmu notifier
2527 * critical section to hold a device specific lock during programming. Once
2528 * programming is complete it should drop the page lock and reference after
2529 * which point CPU access to the page will revoke the exclusive access.
2530 */
2531int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
2532				unsigned long end, struct page **pages,
2533				void *owner)
2534{
2535	long npages = (end - start) >> PAGE_SHIFT;
2536	long i;
2537
2538	npages = get_user_pages_remote(mm, start, npages,
2539				       FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
2540				       pages, NULL);
2541	if (npages < 0)
2542		return npages;
2543
2544	for (i = 0; i < npages; i++, start += PAGE_SIZE) {
2545		struct folio *folio = page_folio(pages[i]);
2546		if (PageTail(pages[i]) || !folio_trylock(folio)) {
2547			folio_put(folio);
2548			pages[i] = NULL;
2549			continue;
2550		}
2551
2552		if (!folio_make_device_exclusive(folio, mm, start, owner)) {
2553			folio_unlock(folio);
2554			folio_put(folio);
2555			pages[i] = NULL;
2556		}
2557	}
2558
2559	return npages;
2560}
2561EXPORT_SYMBOL_GPL(make_device_exclusive_range);
2562#endif
2563
2564void __put_anon_vma(struct anon_vma *anon_vma)
2565{
2566	struct anon_vma *root = anon_vma->root;
2567
2568	anon_vma_free(anon_vma);
2569	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
2570		anon_vma_free(root);
2571}
2572
2573static struct anon_vma *rmap_walk_anon_lock(const struct folio *folio,
2574					    struct rmap_walk_control *rwc)
2575{
2576	struct anon_vma *anon_vma;
2577
2578	if (rwc->anon_lock)
2579		return rwc->anon_lock(folio, rwc);
2580
2581	/*
2582	 * Note: remove_migration_ptes() cannot use folio_lock_anon_vma_read()
2583	 * because that depends on page_mapped(); but not all its usages
2584	 * are holding mmap_lock. Users without mmap_lock are required to
2585	 * take a reference count to prevent the anon_vma disappearing
2586	 */
2587	anon_vma = folio_anon_vma(folio);
2588	if (!anon_vma)
2589		return NULL;
2590
2591	if (anon_vma_trylock_read(anon_vma))
2592		goto out;
2593
2594	if (rwc->try_lock) {
2595		anon_vma = NULL;
2596		rwc->contended = true;
2597		goto out;
2598	}
2599
2600	anon_vma_lock_read(anon_vma);
2601out:
2602	return anon_vma;
2603}
2604
2605/*
2606 * rmap_walk_anon - do something to anonymous page using the object-based
2607 * rmap method
2608 * @folio: the folio to be handled
2609 * @rwc: control variable according to each walk type
2610 * @locked: caller holds relevant rmap lock
2611 *
2612 * Find all the mappings of a folio using the mapping pointer and the vma
2613 * chains contained in the anon_vma struct it points to.
2614 */
2615static void rmap_walk_anon(struct folio *folio,
2616		struct rmap_walk_control *rwc, bool locked)
2617{
2618	struct anon_vma *anon_vma;
2619	pgoff_t pgoff_start, pgoff_end;
2620	struct anon_vma_chain *avc;
2621
2622	if (locked) {
2623		anon_vma = folio_anon_vma(folio);
2624		/* anon_vma disappear under us? */
2625		VM_BUG_ON_FOLIO(!anon_vma, folio);
2626	} else {
2627		anon_vma = rmap_walk_anon_lock(folio, rwc);
2628	}
2629	if (!anon_vma)
2630		return;
2631
2632	pgoff_start = folio_pgoff(folio);
2633	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2634	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
2635			pgoff_start, pgoff_end) {
2636		struct vm_area_struct *vma = avc->vma;
2637		unsigned long address = vma_address(vma, pgoff_start,
2638				folio_nr_pages(folio));
2639
2640		VM_BUG_ON_VMA(address == -EFAULT, vma);
2641		cond_resched();
2642
2643		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2644			continue;
2645
2646		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2647			break;
2648		if (rwc->done && rwc->done(folio))
2649			break;
2650	}
2651
2652	if (!locked)
2653		anon_vma_unlock_read(anon_vma);
2654}
2655
2656/*
2657 * rmap_walk_file - do something to file page using the object-based rmap method
2658 * @folio: the folio to be handled
2659 * @rwc: control variable according to each walk type
2660 * @locked: caller holds relevant rmap lock
2661 *
2662 * Find all the mappings of a folio using the mapping pointer and the vma chains
2663 * contained in the address_space struct it points to.
2664 */
2665static void rmap_walk_file(struct folio *folio,
2666		struct rmap_walk_control *rwc, bool locked)
2667{
2668	struct address_space *mapping = folio_mapping(folio);
2669	pgoff_t pgoff_start, pgoff_end;
2670	struct vm_area_struct *vma;
2671
2672	/*
2673	 * The page lock not only makes sure that page->mapping cannot
2674	 * suddenly be NULLified by truncation, it makes sure that the
2675	 * structure at mapping cannot be freed and reused yet,
2676	 * so we can safely take mapping->i_mmap_rwsem.
2677	 */
2678	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2679
2680	if (!mapping)
2681		return;
2682
2683	pgoff_start = folio_pgoff(folio);
2684	pgoff_end = pgoff_start + folio_nr_pages(folio) - 1;
2685	if (!locked) {
2686		if (i_mmap_trylock_read(mapping))
2687			goto lookup;
2688
2689		if (rwc->try_lock) {
2690			rwc->contended = true;
2691			return;
2692		}
2693
2694		i_mmap_lock_read(mapping);
2695	}
2696lookup:
2697	vma_interval_tree_foreach(vma, &mapping->i_mmap,
2698			pgoff_start, pgoff_end) {
2699		unsigned long address = vma_address(vma, pgoff_start,
2700			       folio_nr_pages(folio));
2701
2702		VM_BUG_ON_VMA(address == -EFAULT, vma);
2703		cond_resched();
2704
2705		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
2706			continue;
2707
2708		if (!rwc->rmap_one(folio, vma, address, rwc->arg))
2709			goto done;
2710		if (rwc->done && rwc->done(folio))
2711			goto done;
2712	}
2713
2714done:
2715	if (!locked)
2716		i_mmap_unlock_read(mapping);
2717}
2718
2719void rmap_walk(struct folio *folio, struct rmap_walk_control *rwc)
2720{
2721	if (unlikely(folio_test_ksm(folio)))
2722		rmap_walk_ksm(folio, rwc);
2723	else if (folio_test_anon(folio))
2724		rmap_walk_anon(folio, rwc, false);
2725	else
2726		rmap_walk_file(folio, rwc, false);
2727}
2728
2729/* Like rmap_walk, but caller holds relevant rmap lock */
2730void rmap_walk_locked(struct folio *folio, struct rmap_walk_control *rwc)
2731{
2732	/* no ksm support for now */
2733	VM_BUG_ON_FOLIO(folio_test_ksm(folio), folio);
2734	if (folio_test_anon(folio))
2735		rmap_walk_anon(folio, rwc, true);
2736	else
2737		rmap_walk_file(folio, rwc, true);
2738}
2739
2740#ifdef CONFIG_HUGETLB_PAGE
2741/*
2742 * The following two functions are for anonymous (private mapped) hugepages.
2743 * Unlike common anonymous pages, anonymous hugepages have no accounting code
2744 * and no lru code, because we handle hugepages differently from common pages.
2745 */
2746void hugetlb_add_anon_rmap(struct folio *folio, struct vm_area_struct *vma,
2747		unsigned long address, rmap_t flags)
2748{
2749	VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2750	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2751
2752	atomic_inc(&folio->_entire_mapcount);
2753	atomic_inc(&folio->_large_mapcount);
2754	if (flags & RMAP_EXCLUSIVE)
2755		SetPageAnonExclusive(&folio->page);
2756	VM_WARN_ON_FOLIO(folio_entire_mapcount(folio) > 1 &&
2757			 PageAnonExclusive(&folio->page), folio);
2758}
2759
2760void hugetlb_add_new_anon_rmap(struct folio *folio,
2761		struct vm_area_struct *vma, unsigned long address)
2762{
2763	VM_WARN_ON_FOLIO(!folio_test_hugetlb(folio), folio);
2764
2765	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2766	/* increment count (starts at -1) */
2767	atomic_set(&folio->_entire_mapcount, 0);
2768	atomic_set(&folio->_large_mapcount, 0);
2769	folio_clear_hugetlb_restore_reserve(folio);
2770	__folio_set_anon(folio, vma, address, true);
2771	SetPageAnonExclusive(&folio->page);
2772}
2773#endif /* CONFIG_HUGETLB_PAGE */